diff options
Diffstat (limited to 'compiler/rustc_codegen_gcc/src')
24 files changed, 14553 insertions, 0 deletions
diff --git a/compiler/rustc_codegen_gcc/src/abi.rs b/compiler/rustc_codegen_gcc/src/abi.rs new file mode 100644 index 000000000..0ed3e1fbe --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/abi.rs @@ -0,0 +1,179 @@ +use gccjit::{ToLValue, ToRValue, Type}; +use rustc_codegen_ssa::traits::{AbiBuilderMethods, BaseTypeMethods}; +use rustc_data_structures::fx::FxHashSet; +use rustc_middle::bug; +use rustc_middle::ty::Ty; +use rustc_target::abi::call::{CastTarget, FnAbi, PassMode, Reg, RegKind}; + +use crate::builder::Builder; +use crate::context::CodegenCx; +use crate::intrinsic::ArgAbiExt; +use crate::type_of::LayoutGccExt; + +impl<'a, 'gcc, 'tcx> AbiBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { + fn apply_attrs_callsite(&mut self, _fn_abi: &FnAbi<'tcx, Ty<'tcx>>, _callsite: Self::Value) { + // TODO(antoyo) + } + + fn get_param(&mut self, index: usize) -> Self::Value { + let func = self.current_func(); + let param = func.get_param(index as i32); + let on_stack = + if let Some(on_stack_param_indices) = self.on_stack_function_params.borrow().get(&func) { + on_stack_param_indices.contains(&index) + } + else { + false + }; + if on_stack { + param.to_lvalue().get_address(None) + } + else { + param.to_rvalue() + } + } +} + +impl GccType for CastTarget { + fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, '_>) -> Type<'gcc> { + let rest_gcc_unit = self.rest.unit.gcc_type(cx); + let (rest_count, rem_bytes) = + if self.rest.unit.size.bytes() == 0 { + (0, 0) + } + else { + (self.rest.total.bytes() / self.rest.unit.size.bytes(), self.rest.total.bytes() % self.rest.unit.size.bytes()) + }; + + if self.prefix.iter().all(|x| x.is_none()) { + // Simplify to a single unit when there is no prefix and size <= unit size + if self.rest.total <= self.rest.unit.size { + return rest_gcc_unit; + } + + // Simplify to array when all chunks are the same size and type + if rem_bytes == 0 { + return cx.type_array(rest_gcc_unit, rest_count); + } + } + + // Create list of fields in the main structure + let mut args: Vec<_> = self + .prefix + .iter() + .flat_map(|option_reg| { + option_reg.map(|reg| reg.gcc_type(cx)) + }) + .chain((0..rest_count).map(|_| rest_gcc_unit)) + .collect(); + + // Append final integer + if rem_bytes != 0 { + // Only integers can be really split further. + assert_eq!(self.rest.unit.kind, RegKind::Integer); + args.push(cx.type_ix(rem_bytes * 8)); + } + + cx.type_struct(&args, false) + } +} + +pub trait GccType { + fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, '_>) -> Type<'gcc>; +} + +impl GccType for Reg { + fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, '_>) -> Type<'gcc> { + match self.kind { + RegKind::Integer => cx.type_ix(self.size.bits()), + RegKind::Float => { + match self.size.bits() { + 32 => cx.type_f32(), + 64 => cx.type_f64(), + _ => bug!("unsupported float: {:?}", self), + } + }, + RegKind::Vector => unimplemented!(), //cx.type_vector(cx.type_i8(), self.size.bytes()), + } + } +} + +pub trait FnAbiGccExt<'gcc, 'tcx> { + // TODO(antoyo): return a function pointer type instead? + fn gcc_type(&self, cx: &CodegenCx<'gcc, 'tcx>) -> (Type<'gcc>, Vec<Type<'gcc>>, bool, FxHashSet<usize>); + fn ptr_to_gcc_type(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; +} + +impl<'gcc, 'tcx> FnAbiGccExt<'gcc, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { + fn gcc_type(&self, cx: &CodegenCx<'gcc, 'tcx>) -> (Type<'gcc>, Vec<Type<'gcc>>, bool, FxHashSet<usize>) { + let mut on_stack_param_indices = FxHashSet::default(); + let args_capacity: usize = self.args.iter().map(|arg| + if arg.pad.is_some() { + 1 + } + else { + 0 + } + + if let PassMode::Pair(_, _) = arg.mode { + 2 + } else { + 1 + } + ).sum(); + let mut argument_tys = Vec::with_capacity( + if let PassMode::Indirect { .. } = self.ret.mode { + 1 + } + else { + 0 + } + args_capacity, + ); + + let return_ty = + match self.ret.mode { + PassMode::Ignore => cx.type_void(), + PassMode::Direct(_) | PassMode::Pair(..) => self.ret.layout.immediate_gcc_type(cx), + PassMode::Cast(cast) => cast.gcc_type(cx), + PassMode::Indirect { .. } => { + argument_tys.push(cx.type_ptr_to(self.ret.memory_ty(cx))); + cx.type_void() + } + }; + + for arg in &self.args { + // add padding + if let Some(ty) = arg.pad { + argument_tys.push(ty.gcc_type(cx)); + } + + let arg_ty = match arg.mode { + PassMode::Ignore => continue, + PassMode::Direct(_) => arg.layout.immediate_gcc_type(cx), + PassMode::Pair(..) => { + argument_tys.push(arg.layout.scalar_pair_element_gcc_type(cx, 0, true)); + argument_tys.push(arg.layout.scalar_pair_element_gcc_type(cx, 1, true)); + continue; + } + PassMode::Indirect { extra_attrs: Some(_), .. } => { + unimplemented!(); + } + PassMode::Cast(cast) => cast.gcc_type(cx), + PassMode::Indirect { extra_attrs: None, on_stack: true, .. } => { + on_stack_param_indices.insert(argument_tys.len()); + arg.memory_ty(cx) + }, + PassMode::Indirect { extra_attrs: None, on_stack: false, .. } => cx.type_ptr_to(arg.memory_ty(cx)), + }; + argument_tys.push(arg_ty); + } + + (return_ty, argument_tys, self.c_variadic, on_stack_param_indices) + } + + fn ptr_to_gcc_type(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> { + let (return_type, params, variadic, on_stack_param_indices) = self.gcc_type(cx); + let pointer_type = cx.context.new_function_pointer_type(None, return_type, ¶ms, variadic); + cx.on_stack_params.borrow_mut().insert(pointer_type.dyncast_function_ptr_type().expect("function ptr type"), on_stack_param_indices); + pointer_type + } +} diff --git a/compiler/rustc_codegen_gcc/src/allocator.rs b/compiler/rustc_codegen_gcc/src/allocator.rs new file mode 100644 index 000000000..58efb81e8 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/allocator.rs @@ -0,0 +1,123 @@ +use gccjit::{FunctionType, GlobalKind, ToRValue}; +use rustc_ast::expand::allocator::{AllocatorKind, AllocatorTy, ALLOCATOR_METHODS}; +use rustc_middle::bug; +use rustc_middle::ty::TyCtxt; +use rustc_session::config::OomStrategy; +use rustc_span::symbol::sym; + +use crate::GccContext; + +pub(crate) unsafe fn codegen(tcx: TyCtxt<'_>, mods: &mut GccContext, _module_name: &str, kind: AllocatorKind, has_alloc_error_handler: bool) { + let context = &mods.context; + let usize = + match tcx.sess.target.pointer_width { + 16 => context.new_type::<u16>(), + 32 => context.new_type::<u32>(), + 64 => context.new_type::<u64>(), + tws => bug!("Unsupported target word size for int: {}", tws), + }; + let i8 = context.new_type::<i8>(); + let i8p = i8.make_pointer(); + let void = context.new_type::<()>(); + + for method in ALLOCATOR_METHODS { + let mut types = Vec::with_capacity(method.inputs.len()); + for ty in method.inputs.iter() { + match *ty { + AllocatorTy::Layout => { + types.push(usize); + types.push(usize); + } + AllocatorTy::Ptr => types.push(i8p), + AllocatorTy::Usize => types.push(usize), + + AllocatorTy::ResultPtr | AllocatorTy::Unit => panic!("invalid allocator arg"), + } + } + let output = match method.output { + AllocatorTy::ResultPtr => Some(i8p), + AllocatorTy::Unit => None, + + AllocatorTy::Layout | AllocatorTy::Usize | AllocatorTy::Ptr => { + panic!("invalid allocator output") + } + }; + let name = format!("__rust_{}", method.name); + + let args: Vec<_> = types.iter().enumerate() + .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) + .collect(); + let func = context.new_function(None, FunctionType::Exported, output.unwrap_or(void), &args, name, false); + + if tcx.sess.target.options.default_hidden_visibility { + // TODO(antoyo): set visibility. + } + if tcx.sess.must_emit_unwind_tables() { + // TODO(antoyo): emit unwind tables. + } + + let callee = kind.fn_name(method.name); + let args: Vec<_> = types.iter().enumerate() + .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) + .collect(); + let callee = context.new_function(None, FunctionType::Extern, output.unwrap_or(void), &args, callee, false); + // TODO(antoyo): set visibility. + + let block = func.new_block("entry"); + + let args = args + .iter() + .enumerate() + .map(|(i, _)| func.get_param(i as i32).to_rvalue()) + .collect::<Vec<_>>(); + let ret = context.new_call(None, callee, &args); + //llvm::LLVMSetTailCall(ret, True); + if output.is_some() { + block.end_with_return(None, ret); + } + else { + block.end_with_void_return(None); + } + + // TODO(@Commeownist): Check if we need to emit some extra debugging info in certain circumstances + // as described in https://github.com/rust-lang/rust/commit/77a96ed5646f7c3ee8897693decc4626fe380643 + } + + let types = [usize, usize]; + let name = "__rust_alloc_error_handler".to_string(); + let args: Vec<_> = types.iter().enumerate() + .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) + .collect(); + let func = context.new_function(None, FunctionType::Exported, void, &args, name, false); + + let kind = + if has_alloc_error_handler { + AllocatorKind::Global + } + else { + AllocatorKind::Default + }; + let callee = kind.fn_name(sym::oom); + let args: Vec<_> = types.iter().enumerate() + .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) + .collect(); + let callee = context.new_function(None, FunctionType::Extern, void, &args, callee, false); + //llvm::LLVMRustSetVisibility(callee, llvm::Visibility::Hidden); + + let block = func.new_block("entry"); + + let args = args + .iter() + .enumerate() + .map(|(i, _)| func.get_param(i as i32).to_rvalue()) + .collect::<Vec<_>>(); + let _ret = context.new_call(None, callee, &args); + //llvm::LLVMSetTailCall(ret, True); + block.end_with_void_return(None); + + let name = OomStrategy::SYMBOL.to_string(); + let global = context.new_global(None, GlobalKind::Exported, i8, name); + let value = tcx.sess.opts.unstable_opts.oom.should_panic(); + let value = context.new_rvalue_from_int(i8, value as i32); + global.global_set_initializer_rvalue(value); +} diff --git a/compiler/rustc_codegen_gcc/src/archive.rs b/compiler/rustc_codegen_gcc/src/archive.rs new file mode 100644 index 000000000..f863abdcc --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/archive.rs @@ -0,0 +1,189 @@ +use std::fs::File; +use std::path::{Path, PathBuf}; + +use rustc_codegen_ssa::back::archive::{ArchiveBuilder, ArchiveBuilderBuilder}; +use rustc_session::Session; + +use rustc_session::cstore::DllImport; + +struct ArchiveConfig<'a> { + sess: &'a Session, + use_native_ar: bool, + use_gnu_style_archive: bool, +} + +#[derive(Debug)] +enum ArchiveEntry { + FromArchive { + archive_index: usize, + entry_index: usize, + }, + File(PathBuf), +} + +pub struct ArArchiveBuilderBuilder; + +impl ArchiveBuilderBuilder for ArArchiveBuilderBuilder { + fn new_archive_builder<'a>(&self, sess: &'a Session) -> Box<dyn ArchiveBuilder<'a> + 'a> { + let config = ArchiveConfig { + sess, + use_native_ar: false, + // FIXME test for linux and System V derivatives instead + use_gnu_style_archive: sess.target.options.archive_format == "gnu", + }; + + Box::new(ArArchiveBuilder { + config, + src_archives: vec![], + entries: vec![], + }) + } + + fn create_dll_import_lib( + &self, + _sess: &Session, + _lib_name: &str, + _dll_imports: &[DllImport], + _tmpdir: &Path, + ) -> PathBuf { + unimplemented!(); + } +} + +pub struct ArArchiveBuilder<'a> { + config: ArchiveConfig<'a>, + src_archives: Vec<(PathBuf, ar::Archive<File>)>, + // Don't use `HashMap` here, as the order is important. `rust.metadata.bin` must always be at + // the end of an archive for linkers to not get confused. + entries: Vec<(String, ArchiveEntry)>, +} + +impl<'a> ArchiveBuilder<'a> for ArArchiveBuilder<'a> { + fn add_file(&mut self, file: &Path) { + self.entries.push(( + file.file_name().unwrap().to_str().unwrap().to_string(), + ArchiveEntry::File(file.to_owned()), + )); + } + + fn add_archive( + &mut self, + archive_path: &Path, + mut skip: Box<dyn FnMut(&str) -> bool + 'static>, + ) -> std::io::Result<()> { + let mut archive = ar::Archive::new(std::fs::File::open(&archive_path)?); + let archive_index = self.src_archives.len(); + + let mut i = 0; + while let Some(entry) = archive.next_entry() { + let entry = entry?; + let file_name = String::from_utf8(entry.header().identifier().to_vec()) + .map_err(|err| std::io::Error::new(std::io::ErrorKind::InvalidData, err))?; + if !skip(&file_name) { + self.entries + .push((file_name, ArchiveEntry::FromArchive { archive_index, entry_index: i })); + } + i += 1; + } + + self.src_archives.push((archive_path.to_owned(), archive)); + Ok(()) + } + + fn build(mut self: Box<Self>, output: &Path) -> bool { + use std::process::Command; + + fn add_file_using_ar(archive: &Path, file: &Path) { + Command::new("ar") + .arg("r") // add or replace file + .arg("-c") // silence created file message + .arg(archive) + .arg(&file) + .status() + .unwrap(); + } + + enum BuilderKind<'a> { + Bsd(ar::Builder<File>), + Gnu(ar::GnuBuilder<File>), + NativeAr(&'a Path), + } + + let mut builder = if self.config.use_native_ar { + BuilderKind::NativeAr(output) + } else if self.config.use_gnu_style_archive { + BuilderKind::Gnu(ar::GnuBuilder::new( + File::create(output).unwrap(), + self.entries + .iter() + .map(|(name, _)| name.as_bytes().to_vec()) + .collect(), + )) + } else { + BuilderKind::Bsd(ar::Builder::new(File::create(output).unwrap())) + }; + + let any_members = !self.entries.is_empty(); + + // Add all files + for (entry_name, entry) in self.entries.into_iter() { + match entry { + ArchiveEntry::FromArchive { + archive_index, + entry_index, + } => { + let (ref src_archive_path, ref mut src_archive) = + self.src_archives[archive_index]; + let entry = src_archive.jump_to_entry(entry_index).unwrap(); + let header = entry.header().clone(); + + match builder { + BuilderKind::Bsd(ref mut builder) => { + builder.append(&header, entry).unwrap() + } + BuilderKind::Gnu(ref mut builder) => { + builder.append(&header, entry).unwrap() + } + BuilderKind::NativeAr(archive_file) => { + Command::new("ar") + .arg("x") + .arg(src_archive_path) + .arg(&entry_name) + .status() + .unwrap(); + add_file_using_ar(archive_file, Path::new(&entry_name)); + std::fs::remove_file(entry_name).unwrap(); + } + } + } + ArchiveEntry::File(file) => + match builder { + BuilderKind::Bsd(ref mut builder) => { + builder + .append_file(entry_name.as_bytes(), &mut File::open(file).expect("file for bsd builder")) + .unwrap() + }, + BuilderKind::Gnu(ref mut builder) => { + builder + .append_file(entry_name.as_bytes(), &mut File::open(&file).expect(&format!("file {:?} for gnu builder", file))) + .unwrap() + }, + BuilderKind::NativeAr(archive_file) => add_file_using_ar(archive_file, &file), + }, + } + } + + // Finalize archive + std::mem::drop(builder); + + // Run ranlib to be able to link the archive + let status = + std::process::Command::new("ranlib").arg(output).status().expect("Couldn't run ranlib"); + + if !status.success() { + self.config.sess.fatal(&format!("Ranlib exited with code {:?}", status.code())); + } + + any_members + } +} diff --git a/compiler/rustc_codegen_gcc/src/asm.rs b/compiler/rustc_codegen_gcc/src/asm.rs new file mode 100644 index 000000000..52fd66af0 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/asm.rs @@ -0,0 +1,817 @@ +use gccjit::{LValue, RValue, ToRValue, Type}; +use rustc_ast::ast::{InlineAsmOptions, InlineAsmTemplatePiece}; +use rustc_codegen_ssa::mir::operand::OperandValue; +use rustc_codegen_ssa::mir::place::PlaceRef; +use rustc_codegen_ssa::traits::{AsmBuilderMethods, AsmMethods, BaseTypeMethods, BuilderMethods, GlobalAsmOperandRef, InlineAsmOperandRef}; + +use rustc_middle::{bug, ty::Instance}; +use rustc_span::Span; +use rustc_target::asm::*; + +use std::borrow::Cow; + +use crate::builder::Builder; +use crate::context::CodegenCx; +use crate::type_of::LayoutGccExt; +use crate::callee::get_fn; + + +// Rust asm! and GCC Extended Asm semantics differ substantially. +// +// 1. Rust asm operands go along as one list of operands. Operands themselves indicate +// if they're "in" or "out". "In" and "out" operands can interleave. One operand can be +// both "in" and "out" (`inout(reg)`). +// +// GCC asm has two different lists for "in" and "out" operands. In terms of gccjit, +// this means that all "out" operands must go before "in" operands. "In" and "out" operands +// cannot interleave. +// +// 2. Operand lists in both Rust and GCC are indexed. Index starts from 0. Indexes are important +// because the asm template refers to operands by index. +// +// Mapping from Rust to GCC index would be 1-1 if it wasn't for... +// +// 3. Clobbers. GCC has a separate list of clobbers, and clobbers don't have indexes. +// Contrary, Rust expresses clobbers through "out" operands that aren't tied to +// a variable (`_`), and such "clobbers" do have index. +// +// 4. Furthermore, GCC Extended Asm does not support explicit register constraints +// (like `out("eax")`) directly, offering so-called "local register variables" +// as a workaround. These variables need to be declared and initialized *before* +// the Extended Asm block but *after* normal local variables +// (see comment in `codegen_inline_asm` for explanation). +// +// With that in mind, let's see how we translate Rust syntax to GCC +// (from now on, `CC` stands for "constraint code"): +// +// * `out(reg_class) var` -> translated to output operand: `"=CC"(var)` +// * `inout(reg_class) var` -> translated to output operand: `"+CC"(var)` +// * `in(reg_class) var` -> translated to input operand: `"CC"(var)` +// +// * `out(reg_class) _` -> translated to one `=r(tmp)`, where "tmp" is a temporary unused variable +// +// * `out("explicit register") _` -> not translated to any operands, register is simply added to clobbers list +// +// * `inout(reg_class) in_var => out_var` -> translated to two operands: +// output: `"=CC"(in_var)` +// input: `"num"(out_var)` where num is the GCC index +// of the corresponding output operand +// +// * `inout(reg_class) in_var => _` -> same as `inout(reg_class) in_var => tmp`, +// where "tmp" is a temporary unused variable +// +// * `out/in/inout("explicit register") var` -> translated to one or two operands as described above +// with `"r"(var)` constraint, +// and one register variable assigned to the desired register. + +const ATT_SYNTAX_INS: &str = ".att_syntax noprefix\n\t"; +const INTEL_SYNTAX_INS: &str = "\n\t.intel_syntax noprefix"; + + +struct AsmOutOperand<'a, 'tcx, 'gcc> { + rust_idx: usize, + constraint: &'a str, + late: bool, + readwrite: bool, + + tmp_var: LValue<'gcc>, + out_place: Option<PlaceRef<'tcx, RValue<'gcc>>> +} + +struct AsmInOperand<'a, 'tcx> { + rust_idx: usize, + constraint: Cow<'a, str>, + val: RValue<'tcx> +} + +impl AsmOutOperand<'_, '_, '_> { + fn to_constraint(&self) -> String { + let mut res = String::with_capacity(self.constraint.len() + self.late as usize + 1); + + let sign = if self.readwrite { '+' } else { '=' }; + res.push(sign); + if !self.late { + res.push('&'); + } + + res.push_str(&self.constraint); + res + } +} + +enum ConstraintOrRegister { + Constraint(&'static str), + Register(&'static str) +} + + +impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { + fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], rust_operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, span: &[Span], _instance: Instance<'_>, _dest_catch_funclet: Option<(Self::BasicBlock, Self::BasicBlock, Option<&Self::Funclet>)>) { + if options.contains(InlineAsmOptions::MAY_UNWIND) { + self.sess() + .struct_span_err(span[0], "GCC backend does not support unwinding from inline asm") + .emit(); + return; + } + + let asm_arch = self.tcx.sess.asm_arch.unwrap(); + let is_x86 = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64); + let att_dialect = is_x86 && options.contains(InlineAsmOptions::ATT_SYNTAX); + + // GCC index of an output operand equals its position in the array + let mut outputs = vec![]; + + // GCC index of an input operand equals its position in the array + // added to `outputs.len()` + let mut inputs = vec![]; + + // Clobbers collected from `out("explicit register") _` and `inout("expl_reg") var => _` + let mut clobbers = vec![]; + + // We're trying to preallocate space for the template + let mut constants_len = 0; + + // There are rules we must adhere to if we want GCC to do the right thing: + // + // * Every local variable that the asm block uses as an output must be declared *before* + // the asm block. + // * There must be no instructions whatsoever between the register variables and the asm. + // + // Therefore, the backend must generate the instructions strictly in this order: + // + // 1. Output variables. + // 2. Register variables. + // 3. The asm block. + // + // We also must make sure that no input operands are emitted before output operands. + // + // This is why we work in passes, first emitting local vars, then local register vars. + // Also, we don't emit any asm operands immediately; we save them to + // the one of the buffers to be emitted later. + + // 1. Normal variables (and saving operands to buffers). + for (rust_idx, op) in rust_operands.iter().enumerate() { + match *op { + InlineAsmOperandRef::Out { reg, late, place } => { + use ConstraintOrRegister::*; + + let (constraint, ty) = match (reg_to_gcc(reg), place) { + (Constraint(constraint), Some(place)) => (constraint, place.layout.gcc_type(self.cx, false)), + // When `reg` is a class and not an explicit register but the out place is not specified, + // we need to create an unused output variable to assign the output to. This var + // needs to be of a type that's "compatible" with the register class, but specific type + // doesn't matter. + (Constraint(constraint), None) => (constraint, dummy_output_type(self.cx, reg.reg_class())), + (Register(_), Some(_)) => { + // left for the next pass + continue + }, + (Register(reg_name), None) => { + // `clobber_abi` can add lots of clobbers that are not supported by the target, + // such as AVX-512 registers, so we just ignore unsupported registers + let is_target_supported = reg.reg_class().supported_types(asm_arch).iter() + .any(|&(_, feature)| { + if let Some(feature) = feature { + self.tcx.sess.target_features.contains(&feature) + } else { + true // Register class is unconditionally supported + } + }); + + if is_target_supported && !clobbers.contains(®_name) { + clobbers.push(reg_name); + } + continue + } + }; + + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + outputs.push(AsmOutOperand { + constraint, + rust_idx, + late, + readwrite: false, + tmp_var, + out_place: place + }); + } + + InlineAsmOperandRef::In { reg, value } => { + if let ConstraintOrRegister::Constraint(constraint) = reg_to_gcc(reg) { + inputs.push(AsmInOperand { + constraint: Cow::Borrowed(constraint), + rust_idx, + val: value.immediate() + }); + } + else { + // left for the next pass + continue + } + } + + InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => { + let constraint = if let ConstraintOrRegister::Constraint(constraint) = reg_to_gcc(reg) { + constraint + } + else { + // left for the next pass + continue + }; + + // Rustc frontend guarantees that input and output types are "compatible", + // so we can just use input var's type for the output variable. + // + // This decision is also backed by the fact that LLVM needs in and out + // values to be of *exactly the same type*, not just "compatible". + // I'm not sure if GCC is so picky too, but better safe than sorry. + let ty = in_value.layout.gcc_type(self.cx, false); + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + + // If the out_place is None (i.e `inout(reg) _` syntax was used), we translate + // it to one "readwrite (+) output variable", otherwise we translate it to two + // "out and tied in" vars as described above. + let readwrite = out_place.is_none(); + outputs.push(AsmOutOperand { + constraint, + rust_idx, + late, + readwrite, + tmp_var, + out_place, + }); + + if !readwrite { + let out_gcc_idx = outputs.len() - 1; + let constraint = Cow::Owned(out_gcc_idx.to_string()); + + inputs.push(AsmInOperand { + constraint, + rust_idx, + val: in_value.immediate() + }); + } + } + + InlineAsmOperandRef::Const { ref string } => { + constants_len += string.len() + att_dialect as usize; + } + + InlineAsmOperandRef::SymFn { instance } => { + // TODO(@Amanieu): Additional mangling is needed on + // some targets to add a leading underscore (Mach-O) + // or byte count suffixes (x86 Windows). + constants_len += self.tcx.symbol_name(instance).name.len(); + } + InlineAsmOperandRef::SymStatic { def_id } => { + // TODO(@Amanieu): Additional mangling is needed on + // some targets to add a leading underscore (Mach-O). + constants_len += self.tcx.symbol_name(Instance::mono(self.tcx, def_id)).name.len(); + } + } + } + + // 2. Register variables. + for (rust_idx, op) in rust_operands.iter().enumerate() { + match *op { + // `out("explicit register") var` + InlineAsmOperandRef::Out { reg, late, place } => { + if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) { + let out_place = if let Some(place) = place { + place + } + else { + // processed in the previous pass + continue + }; + + let ty = out_place.layout.gcc_type(self.cx, false); + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + tmp_var.set_register_name(reg_name); + + outputs.push(AsmOutOperand { + constraint: "r".into(), + rust_idx, + late, + readwrite: false, + tmp_var, + out_place: Some(out_place) + }); + } + + // processed in the previous pass + } + + // `in("explicit register") var` + InlineAsmOperandRef::In { reg, value } => { + if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) { + let ty = value.layout.gcc_type(self.cx, false); + let reg_var = self.current_func().new_local(None, ty, "input_register"); + reg_var.set_register_name(reg_name); + self.llbb().add_assignment(None, reg_var, value.immediate()); + + inputs.push(AsmInOperand { + constraint: "r".into(), + rust_idx, + val: reg_var.to_rvalue() + }); + } + + // processed in the previous pass + } + + // `inout("explicit register") in_var => out_var` + InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => { + if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) { + // See explanation in the first pass. + let ty = in_value.layout.gcc_type(self.cx, false); + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + tmp_var.set_register_name(reg_name); + + outputs.push(AsmOutOperand { + constraint: "r".into(), + rust_idx, + late, + readwrite: false, + tmp_var, + out_place, + }); + + let constraint = Cow::Owned((outputs.len() - 1).to_string()); + inputs.push(AsmInOperand { + constraint, + rust_idx, + val: in_value.immediate() + }); + } + + // processed in the previous pass + } + + InlineAsmOperandRef::SymFn { instance } => { + inputs.push(AsmInOperand { + constraint: "X".into(), + rust_idx, + val: self.cx.rvalue_as_function(get_fn(self.cx, instance)) + .get_address(None), + }); + } + + InlineAsmOperandRef::SymStatic { def_id } => { + inputs.push(AsmInOperand { + constraint: "X".into(), + rust_idx, + val: self.cx.get_static(def_id).get_address(None), + }); + } + + InlineAsmOperandRef::Const { .. } => { + // processed in the previous pass + } + } + } + + // 3. Build the template string + + let mut template_str = String::with_capacity(estimate_template_length(template, constants_len, att_dialect)); + if att_dialect { + template_str.push_str(ATT_SYNTAX_INS); + } + + for piece in template { + match *piece { + InlineAsmTemplatePiece::String(ref string) => { + // TODO(@Commeownist): switch to `Iterator::intersperse` once it's stable + let mut iter = string.split('%'); + if let Some(s) = iter.next() { + template_str.push_str(s); + } + + for s in iter { + template_str.push_str("%%"); + template_str.push_str(s); + } + } + InlineAsmTemplatePiece::Placeholder { operand_idx, modifier, span: _ } => { + let mut push_to_template = |modifier, gcc_idx| { + use std::fmt::Write; + + template_str.push('%'); + if let Some(modifier) = modifier { + template_str.push(modifier); + } + write!(template_str, "{}", gcc_idx).expect("pushing to string failed"); + }; + + match rust_operands[operand_idx] { + InlineAsmOperandRef::Out { reg, .. } => { + let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); + let gcc_index = outputs.iter() + .position(|op| operand_idx == op.rust_idx) + .expect("wrong rust index"); + push_to_template(modifier, gcc_index); + } + + InlineAsmOperandRef::In { reg, .. } => { + let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); + let in_gcc_index = inputs.iter() + .position(|op| operand_idx == op.rust_idx) + .expect("wrong rust index"); + let gcc_index = in_gcc_index + outputs.len(); + push_to_template(modifier, gcc_index); + } + + InlineAsmOperandRef::InOut { reg, .. } => { + let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); + + // The input register is tied to the output, so we can just use the index of the output register + let gcc_index = outputs.iter() + .position(|op| operand_idx == op.rust_idx) + .expect("wrong rust index"); + push_to_template(modifier, gcc_index); + } + + InlineAsmOperandRef::SymFn { instance } => { + // TODO(@Amanieu): Additional mangling is needed on + // some targets to add a leading underscore (Mach-O) + // or byte count suffixes (x86 Windows). + let name = self.tcx.symbol_name(instance).name; + template_str.push_str(name); + } + + InlineAsmOperandRef::SymStatic { def_id } => { + // TODO(@Amanieu): Additional mangling is needed on + // some targets to add a leading underscore (Mach-O). + let instance = Instance::mono(self.tcx, def_id); + let name = self.tcx.symbol_name(instance).name; + template_str.push_str(name); + } + + InlineAsmOperandRef::Const { ref string } => { + // Const operands get injected directly into the template + if att_dialect { + template_str.push('$'); + } + template_str.push_str(string); + } + } + } + } + } + + if att_dialect { + template_str.push_str(INTEL_SYNTAX_INS); + } + + // 4. Generate Extended Asm block + + let block = self.llbb(); + let extended_asm = block.add_extended_asm(None, &template_str); + + for op in &outputs { + extended_asm.add_output_operand(None, &op.to_constraint(), op.tmp_var); + } + + for op in &inputs { + extended_asm.add_input_operand(None, &op.constraint, op.val); + } + + for clobber in clobbers.iter() { + extended_asm.add_clobber(clobber); + } + + if !options.contains(InlineAsmOptions::PRESERVES_FLAGS) { + // TODO(@Commeownist): I'm not 100% sure this one clobber is sufficient + // on all architectures. For instance, what about FP stack? + extended_asm.add_clobber("cc"); + } + if !options.contains(InlineAsmOptions::NOMEM) { + extended_asm.add_clobber("memory"); + } + if !options.contains(InlineAsmOptions::PURE) { + extended_asm.set_volatile_flag(true); + } + if !options.contains(InlineAsmOptions::NOSTACK) { + // TODO(@Commeownist): figure out how to align stack + } + if options.contains(InlineAsmOptions::NORETURN) { + let builtin_unreachable = self.context.get_builtin_function("__builtin_unreachable"); + let builtin_unreachable: RValue<'gcc> = unsafe { std::mem::transmute(builtin_unreachable) }; + self.call(self.type_void(), builtin_unreachable, &[], None); + } + + // Write results to outputs. + // + // We need to do this because: + // 1. Turning `PlaceRef` into `RValue` is error-prone and has nasty edge cases + // (especially with current `rustc_backend_ssa` API). + // 2. Not every output operand has an `out_place`, and it's required by `add_output_operand`. + // + // Instead, we generate a temporary output variable for each output operand, and then this loop, + // generates `out_place = tmp_var;` assignments if out_place exists. + for op in &outputs { + if let Some(place) = op.out_place { + OperandValue::Immediate(op.tmp_var.to_rvalue()).store(self, place); + } + } + + } +} + +fn estimate_template_length(template: &[InlineAsmTemplatePiece], constants_len: usize, att_dialect: bool) -> usize { + let len: usize = template.iter().map(|piece| { + match *piece { + InlineAsmTemplatePiece::String(ref string) => { + string.len() + } + InlineAsmTemplatePiece::Placeholder { .. } => { + // '%' + 1 char modifier + 1 char index + 3 + } + } + }) + .sum(); + + // increase it by 5% to account for possible '%' signs that'll be duplicated + // I pulled the number out of blue, but should be fair enough + // as the upper bound + let mut res = (len as f32 * 1.05) as usize + constants_len; + + if att_dialect { + res += INTEL_SYNTAX_INS.len() + ATT_SYNTAX_INS.len(); + } + res +} + +/// Converts a register class to a GCC constraint code. +fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> ConstraintOrRegister { + let constraint = match reg { + // For vector registers LLVM wants the register name to match the type size. + InlineAsmRegOrRegClass::Reg(reg) => { + match reg { + InlineAsmReg::X86(_) => { + // TODO(antoyo): add support for vector register. + // + // // For explicit registers, we have to create a register variable: https://stackoverflow.com/a/31774784/389119 + return ConstraintOrRegister::Register(match reg.name() { + // Some of registers' names does not map 1-1 from rust to gcc + "st(0)" => "st", + + name => name, + }); + } + + _ => unimplemented!(), + } + }, + InlineAsmRegOrRegClass::RegClass(reg) => match reg { + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => unimplemented!(), + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg) => unimplemented!(), + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) => unimplemented!(), + InlineAsmRegClass::Avr(_) => unimplemented!(), + InlineAsmRegClass::Bpf(_) => unimplemented!(), + InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => unimplemented!(), + InlineAsmRegClass::Msp430(_) => unimplemented!(), + InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => unimplemented!(), + InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => unimplemented!(), + InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => unimplemented!(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg_nonzero) => unimplemented!(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::freg) => unimplemented!(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::cr) + | InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::xer) => { + unreachable!("clobber-only") + }, + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => unimplemented!(), + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => unimplemented!(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) => "r", + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => "Q", + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => "q", + InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg) + | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => "x", + InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => "v", + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => "Yk", + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg0) => unimplemented!(), + InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(), + InlineAsmRegClass::X86( + X86InlineAsmRegClass::x87_reg | X86InlineAsmRegClass::mmx_reg | X86InlineAsmRegClass::tmm_reg, + ) => unreachable!("clobber-only"), + InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => { + bug!("GCC backend does not support SPIR-V") + } + InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => unimplemented!(), + InlineAsmRegClass::Err => unreachable!(), + } + }; + + ConstraintOrRegister::Constraint(constraint) +} + +/// Type to use for outputs that are discarded. It doesn't really matter what +/// the type is, as long as it is valid for the constraint code. +fn dummy_output_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, reg: InlineAsmRegClass) -> Type<'gcc> { + match reg { + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => unimplemented!(), + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg) + | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => { + unimplemented!() + } + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg)=> cx.type_i32(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => cx.type_f32(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => cx.type_f64(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => { + unimplemented!() + } + InlineAsmRegClass::Avr(_) => unimplemented!(), + InlineAsmRegClass::Bpf(_) => unimplemented!(), + InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => cx.type_f32(), + InlineAsmRegClass::Msp430(_) => unimplemented!(), + InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => cx.type_i16(), + InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => cx.type_i32(), + InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => cx.type_i64(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg_nonzero) => cx.type_i32(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::freg) => cx.type_f64(), + InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::cr) + | InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::xer) => { + unreachable!("clobber-only") + }, + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => cx.type_f32(), + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => cx.type_f32(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) + | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => cx.type_i32(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => cx.type_i8(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::mmx_reg) => unimplemented!(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg) + | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) + | InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => cx.type_f32(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg) => unimplemented!(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => cx.type_i16(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg0) => cx.type_i16(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::tmm_reg) => unimplemented!(), + InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => cx.type_i32(), + InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => { + bug!("LLVM backend does not support SPIR-V") + }, + InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => cx.type_f64(), + InlineAsmRegClass::Err => unreachable!(), + } +} + +impl<'gcc, 'tcx> AsmMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn codegen_global_asm(&self, template: &[InlineAsmTemplatePiece], operands: &[GlobalAsmOperandRef<'tcx>], options: InlineAsmOptions, _line_spans: &[Span]) { + let asm_arch = self.tcx.sess.asm_arch.unwrap(); + + // Default to Intel syntax on x86 + let att_dialect = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64) + && options.contains(InlineAsmOptions::ATT_SYNTAX); + + // Build the template string + let mut template_str = String::new(); + for piece in template { + match *piece { + InlineAsmTemplatePiece::String(ref string) => { + for line in string.lines() { + // NOTE: gcc does not allow inline comment, so remove them. + let line = + if let Some(index) = line.rfind("//") { + &line[..index] + } + else { + line + }; + template_str.push_str(line); + template_str.push('\n'); + } + }, + InlineAsmTemplatePiece::Placeholder { operand_idx, modifier: _, span: _ } => { + match operands[operand_idx] { + GlobalAsmOperandRef::Const { ref string } => { + // Const operands get injected directly into the + // template. Note that we don't need to escape % + // here unlike normal inline assembly. + template_str.push_str(string); + } + + GlobalAsmOperandRef::SymFn { instance } => { + // TODO(@Amanieu): Additional mangling is needed on + // some targets to add a leading underscore (Mach-O) + // or byte count suffixes (x86 Windows). + let name = self.tcx.symbol_name(instance).name; + template_str.push_str(name); + } + + GlobalAsmOperandRef::SymStatic { def_id } => { + // TODO(@Amanieu): Additional mangling is needed on + // some targets to add a leading underscore (Mach-O). + let instance = Instance::mono(self.tcx, def_id); + let name = self.tcx.symbol_name(instance).name; + template_str.push_str(name); + } + } + } + } + } + + let template_str = + if att_dialect { + format!(".att_syntax\n\t{}\n\t.intel_syntax noprefix", template_str) + } + else { + template_str + }; + // NOTE: seems like gcc will put the asm in the wrong section, so set it to .text manually. + let template_str = format!(".pushsection .text\n{}\n.popsection", template_str); + self.context.add_top_level_asm(None, &template_str); + } +} + +fn modifier_to_gcc(arch: InlineAsmArch, reg: InlineAsmRegClass, modifier: Option<char>) -> Option<char> { + match reg { + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => modifier, + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => modifier, + InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg) + | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => { + unimplemented!() + } + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => unimplemented!(), + InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8) + | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => { + unimplemented!() + } + InlineAsmRegClass::Avr(_) => unimplemented!(), + InlineAsmRegClass::Bpf(_) => unimplemented!(), + InlineAsmRegClass::Hexagon(_) => unimplemented!(), + InlineAsmRegClass::Mips(_) => unimplemented!(), + InlineAsmRegClass::Msp430(_) => unimplemented!(), + InlineAsmRegClass::Nvptx(_) => unimplemented!(), + InlineAsmRegClass::PowerPC(_) => unimplemented!(), + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) + | InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => unimplemented!(), + InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => unimplemented!(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) + | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => match modifier { + None => if arch == InlineAsmArch::X86_64 { Some('q') } else { Some('k') }, + Some('l') => Some('b'), + Some('h') => Some('h'), + Some('x') => Some('w'), + Some('e') => Some('k'), + Some('r') => Some('q'), + _ => unreachable!(), + }, + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => None, + InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::xmm_reg) + | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::ymm_reg) + | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::zmm_reg) => match (reg, modifier) { + (X86InlineAsmRegClass::xmm_reg, None) => Some('x'), + (X86InlineAsmRegClass::ymm_reg, None) => Some('t'), + (X86InlineAsmRegClass::zmm_reg, None) => Some('g'), + (_, Some('x')) => Some('x'), + (_, Some('y')) => Some('t'), + (_, Some('z')) => Some('g'), + _ => unreachable!(), + }, + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => None, + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg0) => None, + InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg | X86InlineAsmRegClass::mmx_reg | X86InlineAsmRegClass::tmm_reg) => { + unreachable!("clobber-only") + } + InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(), + InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => { + bug!("LLVM backend does not support SPIR-V") + }, + InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => unimplemented!(), + InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => unimplemented!(), + InlineAsmRegClass::Err => unreachable!(), + } +} diff --git a/compiler/rustc_codegen_gcc/src/back/mod.rs b/compiler/rustc_codegen_gcc/src/back/mod.rs new file mode 100644 index 000000000..d692799d7 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/back/mod.rs @@ -0,0 +1 @@ +pub mod write; diff --git a/compiler/rustc_codegen_gcc/src/back/write.rs b/compiler/rustc_codegen_gcc/src/back/write.rs new file mode 100644 index 000000000..efcf18d31 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/back/write.rs @@ -0,0 +1,83 @@ +use std::{env, fs}; + +use gccjit::OutputKind; +use rustc_codegen_ssa::{CompiledModule, ModuleCodegen}; +use rustc_codegen_ssa::back::write::{CodegenContext, EmitObj, ModuleConfig}; +use rustc_errors::Handler; +use rustc_session::config::OutputType; +use rustc_span::fatal_error::FatalError; +use rustc_target::spec::SplitDebuginfo; + +use crate::{GccCodegenBackend, GccContext}; + +pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, module: ModuleCodegen<GccContext>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> { + let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen", &*module.name); + { + let context = &module.module_llvm.context; + + let module_name = module.name.clone(); + let module_name = Some(&module_name[..]); + + let _bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name); + let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name); + + if config.bitcode_needed() { + // TODO(antoyo) + } + + if config.emit_ir { + unimplemented!(); + } + + if config.emit_asm { + let _timer = cgcx + .prof + .generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name); + let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name); + context.compile_to_file(OutputKind::Assembler, path.to_str().expect("path to str")); + } + + match config.emit_obj { + EmitObj::ObjectCode(_) => { + let _timer = cgcx + .prof + .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name); + if env::var("CG_GCCJIT_DUMP_MODULE_NAMES").as_deref() == Ok("1") { + println!("Module {}", module.name); + } + if env::var("CG_GCCJIT_DUMP_ALL_MODULES").as_deref() == Ok("1") || env::var("CG_GCCJIT_DUMP_MODULE").as_deref() == Ok(&module.name) { + println!("Dumping reproducer {}", module.name); + let _ = fs::create_dir("/tmp/reproducers"); + // FIXME(antoyo): segfault in dump_reproducer_to_file() might be caused by + // transmuting an rvalue to an lvalue. + // Segfault is actually in gcc::jit::reproducer::get_identifier_as_lvalue + context.dump_reproducer_to_file(&format!("/tmp/reproducers/{}.c", module.name)); + println!("Dumped reproducer {}", module.name); + } + if env::var("CG_GCCJIT_DUMP_TO_FILE").as_deref() == Ok("1") { + let _ = fs::create_dir("/tmp/gccjit_dumps"); + let path = &format!("/tmp/gccjit_dumps/{}.c", module.name); + context.dump_to_file(path, true); + } + context.compile_to_file(OutputKind::ObjectFile, obj_out.to_str().expect("path to str")); + } + + EmitObj::Bitcode => { + // TODO(antoyo) + } + + EmitObj::None => {} + } + } + + Ok(module.into_compiled_module( + config.emit_obj != EmitObj::None, + cgcx.target_can_use_split_dwarf && cgcx.split_debuginfo == SplitDebuginfo::Unpacked, + config.emit_bc, + &cgcx.output_filenames, + )) +} + +pub(crate) fn link(_cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, mut _modules: Vec<ModuleCodegen<GccContext>>) -> Result<ModuleCodegen<GccContext>, FatalError> { + unimplemented!(); +} diff --git a/compiler/rustc_codegen_gcc/src/base.rs b/compiler/rustc_codegen_gcc/src/base.rs new file mode 100644 index 000000000..8f9f6f98f --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/base.rs @@ -0,0 +1,154 @@ +use std::env; +use std::time::Instant; + +use gccjit::{ + Context, + FunctionType, + GlobalKind, +}; +use rustc_middle::dep_graph; +use rustc_middle::ty::TyCtxt; +use rustc_middle::mir::mono::Linkage; +use rustc_codegen_ssa::{ModuleCodegen, ModuleKind}; +use rustc_codegen_ssa::base::maybe_create_entry_wrapper; +use rustc_codegen_ssa::mono_item::MonoItemExt; +use rustc_codegen_ssa::traits::DebugInfoMethods; +use rustc_session::config::DebugInfo; +use rustc_span::Symbol; + +use crate::GccContext; +use crate::builder::Builder; +use crate::context::CodegenCx; + +pub fn global_linkage_to_gcc(linkage: Linkage) -> GlobalKind { + match linkage { + Linkage::External => GlobalKind::Imported, + Linkage::AvailableExternally => GlobalKind::Imported, + Linkage::LinkOnceAny => unimplemented!(), + Linkage::LinkOnceODR => unimplemented!(), + Linkage::WeakAny => unimplemented!(), + Linkage::WeakODR => unimplemented!(), + Linkage::Appending => unimplemented!(), + Linkage::Internal => GlobalKind::Internal, + Linkage::Private => GlobalKind::Internal, + Linkage::ExternalWeak => GlobalKind::Imported, // TODO(antoyo): should be weak linkage. + Linkage::Common => unimplemented!(), + } +} + +pub fn linkage_to_gcc(linkage: Linkage) -> FunctionType { + match linkage { + Linkage::External => FunctionType::Exported, + Linkage::AvailableExternally => FunctionType::Extern, + Linkage::LinkOnceAny => unimplemented!(), + Linkage::LinkOnceODR => unimplemented!(), + Linkage::WeakAny => FunctionType::Exported, // FIXME(antoyo): should be similar to linkonce. + Linkage::WeakODR => unimplemented!(), + Linkage::Appending => unimplemented!(), + Linkage::Internal => FunctionType::Internal, + Linkage::Private => FunctionType::Internal, + Linkage::ExternalWeak => unimplemented!(), + Linkage::Common => unimplemented!(), + } +} + +pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol, supports_128bit_integers: bool) -> (ModuleCodegen<GccContext>, u64) { + let prof_timer = tcx.prof.generic_activity("codegen_module"); + let start_time = Instant::now(); + + let dep_node = tcx.codegen_unit(cgu_name).codegen_dep_node(tcx); + let (module, _) = tcx.dep_graph.with_task( + dep_node, + tcx, + (cgu_name, supports_128bit_integers), + module_codegen, + Some(dep_graph::hash_result), + ); + let time_to_codegen = start_time.elapsed(); + drop(prof_timer); + + // We assume that the cost to run GCC on a CGU is proportional to + // the time we needed for codegenning it. + let cost = time_to_codegen.as_secs() * 1_000_000_000 + time_to_codegen.subsec_nanos() as u64; + + fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, supports_128bit_integers): (Symbol, bool)) -> ModuleCodegen<GccContext> { + let cgu = tcx.codegen_unit(cgu_name); + // Instantiate monomorphizations without filling out definitions yet... + //let llvm_module = ModuleLlvm::new(tcx, &cgu_name.as_str()); + let context = Context::default(); + // TODO(antoyo): only set on x86 platforms. + context.add_command_line_option("-masm=intel"); + // TODO(antoyo): only add the following cli argument if the feature is supported. + context.add_command_line_option("-msse2"); + context.add_command_line_option("-mavx2"); + context.add_command_line_option("-msha"); + context.add_command_line_option("-mpclmul"); + // FIXME(antoyo): the following causes an illegal instruction on vmovdqu64 in std_example on my CPU. + // Only add if the CPU supports it. + //context.add_command_line_option("-mavx512f"); + for arg in &tcx.sess.opts.cg.llvm_args { + context.add_command_line_option(arg); + } + // NOTE: This is needed to compile the file src/intrinsic/archs.rs during a bootstrap of rustc. + context.add_command_line_option("-fno-var-tracking-assignments"); + // NOTE: an optimization (https://github.com/rust-lang/rustc_codegen_gcc/issues/53). + context.add_command_line_option("-fno-semantic-interposition"); + // NOTE: Rust relies on LLVM not doing TBAA (https://github.com/rust-lang/unsafe-code-guidelines/issues/292). + context.add_command_line_option("-fno-strict-aliasing"); + + if tcx.sess.opts.unstable_opts.function_sections.unwrap_or(tcx.sess.target.function_sections) { + context.add_command_line_option("-ffunction-sections"); + context.add_command_line_option("-fdata-sections"); + } + + if env::var("CG_GCCJIT_DUMP_CODE").as_deref() == Ok("1") { + context.set_dump_code_on_compile(true); + } + if env::var("CG_GCCJIT_DUMP_GIMPLE").as_deref() == Ok("1") { + context.set_dump_initial_gimple(true); + } + context.set_debug_info(true); + if env::var("CG_GCCJIT_DUMP_EVERYTHING").as_deref() == Ok("1") { + context.set_dump_everything(true); + } + if env::var("CG_GCCJIT_KEEP_INTERMEDIATES").as_deref() == Ok("1") { + context.set_keep_intermediates(true); + } + + // TODO(bjorn3): Remove once unwinding is properly implemented + context.set_allow_unreachable_blocks(true); + + { + let cx = CodegenCx::new(&context, cgu, tcx, supports_128bit_integers); + + let mono_items = cgu.items_in_deterministic_order(tcx); + for &(mono_item, (linkage, visibility)) in &mono_items { + mono_item.predefine::<Builder<'_, '_, '_>>(&cx, linkage, visibility); + } + + // ... and now that we have everything pre-defined, fill out those definitions. + for &(mono_item, _) in &mono_items { + mono_item.define::<Builder<'_, '_, '_>>(&cx); + } + + // If this codegen unit contains the main function, also create the + // wrapper here + maybe_create_entry_wrapper::<Builder<'_, '_, '_>>(&cx); + + // Finalize debuginfo + if cx.sess().opts.debuginfo != DebugInfo::None { + cx.debuginfo_finalize(); + } + } + + ModuleCodegen { + name: cgu_name.to_string(), + module_llvm: GccContext { + context + }, + kind: ModuleKind::Regular, + } + } + + (module, cost) +} diff --git a/compiler/rustc_codegen_gcc/src/builder.rs b/compiler/rustc_codegen_gcc/src/builder.rs new file mode 100644 index 000000000..4d40dd099 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/builder.rs @@ -0,0 +1,1561 @@ +use std::borrow::Cow; +use std::cell::Cell; +use std::convert::TryFrom; +use std::ops::Deref; + +use gccjit::{ + BinaryOp, + Block, + ComparisonOp, + Context, + Function, + LValue, + RValue, + ToRValue, + Type, + UnaryOp, +}; +use rustc_codegen_ssa::MemFlags; +use rustc_codegen_ssa::common::{AtomicOrdering, AtomicRmwBinOp, IntPredicate, RealPredicate, SynchronizationScope}; +use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue}; +use rustc_codegen_ssa::mir::place::PlaceRef; +use rustc_codegen_ssa::traits::{ + BackendTypes, + BaseTypeMethods, + BuilderMethods, + ConstMethods, + DerivedTypeMethods, + LayoutTypeMethods, + HasCodegen, + OverflowOp, + StaticBuilderMethods, +}; +use rustc_data_structures::fx::FxHashSet; +use rustc_middle::ty::{ParamEnv, Ty, TyCtxt}; +use rustc_middle::ty::layout::{FnAbiError, FnAbiOfHelpers, FnAbiRequest, HasParamEnv, HasTyCtxt, LayoutError, LayoutOfHelpers, TyAndLayout}; +use rustc_span::Span; +use rustc_span::def_id::DefId; +use rustc_target::abi::{ + self, + call::FnAbi, + Align, + HasDataLayout, + Size, + TargetDataLayout, + WrappingRange, +}; +use rustc_target::spec::{HasTargetSpec, Target}; + +use crate::common::{SignType, TypeReflection, type_is_pointer}; +use crate::context::CodegenCx; +use crate::intrinsic::llvm; +use crate::type_of::LayoutGccExt; + +// TODO(antoyo) +type Funclet = (); + +// TODO(antoyo): remove this variable. +static mut RETURN_VALUE_COUNT: usize = 0; + +enum ExtremumOperation { + Max, + Min, +} + +pub struct Builder<'a: 'gcc, 'gcc, 'tcx> { + pub cx: &'a CodegenCx<'gcc, 'tcx>, + pub block: Block<'gcc>, + stack_var_count: Cell<usize>, +} + +impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { + fn with_cx(cx: &'a CodegenCx<'gcc, 'tcx>, block: Block<'gcc>) -> Self { + Builder { + cx, + block, + stack_var_count: Cell::new(0), + } + } + + fn atomic_extremum(&mut self, operation: ExtremumOperation, dst: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering) -> RValue<'gcc> { + let size = src.get_type().get_size(); + + let func = self.current_func(); + + let load_ordering = + match order { + // TODO(antoyo): does this make sense? + AtomicOrdering::AcquireRelease | AtomicOrdering::Release => AtomicOrdering::Acquire, + _ => order, + }; + let previous_value = self.atomic_load(dst.get_type(), dst, load_ordering, Size::from_bytes(size)); + let previous_var = func.new_local(None, previous_value.get_type(), "previous_value"); + let return_value = func.new_local(None, previous_value.get_type(), "return_value"); + self.llbb().add_assignment(None, previous_var, previous_value); + self.llbb().add_assignment(None, return_value, previous_var.to_rvalue()); + + let while_block = func.new_block("while"); + let after_block = func.new_block("after_while"); + self.llbb().end_with_jump(None, while_block); + + // NOTE: since jumps were added and compare_exchange doesn't expect this, the current block in the + // state need to be updated. + self.switch_to_block(while_block); + + let comparison_operator = + match operation { + ExtremumOperation::Max => ComparisonOp::LessThan, + ExtremumOperation::Min => ComparisonOp::GreaterThan, + }; + + let cond1 = self.context.new_comparison(None, comparison_operator, previous_var.to_rvalue(), self.context.new_cast(None, src, previous_value.get_type())); + let compare_exchange = self.compare_exchange(dst, previous_var, src, order, load_ordering, false); + let cond2 = self.cx.context.new_unary_op(None, UnaryOp::LogicalNegate, compare_exchange.get_type(), compare_exchange); + let cond = self.cx.context.new_binary_op(None, BinaryOp::LogicalAnd, self.cx.bool_type, cond1, cond2); + + while_block.end_with_conditional(None, cond, while_block, after_block); + + // NOTE: since jumps were added in a place rustc does not expect, the current block in the + // state need to be updated. + self.switch_to_block(after_block); + + return_value.to_rvalue() + } + + fn compare_exchange(&self, dst: RValue<'gcc>, cmp: LValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering, failure_order: AtomicOrdering, weak: bool) -> RValue<'gcc> { + let size = src.get_type().get_size(); + let compare_exchange = self.context.get_builtin_function(&format!("__atomic_compare_exchange_{}", size)); + let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc()); + let failure_order = self.context.new_rvalue_from_int(self.i32_type, failure_order.to_gcc()); + let weak = self.context.new_rvalue_from_int(self.bool_type, weak as i32); + + let void_ptr_type = self.context.new_type::<*mut ()>(); + let volatile_void_ptr_type = void_ptr_type.make_volatile(); + let dst = self.context.new_cast(None, dst, volatile_void_ptr_type); + let expected = self.context.new_cast(None, cmp.get_address(None), void_ptr_type); + + // NOTE: not sure why, but we have the wrong type here. + let int_type = compare_exchange.get_param(2).to_rvalue().get_type(); + let src = self.context.new_cast(None, src, int_type); + self.context.new_call(None, compare_exchange, &[dst, expected, src, weak, order, failure_order]) + } + + pub fn assign(&self, lvalue: LValue<'gcc>, value: RValue<'gcc>) { + self.llbb().add_assignment(None, lvalue, value); + } + + fn check_call<'b>(&mut self, _typ: &str, func: Function<'gcc>, args: &'b [RValue<'gcc>]) -> Cow<'b, [RValue<'gcc>]> { + let mut all_args_match = true; + let mut param_types = vec![]; + let param_count = func.get_param_count(); + for (index, arg) in args.iter().enumerate().take(param_count) { + let param = func.get_param(index as i32); + let param = param.to_rvalue().get_type(); + if param != arg.get_type() { + all_args_match = false; + } + param_types.push(param); + } + + if all_args_match { + return Cow::Borrowed(args); + } + + let casted_args: Vec<_> = param_types + .into_iter() + .zip(args.iter()) + .enumerate() + .map(|(_i, (expected_ty, &actual_val))| { + let actual_ty = actual_val.get_type(); + if expected_ty != actual_ty { + self.bitcast(actual_val, expected_ty) + } + else { + actual_val + } + }) + .collect(); + + Cow::Owned(casted_args) + } + + fn check_ptr_call<'b>(&mut self, _typ: &str, func_ptr: RValue<'gcc>, args: &'b [RValue<'gcc>]) -> Cow<'b, [RValue<'gcc>]> { + let mut all_args_match = true; + let mut param_types = vec![]; + let gcc_func = func_ptr.get_type().dyncast_function_ptr_type().expect("function ptr"); + for (index, arg) in args.iter().enumerate().take(gcc_func.get_param_count()) { + let param = gcc_func.get_param_type(index); + if param != arg.get_type() { + all_args_match = false; + } + param_types.push(param); + } + + let mut on_stack_param_indices = FxHashSet::default(); + if let Some(indices) = self.on_stack_params.borrow().get(&gcc_func) { + on_stack_param_indices = indices.clone(); + } + + if all_args_match { + return Cow::Borrowed(args); + } + + let func_name = format!("{:?}", func_ptr); + + let casted_args: Vec<_> = param_types + .into_iter() + .zip(args.iter()) + .enumerate() + .map(|(index, (expected_ty, &actual_val))| { + if llvm::ignore_arg_cast(&func_name, index, args.len()) { + return actual_val; + } + + let actual_ty = actual_val.get_type(); + if expected_ty != actual_ty { + if !actual_ty.is_vector() && !expected_ty.is_vector() && actual_ty.is_integral() && expected_ty.is_integral() && actual_ty.get_size() != expected_ty.get_size() { + self.context.new_cast(None, actual_val, expected_ty) + } + else if on_stack_param_indices.contains(&index) { + actual_val.dereference(None).to_rvalue() + } + else { + assert!(!((actual_ty.is_vector() && !expected_ty.is_vector()) || (!actual_ty.is_vector() && expected_ty.is_vector())), "{:?} ({}) -> {:?} ({}), index: {:?}[{}]", actual_ty, actual_ty.is_vector(), expected_ty, expected_ty.is_vector(), func_ptr, index); + // TODO(antoyo): perhaps use __builtin_convertvector for vector casting. + self.bitcast(actual_val, expected_ty) + } + } + else { + actual_val + } + }) + .collect(); + + Cow::Owned(casted_args) + } + + fn check_store(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>) -> RValue<'gcc> { + let dest_ptr_ty = self.cx.val_ty(ptr).make_pointer(); // TODO(antoyo): make sure make_pointer() is okay here. + let stored_ty = self.cx.val_ty(val); + let stored_ptr_ty = self.cx.type_ptr_to(stored_ty); + + if dest_ptr_ty == stored_ptr_ty { + ptr + } + else { + self.bitcast(ptr, stored_ptr_ty) + } + } + + pub fn current_func(&self) -> Function<'gcc> { + self.block.get_function() + } + + fn function_call(&mut self, func: RValue<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> { + // TODO(antoyo): remove when the API supports a different type for functions. + let func: Function<'gcc> = self.cx.rvalue_as_function(func); + let args = self.check_call("call", func, args); + + // gccjit requires to use the result of functions, even when it's not used. + // That's why we assign the result to a local or call add_eval(). + let return_type = func.get_return_type(); + let void_type = self.context.new_type::<()>(); + let current_func = self.block.get_function(); + if return_type != void_type { + unsafe { RETURN_VALUE_COUNT += 1 }; + let result = current_func.new_local(None, return_type, &format!("returnValue{}", unsafe { RETURN_VALUE_COUNT })); + self.block.add_assignment(None, result, self.cx.context.new_call(None, func, &args)); + result.to_rvalue() + } + else { + self.block.add_eval(None, self.cx.context.new_call(None, func, &args)); + // Return dummy value when not having return value. + self.context.new_rvalue_from_long(self.isize_type, 0) + } + } + + fn function_ptr_call(&mut self, func_ptr: RValue<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> { + let args = self.check_ptr_call("call", func_ptr, args); + + // gccjit requires to use the result of functions, even when it's not used. + // That's why we assign the result to a local or call add_eval(). + let gcc_func = func_ptr.get_type().dyncast_function_ptr_type().expect("function ptr"); + let return_type = gcc_func.get_return_type(); + let void_type = self.context.new_type::<()>(); + let current_func = self.block.get_function(); + + if return_type != void_type { + unsafe { RETURN_VALUE_COUNT += 1 }; + let result = current_func.new_local(None, return_type, &format!("ptrReturnValue{}", unsafe { RETURN_VALUE_COUNT })); + let func_name = format!("{:?}", func_ptr); + let args = llvm::adjust_intrinsic_arguments(&self, gcc_func, args, &func_name); + self.block.add_assignment(None, result, self.cx.context.new_call_through_ptr(None, func_ptr, &args)); + result.to_rvalue() + } + else { + #[cfg(not(feature="master"))] + if gcc_func.get_param_count() == 0 { + // FIXME(antoyo): As a temporary workaround for unsupported LLVM intrinsics. + self.block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &[])); + } + else { + self.block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &args)); + } + #[cfg(feature="master")] + self.block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &args)); + // Return dummy value when not having return value. + let result = current_func.new_local(None, self.isize_type, "dummyValueThatShouldNeverBeUsed"); + self.block.add_assignment(None, result, self.context.new_rvalue_from_long(self.isize_type, 0)); + result.to_rvalue() + } + } + + pub fn overflow_call(&self, func: Function<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> { + // gccjit requires to use the result of functions, even when it's not used. + // That's why we assign the result to a local. + let return_type = self.context.new_type::<bool>(); + let current_func = self.block.get_function(); + // TODO(antoyo): return the new_call() directly? Since the overflow function has no side-effects. + unsafe { RETURN_VALUE_COUNT += 1 }; + let result = current_func.new_local(None, return_type, &format!("overflowReturnValue{}", unsafe { RETURN_VALUE_COUNT })); + self.block.add_assignment(None, result, self.cx.context.new_call(None, func, &args)); + result.to_rvalue() + } +} + +impl<'gcc, 'tcx> HasCodegen<'tcx> for Builder<'_, 'gcc, 'tcx> { + type CodegenCx = CodegenCx<'gcc, 'tcx>; +} + +impl<'tcx> HasTyCtxt<'tcx> for Builder<'_, '_, 'tcx> { + fn tcx(&self) -> TyCtxt<'tcx> { + self.cx.tcx() + } +} + +impl HasDataLayout for Builder<'_, '_, '_> { + fn data_layout(&self) -> &TargetDataLayout { + self.cx.data_layout() + } +} + +impl<'tcx> LayoutOfHelpers<'tcx> for Builder<'_, '_, 'tcx> { + type LayoutOfResult = TyAndLayout<'tcx>; + + #[inline] + fn handle_layout_err(&self, err: LayoutError<'tcx>, span: Span, ty: Ty<'tcx>) -> ! { + self.cx.handle_layout_err(err, span, ty) + } +} + +impl<'tcx> FnAbiOfHelpers<'tcx> for Builder<'_, '_, 'tcx> { + type FnAbiOfResult = &'tcx FnAbi<'tcx, Ty<'tcx>>; + + #[inline] + fn handle_fn_abi_err( + &self, + err: FnAbiError<'tcx>, + span: Span, + fn_abi_request: FnAbiRequest<'tcx>, + ) -> ! { + self.cx.handle_fn_abi_err(err, span, fn_abi_request) + } +} + +impl<'gcc, 'tcx> Deref for Builder<'_, 'gcc, 'tcx> { + type Target = CodegenCx<'gcc, 'tcx>; + + fn deref(&self) -> &Self::Target { + self.cx + } +} + +impl<'gcc, 'tcx> BackendTypes for Builder<'_, 'gcc, 'tcx> { + type Value = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Value; + type Function = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Function; + type BasicBlock = <CodegenCx<'gcc, 'tcx> as BackendTypes>::BasicBlock; + type Type = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Type; + type Funclet = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Funclet; + + type DIScope = <CodegenCx<'gcc, 'tcx> as BackendTypes>::DIScope; + type DILocation = <CodegenCx<'gcc, 'tcx> as BackendTypes>::DILocation; + type DIVariable = <CodegenCx<'gcc, 'tcx> as BackendTypes>::DIVariable; +} + +impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { + fn build(cx: &'a CodegenCx<'gcc, 'tcx>, block: Block<'gcc>) -> Self { + Builder::with_cx(cx, block) + } + + fn llbb(&self) -> Block<'gcc> { + self.block + } + + fn append_block(cx: &'a CodegenCx<'gcc, 'tcx>, func: RValue<'gcc>, name: &str) -> Block<'gcc> { + let func = cx.rvalue_as_function(func); + func.new_block(name) + } + + fn append_sibling_block(&mut self, name: &str) -> Block<'gcc> { + let func = self.current_func(); + func.new_block(name) + } + + fn switch_to_block(&mut self, block: Self::BasicBlock) { + self.block = block; + } + + fn ret_void(&mut self) { + self.llbb().end_with_void_return(None) + } + + fn ret(&mut self, value: RValue<'gcc>) { + let value = + if self.structs_as_pointer.borrow().contains(&value) { + // NOTE: hack to workaround a limitation of the rustc API: see comment on + // CodegenCx.structs_as_pointer + value.dereference(None).to_rvalue() + } + else { + value + }; + self.llbb().end_with_return(None, value); + } + + fn br(&mut self, dest: Block<'gcc>) { + self.llbb().end_with_jump(None, dest) + } + + fn cond_br(&mut self, cond: RValue<'gcc>, then_block: Block<'gcc>, else_block: Block<'gcc>) { + self.llbb().end_with_conditional(None, cond, then_block, else_block) + } + + fn switch(&mut self, value: RValue<'gcc>, default_block: Block<'gcc>, cases: impl ExactSizeIterator<Item = (u128, Block<'gcc>)>) { + let mut gcc_cases = vec![]; + let typ = self.val_ty(value); + for (on_val, dest) in cases { + let on_val = self.const_uint_big(typ, on_val); + gcc_cases.push(self.context.new_case(on_val, on_val, dest)); + } + self.block.end_with_switch(None, value, default_block, &gcc_cases); + } + + fn invoke(&mut self, typ: Type<'gcc>, func: RValue<'gcc>, args: &[RValue<'gcc>], then: Block<'gcc>, catch: Block<'gcc>, _funclet: Option<&Funclet>) -> RValue<'gcc> { + // TODO(bjorn3): Properly implement unwinding. + let call_site = self.call(typ, func, args, None); + let condition = self.context.new_rvalue_from_int(self.bool_type, 1); + self.llbb().end_with_conditional(None, condition, then, catch); + call_site + } + + fn unreachable(&mut self) { + let func = self.context.get_builtin_function("__builtin_unreachable"); + self.block.add_eval(None, self.context.new_call(None, func, &[])); + let return_type = self.block.get_function().get_return_type(); + let void_type = self.context.new_type::<()>(); + if return_type == void_type { + self.block.end_with_void_return(None) + } + else { + let return_value = self.current_func() + .new_local(None, return_type, "unreachableReturn"); + self.block.end_with_return(None, return_value) + } + } + + fn add(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_add(a, b) + } + + fn fadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a + b + } + + fn sub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_sub(a, b) + } + + fn fsub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a - b + } + + fn mul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_mul(a, b) + } + + fn fmul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a * b + } + + fn udiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_udiv(a, b) + } + + fn exactudiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): poison if not exact. + let a_type = a.get_type().to_unsigned(self); + let a = self.gcc_int_cast(a, a_type); + let b_type = b.get_type().to_unsigned(self); + let b = self.gcc_int_cast(b, b_type); + a / b + } + + fn sdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_sdiv(a, b) + } + + fn exactsdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): poison if not exact. + // FIXME(antoyo): rustc_codegen_ssa::mir::intrinsic uses different types for a and b but they + // should be the same. + let typ = a.get_type().to_signed(self); + let b = self.context.new_cast(None, b, typ); + a / b + } + + fn fdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a / b + } + + fn urem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_urem(a, b) + } + + fn srem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_srem(a, b) + } + + fn frem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + if a.get_type().is_compatible_with(self.cx.float_type) { + let fmodf = self.context.get_builtin_function("fmodf"); + // FIXME(antoyo): this seems to produce the wrong result. + return self.context.new_call(None, fmodf, &[a, b]); + } + assert_eq!(a.get_type().unqualified(), self.cx.double_type); + + let fmod = self.context.get_builtin_function("fmod"); + return self.context.new_call(None, fmod, &[a, b]); + } + + fn shl(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_shl(a, b) + } + + fn lshr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_lshr(a, b) + } + + fn ashr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): check whether behavior is an arithmetic shift for >> . + // It seems to be if the value is signed. + self.gcc_lshr(a, b) + } + + fn and(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_and(a, b) + } + + fn or(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.cx.gcc_or(a, b) + } + + fn xor(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_xor(a, b) + } + + fn neg(&mut self, a: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_neg(a) + } + + fn fneg(&mut self, a: RValue<'gcc>) -> RValue<'gcc> { + self.cx.context.new_unary_op(None, UnaryOp::Minus, a.get_type(), a) + } + + fn not(&mut self, a: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_not(a) + } + + fn unchecked_sadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a + b + } + + fn unchecked_uadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_add(a, b) + } + + fn unchecked_ssub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a - b + } + + fn unchecked_usub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): should generate poison value? + self.gcc_sub(a, b) + } + + fn unchecked_smul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a * b + } + + fn unchecked_umul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + a * b + } + + fn fadd_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn fsub_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn fmul_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn fdiv_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn frem_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn checked_binop(&mut self, oop: OverflowOp, typ: Ty<'_>, lhs: Self::Value, rhs: Self::Value) -> (Self::Value, Self::Value) { + self.gcc_checked_binop(oop, typ, lhs, rhs) + } + + fn alloca(&mut self, ty: Type<'gcc>, align: Align) -> RValue<'gcc> { + // FIXME(antoyo): this check that we don't call get_aligned() a second time on a type. + // Ideally, we shouldn't need to do this check. + let aligned_type = + if ty == self.cx.u128_type || ty == self.cx.i128_type { + ty + } + else { + ty.get_aligned(align.bytes()) + }; + // TODO(antoyo): It might be better to return a LValue, but fixing the rustc API is non-trivial. + self.stack_var_count.set(self.stack_var_count.get() + 1); + self.current_func().new_local(None, aligned_type, &format!("stack_var_{}", self.stack_var_count.get())).get_address(None) + } + + fn dynamic_alloca(&mut self, _ty: Type<'gcc>, _align: Align) -> RValue<'gcc> { + unimplemented!(); + } + + fn array_alloca(&mut self, _ty: Type<'gcc>, _len: RValue<'gcc>, _align: Align) -> RValue<'gcc> { + unimplemented!(); + } + + fn load(&mut self, pointee_ty: Type<'gcc>, ptr: RValue<'gcc>, _align: Align) -> RValue<'gcc> { + let block = self.llbb(); + let function = block.get_function(); + // NOTE: instead of returning the dereference here, we have to assign it to a variable in + // the current basic block. Otherwise, it could be used in another basic block, causing a + // dereference after a drop, for instance. + // TODO(antoyo): handle align of the load instruction. + let ptr = self.context.new_cast(None, ptr, pointee_ty.make_pointer()); + let deref = ptr.dereference(None).to_rvalue(); + unsafe { RETURN_VALUE_COUNT += 1 }; + let loaded_value = function.new_local(None, pointee_ty, &format!("loadedValue{}", unsafe { RETURN_VALUE_COUNT })); + block.add_assignment(None, loaded_value, deref); + loaded_value.to_rvalue() + } + + fn volatile_load(&mut self, _ty: Type<'gcc>, ptr: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): use ty. + let ptr = self.context.new_cast(None, ptr, ptr.get_type().make_volatile()); + ptr.dereference(None).to_rvalue() + } + + fn atomic_load(&mut self, _ty: Type<'gcc>, ptr: RValue<'gcc>, order: AtomicOrdering, size: Size) -> RValue<'gcc> { + // TODO(antoyo): use ty. + // TODO(antoyo): handle alignment. + let atomic_load = self.context.get_builtin_function(&format!("__atomic_load_{}", size.bytes())); + let ordering = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc()); + + let volatile_const_void_ptr_type = self.context.new_type::<()>() + .make_const() + .make_volatile() + .make_pointer(); + let ptr = self.context.new_cast(None, ptr, volatile_const_void_ptr_type); + self.context.new_call(None, atomic_load, &[ptr, ordering]) + } + + fn load_operand(&mut self, place: PlaceRef<'tcx, RValue<'gcc>>) -> OperandRef<'tcx, RValue<'gcc>> { + assert_eq!(place.llextra.is_some(), place.layout.is_unsized()); + + if place.layout.is_zst() { + return OperandRef::new_zst(self, place.layout); + } + + fn scalar_load_metadata<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>, load: RValue<'gcc>, scalar: &abi::Scalar) { + let vr = scalar.valid_range(bx); + match scalar.primitive() { + abi::Int(..) => { + if !scalar.is_always_valid(bx) { + bx.range_metadata(load, vr); + } + } + abi::Pointer if vr.start < vr.end && !vr.contains(0) => { + bx.nonnull_metadata(load); + } + _ => {} + } + } + + let val = + if let Some(llextra) = place.llextra { + OperandValue::Ref(place.llval, Some(llextra), place.align) + } + else if place.layout.is_gcc_immediate() { + let load = self.load( + place.layout.gcc_type(self, false), + place.llval, + place.align, + ); + if let abi::Abi::Scalar(ref scalar) = place.layout.abi { + scalar_load_metadata(self, load, scalar); + } + OperandValue::Immediate(self.to_immediate(load, place.layout)) + } + else if let abi::Abi::ScalarPair(ref a, ref b) = place.layout.abi { + let b_offset = a.size(self).align_to(b.align(self).abi); + let pair_type = place.layout.gcc_type(self, false); + + let mut load = |i, scalar: &abi::Scalar, align| { + let llptr = self.struct_gep(pair_type, place.llval, i as u64); + let llty = place.layout.scalar_pair_element_gcc_type(self, i, false); + let load = self.load(llty, llptr, align); + scalar_load_metadata(self, load, scalar); + if scalar.is_bool() { self.trunc(load, self.type_i1()) } else { load } + }; + + OperandValue::Pair( + load(0, a, place.align), + load(1, b, place.align.restrict_for_offset(b_offset)), + ) + } + else { + OperandValue::Ref(place.llval, None, place.align) + }; + + OperandRef { val, layout: place.layout } + } + + fn write_operand_repeatedly(mut self, cg_elem: OperandRef<'tcx, RValue<'gcc>>, count: u64, dest: PlaceRef<'tcx, RValue<'gcc>>) -> Self { + let zero = self.const_usize(0); + let count = self.const_usize(count); + let start = dest.project_index(&mut self, zero).llval; + let end = dest.project_index(&mut self, count).llval; + + let header_bb = self.append_sibling_block("repeat_loop_header"); + let body_bb = self.append_sibling_block("repeat_loop_body"); + let next_bb = self.append_sibling_block("repeat_loop_next"); + + let ptr_type = start.get_type(); + let current = self.llbb().get_function().new_local(None, ptr_type, "loop_var"); + let current_val = current.to_rvalue(); + self.assign(current, start); + + self.br(header_bb); + + self.switch_to_block(header_bb); + let keep_going = self.icmp(IntPredicate::IntNE, current_val, end); + self.cond_br(keep_going, body_bb, next_bb); + + self.switch_to_block(body_bb); + let align = dest.align.restrict_for_offset(dest.layout.field(self.cx(), 0).size); + cg_elem.val.store(&mut self, PlaceRef::new_sized_aligned(current_val, cg_elem.layout, align)); + + let next = self.inbounds_gep(self.backend_type(cg_elem.layout), current.to_rvalue(), &[self.const_usize(1)]); + self.llbb().add_assignment(None, current, next); + self.br(header_bb); + + self.switch_to_block(next_bb); + self + } + + fn range_metadata(&mut self, _load: RValue<'gcc>, _range: WrappingRange) { + // TODO(antoyo) + } + + fn nonnull_metadata(&mut self, _load: RValue<'gcc>) { + // TODO(antoyo) + } + + fn store(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> { + self.store_with_flags(val, ptr, align, MemFlags::empty()) + } + + fn store_with_flags(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, align: Align, _flags: MemFlags) -> RValue<'gcc> { + let ptr = self.check_store(val, ptr); + let destination = ptr.dereference(None); + // NOTE: libgccjit does not support specifying the alignment on the assignment, so we cast + // to type so it gets the proper alignment. + let destination_type = destination.to_rvalue().get_type().unqualified(); + let aligned_type = destination_type.get_aligned(align.bytes()).make_pointer(); + let aligned_destination = self.cx.context.new_bitcast(None, ptr, aligned_type); + let aligned_destination = aligned_destination.dereference(None); + self.llbb().add_assignment(None, aligned_destination, val); + // TODO(antoyo): handle align and flags. + // NOTE: dummy value here since it's never used. FIXME(antoyo): API should not return a value here? + self.cx.context.new_rvalue_zero(self.type_i32()) + } + + fn atomic_store(&mut self, value: RValue<'gcc>, ptr: RValue<'gcc>, order: AtomicOrdering, size: Size) { + // TODO(antoyo): handle alignment. + let atomic_store = self.context.get_builtin_function(&format!("__atomic_store_{}", size.bytes())); + let ordering = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc()); + let volatile_const_void_ptr_type = self.context.new_type::<()>() + .make_volatile() + .make_pointer(); + let ptr = self.context.new_cast(None, ptr, volatile_const_void_ptr_type); + + // FIXME(antoyo): fix libgccjit to allow comparing an integer type with an aligned integer type because + // the following cast is required to avoid this error: + // gcc_jit_context_new_call: mismatching types for argument 2 of function "__atomic_store_4": assignment to param arg1 (type: int) from loadedValue3577 (type: unsigned int __attribute__((aligned(4)))) + let int_type = atomic_store.get_param(1).to_rvalue().get_type(); + let value = self.context.new_cast(None, value, int_type); + self.llbb() + .add_eval(None, self.context.new_call(None, atomic_store, &[ptr, value, ordering])); + } + + fn gep(&mut self, _typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> { + let mut result = ptr; + for index in indices { + result = self.context.new_array_access(None, result, *index).get_address(None).to_rvalue(); + } + result + } + + fn inbounds_gep(&mut self, _typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> { + // FIXME(antoyo): would be safer if doing the same thing (loop) as gep. + // TODO(antoyo): specify inbounds somehow. + match indices.len() { + 1 => { + self.context.new_array_access(None, ptr, indices[0]).get_address(None) + }, + 2 => { + let array = ptr.dereference(None); // TODO(antoyo): assert that first index is 0? + self.context.new_array_access(None, array, indices[1]).get_address(None) + }, + _ => unimplemented!(), + } + } + + fn struct_gep(&mut self, value_type: Type<'gcc>, ptr: RValue<'gcc>, idx: u64) -> RValue<'gcc> { + // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays. + assert_eq!(idx as usize as u64, idx); + let value = ptr.dereference(None).to_rvalue(); + + if value_type.dyncast_array().is_some() { + let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from")); + let element = self.context.new_array_access(None, value, index); + element.get_address(None) + } + else if let Some(vector_type) = value_type.dyncast_vector() { + let array_type = vector_type.get_element_type().make_pointer(); + let array = self.bitcast(ptr, array_type); + let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from")); + let element = self.context.new_array_access(None, array, index); + element.get_address(None) + } + else if let Some(struct_type) = value_type.is_struct() { + ptr.dereference_field(None, struct_type.get_field(idx as i32)).get_address(None) + } + else { + panic!("Unexpected type {:?}", value_type); + } + } + + /* Casts */ + fn trunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): check that it indeed truncate the value. + self.gcc_int_cast(value, dest_ty) + } + + fn sext(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): check that it indeed sign extend the value. + if dest_ty.dyncast_vector().is_some() { + // TODO(antoyo): nothing to do as it is only for LLVM? + return value; + } + self.context.new_cast(None, value, dest_ty) + } + + fn fptoui(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + self.gcc_float_to_uint_cast(value, dest_ty) + } + + fn fptosi(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + self.gcc_float_to_int_cast(value, dest_ty) + } + + fn uitofp(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + self.gcc_uint_to_float_cast(value, dest_ty) + } + + fn sitofp(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + self.gcc_int_to_float_cast(value, dest_ty) + } + + fn fptrunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): make sure it truncates. + self.context.new_cast(None, value, dest_ty) + } + + fn fpext(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + self.context.new_cast(None, value, dest_ty) + } + + fn ptrtoint(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + let usize_value = self.cx.const_bitcast(value, self.cx.type_isize()); + self.intcast(usize_value, dest_ty, false) + } + + fn inttoptr(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + let usize_value = self.intcast(value, self.cx.type_isize(), false); + self.cx.const_bitcast(usize_value, dest_ty) + } + + fn bitcast(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + self.cx.const_bitcast(value, dest_ty) + } + + fn intcast(&mut self, value: RValue<'gcc>, dest_typ: Type<'gcc>, _is_signed: bool) -> RValue<'gcc> { + // NOTE: is_signed is for value, not dest_typ. + self.gcc_int_cast(value, dest_typ) + } + + fn pointercast(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> { + let val_type = value.get_type(); + match (type_is_pointer(val_type), type_is_pointer(dest_ty)) { + (false, true) => { + // NOTE: Projecting a field of a pointer type will attempt a cast from a signed char to + // a pointer, which is not supported by gccjit. + return self.cx.context.new_cast(None, self.inttoptr(value, val_type.make_pointer()), dest_ty); + }, + (false, false) => { + // When they are not pointers, we want a transmute (or reinterpret_cast). + self.bitcast(value, dest_ty) + }, + (true, true) => self.cx.context.new_cast(None, value, dest_ty), + (true, false) => unimplemented!(), + } + } + + /* Comparisons */ + fn icmp(&mut self, op: IntPredicate, lhs: RValue<'gcc>, rhs: RValue<'gcc>) -> RValue<'gcc> { + self.gcc_icmp(op, lhs, rhs) + } + + fn fcmp(&mut self, op: RealPredicate, lhs: RValue<'gcc>, rhs: RValue<'gcc>) -> RValue<'gcc> { + self.context.new_comparison(None, op.to_gcc_comparison(), lhs, rhs) + } + + /* Miscellaneous instructions */ + fn memcpy(&mut self, dst: RValue<'gcc>, _dst_align: Align, src: RValue<'gcc>, _src_align: Align, size: RValue<'gcc>, flags: MemFlags) { + assert!(!flags.contains(MemFlags::NONTEMPORAL), "non-temporal memcpy not supported"); + let size = self.intcast(size, self.type_size_t(), false); + let _is_volatile = flags.contains(MemFlags::VOLATILE); + let dst = self.pointercast(dst, self.type_i8p()); + let src = self.pointercast(src, self.type_ptr_to(self.type_void())); + let memcpy = self.context.get_builtin_function("memcpy"); + // TODO(antoyo): handle aligns and is_volatile. + self.block.add_eval(None, self.context.new_call(None, memcpy, &[dst, src, size])); + } + + fn memmove(&mut self, dst: RValue<'gcc>, dst_align: Align, src: RValue<'gcc>, src_align: Align, size: RValue<'gcc>, flags: MemFlags) { + if flags.contains(MemFlags::NONTEMPORAL) { + // HACK(nox): This is inefficient but there is no nontemporal memmove. + let val = self.load(src.get_type().get_pointee().expect("get_pointee"), src, src_align); + let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val))); + self.store_with_flags(val, ptr, dst_align, flags); + return; + } + let size = self.intcast(size, self.type_size_t(), false); + let _is_volatile = flags.contains(MemFlags::VOLATILE); + let dst = self.pointercast(dst, self.type_i8p()); + let src = self.pointercast(src, self.type_ptr_to(self.type_void())); + + let memmove = self.context.get_builtin_function("memmove"); + // TODO(antoyo): handle is_volatile. + self.block.add_eval(None, self.context.new_call(None, memmove, &[dst, src, size])); + } + + fn memset(&mut self, ptr: RValue<'gcc>, fill_byte: RValue<'gcc>, size: RValue<'gcc>, _align: Align, flags: MemFlags) { + let _is_volatile = flags.contains(MemFlags::VOLATILE); + let ptr = self.pointercast(ptr, self.type_i8p()); + let memset = self.context.get_builtin_function("memset"); + // TODO(antoyo): handle align and is_volatile. + let fill_byte = self.context.new_cast(None, fill_byte, self.i32_type); + let size = self.intcast(size, self.type_size_t(), false); + self.block.add_eval(None, self.context.new_call(None, memset, &[ptr, fill_byte, size])); + } + + fn select(&mut self, cond: RValue<'gcc>, then_val: RValue<'gcc>, mut else_val: RValue<'gcc>) -> RValue<'gcc> { + let func = self.current_func(); + let variable = func.new_local(None, then_val.get_type(), "selectVar"); + let then_block = func.new_block("then"); + let else_block = func.new_block("else"); + let after_block = func.new_block("after"); + self.llbb().end_with_conditional(None, cond, then_block, else_block); + + then_block.add_assignment(None, variable, then_val); + then_block.end_with_jump(None, after_block); + + if !then_val.get_type().is_compatible_with(else_val.get_type()) { + else_val = self.context.new_cast(None, else_val, then_val.get_type()); + } + else_block.add_assignment(None, variable, else_val); + else_block.end_with_jump(None, after_block); + + // NOTE: since jumps were added in a place rustc does not expect, the current block in the + // state need to be updated. + self.switch_to_block(after_block); + + variable.to_rvalue() + } + + #[allow(dead_code)] + fn va_arg(&mut self, _list: RValue<'gcc>, _ty: Type<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn extract_element(&mut self, _vec: RValue<'gcc>, _idx: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn vector_splat(&mut self, _num_elts: usize, _elt: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn extract_value(&mut self, aggregate_value: RValue<'gcc>, idx: u64) -> RValue<'gcc> { + // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays. + assert_eq!(idx as usize as u64, idx); + let value_type = aggregate_value.get_type(); + + if value_type.dyncast_array().is_some() { + let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from")); + let element = self.context.new_array_access(None, aggregate_value, index); + element.get_address(None) + } + else if value_type.dyncast_vector().is_some() { + panic!(); + } + else if let Some(pointer_type) = value_type.get_pointee() { + if let Some(struct_type) = pointer_type.is_struct() { + // NOTE: hack to workaround a limitation of the rustc API: see comment on + // CodegenCx.structs_as_pointer + aggregate_value.dereference_field(None, struct_type.get_field(idx as i32)).to_rvalue() + } + else { + panic!("Unexpected type {:?}", value_type); + } + } + else if let Some(struct_type) = value_type.is_struct() { + aggregate_value.access_field(None, struct_type.get_field(idx as i32)).to_rvalue() + } + else { + panic!("Unexpected type {:?}", value_type); + } + } + + fn insert_value(&mut self, aggregate_value: RValue<'gcc>, value: RValue<'gcc>, idx: u64) -> RValue<'gcc> { + // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays. + assert_eq!(idx as usize as u64, idx); + let value_type = aggregate_value.get_type(); + + let lvalue = + if value_type.dyncast_array().is_some() { + let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from")); + self.context.new_array_access(None, aggregate_value, index) + } + else if value_type.dyncast_vector().is_some() { + panic!(); + } + else if let Some(pointer_type) = value_type.get_pointee() { + if let Some(struct_type) = pointer_type.is_struct() { + // NOTE: hack to workaround a limitation of the rustc API: see comment on + // CodegenCx.structs_as_pointer + aggregate_value.dereference_field(None, struct_type.get_field(idx as i32)) + } + else { + panic!("Unexpected type {:?}", value_type); + } + } + else { + panic!("Unexpected type {:?}", value_type); + }; + + let lvalue_type = lvalue.to_rvalue().get_type(); + let value = + // NOTE: sometimes, rustc will create a value with the wrong type. + if lvalue_type != value.get_type() { + self.context.new_cast(None, value, lvalue_type) + } + else { + value + }; + + self.llbb().add_assignment(None, lvalue, value); + + aggregate_value + } + + fn set_personality_fn(&mut self, _personality: RValue<'gcc>) { + // TODO(antoyo) + } + + fn cleanup_landing_pad(&mut self, _ty: Type<'gcc>, _pers_fn: RValue<'gcc>) -> RValue<'gcc> { + let field1 = self.context.new_field(None, self.u8_type.make_pointer(), "landing_pad_field_1"); + let field2 = self.context.new_field(None, self.i32_type, "landing_pad_field_1"); + let struct_type = self.context.new_struct_type(None, "landing_pad", &[field1, field2]); + self.current_func().new_local(None, struct_type.as_type(), "landing_pad") + .to_rvalue() + // TODO(antoyo): Properly implement unwinding. + // the above is just to make the compilation work as it seems + // rustc_codegen_ssa now calls the unwinding builder methods even on panic=abort. + } + + fn resume(&mut self, _exn: RValue<'gcc>) { + // TODO(bjorn3): Properly implement unwinding. + self.unreachable(); + } + + fn cleanup_pad(&mut self, _parent: Option<RValue<'gcc>>, _args: &[RValue<'gcc>]) -> Funclet { + unimplemented!(); + } + + fn cleanup_ret(&mut self, _funclet: &Funclet, _unwind: Option<Block<'gcc>>) { + unimplemented!(); + } + + fn catch_pad(&mut self, _parent: RValue<'gcc>, _args: &[RValue<'gcc>]) -> Funclet { + unimplemented!(); + } + + fn catch_switch( + &mut self, + _parent: Option<RValue<'gcc>>, + _unwind: Option<Block<'gcc>>, + _handlers: &[Block<'gcc>], + ) -> RValue<'gcc> { + unimplemented!(); + } + + // Atomic Operations + fn atomic_cmpxchg(&mut self, dst: RValue<'gcc>, cmp: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering, failure_order: AtomicOrdering, weak: bool) -> RValue<'gcc> { + let expected = self.current_func().new_local(None, cmp.get_type(), "expected"); + self.llbb().add_assignment(None, expected, cmp); + let success = self.compare_exchange(dst, expected, src, order, failure_order, weak); + + let pair_type = self.cx.type_struct(&[src.get_type(), self.bool_type], false); + let result = self.current_func().new_local(None, pair_type, "atomic_cmpxchg_result"); + let align = Align::from_bits(64).expect("align"); // TODO(antoyo): use good align. + + let value_type = result.to_rvalue().get_type(); + if let Some(struct_type) = value_type.is_struct() { + self.store(success, result.access_field(None, struct_type.get_field(1)).get_address(None), align); + // NOTE: since success contains the call to the intrinsic, it must be stored before + // expected so that we store expected after the call. + self.store(expected.to_rvalue(), result.access_field(None, struct_type.get_field(0)).get_address(None), align); + } + // TODO(antoyo): handle when value is not a struct. + + result.to_rvalue() + } + + fn atomic_rmw(&mut self, op: AtomicRmwBinOp, dst: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering) -> RValue<'gcc> { + let size = src.get_type().get_size(); + let name = + match op { + AtomicRmwBinOp::AtomicXchg => format!("__atomic_exchange_{}", size), + AtomicRmwBinOp::AtomicAdd => format!("__atomic_fetch_add_{}", size), + AtomicRmwBinOp::AtomicSub => format!("__atomic_fetch_sub_{}", size), + AtomicRmwBinOp::AtomicAnd => format!("__atomic_fetch_and_{}", size), + AtomicRmwBinOp::AtomicNand => format!("__atomic_fetch_nand_{}", size), + AtomicRmwBinOp::AtomicOr => format!("__atomic_fetch_or_{}", size), + AtomicRmwBinOp::AtomicXor => format!("__atomic_fetch_xor_{}", size), + AtomicRmwBinOp::AtomicMax => return self.atomic_extremum(ExtremumOperation::Max, dst, src, order), + AtomicRmwBinOp::AtomicMin => return self.atomic_extremum(ExtremumOperation::Min, dst, src, order), + AtomicRmwBinOp::AtomicUMax => return self.atomic_extremum(ExtremumOperation::Max, dst, src, order), + AtomicRmwBinOp::AtomicUMin => return self.atomic_extremum(ExtremumOperation::Min, dst, src, order), + }; + + + let atomic_function = self.context.get_builtin_function(name); + let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc()); + + let void_ptr_type = self.context.new_type::<*mut ()>(); + let volatile_void_ptr_type = void_ptr_type.make_volatile(); + let dst = self.context.new_cast(None, dst, volatile_void_ptr_type); + // FIXME(antoyo): not sure why, but we have the wrong type here. + let new_src_type = atomic_function.get_param(1).to_rvalue().get_type(); + let src = self.context.new_cast(None, src, new_src_type); + let res = self.context.new_call(None, atomic_function, &[dst, src, order]); + self.context.new_cast(None, res, src.get_type()) + } + + fn atomic_fence(&mut self, order: AtomicOrdering, scope: SynchronizationScope) { + let name = + match scope { + SynchronizationScope::SingleThread => "__atomic_signal_fence", + SynchronizationScope::CrossThread => "__atomic_thread_fence", + }; + let thread_fence = self.context.get_builtin_function(name); + let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc()); + self.llbb().add_eval(None, self.context.new_call(None, thread_fence, &[order])); + } + + fn set_invariant_load(&mut self, load: RValue<'gcc>) { + // NOTE: Hack to consider vtable function pointer as non-global-variable function pointer. + self.normal_function_addresses.borrow_mut().insert(load); + // TODO(antoyo) + } + + fn lifetime_start(&mut self, _ptr: RValue<'gcc>, _size: Size) { + // TODO(antoyo) + } + + fn lifetime_end(&mut self, _ptr: RValue<'gcc>, _size: Size) { + // TODO(antoyo) + } + + fn call(&mut self, _typ: Type<'gcc>, func: RValue<'gcc>, args: &[RValue<'gcc>], funclet: Option<&Funclet>) -> RValue<'gcc> { + // FIXME(antoyo): remove when having a proper API. + let gcc_func = unsafe { std::mem::transmute(func) }; + if self.functions.borrow().values().find(|value| **value == gcc_func).is_some() { + self.function_call(func, args, funclet) + } + else { + // If it's a not function that was defined, it's a function pointer. + self.function_ptr_call(func, args, funclet) + } + } + + fn zext(&mut self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + // FIXME(antoyo): this does not zero-extend. + if value.get_type().is_bool() && dest_typ.is_i8(&self.cx) { + // FIXME(antoyo): hack because base::from_immediate converts i1 to i8. + // Fix the code in codegen_ssa::base::from_immediate. + return value; + } + self.gcc_int_cast(value, dest_typ) + } + + fn cx(&self) -> &CodegenCx<'gcc, 'tcx> { + self.cx + } + + fn do_not_inline(&mut self, _llret: RValue<'gcc>) { + // FIXME(bjorn3): implement + } + + fn set_span(&mut self, _span: Span) {} + + fn from_immediate(&mut self, val: Self::Value) -> Self::Value { + if self.cx().val_ty(val) == self.cx().type_i1() { + self.zext(val, self.cx().type_i8()) + } + else { + val + } + } + + fn to_immediate_scalar(&mut self, val: Self::Value, scalar: abi::Scalar) -> Self::Value { + if scalar.is_bool() { + return self.trunc(val, self.cx().type_i1()); + } + val + } + + fn fptoui_sat(&mut self, _val: RValue<'gcc>, _dest_ty: Type<'gcc>) -> Option<RValue<'gcc>> { + None + } + + fn fptosi_sat(&mut self, _val: RValue<'gcc>, _dest_ty: Type<'gcc>) -> Option<RValue<'gcc>> { + None + } + + fn instrprof_increment(&mut self, _fn_name: RValue<'gcc>, _hash: RValue<'gcc>, _num_counters: RValue<'gcc>, _index: RValue<'gcc>) { + unimplemented!(); + } +} + +impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { + #[cfg(feature="master")] + pub fn shuffle_vector(&mut self, v1: RValue<'gcc>, v2: RValue<'gcc>, mask: RValue<'gcc>) -> RValue<'gcc> { + let struct_type = mask.get_type().is_struct().expect("mask of struct type"); + + // TODO(antoyo): use a recursive unqualified() here. + let vector_type = v1.get_type().unqualified().dyncast_vector().expect("vector type"); + let element_type = vector_type.get_element_type(); + let vec_num_units = vector_type.get_num_units(); + + let mask_num_units = struct_type.get_field_count(); + let mut vector_elements = vec![]; + let mask_element_type = + if element_type.is_integral() { + element_type + } + else { + #[cfg(feature="master")] + { + self.cx.type_ix(element_type.get_size() as u64 * 8) + } + #[cfg(not(feature="master"))] + self.int_type + }; + for i in 0..mask_num_units { + let field = struct_type.get_field(i as i32); + vector_elements.push(self.context.new_cast(None, mask.access_field(None, field).to_rvalue(), mask_element_type)); + } + + // NOTE: the mask needs to be the same length as the input vectors, so add the missing + // elements in the mask if needed. + for _ in mask_num_units..vec_num_units { + vector_elements.push(self.context.new_rvalue_zero(mask_element_type)); + } + + let array_type = self.context.new_array_type(None, element_type, vec_num_units as i32); + let result_type = self.context.new_vector_type(element_type, mask_num_units as u64); + let (v1, v2) = + if vec_num_units < mask_num_units { + // NOTE: the mask needs to be the same length as the input vectors, so join the 2 + // vectors and create a dummy second vector. + // TODO(antoyo): switch to using new_vector_access. + let array = self.context.new_bitcast(None, v1, array_type); + let mut elements = vec![]; + for i in 0..vec_num_units { + elements.push(self.context.new_array_access(None, array, self.context.new_rvalue_from_int(self.int_type, i as i32)).to_rvalue()); + } + // TODO(antoyo): switch to using new_vector_access. + let array = self.context.new_bitcast(None, v2, array_type); + for i in 0..(mask_num_units - vec_num_units) { + elements.push(self.context.new_array_access(None, array, self.context.new_rvalue_from_int(self.int_type, i as i32)).to_rvalue()); + } + let v1 = self.context.new_rvalue_from_vector(None, result_type, &elements); + let zero = self.context.new_rvalue_zero(element_type); + let v2 = self.context.new_rvalue_from_vector(None, result_type, &vec![zero; mask_num_units]); + (v1, v2) + } + else { + (v1, v2) + }; + + let new_mask_num_units = std::cmp::max(mask_num_units, vec_num_units); + let mask_type = self.context.new_vector_type(mask_element_type, new_mask_num_units as u64); + let mask = self.context.new_rvalue_from_vector(None, mask_type, &vector_elements); + let result = self.context.new_rvalue_vector_perm(None, v1, v2, mask); + + if vec_num_units != mask_num_units { + // NOTE: if padding was added, only select the number of elements of the masks to + // remove that padding in the result. + let mut elements = vec![]; + // TODO(antoyo): switch to using new_vector_access. + let array = self.context.new_bitcast(None, result, array_type); + for i in 0..mask_num_units { + elements.push(self.context.new_array_access(None, array, self.context.new_rvalue_from_int(self.int_type, i as i32)).to_rvalue()); + } + self.context.new_rvalue_from_vector(None, result_type, &elements) + } + else { + result + } + } + + #[cfg(not(feature="master"))] + pub fn shuffle_vector(&mut self, _v1: RValue<'gcc>, _v2: RValue<'gcc>, _mask: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + #[cfg(feature="master")] + pub fn vector_reduce<F>(&mut self, src: RValue<'gcc>, op: F) -> RValue<'gcc> + where F: Fn(RValue<'gcc>, RValue<'gcc>, &'gcc Context<'gcc>) -> RValue<'gcc> + { + let vector_type = src.get_type().unqualified().dyncast_vector().expect("vector type"); + let element_count = vector_type.get_num_units(); + let mut vector_elements = vec![]; + for i in 0..element_count { + vector_elements.push(i); + } + let mask_type = self.context.new_vector_type(self.int_type, element_count as u64); + let mut shift = 1; + let mut res = src; + while shift < element_count { + let vector_elements: Vec<_> = + vector_elements.iter() + .map(|i| self.context.new_rvalue_from_int(self.int_type, ((i + shift) % element_count) as i32)) + .collect(); + let mask = self.context.new_rvalue_from_vector(None, mask_type, &vector_elements); + let shifted = self.context.new_rvalue_vector_perm(None, res, res, mask); + shift *= 2; + res = op(res, shifted, &self.context); + } + self.context.new_vector_access(None, res, self.context.new_rvalue_zero(self.int_type)) + .to_rvalue() + } + + #[cfg(not(feature="master"))] + pub fn vector_reduce<F>(&mut self, src: RValue<'gcc>, op: F) -> RValue<'gcc> + where F: Fn(RValue<'gcc>, RValue<'gcc>, &'gcc Context<'gcc>) -> RValue<'gcc> + { + unimplemented!(); + } + + pub fn vector_reduce_op(&mut self, src: RValue<'gcc>, op: BinaryOp) -> RValue<'gcc> { + self.vector_reduce(src, |a, b, context| context.new_binary_op(None, op, a.get_type(), a, b)) + } + + pub fn vector_reduce_fadd_fast(&mut self, _acc: RValue<'gcc>, _src: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + pub fn vector_reduce_fmul_fast(&mut self, _acc: RValue<'gcc>, _src: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + // Inspired by Hacker's Delight min implementation. + pub fn vector_reduce_min(&mut self, src: RValue<'gcc>) -> RValue<'gcc> { + self.vector_reduce(src, |a, b, context| { + let differences_or_zeros = difference_or_zero(a, b, context); + context.new_binary_op(None, BinaryOp::Minus, a.get_type(), a, differences_or_zeros) + }) + } + + // Inspired by Hacker's Delight max implementation. + pub fn vector_reduce_max(&mut self, src: RValue<'gcc>) -> RValue<'gcc> { + self.vector_reduce(src, |a, b, context| { + let differences_or_zeros = difference_or_zero(a, b, context); + context.new_binary_op(None, BinaryOp::Plus, b.get_type(), b, differences_or_zeros) + }) + } + + pub fn vector_select(&mut self, cond: RValue<'gcc>, then_val: RValue<'gcc>, else_val: RValue<'gcc>) -> RValue<'gcc> { + // cond is a vector of integers, not of bools. + let cond_type = cond.get_type(); + let vector_type = cond_type.unqualified().dyncast_vector().expect("vector type"); + let num_units = vector_type.get_num_units(); + let element_type = vector_type.get_element_type(); + let zeros = vec![self.context.new_rvalue_zero(element_type); num_units]; + let zeros = self.context.new_rvalue_from_vector(None, cond_type, &zeros); + + let masks = self.context.new_comparison(None, ComparisonOp::NotEquals, cond, zeros); + let then_vals = masks & then_val; + + let ones = vec![self.context.new_rvalue_one(element_type); num_units]; + let ones = self.context.new_rvalue_from_vector(None, cond_type, &ones); + let inverted_masks = masks + ones; + // NOTE: sometimes, the type of else_val can be different than the type of then_val in + // libgccjit (vector of int vs vector of int32_t), but they should be the same for the AND + // operation to work. + let else_val = self.context.new_bitcast(None, else_val, then_val.get_type()); + let else_vals = inverted_masks & else_val; + + then_vals | else_vals + } +} + +fn difference_or_zero<'gcc>(a: RValue<'gcc>, b: RValue<'gcc>, context: &'gcc Context<'gcc>) -> RValue<'gcc> { + let difference = a - b; + let masks = context.new_comparison(None, ComparisonOp::GreaterThanEquals, b, a); + difference & masks +} + +impl<'a, 'gcc, 'tcx> StaticBuilderMethods for Builder<'a, 'gcc, 'tcx> { + fn get_static(&mut self, def_id: DefId) -> RValue<'gcc> { + // Forward to the `get_static` method of `CodegenCx` + self.cx().get_static(def_id).get_address(None) + } +} + +impl<'tcx> HasParamEnv<'tcx> for Builder<'_, '_, 'tcx> { + fn param_env(&self) -> ParamEnv<'tcx> { + self.cx.param_env() + } +} + +impl<'tcx> HasTargetSpec for Builder<'_, '_, 'tcx> { + fn target_spec(&self) -> &Target { + &self.cx.target_spec() + } +} + +pub trait ToGccComp { + fn to_gcc_comparison(&self) -> ComparisonOp; +} + +impl ToGccComp for IntPredicate { + fn to_gcc_comparison(&self) -> ComparisonOp { + match *self { + IntPredicate::IntEQ => ComparisonOp::Equals, + IntPredicate::IntNE => ComparisonOp::NotEquals, + IntPredicate::IntUGT => ComparisonOp::GreaterThan, + IntPredicate::IntUGE => ComparisonOp::GreaterThanEquals, + IntPredicate::IntULT => ComparisonOp::LessThan, + IntPredicate::IntULE => ComparisonOp::LessThanEquals, + IntPredicate::IntSGT => ComparisonOp::GreaterThan, + IntPredicate::IntSGE => ComparisonOp::GreaterThanEquals, + IntPredicate::IntSLT => ComparisonOp::LessThan, + IntPredicate::IntSLE => ComparisonOp::LessThanEquals, + } + } +} + +impl ToGccComp for RealPredicate { + fn to_gcc_comparison(&self) -> ComparisonOp { + // TODO(antoyo): check that ordered vs non-ordered is respected. + match *self { + RealPredicate::RealPredicateFalse => unreachable!(), + RealPredicate::RealOEQ => ComparisonOp::Equals, + RealPredicate::RealOGT => ComparisonOp::GreaterThan, + RealPredicate::RealOGE => ComparisonOp::GreaterThanEquals, + RealPredicate::RealOLT => ComparisonOp::LessThan, + RealPredicate::RealOLE => ComparisonOp::LessThanEquals, + RealPredicate::RealONE => ComparisonOp::NotEquals, + RealPredicate::RealORD => unreachable!(), + RealPredicate::RealUNO => unreachable!(), + RealPredicate::RealUEQ => ComparisonOp::Equals, + RealPredicate::RealUGT => ComparisonOp::GreaterThan, + RealPredicate::RealUGE => ComparisonOp::GreaterThan, + RealPredicate::RealULT => ComparisonOp::LessThan, + RealPredicate::RealULE => ComparisonOp::LessThan, + RealPredicate::RealUNE => ComparisonOp::NotEquals, + RealPredicate::RealPredicateTrue => unreachable!(), + } + } +} + +#[repr(C)] +#[allow(non_camel_case_types)] +enum MemOrdering { + __ATOMIC_RELAXED, + __ATOMIC_CONSUME, + __ATOMIC_ACQUIRE, + __ATOMIC_RELEASE, + __ATOMIC_ACQ_REL, + __ATOMIC_SEQ_CST, +} + +trait ToGccOrdering { + fn to_gcc(self) -> i32; +} + +impl ToGccOrdering for AtomicOrdering { + fn to_gcc(self) -> i32 { + use MemOrdering::*; + + let ordering = + match self { + AtomicOrdering::Unordered => __ATOMIC_RELAXED, + AtomicOrdering::Relaxed => __ATOMIC_RELAXED, // TODO(antoyo): check if that's the same. + AtomicOrdering::Acquire => __ATOMIC_ACQUIRE, + AtomicOrdering::Release => __ATOMIC_RELEASE, + AtomicOrdering::AcquireRelease => __ATOMIC_ACQ_REL, + AtomicOrdering::SequentiallyConsistent => __ATOMIC_SEQ_CST, + }; + ordering as i32 + } +} diff --git a/compiler/rustc_codegen_gcc/src/callee.rs b/compiler/rustc_codegen_gcc/src/callee.rs new file mode 100644 index 000000000..c1041125e --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/callee.rs @@ -0,0 +1,77 @@ +use gccjit::{FunctionType, RValue}; +use rustc_codegen_ssa::traits::BaseTypeMethods; +use rustc_middle::ty::{self, Instance, TypeVisitable}; +use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt}; + +use crate::abi::FnAbiGccExt; +use crate::context::CodegenCx; + +/// Codegens a reference to a fn/method item, monomorphizing and +/// inlining as it goes. +/// +/// # Parameters +/// +/// - `cx`: the crate context +/// - `instance`: the instance to be instantiated +pub fn get_fn<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, instance: Instance<'tcx>) -> RValue<'gcc> { + let tcx = cx.tcx(); + + assert!(!instance.substs.needs_infer()); + assert!(!instance.substs.has_escaping_bound_vars()); + + if let Some(&func) = cx.function_instances.borrow().get(&instance) { + return func; + } + + let sym = tcx.symbol_name(instance).name; + + let fn_abi = cx.fn_abi_of_instance(instance, ty::List::empty()); + + let func = + if let Some(func) = cx.get_declared_value(&sym) { + // Create a fn pointer with the new signature. + let ptrty = fn_abi.ptr_to_gcc_type(cx); + + // This is subtle and surprising, but sometimes we have to bitcast + // the resulting fn pointer. The reason has to do with external + // functions. If you have two crates that both bind the same C + // library, they may not use precisely the same types: for + // example, they will probably each declare their own structs, + // which are distinct types from LLVM's point of view (nominal + // types). + // + // Now, if those two crates are linked into an application, and + // they contain inlined code, you can wind up with a situation + // where both of those functions wind up being loaded into this + // application simultaneously. In that case, the same function + // (from LLVM's point of view) requires two types. But of course + // LLVM won't allow one function to have two types. + // + // What we currently do, therefore, is declare the function with + // one of the two types (whichever happens to come first) and then + // bitcast as needed when the function is referenced to make sure + // it has the type we expect. + // + // This can occur on either a crate-local or crate-external + // reference. It also occurs when testing libcore and in some + // other weird situations. Annoying. + if cx.val_ty(func) != ptrty { + // TODO(antoyo): cast the pointer. + func + } + else { + func + } + } + else { + cx.linkage.set(FunctionType::Extern); + let func = cx.declare_fn(&sym, &fn_abi); + + // TODO(antoyo): set linkage and attributes. + func + }; + + cx.function_instances.borrow_mut().insert(instance, func); + + func +} diff --git a/compiler/rustc_codegen_gcc/src/common.rs b/compiler/rustc_codegen_gcc/src/common.rs new file mode 100644 index 000000000..ccb6cbbc2 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/common.rs @@ -0,0 +1,479 @@ +use gccjit::LValue; +use gccjit::{RValue, Type, ToRValue}; +use rustc_codegen_ssa::mir::place::PlaceRef; +use rustc_codegen_ssa::traits::{ + BaseTypeMethods, + ConstMethods, + DerivedTypeMethods, + MiscMethods, + StaticMethods, +}; +use rustc_middle::mir::Mutability; +use rustc_middle::ty::layout::{TyAndLayout, LayoutOf}; +use rustc_middle::mir::interpret::{ConstAllocation, GlobalAlloc, Scalar}; +use rustc_target::abi::{self, HasDataLayout, Pointer, Size}; + +use crate::consts::const_alloc_to_gcc; +use crate::context::CodegenCx; +use crate::type_of::LayoutGccExt; + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn const_bytes(&self, bytes: &[u8]) -> RValue<'gcc> { + bytes_in_context(self, bytes) + } + + fn global_string(&self, string: &str) -> LValue<'gcc> { + // TODO(antoyo): handle non-null-terminated strings. + let string = self.context.new_string_literal(&*string); + let sym = self.generate_local_symbol_name("str"); + let global = self.declare_private_global(&sym, self.val_ty(string)); + global.global_set_initializer_rvalue(string); + global + // TODO(antoyo): set linkage. + } +} + +pub fn bytes_in_context<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, bytes: &[u8]) -> RValue<'gcc> { + let context = &cx.context; + let byte_type = context.new_type::<u8>(); + let typ = context.new_array_type(None, byte_type, bytes.len() as i32); + let elements: Vec<_> = + bytes.iter() + .map(|&byte| context.new_rvalue_from_int(byte_type, byte as i32)) + .collect(); + context.new_array_constructor(None, typ, &elements) +} + +pub fn type_is_pointer<'gcc>(typ: Type<'gcc>) -> bool { + typ.get_pointee().is_some() +} + +impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn const_null(&self, typ: Type<'gcc>) -> RValue<'gcc> { + if type_is_pointer(typ) { + self.context.new_null(typ) + } + else { + self.const_int(typ, 0) + } + } + + fn const_undef(&self, typ: Type<'gcc>) -> RValue<'gcc> { + let local = self.current_func.borrow().expect("func") + .new_local(None, typ, "undefined"); + if typ.is_struct().is_some() { + // NOTE: hack to workaround a limitation of the rustc API: see comment on + // CodegenCx.structs_as_pointer + let pointer = local.get_address(None); + self.structs_as_pointer.borrow_mut().insert(pointer); + pointer + } + else { + local.to_rvalue() + } + } + + fn const_int(&self, typ: Type<'gcc>, int: i64) -> RValue<'gcc> { + self.gcc_int(typ, int) + } + + fn const_uint(&self, typ: Type<'gcc>, int: u64) -> RValue<'gcc> { + self.gcc_uint(typ, int) + } + + fn const_uint_big(&self, typ: Type<'gcc>, num: u128) -> RValue<'gcc> { + self.gcc_uint_big(typ, num) + } + + fn const_bool(&self, val: bool) -> RValue<'gcc> { + self.const_uint(self.type_i1(), val as u64) + } + + fn const_i16(&self, i: i16) -> RValue<'gcc> { + self.const_int(self.type_i16(), i as i64) + } + + fn const_i32(&self, i: i32) -> RValue<'gcc> { + self.const_int(self.type_i32(), i as i64) + } + + fn const_u32(&self, i: u32) -> RValue<'gcc> { + self.const_uint(self.type_u32(), i as u64) + } + + fn const_u64(&self, i: u64) -> RValue<'gcc> { + self.const_uint(self.type_u64(), i) + } + + fn const_usize(&self, i: u64) -> RValue<'gcc> { + let bit_size = self.data_layout().pointer_size.bits(); + if bit_size < 64 { + // make sure it doesn't overflow + assert!(i < (1 << bit_size)); + } + + self.const_uint(self.usize_type, i) + } + + fn const_u8(&self, _i: u8) -> RValue<'gcc> { + unimplemented!(); + } + + fn const_real(&self, typ: Type<'gcc>, val: f64) -> RValue<'gcc> { + self.context.new_rvalue_from_double(typ, val) + } + + fn const_str(&self, s: &str) -> (RValue<'gcc>, RValue<'gcc>) { + let str_global = *self + .const_str_cache + .borrow_mut() + .raw_entry_mut() + .from_key(s) + .or_insert_with(|| (s.to_owned(), self.global_string(s))) + .1; + let len = s.len(); + let cs = self.const_ptrcast(str_global.get_address(None), + self.type_ptr_to(self.layout_of(self.tcx.types.str_).gcc_type(self, true)), + ); + (cs, self.const_usize(len as u64)) + } + + fn const_struct(&self, values: &[RValue<'gcc>], packed: bool) -> RValue<'gcc> { + let fields: Vec<_> = values.iter() + .map(|value| value.get_type()) + .collect(); + // TODO(antoyo): cache the type? It's anonymous, so probably not. + let typ = self.type_struct(&fields, packed); + let struct_type = typ.is_struct().expect("struct type"); + self.context.new_struct_constructor(None, struct_type.as_type(), None, values) + } + + fn const_to_opt_uint(&self, _v: RValue<'gcc>) -> Option<u64> { + // TODO(antoyo) + None + } + + fn const_to_opt_u128(&self, _v: RValue<'gcc>, _sign_ext: bool) -> Option<u128> { + // TODO(antoyo) + None + } + + fn zst_to_backend(&self, _ty: Type<'gcc>) -> RValue<'gcc> { + self.const_undef(self.type_ix(0)) + } + + fn scalar_to_backend(&self, cv: Scalar, layout: abi::Scalar, ty: Type<'gcc>) -> RValue<'gcc> { + let bitsize = if layout.is_bool() { 1 } else { layout.size(self).bits() }; + match cv { + Scalar::Int(int) => { + let data = int.assert_bits(layout.size(self)); + + // FIXME(antoyo): there's some issues with using the u128 code that follows, so hard-code + // the paths for floating-point values. + if ty == self.float_type { + return self.context.new_rvalue_from_double(ty, f32::from_bits(data as u32) as f64); + } + else if ty == self.double_type { + return self.context.new_rvalue_from_double(ty, f64::from_bits(data as u64)); + } + + let value = self.const_uint_big(self.type_ix(bitsize), data); + // TODO(bjorn3): assert size is correct + self.const_bitcast(value, ty) + } + Scalar::Ptr(ptr, _size) => { + let (alloc_id, offset) = ptr.into_parts(); + let base_addr = + match self.tcx.global_alloc(alloc_id) { + GlobalAlloc::Memory(alloc) => { + let init = const_alloc_to_gcc(self, alloc); + let alloc = alloc.inner(); + let value = + match alloc.mutability { + Mutability::Mut => self.static_addr_of_mut(init, alloc.align, None), + _ => self.static_addr_of(init, alloc.align, None), + }; + if !self.sess().fewer_names() { + // TODO(antoyo): set value name. + } + value + }, + GlobalAlloc::Function(fn_instance) => { + self.get_fn_addr(fn_instance) + }, + GlobalAlloc::VTable(ty, trait_ref) => { + let alloc = self.tcx.global_alloc(self.tcx.vtable_allocation((ty, trait_ref))).unwrap_memory(); + let init = const_alloc_to_gcc(self, alloc); + self.static_addr_of(init, alloc.inner().align, None) + } + GlobalAlloc::Static(def_id) => { + assert!(self.tcx.is_static(def_id)); + self.get_static(def_id).get_address(None) + }, + }; + let ptr_type = base_addr.get_type(); + let base_addr = self.const_bitcast(base_addr, self.usize_type); + let offset = self.context.new_rvalue_from_long(self.usize_type, offset.bytes() as i64); + let ptr = self.const_bitcast(base_addr + offset, ptr_type); + if layout.primitive() != Pointer { + self.const_bitcast(ptr.dereference(None).to_rvalue(), ty) + } + else { + self.const_bitcast(ptr, ty) + } + } + } + } + + fn const_data_from_alloc(&self, alloc: ConstAllocation<'tcx>) -> Self::Value { + const_alloc_to_gcc(self, alloc) + } + + fn from_const_alloc(&self, layout: TyAndLayout<'tcx>, alloc: ConstAllocation<'tcx>, offset: Size) -> PlaceRef<'tcx, RValue<'gcc>> { + assert_eq!(alloc.inner().align, layout.align.abi); + let ty = self.type_ptr_to(layout.gcc_type(self, true)); + let value = + if layout.size == Size::ZERO { + let value = self.const_usize(alloc.inner().align.bytes()); + self.context.new_cast(None, value, ty) + } + else { + let init = const_alloc_to_gcc(self, alloc); + let base_addr = self.static_addr_of(init, alloc.inner().align, None); + + let array = self.const_bitcast(base_addr, self.type_i8p()); + let value = self.context.new_array_access(None, array, self.const_usize(offset.bytes())).get_address(None); + self.const_bitcast(value, ty) + }; + PlaceRef::new_sized(value, layout) + } + + fn const_ptrcast(&self, val: RValue<'gcc>, ty: Type<'gcc>) -> RValue<'gcc> { + self.context.new_cast(None, val, ty) + } +} + +pub trait SignType<'gcc, 'tcx> { + fn is_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_unsigned(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn to_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; + fn to_unsigned(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; +} + +impl<'gcc, 'tcx> SignType<'gcc, 'tcx> for Type<'gcc> { + fn is_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.is_i8(cx) || self.is_i16(cx) || self.is_i32(cx) || self.is_i64(cx) || self.is_i128(cx) + } + + fn is_unsigned(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.is_u8(cx) || self.is_u16(cx) || self.is_u32(cx) || self.is_u64(cx) || self.is_u128(cx) + } + + fn to_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> { + if self.is_u8(cx) { + cx.i8_type + } + else if self.is_u16(cx) { + cx.i16_type + } + else if self.is_u32(cx) { + cx.i32_type + } + else if self.is_u64(cx) { + cx.i64_type + } + else if self.is_u128(cx) { + cx.i128_type + } + else if self.is_uchar(cx) { + cx.char_type + } + else if self.is_ushort(cx) { + cx.short_type + } + else if self.is_uint(cx) { + cx.int_type + } + else if self.is_ulong(cx) { + cx.long_type + } + else if self.is_ulonglong(cx) { + cx.longlong_type + } + else { + self.clone() + } + } + + fn to_unsigned(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> { + if self.is_i8(cx) { + cx.u8_type + } + else if self.is_i16(cx) { + cx.u16_type + } + else if self.is_i32(cx) { + cx.u32_type + } + else if self.is_i64(cx) { + cx.u64_type + } + else if self.is_i128(cx) { + cx.u128_type + } + else if self.is_char(cx) { + cx.uchar_type + } + else if self.is_short(cx) { + cx.ushort_type + } + else if self.is_int(cx) { + cx.uint_type + } + else if self.is_long(cx) { + cx.ulong_type + } + else if self.is_longlong(cx) { + cx.ulonglong_type + } + else { + self.clone() + } + } +} + +pub trait TypeReflection<'gcc, 'tcx> { + fn is_uchar(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_ushort(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_uint(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_ulong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_ulonglong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_char(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_short(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_int(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_long(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_longlong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + + fn is_i8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_u8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_i16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_u16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_i32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_u32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_i64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_u64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_i128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_u128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + + fn is_f32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + fn is_f64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool; + + fn is_vector(&self) -> bool; +} + +impl<'gcc, 'tcx> TypeReflection<'gcc, 'tcx> for Type<'gcc> { + fn is_uchar(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.uchar_type + } + + fn is_ushort(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.ushort_type + } + + fn is_uint(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.uint_type + } + + fn is_ulong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.ulong_type + } + + fn is_ulonglong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.ulonglong_type + } + + fn is_char(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.char_type + } + + fn is_short(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.short_type + } + + fn is_int(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.int_type + } + + fn is_long(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.long_type + } + + fn is_longlong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.longlong_type + } + + fn is_i8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.i8_type + } + + fn is_u8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.u8_type + } + + fn is_i16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.i16_type + } + + fn is_u16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.u16_type + } + + fn is_i32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.i32_type + } + + fn is_u32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.u32_type + } + + fn is_i64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.i64_type + } + + fn is_u64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.u64_type + } + + fn is_i128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.i128_type.unqualified() + } + + fn is_u128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.u128_type.unqualified() + } + + fn is_f32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.context.new_type::<f32>() + } + + fn is_f64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool { + self.unqualified() == cx.context.new_type::<f64>() + } + + fn is_vector(&self) -> bool { + let mut typ = self.clone(); + loop { + if typ.dyncast_vector().is_some() { + return true; + } + + let old_type = typ; + typ = typ.unqualified(); + if old_type == typ { + break; + } + } + + false + } +} diff --git a/compiler/rustc_codegen_gcc/src/consts.rs b/compiler/rustc_codegen_gcc/src/consts.rs new file mode 100644 index 000000000..c0b8d2181 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/consts.rs @@ -0,0 +1,405 @@ +use gccjit::{GlobalKind, LValue, RValue, ToRValue, Type}; +use rustc_codegen_ssa::traits::{BaseTypeMethods, ConstMethods, DerivedTypeMethods, StaticMethods}; +use rustc_hir as hir; +use rustc_hir::Node; +use rustc_middle::{bug, span_bug}; +use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs}; +use rustc_middle::mir::mono::MonoItem; +use rustc_middle::ty::{self, Instance, Ty}; +use rustc_middle::ty::layout::LayoutOf; +use rustc_middle::mir::interpret::{self, ConstAllocation, ErrorHandled, Scalar as InterpScalar, read_target_uint}; +use rustc_span::Span; +use rustc_span::def_id::DefId; +use rustc_target::abi::{self, Align, HasDataLayout, Primitive, Size, WrappingRange}; + +use crate::base; +use crate::context::CodegenCx; +use crate::type_of::LayoutGccExt; + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn const_bitcast(&self, value: RValue<'gcc>, typ: Type<'gcc>) -> RValue<'gcc> { + if value.get_type() == self.bool_type.make_pointer() { + if let Some(pointee) = typ.get_pointee() { + if pointee.dyncast_vector().is_some() { + panic!() + } + } + } + // NOTE: since bitcast makes a value non-constant, don't bitcast if not necessary as some + // SIMD builtins require a constant value. + self.bitcast_if_needed(value, typ) + } +} + +impl<'gcc, 'tcx> StaticMethods for CodegenCx<'gcc, 'tcx> { + fn static_addr_of(&self, cv: RValue<'gcc>, align: Align, kind: Option<&str>) -> RValue<'gcc> { + // TODO(antoyo): implement a proper rvalue comparison in libgccjit instead of doing the + // following: + for (value, variable) in &*self.const_globals.borrow() { + if format!("{:?}", value) == format!("{:?}", cv) { + if let Some(global_variable) = self.global_lvalues.borrow().get(variable) { + let alignment = align.bits() as i32; + if alignment > global_variable.get_alignment() { + global_variable.set_alignment(alignment); + } + } + return *variable; + } + } + let global_value = self.static_addr_of_mut(cv, align, kind); + #[cfg(feature = "master")] + self.global_lvalues.borrow().get(&global_value) + .expect("`static_addr_of_mut` did not add the global to `self.global_lvalues`") + .global_set_readonly(); + self.const_globals.borrow_mut().insert(cv, global_value); + global_value + } + + fn codegen_static(&self, def_id: DefId, is_mutable: bool) { + let attrs = self.tcx.codegen_fn_attrs(def_id); + + let value = + match codegen_static_initializer(&self, def_id) { + Ok((value, _)) => value, + // Error has already been reported + Err(_) => return, + }; + + let global = self.get_static(def_id); + + // boolean SSA values are i1, but they have to be stored in i8 slots, + // otherwise some LLVM optimization passes don't work as expected + let val_llty = self.val_ty(value); + let value = + if val_llty == self.type_i1() { + unimplemented!(); + } + else { + value + }; + + let instance = Instance::mono(self.tcx, def_id); + let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all()); + let gcc_type = self.layout_of(ty).gcc_type(self, true); + + // TODO(antoyo): set alignment. + + let value = self.bitcast_if_needed(value, gcc_type); + global.global_set_initializer_rvalue(value); + + // As an optimization, all shared statics which do not have interior + // mutability are placed into read-only memory. + if !is_mutable { + if self.type_is_freeze(ty) { + #[cfg(feature = "master")] + global.global_set_readonly(); + } + } + + if attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL) { + // Do not allow LLVM to change the alignment of a TLS on macOS. + // + // By default a global's alignment can be freely increased. + // This allows LLVM to generate more performant instructions + // e.g., using load-aligned into a SIMD register. + // + // However, on macOS 10.10 or below, the dynamic linker does not + // respect any alignment given on the TLS (radar 24221680). + // This will violate the alignment assumption, and causing segfault at runtime. + // + // This bug is very easy to trigger. In `println!` and `panic!`, + // the `LOCAL_STDOUT`/`LOCAL_STDERR` handles are stored in a TLS, + // which the values would be `mem::replace`d on initialization. + // The implementation of `mem::replace` will use SIMD + // whenever the size is 32 bytes or higher. LLVM notices SIMD is used + // and tries to align `LOCAL_STDOUT`/`LOCAL_STDERR` to a 32-byte boundary, + // which macOS's dyld disregarded and causing crashes + // (see issues #51794, #51758, #50867, #48866 and #44056). + // + // To workaround the bug, we trick LLVM into not increasing + // the global's alignment by explicitly assigning a section to it + // (equivalent to automatically generating a `#[link_section]` attribute). + // See the comment in the `GlobalValue::canIncreaseAlignment()` function + // of `lib/IR/Globals.cpp` for why this works. + // + // When the alignment is not increased, the optimized `mem::replace` + // will use load-unaligned instructions instead, and thus avoiding the crash. + // + // We could remove this hack whenever we decide to drop macOS 10.10 support. + if self.tcx.sess.target.options.is_like_osx { + // The `inspect` method is okay here because we checked relocations, and + // because we are doing this access to inspect the final interpreter state + // (not as part of the interpreter execution). + // + // FIXME: This check requires that the (arbitrary) value of undefined bytes + // happens to be zero. Instead, we should only check the value of defined bytes + // and set all undefined bytes to zero if this allocation is headed for the + // BSS. + unimplemented!(); + } + } + + // Wasm statics with custom link sections get special treatment as they + // go into custom sections of the wasm executable. + if self.tcx.sess.opts.target_triple.triple().starts_with("wasm32") { + if let Some(_section) = attrs.link_section { + unimplemented!(); + } + } else { + // TODO(antoyo): set link section. + } + + if attrs.flags.contains(CodegenFnAttrFlags::USED) || attrs.flags.contains(CodegenFnAttrFlags::USED_LINKER) { + self.add_used_global(global.to_rvalue()); + } + } + + /// Add a global value to a list to be stored in the `llvm.used` variable, an array of i8*. + fn add_used_global(&self, _global: RValue<'gcc>) { + // TODO(antoyo) + } + + fn add_compiler_used_global(&self, _global: RValue<'gcc>) { + // TODO(antoyo) + } +} + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn static_addr_of_mut(&self, cv: RValue<'gcc>, align: Align, kind: Option<&str>) -> RValue<'gcc> { + let global = + match kind { + Some(kind) if !self.tcx.sess.fewer_names() => { + let name = self.generate_local_symbol_name(kind); + // TODO(antoyo): check if it's okay that no link_section is set. + + let typ = self.val_ty(cv).get_aligned(align.bytes()); + let global = self.declare_private_global(&name[..], typ); + global + } + _ => { + let typ = self.val_ty(cv).get_aligned(align.bytes()); + let global = self.declare_unnamed_global(typ); + global + }, + }; + global.global_set_initializer_rvalue(cv); + // TODO(antoyo): set unnamed address. + let rvalue = global.get_address(None); + self.global_lvalues.borrow_mut().insert(rvalue, global); + rvalue + } + + pub fn get_static(&self, def_id: DefId) -> LValue<'gcc> { + let instance = Instance::mono(self.tcx, def_id); + let fn_attrs = self.tcx.codegen_fn_attrs(def_id); + if let Some(&global) = self.instances.borrow().get(&instance) { + return global; + } + + let defined_in_current_codegen_unit = + self.codegen_unit.items().contains_key(&MonoItem::Static(def_id)); + assert!( + !defined_in_current_codegen_unit, + "consts::get_static() should always hit the cache for \ + statics defined in the same CGU, but did not for `{:?}`", + def_id + ); + + let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all()); + let sym = self.tcx.symbol_name(instance).name; + + let global = + if let Some(def_id) = def_id.as_local() { + let id = self.tcx.hir().local_def_id_to_hir_id(def_id); + let llty = self.layout_of(ty).gcc_type(self, true); + // FIXME: refactor this to work without accessing the HIR + let global = match self.tcx.hir().get(id) { + Node::Item(&hir::Item { span, kind: hir::ItemKind::Static(..), .. }) => { + if let Some(global) = self.get_declared_value(&sym) { + if self.val_ty(global) != self.type_ptr_to(llty) { + span_bug!(span, "Conflicting types for static"); + } + } + + let is_tls = fn_attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL); + let global = self.declare_global( + &sym, + llty, + GlobalKind::Exported, + is_tls, + fn_attrs.link_section, + ); + + if !self.tcx.is_reachable_non_generic(def_id) { + // TODO(antoyo): set visibility. + } + + global + } + + Node::ForeignItem(&hir::ForeignItem { + span, + kind: hir::ForeignItemKind::Static(..), + .. + }) => { + let fn_attrs = self.tcx.codegen_fn_attrs(def_id); + check_and_apply_linkage(&self, &fn_attrs, ty, sym, span) + } + + item => bug!("get_static: expected static, found {:?}", item), + }; + + global + } + else { + // FIXME(nagisa): perhaps the map of externs could be offloaded to llvm somehow? + //debug!("get_static: sym={} item_attr={:?}", sym, self.tcx.item_attrs(def_id)); + + let attrs = self.tcx.codegen_fn_attrs(def_id); + let span = self.tcx.def_span(def_id); + let global = check_and_apply_linkage(&self, &attrs, ty, sym, span); + + let needs_dll_storage_attr = false; // TODO(antoyo) + + // If this assertion triggers, there's something wrong with commandline + // argument validation. + debug_assert!( + !(self.tcx.sess.opts.cg.linker_plugin_lto.enabled() + && self.tcx.sess.target.options.is_like_msvc + && self.tcx.sess.opts.cg.prefer_dynamic) + ); + + if needs_dll_storage_attr { + // This item is external but not foreign, i.e., it originates from an external Rust + // crate. Since we don't know whether this crate will be linked dynamically or + // statically in the final application, we always mark such symbols as 'dllimport'. + // If final linkage happens to be static, we rely on compiler-emitted __imp_ stubs + // to make things work. + // + // However, in some scenarios we defer emission of statics to downstream + // crates, so there are cases where a static with an upstream DefId + // is actually present in the current crate. We can find out via the + // is_codegened_item query. + if !self.tcx.is_codegened_item(def_id) { + unimplemented!(); + } + } + global + }; + + // TODO(antoyo): set dll storage class. + + self.instances.borrow_mut().insert(instance, global); + global + } +} + +pub fn const_alloc_to_gcc<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, alloc: ConstAllocation<'tcx>) -> RValue<'gcc> { + let alloc = alloc.inner(); + let mut llvals = Vec::with_capacity(alloc.relocations().len() + 1); + let dl = cx.data_layout(); + let pointer_size = dl.pointer_size.bytes() as usize; + + let mut next_offset = 0; + for &(offset, alloc_id) in alloc.relocations().iter() { + let offset = offset.bytes(); + assert_eq!(offset as usize as u64, offset); + let offset = offset as usize; + if offset > next_offset { + // This `inspect` is okay since we have checked that it is not within a relocation, it + // is within the bounds of the allocation, and it doesn't affect interpreter execution + // (we inspect the result after interpreter execution). Any undef byte is replaced with + // some arbitrary byte value. + // + // FIXME: relay undef bytes to codegen as undef const bytes + let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(next_offset..offset); + llvals.push(cx.const_bytes(bytes)); + } + let ptr_offset = + read_target_uint( dl.endian, + // This `inspect` is okay since it is within the bounds of the allocation, it doesn't + // affect interpreter execution (we inspect the result after interpreter execution), + // and we properly interpret the relocation as a relocation pointer offset. + alloc.inspect_with_uninit_and_ptr_outside_interpreter(offset..(offset + pointer_size)), + ) + .expect("const_alloc_to_llvm: could not read relocation pointer") + as u64; + llvals.push(cx.scalar_to_backend( + InterpScalar::from_pointer( + interpret::Pointer::new(alloc_id, Size::from_bytes(ptr_offset)), + &cx.tcx, + ), + abi::Scalar::Initialized { value: Primitive::Pointer, valid_range: WrappingRange::full(dl.pointer_size) }, + cx.type_i8p(), + )); + next_offset = offset + pointer_size; + } + if alloc.len() >= next_offset { + let range = next_offset..alloc.len(); + // This `inspect` is okay since we have check that it is after all relocations, it is + // within the bounds of the allocation, and it doesn't affect interpreter execution (we + // inspect the result after interpreter execution). Any undef byte is replaced with some + // arbitrary byte value. + // + // FIXME: relay undef bytes to codegen as undef const bytes + let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(range); + llvals.push(cx.const_bytes(bytes)); + } + + cx.const_struct(&llvals, true) +} + +pub fn codegen_static_initializer<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, def_id: DefId) -> Result<(RValue<'gcc>, ConstAllocation<'tcx>), ErrorHandled> { + let alloc = cx.tcx.eval_static_initializer(def_id)?; + Ok((const_alloc_to_gcc(cx, alloc), alloc)) +} + +fn check_and_apply_linkage<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, attrs: &CodegenFnAttrs, ty: Ty<'tcx>, sym: &str, span: Span) -> LValue<'gcc> { + let is_tls = attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL); + let llty = cx.layout_of(ty).gcc_type(cx, true); + if let Some(linkage) = attrs.linkage { + // If this is a static with a linkage specified, then we need to handle + // it a little specially. The typesystem prevents things like &T and + // extern "C" fn() from being non-null, so we can't just declare a + // static and call it a day. Some linkages (like weak) will make it such + // that the static actually has a null value. + let llty2 = + if let ty::RawPtr(ref mt) = ty.kind() { + cx.layout_of(mt.ty).gcc_type(cx, true) + } + else { + cx.sess().span_fatal( + span, + "must have type `*const T` or `*mut T` due to `#[linkage]` attribute", + ) + }; + // Declare a symbol `foo` with the desired linkage. + let global1 = cx.declare_global_with_linkage(&sym, llty2, base::global_linkage_to_gcc(linkage)); + + // Declare an internal global `extern_with_linkage_foo` which + // is initialized with the address of `foo`. If `foo` is + // discarded during linking (for example, if `foo` has weak + // linkage and there are no definitions), then + // `extern_with_linkage_foo` will instead be initialized to + // zero. + let mut real_name = "_rust_extern_with_linkage_".to_string(); + real_name.push_str(&sym); + let global2 = cx.define_global(&real_name, llty, is_tls, attrs.link_section); + // TODO(antoyo): set linkage. + global2.global_set_initializer_rvalue(global1.get_address(None)); + // TODO(antoyo): use global_set_initializer() when it will work. + global2 + } + else { + // Generate an external declaration. + // FIXME(nagisa): investigate whether it can be changed into define_global + + // Thread-local statics in some other crate need to *always* be linked + // against in a thread-local fashion, so we need to be sure to apply the + // thread-local attribute locally if it was present remotely. If we + // don't do this then linker errors can be generated where the linker + // complains that one object files has a thread local version of the + // symbol and another one doesn't. + cx.declare_global(&sym, llty, GlobalKind::Imported, is_tls, attrs.link_section) + } +} diff --git a/compiler/rustc_codegen_gcc/src/context.rs b/compiler/rustc_codegen_gcc/src/context.rs new file mode 100644 index 000000000..478f6d893 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/context.rs @@ -0,0 +1,553 @@ +use std::cell::{Cell, RefCell}; + +use gccjit::{Block, CType, Context, Function, FunctionPtrType, FunctionType, LValue, RValue, Struct, Type}; +use rustc_codegen_ssa::base::wants_msvc_seh; +use rustc_codegen_ssa::traits::{ + BackendTypes, + MiscMethods, +}; +use rustc_data_structures::base_n; +use rustc_data_structures::fx::{FxHashMap, FxHashSet}; +use rustc_middle::span_bug; +use rustc_middle::mir::mono::CodegenUnit; +use rustc_middle::ty::{self, Instance, ParamEnv, PolyExistentialTraitRef, Ty, TyCtxt}; +use rustc_middle::ty::layout::{FnAbiError, FnAbiOfHelpers, FnAbiRequest, HasParamEnv, HasTyCtxt, LayoutError, TyAndLayout, LayoutOfHelpers}; +use rustc_session::Session; +use rustc_span::Span; +use rustc_target::abi::{call::FnAbi, HasDataLayout, PointeeInfo, Size, TargetDataLayout, VariantIdx}; +use rustc_target::spec::{HasTargetSpec, Target, TlsModel}; + +use crate::callee::get_fn; + +#[derive(Clone)] +pub struct FuncSig<'gcc> { + pub params: Vec<Type<'gcc>>, + pub return_type: Type<'gcc>, +} + +pub struct CodegenCx<'gcc, 'tcx> { + pub check_overflow: bool, + pub codegen_unit: &'tcx CodegenUnit<'tcx>, + pub context: &'gcc Context<'gcc>, + + // TODO(bjorn3): Can this field be removed? + pub current_func: RefCell<Option<Function<'gcc>>>, + pub normal_function_addresses: RefCell<FxHashSet<RValue<'gcc>>>, + + pub functions: RefCell<FxHashMap<String, Function<'gcc>>>, + pub intrinsics: RefCell<FxHashMap<String, Function<'gcc>>>, + + pub tls_model: gccjit::TlsModel, + + pub bool_type: Type<'gcc>, + pub i8_type: Type<'gcc>, + pub i16_type: Type<'gcc>, + pub i32_type: Type<'gcc>, + pub i64_type: Type<'gcc>, + pub i128_type: Type<'gcc>, + pub isize_type: Type<'gcc>, + + pub u8_type: Type<'gcc>, + pub u16_type: Type<'gcc>, + pub u32_type: Type<'gcc>, + pub u64_type: Type<'gcc>, + pub u128_type: Type<'gcc>, + pub usize_type: Type<'gcc>, + + pub char_type: Type<'gcc>, + pub uchar_type: Type<'gcc>, + pub short_type: Type<'gcc>, + pub ushort_type: Type<'gcc>, + pub int_type: Type<'gcc>, + pub uint_type: Type<'gcc>, + pub long_type: Type<'gcc>, + pub ulong_type: Type<'gcc>, + pub longlong_type: Type<'gcc>, + pub ulonglong_type: Type<'gcc>, + pub sizet_type: Type<'gcc>, + + pub supports_128bit_integers: bool, + + pub float_type: Type<'gcc>, + pub double_type: Type<'gcc>, + + pub linkage: Cell<FunctionType>, + pub scalar_types: RefCell<FxHashMap<Ty<'tcx>, Type<'gcc>>>, + pub types: RefCell<FxHashMap<(Ty<'tcx>, Option<VariantIdx>), Type<'gcc>>>, + pub tcx: TyCtxt<'tcx>, + + pub struct_types: RefCell<FxHashMap<Vec<Type<'gcc>>, Type<'gcc>>>, + + pub types_with_fields_to_set: RefCell<FxHashMap<Type<'gcc>, (Struct<'gcc>, TyAndLayout<'tcx>)>>, + + /// Cache instances of monomorphic and polymorphic items + pub instances: RefCell<FxHashMap<Instance<'tcx>, LValue<'gcc>>>, + /// Cache function instances of monomorphic and polymorphic items + pub function_instances: RefCell<FxHashMap<Instance<'tcx>, RValue<'gcc>>>, + /// Cache generated vtables + pub vtables: RefCell<FxHashMap<(Ty<'tcx>, Option<ty::PolyExistentialTraitRef<'tcx>>), RValue<'gcc>>>, + + // TODO(antoyo): improve the SSA API to not require those. + // Mapping from function pointer type to indexes of on stack parameters. + pub on_stack_params: RefCell<FxHashMap<FunctionPtrType<'gcc>, FxHashSet<usize>>>, + // Mapping from function to indexes of on stack parameters. + pub on_stack_function_params: RefCell<FxHashMap<Function<'gcc>, FxHashSet<usize>>>, + + /// Cache of emitted const globals (value -> global) + pub const_globals: RefCell<FxHashMap<RValue<'gcc>, RValue<'gcc>>>, + + /// Map from the address of a global variable (rvalue) to the global variable itself (lvalue). + /// TODO(antoyo): remove when the rustc API is fixed. + pub global_lvalues: RefCell<FxHashMap<RValue<'gcc>, LValue<'gcc>>>, + + /// Cache of constant strings, + pub const_str_cache: RefCell<FxHashMap<String, LValue<'gcc>>>, + + /// Cache of globals. + pub globals: RefCell<FxHashMap<String, RValue<'gcc>>>, + + /// A counter that is used for generating local symbol names + local_gen_sym_counter: Cell<usize>, + + eh_personality: Cell<Option<RValue<'gcc>>>, + + pub pointee_infos: RefCell<FxHashMap<(Ty<'tcx>, Size), Option<PointeeInfo>>>, + + /// NOTE: a hack is used because the rustc API is not suitable to libgccjit and as such, + /// `const_undef()` returns struct as pointer so that they can later be assigned a value. + /// As such, this set remembers which of these pointers were returned by this function so that + /// they can be dereferenced later. + /// FIXME(antoyo): fix the rustc API to avoid having this hack. + pub structs_as_pointer: RefCell<FxHashSet<RValue<'gcc>>>, +} + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn new(context: &'gcc Context<'gcc>, codegen_unit: &'tcx CodegenUnit<'tcx>, tcx: TyCtxt<'tcx>, supports_128bit_integers: bool) -> Self { + let check_overflow = tcx.sess.overflow_checks(); + + let i8_type = context.new_c_type(CType::Int8t); + let i16_type = context.new_c_type(CType::Int16t); + let i32_type = context.new_c_type(CType::Int32t); + let i64_type = context.new_c_type(CType::Int64t); + let u8_type = context.new_c_type(CType::UInt8t); + let u16_type = context.new_c_type(CType::UInt16t); + let u32_type = context.new_c_type(CType::UInt32t); + let u64_type = context.new_c_type(CType::UInt64t); + + let (i128_type, u128_type) = + if supports_128bit_integers { + let i128_type = context.new_c_type(CType::Int128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?; + let u128_type = context.new_c_type(CType::UInt128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?; + (i128_type, u128_type) + } + else { + let i128_type = context.new_array_type(None, i64_type, 2); + let u128_type = context.new_array_type(None, u64_type, 2); + (i128_type, u128_type) + }; + + let tls_model = to_gcc_tls_mode(tcx.sess.tls_model()); + + let float_type = context.new_type::<f32>(); + let double_type = context.new_type::<f64>(); + + let char_type = context.new_c_type(CType::Char); + let uchar_type = context.new_c_type(CType::UChar); + let short_type = context.new_c_type(CType::Short); + let ushort_type = context.new_c_type(CType::UShort); + let int_type = context.new_c_type(CType::Int); + let uint_type = context.new_c_type(CType::UInt); + let long_type = context.new_c_type(CType::Long); + let ulong_type = context.new_c_type(CType::ULong); + let longlong_type = context.new_c_type(CType::LongLong); + let ulonglong_type = context.new_c_type(CType::ULongLong); + let sizet_type = context.new_c_type(CType::SizeT); + + let isize_type = context.new_c_type(CType::LongLong); + let usize_type = context.new_c_type(CType::ULongLong); + let bool_type = context.new_type::<bool>(); + + // TODO(antoyo): only have those assertions on x86_64. + assert_eq!(isize_type.get_size(), i64_type.get_size()); + assert_eq!(usize_type.get_size(), u64_type.get_size()); + + let mut functions = FxHashMap::default(); + let builtins = [ + "__builtin_unreachable", "abort", "__builtin_expect", "__builtin_add_overflow", "__builtin_mul_overflow", + "__builtin_saddll_overflow", /*"__builtin_sadd_overflow",*/ "__builtin_smulll_overflow", /*"__builtin_smul_overflow",*/ + "__builtin_ssubll_overflow", /*"__builtin_ssub_overflow",*/ "__builtin_sub_overflow", "__builtin_uaddll_overflow", + "__builtin_uadd_overflow", "__builtin_umulll_overflow", "__builtin_umul_overflow", "__builtin_usubll_overflow", + "__builtin_usub_overflow", "sqrtf", "sqrt", "__builtin_powif", "__builtin_powi", "sinf", "sin", "cosf", "cos", + "powf", "pow", "expf", "exp", "exp2f", "exp2", "logf", "log", "log10f", "log10", "log2f", "log2", "fmaf", + "fma", "fabsf", "fabs", "fminf", "fmin", "fmaxf", "fmax", "copysignf", "copysign", "floorf", "floor", "ceilf", + "ceil", "truncf", "trunc", "rintf", "rint", "nearbyintf", "nearbyint", "roundf", "round", + "__builtin_expect_with_probability", + ]; + + for builtin in builtins.iter() { + functions.insert(builtin.to_string(), context.get_builtin_function(builtin)); + } + + Self { + check_overflow, + codegen_unit, + context, + current_func: RefCell::new(None), + normal_function_addresses: Default::default(), + functions: RefCell::new(functions), + intrinsics: RefCell::new(FxHashMap::default()), + + tls_model, + + bool_type, + i8_type, + i16_type, + i32_type, + i64_type, + i128_type, + isize_type, + usize_type, + u8_type, + u16_type, + u32_type, + u64_type, + u128_type, + char_type, + uchar_type, + short_type, + ushort_type, + int_type, + uint_type, + long_type, + ulong_type, + longlong_type, + ulonglong_type, + sizet_type, + + supports_128bit_integers, + + float_type, + double_type, + + linkage: Cell::new(FunctionType::Internal), + instances: Default::default(), + function_instances: Default::default(), + on_stack_params: Default::default(), + on_stack_function_params: Default::default(), + vtables: Default::default(), + const_globals: Default::default(), + global_lvalues: Default::default(), + const_str_cache: Default::default(), + globals: Default::default(), + scalar_types: Default::default(), + types: Default::default(), + tcx, + struct_types: Default::default(), + types_with_fields_to_set: Default::default(), + local_gen_sym_counter: Cell::new(0), + eh_personality: Cell::new(None), + pointee_infos: Default::default(), + structs_as_pointer: Default::default(), + } + } + + pub fn rvalue_as_function(&self, value: RValue<'gcc>) -> Function<'gcc> { + let function: Function<'gcc> = unsafe { std::mem::transmute(value) }; + debug_assert!(self.functions.borrow().values().find(|value| **value == function).is_some(), + "{:?} ({:?}) is not a function", value, value.get_type()); + function + } + + pub fn is_native_int_type(&self, typ: Type<'gcc>) -> bool { + let types = [ + self.u8_type, + self.u16_type, + self.u32_type, + self.u64_type, + self.i8_type, + self.i16_type, + self.i32_type, + self.i64_type, + ]; + + for native_type in types { + if native_type.is_compatible_with(typ) { + return true; + } + } + + self.supports_128bit_integers && + (self.u128_type.is_compatible_with(typ) || self.i128_type.is_compatible_with(typ)) + } + + pub fn is_non_native_int_type(&self, typ: Type<'gcc>) -> bool { + !self.supports_128bit_integers && + (self.u128_type.is_compatible_with(typ) || self.i128_type.is_compatible_with(typ)) + } + + pub fn is_native_int_type_or_bool(&self, typ: Type<'gcc>) -> bool { + self.is_native_int_type(typ) || typ.is_compatible_with(self.bool_type) + } + + pub fn is_int_type_or_bool(&self, typ: Type<'gcc>) -> bool { + self.is_native_int_type(typ) || self.is_non_native_int_type(typ) || typ.is_compatible_with(self.bool_type) + } + + pub fn sess(&self) -> &Session { + &self.tcx.sess + } + + pub fn bitcast_if_needed(&self, value: RValue<'gcc>, expected_type: Type<'gcc>) -> RValue<'gcc> { + if value.get_type() != expected_type { + self.context.new_bitcast(None, value, expected_type) + } + else { + value + } + } +} + +impl<'gcc, 'tcx> BackendTypes for CodegenCx<'gcc, 'tcx> { + type Value = RValue<'gcc>; + type Function = RValue<'gcc>; + + type BasicBlock = Block<'gcc>; + type Type = Type<'gcc>; + type Funclet = (); // TODO(antoyo) + + type DIScope = (); // TODO(antoyo) + type DILocation = (); // TODO(antoyo) + type DIVariable = (); // TODO(antoyo) +} + +impl<'gcc, 'tcx> MiscMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn vtables(&self) -> &RefCell<FxHashMap<(Ty<'tcx>, Option<PolyExistentialTraitRef<'tcx>>), RValue<'gcc>>> { + &self.vtables + } + + fn get_fn(&self, instance: Instance<'tcx>) -> RValue<'gcc> { + let func = get_fn(self, instance); + *self.current_func.borrow_mut() = Some(self.rvalue_as_function(func)); + func + } + + fn get_fn_addr(&self, instance: Instance<'tcx>) -> RValue<'gcc> { + let func_name = self.tcx.symbol_name(instance).name; + + let func = + if self.intrinsics.borrow().contains_key(func_name) { + self.intrinsics.borrow()[func_name].clone() + } + else { + let func = get_fn(self, instance); + self.rvalue_as_function(func) + }; + let ptr = func.get_address(None); + + // TODO(antoyo): don't do this twice: i.e. in declare_fn and here. + // FIXME(antoyo): the rustc API seems to call get_fn_addr() when not needed (e.g. for FFI). + + self.normal_function_addresses.borrow_mut().insert(ptr); + + ptr + } + + fn eh_personality(&self) -> RValue<'gcc> { + // The exception handling personality function. + // + // If our compilation unit has the `eh_personality` lang item somewhere + // within it, then we just need to codegen that. Otherwise, we're + // building an rlib which will depend on some upstream implementation of + // this function, so we just codegen a generic reference to it. We don't + // specify any of the types for the function, we just make it a symbol + // that LLVM can later use. + // + // Note that MSVC is a little special here in that we don't use the + // `eh_personality` lang item at all. Currently LLVM has support for + // both Dwarf and SEH unwind mechanisms for MSVC targets and uses the + // *name of the personality function* to decide what kind of unwind side + // tables/landing pads to emit. It looks like Dwarf is used by default, + // injecting a dependency on the `_Unwind_Resume` symbol for resuming + // an "exception", but for MSVC we want to force SEH. This means that we + // can't actually have the personality function be our standard + // `rust_eh_personality` function, but rather we wired it up to the + // CRT's custom personality function, which forces LLVM to consider + // landing pads as "landing pads for SEH". + if let Some(llpersonality) = self.eh_personality.get() { + return llpersonality; + } + let tcx = self.tcx; + let llfn = match tcx.lang_items().eh_personality() { + Some(def_id) if !wants_msvc_seh(self.sess()) => self.get_fn_addr( + ty::Instance::resolve( + tcx, + ty::ParamEnv::reveal_all(), + def_id, + tcx.intern_substs(&[]), + ) + .unwrap().unwrap(), + ), + _ => { + let _name = if wants_msvc_seh(self.sess()) { + "__CxxFrameHandler3" + } else { + "rust_eh_personality" + }; + //let func = self.declare_func(name, self.type_i32(), &[], true); + // FIXME(antoyo): this hack should not be needed. That will probably be removed when + // unwinding support is added. + self.context.new_rvalue_from_int(self.int_type, 0) + } + }; + // TODO(antoyo): apply target cpu attributes. + self.eh_personality.set(Some(llfn)); + llfn + } + + fn sess(&self) -> &Session { + &self.tcx.sess + } + + fn check_overflow(&self) -> bool { + self.check_overflow + } + + fn codegen_unit(&self) -> &'tcx CodegenUnit<'tcx> { + self.codegen_unit + } + + fn used_statics(&self) -> &RefCell<Vec<RValue<'gcc>>> { + unimplemented!(); + } + + fn set_frame_pointer_type(&self, _llfn: RValue<'gcc>) { + // TODO(antoyo) + } + + fn apply_target_cpu_attr(&self, _llfn: RValue<'gcc>) { + // TODO(antoyo) + } + + fn create_used_variable(&self) { + unimplemented!(); + } + + fn declare_c_main(&self, fn_type: Self::Type) -> Option<Self::Function> { + if self.get_declared_value("main").is_none() { + Some(self.declare_cfn("main", fn_type)) + } + else { + // If the symbol already exists, it is an error: for example, the user wrote + // #[no_mangle] extern "C" fn main(..) {..} + // instead of #[start] + None + } + } + + fn compiler_used_statics(&self) -> &RefCell<Vec<RValue<'gcc>>> { + unimplemented!() + } + + fn create_compiler_used_variable(&self) { + unimplemented!() + } +} + +impl<'gcc, 'tcx> HasTyCtxt<'tcx> for CodegenCx<'gcc, 'tcx> { + fn tcx(&self) -> TyCtxt<'tcx> { + self.tcx + } +} + +impl<'gcc, 'tcx> HasDataLayout for CodegenCx<'gcc, 'tcx> { + fn data_layout(&self) -> &TargetDataLayout { + &self.tcx.data_layout + } +} + +impl<'gcc, 'tcx> HasTargetSpec for CodegenCx<'gcc, 'tcx> { + fn target_spec(&self) -> &Target { + &self.tcx.sess.target + } +} + +impl<'gcc, 'tcx> LayoutOfHelpers<'tcx> for CodegenCx<'gcc, 'tcx> { + type LayoutOfResult = TyAndLayout<'tcx>; + + #[inline] + fn handle_layout_err(&self, err: LayoutError<'tcx>, span: Span, ty: Ty<'tcx>) -> ! { + if let LayoutError::SizeOverflow(_) = err { + self.sess().span_fatal(span, &err.to_string()) + } else { + span_bug!(span, "failed to get layout for `{}`: {}", ty, err) + } + } +} + +impl<'gcc, 'tcx> FnAbiOfHelpers<'tcx> for CodegenCx<'gcc, 'tcx> { + type FnAbiOfResult = &'tcx FnAbi<'tcx, Ty<'tcx>>; + + #[inline] + fn handle_fn_abi_err( + &self, + err: FnAbiError<'tcx>, + span: Span, + fn_abi_request: FnAbiRequest<'tcx>, + ) -> ! { + if let FnAbiError::Layout(LayoutError::SizeOverflow(_)) = err { + self.sess().span_fatal(span, &err.to_string()) + } else { + match fn_abi_request { + FnAbiRequest::OfFnPtr { sig, extra_args } => { + span_bug!( + span, + "`fn_abi_of_fn_ptr({}, {:?})` failed: {}", + sig, + extra_args, + err + ); + } + FnAbiRequest::OfInstance { instance, extra_args } => { + span_bug!( + span, + "`fn_abi_of_instance({}, {:?})` failed: {}", + instance, + extra_args, + err + ); + } + } + } + } +} + +impl<'tcx, 'gcc> HasParamEnv<'tcx> for CodegenCx<'gcc, 'tcx> { + fn param_env(&self) -> ParamEnv<'tcx> { + ParamEnv::reveal_all() + } +} + +impl<'b, 'tcx> CodegenCx<'b, 'tcx> { + /// Generates a new symbol name with the given prefix. This symbol name must + /// only be used for definitions with `internal` or `private` linkage. + pub fn generate_local_symbol_name(&self, prefix: &str) -> String { + let idx = self.local_gen_sym_counter.get(); + self.local_gen_sym_counter.set(idx + 1); + // Include a '.' character, so there can be no accidental conflicts with + // user defined names + let mut name = String::with_capacity(prefix.len() + 6); + name.push_str(prefix); + name.push_str("."); + base_n::push_str(idx as u128, base_n::ALPHANUMERIC_ONLY, &mut name); + name + } +} + +fn to_gcc_tls_mode(tls_model: TlsModel) -> gccjit::TlsModel { + match tls_model { + TlsModel::GeneralDynamic => gccjit::TlsModel::GlobalDynamic, + TlsModel::LocalDynamic => gccjit::TlsModel::LocalDynamic, + TlsModel::InitialExec => gccjit::TlsModel::InitialExec, + TlsModel::LocalExec => gccjit::TlsModel::LocalExec, + } +} diff --git a/compiler/rustc_codegen_gcc/src/coverageinfo.rs b/compiler/rustc_codegen_gcc/src/coverageinfo.rs new file mode 100644 index 000000000..872fc2472 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/coverageinfo.rs @@ -0,0 +1,69 @@ +use gccjit::RValue; +use rustc_codegen_ssa::traits::{CoverageInfoBuilderMethods, CoverageInfoMethods}; +use rustc_hir::def_id::DefId; +use rustc_middle::mir::coverage::{ + CodeRegion, + CounterValueReference, + ExpressionOperandId, + InjectedExpressionId, + Op, +}; +use rustc_middle::ty::Instance; + +use crate::builder::Builder; +use crate::context::CodegenCx; + +impl<'a, 'gcc, 'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { + fn set_function_source_hash( + &mut self, + _instance: Instance<'tcx>, + _function_source_hash: u64, + ) -> bool { + unimplemented!(); + } + + fn add_coverage_counter(&mut self, _instance: Instance<'tcx>, _id: CounterValueReference, _region: CodeRegion) -> bool { + // TODO(antoyo) + false + } + + fn add_coverage_counter_expression(&mut self, _instance: Instance<'tcx>, _id: InjectedExpressionId, _lhs: ExpressionOperandId, _op: Op, _rhs: ExpressionOperandId, _region: Option<CodeRegion>) -> bool { + // TODO(antoyo) + false + } + + fn add_coverage_unreachable(&mut self, _instance: Instance<'tcx>, _region: CodeRegion) -> bool { + // TODO(antoyo) + false + } +} + +impl<'gcc, 'tcx> CoverageInfoMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn coverageinfo_finalize(&self) { + // TODO(antoyo) + } + + fn get_pgo_func_name_var(&self, _instance: Instance<'tcx>) -> RValue<'gcc> { + unimplemented!(); + } + + /// Functions with MIR-based coverage are normally codegenned _only_ if + /// called. LLVM coverage tools typically expect every function to be + /// defined (even if unused), with at least one call to LLVM intrinsic + /// `instrprof.increment`. + /// + /// Codegen a small function that will never be called, with one counter + /// that will never be incremented. + /// + /// For used/called functions, the coverageinfo was already added to the + /// `function_coverage_map` (keyed by function `Instance`) during codegen. + /// But in this case, since the unused function was _not_ previously + /// codegenned, collect the coverage `CodeRegion`s from the MIR and add + /// them. The first `CodeRegion` is used to add a single counter, with the + /// same counter ID used in the injected `instrprof.increment` intrinsic + /// call. Since the function is never called, all other `CodeRegion`s can be + /// added as `unreachable_region`s. + fn define_unused_fn(&self, _def_id: DefId) { + unimplemented!(); + } +} diff --git a/compiler/rustc_codegen_gcc/src/debuginfo.rs b/compiler/rustc_codegen_gcc/src/debuginfo.rs new file mode 100644 index 000000000..266759ed6 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/debuginfo.rs @@ -0,0 +1,62 @@ +use gccjit::RValue; +use rustc_codegen_ssa::mir::debuginfo::{FunctionDebugContext, VariableKind}; +use rustc_codegen_ssa::traits::{DebugInfoBuilderMethods, DebugInfoMethods}; +use rustc_middle::mir; +use rustc_middle::ty::{Instance, PolyExistentialTraitRef, Ty}; +use rustc_span::{SourceFile, Span, Symbol}; +use rustc_target::abi::Size; +use rustc_target::abi::call::FnAbi; + +use crate::builder::Builder; +use crate::context::CodegenCx; + +impl<'a, 'gcc, 'tcx> DebugInfoBuilderMethods for Builder<'a, 'gcc, 'tcx> { + // FIXME(eddyb) find a common convention for all of the debuginfo-related + // names (choose between `dbg`, `debug`, `debuginfo`, `debug_info` etc.). + fn dbg_var_addr(&mut self, _dbg_var: Self::DIVariable, _scope_metadata: Self::DIScope, _variable_alloca: Self::Value, _direct_offset: Size, _indirect_offsets: &[Size]) { + unimplemented!(); + } + + fn insert_reference_to_gdb_debug_scripts_section_global(&mut self) { + // TODO(antoyo): insert reference to gdb debug scripts section global. + } + + fn set_var_name(&mut self, _value: RValue<'gcc>, _name: &str) { + unimplemented!(); + } + + fn set_dbg_loc(&mut self, _dbg_loc: Self::DILocation) { + unimplemented!(); + } +} + +impl<'gcc, 'tcx> DebugInfoMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn create_vtable_debuginfo(&self, _ty: Ty<'tcx>, _trait_ref: Option<PolyExistentialTraitRef<'tcx>>, _vtable: Self::Value) { + // TODO(antoyo) + } + + fn create_function_debug_context(&self, _instance: Instance<'tcx>, _fn_abi: &FnAbi<'tcx, Ty<'tcx>>, _llfn: RValue<'gcc>, _mir: &mir::Body<'tcx>) -> Option<FunctionDebugContext<Self::DIScope, Self::DILocation>> { + // TODO(antoyo) + None + } + + fn extend_scope_to_file(&self, _scope_metadata: Self::DIScope, _file: &SourceFile) -> Self::DIScope { + unimplemented!(); + } + + fn debuginfo_finalize(&self) { + // TODO(antoyo) + } + + fn create_dbg_var(&self, _variable_name: Symbol, _variable_type: Ty<'tcx>, _scope_metadata: Self::DIScope, _variable_kind: VariableKind, _span: Span) -> Self::DIVariable { + unimplemented!(); + } + + fn dbg_scope_fn(&self, _instance: Instance<'tcx>, _fn_abi: &FnAbi<'tcx, Ty<'tcx>>, _maybe_definition_llfn: Option<RValue<'gcc>>) -> Self::DIScope { + unimplemented!(); + } + + fn dbg_loc(&self, _scope: Self::DIScope, _inlined_at: Option<Self::DILocation>, _span: Span) -> Self::DILocation { + unimplemented!(); + } +} diff --git a/compiler/rustc_codegen_gcc/src/declare.rs b/compiler/rustc_codegen_gcc/src/declare.rs new file mode 100644 index 000000000..a619e2f77 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/declare.rs @@ -0,0 +1,145 @@ +use gccjit::{Function, FunctionType, GlobalKind, LValue, RValue, Type}; +use rustc_codegen_ssa::traits::BaseTypeMethods; +use rustc_middle::ty::Ty; +use rustc_span::Symbol; +use rustc_target::abi::call::FnAbi; + +use crate::abi::FnAbiGccExt; +use crate::context::CodegenCx; +use crate::intrinsic::llvm; + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn get_or_insert_global(&self, name: &str, ty: Type<'gcc>, is_tls: bool, link_section: Option<Symbol>) -> LValue<'gcc> { + if self.globals.borrow().contains_key(name) { + let typ = self.globals.borrow()[name].get_type(); + let global = self.context.new_global(None, GlobalKind::Imported, typ, name); + if is_tls { + global.set_tls_model(self.tls_model); + } + if let Some(link_section) = link_section { + global.set_link_section(link_section.as_str()); + } + global + } + else { + self.declare_global(name, ty, GlobalKind::Exported, is_tls, link_section) + } + } + + pub fn declare_unnamed_global(&self, ty: Type<'gcc>) -> LValue<'gcc> { + let name = self.generate_local_symbol_name("global"); + self.context.new_global(None, GlobalKind::Internal, ty, &name) + } + + pub fn declare_global_with_linkage(&self, name: &str, ty: Type<'gcc>, linkage: GlobalKind) -> LValue<'gcc> { + let global = self.context.new_global(None, linkage, ty, name); + let global_address = global.get_address(None); + self.globals.borrow_mut().insert(name.to_string(), global_address); + global + } + + /*pub fn declare_func(&self, name: &str, return_type: Type<'gcc>, params: &[Type<'gcc>], variadic: bool) -> RValue<'gcc> { + self.linkage.set(FunctionType::Exported); + let func = declare_raw_fn(self, name, () /*llvm::CCallConv*/, return_type, params, variadic); + // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API. + unsafe { std::mem::transmute(func) } + }*/ + + pub fn declare_global(&self, name: &str, ty: Type<'gcc>, global_kind: GlobalKind, is_tls: bool, link_section: Option<Symbol>) -> LValue<'gcc> { + let global = self.context.new_global(None, global_kind, ty, name); + if is_tls { + global.set_tls_model(self.tls_model); + } + if let Some(link_section) = link_section { + global.set_link_section(link_section.as_str()); + } + let global_address = global.get_address(None); + self.globals.borrow_mut().insert(name.to_string(), global_address); + global + } + + pub fn declare_private_global(&self, name: &str, ty: Type<'gcc>) -> LValue<'gcc> { + let global = self.context.new_global(None, GlobalKind::Internal, ty, name); + let global_address = global.get_address(None); + self.globals.borrow_mut().insert(name.to_string(), global_address); + global + } + + pub fn declare_cfn(&self, name: &str, _fn_type: Type<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): use the fn_type parameter. + let const_string = self.context.new_type::<u8>().make_pointer().make_pointer(); + let return_type = self.type_i32(); + let variadic = false; + self.linkage.set(FunctionType::Exported); + let func = declare_raw_fn(self, name, () /*llvm::CCallConv*/, return_type, &[self.type_i32(), const_string], variadic); + // NOTE: it is needed to set the current_func here as well, because get_fn() is not called + // for the main function. + *self.current_func.borrow_mut() = Some(func); + // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API. + unsafe { std::mem::transmute(func) } + } + + pub fn declare_fn(&self, name: &str, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> RValue<'gcc> { + let (return_type, params, variadic, on_stack_param_indices) = fn_abi.gcc_type(self); + let func = declare_raw_fn(self, name, () /*fn_abi.llvm_cconv()*/, return_type, ¶ms, variadic); + self.on_stack_function_params.borrow_mut().insert(func, on_stack_param_indices); + // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API. + unsafe { std::mem::transmute(func) } + } + + pub fn define_global(&self, name: &str, ty: Type<'gcc>, is_tls: bool, link_section: Option<Symbol>) -> LValue<'gcc> { + self.get_or_insert_global(name, ty, is_tls, link_section) + } + + pub fn get_declared_value(&self, name: &str) -> Option<RValue<'gcc>> { + // TODO(antoyo): use a different field than globals, because this seems to return a function? + self.globals.borrow().get(name).cloned() + } +} + +/// Declare a function. +/// +/// If there’s a value with the same name already declared, the function will +/// update the declaration and return existing Value instead. +fn declare_raw_fn<'gcc>(cx: &CodegenCx<'gcc, '_>, name: &str, _callconv: () /*llvm::CallConv*/, return_type: Type<'gcc>, param_types: &[Type<'gcc>], variadic: bool) -> Function<'gcc> { + if name.starts_with("llvm.") { + let intrinsic = llvm::intrinsic(name, cx); + cx.intrinsics.borrow_mut().insert(name.to_string(), intrinsic); + return intrinsic; + } + let func = + if cx.functions.borrow().contains_key(name) { + cx.functions.borrow()[name] + } + else { + let params: Vec<_> = param_types.into_iter().enumerate() + .map(|(index, param)| cx.context.new_parameter(None, *param, &format!("param{}", index))) // TODO(antoyo): set name. + .collect(); + let func = cx.context.new_function(None, cx.linkage.get(), return_type, ¶ms, mangle_name(name), variadic); + cx.functions.borrow_mut().insert(name.to_string(), func); + func + }; + + // TODO(antoyo): set function calling convention. + // TODO(antoyo): set unnamed address. + // TODO(antoyo): set no red zone function attribute. + // TODO(antoyo): set attributes for optimisation. + // TODO(antoyo): set attributes for non lazy bind. + + // FIXME(antoyo): invalid cast. + func +} + +// FIXME(antoyo): this is a hack because libgccjit currently only supports alpha, num and _. +// Unsupported characters: `$` and `.`. +pub fn mangle_name(name: &str) -> String { + name.replace(|char: char| { + if !char.is_alphanumeric() && char != '_' { + debug_assert!("$.".contains(char), "Unsupported char in function name: {}", char); + true + } + else { + false + } + }, "_") +} diff --git a/compiler/rustc_codegen_gcc/src/int.rs b/compiler/rustc_codegen_gcc/src/int.rs new file mode 100644 index 000000000..0c5dab004 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/int.rs @@ -0,0 +1,742 @@ +//! Module to handle integer operations. +//! This module exists because some integer types are not supported on some gcc platforms, e.g. +//! 128-bit integers on 32-bit platforms and thus require to be handled manually. + +use std::convert::TryFrom; + +use gccjit::{ComparisonOp, FunctionType, RValue, ToRValue, Type, UnaryOp, BinaryOp}; +use rustc_codegen_ssa::common::{IntPredicate, TypeKind}; +use rustc_codegen_ssa::traits::{BackendTypes, BaseTypeMethods, BuilderMethods, OverflowOp}; +use rustc_middle::ty::Ty; + +use crate::builder::ToGccComp; +use crate::{builder::Builder, common::{SignType, TypeReflection}, context::CodegenCx}; + +impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { + pub fn gcc_urem(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // 128-bit unsigned %: __umodti3 + self.multiplicative_operation(BinaryOp::Modulo, "mod", false, a, b) + } + + pub fn gcc_srem(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // 128-bit signed %: __modti3 + self.multiplicative_operation(BinaryOp::Modulo, "mod", true, a, b) + } + + pub fn gcc_not(&self, a: RValue<'gcc>) -> RValue<'gcc> { + let typ = a.get_type(); + if self.is_native_int_type_or_bool(typ) { + let operation = + if typ.is_bool() { + UnaryOp::LogicalNegate + } + else { + UnaryOp::BitwiseNegate + }; + self.cx.context.new_unary_op(None, operation, typ, a) + } + else { + // TODO(antoyo): use __negdi2 and __negti2 instead? + let element_type = typ.dyncast_array().expect("element type"); + let values = [ + self.cx.context.new_unary_op(None, UnaryOp::BitwiseNegate, element_type, self.low(a)), + self.cx.context.new_unary_op(None, UnaryOp::BitwiseNegate, element_type, self.high(a)), + ]; + self.cx.context.new_array_constructor(None, typ, &values) + } + } + + pub fn gcc_neg(&self, a: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + if self.is_native_int_type(a_type) { + self.cx.context.new_unary_op(None, UnaryOp::Minus, a.get_type(), a) + } + else { + let param_a = self.context.new_parameter(None, a_type, "a"); + let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a], "__negti2", false); + self.context.new_call(None, func, &[a]) + } + } + + pub fn gcc_and(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.cx.bitwise_operation(BinaryOp::BitwiseAnd, a, b) + } + + pub fn gcc_lshr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + let b_type = b.get_type(); + let a_native = self.is_native_int_type(a_type); + let b_native = self.is_native_int_type(b_type); + if a_native && b_native { + // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by a signed number. + // TODO(antoyo): cast to unsigned to do a logical shift if that does not work. + if a_type.is_signed(self) != b_type.is_signed(self) { + let b = self.context.new_cast(None, b, a_type); + a >> b + } + else { + a >> b + } + } + else if a_native && !b_native { + self.gcc_lshr(a, self.gcc_int_cast(b, a_type)) + } + else { + // NOTE: we cannot use the lshr builtin because it's calling hi() (to get the most + // significant half of the number) which uses lshr. + + let native_int_type = a_type.dyncast_array().expect("get element type"); + + let func = self.current_func(); + let then_block = func.new_block("then"); + let else_block = func.new_block("else"); + let after_block = func.new_block("after"); + let b0_block = func.new_block("b0"); + let actual_else_block = func.new_block("actual_else"); + + let result = func.new_local(None, a_type, "shiftResult"); + + let sixty_four = self.gcc_int(native_int_type, 64); + let sixty_three = self.gcc_int(native_int_type, 63); + let zero = self.gcc_zero(native_int_type); + let b = self.gcc_int_cast(b, native_int_type); + let condition = self.gcc_icmp(IntPredicate::IntNE, self.gcc_and(b, sixty_four), zero); + self.llbb().end_with_conditional(None, condition, then_block, else_block); + + // TODO(antoyo): take endianness into account. + let shift_value = self.gcc_sub(b, sixty_four); + let high = self.high(a); + let sign = + if a_type.is_signed(self) { + high >> sixty_three + } + else { + zero + }; + let values = [ + high >> shift_value, + sign, + ]; + let array_value = self.context.new_array_constructor(None, a_type, &values); + then_block.add_assignment(None, result, array_value); + then_block.end_with_jump(None, after_block); + + let condition = self.gcc_icmp(IntPredicate::IntEQ, b, zero); + else_block.end_with_conditional(None, condition, b0_block, actual_else_block); + + b0_block.add_assignment(None, result, a); + b0_block.end_with_jump(None, after_block); + + let shift_value = self.gcc_sub(sixty_four, b); + // NOTE: cast low to its unsigned type in order to perform a logical right shift. + let unsigned_type = native_int_type.to_unsigned(&self.cx); + let casted_low = self.context.new_cast(None, self.low(a), unsigned_type); + let shifted_low = casted_low >> self.context.new_cast(None, b, unsigned_type); + let shifted_low = self.context.new_cast(None, shifted_low, native_int_type); + let values = [ + (high << shift_value) | shifted_low, + high >> b, + ]; + let array_value = self.context.new_array_constructor(None, a_type, &values); + actual_else_block.add_assignment(None, result, array_value); + actual_else_block.end_with_jump(None, after_block); + + // NOTE: since jumps were added in a place rustc does not expect, the current block in the + // state need to be updated. + self.switch_to_block(after_block); + + result.to_rvalue() + } + } + + fn additive_operation(&self, operation: BinaryOp, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + let b_type = b.get_type(); + if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) { + if a_type != b_type { + if a_type.is_vector() { + // Vector types need to be bitcast. + // TODO(antoyo): perhaps use __builtin_convertvector for vector casting. + b = self.context.new_bitcast(None, b, a.get_type()); + } + else { + b = self.context.new_cast(None, b, a.get_type()); + } + } + self.context.new_binary_op(None, operation, a_type, a, b) + } + else { + let signed = a_type.is_compatible_with(self.i128_type); + let func_name = + match (operation, signed) { + (BinaryOp::Plus, true) => "__rust_i128_add", + (BinaryOp::Plus, false) => "__rust_u128_add", + (BinaryOp::Minus, true) => "__rust_i128_sub", + (BinaryOp::Minus, false) => "__rust_u128_sub", + _ => unreachable!("unexpected additive operation {:?}", operation), + }; + let param_a = self.context.new_parameter(None, a_type, "a"); + let param_b = self.context.new_parameter(None, b_type, "b"); + let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a, param_b], func_name, false); + self.context.new_call(None, func, &[a, b]) + } + } + + pub fn gcc_add(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.additive_operation(BinaryOp::Plus, a, b) + } + + pub fn gcc_mul(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.multiplicative_operation(BinaryOp::Mult, "mul", true, a, b) + } + + pub fn gcc_sub(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.additive_operation(BinaryOp::Minus, a, b) + } + + fn multiplicative_operation(&self, operation: BinaryOp, operation_name: &str, signed: bool, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + let b_type = b.get_type(); + if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) { + self.context.new_binary_op(None, operation, a_type, a, b) + } + else { + let sign = + if signed { + "" + } + else { + "u" + }; + let func_name = format!("__{}{}ti3", sign, operation_name); + let param_a = self.context.new_parameter(None, a_type, "a"); + let param_b = self.context.new_parameter(None, b_type, "b"); + let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a, param_b], func_name, false); + self.context.new_call(None, func, &[a, b]) + } + } + + pub fn gcc_sdiv(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): check if the types are signed? + // 128-bit, signed: __divti3 + // TODO(antoyo): convert the arguments to signed? + self.multiplicative_operation(BinaryOp::Divide, "div", true, a, b) + } + + pub fn gcc_udiv(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + // 128-bit, unsigned: __udivti3 + self.multiplicative_operation(BinaryOp::Divide, "div", false, a, b) + } + + pub fn gcc_checked_binop(&self, oop: OverflowOp, typ: Ty<'_>, lhs: <Self as BackendTypes>::Value, rhs: <Self as BackendTypes>::Value) -> (<Self as BackendTypes>::Value, <Self as BackendTypes>::Value) { + use rustc_middle::ty::{Int, IntTy::*, Uint, UintTy::*}; + + let new_kind = + match typ.kind() { + Int(t @ Isize) => Int(t.normalize(self.tcx.sess.target.pointer_width)), + Uint(t @ Usize) => Uint(t.normalize(self.tcx.sess.target.pointer_width)), + t @ (Uint(_) | Int(_)) => t.clone(), + _ => panic!("tried to get overflow intrinsic for op applied to non-int type"), + }; + + // TODO(antoyo): remove duplication with intrinsic? + let name = + if self.is_native_int_type(lhs.get_type()) { + match oop { + OverflowOp::Add => + match new_kind { + Int(I8) => "__builtin_add_overflow", + Int(I16) => "__builtin_add_overflow", + Int(I32) => "__builtin_sadd_overflow", + Int(I64) => "__builtin_saddll_overflow", + Int(I128) => "__builtin_add_overflow", + + Uint(U8) => "__builtin_add_overflow", + Uint(U16) => "__builtin_add_overflow", + Uint(U32) => "__builtin_uadd_overflow", + Uint(U64) => "__builtin_uaddll_overflow", + Uint(U128) => "__builtin_add_overflow", + + _ => unreachable!(), + }, + OverflowOp::Sub => + match new_kind { + Int(I8) => "__builtin_sub_overflow", + Int(I16) => "__builtin_sub_overflow", + Int(I32) => "__builtin_ssub_overflow", + Int(I64) => "__builtin_ssubll_overflow", + Int(I128) => "__builtin_sub_overflow", + + Uint(U8) => "__builtin_sub_overflow", + Uint(U16) => "__builtin_sub_overflow", + Uint(U32) => "__builtin_usub_overflow", + Uint(U64) => "__builtin_usubll_overflow", + Uint(U128) => "__builtin_sub_overflow", + + _ => unreachable!(), + }, + OverflowOp::Mul => + match new_kind { + Int(I8) => "__builtin_mul_overflow", + Int(I16) => "__builtin_mul_overflow", + Int(I32) => "__builtin_smul_overflow", + Int(I64) => "__builtin_smulll_overflow", + Int(I128) => "__builtin_mul_overflow", + + Uint(U8) => "__builtin_mul_overflow", + Uint(U16) => "__builtin_mul_overflow", + Uint(U32) => "__builtin_umul_overflow", + Uint(U64) => "__builtin_umulll_overflow", + Uint(U128) => "__builtin_mul_overflow", + + _ => unreachable!(), + }, + } + } + else { + match new_kind { + Int(I128) | Uint(U128) => { + let func_name = + match oop { + OverflowOp::Add => + match new_kind { + Int(I128) => "__rust_i128_addo", + Uint(U128) => "__rust_u128_addo", + _ => unreachable!(), + }, + OverflowOp::Sub => + match new_kind { + Int(I128) => "__rust_i128_subo", + Uint(U128) => "__rust_u128_subo", + _ => unreachable!(), + }, + OverflowOp::Mul => + match new_kind { + Int(I128) => "__rust_i128_mulo", // TODO(antoyo): use __muloti4d instead? + Uint(U128) => "__rust_u128_mulo", + _ => unreachable!(), + }, + }; + let a_type = lhs.get_type(); + let b_type = rhs.get_type(); + let param_a = self.context.new_parameter(None, a_type, "a"); + let param_b = self.context.new_parameter(None, b_type, "b"); + let result_field = self.context.new_field(None, a_type, "result"); + let overflow_field = self.context.new_field(None, self.bool_type, "overflow"); + let return_type = self.context.new_struct_type(None, "result_overflow", &[result_field, overflow_field]); + let func = self.context.new_function(None, FunctionType::Extern, return_type.as_type(), &[param_a, param_b], func_name, false); + let result = self.context.new_call(None, func, &[lhs, rhs]); + let overflow = result.access_field(None, overflow_field); + let int_result = result.access_field(None, result_field); + return (int_result, overflow); + }, + _ => { + match oop { + OverflowOp::Mul => + match new_kind { + Int(I32) => "__mulosi4", + Int(I64) => "__mulodi4", + _ => unreachable!(), + }, + _ => unimplemented!("overflow operation for {:?}", new_kind), + } + } + } + }; + + let intrinsic = self.context.get_builtin_function(&name); + let res = self.current_func() + // TODO(antoyo): is it correct to use rhs type instead of the parameter typ? + .new_local(None, rhs.get_type(), "binopResult") + .get_address(None); + let overflow = self.overflow_call(intrinsic, &[lhs, rhs, res], None); + (res.dereference(None).to_rvalue(), overflow) + } + + pub fn gcc_icmp(&self, op: IntPredicate, mut lhs: RValue<'gcc>, mut rhs: RValue<'gcc>) -> RValue<'gcc> { + let a_type = lhs.get_type(); + let b_type = rhs.get_type(); + if self.is_non_native_int_type(a_type) || self.is_non_native_int_type(b_type) { + let signed = a_type.is_compatible_with(self.i128_type); + let sign = + if signed { + "" + } + else { + "u" + }; + let func_name = format!("__{}cmpti2", sign); + let param_a = self.context.new_parameter(None, a_type, "a"); + let param_b = self.context.new_parameter(None, b_type, "b"); + let func = self.context.new_function(None, FunctionType::Extern, self.int_type, &[param_a, param_b], func_name, false); + let cmp = self.context.new_call(None, func, &[lhs, rhs]); + let (op, limit) = + match op { + IntPredicate::IntEQ => { + return self.context.new_comparison(None, ComparisonOp::Equals, cmp, self.context.new_rvalue_one(self.int_type)); + }, + IntPredicate::IntNE => { + return self.context.new_comparison(None, ComparisonOp::NotEquals, cmp, self.context.new_rvalue_one(self.int_type)); + }, + IntPredicate::IntUGT => (ComparisonOp::Equals, 2), + IntPredicate::IntUGE => (ComparisonOp::GreaterThanEquals, 1), + IntPredicate::IntULT => (ComparisonOp::Equals, 0), + IntPredicate::IntULE => (ComparisonOp::LessThanEquals, 1), + IntPredicate::IntSGT => (ComparisonOp::Equals, 2), + IntPredicate::IntSGE => (ComparisonOp::GreaterThanEquals, 1), + IntPredicate::IntSLT => (ComparisonOp::Equals, 0), + IntPredicate::IntSLE => (ComparisonOp::LessThanEquals, 1), + }; + self.context.new_comparison(None, op, cmp, self.context.new_rvalue_from_int(self.int_type, limit)) + } + else { + let left_type = lhs.get_type(); + let right_type = rhs.get_type(); + if left_type != right_type { + // NOTE: because libgccjit cannot compare function pointers. + if left_type.dyncast_function_ptr_type().is_some() && right_type.dyncast_function_ptr_type().is_some() { + lhs = self.context.new_cast(None, lhs, self.usize_type.make_pointer()); + rhs = self.context.new_cast(None, rhs, self.usize_type.make_pointer()); + } + // NOTE: hack because we try to cast a vector type to the same vector type. + else if format!("{:?}", left_type) != format!("{:?}", right_type) { + rhs = self.context.new_cast(None, rhs, left_type); + } + } + self.context.new_comparison(None, op.to_gcc_comparison(), lhs, rhs) + } + } + + pub fn gcc_xor(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + let b_type = b.get_type(); + if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) { + a ^ b + } + else { + let values = [ + self.low(a) ^ self.low(b), + self.high(a) ^ self.high(b), + ]; + self.context.new_array_constructor(None, a_type, &values) + } + } + + pub fn gcc_shl(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + let b_type = b.get_type(); + let a_native = self.is_native_int_type(a_type); + let b_native = self.is_native_int_type(b_type); + if a_native && b_native { + // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number. + if a_type.is_unsigned(self) && b_type.is_signed(self) { + let a = self.context.new_cast(None, a, b_type); + let result = a << b; + self.context.new_cast(None, result, a_type) + } + else if a_type.is_signed(self) && b_type.is_unsigned(self) { + let b = self.context.new_cast(None, b, a_type); + a << b + } + else { + a << b + } + } + else if a_native && !b_native { + self.gcc_shl(a, self.gcc_int_cast(b, a_type)) + } + else { + // NOTE: we cannot use the ashl builtin because it's calling widen_hi() which uses ashl. + let native_int_type = a_type.dyncast_array().expect("get element type"); + + let func = self.current_func(); + let then_block = func.new_block("then"); + let else_block = func.new_block("else"); + let after_block = func.new_block("after"); + let b0_block = func.new_block("b0"); + let actual_else_block = func.new_block("actual_else"); + + let result = func.new_local(None, a_type, "shiftResult"); + + let b = self.gcc_int_cast(b, native_int_type); + let sixty_four = self.gcc_int(native_int_type, 64); + let zero = self.gcc_zero(native_int_type); + let condition = self.gcc_icmp(IntPredicate::IntNE, self.gcc_and(b, sixty_four), zero); + self.llbb().end_with_conditional(None, condition, then_block, else_block); + + // TODO(antoyo): take endianness into account. + let values = [ + zero, + self.low(a) << (b - sixty_four), + ]; + let array_value = self.context.new_array_constructor(None, a_type, &values); + then_block.add_assignment(None, result, array_value); + then_block.end_with_jump(None, after_block); + + let condition = self.gcc_icmp(IntPredicate::IntEQ, b, zero); + else_block.end_with_conditional(None, condition, b0_block, actual_else_block); + + b0_block.add_assignment(None, result, a); + b0_block.end_with_jump(None, after_block); + + // NOTE: cast low to its unsigned type in order to perform a logical right shift. + let unsigned_type = native_int_type.to_unsigned(&self.cx); + let casted_low = self.context.new_cast(None, self.low(a), unsigned_type); + let shift_value = self.context.new_cast(None, sixty_four - b, unsigned_type); + let high_low = self.context.new_cast(None, casted_low >> shift_value, native_int_type); + let values = [ + self.low(a) << b, + (self.high(a) << b) | high_low, + ]; + + let array_value = self.context.new_array_constructor(None, a_type, &values); + actual_else_block.add_assignment(None, result, array_value); + actual_else_block.end_with_jump(None, after_block); + + // NOTE: since jumps were added in a place rustc does not expect, the current block in the + // state need to be updated. + self.switch_to_block(after_block); + + result.to_rvalue() + } + } + + pub fn gcc_bswap(&mut self, mut arg: RValue<'gcc>, width: u64) -> RValue<'gcc> { + let arg_type = arg.get_type(); + if !self.is_native_int_type(arg_type) { + let native_int_type = arg_type.dyncast_array().expect("get element type"); + let lsb = self.context.new_array_access(None, arg, self.context.new_rvalue_from_int(self.int_type, 0)).to_rvalue(); + let swapped_lsb = self.gcc_bswap(lsb, width / 2); + let swapped_lsb = self.context.new_cast(None, swapped_lsb, native_int_type); + let msb = self.context.new_array_access(None, arg, self.context.new_rvalue_from_int(self.int_type, 1)).to_rvalue(); + let swapped_msb = self.gcc_bswap(msb, width / 2); + let swapped_msb = self.context.new_cast(None, swapped_msb, native_int_type); + + // NOTE: we also need to swap the two elements here, in addition to swapping inside + // the elements themselves like done above. + return self.context.new_array_constructor(None, arg_type, &[swapped_msb, swapped_lsb]); + } + + // TODO(antoyo): check if it's faster to use string literals and a + // match instead of format!. + let bswap = self.cx.context.get_builtin_function(&format!("__builtin_bswap{}", width)); + // FIXME(antoyo): this cast should not be necessary. Remove + // when having proper sized integer types. + let param_type = bswap.get_param(0).to_rvalue().get_type(); + if param_type != arg_type { + arg = self.bitcast(arg, param_type); + } + self.cx.context.new_call(None, bswap, &[arg]) + } +} + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn gcc_int(&self, typ: Type<'gcc>, int: i64) -> RValue<'gcc> { + if self.is_native_int_type_or_bool(typ) { + self.context.new_rvalue_from_long(typ, i64::try_from(int).expect("i64::try_from")) + } + else { + // NOTE: set the sign in high. + self.from_low_high(typ, int, -(int.is_negative() as i64)) + } + } + + pub fn gcc_uint(&self, typ: Type<'gcc>, int: u64) -> RValue<'gcc> { + if self.is_native_int_type_or_bool(typ) { + self.context.new_rvalue_from_long(typ, u64::try_from(int).expect("u64::try_from") as i64) + } + else { + self.from_low_high(typ, int as i64, 0) + } + } + + pub fn gcc_uint_big(&self, typ: Type<'gcc>, num: u128) -> RValue<'gcc> { + let low = num as u64; + let high = (num >> 64) as u64; + if num >> 64 != 0 { + // FIXME(antoyo): use a new function new_rvalue_from_unsigned_long()? + if self.is_native_int_type(typ) { + let low = self.context.new_rvalue_from_long(self.u64_type, low as i64); + let high = self.context.new_rvalue_from_long(typ, high as i64); + + let sixty_four = self.context.new_rvalue_from_long(typ, 64); + let shift = high << sixty_four; + shift | self.context.new_cast(None, low, typ) + } + else { + self.from_low_high(typ, low as i64, high as i64) + } + } + else if typ.is_i128(self) { + let num = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64); + self.gcc_int_cast(num, typ) + } + else { + self.gcc_uint(typ, num as u64) + } + } + + pub fn gcc_zero(&self, typ: Type<'gcc>) -> RValue<'gcc> { + if self.is_native_int_type_or_bool(typ) { + self.context.new_rvalue_zero(typ) + } + else { + self.from_low_high(typ, 0, 0) + } + } + + pub fn gcc_int_width(&self, typ: Type<'gcc>) -> u64 { + if self.is_native_int_type_or_bool(typ) { + typ.get_size() as u64 * 8 + } + else { + // NOTE: the only unsupported types are u128 and i128. + 128 + } + } + + fn bitwise_operation(&self, operation: BinaryOp, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> { + let a_type = a.get_type(); + let b_type = b.get_type(); + let a_native = self.is_native_int_type_or_bool(a_type); + let b_native = self.is_native_int_type_or_bool(b_type); + if a_type.is_vector() && b_type.is_vector() { + self.context.new_binary_op(None, operation, a_type, a, b) + } + else if a_native && b_native { + if a_type != b_type { + b = self.context.new_cast(None, b, a_type); + } + self.context.new_binary_op(None, operation, a_type, a, b) + } + else { + assert!(!a_native && !b_native, "both types should either be native or non-native for or operation"); + let native_int_type = a_type.dyncast_array().expect("get element type"); + let values = [ + self.context.new_binary_op(None, operation, native_int_type, self.low(a), self.low(b)), + self.context.new_binary_op(None, operation, native_int_type, self.high(a), self.high(b)), + ]; + self.context.new_array_constructor(None, a_type, &values) + } + } + + pub fn gcc_or(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { + self.bitwise_operation(BinaryOp::BitwiseOr, a, b) + } + + // TODO(antoyo): can we use https://github.com/rust-lang/compiler-builtins/blob/master/src/int/mod.rs#L379 instead? + pub fn gcc_int_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + let value_type = value.get_type(); + if self.is_native_int_type_or_bool(dest_typ) && self.is_native_int_type_or_bool(value_type) { + self.context.new_cast(None, value, dest_typ) + } + else if self.is_native_int_type_or_bool(dest_typ) { + self.context.new_cast(None, self.low(value), dest_typ) + } + else if self.is_native_int_type_or_bool(value_type) { + let dest_element_type = dest_typ.dyncast_array().expect("get element type"); + + // NOTE: set the sign of the value. + let zero = self.context.new_rvalue_zero(value_type); + let is_negative = self.context.new_comparison(None, ComparisonOp::LessThan, value, zero); + let is_negative = self.gcc_int_cast(is_negative, dest_element_type); + let values = [ + self.context.new_cast(None, value, dest_element_type), + self.context.new_unary_op(None, UnaryOp::Minus, dest_element_type, is_negative), + ]; + self.context.new_array_constructor(None, dest_typ, &values) + } + else { + // Since u128 and i128 are the only types that can be unsupported, we know the type of + // value and the destination type have the same size, so a bitcast is fine. + + // TODO(antoyo): perhaps use __builtin_convertvector for vector casting. + self.context.new_bitcast(None, value, dest_typ) + } + } + + fn int_to_float_cast(&self, signed: bool, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + let value_type = value.get_type(); + if self.is_native_int_type_or_bool(value_type) { + return self.context.new_cast(None, value, dest_typ); + } + + let name_suffix = + match self.type_kind(dest_typ) { + TypeKind::Float => "tisf", + TypeKind::Double => "tidf", + kind => panic!("cannot cast a non-native integer to type {:?}", kind), + }; + let sign = + if signed { + "" + } + else { + "un" + }; + let func_name = format!("__float{}{}", sign, name_suffix); + let param = self.context.new_parameter(None, value_type, "n"); + let func = self.context.new_function(None, FunctionType::Extern, dest_typ, &[param], func_name, false); + self.context.new_call(None, func, &[value]) + } + + pub fn gcc_int_to_float_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + self.int_to_float_cast(true, value, dest_typ) + } + + pub fn gcc_uint_to_float_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + self.int_to_float_cast(false, value, dest_typ) + } + + fn float_to_int_cast(&self, signed: bool, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + let value_type = value.get_type(); + if self.is_native_int_type_or_bool(dest_typ) { + return self.context.new_cast(None, value, dest_typ); + } + + let name_suffix = + match self.type_kind(value_type) { + TypeKind::Float => "sfti", + TypeKind::Double => "dfti", + kind => panic!("cannot cast a {:?} to non-native integer", kind), + }; + let sign = + if signed { + "" + } + else { + "uns" + }; + let func_name = format!("__fix{}{}", sign, name_suffix); + let param = self.context.new_parameter(None, value_type, "n"); + let func = self.context.new_function(None, FunctionType::Extern, dest_typ, &[param], func_name, false); + self.context.new_call(None, func, &[value]) + } + + pub fn gcc_float_to_int_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + self.float_to_int_cast(true, value, dest_typ) + } + + pub fn gcc_float_to_uint_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> { + self.float_to_int_cast(false, value, dest_typ) + } + + fn high(&self, value: RValue<'gcc>) -> RValue<'gcc> { + self.context.new_array_access(None, value, self.context.new_rvalue_from_int(self.int_type, 1)) + .to_rvalue() + } + + fn low(&self, value: RValue<'gcc>) -> RValue<'gcc> { + self.context.new_array_access(None, value, self.context.new_rvalue_from_int(self.int_type, 0)) + .to_rvalue() + } + + fn from_low_high(&self, typ: Type<'gcc>, low: i64, high: i64) -> RValue<'gcc> { + let native_int_type = typ.dyncast_array().expect("get element type"); + let values = [ + self.context.new_rvalue_from_long(native_int_type, low), + self.context.new_rvalue_from_long(native_int_type, high), + ]; + self.context.new_array_constructor(None, typ, &values) + } +} diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs b/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs new file mode 100644 index 000000000..fb6c38fa0 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs @@ -0,0 +1,5722 @@ +// File generated by `rustc_codegen_gcc/tools/generate_intrinsics.py` +// DO NOT EDIT IT! +match name { + // AMDGPU + "llvm.AMDGPU.div.fixup.f32" => "__builtin_amdgpu_div_fixup", + "llvm.AMDGPU.div.fixup.f64" => "__builtin_amdgpu_div_fixup", + "llvm.AMDGPU.div.fixup.v2f64" => "__builtin_amdgpu_div_fixup", + "llvm.AMDGPU.div.fixup.v4f32" => "__builtin_amdgpu_div_fixup", + "llvm.AMDGPU.div.fmas.f32" => "__builtin_amdgpu_div_fmas", + "llvm.AMDGPU.div.fmas.f64" => "__builtin_amdgpu_div_fmas", + "llvm.AMDGPU.div.fmas.v2f64" => "__builtin_amdgpu_div_fmas", + "llvm.AMDGPU.div.fmas.v4f32" => "__builtin_amdgpu_div_fmas", + "llvm.AMDGPU.ldexp.f32" => "__builtin_amdgpu_ldexp", + "llvm.AMDGPU.ldexp.f64" => "__builtin_amdgpu_ldexp", + "llvm.AMDGPU.ldexp.v2f64" => "__builtin_amdgpu_ldexp", + "llvm.AMDGPU.ldexp.v4f32" => "__builtin_amdgpu_ldexp", + "llvm.AMDGPU.rcp.f32" => "__builtin_amdgpu_rcp", + "llvm.AMDGPU.rcp.f64" => "__builtin_amdgpu_rcp", + "llvm.AMDGPU.rcp.v2f64" => "__builtin_amdgpu_rcp", + "llvm.AMDGPU.rcp.v4f32" => "__builtin_amdgpu_rcp", + "llvm.AMDGPU.rsq.clamped.f32" => "__builtin_amdgpu_rsq_clamped", + "llvm.AMDGPU.rsq.clamped.f64" => "__builtin_amdgpu_rsq_clamped", + "llvm.AMDGPU.rsq.clamped.v2f64" => "__builtin_amdgpu_rsq_clamped", + "llvm.AMDGPU.rsq.clamped.v4f32" => "__builtin_amdgpu_rsq_clamped", + "llvm.AMDGPU.rsq.f32" => "__builtin_amdgpu_rsq", + "llvm.AMDGPU.rsq.f64" => "__builtin_amdgpu_rsq", + "llvm.AMDGPU.rsq.v2f64" => "__builtin_amdgpu_rsq", + "llvm.AMDGPU.rsq.v4f32" => "__builtin_amdgpu_rsq", + "llvm.AMDGPU.trig.preop.f32" => "__builtin_amdgpu_trig_preop", + "llvm.AMDGPU.trig.preop.f64" => "__builtin_amdgpu_trig_preop", + "llvm.AMDGPU.trig.preop.v2f64" => "__builtin_amdgpu_trig_preop", + "llvm.AMDGPU.trig.preop.v4f32" => "__builtin_amdgpu_trig_preop", + // aarch64 + "llvm.aarch64.dmb" => "__builtin_arm_dmb", + "llvm.aarch64.dsb" => "__builtin_arm_dsb", + "llvm.aarch64.isb" => "__builtin_arm_isb", + "llvm.aarch64.sve.aesd" => "__builtin_sve_svaesd_u8", + "llvm.aarch64.sve.aese" => "__builtin_sve_svaese_u8", + "llvm.aarch64.sve.aesimc" => "__builtin_sve_svaesimc_u8", + "llvm.aarch64.sve.aesmc" => "__builtin_sve_svaesmc_u8", + "llvm.aarch64.sve.rax1" => "__builtin_sve_svrax1_u64", + "llvm.aarch64.sve.rdffr" => "__builtin_sve_svrdffr", + "llvm.aarch64.sve.rdffr.z" => "__builtin_sve_svrdffr_z", + "llvm.aarch64.sve.setffr" => "__builtin_sve_svsetffr", + "llvm.aarch64.sve.sm4e" => "__builtin_sve_svsm4e_u32", + "llvm.aarch64.sve.sm4ekey" => "__builtin_sve_svsm4ekey_u32", + "llvm.aarch64.sve.wrffr" => "__builtin_sve_svwrffr", + "llvm.aarch64.tcancel" => "__builtin_arm_tcancel", + "llvm.aarch64.tcommit" => "__builtin_arm_tcommit", + "llvm.aarch64.tstart" => "__builtin_arm_tstart", + "llvm.aarch64.ttest" => "__builtin_arm_ttest", + // amdgcn + "llvm.amdgcn.alignbyte" => "__builtin_amdgcn_alignbyte", + "llvm.amdgcn.buffer.wbinvl1" => "__builtin_amdgcn_buffer_wbinvl1", + "llvm.amdgcn.buffer.wbinvl1.sc" => "__builtin_amdgcn_buffer_wbinvl1_sc", + "llvm.amdgcn.buffer.wbinvl1.vol" => "__builtin_amdgcn_buffer_wbinvl1_vol", + "llvm.amdgcn.cubeid" => "__builtin_amdgcn_cubeid", + "llvm.amdgcn.cubema" => "__builtin_amdgcn_cubema", + "llvm.amdgcn.cubesc" => "__builtin_amdgcn_cubesc", + "llvm.amdgcn.cubetc" => "__builtin_amdgcn_cubetc", + "llvm.amdgcn.cvt.pk.i16" => "__builtin_amdgcn_cvt_pk_i16", + "llvm.amdgcn.cvt.pk.u16" => "__builtin_amdgcn_cvt_pk_u16", + "llvm.amdgcn.cvt.pk.u8.f32" => "__builtin_amdgcn_cvt_pk_u8_f32", + "llvm.amdgcn.cvt.pknorm.i16" => "__builtin_amdgcn_cvt_pknorm_i16", + "llvm.amdgcn.cvt.pknorm.u16" => "__builtin_amdgcn_cvt_pknorm_u16", + "llvm.amdgcn.cvt.pkrtz" => "__builtin_amdgcn_cvt_pkrtz", + "llvm.amdgcn.dispatch.id" => "__builtin_amdgcn_dispatch_id", + "llvm.amdgcn.ds.bpermute" => "__builtin_amdgcn_ds_bpermute", + "llvm.amdgcn.ds.fadd.v2bf16" => "__builtin_amdgcn_ds_atomic_fadd_v2bf16", + "llvm.amdgcn.ds.gws.barrier" => "__builtin_amdgcn_ds_gws_barrier", + "llvm.amdgcn.ds.gws.init" => "__builtin_amdgcn_ds_gws_init", + "llvm.amdgcn.ds.gws.sema.br" => "__builtin_amdgcn_ds_gws_sema_br", + "llvm.amdgcn.ds.gws.sema.p" => "__builtin_amdgcn_ds_gws_sema_p", + "llvm.amdgcn.ds.gws.sema.release.all" => "__builtin_amdgcn_ds_gws_sema_release_all", + "llvm.amdgcn.ds.gws.sema.v" => "__builtin_amdgcn_ds_gws_sema_v", + "llvm.amdgcn.ds.permute" => "__builtin_amdgcn_ds_permute", + "llvm.amdgcn.ds.swizzle" => "__builtin_amdgcn_ds_swizzle", + "llvm.amdgcn.endpgm" => "__builtin_amdgcn_endpgm", + "llvm.amdgcn.fdot2" => "__builtin_amdgcn_fdot2", + "llvm.amdgcn.fmed3" => "__builtin_amdgcn_fmed3", + "llvm.amdgcn.fmul.legacy" => "__builtin_amdgcn_fmul_legacy", + "llvm.amdgcn.groupstaticsize" => "__builtin_amdgcn_groupstaticsize", + "llvm.amdgcn.implicit.buffer.ptr" => "__builtin_amdgcn_implicit_buffer_ptr", + "llvm.amdgcn.implicitarg.ptr" => "__builtin_amdgcn_implicitarg_ptr", + "llvm.amdgcn.interp.mov" => "__builtin_amdgcn_interp_mov", + "llvm.amdgcn.interp.p1" => "__builtin_amdgcn_interp_p1", + "llvm.amdgcn.interp.p1.f16" => "__builtin_amdgcn_interp_p1_f16", + "llvm.amdgcn.interp.p2" => "__builtin_amdgcn_interp_p2", + "llvm.amdgcn.interp.p2.f16" => "__builtin_amdgcn_interp_p2_f16", + "llvm.amdgcn.is.private" => "__builtin_amdgcn_is_private", + "llvm.amdgcn.is.shared" => "__builtin_amdgcn_is_shared", + "llvm.amdgcn.kernarg.segment.ptr" => "__builtin_amdgcn_kernarg_segment_ptr", + "llvm.amdgcn.lerp" => "__builtin_amdgcn_lerp", + "llvm.amdgcn.mbcnt.hi" => "__builtin_amdgcn_mbcnt_hi", + "llvm.amdgcn.mbcnt.lo" => "__builtin_amdgcn_mbcnt_lo", + "llvm.amdgcn.mqsad.pk.u16.u8" => "__builtin_amdgcn_mqsad_pk_u16_u8", + "llvm.amdgcn.mqsad.u32.u8" => "__builtin_amdgcn_mqsad_u32_u8", + "llvm.amdgcn.msad.u8" => "__builtin_amdgcn_msad_u8", + "llvm.amdgcn.perm" => "__builtin_amdgcn_perm", + "llvm.amdgcn.permlane16" => "__builtin_amdgcn_permlane16", + "llvm.amdgcn.permlanex16" => "__builtin_amdgcn_permlanex16", + "llvm.amdgcn.qsad.pk.u16.u8" => "__builtin_amdgcn_qsad_pk_u16_u8", + "llvm.amdgcn.queue.ptr" => "__builtin_amdgcn_queue_ptr", + "llvm.amdgcn.rcp.legacy" => "__builtin_amdgcn_rcp_legacy", + "llvm.amdgcn.readfirstlane" => "__builtin_amdgcn_readfirstlane", + "llvm.amdgcn.readlane" => "__builtin_amdgcn_readlane", + "llvm.amdgcn.rsq.legacy" => "__builtin_amdgcn_rsq_legacy", + "llvm.amdgcn.s.barrier" => "__builtin_amdgcn_s_barrier", + "llvm.amdgcn.s.dcache.inv" => "__builtin_amdgcn_s_dcache_inv", + "llvm.amdgcn.s.dcache.inv.vol" => "__builtin_amdgcn_s_dcache_inv_vol", + "llvm.amdgcn.s.dcache.wb" => "__builtin_amdgcn_s_dcache_wb", + "llvm.amdgcn.s.dcache.wb.vol" => "__builtin_amdgcn_s_dcache_wb_vol", + "llvm.amdgcn.s.decperflevel" => "__builtin_amdgcn_s_decperflevel", + "llvm.amdgcn.s.get.waveid.in.workgroup" => "__builtin_amdgcn_s_get_waveid_in_workgroup", + "llvm.amdgcn.s.getpc" => "__builtin_amdgcn_s_getpc", + "llvm.amdgcn.s.getreg" => "__builtin_amdgcn_s_getreg", + "llvm.amdgcn.s.incperflevel" => "__builtin_amdgcn_s_incperflevel", + "llvm.amdgcn.s.memrealtime" => "__builtin_amdgcn_s_memrealtime", + "llvm.amdgcn.s.memtime" => "__builtin_amdgcn_s_memtime", + "llvm.amdgcn.s.sendmsg" => "__builtin_amdgcn_s_sendmsg", + "llvm.amdgcn.s.sendmsghalt" => "__builtin_amdgcn_s_sendmsghalt", + "llvm.amdgcn.s.setprio" => "__builtin_amdgcn_s_setprio", + "llvm.amdgcn.s.setreg" => "__builtin_amdgcn_s_setreg", + "llvm.amdgcn.s.sleep" => "__builtin_amdgcn_s_sleep", + "llvm.amdgcn.s.waitcnt" => "__builtin_amdgcn_s_waitcnt", + "llvm.amdgcn.sad.hi.u8" => "__builtin_amdgcn_sad_hi_u8", + "llvm.amdgcn.sad.u16" => "__builtin_amdgcn_sad_u16", + "llvm.amdgcn.sad.u8" => "__builtin_amdgcn_sad_u8", + "llvm.amdgcn.sched.barrier" => "__builtin_amdgcn_sched_barrier", + "llvm.amdgcn.sdot2" => "__builtin_amdgcn_sdot2", + "llvm.amdgcn.sdot4" => "__builtin_amdgcn_sdot4", + "llvm.amdgcn.sdot8" => "__builtin_amdgcn_sdot8", + "llvm.amdgcn.udot2" => "__builtin_amdgcn_udot2", + "llvm.amdgcn.udot4" => "__builtin_amdgcn_udot4", + "llvm.amdgcn.udot8" => "__builtin_amdgcn_udot8", + "llvm.amdgcn.wave.barrier" => "__builtin_amdgcn_wave_barrier", + "llvm.amdgcn.wavefrontsize" => "__builtin_amdgcn_wavefrontsize", + "llvm.amdgcn.writelane" => "__builtin_amdgcn_writelane", + // arm + "llvm.arm.cdp" => "__builtin_arm_cdp", + "llvm.arm.cdp2" => "__builtin_arm_cdp2", + "llvm.arm.cmse.tt" => "__builtin_arm_cmse_TT", + "llvm.arm.cmse.tta" => "__builtin_arm_cmse_TTA", + "llvm.arm.cmse.ttat" => "__builtin_arm_cmse_TTAT", + "llvm.arm.cmse.ttt" => "__builtin_arm_cmse_TTT", + "llvm.arm.dmb" => "__builtin_arm_dmb", + "llvm.arm.dsb" => "__builtin_arm_dsb", + "llvm.arm.get.fpscr" => "__builtin_arm_get_fpscr", + "llvm.arm.isb" => "__builtin_arm_isb", + "llvm.arm.ldc" => "__builtin_arm_ldc", + "llvm.arm.ldc2" => "__builtin_arm_ldc2", + "llvm.arm.ldc2l" => "__builtin_arm_ldc2l", + "llvm.arm.ldcl" => "__builtin_arm_ldcl", + "llvm.arm.mcr" => "__builtin_arm_mcr", + "llvm.arm.mcr2" => "__builtin_arm_mcr2", + "llvm.arm.mcrr" => "__builtin_arm_mcrr", + "llvm.arm.mcrr2" => "__builtin_arm_mcrr2", + "llvm.arm.mrc" => "__builtin_arm_mrc", + "llvm.arm.mrc2" => "__builtin_arm_mrc2", + "llvm.arm.qadd" => "__builtin_arm_qadd", + "llvm.arm.qadd16" => "__builtin_arm_qadd16", + "llvm.arm.qadd8" => "__builtin_arm_qadd8", + "llvm.arm.qasx" => "__builtin_arm_qasx", + "llvm.arm.qsax" => "__builtin_arm_qsax", + "llvm.arm.qsub" => "__builtin_arm_qsub", + "llvm.arm.qsub16" => "__builtin_arm_qsub16", + "llvm.arm.qsub8" => "__builtin_arm_qsub8", + "llvm.arm.sadd16" => "__builtin_arm_sadd16", + "llvm.arm.sadd8" => "__builtin_arm_sadd8", + "llvm.arm.sasx" => "__builtin_arm_sasx", + "llvm.arm.sel" => "__builtin_arm_sel", + "llvm.arm.set.fpscr" => "__builtin_arm_set_fpscr", + "llvm.arm.shadd16" => "__builtin_arm_shadd16", + "llvm.arm.shadd8" => "__builtin_arm_shadd8", + "llvm.arm.shasx" => "__builtin_arm_shasx", + "llvm.arm.shsax" => "__builtin_arm_shsax", + "llvm.arm.shsub16" => "__builtin_arm_shsub16", + "llvm.arm.shsub8" => "__builtin_arm_shsub8", + "llvm.arm.smlabb" => "__builtin_arm_smlabb", + "llvm.arm.smlabt" => "__builtin_arm_smlabt", + "llvm.arm.smlad" => "__builtin_arm_smlad", + "llvm.arm.smladx" => "__builtin_arm_smladx", + "llvm.arm.smlald" => "__builtin_arm_smlald", + "llvm.arm.smlaldx" => "__builtin_arm_smlaldx", + "llvm.arm.smlatb" => "__builtin_arm_smlatb", + "llvm.arm.smlatt" => "__builtin_arm_smlatt", + "llvm.arm.smlawb" => "__builtin_arm_smlawb", + "llvm.arm.smlawt" => "__builtin_arm_smlawt", + "llvm.arm.smlsd" => "__builtin_arm_smlsd", + "llvm.arm.smlsdx" => "__builtin_arm_smlsdx", + "llvm.arm.smlsld" => "__builtin_arm_smlsld", + "llvm.arm.smlsldx" => "__builtin_arm_smlsldx", + "llvm.arm.smuad" => "__builtin_arm_smuad", + "llvm.arm.smuadx" => "__builtin_arm_smuadx", + "llvm.arm.smulbb" => "__builtin_arm_smulbb", + "llvm.arm.smulbt" => "__builtin_arm_smulbt", + "llvm.arm.smultb" => "__builtin_arm_smultb", + "llvm.arm.smultt" => "__builtin_arm_smultt", + "llvm.arm.smulwb" => "__builtin_arm_smulwb", + "llvm.arm.smulwt" => "__builtin_arm_smulwt", + "llvm.arm.smusd" => "__builtin_arm_smusd", + "llvm.arm.smusdx" => "__builtin_arm_smusdx", + "llvm.arm.ssat" => "__builtin_arm_ssat", + "llvm.arm.ssat16" => "__builtin_arm_ssat16", + "llvm.arm.ssax" => "__builtin_arm_ssax", + "llvm.arm.ssub16" => "__builtin_arm_ssub16", + "llvm.arm.ssub8" => "__builtin_arm_ssub8", + "llvm.arm.stc" => "__builtin_arm_stc", + "llvm.arm.stc2" => "__builtin_arm_stc2", + "llvm.arm.stc2l" => "__builtin_arm_stc2l", + "llvm.arm.stcl" => "__builtin_arm_stcl", + "llvm.arm.sxtab16" => "__builtin_arm_sxtab16", + "llvm.arm.sxtb16" => "__builtin_arm_sxtb16", + "llvm.arm.thread.pointer" => "__builtin_thread_pointer", + "llvm.arm.uadd16" => "__builtin_arm_uadd16", + "llvm.arm.uadd8" => "__builtin_arm_uadd8", + "llvm.arm.uasx" => "__builtin_arm_uasx", + "llvm.arm.uhadd16" => "__builtin_arm_uhadd16", + "llvm.arm.uhadd8" => "__builtin_arm_uhadd8", + "llvm.arm.uhasx" => "__builtin_arm_uhasx", + "llvm.arm.uhsax" => "__builtin_arm_uhsax", + "llvm.arm.uhsub16" => "__builtin_arm_uhsub16", + "llvm.arm.uhsub8" => "__builtin_arm_uhsub8", + "llvm.arm.uqadd16" => "__builtin_arm_uqadd16", + "llvm.arm.uqadd8" => "__builtin_arm_uqadd8", + "llvm.arm.uqasx" => "__builtin_arm_uqasx", + "llvm.arm.uqsax" => "__builtin_arm_uqsax", + "llvm.arm.uqsub16" => "__builtin_arm_uqsub16", + "llvm.arm.uqsub8" => "__builtin_arm_uqsub8", + "llvm.arm.usad8" => "__builtin_arm_usad8", + "llvm.arm.usada8" => "__builtin_arm_usada8", + "llvm.arm.usat" => "__builtin_arm_usat", + "llvm.arm.usat16" => "__builtin_arm_usat16", + "llvm.arm.usax" => "__builtin_arm_usax", + "llvm.arm.usub16" => "__builtin_arm_usub16", + "llvm.arm.usub8" => "__builtin_arm_usub8", + "llvm.arm.uxtab16" => "__builtin_arm_uxtab16", + "llvm.arm.uxtb16" => "__builtin_arm_uxtb16", + // bpf + "llvm.bpf.btf.type.id" => "__builtin_bpf_btf_type_id", + "llvm.bpf.compare" => "__builtin_bpf_compare", + "llvm.bpf.load.byte" => "__builtin_bpf_load_byte", + "llvm.bpf.load.half" => "__builtin_bpf_load_half", + "llvm.bpf.load.word" => "__builtin_bpf_load_word", + "llvm.bpf.passthrough" => "__builtin_bpf_passthrough", + "llvm.bpf.preserve.enum.value" => "__builtin_bpf_preserve_enum_value", + "llvm.bpf.preserve.field.info" => "__builtin_bpf_preserve_field_info", + "llvm.bpf.preserve.type.info" => "__builtin_bpf_preserve_type_info", + "llvm.bpf.pseudo" => "__builtin_bpf_pseudo", + // cuda + "llvm.cuda.syncthreads" => "__syncthreads", + // hexagon + "llvm.hexagon.A2.abs" => "__builtin_HEXAGON_A2_abs", + "llvm.hexagon.A2.absp" => "__builtin_HEXAGON_A2_absp", + "llvm.hexagon.A2.abssat" => "__builtin_HEXAGON_A2_abssat", + "llvm.hexagon.A2.add" => "__builtin_HEXAGON_A2_add", + "llvm.hexagon.A2.addh.h16.hh" => "__builtin_HEXAGON_A2_addh_h16_hh", + "llvm.hexagon.A2.addh.h16.hl" => "__builtin_HEXAGON_A2_addh_h16_hl", + "llvm.hexagon.A2.addh.h16.lh" => "__builtin_HEXAGON_A2_addh_h16_lh", + "llvm.hexagon.A2.addh.h16.ll" => "__builtin_HEXAGON_A2_addh_h16_ll", + "llvm.hexagon.A2.addh.h16.sat.hh" => "__builtin_HEXAGON_A2_addh_h16_sat_hh", + "llvm.hexagon.A2.addh.h16.sat.hl" => "__builtin_HEXAGON_A2_addh_h16_sat_hl", + "llvm.hexagon.A2.addh.h16.sat.lh" => "__builtin_HEXAGON_A2_addh_h16_sat_lh", + "llvm.hexagon.A2.addh.h16.sat.ll" => "__builtin_HEXAGON_A2_addh_h16_sat_ll", + "llvm.hexagon.A2.addh.l16.hl" => "__builtin_HEXAGON_A2_addh_l16_hl", + "llvm.hexagon.A2.addh.l16.ll" => "__builtin_HEXAGON_A2_addh_l16_ll", + "llvm.hexagon.A2.addh.l16.sat.hl" => "__builtin_HEXAGON_A2_addh_l16_sat_hl", + "llvm.hexagon.A2.addh.l16.sat.ll" => "__builtin_HEXAGON_A2_addh_l16_sat_ll", + "llvm.hexagon.A2.addi" => "__builtin_HEXAGON_A2_addi", + "llvm.hexagon.A2.addp" => "__builtin_HEXAGON_A2_addp", + "llvm.hexagon.A2.addpsat" => "__builtin_HEXAGON_A2_addpsat", + "llvm.hexagon.A2.addsat" => "__builtin_HEXAGON_A2_addsat", + "llvm.hexagon.A2.addsp" => "__builtin_HEXAGON_A2_addsp", + "llvm.hexagon.A2.and" => "__builtin_HEXAGON_A2_and", + "llvm.hexagon.A2.andir" => "__builtin_HEXAGON_A2_andir", + "llvm.hexagon.A2.andp" => "__builtin_HEXAGON_A2_andp", + "llvm.hexagon.A2.aslh" => "__builtin_HEXAGON_A2_aslh", + "llvm.hexagon.A2.asrh" => "__builtin_HEXAGON_A2_asrh", + "llvm.hexagon.A2.combine.hh" => "__builtin_HEXAGON_A2_combine_hh", + "llvm.hexagon.A2.combine.hl" => "__builtin_HEXAGON_A2_combine_hl", + "llvm.hexagon.A2.combine.lh" => "__builtin_HEXAGON_A2_combine_lh", + "llvm.hexagon.A2.combine.ll" => "__builtin_HEXAGON_A2_combine_ll", + "llvm.hexagon.A2.combineii" => "__builtin_HEXAGON_A2_combineii", + "llvm.hexagon.A2.combinew" => "__builtin_HEXAGON_A2_combinew", + "llvm.hexagon.A2.max" => "__builtin_HEXAGON_A2_max", + "llvm.hexagon.A2.maxp" => "__builtin_HEXAGON_A2_maxp", + "llvm.hexagon.A2.maxu" => "__builtin_HEXAGON_A2_maxu", + "llvm.hexagon.A2.maxup" => "__builtin_HEXAGON_A2_maxup", + "llvm.hexagon.A2.min" => "__builtin_HEXAGON_A2_min", + "llvm.hexagon.A2.minp" => "__builtin_HEXAGON_A2_minp", + "llvm.hexagon.A2.minu" => "__builtin_HEXAGON_A2_minu", + "llvm.hexagon.A2.minup" => "__builtin_HEXAGON_A2_minup", + "llvm.hexagon.A2.neg" => "__builtin_HEXAGON_A2_neg", + "llvm.hexagon.A2.negp" => "__builtin_HEXAGON_A2_negp", + "llvm.hexagon.A2.negsat" => "__builtin_HEXAGON_A2_negsat", + "llvm.hexagon.A2.not" => "__builtin_HEXAGON_A2_not", + "llvm.hexagon.A2.notp" => "__builtin_HEXAGON_A2_notp", + "llvm.hexagon.A2.or" => "__builtin_HEXAGON_A2_or", + "llvm.hexagon.A2.orir" => "__builtin_HEXAGON_A2_orir", + "llvm.hexagon.A2.orp" => "__builtin_HEXAGON_A2_orp", + "llvm.hexagon.A2.roundsat" => "__builtin_HEXAGON_A2_roundsat", + "llvm.hexagon.A2.sat" => "__builtin_HEXAGON_A2_sat", + "llvm.hexagon.A2.satb" => "__builtin_HEXAGON_A2_satb", + "llvm.hexagon.A2.sath" => "__builtin_HEXAGON_A2_sath", + "llvm.hexagon.A2.satub" => "__builtin_HEXAGON_A2_satub", + "llvm.hexagon.A2.satuh" => "__builtin_HEXAGON_A2_satuh", + "llvm.hexagon.A2.sub" => "__builtin_HEXAGON_A2_sub", + "llvm.hexagon.A2.subh.h16.hh" => "__builtin_HEXAGON_A2_subh_h16_hh", + "llvm.hexagon.A2.subh.h16.hl" => "__builtin_HEXAGON_A2_subh_h16_hl", + "llvm.hexagon.A2.subh.h16.lh" => "__builtin_HEXAGON_A2_subh_h16_lh", + "llvm.hexagon.A2.subh.h16.ll" => "__builtin_HEXAGON_A2_subh_h16_ll", + "llvm.hexagon.A2.subh.h16.sat.hh" => "__builtin_HEXAGON_A2_subh_h16_sat_hh", + "llvm.hexagon.A2.subh.h16.sat.hl" => "__builtin_HEXAGON_A2_subh_h16_sat_hl", + "llvm.hexagon.A2.subh.h16.sat.lh" => "__builtin_HEXAGON_A2_subh_h16_sat_lh", + "llvm.hexagon.A2.subh.h16.sat.ll" => "__builtin_HEXAGON_A2_subh_h16_sat_ll", + "llvm.hexagon.A2.subh.l16.hl" => "__builtin_HEXAGON_A2_subh_l16_hl", + "llvm.hexagon.A2.subh.l16.ll" => "__builtin_HEXAGON_A2_subh_l16_ll", + "llvm.hexagon.A2.subh.l16.sat.hl" => "__builtin_HEXAGON_A2_subh_l16_sat_hl", + "llvm.hexagon.A2.subh.l16.sat.ll" => "__builtin_HEXAGON_A2_subh_l16_sat_ll", + "llvm.hexagon.A2.subp" => "__builtin_HEXAGON_A2_subp", + "llvm.hexagon.A2.subri" => "__builtin_HEXAGON_A2_subri", + "llvm.hexagon.A2.subsat" => "__builtin_HEXAGON_A2_subsat", + "llvm.hexagon.A2.svaddh" => "__builtin_HEXAGON_A2_svaddh", + "llvm.hexagon.A2.svaddhs" => "__builtin_HEXAGON_A2_svaddhs", + "llvm.hexagon.A2.svadduhs" => "__builtin_HEXAGON_A2_svadduhs", + "llvm.hexagon.A2.svavgh" => "__builtin_HEXAGON_A2_svavgh", + "llvm.hexagon.A2.svavghs" => "__builtin_HEXAGON_A2_svavghs", + "llvm.hexagon.A2.svnavgh" => "__builtin_HEXAGON_A2_svnavgh", + "llvm.hexagon.A2.svsubh" => "__builtin_HEXAGON_A2_svsubh", + "llvm.hexagon.A2.svsubhs" => "__builtin_HEXAGON_A2_svsubhs", + "llvm.hexagon.A2.svsubuhs" => "__builtin_HEXAGON_A2_svsubuhs", + "llvm.hexagon.A2.swiz" => "__builtin_HEXAGON_A2_swiz", + "llvm.hexagon.A2.sxtb" => "__builtin_HEXAGON_A2_sxtb", + "llvm.hexagon.A2.sxth" => "__builtin_HEXAGON_A2_sxth", + "llvm.hexagon.A2.sxtw" => "__builtin_HEXAGON_A2_sxtw", + "llvm.hexagon.A2.tfr" => "__builtin_HEXAGON_A2_tfr", + "llvm.hexagon.A2.tfrih" => "__builtin_HEXAGON_A2_tfrih", + "llvm.hexagon.A2.tfril" => "__builtin_HEXAGON_A2_tfril", + "llvm.hexagon.A2.tfrp" => "__builtin_HEXAGON_A2_tfrp", + "llvm.hexagon.A2.tfrpi" => "__builtin_HEXAGON_A2_tfrpi", + "llvm.hexagon.A2.tfrsi" => "__builtin_HEXAGON_A2_tfrsi", + "llvm.hexagon.A2.vabsh" => "__builtin_HEXAGON_A2_vabsh", + "llvm.hexagon.A2.vabshsat" => "__builtin_HEXAGON_A2_vabshsat", + "llvm.hexagon.A2.vabsw" => "__builtin_HEXAGON_A2_vabsw", + "llvm.hexagon.A2.vabswsat" => "__builtin_HEXAGON_A2_vabswsat", + "llvm.hexagon.A2.vaddb.map" => "__builtin_HEXAGON_A2_vaddb_map", + "llvm.hexagon.A2.vaddh" => "__builtin_HEXAGON_A2_vaddh", + "llvm.hexagon.A2.vaddhs" => "__builtin_HEXAGON_A2_vaddhs", + "llvm.hexagon.A2.vaddub" => "__builtin_HEXAGON_A2_vaddub", + "llvm.hexagon.A2.vaddubs" => "__builtin_HEXAGON_A2_vaddubs", + "llvm.hexagon.A2.vadduhs" => "__builtin_HEXAGON_A2_vadduhs", + "llvm.hexagon.A2.vaddw" => "__builtin_HEXAGON_A2_vaddw", + "llvm.hexagon.A2.vaddws" => "__builtin_HEXAGON_A2_vaddws", + "llvm.hexagon.A2.vavgh" => "__builtin_HEXAGON_A2_vavgh", + "llvm.hexagon.A2.vavghcr" => "__builtin_HEXAGON_A2_vavghcr", + "llvm.hexagon.A2.vavghr" => "__builtin_HEXAGON_A2_vavghr", + "llvm.hexagon.A2.vavgub" => "__builtin_HEXAGON_A2_vavgub", + "llvm.hexagon.A2.vavgubr" => "__builtin_HEXAGON_A2_vavgubr", + "llvm.hexagon.A2.vavguh" => "__builtin_HEXAGON_A2_vavguh", + "llvm.hexagon.A2.vavguhr" => "__builtin_HEXAGON_A2_vavguhr", + "llvm.hexagon.A2.vavguw" => "__builtin_HEXAGON_A2_vavguw", + "llvm.hexagon.A2.vavguwr" => "__builtin_HEXAGON_A2_vavguwr", + "llvm.hexagon.A2.vavgw" => "__builtin_HEXAGON_A2_vavgw", + "llvm.hexagon.A2.vavgwcr" => "__builtin_HEXAGON_A2_vavgwcr", + "llvm.hexagon.A2.vavgwr" => "__builtin_HEXAGON_A2_vavgwr", + "llvm.hexagon.A2.vcmpbeq" => "__builtin_HEXAGON_A2_vcmpbeq", + "llvm.hexagon.A2.vcmpbgtu" => "__builtin_HEXAGON_A2_vcmpbgtu", + "llvm.hexagon.A2.vcmpheq" => "__builtin_HEXAGON_A2_vcmpheq", + "llvm.hexagon.A2.vcmphgt" => "__builtin_HEXAGON_A2_vcmphgt", + "llvm.hexagon.A2.vcmphgtu" => "__builtin_HEXAGON_A2_vcmphgtu", + "llvm.hexagon.A2.vcmpweq" => "__builtin_HEXAGON_A2_vcmpweq", + "llvm.hexagon.A2.vcmpwgt" => "__builtin_HEXAGON_A2_vcmpwgt", + "llvm.hexagon.A2.vcmpwgtu" => "__builtin_HEXAGON_A2_vcmpwgtu", + "llvm.hexagon.A2.vconj" => "__builtin_HEXAGON_A2_vconj", + "llvm.hexagon.A2.vmaxb" => "__builtin_HEXAGON_A2_vmaxb", + "llvm.hexagon.A2.vmaxh" => "__builtin_HEXAGON_A2_vmaxh", + "llvm.hexagon.A2.vmaxub" => "__builtin_HEXAGON_A2_vmaxub", + "llvm.hexagon.A2.vmaxuh" => "__builtin_HEXAGON_A2_vmaxuh", + "llvm.hexagon.A2.vmaxuw" => "__builtin_HEXAGON_A2_vmaxuw", + "llvm.hexagon.A2.vmaxw" => "__builtin_HEXAGON_A2_vmaxw", + "llvm.hexagon.A2.vminb" => "__builtin_HEXAGON_A2_vminb", + "llvm.hexagon.A2.vminh" => "__builtin_HEXAGON_A2_vminh", + "llvm.hexagon.A2.vminub" => "__builtin_HEXAGON_A2_vminub", + "llvm.hexagon.A2.vminuh" => "__builtin_HEXAGON_A2_vminuh", + "llvm.hexagon.A2.vminuw" => "__builtin_HEXAGON_A2_vminuw", + "llvm.hexagon.A2.vminw" => "__builtin_HEXAGON_A2_vminw", + "llvm.hexagon.A2.vnavgh" => "__builtin_HEXAGON_A2_vnavgh", + "llvm.hexagon.A2.vnavghcr" => "__builtin_HEXAGON_A2_vnavghcr", + "llvm.hexagon.A2.vnavghr" => "__builtin_HEXAGON_A2_vnavghr", + "llvm.hexagon.A2.vnavgw" => "__builtin_HEXAGON_A2_vnavgw", + "llvm.hexagon.A2.vnavgwcr" => "__builtin_HEXAGON_A2_vnavgwcr", + "llvm.hexagon.A2.vnavgwr" => "__builtin_HEXAGON_A2_vnavgwr", + "llvm.hexagon.A2.vraddub" => "__builtin_HEXAGON_A2_vraddub", + "llvm.hexagon.A2.vraddub.acc" => "__builtin_HEXAGON_A2_vraddub_acc", + "llvm.hexagon.A2.vrsadub" => "__builtin_HEXAGON_A2_vrsadub", + "llvm.hexagon.A2.vrsadub.acc" => "__builtin_HEXAGON_A2_vrsadub_acc", + "llvm.hexagon.A2.vsubb.map" => "__builtin_HEXAGON_A2_vsubb_map", + "llvm.hexagon.A2.vsubh" => "__builtin_HEXAGON_A2_vsubh", + "llvm.hexagon.A2.vsubhs" => "__builtin_HEXAGON_A2_vsubhs", + "llvm.hexagon.A2.vsubub" => "__builtin_HEXAGON_A2_vsubub", + "llvm.hexagon.A2.vsububs" => "__builtin_HEXAGON_A2_vsububs", + "llvm.hexagon.A2.vsubuhs" => "__builtin_HEXAGON_A2_vsubuhs", + "llvm.hexagon.A2.vsubw" => "__builtin_HEXAGON_A2_vsubw", + "llvm.hexagon.A2.vsubws" => "__builtin_HEXAGON_A2_vsubws", + "llvm.hexagon.A2.xor" => "__builtin_HEXAGON_A2_xor", + "llvm.hexagon.A2.xorp" => "__builtin_HEXAGON_A2_xorp", + "llvm.hexagon.A2.zxtb" => "__builtin_HEXAGON_A2_zxtb", + "llvm.hexagon.A2.zxth" => "__builtin_HEXAGON_A2_zxth", + "llvm.hexagon.A4.andn" => "__builtin_HEXAGON_A4_andn", + "llvm.hexagon.A4.andnp" => "__builtin_HEXAGON_A4_andnp", + "llvm.hexagon.A4.bitsplit" => "__builtin_HEXAGON_A4_bitsplit", + "llvm.hexagon.A4.bitspliti" => "__builtin_HEXAGON_A4_bitspliti", + "llvm.hexagon.A4.boundscheck" => "__builtin_HEXAGON_A4_boundscheck", + "llvm.hexagon.A4.cmpbeq" => "__builtin_HEXAGON_A4_cmpbeq", + "llvm.hexagon.A4.cmpbeqi" => "__builtin_HEXAGON_A4_cmpbeqi", + "llvm.hexagon.A4.cmpbgt" => "__builtin_HEXAGON_A4_cmpbgt", + "llvm.hexagon.A4.cmpbgti" => "__builtin_HEXAGON_A4_cmpbgti", + "llvm.hexagon.A4.cmpbgtu" => "__builtin_HEXAGON_A4_cmpbgtu", + "llvm.hexagon.A4.cmpbgtui" => "__builtin_HEXAGON_A4_cmpbgtui", + "llvm.hexagon.A4.cmpheq" => "__builtin_HEXAGON_A4_cmpheq", + "llvm.hexagon.A4.cmpheqi" => "__builtin_HEXAGON_A4_cmpheqi", + "llvm.hexagon.A4.cmphgt" => "__builtin_HEXAGON_A4_cmphgt", + "llvm.hexagon.A4.cmphgti" => "__builtin_HEXAGON_A4_cmphgti", + "llvm.hexagon.A4.cmphgtu" => "__builtin_HEXAGON_A4_cmphgtu", + "llvm.hexagon.A4.cmphgtui" => "__builtin_HEXAGON_A4_cmphgtui", + "llvm.hexagon.A4.combineir" => "__builtin_HEXAGON_A4_combineir", + "llvm.hexagon.A4.combineri" => "__builtin_HEXAGON_A4_combineri", + "llvm.hexagon.A4.cround.ri" => "__builtin_HEXAGON_A4_cround_ri", + "llvm.hexagon.A4.cround.rr" => "__builtin_HEXAGON_A4_cround_rr", + "llvm.hexagon.A4.modwrapu" => "__builtin_HEXAGON_A4_modwrapu", + "llvm.hexagon.A4.orn" => "__builtin_HEXAGON_A4_orn", + "llvm.hexagon.A4.ornp" => "__builtin_HEXAGON_A4_ornp", + "llvm.hexagon.A4.rcmpeq" => "__builtin_HEXAGON_A4_rcmpeq", + "llvm.hexagon.A4.rcmpeqi" => "__builtin_HEXAGON_A4_rcmpeqi", + "llvm.hexagon.A4.rcmpneq" => "__builtin_HEXAGON_A4_rcmpneq", + "llvm.hexagon.A4.rcmpneqi" => "__builtin_HEXAGON_A4_rcmpneqi", + "llvm.hexagon.A4.round.ri" => "__builtin_HEXAGON_A4_round_ri", + "llvm.hexagon.A4.round.ri.sat" => "__builtin_HEXAGON_A4_round_ri_sat", + "llvm.hexagon.A4.round.rr" => "__builtin_HEXAGON_A4_round_rr", + "llvm.hexagon.A4.round.rr.sat" => "__builtin_HEXAGON_A4_round_rr_sat", + "llvm.hexagon.A4.tlbmatch" => "__builtin_HEXAGON_A4_tlbmatch", + "llvm.hexagon.A4.vcmpbeq.any" => "__builtin_HEXAGON_A4_vcmpbeq_any", + "llvm.hexagon.A4.vcmpbeqi" => "__builtin_HEXAGON_A4_vcmpbeqi", + "llvm.hexagon.A4.vcmpbgt" => "__builtin_HEXAGON_A4_vcmpbgt", + "llvm.hexagon.A4.vcmpbgti" => "__builtin_HEXAGON_A4_vcmpbgti", + "llvm.hexagon.A4.vcmpbgtui" => "__builtin_HEXAGON_A4_vcmpbgtui", + "llvm.hexagon.A4.vcmpheqi" => "__builtin_HEXAGON_A4_vcmpheqi", + "llvm.hexagon.A4.vcmphgti" => "__builtin_HEXAGON_A4_vcmphgti", + "llvm.hexagon.A4.vcmphgtui" => "__builtin_HEXAGON_A4_vcmphgtui", + "llvm.hexagon.A4.vcmpweqi" => "__builtin_HEXAGON_A4_vcmpweqi", + "llvm.hexagon.A4.vcmpwgti" => "__builtin_HEXAGON_A4_vcmpwgti", + "llvm.hexagon.A4.vcmpwgtui" => "__builtin_HEXAGON_A4_vcmpwgtui", + "llvm.hexagon.A4.vrmaxh" => "__builtin_HEXAGON_A4_vrmaxh", + "llvm.hexagon.A4.vrmaxuh" => "__builtin_HEXAGON_A4_vrmaxuh", + "llvm.hexagon.A4.vrmaxuw" => "__builtin_HEXAGON_A4_vrmaxuw", + "llvm.hexagon.A4.vrmaxw" => "__builtin_HEXAGON_A4_vrmaxw", + "llvm.hexagon.A4.vrminh" => "__builtin_HEXAGON_A4_vrminh", + "llvm.hexagon.A4.vrminuh" => "__builtin_HEXAGON_A4_vrminuh", + "llvm.hexagon.A4.vrminuw" => "__builtin_HEXAGON_A4_vrminuw", + "llvm.hexagon.A4.vrminw" => "__builtin_HEXAGON_A4_vrminw", + "llvm.hexagon.A5.vaddhubs" => "__builtin_HEXAGON_A5_vaddhubs", + "llvm.hexagon.C2.all8" => "__builtin_HEXAGON_C2_all8", + "llvm.hexagon.C2.and" => "__builtin_HEXAGON_C2_and", + "llvm.hexagon.C2.andn" => "__builtin_HEXAGON_C2_andn", + "llvm.hexagon.C2.any8" => "__builtin_HEXAGON_C2_any8", + "llvm.hexagon.C2.bitsclr" => "__builtin_HEXAGON_C2_bitsclr", + "llvm.hexagon.C2.bitsclri" => "__builtin_HEXAGON_C2_bitsclri", + "llvm.hexagon.C2.bitsset" => "__builtin_HEXAGON_C2_bitsset", + "llvm.hexagon.C2.cmpeq" => "__builtin_HEXAGON_C2_cmpeq", + "llvm.hexagon.C2.cmpeqi" => "__builtin_HEXAGON_C2_cmpeqi", + "llvm.hexagon.C2.cmpeqp" => "__builtin_HEXAGON_C2_cmpeqp", + "llvm.hexagon.C2.cmpgei" => "__builtin_HEXAGON_C2_cmpgei", + "llvm.hexagon.C2.cmpgeui" => "__builtin_HEXAGON_C2_cmpgeui", + "llvm.hexagon.C2.cmpgt" => "__builtin_HEXAGON_C2_cmpgt", + "llvm.hexagon.C2.cmpgti" => "__builtin_HEXAGON_C2_cmpgti", + "llvm.hexagon.C2.cmpgtp" => "__builtin_HEXAGON_C2_cmpgtp", + "llvm.hexagon.C2.cmpgtu" => "__builtin_HEXAGON_C2_cmpgtu", + "llvm.hexagon.C2.cmpgtui" => "__builtin_HEXAGON_C2_cmpgtui", + "llvm.hexagon.C2.cmpgtup" => "__builtin_HEXAGON_C2_cmpgtup", + "llvm.hexagon.C2.cmplt" => "__builtin_HEXAGON_C2_cmplt", + "llvm.hexagon.C2.cmpltu" => "__builtin_HEXAGON_C2_cmpltu", + "llvm.hexagon.C2.mask" => "__builtin_HEXAGON_C2_mask", + "llvm.hexagon.C2.mux" => "__builtin_HEXAGON_C2_mux", + "llvm.hexagon.C2.muxii" => "__builtin_HEXAGON_C2_muxii", + "llvm.hexagon.C2.muxir" => "__builtin_HEXAGON_C2_muxir", + "llvm.hexagon.C2.muxri" => "__builtin_HEXAGON_C2_muxri", + "llvm.hexagon.C2.not" => "__builtin_HEXAGON_C2_not", + "llvm.hexagon.C2.or" => "__builtin_HEXAGON_C2_or", + "llvm.hexagon.C2.orn" => "__builtin_HEXAGON_C2_orn", + "llvm.hexagon.C2.pxfer.map" => "__builtin_HEXAGON_C2_pxfer_map", + "llvm.hexagon.C2.tfrpr" => "__builtin_HEXAGON_C2_tfrpr", + "llvm.hexagon.C2.tfrrp" => "__builtin_HEXAGON_C2_tfrrp", + "llvm.hexagon.C2.vitpack" => "__builtin_HEXAGON_C2_vitpack", + "llvm.hexagon.C2.vmux" => "__builtin_HEXAGON_C2_vmux", + "llvm.hexagon.C2.xor" => "__builtin_HEXAGON_C2_xor", + "llvm.hexagon.C4.and.and" => "__builtin_HEXAGON_C4_and_and", + "llvm.hexagon.C4.and.andn" => "__builtin_HEXAGON_C4_and_andn", + "llvm.hexagon.C4.and.or" => "__builtin_HEXAGON_C4_and_or", + "llvm.hexagon.C4.and.orn" => "__builtin_HEXAGON_C4_and_orn", + "llvm.hexagon.C4.cmplte" => "__builtin_HEXAGON_C4_cmplte", + "llvm.hexagon.C4.cmpltei" => "__builtin_HEXAGON_C4_cmpltei", + "llvm.hexagon.C4.cmplteu" => "__builtin_HEXAGON_C4_cmplteu", + "llvm.hexagon.C4.cmplteui" => "__builtin_HEXAGON_C4_cmplteui", + "llvm.hexagon.C4.cmpneq" => "__builtin_HEXAGON_C4_cmpneq", + "llvm.hexagon.C4.cmpneqi" => "__builtin_HEXAGON_C4_cmpneqi", + "llvm.hexagon.C4.fastcorner9" => "__builtin_HEXAGON_C4_fastcorner9", + "llvm.hexagon.C4.fastcorner9.not" => "__builtin_HEXAGON_C4_fastcorner9_not", + "llvm.hexagon.C4.nbitsclr" => "__builtin_HEXAGON_C4_nbitsclr", + "llvm.hexagon.C4.nbitsclri" => "__builtin_HEXAGON_C4_nbitsclri", + "llvm.hexagon.C4.nbitsset" => "__builtin_HEXAGON_C4_nbitsset", + "llvm.hexagon.C4.or.and" => "__builtin_HEXAGON_C4_or_and", + "llvm.hexagon.C4.or.andn" => "__builtin_HEXAGON_C4_or_andn", + "llvm.hexagon.C4.or.or" => "__builtin_HEXAGON_C4_or_or", + "llvm.hexagon.C4.or.orn" => "__builtin_HEXAGON_C4_or_orn", + "llvm.hexagon.F2.conv.d2df" => "__builtin_HEXAGON_F2_conv_d2df", + "llvm.hexagon.F2.conv.d2sf" => "__builtin_HEXAGON_F2_conv_d2sf", + "llvm.hexagon.F2.conv.df2d" => "__builtin_HEXAGON_F2_conv_df2d", + "llvm.hexagon.F2.conv.df2d.chop" => "__builtin_HEXAGON_F2_conv_df2d_chop", + "llvm.hexagon.F2.conv.df2sf" => "__builtin_HEXAGON_F2_conv_df2sf", + "llvm.hexagon.F2.conv.df2ud" => "__builtin_HEXAGON_F2_conv_df2ud", + "llvm.hexagon.F2.conv.df2ud.chop" => "__builtin_HEXAGON_F2_conv_df2ud_chop", + "llvm.hexagon.F2.conv.df2uw" => "__builtin_HEXAGON_F2_conv_df2uw", + "llvm.hexagon.F2.conv.df2uw.chop" => "__builtin_HEXAGON_F2_conv_df2uw_chop", + "llvm.hexagon.F2.conv.df2w" => "__builtin_HEXAGON_F2_conv_df2w", + "llvm.hexagon.F2.conv.df2w.chop" => "__builtin_HEXAGON_F2_conv_df2w_chop", + "llvm.hexagon.F2.conv.sf2d" => "__builtin_HEXAGON_F2_conv_sf2d", + "llvm.hexagon.F2.conv.sf2d.chop" => "__builtin_HEXAGON_F2_conv_sf2d_chop", + "llvm.hexagon.F2.conv.sf2df" => "__builtin_HEXAGON_F2_conv_sf2df", + "llvm.hexagon.F2.conv.sf2ud" => "__builtin_HEXAGON_F2_conv_sf2ud", + "llvm.hexagon.F2.conv.sf2ud.chop" => "__builtin_HEXAGON_F2_conv_sf2ud_chop", + "llvm.hexagon.F2.conv.sf2uw" => "__builtin_HEXAGON_F2_conv_sf2uw", + "llvm.hexagon.F2.conv.sf2uw.chop" => "__builtin_HEXAGON_F2_conv_sf2uw_chop", + "llvm.hexagon.F2.conv.sf2w" => "__builtin_HEXAGON_F2_conv_sf2w", + "llvm.hexagon.F2.conv.sf2w.chop" => "__builtin_HEXAGON_F2_conv_sf2w_chop", + "llvm.hexagon.F2.conv.ud2df" => "__builtin_HEXAGON_F2_conv_ud2df", + "llvm.hexagon.F2.conv.ud2sf" => "__builtin_HEXAGON_F2_conv_ud2sf", + "llvm.hexagon.F2.conv.uw2df" => "__builtin_HEXAGON_F2_conv_uw2df", + "llvm.hexagon.F2.conv.uw2sf" => "__builtin_HEXAGON_F2_conv_uw2sf", + "llvm.hexagon.F2.conv.w2df" => "__builtin_HEXAGON_F2_conv_w2df", + "llvm.hexagon.F2.conv.w2sf" => "__builtin_HEXAGON_F2_conv_w2sf", + "llvm.hexagon.F2.dfadd" => "__builtin_HEXAGON_F2_dfadd", + "llvm.hexagon.F2.dfclass" => "__builtin_HEXAGON_F2_dfclass", + "llvm.hexagon.F2.dfcmpeq" => "__builtin_HEXAGON_F2_dfcmpeq", + "llvm.hexagon.F2.dfcmpge" => "__builtin_HEXAGON_F2_dfcmpge", + "llvm.hexagon.F2.dfcmpgt" => "__builtin_HEXAGON_F2_dfcmpgt", + "llvm.hexagon.F2.dfcmpuo" => "__builtin_HEXAGON_F2_dfcmpuo", + "llvm.hexagon.F2.dffixupd" => "__builtin_HEXAGON_F2_dffixupd", + "llvm.hexagon.F2.dffixupn" => "__builtin_HEXAGON_F2_dffixupn", + "llvm.hexagon.F2.dffixupr" => "__builtin_HEXAGON_F2_dffixupr", + "llvm.hexagon.F2.dffma" => "__builtin_HEXAGON_F2_dffma", + "llvm.hexagon.F2.dffma.lib" => "__builtin_HEXAGON_F2_dffma_lib", + "llvm.hexagon.F2.dffma.sc" => "__builtin_HEXAGON_F2_dffma_sc", + "llvm.hexagon.F2.dffms" => "__builtin_HEXAGON_F2_dffms", + "llvm.hexagon.F2.dffms.lib" => "__builtin_HEXAGON_F2_dffms_lib", + "llvm.hexagon.F2.dfimm.n" => "__builtin_HEXAGON_F2_dfimm_n", + "llvm.hexagon.F2.dfimm.p" => "__builtin_HEXAGON_F2_dfimm_p", + "llvm.hexagon.F2.dfmax" => "__builtin_HEXAGON_F2_dfmax", + "llvm.hexagon.F2.dfmin" => "__builtin_HEXAGON_F2_dfmin", + "llvm.hexagon.F2.dfmpy" => "__builtin_HEXAGON_F2_dfmpy", + "llvm.hexagon.F2.dfsub" => "__builtin_HEXAGON_F2_dfsub", + "llvm.hexagon.F2.sfadd" => "__builtin_HEXAGON_F2_sfadd", + "llvm.hexagon.F2.sfclass" => "__builtin_HEXAGON_F2_sfclass", + "llvm.hexagon.F2.sfcmpeq" => "__builtin_HEXAGON_F2_sfcmpeq", + "llvm.hexagon.F2.sfcmpge" => "__builtin_HEXAGON_F2_sfcmpge", + "llvm.hexagon.F2.sfcmpgt" => "__builtin_HEXAGON_F2_sfcmpgt", + "llvm.hexagon.F2.sfcmpuo" => "__builtin_HEXAGON_F2_sfcmpuo", + "llvm.hexagon.F2.sffixupd" => "__builtin_HEXAGON_F2_sffixupd", + "llvm.hexagon.F2.sffixupn" => "__builtin_HEXAGON_F2_sffixupn", + "llvm.hexagon.F2.sffixupr" => "__builtin_HEXAGON_F2_sffixupr", + "llvm.hexagon.F2.sffma" => "__builtin_HEXAGON_F2_sffma", + "llvm.hexagon.F2.sffma.lib" => "__builtin_HEXAGON_F2_sffma_lib", + "llvm.hexagon.F2.sffma.sc" => "__builtin_HEXAGON_F2_sffma_sc", + "llvm.hexagon.F2.sffms" => "__builtin_HEXAGON_F2_sffms", + "llvm.hexagon.F2.sffms.lib" => "__builtin_HEXAGON_F2_sffms_lib", + "llvm.hexagon.F2.sfimm.n" => "__builtin_HEXAGON_F2_sfimm_n", + "llvm.hexagon.F2.sfimm.p" => "__builtin_HEXAGON_F2_sfimm_p", + "llvm.hexagon.F2.sfmax" => "__builtin_HEXAGON_F2_sfmax", + "llvm.hexagon.F2.sfmin" => "__builtin_HEXAGON_F2_sfmin", + "llvm.hexagon.F2.sfmpy" => "__builtin_HEXAGON_F2_sfmpy", + "llvm.hexagon.F2.sfsub" => "__builtin_HEXAGON_F2_sfsub", + "llvm.hexagon.M2.acci" => "__builtin_HEXAGON_M2_acci", + "llvm.hexagon.M2.accii" => "__builtin_HEXAGON_M2_accii", + "llvm.hexagon.M2.cmaci.s0" => "__builtin_HEXAGON_M2_cmaci_s0", + "llvm.hexagon.M2.cmacr.s0" => "__builtin_HEXAGON_M2_cmacr_s0", + "llvm.hexagon.M2.cmacs.s0" => "__builtin_HEXAGON_M2_cmacs_s0", + "llvm.hexagon.M2.cmacs.s1" => "__builtin_HEXAGON_M2_cmacs_s1", + "llvm.hexagon.M2.cmacsc.s0" => "__builtin_HEXAGON_M2_cmacsc_s0", + "llvm.hexagon.M2.cmacsc.s1" => "__builtin_HEXAGON_M2_cmacsc_s1", + "llvm.hexagon.M2.cmpyi.s0" => "__builtin_HEXAGON_M2_cmpyi_s0", + "llvm.hexagon.M2.cmpyr.s0" => "__builtin_HEXAGON_M2_cmpyr_s0", + "llvm.hexagon.M2.cmpyrs.s0" => "__builtin_HEXAGON_M2_cmpyrs_s0", + "llvm.hexagon.M2.cmpyrs.s1" => "__builtin_HEXAGON_M2_cmpyrs_s1", + "llvm.hexagon.M2.cmpyrsc.s0" => "__builtin_HEXAGON_M2_cmpyrsc_s0", + "llvm.hexagon.M2.cmpyrsc.s1" => "__builtin_HEXAGON_M2_cmpyrsc_s1", + "llvm.hexagon.M2.cmpys.s0" => "__builtin_HEXAGON_M2_cmpys_s0", + "llvm.hexagon.M2.cmpys.s1" => "__builtin_HEXAGON_M2_cmpys_s1", + "llvm.hexagon.M2.cmpysc.s0" => "__builtin_HEXAGON_M2_cmpysc_s0", + "llvm.hexagon.M2.cmpysc.s1" => "__builtin_HEXAGON_M2_cmpysc_s1", + "llvm.hexagon.M2.cnacs.s0" => "__builtin_HEXAGON_M2_cnacs_s0", + "llvm.hexagon.M2.cnacs.s1" => "__builtin_HEXAGON_M2_cnacs_s1", + "llvm.hexagon.M2.cnacsc.s0" => "__builtin_HEXAGON_M2_cnacsc_s0", + "llvm.hexagon.M2.cnacsc.s1" => "__builtin_HEXAGON_M2_cnacsc_s1", + "llvm.hexagon.M2.dpmpyss.acc.s0" => "__builtin_HEXAGON_M2_dpmpyss_acc_s0", + "llvm.hexagon.M2.dpmpyss.nac.s0" => "__builtin_HEXAGON_M2_dpmpyss_nac_s0", + "llvm.hexagon.M2.dpmpyss.rnd.s0" => "__builtin_HEXAGON_M2_dpmpyss_rnd_s0", + "llvm.hexagon.M2.dpmpyss.s0" => "__builtin_HEXAGON_M2_dpmpyss_s0", + "llvm.hexagon.M2.dpmpyuu.acc.s0" => "__builtin_HEXAGON_M2_dpmpyuu_acc_s0", + "llvm.hexagon.M2.dpmpyuu.nac.s0" => "__builtin_HEXAGON_M2_dpmpyuu_nac_s0", + "llvm.hexagon.M2.dpmpyuu.s0" => "__builtin_HEXAGON_M2_dpmpyuu_s0", + "llvm.hexagon.M2.hmmpyh.rs1" => "__builtin_HEXAGON_M2_hmmpyh_rs1", + "llvm.hexagon.M2.hmmpyh.s1" => "__builtin_HEXAGON_M2_hmmpyh_s1", + "llvm.hexagon.M2.hmmpyl.rs1" => "__builtin_HEXAGON_M2_hmmpyl_rs1", + "llvm.hexagon.M2.hmmpyl.s1" => "__builtin_HEXAGON_M2_hmmpyl_s1", + "llvm.hexagon.M2.maci" => "__builtin_HEXAGON_M2_maci", + "llvm.hexagon.M2.macsin" => "__builtin_HEXAGON_M2_macsin", + "llvm.hexagon.M2.macsip" => "__builtin_HEXAGON_M2_macsip", + "llvm.hexagon.M2.mmachs.rs0" => "__builtin_HEXAGON_M2_mmachs_rs0", + "llvm.hexagon.M2.mmachs.rs1" => "__builtin_HEXAGON_M2_mmachs_rs1", + "llvm.hexagon.M2.mmachs.s0" => "__builtin_HEXAGON_M2_mmachs_s0", + "llvm.hexagon.M2.mmachs.s1" => "__builtin_HEXAGON_M2_mmachs_s1", + "llvm.hexagon.M2.mmacls.rs0" => "__builtin_HEXAGON_M2_mmacls_rs0", + "llvm.hexagon.M2.mmacls.rs1" => "__builtin_HEXAGON_M2_mmacls_rs1", + "llvm.hexagon.M2.mmacls.s0" => "__builtin_HEXAGON_M2_mmacls_s0", + "llvm.hexagon.M2.mmacls.s1" => "__builtin_HEXAGON_M2_mmacls_s1", + "llvm.hexagon.M2.mmacuhs.rs0" => "__builtin_HEXAGON_M2_mmacuhs_rs0", + "llvm.hexagon.M2.mmacuhs.rs1" => "__builtin_HEXAGON_M2_mmacuhs_rs1", + "llvm.hexagon.M2.mmacuhs.s0" => "__builtin_HEXAGON_M2_mmacuhs_s0", + "llvm.hexagon.M2.mmacuhs.s1" => "__builtin_HEXAGON_M2_mmacuhs_s1", + "llvm.hexagon.M2.mmaculs.rs0" => "__builtin_HEXAGON_M2_mmaculs_rs0", + "llvm.hexagon.M2.mmaculs.rs1" => "__builtin_HEXAGON_M2_mmaculs_rs1", + "llvm.hexagon.M2.mmaculs.s0" => "__builtin_HEXAGON_M2_mmaculs_s0", + "llvm.hexagon.M2.mmaculs.s1" => "__builtin_HEXAGON_M2_mmaculs_s1", + "llvm.hexagon.M2.mmpyh.rs0" => "__builtin_HEXAGON_M2_mmpyh_rs0", + "llvm.hexagon.M2.mmpyh.rs1" => "__builtin_HEXAGON_M2_mmpyh_rs1", + "llvm.hexagon.M2.mmpyh.s0" => "__builtin_HEXAGON_M2_mmpyh_s0", + "llvm.hexagon.M2.mmpyh.s1" => "__builtin_HEXAGON_M2_mmpyh_s1", + "llvm.hexagon.M2.mmpyl.rs0" => "__builtin_HEXAGON_M2_mmpyl_rs0", + "llvm.hexagon.M2.mmpyl.rs1" => "__builtin_HEXAGON_M2_mmpyl_rs1", + "llvm.hexagon.M2.mmpyl.s0" => "__builtin_HEXAGON_M2_mmpyl_s0", + "llvm.hexagon.M2.mmpyl.s1" => "__builtin_HEXAGON_M2_mmpyl_s1", + "llvm.hexagon.M2.mmpyuh.rs0" => "__builtin_HEXAGON_M2_mmpyuh_rs0", + "llvm.hexagon.M2.mmpyuh.rs1" => "__builtin_HEXAGON_M2_mmpyuh_rs1", + "llvm.hexagon.M2.mmpyuh.s0" => "__builtin_HEXAGON_M2_mmpyuh_s0", + "llvm.hexagon.M2.mmpyuh.s1" => "__builtin_HEXAGON_M2_mmpyuh_s1", + "llvm.hexagon.M2.mmpyul.rs0" => "__builtin_HEXAGON_M2_mmpyul_rs0", + "llvm.hexagon.M2.mmpyul.rs1" => "__builtin_HEXAGON_M2_mmpyul_rs1", + "llvm.hexagon.M2.mmpyul.s0" => "__builtin_HEXAGON_M2_mmpyul_s0", + "llvm.hexagon.M2.mmpyul.s1" => "__builtin_HEXAGON_M2_mmpyul_s1", + "llvm.hexagon.M2.mpy.acc.hh.s0" => "__builtin_HEXAGON_M2_mpy_acc_hh_s0", + "llvm.hexagon.M2.mpy.acc.hh.s1" => "__builtin_HEXAGON_M2_mpy_acc_hh_s1", + "llvm.hexagon.M2.mpy.acc.hl.s0" => "__builtin_HEXAGON_M2_mpy_acc_hl_s0", + "llvm.hexagon.M2.mpy.acc.hl.s1" => "__builtin_HEXAGON_M2_mpy_acc_hl_s1", + "llvm.hexagon.M2.mpy.acc.lh.s0" => "__builtin_HEXAGON_M2_mpy_acc_lh_s0", + "llvm.hexagon.M2.mpy.acc.lh.s1" => "__builtin_HEXAGON_M2_mpy_acc_lh_s1", + "llvm.hexagon.M2.mpy.acc.ll.s0" => "__builtin_HEXAGON_M2_mpy_acc_ll_s0", + "llvm.hexagon.M2.mpy.acc.ll.s1" => "__builtin_HEXAGON_M2_mpy_acc_ll_s1", + "llvm.hexagon.M2.mpy.acc.sat.hh.s0" => "__builtin_HEXAGON_M2_mpy_acc_sat_hh_s0", + "llvm.hexagon.M2.mpy.acc.sat.hh.s1" => "__builtin_HEXAGON_M2_mpy_acc_sat_hh_s1", + "llvm.hexagon.M2.mpy.acc.sat.hl.s0" => "__builtin_HEXAGON_M2_mpy_acc_sat_hl_s0", + "llvm.hexagon.M2.mpy.acc.sat.hl.s1" => "__builtin_HEXAGON_M2_mpy_acc_sat_hl_s1", + "llvm.hexagon.M2.mpy.acc.sat.lh.s0" => "__builtin_HEXAGON_M2_mpy_acc_sat_lh_s0", + "llvm.hexagon.M2.mpy.acc.sat.lh.s1" => "__builtin_HEXAGON_M2_mpy_acc_sat_lh_s1", + "llvm.hexagon.M2.mpy.acc.sat.ll.s0" => "__builtin_HEXAGON_M2_mpy_acc_sat_ll_s0", + "llvm.hexagon.M2.mpy.acc.sat.ll.s1" => "__builtin_HEXAGON_M2_mpy_acc_sat_ll_s1", + "llvm.hexagon.M2.mpy.hh.s0" => "__builtin_HEXAGON_M2_mpy_hh_s0", + "llvm.hexagon.M2.mpy.hh.s1" => "__builtin_HEXAGON_M2_mpy_hh_s1", + "llvm.hexagon.M2.mpy.hl.s0" => "__builtin_HEXAGON_M2_mpy_hl_s0", + "llvm.hexagon.M2.mpy.hl.s1" => "__builtin_HEXAGON_M2_mpy_hl_s1", + "llvm.hexagon.M2.mpy.lh.s0" => "__builtin_HEXAGON_M2_mpy_lh_s0", + "llvm.hexagon.M2.mpy.lh.s1" => "__builtin_HEXAGON_M2_mpy_lh_s1", + "llvm.hexagon.M2.mpy.ll.s0" => "__builtin_HEXAGON_M2_mpy_ll_s0", + "llvm.hexagon.M2.mpy.ll.s1" => "__builtin_HEXAGON_M2_mpy_ll_s1", + "llvm.hexagon.M2.mpy.nac.hh.s0" => "__builtin_HEXAGON_M2_mpy_nac_hh_s0", + "llvm.hexagon.M2.mpy.nac.hh.s1" => "__builtin_HEXAGON_M2_mpy_nac_hh_s1", + "llvm.hexagon.M2.mpy.nac.hl.s0" => "__builtin_HEXAGON_M2_mpy_nac_hl_s0", + "llvm.hexagon.M2.mpy.nac.hl.s1" => "__builtin_HEXAGON_M2_mpy_nac_hl_s1", + "llvm.hexagon.M2.mpy.nac.lh.s0" => "__builtin_HEXAGON_M2_mpy_nac_lh_s0", + "llvm.hexagon.M2.mpy.nac.lh.s1" => "__builtin_HEXAGON_M2_mpy_nac_lh_s1", + "llvm.hexagon.M2.mpy.nac.ll.s0" => "__builtin_HEXAGON_M2_mpy_nac_ll_s0", + "llvm.hexagon.M2.mpy.nac.ll.s1" => "__builtin_HEXAGON_M2_mpy_nac_ll_s1", + "llvm.hexagon.M2.mpy.nac.sat.hh.s0" => "__builtin_HEXAGON_M2_mpy_nac_sat_hh_s0", + "llvm.hexagon.M2.mpy.nac.sat.hh.s1" => "__builtin_HEXAGON_M2_mpy_nac_sat_hh_s1", + "llvm.hexagon.M2.mpy.nac.sat.hl.s0" => "__builtin_HEXAGON_M2_mpy_nac_sat_hl_s0", + "llvm.hexagon.M2.mpy.nac.sat.hl.s1" => "__builtin_HEXAGON_M2_mpy_nac_sat_hl_s1", + "llvm.hexagon.M2.mpy.nac.sat.lh.s0" => "__builtin_HEXAGON_M2_mpy_nac_sat_lh_s0", + "llvm.hexagon.M2.mpy.nac.sat.lh.s1" => "__builtin_HEXAGON_M2_mpy_nac_sat_lh_s1", + "llvm.hexagon.M2.mpy.nac.sat.ll.s0" => "__builtin_HEXAGON_M2_mpy_nac_sat_ll_s0", + "llvm.hexagon.M2.mpy.nac.sat.ll.s1" => "__builtin_HEXAGON_M2_mpy_nac_sat_ll_s1", + "llvm.hexagon.M2.mpy.rnd.hh.s0" => "__builtin_HEXAGON_M2_mpy_rnd_hh_s0", + "llvm.hexagon.M2.mpy.rnd.hh.s1" => "__builtin_HEXAGON_M2_mpy_rnd_hh_s1", + "llvm.hexagon.M2.mpy.rnd.hl.s0" => "__builtin_HEXAGON_M2_mpy_rnd_hl_s0", + "llvm.hexagon.M2.mpy.rnd.hl.s1" => "__builtin_HEXAGON_M2_mpy_rnd_hl_s1", + "llvm.hexagon.M2.mpy.rnd.lh.s0" => "__builtin_HEXAGON_M2_mpy_rnd_lh_s0", + "llvm.hexagon.M2.mpy.rnd.lh.s1" => "__builtin_HEXAGON_M2_mpy_rnd_lh_s1", + "llvm.hexagon.M2.mpy.rnd.ll.s0" => "__builtin_HEXAGON_M2_mpy_rnd_ll_s0", + "llvm.hexagon.M2.mpy.rnd.ll.s1" => "__builtin_HEXAGON_M2_mpy_rnd_ll_s1", + "llvm.hexagon.M2.mpy.sat.hh.s0" => "__builtin_HEXAGON_M2_mpy_sat_hh_s0", + "llvm.hexagon.M2.mpy.sat.hh.s1" => "__builtin_HEXAGON_M2_mpy_sat_hh_s1", + "llvm.hexagon.M2.mpy.sat.hl.s0" => "__builtin_HEXAGON_M2_mpy_sat_hl_s0", + "llvm.hexagon.M2.mpy.sat.hl.s1" => "__builtin_HEXAGON_M2_mpy_sat_hl_s1", + "llvm.hexagon.M2.mpy.sat.lh.s0" => "__builtin_HEXAGON_M2_mpy_sat_lh_s0", + "llvm.hexagon.M2.mpy.sat.lh.s1" => "__builtin_HEXAGON_M2_mpy_sat_lh_s1", + "llvm.hexagon.M2.mpy.sat.ll.s0" => "__builtin_HEXAGON_M2_mpy_sat_ll_s0", + "llvm.hexagon.M2.mpy.sat.ll.s1" => "__builtin_HEXAGON_M2_mpy_sat_ll_s1", + "llvm.hexagon.M2.mpy.sat.rnd.hh.s0" => "__builtin_HEXAGON_M2_mpy_sat_rnd_hh_s0", + "llvm.hexagon.M2.mpy.sat.rnd.hh.s1" => "__builtin_HEXAGON_M2_mpy_sat_rnd_hh_s1", + "llvm.hexagon.M2.mpy.sat.rnd.hl.s0" => "__builtin_HEXAGON_M2_mpy_sat_rnd_hl_s0", + "llvm.hexagon.M2.mpy.sat.rnd.hl.s1" => "__builtin_HEXAGON_M2_mpy_sat_rnd_hl_s1", + "llvm.hexagon.M2.mpy.sat.rnd.lh.s0" => "__builtin_HEXAGON_M2_mpy_sat_rnd_lh_s0", + "llvm.hexagon.M2.mpy.sat.rnd.lh.s1" => "__builtin_HEXAGON_M2_mpy_sat_rnd_lh_s1", + "llvm.hexagon.M2.mpy.sat.rnd.ll.s0" => "__builtin_HEXAGON_M2_mpy_sat_rnd_ll_s0", + "llvm.hexagon.M2.mpy.sat.rnd.ll.s1" => "__builtin_HEXAGON_M2_mpy_sat_rnd_ll_s1", + "llvm.hexagon.M2.mpy.up" => "__builtin_HEXAGON_M2_mpy_up", + "llvm.hexagon.M2.mpy.up.s1" => "__builtin_HEXAGON_M2_mpy_up_s1", + "llvm.hexagon.M2.mpy.up.s1.sat" => "__builtin_HEXAGON_M2_mpy_up_s1_sat", + "llvm.hexagon.M2.mpyd.acc.hh.s0" => "__builtin_HEXAGON_M2_mpyd_acc_hh_s0", + "llvm.hexagon.M2.mpyd.acc.hh.s1" => "__builtin_HEXAGON_M2_mpyd_acc_hh_s1", + "llvm.hexagon.M2.mpyd.acc.hl.s0" => "__builtin_HEXAGON_M2_mpyd_acc_hl_s0", + "llvm.hexagon.M2.mpyd.acc.hl.s1" => "__builtin_HEXAGON_M2_mpyd_acc_hl_s1", + "llvm.hexagon.M2.mpyd.acc.lh.s0" => "__builtin_HEXAGON_M2_mpyd_acc_lh_s0", + "llvm.hexagon.M2.mpyd.acc.lh.s1" => "__builtin_HEXAGON_M2_mpyd_acc_lh_s1", + "llvm.hexagon.M2.mpyd.acc.ll.s0" => "__builtin_HEXAGON_M2_mpyd_acc_ll_s0", + "llvm.hexagon.M2.mpyd.acc.ll.s1" => "__builtin_HEXAGON_M2_mpyd_acc_ll_s1", + "llvm.hexagon.M2.mpyd.hh.s0" => "__builtin_HEXAGON_M2_mpyd_hh_s0", + "llvm.hexagon.M2.mpyd.hh.s1" => "__builtin_HEXAGON_M2_mpyd_hh_s1", + "llvm.hexagon.M2.mpyd.hl.s0" => "__builtin_HEXAGON_M2_mpyd_hl_s0", + "llvm.hexagon.M2.mpyd.hl.s1" => "__builtin_HEXAGON_M2_mpyd_hl_s1", + "llvm.hexagon.M2.mpyd.lh.s0" => "__builtin_HEXAGON_M2_mpyd_lh_s0", + "llvm.hexagon.M2.mpyd.lh.s1" => "__builtin_HEXAGON_M2_mpyd_lh_s1", + "llvm.hexagon.M2.mpyd.ll.s0" => "__builtin_HEXAGON_M2_mpyd_ll_s0", + "llvm.hexagon.M2.mpyd.ll.s1" => "__builtin_HEXAGON_M2_mpyd_ll_s1", + "llvm.hexagon.M2.mpyd.nac.hh.s0" => "__builtin_HEXAGON_M2_mpyd_nac_hh_s0", + "llvm.hexagon.M2.mpyd.nac.hh.s1" => "__builtin_HEXAGON_M2_mpyd_nac_hh_s1", + "llvm.hexagon.M2.mpyd.nac.hl.s0" => "__builtin_HEXAGON_M2_mpyd_nac_hl_s0", + "llvm.hexagon.M2.mpyd.nac.hl.s1" => "__builtin_HEXAGON_M2_mpyd_nac_hl_s1", + "llvm.hexagon.M2.mpyd.nac.lh.s0" => "__builtin_HEXAGON_M2_mpyd_nac_lh_s0", + "llvm.hexagon.M2.mpyd.nac.lh.s1" => "__builtin_HEXAGON_M2_mpyd_nac_lh_s1", + "llvm.hexagon.M2.mpyd.nac.ll.s0" => "__builtin_HEXAGON_M2_mpyd_nac_ll_s0", + "llvm.hexagon.M2.mpyd.nac.ll.s1" => "__builtin_HEXAGON_M2_mpyd_nac_ll_s1", + "llvm.hexagon.M2.mpyd.rnd.hh.s0" => "__builtin_HEXAGON_M2_mpyd_rnd_hh_s0", + "llvm.hexagon.M2.mpyd.rnd.hh.s1" => "__builtin_HEXAGON_M2_mpyd_rnd_hh_s1", + "llvm.hexagon.M2.mpyd.rnd.hl.s0" => "__builtin_HEXAGON_M2_mpyd_rnd_hl_s0", + "llvm.hexagon.M2.mpyd.rnd.hl.s1" => "__builtin_HEXAGON_M2_mpyd_rnd_hl_s1", + "llvm.hexagon.M2.mpyd.rnd.lh.s0" => "__builtin_HEXAGON_M2_mpyd_rnd_lh_s0", + "llvm.hexagon.M2.mpyd.rnd.lh.s1" => "__builtin_HEXAGON_M2_mpyd_rnd_lh_s1", + "llvm.hexagon.M2.mpyd.rnd.ll.s0" => "__builtin_HEXAGON_M2_mpyd_rnd_ll_s0", + "llvm.hexagon.M2.mpyd.rnd.ll.s1" => "__builtin_HEXAGON_M2_mpyd_rnd_ll_s1", + "llvm.hexagon.M2.mpyi" => "__builtin_HEXAGON_M2_mpyi", + "llvm.hexagon.M2.mpysmi" => "__builtin_HEXAGON_M2_mpysmi", + "llvm.hexagon.M2.mpysu.up" => "__builtin_HEXAGON_M2_mpysu_up", + "llvm.hexagon.M2.mpyu.acc.hh.s0" => "__builtin_HEXAGON_M2_mpyu_acc_hh_s0", + "llvm.hexagon.M2.mpyu.acc.hh.s1" => "__builtin_HEXAGON_M2_mpyu_acc_hh_s1", + "llvm.hexagon.M2.mpyu.acc.hl.s0" => "__builtin_HEXAGON_M2_mpyu_acc_hl_s0", + "llvm.hexagon.M2.mpyu.acc.hl.s1" => "__builtin_HEXAGON_M2_mpyu_acc_hl_s1", + "llvm.hexagon.M2.mpyu.acc.lh.s0" => "__builtin_HEXAGON_M2_mpyu_acc_lh_s0", + "llvm.hexagon.M2.mpyu.acc.lh.s1" => "__builtin_HEXAGON_M2_mpyu_acc_lh_s1", + "llvm.hexagon.M2.mpyu.acc.ll.s0" => "__builtin_HEXAGON_M2_mpyu_acc_ll_s0", + "llvm.hexagon.M2.mpyu.acc.ll.s1" => "__builtin_HEXAGON_M2_mpyu_acc_ll_s1", + "llvm.hexagon.M2.mpyu.hh.s0" => "__builtin_HEXAGON_M2_mpyu_hh_s0", + "llvm.hexagon.M2.mpyu.hh.s1" => "__builtin_HEXAGON_M2_mpyu_hh_s1", + "llvm.hexagon.M2.mpyu.hl.s0" => "__builtin_HEXAGON_M2_mpyu_hl_s0", + "llvm.hexagon.M2.mpyu.hl.s1" => "__builtin_HEXAGON_M2_mpyu_hl_s1", + "llvm.hexagon.M2.mpyu.lh.s0" => "__builtin_HEXAGON_M2_mpyu_lh_s0", + "llvm.hexagon.M2.mpyu.lh.s1" => "__builtin_HEXAGON_M2_mpyu_lh_s1", + "llvm.hexagon.M2.mpyu.ll.s0" => "__builtin_HEXAGON_M2_mpyu_ll_s0", + "llvm.hexagon.M2.mpyu.ll.s1" => "__builtin_HEXAGON_M2_mpyu_ll_s1", + "llvm.hexagon.M2.mpyu.nac.hh.s0" => "__builtin_HEXAGON_M2_mpyu_nac_hh_s0", + "llvm.hexagon.M2.mpyu.nac.hh.s1" => "__builtin_HEXAGON_M2_mpyu_nac_hh_s1", + "llvm.hexagon.M2.mpyu.nac.hl.s0" => "__builtin_HEXAGON_M2_mpyu_nac_hl_s0", + "llvm.hexagon.M2.mpyu.nac.hl.s1" => "__builtin_HEXAGON_M2_mpyu_nac_hl_s1", + "llvm.hexagon.M2.mpyu.nac.lh.s0" => "__builtin_HEXAGON_M2_mpyu_nac_lh_s0", + "llvm.hexagon.M2.mpyu.nac.lh.s1" => "__builtin_HEXAGON_M2_mpyu_nac_lh_s1", + "llvm.hexagon.M2.mpyu.nac.ll.s0" => "__builtin_HEXAGON_M2_mpyu_nac_ll_s0", + "llvm.hexagon.M2.mpyu.nac.ll.s1" => "__builtin_HEXAGON_M2_mpyu_nac_ll_s1", + "llvm.hexagon.M2.mpyu.up" => "__builtin_HEXAGON_M2_mpyu_up", + "llvm.hexagon.M2.mpyud.acc.hh.s0" => "__builtin_HEXAGON_M2_mpyud_acc_hh_s0", + "llvm.hexagon.M2.mpyud.acc.hh.s1" => "__builtin_HEXAGON_M2_mpyud_acc_hh_s1", + "llvm.hexagon.M2.mpyud.acc.hl.s0" => "__builtin_HEXAGON_M2_mpyud_acc_hl_s0", + "llvm.hexagon.M2.mpyud.acc.hl.s1" => "__builtin_HEXAGON_M2_mpyud_acc_hl_s1", + "llvm.hexagon.M2.mpyud.acc.lh.s0" => "__builtin_HEXAGON_M2_mpyud_acc_lh_s0", + "llvm.hexagon.M2.mpyud.acc.lh.s1" => "__builtin_HEXAGON_M2_mpyud_acc_lh_s1", + "llvm.hexagon.M2.mpyud.acc.ll.s0" => "__builtin_HEXAGON_M2_mpyud_acc_ll_s0", + "llvm.hexagon.M2.mpyud.acc.ll.s1" => "__builtin_HEXAGON_M2_mpyud_acc_ll_s1", + "llvm.hexagon.M2.mpyud.hh.s0" => "__builtin_HEXAGON_M2_mpyud_hh_s0", + "llvm.hexagon.M2.mpyud.hh.s1" => "__builtin_HEXAGON_M2_mpyud_hh_s1", + "llvm.hexagon.M2.mpyud.hl.s0" => "__builtin_HEXAGON_M2_mpyud_hl_s0", + "llvm.hexagon.M2.mpyud.hl.s1" => "__builtin_HEXAGON_M2_mpyud_hl_s1", + "llvm.hexagon.M2.mpyud.lh.s0" => "__builtin_HEXAGON_M2_mpyud_lh_s0", + "llvm.hexagon.M2.mpyud.lh.s1" => "__builtin_HEXAGON_M2_mpyud_lh_s1", + "llvm.hexagon.M2.mpyud.ll.s0" => "__builtin_HEXAGON_M2_mpyud_ll_s0", + "llvm.hexagon.M2.mpyud.ll.s1" => "__builtin_HEXAGON_M2_mpyud_ll_s1", + "llvm.hexagon.M2.mpyud.nac.hh.s0" => "__builtin_HEXAGON_M2_mpyud_nac_hh_s0", + "llvm.hexagon.M2.mpyud.nac.hh.s1" => "__builtin_HEXAGON_M2_mpyud_nac_hh_s1", + "llvm.hexagon.M2.mpyud.nac.hl.s0" => "__builtin_HEXAGON_M2_mpyud_nac_hl_s0", + "llvm.hexagon.M2.mpyud.nac.hl.s1" => "__builtin_HEXAGON_M2_mpyud_nac_hl_s1", + "llvm.hexagon.M2.mpyud.nac.lh.s0" => "__builtin_HEXAGON_M2_mpyud_nac_lh_s0", + "llvm.hexagon.M2.mpyud.nac.lh.s1" => "__builtin_HEXAGON_M2_mpyud_nac_lh_s1", + "llvm.hexagon.M2.mpyud.nac.ll.s0" => "__builtin_HEXAGON_M2_mpyud_nac_ll_s0", + "llvm.hexagon.M2.mpyud.nac.ll.s1" => "__builtin_HEXAGON_M2_mpyud_nac_ll_s1", + "llvm.hexagon.M2.mpyui" => "__builtin_HEXAGON_M2_mpyui", + "llvm.hexagon.M2.nacci" => "__builtin_HEXAGON_M2_nacci", + "llvm.hexagon.M2.naccii" => "__builtin_HEXAGON_M2_naccii", + "llvm.hexagon.M2.subacc" => "__builtin_HEXAGON_M2_subacc", + "llvm.hexagon.M2.vabsdiffh" => "__builtin_HEXAGON_M2_vabsdiffh", + "llvm.hexagon.M2.vabsdiffw" => "__builtin_HEXAGON_M2_vabsdiffw", + "llvm.hexagon.M2.vcmac.s0.sat.i" => "__builtin_HEXAGON_M2_vcmac_s0_sat_i", + "llvm.hexagon.M2.vcmac.s0.sat.r" => "__builtin_HEXAGON_M2_vcmac_s0_sat_r", + "llvm.hexagon.M2.vcmpy.s0.sat.i" => "__builtin_HEXAGON_M2_vcmpy_s0_sat_i", + "llvm.hexagon.M2.vcmpy.s0.sat.r" => "__builtin_HEXAGON_M2_vcmpy_s0_sat_r", + "llvm.hexagon.M2.vcmpy.s1.sat.i" => "__builtin_HEXAGON_M2_vcmpy_s1_sat_i", + "llvm.hexagon.M2.vcmpy.s1.sat.r" => "__builtin_HEXAGON_M2_vcmpy_s1_sat_r", + "llvm.hexagon.M2.vdmacs.s0" => "__builtin_HEXAGON_M2_vdmacs_s0", + "llvm.hexagon.M2.vdmacs.s1" => "__builtin_HEXAGON_M2_vdmacs_s1", + "llvm.hexagon.M2.vdmpyrs.s0" => "__builtin_HEXAGON_M2_vdmpyrs_s0", + "llvm.hexagon.M2.vdmpyrs.s1" => "__builtin_HEXAGON_M2_vdmpyrs_s1", + "llvm.hexagon.M2.vdmpys.s0" => "__builtin_HEXAGON_M2_vdmpys_s0", + "llvm.hexagon.M2.vdmpys.s1" => "__builtin_HEXAGON_M2_vdmpys_s1", + "llvm.hexagon.M2.vmac2" => "__builtin_HEXAGON_M2_vmac2", + "llvm.hexagon.M2.vmac2es" => "__builtin_HEXAGON_M2_vmac2es", + "llvm.hexagon.M2.vmac2es.s0" => "__builtin_HEXAGON_M2_vmac2es_s0", + "llvm.hexagon.M2.vmac2es.s1" => "__builtin_HEXAGON_M2_vmac2es_s1", + "llvm.hexagon.M2.vmac2s.s0" => "__builtin_HEXAGON_M2_vmac2s_s0", + "llvm.hexagon.M2.vmac2s.s1" => "__builtin_HEXAGON_M2_vmac2s_s1", + "llvm.hexagon.M2.vmac2su.s0" => "__builtin_HEXAGON_M2_vmac2su_s0", + "llvm.hexagon.M2.vmac2su.s1" => "__builtin_HEXAGON_M2_vmac2su_s1", + "llvm.hexagon.M2.vmpy2es.s0" => "__builtin_HEXAGON_M2_vmpy2es_s0", + "llvm.hexagon.M2.vmpy2es.s1" => "__builtin_HEXAGON_M2_vmpy2es_s1", + "llvm.hexagon.M2.vmpy2s.s0" => "__builtin_HEXAGON_M2_vmpy2s_s0", + "llvm.hexagon.M2.vmpy2s.s0pack" => "__builtin_HEXAGON_M2_vmpy2s_s0pack", + "llvm.hexagon.M2.vmpy2s.s1" => "__builtin_HEXAGON_M2_vmpy2s_s1", + "llvm.hexagon.M2.vmpy2s.s1pack" => "__builtin_HEXAGON_M2_vmpy2s_s1pack", + "llvm.hexagon.M2.vmpy2su.s0" => "__builtin_HEXAGON_M2_vmpy2su_s0", + "llvm.hexagon.M2.vmpy2su.s1" => "__builtin_HEXAGON_M2_vmpy2su_s1", + "llvm.hexagon.M2.vraddh" => "__builtin_HEXAGON_M2_vraddh", + "llvm.hexagon.M2.vradduh" => "__builtin_HEXAGON_M2_vradduh", + "llvm.hexagon.M2.vrcmaci.s0" => "__builtin_HEXAGON_M2_vrcmaci_s0", + "llvm.hexagon.M2.vrcmaci.s0c" => "__builtin_HEXAGON_M2_vrcmaci_s0c", + "llvm.hexagon.M2.vrcmacr.s0" => "__builtin_HEXAGON_M2_vrcmacr_s0", + "llvm.hexagon.M2.vrcmacr.s0c" => "__builtin_HEXAGON_M2_vrcmacr_s0c", + "llvm.hexagon.M2.vrcmpyi.s0" => "__builtin_HEXAGON_M2_vrcmpyi_s0", + "llvm.hexagon.M2.vrcmpyi.s0c" => "__builtin_HEXAGON_M2_vrcmpyi_s0c", + "llvm.hexagon.M2.vrcmpyr.s0" => "__builtin_HEXAGON_M2_vrcmpyr_s0", + "llvm.hexagon.M2.vrcmpyr.s0c" => "__builtin_HEXAGON_M2_vrcmpyr_s0c", + "llvm.hexagon.M2.vrcmpys.acc.s1" => "__builtin_HEXAGON_M2_vrcmpys_acc_s1", + "llvm.hexagon.M2.vrcmpys.s1" => "__builtin_HEXAGON_M2_vrcmpys_s1", + "llvm.hexagon.M2.vrcmpys.s1rp" => "__builtin_HEXAGON_M2_vrcmpys_s1rp", + "llvm.hexagon.M2.vrmac.s0" => "__builtin_HEXAGON_M2_vrmac_s0", + "llvm.hexagon.M2.vrmpy.s0" => "__builtin_HEXAGON_M2_vrmpy_s0", + "llvm.hexagon.M2.xor.xacc" => "__builtin_HEXAGON_M2_xor_xacc", + "llvm.hexagon.M4.and.and" => "__builtin_HEXAGON_M4_and_and", + "llvm.hexagon.M4.and.andn" => "__builtin_HEXAGON_M4_and_andn", + "llvm.hexagon.M4.and.or" => "__builtin_HEXAGON_M4_and_or", + "llvm.hexagon.M4.and.xor" => "__builtin_HEXAGON_M4_and_xor", + "llvm.hexagon.M4.cmpyi.wh" => "__builtin_HEXAGON_M4_cmpyi_wh", + "llvm.hexagon.M4.cmpyi.whc" => "__builtin_HEXAGON_M4_cmpyi_whc", + "llvm.hexagon.M4.cmpyr.wh" => "__builtin_HEXAGON_M4_cmpyr_wh", + "llvm.hexagon.M4.cmpyr.whc" => "__builtin_HEXAGON_M4_cmpyr_whc", + "llvm.hexagon.M4.mac.up.s1.sat" => "__builtin_HEXAGON_M4_mac_up_s1_sat", + "llvm.hexagon.M4.mpyri.addi" => "__builtin_HEXAGON_M4_mpyri_addi", + "llvm.hexagon.M4.mpyri.addr" => "__builtin_HEXAGON_M4_mpyri_addr", + "llvm.hexagon.M4.mpyri.addr.u2" => "__builtin_HEXAGON_M4_mpyri_addr_u2", + "llvm.hexagon.M4.mpyrr.addi" => "__builtin_HEXAGON_M4_mpyrr_addi", + "llvm.hexagon.M4.mpyrr.addr" => "__builtin_HEXAGON_M4_mpyrr_addr", + "llvm.hexagon.M4.nac.up.s1.sat" => "__builtin_HEXAGON_M4_nac_up_s1_sat", + "llvm.hexagon.M4.or.and" => "__builtin_HEXAGON_M4_or_and", + "llvm.hexagon.M4.or.andn" => "__builtin_HEXAGON_M4_or_andn", + "llvm.hexagon.M4.or.or" => "__builtin_HEXAGON_M4_or_or", + "llvm.hexagon.M4.or.xor" => "__builtin_HEXAGON_M4_or_xor", + "llvm.hexagon.M4.pmpyw" => "__builtin_HEXAGON_M4_pmpyw", + "llvm.hexagon.M4.pmpyw.acc" => "__builtin_HEXAGON_M4_pmpyw_acc", + "llvm.hexagon.M4.vpmpyh" => "__builtin_HEXAGON_M4_vpmpyh", + "llvm.hexagon.M4.vpmpyh.acc" => "__builtin_HEXAGON_M4_vpmpyh_acc", + "llvm.hexagon.M4.vrmpyeh.acc.s0" => "__builtin_HEXAGON_M4_vrmpyeh_acc_s0", + "llvm.hexagon.M4.vrmpyeh.acc.s1" => "__builtin_HEXAGON_M4_vrmpyeh_acc_s1", + "llvm.hexagon.M4.vrmpyeh.s0" => "__builtin_HEXAGON_M4_vrmpyeh_s0", + "llvm.hexagon.M4.vrmpyeh.s1" => "__builtin_HEXAGON_M4_vrmpyeh_s1", + "llvm.hexagon.M4.vrmpyoh.acc.s0" => "__builtin_HEXAGON_M4_vrmpyoh_acc_s0", + "llvm.hexagon.M4.vrmpyoh.acc.s1" => "__builtin_HEXAGON_M4_vrmpyoh_acc_s1", + "llvm.hexagon.M4.vrmpyoh.s0" => "__builtin_HEXAGON_M4_vrmpyoh_s0", + "llvm.hexagon.M4.vrmpyoh.s1" => "__builtin_HEXAGON_M4_vrmpyoh_s1", + "llvm.hexagon.M4.xor.and" => "__builtin_HEXAGON_M4_xor_and", + "llvm.hexagon.M4.xor.andn" => "__builtin_HEXAGON_M4_xor_andn", + "llvm.hexagon.M4.xor.or" => "__builtin_HEXAGON_M4_xor_or", + "llvm.hexagon.M4.xor.xacc" => "__builtin_HEXAGON_M4_xor_xacc", + "llvm.hexagon.M5.vdmacbsu" => "__builtin_HEXAGON_M5_vdmacbsu", + "llvm.hexagon.M5.vdmpybsu" => "__builtin_HEXAGON_M5_vdmpybsu", + "llvm.hexagon.M5.vmacbsu" => "__builtin_HEXAGON_M5_vmacbsu", + "llvm.hexagon.M5.vmacbuu" => "__builtin_HEXAGON_M5_vmacbuu", + "llvm.hexagon.M5.vmpybsu" => "__builtin_HEXAGON_M5_vmpybsu", + "llvm.hexagon.M5.vmpybuu" => "__builtin_HEXAGON_M5_vmpybuu", + "llvm.hexagon.M5.vrmacbsu" => "__builtin_HEXAGON_M5_vrmacbsu", + "llvm.hexagon.M5.vrmacbuu" => "__builtin_HEXAGON_M5_vrmacbuu", + "llvm.hexagon.M5.vrmpybsu" => "__builtin_HEXAGON_M5_vrmpybsu", + "llvm.hexagon.M5.vrmpybuu" => "__builtin_HEXAGON_M5_vrmpybuu", + "llvm.hexagon.M6.vabsdiffb" => "__builtin_HEXAGON_M6_vabsdiffb", + "llvm.hexagon.M6.vabsdiffub" => "__builtin_HEXAGON_M6_vabsdiffub", + "llvm.hexagon.S2.addasl.rrri" => "__builtin_HEXAGON_S2_addasl_rrri", + "llvm.hexagon.S2.asl.i.p" => "__builtin_HEXAGON_S2_asl_i_p", + "llvm.hexagon.S2.asl.i.p.acc" => "__builtin_HEXAGON_S2_asl_i_p_acc", + "llvm.hexagon.S2.asl.i.p.and" => "__builtin_HEXAGON_S2_asl_i_p_and", + "llvm.hexagon.S2.asl.i.p.nac" => "__builtin_HEXAGON_S2_asl_i_p_nac", + "llvm.hexagon.S2.asl.i.p.or" => "__builtin_HEXAGON_S2_asl_i_p_or", + "llvm.hexagon.S2.asl.i.p.xacc" => "__builtin_HEXAGON_S2_asl_i_p_xacc", + "llvm.hexagon.S2.asl.i.r" => "__builtin_HEXAGON_S2_asl_i_r", + "llvm.hexagon.S2.asl.i.r.acc" => "__builtin_HEXAGON_S2_asl_i_r_acc", + "llvm.hexagon.S2.asl.i.r.and" => "__builtin_HEXAGON_S2_asl_i_r_and", + "llvm.hexagon.S2.asl.i.r.nac" => "__builtin_HEXAGON_S2_asl_i_r_nac", + "llvm.hexagon.S2.asl.i.r.or" => "__builtin_HEXAGON_S2_asl_i_r_or", + "llvm.hexagon.S2.asl.i.r.sat" => "__builtin_HEXAGON_S2_asl_i_r_sat", + "llvm.hexagon.S2.asl.i.r.xacc" => "__builtin_HEXAGON_S2_asl_i_r_xacc", + "llvm.hexagon.S2.asl.i.vh" => "__builtin_HEXAGON_S2_asl_i_vh", + "llvm.hexagon.S2.asl.i.vw" => "__builtin_HEXAGON_S2_asl_i_vw", + "llvm.hexagon.S2.asl.r.p" => "__builtin_HEXAGON_S2_asl_r_p", + "llvm.hexagon.S2.asl.r.p.acc" => "__builtin_HEXAGON_S2_asl_r_p_acc", + "llvm.hexagon.S2.asl.r.p.and" => "__builtin_HEXAGON_S2_asl_r_p_and", + "llvm.hexagon.S2.asl.r.p.nac" => "__builtin_HEXAGON_S2_asl_r_p_nac", + "llvm.hexagon.S2.asl.r.p.or" => "__builtin_HEXAGON_S2_asl_r_p_or", + "llvm.hexagon.S2.asl.r.p.xor" => "__builtin_HEXAGON_S2_asl_r_p_xor", + "llvm.hexagon.S2.asl.r.r" => "__builtin_HEXAGON_S2_asl_r_r", + "llvm.hexagon.S2.asl.r.r.acc" => "__builtin_HEXAGON_S2_asl_r_r_acc", + "llvm.hexagon.S2.asl.r.r.and" => "__builtin_HEXAGON_S2_asl_r_r_and", + "llvm.hexagon.S2.asl.r.r.nac" => "__builtin_HEXAGON_S2_asl_r_r_nac", + "llvm.hexagon.S2.asl.r.r.or" => "__builtin_HEXAGON_S2_asl_r_r_or", + "llvm.hexagon.S2.asl.r.r.sat" => "__builtin_HEXAGON_S2_asl_r_r_sat", + "llvm.hexagon.S2.asl.r.vh" => "__builtin_HEXAGON_S2_asl_r_vh", + "llvm.hexagon.S2.asl.r.vw" => "__builtin_HEXAGON_S2_asl_r_vw", + "llvm.hexagon.S2.asr.i.p" => "__builtin_HEXAGON_S2_asr_i_p", + "llvm.hexagon.S2.asr.i.p.acc" => "__builtin_HEXAGON_S2_asr_i_p_acc", + "llvm.hexagon.S2.asr.i.p.and" => "__builtin_HEXAGON_S2_asr_i_p_and", + "llvm.hexagon.S2.asr.i.p.nac" => "__builtin_HEXAGON_S2_asr_i_p_nac", + "llvm.hexagon.S2.asr.i.p.or" => "__builtin_HEXAGON_S2_asr_i_p_or", + "llvm.hexagon.S2.asr.i.p.rnd" => "__builtin_HEXAGON_S2_asr_i_p_rnd", + "llvm.hexagon.S2.asr.i.p.rnd.goodsyntax" => "__builtin_HEXAGON_S2_asr_i_p_rnd_goodsyntax", + "llvm.hexagon.S2.asr.i.r" => "__builtin_HEXAGON_S2_asr_i_r", + "llvm.hexagon.S2.asr.i.r.acc" => "__builtin_HEXAGON_S2_asr_i_r_acc", + "llvm.hexagon.S2.asr.i.r.and" => "__builtin_HEXAGON_S2_asr_i_r_and", + "llvm.hexagon.S2.asr.i.r.nac" => "__builtin_HEXAGON_S2_asr_i_r_nac", + "llvm.hexagon.S2.asr.i.r.or" => "__builtin_HEXAGON_S2_asr_i_r_or", + "llvm.hexagon.S2.asr.i.r.rnd" => "__builtin_HEXAGON_S2_asr_i_r_rnd", + "llvm.hexagon.S2.asr.i.r.rnd.goodsyntax" => "__builtin_HEXAGON_S2_asr_i_r_rnd_goodsyntax", + "llvm.hexagon.S2.asr.i.svw.trun" => "__builtin_HEXAGON_S2_asr_i_svw_trun", + "llvm.hexagon.S2.asr.i.vh" => "__builtin_HEXAGON_S2_asr_i_vh", + "llvm.hexagon.S2.asr.i.vw" => "__builtin_HEXAGON_S2_asr_i_vw", + "llvm.hexagon.S2.asr.r.p" => "__builtin_HEXAGON_S2_asr_r_p", + "llvm.hexagon.S2.asr.r.p.acc" => "__builtin_HEXAGON_S2_asr_r_p_acc", + "llvm.hexagon.S2.asr.r.p.and" => "__builtin_HEXAGON_S2_asr_r_p_and", + "llvm.hexagon.S2.asr.r.p.nac" => "__builtin_HEXAGON_S2_asr_r_p_nac", + "llvm.hexagon.S2.asr.r.p.or" => "__builtin_HEXAGON_S2_asr_r_p_or", + "llvm.hexagon.S2.asr.r.p.xor" => "__builtin_HEXAGON_S2_asr_r_p_xor", + "llvm.hexagon.S2.asr.r.r" => "__builtin_HEXAGON_S2_asr_r_r", + "llvm.hexagon.S2.asr.r.r.acc" => "__builtin_HEXAGON_S2_asr_r_r_acc", + "llvm.hexagon.S2.asr.r.r.and" => "__builtin_HEXAGON_S2_asr_r_r_and", + "llvm.hexagon.S2.asr.r.r.nac" => "__builtin_HEXAGON_S2_asr_r_r_nac", + "llvm.hexagon.S2.asr.r.r.or" => "__builtin_HEXAGON_S2_asr_r_r_or", + "llvm.hexagon.S2.asr.r.r.sat" => "__builtin_HEXAGON_S2_asr_r_r_sat", + "llvm.hexagon.S2.asr.r.svw.trun" => "__builtin_HEXAGON_S2_asr_r_svw_trun", + "llvm.hexagon.S2.asr.r.vh" => "__builtin_HEXAGON_S2_asr_r_vh", + "llvm.hexagon.S2.asr.r.vw" => "__builtin_HEXAGON_S2_asr_r_vw", + "llvm.hexagon.S2.brev" => "__builtin_HEXAGON_S2_brev", + "llvm.hexagon.S2.brevp" => "__builtin_HEXAGON_S2_brevp", + "llvm.hexagon.S2.cabacencbin" => "__builtin_HEXAGON_S2_cabacencbin", + "llvm.hexagon.S2.cl0" => "__builtin_HEXAGON_S2_cl0", + "llvm.hexagon.S2.cl0p" => "__builtin_HEXAGON_S2_cl0p", + "llvm.hexagon.S2.cl1" => "__builtin_HEXAGON_S2_cl1", + "llvm.hexagon.S2.cl1p" => "__builtin_HEXAGON_S2_cl1p", + "llvm.hexagon.S2.clb" => "__builtin_HEXAGON_S2_clb", + "llvm.hexagon.S2.clbnorm" => "__builtin_HEXAGON_S2_clbnorm", + "llvm.hexagon.S2.clbp" => "__builtin_HEXAGON_S2_clbp", + "llvm.hexagon.S2.clrbit.i" => "__builtin_HEXAGON_S2_clrbit_i", + "llvm.hexagon.S2.clrbit.r" => "__builtin_HEXAGON_S2_clrbit_r", + "llvm.hexagon.S2.ct0" => "__builtin_HEXAGON_S2_ct0", + "llvm.hexagon.S2.ct0p" => "__builtin_HEXAGON_S2_ct0p", + "llvm.hexagon.S2.ct1" => "__builtin_HEXAGON_S2_ct1", + "llvm.hexagon.S2.ct1p" => "__builtin_HEXAGON_S2_ct1p", + "llvm.hexagon.S2.deinterleave" => "__builtin_HEXAGON_S2_deinterleave", + "llvm.hexagon.S2.extractu" => "__builtin_HEXAGON_S2_extractu", + "llvm.hexagon.S2.extractu.rp" => "__builtin_HEXAGON_S2_extractu_rp", + "llvm.hexagon.S2.extractup" => "__builtin_HEXAGON_S2_extractup", + "llvm.hexagon.S2.extractup.rp" => "__builtin_HEXAGON_S2_extractup_rp", + "llvm.hexagon.S2.insert" => "__builtin_HEXAGON_S2_insert", + "llvm.hexagon.S2.insert.rp" => "__builtin_HEXAGON_S2_insert_rp", + "llvm.hexagon.S2.insertp" => "__builtin_HEXAGON_S2_insertp", + "llvm.hexagon.S2.insertp.rp" => "__builtin_HEXAGON_S2_insertp_rp", + "llvm.hexagon.S2.interleave" => "__builtin_HEXAGON_S2_interleave", + "llvm.hexagon.S2.lfsp" => "__builtin_HEXAGON_S2_lfsp", + "llvm.hexagon.S2.lsl.r.p" => "__builtin_HEXAGON_S2_lsl_r_p", + "llvm.hexagon.S2.lsl.r.p.acc" => "__builtin_HEXAGON_S2_lsl_r_p_acc", + "llvm.hexagon.S2.lsl.r.p.and" => "__builtin_HEXAGON_S2_lsl_r_p_and", + "llvm.hexagon.S2.lsl.r.p.nac" => "__builtin_HEXAGON_S2_lsl_r_p_nac", + "llvm.hexagon.S2.lsl.r.p.or" => "__builtin_HEXAGON_S2_lsl_r_p_or", + "llvm.hexagon.S2.lsl.r.p.xor" => "__builtin_HEXAGON_S2_lsl_r_p_xor", + "llvm.hexagon.S2.lsl.r.r" => "__builtin_HEXAGON_S2_lsl_r_r", + "llvm.hexagon.S2.lsl.r.r.acc" => "__builtin_HEXAGON_S2_lsl_r_r_acc", + "llvm.hexagon.S2.lsl.r.r.and" => "__builtin_HEXAGON_S2_lsl_r_r_and", + "llvm.hexagon.S2.lsl.r.r.nac" => "__builtin_HEXAGON_S2_lsl_r_r_nac", + "llvm.hexagon.S2.lsl.r.r.or" => "__builtin_HEXAGON_S2_lsl_r_r_or", + "llvm.hexagon.S2.lsl.r.vh" => "__builtin_HEXAGON_S2_lsl_r_vh", + "llvm.hexagon.S2.lsl.r.vw" => "__builtin_HEXAGON_S2_lsl_r_vw", + "llvm.hexagon.S2.lsr.i.p" => "__builtin_HEXAGON_S2_lsr_i_p", + "llvm.hexagon.S2.lsr.i.p.acc" => "__builtin_HEXAGON_S2_lsr_i_p_acc", + "llvm.hexagon.S2.lsr.i.p.and" => "__builtin_HEXAGON_S2_lsr_i_p_and", + "llvm.hexagon.S2.lsr.i.p.nac" => "__builtin_HEXAGON_S2_lsr_i_p_nac", + "llvm.hexagon.S2.lsr.i.p.or" => "__builtin_HEXAGON_S2_lsr_i_p_or", + "llvm.hexagon.S2.lsr.i.p.xacc" => "__builtin_HEXAGON_S2_lsr_i_p_xacc", + "llvm.hexagon.S2.lsr.i.r" => "__builtin_HEXAGON_S2_lsr_i_r", + "llvm.hexagon.S2.lsr.i.r.acc" => "__builtin_HEXAGON_S2_lsr_i_r_acc", + "llvm.hexagon.S2.lsr.i.r.and" => "__builtin_HEXAGON_S2_lsr_i_r_and", + "llvm.hexagon.S2.lsr.i.r.nac" => "__builtin_HEXAGON_S2_lsr_i_r_nac", + "llvm.hexagon.S2.lsr.i.r.or" => "__builtin_HEXAGON_S2_lsr_i_r_or", + "llvm.hexagon.S2.lsr.i.r.xacc" => "__builtin_HEXAGON_S2_lsr_i_r_xacc", + "llvm.hexagon.S2.lsr.i.vh" => "__builtin_HEXAGON_S2_lsr_i_vh", + "llvm.hexagon.S2.lsr.i.vw" => "__builtin_HEXAGON_S2_lsr_i_vw", + "llvm.hexagon.S2.lsr.r.p" => "__builtin_HEXAGON_S2_lsr_r_p", + "llvm.hexagon.S2.lsr.r.p.acc" => "__builtin_HEXAGON_S2_lsr_r_p_acc", + "llvm.hexagon.S2.lsr.r.p.and" => "__builtin_HEXAGON_S2_lsr_r_p_and", + "llvm.hexagon.S2.lsr.r.p.nac" => "__builtin_HEXAGON_S2_lsr_r_p_nac", + "llvm.hexagon.S2.lsr.r.p.or" => "__builtin_HEXAGON_S2_lsr_r_p_or", + "llvm.hexagon.S2.lsr.r.p.xor" => "__builtin_HEXAGON_S2_lsr_r_p_xor", + "llvm.hexagon.S2.lsr.r.r" => "__builtin_HEXAGON_S2_lsr_r_r", + "llvm.hexagon.S2.lsr.r.r.acc" => "__builtin_HEXAGON_S2_lsr_r_r_acc", + "llvm.hexagon.S2.lsr.r.r.and" => "__builtin_HEXAGON_S2_lsr_r_r_and", + "llvm.hexagon.S2.lsr.r.r.nac" => "__builtin_HEXAGON_S2_lsr_r_r_nac", + "llvm.hexagon.S2.lsr.r.r.or" => "__builtin_HEXAGON_S2_lsr_r_r_or", + "llvm.hexagon.S2.lsr.r.vh" => "__builtin_HEXAGON_S2_lsr_r_vh", + "llvm.hexagon.S2.lsr.r.vw" => "__builtin_HEXAGON_S2_lsr_r_vw", + "llvm.hexagon.S2.packhl" => "__builtin_HEXAGON_S2_packhl", + "llvm.hexagon.S2.parityp" => "__builtin_HEXAGON_S2_parityp", + "llvm.hexagon.S2.setbit.i" => "__builtin_HEXAGON_S2_setbit_i", + "llvm.hexagon.S2.setbit.r" => "__builtin_HEXAGON_S2_setbit_r", + "llvm.hexagon.S2.shuffeb" => "__builtin_HEXAGON_S2_shuffeb", + "llvm.hexagon.S2.shuffeh" => "__builtin_HEXAGON_S2_shuffeh", + "llvm.hexagon.S2.shuffob" => "__builtin_HEXAGON_S2_shuffob", + "llvm.hexagon.S2.shuffoh" => "__builtin_HEXAGON_S2_shuffoh", + "llvm.hexagon.S2.svsathb" => "__builtin_HEXAGON_S2_svsathb", + "llvm.hexagon.S2.svsathub" => "__builtin_HEXAGON_S2_svsathub", + "llvm.hexagon.S2.tableidxb.goodsyntax" => "__builtin_HEXAGON_S2_tableidxb_goodsyntax", + "llvm.hexagon.S2.tableidxd.goodsyntax" => "__builtin_HEXAGON_S2_tableidxd_goodsyntax", + "llvm.hexagon.S2.tableidxh.goodsyntax" => "__builtin_HEXAGON_S2_tableidxh_goodsyntax", + "llvm.hexagon.S2.tableidxw.goodsyntax" => "__builtin_HEXAGON_S2_tableidxw_goodsyntax", + "llvm.hexagon.S2.togglebit.i" => "__builtin_HEXAGON_S2_togglebit_i", + "llvm.hexagon.S2.togglebit.r" => "__builtin_HEXAGON_S2_togglebit_r", + "llvm.hexagon.S2.tstbit.i" => "__builtin_HEXAGON_S2_tstbit_i", + "llvm.hexagon.S2.tstbit.r" => "__builtin_HEXAGON_S2_tstbit_r", + "llvm.hexagon.S2.valignib" => "__builtin_HEXAGON_S2_valignib", + "llvm.hexagon.S2.valignrb" => "__builtin_HEXAGON_S2_valignrb", + "llvm.hexagon.S2.vcnegh" => "__builtin_HEXAGON_S2_vcnegh", + "llvm.hexagon.S2.vcrotate" => "__builtin_HEXAGON_S2_vcrotate", + "llvm.hexagon.S2.vrcnegh" => "__builtin_HEXAGON_S2_vrcnegh", + "llvm.hexagon.S2.vrndpackwh" => "__builtin_HEXAGON_S2_vrndpackwh", + "llvm.hexagon.S2.vrndpackwhs" => "__builtin_HEXAGON_S2_vrndpackwhs", + "llvm.hexagon.S2.vsathb" => "__builtin_HEXAGON_S2_vsathb", + "llvm.hexagon.S2.vsathb.nopack" => "__builtin_HEXAGON_S2_vsathb_nopack", + "llvm.hexagon.S2.vsathub" => "__builtin_HEXAGON_S2_vsathub", + "llvm.hexagon.S2.vsathub.nopack" => "__builtin_HEXAGON_S2_vsathub_nopack", + "llvm.hexagon.S2.vsatwh" => "__builtin_HEXAGON_S2_vsatwh", + "llvm.hexagon.S2.vsatwh.nopack" => "__builtin_HEXAGON_S2_vsatwh_nopack", + "llvm.hexagon.S2.vsatwuh" => "__builtin_HEXAGON_S2_vsatwuh", + "llvm.hexagon.S2.vsatwuh.nopack" => "__builtin_HEXAGON_S2_vsatwuh_nopack", + "llvm.hexagon.S2.vsplatrb" => "__builtin_HEXAGON_S2_vsplatrb", + "llvm.hexagon.S2.vsplatrh" => "__builtin_HEXAGON_S2_vsplatrh", + "llvm.hexagon.S2.vspliceib" => "__builtin_HEXAGON_S2_vspliceib", + "llvm.hexagon.S2.vsplicerb" => "__builtin_HEXAGON_S2_vsplicerb", + "llvm.hexagon.S2.vsxtbh" => "__builtin_HEXAGON_S2_vsxtbh", + "llvm.hexagon.S2.vsxthw" => "__builtin_HEXAGON_S2_vsxthw", + "llvm.hexagon.S2.vtrunehb" => "__builtin_HEXAGON_S2_vtrunehb", + "llvm.hexagon.S2.vtrunewh" => "__builtin_HEXAGON_S2_vtrunewh", + "llvm.hexagon.S2.vtrunohb" => "__builtin_HEXAGON_S2_vtrunohb", + "llvm.hexagon.S2.vtrunowh" => "__builtin_HEXAGON_S2_vtrunowh", + "llvm.hexagon.S2.vzxtbh" => "__builtin_HEXAGON_S2_vzxtbh", + "llvm.hexagon.S2.vzxthw" => "__builtin_HEXAGON_S2_vzxthw", + "llvm.hexagon.S4.addaddi" => "__builtin_HEXAGON_S4_addaddi", + "llvm.hexagon.S4.addi.asl.ri" => "__builtin_HEXAGON_S4_addi_asl_ri", + "llvm.hexagon.S4.addi.lsr.ri" => "__builtin_HEXAGON_S4_addi_lsr_ri", + "llvm.hexagon.S4.andi.asl.ri" => "__builtin_HEXAGON_S4_andi_asl_ri", + "llvm.hexagon.S4.andi.lsr.ri" => "__builtin_HEXAGON_S4_andi_lsr_ri", + "llvm.hexagon.S4.clbaddi" => "__builtin_HEXAGON_S4_clbaddi", + "llvm.hexagon.S4.clbpaddi" => "__builtin_HEXAGON_S4_clbpaddi", + "llvm.hexagon.S4.clbpnorm" => "__builtin_HEXAGON_S4_clbpnorm", + "llvm.hexagon.S4.extract" => "__builtin_HEXAGON_S4_extract", + "llvm.hexagon.S4.extract.rp" => "__builtin_HEXAGON_S4_extract_rp", + "llvm.hexagon.S4.extractp" => "__builtin_HEXAGON_S4_extractp", + "llvm.hexagon.S4.extractp.rp" => "__builtin_HEXAGON_S4_extractp_rp", + "llvm.hexagon.S4.lsli" => "__builtin_HEXAGON_S4_lsli", + "llvm.hexagon.S4.ntstbit.i" => "__builtin_HEXAGON_S4_ntstbit_i", + "llvm.hexagon.S4.ntstbit.r" => "__builtin_HEXAGON_S4_ntstbit_r", + "llvm.hexagon.S4.or.andi" => "__builtin_HEXAGON_S4_or_andi", + "llvm.hexagon.S4.or.andix" => "__builtin_HEXAGON_S4_or_andix", + "llvm.hexagon.S4.or.ori" => "__builtin_HEXAGON_S4_or_ori", + "llvm.hexagon.S4.ori.asl.ri" => "__builtin_HEXAGON_S4_ori_asl_ri", + "llvm.hexagon.S4.ori.lsr.ri" => "__builtin_HEXAGON_S4_ori_lsr_ri", + "llvm.hexagon.S4.parity" => "__builtin_HEXAGON_S4_parity", + "llvm.hexagon.S4.subaddi" => "__builtin_HEXAGON_S4_subaddi", + "llvm.hexagon.S4.subi.asl.ri" => "__builtin_HEXAGON_S4_subi_asl_ri", + "llvm.hexagon.S4.subi.lsr.ri" => "__builtin_HEXAGON_S4_subi_lsr_ri", + "llvm.hexagon.S4.vrcrotate" => "__builtin_HEXAGON_S4_vrcrotate", + "llvm.hexagon.S4.vrcrotate.acc" => "__builtin_HEXAGON_S4_vrcrotate_acc", + "llvm.hexagon.S4.vxaddsubh" => "__builtin_HEXAGON_S4_vxaddsubh", + "llvm.hexagon.S4.vxaddsubhr" => "__builtin_HEXAGON_S4_vxaddsubhr", + "llvm.hexagon.S4.vxaddsubw" => "__builtin_HEXAGON_S4_vxaddsubw", + "llvm.hexagon.S4.vxsubaddh" => "__builtin_HEXAGON_S4_vxsubaddh", + "llvm.hexagon.S4.vxsubaddhr" => "__builtin_HEXAGON_S4_vxsubaddhr", + "llvm.hexagon.S4.vxsubaddw" => "__builtin_HEXAGON_S4_vxsubaddw", + "llvm.hexagon.S5.asrhub.rnd.sat.goodsyntax" => "__builtin_HEXAGON_S5_asrhub_rnd_sat_goodsyntax", + "llvm.hexagon.S5.asrhub.sat" => "__builtin_HEXAGON_S5_asrhub_sat", + "llvm.hexagon.S5.popcountp" => "__builtin_HEXAGON_S5_popcountp", + "llvm.hexagon.S5.vasrhrnd.goodsyntax" => "__builtin_HEXAGON_S5_vasrhrnd_goodsyntax", + "llvm.hexagon.S6.rol.i.p" => "__builtin_HEXAGON_S6_rol_i_p", + "llvm.hexagon.S6.rol.i.p.acc" => "__builtin_HEXAGON_S6_rol_i_p_acc", + "llvm.hexagon.S6.rol.i.p.and" => "__builtin_HEXAGON_S6_rol_i_p_and", + "llvm.hexagon.S6.rol.i.p.nac" => "__builtin_HEXAGON_S6_rol_i_p_nac", + "llvm.hexagon.S6.rol.i.p.or" => "__builtin_HEXAGON_S6_rol_i_p_or", + "llvm.hexagon.S6.rol.i.p.xacc" => "__builtin_HEXAGON_S6_rol_i_p_xacc", + "llvm.hexagon.S6.rol.i.r" => "__builtin_HEXAGON_S6_rol_i_r", + "llvm.hexagon.S6.rol.i.r.acc" => "__builtin_HEXAGON_S6_rol_i_r_acc", + "llvm.hexagon.S6.rol.i.r.and" => "__builtin_HEXAGON_S6_rol_i_r_and", + "llvm.hexagon.S6.rol.i.r.nac" => "__builtin_HEXAGON_S6_rol_i_r_nac", + "llvm.hexagon.S6.rol.i.r.or" => "__builtin_HEXAGON_S6_rol_i_r_or", + "llvm.hexagon.S6.rol.i.r.xacc" => "__builtin_HEXAGON_S6_rol_i_r_xacc", + "llvm.hexagon.S6.vsplatrbp" => "__builtin_HEXAGON_S6_vsplatrbp", + "llvm.hexagon.S6.vtrunehb.ppp" => "__builtin_HEXAGON_S6_vtrunehb_ppp", + "llvm.hexagon.S6.vtrunohb.ppp" => "__builtin_HEXAGON_S6_vtrunohb_ppp", + "llvm.hexagon.SI.to.SXTHI.asrh" => "__builtin_SI_to_SXTHI_asrh", + "llvm.hexagon.V6.extractw" => "__builtin_HEXAGON_V6_extractw", + "llvm.hexagon.V6.extractw.128B" => "__builtin_HEXAGON_V6_extractw_128B", + "llvm.hexagon.V6.hi" => "__builtin_HEXAGON_V6_hi", + "llvm.hexagon.V6.hi.128B" => "__builtin_HEXAGON_V6_hi_128B", + "llvm.hexagon.V6.lo" => "__builtin_HEXAGON_V6_lo", + "llvm.hexagon.V6.lo.128B" => "__builtin_HEXAGON_V6_lo_128B", + "llvm.hexagon.V6.lvsplatw" => "__builtin_HEXAGON_V6_lvsplatw", + "llvm.hexagon.V6.lvsplatw.128B" => "__builtin_HEXAGON_V6_lvsplatw_128B", + "llvm.hexagon.V6.vabsdiffh" => "__builtin_HEXAGON_V6_vabsdiffh", + "llvm.hexagon.V6.vabsdiffh.128B" => "__builtin_HEXAGON_V6_vabsdiffh_128B", + "llvm.hexagon.V6.vabsdiffub" => "__builtin_HEXAGON_V6_vabsdiffub", + "llvm.hexagon.V6.vabsdiffub.128B" => "__builtin_HEXAGON_V6_vabsdiffub_128B", + "llvm.hexagon.V6.vabsdiffuh" => "__builtin_HEXAGON_V6_vabsdiffuh", + "llvm.hexagon.V6.vabsdiffuh.128B" => "__builtin_HEXAGON_V6_vabsdiffuh_128B", + "llvm.hexagon.V6.vabsdiffw" => "__builtin_HEXAGON_V6_vabsdiffw", + "llvm.hexagon.V6.vabsdiffw.128B" => "__builtin_HEXAGON_V6_vabsdiffw_128B", + "llvm.hexagon.V6.vabsh" => "__builtin_HEXAGON_V6_vabsh", + "llvm.hexagon.V6.vabsh.128B" => "__builtin_HEXAGON_V6_vabsh_128B", + "llvm.hexagon.V6.vabsh.sat" => "__builtin_HEXAGON_V6_vabsh_sat", + "llvm.hexagon.V6.vabsh.sat.128B" => "__builtin_HEXAGON_V6_vabsh_sat_128B", + "llvm.hexagon.V6.vabsw" => "__builtin_HEXAGON_V6_vabsw", + "llvm.hexagon.V6.vabsw.128B" => "__builtin_HEXAGON_V6_vabsw_128B", + "llvm.hexagon.V6.vabsw.sat" => "__builtin_HEXAGON_V6_vabsw_sat", + "llvm.hexagon.V6.vabsw.sat.128B" => "__builtin_HEXAGON_V6_vabsw_sat_128B", + "llvm.hexagon.V6.vaddb" => "__builtin_HEXAGON_V6_vaddb", + "llvm.hexagon.V6.vaddb.128B" => "__builtin_HEXAGON_V6_vaddb_128B", + "llvm.hexagon.V6.vaddb.dv" => "__builtin_HEXAGON_V6_vaddb_dv", + "llvm.hexagon.V6.vaddb.dv.128B" => "__builtin_HEXAGON_V6_vaddb_dv_128B", + "llvm.hexagon.V6.vaddh" => "__builtin_HEXAGON_V6_vaddh", + "llvm.hexagon.V6.vaddh.128B" => "__builtin_HEXAGON_V6_vaddh_128B", + "llvm.hexagon.V6.vaddh.dv" => "__builtin_HEXAGON_V6_vaddh_dv", + "llvm.hexagon.V6.vaddh.dv.128B" => "__builtin_HEXAGON_V6_vaddh_dv_128B", + "llvm.hexagon.V6.vaddhsat" => "__builtin_HEXAGON_V6_vaddhsat", + "llvm.hexagon.V6.vaddhsat.128B" => "__builtin_HEXAGON_V6_vaddhsat_128B", + "llvm.hexagon.V6.vaddhsat.dv" => "__builtin_HEXAGON_V6_vaddhsat_dv", + "llvm.hexagon.V6.vaddhsat.dv.128B" => "__builtin_HEXAGON_V6_vaddhsat_dv_128B", + "llvm.hexagon.V6.vaddhw" => "__builtin_HEXAGON_V6_vaddhw", + "llvm.hexagon.V6.vaddhw.128B" => "__builtin_HEXAGON_V6_vaddhw_128B", + "llvm.hexagon.V6.vaddubh" => "__builtin_HEXAGON_V6_vaddubh", + "llvm.hexagon.V6.vaddubh.128B" => "__builtin_HEXAGON_V6_vaddubh_128B", + "llvm.hexagon.V6.vaddubsat" => "__builtin_HEXAGON_V6_vaddubsat", + "llvm.hexagon.V6.vaddubsat.128B" => "__builtin_HEXAGON_V6_vaddubsat_128B", + "llvm.hexagon.V6.vaddubsat.dv" => "__builtin_HEXAGON_V6_vaddubsat_dv", + "llvm.hexagon.V6.vaddubsat.dv.128B" => "__builtin_HEXAGON_V6_vaddubsat_dv_128B", + "llvm.hexagon.V6.vadduhsat" => "__builtin_HEXAGON_V6_vadduhsat", + "llvm.hexagon.V6.vadduhsat.128B" => "__builtin_HEXAGON_V6_vadduhsat_128B", + "llvm.hexagon.V6.vadduhsat.dv" => "__builtin_HEXAGON_V6_vadduhsat_dv", + "llvm.hexagon.V6.vadduhsat.dv.128B" => "__builtin_HEXAGON_V6_vadduhsat_dv_128B", + "llvm.hexagon.V6.vadduhw" => "__builtin_HEXAGON_V6_vadduhw", + "llvm.hexagon.V6.vadduhw.128B" => "__builtin_HEXAGON_V6_vadduhw_128B", + "llvm.hexagon.V6.vaddw" => "__builtin_HEXAGON_V6_vaddw", + "llvm.hexagon.V6.vaddw.128B" => "__builtin_HEXAGON_V6_vaddw_128B", + "llvm.hexagon.V6.vaddw.dv" => "__builtin_HEXAGON_V6_vaddw_dv", + "llvm.hexagon.V6.vaddw.dv.128B" => "__builtin_HEXAGON_V6_vaddw_dv_128B", + "llvm.hexagon.V6.vaddwsat" => "__builtin_HEXAGON_V6_vaddwsat", + "llvm.hexagon.V6.vaddwsat.128B" => "__builtin_HEXAGON_V6_vaddwsat_128B", + "llvm.hexagon.V6.vaddwsat.dv" => "__builtin_HEXAGON_V6_vaddwsat_dv", + "llvm.hexagon.V6.vaddwsat.dv.128B" => "__builtin_HEXAGON_V6_vaddwsat_dv_128B", + "llvm.hexagon.V6.valignb" => "__builtin_HEXAGON_V6_valignb", + "llvm.hexagon.V6.valignb.128B" => "__builtin_HEXAGON_V6_valignb_128B", + "llvm.hexagon.V6.valignbi" => "__builtin_HEXAGON_V6_valignbi", + "llvm.hexagon.V6.valignbi.128B" => "__builtin_HEXAGON_V6_valignbi_128B", + "llvm.hexagon.V6.vand" => "__builtin_HEXAGON_V6_vand", + "llvm.hexagon.V6.vand.128B" => "__builtin_HEXAGON_V6_vand_128B", + "llvm.hexagon.V6.vaslh" => "__builtin_HEXAGON_V6_vaslh", + "llvm.hexagon.V6.vaslh.128B" => "__builtin_HEXAGON_V6_vaslh_128B", + "llvm.hexagon.V6.vaslhv" => "__builtin_HEXAGON_V6_vaslhv", + "llvm.hexagon.V6.vaslhv.128B" => "__builtin_HEXAGON_V6_vaslhv_128B", + "llvm.hexagon.V6.vaslw" => "__builtin_HEXAGON_V6_vaslw", + "llvm.hexagon.V6.vaslw.128B" => "__builtin_HEXAGON_V6_vaslw_128B", + "llvm.hexagon.V6.vaslw.acc" => "__builtin_HEXAGON_V6_vaslw_acc", + "llvm.hexagon.V6.vaslw.acc.128B" => "__builtin_HEXAGON_V6_vaslw_acc_128B", + "llvm.hexagon.V6.vaslwv" => "__builtin_HEXAGON_V6_vaslwv", + "llvm.hexagon.V6.vaslwv.128B" => "__builtin_HEXAGON_V6_vaslwv_128B", + "llvm.hexagon.V6.vasrh" => "__builtin_HEXAGON_V6_vasrh", + "llvm.hexagon.V6.vasrh.128B" => "__builtin_HEXAGON_V6_vasrh_128B", + "llvm.hexagon.V6.vasrhbrndsat" => "__builtin_HEXAGON_V6_vasrhbrndsat", + "llvm.hexagon.V6.vasrhbrndsat.128B" => "__builtin_HEXAGON_V6_vasrhbrndsat_128B", + "llvm.hexagon.V6.vasrhubrndsat" => "__builtin_HEXAGON_V6_vasrhubrndsat", + "llvm.hexagon.V6.vasrhubrndsat.128B" => "__builtin_HEXAGON_V6_vasrhubrndsat_128B", + "llvm.hexagon.V6.vasrhubsat" => "__builtin_HEXAGON_V6_vasrhubsat", + "llvm.hexagon.V6.vasrhubsat.128B" => "__builtin_HEXAGON_V6_vasrhubsat_128B", + "llvm.hexagon.V6.vasrhv" => "__builtin_HEXAGON_V6_vasrhv", + "llvm.hexagon.V6.vasrhv.128B" => "__builtin_HEXAGON_V6_vasrhv_128B", + "llvm.hexagon.V6.vasrw" => "__builtin_HEXAGON_V6_vasrw", + "llvm.hexagon.V6.vasrw.128B" => "__builtin_HEXAGON_V6_vasrw_128B", + "llvm.hexagon.V6.vasrw.acc" => "__builtin_HEXAGON_V6_vasrw_acc", + "llvm.hexagon.V6.vasrw.acc.128B" => "__builtin_HEXAGON_V6_vasrw_acc_128B", + "llvm.hexagon.V6.vasrwh" => "__builtin_HEXAGON_V6_vasrwh", + "llvm.hexagon.V6.vasrwh.128B" => "__builtin_HEXAGON_V6_vasrwh_128B", + "llvm.hexagon.V6.vasrwhrndsat" => "__builtin_HEXAGON_V6_vasrwhrndsat", + "llvm.hexagon.V6.vasrwhrndsat.128B" => "__builtin_HEXAGON_V6_vasrwhrndsat_128B", + "llvm.hexagon.V6.vasrwhsat" => "__builtin_HEXAGON_V6_vasrwhsat", + "llvm.hexagon.V6.vasrwhsat.128B" => "__builtin_HEXAGON_V6_vasrwhsat_128B", + "llvm.hexagon.V6.vasrwuhsat" => "__builtin_HEXAGON_V6_vasrwuhsat", + "llvm.hexagon.V6.vasrwuhsat.128B" => "__builtin_HEXAGON_V6_vasrwuhsat_128B", + "llvm.hexagon.V6.vasrwv" => "__builtin_HEXAGON_V6_vasrwv", + "llvm.hexagon.V6.vasrwv.128B" => "__builtin_HEXAGON_V6_vasrwv_128B", + "llvm.hexagon.V6.vassign" => "__builtin_HEXAGON_V6_vassign", + "llvm.hexagon.V6.vassign.128B" => "__builtin_HEXAGON_V6_vassign_128B", + "llvm.hexagon.V6.vassignp" => "__builtin_HEXAGON_V6_vassignp", + "llvm.hexagon.V6.vassignp.128B" => "__builtin_HEXAGON_V6_vassignp_128B", + "llvm.hexagon.V6.vavgh" => "__builtin_HEXAGON_V6_vavgh", + "llvm.hexagon.V6.vavgh.128B" => "__builtin_HEXAGON_V6_vavgh_128B", + "llvm.hexagon.V6.vavghrnd" => "__builtin_HEXAGON_V6_vavghrnd", + "llvm.hexagon.V6.vavghrnd.128B" => "__builtin_HEXAGON_V6_vavghrnd_128B", + "llvm.hexagon.V6.vavgub" => "__builtin_HEXAGON_V6_vavgub", + "llvm.hexagon.V6.vavgub.128B" => "__builtin_HEXAGON_V6_vavgub_128B", + "llvm.hexagon.V6.vavgubrnd" => "__builtin_HEXAGON_V6_vavgubrnd", + "llvm.hexagon.V6.vavgubrnd.128B" => "__builtin_HEXAGON_V6_vavgubrnd_128B", + "llvm.hexagon.V6.vavguh" => "__builtin_HEXAGON_V6_vavguh", + "llvm.hexagon.V6.vavguh.128B" => "__builtin_HEXAGON_V6_vavguh_128B", + "llvm.hexagon.V6.vavguhrnd" => "__builtin_HEXAGON_V6_vavguhrnd", + "llvm.hexagon.V6.vavguhrnd.128B" => "__builtin_HEXAGON_V6_vavguhrnd_128B", + "llvm.hexagon.V6.vavgw" => "__builtin_HEXAGON_V6_vavgw", + "llvm.hexagon.V6.vavgw.128B" => "__builtin_HEXAGON_V6_vavgw_128B", + "llvm.hexagon.V6.vavgwrnd" => "__builtin_HEXAGON_V6_vavgwrnd", + "llvm.hexagon.V6.vavgwrnd.128B" => "__builtin_HEXAGON_V6_vavgwrnd_128B", + "llvm.hexagon.V6.vcl0h" => "__builtin_HEXAGON_V6_vcl0h", + "llvm.hexagon.V6.vcl0h.128B" => "__builtin_HEXAGON_V6_vcl0h_128B", + "llvm.hexagon.V6.vcl0w" => "__builtin_HEXAGON_V6_vcl0w", + "llvm.hexagon.V6.vcl0w.128B" => "__builtin_HEXAGON_V6_vcl0w_128B", + "llvm.hexagon.V6.vcombine" => "__builtin_HEXAGON_V6_vcombine", + "llvm.hexagon.V6.vcombine.128B" => "__builtin_HEXAGON_V6_vcombine_128B", + "llvm.hexagon.V6.vd0" => "__builtin_HEXAGON_V6_vd0", + "llvm.hexagon.V6.vd0.128B" => "__builtin_HEXAGON_V6_vd0_128B", + "llvm.hexagon.V6.vdealb" => "__builtin_HEXAGON_V6_vdealb", + "llvm.hexagon.V6.vdealb.128B" => "__builtin_HEXAGON_V6_vdealb_128B", + "llvm.hexagon.V6.vdealb4w" => "__builtin_HEXAGON_V6_vdealb4w", + "llvm.hexagon.V6.vdealb4w.128B" => "__builtin_HEXAGON_V6_vdealb4w_128B", + "llvm.hexagon.V6.vdealh" => "__builtin_HEXAGON_V6_vdealh", + "llvm.hexagon.V6.vdealh.128B" => "__builtin_HEXAGON_V6_vdealh_128B", + "llvm.hexagon.V6.vdealvdd" => "__builtin_HEXAGON_V6_vdealvdd", + "llvm.hexagon.V6.vdealvdd.128B" => "__builtin_HEXAGON_V6_vdealvdd_128B", + "llvm.hexagon.V6.vdelta" => "__builtin_HEXAGON_V6_vdelta", + "llvm.hexagon.V6.vdelta.128B" => "__builtin_HEXAGON_V6_vdelta_128B", + "llvm.hexagon.V6.vdmpybus" => "__builtin_HEXAGON_V6_vdmpybus", + "llvm.hexagon.V6.vdmpybus.128B" => "__builtin_HEXAGON_V6_vdmpybus_128B", + "llvm.hexagon.V6.vdmpybus.acc" => "__builtin_HEXAGON_V6_vdmpybus_acc", + "llvm.hexagon.V6.vdmpybus.acc.128B" => "__builtin_HEXAGON_V6_vdmpybus_acc_128B", + "llvm.hexagon.V6.vdmpybus.dv" => "__builtin_HEXAGON_V6_vdmpybus_dv", + "llvm.hexagon.V6.vdmpybus.dv.128B" => "__builtin_HEXAGON_V6_vdmpybus_dv_128B", + "llvm.hexagon.V6.vdmpybus.dv.acc" => "__builtin_HEXAGON_V6_vdmpybus_dv_acc", + "llvm.hexagon.V6.vdmpybus.dv.acc.128B" => "__builtin_HEXAGON_V6_vdmpybus_dv_acc_128B", + "llvm.hexagon.V6.vdmpyhb" => "__builtin_HEXAGON_V6_vdmpyhb", + "llvm.hexagon.V6.vdmpyhb.128B" => "__builtin_HEXAGON_V6_vdmpyhb_128B", + "llvm.hexagon.V6.vdmpyhb.acc" => "__builtin_HEXAGON_V6_vdmpyhb_acc", + "llvm.hexagon.V6.vdmpyhb.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhb_acc_128B", + "llvm.hexagon.V6.vdmpyhb.dv" => "__builtin_HEXAGON_V6_vdmpyhb_dv", + "llvm.hexagon.V6.vdmpyhb.dv.128B" => "__builtin_HEXAGON_V6_vdmpyhb_dv_128B", + "llvm.hexagon.V6.vdmpyhb.dv.acc" => "__builtin_HEXAGON_V6_vdmpyhb_dv_acc", + "llvm.hexagon.V6.vdmpyhb.dv.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhb_dv_acc_128B", + "llvm.hexagon.V6.vdmpyhisat" => "__builtin_HEXAGON_V6_vdmpyhisat", + "llvm.hexagon.V6.vdmpyhisat.128B" => "__builtin_HEXAGON_V6_vdmpyhisat_128B", + "llvm.hexagon.V6.vdmpyhisat.acc" => "__builtin_HEXAGON_V6_vdmpyhisat_acc", + "llvm.hexagon.V6.vdmpyhisat.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhisat_acc_128B", + "llvm.hexagon.V6.vdmpyhsat" => "__builtin_HEXAGON_V6_vdmpyhsat", + "llvm.hexagon.V6.vdmpyhsat.128B" => "__builtin_HEXAGON_V6_vdmpyhsat_128B", + "llvm.hexagon.V6.vdmpyhsat.acc" => "__builtin_HEXAGON_V6_vdmpyhsat_acc", + "llvm.hexagon.V6.vdmpyhsat.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhsat_acc_128B", + "llvm.hexagon.V6.vdmpyhsuisat" => "__builtin_HEXAGON_V6_vdmpyhsuisat", + "llvm.hexagon.V6.vdmpyhsuisat.128B" => "__builtin_HEXAGON_V6_vdmpyhsuisat_128B", + "llvm.hexagon.V6.vdmpyhsuisat.acc" => "__builtin_HEXAGON_V6_vdmpyhsuisat_acc", + "llvm.hexagon.V6.vdmpyhsuisat.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhsuisat_acc_128B", + "llvm.hexagon.V6.vdmpyhsusat" => "__builtin_HEXAGON_V6_vdmpyhsusat", + "llvm.hexagon.V6.vdmpyhsusat.128B" => "__builtin_HEXAGON_V6_vdmpyhsusat_128B", + "llvm.hexagon.V6.vdmpyhsusat.acc" => "__builtin_HEXAGON_V6_vdmpyhsusat_acc", + "llvm.hexagon.V6.vdmpyhsusat.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhsusat_acc_128B", + "llvm.hexagon.V6.vdmpyhvsat" => "__builtin_HEXAGON_V6_vdmpyhvsat", + "llvm.hexagon.V6.vdmpyhvsat.128B" => "__builtin_HEXAGON_V6_vdmpyhvsat_128B", + "llvm.hexagon.V6.vdmpyhvsat.acc" => "__builtin_HEXAGON_V6_vdmpyhvsat_acc", + "llvm.hexagon.V6.vdmpyhvsat.acc.128B" => "__builtin_HEXAGON_V6_vdmpyhvsat_acc_128B", + "llvm.hexagon.V6.vdsaduh" => "__builtin_HEXAGON_V6_vdsaduh", + "llvm.hexagon.V6.vdsaduh.128B" => "__builtin_HEXAGON_V6_vdsaduh_128B", + "llvm.hexagon.V6.vdsaduh.acc" => "__builtin_HEXAGON_V6_vdsaduh_acc", + "llvm.hexagon.V6.vdsaduh.acc.128B" => "__builtin_HEXAGON_V6_vdsaduh_acc_128B", + "llvm.hexagon.V6.vinsertwr" => "__builtin_HEXAGON_V6_vinsertwr", + "llvm.hexagon.V6.vinsertwr.128B" => "__builtin_HEXAGON_V6_vinsertwr_128B", + "llvm.hexagon.V6.vlalignb" => "__builtin_HEXAGON_V6_vlalignb", + "llvm.hexagon.V6.vlalignb.128B" => "__builtin_HEXAGON_V6_vlalignb_128B", + "llvm.hexagon.V6.vlalignbi" => "__builtin_HEXAGON_V6_vlalignbi", + "llvm.hexagon.V6.vlalignbi.128B" => "__builtin_HEXAGON_V6_vlalignbi_128B", + "llvm.hexagon.V6.vlsrh" => "__builtin_HEXAGON_V6_vlsrh", + "llvm.hexagon.V6.vlsrh.128B" => "__builtin_HEXAGON_V6_vlsrh_128B", + "llvm.hexagon.V6.vlsrhv" => "__builtin_HEXAGON_V6_vlsrhv", + "llvm.hexagon.V6.vlsrhv.128B" => "__builtin_HEXAGON_V6_vlsrhv_128B", + "llvm.hexagon.V6.vlsrw" => "__builtin_HEXAGON_V6_vlsrw", + "llvm.hexagon.V6.vlsrw.128B" => "__builtin_HEXAGON_V6_vlsrw_128B", + "llvm.hexagon.V6.vlsrwv" => "__builtin_HEXAGON_V6_vlsrwv", + "llvm.hexagon.V6.vlsrwv.128B" => "__builtin_HEXAGON_V6_vlsrwv_128B", + "llvm.hexagon.V6.vlutb" => "__builtin_HEXAGON_V6_vlutb", + "llvm.hexagon.V6.vlutb.128B" => "__builtin_HEXAGON_V6_vlutb_128B", + "llvm.hexagon.V6.vlutb.acc" => "__builtin_HEXAGON_V6_vlutb_acc", + "llvm.hexagon.V6.vlutb.acc.128B" => "__builtin_HEXAGON_V6_vlutb_acc_128B", + "llvm.hexagon.V6.vlutb.dv" => "__builtin_HEXAGON_V6_vlutb_dv", + "llvm.hexagon.V6.vlutb.dv.128B" => "__builtin_HEXAGON_V6_vlutb_dv_128B", + "llvm.hexagon.V6.vlutb.dv.acc" => "__builtin_HEXAGON_V6_vlutb_dv_acc", + "llvm.hexagon.V6.vlutb.dv.acc.128B" => "__builtin_HEXAGON_V6_vlutb_dv_acc_128B", + "llvm.hexagon.V6.vlutvvb" => "__builtin_HEXAGON_V6_vlutvvb", + "llvm.hexagon.V6.vlutvvb.128B" => "__builtin_HEXAGON_V6_vlutvvb_128B", + "llvm.hexagon.V6.vlutvvb.oracc" => "__builtin_HEXAGON_V6_vlutvvb_oracc", + "llvm.hexagon.V6.vlutvvb.oracc.128B" => "__builtin_HEXAGON_V6_vlutvvb_oracc_128B", + "llvm.hexagon.V6.vlutvwh" => "__builtin_HEXAGON_V6_vlutvwh", + "llvm.hexagon.V6.vlutvwh.128B" => "__builtin_HEXAGON_V6_vlutvwh_128B", + "llvm.hexagon.V6.vlutvwh.oracc" => "__builtin_HEXAGON_V6_vlutvwh_oracc", + "llvm.hexagon.V6.vlutvwh.oracc.128B" => "__builtin_HEXAGON_V6_vlutvwh_oracc_128B", + "llvm.hexagon.V6.vmaxh" => "__builtin_HEXAGON_V6_vmaxh", + "llvm.hexagon.V6.vmaxh.128B" => "__builtin_HEXAGON_V6_vmaxh_128B", + "llvm.hexagon.V6.vmaxub" => "__builtin_HEXAGON_V6_vmaxub", + "llvm.hexagon.V6.vmaxub.128B" => "__builtin_HEXAGON_V6_vmaxub_128B", + "llvm.hexagon.V6.vmaxuh" => "__builtin_HEXAGON_V6_vmaxuh", + "llvm.hexagon.V6.vmaxuh.128B" => "__builtin_HEXAGON_V6_vmaxuh_128B", + "llvm.hexagon.V6.vmaxw" => "__builtin_HEXAGON_V6_vmaxw", + "llvm.hexagon.V6.vmaxw.128B" => "__builtin_HEXAGON_V6_vmaxw_128B", + "llvm.hexagon.V6.vminh" => "__builtin_HEXAGON_V6_vminh", + "llvm.hexagon.V6.vminh.128B" => "__builtin_HEXAGON_V6_vminh_128B", + "llvm.hexagon.V6.vminub" => "__builtin_HEXAGON_V6_vminub", + "llvm.hexagon.V6.vminub.128B" => "__builtin_HEXAGON_V6_vminub_128B", + "llvm.hexagon.V6.vminuh" => "__builtin_HEXAGON_V6_vminuh", + "llvm.hexagon.V6.vminuh.128B" => "__builtin_HEXAGON_V6_vminuh_128B", + "llvm.hexagon.V6.vminw" => "__builtin_HEXAGON_V6_vminw", + "llvm.hexagon.V6.vminw.128B" => "__builtin_HEXAGON_V6_vminw_128B", + "llvm.hexagon.V6.vmpabus" => "__builtin_HEXAGON_V6_vmpabus", + "llvm.hexagon.V6.vmpabus.128B" => "__builtin_HEXAGON_V6_vmpabus_128B", + "llvm.hexagon.V6.vmpabus.acc" => "__builtin_HEXAGON_V6_vmpabus_acc", + "llvm.hexagon.V6.vmpabus.acc.128B" => "__builtin_HEXAGON_V6_vmpabus_acc_128B", + "llvm.hexagon.V6.vmpabusv" => "__builtin_HEXAGON_V6_vmpabusv", + "llvm.hexagon.V6.vmpabusv.128B" => "__builtin_HEXAGON_V6_vmpabusv_128B", + "llvm.hexagon.V6.vmpabuuv" => "__builtin_HEXAGON_V6_vmpabuuv", + "llvm.hexagon.V6.vmpabuuv.128B" => "__builtin_HEXAGON_V6_vmpabuuv_128B", + "llvm.hexagon.V6.vmpahb" => "__builtin_HEXAGON_V6_vmpahb", + "llvm.hexagon.V6.vmpahb.128B" => "__builtin_HEXAGON_V6_vmpahb_128B", + "llvm.hexagon.V6.vmpahb.acc" => "__builtin_HEXAGON_V6_vmpahb_acc", + "llvm.hexagon.V6.vmpahb.acc.128B" => "__builtin_HEXAGON_V6_vmpahb_acc_128B", + "llvm.hexagon.V6.vmpybus" => "__builtin_HEXAGON_V6_vmpybus", + "llvm.hexagon.V6.vmpybus.128B" => "__builtin_HEXAGON_V6_vmpybus_128B", + "llvm.hexagon.V6.vmpybus.acc" => "__builtin_HEXAGON_V6_vmpybus_acc", + "llvm.hexagon.V6.vmpybus.acc.128B" => "__builtin_HEXAGON_V6_vmpybus_acc_128B", + "llvm.hexagon.V6.vmpybusv" => "__builtin_HEXAGON_V6_vmpybusv", + "llvm.hexagon.V6.vmpybusv.128B" => "__builtin_HEXAGON_V6_vmpybusv_128B", + "llvm.hexagon.V6.vmpybusv.acc" => "__builtin_HEXAGON_V6_vmpybusv_acc", + "llvm.hexagon.V6.vmpybusv.acc.128B" => "__builtin_HEXAGON_V6_vmpybusv_acc_128B", + "llvm.hexagon.V6.vmpybv" => "__builtin_HEXAGON_V6_vmpybv", + "llvm.hexagon.V6.vmpybv.128B" => "__builtin_HEXAGON_V6_vmpybv_128B", + "llvm.hexagon.V6.vmpybv.acc" => "__builtin_HEXAGON_V6_vmpybv_acc", + "llvm.hexagon.V6.vmpybv.acc.128B" => "__builtin_HEXAGON_V6_vmpybv_acc_128B", + "llvm.hexagon.V6.vmpyewuh" => "__builtin_HEXAGON_V6_vmpyewuh", + "llvm.hexagon.V6.vmpyewuh.128B" => "__builtin_HEXAGON_V6_vmpyewuh_128B", + "llvm.hexagon.V6.vmpyh" => "__builtin_HEXAGON_V6_vmpyh", + "llvm.hexagon.V6.vmpyh.128B" => "__builtin_HEXAGON_V6_vmpyh_128B", + "llvm.hexagon.V6.vmpyhsat.acc" => "__builtin_HEXAGON_V6_vmpyhsat_acc", + "llvm.hexagon.V6.vmpyhsat.acc.128B" => "__builtin_HEXAGON_V6_vmpyhsat_acc_128B", + "llvm.hexagon.V6.vmpyhsrs" => "__builtin_HEXAGON_V6_vmpyhsrs", + "llvm.hexagon.V6.vmpyhsrs.128B" => "__builtin_HEXAGON_V6_vmpyhsrs_128B", + "llvm.hexagon.V6.vmpyhss" => "__builtin_HEXAGON_V6_vmpyhss", + "llvm.hexagon.V6.vmpyhss.128B" => "__builtin_HEXAGON_V6_vmpyhss_128B", + "llvm.hexagon.V6.vmpyhus" => "__builtin_HEXAGON_V6_vmpyhus", + "llvm.hexagon.V6.vmpyhus.128B" => "__builtin_HEXAGON_V6_vmpyhus_128B", + "llvm.hexagon.V6.vmpyhus.acc" => "__builtin_HEXAGON_V6_vmpyhus_acc", + "llvm.hexagon.V6.vmpyhus.acc.128B" => "__builtin_HEXAGON_V6_vmpyhus_acc_128B", + "llvm.hexagon.V6.vmpyhv" => "__builtin_HEXAGON_V6_vmpyhv", + "llvm.hexagon.V6.vmpyhv.128B" => "__builtin_HEXAGON_V6_vmpyhv_128B", + "llvm.hexagon.V6.vmpyhv.acc" => "__builtin_HEXAGON_V6_vmpyhv_acc", + "llvm.hexagon.V6.vmpyhv.acc.128B" => "__builtin_HEXAGON_V6_vmpyhv_acc_128B", + "llvm.hexagon.V6.vmpyhvsrs" => "__builtin_HEXAGON_V6_vmpyhvsrs", + "llvm.hexagon.V6.vmpyhvsrs.128B" => "__builtin_HEXAGON_V6_vmpyhvsrs_128B", + "llvm.hexagon.V6.vmpyieoh" => "__builtin_HEXAGON_V6_vmpyieoh", + "llvm.hexagon.V6.vmpyieoh.128B" => "__builtin_HEXAGON_V6_vmpyieoh_128B", + "llvm.hexagon.V6.vmpyiewh.acc" => "__builtin_HEXAGON_V6_vmpyiewh_acc", + "llvm.hexagon.V6.vmpyiewh.acc.128B" => "__builtin_HEXAGON_V6_vmpyiewh_acc_128B", + "llvm.hexagon.V6.vmpyiewuh" => "__builtin_HEXAGON_V6_vmpyiewuh", + "llvm.hexagon.V6.vmpyiewuh.128B" => "__builtin_HEXAGON_V6_vmpyiewuh_128B", + "llvm.hexagon.V6.vmpyiewuh.acc" => "__builtin_HEXAGON_V6_vmpyiewuh_acc", + "llvm.hexagon.V6.vmpyiewuh.acc.128B" => "__builtin_HEXAGON_V6_vmpyiewuh_acc_128B", + "llvm.hexagon.V6.vmpyih" => "__builtin_HEXAGON_V6_vmpyih", + "llvm.hexagon.V6.vmpyih.128B" => "__builtin_HEXAGON_V6_vmpyih_128B", + "llvm.hexagon.V6.vmpyih.acc" => "__builtin_HEXAGON_V6_vmpyih_acc", + "llvm.hexagon.V6.vmpyih.acc.128B" => "__builtin_HEXAGON_V6_vmpyih_acc_128B", + "llvm.hexagon.V6.vmpyihb" => "__builtin_HEXAGON_V6_vmpyihb", + "llvm.hexagon.V6.vmpyihb.128B" => "__builtin_HEXAGON_V6_vmpyihb_128B", + "llvm.hexagon.V6.vmpyihb.acc" => "__builtin_HEXAGON_V6_vmpyihb_acc", + "llvm.hexagon.V6.vmpyihb.acc.128B" => "__builtin_HEXAGON_V6_vmpyihb_acc_128B", + "llvm.hexagon.V6.vmpyiowh" => "__builtin_HEXAGON_V6_vmpyiowh", + "llvm.hexagon.V6.vmpyiowh.128B" => "__builtin_HEXAGON_V6_vmpyiowh_128B", + "llvm.hexagon.V6.vmpyiwb" => "__builtin_HEXAGON_V6_vmpyiwb", + "llvm.hexagon.V6.vmpyiwb.128B" => "__builtin_HEXAGON_V6_vmpyiwb_128B", + "llvm.hexagon.V6.vmpyiwb.acc" => "__builtin_HEXAGON_V6_vmpyiwb_acc", + "llvm.hexagon.V6.vmpyiwb.acc.128B" => "__builtin_HEXAGON_V6_vmpyiwb_acc_128B", + "llvm.hexagon.V6.vmpyiwh" => "__builtin_HEXAGON_V6_vmpyiwh", + "llvm.hexagon.V6.vmpyiwh.128B" => "__builtin_HEXAGON_V6_vmpyiwh_128B", + "llvm.hexagon.V6.vmpyiwh.acc" => "__builtin_HEXAGON_V6_vmpyiwh_acc", + "llvm.hexagon.V6.vmpyiwh.acc.128B" => "__builtin_HEXAGON_V6_vmpyiwh_acc_128B", + "llvm.hexagon.V6.vmpyowh" => "__builtin_HEXAGON_V6_vmpyowh", + "llvm.hexagon.V6.vmpyowh.128B" => "__builtin_HEXAGON_V6_vmpyowh_128B", + "llvm.hexagon.V6.vmpyowh.rnd" => "__builtin_HEXAGON_V6_vmpyowh_rnd", + "llvm.hexagon.V6.vmpyowh.rnd.128B" => "__builtin_HEXAGON_V6_vmpyowh_rnd_128B", + "llvm.hexagon.V6.vmpyowh.rnd.sacc" => "__builtin_HEXAGON_V6_vmpyowh_rnd_sacc", + "llvm.hexagon.V6.vmpyowh.rnd.sacc.128B" => "__builtin_HEXAGON_V6_vmpyowh_rnd_sacc_128B", + "llvm.hexagon.V6.vmpyowh.sacc" => "__builtin_HEXAGON_V6_vmpyowh_sacc", + "llvm.hexagon.V6.vmpyowh.sacc.128B" => "__builtin_HEXAGON_V6_vmpyowh_sacc_128B", + "llvm.hexagon.V6.vmpyub" => "__builtin_HEXAGON_V6_vmpyub", + "llvm.hexagon.V6.vmpyub.128B" => "__builtin_HEXAGON_V6_vmpyub_128B", + "llvm.hexagon.V6.vmpyub.acc" => "__builtin_HEXAGON_V6_vmpyub_acc", + "llvm.hexagon.V6.vmpyub.acc.128B" => "__builtin_HEXAGON_V6_vmpyub_acc_128B", + "llvm.hexagon.V6.vmpyubv" => "__builtin_HEXAGON_V6_vmpyubv", + "llvm.hexagon.V6.vmpyubv.128B" => "__builtin_HEXAGON_V6_vmpyubv_128B", + "llvm.hexagon.V6.vmpyubv.acc" => "__builtin_HEXAGON_V6_vmpyubv_acc", + "llvm.hexagon.V6.vmpyubv.acc.128B" => "__builtin_HEXAGON_V6_vmpyubv_acc_128B", + "llvm.hexagon.V6.vmpyuh" => "__builtin_HEXAGON_V6_vmpyuh", + "llvm.hexagon.V6.vmpyuh.128B" => "__builtin_HEXAGON_V6_vmpyuh_128B", + "llvm.hexagon.V6.vmpyuh.acc" => "__builtin_HEXAGON_V6_vmpyuh_acc", + "llvm.hexagon.V6.vmpyuh.acc.128B" => "__builtin_HEXAGON_V6_vmpyuh_acc_128B", + "llvm.hexagon.V6.vmpyuhv" => "__builtin_HEXAGON_V6_vmpyuhv", + "llvm.hexagon.V6.vmpyuhv.128B" => "__builtin_HEXAGON_V6_vmpyuhv_128B", + "llvm.hexagon.V6.vmpyuhv.acc" => "__builtin_HEXAGON_V6_vmpyuhv_acc", + "llvm.hexagon.V6.vmpyuhv.acc.128B" => "__builtin_HEXAGON_V6_vmpyuhv_acc_128B", + "llvm.hexagon.V6.vnavgh" => "__builtin_HEXAGON_V6_vnavgh", + "llvm.hexagon.V6.vnavgh.128B" => "__builtin_HEXAGON_V6_vnavgh_128B", + "llvm.hexagon.V6.vnavgub" => "__builtin_HEXAGON_V6_vnavgub", + "llvm.hexagon.V6.vnavgub.128B" => "__builtin_HEXAGON_V6_vnavgub_128B", + "llvm.hexagon.V6.vnavgw" => "__builtin_HEXAGON_V6_vnavgw", + "llvm.hexagon.V6.vnavgw.128B" => "__builtin_HEXAGON_V6_vnavgw_128B", + "llvm.hexagon.V6.vnormamth" => "__builtin_HEXAGON_V6_vnormamth", + "llvm.hexagon.V6.vnormamth.128B" => "__builtin_HEXAGON_V6_vnormamth_128B", + "llvm.hexagon.V6.vnormamtw" => "__builtin_HEXAGON_V6_vnormamtw", + "llvm.hexagon.V6.vnormamtw.128B" => "__builtin_HEXAGON_V6_vnormamtw_128B", + "llvm.hexagon.V6.vnot" => "__builtin_HEXAGON_V6_vnot", + "llvm.hexagon.V6.vnot.128B" => "__builtin_HEXAGON_V6_vnot_128B", + "llvm.hexagon.V6.vor" => "__builtin_HEXAGON_V6_vor", + "llvm.hexagon.V6.vor.128B" => "__builtin_HEXAGON_V6_vor_128B", + "llvm.hexagon.V6.vpackeb" => "__builtin_HEXAGON_V6_vpackeb", + "llvm.hexagon.V6.vpackeb.128B" => "__builtin_HEXAGON_V6_vpackeb_128B", + "llvm.hexagon.V6.vpackeh" => "__builtin_HEXAGON_V6_vpackeh", + "llvm.hexagon.V6.vpackeh.128B" => "__builtin_HEXAGON_V6_vpackeh_128B", + "llvm.hexagon.V6.vpackhb.sat" => "__builtin_HEXAGON_V6_vpackhb_sat", + "llvm.hexagon.V6.vpackhb.sat.128B" => "__builtin_HEXAGON_V6_vpackhb_sat_128B", + "llvm.hexagon.V6.vpackhub.sat" => "__builtin_HEXAGON_V6_vpackhub_sat", + "llvm.hexagon.V6.vpackhub.sat.128B" => "__builtin_HEXAGON_V6_vpackhub_sat_128B", + "llvm.hexagon.V6.vpackob" => "__builtin_HEXAGON_V6_vpackob", + "llvm.hexagon.V6.vpackob.128B" => "__builtin_HEXAGON_V6_vpackob_128B", + "llvm.hexagon.V6.vpackoh" => "__builtin_HEXAGON_V6_vpackoh", + "llvm.hexagon.V6.vpackoh.128B" => "__builtin_HEXAGON_V6_vpackoh_128B", + "llvm.hexagon.V6.vpackwh.sat" => "__builtin_HEXAGON_V6_vpackwh_sat", + "llvm.hexagon.V6.vpackwh.sat.128B" => "__builtin_HEXAGON_V6_vpackwh_sat_128B", + "llvm.hexagon.V6.vpackwuh.sat" => "__builtin_HEXAGON_V6_vpackwuh_sat", + "llvm.hexagon.V6.vpackwuh.sat.128B" => "__builtin_HEXAGON_V6_vpackwuh_sat_128B", + "llvm.hexagon.V6.vpopcounth" => "__builtin_HEXAGON_V6_vpopcounth", + "llvm.hexagon.V6.vpopcounth.128B" => "__builtin_HEXAGON_V6_vpopcounth_128B", + "llvm.hexagon.V6.vrdelta" => "__builtin_HEXAGON_V6_vrdelta", + "llvm.hexagon.V6.vrdelta.128B" => "__builtin_HEXAGON_V6_vrdelta_128B", + "llvm.hexagon.V6.vrmpybus" => "__builtin_HEXAGON_V6_vrmpybus", + "llvm.hexagon.V6.vrmpybus.128B" => "__builtin_HEXAGON_V6_vrmpybus_128B", + "llvm.hexagon.V6.vrmpybus.acc" => "__builtin_HEXAGON_V6_vrmpybus_acc", + "llvm.hexagon.V6.vrmpybus.acc.128B" => "__builtin_HEXAGON_V6_vrmpybus_acc_128B", + "llvm.hexagon.V6.vrmpybusi" => "__builtin_HEXAGON_V6_vrmpybusi", + "llvm.hexagon.V6.vrmpybusi.128B" => "__builtin_HEXAGON_V6_vrmpybusi_128B", + "llvm.hexagon.V6.vrmpybusi.acc" => "__builtin_HEXAGON_V6_vrmpybusi_acc", + "llvm.hexagon.V6.vrmpybusi.acc.128B" => "__builtin_HEXAGON_V6_vrmpybusi_acc_128B", + "llvm.hexagon.V6.vrmpybusv" => "__builtin_HEXAGON_V6_vrmpybusv", + "llvm.hexagon.V6.vrmpybusv.128B" => "__builtin_HEXAGON_V6_vrmpybusv_128B", + "llvm.hexagon.V6.vrmpybusv.acc" => "__builtin_HEXAGON_V6_vrmpybusv_acc", + "llvm.hexagon.V6.vrmpybusv.acc.128B" => "__builtin_HEXAGON_V6_vrmpybusv_acc_128B", + "llvm.hexagon.V6.vrmpybv" => "__builtin_HEXAGON_V6_vrmpybv", + "llvm.hexagon.V6.vrmpybv.128B" => "__builtin_HEXAGON_V6_vrmpybv_128B", + "llvm.hexagon.V6.vrmpybv.acc" => "__builtin_HEXAGON_V6_vrmpybv_acc", + "llvm.hexagon.V6.vrmpybv.acc.128B" => "__builtin_HEXAGON_V6_vrmpybv_acc_128B", + "llvm.hexagon.V6.vrmpyub" => "__builtin_HEXAGON_V6_vrmpyub", + "llvm.hexagon.V6.vrmpyub.128B" => "__builtin_HEXAGON_V6_vrmpyub_128B", + "llvm.hexagon.V6.vrmpyub.acc" => "__builtin_HEXAGON_V6_vrmpyub_acc", + "llvm.hexagon.V6.vrmpyub.acc.128B" => "__builtin_HEXAGON_V6_vrmpyub_acc_128B", + "llvm.hexagon.V6.vrmpyubi" => "__builtin_HEXAGON_V6_vrmpyubi", + "llvm.hexagon.V6.vrmpyubi.128B" => "__builtin_HEXAGON_V6_vrmpyubi_128B", + "llvm.hexagon.V6.vrmpyubi.acc" => "__builtin_HEXAGON_V6_vrmpyubi_acc", + "llvm.hexagon.V6.vrmpyubi.acc.128B" => "__builtin_HEXAGON_V6_vrmpyubi_acc_128B", + "llvm.hexagon.V6.vrmpyubv" => "__builtin_HEXAGON_V6_vrmpyubv", + "llvm.hexagon.V6.vrmpyubv.128B" => "__builtin_HEXAGON_V6_vrmpyubv_128B", + "llvm.hexagon.V6.vrmpyubv.acc" => "__builtin_HEXAGON_V6_vrmpyubv_acc", + "llvm.hexagon.V6.vrmpyubv.acc.128B" => "__builtin_HEXAGON_V6_vrmpyubv_acc_128B", + "llvm.hexagon.V6.vror" => "__builtin_HEXAGON_V6_vror", + "llvm.hexagon.V6.vror.128B" => "__builtin_HEXAGON_V6_vror_128B", + "llvm.hexagon.V6.vroundhb" => "__builtin_HEXAGON_V6_vroundhb", + "llvm.hexagon.V6.vroundhb.128B" => "__builtin_HEXAGON_V6_vroundhb_128B", + "llvm.hexagon.V6.vroundhub" => "__builtin_HEXAGON_V6_vroundhub", + "llvm.hexagon.V6.vroundhub.128B" => "__builtin_HEXAGON_V6_vroundhub_128B", + "llvm.hexagon.V6.vroundwh" => "__builtin_HEXAGON_V6_vroundwh", + "llvm.hexagon.V6.vroundwh.128B" => "__builtin_HEXAGON_V6_vroundwh_128B", + "llvm.hexagon.V6.vroundwuh" => "__builtin_HEXAGON_V6_vroundwuh", + "llvm.hexagon.V6.vroundwuh.128B" => "__builtin_HEXAGON_V6_vroundwuh_128B", + "llvm.hexagon.V6.vrsadubi" => "__builtin_HEXAGON_V6_vrsadubi", + "llvm.hexagon.V6.vrsadubi.128B" => "__builtin_HEXAGON_V6_vrsadubi_128B", + "llvm.hexagon.V6.vrsadubi.acc" => "__builtin_HEXAGON_V6_vrsadubi_acc", + "llvm.hexagon.V6.vrsadubi.acc.128B" => "__builtin_HEXAGON_V6_vrsadubi_acc_128B", + "llvm.hexagon.V6.vsathub" => "__builtin_HEXAGON_V6_vsathub", + "llvm.hexagon.V6.vsathub.128B" => "__builtin_HEXAGON_V6_vsathub_128B", + "llvm.hexagon.V6.vsatwh" => "__builtin_HEXAGON_V6_vsatwh", + "llvm.hexagon.V6.vsatwh.128B" => "__builtin_HEXAGON_V6_vsatwh_128B", + "llvm.hexagon.V6.vsb" => "__builtin_HEXAGON_V6_vsb", + "llvm.hexagon.V6.vsb.128B" => "__builtin_HEXAGON_V6_vsb_128B", + "llvm.hexagon.V6.vsh" => "__builtin_HEXAGON_V6_vsh", + "llvm.hexagon.V6.vsh.128B" => "__builtin_HEXAGON_V6_vsh_128B", + "llvm.hexagon.V6.vshufeh" => "__builtin_HEXAGON_V6_vshufeh", + "llvm.hexagon.V6.vshufeh.128B" => "__builtin_HEXAGON_V6_vshufeh_128B", + "llvm.hexagon.V6.vshuffb" => "__builtin_HEXAGON_V6_vshuffb", + "llvm.hexagon.V6.vshuffb.128B" => "__builtin_HEXAGON_V6_vshuffb_128B", + "llvm.hexagon.V6.vshuffeb" => "__builtin_HEXAGON_V6_vshuffeb", + "llvm.hexagon.V6.vshuffeb.128B" => "__builtin_HEXAGON_V6_vshuffeb_128B", + "llvm.hexagon.V6.vshuffh" => "__builtin_HEXAGON_V6_vshuffh", + "llvm.hexagon.V6.vshuffh.128B" => "__builtin_HEXAGON_V6_vshuffh_128B", + "llvm.hexagon.V6.vshuffob" => "__builtin_HEXAGON_V6_vshuffob", + "llvm.hexagon.V6.vshuffob.128B" => "__builtin_HEXAGON_V6_vshuffob_128B", + "llvm.hexagon.V6.vshuffvdd" => "__builtin_HEXAGON_V6_vshuffvdd", + "llvm.hexagon.V6.vshuffvdd.128B" => "__builtin_HEXAGON_V6_vshuffvdd_128B", + "llvm.hexagon.V6.vshufoeb" => "__builtin_HEXAGON_V6_vshufoeb", + "llvm.hexagon.V6.vshufoeb.128B" => "__builtin_HEXAGON_V6_vshufoeb_128B", + "llvm.hexagon.V6.vshufoeh" => "__builtin_HEXAGON_V6_vshufoeh", + "llvm.hexagon.V6.vshufoeh.128B" => "__builtin_HEXAGON_V6_vshufoeh_128B", + "llvm.hexagon.V6.vshufoh" => "__builtin_HEXAGON_V6_vshufoh", + "llvm.hexagon.V6.vshufoh.128B" => "__builtin_HEXAGON_V6_vshufoh_128B", + "llvm.hexagon.V6.vsubb" => "__builtin_HEXAGON_V6_vsubb", + "llvm.hexagon.V6.vsubb.128B" => "__builtin_HEXAGON_V6_vsubb_128B", + "llvm.hexagon.V6.vsubb.dv" => "__builtin_HEXAGON_V6_vsubb_dv", + "llvm.hexagon.V6.vsubb.dv.128B" => "__builtin_HEXAGON_V6_vsubb_dv_128B", + "llvm.hexagon.V6.vsubh" => "__builtin_HEXAGON_V6_vsubh", + "llvm.hexagon.V6.vsubh.128B" => "__builtin_HEXAGON_V6_vsubh_128B", + "llvm.hexagon.V6.vsubh.dv" => "__builtin_HEXAGON_V6_vsubh_dv", + "llvm.hexagon.V6.vsubh.dv.128B" => "__builtin_HEXAGON_V6_vsubh_dv_128B", + "llvm.hexagon.V6.vsubhsat" => "__builtin_HEXAGON_V6_vsubhsat", + "llvm.hexagon.V6.vsubhsat.128B" => "__builtin_HEXAGON_V6_vsubhsat_128B", + "llvm.hexagon.V6.vsubhsat.dv" => "__builtin_HEXAGON_V6_vsubhsat_dv", + "llvm.hexagon.V6.vsubhsat.dv.128B" => "__builtin_HEXAGON_V6_vsubhsat_dv_128B", + "llvm.hexagon.V6.vsubhw" => "__builtin_HEXAGON_V6_vsubhw", + "llvm.hexagon.V6.vsubhw.128B" => "__builtin_HEXAGON_V6_vsubhw_128B", + "llvm.hexagon.V6.vsububh" => "__builtin_HEXAGON_V6_vsububh", + "llvm.hexagon.V6.vsububh.128B" => "__builtin_HEXAGON_V6_vsububh_128B", + "llvm.hexagon.V6.vsububsat" => "__builtin_HEXAGON_V6_vsububsat", + "llvm.hexagon.V6.vsububsat.128B" => "__builtin_HEXAGON_V6_vsububsat_128B", + "llvm.hexagon.V6.vsububsat.dv" => "__builtin_HEXAGON_V6_vsububsat_dv", + "llvm.hexagon.V6.vsububsat.dv.128B" => "__builtin_HEXAGON_V6_vsububsat_dv_128B", + "llvm.hexagon.V6.vsubuhsat" => "__builtin_HEXAGON_V6_vsubuhsat", + "llvm.hexagon.V6.vsubuhsat.128B" => "__builtin_HEXAGON_V6_vsubuhsat_128B", + "llvm.hexagon.V6.vsubuhsat.dv" => "__builtin_HEXAGON_V6_vsubuhsat_dv", + "llvm.hexagon.V6.vsubuhsat.dv.128B" => "__builtin_HEXAGON_V6_vsubuhsat_dv_128B", + "llvm.hexagon.V6.vsubuhw" => "__builtin_HEXAGON_V6_vsubuhw", + "llvm.hexagon.V6.vsubuhw.128B" => "__builtin_HEXAGON_V6_vsubuhw_128B", + "llvm.hexagon.V6.vsubw" => "__builtin_HEXAGON_V6_vsubw", + "llvm.hexagon.V6.vsubw.128B" => "__builtin_HEXAGON_V6_vsubw_128B", + "llvm.hexagon.V6.vsubw.dv" => "__builtin_HEXAGON_V6_vsubw_dv", + "llvm.hexagon.V6.vsubw.dv.128B" => "__builtin_HEXAGON_V6_vsubw_dv_128B", + "llvm.hexagon.V6.vsubwsat" => "__builtin_HEXAGON_V6_vsubwsat", + "llvm.hexagon.V6.vsubwsat.128B" => "__builtin_HEXAGON_V6_vsubwsat_128B", + "llvm.hexagon.V6.vsubwsat.dv" => "__builtin_HEXAGON_V6_vsubwsat_dv", + "llvm.hexagon.V6.vsubwsat.dv.128B" => "__builtin_HEXAGON_V6_vsubwsat_dv_128B", + "llvm.hexagon.V6.vtmpyb" => "__builtin_HEXAGON_V6_vtmpyb", + "llvm.hexagon.V6.vtmpyb.128B" => "__builtin_HEXAGON_V6_vtmpyb_128B", + "llvm.hexagon.V6.vtmpyb.acc" => "__builtin_HEXAGON_V6_vtmpyb_acc", + "llvm.hexagon.V6.vtmpyb.acc.128B" => "__builtin_HEXAGON_V6_vtmpyb_acc_128B", + "llvm.hexagon.V6.vtmpybus" => "__builtin_HEXAGON_V6_vtmpybus", + "llvm.hexagon.V6.vtmpybus.128B" => "__builtin_HEXAGON_V6_vtmpybus_128B", + "llvm.hexagon.V6.vtmpybus.acc" => "__builtin_HEXAGON_V6_vtmpybus_acc", + "llvm.hexagon.V6.vtmpybus.acc.128B" => "__builtin_HEXAGON_V6_vtmpybus_acc_128B", + "llvm.hexagon.V6.vtmpyhb" => "__builtin_HEXAGON_V6_vtmpyhb", + "llvm.hexagon.V6.vtmpyhb.128B" => "__builtin_HEXAGON_V6_vtmpyhb_128B", + "llvm.hexagon.V6.vtmpyhb.acc" => "__builtin_HEXAGON_V6_vtmpyhb_acc", + "llvm.hexagon.V6.vtmpyhb.acc.128B" => "__builtin_HEXAGON_V6_vtmpyhb_acc_128B", + "llvm.hexagon.V6.vunpackb" => "__builtin_HEXAGON_V6_vunpackb", + "llvm.hexagon.V6.vunpackb.128B" => "__builtin_HEXAGON_V6_vunpackb_128B", + "llvm.hexagon.V6.vunpackh" => "__builtin_HEXAGON_V6_vunpackh", + "llvm.hexagon.V6.vunpackh.128B" => "__builtin_HEXAGON_V6_vunpackh_128B", + "llvm.hexagon.V6.vunpackob" => "__builtin_HEXAGON_V6_vunpackob", + "llvm.hexagon.V6.vunpackob.128B" => "__builtin_HEXAGON_V6_vunpackob_128B", + "llvm.hexagon.V6.vunpackoh" => "__builtin_HEXAGON_V6_vunpackoh", + "llvm.hexagon.V6.vunpackoh.128B" => "__builtin_HEXAGON_V6_vunpackoh_128B", + "llvm.hexagon.V6.vunpackub" => "__builtin_HEXAGON_V6_vunpackub", + "llvm.hexagon.V6.vunpackub.128B" => "__builtin_HEXAGON_V6_vunpackub_128B", + "llvm.hexagon.V6.vunpackuh" => "__builtin_HEXAGON_V6_vunpackuh", + "llvm.hexagon.V6.vunpackuh.128B" => "__builtin_HEXAGON_V6_vunpackuh_128B", + "llvm.hexagon.V6.vxor" => "__builtin_HEXAGON_V6_vxor", + "llvm.hexagon.V6.vxor.128B" => "__builtin_HEXAGON_V6_vxor_128B", + "llvm.hexagon.V6.vzb" => "__builtin_HEXAGON_V6_vzb", + "llvm.hexagon.V6.vzb.128B" => "__builtin_HEXAGON_V6_vzb_128B", + "llvm.hexagon.V6.vzh" => "__builtin_HEXAGON_V6_vzh", + "llvm.hexagon.V6.vzh.128B" => "__builtin_HEXAGON_V6_vzh_128B", + "llvm.hexagon.brev.ldb" => "__builtin_brev_ldb", + "llvm.hexagon.brev.ldd" => "__builtin_brev_ldd", + "llvm.hexagon.brev.ldh" => "__builtin_brev_ldh", + "llvm.hexagon.brev.ldub" => "__builtin_brev_ldub", + "llvm.hexagon.brev.lduh" => "__builtin_brev_lduh", + "llvm.hexagon.brev.ldw" => "__builtin_brev_ldw", + "llvm.hexagon.brev.stb" => "__builtin_brev_stb", + "llvm.hexagon.brev.std" => "__builtin_brev_std", + "llvm.hexagon.brev.sth" => "__builtin_brev_sth", + "llvm.hexagon.brev.sthhi" => "__builtin_brev_sthhi", + "llvm.hexagon.brev.stw" => "__builtin_brev_stw", + "llvm.hexagon.circ.ldb" => "__builtin_circ_ldb", + "llvm.hexagon.circ.ldd" => "__builtin_circ_ldd", + "llvm.hexagon.circ.ldh" => "__builtin_circ_ldh", + "llvm.hexagon.circ.ldub" => "__builtin_circ_ldub", + "llvm.hexagon.circ.lduh" => "__builtin_circ_lduh", + "llvm.hexagon.circ.ldw" => "__builtin_circ_ldw", + "llvm.hexagon.circ.stb" => "__builtin_circ_stb", + "llvm.hexagon.circ.std" => "__builtin_circ_std", + "llvm.hexagon.circ.sth" => "__builtin_circ_sth", + "llvm.hexagon.circ.sthhi" => "__builtin_circ_sthhi", + "llvm.hexagon.circ.stw" => "__builtin_circ_stw", + "llvm.hexagon.mm256i.vaddw" => "__builtin__mm256i_vaddw", + "llvm.hexagon.prefetch" => "__builtin_HEXAGON_prefetch", + // mips + "llvm.mips.absq.s.ph" => "__builtin_mips_absq_s_ph", + "llvm.mips.absq.s.qb" => "__builtin_mips_absq_s_qb", + "llvm.mips.absq.s.w" => "__builtin_mips_absq_s_w", + "llvm.mips.add.a.b" => "__builtin_msa_add_a_b", + "llvm.mips.add.a.d" => "__builtin_msa_add_a_d", + "llvm.mips.add.a.h" => "__builtin_msa_add_a_h", + "llvm.mips.add.a.w" => "__builtin_msa_add_a_w", + "llvm.mips.addq.ph" => "__builtin_mips_addq_ph", + "llvm.mips.addq.s.ph" => "__builtin_mips_addq_s_ph", + "llvm.mips.addq.s.w" => "__builtin_mips_addq_s_w", + "llvm.mips.addqh.ph" => "__builtin_mips_addqh_ph", + "llvm.mips.addqh.r.ph" => "__builtin_mips_addqh_r_ph", + "llvm.mips.addqh.r.w" => "__builtin_mips_addqh_r_w", + "llvm.mips.addqh.w" => "__builtin_mips_addqh_w", + "llvm.mips.adds.a.b" => "__builtin_msa_adds_a_b", + "llvm.mips.adds.a.d" => "__builtin_msa_adds_a_d", + "llvm.mips.adds.a.h" => "__builtin_msa_adds_a_h", + "llvm.mips.adds.a.w" => "__builtin_msa_adds_a_w", + "llvm.mips.adds.s.b" => "__builtin_msa_adds_s_b", + "llvm.mips.adds.s.d" => "__builtin_msa_adds_s_d", + "llvm.mips.adds.s.h" => "__builtin_msa_adds_s_h", + "llvm.mips.adds.s.w" => "__builtin_msa_adds_s_w", + "llvm.mips.adds.u.b" => "__builtin_msa_adds_u_b", + "llvm.mips.adds.u.d" => "__builtin_msa_adds_u_d", + "llvm.mips.adds.u.h" => "__builtin_msa_adds_u_h", + "llvm.mips.adds.u.w" => "__builtin_msa_adds_u_w", + "llvm.mips.addsc" => "__builtin_mips_addsc", + "llvm.mips.addu.ph" => "__builtin_mips_addu_ph", + "llvm.mips.addu.qb" => "__builtin_mips_addu_qb", + "llvm.mips.addu.s.ph" => "__builtin_mips_addu_s_ph", + "llvm.mips.addu.s.qb" => "__builtin_mips_addu_s_qb", + "llvm.mips.adduh.qb" => "__builtin_mips_adduh_qb", + "llvm.mips.adduh.r.qb" => "__builtin_mips_adduh_r_qb", + "llvm.mips.addv.b" => "__builtin_msa_addv_b", + "llvm.mips.addv.d" => "__builtin_msa_addv_d", + "llvm.mips.addv.h" => "__builtin_msa_addv_h", + "llvm.mips.addv.w" => "__builtin_msa_addv_w", + "llvm.mips.addvi.b" => "__builtin_msa_addvi_b", + "llvm.mips.addvi.d" => "__builtin_msa_addvi_d", + "llvm.mips.addvi.h" => "__builtin_msa_addvi_h", + "llvm.mips.addvi.w" => "__builtin_msa_addvi_w", + "llvm.mips.addwc" => "__builtin_mips_addwc", + "llvm.mips.and.v" => "__builtin_msa_and_v", + "llvm.mips.andi.b" => "__builtin_msa_andi_b", + "llvm.mips.append" => "__builtin_mips_append", + "llvm.mips.asub.s.b" => "__builtin_msa_asub_s_b", + "llvm.mips.asub.s.d" => "__builtin_msa_asub_s_d", + "llvm.mips.asub.s.h" => "__builtin_msa_asub_s_h", + "llvm.mips.asub.s.w" => "__builtin_msa_asub_s_w", + "llvm.mips.asub.u.b" => "__builtin_msa_asub_u_b", + "llvm.mips.asub.u.d" => "__builtin_msa_asub_u_d", + "llvm.mips.asub.u.h" => "__builtin_msa_asub_u_h", + "llvm.mips.asub.u.w" => "__builtin_msa_asub_u_w", + "llvm.mips.ave.s.b" => "__builtin_msa_ave_s_b", + "llvm.mips.ave.s.d" => "__builtin_msa_ave_s_d", + "llvm.mips.ave.s.h" => "__builtin_msa_ave_s_h", + "llvm.mips.ave.s.w" => "__builtin_msa_ave_s_w", + "llvm.mips.ave.u.b" => "__builtin_msa_ave_u_b", + "llvm.mips.ave.u.d" => "__builtin_msa_ave_u_d", + "llvm.mips.ave.u.h" => "__builtin_msa_ave_u_h", + "llvm.mips.ave.u.w" => "__builtin_msa_ave_u_w", + "llvm.mips.aver.s.b" => "__builtin_msa_aver_s_b", + "llvm.mips.aver.s.d" => "__builtin_msa_aver_s_d", + "llvm.mips.aver.s.h" => "__builtin_msa_aver_s_h", + "llvm.mips.aver.s.w" => "__builtin_msa_aver_s_w", + "llvm.mips.aver.u.b" => "__builtin_msa_aver_u_b", + "llvm.mips.aver.u.d" => "__builtin_msa_aver_u_d", + "llvm.mips.aver.u.h" => "__builtin_msa_aver_u_h", + "llvm.mips.aver.u.w" => "__builtin_msa_aver_u_w", + "llvm.mips.balign" => "__builtin_mips_balign", + "llvm.mips.bclr.b" => "__builtin_msa_bclr_b", + "llvm.mips.bclr.d" => "__builtin_msa_bclr_d", + "llvm.mips.bclr.h" => "__builtin_msa_bclr_h", + "llvm.mips.bclr.w" => "__builtin_msa_bclr_w", + "llvm.mips.bclri.b" => "__builtin_msa_bclri_b", + "llvm.mips.bclri.d" => "__builtin_msa_bclri_d", + "llvm.mips.bclri.h" => "__builtin_msa_bclri_h", + "llvm.mips.bclri.w" => "__builtin_msa_bclri_w", + "llvm.mips.binsl.b" => "__builtin_msa_binsl_b", + "llvm.mips.binsl.d" => "__builtin_msa_binsl_d", + "llvm.mips.binsl.h" => "__builtin_msa_binsl_h", + "llvm.mips.binsl.w" => "__builtin_msa_binsl_w", + "llvm.mips.binsli.b" => "__builtin_msa_binsli_b", + "llvm.mips.binsli.d" => "__builtin_msa_binsli_d", + "llvm.mips.binsli.h" => "__builtin_msa_binsli_h", + "llvm.mips.binsli.w" => "__builtin_msa_binsli_w", + "llvm.mips.binsr.b" => "__builtin_msa_binsr_b", + "llvm.mips.binsr.d" => "__builtin_msa_binsr_d", + "llvm.mips.binsr.h" => "__builtin_msa_binsr_h", + "llvm.mips.binsr.w" => "__builtin_msa_binsr_w", + "llvm.mips.binsri.b" => "__builtin_msa_binsri_b", + "llvm.mips.binsri.d" => "__builtin_msa_binsri_d", + "llvm.mips.binsri.h" => "__builtin_msa_binsri_h", + "llvm.mips.binsri.w" => "__builtin_msa_binsri_w", + "llvm.mips.bitrev" => "__builtin_mips_bitrev", + "llvm.mips.bmnz.v" => "__builtin_msa_bmnz_v", + "llvm.mips.bmnzi.b" => "__builtin_msa_bmnzi_b", + "llvm.mips.bmz.v" => "__builtin_msa_bmz_v", + "llvm.mips.bmzi.b" => "__builtin_msa_bmzi_b", + "llvm.mips.bneg.b" => "__builtin_msa_bneg_b", + "llvm.mips.bneg.d" => "__builtin_msa_bneg_d", + "llvm.mips.bneg.h" => "__builtin_msa_bneg_h", + "llvm.mips.bneg.w" => "__builtin_msa_bneg_w", + "llvm.mips.bnegi.b" => "__builtin_msa_bnegi_b", + "llvm.mips.bnegi.d" => "__builtin_msa_bnegi_d", + "llvm.mips.bnegi.h" => "__builtin_msa_bnegi_h", + "llvm.mips.bnegi.w" => "__builtin_msa_bnegi_w", + "llvm.mips.bnz.b" => "__builtin_msa_bnz_b", + "llvm.mips.bnz.d" => "__builtin_msa_bnz_d", + "llvm.mips.bnz.h" => "__builtin_msa_bnz_h", + "llvm.mips.bnz.v" => "__builtin_msa_bnz_v", + "llvm.mips.bnz.w" => "__builtin_msa_bnz_w", + "llvm.mips.bposge32" => "__builtin_mips_bposge32", + "llvm.mips.bsel.v" => "__builtin_msa_bsel_v", + "llvm.mips.bseli.b" => "__builtin_msa_bseli_b", + "llvm.mips.bset.b" => "__builtin_msa_bset_b", + "llvm.mips.bset.d" => "__builtin_msa_bset_d", + "llvm.mips.bset.h" => "__builtin_msa_bset_h", + "llvm.mips.bset.w" => "__builtin_msa_bset_w", + "llvm.mips.bseti.b" => "__builtin_msa_bseti_b", + "llvm.mips.bseti.d" => "__builtin_msa_bseti_d", + "llvm.mips.bseti.h" => "__builtin_msa_bseti_h", + "llvm.mips.bseti.w" => "__builtin_msa_bseti_w", + "llvm.mips.bz.b" => "__builtin_msa_bz_b", + "llvm.mips.bz.d" => "__builtin_msa_bz_d", + "llvm.mips.bz.h" => "__builtin_msa_bz_h", + "llvm.mips.bz.v" => "__builtin_msa_bz_v", + "llvm.mips.bz.w" => "__builtin_msa_bz_w", + "llvm.mips.ceq.b" => "__builtin_msa_ceq_b", + "llvm.mips.ceq.d" => "__builtin_msa_ceq_d", + "llvm.mips.ceq.h" => "__builtin_msa_ceq_h", + "llvm.mips.ceq.w" => "__builtin_msa_ceq_w", + "llvm.mips.ceqi.b" => "__builtin_msa_ceqi_b", + "llvm.mips.ceqi.d" => "__builtin_msa_ceqi_d", + "llvm.mips.ceqi.h" => "__builtin_msa_ceqi_h", + "llvm.mips.ceqi.w" => "__builtin_msa_ceqi_w", + "llvm.mips.cfcmsa" => "__builtin_msa_cfcmsa", + "llvm.mips.cle.s.b" => "__builtin_msa_cle_s_b", + "llvm.mips.cle.s.d" => "__builtin_msa_cle_s_d", + "llvm.mips.cle.s.h" => "__builtin_msa_cle_s_h", + "llvm.mips.cle.s.w" => "__builtin_msa_cle_s_w", + "llvm.mips.cle.u.b" => "__builtin_msa_cle_u_b", + "llvm.mips.cle.u.d" => "__builtin_msa_cle_u_d", + "llvm.mips.cle.u.h" => "__builtin_msa_cle_u_h", + "llvm.mips.cle.u.w" => "__builtin_msa_cle_u_w", + "llvm.mips.clei.s.b" => "__builtin_msa_clei_s_b", + "llvm.mips.clei.s.d" => "__builtin_msa_clei_s_d", + "llvm.mips.clei.s.h" => "__builtin_msa_clei_s_h", + "llvm.mips.clei.s.w" => "__builtin_msa_clei_s_w", + "llvm.mips.clei.u.b" => "__builtin_msa_clei_u_b", + "llvm.mips.clei.u.d" => "__builtin_msa_clei_u_d", + "llvm.mips.clei.u.h" => "__builtin_msa_clei_u_h", + "llvm.mips.clei.u.w" => "__builtin_msa_clei_u_w", + "llvm.mips.clt.s.b" => "__builtin_msa_clt_s_b", + "llvm.mips.clt.s.d" => "__builtin_msa_clt_s_d", + "llvm.mips.clt.s.h" => "__builtin_msa_clt_s_h", + "llvm.mips.clt.s.w" => "__builtin_msa_clt_s_w", + "llvm.mips.clt.u.b" => "__builtin_msa_clt_u_b", + "llvm.mips.clt.u.d" => "__builtin_msa_clt_u_d", + "llvm.mips.clt.u.h" => "__builtin_msa_clt_u_h", + "llvm.mips.clt.u.w" => "__builtin_msa_clt_u_w", + "llvm.mips.clti.s.b" => "__builtin_msa_clti_s_b", + "llvm.mips.clti.s.d" => "__builtin_msa_clti_s_d", + "llvm.mips.clti.s.h" => "__builtin_msa_clti_s_h", + "llvm.mips.clti.s.w" => "__builtin_msa_clti_s_w", + "llvm.mips.clti.u.b" => "__builtin_msa_clti_u_b", + "llvm.mips.clti.u.d" => "__builtin_msa_clti_u_d", + "llvm.mips.clti.u.h" => "__builtin_msa_clti_u_h", + "llvm.mips.clti.u.w" => "__builtin_msa_clti_u_w", + "llvm.mips.cmp.eq.ph" => "__builtin_mips_cmp_eq_ph", + "llvm.mips.cmp.le.ph" => "__builtin_mips_cmp_le_ph", + "llvm.mips.cmp.lt.ph" => "__builtin_mips_cmp_lt_ph", + "llvm.mips.cmpgdu.eq.qb" => "__builtin_mips_cmpgdu_eq_qb", + "llvm.mips.cmpgdu.le.qb" => "__builtin_mips_cmpgdu_le_qb", + "llvm.mips.cmpgdu.lt.qb" => "__builtin_mips_cmpgdu_lt_qb", + "llvm.mips.cmpgu.eq.qb" => "__builtin_mips_cmpgu_eq_qb", + "llvm.mips.cmpgu.le.qb" => "__builtin_mips_cmpgu_le_qb", + "llvm.mips.cmpgu.lt.qb" => "__builtin_mips_cmpgu_lt_qb", + "llvm.mips.cmpu.eq.qb" => "__builtin_mips_cmpu_eq_qb", + "llvm.mips.cmpu.le.qb" => "__builtin_mips_cmpu_le_qb", + "llvm.mips.cmpu.lt.qb" => "__builtin_mips_cmpu_lt_qb", + "llvm.mips.copy.s.b" => "__builtin_msa_copy_s_b", + "llvm.mips.copy.s.d" => "__builtin_msa_copy_s_d", + "llvm.mips.copy.s.h" => "__builtin_msa_copy_s_h", + "llvm.mips.copy.s.w" => "__builtin_msa_copy_s_w", + "llvm.mips.copy.u.b" => "__builtin_msa_copy_u_b", + "llvm.mips.copy.u.d" => "__builtin_msa_copy_u_d", + "llvm.mips.copy.u.h" => "__builtin_msa_copy_u_h", + "llvm.mips.copy.u.w" => "__builtin_msa_copy_u_w", + "llvm.mips.ctcmsa" => "__builtin_msa_ctcmsa", + "llvm.mips.div.s.b" => "__builtin_msa_div_s_b", + "llvm.mips.div.s.d" => "__builtin_msa_div_s_d", + "llvm.mips.div.s.h" => "__builtin_msa_div_s_h", + "llvm.mips.div.s.w" => "__builtin_msa_div_s_w", + "llvm.mips.div.u.b" => "__builtin_msa_div_u_b", + "llvm.mips.div.u.d" => "__builtin_msa_div_u_d", + "llvm.mips.div.u.h" => "__builtin_msa_div_u_h", + "llvm.mips.div.u.w" => "__builtin_msa_div_u_w", + "llvm.mips.dlsa" => "__builtin_mips_dlsa", + "llvm.mips.dotp.s.d" => "__builtin_msa_dotp_s_d", + "llvm.mips.dotp.s.h" => "__builtin_msa_dotp_s_h", + "llvm.mips.dotp.s.w" => "__builtin_msa_dotp_s_w", + "llvm.mips.dotp.u.d" => "__builtin_msa_dotp_u_d", + "llvm.mips.dotp.u.h" => "__builtin_msa_dotp_u_h", + "llvm.mips.dotp.u.w" => "__builtin_msa_dotp_u_w", + "llvm.mips.dpa.w.ph" => "__builtin_mips_dpa_w_ph", + "llvm.mips.dpadd.s.d" => "__builtin_msa_dpadd_s_d", + "llvm.mips.dpadd.s.h" => "__builtin_msa_dpadd_s_h", + "llvm.mips.dpadd.s.w" => "__builtin_msa_dpadd_s_w", + "llvm.mips.dpadd.u.d" => "__builtin_msa_dpadd_u_d", + "llvm.mips.dpadd.u.h" => "__builtin_msa_dpadd_u_h", + "llvm.mips.dpadd.u.w" => "__builtin_msa_dpadd_u_w", + "llvm.mips.dpaq.s.w.ph" => "__builtin_mips_dpaq_s_w_ph", + "llvm.mips.dpaq.sa.l.w" => "__builtin_mips_dpaq_sa_l_w", + "llvm.mips.dpaqx.s.w.ph" => "__builtin_mips_dpaqx_s_w_ph", + "llvm.mips.dpaqx.sa.w.ph" => "__builtin_mips_dpaqx_sa_w_ph", + "llvm.mips.dpau.h.qbl" => "__builtin_mips_dpau_h_qbl", + "llvm.mips.dpau.h.qbr" => "__builtin_mips_dpau_h_qbr", + "llvm.mips.dpax.w.ph" => "__builtin_mips_dpax_w_ph", + "llvm.mips.dps.w.ph" => "__builtin_mips_dps_w_ph", + "llvm.mips.dpsq.s.w.ph" => "__builtin_mips_dpsq_s_w_ph", + "llvm.mips.dpsq.sa.l.w" => "__builtin_mips_dpsq_sa_l_w", + "llvm.mips.dpsqx.s.w.ph" => "__builtin_mips_dpsqx_s_w_ph", + "llvm.mips.dpsqx.sa.w.ph" => "__builtin_mips_dpsqx_sa_w_ph", + "llvm.mips.dpsu.h.qbl" => "__builtin_mips_dpsu_h_qbl", + "llvm.mips.dpsu.h.qbr" => "__builtin_mips_dpsu_h_qbr", + "llvm.mips.dpsub.s.d" => "__builtin_msa_dpsub_s_d", + "llvm.mips.dpsub.s.h" => "__builtin_msa_dpsub_s_h", + "llvm.mips.dpsub.s.w" => "__builtin_msa_dpsub_s_w", + "llvm.mips.dpsub.u.d" => "__builtin_msa_dpsub_u_d", + "llvm.mips.dpsub.u.h" => "__builtin_msa_dpsub_u_h", + "llvm.mips.dpsub.u.w" => "__builtin_msa_dpsub_u_w", + "llvm.mips.dpsx.w.ph" => "__builtin_mips_dpsx_w_ph", + "llvm.mips.extp" => "__builtin_mips_extp", + "llvm.mips.extpdp" => "__builtin_mips_extpdp", + "llvm.mips.extr.r.w" => "__builtin_mips_extr_r_w", + "llvm.mips.extr.rs.w" => "__builtin_mips_extr_rs_w", + "llvm.mips.extr.s.h" => "__builtin_mips_extr_s_h", + "llvm.mips.extr.w" => "__builtin_mips_extr_w", + "llvm.mips.fadd.d" => "__builtin_msa_fadd_d", + "llvm.mips.fadd.w" => "__builtin_msa_fadd_w", + "llvm.mips.fcaf.d" => "__builtin_msa_fcaf_d", + "llvm.mips.fcaf.w" => "__builtin_msa_fcaf_w", + "llvm.mips.fceq.d" => "__builtin_msa_fceq_d", + "llvm.mips.fceq.w" => "__builtin_msa_fceq_w", + "llvm.mips.fclass.d" => "__builtin_msa_fclass_d", + "llvm.mips.fclass.w" => "__builtin_msa_fclass_w", + "llvm.mips.fcle.d" => "__builtin_msa_fcle_d", + "llvm.mips.fcle.w" => "__builtin_msa_fcle_w", + "llvm.mips.fclt.d" => "__builtin_msa_fclt_d", + "llvm.mips.fclt.w" => "__builtin_msa_fclt_w", + "llvm.mips.fcne.d" => "__builtin_msa_fcne_d", + "llvm.mips.fcne.w" => "__builtin_msa_fcne_w", + "llvm.mips.fcor.d" => "__builtin_msa_fcor_d", + "llvm.mips.fcor.w" => "__builtin_msa_fcor_w", + "llvm.mips.fcueq.d" => "__builtin_msa_fcueq_d", + "llvm.mips.fcueq.w" => "__builtin_msa_fcueq_w", + "llvm.mips.fcule.d" => "__builtin_msa_fcule_d", + "llvm.mips.fcule.w" => "__builtin_msa_fcule_w", + "llvm.mips.fcult.d" => "__builtin_msa_fcult_d", + "llvm.mips.fcult.w" => "__builtin_msa_fcult_w", + "llvm.mips.fcun.d" => "__builtin_msa_fcun_d", + "llvm.mips.fcun.w" => "__builtin_msa_fcun_w", + "llvm.mips.fcune.d" => "__builtin_msa_fcune_d", + "llvm.mips.fcune.w" => "__builtin_msa_fcune_w", + "llvm.mips.fdiv.d" => "__builtin_msa_fdiv_d", + "llvm.mips.fdiv.w" => "__builtin_msa_fdiv_w", + "llvm.mips.fexdo.h" => "__builtin_msa_fexdo_h", + "llvm.mips.fexdo.w" => "__builtin_msa_fexdo_w", + "llvm.mips.fexp2.d" => "__builtin_msa_fexp2_d", + "llvm.mips.fexp2.w" => "__builtin_msa_fexp2_w", + "llvm.mips.fexupl.d" => "__builtin_msa_fexupl_d", + "llvm.mips.fexupl.w" => "__builtin_msa_fexupl_w", + "llvm.mips.fexupr.d" => "__builtin_msa_fexupr_d", + "llvm.mips.fexupr.w" => "__builtin_msa_fexupr_w", + "llvm.mips.ffint.s.d" => "__builtin_msa_ffint_s_d", + "llvm.mips.ffint.s.w" => "__builtin_msa_ffint_s_w", + "llvm.mips.ffint.u.d" => "__builtin_msa_ffint_u_d", + "llvm.mips.ffint.u.w" => "__builtin_msa_ffint_u_w", + "llvm.mips.ffql.d" => "__builtin_msa_ffql_d", + "llvm.mips.ffql.w" => "__builtin_msa_ffql_w", + "llvm.mips.ffqr.d" => "__builtin_msa_ffqr_d", + "llvm.mips.ffqr.w" => "__builtin_msa_ffqr_w", + "llvm.mips.fill.b" => "__builtin_msa_fill_b", + "llvm.mips.fill.d" => "__builtin_msa_fill_d", + "llvm.mips.fill.h" => "__builtin_msa_fill_h", + "llvm.mips.fill.w" => "__builtin_msa_fill_w", + "llvm.mips.flog2.d" => "__builtin_msa_flog2_d", + "llvm.mips.flog2.w" => "__builtin_msa_flog2_w", + "llvm.mips.fmadd.d" => "__builtin_msa_fmadd_d", + "llvm.mips.fmadd.w" => "__builtin_msa_fmadd_w", + "llvm.mips.fmax.a.d" => "__builtin_msa_fmax_a_d", + "llvm.mips.fmax.a.w" => "__builtin_msa_fmax_a_w", + "llvm.mips.fmax.d" => "__builtin_msa_fmax_d", + "llvm.mips.fmax.w" => "__builtin_msa_fmax_w", + "llvm.mips.fmin.a.d" => "__builtin_msa_fmin_a_d", + "llvm.mips.fmin.a.w" => "__builtin_msa_fmin_a_w", + "llvm.mips.fmin.d" => "__builtin_msa_fmin_d", + "llvm.mips.fmin.w" => "__builtin_msa_fmin_w", + "llvm.mips.fmsub.d" => "__builtin_msa_fmsub_d", + "llvm.mips.fmsub.w" => "__builtin_msa_fmsub_w", + "llvm.mips.fmul.d" => "__builtin_msa_fmul_d", + "llvm.mips.fmul.w" => "__builtin_msa_fmul_w", + "llvm.mips.frcp.d" => "__builtin_msa_frcp_d", + "llvm.mips.frcp.w" => "__builtin_msa_frcp_w", + "llvm.mips.frint.d" => "__builtin_msa_frint_d", + "llvm.mips.frint.w" => "__builtin_msa_frint_w", + "llvm.mips.frsqrt.d" => "__builtin_msa_frsqrt_d", + "llvm.mips.frsqrt.w" => "__builtin_msa_frsqrt_w", + "llvm.mips.fsaf.d" => "__builtin_msa_fsaf_d", + "llvm.mips.fsaf.w" => "__builtin_msa_fsaf_w", + "llvm.mips.fseq.d" => "__builtin_msa_fseq_d", + "llvm.mips.fseq.w" => "__builtin_msa_fseq_w", + "llvm.mips.fsle.d" => "__builtin_msa_fsle_d", + "llvm.mips.fsle.w" => "__builtin_msa_fsle_w", + "llvm.mips.fslt.d" => "__builtin_msa_fslt_d", + "llvm.mips.fslt.w" => "__builtin_msa_fslt_w", + "llvm.mips.fsne.d" => "__builtin_msa_fsne_d", + "llvm.mips.fsne.w" => "__builtin_msa_fsne_w", + "llvm.mips.fsor.d" => "__builtin_msa_fsor_d", + "llvm.mips.fsor.w" => "__builtin_msa_fsor_w", + "llvm.mips.fsqrt.d" => "__builtin_msa_fsqrt_d", + "llvm.mips.fsqrt.w" => "__builtin_msa_fsqrt_w", + "llvm.mips.fsub.d" => "__builtin_msa_fsub_d", + "llvm.mips.fsub.w" => "__builtin_msa_fsub_w", + "llvm.mips.fsueq.d" => "__builtin_msa_fsueq_d", + "llvm.mips.fsueq.w" => "__builtin_msa_fsueq_w", + "llvm.mips.fsule.d" => "__builtin_msa_fsule_d", + "llvm.mips.fsule.w" => "__builtin_msa_fsule_w", + "llvm.mips.fsult.d" => "__builtin_msa_fsult_d", + "llvm.mips.fsult.w" => "__builtin_msa_fsult_w", + "llvm.mips.fsun.d" => "__builtin_msa_fsun_d", + "llvm.mips.fsun.w" => "__builtin_msa_fsun_w", + "llvm.mips.fsune.d" => "__builtin_msa_fsune_d", + "llvm.mips.fsune.w" => "__builtin_msa_fsune_w", + "llvm.mips.ftint.s.d" => "__builtin_msa_ftint_s_d", + "llvm.mips.ftint.s.w" => "__builtin_msa_ftint_s_w", + "llvm.mips.ftint.u.d" => "__builtin_msa_ftint_u_d", + "llvm.mips.ftint.u.w" => "__builtin_msa_ftint_u_w", + "llvm.mips.ftq.h" => "__builtin_msa_ftq_h", + "llvm.mips.ftq.w" => "__builtin_msa_ftq_w", + "llvm.mips.ftrunc.s.d" => "__builtin_msa_ftrunc_s_d", + "llvm.mips.ftrunc.s.w" => "__builtin_msa_ftrunc_s_w", + "llvm.mips.ftrunc.u.d" => "__builtin_msa_ftrunc_u_d", + "llvm.mips.ftrunc.u.w" => "__builtin_msa_ftrunc_u_w", + "llvm.mips.hadd.s.d" => "__builtin_msa_hadd_s_d", + "llvm.mips.hadd.s.h" => "__builtin_msa_hadd_s_h", + "llvm.mips.hadd.s.w" => "__builtin_msa_hadd_s_w", + "llvm.mips.hadd.u.d" => "__builtin_msa_hadd_u_d", + "llvm.mips.hadd.u.h" => "__builtin_msa_hadd_u_h", + "llvm.mips.hadd.u.w" => "__builtin_msa_hadd_u_w", + "llvm.mips.hsub.s.d" => "__builtin_msa_hsub_s_d", + "llvm.mips.hsub.s.h" => "__builtin_msa_hsub_s_h", + "llvm.mips.hsub.s.w" => "__builtin_msa_hsub_s_w", + "llvm.mips.hsub.u.d" => "__builtin_msa_hsub_u_d", + "llvm.mips.hsub.u.h" => "__builtin_msa_hsub_u_h", + "llvm.mips.hsub.u.w" => "__builtin_msa_hsub_u_w", + "llvm.mips.ilvev.b" => "__builtin_msa_ilvev_b", + "llvm.mips.ilvev.d" => "__builtin_msa_ilvev_d", + "llvm.mips.ilvev.h" => "__builtin_msa_ilvev_h", + "llvm.mips.ilvev.w" => "__builtin_msa_ilvev_w", + "llvm.mips.ilvl.b" => "__builtin_msa_ilvl_b", + "llvm.mips.ilvl.d" => "__builtin_msa_ilvl_d", + "llvm.mips.ilvl.h" => "__builtin_msa_ilvl_h", + "llvm.mips.ilvl.w" => "__builtin_msa_ilvl_w", + "llvm.mips.ilvod.b" => "__builtin_msa_ilvod_b", + "llvm.mips.ilvod.d" => "__builtin_msa_ilvod_d", + "llvm.mips.ilvod.h" => "__builtin_msa_ilvod_h", + "llvm.mips.ilvod.w" => "__builtin_msa_ilvod_w", + "llvm.mips.ilvr.b" => "__builtin_msa_ilvr_b", + "llvm.mips.ilvr.d" => "__builtin_msa_ilvr_d", + "llvm.mips.ilvr.h" => "__builtin_msa_ilvr_h", + "llvm.mips.ilvr.w" => "__builtin_msa_ilvr_w", + "llvm.mips.insert.b" => "__builtin_msa_insert_b", + "llvm.mips.insert.d" => "__builtin_msa_insert_d", + "llvm.mips.insert.h" => "__builtin_msa_insert_h", + "llvm.mips.insert.w" => "__builtin_msa_insert_w", + "llvm.mips.insv" => "__builtin_mips_insv", + "llvm.mips.insve.b" => "__builtin_msa_insve_b", + "llvm.mips.insve.d" => "__builtin_msa_insve_d", + "llvm.mips.insve.h" => "__builtin_msa_insve_h", + "llvm.mips.insve.w" => "__builtin_msa_insve_w", + "llvm.mips.lbux" => "__builtin_mips_lbux", + "llvm.mips.ld.b" => "__builtin_msa_ld_b", + "llvm.mips.ld.d" => "__builtin_msa_ld_d", + "llvm.mips.ld.h" => "__builtin_msa_ld_h", + "llvm.mips.ld.w" => "__builtin_msa_ld_w", + "llvm.mips.ldi.b" => "__builtin_msa_ldi_b", + "llvm.mips.ldi.d" => "__builtin_msa_ldi_d", + "llvm.mips.ldi.h" => "__builtin_msa_ldi_h", + "llvm.mips.ldi.w" => "__builtin_msa_ldi_w", + "llvm.mips.ldr.d" => "__builtin_msa_ldr_d", + "llvm.mips.ldr.w" => "__builtin_msa_ldr_w", + "llvm.mips.lhx" => "__builtin_mips_lhx", + "llvm.mips.lsa" => "__builtin_mips_lsa", + "llvm.mips.lwx" => "__builtin_mips_lwx", + "llvm.mips.madd" => "__builtin_mips_madd", + "llvm.mips.madd.q.h" => "__builtin_msa_madd_q_h", + "llvm.mips.madd.q.w" => "__builtin_msa_madd_q_w", + "llvm.mips.maddr.q.h" => "__builtin_msa_maddr_q_h", + "llvm.mips.maddr.q.w" => "__builtin_msa_maddr_q_w", + "llvm.mips.maddu" => "__builtin_mips_maddu", + "llvm.mips.maddv.b" => "__builtin_msa_maddv_b", + "llvm.mips.maddv.d" => "__builtin_msa_maddv_d", + "llvm.mips.maddv.h" => "__builtin_msa_maddv_h", + "llvm.mips.maddv.w" => "__builtin_msa_maddv_w", + "llvm.mips.maq.s.w.phl" => "__builtin_mips_maq_s_w_phl", + "llvm.mips.maq.s.w.phr" => "__builtin_mips_maq_s_w_phr", + "llvm.mips.maq.sa.w.phl" => "__builtin_mips_maq_sa_w_phl", + "llvm.mips.maq.sa.w.phr" => "__builtin_mips_maq_sa_w_phr", + "llvm.mips.max.a.b" => "__builtin_msa_max_a_b", + "llvm.mips.max.a.d" => "__builtin_msa_max_a_d", + "llvm.mips.max.a.h" => "__builtin_msa_max_a_h", + "llvm.mips.max.a.w" => "__builtin_msa_max_a_w", + "llvm.mips.max.s.b" => "__builtin_msa_max_s_b", + "llvm.mips.max.s.d" => "__builtin_msa_max_s_d", + "llvm.mips.max.s.h" => "__builtin_msa_max_s_h", + "llvm.mips.max.s.w" => "__builtin_msa_max_s_w", + "llvm.mips.max.u.b" => "__builtin_msa_max_u_b", + "llvm.mips.max.u.d" => "__builtin_msa_max_u_d", + "llvm.mips.max.u.h" => "__builtin_msa_max_u_h", + "llvm.mips.max.u.w" => "__builtin_msa_max_u_w", + "llvm.mips.maxi.s.b" => "__builtin_msa_maxi_s_b", + "llvm.mips.maxi.s.d" => "__builtin_msa_maxi_s_d", + "llvm.mips.maxi.s.h" => "__builtin_msa_maxi_s_h", + "llvm.mips.maxi.s.w" => "__builtin_msa_maxi_s_w", + "llvm.mips.maxi.u.b" => "__builtin_msa_maxi_u_b", + "llvm.mips.maxi.u.d" => "__builtin_msa_maxi_u_d", + "llvm.mips.maxi.u.h" => "__builtin_msa_maxi_u_h", + "llvm.mips.maxi.u.w" => "__builtin_msa_maxi_u_w", + "llvm.mips.min.a.b" => "__builtin_msa_min_a_b", + "llvm.mips.min.a.d" => "__builtin_msa_min_a_d", + "llvm.mips.min.a.h" => "__builtin_msa_min_a_h", + "llvm.mips.min.a.w" => "__builtin_msa_min_a_w", + "llvm.mips.min.s.b" => "__builtin_msa_min_s_b", + "llvm.mips.min.s.d" => "__builtin_msa_min_s_d", + "llvm.mips.min.s.h" => "__builtin_msa_min_s_h", + "llvm.mips.min.s.w" => "__builtin_msa_min_s_w", + "llvm.mips.min.u.b" => "__builtin_msa_min_u_b", + "llvm.mips.min.u.d" => "__builtin_msa_min_u_d", + "llvm.mips.min.u.h" => "__builtin_msa_min_u_h", + "llvm.mips.min.u.w" => "__builtin_msa_min_u_w", + "llvm.mips.mini.s.b" => "__builtin_msa_mini_s_b", + "llvm.mips.mini.s.d" => "__builtin_msa_mini_s_d", + "llvm.mips.mini.s.h" => "__builtin_msa_mini_s_h", + "llvm.mips.mini.s.w" => "__builtin_msa_mini_s_w", + "llvm.mips.mini.u.b" => "__builtin_msa_mini_u_b", + "llvm.mips.mini.u.d" => "__builtin_msa_mini_u_d", + "llvm.mips.mini.u.h" => "__builtin_msa_mini_u_h", + "llvm.mips.mini.u.w" => "__builtin_msa_mini_u_w", + "llvm.mips.mod.s.b" => "__builtin_msa_mod_s_b", + "llvm.mips.mod.s.d" => "__builtin_msa_mod_s_d", + "llvm.mips.mod.s.h" => "__builtin_msa_mod_s_h", + "llvm.mips.mod.s.w" => "__builtin_msa_mod_s_w", + "llvm.mips.mod.u.b" => "__builtin_msa_mod_u_b", + "llvm.mips.mod.u.d" => "__builtin_msa_mod_u_d", + "llvm.mips.mod.u.h" => "__builtin_msa_mod_u_h", + "llvm.mips.mod.u.w" => "__builtin_msa_mod_u_w", + "llvm.mips.modsub" => "__builtin_mips_modsub", + "llvm.mips.move.v" => "__builtin_msa_move_v", + "llvm.mips.msub" => "__builtin_mips_msub", + "llvm.mips.msub.q.h" => "__builtin_msa_msub_q_h", + "llvm.mips.msub.q.w" => "__builtin_msa_msub_q_w", + "llvm.mips.msubr.q.h" => "__builtin_msa_msubr_q_h", + "llvm.mips.msubr.q.w" => "__builtin_msa_msubr_q_w", + "llvm.mips.msubu" => "__builtin_mips_msubu", + "llvm.mips.msubv.b" => "__builtin_msa_msubv_b", + "llvm.mips.msubv.d" => "__builtin_msa_msubv_d", + "llvm.mips.msubv.h" => "__builtin_msa_msubv_h", + "llvm.mips.msubv.w" => "__builtin_msa_msubv_w", + "llvm.mips.mthlip" => "__builtin_mips_mthlip", + "llvm.mips.mul.ph" => "__builtin_mips_mul_ph", + "llvm.mips.mul.q.h" => "__builtin_msa_mul_q_h", + "llvm.mips.mul.q.w" => "__builtin_msa_mul_q_w", + "llvm.mips.mul.s.ph" => "__builtin_mips_mul_s_ph", + "llvm.mips.muleq.s.w.phl" => "__builtin_mips_muleq_s_w_phl", + "llvm.mips.muleq.s.w.phr" => "__builtin_mips_muleq_s_w_phr", + "llvm.mips.muleu.s.ph.qbl" => "__builtin_mips_muleu_s_ph_qbl", + "llvm.mips.muleu.s.ph.qbr" => "__builtin_mips_muleu_s_ph_qbr", + "llvm.mips.mulq.rs.ph" => "__builtin_mips_mulq_rs_ph", + "llvm.mips.mulq.rs.w" => "__builtin_mips_mulq_rs_w", + "llvm.mips.mulq.s.ph" => "__builtin_mips_mulq_s_ph", + "llvm.mips.mulq.s.w" => "__builtin_mips_mulq_s_w", + "llvm.mips.mulr.q.h" => "__builtin_msa_mulr_q_h", + "llvm.mips.mulr.q.w" => "__builtin_msa_mulr_q_w", + "llvm.mips.mulsa.w.ph" => "__builtin_mips_mulsa_w_ph", + "llvm.mips.mulsaq.s.w.ph" => "__builtin_mips_mulsaq_s_w_ph", + "llvm.mips.mult" => "__builtin_mips_mult", + "llvm.mips.multu" => "__builtin_mips_multu", + "llvm.mips.mulv.b" => "__builtin_msa_mulv_b", + "llvm.mips.mulv.d" => "__builtin_msa_mulv_d", + "llvm.mips.mulv.h" => "__builtin_msa_mulv_h", + "llvm.mips.mulv.w" => "__builtin_msa_mulv_w", + "llvm.mips.nloc.b" => "__builtin_msa_nloc_b", + "llvm.mips.nloc.d" => "__builtin_msa_nloc_d", + "llvm.mips.nloc.h" => "__builtin_msa_nloc_h", + "llvm.mips.nloc.w" => "__builtin_msa_nloc_w", + "llvm.mips.nlzc.b" => "__builtin_msa_nlzc_b", + "llvm.mips.nlzc.d" => "__builtin_msa_nlzc_d", + "llvm.mips.nlzc.h" => "__builtin_msa_nlzc_h", + "llvm.mips.nlzc.w" => "__builtin_msa_nlzc_w", + "llvm.mips.nor.v" => "__builtin_msa_nor_v", + "llvm.mips.nori.b" => "__builtin_msa_nori_b", + "llvm.mips.or.v" => "__builtin_msa_or_v", + "llvm.mips.ori.b" => "__builtin_msa_ori_b", + "llvm.mips.packrl.ph" => "__builtin_mips_packrl_ph", + "llvm.mips.pckev.b" => "__builtin_msa_pckev_b", + "llvm.mips.pckev.d" => "__builtin_msa_pckev_d", + "llvm.mips.pckev.h" => "__builtin_msa_pckev_h", + "llvm.mips.pckev.w" => "__builtin_msa_pckev_w", + "llvm.mips.pckod.b" => "__builtin_msa_pckod_b", + "llvm.mips.pckod.d" => "__builtin_msa_pckod_d", + "llvm.mips.pckod.h" => "__builtin_msa_pckod_h", + "llvm.mips.pckod.w" => "__builtin_msa_pckod_w", + "llvm.mips.pcnt.b" => "__builtin_msa_pcnt_b", + "llvm.mips.pcnt.d" => "__builtin_msa_pcnt_d", + "llvm.mips.pcnt.h" => "__builtin_msa_pcnt_h", + "llvm.mips.pcnt.w" => "__builtin_msa_pcnt_w", + "llvm.mips.pick.ph" => "__builtin_mips_pick_ph", + "llvm.mips.pick.qb" => "__builtin_mips_pick_qb", + "llvm.mips.preceq.w.phl" => "__builtin_mips_preceq_w_phl", + "llvm.mips.preceq.w.phr" => "__builtin_mips_preceq_w_phr", + "llvm.mips.precequ.ph.qbl" => "__builtin_mips_precequ_ph_qbl", + "llvm.mips.precequ.ph.qbla" => "__builtin_mips_precequ_ph_qbla", + "llvm.mips.precequ.ph.qbr" => "__builtin_mips_precequ_ph_qbr", + "llvm.mips.precequ.ph.qbra" => "__builtin_mips_precequ_ph_qbra", + "llvm.mips.preceu.ph.qbl" => "__builtin_mips_preceu_ph_qbl", + "llvm.mips.preceu.ph.qbla" => "__builtin_mips_preceu_ph_qbla", + "llvm.mips.preceu.ph.qbr" => "__builtin_mips_preceu_ph_qbr", + "llvm.mips.preceu.ph.qbra" => "__builtin_mips_preceu_ph_qbra", + "llvm.mips.precr.qb.ph" => "__builtin_mips_precr_qb_ph", + "llvm.mips.precr.sra.ph.w" => "__builtin_mips_precr_sra_ph_w", + "llvm.mips.precr.sra.r.ph.w" => "__builtin_mips_precr_sra_r_ph_w", + "llvm.mips.precrq.ph.w" => "__builtin_mips_precrq_ph_w", + "llvm.mips.precrq.qb.ph" => "__builtin_mips_precrq_qb_ph", + "llvm.mips.precrq.rs.ph.w" => "__builtin_mips_precrq_rs_ph_w", + "llvm.mips.precrqu.s.qb.ph" => "__builtin_mips_precrqu_s_qb_ph", + "llvm.mips.prepend" => "__builtin_mips_prepend", + "llvm.mips.raddu.w.qb" => "__builtin_mips_raddu_w_qb", + "llvm.mips.rddsp" => "__builtin_mips_rddsp", + "llvm.mips.repl.ph" => "__builtin_mips_repl_ph", + "llvm.mips.repl.qb" => "__builtin_mips_repl_qb", + "llvm.mips.sat.s.b" => "__builtin_msa_sat_s_b", + "llvm.mips.sat.s.d" => "__builtin_msa_sat_s_d", + "llvm.mips.sat.s.h" => "__builtin_msa_sat_s_h", + "llvm.mips.sat.s.w" => "__builtin_msa_sat_s_w", + "llvm.mips.sat.u.b" => "__builtin_msa_sat_u_b", + "llvm.mips.sat.u.d" => "__builtin_msa_sat_u_d", + "llvm.mips.sat.u.h" => "__builtin_msa_sat_u_h", + "llvm.mips.sat.u.w" => "__builtin_msa_sat_u_w", + "llvm.mips.shf.b" => "__builtin_msa_shf_b", + "llvm.mips.shf.h" => "__builtin_msa_shf_h", + "llvm.mips.shf.w" => "__builtin_msa_shf_w", + "llvm.mips.shilo" => "__builtin_mips_shilo", + "llvm.mips.shll.ph" => "__builtin_mips_shll_ph", + "llvm.mips.shll.qb" => "__builtin_mips_shll_qb", + "llvm.mips.shll.s.ph" => "__builtin_mips_shll_s_ph", + "llvm.mips.shll.s.w" => "__builtin_mips_shll_s_w", + "llvm.mips.shra.ph" => "__builtin_mips_shra_ph", + "llvm.mips.shra.qb" => "__builtin_mips_shra_qb", + "llvm.mips.shra.r.ph" => "__builtin_mips_shra_r_ph", + "llvm.mips.shra.r.qb" => "__builtin_mips_shra_r_qb", + "llvm.mips.shra.r.w" => "__builtin_mips_shra_r_w", + "llvm.mips.shrl.ph" => "__builtin_mips_shrl_ph", + "llvm.mips.shrl.qb" => "__builtin_mips_shrl_qb", + "llvm.mips.sld.b" => "__builtin_msa_sld_b", + "llvm.mips.sld.d" => "__builtin_msa_sld_d", + "llvm.mips.sld.h" => "__builtin_msa_sld_h", + "llvm.mips.sld.w" => "__builtin_msa_sld_w", + "llvm.mips.sldi.b" => "__builtin_msa_sldi_b", + "llvm.mips.sldi.d" => "__builtin_msa_sldi_d", + "llvm.mips.sldi.h" => "__builtin_msa_sldi_h", + "llvm.mips.sldi.w" => "__builtin_msa_sldi_w", + "llvm.mips.sll.b" => "__builtin_msa_sll_b", + "llvm.mips.sll.d" => "__builtin_msa_sll_d", + "llvm.mips.sll.h" => "__builtin_msa_sll_h", + "llvm.mips.sll.w" => "__builtin_msa_sll_w", + "llvm.mips.slli.b" => "__builtin_msa_slli_b", + "llvm.mips.slli.d" => "__builtin_msa_slli_d", + "llvm.mips.slli.h" => "__builtin_msa_slli_h", + "llvm.mips.slli.w" => "__builtin_msa_slli_w", + "llvm.mips.splat.b" => "__builtin_msa_splat_b", + "llvm.mips.splat.d" => "__builtin_msa_splat_d", + "llvm.mips.splat.h" => "__builtin_msa_splat_h", + "llvm.mips.splat.w" => "__builtin_msa_splat_w", + "llvm.mips.splati.b" => "__builtin_msa_splati_b", + "llvm.mips.splati.d" => "__builtin_msa_splati_d", + "llvm.mips.splati.h" => "__builtin_msa_splati_h", + "llvm.mips.splati.w" => "__builtin_msa_splati_w", + "llvm.mips.sra.b" => "__builtin_msa_sra_b", + "llvm.mips.sra.d" => "__builtin_msa_sra_d", + "llvm.mips.sra.h" => "__builtin_msa_sra_h", + "llvm.mips.sra.w" => "__builtin_msa_sra_w", + "llvm.mips.srai.b" => "__builtin_msa_srai_b", + "llvm.mips.srai.d" => "__builtin_msa_srai_d", + "llvm.mips.srai.h" => "__builtin_msa_srai_h", + "llvm.mips.srai.w" => "__builtin_msa_srai_w", + "llvm.mips.srar.b" => "__builtin_msa_srar_b", + "llvm.mips.srar.d" => "__builtin_msa_srar_d", + "llvm.mips.srar.h" => "__builtin_msa_srar_h", + "llvm.mips.srar.w" => "__builtin_msa_srar_w", + "llvm.mips.srari.b" => "__builtin_msa_srari_b", + "llvm.mips.srari.d" => "__builtin_msa_srari_d", + "llvm.mips.srari.h" => "__builtin_msa_srari_h", + "llvm.mips.srari.w" => "__builtin_msa_srari_w", + "llvm.mips.srl.b" => "__builtin_msa_srl_b", + "llvm.mips.srl.d" => "__builtin_msa_srl_d", + "llvm.mips.srl.h" => "__builtin_msa_srl_h", + "llvm.mips.srl.w" => "__builtin_msa_srl_w", + "llvm.mips.srli.b" => "__builtin_msa_srli_b", + "llvm.mips.srli.d" => "__builtin_msa_srli_d", + "llvm.mips.srli.h" => "__builtin_msa_srli_h", + "llvm.mips.srli.w" => "__builtin_msa_srli_w", + "llvm.mips.srlr.b" => "__builtin_msa_srlr_b", + "llvm.mips.srlr.d" => "__builtin_msa_srlr_d", + "llvm.mips.srlr.h" => "__builtin_msa_srlr_h", + "llvm.mips.srlr.w" => "__builtin_msa_srlr_w", + "llvm.mips.srlri.b" => "__builtin_msa_srlri_b", + "llvm.mips.srlri.d" => "__builtin_msa_srlri_d", + "llvm.mips.srlri.h" => "__builtin_msa_srlri_h", + "llvm.mips.srlri.w" => "__builtin_msa_srlri_w", + "llvm.mips.st.b" => "__builtin_msa_st_b", + "llvm.mips.st.d" => "__builtin_msa_st_d", + "llvm.mips.st.h" => "__builtin_msa_st_h", + "llvm.mips.st.w" => "__builtin_msa_st_w", + "llvm.mips.str.d" => "__builtin_msa_str_d", + "llvm.mips.str.w" => "__builtin_msa_str_w", + "llvm.mips.subq.ph" => "__builtin_mips_subq_ph", + "llvm.mips.subq.s.ph" => "__builtin_mips_subq_s_ph", + "llvm.mips.subq.s.w" => "__builtin_mips_subq_s_w", + "llvm.mips.subqh.ph" => "__builtin_mips_subqh_ph", + "llvm.mips.subqh.r.ph" => "__builtin_mips_subqh_r_ph", + "llvm.mips.subqh.r.w" => "__builtin_mips_subqh_r_w", + "llvm.mips.subqh.w" => "__builtin_mips_subqh_w", + "llvm.mips.subs.s.b" => "__builtin_msa_subs_s_b", + "llvm.mips.subs.s.d" => "__builtin_msa_subs_s_d", + "llvm.mips.subs.s.h" => "__builtin_msa_subs_s_h", + "llvm.mips.subs.s.w" => "__builtin_msa_subs_s_w", + "llvm.mips.subs.u.b" => "__builtin_msa_subs_u_b", + "llvm.mips.subs.u.d" => "__builtin_msa_subs_u_d", + "llvm.mips.subs.u.h" => "__builtin_msa_subs_u_h", + "llvm.mips.subs.u.w" => "__builtin_msa_subs_u_w", + "llvm.mips.subsus.u.b" => "__builtin_msa_subsus_u_b", + "llvm.mips.subsus.u.d" => "__builtin_msa_subsus_u_d", + "llvm.mips.subsus.u.h" => "__builtin_msa_subsus_u_h", + "llvm.mips.subsus.u.w" => "__builtin_msa_subsus_u_w", + "llvm.mips.subsuu.s.b" => "__builtin_msa_subsuu_s_b", + "llvm.mips.subsuu.s.d" => "__builtin_msa_subsuu_s_d", + "llvm.mips.subsuu.s.h" => "__builtin_msa_subsuu_s_h", + "llvm.mips.subsuu.s.w" => "__builtin_msa_subsuu_s_w", + "llvm.mips.subu.ph" => "__builtin_mips_subu_ph", + "llvm.mips.subu.qb" => "__builtin_mips_subu_qb", + "llvm.mips.subu.s.ph" => "__builtin_mips_subu_s_ph", + "llvm.mips.subu.s.qb" => "__builtin_mips_subu_s_qb", + "llvm.mips.subuh.qb" => "__builtin_mips_subuh_qb", + "llvm.mips.subuh.r.qb" => "__builtin_mips_subuh_r_qb", + "llvm.mips.subv.b" => "__builtin_msa_subv_b", + "llvm.mips.subv.d" => "__builtin_msa_subv_d", + "llvm.mips.subv.h" => "__builtin_msa_subv_h", + "llvm.mips.subv.w" => "__builtin_msa_subv_w", + "llvm.mips.subvi.b" => "__builtin_msa_subvi_b", + "llvm.mips.subvi.d" => "__builtin_msa_subvi_d", + "llvm.mips.subvi.h" => "__builtin_msa_subvi_h", + "llvm.mips.subvi.w" => "__builtin_msa_subvi_w", + "llvm.mips.vshf.b" => "__builtin_msa_vshf_b", + "llvm.mips.vshf.d" => "__builtin_msa_vshf_d", + "llvm.mips.vshf.h" => "__builtin_msa_vshf_h", + "llvm.mips.vshf.w" => "__builtin_msa_vshf_w", + "llvm.mips.wrdsp" => "__builtin_mips_wrdsp", + "llvm.mips.xor.v" => "__builtin_msa_xor_v", + "llvm.mips.xori.b" => "__builtin_msa_xori_b", + // nvvm + "llvm.nvvm.abs.i" => "__nvvm_abs_i", + "llvm.nvvm.abs.ll" => "__nvvm_abs_ll", + "llvm.nvvm.add.rm.d" => "__nvvm_add_rm_d", + "llvm.nvvm.add.rm.f" => "__nvvm_add_rm_f", + "llvm.nvvm.add.rm.ftz.f" => "__nvvm_add_rm_ftz_f", + "llvm.nvvm.add.rn.d" => "__nvvm_add_rn_d", + "llvm.nvvm.add.rn.f" => "__nvvm_add_rn_f", + "llvm.nvvm.add.rn.ftz.f" => "__nvvm_add_rn_ftz_f", + "llvm.nvvm.add.rp.d" => "__nvvm_add_rp_d", + "llvm.nvvm.add.rp.f" => "__nvvm_add_rp_f", + "llvm.nvvm.add.rp.ftz.f" => "__nvvm_add_rp_ftz_f", + "llvm.nvvm.add.rz.d" => "__nvvm_add_rz_d", + "llvm.nvvm.add.rz.f" => "__nvvm_add_rz_f", + "llvm.nvvm.add.rz.ftz.f" => "__nvvm_add_rz_ftz_f", + "llvm.nvvm.bar.sync" => "__nvvm_bar_sync", + "llvm.nvvm.barrier0" => "__nvvm_bar0", + // [DUPLICATE]: "llvm.nvvm.barrier0" => "__syncthreads", + "llvm.nvvm.barrier0.and" => "__nvvm_bar0_and", + "llvm.nvvm.barrier0.or" => "__nvvm_bar0_or", + "llvm.nvvm.barrier0.popc" => "__nvvm_bar0_popc", + "llvm.nvvm.bitcast.d2ll" => "__nvvm_bitcast_d2ll", + "llvm.nvvm.bitcast.f2i" => "__nvvm_bitcast_f2i", + "llvm.nvvm.bitcast.i2f" => "__nvvm_bitcast_i2f", + "llvm.nvvm.bitcast.ll2d" => "__nvvm_bitcast_ll2d", + "llvm.nvvm.brev32" => "__nvvm_brev32", + "llvm.nvvm.brev64" => "__nvvm_brev64", + "llvm.nvvm.ceil.d" => "__nvvm_ceil_d", + "llvm.nvvm.ceil.f" => "__nvvm_ceil_f", + "llvm.nvvm.ceil.ftz.f" => "__nvvm_ceil_ftz_f", + "llvm.nvvm.clz.i" => "__nvvm_clz_i", + "llvm.nvvm.clz.ll" => "__nvvm_clz_ll", + "llvm.nvvm.cos.approx.f" => "__nvvm_cos_approx_f", + "llvm.nvvm.cos.approx.ftz.f" => "__nvvm_cos_approx_ftz_f", + "llvm.nvvm.d2f.rm" => "__nvvm_d2f_rm", + "llvm.nvvm.d2f.rm.ftz" => "__nvvm_d2f_rm_ftz", + "llvm.nvvm.d2f.rn" => "__nvvm_d2f_rn", + "llvm.nvvm.d2f.rn.ftz" => "__nvvm_d2f_rn_ftz", + "llvm.nvvm.d2f.rp" => "__nvvm_d2f_rp", + "llvm.nvvm.d2f.rp.ftz" => "__nvvm_d2f_rp_ftz", + "llvm.nvvm.d2f.rz" => "__nvvm_d2f_rz", + "llvm.nvvm.d2f.rz.ftz" => "__nvvm_d2f_rz_ftz", + "llvm.nvvm.d2i.hi" => "__nvvm_d2i_hi", + "llvm.nvvm.d2i.lo" => "__nvvm_d2i_lo", + "llvm.nvvm.d2i.rm" => "__nvvm_d2i_rm", + "llvm.nvvm.d2i.rn" => "__nvvm_d2i_rn", + "llvm.nvvm.d2i.rp" => "__nvvm_d2i_rp", + "llvm.nvvm.d2i.rz" => "__nvvm_d2i_rz", + "llvm.nvvm.d2ll.rm" => "__nvvm_d2ll_rm", + "llvm.nvvm.d2ll.rn" => "__nvvm_d2ll_rn", + "llvm.nvvm.d2ll.rp" => "__nvvm_d2ll_rp", + "llvm.nvvm.d2ll.rz" => "__nvvm_d2ll_rz", + "llvm.nvvm.d2ui.rm" => "__nvvm_d2ui_rm", + "llvm.nvvm.d2ui.rn" => "__nvvm_d2ui_rn", + "llvm.nvvm.d2ui.rp" => "__nvvm_d2ui_rp", + "llvm.nvvm.d2ui.rz" => "__nvvm_d2ui_rz", + "llvm.nvvm.d2ull.rm" => "__nvvm_d2ull_rm", + "llvm.nvvm.d2ull.rn" => "__nvvm_d2ull_rn", + "llvm.nvvm.d2ull.rp" => "__nvvm_d2ull_rp", + "llvm.nvvm.d2ull.rz" => "__nvvm_d2ull_rz", + "llvm.nvvm.div.approx.f" => "__nvvm_div_approx_f", + "llvm.nvvm.div.approx.ftz.f" => "__nvvm_div_approx_ftz_f", + "llvm.nvvm.div.rm.d" => "__nvvm_div_rm_d", + "llvm.nvvm.div.rm.f" => "__nvvm_div_rm_f", + "llvm.nvvm.div.rm.ftz.f" => "__nvvm_div_rm_ftz_f", + "llvm.nvvm.div.rn.d" => "__nvvm_div_rn_d", + "llvm.nvvm.div.rn.f" => "__nvvm_div_rn_f", + "llvm.nvvm.div.rn.ftz.f" => "__nvvm_div_rn_ftz_f", + "llvm.nvvm.div.rp.d" => "__nvvm_div_rp_d", + "llvm.nvvm.div.rp.f" => "__nvvm_div_rp_f", + "llvm.nvvm.div.rp.ftz.f" => "__nvvm_div_rp_ftz_f", + "llvm.nvvm.div.rz.d" => "__nvvm_div_rz_d", + "llvm.nvvm.div.rz.f" => "__nvvm_div_rz_f", + "llvm.nvvm.div.rz.ftz.f" => "__nvvm_div_rz_ftz_f", + "llvm.nvvm.ex2.approx.d" => "__nvvm_ex2_approx_d", + "llvm.nvvm.ex2.approx.f" => "__nvvm_ex2_approx_f", + "llvm.nvvm.ex2.approx.ftz.f" => "__nvvm_ex2_approx_ftz_f", + "llvm.nvvm.f2h.rn" => "__nvvm_f2h_rn", + "llvm.nvvm.f2h.rn.ftz" => "__nvvm_f2h_rn_ftz", + "llvm.nvvm.f2i.rm" => "__nvvm_f2i_rm", + "llvm.nvvm.f2i.rm.ftz" => "__nvvm_f2i_rm_ftz", + "llvm.nvvm.f2i.rn" => "__nvvm_f2i_rn", + "llvm.nvvm.f2i.rn.ftz" => "__nvvm_f2i_rn_ftz", + "llvm.nvvm.f2i.rp" => "__nvvm_f2i_rp", + "llvm.nvvm.f2i.rp.ftz" => "__nvvm_f2i_rp_ftz", + "llvm.nvvm.f2i.rz" => "__nvvm_f2i_rz", + "llvm.nvvm.f2i.rz.ftz" => "__nvvm_f2i_rz_ftz", + "llvm.nvvm.f2ll.rm" => "__nvvm_f2ll_rm", + "llvm.nvvm.f2ll.rm.ftz" => "__nvvm_f2ll_rm_ftz", + "llvm.nvvm.f2ll.rn" => "__nvvm_f2ll_rn", + "llvm.nvvm.f2ll.rn.ftz" => "__nvvm_f2ll_rn_ftz", + "llvm.nvvm.f2ll.rp" => "__nvvm_f2ll_rp", + "llvm.nvvm.f2ll.rp.ftz" => "__nvvm_f2ll_rp_ftz", + "llvm.nvvm.f2ll.rz" => "__nvvm_f2ll_rz", + "llvm.nvvm.f2ll.rz.ftz" => "__nvvm_f2ll_rz_ftz", + "llvm.nvvm.f2ui.rm" => "__nvvm_f2ui_rm", + "llvm.nvvm.f2ui.rm.ftz" => "__nvvm_f2ui_rm_ftz", + "llvm.nvvm.f2ui.rn" => "__nvvm_f2ui_rn", + "llvm.nvvm.f2ui.rn.ftz" => "__nvvm_f2ui_rn_ftz", + "llvm.nvvm.f2ui.rp" => "__nvvm_f2ui_rp", + "llvm.nvvm.f2ui.rp.ftz" => "__nvvm_f2ui_rp_ftz", + "llvm.nvvm.f2ui.rz" => "__nvvm_f2ui_rz", + "llvm.nvvm.f2ui.rz.ftz" => "__nvvm_f2ui_rz_ftz", + "llvm.nvvm.f2ull.rm" => "__nvvm_f2ull_rm", + "llvm.nvvm.f2ull.rm.ftz" => "__nvvm_f2ull_rm_ftz", + "llvm.nvvm.f2ull.rn" => "__nvvm_f2ull_rn", + "llvm.nvvm.f2ull.rn.ftz" => "__nvvm_f2ull_rn_ftz", + "llvm.nvvm.f2ull.rp" => "__nvvm_f2ull_rp", + "llvm.nvvm.f2ull.rp.ftz" => "__nvvm_f2ull_rp_ftz", + "llvm.nvvm.f2ull.rz" => "__nvvm_f2ull_rz", + "llvm.nvvm.f2ull.rz.ftz" => "__nvvm_f2ull_rz_ftz", + "llvm.nvvm.fabs.d" => "__nvvm_fabs_d", + "llvm.nvvm.fabs.f" => "__nvvm_fabs_f", + "llvm.nvvm.fabs.ftz.f" => "__nvvm_fabs_ftz_f", + "llvm.nvvm.floor.d" => "__nvvm_floor_d", + "llvm.nvvm.floor.f" => "__nvvm_floor_f", + "llvm.nvvm.floor.ftz.f" => "__nvvm_floor_ftz_f", + "llvm.nvvm.fma.rm.d" => "__nvvm_fma_rm_d", + "llvm.nvvm.fma.rm.f" => "__nvvm_fma_rm_f", + "llvm.nvvm.fma.rm.ftz.f" => "__nvvm_fma_rm_ftz_f", + "llvm.nvvm.fma.rn.d" => "__nvvm_fma_rn_d", + "llvm.nvvm.fma.rn.f" => "__nvvm_fma_rn_f", + "llvm.nvvm.fma.rn.ftz.f" => "__nvvm_fma_rn_ftz_f", + "llvm.nvvm.fma.rp.d" => "__nvvm_fma_rp_d", + "llvm.nvvm.fma.rp.f" => "__nvvm_fma_rp_f", + "llvm.nvvm.fma.rp.ftz.f" => "__nvvm_fma_rp_ftz_f", + "llvm.nvvm.fma.rz.d" => "__nvvm_fma_rz_d", + "llvm.nvvm.fma.rz.f" => "__nvvm_fma_rz_f", + "llvm.nvvm.fma.rz.ftz.f" => "__nvvm_fma_rz_ftz_f", + "llvm.nvvm.fmax.d" => "__nvvm_fmax_d", + "llvm.nvvm.fmax.f" => "__nvvm_fmax_f", + "llvm.nvvm.fmax.ftz.f" => "__nvvm_fmax_ftz_f", + "llvm.nvvm.fmin.d" => "__nvvm_fmin_d", + "llvm.nvvm.fmin.f" => "__nvvm_fmin_f", + "llvm.nvvm.fmin.ftz.f" => "__nvvm_fmin_ftz_f", + "llvm.nvvm.h2f" => "__nvvm_h2f", + "llvm.nvvm.i2d.rm" => "__nvvm_i2d_rm", + "llvm.nvvm.i2d.rn" => "__nvvm_i2d_rn", + "llvm.nvvm.i2d.rp" => "__nvvm_i2d_rp", + "llvm.nvvm.i2d.rz" => "__nvvm_i2d_rz", + "llvm.nvvm.i2f.rm" => "__nvvm_i2f_rm", + "llvm.nvvm.i2f.rn" => "__nvvm_i2f_rn", + "llvm.nvvm.i2f.rp" => "__nvvm_i2f_rp", + "llvm.nvvm.i2f.rz" => "__nvvm_i2f_rz", + "llvm.nvvm.isspacep.const" => "__nvvm_isspacep_const", + "llvm.nvvm.isspacep.global" => "__nvvm_isspacep_global", + "llvm.nvvm.isspacep.local" => "__nvvm_isspacep_local", + "llvm.nvvm.isspacep.shared" => "__nvvm_isspacep_shared", + "llvm.nvvm.istypep.sampler" => "__nvvm_istypep_sampler", + "llvm.nvvm.istypep.surface" => "__nvvm_istypep_surface", + "llvm.nvvm.istypep.texture" => "__nvvm_istypep_texture", + "llvm.nvvm.lg2.approx.d" => "__nvvm_lg2_approx_d", + "llvm.nvvm.lg2.approx.f" => "__nvvm_lg2_approx_f", + "llvm.nvvm.lg2.approx.ftz.f" => "__nvvm_lg2_approx_ftz_f", + "llvm.nvvm.ll2d.rm" => "__nvvm_ll2d_rm", + "llvm.nvvm.ll2d.rn" => "__nvvm_ll2d_rn", + "llvm.nvvm.ll2d.rp" => "__nvvm_ll2d_rp", + "llvm.nvvm.ll2d.rz" => "__nvvm_ll2d_rz", + "llvm.nvvm.ll2f.rm" => "__nvvm_ll2f_rm", + "llvm.nvvm.ll2f.rn" => "__nvvm_ll2f_rn", + "llvm.nvvm.ll2f.rp" => "__nvvm_ll2f_rp", + "llvm.nvvm.ll2f.rz" => "__nvvm_ll2f_rz", + "llvm.nvvm.lohi.i2d" => "__nvvm_lohi_i2d", + "llvm.nvvm.max.i" => "__nvvm_max_i", + "llvm.nvvm.max.ll" => "__nvvm_max_ll", + "llvm.nvvm.max.ui" => "__nvvm_max_ui", + "llvm.nvvm.max.ull" => "__nvvm_max_ull", + "llvm.nvvm.membar.cta" => "__nvvm_membar_cta", + "llvm.nvvm.membar.gl" => "__nvvm_membar_gl", + "llvm.nvvm.membar.sys" => "__nvvm_membar_sys", + "llvm.nvvm.min.i" => "__nvvm_min_i", + "llvm.nvvm.min.ll" => "__nvvm_min_ll", + "llvm.nvvm.min.ui" => "__nvvm_min_ui", + "llvm.nvvm.min.ull" => "__nvvm_min_ull", + "llvm.nvvm.mul.rm.d" => "__nvvm_mul_rm_d", + "llvm.nvvm.mul.rm.f" => "__nvvm_mul_rm_f", + "llvm.nvvm.mul.rm.ftz.f" => "__nvvm_mul_rm_ftz_f", + "llvm.nvvm.mul.rn.d" => "__nvvm_mul_rn_d", + "llvm.nvvm.mul.rn.f" => "__nvvm_mul_rn_f", + "llvm.nvvm.mul.rn.ftz.f" => "__nvvm_mul_rn_ftz_f", + "llvm.nvvm.mul.rp.d" => "__nvvm_mul_rp_d", + "llvm.nvvm.mul.rp.f" => "__nvvm_mul_rp_f", + "llvm.nvvm.mul.rp.ftz.f" => "__nvvm_mul_rp_ftz_f", + "llvm.nvvm.mul.rz.d" => "__nvvm_mul_rz_d", + "llvm.nvvm.mul.rz.f" => "__nvvm_mul_rz_f", + "llvm.nvvm.mul.rz.ftz.f" => "__nvvm_mul_rz_ftz_f", + "llvm.nvvm.mul24.i" => "__nvvm_mul24_i", + "llvm.nvvm.mul24.ui" => "__nvvm_mul24_ui", + "llvm.nvvm.mulhi.i" => "__nvvm_mulhi_i", + "llvm.nvvm.mulhi.ll" => "__nvvm_mulhi_ll", + "llvm.nvvm.mulhi.ui" => "__nvvm_mulhi_ui", + "llvm.nvvm.mulhi.ull" => "__nvvm_mulhi_ull", + "llvm.nvvm.popc.i" => "__nvvm_popc_i", + "llvm.nvvm.popc.ll" => "__nvvm_popc_ll", + "llvm.nvvm.prmt" => "__nvvm_prmt", + "llvm.nvvm.rcp.approx.ftz.d" => "__nvvm_rcp_approx_ftz_d", + "llvm.nvvm.rcp.rm.d" => "__nvvm_rcp_rm_d", + "llvm.nvvm.rcp.rm.f" => "__nvvm_rcp_rm_f", + "llvm.nvvm.rcp.rm.ftz.f" => "__nvvm_rcp_rm_ftz_f", + "llvm.nvvm.rcp.rn.d" => "__nvvm_rcp_rn_d", + "llvm.nvvm.rcp.rn.f" => "__nvvm_rcp_rn_f", + "llvm.nvvm.rcp.rn.ftz.f" => "__nvvm_rcp_rn_ftz_f", + "llvm.nvvm.rcp.rp.d" => "__nvvm_rcp_rp_d", + "llvm.nvvm.rcp.rp.f" => "__nvvm_rcp_rp_f", + "llvm.nvvm.rcp.rp.ftz.f" => "__nvvm_rcp_rp_ftz_f", + "llvm.nvvm.rcp.rz.d" => "__nvvm_rcp_rz_d", + "llvm.nvvm.rcp.rz.f" => "__nvvm_rcp_rz_f", + "llvm.nvvm.rcp.rz.ftz.f" => "__nvvm_rcp_rz_ftz_f", + "llvm.nvvm.read.ptx.sreg.clock" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.clock64" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.ctaid.x" => "__nvvm_read_ptx_sreg_ctaid_x", + "llvm.nvvm.read.ptx.sreg.ctaid.y" => "__nvvm_read_ptx_sreg_ctaid_y", + "llvm.nvvm.read.ptx.sreg.ctaid.z" => "__nvvm_read_ptx_sreg_ctaid_z", + "llvm.nvvm.read.ptx.sreg.envreg0" => "__nvvm_read_ptx_sreg_envreg0", + "llvm.nvvm.read.ptx.sreg.envreg1" => "__nvvm_read_ptx_sreg_envreg1", + "llvm.nvvm.read.ptx.sreg.envreg10" => "__nvvm_read_ptx_sreg_envreg10", + "llvm.nvvm.read.ptx.sreg.envreg11" => "__nvvm_read_ptx_sreg_envreg11", + "llvm.nvvm.read.ptx.sreg.envreg12" => "__nvvm_read_ptx_sreg_envreg12", + "llvm.nvvm.read.ptx.sreg.envreg13" => "__nvvm_read_ptx_sreg_envreg13", + "llvm.nvvm.read.ptx.sreg.envreg14" => "__nvvm_read_ptx_sreg_envreg14", + "llvm.nvvm.read.ptx.sreg.envreg15" => "__nvvm_read_ptx_sreg_envreg15", + "llvm.nvvm.read.ptx.sreg.envreg16" => "__nvvm_read_ptx_sreg_envreg16", + "llvm.nvvm.read.ptx.sreg.envreg17" => "__nvvm_read_ptx_sreg_envreg17", + "llvm.nvvm.read.ptx.sreg.envreg18" => "__nvvm_read_ptx_sreg_envreg18", + "llvm.nvvm.read.ptx.sreg.envreg19" => "__nvvm_read_ptx_sreg_envreg19", + "llvm.nvvm.read.ptx.sreg.envreg2" => "__nvvm_read_ptx_sreg_envreg2", + "llvm.nvvm.read.ptx.sreg.envreg20" => "__nvvm_read_ptx_sreg_envreg20", + "llvm.nvvm.read.ptx.sreg.envreg21" => "__nvvm_read_ptx_sreg_envreg21", + "llvm.nvvm.read.ptx.sreg.envreg22" => "__nvvm_read_ptx_sreg_envreg22", + "llvm.nvvm.read.ptx.sreg.envreg23" => "__nvvm_read_ptx_sreg_envreg23", + "llvm.nvvm.read.ptx.sreg.envreg24" => "__nvvm_read_ptx_sreg_envreg24", + "llvm.nvvm.read.ptx.sreg.envreg25" => "__nvvm_read_ptx_sreg_envreg25", + "llvm.nvvm.read.ptx.sreg.envreg26" => "__nvvm_read_ptx_sreg_envreg26", + "llvm.nvvm.read.ptx.sreg.envreg27" => "__nvvm_read_ptx_sreg_envreg27", + "llvm.nvvm.read.ptx.sreg.envreg28" => "__nvvm_read_ptx_sreg_envreg28", + "llvm.nvvm.read.ptx.sreg.envreg29" => "__nvvm_read_ptx_sreg_envreg29", + "llvm.nvvm.read.ptx.sreg.envreg3" => "__nvvm_read_ptx_sreg_envreg3", + "llvm.nvvm.read.ptx.sreg.envreg30" => "__nvvm_read_ptx_sreg_envreg30", + "llvm.nvvm.read.ptx.sreg.envreg31" => "__nvvm_read_ptx_sreg_envreg31", + "llvm.nvvm.read.ptx.sreg.envreg4" => "__nvvm_read_ptx_sreg_envreg4", + "llvm.nvvm.read.ptx.sreg.envreg5" => "__nvvm_read_ptx_sreg_envreg5", + "llvm.nvvm.read.ptx.sreg.envreg6" => "__nvvm_read_ptx_sreg_envreg6", + "llvm.nvvm.read.ptx.sreg.envreg7" => "__nvvm_read_ptx_sreg_envreg7", + "llvm.nvvm.read.ptx.sreg.envreg8" => "__nvvm_read_ptx_sreg_envreg8", + "llvm.nvvm.read.ptx.sreg.envreg9" => "__nvvm_read_ptx_sreg_envreg9", + "llvm.nvvm.read.ptx.sreg.gridid" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.laneid" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.lanemask.eq" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.lanemask.ge" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.lanemask.gt" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.lanemask.le" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.lanemask.lt" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.nctaid.x" => "__nvvm_read_ptx_sreg_nctaid_x", + "llvm.nvvm.read.ptx.sreg.nctaid.y" => "__nvvm_read_ptx_sreg_nctaid_y", + "llvm.nvvm.read.ptx.sreg.nctaid.z" => "__nvvm_read_ptx_sreg_nctaid_z", + "llvm.nvvm.read.ptx.sreg.nsmid" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.ntid.x" => "__nvvm_read_ptx_sreg_ntid_x", + "llvm.nvvm.read.ptx.sreg.ntid.y" => "__nvvm_read_ptx_sreg_ntid_y", + "llvm.nvvm.read.ptx.sreg.ntid.z" => "__nvvm_read_ptx_sreg_ntid_z", + "llvm.nvvm.read.ptx.sreg.nwarpid" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.pm0" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.pm1" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.pm2" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.pm3" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.smid" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.tid.x" => "__nvvm_read_ptx_sreg_tid_x", + "llvm.nvvm.read.ptx.sreg.tid.y" => "__nvvm_read_ptx_sreg_tid_y", + "llvm.nvvm.read.ptx.sreg.tid.z" => "__nvvm_read_ptx_sreg_tid_z", + "llvm.nvvm.read.ptx.sreg.warpid" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.read.ptx.sreg.warpsize" => "__nvvm_read_ptx_sreg_warpsize", + // [DUPLICATE]: "llvm.nvvm.read.ptx.sreg.warpsize" => "__nvvm_read_ptx_sreg_", + "llvm.nvvm.rotate.b32" => "__nvvm_rotate_b32", + "llvm.nvvm.rotate.b64" => "__nvvm_rotate_b64", + "llvm.nvvm.rotate.right.b64" => "__nvvm_rotate_right_b64", + "llvm.nvvm.round.d" => "__nvvm_round_d", + "llvm.nvvm.round.f" => "__nvvm_round_f", + "llvm.nvvm.round.ftz.f" => "__nvvm_round_ftz_f", + "llvm.nvvm.rsqrt.approx.d" => "__nvvm_rsqrt_approx_d", + "llvm.nvvm.rsqrt.approx.f" => "__nvvm_rsqrt_approx_f", + "llvm.nvvm.rsqrt.approx.ftz.f" => "__nvvm_rsqrt_approx_ftz_f", + "llvm.nvvm.sad.i" => "__nvvm_sad_i", + "llvm.nvvm.sad.ui" => "__nvvm_sad_ui", + "llvm.nvvm.saturate.d" => "__nvvm_saturate_d", + "llvm.nvvm.saturate.f" => "__nvvm_saturate_f", + "llvm.nvvm.saturate.ftz.f" => "__nvvm_saturate_ftz_f", + "llvm.nvvm.shfl.bfly.f32" => "__nvvm_shfl_bfly_f32", + "llvm.nvvm.shfl.bfly.i32" => "__nvvm_shfl_bfly_i32", + "llvm.nvvm.shfl.down.f32" => "__nvvm_shfl_down_f32", + "llvm.nvvm.shfl.down.i32" => "__nvvm_shfl_down_i32", + "llvm.nvvm.shfl.idx.f32" => "__nvvm_shfl_idx_f32", + "llvm.nvvm.shfl.idx.i32" => "__nvvm_shfl_idx_i32", + "llvm.nvvm.shfl.up.f32" => "__nvvm_shfl_up_f32", + "llvm.nvvm.shfl.up.i32" => "__nvvm_shfl_up_i32", + "llvm.nvvm.sin.approx.f" => "__nvvm_sin_approx_f", + "llvm.nvvm.sin.approx.ftz.f" => "__nvvm_sin_approx_ftz_f", + "llvm.nvvm.sqrt.approx.f" => "__nvvm_sqrt_approx_f", + "llvm.nvvm.sqrt.approx.ftz.f" => "__nvvm_sqrt_approx_ftz_f", + "llvm.nvvm.sqrt.f" => "__nvvm_sqrt_f", + "llvm.nvvm.sqrt.rm.d" => "__nvvm_sqrt_rm_d", + "llvm.nvvm.sqrt.rm.f" => "__nvvm_sqrt_rm_f", + "llvm.nvvm.sqrt.rm.ftz.f" => "__nvvm_sqrt_rm_ftz_f", + "llvm.nvvm.sqrt.rn.d" => "__nvvm_sqrt_rn_d", + "llvm.nvvm.sqrt.rn.f" => "__nvvm_sqrt_rn_f", + "llvm.nvvm.sqrt.rn.ftz.f" => "__nvvm_sqrt_rn_ftz_f", + "llvm.nvvm.sqrt.rp.d" => "__nvvm_sqrt_rp_d", + "llvm.nvvm.sqrt.rp.f" => "__nvvm_sqrt_rp_f", + "llvm.nvvm.sqrt.rp.ftz.f" => "__nvvm_sqrt_rp_ftz_f", + "llvm.nvvm.sqrt.rz.d" => "__nvvm_sqrt_rz_d", + "llvm.nvvm.sqrt.rz.f" => "__nvvm_sqrt_rz_f", + "llvm.nvvm.sqrt.rz.ftz.f" => "__nvvm_sqrt_rz_ftz_f", + "llvm.nvvm.suq.array.size" => "__nvvm_suq_array_size", + "llvm.nvvm.suq.channel.data.type" => "__nvvm_suq_channel_data_type", + "llvm.nvvm.suq.channel.order" => "__nvvm_suq_channel_order", + "llvm.nvvm.suq.depth" => "__nvvm_suq_depth", + "llvm.nvvm.suq.height" => "__nvvm_suq_height", + "llvm.nvvm.suq.width" => "__nvvm_suq_width", + "llvm.nvvm.sust.b.1d.array.i16.clamp" => "__nvvm_sust_b_1d_array_i16_clamp", + "llvm.nvvm.sust.b.1d.array.i16.trap" => "__nvvm_sust_b_1d_array_i16_trap", + "llvm.nvvm.sust.b.1d.array.i16.zero" => "__nvvm_sust_b_1d_array_i16_zero", + "llvm.nvvm.sust.b.1d.array.i32.clamp" => "__nvvm_sust_b_1d_array_i32_clamp", + "llvm.nvvm.sust.b.1d.array.i32.trap" => "__nvvm_sust_b_1d_array_i32_trap", + "llvm.nvvm.sust.b.1d.array.i32.zero" => "__nvvm_sust_b_1d_array_i32_zero", + "llvm.nvvm.sust.b.1d.array.i64.clamp" => "__nvvm_sust_b_1d_array_i64_clamp", + "llvm.nvvm.sust.b.1d.array.i64.trap" => "__nvvm_sust_b_1d_array_i64_trap", + "llvm.nvvm.sust.b.1d.array.i64.zero" => "__nvvm_sust_b_1d_array_i64_zero", + "llvm.nvvm.sust.b.1d.array.i8.clamp" => "__nvvm_sust_b_1d_array_i8_clamp", + "llvm.nvvm.sust.b.1d.array.i8.trap" => "__nvvm_sust_b_1d_array_i8_trap", + "llvm.nvvm.sust.b.1d.array.i8.zero" => "__nvvm_sust_b_1d_array_i8_zero", + "llvm.nvvm.sust.b.1d.array.v2i16.clamp" => "__nvvm_sust_b_1d_array_v2i16_clamp", + "llvm.nvvm.sust.b.1d.array.v2i16.trap" => "__nvvm_sust_b_1d_array_v2i16_trap", + "llvm.nvvm.sust.b.1d.array.v2i16.zero" => "__nvvm_sust_b_1d_array_v2i16_zero", + "llvm.nvvm.sust.b.1d.array.v2i32.clamp" => "__nvvm_sust_b_1d_array_v2i32_clamp", + "llvm.nvvm.sust.b.1d.array.v2i32.trap" => "__nvvm_sust_b_1d_array_v2i32_trap", + "llvm.nvvm.sust.b.1d.array.v2i32.zero" => "__nvvm_sust_b_1d_array_v2i32_zero", + "llvm.nvvm.sust.b.1d.array.v2i64.clamp" => "__nvvm_sust_b_1d_array_v2i64_clamp", + "llvm.nvvm.sust.b.1d.array.v2i64.trap" => "__nvvm_sust_b_1d_array_v2i64_trap", + "llvm.nvvm.sust.b.1d.array.v2i64.zero" => "__nvvm_sust_b_1d_array_v2i64_zero", + "llvm.nvvm.sust.b.1d.array.v2i8.clamp" => "__nvvm_sust_b_1d_array_v2i8_clamp", + "llvm.nvvm.sust.b.1d.array.v2i8.trap" => "__nvvm_sust_b_1d_array_v2i8_trap", + "llvm.nvvm.sust.b.1d.array.v2i8.zero" => "__nvvm_sust_b_1d_array_v2i8_zero", + "llvm.nvvm.sust.b.1d.array.v4i16.clamp" => "__nvvm_sust_b_1d_array_v4i16_clamp", + "llvm.nvvm.sust.b.1d.array.v4i16.trap" => "__nvvm_sust_b_1d_array_v4i16_trap", + "llvm.nvvm.sust.b.1d.array.v4i16.zero" => "__nvvm_sust_b_1d_array_v4i16_zero", + "llvm.nvvm.sust.b.1d.array.v4i32.clamp" => "__nvvm_sust_b_1d_array_v4i32_clamp", + "llvm.nvvm.sust.b.1d.array.v4i32.trap" => "__nvvm_sust_b_1d_array_v4i32_trap", + "llvm.nvvm.sust.b.1d.array.v4i32.zero" => "__nvvm_sust_b_1d_array_v4i32_zero", + "llvm.nvvm.sust.b.1d.array.v4i8.clamp" => "__nvvm_sust_b_1d_array_v4i8_clamp", + "llvm.nvvm.sust.b.1d.array.v4i8.trap" => "__nvvm_sust_b_1d_array_v4i8_trap", + "llvm.nvvm.sust.b.1d.array.v4i8.zero" => "__nvvm_sust_b_1d_array_v4i8_zero", + "llvm.nvvm.sust.b.1d.i16.clamp" => "__nvvm_sust_b_1d_i16_clamp", + "llvm.nvvm.sust.b.1d.i16.trap" => "__nvvm_sust_b_1d_i16_trap", + "llvm.nvvm.sust.b.1d.i16.zero" => "__nvvm_sust_b_1d_i16_zero", + "llvm.nvvm.sust.b.1d.i32.clamp" => "__nvvm_sust_b_1d_i32_clamp", + "llvm.nvvm.sust.b.1d.i32.trap" => "__nvvm_sust_b_1d_i32_trap", + "llvm.nvvm.sust.b.1d.i32.zero" => "__nvvm_sust_b_1d_i32_zero", + "llvm.nvvm.sust.b.1d.i64.clamp" => "__nvvm_sust_b_1d_i64_clamp", + "llvm.nvvm.sust.b.1d.i64.trap" => "__nvvm_sust_b_1d_i64_trap", + "llvm.nvvm.sust.b.1d.i64.zero" => "__nvvm_sust_b_1d_i64_zero", + "llvm.nvvm.sust.b.1d.i8.clamp" => "__nvvm_sust_b_1d_i8_clamp", + "llvm.nvvm.sust.b.1d.i8.trap" => "__nvvm_sust_b_1d_i8_trap", + "llvm.nvvm.sust.b.1d.i8.zero" => "__nvvm_sust_b_1d_i8_zero", + "llvm.nvvm.sust.b.1d.v2i16.clamp" => "__nvvm_sust_b_1d_v2i16_clamp", + "llvm.nvvm.sust.b.1d.v2i16.trap" => "__nvvm_sust_b_1d_v2i16_trap", + "llvm.nvvm.sust.b.1d.v2i16.zero" => "__nvvm_sust_b_1d_v2i16_zero", + "llvm.nvvm.sust.b.1d.v2i32.clamp" => "__nvvm_sust_b_1d_v2i32_clamp", + "llvm.nvvm.sust.b.1d.v2i32.trap" => "__nvvm_sust_b_1d_v2i32_trap", + "llvm.nvvm.sust.b.1d.v2i32.zero" => "__nvvm_sust_b_1d_v2i32_zero", + "llvm.nvvm.sust.b.1d.v2i64.clamp" => "__nvvm_sust_b_1d_v2i64_clamp", + "llvm.nvvm.sust.b.1d.v2i64.trap" => "__nvvm_sust_b_1d_v2i64_trap", + "llvm.nvvm.sust.b.1d.v2i64.zero" => "__nvvm_sust_b_1d_v2i64_zero", + "llvm.nvvm.sust.b.1d.v2i8.clamp" => "__nvvm_sust_b_1d_v2i8_clamp", + "llvm.nvvm.sust.b.1d.v2i8.trap" => "__nvvm_sust_b_1d_v2i8_trap", + "llvm.nvvm.sust.b.1d.v2i8.zero" => "__nvvm_sust_b_1d_v2i8_zero", + "llvm.nvvm.sust.b.1d.v4i16.clamp" => "__nvvm_sust_b_1d_v4i16_clamp", + "llvm.nvvm.sust.b.1d.v4i16.trap" => "__nvvm_sust_b_1d_v4i16_trap", + "llvm.nvvm.sust.b.1d.v4i16.zero" => "__nvvm_sust_b_1d_v4i16_zero", + "llvm.nvvm.sust.b.1d.v4i32.clamp" => "__nvvm_sust_b_1d_v4i32_clamp", + "llvm.nvvm.sust.b.1d.v4i32.trap" => "__nvvm_sust_b_1d_v4i32_trap", + "llvm.nvvm.sust.b.1d.v4i32.zero" => "__nvvm_sust_b_1d_v4i32_zero", + "llvm.nvvm.sust.b.1d.v4i8.clamp" => "__nvvm_sust_b_1d_v4i8_clamp", + "llvm.nvvm.sust.b.1d.v4i8.trap" => "__nvvm_sust_b_1d_v4i8_trap", + "llvm.nvvm.sust.b.1d.v4i8.zero" => "__nvvm_sust_b_1d_v4i8_zero", + "llvm.nvvm.sust.b.2d.array.i16.clamp" => "__nvvm_sust_b_2d_array_i16_clamp", + "llvm.nvvm.sust.b.2d.array.i16.trap" => "__nvvm_sust_b_2d_array_i16_trap", + "llvm.nvvm.sust.b.2d.array.i16.zero" => "__nvvm_sust_b_2d_array_i16_zero", + "llvm.nvvm.sust.b.2d.array.i32.clamp" => "__nvvm_sust_b_2d_array_i32_clamp", + "llvm.nvvm.sust.b.2d.array.i32.trap" => "__nvvm_sust_b_2d_array_i32_trap", + "llvm.nvvm.sust.b.2d.array.i32.zero" => "__nvvm_sust_b_2d_array_i32_zero", + "llvm.nvvm.sust.b.2d.array.i64.clamp" => "__nvvm_sust_b_2d_array_i64_clamp", + "llvm.nvvm.sust.b.2d.array.i64.trap" => "__nvvm_sust_b_2d_array_i64_trap", + "llvm.nvvm.sust.b.2d.array.i64.zero" => "__nvvm_sust_b_2d_array_i64_zero", + "llvm.nvvm.sust.b.2d.array.i8.clamp" => "__nvvm_sust_b_2d_array_i8_clamp", + "llvm.nvvm.sust.b.2d.array.i8.trap" => "__nvvm_sust_b_2d_array_i8_trap", + "llvm.nvvm.sust.b.2d.array.i8.zero" => "__nvvm_sust_b_2d_array_i8_zero", + "llvm.nvvm.sust.b.2d.array.v2i16.clamp" => "__nvvm_sust_b_2d_array_v2i16_clamp", + "llvm.nvvm.sust.b.2d.array.v2i16.trap" => "__nvvm_sust_b_2d_array_v2i16_trap", + "llvm.nvvm.sust.b.2d.array.v2i16.zero" => "__nvvm_sust_b_2d_array_v2i16_zero", + "llvm.nvvm.sust.b.2d.array.v2i32.clamp" => "__nvvm_sust_b_2d_array_v2i32_clamp", + "llvm.nvvm.sust.b.2d.array.v2i32.trap" => "__nvvm_sust_b_2d_array_v2i32_trap", + "llvm.nvvm.sust.b.2d.array.v2i32.zero" => "__nvvm_sust_b_2d_array_v2i32_zero", + "llvm.nvvm.sust.b.2d.array.v2i64.clamp" => "__nvvm_sust_b_2d_array_v2i64_clamp", + "llvm.nvvm.sust.b.2d.array.v2i64.trap" => "__nvvm_sust_b_2d_array_v2i64_trap", + "llvm.nvvm.sust.b.2d.array.v2i64.zero" => "__nvvm_sust_b_2d_array_v2i64_zero", + "llvm.nvvm.sust.b.2d.array.v2i8.clamp" => "__nvvm_sust_b_2d_array_v2i8_clamp", + "llvm.nvvm.sust.b.2d.array.v2i8.trap" => "__nvvm_sust_b_2d_array_v2i8_trap", + "llvm.nvvm.sust.b.2d.array.v2i8.zero" => "__nvvm_sust_b_2d_array_v2i8_zero", + "llvm.nvvm.sust.b.2d.array.v4i16.clamp" => "__nvvm_sust_b_2d_array_v4i16_clamp", + "llvm.nvvm.sust.b.2d.array.v4i16.trap" => "__nvvm_sust_b_2d_array_v4i16_trap", + "llvm.nvvm.sust.b.2d.array.v4i16.zero" => "__nvvm_sust_b_2d_array_v4i16_zero", + "llvm.nvvm.sust.b.2d.array.v4i32.clamp" => "__nvvm_sust_b_2d_array_v4i32_clamp", + "llvm.nvvm.sust.b.2d.array.v4i32.trap" => "__nvvm_sust_b_2d_array_v4i32_trap", + "llvm.nvvm.sust.b.2d.array.v4i32.zero" => "__nvvm_sust_b_2d_array_v4i32_zero", + "llvm.nvvm.sust.b.2d.array.v4i8.clamp" => "__nvvm_sust_b_2d_array_v4i8_clamp", + "llvm.nvvm.sust.b.2d.array.v4i8.trap" => "__nvvm_sust_b_2d_array_v4i8_trap", + "llvm.nvvm.sust.b.2d.array.v4i8.zero" => "__nvvm_sust_b_2d_array_v4i8_zero", + "llvm.nvvm.sust.b.2d.i16.clamp" => "__nvvm_sust_b_2d_i16_clamp", + "llvm.nvvm.sust.b.2d.i16.trap" => "__nvvm_sust_b_2d_i16_trap", + "llvm.nvvm.sust.b.2d.i16.zero" => "__nvvm_sust_b_2d_i16_zero", + "llvm.nvvm.sust.b.2d.i32.clamp" => "__nvvm_sust_b_2d_i32_clamp", + "llvm.nvvm.sust.b.2d.i32.trap" => "__nvvm_sust_b_2d_i32_trap", + "llvm.nvvm.sust.b.2d.i32.zero" => "__nvvm_sust_b_2d_i32_zero", + "llvm.nvvm.sust.b.2d.i64.clamp" => "__nvvm_sust_b_2d_i64_clamp", + "llvm.nvvm.sust.b.2d.i64.trap" => "__nvvm_sust_b_2d_i64_trap", + "llvm.nvvm.sust.b.2d.i64.zero" => "__nvvm_sust_b_2d_i64_zero", + "llvm.nvvm.sust.b.2d.i8.clamp" => "__nvvm_sust_b_2d_i8_clamp", + "llvm.nvvm.sust.b.2d.i8.trap" => "__nvvm_sust_b_2d_i8_trap", + "llvm.nvvm.sust.b.2d.i8.zero" => "__nvvm_sust_b_2d_i8_zero", + "llvm.nvvm.sust.b.2d.v2i16.clamp" => "__nvvm_sust_b_2d_v2i16_clamp", + "llvm.nvvm.sust.b.2d.v2i16.trap" => "__nvvm_sust_b_2d_v2i16_trap", + "llvm.nvvm.sust.b.2d.v2i16.zero" => "__nvvm_sust_b_2d_v2i16_zero", + "llvm.nvvm.sust.b.2d.v2i32.clamp" => "__nvvm_sust_b_2d_v2i32_clamp", + "llvm.nvvm.sust.b.2d.v2i32.trap" => "__nvvm_sust_b_2d_v2i32_trap", + "llvm.nvvm.sust.b.2d.v2i32.zero" => "__nvvm_sust_b_2d_v2i32_zero", + "llvm.nvvm.sust.b.2d.v2i64.clamp" => "__nvvm_sust_b_2d_v2i64_clamp", + "llvm.nvvm.sust.b.2d.v2i64.trap" => "__nvvm_sust_b_2d_v2i64_trap", + "llvm.nvvm.sust.b.2d.v2i64.zero" => "__nvvm_sust_b_2d_v2i64_zero", + "llvm.nvvm.sust.b.2d.v2i8.clamp" => "__nvvm_sust_b_2d_v2i8_clamp", + "llvm.nvvm.sust.b.2d.v2i8.trap" => "__nvvm_sust_b_2d_v2i8_trap", + "llvm.nvvm.sust.b.2d.v2i8.zero" => "__nvvm_sust_b_2d_v2i8_zero", + "llvm.nvvm.sust.b.2d.v4i16.clamp" => "__nvvm_sust_b_2d_v4i16_clamp", + "llvm.nvvm.sust.b.2d.v4i16.trap" => "__nvvm_sust_b_2d_v4i16_trap", + "llvm.nvvm.sust.b.2d.v4i16.zero" => "__nvvm_sust_b_2d_v4i16_zero", + "llvm.nvvm.sust.b.2d.v4i32.clamp" => "__nvvm_sust_b_2d_v4i32_clamp", + "llvm.nvvm.sust.b.2d.v4i32.trap" => "__nvvm_sust_b_2d_v4i32_trap", + "llvm.nvvm.sust.b.2d.v4i32.zero" => "__nvvm_sust_b_2d_v4i32_zero", + "llvm.nvvm.sust.b.2d.v4i8.clamp" => "__nvvm_sust_b_2d_v4i8_clamp", + "llvm.nvvm.sust.b.2d.v4i8.trap" => "__nvvm_sust_b_2d_v4i8_trap", + "llvm.nvvm.sust.b.2d.v4i8.zero" => "__nvvm_sust_b_2d_v4i8_zero", + "llvm.nvvm.sust.b.3d.i16.clamp" => "__nvvm_sust_b_3d_i16_clamp", + "llvm.nvvm.sust.b.3d.i16.trap" => "__nvvm_sust_b_3d_i16_trap", + "llvm.nvvm.sust.b.3d.i16.zero" => "__nvvm_sust_b_3d_i16_zero", + "llvm.nvvm.sust.b.3d.i32.clamp" => "__nvvm_sust_b_3d_i32_clamp", + "llvm.nvvm.sust.b.3d.i32.trap" => "__nvvm_sust_b_3d_i32_trap", + "llvm.nvvm.sust.b.3d.i32.zero" => "__nvvm_sust_b_3d_i32_zero", + "llvm.nvvm.sust.b.3d.i64.clamp" => "__nvvm_sust_b_3d_i64_clamp", + "llvm.nvvm.sust.b.3d.i64.trap" => "__nvvm_sust_b_3d_i64_trap", + "llvm.nvvm.sust.b.3d.i64.zero" => "__nvvm_sust_b_3d_i64_zero", + "llvm.nvvm.sust.b.3d.i8.clamp" => "__nvvm_sust_b_3d_i8_clamp", + "llvm.nvvm.sust.b.3d.i8.trap" => "__nvvm_sust_b_3d_i8_trap", + "llvm.nvvm.sust.b.3d.i8.zero" => "__nvvm_sust_b_3d_i8_zero", + "llvm.nvvm.sust.b.3d.v2i16.clamp" => "__nvvm_sust_b_3d_v2i16_clamp", + "llvm.nvvm.sust.b.3d.v2i16.trap" => "__nvvm_sust_b_3d_v2i16_trap", + "llvm.nvvm.sust.b.3d.v2i16.zero" => "__nvvm_sust_b_3d_v2i16_zero", + "llvm.nvvm.sust.b.3d.v2i32.clamp" => "__nvvm_sust_b_3d_v2i32_clamp", + "llvm.nvvm.sust.b.3d.v2i32.trap" => "__nvvm_sust_b_3d_v2i32_trap", + "llvm.nvvm.sust.b.3d.v2i32.zero" => "__nvvm_sust_b_3d_v2i32_zero", + "llvm.nvvm.sust.b.3d.v2i64.clamp" => "__nvvm_sust_b_3d_v2i64_clamp", + "llvm.nvvm.sust.b.3d.v2i64.trap" => "__nvvm_sust_b_3d_v2i64_trap", + "llvm.nvvm.sust.b.3d.v2i64.zero" => "__nvvm_sust_b_3d_v2i64_zero", + "llvm.nvvm.sust.b.3d.v2i8.clamp" => "__nvvm_sust_b_3d_v2i8_clamp", + "llvm.nvvm.sust.b.3d.v2i8.trap" => "__nvvm_sust_b_3d_v2i8_trap", + "llvm.nvvm.sust.b.3d.v2i8.zero" => "__nvvm_sust_b_3d_v2i8_zero", + "llvm.nvvm.sust.b.3d.v4i16.clamp" => "__nvvm_sust_b_3d_v4i16_clamp", + "llvm.nvvm.sust.b.3d.v4i16.trap" => "__nvvm_sust_b_3d_v4i16_trap", + "llvm.nvvm.sust.b.3d.v4i16.zero" => "__nvvm_sust_b_3d_v4i16_zero", + "llvm.nvvm.sust.b.3d.v4i32.clamp" => "__nvvm_sust_b_3d_v4i32_clamp", + "llvm.nvvm.sust.b.3d.v4i32.trap" => "__nvvm_sust_b_3d_v4i32_trap", + "llvm.nvvm.sust.b.3d.v4i32.zero" => "__nvvm_sust_b_3d_v4i32_zero", + "llvm.nvvm.sust.b.3d.v4i8.clamp" => "__nvvm_sust_b_3d_v4i8_clamp", + "llvm.nvvm.sust.b.3d.v4i8.trap" => "__nvvm_sust_b_3d_v4i8_trap", + "llvm.nvvm.sust.b.3d.v4i8.zero" => "__nvvm_sust_b_3d_v4i8_zero", + "llvm.nvvm.sust.p.1d.array.i16.trap" => "__nvvm_sust_p_1d_array_i16_trap", + "llvm.nvvm.sust.p.1d.array.i32.trap" => "__nvvm_sust_p_1d_array_i32_trap", + "llvm.nvvm.sust.p.1d.array.i8.trap" => "__nvvm_sust_p_1d_array_i8_trap", + "llvm.nvvm.sust.p.1d.array.v2i16.trap" => "__nvvm_sust_p_1d_array_v2i16_trap", + "llvm.nvvm.sust.p.1d.array.v2i32.trap" => "__nvvm_sust_p_1d_array_v2i32_trap", + "llvm.nvvm.sust.p.1d.array.v2i8.trap" => "__nvvm_sust_p_1d_array_v2i8_trap", + "llvm.nvvm.sust.p.1d.array.v4i16.trap" => "__nvvm_sust_p_1d_array_v4i16_trap", + "llvm.nvvm.sust.p.1d.array.v4i32.trap" => "__nvvm_sust_p_1d_array_v4i32_trap", + "llvm.nvvm.sust.p.1d.array.v4i8.trap" => "__nvvm_sust_p_1d_array_v4i8_trap", + "llvm.nvvm.sust.p.1d.i16.trap" => "__nvvm_sust_p_1d_i16_trap", + "llvm.nvvm.sust.p.1d.i32.trap" => "__nvvm_sust_p_1d_i32_trap", + "llvm.nvvm.sust.p.1d.i8.trap" => "__nvvm_sust_p_1d_i8_trap", + "llvm.nvvm.sust.p.1d.v2i16.trap" => "__nvvm_sust_p_1d_v2i16_trap", + "llvm.nvvm.sust.p.1d.v2i32.trap" => "__nvvm_sust_p_1d_v2i32_trap", + "llvm.nvvm.sust.p.1d.v2i8.trap" => "__nvvm_sust_p_1d_v2i8_trap", + "llvm.nvvm.sust.p.1d.v4i16.trap" => "__nvvm_sust_p_1d_v4i16_trap", + "llvm.nvvm.sust.p.1d.v4i32.trap" => "__nvvm_sust_p_1d_v4i32_trap", + "llvm.nvvm.sust.p.1d.v4i8.trap" => "__nvvm_sust_p_1d_v4i8_trap", + "llvm.nvvm.sust.p.2d.array.i16.trap" => "__nvvm_sust_p_2d_array_i16_trap", + "llvm.nvvm.sust.p.2d.array.i32.trap" => "__nvvm_sust_p_2d_array_i32_trap", + "llvm.nvvm.sust.p.2d.array.i8.trap" => "__nvvm_sust_p_2d_array_i8_trap", + "llvm.nvvm.sust.p.2d.array.v2i16.trap" => "__nvvm_sust_p_2d_array_v2i16_trap", + "llvm.nvvm.sust.p.2d.array.v2i32.trap" => "__nvvm_sust_p_2d_array_v2i32_trap", + "llvm.nvvm.sust.p.2d.array.v2i8.trap" => "__nvvm_sust_p_2d_array_v2i8_trap", + "llvm.nvvm.sust.p.2d.array.v4i16.trap" => "__nvvm_sust_p_2d_array_v4i16_trap", + "llvm.nvvm.sust.p.2d.array.v4i32.trap" => "__nvvm_sust_p_2d_array_v4i32_trap", + "llvm.nvvm.sust.p.2d.array.v4i8.trap" => "__nvvm_sust_p_2d_array_v4i8_trap", + "llvm.nvvm.sust.p.2d.i16.trap" => "__nvvm_sust_p_2d_i16_trap", + "llvm.nvvm.sust.p.2d.i32.trap" => "__nvvm_sust_p_2d_i32_trap", + "llvm.nvvm.sust.p.2d.i8.trap" => "__nvvm_sust_p_2d_i8_trap", + "llvm.nvvm.sust.p.2d.v2i16.trap" => "__nvvm_sust_p_2d_v2i16_trap", + "llvm.nvvm.sust.p.2d.v2i32.trap" => "__nvvm_sust_p_2d_v2i32_trap", + "llvm.nvvm.sust.p.2d.v2i8.trap" => "__nvvm_sust_p_2d_v2i8_trap", + "llvm.nvvm.sust.p.2d.v4i16.trap" => "__nvvm_sust_p_2d_v4i16_trap", + "llvm.nvvm.sust.p.2d.v4i32.trap" => "__nvvm_sust_p_2d_v4i32_trap", + "llvm.nvvm.sust.p.2d.v4i8.trap" => "__nvvm_sust_p_2d_v4i8_trap", + "llvm.nvvm.sust.p.3d.i16.trap" => "__nvvm_sust_p_3d_i16_trap", + "llvm.nvvm.sust.p.3d.i32.trap" => "__nvvm_sust_p_3d_i32_trap", + "llvm.nvvm.sust.p.3d.i8.trap" => "__nvvm_sust_p_3d_i8_trap", + "llvm.nvvm.sust.p.3d.v2i16.trap" => "__nvvm_sust_p_3d_v2i16_trap", + "llvm.nvvm.sust.p.3d.v2i32.trap" => "__nvvm_sust_p_3d_v2i32_trap", + "llvm.nvvm.sust.p.3d.v2i8.trap" => "__nvvm_sust_p_3d_v2i8_trap", + "llvm.nvvm.sust.p.3d.v4i16.trap" => "__nvvm_sust_p_3d_v4i16_trap", + "llvm.nvvm.sust.p.3d.v4i32.trap" => "__nvvm_sust_p_3d_v4i32_trap", + "llvm.nvvm.sust.p.3d.v4i8.trap" => "__nvvm_sust_p_3d_v4i8_trap", + "llvm.nvvm.swap.lo.hi.b64" => "__nvvm_swap_lo_hi_b64", + "llvm.nvvm.trunc.d" => "__nvvm_trunc_d", + "llvm.nvvm.trunc.f" => "__nvvm_trunc_f", + "llvm.nvvm.trunc.ftz.f" => "__nvvm_trunc_ftz_f", + "llvm.nvvm.txq.array.size" => "__nvvm_txq_array_size", + "llvm.nvvm.txq.channel.data.type" => "__nvvm_txq_channel_data_type", + "llvm.nvvm.txq.channel.order" => "__nvvm_txq_channel_order", + "llvm.nvvm.txq.depth" => "__nvvm_txq_depth", + "llvm.nvvm.txq.height" => "__nvvm_txq_height", + "llvm.nvvm.txq.num.mipmap.levels" => "__nvvm_txq_num_mipmap_levels", + "llvm.nvvm.txq.num.samples" => "__nvvm_txq_num_samples", + "llvm.nvvm.txq.width" => "__nvvm_txq_width", + "llvm.nvvm.ui2d.rm" => "__nvvm_ui2d_rm", + "llvm.nvvm.ui2d.rn" => "__nvvm_ui2d_rn", + "llvm.nvvm.ui2d.rp" => "__nvvm_ui2d_rp", + "llvm.nvvm.ui2d.rz" => "__nvvm_ui2d_rz", + "llvm.nvvm.ui2f.rm" => "__nvvm_ui2f_rm", + "llvm.nvvm.ui2f.rn" => "__nvvm_ui2f_rn", + "llvm.nvvm.ui2f.rp" => "__nvvm_ui2f_rp", + "llvm.nvvm.ui2f.rz" => "__nvvm_ui2f_rz", + "llvm.nvvm.ull2d.rm" => "__nvvm_ull2d_rm", + "llvm.nvvm.ull2d.rn" => "__nvvm_ull2d_rn", + "llvm.nvvm.ull2d.rp" => "__nvvm_ull2d_rp", + "llvm.nvvm.ull2d.rz" => "__nvvm_ull2d_rz", + "llvm.nvvm.ull2f.rm" => "__nvvm_ull2f_rm", + "llvm.nvvm.ull2f.rn" => "__nvvm_ull2f_rn", + "llvm.nvvm.ull2f.rp" => "__nvvm_ull2f_rp", + "llvm.nvvm.ull2f.rz" => "__nvvm_ull2f_rz", + // ppc + "llvm.ppc.addex" => "__builtin_ppc_addex", + "llvm.ppc.addf128.round.to.odd" => "__builtin_addf128_round_to_odd", + "llvm.ppc.altivec.crypto.vcipher" => "__builtin_altivec_crypto_vcipher", + "llvm.ppc.altivec.crypto.vcipherlast" => "__builtin_altivec_crypto_vcipherlast", + "llvm.ppc.altivec.crypto.vncipher" => "__builtin_altivec_crypto_vncipher", + "llvm.ppc.altivec.crypto.vncipherlast" => "__builtin_altivec_crypto_vncipherlast", + "llvm.ppc.altivec.crypto.vpermxor" => "__builtin_altivec_crypto_vpermxor", + "llvm.ppc.altivec.crypto.vpermxor.be" => "__builtin_altivec_crypto_vpermxor_be", + "llvm.ppc.altivec.crypto.vpmsumb" => "__builtin_altivec_crypto_vpmsumb", + "llvm.ppc.altivec.crypto.vpmsumd" => "__builtin_altivec_crypto_vpmsumd", + "llvm.ppc.altivec.crypto.vpmsumh" => "__builtin_altivec_crypto_vpmsumh", + "llvm.ppc.altivec.crypto.vpmsumw" => "__builtin_altivec_crypto_vpmsumw", + "llvm.ppc.altivec.crypto.vsbox" => "__builtin_altivec_crypto_vsbox", + "llvm.ppc.altivec.crypto.vshasigmad" => "__builtin_altivec_crypto_vshasigmad", + "llvm.ppc.altivec.crypto.vshasigmaw" => "__builtin_altivec_crypto_vshasigmaw", + "llvm.ppc.altivec.dss" => "__builtin_altivec_dss", + "llvm.ppc.altivec.dssall" => "__builtin_altivec_dssall", + "llvm.ppc.altivec.dst" => "__builtin_altivec_dst", + "llvm.ppc.altivec.dstst" => "__builtin_altivec_dstst", + "llvm.ppc.altivec.dststt" => "__builtin_altivec_dststt", + "llvm.ppc.altivec.dstt" => "__builtin_altivec_dstt", + "llvm.ppc.altivec.mfvscr" => "__builtin_altivec_mfvscr", + "llvm.ppc.altivec.mtvscr" => "__builtin_altivec_mtvscr", + "llvm.ppc.altivec.mtvsrbm" => "__builtin_altivec_mtvsrbm", + "llvm.ppc.altivec.mtvsrdm" => "__builtin_altivec_mtvsrdm", + "llvm.ppc.altivec.mtvsrhm" => "__builtin_altivec_mtvsrhm", + "llvm.ppc.altivec.mtvsrqm" => "__builtin_altivec_mtvsrqm", + "llvm.ppc.altivec.mtvsrwm" => "__builtin_altivec_mtvsrwm", + "llvm.ppc.altivec.vaddcuw" => "__builtin_altivec_vaddcuw", + "llvm.ppc.altivec.vaddecuq" => "__builtin_altivec_vaddecuq", + "llvm.ppc.altivec.vaddeuqm" => "__builtin_altivec_vaddeuqm", + "llvm.ppc.altivec.vaddsbs" => "__builtin_altivec_vaddsbs", + "llvm.ppc.altivec.vaddshs" => "__builtin_altivec_vaddshs", + "llvm.ppc.altivec.vaddsws" => "__builtin_altivec_vaddsws", + "llvm.ppc.altivec.vaddubs" => "__builtin_altivec_vaddubs", + "llvm.ppc.altivec.vadduhs" => "__builtin_altivec_vadduhs", + "llvm.ppc.altivec.vadduws" => "__builtin_altivec_vadduws", + "llvm.ppc.altivec.vavgsb" => "__builtin_altivec_vavgsb", + "llvm.ppc.altivec.vavgsh" => "__builtin_altivec_vavgsh", + "llvm.ppc.altivec.vavgsw" => "__builtin_altivec_vavgsw", + "llvm.ppc.altivec.vavgub" => "__builtin_altivec_vavgub", + "llvm.ppc.altivec.vavguh" => "__builtin_altivec_vavguh", + "llvm.ppc.altivec.vavguw" => "__builtin_altivec_vavguw", + "llvm.ppc.altivec.vbpermd" => "__builtin_altivec_vbpermd", + "llvm.ppc.altivec.vbpermq" => "__builtin_altivec_vbpermq", + "llvm.ppc.altivec.vcfsx" => "__builtin_altivec_vcfsx", + "llvm.ppc.altivec.vcfuged" => "__builtin_altivec_vcfuged", + "llvm.ppc.altivec.vcfux" => "__builtin_altivec_vcfux", + "llvm.ppc.altivec.vclrlb" => "__builtin_altivec_vclrlb", + "llvm.ppc.altivec.vclrrb" => "__builtin_altivec_vclrrb", + "llvm.ppc.altivec.vclzdm" => "__builtin_altivec_vclzdm", + "llvm.ppc.altivec.vclzlsbb" => "__builtin_altivec_vclzlsbb", + "llvm.ppc.altivec.vcmpbfp" => "__builtin_altivec_vcmpbfp", + "llvm.ppc.altivec.vcmpbfp.p" => "__builtin_altivec_vcmpbfp_p", + "llvm.ppc.altivec.vcmpeqfp" => "__builtin_altivec_vcmpeqfp", + "llvm.ppc.altivec.vcmpeqfp.p" => "__builtin_altivec_vcmpeqfp_p", + "llvm.ppc.altivec.vcmpequb" => "__builtin_altivec_vcmpequb", + "llvm.ppc.altivec.vcmpequb.p" => "__builtin_altivec_vcmpequb_p", + "llvm.ppc.altivec.vcmpequd" => "__builtin_altivec_vcmpequd", + "llvm.ppc.altivec.vcmpequd.p" => "__builtin_altivec_vcmpequd_p", + "llvm.ppc.altivec.vcmpequh" => "__builtin_altivec_vcmpequh", + "llvm.ppc.altivec.vcmpequh.p" => "__builtin_altivec_vcmpequh_p", + "llvm.ppc.altivec.vcmpequq" => "__builtin_altivec_vcmpequq", + "llvm.ppc.altivec.vcmpequq.p" => "__builtin_altivec_vcmpequq_p", + "llvm.ppc.altivec.vcmpequw" => "__builtin_altivec_vcmpequw", + "llvm.ppc.altivec.vcmpequw.p" => "__builtin_altivec_vcmpequw_p", + "llvm.ppc.altivec.vcmpgefp" => "__builtin_altivec_vcmpgefp", + "llvm.ppc.altivec.vcmpgefp.p" => "__builtin_altivec_vcmpgefp_p", + "llvm.ppc.altivec.vcmpgtfp" => "__builtin_altivec_vcmpgtfp", + "llvm.ppc.altivec.vcmpgtfp.p" => "__builtin_altivec_vcmpgtfp_p", + "llvm.ppc.altivec.vcmpgtsb" => "__builtin_altivec_vcmpgtsb", + "llvm.ppc.altivec.vcmpgtsb.p" => "__builtin_altivec_vcmpgtsb_p", + "llvm.ppc.altivec.vcmpgtsd" => "__builtin_altivec_vcmpgtsd", + "llvm.ppc.altivec.vcmpgtsd.p" => "__builtin_altivec_vcmpgtsd_p", + "llvm.ppc.altivec.vcmpgtsh" => "__builtin_altivec_vcmpgtsh", + "llvm.ppc.altivec.vcmpgtsh.p" => "__builtin_altivec_vcmpgtsh_p", + "llvm.ppc.altivec.vcmpgtsq" => "__builtin_altivec_vcmpgtsq", + "llvm.ppc.altivec.vcmpgtsq.p" => "__builtin_altivec_vcmpgtsq_p", + "llvm.ppc.altivec.vcmpgtsw" => "__builtin_altivec_vcmpgtsw", + "llvm.ppc.altivec.vcmpgtsw.p" => "__builtin_altivec_vcmpgtsw_p", + "llvm.ppc.altivec.vcmpgtub" => "__builtin_altivec_vcmpgtub", + "llvm.ppc.altivec.vcmpgtub.p" => "__builtin_altivec_vcmpgtub_p", + "llvm.ppc.altivec.vcmpgtud" => "__builtin_altivec_vcmpgtud", + "llvm.ppc.altivec.vcmpgtud.p" => "__builtin_altivec_vcmpgtud_p", + "llvm.ppc.altivec.vcmpgtuh" => "__builtin_altivec_vcmpgtuh", + "llvm.ppc.altivec.vcmpgtuh.p" => "__builtin_altivec_vcmpgtuh_p", + "llvm.ppc.altivec.vcmpgtuq" => "__builtin_altivec_vcmpgtuq", + "llvm.ppc.altivec.vcmpgtuq.p" => "__builtin_altivec_vcmpgtuq_p", + "llvm.ppc.altivec.vcmpgtuw" => "__builtin_altivec_vcmpgtuw", + "llvm.ppc.altivec.vcmpgtuw.p" => "__builtin_altivec_vcmpgtuw_p", + "llvm.ppc.altivec.vcmpneb" => "__builtin_altivec_vcmpneb", + "llvm.ppc.altivec.vcmpneb.p" => "__builtin_altivec_vcmpneb_p", + "llvm.ppc.altivec.vcmpneh" => "__builtin_altivec_vcmpneh", + "llvm.ppc.altivec.vcmpneh.p" => "__builtin_altivec_vcmpneh_p", + "llvm.ppc.altivec.vcmpnew" => "__builtin_altivec_vcmpnew", + "llvm.ppc.altivec.vcmpnew.p" => "__builtin_altivec_vcmpnew_p", + "llvm.ppc.altivec.vcmpnezb" => "__builtin_altivec_vcmpnezb", + "llvm.ppc.altivec.vcmpnezb.p" => "__builtin_altivec_vcmpnezb_p", + "llvm.ppc.altivec.vcmpnezh" => "__builtin_altivec_vcmpnezh", + "llvm.ppc.altivec.vcmpnezh.p" => "__builtin_altivec_vcmpnezh_p", + "llvm.ppc.altivec.vcmpnezw" => "__builtin_altivec_vcmpnezw", + "llvm.ppc.altivec.vcmpnezw.p" => "__builtin_altivec_vcmpnezw_p", + "llvm.ppc.altivec.vcntmbb" => "__builtin_altivec_vcntmbb", + "llvm.ppc.altivec.vcntmbd" => "__builtin_altivec_vcntmbd", + "llvm.ppc.altivec.vcntmbh" => "__builtin_altivec_vcntmbh", + "llvm.ppc.altivec.vcntmbw" => "__builtin_altivec_vcntmbw", + "llvm.ppc.altivec.vctsxs" => "__builtin_altivec_vctsxs", + "llvm.ppc.altivec.vctuxs" => "__builtin_altivec_vctuxs", + "llvm.ppc.altivec.vctzdm" => "__builtin_altivec_vctzdm", + "llvm.ppc.altivec.vctzlsbb" => "__builtin_altivec_vctzlsbb", + "llvm.ppc.altivec.vexpandbm" => "__builtin_altivec_vexpandbm", + "llvm.ppc.altivec.vexpanddm" => "__builtin_altivec_vexpanddm", + "llvm.ppc.altivec.vexpandhm" => "__builtin_altivec_vexpandhm", + "llvm.ppc.altivec.vexpandqm" => "__builtin_altivec_vexpandqm", + "llvm.ppc.altivec.vexpandwm" => "__builtin_altivec_vexpandwm", + "llvm.ppc.altivec.vexptefp" => "__builtin_altivec_vexptefp", + "llvm.ppc.altivec.vextddvlx" => "__builtin_altivec_vextddvlx", + "llvm.ppc.altivec.vextddvrx" => "__builtin_altivec_vextddvrx", + "llvm.ppc.altivec.vextdubvlx" => "__builtin_altivec_vextdubvlx", + "llvm.ppc.altivec.vextdubvrx" => "__builtin_altivec_vextdubvrx", + "llvm.ppc.altivec.vextduhvlx" => "__builtin_altivec_vextduhvlx", + "llvm.ppc.altivec.vextduhvrx" => "__builtin_altivec_vextduhvrx", + "llvm.ppc.altivec.vextduwvlx" => "__builtin_altivec_vextduwvlx", + "llvm.ppc.altivec.vextduwvrx" => "__builtin_altivec_vextduwvrx", + "llvm.ppc.altivec.vextractbm" => "__builtin_altivec_vextractbm", + "llvm.ppc.altivec.vextractdm" => "__builtin_altivec_vextractdm", + "llvm.ppc.altivec.vextracthm" => "__builtin_altivec_vextracthm", + "llvm.ppc.altivec.vextractqm" => "__builtin_altivec_vextractqm", + "llvm.ppc.altivec.vextractwm" => "__builtin_altivec_vextractwm", + "llvm.ppc.altivec.vextsb2d" => "__builtin_altivec_vextsb2d", + "llvm.ppc.altivec.vextsb2w" => "__builtin_altivec_vextsb2w", + "llvm.ppc.altivec.vextsd2q" => "__builtin_altivec_vextsd2q", + "llvm.ppc.altivec.vextsh2d" => "__builtin_altivec_vextsh2d", + "llvm.ppc.altivec.vextsh2w" => "__builtin_altivec_vextsh2w", + "llvm.ppc.altivec.vextsw2d" => "__builtin_altivec_vextsw2d", + "llvm.ppc.altivec.vgbbd" => "__builtin_altivec_vgbbd", + "llvm.ppc.altivec.vgnb" => "__builtin_altivec_vgnb", + "llvm.ppc.altivec.vinsblx" => "__builtin_altivec_vinsblx", + "llvm.ppc.altivec.vinsbrx" => "__builtin_altivec_vinsbrx", + "llvm.ppc.altivec.vinsbvlx" => "__builtin_altivec_vinsbvlx", + "llvm.ppc.altivec.vinsbvrx" => "__builtin_altivec_vinsbvrx", + "llvm.ppc.altivec.vinsdlx" => "__builtin_altivec_vinsdlx", + "llvm.ppc.altivec.vinsdrx" => "__builtin_altivec_vinsdrx", + "llvm.ppc.altivec.vinshlx" => "__builtin_altivec_vinshlx", + "llvm.ppc.altivec.vinshrx" => "__builtin_altivec_vinshrx", + "llvm.ppc.altivec.vinshvlx" => "__builtin_altivec_vinshvlx", + "llvm.ppc.altivec.vinshvrx" => "__builtin_altivec_vinshvrx", + "llvm.ppc.altivec.vinswlx" => "__builtin_altivec_vinswlx", + "llvm.ppc.altivec.vinswrx" => "__builtin_altivec_vinswrx", + "llvm.ppc.altivec.vinswvlx" => "__builtin_altivec_vinswvlx", + "llvm.ppc.altivec.vinswvrx" => "__builtin_altivec_vinswvrx", + "llvm.ppc.altivec.vlogefp" => "__builtin_altivec_vlogefp", + "llvm.ppc.altivec.vmaddfp" => "__builtin_altivec_vmaddfp", + "llvm.ppc.altivec.vmaxfp" => "__builtin_altivec_vmaxfp", + "llvm.ppc.altivec.vmaxsb" => "__builtin_altivec_vmaxsb", + "llvm.ppc.altivec.vmaxsd" => "__builtin_altivec_vmaxsd", + "llvm.ppc.altivec.vmaxsh" => "__builtin_altivec_vmaxsh", + "llvm.ppc.altivec.vmaxsw" => "__builtin_altivec_vmaxsw", + "llvm.ppc.altivec.vmaxub" => "__builtin_altivec_vmaxub", + "llvm.ppc.altivec.vmaxud" => "__builtin_altivec_vmaxud", + "llvm.ppc.altivec.vmaxuh" => "__builtin_altivec_vmaxuh", + "llvm.ppc.altivec.vmaxuw" => "__builtin_altivec_vmaxuw", + "llvm.ppc.altivec.vmhaddshs" => "__builtin_altivec_vmhaddshs", + "llvm.ppc.altivec.vmhraddshs" => "__builtin_altivec_vmhraddshs", + "llvm.ppc.altivec.vminfp" => "__builtin_altivec_vminfp", + "llvm.ppc.altivec.vminsb" => "__builtin_altivec_vminsb", + "llvm.ppc.altivec.vminsd" => "__builtin_altivec_vminsd", + "llvm.ppc.altivec.vminsh" => "__builtin_altivec_vminsh", + "llvm.ppc.altivec.vminsw" => "__builtin_altivec_vminsw", + "llvm.ppc.altivec.vminub" => "__builtin_altivec_vminub", + "llvm.ppc.altivec.vminud" => "__builtin_altivec_vminud", + "llvm.ppc.altivec.vminuh" => "__builtin_altivec_vminuh", + "llvm.ppc.altivec.vminuw" => "__builtin_altivec_vminuw", + "llvm.ppc.altivec.vmladduhm" => "__builtin_altivec_vmladduhm", + "llvm.ppc.altivec.vmsumcud" => "__builtin_altivec_vmsumcud", + "llvm.ppc.altivec.vmsummbm" => "__builtin_altivec_vmsummbm", + "llvm.ppc.altivec.vmsumshm" => "__builtin_altivec_vmsumshm", + "llvm.ppc.altivec.vmsumshs" => "__builtin_altivec_vmsumshs", + "llvm.ppc.altivec.vmsumubm" => "__builtin_altivec_vmsumubm", + "llvm.ppc.altivec.vmsumudm" => "__builtin_altivec_vmsumudm", + "llvm.ppc.altivec.vmsumuhm" => "__builtin_altivec_vmsumuhm", + "llvm.ppc.altivec.vmsumuhs" => "__builtin_altivec_vmsumuhs", + "llvm.ppc.altivec.vmulesb" => "__builtin_altivec_vmulesb", + "llvm.ppc.altivec.vmulesh" => "__builtin_altivec_vmulesh", + "llvm.ppc.altivec.vmulesw" => "__builtin_altivec_vmulesw", + "llvm.ppc.altivec.vmuleub" => "__builtin_altivec_vmuleub", + "llvm.ppc.altivec.vmuleuh" => "__builtin_altivec_vmuleuh", + "llvm.ppc.altivec.vmuleuw" => "__builtin_altivec_vmuleuw", + "llvm.ppc.altivec.vmulosb" => "__builtin_altivec_vmulosb", + "llvm.ppc.altivec.vmulosh" => "__builtin_altivec_vmulosh", + "llvm.ppc.altivec.vmulosw" => "__builtin_altivec_vmulosw", + "llvm.ppc.altivec.vmuloub" => "__builtin_altivec_vmuloub", + "llvm.ppc.altivec.vmulouh" => "__builtin_altivec_vmulouh", + "llvm.ppc.altivec.vmulouw" => "__builtin_altivec_vmulouw", + "llvm.ppc.altivec.vnmsubfp" => "__builtin_altivec_vnmsubfp", + "llvm.ppc.altivec.vpdepd" => "__builtin_altivec_vpdepd", + "llvm.ppc.altivec.vperm" => "__builtin_altivec_vperm_4si", + "llvm.ppc.altivec.vpextd" => "__builtin_altivec_vpextd", + "llvm.ppc.altivec.vpkpx" => "__builtin_altivec_vpkpx", + "llvm.ppc.altivec.vpksdss" => "__builtin_altivec_vpksdss", + "llvm.ppc.altivec.vpksdus" => "__builtin_altivec_vpksdus", + "llvm.ppc.altivec.vpkshss" => "__builtin_altivec_vpkshss", + "llvm.ppc.altivec.vpkshus" => "__builtin_altivec_vpkshus", + "llvm.ppc.altivec.vpkswss" => "__builtin_altivec_vpkswss", + "llvm.ppc.altivec.vpkswus" => "__builtin_altivec_vpkswus", + "llvm.ppc.altivec.vpkudus" => "__builtin_altivec_vpkudus", + "llvm.ppc.altivec.vpkuhus" => "__builtin_altivec_vpkuhus", + "llvm.ppc.altivec.vpkuwus" => "__builtin_altivec_vpkuwus", + "llvm.ppc.altivec.vprtybd" => "__builtin_altivec_vprtybd", + "llvm.ppc.altivec.vprtybq" => "__builtin_altivec_vprtybq", + "llvm.ppc.altivec.vprtybw" => "__builtin_altivec_vprtybw", + "llvm.ppc.altivec.vrefp" => "__builtin_altivec_vrefp", + "llvm.ppc.altivec.vrfim" => "__builtin_altivec_vrfim", + "llvm.ppc.altivec.vrfin" => "__builtin_altivec_vrfin", + "llvm.ppc.altivec.vrfip" => "__builtin_altivec_vrfip", + "llvm.ppc.altivec.vrfiz" => "__builtin_altivec_vrfiz", + "llvm.ppc.altivec.vrlb" => "__builtin_altivec_vrlb", + "llvm.ppc.altivec.vrld" => "__builtin_altivec_vrld", + "llvm.ppc.altivec.vrlh" => "__builtin_altivec_vrlh", + "llvm.ppc.altivec.vrlw" => "__builtin_altivec_vrlw", + "llvm.ppc.altivec.vrsqrtefp" => "__builtin_altivec_vrsqrtefp", + "llvm.ppc.altivec.vsel" => "__builtin_altivec_vsel_4si", + "llvm.ppc.altivec.vsl" => "__builtin_altivec_vsl", + "llvm.ppc.altivec.vslb" => "__builtin_altivec_vslb", + "llvm.ppc.altivec.vsldbi" => "__builtin_altivec_vsldbi", + "llvm.ppc.altivec.vslh" => "__builtin_altivec_vslh", + "llvm.ppc.altivec.vslo" => "__builtin_altivec_vslo", + "llvm.ppc.altivec.vslw" => "__builtin_altivec_vslw", + "llvm.ppc.altivec.vsr" => "__builtin_altivec_vsr", + "llvm.ppc.altivec.vsrab" => "__builtin_altivec_vsrab", + "llvm.ppc.altivec.vsrah" => "__builtin_altivec_vsrah", + "llvm.ppc.altivec.vsraw" => "__builtin_altivec_vsraw", + "llvm.ppc.altivec.vsrb" => "__builtin_altivec_vsrb", + "llvm.ppc.altivec.vsrdbi" => "__builtin_altivec_vsrdbi", + "llvm.ppc.altivec.vsrh" => "__builtin_altivec_vsrh", + "llvm.ppc.altivec.vsro" => "__builtin_altivec_vsro", + "llvm.ppc.altivec.vsrw" => "__builtin_altivec_vsrw", + "llvm.ppc.altivec.vstribl" => "__builtin_altivec_vstribl", + "llvm.ppc.altivec.vstribl.p" => "__builtin_altivec_vstribl_p", + "llvm.ppc.altivec.vstribr" => "__builtin_altivec_vstribr", + "llvm.ppc.altivec.vstribr.p" => "__builtin_altivec_vstribr_p", + "llvm.ppc.altivec.vstrihl" => "__builtin_altivec_vstrihl", + "llvm.ppc.altivec.vstrihl.p" => "__builtin_altivec_vstrihl_p", + "llvm.ppc.altivec.vstrihr" => "__builtin_altivec_vstrihr", + "llvm.ppc.altivec.vstrihr.p" => "__builtin_altivec_vstrihr_p", + "llvm.ppc.altivec.vsubcuw" => "__builtin_altivec_vsubcuw", + "llvm.ppc.altivec.vsubecuq" => "__builtin_altivec_vsubecuq", + "llvm.ppc.altivec.vsubeuqm" => "__builtin_altivec_vsubeuqm", + "llvm.ppc.altivec.vsubsbs" => "__builtin_altivec_vsubsbs", + "llvm.ppc.altivec.vsubshs" => "__builtin_altivec_vsubshs", + "llvm.ppc.altivec.vsubsws" => "__builtin_altivec_vsubsws", + "llvm.ppc.altivec.vsububs" => "__builtin_altivec_vsububs", + "llvm.ppc.altivec.vsubuhs" => "__builtin_altivec_vsubuhs", + "llvm.ppc.altivec.vsubuws" => "__builtin_altivec_vsubuws", + "llvm.ppc.altivec.vsum2sws" => "__builtin_altivec_vsum2sws", + "llvm.ppc.altivec.vsum4sbs" => "__builtin_altivec_vsum4sbs", + "llvm.ppc.altivec.vsum4shs" => "__builtin_altivec_vsum4shs", + "llvm.ppc.altivec.vsum4ubs" => "__builtin_altivec_vsum4ubs", + "llvm.ppc.altivec.vsumsws" => "__builtin_altivec_vsumsws", + "llvm.ppc.altivec.vupkhpx" => "__builtin_altivec_vupkhpx", + "llvm.ppc.altivec.vupkhsb" => "__builtin_altivec_vupkhsb", + "llvm.ppc.altivec.vupkhsh" => "__builtin_altivec_vupkhsh", + "llvm.ppc.altivec.vupkhsw" => "__builtin_altivec_vupkhsw", + "llvm.ppc.altivec.vupklpx" => "__builtin_altivec_vupklpx", + "llvm.ppc.altivec.vupklsb" => "__builtin_altivec_vupklsb", + "llvm.ppc.altivec.vupklsh" => "__builtin_altivec_vupklsh", + "llvm.ppc.altivec.vupklsw" => "__builtin_altivec_vupklsw", + "llvm.ppc.bcdadd" => "__builtin_ppc_bcdadd", + "llvm.ppc.bcdadd.p" => "__builtin_ppc_bcdadd_p", + "llvm.ppc.bcdsub" => "__builtin_ppc_bcdsub", + "llvm.ppc.bcdsub.p" => "__builtin_ppc_bcdsub_p", + "llvm.ppc.bpermd" => "__builtin_bpermd", + "llvm.ppc.cfuged" => "__builtin_cfuged", + "llvm.ppc.cmpeqb" => "__builtin_ppc_cmpeqb", + "llvm.ppc.cmprb" => "__builtin_ppc_cmprb", + "llvm.ppc.cntlzdm" => "__builtin_cntlzdm", + "llvm.ppc.cnttzdm" => "__builtin_cnttzdm", + "llvm.ppc.compare.exp.eq" => "__builtin_ppc_compare_exp_eq", + "llvm.ppc.compare.exp.gt" => "__builtin_ppc_compare_exp_gt", + "llvm.ppc.compare.exp.lt" => "__builtin_ppc_compare_exp_lt", + "llvm.ppc.compare.exp.uo" => "__builtin_ppc_compare_exp_uo", + "llvm.ppc.darn" => "__builtin_darn", + "llvm.ppc.darn32" => "__builtin_darn_32", + "llvm.ppc.darnraw" => "__builtin_darn_raw", + "llvm.ppc.dcbf" => "__builtin_dcbf", + "llvm.ppc.dcbfl" => "__builtin_ppc_dcbfl", + "llvm.ppc.dcbflp" => "__builtin_ppc_dcbflp", + "llvm.ppc.dcbst" => "__builtin_ppc_dcbst", + "llvm.ppc.dcbt" => "__builtin_ppc_dcbt", + "llvm.ppc.dcbtst" => "__builtin_ppc_dcbtst", + "llvm.ppc.dcbtstt" => "__builtin_ppc_dcbtstt", + "llvm.ppc.dcbtt" => "__builtin_ppc_dcbtt", + "llvm.ppc.dcbz" => "__builtin_ppc_dcbz", + "llvm.ppc.divde" => "__builtin_divde", + "llvm.ppc.divdeu" => "__builtin_divdeu", + "llvm.ppc.divf128.round.to.odd" => "__builtin_divf128_round_to_odd", + "llvm.ppc.divwe" => "__builtin_divwe", + "llvm.ppc.divweu" => "__builtin_divweu", + "llvm.ppc.eieio" => "__builtin_ppc_eieio", + "llvm.ppc.extract.exp" => "__builtin_ppc_extract_exp", + "llvm.ppc.extract.sig" => "__builtin_ppc_extract_sig", + "llvm.ppc.fcfid" => "__builtin_ppc_fcfid", + "llvm.ppc.fcfud" => "__builtin_ppc_fcfud", + "llvm.ppc.fctid" => "__builtin_ppc_fctid", + "llvm.ppc.fctidz" => "__builtin_ppc_fctidz", + "llvm.ppc.fctiw" => "__builtin_ppc_fctiw", + "llvm.ppc.fctiwz" => "__builtin_ppc_fctiwz", + "llvm.ppc.fctudz" => "__builtin_ppc_fctudz", + "llvm.ppc.fctuwz" => "__builtin_ppc_fctuwz", + "llvm.ppc.fmaf128.round.to.odd" => "__builtin_fmaf128_round_to_odd", + "llvm.ppc.fmsub" => "__builtin_ppc_fmsub", + "llvm.ppc.fmsubs" => "__builtin_ppc_fmsubs", + "llvm.ppc.fnmadd" => "__builtin_ppc_fnmadd", + "llvm.ppc.fnmadds" => "__builtin_ppc_fnmadds", + "llvm.ppc.fre" => "__builtin_ppc_fre", + "llvm.ppc.fres" => "__builtin_ppc_fres", + "llvm.ppc.frsqrte" => "__builtin_ppc_frsqrte", + "llvm.ppc.frsqrtes" => "__builtin_ppc_frsqrtes", + "llvm.ppc.fsel" => "__builtin_ppc_fsel", + "llvm.ppc.fsels" => "__builtin_ppc_fsels", + "llvm.ppc.get.texasr" => "__builtin_get_texasr", + "llvm.ppc.get.texasru" => "__builtin_get_texasru", + "llvm.ppc.get.tfhar" => "__builtin_get_tfhar", + "llvm.ppc.get.tfiar" => "__builtin_get_tfiar", + "llvm.ppc.icbt" => "__builtin_ppc_icbt", + "llvm.ppc.insert.exp" => "__builtin_ppc_insert_exp", + "llvm.ppc.iospace.eieio" => "__builtin_ppc_iospace_eieio", + "llvm.ppc.iospace.lwsync" => "__builtin_ppc_iospace_lwsync", + "llvm.ppc.iospace.sync" => "__builtin_ppc_iospace_sync", + "llvm.ppc.isync" => "__builtin_ppc_isync", + "llvm.ppc.load4r" => "__builtin_ppc_load4r", + "llvm.ppc.load8r" => "__builtin_ppc_load8r", + "llvm.ppc.lwsync" => "__builtin_ppc_lwsync", + "llvm.ppc.maddhd" => "__builtin_ppc_maddhd", + "llvm.ppc.maddhdu" => "__builtin_ppc_maddhdu", + "llvm.ppc.maddld" => "__builtin_ppc_maddld", + "llvm.ppc.mfmsr" => "__builtin_ppc_mfmsr", + "llvm.ppc.mftbu" => "__builtin_ppc_mftbu", + "llvm.ppc.mtfsb0" => "__builtin_ppc_mtfsb0", + "llvm.ppc.mtfsb1" => "__builtin_ppc_mtfsb1", + "llvm.ppc.mtfsfi" => "__builtin_ppc_mtfsfi", + "llvm.ppc.mtmsr" => "__builtin_ppc_mtmsr", + "llvm.ppc.mulf128.round.to.odd" => "__builtin_mulf128_round_to_odd", + "llvm.ppc.mulhd" => "__builtin_ppc_mulhd", + "llvm.ppc.mulhdu" => "__builtin_ppc_mulhdu", + "llvm.ppc.mulhw" => "__builtin_ppc_mulhw", + "llvm.ppc.mulhwu" => "__builtin_ppc_mulhwu", + "llvm.ppc.pack.longdouble" => "__builtin_pack_longdouble", + "llvm.ppc.pdepd" => "__builtin_pdepd", + "llvm.ppc.pextd" => "__builtin_pextd", + "llvm.ppc.qpx.qvfabs" => "__builtin_qpx_qvfabs", + "llvm.ppc.qpx.qvfadd" => "__builtin_qpx_qvfadd", + "llvm.ppc.qpx.qvfadds" => "__builtin_qpx_qvfadds", + "llvm.ppc.qpx.qvfcfid" => "__builtin_qpx_qvfcfid", + "llvm.ppc.qpx.qvfcfids" => "__builtin_qpx_qvfcfids", + "llvm.ppc.qpx.qvfcfidu" => "__builtin_qpx_qvfcfidu", + "llvm.ppc.qpx.qvfcfidus" => "__builtin_qpx_qvfcfidus", + "llvm.ppc.qpx.qvfcmpeq" => "__builtin_qpx_qvfcmpeq", + "llvm.ppc.qpx.qvfcmpgt" => "__builtin_qpx_qvfcmpgt", + "llvm.ppc.qpx.qvfcmplt" => "__builtin_qpx_qvfcmplt", + "llvm.ppc.qpx.qvfcpsgn" => "__builtin_qpx_qvfcpsgn", + "llvm.ppc.qpx.qvfctid" => "__builtin_qpx_qvfctid", + "llvm.ppc.qpx.qvfctidu" => "__builtin_qpx_qvfctidu", + "llvm.ppc.qpx.qvfctiduz" => "__builtin_qpx_qvfctiduz", + "llvm.ppc.qpx.qvfctidz" => "__builtin_qpx_qvfctidz", + "llvm.ppc.qpx.qvfctiw" => "__builtin_qpx_qvfctiw", + "llvm.ppc.qpx.qvfctiwu" => "__builtin_qpx_qvfctiwu", + "llvm.ppc.qpx.qvfctiwuz" => "__builtin_qpx_qvfctiwuz", + "llvm.ppc.qpx.qvfctiwz" => "__builtin_qpx_qvfctiwz", + "llvm.ppc.qpx.qvflogical" => "__builtin_qpx_qvflogical", + "llvm.ppc.qpx.qvfmadd" => "__builtin_qpx_qvfmadd", + "llvm.ppc.qpx.qvfmadds" => "__builtin_qpx_qvfmadds", + "llvm.ppc.qpx.qvfmsub" => "__builtin_qpx_qvfmsub", + "llvm.ppc.qpx.qvfmsubs" => "__builtin_qpx_qvfmsubs", + "llvm.ppc.qpx.qvfmul" => "__builtin_qpx_qvfmul", + "llvm.ppc.qpx.qvfmuls" => "__builtin_qpx_qvfmuls", + "llvm.ppc.qpx.qvfnabs" => "__builtin_qpx_qvfnabs", + "llvm.ppc.qpx.qvfneg" => "__builtin_qpx_qvfneg", + "llvm.ppc.qpx.qvfnmadd" => "__builtin_qpx_qvfnmadd", + "llvm.ppc.qpx.qvfnmadds" => "__builtin_qpx_qvfnmadds", + "llvm.ppc.qpx.qvfnmsub" => "__builtin_qpx_qvfnmsub", + "llvm.ppc.qpx.qvfnmsubs" => "__builtin_qpx_qvfnmsubs", + "llvm.ppc.qpx.qvfperm" => "__builtin_qpx_qvfperm", + "llvm.ppc.qpx.qvfre" => "__builtin_qpx_qvfre", + "llvm.ppc.qpx.qvfres" => "__builtin_qpx_qvfres", + "llvm.ppc.qpx.qvfrim" => "__builtin_qpx_qvfrim", + "llvm.ppc.qpx.qvfrin" => "__builtin_qpx_qvfrin", + "llvm.ppc.qpx.qvfrip" => "__builtin_qpx_qvfrip", + "llvm.ppc.qpx.qvfriz" => "__builtin_qpx_qvfriz", + "llvm.ppc.qpx.qvfrsp" => "__builtin_qpx_qvfrsp", + "llvm.ppc.qpx.qvfrsqrte" => "__builtin_qpx_qvfrsqrte", + "llvm.ppc.qpx.qvfrsqrtes" => "__builtin_qpx_qvfrsqrtes", + "llvm.ppc.qpx.qvfsel" => "__builtin_qpx_qvfsel", + "llvm.ppc.qpx.qvfsub" => "__builtin_qpx_qvfsub", + "llvm.ppc.qpx.qvfsubs" => "__builtin_qpx_qvfsubs", + "llvm.ppc.qpx.qvftstnan" => "__builtin_qpx_qvftstnan", + "llvm.ppc.qpx.qvfxmadd" => "__builtin_qpx_qvfxmadd", + "llvm.ppc.qpx.qvfxmadds" => "__builtin_qpx_qvfxmadds", + "llvm.ppc.qpx.qvfxmul" => "__builtin_qpx_qvfxmul", + "llvm.ppc.qpx.qvfxmuls" => "__builtin_qpx_qvfxmuls", + "llvm.ppc.qpx.qvfxxcpnmadd" => "__builtin_qpx_qvfxxcpnmadd", + "llvm.ppc.qpx.qvfxxcpnmadds" => "__builtin_qpx_qvfxxcpnmadds", + "llvm.ppc.qpx.qvfxxmadd" => "__builtin_qpx_qvfxxmadd", + "llvm.ppc.qpx.qvfxxmadds" => "__builtin_qpx_qvfxxmadds", + "llvm.ppc.qpx.qvfxxnpmadd" => "__builtin_qpx_qvfxxnpmadd", + "llvm.ppc.qpx.qvfxxnpmadds" => "__builtin_qpx_qvfxxnpmadds", + "llvm.ppc.qpx.qvgpci" => "__builtin_qpx_qvgpci", + "llvm.ppc.qpx.qvlfcd" => "__builtin_qpx_qvlfcd", + "llvm.ppc.qpx.qvlfcda" => "__builtin_qpx_qvlfcda", + "llvm.ppc.qpx.qvlfcs" => "__builtin_qpx_qvlfcs", + "llvm.ppc.qpx.qvlfcsa" => "__builtin_qpx_qvlfcsa", + "llvm.ppc.qpx.qvlfd" => "__builtin_qpx_qvlfd", + "llvm.ppc.qpx.qvlfda" => "__builtin_qpx_qvlfda", + "llvm.ppc.qpx.qvlfiwa" => "__builtin_qpx_qvlfiwa", + "llvm.ppc.qpx.qvlfiwaa" => "__builtin_qpx_qvlfiwaa", + "llvm.ppc.qpx.qvlfiwz" => "__builtin_qpx_qvlfiwz", + "llvm.ppc.qpx.qvlfiwza" => "__builtin_qpx_qvlfiwza", + "llvm.ppc.qpx.qvlfs" => "__builtin_qpx_qvlfs", + "llvm.ppc.qpx.qvlfsa" => "__builtin_qpx_qvlfsa", + "llvm.ppc.qpx.qvlpcld" => "__builtin_qpx_qvlpcld", + "llvm.ppc.qpx.qvlpcls" => "__builtin_qpx_qvlpcls", + "llvm.ppc.qpx.qvlpcrd" => "__builtin_qpx_qvlpcrd", + "llvm.ppc.qpx.qvlpcrs" => "__builtin_qpx_qvlpcrs", + "llvm.ppc.qpx.qvstfcd" => "__builtin_qpx_qvstfcd", + "llvm.ppc.qpx.qvstfcda" => "__builtin_qpx_qvstfcda", + "llvm.ppc.qpx.qvstfcs" => "__builtin_qpx_qvstfcs", + "llvm.ppc.qpx.qvstfcsa" => "__builtin_qpx_qvstfcsa", + "llvm.ppc.qpx.qvstfd" => "__builtin_qpx_qvstfd", + "llvm.ppc.qpx.qvstfda" => "__builtin_qpx_qvstfda", + "llvm.ppc.qpx.qvstfiw" => "__builtin_qpx_qvstfiw", + "llvm.ppc.qpx.qvstfiwa" => "__builtin_qpx_qvstfiwa", + "llvm.ppc.qpx.qvstfs" => "__builtin_qpx_qvstfs", + "llvm.ppc.qpx.qvstfsa" => "__builtin_qpx_qvstfsa", + "llvm.ppc.readflm" => "__builtin_readflm", + "llvm.ppc.scalar.extract.expq" => "__builtin_vsx_scalar_extract_expq", + "llvm.ppc.scalar.insert.exp.qp" => "__builtin_vsx_scalar_insert_exp_qp", + "llvm.ppc.set.texasr" => "__builtin_set_texasr", + "llvm.ppc.set.texasru" => "__builtin_set_texasru", + "llvm.ppc.set.tfhar" => "__builtin_set_tfhar", + "llvm.ppc.set.tfiar" => "__builtin_set_tfiar", + "llvm.ppc.setb" => "__builtin_ppc_setb", + "llvm.ppc.setflm" => "__builtin_setflm", + "llvm.ppc.setrnd" => "__builtin_setrnd", + "llvm.ppc.sqrtf128.round.to.odd" => "__builtin_sqrtf128_round_to_odd", + "llvm.ppc.stbcx" => "__builtin_ppc_stbcx", + "llvm.ppc.stdcx" => "__builtin_ppc_stdcx", + "llvm.ppc.stfiw" => "__builtin_ppc_stfiw", + "llvm.ppc.store2r" => "__builtin_ppc_store2r", + "llvm.ppc.store4r" => "__builtin_ppc_store4r", + "llvm.ppc.store8r" => "__builtin_ppc_store8r", + "llvm.ppc.stwcx" => "__builtin_ppc_stwcx", + "llvm.ppc.subf128.round.to.odd" => "__builtin_subf128_round_to_odd", + "llvm.ppc.sync" => "__builtin_ppc_sync", + "llvm.ppc.tabort" => "__builtin_tabort", + "llvm.ppc.tabortdc" => "__builtin_tabortdc", + "llvm.ppc.tabortdci" => "__builtin_tabortdci", + "llvm.ppc.tabortwc" => "__builtin_tabortwc", + "llvm.ppc.tabortwci" => "__builtin_tabortwci", + "llvm.ppc.tbegin" => "__builtin_tbegin", + "llvm.ppc.tcheck" => "__builtin_tcheck", + "llvm.ppc.tdw" => "__builtin_ppc_tdw", + "llvm.ppc.tend" => "__builtin_tend", + "llvm.ppc.tendall" => "__builtin_tendall", + "llvm.ppc.trap" => "__builtin_ppc_trap", + "llvm.ppc.trapd" => "__builtin_ppc_trapd", + "llvm.ppc.trechkpt" => "__builtin_trechkpt", + "llvm.ppc.treclaim" => "__builtin_treclaim", + "llvm.ppc.tresume" => "__builtin_tresume", + "llvm.ppc.truncf128.round.to.odd" => "__builtin_truncf128_round_to_odd", + "llvm.ppc.tsr" => "__builtin_tsr", + "llvm.ppc.tsuspend" => "__builtin_tsuspend", + "llvm.ppc.ttest" => "__builtin_ttest", + "llvm.ppc.tw" => "__builtin_ppc_tw", + "llvm.ppc.unpack.longdouble" => "__builtin_unpack_longdouble", + "llvm.ppc.vsx.xsmaxdp" => "__builtin_vsx_xsmaxdp", + "llvm.ppc.vsx.xsmindp" => "__builtin_vsx_xsmindp", + "llvm.ppc.vsx.xvcmpeqdp" => "__builtin_vsx_xvcmpeqdp", + "llvm.ppc.vsx.xvcmpeqdp.p" => "__builtin_vsx_xvcmpeqdp_p", + "llvm.ppc.vsx.xvcmpeqsp" => "__builtin_vsx_xvcmpeqsp", + "llvm.ppc.vsx.xvcmpeqsp.p" => "__builtin_vsx_xvcmpeqsp_p", + "llvm.ppc.vsx.xvcmpgedp" => "__builtin_vsx_xvcmpgedp", + "llvm.ppc.vsx.xvcmpgedp.p" => "__builtin_vsx_xvcmpgedp_p", + "llvm.ppc.vsx.xvcmpgesp" => "__builtin_vsx_xvcmpgesp", + "llvm.ppc.vsx.xvcmpgesp.p" => "__builtin_vsx_xvcmpgesp_p", + "llvm.ppc.vsx.xvcmpgtdp" => "__builtin_vsx_xvcmpgtdp", + "llvm.ppc.vsx.xvcmpgtdp.p" => "__builtin_vsx_xvcmpgtdp_p", + "llvm.ppc.vsx.xvcmpgtsp" => "__builtin_vsx_xvcmpgtsp", + "llvm.ppc.vsx.xvcmpgtsp.p" => "__builtin_vsx_xvcmpgtsp_p", + "llvm.ppc.vsx.xvdivdp" => "__builtin_vsx_xvdivdp", + "llvm.ppc.vsx.xvdivsp" => "__builtin_vsx_xvdivsp", + "llvm.ppc.vsx.xvmaxdp" => "__builtin_vsx_xvmaxdp", + "llvm.ppc.vsx.xvmaxsp" => "__builtin_vsx_xvmaxsp", + "llvm.ppc.vsx.xvmindp" => "__builtin_vsx_xvmindp", + "llvm.ppc.vsx.xvminsp" => "__builtin_vsx_xvminsp", + "llvm.ppc.vsx.xvredp" => "__builtin_vsx_xvredp", + "llvm.ppc.vsx.xvresp" => "__builtin_vsx_xvresp", + "llvm.ppc.vsx.xvrsqrtedp" => "__builtin_vsx_xvrsqrtedp", + "llvm.ppc.vsx.xvrsqrtesp" => "__builtin_vsx_xvrsqrtesp", + "llvm.ppc.vsx.xxblendvb" => "__builtin_vsx_xxblendvb", + "llvm.ppc.vsx.xxblendvd" => "__builtin_vsx_xxblendvd", + "llvm.ppc.vsx.xxblendvh" => "__builtin_vsx_xxblendvh", + "llvm.ppc.vsx.xxblendvw" => "__builtin_vsx_xxblendvw", + "llvm.ppc.vsx.xxleqv" => "__builtin_vsx_xxleqv", + "llvm.ppc.vsx.xxpermx" => "__builtin_vsx_xxpermx", + // ptx + "llvm.ptx.bar.sync" => "__builtin_ptx_bar_sync", + "llvm.ptx.read.clock" => "__builtin_ptx_read_clock", + "llvm.ptx.read.clock64" => "__builtin_ptx_read_clock64", + "llvm.ptx.read.gridid" => "__builtin_ptx_read_gridid", + "llvm.ptx.read.laneid" => "__builtin_ptx_read_laneid", + "llvm.ptx.read.lanemask.eq" => "__builtin_ptx_read_lanemask_eq", + "llvm.ptx.read.lanemask.ge" => "__builtin_ptx_read_lanemask_ge", + "llvm.ptx.read.lanemask.gt" => "__builtin_ptx_read_lanemask_gt", + "llvm.ptx.read.lanemask.le" => "__builtin_ptx_read_lanemask_le", + "llvm.ptx.read.lanemask.lt" => "__builtin_ptx_read_lanemask_lt", + "llvm.ptx.read.nsmid" => "__builtin_ptx_read_nsmid", + "llvm.ptx.read.nwarpid" => "__builtin_ptx_read_nwarpid", + "llvm.ptx.read.pm0" => "__builtin_ptx_read_pm0", + "llvm.ptx.read.pm1" => "__builtin_ptx_read_pm1", + "llvm.ptx.read.pm2" => "__builtin_ptx_read_pm2", + "llvm.ptx.read.pm3" => "__builtin_ptx_read_pm3", + "llvm.ptx.read.smid" => "__builtin_ptx_read_smid", + "llvm.ptx.read.warpid" => "__builtin_ptx_read_warpid", + // s390 + "llvm.s390.efpc" => "__builtin_s390_efpc", + "llvm.s390.etnd" => "__builtin_tx_nesting_depth", + "llvm.s390.lcbb" => "__builtin_s390_lcbb", + "llvm.s390.ppa.txassist" => "__builtin_tx_assist", + "llvm.s390.sfpc" => "__builtin_s390_sfpc", + "llvm.s390.tend" => "__builtin_tend", + "llvm.s390.vcfn" => "__builtin_s390_vcfn", + "llvm.s390.vclfnhs" => "__builtin_s390_vclfnhs", + "llvm.s390.vclfnls" => "__builtin_s390_vclfnls", + "llvm.s390.vcnf" => "__builtin_s390_vcnf", + "llvm.s390.vcrnfs" => "__builtin_s390_vcrnfs", + "llvm.s390.vlbb" => "__builtin_s390_vlbb", + "llvm.s390.vll" => "__builtin_s390_vll", + "llvm.s390.vlrl" => "__builtin_s390_vlrl", + "llvm.s390.vmslg" => "__builtin_s390_vmslg", + "llvm.s390.vpdi" => "__builtin_s390_vpdi", + "llvm.s390.vperm" => "__builtin_s390_vperm", + "llvm.s390.vsld" => "__builtin_s390_vsld", + "llvm.s390.vsldb" => "__builtin_s390_vsldb", + "llvm.s390.vsrd" => "__builtin_s390_vsrd", + "llvm.s390.vstl" => "__builtin_s390_vstl", + "llvm.s390.vstrl" => "__builtin_s390_vstrl", + // ve + "llvm.ve.vl.extract.vm512l" => "__builtin_ve_vl_extract_vm512l", + "llvm.ve.vl.extract.vm512u" => "__builtin_ve_vl_extract_vm512u", + "llvm.ve.vl.insert.vm512l" => "__builtin_ve_vl_insert_vm512l", + "llvm.ve.vl.insert.vm512u" => "__builtin_ve_vl_insert_vm512u", + "llvm.ve.vl.pack.f32a" => "__builtin_ve_vl_pack_f32a", + "llvm.ve.vl.pack.f32p" => "__builtin_ve_vl_pack_f32p", + // x86 + "llvm.x86.3dnow.pavgusb" => "__builtin_ia32_pavgusb", + "llvm.x86.3dnow.pf2id" => "__builtin_ia32_pf2id", + "llvm.x86.3dnow.pfacc" => "__builtin_ia32_pfacc", + "llvm.x86.3dnow.pfadd" => "__builtin_ia32_pfadd", + "llvm.x86.3dnow.pfcmpeq" => "__builtin_ia32_pfcmpeq", + "llvm.x86.3dnow.pfcmpge" => "__builtin_ia32_pfcmpge", + "llvm.x86.3dnow.pfcmpgt" => "__builtin_ia32_pfcmpgt", + "llvm.x86.3dnow.pfmax" => "__builtin_ia32_pfmax", + "llvm.x86.3dnow.pfmin" => "__builtin_ia32_pfmin", + "llvm.x86.3dnow.pfmul" => "__builtin_ia32_pfmul", + "llvm.x86.3dnow.pfrcp" => "__builtin_ia32_pfrcp", + "llvm.x86.3dnow.pfrcpit1" => "__builtin_ia32_pfrcpit1", + "llvm.x86.3dnow.pfrcpit2" => "__builtin_ia32_pfrcpit2", + "llvm.x86.3dnow.pfrsqit1" => "__builtin_ia32_pfrsqit1", + "llvm.x86.3dnow.pfrsqrt" => "__builtin_ia32_pfrsqrt", + "llvm.x86.3dnow.pfsub" => "__builtin_ia32_pfsub", + "llvm.x86.3dnow.pfsubr" => "__builtin_ia32_pfsubr", + "llvm.x86.3dnow.pi2fd" => "__builtin_ia32_pi2fd", + "llvm.x86.3dnow.pmulhrw" => "__builtin_ia32_pmulhrw", + "llvm.x86.3dnowa.pf2iw" => "__builtin_ia32_pf2iw", + "llvm.x86.3dnowa.pfnacc" => "__builtin_ia32_pfnacc", + "llvm.x86.3dnowa.pfpnacc" => "__builtin_ia32_pfpnacc", + "llvm.x86.3dnowa.pi2fw" => "__builtin_ia32_pi2fw", + "llvm.x86.addcarry.u32" => "__builtin_ia32_addcarry_u32", + "llvm.x86.addcarry.u64" => "__builtin_ia32_addcarry_u64", + "llvm.x86.addcarryx.u32" => "__builtin_ia32_addcarryx_u32", + "llvm.x86.addcarryx.u64" => "__builtin_ia32_addcarryx_u64", + "llvm.x86.aesni.aesdec" => "__builtin_ia32_aesdec128", + "llvm.x86.aesni.aesdec.256" => "__builtin_ia32_aesdec256", + "llvm.x86.aesni.aesdec.512" => "__builtin_ia32_aesdec512", + "llvm.x86.aesni.aesdeclast" => "__builtin_ia32_aesdeclast128", + "llvm.x86.aesni.aesdeclast.256" => "__builtin_ia32_aesdeclast256", + "llvm.x86.aesni.aesdeclast.512" => "__builtin_ia32_aesdeclast512", + "llvm.x86.aesni.aesenc" => "__builtin_ia32_aesenc128", + "llvm.x86.aesni.aesenc.256" => "__builtin_ia32_aesenc256", + "llvm.x86.aesni.aesenc.512" => "__builtin_ia32_aesenc512", + "llvm.x86.aesni.aesenclast" => "__builtin_ia32_aesenclast128", + "llvm.x86.aesni.aesenclast.256" => "__builtin_ia32_aesenclast256", + "llvm.x86.aesni.aesenclast.512" => "__builtin_ia32_aesenclast512", + "llvm.x86.aesni.aesimc" => "__builtin_ia32_aesimc128", + "llvm.x86.aesni.aeskeygenassist" => "__builtin_ia32_aeskeygenassist128", + "llvm.x86.avx.addsub.pd.256" => "__builtin_ia32_addsubpd256", + "llvm.x86.avx.addsub.ps.256" => "__builtin_ia32_addsubps256", + "llvm.x86.avx.blend.pd.256" => "__builtin_ia32_blendpd256", + "llvm.x86.avx.blend.ps.256" => "__builtin_ia32_blendps256", + "llvm.x86.avx.blendv.pd.256" => "__builtin_ia32_blendvpd256", + "llvm.x86.avx.blendv.ps.256" => "__builtin_ia32_blendvps256", + "llvm.x86.avx.cmp.pd.256" => "__builtin_ia32_cmppd256", + "llvm.x86.avx.cmp.ps.256" => "__builtin_ia32_cmpps256", + "llvm.x86.avx.cvt.pd2.ps.256" => "__builtin_ia32_cvtpd2ps256", + "llvm.x86.avx.cvt.pd2dq.256" => "__builtin_ia32_cvtpd2dq256", + "llvm.x86.avx.cvt.ps2.pd.256" => "__builtin_ia32_cvtps2pd256", + "llvm.x86.avx.cvt.ps2dq.256" => "__builtin_ia32_cvtps2dq256", + "llvm.x86.avx.cvtdq2.pd.256" => "__builtin_ia32_cvtdq2pd256", + "llvm.x86.avx.cvtdq2.ps.256" => "__builtin_ia32_cvtdq2ps256", + "llvm.x86.avx.cvtt.pd2dq.256" => "__builtin_ia32_cvttpd2dq256", + "llvm.x86.avx.cvtt.ps2dq.256" => "__builtin_ia32_cvttps2dq256", + "llvm.x86.avx.dp.ps.256" => "__builtin_ia32_dpps256", + "llvm.x86.avx.hadd.pd.256" => "__builtin_ia32_haddpd256", + "llvm.x86.avx.hadd.ps.256" => "__builtin_ia32_haddps256", + "llvm.x86.avx.hsub.pd.256" => "__builtin_ia32_hsubpd256", + "llvm.x86.avx.hsub.ps.256" => "__builtin_ia32_hsubps256", + "llvm.x86.avx.ldu.dq.256" => "__builtin_ia32_lddqu256", + "llvm.x86.avx.maskload.pd" => "__builtin_ia32_maskloadpd", + "llvm.x86.avx.maskload.pd.256" => "__builtin_ia32_maskloadpd256", + "llvm.x86.avx.maskload.ps" => "__builtin_ia32_maskloadps", + "llvm.x86.avx.maskload.ps.256" => "__builtin_ia32_maskloadps256", + "llvm.x86.avx.maskstore.pd" => "__builtin_ia32_maskstorepd", + "llvm.x86.avx.maskstore.pd.256" => "__builtin_ia32_maskstorepd256", + "llvm.x86.avx.maskstore.ps" => "__builtin_ia32_maskstoreps", + "llvm.x86.avx.maskstore.ps.256" => "__builtin_ia32_maskstoreps256", + "llvm.x86.avx.max.pd.256" => "__builtin_ia32_maxpd256", + "llvm.x86.avx.max.ps.256" => "__builtin_ia32_maxps256", + "llvm.x86.avx.min.pd.256" => "__builtin_ia32_minpd256", + "llvm.x86.avx.min.ps.256" => "__builtin_ia32_minps256", + "llvm.x86.avx.movmsk.pd.256" => "__builtin_ia32_movmskpd256", + "llvm.x86.avx.movmsk.ps.256" => "__builtin_ia32_movmskps256", + "llvm.x86.avx.ptestc.256" => "__builtin_ia32_ptestc256", + "llvm.x86.avx.ptestnzc.256" => "__builtin_ia32_ptestnzc256", + "llvm.x86.avx.ptestz.256" => "__builtin_ia32_ptestz256", + "llvm.x86.avx.rcp.ps.256" => "__builtin_ia32_rcpps256", + "llvm.x86.avx.round.pd.256" => "__builtin_ia32_roundpd256", + "llvm.x86.avx.round.ps.256" => "__builtin_ia32_roundps256", + "llvm.x86.avx.rsqrt.ps.256" => "__builtin_ia32_rsqrtps256", + "llvm.x86.avx.sqrt.pd.256" => "__builtin_ia32_sqrtpd256", + "llvm.x86.avx.sqrt.ps.256" => "__builtin_ia32_sqrtps256", + "llvm.x86.avx.storeu.dq.256" => "__builtin_ia32_storedqu256", + "llvm.x86.avx.storeu.pd.256" => "__builtin_ia32_storeupd256", + "llvm.x86.avx.storeu.ps.256" => "__builtin_ia32_storeups256", + "llvm.x86.avx.vbroadcastf128.pd.256" => "__builtin_ia32_vbroadcastf128_pd256", + "llvm.x86.avx.vbroadcastf128.ps.256" => "__builtin_ia32_vbroadcastf128_ps256", + "llvm.x86.avx.vextractf128.pd.256" => "__builtin_ia32_vextractf128_pd256", + "llvm.x86.avx.vextractf128.ps.256" => "__builtin_ia32_vextractf128_ps256", + "llvm.x86.avx.vextractf128.si.256" => "__builtin_ia32_vextractf128_si256", + "llvm.x86.avx.vinsertf128.pd.256" => "__builtin_ia32_vinsertf128_pd256", + "llvm.x86.avx.vinsertf128.ps.256" => "__builtin_ia32_vinsertf128_ps256", + "llvm.x86.avx.vinsertf128.si.256" => "__builtin_ia32_vinsertf128_si256", + "llvm.x86.avx.vperm2f128.pd.256" => "__builtin_ia32_vperm2f128_pd256", + "llvm.x86.avx.vperm2f128.ps.256" => "__builtin_ia32_vperm2f128_ps256", + "llvm.x86.avx.vperm2f128.si.256" => "__builtin_ia32_vperm2f128_si256", + "llvm.x86.avx.vpermilvar.pd" => "__builtin_ia32_vpermilvarpd", + "llvm.x86.avx.vpermilvar.pd.256" => "__builtin_ia32_vpermilvarpd256", + "llvm.x86.avx.vpermilvar.ps" => "__builtin_ia32_vpermilvarps", + "llvm.x86.avx.vpermilvar.ps.256" => "__builtin_ia32_vpermilvarps256", + "llvm.x86.avx.vtestc.pd" => "__builtin_ia32_vtestcpd", + "llvm.x86.avx.vtestc.pd.256" => "__builtin_ia32_vtestcpd256", + "llvm.x86.avx.vtestc.ps" => "__builtin_ia32_vtestcps", + "llvm.x86.avx.vtestc.ps.256" => "__builtin_ia32_vtestcps256", + "llvm.x86.avx.vtestnzc.pd" => "__builtin_ia32_vtestnzcpd", + "llvm.x86.avx.vtestnzc.pd.256" => "__builtin_ia32_vtestnzcpd256", + "llvm.x86.avx.vtestnzc.ps" => "__builtin_ia32_vtestnzcps", + "llvm.x86.avx.vtestnzc.ps.256" => "__builtin_ia32_vtestnzcps256", + "llvm.x86.avx.vtestz.pd" => "__builtin_ia32_vtestzpd", + "llvm.x86.avx.vtestz.pd.256" => "__builtin_ia32_vtestzpd256", + "llvm.x86.avx.vtestz.ps" => "__builtin_ia32_vtestzps", + "llvm.x86.avx.vtestz.ps.256" => "__builtin_ia32_vtestzps256", + "llvm.x86.avx.vzeroall" => "__builtin_ia32_vzeroall", + "llvm.x86.avx.vzeroupper" => "__builtin_ia32_vzeroupper", + "llvm.x86.avx2.gather.d.d" => "__builtin_ia32_gatherd_d", + "llvm.x86.avx2.gather.d.d.256" => "__builtin_ia32_gatherd_d256", + "llvm.x86.avx2.gather.d.pd" => "__builtin_ia32_gatherd_pd", + "llvm.x86.avx2.gather.d.pd.256" => "__builtin_ia32_gatherd_pd256", + "llvm.x86.avx2.gather.d.ps" => "__builtin_ia32_gatherd_ps", + "llvm.x86.avx2.gather.d.ps.256" => "__builtin_ia32_gatherd_ps256", + "llvm.x86.avx2.gather.d.q" => "__builtin_ia32_gatherd_q", + "llvm.x86.avx2.gather.d.q.256" => "__builtin_ia32_gatherd_q256", + "llvm.x86.avx2.gather.q.d" => "__builtin_ia32_gatherq_d", + "llvm.x86.avx2.gather.q.d.256" => "__builtin_ia32_gatherq_d256", + "llvm.x86.avx2.gather.q.pd" => "__builtin_ia32_gatherq_pd", + "llvm.x86.avx2.gather.q.pd.256" => "__builtin_ia32_gatherq_pd256", + "llvm.x86.avx2.gather.q.ps" => "__builtin_ia32_gatherq_ps", + "llvm.x86.avx2.gather.q.ps.256" => "__builtin_ia32_gatherq_ps256", + "llvm.x86.avx2.gather.q.q" => "__builtin_ia32_gatherq_q", + "llvm.x86.avx2.gather.q.q.256" => "__builtin_ia32_gatherq_q256", + "llvm.x86.avx2.maskload.d" => "__builtin_ia32_maskloadd", + "llvm.x86.avx2.maskload.d.256" => "__builtin_ia32_maskloadd256", + "llvm.x86.avx2.maskload.q" => "__builtin_ia32_maskloadq", + "llvm.x86.avx2.maskload.q.256" => "__builtin_ia32_maskloadq256", + "llvm.x86.avx2.maskstore.d" => "__builtin_ia32_maskstored", + "llvm.x86.avx2.maskstore.d.256" => "__builtin_ia32_maskstored256", + "llvm.x86.avx2.maskstore.q" => "__builtin_ia32_maskstoreq", + "llvm.x86.avx2.maskstore.q.256" => "__builtin_ia32_maskstoreq256", + "llvm.x86.avx2.movntdqa" => "__builtin_ia32_movntdqa256", + "llvm.x86.avx2.mpsadbw" => "__builtin_ia32_mpsadbw256", + "llvm.x86.avx2.pabs.b" => "__builtin_ia32_pabsb256", + "llvm.x86.avx2.pabs.d" => "__builtin_ia32_pabsd256", + "llvm.x86.avx2.pabs.w" => "__builtin_ia32_pabsw256", + "llvm.x86.avx2.packssdw" => "__builtin_ia32_packssdw256", + "llvm.x86.avx2.packsswb" => "__builtin_ia32_packsswb256", + "llvm.x86.avx2.packusdw" => "__builtin_ia32_packusdw256", + "llvm.x86.avx2.packuswb" => "__builtin_ia32_packuswb256", + "llvm.x86.avx2.padds.b" => "__builtin_ia32_paddsb256", + "llvm.x86.avx2.padds.w" => "__builtin_ia32_paddsw256", + "llvm.x86.avx2.paddus.b" => "__builtin_ia32_paddusb256", + "llvm.x86.avx2.paddus.w" => "__builtin_ia32_paddusw256", + "llvm.x86.avx2.pavg.b" => "__builtin_ia32_pavgb256", + "llvm.x86.avx2.pavg.w" => "__builtin_ia32_pavgw256", + "llvm.x86.avx2.pblendd.128" => "__builtin_ia32_pblendd128", + "llvm.x86.avx2.pblendd.256" => "__builtin_ia32_pblendd256", + "llvm.x86.avx2.pblendvb" => "__builtin_ia32_pblendvb256", + "llvm.x86.avx2.pblendw" => "__builtin_ia32_pblendw256", + "llvm.x86.avx2.pbroadcastb.128" => "__builtin_ia32_pbroadcastb128", + "llvm.x86.avx2.pbroadcastb.256" => "__builtin_ia32_pbroadcastb256", + "llvm.x86.avx2.pbroadcastd.128" => "__builtin_ia32_pbroadcastd128", + "llvm.x86.avx2.pbroadcastd.256" => "__builtin_ia32_pbroadcastd256", + "llvm.x86.avx2.pbroadcastq.128" => "__builtin_ia32_pbroadcastq128", + "llvm.x86.avx2.pbroadcastq.256" => "__builtin_ia32_pbroadcastq256", + "llvm.x86.avx2.pbroadcastw.128" => "__builtin_ia32_pbroadcastw128", + "llvm.x86.avx2.pbroadcastw.256" => "__builtin_ia32_pbroadcastw256", + "llvm.x86.avx2.permd" => "__builtin_ia32_permvarsi256", + "llvm.x86.avx2.permps" => "__builtin_ia32_permvarsf256", + "llvm.x86.avx2.phadd.d" => "__builtin_ia32_phaddd256", + "llvm.x86.avx2.phadd.sw" => "__builtin_ia32_phaddsw256", + "llvm.x86.avx2.phadd.w" => "__builtin_ia32_phaddw256", + "llvm.x86.avx2.phsub.d" => "__builtin_ia32_phsubd256", + "llvm.x86.avx2.phsub.sw" => "__builtin_ia32_phsubsw256", + "llvm.x86.avx2.phsub.w" => "__builtin_ia32_phsubw256", + "llvm.x86.avx2.pmadd.ub.sw" => "__builtin_ia32_pmaddubsw256", + "llvm.x86.avx2.pmadd.wd" => "__builtin_ia32_pmaddwd256", + "llvm.x86.avx2.pmaxs.b" => "__builtin_ia32_pmaxsb256", + "llvm.x86.avx2.pmaxs.d" => "__builtin_ia32_pmaxsd256", + "llvm.x86.avx2.pmaxs.w" => "__builtin_ia32_pmaxsw256", + "llvm.x86.avx2.pmaxu.b" => "__builtin_ia32_pmaxub256", + "llvm.x86.avx2.pmaxu.d" => "__builtin_ia32_pmaxud256", + "llvm.x86.avx2.pmaxu.w" => "__builtin_ia32_pmaxuw256", + "llvm.x86.avx2.pmins.b" => "__builtin_ia32_pminsb256", + "llvm.x86.avx2.pmins.d" => "__builtin_ia32_pminsd256", + "llvm.x86.avx2.pmins.w" => "__builtin_ia32_pminsw256", + "llvm.x86.avx2.pminu.b" => "__builtin_ia32_pminub256", + "llvm.x86.avx2.pminu.d" => "__builtin_ia32_pminud256", + "llvm.x86.avx2.pminu.w" => "__builtin_ia32_pminuw256", + "llvm.x86.avx2.pmovmskb" => "__builtin_ia32_pmovmskb256", + "llvm.x86.avx2.pmovsxbd" => "__builtin_ia32_pmovsxbd256", + "llvm.x86.avx2.pmovsxbq" => "__builtin_ia32_pmovsxbq256", + "llvm.x86.avx2.pmovsxbw" => "__builtin_ia32_pmovsxbw256", + "llvm.x86.avx2.pmovsxdq" => "__builtin_ia32_pmovsxdq256", + "llvm.x86.avx2.pmovsxwd" => "__builtin_ia32_pmovsxwd256", + "llvm.x86.avx2.pmovsxwq" => "__builtin_ia32_pmovsxwq256", + "llvm.x86.avx2.pmovzxbd" => "__builtin_ia32_pmovzxbd256", + "llvm.x86.avx2.pmovzxbq" => "__builtin_ia32_pmovzxbq256", + "llvm.x86.avx2.pmovzxbw" => "__builtin_ia32_pmovzxbw256", + "llvm.x86.avx2.pmovzxdq" => "__builtin_ia32_pmovzxdq256", + "llvm.x86.avx2.pmovzxwd" => "__builtin_ia32_pmovzxwd256", + "llvm.x86.avx2.pmovzxwq" => "__builtin_ia32_pmovzxwq256", + "llvm.x86.avx2.pmul.dq" => "__builtin_ia32_pmuldq256", + "llvm.x86.avx2.pmul.hr.sw" => "__builtin_ia32_pmulhrsw256", + "llvm.x86.avx2.pmulh.w" => "__builtin_ia32_pmulhw256", + "llvm.x86.avx2.pmulhu.w" => "__builtin_ia32_pmulhuw256", + "llvm.x86.avx2.pmulu.dq" => "__builtin_ia32_pmuludq256", + "llvm.x86.avx2.psad.bw" => "__builtin_ia32_psadbw256", + "llvm.x86.avx2.pshuf.b" => "__builtin_ia32_pshufb256", + "llvm.x86.avx2.psign.b" => "__builtin_ia32_psignb256", + "llvm.x86.avx2.psign.d" => "__builtin_ia32_psignd256", + "llvm.x86.avx2.psign.w" => "__builtin_ia32_psignw256", + "llvm.x86.avx2.psll.d" => "__builtin_ia32_pslld256", + "llvm.x86.avx2.psll.dq" => "__builtin_ia32_pslldqi256", + "llvm.x86.avx2.psll.dq.bs" => "__builtin_ia32_pslldqi256_byteshift", + "llvm.x86.avx2.psll.q" => "__builtin_ia32_psllq256", + "llvm.x86.avx2.psll.w" => "__builtin_ia32_psllw256", + "llvm.x86.avx2.pslli.d" => "__builtin_ia32_pslldi256", + "llvm.x86.avx2.pslli.q" => "__builtin_ia32_psllqi256", + "llvm.x86.avx2.pslli.w" => "__builtin_ia32_psllwi256", + "llvm.x86.avx2.psllv.d" => "__builtin_ia32_psllv4si", + "llvm.x86.avx2.psllv.d.256" => "__builtin_ia32_psllv8si", + "llvm.x86.avx2.psllv.q" => "__builtin_ia32_psllv2di", + "llvm.x86.avx2.psllv.q.256" => "__builtin_ia32_psllv4di", + "llvm.x86.avx2.psra.d" => "__builtin_ia32_psrad256", + "llvm.x86.avx2.psra.w" => "__builtin_ia32_psraw256", + "llvm.x86.avx2.psrai.d" => "__builtin_ia32_psradi256", + "llvm.x86.avx2.psrai.w" => "__builtin_ia32_psrawi256", + "llvm.x86.avx2.psrav.d" => "__builtin_ia32_psrav4si", + "llvm.x86.avx2.psrav.d.256" => "__builtin_ia32_psrav8si", + "llvm.x86.avx2.psrl.d" => "__builtin_ia32_psrld256", + "llvm.x86.avx2.psrl.dq" => "__builtin_ia32_psrldqi256", + "llvm.x86.avx2.psrl.dq.bs" => "__builtin_ia32_psrldqi256_byteshift", + "llvm.x86.avx2.psrl.q" => "__builtin_ia32_psrlq256", + "llvm.x86.avx2.psrl.w" => "__builtin_ia32_psrlw256", + "llvm.x86.avx2.psrli.d" => "__builtin_ia32_psrldi256", + "llvm.x86.avx2.psrli.q" => "__builtin_ia32_psrlqi256", + "llvm.x86.avx2.psrli.w" => "__builtin_ia32_psrlwi256", + "llvm.x86.avx2.psrlv.d" => "__builtin_ia32_psrlv4si", + "llvm.x86.avx2.psrlv.d.256" => "__builtin_ia32_psrlv8si", + "llvm.x86.avx2.psrlv.q" => "__builtin_ia32_psrlv2di", + "llvm.x86.avx2.psrlv.q.256" => "__builtin_ia32_psrlv4di", + "llvm.x86.avx2.psubs.b" => "__builtin_ia32_psubsb256", + "llvm.x86.avx2.psubs.w" => "__builtin_ia32_psubsw256", + "llvm.x86.avx2.psubus.b" => "__builtin_ia32_psubusb256", + "llvm.x86.avx2.psubus.w" => "__builtin_ia32_psubusw256", + "llvm.x86.avx2.vbroadcast.sd.pd.256" => "__builtin_ia32_vbroadcastsd_pd256", + "llvm.x86.avx2.vbroadcast.ss.ps" => "__builtin_ia32_vbroadcastss_ps", + "llvm.x86.avx2.vbroadcast.ss.ps.256" => "__builtin_ia32_vbroadcastss_ps256", + "llvm.x86.avx2.vextracti128" => "__builtin_ia32_extract128i256", + "llvm.x86.avx2.vinserti128" => "__builtin_ia32_insert128i256", + "llvm.x86.avx2.vperm2i128" => "__builtin_ia32_permti256", + "llvm.x86.avx512.add.pd.512" => "__builtin_ia32_addpd512", + "llvm.x86.avx512.add.ps.512" => "__builtin_ia32_addps512", + "llvm.x86.avx512.broadcastmb.128" => "__builtin_ia32_broadcastmb128", + "llvm.x86.avx512.broadcastmb.256" => "__builtin_ia32_broadcastmb256", + "llvm.x86.avx512.broadcastmb.512" => "__builtin_ia32_broadcastmb512", + "llvm.x86.avx512.broadcastmw.128" => "__builtin_ia32_broadcastmw128", + "llvm.x86.avx512.broadcastmw.256" => "__builtin_ia32_broadcastmw256", + "llvm.x86.avx512.broadcastmw.512" => "__builtin_ia32_broadcastmw512", + "llvm.x86.avx512.conflict.d.128" => "__builtin_ia32_vpconflictsi_128", + "llvm.x86.avx512.conflict.d.256" => "__builtin_ia32_vpconflictsi_256", + "llvm.x86.avx512.conflict.d.512" => "__builtin_ia32_vpconflictsi_512", + "llvm.x86.avx512.conflict.q.128" => "__builtin_ia32_vpconflictdi_128", + "llvm.x86.avx512.conflict.q.256" => "__builtin_ia32_vpconflictdi_256", + "llvm.x86.avx512.conflict.q.512" => "__builtin_ia32_vpconflictdi_512", + "llvm.x86.avx512.cvtb2mask.128" => "__builtin_ia32_cvtb2mask128", + "llvm.x86.avx512.cvtb2mask.256" => "__builtin_ia32_cvtb2mask256", + "llvm.x86.avx512.cvtb2mask.512" => "__builtin_ia32_cvtb2mask512", + "llvm.x86.avx512.cvtd2mask.128" => "__builtin_ia32_cvtd2mask128", + "llvm.x86.avx512.cvtd2mask.256" => "__builtin_ia32_cvtd2mask256", + "llvm.x86.avx512.cvtd2mask.512" => "__builtin_ia32_cvtd2mask512", + "llvm.x86.avx512.cvtmask2b.128" => "__builtin_ia32_cvtmask2b128", + "llvm.x86.avx512.cvtmask2b.256" => "__builtin_ia32_cvtmask2b256", + "llvm.x86.avx512.cvtmask2b.512" => "__builtin_ia32_cvtmask2b512", + "llvm.x86.avx512.cvtmask2d.128" => "__builtin_ia32_cvtmask2d128", + "llvm.x86.avx512.cvtmask2d.256" => "__builtin_ia32_cvtmask2d256", + "llvm.x86.avx512.cvtmask2d.512" => "__builtin_ia32_cvtmask2d512", + "llvm.x86.avx512.cvtmask2q.128" => "__builtin_ia32_cvtmask2q128", + "llvm.x86.avx512.cvtmask2q.256" => "__builtin_ia32_cvtmask2q256", + "llvm.x86.avx512.cvtmask2q.512" => "__builtin_ia32_cvtmask2q512", + "llvm.x86.avx512.cvtmask2w.128" => "__builtin_ia32_cvtmask2w128", + "llvm.x86.avx512.cvtmask2w.256" => "__builtin_ia32_cvtmask2w256", + "llvm.x86.avx512.cvtmask2w.512" => "__builtin_ia32_cvtmask2w512", + "llvm.x86.avx512.cvtq2mask.128" => "__builtin_ia32_cvtq2mask128", + "llvm.x86.avx512.cvtq2mask.256" => "__builtin_ia32_cvtq2mask256", + "llvm.x86.avx512.cvtq2mask.512" => "__builtin_ia32_cvtq2mask512", + "llvm.x86.avx512.cvtsd2usi" => "__builtin_ia32_cvtsd2usi", + "llvm.x86.avx512.cvtsd2usi64" => "__builtin_ia32_cvtsd2usi64", + "llvm.x86.avx512.cvtsi2sd32" => "__builtin_ia32_cvtsi2sd32", + "llvm.x86.avx512.cvtsi2sd64" => "__builtin_ia32_cvtsi2sd64", + "llvm.x86.avx512.cvtsi2ss32" => "__builtin_ia32_cvtsi2ss32", + "llvm.x86.avx512.cvtsi2ss64" => "__builtin_ia32_cvtsi2ss64", + "llvm.x86.avx512.cvtss2usi" => "__builtin_ia32_cvtss2usi", + "llvm.x86.avx512.cvtss2usi64" => "__builtin_ia32_cvtss2usi64", + "llvm.x86.avx512.cvttsd2si" => "__builtin_ia32_vcvttsd2si32", + "llvm.x86.avx512.cvttsd2si64" => "__builtin_ia32_vcvttsd2si64", + "llvm.x86.avx512.cvttsd2usi" => "__builtin_ia32_vcvttsd2usi32", + // [DUPLICATE]: "llvm.x86.avx512.cvttsd2usi" => "__builtin_ia32_cvttsd2usi", + "llvm.x86.avx512.cvttsd2usi64" => "__builtin_ia32_vcvttsd2usi64", + // [DUPLICATE]: "llvm.x86.avx512.cvttsd2usi64" => "__builtin_ia32_cvttsd2usi64", + "llvm.x86.avx512.cvttss2si" => "__builtin_ia32_vcvttss2si32", + "llvm.x86.avx512.cvttss2si64" => "__builtin_ia32_vcvttss2si64", + "llvm.x86.avx512.cvttss2usi" => "__builtin_ia32_vcvttss2usi32", + // [DUPLICATE]: "llvm.x86.avx512.cvttss2usi" => "__builtin_ia32_cvttss2usi", + "llvm.x86.avx512.cvttss2usi64" => "__builtin_ia32_vcvttss2usi64", + // [DUPLICATE]: "llvm.x86.avx512.cvttss2usi64" => "__builtin_ia32_cvttss2usi64", + "llvm.x86.avx512.cvtusi2sd" => "__builtin_ia32_cvtusi2sd", + // [DUPLICATE]: "llvm.x86.avx512.cvtusi2sd" => "__builtin_ia32_cvtusi2sd32", + "llvm.x86.avx512.cvtusi2ss" => "__builtin_ia32_cvtusi2ss32", + // [DUPLICATE]: "llvm.x86.avx512.cvtusi2ss" => "__builtin_ia32_cvtusi2ss", + "llvm.x86.avx512.cvtusi642sd" => "__builtin_ia32_cvtusi2sd64", + // [DUPLICATE]: "llvm.x86.avx512.cvtusi642sd" => "__builtin_ia32_cvtusi642sd", + "llvm.x86.avx512.cvtusi642ss" => "__builtin_ia32_cvtusi2ss64", + // [DUPLICATE]: "llvm.x86.avx512.cvtusi642ss" => "__builtin_ia32_cvtusi642ss", + "llvm.x86.avx512.cvtw2mask.128" => "__builtin_ia32_cvtw2mask128", + "llvm.x86.avx512.cvtw2mask.256" => "__builtin_ia32_cvtw2mask256", + "llvm.x86.avx512.cvtw2mask.512" => "__builtin_ia32_cvtw2mask512", + "llvm.x86.avx512.dbpsadbw.128" => "__builtin_ia32_dbpsadbw128", + "llvm.x86.avx512.dbpsadbw.256" => "__builtin_ia32_dbpsadbw256", + "llvm.x86.avx512.dbpsadbw.512" => "__builtin_ia32_dbpsadbw512", + "llvm.x86.avx512.div.pd.512" => "__builtin_ia32_divpd512", + "llvm.x86.avx512.div.ps.512" => "__builtin_ia32_divps512", + "llvm.x86.avx512.exp2.pd" => "__builtin_ia32_exp2pd_mask", + "llvm.x86.avx512.exp2.ps" => "__builtin_ia32_exp2ps_mask", + "llvm.x86.avx512.gather.dpd.512" => "__builtin_ia32_gathersiv8df", + "llvm.x86.avx512.gather.dpi.512" => "__builtin_ia32_gathersiv16si", + "llvm.x86.avx512.gather.dpq.512" => "__builtin_ia32_gathersiv8di", + "llvm.x86.avx512.gather.dps.512" => "__builtin_ia32_gathersiv16sf", + "llvm.x86.avx512.gather.qpd.512" => "__builtin_ia32_gatherdiv8df", + "llvm.x86.avx512.gather.qpi.512" => "__builtin_ia32_gatherdiv16si", + "llvm.x86.avx512.gather.qpq.512" => "__builtin_ia32_gatherdiv8di", + "llvm.x86.avx512.gather.qps.512" => "__builtin_ia32_gatherdiv16sf", + "llvm.x86.avx512.gather3div2.df" => "__builtin_ia32_gather3div2df", + "llvm.x86.avx512.gather3div2.di" => "__builtin_ia32_gather3div2di", + "llvm.x86.avx512.gather3div4.df" => "__builtin_ia32_gather3div4df", + "llvm.x86.avx512.gather3div4.di" => "__builtin_ia32_gather3div4di", + "llvm.x86.avx512.gather3div4.sf" => "__builtin_ia32_gather3div4sf", + "llvm.x86.avx512.gather3div4.si" => "__builtin_ia32_gather3div4si", + "llvm.x86.avx512.gather3div8.sf" => "__builtin_ia32_gather3div8sf", + "llvm.x86.avx512.gather3div8.si" => "__builtin_ia32_gather3div8si", + "llvm.x86.avx512.gather3siv2.df" => "__builtin_ia32_gather3siv2df", + "llvm.x86.avx512.gather3siv2.di" => "__builtin_ia32_gather3siv2di", + "llvm.x86.avx512.gather3siv4.df" => "__builtin_ia32_gather3siv4df", + "llvm.x86.avx512.gather3siv4.di" => "__builtin_ia32_gather3siv4di", + "llvm.x86.avx512.gather3siv4.sf" => "__builtin_ia32_gather3siv4sf", + "llvm.x86.avx512.gather3siv4.si" => "__builtin_ia32_gather3siv4si", + "llvm.x86.avx512.gather3siv8.sf" => "__builtin_ia32_gather3siv8sf", + "llvm.x86.avx512.gather3siv8.si" => "__builtin_ia32_gather3siv8si", + "llvm.x86.avx512.gatherpf.dpd.512" => "__builtin_ia32_gatherpfdpd", + "llvm.x86.avx512.gatherpf.dps.512" => "__builtin_ia32_gatherpfdps", + "llvm.x86.avx512.gatherpf.qpd.512" => "__builtin_ia32_gatherpfqpd", + "llvm.x86.avx512.gatherpf.qps.512" => "__builtin_ia32_gatherpfqps", + "llvm.x86.avx512.kand.w" => "__builtin_ia32_kandhi", + "llvm.x86.avx512.kandn.w" => "__builtin_ia32_kandnhi", + "llvm.x86.avx512.knot.w" => "__builtin_ia32_knothi", + "llvm.x86.avx512.kor.w" => "__builtin_ia32_korhi", + "llvm.x86.avx512.kortestc.w" => "__builtin_ia32_kortestchi", + "llvm.x86.avx512.kortestz.w" => "__builtin_ia32_kortestzhi", + "llvm.x86.avx512.kunpck.bw" => "__builtin_ia32_kunpckhi", + "llvm.x86.avx512.kunpck.dq" => "__builtin_ia32_kunpckdi", + "llvm.x86.avx512.kunpck.wd" => "__builtin_ia32_kunpcksi", + "llvm.x86.avx512.kxnor.w" => "__builtin_ia32_kxnorhi", + "llvm.x86.avx512.kxor.w" => "__builtin_ia32_kxorhi", + "llvm.x86.avx512.mask.add.pd.128" => "__builtin_ia32_addpd128_mask", + "llvm.x86.avx512.mask.add.pd.256" => "__builtin_ia32_addpd256_mask", + "llvm.x86.avx512.mask.add.pd.512" => "__builtin_ia32_addpd512_mask", + "llvm.x86.avx512.mask.add.ps.128" => "__builtin_ia32_addps128_mask", + "llvm.x86.avx512.mask.add.ps.256" => "__builtin_ia32_addps256_mask", + "llvm.x86.avx512.mask.add.ps.512" => "__builtin_ia32_addps512_mask", + "llvm.x86.avx512.mask.add.sd.round" => "__builtin_ia32_addsd_round_mask", + "llvm.x86.avx512.mask.add.ss.round" => "__builtin_ia32_addss_round_mask", + "llvm.x86.avx512.mask.and.pd.128" => "__builtin_ia32_andpd128_mask", + "llvm.x86.avx512.mask.and.pd.256" => "__builtin_ia32_andpd256_mask", + "llvm.x86.avx512.mask.and.pd.512" => "__builtin_ia32_andpd512_mask", + "llvm.x86.avx512.mask.and.ps.128" => "__builtin_ia32_andps128_mask", + "llvm.x86.avx512.mask.and.ps.256" => "__builtin_ia32_andps256_mask", + "llvm.x86.avx512.mask.and.ps.512" => "__builtin_ia32_andps512_mask", + "llvm.x86.avx512.mask.andn.pd.128" => "__builtin_ia32_andnpd128_mask", + "llvm.x86.avx512.mask.andn.pd.256" => "__builtin_ia32_andnpd256_mask", + "llvm.x86.avx512.mask.andn.pd.512" => "__builtin_ia32_andnpd512_mask", + "llvm.x86.avx512.mask.andn.ps.128" => "__builtin_ia32_andnps128_mask", + "llvm.x86.avx512.mask.andn.ps.256" => "__builtin_ia32_andnps256_mask", + "llvm.x86.avx512.mask.andn.ps.512" => "__builtin_ia32_andnps512_mask", + "llvm.x86.avx512.mask.blend.d.512" => "__builtin_ia32_blendmd_512_mask", + "llvm.x86.avx512.mask.blend.pd.512" => "__builtin_ia32_blendmpd_512_mask", + "llvm.x86.avx512.mask.blend.ps.512" => "__builtin_ia32_blendmps_512_mask", + "llvm.x86.avx512.mask.blend.q.512" => "__builtin_ia32_blendmq_512_mask", + "llvm.x86.avx512.mask.broadcastf32x2.256" => "__builtin_ia32_broadcastf32x2_256_mask", + "llvm.x86.avx512.mask.broadcastf32x2.512" => "__builtin_ia32_broadcastf32x2_512_mask", + "llvm.x86.avx512.mask.broadcastf32x4.256" => "__builtin_ia32_broadcastf32x4_256_mask", + "llvm.x86.avx512.mask.broadcastf32x4.512" => "__builtin_ia32_broadcastf32x4_512", + "llvm.x86.avx512.mask.broadcastf32x8.512" => "__builtin_ia32_broadcastf32x8_512_mask", + "llvm.x86.avx512.mask.broadcastf64x2.256" => "__builtin_ia32_broadcastf64x2_256_mask", + "llvm.x86.avx512.mask.broadcastf64x2.512" => "__builtin_ia32_broadcastf64x2_512_mask", + "llvm.x86.avx512.mask.broadcastf64x4.512" => "__builtin_ia32_broadcastf64x4_512", + "llvm.x86.avx512.mask.broadcasti32x2.128" => "__builtin_ia32_broadcasti32x2_128_mask", + "llvm.x86.avx512.mask.broadcasti32x2.256" => "__builtin_ia32_broadcasti32x2_256_mask", + "llvm.x86.avx512.mask.broadcasti32x2.512" => "__builtin_ia32_broadcasti32x2_512_mask", + "llvm.x86.avx512.mask.broadcasti32x4.256" => "__builtin_ia32_broadcasti32x4_256_mask", + "llvm.x86.avx512.mask.broadcasti32x4.512" => "__builtin_ia32_broadcasti32x4_512", + "llvm.x86.avx512.mask.broadcasti32x8.512" => "__builtin_ia32_broadcasti32x8_512_mask", + "llvm.x86.avx512.mask.broadcasti64x2.256" => "__builtin_ia32_broadcasti64x2_256_mask", + "llvm.x86.avx512.mask.broadcasti64x2.512" => "__builtin_ia32_broadcasti64x2_512_mask", + "llvm.x86.avx512.mask.broadcasti64x4.512" => "__builtin_ia32_broadcasti64x4_512", + "llvm.x86.avx512.mask.cmp.pd.128" => "__builtin_ia32_cmppd128_mask", + "llvm.x86.avx512.mask.cmp.pd.256" => "__builtin_ia32_cmppd256_mask", + "llvm.x86.avx512.mask.cmp.pd.512" => "__builtin_ia32_cmppd512_mask", + "llvm.x86.avx512.mask.cmp.ps.128" => "__builtin_ia32_cmpps128_mask", + "llvm.x86.avx512.mask.cmp.ps.256" => "__builtin_ia32_cmpps256_mask", + "llvm.x86.avx512.mask.cmp.ps.512" => "__builtin_ia32_cmpps512_mask", + "llvm.x86.avx512.mask.cmp.sd" => "__builtin_ia32_cmpsd_mask", + "llvm.x86.avx512.mask.cmp.ss" => "__builtin_ia32_cmpss_mask", + "llvm.x86.avx512.mask.compress.d.128" => "__builtin_ia32_compresssi128_mask", + "llvm.x86.avx512.mask.compress.d.256" => "__builtin_ia32_compresssi256_mask", + "llvm.x86.avx512.mask.compress.d.512" => "__builtin_ia32_compresssi512_mask", + "llvm.x86.avx512.mask.compress.pd.128" => "__builtin_ia32_compressdf128_mask", + "llvm.x86.avx512.mask.compress.pd.256" => "__builtin_ia32_compressdf256_mask", + "llvm.x86.avx512.mask.compress.pd.512" => "__builtin_ia32_compressdf512_mask", + "llvm.x86.avx512.mask.compress.ps.128" => "__builtin_ia32_compresssf128_mask", + "llvm.x86.avx512.mask.compress.ps.256" => "__builtin_ia32_compresssf256_mask", + "llvm.x86.avx512.mask.compress.ps.512" => "__builtin_ia32_compresssf512_mask", + "llvm.x86.avx512.mask.compress.q.128" => "__builtin_ia32_compressdi128_mask", + "llvm.x86.avx512.mask.compress.q.256" => "__builtin_ia32_compressdi256_mask", + "llvm.x86.avx512.mask.compress.q.512" => "__builtin_ia32_compressdi512_mask", + "llvm.x86.avx512.mask.compress.store.d.128" => "__builtin_ia32_compressstoresi128_mask", + "llvm.x86.avx512.mask.compress.store.d.256" => "__builtin_ia32_compressstoresi256_mask", + "llvm.x86.avx512.mask.compress.store.d.512" => "__builtin_ia32_compressstoresi512_mask", + "llvm.x86.avx512.mask.compress.store.pd.128" => "__builtin_ia32_compressstoredf128_mask", + "llvm.x86.avx512.mask.compress.store.pd.256" => "__builtin_ia32_compressstoredf256_mask", + "llvm.x86.avx512.mask.compress.store.pd.512" => "__builtin_ia32_compressstoredf512_mask", + "llvm.x86.avx512.mask.compress.store.ps.128" => "__builtin_ia32_compressstoresf128_mask", + "llvm.x86.avx512.mask.compress.store.ps.256" => "__builtin_ia32_compressstoresf256_mask", + "llvm.x86.avx512.mask.compress.store.ps.512" => "__builtin_ia32_compressstoresf512_mask", + "llvm.x86.avx512.mask.compress.store.q.128" => "__builtin_ia32_compressstoredi128_mask", + "llvm.x86.avx512.mask.compress.store.q.256" => "__builtin_ia32_compressstoredi256_mask", + "llvm.x86.avx512.mask.compress.store.q.512" => "__builtin_ia32_compressstoredi512_mask", + "llvm.x86.avx512.mask.conflict.d.128" => "__builtin_ia32_vpconflictsi_128_mask", + "llvm.x86.avx512.mask.conflict.d.256" => "__builtin_ia32_vpconflictsi_256_mask", + "llvm.x86.avx512.mask.conflict.d.512" => "__builtin_ia32_vpconflictsi_512_mask", + "llvm.x86.avx512.mask.conflict.q.128" => "__builtin_ia32_vpconflictdi_128_mask", + "llvm.x86.avx512.mask.conflict.q.256" => "__builtin_ia32_vpconflictdi_256_mask", + "llvm.x86.avx512.mask.conflict.q.512" => "__builtin_ia32_vpconflictdi_512_mask", + "llvm.x86.avx512.mask.cvtdq2pd.128" => "__builtin_ia32_cvtdq2pd128_mask", + "llvm.x86.avx512.mask.cvtdq2pd.256" => "__builtin_ia32_cvtdq2pd256_mask", + "llvm.x86.avx512.mask.cvtdq2pd.512" => "__builtin_ia32_cvtdq2pd512_mask", + "llvm.x86.avx512.mask.cvtdq2ps.128" => "__builtin_ia32_cvtdq2ps128_mask", + "llvm.x86.avx512.mask.cvtdq2ps.256" => "__builtin_ia32_cvtdq2ps256_mask", + "llvm.x86.avx512.mask.cvtdq2ps.512" => "__builtin_ia32_cvtdq2ps512_mask", + "llvm.x86.avx512.mask.cvtpd2dq.128" => "__builtin_ia32_cvtpd2dq128_mask", + "llvm.x86.avx512.mask.cvtpd2dq.256" => "__builtin_ia32_cvtpd2dq256_mask", + "llvm.x86.avx512.mask.cvtpd2dq.512" => "__builtin_ia32_cvtpd2dq512_mask", + "llvm.x86.avx512.mask.cvtpd2ps" => "__builtin_ia32_cvtpd2ps_mask", + "llvm.x86.avx512.mask.cvtpd2ps.256" => "__builtin_ia32_cvtpd2ps256_mask", + "llvm.x86.avx512.mask.cvtpd2ps.512" => "__builtin_ia32_cvtpd2ps512_mask", + "llvm.x86.avx512.mask.cvtpd2qq.128" => "__builtin_ia32_cvtpd2qq128_mask", + "llvm.x86.avx512.mask.cvtpd2qq.256" => "__builtin_ia32_cvtpd2qq256_mask", + "llvm.x86.avx512.mask.cvtpd2qq.512" => "__builtin_ia32_cvtpd2qq512_mask", + "llvm.x86.avx512.mask.cvtpd2udq.128" => "__builtin_ia32_cvtpd2udq128_mask", + "llvm.x86.avx512.mask.cvtpd2udq.256" => "__builtin_ia32_cvtpd2udq256_mask", + "llvm.x86.avx512.mask.cvtpd2udq.512" => "__builtin_ia32_cvtpd2udq512_mask", + "llvm.x86.avx512.mask.cvtpd2uqq.128" => "__builtin_ia32_cvtpd2uqq128_mask", + "llvm.x86.avx512.mask.cvtpd2uqq.256" => "__builtin_ia32_cvtpd2uqq256_mask", + "llvm.x86.avx512.mask.cvtpd2uqq.512" => "__builtin_ia32_cvtpd2uqq512_mask", + "llvm.x86.avx512.mask.cvtps2dq.128" => "__builtin_ia32_cvtps2dq128_mask", + "llvm.x86.avx512.mask.cvtps2dq.256" => "__builtin_ia32_cvtps2dq256_mask", + "llvm.x86.avx512.mask.cvtps2dq.512" => "__builtin_ia32_cvtps2dq512_mask", + "llvm.x86.avx512.mask.cvtps2pd.128" => "__builtin_ia32_cvtps2pd128_mask", + "llvm.x86.avx512.mask.cvtps2pd.256" => "__builtin_ia32_cvtps2pd256_mask", + "llvm.x86.avx512.mask.cvtps2pd.512" => "__builtin_ia32_cvtps2pd512_mask", + "llvm.x86.avx512.mask.cvtps2qq.128" => "__builtin_ia32_cvtps2qq128_mask", + "llvm.x86.avx512.mask.cvtps2qq.256" => "__builtin_ia32_cvtps2qq256_mask", + "llvm.x86.avx512.mask.cvtps2qq.512" => "__builtin_ia32_cvtps2qq512_mask", + "llvm.x86.avx512.mask.cvtps2udq.128" => "__builtin_ia32_cvtps2udq128_mask", + "llvm.x86.avx512.mask.cvtps2udq.256" => "__builtin_ia32_cvtps2udq256_mask", + "llvm.x86.avx512.mask.cvtps2udq.512" => "__builtin_ia32_cvtps2udq512_mask", + "llvm.x86.avx512.mask.cvtps2uqq.128" => "__builtin_ia32_cvtps2uqq128_mask", + "llvm.x86.avx512.mask.cvtps2uqq.256" => "__builtin_ia32_cvtps2uqq256_mask", + "llvm.x86.avx512.mask.cvtps2uqq.512" => "__builtin_ia32_cvtps2uqq512_mask", + "llvm.x86.avx512.mask.cvtqq2pd.128" => "__builtin_ia32_cvtqq2pd128_mask", + "llvm.x86.avx512.mask.cvtqq2pd.256" => "__builtin_ia32_cvtqq2pd256_mask", + "llvm.x86.avx512.mask.cvtqq2pd.512" => "__builtin_ia32_cvtqq2pd512_mask", + "llvm.x86.avx512.mask.cvtqq2ps.128" => "__builtin_ia32_cvtqq2ps128_mask", + "llvm.x86.avx512.mask.cvtqq2ps.256" => "__builtin_ia32_cvtqq2ps256_mask", + "llvm.x86.avx512.mask.cvtqq2ps.512" => "__builtin_ia32_cvtqq2ps512_mask", + "llvm.x86.avx512.mask.cvtsd2ss.round" => "__builtin_ia32_cvtsd2ss_round_mask", + "llvm.x86.avx512.mask.cvtss2sd.round" => "__builtin_ia32_cvtss2sd_round_mask", + "llvm.x86.avx512.mask.cvttpd2dq.128" => "__builtin_ia32_cvttpd2dq128_mask", + "llvm.x86.avx512.mask.cvttpd2dq.256" => "__builtin_ia32_cvttpd2dq256_mask", + "llvm.x86.avx512.mask.cvttpd2dq.512" => "__builtin_ia32_cvttpd2dq512_mask", + "llvm.x86.avx512.mask.cvttpd2qq.128" => "__builtin_ia32_cvttpd2qq128_mask", + "llvm.x86.avx512.mask.cvttpd2qq.256" => "__builtin_ia32_cvttpd2qq256_mask", + "llvm.x86.avx512.mask.cvttpd2qq.512" => "__builtin_ia32_cvttpd2qq512_mask", + "llvm.x86.avx512.mask.cvttpd2udq.128" => "__builtin_ia32_cvttpd2udq128_mask", + "llvm.x86.avx512.mask.cvttpd2udq.256" => "__builtin_ia32_cvttpd2udq256_mask", + "llvm.x86.avx512.mask.cvttpd2udq.512" => "__builtin_ia32_cvttpd2udq512_mask", + "llvm.x86.avx512.mask.cvttpd2uqq.128" => "__builtin_ia32_cvttpd2uqq128_mask", + "llvm.x86.avx512.mask.cvttpd2uqq.256" => "__builtin_ia32_cvttpd2uqq256_mask", + "llvm.x86.avx512.mask.cvttpd2uqq.512" => "__builtin_ia32_cvttpd2uqq512_mask", + "llvm.x86.avx512.mask.cvttps2dq.128" => "__builtin_ia32_cvttps2dq128_mask", + "llvm.x86.avx512.mask.cvttps2dq.256" => "__builtin_ia32_cvttps2dq256_mask", + "llvm.x86.avx512.mask.cvttps2dq.512" => "__builtin_ia32_cvttps2dq512_mask", + "llvm.x86.avx512.mask.cvttps2qq.128" => "__builtin_ia32_cvttps2qq128_mask", + "llvm.x86.avx512.mask.cvttps2qq.256" => "__builtin_ia32_cvttps2qq256_mask", + "llvm.x86.avx512.mask.cvttps2qq.512" => "__builtin_ia32_cvttps2qq512_mask", + "llvm.x86.avx512.mask.cvttps2udq.128" => "__builtin_ia32_cvttps2udq128_mask", + "llvm.x86.avx512.mask.cvttps2udq.256" => "__builtin_ia32_cvttps2udq256_mask", + "llvm.x86.avx512.mask.cvttps2udq.512" => "__builtin_ia32_cvttps2udq512_mask", + "llvm.x86.avx512.mask.cvttps2uqq.128" => "__builtin_ia32_cvttps2uqq128_mask", + "llvm.x86.avx512.mask.cvttps2uqq.256" => "__builtin_ia32_cvttps2uqq256_mask", + "llvm.x86.avx512.mask.cvttps2uqq.512" => "__builtin_ia32_cvttps2uqq512_mask", + "llvm.x86.avx512.mask.cvtudq2pd.128" => "__builtin_ia32_cvtudq2pd128_mask", + "llvm.x86.avx512.mask.cvtudq2pd.256" => "__builtin_ia32_cvtudq2pd256_mask", + "llvm.x86.avx512.mask.cvtudq2pd.512" => "__builtin_ia32_cvtudq2pd512_mask", + "llvm.x86.avx512.mask.cvtudq2ps.128" => "__builtin_ia32_cvtudq2ps128_mask", + "llvm.x86.avx512.mask.cvtudq2ps.256" => "__builtin_ia32_cvtudq2ps256_mask", + "llvm.x86.avx512.mask.cvtudq2ps.512" => "__builtin_ia32_cvtudq2ps512_mask", + "llvm.x86.avx512.mask.cvtuqq2pd.128" => "__builtin_ia32_cvtuqq2pd128_mask", + "llvm.x86.avx512.mask.cvtuqq2pd.256" => "__builtin_ia32_cvtuqq2pd256_mask", + "llvm.x86.avx512.mask.cvtuqq2pd.512" => "__builtin_ia32_cvtuqq2pd512_mask", + "llvm.x86.avx512.mask.cvtuqq2ps.128" => "__builtin_ia32_cvtuqq2ps128_mask", + "llvm.x86.avx512.mask.cvtuqq2ps.256" => "__builtin_ia32_cvtuqq2ps256_mask", + "llvm.x86.avx512.mask.cvtuqq2ps.512" => "__builtin_ia32_cvtuqq2ps512_mask", + "llvm.x86.avx512.mask.dbpsadbw.128" => "__builtin_ia32_dbpsadbw128_mask", + "llvm.x86.avx512.mask.dbpsadbw.256" => "__builtin_ia32_dbpsadbw256_mask", + "llvm.x86.avx512.mask.dbpsadbw.512" => "__builtin_ia32_dbpsadbw512_mask", + "llvm.x86.avx512.mask.div.pd.128" => "__builtin_ia32_divpd_mask", + "llvm.x86.avx512.mask.div.pd.256" => "__builtin_ia32_divpd256_mask", + "llvm.x86.avx512.mask.div.pd.512" => "__builtin_ia32_divpd512_mask", + "llvm.x86.avx512.mask.div.ps.128" => "__builtin_ia32_divps_mask", + "llvm.x86.avx512.mask.div.ps.256" => "__builtin_ia32_divps256_mask", + "llvm.x86.avx512.mask.div.ps.512" => "__builtin_ia32_divps512_mask", + "llvm.x86.avx512.mask.div.sd.round" => "__builtin_ia32_divsd_round_mask", + "llvm.x86.avx512.mask.div.ss.round" => "__builtin_ia32_divss_round_mask", + "llvm.x86.avx512.mask.expand.d.128" => "__builtin_ia32_expandsi128_mask", + "llvm.x86.avx512.mask.expand.d.256" => "__builtin_ia32_expandsi256_mask", + "llvm.x86.avx512.mask.expand.d.512" => "__builtin_ia32_expandsi512_mask", + "llvm.x86.avx512.mask.expand.load.d.128" => "__builtin_ia32_expandloadsi128_mask", + "llvm.x86.avx512.mask.expand.load.d.256" => "__builtin_ia32_expandloadsi256_mask", + "llvm.x86.avx512.mask.expand.load.d.512" => "__builtin_ia32_expandloadsi512_mask", + "llvm.x86.avx512.mask.expand.load.pd.128" => "__builtin_ia32_expandloaddf128_mask", + "llvm.x86.avx512.mask.expand.load.pd.256" => "__builtin_ia32_expandloaddf256_mask", + "llvm.x86.avx512.mask.expand.load.pd.512" => "__builtin_ia32_expandloaddf512_mask", + "llvm.x86.avx512.mask.expand.load.ps.128" => "__builtin_ia32_expandloadsf128_mask", + "llvm.x86.avx512.mask.expand.load.ps.256" => "__builtin_ia32_expandloadsf256_mask", + "llvm.x86.avx512.mask.expand.load.ps.512" => "__builtin_ia32_expandloadsf512_mask", + "llvm.x86.avx512.mask.expand.load.q.128" => "__builtin_ia32_expandloaddi128_mask", + "llvm.x86.avx512.mask.expand.load.q.256" => "__builtin_ia32_expandloaddi256_mask", + "llvm.x86.avx512.mask.expand.load.q.512" => "__builtin_ia32_expandloaddi512_mask", + "llvm.x86.avx512.mask.expand.pd.128" => "__builtin_ia32_expanddf128_mask", + "llvm.x86.avx512.mask.expand.pd.256" => "__builtin_ia32_expanddf256_mask", + "llvm.x86.avx512.mask.expand.pd.512" => "__builtin_ia32_expanddf512_mask", + "llvm.x86.avx512.mask.expand.ps.128" => "__builtin_ia32_expandsf128_mask", + "llvm.x86.avx512.mask.expand.ps.256" => "__builtin_ia32_expandsf256_mask", + "llvm.x86.avx512.mask.expand.ps.512" => "__builtin_ia32_expandsf512_mask", + "llvm.x86.avx512.mask.expand.q.128" => "__builtin_ia32_expanddi128_mask", + "llvm.x86.avx512.mask.expand.q.256" => "__builtin_ia32_expanddi256_mask", + "llvm.x86.avx512.mask.expand.q.512" => "__builtin_ia32_expanddi512_mask", + "llvm.x86.avx512.mask.fixupimm.pd.128" => "__builtin_ia32_fixupimmpd128_mask", + "llvm.x86.avx512.mask.fixupimm.pd.256" => "__builtin_ia32_fixupimmpd256_mask", + "llvm.x86.avx512.mask.fixupimm.pd.512" => "__builtin_ia32_fixupimmpd512_mask", + "llvm.x86.avx512.mask.fixupimm.ps.128" => "__builtin_ia32_fixupimmps128_mask", + "llvm.x86.avx512.mask.fixupimm.ps.256" => "__builtin_ia32_fixupimmps256_mask", + "llvm.x86.avx512.mask.fixupimm.ps.512" => "__builtin_ia32_fixupimmps512_mask", + "llvm.x86.avx512.mask.fixupimm.sd" => "__builtin_ia32_fixupimmsd_mask", + "llvm.x86.avx512.mask.fixupimm.ss" => "__builtin_ia32_fixupimmss_mask", + "llvm.x86.avx512.mask.fpclass.pd.128" => "__builtin_ia32_fpclasspd128_mask", + "llvm.x86.avx512.mask.fpclass.pd.256" => "__builtin_ia32_fpclasspd256_mask", + "llvm.x86.avx512.mask.fpclass.pd.512" => "__builtin_ia32_fpclasspd512_mask", + "llvm.x86.avx512.mask.fpclass.ps.128" => "__builtin_ia32_fpclassps128_mask", + "llvm.x86.avx512.mask.fpclass.ps.256" => "__builtin_ia32_fpclassps256_mask", + "llvm.x86.avx512.mask.fpclass.ps.512" => "__builtin_ia32_fpclassps512_mask", + "llvm.x86.avx512.mask.fpclass.sd" => "__builtin_ia32_fpclasssd_mask", + "llvm.x86.avx512.mask.fpclass.ss" => "__builtin_ia32_fpclassss_mask", + "llvm.x86.avx512.mask.getexp.pd.128" => "__builtin_ia32_getexppd128_mask", + "llvm.x86.avx512.mask.getexp.pd.256" => "__builtin_ia32_getexppd256_mask", + "llvm.x86.avx512.mask.getexp.pd.512" => "__builtin_ia32_getexppd512_mask", + "llvm.x86.avx512.mask.getexp.ps.128" => "__builtin_ia32_getexpps128_mask", + "llvm.x86.avx512.mask.getexp.ps.256" => "__builtin_ia32_getexpps256_mask", + "llvm.x86.avx512.mask.getexp.ps.512" => "__builtin_ia32_getexpps512_mask", + "llvm.x86.avx512.mask.getexp.sd" => "__builtin_ia32_getexpsd128_round_mask", + "llvm.x86.avx512.mask.getexp.ss" => "__builtin_ia32_getexpss128_round_mask", + "llvm.x86.avx512.mask.getmant.pd.128" => "__builtin_ia32_getmantpd128_mask", + "llvm.x86.avx512.mask.getmant.pd.256" => "__builtin_ia32_getmantpd256_mask", + "llvm.x86.avx512.mask.getmant.pd.512" => "__builtin_ia32_getmantpd512_mask", + "llvm.x86.avx512.mask.getmant.ps.128" => "__builtin_ia32_getmantps128_mask", + "llvm.x86.avx512.mask.getmant.ps.256" => "__builtin_ia32_getmantps256_mask", + "llvm.x86.avx512.mask.getmant.ps.512" => "__builtin_ia32_getmantps512_mask", + "llvm.x86.avx512.mask.getmant.sd" => "__builtin_ia32_getmantsd_round_mask", + "llvm.x86.avx512.mask.getmant.ss" => "__builtin_ia32_getmantss_round_mask", + "llvm.x86.avx512.mask.insertf32x4.256" => "__builtin_ia32_insertf32x4_256_mask", + "llvm.x86.avx512.mask.insertf32x4.512" => "__builtin_ia32_insertf32x4_mask", + "llvm.x86.avx512.mask.insertf32x8.512" => "__builtin_ia32_insertf32x8_mask", + "llvm.x86.avx512.mask.insertf64x2.256" => "__builtin_ia32_insertf64x2_256_mask", + "llvm.x86.avx512.mask.insertf64x2.512" => "__builtin_ia32_insertf64x2_512_mask", + "llvm.x86.avx512.mask.insertf64x4.512" => "__builtin_ia32_insertf64x4_mask", + "llvm.x86.avx512.mask.inserti32x4.256" => "__builtin_ia32_inserti32x4_256_mask", + "llvm.x86.avx512.mask.inserti32x4.512" => "__builtin_ia32_inserti32x4_mask", + "llvm.x86.avx512.mask.inserti32x8.512" => "__builtin_ia32_inserti32x8_mask", + "llvm.x86.avx512.mask.inserti64x2.256" => "__builtin_ia32_inserti64x2_256_mask", + "llvm.x86.avx512.mask.inserti64x2.512" => "__builtin_ia32_inserti64x2_512_mask", + "llvm.x86.avx512.mask.inserti64x4.512" => "__builtin_ia32_inserti64x4_mask", + "llvm.x86.avx512.mask.loadu.d.512" => "__builtin_ia32_loaddqusi512_mask", + "llvm.x86.avx512.mask.loadu.pd.512" => "__builtin_ia32_loadupd512_mask", + "llvm.x86.avx512.mask.loadu.ps.512" => "__builtin_ia32_loadups512_mask", + "llvm.x86.avx512.mask.loadu.q.512" => "__builtin_ia32_loaddqudi512_mask", + "llvm.x86.avx512.mask.lzcnt.d.512" => "__builtin_ia32_vplzcntd_512_mask", + "llvm.x86.avx512.mask.lzcnt.q.512" => "__builtin_ia32_vplzcntq_512_mask", + "llvm.x86.avx512.mask.max.pd.128" => "__builtin_ia32_maxpd_mask", + "llvm.x86.avx512.mask.max.pd.256" => "__builtin_ia32_maxpd256_mask", + "llvm.x86.avx512.mask.max.pd.512" => "__builtin_ia32_maxpd512_mask", + "llvm.x86.avx512.mask.max.ps.128" => "__builtin_ia32_maxps_mask", + "llvm.x86.avx512.mask.max.ps.256" => "__builtin_ia32_maxps256_mask", + "llvm.x86.avx512.mask.max.ps.512" => "__builtin_ia32_maxps512_mask", + "llvm.x86.avx512.mask.max.sd.round" => "__builtin_ia32_maxsd_round_mask", + "llvm.x86.avx512.mask.max.ss.round" => "__builtin_ia32_maxss_round_mask", + "llvm.x86.avx512.mask.min.pd.128" => "__builtin_ia32_minpd_mask", + "llvm.x86.avx512.mask.min.pd.256" => "__builtin_ia32_minpd256_mask", + "llvm.x86.avx512.mask.min.pd.512" => "__builtin_ia32_minpd512_mask", + "llvm.x86.avx512.mask.min.ps.128" => "__builtin_ia32_minps_mask", + "llvm.x86.avx512.mask.min.ps.256" => "__builtin_ia32_minps256_mask", + "llvm.x86.avx512.mask.min.ps.512" => "__builtin_ia32_minps512_mask", + "llvm.x86.avx512.mask.min.sd.round" => "__builtin_ia32_minsd_round_mask", + "llvm.x86.avx512.mask.min.ss.round" => "__builtin_ia32_minss_round_mask", + "llvm.x86.avx512.mask.move.sd" => "__builtin_ia32_movsd_mask", + "llvm.x86.avx512.mask.move.ss" => "__builtin_ia32_movss_mask", + "llvm.x86.avx512.mask.mul.pd.128" => "__builtin_ia32_mulpd_mask", + "llvm.x86.avx512.mask.mul.pd.256" => "__builtin_ia32_mulpd256_mask", + "llvm.x86.avx512.mask.mul.pd.512" => "__builtin_ia32_mulpd512_mask", + "llvm.x86.avx512.mask.mul.ps.128" => "__builtin_ia32_mulps_mask", + "llvm.x86.avx512.mask.mul.ps.256" => "__builtin_ia32_mulps256_mask", + "llvm.x86.avx512.mask.mul.ps.512" => "__builtin_ia32_mulps512_mask", + "llvm.x86.avx512.mask.mul.sd.round" => "__builtin_ia32_mulsd_round_mask", + "llvm.x86.avx512.mask.mul.ss.round" => "__builtin_ia32_mulss_round_mask", + "llvm.x86.avx512.mask.or.pd.128" => "__builtin_ia32_orpd128_mask", + "llvm.x86.avx512.mask.or.pd.256" => "__builtin_ia32_orpd256_mask", + "llvm.x86.avx512.mask.or.pd.512" => "__builtin_ia32_orpd512_mask", + "llvm.x86.avx512.mask.or.ps.128" => "__builtin_ia32_orps128_mask", + "llvm.x86.avx512.mask.or.ps.256" => "__builtin_ia32_orps256_mask", + "llvm.x86.avx512.mask.or.ps.512" => "__builtin_ia32_orps512_mask", + "llvm.x86.avx512.mask.pabs.b.128" => "__builtin_ia32_pabsb128_mask", + "llvm.x86.avx512.mask.pabs.b.256" => "__builtin_ia32_pabsb256_mask", + "llvm.x86.avx512.mask.pabs.b.512" => "__builtin_ia32_pabsb512_mask", + "llvm.x86.avx512.mask.pabs.d.128" => "__builtin_ia32_pabsd128_mask", + "llvm.x86.avx512.mask.pabs.d.256" => "__builtin_ia32_pabsd256_mask", + "llvm.x86.avx512.mask.pabs.d.512" => "__builtin_ia32_pabsd512_mask", + "llvm.x86.avx512.mask.pabs.q.128" => "__builtin_ia32_pabsq128_mask", + "llvm.x86.avx512.mask.pabs.q.256" => "__builtin_ia32_pabsq256_mask", + "llvm.x86.avx512.mask.pabs.q.512" => "__builtin_ia32_pabsq512_mask", + "llvm.x86.avx512.mask.pabs.w.128" => "__builtin_ia32_pabsw128_mask", + "llvm.x86.avx512.mask.pabs.w.256" => "__builtin_ia32_pabsw256_mask", + "llvm.x86.avx512.mask.pabs.w.512" => "__builtin_ia32_pabsw512_mask", + "llvm.x86.avx512.mask.packssdw.128" => "__builtin_ia32_packssdw128_mask", + "llvm.x86.avx512.mask.packssdw.256" => "__builtin_ia32_packssdw256_mask", + "llvm.x86.avx512.mask.packssdw.512" => "__builtin_ia32_packssdw512_mask", + "llvm.x86.avx512.mask.packsswb.128" => "__builtin_ia32_packsswb128_mask", + "llvm.x86.avx512.mask.packsswb.256" => "__builtin_ia32_packsswb256_mask", + "llvm.x86.avx512.mask.packsswb.512" => "__builtin_ia32_packsswb512_mask", + "llvm.x86.avx512.mask.packusdw.128" => "__builtin_ia32_packusdw128_mask", + "llvm.x86.avx512.mask.packusdw.256" => "__builtin_ia32_packusdw256_mask", + "llvm.x86.avx512.mask.packusdw.512" => "__builtin_ia32_packusdw512_mask", + "llvm.x86.avx512.mask.packuswb.128" => "__builtin_ia32_packuswb128_mask", + "llvm.x86.avx512.mask.packuswb.256" => "__builtin_ia32_packuswb256_mask", + "llvm.x86.avx512.mask.packuswb.512" => "__builtin_ia32_packuswb512_mask", + "llvm.x86.avx512.mask.padd.b.128" => "__builtin_ia32_paddb128_mask", + "llvm.x86.avx512.mask.padd.b.256" => "__builtin_ia32_paddb256_mask", + "llvm.x86.avx512.mask.padd.b.512" => "__builtin_ia32_paddb512_mask", + "llvm.x86.avx512.mask.padd.d.128" => "__builtin_ia32_paddd128_mask", + "llvm.x86.avx512.mask.padd.d.256" => "__builtin_ia32_paddd256_mask", + "llvm.x86.avx512.mask.padd.d.512" => "__builtin_ia32_paddd512_mask", + "llvm.x86.avx512.mask.padd.q.128" => "__builtin_ia32_paddq128_mask", + "llvm.x86.avx512.mask.padd.q.256" => "__builtin_ia32_paddq256_mask", + "llvm.x86.avx512.mask.padd.q.512" => "__builtin_ia32_paddq512_mask", + "llvm.x86.avx512.mask.padd.w.128" => "__builtin_ia32_paddw128_mask", + "llvm.x86.avx512.mask.padd.w.256" => "__builtin_ia32_paddw256_mask", + "llvm.x86.avx512.mask.padd.w.512" => "__builtin_ia32_paddw512_mask", + "llvm.x86.avx512.mask.padds.b.128" => "__builtin_ia32_paddsb128_mask", + "llvm.x86.avx512.mask.padds.b.256" => "__builtin_ia32_paddsb256_mask", + "llvm.x86.avx512.mask.padds.b.512" => "__builtin_ia32_paddsb512_mask", + "llvm.x86.avx512.mask.padds.w.128" => "__builtin_ia32_paddsw128_mask", + "llvm.x86.avx512.mask.padds.w.256" => "__builtin_ia32_paddsw256_mask", + "llvm.x86.avx512.mask.padds.w.512" => "__builtin_ia32_paddsw512_mask", + "llvm.x86.avx512.mask.paddus.b.128" => "__builtin_ia32_paddusb128_mask", + "llvm.x86.avx512.mask.paddus.b.256" => "__builtin_ia32_paddusb256_mask", + "llvm.x86.avx512.mask.paddus.b.512" => "__builtin_ia32_paddusb512_mask", + "llvm.x86.avx512.mask.paddus.w.128" => "__builtin_ia32_paddusw128_mask", + "llvm.x86.avx512.mask.paddus.w.256" => "__builtin_ia32_paddusw256_mask", + "llvm.x86.avx512.mask.paddus.w.512" => "__builtin_ia32_paddusw512_mask", + "llvm.x86.avx512.mask.pand.d.512" => "__builtin_ia32_pandd512_mask", + "llvm.x86.avx512.mask.pand.q.512" => "__builtin_ia32_pandq512_mask", + "llvm.x86.avx512.mask.pavg.b.128" => "__builtin_ia32_pavgb128_mask", + "llvm.x86.avx512.mask.pavg.b.256" => "__builtin_ia32_pavgb256_mask", + "llvm.x86.avx512.mask.pavg.b.512" => "__builtin_ia32_pavgb512_mask", + "llvm.x86.avx512.mask.pavg.w.128" => "__builtin_ia32_pavgw128_mask", + "llvm.x86.avx512.mask.pavg.w.256" => "__builtin_ia32_pavgw256_mask", + "llvm.x86.avx512.mask.pavg.w.512" => "__builtin_ia32_pavgw512_mask", + "llvm.x86.avx512.mask.pbroadcast.b.gpr.128" => "__builtin_ia32_pbroadcastb128_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.b.gpr.256" => "__builtin_ia32_pbroadcastb256_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.b.gpr.512" => "__builtin_ia32_pbroadcastb512_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.d.gpr.128" => "__builtin_ia32_pbroadcastd128_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.d.gpr.256" => "__builtin_ia32_pbroadcastd256_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.d.gpr.512" => "__builtin_ia32_pbroadcastd512_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.q.gpr.128" => "__builtin_ia32_pbroadcastq128_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.q.gpr.256" => "__builtin_ia32_pbroadcastq256_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.q.gpr.512" => "__builtin_ia32_pbroadcastq512_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.q.mem.512" => "__builtin_ia32_pbroadcastq512_mem_mask", + "llvm.x86.avx512.mask.pbroadcast.w.gpr.128" => "__builtin_ia32_pbroadcastw128_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.w.gpr.256" => "__builtin_ia32_pbroadcastw256_gpr_mask", + "llvm.x86.avx512.mask.pbroadcast.w.gpr.512" => "__builtin_ia32_pbroadcastw512_gpr_mask", + "llvm.x86.avx512.mask.pcmpeq.b.128" => "__builtin_ia32_pcmpeqb128_mask", + "llvm.x86.avx512.mask.pcmpeq.b.256" => "__builtin_ia32_pcmpeqb256_mask", + "llvm.x86.avx512.mask.pcmpeq.b.512" => "__builtin_ia32_pcmpeqb512_mask", + "llvm.x86.avx512.mask.pcmpeq.d.128" => "__builtin_ia32_pcmpeqd128_mask", + "llvm.x86.avx512.mask.pcmpeq.d.256" => "__builtin_ia32_pcmpeqd256_mask", + "llvm.x86.avx512.mask.pcmpeq.d.512" => "__builtin_ia32_pcmpeqd512_mask", + "llvm.x86.avx512.mask.pcmpeq.q.128" => "__builtin_ia32_pcmpeqq128_mask", + "llvm.x86.avx512.mask.pcmpeq.q.256" => "__builtin_ia32_pcmpeqq256_mask", + "llvm.x86.avx512.mask.pcmpeq.q.512" => "__builtin_ia32_pcmpeqq512_mask", + "llvm.x86.avx512.mask.pcmpeq.w.128" => "__builtin_ia32_pcmpeqw128_mask", + "llvm.x86.avx512.mask.pcmpeq.w.256" => "__builtin_ia32_pcmpeqw256_mask", + "llvm.x86.avx512.mask.pcmpeq.w.512" => "__builtin_ia32_pcmpeqw512_mask", + "llvm.x86.avx512.mask.pcmpgt.b.128" => "__builtin_ia32_pcmpgtb128_mask", + "llvm.x86.avx512.mask.pcmpgt.b.256" => "__builtin_ia32_pcmpgtb256_mask", + "llvm.x86.avx512.mask.pcmpgt.b.512" => "__builtin_ia32_pcmpgtb512_mask", + "llvm.x86.avx512.mask.pcmpgt.d.128" => "__builtin_ia32_pcmpgtd128_mask", + "llvm.x86.avx512.mask.pcmpgt.d.256" => "__builtin_ia32_pcmpgtd256_mask", + "llvm.x86.avx512.mask.pcmpgt.d.512" => "__builtin_ia32_pcmpgtd512_mask", + "llvm.x86.avx512.mask.pcmpgt.q.128" => "__builtin_ia32_pcmpgtq128_mask", + "llvm.x86.avx512.mask.pcmpgt.q.256" => "__builtin_ia32_pcmpgtq256_mask", + "llvm.x86.avx512.mask.pcmpgt.q.512" => "__builtin_ia32_pcmpgtq512_mask", + "llvm.x86.avx512.mask.pcmpgt.w.128" => "__builtin_ia32_pcmpgtw128_mask", + "llvm.x86.avx512.mask.pcmpgt.w.256" => "__builtin_ia32_pcmpgtw256_mask", + "llvm.x86.avx512.mask.pcmpgt.w.512" => "__builtin_ia32_pcmpgtw512_mask", + "llvm.x86.avx512.mask.permvar.df.256" => "__builtin_ia32_permvardf256_mask", + "llvm.x86.avx512.mask.permvar.df.512" => "__builtin_ia32_permvardf512_mask", + "llvm.x86.avx512.mask.permvar.di.256" => "__builtin_ia32_permvardi256_mask", + "llvm.x86.avx512.mask.permvar.di.512" => "__builtin_ia32_permvardi512_mask", + "llvm.x86.avx512.mask.permvar.hi.128" => "__builtin_ia32_permvarhi128_mask", + "llvm.x86.avx512.mask.permvar.hi.256" => "__builtin_ia32_permvarhi256_mask", + "llvm.x86.avx512.mask.permvar.hi.512" => "__builtin_ia32_permvarhi512_mask", + "llvm.x86.avx512.mask.permvar.qi.128" => "__builtin_ia32_permvarqi128_mask", + "llvm.x86.avx512.mask.permvar.qi.256" => "__builtin_ia32_permvarqi256_mask", + "llvm.x86.avx512.mask.permvar.qi.512" => "__builtin_ia32_permvarqi512_mask", + "llvm.x86.avx512.mask.permvar.sf.256" => "__builtin_ia32_permvarsf256_mask", + "llvm.x86.avx512.mask.permvar.sf.512" => "__builtin_ia32_permvarsf512_mask", + "llvm.x86.avx512.mask.permvar.si.256" => "__builtin_ia32_permvarsi256_mask", + "llvm.x86.avx512.mask.permvar.si.512" => "__builtin_ia32_permvarsi512_mask", + "llvm.x86.avx512.mask.pmaddubs.w.128" => "__builtin_ia32_pmaddubsw128_mask", + "llvm.x86.avx512.mask.pmaddubs.w.256" => "__builtin_ia32_pmaddubsw256_mask", + "llvm.x86.avx512.mask.pmaddubs.w.512" => "__builtin_ia32_pmaddubsw512_mask", + "llvm.x86.avx512.mask.pmaddw.d.128" => "__builtin_ia32_pmaddwd128_mask", + "llvm.x86.avx512.mask.pmaddw.d.256" => "__builtin_ia32_pmaddwd256_mask", + "llvm.x86.avx512.mask.pmaddw.d.512" => "__builtin_ia32_pmaddwd512_mask", + "llvm.x86.avx512.mask.pmaxs.b.128" => "__builtin_ia32_pmaxsb128_mask", + "llvm.x86.avx512.mask.pmaxs.b.256" => "__builtin_ia32_pmaxsb256_mask", + "llvm.x86.avx512.mask.pmaxs.b.512" => "__builtin_ia32_pmaxsb512_mask", + "llvm.x86.avx512.mask.pmaxs.d.128" => "__builtin_ia32_pmaxsd128_mask", + "llvm.x86.avx512.mask.pmaxs.d.256" => "__builtin_ia32_pmaxsd256_mask", + "llvm.x86.avx512.mask.pmaxs.d.512" => "__builtin_ia32_pmaxsd512_mask", + "llvm.x86.avx512.mask.pmaxs.q.128" => "__builtin_ia32_pmaxsq128_mask", + "llvm.x86.avx512.mask.pmaxs.q.256" => "__builtin_ia32_pmaxsq256_mask", + "llvm.x86.avx512.mask.pmaxs.q.512" => "__builtin_ia32_pmaxsq512_mask", + "llvm.x86.avx512.mask.pmaxs.w.128" => "__builtin_ia32_pmaxsw128_mask", + "llvm.x86.avx512.mask.pmaxs.w.256" => "__builtin_ia32_pmaxsw256_mask", + "llvm.x86.avx512.mask.pmaxs.w.512" => "__builtin_ia32_pmaxsw512_mask", + "llvm.x86.avx512.mask.pmaxu.b.128" => "__builtin_ia32_pmaxub128_mask", + "llvm.x86.avx512.mask.pmaxu.b.256" => "__builtin_ia32_pmaxub256_mask", + "llvm.x86.avx512.mask.pmaxu.b.512" => "__builtin_ia32_pmaxub512_mask", + "llvm.x86.avx512.mask.pmaxu.d.128" => "__builtin_ia32_pmaxud128_mask", + "llvm.x86.avx512.mask.pmaxu.d.256" => "__builtin_ia32_pmaxud256_mask", + "llvm.x86.avx512.mask.pmaxu.d.512" => "__builtin_ia32_pmaxud512_mask", + "llvm.x86.avx512.mask.pmaxu.q.128" => "__builtin_ia32_pmaxuq128_mask", + "llvm.x86.avx512.mask.pmaxu.q.256" => "__builtin_ia32_pmaxuq256_mask", + "llvm.x86.avx512.mask.pmaxu.q.512" => "__builtin_ia32_pmaxuq512_mask", + "llvm.x86.avx512.mask.pmaxu.w.128" => "__builtin_ia32_pmaxuw128_mask", + "llvm.x86.avx512.mask.pmaxu.w.256" => "__builtin_ia32_pmaxuw256_mask", + "llvm.x86.avx512.mask.pmaxu.w.512" => "__builtin_ia32_pmaxuw512_mask", + "llvm.x86.avx512.mask.pmins.b.128" => "__builtin_ia32_pminsb128_mask", + "llvm.x86.avx512.mask.pmins.b.256" => "__builtin_ia32_pminsb256_mask", + "llvm.x86.avx512.mask.pmins.b.512" => "__builtin_ia32_pminsb512_mask", + "llvm.x86.avx512.mask.pmins.d.128" => "__builtin_ia32_pminsd128_mask", + "llvm.x86.avx512.mask.pmins.d.256" => "__builtin_ia32_pminsd256_mask", + "llvm.x86.avx512.mask.pmins.d.512" => "__builtin_ia32_pminsd512_mask", + "llvm.x86.avx512.mask.pmins.q.128" => "__builtin_ia32_pminsq128_mask", + "llvm.x86.avx512.mask.pmins.q.256" => "__builtin_ia32_pminsq256_mask", + "llvm.x86.avx512.mask.pmins.q.512" => "__builtin_ia32_pminsq512_mask", + "llvm.x86.avx512.mask.pmins.w.128" => "__builtin_ia32_pminsw128_mask", + "llvm.x86.avx512.mask.pmins.w.256" => "__builtin_ia32_pminsw256_mask", + "llvm.x86.avx512.mask.pmins.w.512" => "__builtin_ia32_pminsw512_mask", + "llvm.x86.avx512.mask.pminu.b.128" => "__builtin_ia32_pminub128_mask", + "llvm.x86.avx512.mask.pminu.b.256" => "__builtin_ia32_pminub256_mask", + "llvm.x86.avx512.mask.pminu.b.512" => "__builtin_ia32_pminub512_mask", + "llvm.x86.avx512.mask.pminu.d.128" => "__builtin_ia32_pminud128_mask", + "llvm.x86.avx512.mask.pminu.d.256" => "__builtin_ia32_pminud256_mask", + "llvm.x86.avx512.mask.pminu.d.512" => "__builtin_ia32_pminud512_mask", + "llvm.x86.avx512.mask.pminu.q.128" => "__builtin_ia32_pminuq128_mask", + "llvm.x86.avx512.mask.pminu.q.256" => "__builtin_ia32_pminuq256_mask", + "llvm.x86.avx512.mask.pminu.q.512" => "__builtin_ia32_pminuq512_mask", + "llvm.x86.avx512.mask.pminu.w.128" => "__builtin_ia32_pminuw128_mask", + "llvm.x86.avx512.mask.pminu.w.256" => "__builtin_ia32_pminuw256_mask", + "llvm.x86.avx512.mask.pminu.w.512" => "__builtin_ia32_pminuw512_mask", + "llvm.x86.avx512.mask.pmov.db.128" => "__builtin_ia32_pmovdb128_mask", + "llvm.x86.avx512.mask.pmov.db.256" => "__builtin_ia32_pmovdb256_mask", + "llvm.x86.avx512.mask.pmov.db.512" => "__builtin_ia32_pmovdb512_mask", + "llvm.x86.avx512.mask.pmov.db.mem.128" => "__builtin_ia32_pmovdb128mem_mask", + "llvm.x86.avx512.mask.pmov.db.mem.256" => "__builtin_ia32_pmovdb256mem_mask", + "llvm.x86.avx512.mask.pmov.db.mem.512" => "__builtin_ia32_pmovdb512mem_mask", + "llvm.x86.avx512.mask.pmov.dw.128" => "__builtin_ia32_pmovdw128_mask", + "llvm.x86.avx512.mask.pmov.dw.256" => "__builtin_ia32_pmovdw256_mask", + "llvm.x86.avx512.mask.pmov.dw.512" => "__builtin_ia32_pmovdw512_mask", + "llvm.x86.avx512.mask.pmov.dw.mem.128" => "__builtin_ia32_pmovdw128mem_mask", + "llvm.x86.avx512.mask.pmov.dw.mem.256" => "__builtin_ia32_pmovdw256mem_mask", + "llvm.x86.avx512.mask.pmov.dw.mem.512" => "__builtin_ia32_pmovdw512mem_mask", + "llvm.x86.avx512.mask.pmov.qb.128" => "__builtin_ia32_pmovqb128_mask", + "llvm.x86.avx512.mask.pmov.qb.256" => "__builtin_ia32_pmovqb256_mask", + "llvm.x86.avx512.mask.pmov.qb.512" => "__builtin_ia32_pmovqb512_mask", + "llvm.x86.avx512.mask.pmov.qb.mem.128" => "__builtin_ia32_pmovqb128mem_mask", + "llvm.x86.avx512.mask.pmov.qb.mem.256" => "__builtin_ia32_pmovqb256mem_mask", + "llvm.x86.avx512.mask.pmov.qb.mem.512" => "__builtin_ia32_pmovqb512mem_mask", + "llvm.x86.avx512.mask.pmov.qd.128" => "__builtin_ia32_pmovqd128_mask", + "llvm.x86.avx512.mask.pmov.qd.256" => "__builtin_ia32_pmovqd256_mask", + "llvm.x86.avx512.mask.pmov.qd.512" => "__builtin_ia32_pmovqd512_mask", + "llvm.x86.avx512.mask.pmov.qd.mem.128" => "__builtin_ia32_pmovqd128mem_mask", + "llvm.x86.avx512.mask.pmov.qd.mem.256" => "__builtin_ia32_pmovqd256mem_mask", + "llvm.x86.avx512.mask.pmov.qd.mem.512" => "__builtin_ia32_pmovqd512mem_mask", + "llvm.x86.avx512.mask.pmov.qw.128" => "__builtin_ia32_pmovqw128_mask", + "llvm.x86.avx512.mask.pmov.qw.256" => "__builtin_ia32_pmovqw256_mask", + "llvm.x86.avx512.mask.pmov.qw.512" => "__builtin_ia32_pmovqw512_mask", + "llvm.x86.avx512.mask.pmov.qw.mem.128" => "__builtin_ia32_pmovqw128mem_mask", + "llvm.x86.avx512.mask.pmov.qw.mem.256" => "__builtin_ia32_pmovqw256mem_mask", + "llvm.x86.avx512.mask.pmov.qw.mem.512" => "__builtin_ia32_pmovqw512mem_mask", + "llvm.x86.avx512.mask.pmov.wb.128" => "__builtin_ia32_pmovwb128_mask", + "llvm.x86.avx512.mask.pmov.wb.256" => "__builtin_ia32_pmovwb256_mask", + "llvm.x86.avx512.mask.pmov.wb.512" => "__builtin_ia32_pmovwb512_mask", + "llvm.x86.avx512.mask.pmov.wb.mem.128" => "__builtin_ia32_pmovwb128mem_mask", + "llvm.x86.avx512.mask.pmov.wb.mem.256" => "__builtin_ia32_pmovwb256mem_mask", + "llvm.x86.avx512.mask.pmov.wb.mem.512" => "__builtin_ia32_pmovwb512mem_mask", + "llvm.x86.avx512.mask.pmovs.db.128" => "__builtin_ia32_pmovsdb128_mask", + "llvm.x86.avx512.mask.pmovs.db.256" => "__builtin_ia32_pmovsdb256_mask", + "llvm.x86.avx512.mask.pmovs.db.512" => "__builtin_ia32_pmovsdb512_mask", + "llvm.x86.avx512.mask.pmovs.db.mem.128" => "__builtin_ia32_pmovsdb128mem_mask", + "llvm.x86.avx512.mask.pmovs.db.mem.256" => "__builtin_ia32_pmovsdb256mem_mask", + "llvm.x86.avx512.mask.pmovs.db.mem.512" => "__builtin_ia32_pmovsdb512mem_mask", + "llvm.x86.avx512.mask.pmovs.dw.128" => "__builtin_ia32_pmovsdw128_mask", + "llvm.x86.avx512.mask.pmovs.dw.256" => "__builtin_ia32_pmovsdw256_mask", + "llvm.x86.avx512.mask.pmovs.dw.512" => "__builtin_ia32_pmovsdw512_mask", + "llvm.x86.avx512.mask.pmovs.dw.mem.128" => "__builtin_ia32_pmovsdw128mem_mask", + "llvm.x86.avx512.mask.pmovs.dw.mem.256" => "__builtin_ia32_pmovsdw256mem_mask", + "llvm.x86.avx512.mask.pmovs.dw.mem.512" => "__builtin_ia32_pmovsdw512mem_mask", + "llvm.x86.avx512.mask.pmovs.qb.128" => "__builtin_ia32_pmovsqb128_mask", + "llvm.x86.avx512.mask.pmovs.qb.256" => "__builtin_ia32_pmovsqb256_mask", + "llvm.x86.avx512.mask.pmovs.qb.512" => "__builtin_ia32_pmovsqb512_mask", + "llvm.x86.avx512.mask.pmovs.qb.mem.128" => "__builtin_ia32_pmovsqb128mem_mask", + "llvm.x86.avx512.mask.pmovs.qb.mem.256" => "__builtin_ia32_pmovsqb256mem_mask", + "llvm.x86.avx512.mask.pmovs.qb.mem.512" => "__builtin_ia32_pmovsqb512mem_mask", + "llvm.x86.avx512.mask.pmovs.qd.128" => "__builtin_ia32_pmovsqd128_mask", + "llvm.x86.avx512.mask.pmovs.qd.256" => "__builtin_ia32_pmovsqd256_mask", + "llvm.x86.avx512.mask.pmovs.qd.512" => "__builtin_ia32_pmovsqd512_mask", + "llvm.x86.avx512.mask.pmovs.qd.mem.128" => "__builtin_ia32_pmovsqd128mem_mask", + "llvm.x86.avx512.mask.pmovs.qd.mem.256" => "__builtin_ia32_pmovsqd256mem_mask", + "llvm.x86.avx512.mask.pmovs.qd.mem.512" => "__builtin_ia32_pmovsqd512mem_mask", + "llvm.x86.avx512.mask.pmovs.qw.128" => "__builtin_ia32_pmovsqw128_mask", + "llvm.x86.avx512.mask.pmovs.qw.256" => "__builtin_ia32_pmovsqw256_mask", + "llvm.x86.avx512.mask.pmovs.qw.512" => "__builtin_ia32_pmovsqw512_mask", + "llvm.x86.avx512.mask.pmovs.qw.mem.128" => "__builtin_ia32_pmovsqw128mem_mask", + "llvm.x86.avx512.mask.pmovs.qw.mem.256" => "__builtin_ia32_pmovsqw256mem_mask", + "llvm.x86.avx512.mask.pmovs.qw.mem.512" => "__builtin_ia32_pmovsqw512mem_mask", + "llvm.x86.avx512.mask.pmovs.wb.128" => "__builtin_ia32_pmovswb128_mask", + "llvm.x86.avx512.mask.pmovs.wb.256" => "__builtin_ia32_pmovswb256_mask", + "llvm.x86.avx512.mask.pmovs.wb.512" => "__builtin_ia32_pmovswb512_mask", + "llvm.x86.avx512.mask.pmovs.wb.mem.128" => "__builtin_ia32_pmovswb128mem_mask", + "llvm.x86.avx512.mask.pmovs.wb.mem.256" => "__builtin_ia32_pmovswb256mem_mask", + "llvm.x86.avx512.mask.pmovs.wb.mem.512" => "__builtin_ia32_pmovswb512mem_mask", + "llvm.x86.avx512.mask.pmovsxb.d.128" => "__builtin_ia32_pmovsxbd128_mask", + "llvm.x86.avx512.mask.pmovsxb.d.256" => "__builtin_ia32_pmovsxbd256_mask", + "llvm.x86.avx512.mask.pmovsxb.d.512" => "__builtin_ia32_pmovsxbd512_mask", + "llvm.x86.avx512.mask.pmovsxb.q.128" => "__builtin_ia32_pmovsxbq128_mask", + "llvm.x86.avx512.mask.pmovsxb.q.256" => "__builtin_ia32_pmovsxbq256_mask", + "llvm.x86.avx512.mask.pmovsxb.q.512" => "__builtin_ia32_pmovsxbq512_mask", + "llvm.x86.avx512.mask.pmovsxb.w.128" => "__builtin_ia32_pmovsxbw128_mask", + "llvm.x86.avx512.mask.pmovsxb.w.256" => "__builtin_ia32_pmovsxbw256_mask", + "llvm.x86.avx512.mask.pmovsxb.w.512" => "__builtin_ia32_pmovsxbw512_mask", + "llvm.x86.avx512.mask.pmovsxd.q.128" => "__builtin_ia32_pmovsxdq128_mask", + "llvm.x86.avx512.mask.pmovsxd.q.256" => "__builtin_ia32_pmovsxdq256_mask", + "llvm.x86.avx512.mask.pmovsxd.q.512" => "__builtin_ia32_pmovsxdq512_mask", + "llvm.x86.avx512.mask.pmovsxw.d.128" => "__builtin_ia32_pmovsxwd128_mask", + "llvm.x86.avx512.mask.pmovsxw.d.256" => "__builtin_ia32_pmovsxwd256_mask", + "llvm.x86.avx512.mask.pmovsxw.d.512" => "__builtin_ia32_pmovsxwd512_mask", + "llvm.x86.avx512.mask.pmovsxw.q.128" => "__builtin_ia32_pmovsxwq128_mask", + "llvm.x86.avx512.mask.pmovsxw.q.256" => "__builtin_ia32_pmovsxwq256_mask", + "llvm.x86.avx512.mask.pmovsxw.q.512" => "__builtin_ia32_pmovsxwq512_mask", + "llvm.x86.avx512.mask.pmovus.db.128" => "__builtin_ia32_pmovusdb128_mask", + "llvm.x86.avx512.mask.pmovus.db.256" => "__builtin_ia32_pmovusdb256_mask", + "llvm.x86.avx512.mask.pmovus.db.512" => "__builtin_ia32_pmovusdb512_mask", + "llvm.x86.avx512.mask.pmovus.db.mem.128" => "__builtin_ia32_pmovusdb128mem_mask", + "llvm.x86.avx512.mask.pmovus.db.mem.256" => "__builtin_ia32_pmovusdb256mem_mask", + "llvm.x86.avx512.mask.pmovus.db.mem.512" => "__builtin_ia32_pmovusdb512mem_mask", + "llvm.x86.avx512.mask.pmovus.dw.128" => "__builtin_ia32_pmovusdw128_mask", + "llvm.x86.avx512.mask.pmovus.dw.256" => "__builtin_ia32_pmovusdw256_mask", + "llvm.x86.avx512.mask.pmovus.dw.512" => "__builtin_ia32_pmovusdw512_mask", + "llvm.x86.avx512.mask.pmovus.dw.mem.128" => "__builtin_ia32_pmovusdw128mem_mask", + "llvm.x86.avx512.mask.pmovus.dw.mem.256" => "__builtin_ia32_pmovusdw256mem_mask", + "llvm.x86.avx512.mask.pmovus.dw.mem.512" => "__builtin_ia32_pmovusdw512mem_mask", + "llvm.x86.avx512.mask.pmovus.qb.128" => "__builtin_ia32_pmovusqb128_mask", + "llvm.x86.avx512.mask.pmovus.qb.256" => "__builtin_ia32_pmovusqb256_mask", + "llvm.x86.avx512.mask.pmovus.qb.512" => "__builtin_ia32_pmovusqb512_mask", + "llvm.x86.avx512.mask.pmovus.qb.mem.128" => "__builtin_ia32_pmovusqb128mem_mask", + "llvm.x86.avx512.mask.pmovus.qb.mem.256" => "__builtin_ia32_pmovusqb256mem_mask", + "llvm.x86.avx512.mask.pmovus.qb.mem.512" => "__builtin_ia32_pmovusqb512mem_mask", + "llvm.x86.avx512.mask.pmovus.qd.128" => "__builtin_ia32_pmovusqd128_mask", + "llvm.x86.avx512.mask.pmovus.qd.256" => "__builtin_ia32_pmovusqd256_mask", + "llvm.x86.avx512.mask.pmovus.qd.512" => "__builtin_ia32_pmovusqd512_mask", + "llvm.x86.avx512.mask.pmovus.qd.mem.128" => "__builtin_ia32_pmovusqd128mem_mask", + "llvm.x86.avx512.mask.pmovus.qd.mem.256" => "__builtin_ia32_pmovusqd256mem_mask", + "llvm.x86.avx512.mask.pmovus.qd.mem.512" => "__builtin_ia32_pmovusqd512mem_mask", + "llvm.x86.avx512.mask.pmovus.qw.128" => "__builtin_ia32_pmovusqw128_mask", + "llvm.x86.avx512.mask.pmovus.qw.256" => "__builtin_ia32_pmovusqw256_mask", + "llvm.x86.avx512.mask.pmovus.qw.512" => "__builtin_ia32_pmovusqw512_mask", + "llvm.x86.avx512.mask.pmovus.qw.mem.128" => "__builtin_ia32_pmovusqw128mem_mask", + "llvm.x86.avx512.mask.pmovus.qw.mem.256" => "__builtin_ia32_pmovusqw256mem_mask", + "llvm.x86.avx512.mask.pmovus.qw.mem.512" => "__builtin_ia32_pmovusqw512mem_mask", + "llvm.x86.avx512.mask.pmovus.wb.128" => "__builtin_ia32_pmovuswb128_mask", + "llvm.x86.avx512.mask.pmovus.wb.256" => "__builtin_ia32_pmovuswb256_mask", + "llvm.x86.avx512.mask.pmovus.wb.512" => "__builtin_ia32_pmovuswb512_mask", + "llvm.x86.avx512.mask.pmovus.wb.mem.128" => "__builtin_ia32_pmovuswb128mem_mask", + "llvm.x86.avx512.mask.pmovus.wb.mem.256" => "__builtin_ia32_pmovuswb256mem_mask", + "llvm.x86.avx512.mask.pmovus.wb.mem.512" => "__builtin_ia32_pmovuswb512mem_mask", + "llvm.x86.avx512.mask.pmovzxb.d.128" => "__builtin_ia32_pmovzxbd128_mask", + "llvm.x86.avx512.mask.pmovzxb.d.256" => "__builtin_ia32_pmovzxbd256_mask", + "llvm.x86.avx512.mask.pmovzxb.d.512" => "__builtin_ia32_pmovzxbd512_mask", + "llvm.x86.avx512.mask.pmovzxb.q.128" => "__builtin_ia32_pmovzxbq128_mask", + "llvm.x86.avx512.mask.pmovzxb.q.256" => "__builtin_ia32_pmovzxbq256_mask", + "llvm.x86.avx512.mask.pmovzxb.q.512" => "__builtin_ia32_pmovzxbq512_mask", + "llvm.x86.avx512.mask.pmovzxb.w.128" => "__builtin_ia32_pmovzxbw128_mask", + "llvm.x86.avx512.mask.pmovzxb.w.256" => "__builtin_ia32_pmovzxbw256_mask", + "llvm.x86.avx512.mask.pmovzxb.w.512" => "__builtin_ia32_pmovzxbw512_mask", + "llvm.x86.avx512.mask.pmovzxd.q.128" => "__builtin_ia32_pmovzxdq128_mask", + "llvm.x86.avx512.mask.pmovzxd.q.256" => "__builtin_ia32_pmovzxdq256_mask", + "llvm.x86.avx512.mask.pmovzxd.q.512" => "__builtin_ia32_pmovzxdq512_mask", + "llvm.x86.avx512.mask.pmovzxw.d.128" => "__builtin_ia32_pmovzxwd128_mask", + "llvm.x86.avx512.mask.pmovzxw.d.256" => "__builtin_ia32_pmovzxwd256_mask", + "llvm.x86.avx512.mask.pmovzxw.d.512" => "__builtin_ia32_pmovzxwd512_mask", + "llvm.x86.avx512.mask.pmovzxw.q.128" => "__builtin_ia32_pmovzxwq128_mask", + "llvm.x86.avx512.mask.pmovzxw.q.256" => "__builtin_ia32_pmovzxwq256_mask", + "llvm.x86.avx512.mask.pmovzxw.q.512" => "__builtin_ia32_pmovzxwq512_mask", + "llvm.x86.avx512.mask.pmul.dq.128" => "__builtin_ia32_pmuldq128_mask", + "llvm.x86.avx512.mask.pmul.dq.256" => "__builtin_ia32_pmuldq256_mask", + "llvm.x86.avx512.mask.pmul.dq.512" => "__builtin_ia32_pmuldq512_mask", + "llvm.x86.avx512.mask.pmul.hr.sw.128" => "__builtin_ia32_pmulhrsw128_mask", + "llvm.x86.avx512.mask.pmul.hr.sw.256" => "__builtin_ia32_pmulhrsw256_mask", + "llvm.x86.avx512.mask.pmul.hr.sw.512" => "__builtin_ia32_pmulhrsw512_mask", + "llvm.x86.avx512.mask.pmulh.w.128" => "__builtin_ia32_pmulhw128_mask", + "llvm.x86.avx512.mask.pmulh.w.256" => "__builtin_ia32_pmulhw256_mask", + "llvm.x86.avx512.mask.pmulh.w.512" => "__builtin_ia32_pmulhw512_mask", + "llvm.x86.avx512.mask.pmulhu.w.128" => "__builtin_ia32_pmulhuw128_mask", + "llvm.x86.avx512.mask.pmulhu.w.256" => "__builtin_ia32_pmulhuw256_mask", + "llvm.x86.avx512.mask.pmulhu.w.512" => "__builtin_ia32_pmulhuw512_mask", + "llvm.x86.avx512.mask.pmull.d.128" => "__builtin_ia32_pmulld128_mask", + "llvm.x86.avx512.mask.pmull.d.256" => "__builtin_ia32_pmulld256_mask", + "llvm.x86.avx512.mask.pmull.d.512" => "__builtin_ia32_pmulld512_mask", + "llvm.x86.avx512.mask.pmull.q.128" => "__builtin_ia32_pmullq128_mask", + "llvm.x86.avx512.mask.pmull.q.256" => "__builtin_ia32_pmullq256_mask", + "llvm.x86.avx512.mask.pmull.q.512" => "__builtin_ia32_pmullq512_mask", + "llvm.x86.avx512.mask.pmull.w.128" => "__builtin_ia32_pmullw128_mask", + "llvm.x86.avx512.mask.pmull.w.256" => "__builtin_ia32_pmullw256_mask", + "llvm.x86.avx512.mask.pmull.w.512" => "__builtin_ia32_pmullw512_mask", + "llvm.x86.avx512.mask.pmultishift.qb.128" => "__builtin_ia32_vpmultishiftqb128_mask", + "llvm.x86.avx512.mask.pmultishift.qb.256" => "__builtin_ia32_vpmultishiftqb256_mask", + "llvm.x86.avx512.mask.pmultishift.qb.512" => "__builtin_ia32_vpmultishiftqb512_mask", + "llvm.x86.avx512.mask.pmulu.dq.128" => "__builtin_ia32_pmuludq128_mask", + "llvm.x86.avx512.mask.pmulu.dq.256" => "__builtin_ia32_pmuludq256_mask", + "llvm.x86.avx512.mask.pmulu.dq.512" => "__builtin_ia32_pmuludq512_mask", + "llvm.x86.avx512.mask.prol.d.128" => "__builtin_ia32_prold128_mask", + "llvm.x86.avx512.mask.prol.d.256" => "__builtin_ia32_prold256_mask", + "llvm.x86.avx512.mask.prol.d.512" => "__builtin_ia32_prold512_mask", + "llvm.x86.avx512.mask.prol.q.128" => "__builtin_ia32_prolq128_mask", + "llvm.x86.avx512.mask.prol.q.256" => "__builtin_ia32_prolq256_mask", + "llvm.x86.avx512.mask.prol.q.512" => "__builtin_ia32_prolq512_mask", + "llvm.x86.avx512.mask.prolv.d.128" => "__builtin_ia32_prolvd128_mask", + "llvm.x86.avx512.mask.prolv.d.256" => "__builtin_ia32_prolvd256_mask", + "llvm.x86.avx512.mask.prolv.d.512" => "__builtin_ia32_prolvd512_mask", + "llvm.x86.avx512.mask.prolv.q.128" => "__builtin_ia32_prolvq128_mask", + "llvm.x86.avx512.mask.prolv.q.256" => "__builtin_ia32_prolvq256_mask", + "llvm.x86.avx512.mask.prolv.q.512" => "__builtin_ia32_prolvq512_mask", + "llvm.x86.avx512.mask.pror.d.128" => "__builtin_ia32_prord128_mask", + "llvm.x86.avx512.mask.pror.d.256" => "__builtin_ia32_prord256_mask", + "llvm.x86.avx512.mask.pror.d.512" => "__builtin_ia32_prord512_mask", + "llvm.x86.avx512.mask.pror.q.128" => "__builtin_ia32_prorq128_mask", + "llvm.x86.avx512.mask.pror.q.256" => "__builtin_ia32_prorq256_mask", + "llvm.x86.avx512.mask.pror.q.512" => "__builtin_ia32_prorq512_mask", + "llvm.x86.avx512.mask.prorv.d.128" => "__builtin_ia32_prorvd128_mask", + "llvm.x86.avx512.mask.prorv.d.256" => "__builtin_ia32_prorvd256_mask", + "llvm.x86.avx512.mask.prorv.d.512" => "__builtin_ia32_prorvd512_mask", + "llvm.x86.avx512.mask.prorv.q.128" => "__builtin_ia32_prorvq128_mask", + "llvm.x86.avx512.mask.prorv.q.256" => "__builtin_ia32_prorvq256_mask", + "llvm.x86.avx512.mask.prorv.q.512" => "__builtin_ia32_prorvq512_mask", + "llvm.x86.avx512.mask.pshuf.b.128" => "__builtin_ia32_pshufb128_mask", + "llvm.x86.avx512.mask.pshuf.b.256" => "__builtin_ia32_pshufb256_mask", + "llvm.x86.avx512.mask.pshuf.b.512" => "__builtin_ia32_pshufb512_mask", + "llvm.x86.avx512.mask.psll.d" => "__builtin_ia32_pslld512_mask", + "llvm.x86.avx512.mask.psll.d.128" => "__builtin_ia32_pslld128_mask", + "llvm.x86.avx512.mask.psll.d.256" => "__builtin_ia32_pslld256_mask", + "llvm.x86.avx512.mask.psll.di.128" => "__builtin_ia32_pslldi128_mask", + "llvm.x86.avx512.mask.psll.di.256" => "__builtin_ia32_pslldi256_mask", + "llvm.x86.avx512.mask.psll.di.512" => "__builtin_ia32_pslldi512_mask", + "llvm.x86.avx512.mask.psll.q" => "__builtin_ia32_psllq512_mask", + "llvm.x86.avx512.mask.psll.q.128" => "__builtin_ia32_psllq128_mask", + "llvm.x86.avx512.mask.psll.q.256" => "__builtin_ia32_psllq256_mask", + "llvm.x86.avx512.mask.psll.qi.128" => "__builtin_ia32_psllqi128_mask", + "llvm.x86.avx512.mask.psll.qi.256" => "__builtin_ia32_psllqi256_mask", + "llvm.x86.avx512.mask.psll.qi.512" => "__builtin_ia32_psllqi512_mask", + "llvm.x86.avx512.mask.psll.w.128" => "__builtin_ia32_psllw128_mask", + "llvm.x86.avx512.mask.psll.w.256" => "__builtin_ia32_psllw256_mask", + "llvm.x86.avx512.mask.psll.w.512" => "__builtin_ia32_psllw512_mask", + "llvm.x86.avx512.mask.psll.wi.128" => "__builtin_ia32_psllwi128_mask", + "llvm.x86.avx512.mask.psll.wi.256" => "__builtin_ia32_psllwi256_mask", + "llvm.x86.avx512.mask.psll.wi.512" => "__builtin_ia32_psllwi512_mask", + "llvm.x86.avx512.mask.psllv.d" => "__builtin_ia32_psllv16si_mask", + "llvm.x86.avx512.mask.psllv.q" => "__builtin_ia32_psllv8di_mask", + "llvm.x86.avx512.mask.psllv16.hi" => "__builtin_ia32_psllv16hi_mask", + "llvm.x86.avx512.mask.psllv2.di" => "__builtin_ia32_psllv2di_mask", + "llvm.x86.avx512.mask.psllv32hi" => "__builtin_ia32_psllv32hi_mask", + "llvm.x86.avx512.mask.psllv4.di" => "__builtin_ia32_psllv4di_mask", + "llvm.x86.avx512.mask.psllv4.si" => "__builtin_ia32_psllv4si_mask", + "llvm.x86.avx512.mask.psllv8.hi" => "__builtin_ia32_psllv8hi_mask", + "llvm.x86.avx512.mask.psllv8.si" => "__builtin_ia32_psllv8si_mask", + "llvm.x86.avx512.mask.psra.d" => "__builtin_ia32_psrad512_mask", + "llvm.x86.avx512.mask.psra.d.128" => "__builtin_ia32_psrad128_mask", + "llvm.x86.avx512.mask.psra.d.256" => "__builtin_ia32_psrad256_mask", + "llvm.x86.avx512.mask.psra.di.128" => "__builtin_ia32_psradi128_mask", + "llvm.x86.avx512.mask.psra.di.256" => "__builtin_ia32_psradi256_mask", + "llvm.x86.avx512.mask.psra.di.512" => "__builtin_ia32_psradi512_mask", + "llvm.x86.avx512.mask.psra.q" => "__builtin_ia32_psraq512_mask", + "llvm.x86.avx512.mask.psra.q.128" => "__builtin_ia32_psraq128_mask", + "llvm.x86.avx512.mask.psra.q.256" => "__builtin_ia32_psraq256_mask", + "llvm.x86.avx512.mask.psra.qi.128" => "__builtin_ia32_psraqi128_mask", + "llvm.x86.avx512.mask.psra.qi.256" => "__builtin_ia32_psraqi256_mask", + "llvm.x86.avx512.mask.psra.qi.512" => "__builtin_ia32_psraqi512_mask", + "llvm.x86.avx512.mask.psra.w.128" => "__builtin_ia32_psraw128_mask", + "llvm.x86.avx512.mask.psra.w.256" => "__builtin_ia32_psraw256_mask", + "llvm.x86.avx512.mask.psra.w.512" => "__builtin_ia32_psraw512_mask", + "llvm.x86.avx512.mask.psra.wi.128" => "__builtin_ia32_psrawi128_mask", + "llvm.x86.avx512.mask.psra.wi.256" => "__builtin_ia32_psrawi256_mask", + "llvm.x86.avx512.mask.psra.wi.512" => "__builtin_ia32_psrawi512_mask", + "llvm.x86.avx512.mask.psrav.d" => "__builtin_ia32_psrav16si_mask", + "llvm.x86.avx512.mask.psrav.q" => "__builtin_ia32_psrav8di_mask", + "llvm.x86.avx512.mask.psrav.q.128" => "__builtin_ia32_psravq128_mask", + "llvm.x86.avx512.mask.psrav.q.256" => "__builtin_ia32_psravq256_mask", + "llvm.x86.avx512.mask.psrav16.hi" => "__builtin_ia32_psrav16hi_mask", + "llvm.x86.avx512.mask.psrav32.hi" => "__builtin_ia32_psrav32hi_mask", + "llvm.x86.avx512.mask.psrav4.si" => "__builtin_ia32_psrav4si_mask", + "llvm.x86.avx512.mask.psrav8.hi" => "__builtin_ia32_psrav8hi_mask", + "llvm.x86.avx512.mask.psrav8.si" => "__builtin_ia32_psrav8si_mask", + "llvm.x86.avx512.mask.psrl.d" => "__builtin_ia32_psrld512_mask", + "llvm.x86.avx512.mask.psrl.d.128" => "__builtin_ia32_psrld128_mask", + "llvm.x86.avx512.mask.psrl.d.256" => "__builtin_ia32_psrld256_mask", + "llvm.x86.avx512.mask.psrl.di.128" => "__builtin_ia32_psrldi128_mask", + "llvm.x86.avx512.mask.psrl.di.256" => "__builtin_ia32_psrldi256_mask", + "llvm.x86.avx512.mask.psrl.di.512" => "__builtin_ia32_psrldi512_mask", + "llvm.x86.avx512.mask.psrl.q" => "__builtin_ia32_psrlq512_mask", + "llvm.x86.avx512.mask.psrl.q.128" => "__builtin_ia32_psrlq128_mask", + "llvm.x86.avx512.mask.psrl.q.256" => "__builtin_ia32_psrlq256_mask", + "llvm.x86.avx512.mask.psrl.qi.128" => "__builtin_ia32_psrlqi128_mask", + "llvm.x86.avx512.mask.psrl.qi.256" => "__builtin_ia32_psrlqi256_mask", + "llvm.x86.avx512.mask.psrl.qi.512" => "__builtin_ia32_psrlqi512_mask", + "llvm.x86.avx512.mask.psrl.w.128" => "__builtin_ia32_psrlw128_mask", + "llvm.x86.avx512.mask.psrl.w.256" => "__builtin_ia32_psrlw256_mask", + "llvm.x86.avx512.mask.psrl.w.512" => "__builtin_ia32_psrlw512_mask", + "llvm.x86.avx512.mask.psrl.wi.128" => "__builtin_ia32_psrlwi128_mask", + "llvm.x86.avx512.mask.psrl.wi.256" => "__builtin_ia32_psrlwi256_mask", + "llvm.x86.avx512.mask.psrl.wi.512" => "__builtin_ia32_psrlwi512_mask", + "llvm.x86.avx512.mask.psrlv.d" => "__builtin_ia32_psrlv16si_mask", + "llvm.x86.avx512.mask.psrlv.q" => "__builtin_ia32_psrlv8di_mask", + "llvm.x86.avx512.mask.psrlv16.hi" => "__builtin_ia32_psrlv16hi_mask", + "llvm.x86.avx512.mask.psrlv2.di" => "__builtin_ia32_psrlv2di_mask", + "llvm.x86.avx512.mask.psrlv32hi" => "__builtin_ia32_psrlv32hi_mask", + "llvm.x86.avx512.mask.psrlv4.di" => "__builtin_ia32_psrlv4di_mask", + "llvm.x86.avx512.mask.psrlv4.si" => "__builtin_ia32_psrlv4si_mask", + "llvm.x86.avx512.mask.psrlv8.hi" => "__builtin_ia32_psrlv8hi_mask", + "llvm.x86.avx512.mask.psrlv8.si" => "__builtin_ia32_psrlv8si_mask", + "llvm.x86.avx512.mask.psub.b.128" => "__builtin_ia32_psubb128_mask", + "llvm.x86.avx512.mask.psub.b.256" => "__builtin_ia32_psubb256_mask", + "llvm.x86.avx512.mask.psub.b.512" => "__builtin_ia32_psubb512_mask", + "llvm.x86.avx512.mask.psub.d.128" => "__builtin_ia32_psubd128_mask", + "llvm.x86.avx512.mask.psub.d.256" => "__builtin_ia32_psubd256_mask", + "llvm.x86.avx512.mask.psub.d.512" => "__builtin_ia32_psubd512_mask", + "llvm.x86.avx512.mask.psub.q.128" => "__builtin_ia32_psubq128_mask", + "llvm.x86.avx512.mask.psub.q.256" => "__builtin_ia32_psubq256_mask", + "llvm.x86.avx512.mask.psub.q.512" => "__builtin_ia32_psubq512_mask", + "llvm.x86.avx512.mask.psub.w.128" => "__builtin_ia32_psubw128_mask", + "llvm.x86.avx512.mask.psub.w.256" => "__builtin_ia32_psubw256_mask", + "llvm.x86.avx512.mask.psub.w.512" => "__builtin_ia32_psubw512_mask", + "llvm.x86.avx512.mask.psubs.b.128" => "__builtin_ia32_psubsb128_mask", + "llvm.x86.avx512.mask.psubs.b.256" => "__builtin_ia32_psubsb256_mask", + "llvm.x86.avx512.mask.psubs.b.512" => "__builtin_ia32_psubsb512_mask", + "llvm.x86.avx512.mask.psubs.w.128" => "__builtin_ia32_psubsw128_mask", + "llvm.x86.avx512.mask.psubs.w.256" => "__builtin_ia32_psubsw256_mask", + "llvm.x86.avx512.mask.psubs.w.512" => "__builtin_ia32_psubsw512_mask", + "llvm.x86.avx512.mask.psubus.b.128" => "__builtin_ia32_psubusb128_mask", + "llvm.x86.avx512.mask.psubus.b.256" => "__builtin_ia32_psubusb256_mask", + "llvm.x86.avx512.mask.psubus.b.512" => "__builtin_ia32_psubusb512_mask", + "llvm.x86.avx512.mask.psubus.w.128" => "__builtin_ia32_psubusw128_mask", + "llvm.x86.avx512.mask.psubus.w.256" => "__builtin_ia32_psubusw256_mask", + "llvm.x86.avx512.mask.psubus.w.512" => "__builtin_ia32_psubusw512_mask", + "llvm.x86.avx512.mask.pternlog.d.128" => "__builtin_ia32_pternlogd128_mask", + "llvm.x86.avx512.mask.pternlog.d.256" => "__builtin_ia32_pternlogd256_mask", + "llvm.x86.avx512.mask.pternlog.d.512" => "__builtin_ia32_pternlogd512_mask", + "llvm.x86.avx512.mask.pternlog.q.128" => "__builtin_ia32_pternlogq128_mask", + "llvm.x86.avx512.mask.pternlog.q.256" => "__builtin_ia32_pternlogq256_mask", + "llvm.x86.avx512.mask.pternlog.q.512" => "__builtin_ia32_pternlogq512_mask", + "llvm.x86.avx512.mask.ptestm.d.512" => "__builtin_ia32_ptestmd512", + "llvm.x86.avx512.mask.ptestm.q.512" => "__builtin_ia32_ptestmq512", + "llvm.x86.avx512.mask.range.pd.128" => "__builtin_ia32_rangepd128_mask", + "llvm.x86.avx512.mask.range.pd.256" => "__builtin_ia32_rangepd256_mask", + "llvm.x86.avx512.mask.range.pd.512" => "__builtin_ia32_rangepd512_mask", + "llvm.x86.avx512.mask.range.ps.128" => "__builtin_ia32_rangeps128_mask", + "llvm.x86.avx512.mask.range.ps.256" => "__builtin_ia32_rangeps256_mask", + "llvm.x86.avx512.mask.range.ps.512" => "__builtin_ia32_rangeps512_mask", + "llvm.x86.avx512.mask.range.sd" => "__builtin_ia32_rangesd128_round_mask", + "llvm.x86.avx512.mask.range.ss" => "__builtin_ia32_rangess128_round_mask", + "llvm.x86.avx512.mask.reduce.pd.128" => "__builtin_ia32_reducepd128_mask", + "llvm.x86.avx512.mask.reduce.pd.256" => "__builtin_ia32_reducepd256_mask", + "llvm.x86.avx512.mask.reduce.pd.512" => "__builtin_ia32_reducepd512_mask", + "llvm.x86.avx512.mask.reduce.ps.128" => "__builtin_ia32_reduceps128_mask", + "llvm.x86.avx512.mask.reduce.ps.256" => "__builtin_ia32_reduceps256_mask", + "llvm.x86.avx512.mask.reduce.ps.512" => "__builtin_ia32_reduceps512_mask", + "llvm.x86.avx512.mask.reduce.sd" => "__builtin_ia32_reducesd_mask", + "llvm.x86.avx512.mask.reduce.ss" => "__builtin_ia32_reducess_mask", + "llvm.x86.avx512.mask.rndscale.pd.128" => "__builtin_ia32_rndscalepd_128_mask", + "llvm.x86.avx512.mask.rndscale.pd.256" => "__builtin_ia32_rndscalepd_256_mask", + "llvm.x86.avx512.mask.rndscale.pd.512" => "__builtin_ia32_rndscalepd_mask", + "llvm.x86.avx512.mask.rndscale.ps.128" => "__builtin_ia32_rndscaleps_128_mask", + "llvm.x86.avx512.mask.rndscale.ps.256" => "__builtin_ia32_rndscaleps_256_mask", + "llvm.x86.avx512.mask.rndscale.ps.512" => "__builtin_ia32_rndscaleps_mask", + "llvm.x86.avx512.mask.rndscale.sd" => "__builtin_ia32_rndscalesd_round_mask", + "llvm.x86.avx512.mask.rndscale.ss" => "__builtin_ia32_rndscaless_round_mask", + "llvm.x86.avx512.mask.scalef.pd.128" => "__builtin_ia32_scalefpd128_mask", + "llvm.x86.avx512.mask.scalef.pd.256" => "__builtin_ia32_scalefpd256_mask", + "llvm.x86.avx512.mask.scalef.pd.512" => "__builtin_ia32_scalefpd512_mask", + "llvm.x86.avx512.mask.scalef.ps.128" => "__builtin_ia32_scalefps128_mask", + "llvm.x86.avx512.mask.scalef.ps.256" => "__builtin_ia32_scalefps256_mask", + "llvm.x86.avx512.mask.scalef.ps.512" => "__builtin_ia32_scalefps512_mask", + "llvm.x86.avx512.mask.scalef.sd" => "__builtin_ia32_scalefsd_round_mask", + "llvm.x86.avx512.mask.scalef.ss" => "__builtin_ia32_scalefss_round_mask", + "llvm.x86.avx512.mask.shuf.f32x4" => "__builtin_ia32_shuf_f32x4_mask", + "llvm.x86.avx512.mask.shuf.f32x4.256" => "__builtin_ia32_shuf_f32x4_256_mask", + "llvm.x86.avx512.mask.shuf.f64x2" => "__builtin_ia32_shuf_f64x2_mask", + "llvm.x86.avx512.mask.shuf.f64x2.256" => "__builtin_ia32_shuf_f64x2_256_mask", + "llvm.x86.avx512.mask.shuf.i32x4" => "__builtin_ia32_shuf_i32x4_mask", + "llvm.x86.avx512.mask.shuf.i32x4.256" => "__builtin_ia32_shuf_i32x4_256_mask", + "llvm.x86.avx512.mask.shuf.i64x2" => "__builtin_ia32_shuf_i64x2_mask", + "llvm.x86.avx512.mask.shuf.i64x2.256" => "__builtin_ia32_shuf_i64x2_256_mask", + "llvm.x86.avx512.mask.shuf.pd.128" => "__builtin_ia32_shufpd128_mask", + "llvm.x86.avx512.mask.shuf.pd.256" => "__builtin_ia32_shufpd256_mask", + "llvm.x86.avx512.mask.shuf.pd.512" => "__builtin_ia32_shufpd512_mask", + "llvm.x86.avx512.mask.shuf.ps.128" => "__builtin_ia32_shufps128_mask", + "llvm.x86.avx512.mask.shuf.ps.256" => "__builtin_ia32_shufps256_mask", + "llvm.x86.avx512.mask.shuf.ps.512" => "__builtin_ia32_shufps512_mask", + "llvm.x86.avx512.mask.sqrt.pd.128" => "__builtin_ia32_sqrtpd128_mask", + "llvm.x86.avx512.mask.sqrt.pd.256" => "__builtin_ia32_sqrtpd256_mask", + "llvm.x86.avx512.mask.sqrt.pd.512" => "__builtin_ia32_sqrtpd512_mask", + "llvm.x86.avx512.mask.sqrt.ps.128" => "__builtin_ia32_sqrtps128_mask", + "llvm.x86.avx512.mask.sqrt.ps.256" => "__builtin_ia32_sqrtps256_mask", + "llvm.x86.avx512.mask.sqrt.ps.512" => "__builtin_ia32_sqrtps512_mask", + "llvm.x86.avx512.mask.sqrt.sd" => "__builtin_ia32_sqrtsd_round_mask", + "llvm.x86.avx512.mask.sqrt.ss" => "__builtin_ia32_sqrtss_round_mask", + "llvm.x86.avx512.mask.store.ss" => "__builtin_ia32_storess_mask", + "llvm.x86.avx512.mask.storeu.d.512" => "__builtin_ia32_storedqusi512_mask", + "llvm.x86.avx512.mask.storeu.pd.512" => "__builtin_ia32_storeupd512_mask", + "llvm.x86.avx512.mask.storeu.ps.512" => "__builtin_ia32_storeups512_mask", + "llvm.x86.avx512.mask.storeu.q.512" => "__builtin_ia32_storedqudi512_mask", + "llvm.x86.avx512.mask.sub.pd.128" => "__builtin_ia32_subpd128_mask", + "llvm.x86.avx512.mask.sub.pd.256" => "__builtin_ia32_subpd256_mask", + "llvm.x86.avx512.mask.sub.pd.512" => "__builtin_ia32_subpd512_mask", + "llvm.x86.avx512.mask.sub.ps.128" => "__builtin_ia32_subps128_mask", + "llvm.x86.avx512.mask.sub.ps.256" => "__builtin_ia32_subps256_mask", + "llvm.x86.avx512.mask.sub.ps.512" => "__builtin_ia32_subps512_mask", + "llvm.x86.avx512.mask.sub.sd.round" => "__builtin_ia32_subsd_round_mask", + "llvm.x86.avx512.mask.sub.ss.round" => "__builtin_ia32_subss_round_mask", + "llvm.x86.avx512.mask.valign.d.128" => "__builtin_ia32_alignd128_mask", + "llvm.x86.avx512.mask.valign.d.256" => "__builtin_ia32_alignd256_mask", + "llvm.x86.avx512.mask.valign.d.512" => "__builtin_ia32_alignd512_mask", + "llvm.x86.avx512.mask.valign.q.128" => "__builtin_ia32_alignq128_mask", + "llvm.x86.avx512.mask.valign.q.256" => "__builtin_ia32_alignq256_mask", + "llvm.x86.avx512.mask.valign.q.512" => "__builtin_ia32_alignq512_mask", + "llvm.x86.avx512.mask.vcvtph2ps.128" => "__builtin_ia32_vcvtph2ps_mask", + "llvm.x86.avx512.mask.vcvtph2ps.256" => "__builtin_ia32_vcvtph2ps256_mask", + "llvm.x86.avx512.mask.vcvtph2ps.512" => "__builtin_ia32_vcvtph2ps512_mask", + "llvm.x86.avx512.mask.vcvtps2ph.128" => "__builtin_ia32_vcvtps2ph_mask", + "llvm.x86.avx512.mask.vcvtps2ph.256" => "__builtin_ia32_vcvtps2ph256_mask", + "llvm.x86.avx512.mask.vcvtps2ph.512" => "__builtin_ia32_vcvtps2ph512_mask", + "llvm.x86.avx512.mask.vextractf32x4.256" => "__builtin_ia32_extractf32x4_256_mask", + "llvm.x86.avx512.mask.vextractf32x4.512" => "__builtin_ia32_extractf32x4_mask", + "llvm.x86.avx512.mask.vextractf32x8.512" => "__builtin_ia32_extractf32x8_mask", + "llvm.x86.avx512.mask.vextractf64x2.256" => "__builtin_ia32_extractf64x2_256_mask", + "llvm.x86.avx512.mask.vextractf64x2.512" => "__builtin_ia32_extractf64x2_512_mask", + "llvm.x86.avx512.mask.vextractf64x4.512" => "__builtin_ia32_extractf64x4_mask", + "llvm.x86.avx512.mask.vextracti32x4.256" => "__builtin_ia32_extracti32x4_256_mask", + "llvm.x86.avx512.mask.vextracti32x4.512" => "__builtin_ia32_extracti32x4_mask", + "llvm.x86.avx512.mask.vextracti32x8.512" => "__builtin_ia32_extracti32x8_mask", + "llvm.x86.avx512.mask.vextracti64x2.256" => "__builtin_ia32_extracti64x2_256_mask", + "llvm.x86.avx512.mask.vextracti64x2.512" => "__builtin_ia32_extracti64x2_512_mask", + "llvm.x86.avx512.mask.vextracti64x4.512" => "__builtin_ia32_extracti64x4_mask", + "llvm.x86.avx512.mask.vfmadd.pd.128" => "__builtin_ia32_vfmaddpd128_mask", + "llvm.x86.avx512.mask.vfmadd.pd.256" => "__builtin_ia32_vfmaddpd256_mask", + "llvm.x86.avx512.mask.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_mask", + "llvm.x86.avx512.mask.vfmadd.ps.128" => "__builtin_ia32_vfmaddps128_mask", + "llvm.x86.avx512.mask.vfmadd.ps.256" => "__builtin_ia32_vfmaddps256_mask", + "llvm.x86.avx512.mask.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_mask", + "llvm.x86.avx512.mask.vfmadd.sd" => "__builtin_ia32_vfmaddsd3_mask", + "llvm.x86.avx512.mask.vfmadd.ss" => "__builtin_ia32_vfmaddss3_mask", + "llvm.x86.avx512.mask.vfmaddsub.pd.128" => "__builtin_ia32_vfmaddsubpd128_mask", + "llvm.x86.avx512.mask.vfmaddsub.pd.256" => "__builtin_ia32_vfmaddsubpd256_mask", + "llvm.x86.avx512.mask.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_mask", + "llvm.x86.avx512.mask.vfmaddsub.ps.128" => "__builtin_ia32_vfmaddsubps128_mask", + "llvm.x86.avx512.mask.vfmaddsub.ps.256" => "__builtin_ia32_vfmaddsubps256_mask", + "llvm.x86.avx512.mask.vfmaddsub.ps.512" => "__builtin_ia32_vfmaddsubps512_mask", + "llvm.x86.avx512.mask.vfnmadd.pd.128" => "__builtin_ia32_vfnmaddpd128_mask", + "llvm.x86.avx512.mask.vfnmadd.pd.256" => "__builtin_ia32_vfnmaddpd256_mask", + "llvm.x86.avx512.mask.vfnmadd.pd.512" => "__builtin_ia32_vfnmaddpd512_mask", + "llvm.x86.avx512.mask.vfnmadd.ps.128" => "__builtin_ia32_vfnmaddps128_mask", + "llvm.x86.avx512.mask.vfnmadd.ps.256" => "__builtin_ia32_vfnmaddps256_mask", + "llvm.x86.avx512.mask.vfnmadd.ps.512" => "__builtin_ia32_vfnmaddps512_mask", + "llvm.x86.avx512.mask.vfnmsub.pd.128" => "__builtin_ia32_vfnmsubpd128_mask", + "llvm.x86.avx512.mask.vfnmsub.pd.256" => "__builtin_ia32_vfnmsubpd256_mask", + "llvm.x86.avx512.mask.vfnmsub.pd.512" => "__builtin_ia32_vfnmsubpd512_mask", + "llvm.x86.avx512.mask.vfnmsub.ps.128" => "__builtin_ia32_vfnmsubps128_mask", + "llvm.x86.avx512.mask.vfnmsub.ps.256" => "__builtin_ia32_vfnmsubps256_mask", + "llvm.x86.avx512.mask.vfnmsub.ps.512" => "__builtin_ia32_vfnmsubps512_mask", + "llvm.x86.avx512.mask.vpermi2var.d.128" => "__builtin_ia32_vpermi2vard128_mask", + "llvm.x86.avx512.mask.vpermi2var.d.256" => "__builtin_ia32_vpermi2vard256_mask", + "llvm.x86.avx512.mask.vpermi2var.d.512" => "__builtin_ia32_vpermi2vard512_mask", + "llvm.x86.avx512.mask.vpermi2var.hi.128" => "__builtin_ia32_vpermi2varhi128_mask", + "llvm.x86.avx512.mask.vpermi2var.hi.256" => "__builtin_ia32_vpermi2varhi256_mask", + "llvm.x86.avx512.mask.vpermi2var.hi.512" => "__builtin_ia32_vpermi2varhi512_mask", + "llvm.x86.avx512.mask.vpermi2var.pd.128" => "__builtin_ia32_vpermi2varpd128_mask", + "llvm.x86.avx512.mask.vpermi2var.pd.256" => "__builtin_ia32_vpermi2varpd256_mask", + "llvm.x86.avx512.mask.vpermi2var.pd.512" => "__builtin_ia32_vpermi2varpd512_mask", + "llvm.x86.avx512.mask.vpermi2var.ps.128" => "__builtin_ia32_vpermi2varps128_mask", + "llvm.x86.avx512.mask.vpermi2var.ps.256" => "__builtin_ia32_vpermi2varps256_mask", + "llvm.x86.avx512.mask.vpermi2var.ps.512" => "__builtin_ia32_vpermi2varps512_mask", + "llvm.x86.avx512.mask.vpermi2var.q.128" => "__builtin_ia32_vpermi2varq128_mask", + "llvm.x86.avx512.mask.vpermi2var.q.256" => "__builtin_ia32_vpermi2varq256_mask", + "llvm.x86.avx512.mask.vpermi2var.q.512" => "__builtin_ia32_vpermi2varq512_mask", + "llvm.x86.avx512.mask.vpermi2var.qi.128" => "__builtin_ia32_vpermi2varqi128_mask", + "llvm.x86.avx512.mask.vpermi2var.qi.256" => "__builtin_ia32_vpermi2varqi256_mask", + "llvm.x86.avx512.mask.vpermi2var.qi.512" => "__builtin_ia32_vpermi2varqi512_mask", + "llvm.x86.avx512.mask.vpermilvar.pd.128" => "__builtin_ia32_vpermilvarpd_mask", + "llvm.x86.avx512.mask.vpermilvar.pd.256" => "__builtin_ia32_vpermilvarpd256_mask", + "llvm.x86.avx512.mask.vpermilvar.pd.512" => "__builtin_ia32_vpermilvarpd512_mask", + "llvm.x86.avx512.mask.vpermilvar.ps.128" => "__builtin_ia32_vpermilvarps_mask", + "llvm.x86.avx512.mask.vpermilvar.ps.256" => "__builtin_ia32_vpermilvarps256_mask", + "llvm.x86.avx512.mask.vpermilvar.ps.512" => "__builtin_ia32_vpermilvarps512_mask", + "llvm.x86.avx512.mask.vpermt.d.512" => "__builtin_ia32_vpermt2vard512_mask", + "llvm.x86.avx512.mask.vpermt.pd.512" => "__builtin_ia32_vpermt2varpd512_mask", + "llvm.x86.avx512.mask.vpermt.ps.512" => "__builtin_ia32_vpermt2varps512_mask", + "llvm.x86.avx512.mask.vpermt.q.512" => "__builtin_ia32_vpermt2varq512_mask", + "llvm.x86.avx512.mask.vpermt2var.d.128" => "__builtin_ia32_vpermt2vard128_mask", + "llvm.x86.avx512.mask.vpermt2var.d.256" => "__builtin_ia32_vpermt2vard256_mask", + "llvm.x86.avx512.mask.vpermt2var.d.512" => "__builtin_ia32_vpermt2vard512_mask", + "llvm.x86.avx512.mask.vpermt2var.hi.128" => "__builtin_ia32_vpermt2varhi128_mask", + "llvm.x86.avx512.mask.vpermt2var.hi.256" => "__builtin_ia32_vpermt2varhi256_mask", + "llvm.x86.avx512.mask.vpermt2var.hi.512" => "__builtin_ia32_vpermt2varhi512_mask", + "llvm.x86.avx512.mask.vpermt2var.pd.128" => "__builtin_ia32_vpermt2varpd128_mask", + "llvm.x86.avx512.mask.vpermt2var.pd.256" => "__builtin_ia32_vpermt2varpd256_mask", + "llvm.x86.avx512.mask.vpermt2var.pd.512" => "__builtin_ia32_vpermt2varpd512_mask", + "llvm.x86.avx512.mask.vpermt2var.ps.128" => "__builtin_ia32_vpermt2varps128_mask", + "llvm.x86.avx512.mask.vpermt2var.ps.256" => "__builtin_ia32_vpermt2varps256_mask", + "llvm.x86.avx512.mask.vpermt2var.ps.512" => "__builtin_ia32_vpermt2varps512_mask", + "llvm.x86.avx512.mask.vpermt2var.q.128" => "__builtin_ia32_vpermt2varq128_mask", + "llvm.x86.avx512.mask.vpermt2var.q.256" => "__builtin_ia32_vpermt2varq256_mask", + "llvm.x86.avx512.mask.vpermt2var.q.512" => "__builtin_ia32_vpermt2varq512_mask", + "llvm.x86.avx512.mask.vpermt2var.qi.128" => "__builtin_ia32_vpermt2varqi128_mask", + "llvm.x86.avx512.mask.vpermt2var.qi.256" => "__builtin_ia32_vpermt2varqi256_mask", + "llvm.x86.avx512.mask.vpermt2var.qi.512" => "__builtin_ia32_vpermt2varqi512_mask", + "llvm.x86.avx512.mask.vpmadd52h.uq.128" => "__builtin_ia32_vpmadd52huq128_mask", + "llvm.x86.avx512.mask.vpmadd52h.uq.256" => "__builtin_ia32_vpmadd52huq256_mask", + "llvm.x86.avx512.mask.vpmadd52h.uq.512" => "__builtin_ia32_vpmadd52huq512_mask", + "llvm.x86.avx512.mask.vpmadd52l.uq.128" => "__builtin_ia32_vpmadd52luq128_mask", + "llvm.x86.avx512.mask.vpmadd52l.uq.256" => "__builtin_ia32_vpmadd52luq256_mask", + "llvm.x86.avx512.mask.vpmadd52l.uq.512" => "__builtin_ia32_vpmadd52luq512_mask", + "llvm.x86.avx512.mask.xor.pd.128" => "__builtin_ia32_xorpd128_mask", + "llvm.x86.avx512.mask.xor.pd.256" => "__builtin_ia32_xorpd256_mask", + "llvm.x86.avx512.mask.xor.pd.512" => "__builtin_ia32_xorpd512_mask", + "llvm.x86.avx512.mask.xor.ps.128" => "__builtin_ia32_xorps128_mask", + "llvm.x86.avx512.mask.xor.ps.256" => "__builtin_ia32_xorps256_mask", + "llvm.x86.avx512.mask.xor.ps.512" => "__builtin_ia32_xorps512_mask", + "llvm.x86.avx512.mask3.vfmadd.pd.128" => "__builtin_ia32_vfmaddpd128_mask3", + "llvm.x86.avx512.mask3.vfmadd.pd.256" => "__builtin_ia32_vfmaddpd256_mask3", + "llvm.x86.avx512.mask3.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_mask3", + "llvm.x86.avx512.mask3.vfmadd.ps.128" => "__builtin_ia32_vfmaddps128_mask3", + "llvm.x86.avx512.mask3.vfmadd.ps.256" => "__builtin_ia32_vfmaddps256_mask3", + "llvm.x86.avx512.mask3.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_mask3", + "llvm.x86.avx512.mask3.vfmadd.sd" => "__builtin_ia32_vfmaddsd3_mask3", + "llvm.x86.avx512.mask3.vfmadd.ss" => "__builtin_ia32_vfmaddss3_mask3", + "llvm.x86.avx512.mask3.vfmaddsub.pd.128" => "__builtin_ia32_vfmaddsubpd128_mask3", + "llvm.x86.avx512.mask3.vfmaddsub.pd.256" => "__builtin_ia32_vfmaddsubpd256_mask3", + "llvm.x86.avx512.mask3.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_mask3", + "llvm.x86.avx512.mask3.vfmaddsub.ps.128" => "__builtin_ia32_vfmaddsubps128_mask3", + "llvm.x86.avx512.mask3.vfmaddsub.ps.256" => "__builtin_ia32_vfmaddsubps256_mask3", + "llvm.x86.avx512.mask3.vfmaddsub.ps.512" => "__builtin_ia32_vfmaddsubps512_mask3", + "llvm.x86.avx512.mask3.vfmsub.pd.128" => "__builtin_ia32_vfmsubpd128_mask3", + "llvm.x86.avx512.mask3.vfmsub.pd.256" => "__builtin_ia32_vfmsubpd256_mask3", + "llvm.x86.avx512.mask3.vfmsub.pd.512" => "__builtin_ia32_vfmsubpd512_mask3", + "llvm.x86.avx512.mask3.vfmsub.ps.128" => "__builtin_ia32_vfmsubps128_mask3", + "llvm.x86.avx512.mask3.vfmsub.ps.256" => "__builtin_ia32_vfmsubps256_mask3", + "llvm.x86.avx512.mask3.vfmsub.ps.512" => "__builtin_ia32_vfmsubps512_mask3", + "llvm.x86.avx512.mask3.vfmsubadd.pd.128" => "__builtin_ia32_vfmsubaddpd128_mask3", + "llvm.x86.avx512.mask3.vfmsubadd.pd.256" => "__builtin_ia32_vfmsubaddpd256_mask3", + "llvm.x86.avx512.mask3.vfmsubadd.pd.512" => "__builtin_ia32_vfmsubaddpd512_mask3", + "llvm.x86.avx512.mask3.vfmsubadd.ps.128" => "__builtin_ia32_vfmsubaddps128_mask3", + "llvm.x86.avx512.mask3.vfmsubadd.ps.256" => "__builtin_ia32_vfmsubaddps256_mask3", + "llvm.x86.avx512.mask3.vfmsubadd.ps.512" => "__builtin_ia32_vfmsubaddps512_mask3", + "llvm.x86.avx512.mask3.vfnmsub.pd.128" => "__builtin_ia32_vfnmsubpd128_mask3", + "llvm.x86.avx512.mask3.vfnmsub.pd.256" => "__builtin_ia32_vfnmsubpd256_mask3", + "llvm.x86.avx512.mask3.vfnmsub.pd.512" => "__builtin_ia32_vfnmsubpd512_mask3", + "llvm.x86.avx512.mask3.vfnmsub.ps.128" => "__builtin_ia32_vfnmsubps128_mask3", + "llvm.x86.avx512.mask3.vfnmsub.ps.256" => "__builtin_ia32_vfnmsubps256_mask3", + "llvm.x86.avx512.mask3.vfnmsub.ps.512" => "__builtin_ia32_vfnmsubps512_mask3", + "llvm.x86.avx512.maskz.fixupimm.pd.128" => "__builtin_ia32_fixupimmpd128_maskz", + "llvm.x86.avx512.maskz.fixupimm.pd.256" => "__builtin_ia32_fixupimmpd256_maskz", + "llvm.x86.avx512.maskz.fixupimm.pd.512" => "__builtin_ia32_fixupimmpd512_maskz", + "llvm.x86.avx512.maskz.fixupimm.ps.128" => "__builtin_ia32_fixupimmps128_maskz", + "llvm.x86.avx512.maskz.fixupimm.ps.256" => "__builtin_ia32_fixupimmps256_maskz", + "llvm.x86.avx512.maskz.fixupimm.ps.512" => "__builtin_ia32_fixupimmps512_maskz", + "llvm.x86.avx512.maskz.fixupimm.sd" => "__builtin_ia32_fixupimmsd_maskz", + "llvm.x86.avx512.maskz.fixupimm.ss" => "__builtin_ia32_fixupimmss_maskz", + "llvm.x86.avx512.maskz.pternlog.d.128" => "__builtin_ia32_pternlogd128_maskz", + "llvm.x86.avx512.maskz.pternlog.d.256" => "__builtin_ia32_pternlogd256_maskz", + "llvm.x86.avx512.maskz.pternlog.d.512" => "__builtin_ia32_pternlogd512_maskz", + "llvm.x86.avx512.maskz.pternlog.q.128" => "__builtin_ia32_pternlogq128_maskz", + "llvm.x86.avx512.maskz.pternlog.q.256" => "__builtin_ia32_pternlogq256_maskz", + "llvm.x86.avx512.maskz.pternlog.q.512" => "__builtin_ia32_pternlogq512_maskz", + "llvm.x86.avx512.maskz.vfmadd.pd.128" => "__builtin_ia32_vfmaddpd128_maskz", + "llvm.x86.avx512.maskz.vfmadd.pd.256" => "__builtin_ia32_vfmaddpd256_maskz", + "llvm.x86.avx512.maskz.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_maskz", + "llvm.x86.avx512.maskz.vfmadd.ps.128" => "__builtin_ia32_vfmaddps128_maskz", + "llvm.x86.avx512.maskz.vfmadd.ps.256" => "__builtin_ia32_vfmaddps256_maskz", + "llvm.x86.avx512.maskz.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_maskz", + "llvm.x86.avx512.maskz.vfmadd.sd" => "__builtin_ia32_vfmaddsd3_maskz", + "llvm.x86.avx512.maskz.vfmadd.ss" => "__builtin_ia32_vfmaddss3_maskz", + "llvm.x86.avx512.maskz.vfmaddsub.pd.128" => "__builtin_ia32_vfmaddsubpd128_maskz", + "llvm.x86.avx512.maskz.vfmaddsub.pd.256" => "__builtin_ia32_vfmaddsubpd256_maskz", + "llvm.x86.avx512.maskz.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_maskz", + "llvm.x86.avx512.maskz.vfmaddsub.ps.128" => "__builtin_ia32_vfmaddsubps128_maskz", + "llvm.x86.avx512.maskz.vfmaddsub.ps.256" => "__builtin_ia32_vfmaddsubps256_maskz", + "llvm.x86.avx512.maskz.vfmaddsub.ps.512" => "__builtin_ia32_vfmaddsubps512_maskz", + "llvm.x86.avx512.maskz.vpermt2var.d.128" => "__builtin_ia32_vpermt2vard128_maskz", + "llvm.x86.avx512.maskz.vpermt2var.d.256" => "__builtin_ia32_vpermt2vard256_maskz", + "llvm.x86.avx512.maskz.vpermt2var.d.512" => "__builtin_ia32_vpermt2vard512_maskz", + "llvm.x86.avx512.maskz.vpermt2var.hi.128" => "__builtin_ia32_vpermt2varhi128_maskz", + "llvm.x86.avx512.maskz.vpermt2var.hi.256" => "__builtin_ia32_vpermt2varhi256_maskz", + "llvm.x86.avx512.maskz.vpermt2var.hi.512" => "__builtin_ia32_vpermt2varhi512_maskz", + "llvm.x86.avx512.maskz.vpermt2var.pd.128" => "__builtin_ia32_vpermt2varpd128_maskz", + "llvm.x86.avx512.maskz.vpermt2var.pd.256" => "__builtin_ia32_vpermt2varpd256_maskz", + "llvm.x86.avx512.maskz.vpermt2var.pd.512" => "__builtin_ia32_vpermt2varpd512_maskz", + "llvm.x86.avx512.maskz.vpermt2var.ps.128" => "__builtin_ia32_vpermt2varps128_maskz", + "llvm.x86.avx512.maskz.vpermt2var.ps.256" => "__builtin_ia32_vpermt2varps256_maskz", + "llvm.x86.avx512.maskz.vpermt2var.ps.512" => "__builtin_ia32_vpermt2varps512_maskz", + "llvm.x86.avx512.maskz.vpermt2var.q.128" => "__builtin_ia32_vpermt2varq128_maskz", + "llvm.x86.avx512.maskz.vpermt2var.q.256" => "__builtin_ia32_vpermt2varq256_maskz", + "llvm.x86.avx512.maskz.vpermt2var.q.512" => "__builtin_ia32_vpermt2varq512_maskz", + "llvm.x86.avx512.maskz.vpermt2var.qi.128" => "__builtin_ia32_vpermt2varqi128_maskz", + "llvm.x86.avx512.maskz.vpermt2var.qi.256" => "__builtin_ia32_vpermt2varqi256_maskz", + "llvm.x86.avx512.maskz.vpermt2var.qi.512" => "__builtin_ia32_vpermt2varqi512_maskz", + "llvm.x86.avx512.maskz.vpmadd52h.uq.128" => "__builtin_ia32_vpmadd52huq128_maskz", + "llvm.x86.avx512.maskz.vpmadd52h.uq.256" => "__builtin_ia32_vpmadd52huq256_maskz", + "llvm.x86.avx512.maskz.vpmadd52h.uq.512" => "__builtin_ia32_vpmadd52huq512_maskz", + "llvm.x86.avx512.maskz.vpmadd52l.uq.128" => "__builtin_ia32_vpmadd52luq128_maskz", + "llvm.x86.avx512.maskz.vpmadd52l.uq.256" => "__builtin_ia32_vpmadd52luq256_maskz", + "llvm.x86.avx512.maskz.vpmadd52l.uq.512" => "__builtin_ia32_vpmadd52luq512_maskz", + "llvm.x86.avx512.max.pd.512" => "__builtin_ia32_maxpd512", + "llvm.x86.avx512.max.ps.512" => "__builtin_ia32_maxps512", + "llvm.x86.avx512.min.pd.512" => "__builtin_ia32_minpd512", + "llvm.x86.avx512.min.ps.512" => "__builtin_ia32_minps512", + "llvm.x86.avx512.movntdqa" => "__builtin_ia32_movntdqa512", + "llvm.x86.avx512.mul.pd.512" => "__builtin_ia32_mulpd512", + "llvm.x86.avx512.mul.ps.512" => "__builtin_ia32_mulps512", + "llvm.x86.avx512.packssdw.512" => "__builtin_ia32_packssdw512", + "llvm.x86.avx512.packsswb.512" => "__builtin_ia32_packsswb512", + "llvm.x86.avx512.packusdw.512" => "__builtin_ia32_packusdw512", + "llvm.x86.avx512.packuswb.512" => "__builtin_ia32_packuswb512", + "llvm.x86.avx512.pavg.b.512" => "__builtin_ia32_pavgb512", + "llvm.x86.avx512.pavg.w.512" => "__builtin_ia32_pavgw512", + "llvm.x86.avx512.pbroadcastd.512" => "__builtin_ia32_pbroadcastd512", + "llvm.x86.avx512.pbroadcastq.512" => "__builtin_ia32_pbroadcastq512", + "llvm.x86.avx512.permvar.df.256" => "__builtin_ia32_permvardf256", + "llvm.x86.avx512.permvar.df.512" => "__builtin_ia32_permvardf512", + "llvm.x86.avx512.permvar.di.256" => "__builtin_ia32_permvardi256", + "llvm.x86.avx512.permvar.di.512" => "__builtin_ia32_permvardi512", + "llvm.x86.avx512.permvar.hi.128" => "__builtin_ia32_permvarhi128", + "llvm.x86.avx512.permvar.hi.256" => "__builtin_ia32_permvarhi256", + "llvm.x86.avx512.permvar.hi.512" => "__builtin_ia32_permvarhi512", + "llvm.x86.avx512.permvar.qi.128" => "__builtin_ia32_permvarqi128", + "llvm.x86.avx512.permvar.qi.256" => "__builtin_ia32_permvarqi256", + "llvm.x86.avx512.permvar.qi.512" => "__builtin_ia32_permvarqi512", + "llvm.x86.avx512.permvar.sf.512" => "__builtin_ia32_permvarsf512", + "llvm.x86.avx512.permvar.si.512" => "__builtin_ia32_permvarsi512", + "llvm.x86.avx512.pmaddubs.w.512" => "__builtin_ia32_pmaddubsw512", + "llvm.x86.avx512.pmaddw.d.512" => "__builtin_ia32_pmaddwd512", + "llvm.x86.avx512.pmovzxbd" => "__builtin_ia32_pmovzxbd512", + "llvm.x86.avx512.pmovzxbq" => "__builtin_ia32_pmovzxbq512", + "llvm.x86.avx512.pmovzxdq" => "__builtin_ia32_pmovzxdq512", + "llvm.x86.avx512.pmovzxwd" => "__builtin_ia32_pmovzxwd512", + "llvm.x86.avx512.pmovzxwq" => "__builtin_ia32_pmovzxwq512", + "llvm.x86.avx512.pmul.hr.sw.512" => "__builtin_ia32_pmulhrsw512", + "llvm.x86.avx512.pmulh.w.512" => "__builtin_ia32_pmulhw512", + "llvm.x86.avx512.pmulhu.w.512" => "__builtin_ia32_pmulhuw512", + "llvm.x86.avx512.pmultishift.qb.128" => "__builtin_ia32_vpmultishiftqb128", + "llvm.x86.avx512.pmultishift.qb.256" => "__builtin_ia32_vpmultishiftqb256", + "llvm.x86.avx512.pmultishift.qb.512" => "__builtin_ia32_vpmultishiftqb512", + "llvm.x86.avx512.psad.bw.512" => "__builtin_ia32_psadbw512", + "llvm.x86.avx512.pshuf.b.512" => "__builtin_ia32_pshufb512", + "llvm.x86.avx512.psll.d.512" => "__builtin_ia32_pslld512", + "llvm.x86.avx512.psll.dq" => "__builtin_ia32_pslldqi512", + "llvm.x86.avx512.psll.dq.bs" => "__builtin_ia32_pslldqi512_byteshift", + "llvm.x86.avx512.psll.q.512" => "__builtin_ia32_psllq512", + "llvm.x86.avx512.psll.w.512" => "__builtin_ia32_psllw512", + "llvm.x86.avx512.pslli.d.512" => "__builtin_ia32_pslldi512", + "llvm.x86.avx512.pslli.q.512" => "__builtin_ia32_psllqi512", + "llvm.x86.avx512.pslli.w.512" => "__builtin_ia32_psllwi512", + "llvm.x86.avx512.psllv.d.512" => "__builtin_ia32_psllv16si", + "llvm.x86.avx512.psllv.q.512" => "__builtin_ia32_psllv8di", + "llvm.x86.avx512.psllv.w.128" => "__builtin_ia32_psllv8hi", + "llvm.x86.avx512.psllv.w.256" => "__builtin_ia32_psllv16hi", + "llvm.x86.avx512.psllv.w.512" => "__builtin_ia32_psllv32hi", + "llvm.x86.avx512.psra.d.512" => "__builtin_ia32_psrad512", + "llvm.x86.avx512.psra.q.128" => "__builtin_ia32_psraq128", + "llvm.x86.avx512.psra.q.256" => "__builtin_ia32_psraq256", + "llvm.x86.avx512.psra.q.512" => "__builtin_ia32_psraq512", + "llvm.x86.avx512.psra.w.512" => "__builtin_ia32_psraw512", + "llvm.x86.avx512.psrai.d.512" => "__builtin_ia32_psradi512", + "llvm.x86.avx512.psrai.q.128" => "__builtin_ia32_psraqi128", + "llvm.x86.avx512.psrai.q.256" => "__builtin_ia32_psraqi256", + "llvm.x86.avx512.psrai.q.512" => "__builtin_ia32_psraqi512", + "llvm.x86.avx512.psrai.w.512" => "__builtin_ia32_psrawi512", + "llvm.x86.avx512.psrav.d.512" => "__builtin_ia32_psrav16si", + "llvm.x86.avx512.psrav.q.128" => "__builtin_ia32_psravq128", + "llvm.x86.avx512.psrav.q.256" => "__builtin_ia32_psravq256", + "llvm.x86.avx512.psrav.q.512" => "__builtin_ia32_psrav8di", + "llvm.x86.avx512.psrav.w.128" => "__builtin_ia32_psrav8hi", + "llvm.x86.avx512.psrav.w.256" => "__builtin_ia32_psrav16hi", + "llvm.x86.avx512.psrav.w.512" => "__builtin_ia32_psrav32hi", + "llvm.x86.avx512.psrl.d.512" => "__builtin_ia32_psrld512", + "llvm.x86.avx512.psrl.dq" => "__builtin_ia32_psrldqi512", + "llvm.x86.avx512.psrl.dq.bs" => "__builtin_ia32_psrldqi512_byteshift", + "llvm.x86.avx512.psrl.q.512" => "__builtin_ia32_psrlq512", + "llvm.x86.avx512.psrl.w.512" => "__builtin_ia32_psrlw512", + "llvm.x86.avx512.psrli.d.512" => "__builtin_ia32_psrldi512", + "llvm.x86.avx512.psrli.q.512" => "__builtin_ia32_psrlqi512", + "llvm.x86.avx512.psrli.w.512" => "__builtin_ia32_psrlwi512", + "llvm.x86.avx512.psrlv.d.512" => "__builtin_ia32_psrlv16si", + "llvm.x86.avx512.psrlv.q.512" => "__builtin_ia32_psrlv8di", + "llvm.x86.avx512.psrlv.w.128" => "__builtin_ia32_psrlv8hi", + "llvm.x86.avx512.psrlv.w.256" => "__builtin_ia32_psrlv16hi", + "llvm.x86.avx512.psrlv.w.512" => "__builtin_ia32_psrlv32hi", + "llvm.x86.avx512.pternlog.d.128" => "__builtin_ia32_pternlogd128", + "llvm.x86.avx512.pternlog.d.256" => "__builtin_ia32_pternlogd256", + "llvm.x86.avx512.pternlog.d.512" => "__builtin_ia32_pternlogd512", + "llvm.x86.avx512.pternlog.q.128" => "__builtin_ia32_pternlogq128", + "llvm.x86.avx512.pternlog.q.256" => "__builtin_ia32_pternlogq256", + "llvm.x86.avx512.pternlog.q.512" => "__builtin_ia32_pternlogq512", + "llvm.x86.avx512.ptestm.b.128" => "__builtin_ia32_ptestmb128", + "llvm.x86.avx512.ptestm.b.256" => "__builtin_ia32_ptestmb256", + "llvm.x86.avx512.ptestm.b.512" => "__builtin_ia32_ptestmb512", + "llvm.x86.avx512.ptestm.d.128" => "__builtin_ia32_ptestmd128", + "llvm.x86.avx512.ptestm.d.256" => "__builtin_ia32_ptestmd256", + "llvm.x86.avx512.ptestm.d.512" => "__builtin_ia32_ptestmd512", + "llvm.x86.avx512.ptestm.q.128" => "__builtin_ia32_ptestmq128", + "llvm.x86.avx512.ptestm.q.256" => "__builtin_ia32_ptestmq256", + "llvm.x86.avx512.ptestm.q.512" => "__builtin_ia32_ptestmq512", + "llvm.x86.avx512.ptestm.w.128" => "__builtin_ia32_ptestmw128", + "llvm.x86.avx512.ptestm.w.256" => "__builtin_ia32_ptestmw256", + "llvm.x86.avx512.ptestm.w.512" => "__builtin_ia32_ptestmw512", + "llvm.x86.avx512.ptestnm.b.128" => "__builtin_ia32_ptestnmb128", + "llvm.x86.avx512.ptestnm.b.256" => "__builtin_ia32_ptestnmb256", + "llvm.x86.avx512.ptestnm.b.512" => "__builtin_ia32_ptestnmb512", + "llvm.x86.avx512.ptestnm.d.128" => "__builtin_ia32_ptestnmd128", + "llvm.x86.avx512.ptestnm.d.256" => "__builtin_ia32_ptestnmd256", + "llvm.x86.avx512.ptestnm.d.512" => "__builtin_ia32_ptestnmd512", + "llvm.x86.avx512.ptestnm.q.128" => "__builtin_ia32_ptestnmq128", + "llvm.x86.avx512.ptestnm.q.256" => "__builtin_ia32_ptestnmq256", + "llvm.x86.avx512.ptestnm.q.512" => "__builtin_ia32_ptestnmq512", + "llvm.x86.avx512.ptestnm.w.128" => "__builtin_ia32_ptestnmw128", + "llvm.x86.avx512.ptestnm.w.256" => "__builtin_ia32_ptestnmw256", + "llvm.x86.avx512.ptestnm.w.512" => "__builtin_ia32_ptestnmw512", + "llvm.x86.avx512.rcp14.pd.128" => "__builtin_ia32_rcp14pd128_mask", + "llvm.x86.avx512.rcp14.pd.256" => "__builtin_ia32_rcp14pd256_mask", + "llvm.x86.avx512.rcp14.pd.512" => "__builtin_ia32_rcp14pd512_mask", + "llvm.x86.avx512.rcp14.ps.128" => "__builtin_ia32_rcp14ps128_mask", + "llvm.x86.avx512.rcp14.ps.256" => "__builtin_ia32_rcp14ps256_mask", + "llvm.x86.avx512.rcp14.ps.512" => "__builtin_ia32_rcp14ps512_mask", + "llvm.x86.avx512.rcp14.sd" => "__builtin_ia32_rcp14sd_mask", + "llvm.x86.avx512.rcp14.ss" => "__builtin_ia32_rcp14ss_mask", + "llvm.x86.avx512.rcp28.pd" => "__builtin_ia32_rcp28pd_mask", + "llvm.x86.avx512.rcp28.ps" => "__builtin_ia32_rcp28ps_mask", + "llvm.x86.avx512.rcp28.sd" => "__builtin_ia32_rcp28sd_round_mask", + // [DUPLICATE]: "llvm.x86.avx512.rcp28.sd" => "__builtin_ia32_rcp28sd_mask", + "llvm.x86.avx512.rcp28.ss" => "__builtin_ia32_rcp28ss_round_mask", + // [DUPLICATE]: "llvm.x86.avx512.rcp28.ss" => "__builtin_ia32_rcp28ss_mask", + "llvm.x86.avx512.rndscale.sd" => "__builtin_ia32_rndscalesd", + "llvm.x86.avx512.rndscale.ss" => "__builtin_ia32_rndscaless", + "llvm.x86.avx512.rsqrt14.pd.128" => "__builtin_ia32_rsqrt14pd128_mask", + "llvm.x86.avx512.rsqrt14.pd.256" => "__builtin_ia32_rsqrt14pd256_mask", + "llvm.x86.avx512.rsqrt14.pd.512" => "__builtin_ia32_rsqrt14pd512_mask", + "llvm.x86.avx512.rsqrt14.ps.128" => "__builtin_ia32_rsqrt14ps128_mask", + "llvm.x86.avx512.rsqrt14.ps.256" => "__builtin_ia32_rsqrt14ps256_mask", + "llvm.x86.avx512.rsqrt14.ps.512" => "__builtin_ia32_rsqrt14ps512_mask", + "llvm.x86.avx512.rsqrt14.sd" => "__builtin_ia32_rsqrt14sd_mask", + "llvm.x86.avx512.rsqrt14.ss" => "__builtin_ia32_rsqrt14ss_mask", + "llvm.x86.avx512.rsqrt28.pd" => "__builtin_ia32_rsqrt28pd_mask", + "llvm.x86.avx512.rsqrt28.ps" => "__builtin_ia32_rsqrt28ps_mask", + "llvm.x86.avx512.rsqrt28.sd" => "__builtin_ia32_rsqrt28sd_round_mask", + // [DUPLICATE]: "llvm.x86.avx512.rsqrt28.sd" => "__builtin_ia32_rsqrt28sd_mask", + "llvm.x86.avx512.rsqrt28.ss" => "__builtin_ia32_rsqrt28ss_round_mask", + // [DUPLICATE]: "llvm.x86.avx512.rsqrt28.ss" => "__builtin_ia32_rsqrt28ss_mask", + "llvm.x86.avx512.scatter.dpd.512" => "__builtin_ia32_scattersiv8df", + "llvm.x86.avx512.scatter.dpi.512" => "__builtin_ia32_scattersiv16si", + "llvm.x86.avx512.scatter.dpq.512" => "__builtin_ia32_scattersiv8di", + "llvm.x86.avx512.scatter.dps.512" => "__builtin_ia32_scattersiv16sf", + "llvm.x86.avx512.scatter.qpd.512" => "__builtin_ia32_scatterdiv8df", + "llvm.x86.avx512.scatter.qpi.512" => "__builtin_ia32_scatterdiv16si", + "llvm.x86.avx512.scatter.qpq.512" => "__builtin_ia32_scatterdiv8di", + "llvm.x86.avx512.scatter.qps.512" => "__builtin_ia32_scatterdiv16sf", + "llvm.x86.avx512.scatterdiv2.df" => "__builtin_ia32_scatterdiv2df", + "llvm.x86.avx512.scatterdiv2.di" => "__builtin_ia32_scatterdiv2di", + "llvm.x86.avx512.scatterdiv4.df" => "__builtin_ia32_scatterdiv4df", + "llvm.x86.avx512.scatterdiv4.di" => "__builtin_ia32_scatterdiv4di", + "llvm.x86.avx512.scatterdiv4.sf" => "__builtin_ia32_scatterdiv4sf", + "llvm.x86.avx512.scatterdiv4.si" => "__builtin_ia32_scatterdiv4si", + "llvm.x86.avx512.scatterdiv8.sf" => "__builtin_ia32_scatterdiv8sf", + "llvm.x86.avx512.scatterdiv8.si" => "__builtin_ia32_scatterdiv8si", + "llvm.x86.avx512.scatterpf.dpd.512" => "__builtin_ia32_scatterpfdpd", + "llvm.x86.avx512.scatterpf.dps.512" => "__builtin_ia32_scatterpfdps", + "llvm.x86.avx512.scatterpf.qpd.512" => "__builtin_ia32_scatterpfqpd", + "llvm.x86.avx512.scatterpf.qps.512" => "__builtin_ia32_scatterpfqps", + "llvm.x86.avx512.scattersiv2.df" => "__builtin_ia32_scattersiv2df", + "llvm.x86.avx512.scattersiv2.di" => "__builtin_ia32_scattersiv2di", + "llvm.x86.avx512.scattersiv4.df" => "__builtin_ia32_scattersiv4df", + "llvm.x86.avx512.scattersiv4.di" => "__builtin_ia32_scattersiv4di", + "llvm.x86.avx512.scattersiv4.sf" => "__builtin_ia32_scattersiv4sf", + "llvm.x86.avx512.scattersiv4.si" => "__builtin_ia32_scattersiv4si", + "llvm.x86.avx512.scattersiv8.sf" => "__builtin_ia32_scattersiv8sf", + "llvm.x86.avx512.scattersiv8.si" => "__builtin_ia32_scattersiv8si", + "llvm.x86.avx512.sqrt.pd.512" => "__builtin_ia32_sqrtpd512_mask", + "llvm.x86.avx512.sqrt.ps.512" => "__builtin_ia32_sqrtps512_mask", + "llvm.x86.avx512.sqrt.sd" => "__builtin_ia32_sqrtrndsd", + "llvm.x86.avx512.sqrt.ss" => "__builtin_ia32_sqrtrndss", + "llvm.x86.avx512.sub.pd.512" => "__builtin_ia32_subpd512", + "llvm.x86.avx512.sub.ps.512" => "__builtin_ia32_subps512", + "llvm.x86.avx512.vbroadcast.sd.512" => "__builtin_ia32_vbroadcastsd512", + "llvm.x86.avx512.vbroadcast.sd.pd.512" => "__builtin_ia32_vbroadcastsd_pd512", + "llvm.x86.avx512.vbroadcast.ss.512" => "__builtin_ia32_vbroadcastss512", + "llvm.x86.avx512.vbroadcast.ss.ps.512" => "__builtin_ia32_vbroadcastss_ps512", + "llvm.x86.avx512.vcomi.sd" => "__builtin_ia32_vcomisd", + "llvm.x86.avx512.vcomi.ss" => "__builtin_ia32_vcomiss", + "llvm.x86.avx512.vcvtsd2si32" => "__builtin_ia32_vcvtsd2si32", + "llvm.x86.avx512.vcvtsd2si64" => "__builtin_ia32_vcvtsd2si64", + "llvm.x86.avx512.vcvtsd2usi32" => "__builtin_ia32_vcvtsd2usi32", + "llvm.x86.avx512.vcvtsd2usi64" => "__builtin_ia32_vcvtsd2usi64", + "llvm.x86.avx512.vcvtss2si32" => "__builtin_ia32_vcvtss2si32", + "llvm.x86.avx512.vcvtss2si64" => "__builtin_ia32_vcvtss2si64", + "llvm.x86.avx512.vcvtss2usi32" => "__builtin_ia32_vcvtss2usi32", + "llvm.x86.avx512.vcvtss2usi64" => "__builtin_ia32_vcvtss2usi64", + "llvm.x86.avx512.vpdpbusd.128" => "__builtin_ia32_vpdpbusd128", + "llvm.x86.avx512.vpdpbusd.256" => "__builtin_ia32_vpdpbusd256", + "llvm.x86.avx512.vpdpbusd.512" => "__builtin_ia32_vpdpbusd512", + "llvm.x86.avx512.vpdpbusds.128" => "__builtin_ia32_vpdpbusds128", + "llvm.x86.avx512.vpdpbusds.256" => "__builtin_ia32_vpdpbusds256", + "llvm.x86.avx512.vpdpbusds.512" => "__builtin_ia32_vpdpbusds512", + "llvm.x86.avx512.vpdpwssd.128" => "__builtin_ia32_vpdpwssd128", + "llvm.x86.avx512.vpdpwssd.256" => "__builtin_ia32_vpdpwssd256", + "llvm.x86.avx512.vpdpwssd.512" => "__builtin_ia32_vpdpwssd512", + "llvm.x86.avx512.vpdpwssds.128" => "__builtin_ia32_vpdpwssds128", + "llvm.x86.avx512.vpdpwssds.256" => "__builtin_ia32_vpdpwssds256", + "llvm.x86.avx512.vpdpwssds.512" => "__builtin_ia32_vpdpwssds512", + "llvm.x86.avx512.vpermi2var.d.128" => "__builtin_ia32_vpermi2vard128", + "llvm.x86.avx512.vpermi2var.d.256" => "__builtin_ia32_vpermi2vard256", + "llvm.x86.avx512.vpermi2var.d.512" => "__builtin_ia32_vpermi2vard512", + "llvm.x86.avx512.vpermi2var.hi.128" => "__builtin_ia32_vpermi2varhi128", + "llvm.x86.avx512.vpermi2var.hi.256" => "__builtin_ia32_vpermi2varhi256", + "llvm.x86.avx512.vpermi2var.hi.512" => "__builtin_ia32_vpermi2varhi512", + "llvm.x86.avx512.vpermi2var.pd.128" => "__builtin_ia32_vpermi2varpd128", + "llvm.x86.avx512.vpermi2var.pd.256" => "__builtin_ia32_vpermi2varpd256", + "llvm.x86.avx512.vpermi2var.pd.512" => "__builtin_ia32_vpermi2varpd512", + "llvm.x86.avx512.vpermi2var.ps.128" => "__builtin_ia32_vpermi2varps128", + "llvm.x86.avx512.vpermi2var.ps.256" => "__builtin_ia32_vpermi2varps256", + "llvm.x86.avx512.vpermi2var.ps.512" => "__builtin_ia32_vpermi2varps512", + "llvm.x86.avx512.vpermi2var.q.128" => "__builtin_ia32_vpermi2varq128", + "llvm.x86.avx512.vpermi2var.q.256" => "__builtin_ia32_vpermi2varq256", + "llvm.x86.avx512.vpermi2var.q.512" => "__builtin_ia32_vpermi2varq512", + "llvm.x86.avx512.vpermi2var.qi.128" => "__builtin_ia32_vpermi2varqi128", + "llvm.x86.avx512.vpermi2var.qi.256" => "__builtin_ia32_vpermi2varqi256", + "llvm.x86.avx512.vpermi2var.qi.512" => "__builtin_ia32_vpermi2varqi512", + "llvm.x86.avx512.vpermilvar.pd.512" => "__builtin_ia32_vpermilvarpd512", + "llvm.x86.avx512.vpermilvar.ps.512" => "__builtin_ia32_vpermilvarps512", + "llvm.x86.avx512.vpmadd52h.uq.128" => "__builtin_ia32_vpmadd52huq128", + "llvm.x86.avx512.vpmadd52h.uq.256" => "__builtin_ia32_vpmadd52huq256", + "llvm.x86.avx512.vpmadd52h.uq.512" => "__builtin_ia32_vpmadd52huq512", + "llvm.x86.avx512.vpmadd52l.uq.128" => "__builtin_ia32_vpmadd52luq128", + "llvm.x86.avx512.vpmadd52l.uq.256" => "__builtin_ia32_vpmadd52luq256", + "llvm.x86.avx512.vpmadd52l.uq.512" => "__builtin_ia32_vpmadd52luq512", + "llvm.x86.avx512bf16.cvtne2ps2bf16.128" => "__builtin_ia32_cvtne2ps2bf16_128", + "llvm.x86.avx512bf16.cvtne2ps2bf16.256" => "__builtin_ia32_cvtne2ps2bf16_256", + "llvm.x86.avx512bf16.cvtne2ps2bf16.512" => "__builtin_ia32_cvtne2ps2bf16_512", + "llvm.x86.avx512bf16.cvtneps2bf16.256" => "__builtin_ia32_cvtneps2bf16_256", + "llvm.x86.avx512bf16.cvtneps2bf16.512" => "__builtin_ia32_cvtneps2bf16_512", + "llvm.x86.avx512bf16.dpbf16ps.128" => "__builtin_ia32_dpbf16ps_128", + "llvm.x86.avx512bf16.dpbf16ps.256" => "__builtin_ia32_dpbf16ps_256", + "llvm.x86.avx512bf16.dpbf16ps.512" => "__builtin_ia32_dpbf16ps_512", + "llvm.x86.avx512fp16.add.ph.512" => "__builtin_ia32_addph512", + "llvm.x86.avx512fp16.div.ph.512" => "__builtin_ia32_divph512", + "llvm.x86.avx512fp16.mask.add.sh.round" => "__builtin_ia32_addsh_round_mask", + "llvm.x86.avx512fp16.mask.cmp.sh" => "__builtin_ia32_cmpsh_mask", + "llvm.x86.avx512fp16.mask.div.sh.round" => "__builtin_ia32_divsh_round_mask", + "llvm.x86.avx512fp16.mask.fpclass.sh" => "__builtin_ia32_fpclasssh_mask", + "llvm.x86.avx512fp16.mask.getexp.ph.128" => "__builtin_ia32_getexpph128_mask", + "llvm.x86.avx512fp16.mask.getexp.ph.256" => "__builtin_ia32_getexpph256_mask", + "llvm.x86.avx512fp16.mask.getexp.ph.512" => "__builtin_ia32_getexpph512_mask", + "llvm.x86.avx512fp16.mask.getexp.sh" => "__builtin_ia32_getexpsh128_round_mask", + "llvm.x86.avx512fp16.mask.getmant.ph.128" => "__builtin_ia32_getmantph128_mask", + "llvm.x86.avx512fp16.mask.getmant.ph.256" => "__builtin_ia32_getmantph256_mask", + "llvm.x86.avx512fp16.mask.getmant.ph.512" => "__builtin_ia32_getmantph512_mask", + "llvm.x86.avx512fp16.mask.getmant.sh" => "__builtin_ia32_getmantsh_round_mask", + "llvm.x86.avx512fp16.mask.max.sh.round" => "__builtin_ia32_maxsh_round_mask", + "llvm.x86.avx512fp16.mask.min.sh.round" => "__builtin_ia32_minsh_round_mask", + "llvm.x86.avx512fp16.mask.mul.sh.round" => "__builtin_ia32_mulsh_round_mask", + "llvm.x86.avx512fp16.mask.rcp.ph.128" => "__builtin_ia32_rcpph128_mask", + "llvm.x86.avx512fp16.mask.rcp.ph.256" => "__builtin_ia32_rcpph256_mask", + "llvm.x86.avx512fp16.mask.rcp.ph.512" => "__builtin_ia32_rcpph512_mask", + "llvm.x86.avx512fp16.mask.rcp.sh" => "__builtin_ia32_rcpsh_mask", + "llvm.x86.avx512fp16.mask.reduce.ph.128" => "__builtin_ia32_reduceph128_mask", + "llvm.x86.avx512fp16.mask.reduce.ph.256" => "__builtin_ia32_reduceph256_mask", + "llvm.x86.avx512fp16.mask.reduce.ph.512" => "__builtin_ia32_reduceph512_mask", + "llvm.x86.avx512fp16.mask.reduce.sh" => "__builtin_ia32_reducesh_mask", + "llvm.x86.avx512fp16.mask.rndscale.ph.128" => "__builtin_ia32_rndscaleph_128_mask", + "llvm.x86.avx512fp16.mask.rndscale.ph.256" => "__builtin_ia32_rndscaleph_256_mask", + "llvm.x86.avx512fp16.mask.rndscale.ph.512" => "__builtin_ia32_rndscaleph_mask", + "llvm.x86.avx512fp16.mask.rndscale.sh" => "__builtin_ia32_rndscalesh_round_mask", + "llvm.x86.avx512fp16.mask.rsqrt.ph.128" => "__builtin_ia32_rsqrtph128_mask", + "llvm.x86.avx512fp16.mask.rsqrt.ph.256" => "__builtin_ia32_rsqrtph256_mask", + "llvm.x86.avx512fp16.mask.rsqrt.ph.512" => "__builtin_ia32_rsqrtph512_mask", + "llvm.x86.avx512fp16.mask.rsqrt.sh" => "__builtin_ia32_rsqrtsh_mask", + "llvm.x86.avx512fp16.mask.scalef.ph.128" => "__builtin_ia32_scalefph128_mask", + "llvm.x86.avx512fp16.mask.scalef.ph.256" => "__builtin_ia32_scalefph256_mask", + "llvm.x86.avx512fp16.mask.scalef.ph.512" => "__builtin_ia32_scalefph512_mask", + "llvm.x86.avx512fp16.mask.scalef.sh" => "__builtin_ia32_scalefsh_round_mask", + "llvm.x86.avx512fp16.mask.sub.sh.round" => "__builtin_ia32_subsh_round_mask", + "llvm.x86.avx512fp16.mask.vcvtdq2ph.128" => "__builtin_ia32_vcvtdq2ph128_mask", + "llvm.x86.avx512fp16.mask.vcvtpd2ph.128" => "__builtin_ia32_vcvtpd2ph128_mask", + "llvm.x86.avx512fp16.mask.vcvtpd2ph.256" => "__builtin_ia32_vcvtpd2ph256_mask", + "llvm.x86.avx512fp16.mask.vcvtpd2ph.512" => "__builtin_ia32_vcvtpd2ph512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2dq.128" => "__builtin_ia32_vcvtph2dq128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2dq.256" => "__builtin_ia32_vcvtph2dq256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2dq.512" => "__builtin_ia32_vcvtph2dq512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2pd.128" => "__builtin_ia32_vcvtph2pd128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2pd.256" => "__builtin_ia32_vcvtph2pd256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2pd.512" => "__builtin_ia32_vcvtph2pd512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2psx.128" => "__builtin_ia32_vcvtph2psx128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2psx.256" => "__builtin_ia32_vcvtph2psx256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2psx.512" => "__builtin_ia32_vcvtph2psx512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2qq.128" => "__builtin_ia32_vcvtph2qq128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2qq.256" => "__builtin_ia32_vcvtph2qq256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2qq.512" => "__builtin_ia32_vcvtph2qq512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2udq.128" => "__builtin_ia32_vcvtph2udq128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2udq.256" => "__builtin_ia32_vcvtph2udq256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2udq.512" => "__builtin_ia32_vcvtph2udq512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2uqq.128" => "__builtin_ia32_vcvtph2uqq128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2uqq.256" => "__builtin_ia32_vcvtph2uqq256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2uqq.512" => "__builtin_ia32_vcvtph2uqq512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2uw.128" => "__builtin_ia32_vcvtph2uw128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2uw.256" => "__builtin_ia32_vcvtph2uw256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2uw.512" => "__builtin_ia32_vcvtph2uw512_mask", + "llvm.x86.avx512fp16.mask.vcvtph2w.128" => "__builtin_ia32_vcvtph2w128_mask", + "llvm.x86.avx512fp16.mask.vcvtph2w.256" => "__builtin_ia32_vcvtph2w256_mask", + "llvm.x86.avx512fp16.mask.vcvtph2w.512" => "__builtin_ia32_vcvtph2w512_mask", + "llvm.x86.avx512fp16.mask.vcvtps2phx.128" => "__builtin_ia32_vcvtps2phx128_mask", + "llvm.x86.avx512fp16.mask.vcvtps2phx.256" => "__builtin_ia32_vcvtps2phx256_mask", + "llvm.x86.avx512fp16.mask.vcvtps2phx.512" => "__builtin_ia32_vcvtps2phx512_mask", + "llvm.x86.avx512fp16.mask.vcvtqq2ph.128" => "__builtin_ia32_vcvtqq2ph128_mask", + "llvm.x86.avx512fp16.mask.vcvtqq2ph.256" => "__builtin_ia32_vcvtqq2ph256_mask", + "llvm.x86.avx512fp16.mask.vcvtsd2sh.round" => "__builtin_ia32_vcvtsd2sh_round_mask", + "llvm.x86.avx512fp16.mask.vcvtsh2sd.round" => "__builtin_ia32_vcvtsh2sd_round_mask", + "llvm.x86.avx512fp16.mask.vcvtsh2ss.round" => "__builtin_ia32_vcvtsh2ss_round_mask", + "llvm.x86.avx512fp16.mask.vcvtss2sh.round" => "__builtin_ia32_vcvtss2sh_round_mask", + "llvm.x86.avx512fp16.mask.vcvttph2dq.128" => "__builtin_ia32_vcvttph2dq128_mask", + "llvm.x86.avx512fp16.mask.vcvttph2dq.256" => "__builtin_ia32_vcvttph2dq256_mask", + "llvm.x86.avx512fp16.mask.vcvttph2dq.512" => "__builtin_ia32_vcvttph2dq512_mask", + "llvm.x86.avx512fp16.mask.vcvttph2qq.128" => "__builtin_ia32_vcvttph2qq128_mask", + "llvm.x86.avx512fp16.mask.vcvttph2qq.256" => "__builtin_ia32_vcvttph2qq256_mask", + "llvm.x86.avx512fp16.mask.vcvttph2qq.512" => "__builtin_ia32_vcvttph2qq512_mask", + "llvm.x86.avx512fp16.mask.vcvttph2udq.128" => "__builtin_ia32_vcvttph2udq128_mask", + "llvm.x86.avx512fp16.mask.vcvttph2udq.256" => "__builtin_ia32_vcvttph2udq256_mask", + "llvm.x86.avx512fp16.mask.vcvttph2udq.512" => "__builtin_ia32_vcvttph2udq512_mask", + "llvm.x86.avx512fp16.mask.vcvttph2uqq.128" => "__builtin_ia32_vcvttph2uqq128_mask", + "llvm.x86.avx512fp16.mask.vcvttph2uqq.256" => "__builtin_ia32_vcvttph2uqq256_mask", + "llvm.x86.avx512fp16.mask.vcvttph2uqq.512" => "__builtin_ia32_vcvttph2uqq512_mask", + "llvm.x86.avx512fp16.mask.vcvttph2uw.128" => "__builtin_ia32_vcvttph2uw128_mask", + "llvm.x86.avx512fp16.mask.vcvttph2uw.256" => "__builtin_ia32_vcvttph2uw256_mask", + "llvm.x86.avx512fp16.mask.vcvttph2uw.512" => "__builtin_ia32_vcvttph2uw512_mask", + "llvm.x86.avx512fp16.mask.vcvttph2w.128" => "__builtin_ia32_vcvttph2w128_mask", + "llvm.x86.avx512fp16.mask.vcvttph2w.256" => "__builtin_ia32_vcvttph2w256_mask", + "llvm.x86.avx512fp16.mask.vcvttph2w.512" => "__builtin_ia32_vcvttph2w512_mask", + "llvm.x86.avx512fp16.mask.vcvtudq2ph.128" => "__builtin_ia32_vcvtudq2ph128_mask", + "llvm.x86.avx512fp16.mask.vcvtuqq2ph.128" => "__builtin_ia32_vcvtuqq2ph128_mask", + "llvm.x86.avx512fp16.mask.vcvtuqq2ph.256" => "__builtin_ia32_vcvtuqq2ph256_mask", + "llvm.x86.avx512fp16.mask.vfcmadd.cph.128" => "__builtin_ia32_vfcmaddcph128_mask", + "llvm.x86.avx512fp16.mask.vfcmadd.cph.256" => "__builtin_ia32_vfcmaddcph256_mask", + "llvm.x86.avx512fp16.mask.vfcmadd.cph.512" => "__builtin_ia32_vfcmaddcph512_mask3", + "llvm.x86.avx512fp16.mask.vfcmadd.csh" => "__builtin_ia32_vfcmaddcsh_mask", + "llvm.x86.avx512fp16.mask.vfcmul.cph.128" => "__builtin_ia32_vfcmulcph128_mask", + "llvm.x86.avx512fp16.mask.vfcmul.cph.256" => "__builtin_ia32_vfcmulcph256_mask", + "llvm.x86.avx512fp16.mask.vfcmul.cph.512" => "__builtin_ia32_vfcmulcph512_mask", + "llvm.x86.avx512fp16.mask.vfcmul.csh" => "__builtin_ia32_vfcmulcsh_mask", + "llvm.x86.avx512fp16.mask.vfmadd.cph.128" => "__builtin_ia32_vfmaddcph128_mask", + "llvm.x86.avx512fp16.mask.vfmadd.cph.256" => "__builtin_ia32_vfmaddcph256_mask", + "llvm.x86.avx512fp16.mask.vfmadd.cph.512" => "__builtin_ia32_vfmaddcph512_mask3", + "llvm.x86.avx512fp16.mask.vfmadd.csh" => "__builtin_ia32_vfmaddcsh_mask", + "llvm.x86.avx512fp16.mask.vfmul.cph.128" => "__builtin_ia32_vfmulcph128_mask", + "llvm.x86.avx512fp16.mask.vfmul.cph.256" => "__builtin_ia32_vfmulcph256_mask", + "llvm.x86.avx512fp16.mask.vfmul.cph.512" => "__builtin_ia32_vfmulcph512_mask", + "llvm.x86.avx512fp16.mask.vfmul.csh" => "__builtin_ia32_vfmulcsh_mask", + "llvm.x86.avx512fp16.maskz.vfcmadd.cph.128" => "__builtin_ia32_vfcmaddcph128_maskz", + "llvm.x86.avx512fp16.maskz.vfcmadd.cph.256" => "__builtin_ia32_vfcmaddcph256_maskz", + "llvm.x86.avx512fp16.maskz.vfcmadd.cph.512" => "__builtin_ia32_vfcmaddcph512_maskz", + "llvm.x86.avx512fp16.maskz.vfcmadd.csh" => "__builtin_ia32_vfcmaddcsh_maskz", + "llvm.x86.avx512fp16.maskz.vfmadd.cph.128" => "__builtin_ia32_vfmaddcph128_maskz", + "llvm.x86.avx512fp16.maskz.vfmadd.cph.256" => "__builtin_ia32_vfmaddcph256_maskz", + "llvm.x86.avx512fp16.maskz.vfmadd.cph.512" => "__builtin_ia32_vfmaddcph512_maskz", + "llvm.x86.avx512fp16.maskz.vfmadd.csh" => "__builtin_ia32_vfmaddcsh_maskz", + "llvm.x86.avx512fp16.max.ph.128" => "__builtin_ia32_maxph128", + "llvm.x86.avx512fp16.max.ph.256" => "__builtin_ia32_maxph256", + "llvm.x86.avx512fp16.max.ph.512" => "__builtin_ia32_maxph512", + "llvm.x86.avx512fp16.min.ph.128" => "__builtin_ia32_minph128", + "llvm.x86.avx512fp16.min.ph.256" => "__builtin_ia32_minph256", + "llvm.x86.avx512fp16.min.ph.512" => "__builtin_ia32_minph512", + "llvm.x86.avx512fp16.mul.ph.512" => "__builtin_ia32_mulph512", + "llvm.x86.avx512fp16.sub.ph.512" => "__builtin_ia32_subph512", + "llvm.x86.avx512fp16.vcomi.sh" => "__builtin_ia32_vcomish", + "llvm.x86.avx512fp16.vcvtsh2si32" => "__builtin_ia32_vcvtsh2si32", + "llvm.x86.avx512fp16.vcvtsh2si64" => "__builtin_ia32_vcvtsh2si64", + "llvm.x86.avx512fp16.vcvtsh2usi32" => "__builtin_ia32_vcvtsh2usi32", + "llvm.x86.avx512fp16.vcvtsh2usi64" => "__builtin_ia32_vcvtsh2usi64", + "llvm.x86.avx512fp16.vcvtsi2sh" => "__builtin_ia32_vcvtsi2sh", + "llvm.x86.avx512fp16.vcvtsi642sh" => "__builtin_ia32_vcvtsi642sh", + "llvm.x86.avx512fp16.vcvttsh2si32" => "__builtin_ia32_vcvttsh2si32", + "llvm.x86.avx512fp16.vcvttsh2si64" => "__builtin_ia32_vcvttsh2si64", + "llvm.x86.avx512fp16.vcvttsh2usi32" => "__builtin_ia32_vcvttsh2usi32", + "llvm.x86.avx512fp16.vcvttsh2usi64" => "__builtin_ia32_vcvttsh2usi64", + "llvm.x86.avx512fp16.vcvtusi2sh" => "__builtin_ia32_vcvtusi2sh", + "llvm.x86.avx512fp16.vcvtusi642sh" => "__builtin_ia32_vcvtusi642sh", + "llvm.x86.avx512fp16.vfmaddsub.ph.128" => "__builtin_ia32_vfmaddsubph", + "llvm.x86.avx512fp16.vfmaddsub.ph.256" => "__builtin_ia32_vfmaddsubph256", + "llvm.x86.bmi.bextr.32" => "__builtin_ia32_bextr_u32", + "llvm.x86.bmi.bextr.64" => "__builtin_ia32_bextr_u64", + "llvm.x86.bmi.bzhi.32" => "__builtin_ia32_bzhi_si", + "llvm.x86.bmi.bzhi.64" => "__builtin_ia32_bzhi_di", + "llvm.x86.bmi.pdep.32" => "__builtin_ia32_pdep_si", + "llvm.x86.bmi.pdep.64" => "__builtin_ia32_pdep_di", + "llvm.x86.bmi.pext.32" => "__builtin_ia32_pext_si", + "llvm.x86.bmi.pext.64" => "__builtin_ia32_pext_di", + "llvm.x86.cldemote" => "__builtin_ia32_cldemote", + "llvm.x86.clflushopt" => "__builtin_ia32_clflushopt", + "llvm.x86.clrssbsy" => "__builtin_ia32_clrssbsy", + "llvm.x86.clui" => "__builtin_ia32_clui", + "llvm.x86.clwb" => "__builtin_ia32_clwb", + "llvm.x86.clzero" => "__builtin_ia32_clzero", + "llvm.x86.directstore32" => "__builtin_ia32_directstore_u32", + "llvm.x86.directstore64" => "__builtin_ia32_directstore_u64", + "llvm.x86.enqcmd" => "__builtin_ia32_enqcmd", + "llvm.x86.enqcmds" => "__builtin_ia32_enqcmds", + "llvm.x86.flags.read.u32" => "__builtin_ia32_readeflags_u32", + "llvm.x86.flags.read.u64" => "__builtin_ia32_readeflags_u64", + "llvm.x86.flags.write.u32" => "__builtin_ia32_writeeflags_u32", + "llvm.x86.flags.write.u64" => "__builtin_ia32_writeeflags_u64", + "llvm.x86.fma.mask.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_mask", + "llvm.x86.fma.mask.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_mask", + "llvm.x86.fma.mask.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_mask", + "llvm.x86.fma.mask.vfmaddsub.ps.512" => "__builtin_ia32_vfmaddsubps512_mask", + "llvm.x86.fma.mask.vfmsub.pd.512" => "__builtin_ia32_vfmsubpd512_mask", + "llvm.x86.fma.mask.vfmsub.ps.512" => "__builtin_ia32_vfmsubps512_mask", + "llvm.x86.fma.mask.vfmsubadd.pd.512" => "__builtin_ia32_vfmsubaddpd512_mask", + "llvm.x86.fma.mask.vfmsubadd.ps.512" => "__builtin_ia32_vfmsubaddps512_mask", + "llvm.x86.fma.mask.vfnmadd.pd.512" => "__builtin_ia32_vfnmaddpd512_mask", + "llvm.x86.fma.mask.vfnmadd.ps.512" => "__builtin_ia32_vfnmaddps512_mask", + "llvm.x86.fma.mask.vfnmsub.pd.512" => "__builtin_ia32_vfnmsubpd512_mask", + "llvm.x86.fma.mask.vfnmsub.ps.512" => "__builtin_ia32_vfnmsubps512_mask", + "llvm.x86.fma.vfmadd.pd" => "__builtin_ia32_vfmaddpd", + "llvm.x86.fma.vfmadd.pd.256" => "__builtin_ia32_vfmaddpd256", + "llvm.x86.fma.vfmadd.ps" => "__builtin_ia32_vfmaddps", + "llvm.x86.fma.vfmadd.ps.256" => "__builtin_ia32_vfmaddps256", + "llvm.x86.fma.vfmadd.sd" => "__builtin_ia32_vfmaddsd", + "llvm.x86.fma.vfmadd.ss" => "__builtin_ia32_vfmaddss", + "llvm.x86.fma.vfmaddsub.pd" => "__builtin_ia32_vfmaddsubpd", + "llvm.x86.fma.vfmaddsub.pd.256" => "__builtin_ia32_vfmaddsubpd256", + "llvm.x86.fma.vfmaddsub.ps" => "__builtin_ia32_vfmaddsubps", + "llvm.x86.fma.vfmaddsub.ps.256" => "__builtin_ia32_vfmaddsubps256", + "llvm.x86.fma.vfmsub.pd" => "__builtin_ia32_vfmsubpd", + "llvm.x86.fma.vfmsub.pd.256" => "__builtin_ia32_vfmsubpd256", + "llvm.x86.fma.vfmsub.ps" => "__builtin_ia32_vfmsubps", + "llvm.x86.fma.vfmsub.ps.256" => "__builtin_ia32_vfmsubps256", + "llvm.x86.fma.vfmsub.sd" => "__builtin_ia32_vfmsubsd", + "llvm.x86.fma.vfmsub.ss" => "__builtin_ia32_vfmsubss", + "llvm.x86.fma.vfmsubadd.pd" => "__builtin_ia32_vfmsubaddpd", + "llvm.x86.fma.vfmsubadd.pd.256" => "__builtin_ia32_vfmsubaddpd256", + "llvm.x86.fma.vfmsubadd.ps" => "__builtin_ia32_vfmsubaddps", + "llvm.x86.fma.vfmsubadd.ps.256" => "__builtin_ia32_vfmsubaddps256", + "llvm.x86.fma.vfnmadd.pd" => "__builtin_ia32_vfnmaddpd", + "llvm.x86.fma.vfnmadd.pd.256" => "__builtin_ia32_vfnmaddpd256", + "llvm.x86.fma.vfnmadd.ps" => "__builtin_ia32_vfnmaddps", + "llvm.x86.fma.vfnmadd.ps.256" => "__builtin_ia32_vfnmaddps256", + "llvm.x86.fma.vfnmadd.sd" => "__builtin_ia32_vfnmaddsd", + "llvm.x86.fma.vfnmadd.ss" => "__builtin_ia32_vfnmaddss", + "llvm.x86.fma.vfnmsub.pd" => "__builtin_ia32_vfnmsubpd", + "llvm.x86.fma.vfnmsub.pd.256" => "__builtin_ia32_vfnmsubpd256", + "llvm.x86.fma.vfnmsub.ps" => "__builtin_ia32_vfnmsubps", + "llvm.x86.fma.vfnmsub.ps.256" => "__builtin_ia32_vfnmsubps256", + "llvm.x86.fma.vfnmsub.sd" => "__builtin_ia32_vfnmsubsd", + "llvm.x86.fma.vfnmsub.ss" => "__builtin_ia32_vfnmsubss", + "llvm.x86.fxrstor" => "__builtin_ia32_fxrstor", + "llvm.x86.fxrstor64" => "__builtin_ia32_fxrstor64", + "llvm.x86.fxsave" => "__builtin_ia32_fxsave", + "llvm.x86.fxsave64" => "__builtin_ia32_fxsave64", + "llvm.x86.incsspd" => "__builtin_ia32_incsspd", + "llvm.x86.incsspq" => "__builtin_ia32_incsspq", + "llvm.x86.invpcid" => "__builtin_ia32_invpcid", + "llvm.x86.ldtilecfg" => "__builtin_ia32_tile_loadconfig", + "llvm.x86.ldtilecfg.internal" => "__builtin_ia32_tile_loadconfig_internal", + "llvm.x86.llwpcb" => "__builtin_ia32_llwpcb", + "llvm.x86.loadiwkey" => "__builtin_ia32_loadiwkey", + "llvm.x86.lwpins32" => "__builtin_ia32_lwpins32", + "llvm.x86.lwpins64" => "__builtin_ia32_lwpins64", + "llvm.x86.lwpval32" => "__builtin_ia32_lwpval32", + "llvm.x86.lwpval64" => "__builtin_ia32_lwpval64", + "llvm.x86.mmx.emms" => "__builtin_ia32_emms", + "llvm.x86.mmx.femms" => "__builtin_ia32_femms", + "llvm.x86.mmx.maskmovq" => "__builtin_ia32_maskmovq", + "llvm.x86.mmx.movnt.dq" => "__builtin_ia32_movntq", + "llvm.x86.mmx.packssdw" => "__builtin_ia32_packssdw", + "llvm.x86.mmx.packsswb" => "__builtin_ia32_packsswb", + "llvm.x86.mmx.packuswb" => "__builtin_ia32_packuswb", + "llvm.x86.mmx.padd.b" => "__builtin_ia32_paddb", + "llvm.x86.mmx.padd.d" => "__builtin_ia32_paddd", + "llvm.x86.mmx.padd.q" => "__builtin_ia32_paddq", + "llvm.x86.mmx.padd.w" => "__builtin_ia32_paddw", + "llvm.x86.mmx.padds.b" => "__builtin_ia32_paddsb", + "llvm.x86.mmx.padds.w" => "__builtin_ia32_paddsw", + "llvm.x86.mmx.paddus.b" => "__builtin_ia32_paddusb", + "llvm.x86.mmx.paddus.w" => "__builtin_ia32_paddusw", + "llvm.x86.mmx.palignr.b" => "__builtin_ia32_palignr", + "llvm.x86.mmx.pand" => "__builtin_ia32_pand", + "llvm.x86.mmx.pandn" => "__builtin_ia32_pandn", + "llvm.x86.mmx.pavg.b" => "__builtin_ia32_pavgb", + "llvm.x86.mmx.pavg.w" => "__builtin_ia32_pavgw", + "llvm.x86.mmx.pcmpeq.b" => "__builtin_ia32_pcmpeqb", + "llvm.x86.mmx.pcmpeq.d" => "__builtin_ia32_pcmpeqd", + "llvm.x86.mmx.pcmpeq.w" => "__builtin_ia32_pcmpeqw", + "llvm.x86.mmx.pcmpgt.b" => "__builtin_ia32_pcmpgtb", + "llvm.x86.mmx.pcmpgt.d" => "__builtin_ia32_pcmpgtd", + "llvm.x86.mmx.pcmpgt.w" => "__builtin_ia32_pcmpgtw", + "llvm.x86.mmx.pextr.w" => "__builtin_ia32_vec_ext_v4hi", + "llvm.x86.mmx.pinsr.w" => "__builtin_ia32_vec_set_v4hi", + "llvm.x86.mmx.pmadd.wd" => "__builtin_ia32_pmaddwd", + "llvm.x86.mmx.pmaxs.w" => "__builtin_ia32_pmaxsw", + "llvm.x86.mmx.pmaxu.b" => "__builtin_ia32_pmaxub", + "llvm.x86.mmx.pmins.w" => "__builtin_ia32_pminsw", + "llvm.x86.mmx.pminu.b" => "__builtin_ia32_pminub", + "llvm.x86.mmx.pmovmskb" => "__builtin_ia32_pmovmskb", + "llvm.x86.mmx.pmulh.w" => "__builtin_ia32_pmulhw", + "llvm.x86.mmx.pmulhu.w" => "__builtin_ia32_pmulhuw", + "llvm.x86.mmx.pmull.w" => "__builtin_ia32_pmullw", + "llvm.x86.mmx.pmulu.dq" => "__builtin_ia32_pmuludq", + "llvm.x86.mmx.por" => "__builtin_ia32_por", + "llvm.x86.mmx.psad.bw" => "__builtin_ia32_psadbw", + "llvm.x86.mmx.psll.d" => "__builtin_ia32_pslld", + "llvm.x86.mmx.psll.q" => "__builtin_ia32_psllq", + "llvm.x86.mmx.psll.w" => "__builtin_ia32_psllw", + "llvm.x86.mmx.pslli.d" => "__builtin_ia32_pslldi", + "llvm.x86.mmx.pslli.q" => "__builtin_ia32_psllqi", + "llvm.x86.mmx.pslli.w" => "__builtin_ia32_psllwi", + "llvm.x86.mmx.psra.d" => "__builtin_ia32_psrad", + "llvm.x86.mmx.psra.w" => "__builtin_ia32_psraw", + "llvm.x86.mmx.psrai.d" => "__builtin_ia32_psradi", + "llvm.x86.mmx.psrai.w" => "__builtin_ia32_psrawi", + "llvm.x86.mmx.psrl.d" => "__builtin_ia32_psrld", + "llvm.x86.mmx.psrl.q" => "__builtin_ia32_psrlq", + "llvm.x86.mmx.psrl.w" => "__builtin_ia32_psrlw", + "llvm.x86.mmx.psrli.d" => "__builtin_ia32_psrldi", + "llvm.x86.mmx.psrli.q" => "__builtin_ia32_psrlqi", + "llvm.x86.mmx.psrli.w" => "__builtin_ia32_psrlwi", + "llvm.x86.mmx.psub.b" => "__builtin_ia32_psubb", + "llvm.x86.mmx.psub.d" => "__builtin_ia32_psubd", + "llvm.x86.mmx.psub.q" => "__builtin_ia32_psubq", + "llvm.x86.mmx.psub.w" => "__builtin_ia32_psubw", + "llvm.x86.mmx.psubs.b" => "__builtin_ia32_psubsb", + "llvm.x86.mmx.psubs.w" => "__builtin_ia32_psubsw", + "llvm.x86.mmx.psubus.b" => "__builtin_ia32_psubusb", + "llvm.x86.mmx.psubus.w" => "__builtin_ia32_psubusw", + "llvm.x86.mmx.punpckhbw" => "__builtin_ia32_punpckhbw", + "llvm.x86.mmx.punpckhdq" => "__builtin_ia32_punpckhdq", + "llvm.x86.mmx.punpckhwd" => "__builtin_ia32_punpckhwd", + "llvm.x86.mmx.punpcklbw" => "__builtin_ia32_punpcklbw", + "llvm.x86.mmx.punpckldq" => "__builtin_ia32_punpckldq", + "llvm.x86.mmx.punpcklwd" => "__builtin_ia32_punpcklwd", + "llvm.x86.mmx.pxor" => "__builtin_ia32_pxor", + "llvm.x86.monitorx" => "__builtin_ia32_monitorx", + "llvm.x86.movdir64b" => "__builtin_ia32_movdir64b", + "llvm.x86.mwaitx" => "__builtin_ia32_mwaitx", + "llvm.x86.pclmulqdq" => "__builtin_ia32_pclmulqdq128", + "llvm.x86.pclmulqdq.256" => "__builtin_ia32_pclmulqdq256", + "llvm.x86.pclmulqdq.512" => "__builtin_ia32_pclmulqdq512", + "llvm.x86.ptwrite32" => "__builtin_ia32_ptwrite32", + "llvm.x86.ptwrite64" => "__builtin_ia32_ptwrite64", + "llvm.x86.rdfsbase.32" => "__builtin_ia32_rdfsbase32", + "llvm.x86.rdfsbase.64" => "__builtin_ia32_rdfsbase64", + "llvm.x86.rdgsbase.32" => "__builtin_ia32_rdgsbase32", + "llvm.x86.rdgsbase.64" => "__builtin_ia32_rdgsbase64", + "llvm.x86.rdpid" => "__builtin_ia32_rdpid", + "llvm.x86.rdpkru" => "__builtin_ia32_rdpkru", + "llvm.x86.rdpmc" => "__builtin_ia32_rdpmc", + "llvm.x86.rdsspd" => "__builtin_ia32_rdsspd", + "llvm.x86.rdsspq" => "__builtin_ia32_rdsspq", + "llvm.x86.rdtsc" => "__builtin_ia32_rdtsc", + "llvm.x86.rdtscp" => "__builtin_ia32_rdtscp", + "llvm.x86.rstorssp" => "__builtin_ia32_rstorssp", + "llvm.x86.saveprevssp" => "__builtin_ia32_saveprevssp", + "llvm.x86.senduipi" => "__builtin_ia32_senduipi", + "llvm.x86.serialize" => "__builtin_ia32_serialize", + "llvm.x86.setssbsy" => "__builtin_ia32_setssbsy", + "llvm.x86.sha1msg1" => "__builtin_ia32_sha1msg1", + "llvm.x86.sha1msg2" => "__builtin_ia32_sha1msg2", + "llvm.x86.sha1nexte" => "__builtin_ia32_sha1nexte", + "llvm.x86.sha1rnds4" => "__builtin_ia32_sha1rnds4", + "llvm.x86.sha256msg1" => "__builtin_ia32_sha256msg1", + "llvm.x86.sha256msg2" => "__builtin_ia32_sha256msg2", + "llvm.x86.sha256rnds2" => "__builtin_ia32_sha256rnds2", + "llvm.x86.slwpcb" => "__builtin_ia32_slwpcb", + "llvm.x86.sse.add.ss" => "__builtin_ia32_addss", + "llvm.x86.sse.cmp.ps" => "__builtin_ia32_cmpps", + "llvm.x86.sse.cmp.ss" => "__builtin_ia32_cmpss", + "llvm.x86.sse.comieq.ss" => "__builtin_ia32_comieq", + "llvm.x86.sse.comige.ss" => "__builtin_ia32_comige", + "llvm.x86.sse.comigt.ss" => "__builtin_ia32_comigt", + "llvm.x86.sse.comile.ss" => "__builtin_ia32_comile", + "llvm.x86.sse.comilt.ss" => "__builtin_ia32_comilt", + "llvm.x86.sse.comineq.ss" => "__builtin_ia32_comineq", + "llvm.x86.sse.cvtpd2pi" => "__builtin_ia32_cvtpd2pi", + "llvm.x86.sse.cvtpi2pd" => "__builtin_ia32_cvtpi2pd", + "llvm.x86.sse.cvtpi2ps" => "__builtin_ia32_cvtpi2ps", + "llvm.x86.sse.cvtps2pi" => "__builtin_ia32_cvtps2pi", + "llvm.x86.sse.cvtsi2ss" => "__builtin_ia32_cvtsi2ss", + "llvm.x86.sse.cvtsi642ss" => "__builtin_ia32_cvtsi642ss", + "llvm.x86.sse.cvtss2si" => "__builtin_ia32_cvtss2si", + "llvm.x86.sse.cvtss2si64" => "__builtin_ia32_cvtss2si64", + "llvm.x86.sse.cvttpd2pi" => "__builtin_ia32_cvttpd2pi", + "llvm.x86.sse.cvttps2pi" => "__builtin_ia32_cvttps2pi", + "llvm.x86.sse.cvttss2si" => "__builtin_ia32_cvttss2si", + "llvm.x86.sse.cvttss2si64" => "__builtin_ia32_cvttss2si64", + "llvm.x86.sse.div.ss" => "__builtin_ia32_divss", + "llvm.x86.sse.max.ps" => "__builtin_ia32_maxps", + "llvm.x86.sse.max.ss" => "__builtin_ia32_maxss", + "llvm.x86.sse.min.ps" => "__builtin_ia32_minps", + "llvm.x86.sse.min.ss" => "__builtin_ia32_minss", + "llvm.x86.sse.movmsk.ps" => "__builtin_ia32_movmskps", + "llvm.x86.sse.mul.ss" => "__builtin_ia32_mulss", + "llvm.x86.sse.pshuf.w" => "__builtin_ia32_pshufw", + "llvm.x86.sse.rcp.ps" => "__builtin_ia32_rcpps", + "llvm.x86.sse.rcp.ss" => "__builtin_ia32_rcpss", + "llvm.x86.sse.rsqrt.ps" => "__builtin_ia32_rsqrtps", + "llvm.x86.sse.rsqrt.ss" => "__builtin_ia32_rsqrtss", + "llvm.x86.sse.sfence" => "__builtin_ia32_sfence", + "llvm.x86.sse.sqrt.ps" => "__builtin_ia32_sqrtps", + "llvm.x86.sse.sqrt.ss" => "__builtin_ia32_sqrtss", + "llvm.x86.sse.storeu.ps" => "__builtin_ia32_storeups", + "llvm.x86.sse.sub.ss" => "__builtin_ia32_subss", + "llvm.x86.sse.ucomieq.ss" => "__builtin_ia32_ucomieq", + "llvm.x86.sse.ucomige.ss" => "__builtin_ia32_ucomige", + "llvm.x86.sse.ucomigt.ss" => "__builtin_ia32_ucomigt", + "llvm.x86.sse.ucomile.ss" => "__builtin_ia32_ucomile", + "llvm.x86.sse.ucomilt.ss" => "__builtin_ia32_ucomilt", + "llvm.x86.sse.ucomineq.ss" => "__builtin_ia32_ucomineq", + "llvm.x86.sse2.add.sd" => "__builtin_ia32_addsd", + "llvm.x86.sse2.clflush" => "__builtin_ia32_clflush", + "llvm.x86.sse2.cmp.pd" => "__builtin_ia32_cmppd", + "llvm.x86.sse2.cmp.sd" => "__builtin_ia32_cmpsd", + "llvm.x86.sse2.comieq.sd" => "__builtin_ia32_comisdeq", + "llvm.x86.sse2.comige.sd" => "__builtin_ia32_comisdge", + "llvm.x86.sse2.comigt.sd" => "__builtin_ia32_comisdgt", + "llvm.x86.sse2.comile.sd" => "__builtin_ia32_comisdle", + "llvm.x86.sse2.comilt.sd" => "__builtin_ia32_comisdlt", + "llvm.x86.sse2.comineq.sd" => "__builtin_ia32_comisdneq", + "llvm.x86.sse2.cvtdq2pd" => "__builtin_ia32_cvtdq2pd", + "llvm.x86.sse2.cvtdq2ps" => "__builtin_ia32_cvtdq2ps", + "llvm.x86.sse2.cvtpd2dq" => "__builtin_ia32_cvtpd2dq", + "llvm.x86.sse2.cvtpd2ps" => "__builtin_ia32_cvtpd2ps", + "llvm.x86.sse2.cvtps2dq" => "__builtin_ia32_cvtps2dq", + "llvm.x86.sse2.cvtps2pd" => "__builtin_ia32_cvtps2pd", + "llvm.x86.sse2.cvtsd2si" => "__builtin_ia32_cvtsd2si", + "llvm.x86.sse2.cvtsd2si64" => "__builtin_ia32_cvtsd2si64", + "llvm.x86.sse2.cvtsd2ss" => "__builtin_ia32_cvtsd2ss", + "llvm.x86.sse2.cvtsi2sd" => "__builtin_ia32_cvtsi2sd", + "llvm.x86.sse2.cvtsi642sd" => "__builtin_ia32_cvtsi642sd", + "llvm.x86.sse2.cvtss2sd" => "__builtin_ia32_cvtss2sd", + "llvm.x86.sse2.cvttpd2dq" => "__builtin_ia32_cvttpd2dq", + "llvm.x86.sse2.cvttps2dq" => "__builtin_ia32_cvttps2dq", + "llvm.x86.sse2.cvttsd2si" => "__builtin_ia32_cvttsd2si", + "llvm.x86.sse2.cvttsd2si64" => "__builtin_ia32_cvttsd2si64", + "llvm.x86.sse2.div.sd" => "__builtin_ia32_divsd", + "llvm.x86.sse2.lfence" => "__builtin_ia32_lfence", + "llvm.x86.sse2.maskmov.dqu" => "__builtin_ia32_maskmovdqu", + "llvm.x86.sse2.max.pd" => "__builtin_ia32_maxpd", + "llvm.x86.sse2.max.sd" => "__builtin_ia32_maxsd", + "llvm.x86.sse2.mfence" => "__builtin_ia32_mfence", + "llvm.x86.sse2.min.pd" => "__builtin_ia32_minpd", + "llvm.x86.sse2.min.sd" => "__builtin_ia32_minsd", + "llvm.x86.sse2.movmsk.pd" => "__builtin_ia32_movmskpd", + "llvm.x86.sse2.mul.sd" => "__builtin_ia32_mulsd", + "llvm.x86.sse2.packssdw.128" => "__builtin_ia32_packssdw128", + "llvm.x86.sse2.packsswb.128" => "__builtin_ia32_packsswb128", + "llvm.x86.sse2.packuswb.128" => "__builtin_ia32_packuswb128", + "llvm.x86.sse2.padds.b" => "__builtin_ia32_paddsb128", + "llvm.x86.sse2.padds.w" => "__builtin_ia32_paddsw128", + "llvm.x86.sse2.paddus.b" => "__builtin_ia32_paddusb128", + "llvm.x86.sse2.paddus.w" => "__builtin_ia32_paddusw128", + "llvm.x86.sse2.pause" => "__builtin_ia32_pause", + "llvm.x86.sse2.pavg.b" => "__builtin_ia32_pavgb128", + "llvm.x86.sse2.pavg.w" => "__builtin_ia32_pavgw128", + "llvm.x86.sse2.pmadd.wd" => "__builtin_ia32_pmaddwd128", + "llvm.x86.sse2.pmaxs.w" => "__builtin_ia32_pmaxsw128", + "llvm.x86.sse2.pmaxu.b" => "__builtin_ia32_pmaxub128", + "llvm.x86.sse2.pmins.w" => "__builtin_ia32_pminsw128", + "llvm.x86.sse2.pminu.b" => "__builtin_ia32_pminub128", + "llvm.x86.sse2.pmovmskb.128" => "__builtin_ia32_pmovmskb128", + "llvm.x86.sse2.pmulh.w" => "__builtin_ia32_pmulhw128", + "llvm.x86.sse2.pmulhu.w" => "__builtin_ia32_pmulhuw128", + "llvm.x86.sse2.pmulu.dq" => "__builtin_ia32_pmuludq128", + "llvm.x86.sse2.psad.bw" => "__builtin_ia32_psadbw128", + "llvm.x86.sse2.pshuf.d" => "__builtin_ia32_pshufd", + "llvm.x86.sse2.pshufh.w" => "__builtin_ia32_pshufhw", + "llvm.x86.sse2.pshufl.w" => "__builtin_ia32_pshuflw", + "llvm.x86.sse2.psll.d" => "__builtin_ia32_pslld128", + "llvm.x86.sse2.psll.dq" => "__builtin_ia32_pslldqi128", + "llvm.x86.sse2.psll.dq.bs" => "__builtin_ia32_pslldqi128_byteshift", + "llvm.x86.sse2.psll.q" => "__builtin_ia32_psllq128", + "llvm.x86.sse2.psll.w" => "__builtin_ia32_psllw128", + "llvm.x86.sse2.pslli.d" => "__builtin_ia32_pslldi128", + "llvm.x86.sse2.pslli.q" => "__builtin_ia32_psllqi128", + "llvm.x86.sse2.pslli.w" => "__builtin_ia32_psllwi128", + "llvm.x86.sse2.psra.d" => "__builtin_ia32_psrad128", + "llvm.x86.sse2.psra.w" => "__builtin_ia32_psraw128", + "llvm.x86.sse2.psrai.d" => "__builtin_ia32_psradi128", + "llvm.x86.sse2.psrai.w" => "__builtin_ia32_psrawi128", + "llvm.x86.sse2.psrl.d" => "__builtin_ia32_psrld128", + "llvm.x86.sse2.psrl.dq" => "__builtin_ia32_psrldqi128", + "llvm.x86.sse2.psrl.dq.bs" => "__builtin_ia32_psrldqi128_byteshift", + "llvm.x86.sse2.psrl.q" => "__builtin_ia32_psrlq128", + "llvm.x86.sse2.psrl.w" => "__builtin_ia32_psrlw128", + "llvm.x86.sse2.psrli.d" => "__builtin_ia32_psrldi128", + "llvm.x86.sse2.psrli.q" => "__builtin_ia32_psrlqi128", + "llvm.x86.sse2.psrli.w" => "__builtin_ia32_psrlwi128", + "llvm.x86.sse2.psubs.b" => "__builtin_ia32_psubsb128", + "llvm.x86.sse2.psubs.w" => "__builtin_ia32_psubsw128", + "llvm.x86.sse2.psubus.b" => "__builtin_ia32_psubusb128", + "llvm.x86.sse2.psubus.w" => "__builtin_ia32_psubusw128", + "llvm.x86.sse2.sqrt.pd" => "__builtin_ia32_sqrtpd", + "llvm.x86.sse2.sqrt.sd" => "__builtin_ia32_sqrtsd", + "llvm.x86.sse2.storel.dq" => "__builtin_ia32_storelv4si", + "llvm.x86.sse2.storeu.dq" => "__builtin_ia32_storedqu", + "llvm.x86.sse2.storeu.pd" => "__builtin_ia32_storeupd", + "llvm.x86.sse2.sub.sd" => "__builtin_ia32_subsd", + "llvm.x86.sse2.ucomieq.sd" => "__builtin_ia32_ucomisdeq", + "llvm.x86.sse2.ucomige.sd" => "__builtin_ia32_ucomisdge", + "llvm.x86.sse2.ucomigt.sd" => "__builtin_ia32_ucomisdgt", + "llvm.x86.sse2.ucomile.sd" => "__builtin_ia32_ucomisdle", + "llvm.x86.sse2.ucomilt.sd" => "__builtin_ia32_ucomisdlt", + "llvm.x86.sse2.ucomineq.sd" => "__builtin_ia32_ucomisdneq", + "llvm.x86.sse3.addsub.pd" => "__builtin_ia32_addsubpd", + "llvm.x86.sse3.addsub.ps" => "__builtin_ia32_addsubps", + "llvm.x86.sse3.hadd.pd" => "__builtin_ia32_haddpd", + "llvm.x86.sse3.hadd.ps" => "__builtin_ia32_haddps", + "llvm.x86.sse3.hsub.pd" => "__builtin_ia32_hsubpd", + "llvm.x86.sse3.hsub.ps" => "__builtin_ia32_hsubps", + "llvm.x86.sse3.ldu.dq" => "__builtin_ia32_lddqu", + "llvm.x86.sse3.monitor" => "__builtin_ia32_monitor", + "llvm.x86.sse3.mwait" => "__builtin_ia32_mwait", + "llvm.x86.sse41.blendpd" => "__builtin_ia32_blendpd", + "llvm.x86.sse41.blendps" => "__builtin_ia32_blendps", + "llvm.x86.sse41.blendvpd" => "__builtin_ia32_blendvpd", + "llvm.x86.sse41.blendvps" => "__builtin_ia32_blendvps", + "llvm.x86.sse41.dppd" => "__builtin_ia32_dppd", + "llvm.x86.sse41.dpps" => "__builtin_ia32_dpps", + "llvm.x86.sse41.extractps" => "__builtin_ia32_extractps128", + "llvm.x86.sse41.insertps" => "__builtin_ia32_insertps128", + "llvm.x86.sse41.movntdqa" => "__builtin_ia32_movntdqa", + "llvm.x86.sse41.mpsadbw" => "__builtin_ia32_mpsadbw128", + "llvm.x86.sse41.packusdw" => "__builtin_ia32_packusdw128", + "llvm.x86.sse41.pblendvb" => "__builtin_ia32_pblendvb128", + "llvm.x86.sse41.pblendw" => "__builtin_ia32_pblendw128", + "llvm.x86.sse41.phminposuw" => "__builtin_ia32_phminposuw128", + "llvm.x86.sse41.pmaxsb" => "__builtin_ia32_pmaxsb128", + "llvm.x86.sse41.pmaxsd" => "__builtin_ia32_pmaxsd128", + "llvm.x86.sse41.pmaxud" => "__builtin_ia32_pmaxud128", + "llvm.x86.sse41.pmaxuw" => "__builtin_ia32_pmaxuw128", + "llvm.x86.sse41.pminsb" => "__builtin_ia32_pminsb128", + "llvm.x86.sse41.pminsd" => "__builtin_ia32_pminsd128", + "llvm.x86.sse41.pminud" => "__builtin_ia32_pminud128", + "llvm.x86.sse41.pminuw" => "__builtin_ia32_pminuw128", + "llvm.x86.sse41.pmovsxbd" => "__builtin_ia32_pmovsxbd128", + "llvm.x86.sse41.pmovsxbq" => "__builtin_ia32_pmovsxbq128", + "llvm.x86.sse41.pmovsxbw" => "__builtin_ia32_pmovsxbw128", + "llvm.x86.sse41.pmovsxdq" => "__builtin_ia32_pmovsxdq128", + "llvm.x86.sse41.pmovsxwd" => "__builtin_ia32_pmovsxwd128", + "llvm.x86.sse41.pmovsxwq" => "__builtin_ia32_pmovsxwq128", + "llvm.x86.sse41.pmovzxbd" => "__builtin_ia32_pmovzxbd128", + "llvm.x86.sse41.pmovzxbq" => "__builtin_ia32_pmovzxbq128", + "llvm.x86.sse41.pmovzxbw" => "__builtin_ia32_pmovzxbw128", + "llvm.x86.sse41.pmovzxdq" => "__builtin_ia32_pmovzxdq128", + "llvm.x86.sse41.pmovzxwd" => "__builtin_ia32_pmovzxwd128", + "llvm.x86.sse41.pmovzxwq" => "__builtin_ia32_pmovzxwq128", + "llvm.x86.sse41.pmuldq" => "__builtin_ia32_pmuldq128", + "llvm.x86.sse41.ptestc" => "__builtin_ia32_ptestc128", + "llvm.x86.sse41.ptestnzc" => "__builtin_ia32_ptestnzc128", + "llvm.x86.sse41.ptestz" => "__builtin_ia32_ptestz128", + "llvm.x86.sse41.round.pd" => "__builtin_ia32_roundpd", + "llvm.x86.sse41.round.ps" => "__builtin_ia32_roundps", + "llvm.x86.sse41.round.sd" => "__builtin_ia32_roundsd", + "llvm.x86.sse41.round.ss" => "__builtin_ia32_roundss", + "llvm.x86.sse42.crc32.32.16" => "__builtin_ia32_crc32hi", + "llvm.x86.sse42.crc32.32.32" => "__builtin_ia32_crc32si", + "llvm.x86.sse42.crc32.32.8" => "__builtin_ia32_crc32qi", + "llvm.x86.sse42.crc32.64.64" => "__builtin_ia32_crc32di", + "llvm.x86.sse42.pcmpestri128" => "__builtin_ia32_pcmpestri128", + "llvm.x86.sse42.pcmpestria128" => "__builtin_ia32_pcmpestria128", + "llvm.x86.sse42.pcmpestric128" => "__builtin_ia32_pcmpestric128", + "llvm.x86.sse42.pcmpestrio128" => "__builtin_ia32_pcmpestrio128", + "llvm.x86.sse42.pcmpestris128" => "__builtin_ia32_pcmpestris128", + "llvm.x86.sse42.pcmpestriz128" => "__builtin_ia32_pcmpestriz128", + "llvm.x86.sse42.pcmpestrm128" => "__builtin_ia32_pcmpestrm128", + "llvm.x86.sse42.pcmpistri128" => "__builtin_ia32_pcmpistri128", + "llvm.x86.sse42.pcmpistria128" => "__builtin_ia32_pcmpistria128", + "llvm.x86.sse42.pcmpistric128" => "__builtin_ia32_pcmpistric128", + "llvm.x86.sse42.pcmpistrio128" => "__builtin_ia32_pcmpistrio128", + "llvm.x86.sse42.pcmpistris128" => "__builtin_ia32_pcmpistris128", + "llvm.x86.sse42.pcmpistriz128" => "__builtin_ia32_pcmpistriz128", + "llvm.x86.sse42.pcmpistrm128" => "__builtin_ia32_pcmpistrm128", + "llvm.x86.sse4a.extrq" => "__builtin_ia32_extrq", + "llvm.x86.sse4a.extrqi" => "__builtin_ia32_extrqi", + "llvm.x86.sse4a.insertq" => "__builtin_ia32_insertq", + "llvm.x86.sse4a.insertqi" => "__builtin_ia32_insertqi", + "llvm.x86.sse4a.movnt.sd" => "__builtin_ia32_movntsd", + "llvm.x86.sse4a.movnt.ss" => "__builtin_ia32_movntss", + "llvm.x86.ssse3.pabs.b" => "__builtin_ia32_pabsb", + "llvm.x86.ssse3.pabs.b.128" => "__builtin_ia32_pabsb128", + "llvm.x86.ssse3.pabs.d" => "__builtin_ia32_pabsd", + "llvm.x86.ssse3.pabs.d.128" => "__builtin_ia32_pabsd128", + "llvm.x86.ssse3.pabs.w" => "__builtin_ia32_pabsw", + "llvm.x86.ssse3.pabs.w.128" => "__builtin_ia32_pabsw128", + "llvm.x86.ssse3.phadd.d" => "__builtin_ia32_phaddd", + "llvm.x86.ssse3.phadd.d.128" => "__builtin_ia32_phaddd128", + "llvm.x86.ssse3.phadd.sw" => "__builtin_ia32_phaddsw", + "llvm.x86.ssse3.phadd.sw.128" => "__builtin_ia32_phaddsw128", + "llvm.x86.ssse3.phadd.w" => "__builtin_ia32_phaddw", + "llvm.x86.ssse3.phadd.w.128" => "__builtin_ia32_phaddw128", + "llvm.x86.ssse3.phsub.d" => "__builtin_ia32_phsubd", + "llvm.x86.ssse3.phsub.d.128" => "__builtin_ia32_phsubd128", + "llvm.x86.ssse3.phsub.sw" => "__builtin_ia32_phsubsw", + "llvm.x86.ssse3.phsub.sw.128" => "__builtin_ia32_phsubsw128", + "llvm.x86.ssse3.phsub.w" => "__builtin_ia32_phsubw", + "llvm.x86.ssse3.phsub.w.128" => "__builtin_ia32_phsubw128", + "llvm.x86.ssse3.pmadd.ub.sw" => "__builtin_ia32_pmaddubsw", + "llvm.x86.ssse3.pmadd.ub.sw.128" => "__builtin_ia32_pmaddubsw128", + "llvm.x86.ssse3.pmul.hr.sw" => "__builtin_ia32_pmulhrsw", + "llvm.x86.ssse3.pmul.hr.sw.128" => "__builtin_ia32_pmulhrsw128", + "llvm.x86.ssse3.pshuf.b" => "__builtin_ia32_pshufb", + "llvm.x86.ssse3.pshuf.b.128" => "__builtin_ia32_pshufb128", + "llvm.x86.ssse3.psign.b" => "__builtin_ia32_psignb", + "llvm.x86.ssse3.psign.b.128" => "__builtin_ia32_psignb128", + "llvm.x86.ssse3.psign.d" => "__builtin_ia32_psignd", + "llvm.x86.ssse3.psign.d.128" => "__builtin_ia32_psignd128", + "llvm.x86.ssse3.psign.w" => "__builtin_ia32_psignw", + "llvm.x86.ssse3.psign.w.128" => "__builtin_ia32_psignw128", + "llvm.x86.sttilecfg" => "__builtin_ia32_tile_storeconfig", + "llvm.x86.stui" => "__builtin_ia32_stui", + "llvm.x86.subborrow.u32" => "__builtin_ia32_subborrow_u32", + "llvm.x86.subborrow.u64" => "__builtin_ia32_subborrow_u64", + "llvm.x86.tbm.bextri.u32" => "__builtin_ia32_bextri_u32", + "llvm.x86.tbm.bextri.u64" => "__builtin_ia32_bextri_u64", + "llvm.x86.tdpbf16ps" => "__builtin_ia32_tdpbf16ps", + "llvm.x86.tdpbf16ps.internal" => "__builtin_ia32_tdpbf16ps_internal", + "llvm.x86.tdpbssd" => "__builtin_ia32_tdpbssd", + "llvm.x86.tdpbssd.internal" => "__builtin_ia32_tdpbssd_internal", + "llvm.x86.tdpbsud" => "__builtin_ia32_tdpbsud", + "llvm.x86.tdpbsud.internal" => "__builtin_ia32_tdpbsud_internal", + "llvm.x86.tdpbusd" => "__builtin_ia32_tdpbusd", + "llvm.x86.tdpbusd.internal" => "__builtin_ia32_tdpbusd_internal", + "llvm.x86.tdpbuud" => "__builtin_ia32_tdpbuud", + "llvm.x86.tdpbuud.internal" => "__builtin_ia32_tdpbuud_internal", + "llvm.x86.testui" => "__builtin_ia32_testui", + "llvm.x86.tileloadd64" => "__builtin_ia32_tileloadd64", + "llvm.x86.tileloadd64.internal" => "__builtin_ia32_tileloadd64_internal", + "llvm.x86.tileloaddt164" => "__builtin_ia32_tileloaddt164", + "llvm.x86.tileloaddt164.internal" => "__builtin_ia32_tileloaddt164_internal", + "llvm.x86.tilerelease" => "__builtin_ia32_tilerelease", + "llvm.x86.tilestored64" => "__builtin_ia32_tilestored64", + "llvm.x86.tilestored64.internal" => "__builtin_ia32_tilestored64_internal", + "llvm.x86.tilezero" => "__builtin_ia32_tilezero", + "llvm.x86.tilezero.internal" => "__builtin_ia32_tilezero_internal", + "llvm.x86.tpause" => "__builtin_ia32_tpause", + "llvm.x86.umonitor" => "__builtin_ia32_umonitor", + "llvm.x86.umwait" => "__builtin_ia32_umwait", + "llvm.x86.vcvtph2ps.128" => "__builtin_ia32_vcvtph2ps", + "llvm.x86.vcvtph2ps.256" => "__builtin_ia32_vcvtph2ps256", + "llvm.x86.vcvtps2ph.128" => "__builtin_ia32_vcvtps2ph", + "llvm.x86.vcvtps2ph.256" => "__builtin_ia32_vcvtps2ph256", + "llvm.x86.vgf2p8affineinvqb.128" => "__builtin_ia32_vgf2p8affineinvqb_v16qi", + "llvm.x86.vgf2p8affineinvqb.256" => "__builtin_ia32_vgf2p8affineinvqb_v32qi", + "llvm.x86.vgf2p8affineinvqb.512" => "__builtin_ia32_vgf2p8affineinvqb_v64qi", + "llvm.x86.vgf2p8affineqb.128" => "__builtin_ia32_vgf2p8affineqb_v16qi", + "llvm.x86.vgf2p8affineqb.256" => "__builtin_ia32_vgf2p8affineqb_v32qi", + "llvm.x86.vgf2p8affineqb.512" => "__builtin_ia32_vgf2p8affineqb_v64qi", + "llvm.x86.vgf2p8mulb.128" => "__builtin_ia32_vgf2p8mulb_v16qi", + "llvm.x86.vgf2p8mulb.256" => "__builtin_ia32_vgf2p8mulb_v32qi", + "llvm.x86.vgf2p8mulb.512" => "__builtin_ia32_vgf2p8mulb_v64qi", + "llvm.x86.wbinvd" => "__builtin_ia32_wbinvd", + "llvm.x86.wbnoinvd" => "__builtin_ia32_wbnoinvd", + "llvm.x86.wrfsbase.32" => "__builtin_ia32_wrfsbase32", + "llvm.x86.wrfsbase.64" => "__builtin_ia32_wrfsbase64", + "llvm.x86.wrgsbase.32" => "__builtin_ia32_wrgsbase32", + "llvm.x86.wrgsbase.64" => "__builtin_ia32_wrgsbase64", + "llvm.x86.wrpkru" => "__builtin_ia32_wrpkru", + "llvm.x86.wrssd" => "__builtin_ia32_wrssd", + "llvm.x86.wrssq" => "__builtin_ia32_wrssq", + "llvm.x86.wrussd" => "__builtin_ia32_wrussd", + "llvm.x86.wrussq" => "__builtin_ia32_wrussq", + "llvm.x86.xabort" => "__builtin_ia32_xabort", + "llvm.x86.xbegin" => "__builtin_ia32_xbegin", + "llvm.x86.xend" => "__builtin_ia32_xend", + "llvm.x86.xop.vfrcz.pd" => "__builtin_ia32_vfrczpd", + "llvm.x86.xop.vfrcz.pd.256" => "__builtin_ia32_vfrczpd256", + "llvm.x86.xop.vfrcz.ps" => "__builtin_ia32_vfrczps", + "llvm.x86.xop.vfrcz.ps.256" => "__builtin_ia32_vfrczps256", + "llvm.x86.xop.vfrcz.sd" => "__builtin_ia32_vfrczsd", + "llvm.x86.xop.vfrcz.ss" => "__builtin_ia32_vfrczss", + "llvm.x86.xop.vpcmov" => "__builtin_ia32_vpcmov", + "llvm.x86.xop.vpcmov.256" => "__builtin_ia32_vpcmov_256", + "llvm.x86.xop.vpcomb" => "__builtin_ia32_vpcomb", + "llvm.x86.xop.vpcomd" => "__builtin_ia32_vpcomd", + "llvm.x86.xop.vpcomq" => "__builtin_ia32_vpcomq", + "llvm.x86.xop.vpcomub" => "__builtin_ia32_vpcomub", + "llvm.x86.xop.vpcomud" => "__builtin_ia32_vpcomud", + "llvm.x86.xop.vpcomuq" => "__builtin_ia32_vpcomuq", + "llvm.x86.xop.vpcomuw" => "__builtin_ia32_vpcomuw", + "llvm.x86.xop.vpcomw" => "__builtin_ia32_vpcomw", + "llvm.x86.xop.vpermil2pd" => "__builtin_ia32_vpermil2pd", + "llvm.x86.xop.vpermil2pd.256" => "__builtin_ia32_vpermil2pd256", + "llvm.x86.xop.vpermil2ps" => "__builtin_ia32_vpermil2ps", + "llvm.x86.xop.vpermil2ps.256" => "__builtin_ia32_vpermil2ps256", + "llvm.x86.xop.vphaddbd" => "__builtin_ia32_vphaddbd", + "llvm.x86.xop.vphaddbq" => "__builtin_ia32_vphaddbq", + "llvm.x86.xop.vphaddbw" => "__builtin_ia32_vphaddbw", + "llvm.x86.xop.vphadddq" => "__builtin_ia32_vphadddq", + "llvm.x86.xop.vphaddubd" => "__builtin_ia32_vphaddubd", + "llvm.x86.xop.vphaddubq" => "__builtin_ia32_vphaddubq", + "llvm.x86.xop.vphaddubw" => "__builtin_ia32_vphaddubw", + "llvm.x86.xop.vphaddudq" => "__builtin_ia32_vphaddudq", + "llvm.x86.xop.vphadduwd" => "__builtin_ia32_vphadduwd", + "llvm.x86.xop.vphadduwq" => "__builtin_ia32_vphadduwq", + "llvm.x86.xop.vphaddwd" => "__builtin_ia32_vphaddwd", + "llvm.x86.xop.vphaddwq" => "__builtin_ia32_vphaddwq", + "llvm.x86.xop.vphsubbw" => "__builtin_ia32_vphsubbw", + "llvm.x86.xop.vphsubdq" => "__builtin_ia32_vphsubdq", + "llvm.x86.xop.vphsubwd" => "__builtin_ia32_vphsubwd", + "llvm.x86.xop.vpmacsdd" => "__builtin_ia32_vpmacsdd", + "llvm.x86.xop.vpmacsdqh" => "__builtin_ia32_vpmacsdqh", + "llvm.x86.xop.vpmacsdql" => "__builtin_ia32_vpmacsdql", + "llvm.x86.xop.vpmacssdd" => "__builtin_ia32_vpmacssdd", + "llvm.x86.xop.vpmacssdqh" => "__builtin_ia32_vpmacssdqh", + "llvm.x86.xop.vpmacssdql" => "__builtin_ia32_vpmacssdql", + "llvm.x86.xop.vpmacsswd" => "__builtin_ia32_vpmacsswd", + "llvm.x86.xop.vpmacssww" => "__builtin_ia32_vpmacssww", + "llvm.x86.xop.vpmacswd" => "__builtin_ia32_vpmacswd", + "llvm.x86.xop.vpmacsww" => "__builtin_ia32_vpmacsww", + "llvm.x86.xop.vpmadcsswd" => "__builtin_ia32_vpmadcsswd", + "llvm.x86.xop.vpmadcswd" => "__builtin_ia32_vpmadcswd", + "llvm.x86.xop.vpperm" => "__builtin_ia32_vpperm", + "llvm.x86.xop.vprotb" => "__builtin_ia32_vprotb", + "llvm.x86.xop.vprotbi" => "__builtin_ia32_vprotbi", + "llvm.x86.xop.vprotd" => "__builtin_ia32_vprotd", + "llvm.x86.xop.vprotdi" => "__builtin_ia32_vprotdi", + "llvm.x86.xop.vprotq" => "__builtin_ia32_vprotq", + "llvm.x86.xop.vprotqi" => "__builtin_ia32_vprotqi", + "llvm.x86.xop.vprotw" => "__builtin_ia32_vprotw", + "llvm.x86.xop.vprotwi" => "__builtin_ia32_vprotwi", + "llvm.x86.xop.vpshab" => "__builtin_ia32_vpshab", + "llvm.x86.xop.vpshad" => "__builtin_ia32_vpshad", + "llvm.x86.xop.vpshaq" => "__builtin_ia32_vpshaq", + "llvm.x86.xop.vpshaw" => "__builtin_ia32_vpshaw", + "llvm.x86.xop.vpshlb" => "__builtin_ia32_vpshlb", + "llvm.x86.xop.vpshld" => "__builtin_ia32_vpshld", + "llvm.x86.xop.vpshlq" => "__builtin_ia32_vpshlq", + "llvm.x86.xop.vpshlw" => "__builtin_ia32_vpshlw", + "llvm.x86.xresldtrk" => "__builtin_ia32_xresldtrk", + "llvm.x86.xsusldtrk" => "__builtin_ia32_xsusldtrk", + "llvm.x86.xtest" => "__builtin_ia32_xtest", + // xcore + "llvm.xcore.bitrev" => "__builtin_bitrev", + "llvm.xcore.getid" => "__builtin_getid", + "llvm.xcore.getps" => "__builtin_getps", + "llvm.xcore.setps" => "__builtin_setps", + _ => unimplemented!("***** unsupported LLVM intrinsic {}", name), +} diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs b/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs new file mode 100644 index 000000000..1b089f08f --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs @@ -0,0 +1,250 @@ +use std::borrow::Cow; + +use gccjit::{Function, FunctionPtrType, RValue, ToRValue}; + +use crate::{context::CodegenCx, builder::Builder}; + +pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc, 'tcx>, gcc_func: FunctionPtrType<'gcc>, mut args: Cow<'b, [RValue<'gcc>]>, func_name: &str) -> Cow<'b, [RValue<'gcc>]> { + // Some LLVM intrinsics do not map 1-to-1 to GCC intrinsics, so we add the missing + // arguments here. + if gcc_func.get_param_count() != args.len() { + match &*func_name { + "__builtin_ia32_pmuldq512_mask" | "__builtin_ia32_pmuludq512_mask" + // FIXME(antoyo): the following intrinsics has 4 (or 5) arguments according to the doc, but is defined with 2 (or 3) arguments in library/stdarch/crates/core_arch/src/x86/avx512f.rs. + | "__builtin_ia32_pmaxsd512_mask" | "__builtin_ia32_pmaxsq512_mask" | "__builtin_ia32_pmaxsq256_mask" + | "__builtin_ia32_pmaxsq128_mask" | "__builtin_ia32_maxps512_mask" | "__builtin_ia32_maxpd512_mask" + | "__builtin_ia32_pmaxud512_mask" | "__builtin_ia32_pmaxuq512_mask" | "__builtin_ia32_pmaxuq256_mask" + | "__builtin_ia32_pmaxuq128_mask" + | "__builtin_ia32_pminsd512_mask" | "__builtin_ia32_pminsq512_mask" | "__builtin_ia32_pminsq256_mask" + | "__builtin_ia32_pminsq128_mask" | "__builtin_ia32_minps512_mask" | "__builtin_ia32_minpd512_mask" + | "__builtin_ia32_pminud512_mask" | "__builtin_ia32_pminuq512_mask" | "__builtin_ia32_pminuq256_mask" + | "__builtin_ia32_pminuq128_mask" | "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" + => { + // TODO: refactor by separating those intrinsics outside of this branch. + let add_before_last_arg = + match &*func_name { + "__builtin_ia32_maxps512_mask" | "__builtin_ia32_maxpd512_mask" + | "__builtin_ia32_minps512_mask" | "__builtin_ia32_minpd512_mask" + | "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" => true, + _ => false, + }; + let new_first_arg_is_zero = + match &*func_name { + "__builtin_ia32_pmaxuq256_mask" | "__builtin_ia32_pmaxuq128_mask" + | "__builtin_ia32_pminuq256_mask" | "__builtin_ia32_pminuq128_mask" => true, + _ => false + }; + let arg3_index = + match &*func_name { + "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" => 1, + _ => 2, + }; + let mut new_args = args.to_vec(); + let arg3_type = gcc_func.get_param_type(arg3_index); + let first_arg = + if new_first_arg_is_zero { + let vector_type = arg3_type.dyncast_vector().expect("vector type"); + let zero = builder.context.new_rvalue_zero(vector_type.get_element_type()); + let num_units = vector_type.get_num_units(); + builder.context.new_rvalue_from_vector(None, arg3_type, &vec![zero; num_units]) + } + else { + builder.current_func().new_local(None, arg3_type, "undefined_for_intrinsic").to_rvalue() + }; + if add_before_last_arg { + new_args.insert(new_args.len() - 1, first_arg); + } + else { + new_args.push(first_arg); + } + let arg4_index = + match &*func_name { + "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" => 2, + _ => 3, + }; + let arg4_type = gcc_func.get_param_type(arg4_index); + let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1); + if add_before_last_arg { + new_args.insert(new_args.len() - 1, minus_one); + } + else { + new_args.push(minus_one); + } + args = new_args.into(); + }, + "__builtin_ia32_pternlogd512_mask" | "__builtin_ia32_pternlogd256_mask" + | "__builtin_ia32_pternlogd128_mask" | "__builtin_ia32_pternlogq512_mask" + | "__builtin_ia32_pternlogq256_mask" | "__builtin_ia32_pternlogq128_mask" => { + let mut new_args = args.to_vec(); + let arg5_type = gcc_func.get_param_type(4); + let minus_one = builder.context.new_rvalue_from_int(arg5_type, -1); + new_args.push(minus_one); + args = new_args.into(); + }, + "__builtin_ia32_vfmaddps512_mask" | "__builtin_ia32_vfmaddpd512_mask" => { + let mut new_args = args.to_vec(); + + let mut last_arg = None; + if args.len() == 4 { + last_arg = new_args.pop(); + } + + let arg4_type = gcc_func.get_param_type(3); + let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1); + new_args.push(minus_one); + + if args.len() == 3 { + // Both llvm.fma.v16f32 and llvm.x86.avx512.vfmadd.ps.512 maps to + // the same GCC intrinsic, but the former has 3 parameters and the + // latter has 4 so it doesn't require this additional argument. + let arg5_type = gcc_func.get_param_type(4); + new_args.push(builder.context.new_rvalue_from_int(arg5_type, 4)); + } + + if let Some(last_arg) = last_arg { + new_args.push(last_arg); + } + + args = new_args.into(); + }, + "__builtin_ia32_addps512_mask" | "__builtin_ia32_addpd512_mask" + | "__builtin_ia32_subps512_mask" | "__builtin_ia32_subpd512_mask" + | "__builtin_ia32_mulps512_mask" | "__builtin_ia32_mulpd512_mask" + | "__builtin_ia32_divps512_mask" | "__builtin_ia32_divpd512_mask" => { + let mut new_args = args.to_vec(); + let last_arg = new_args.pop().expect("last arg"); + let arg3_type = gcc_func.get_param_type(2); + let undefined = builder.current_func().new_local(None, arg3_type, "undefined_for_intrinsic").to_rvalue(); + new_args.push(undefined); + let arg4_type = gcc_func.get_param_type(3); + let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1); + new_args.push(minus_one); + new_args.push(last_arg); + args = new_args.into(); + }, + "__builtin_ia32_vfmaddsubps512_mask" | "__builtin_ia32_vfmaddsubpd512_mask" => { + let mut new_args = args.to_vec(); + let last_arg = new_args.pop().expect("last arg"); + let arg4_type = gcc_func.get_param_type(3); + let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1); + new_args.push(minus_one); + new_args.push(last_arg); + args = new_args.into(); + }, + _ => (), + } + } + + args +} + +pub fn ignore_arg_cast(func_name: &str, index: usize, args_len: usize) -> bool { + // NOTE: these intrinsics have missing parameters before the last one, so ignore the + // last argument type check. + // FIXME(antoyo): find a way to refactor in order to avoid this hack. + match func_name { + "__builtin_ia32_maxps512_mask" | "__builtin_ia32_maxpd512_mask" + | "__builtin_ia32_minps512_mask" | "__builtin_ia32_minpd512_mask" | "__builtin_ia32_sqrtps512_mask" + | "__builtin_ia32_sqrtpd512_mask" | "__builtin_ia32_addps512_mask" | "__builtin_ia32_addpd512_mask" + | "__builtin_ia32_subps512_mask" | "__builtin_ia32_subpd512_mask" + | "__builtin_ia32_mulps512_mask" | "__builtin_ia32_mulpd512_mask" + | "__builtin_ia32_divps512_mask" | "__builtin_ia32_divpd512_mask" + | "__builtin_ia32_vfmaddsubps512_mask" | "__builtin_ia32_vfmaddsubpd512_mask" => { + if index == args_len - 1 { + return true; + } + }, + "__builtin_ia32_vfmaddps512_mask" | "__builtin_ia32_vfmaddpd512_mask" => { + // Since there are two LLVM intrinsics that map to each of these GCC builtins and only + // one of them has a missing parameter before the last one, we check the number of + // arguments to distinguish those cases. + if args_len == 4 && index == args_len - 1 { + return true; + } + }, + _ => (), + } + + false +} + +#[cfg(not(feature="master"))] +pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function<'gcc> { + match name { + "llvm.x86.xgetbv" => { + let gcc_name = "__builtin_trap"; + let func = cx.context.get_builtin_function(gcc_name); + cx.functions.borrow_mut().insert(gcc_name.to_string(), func); + return func; + }, + _ => unimplemented!("unsupported LLVM intrinsic {}", name), + } +} + +#[cfg(feature="master")] +pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function<'gcc> { + let gcc_name = match name { + "llvm.x86.xgetbv" => "__builtin_ia32_xgetbv", + // NOTE: this doc specifies the equivalent GCC builtins: http://huonw.github.io/llvmint/llvmint/x86/index.html + "llvm.sqrt.v2f64" => "__builtin_ia32_sqrtpd", + "llvm.x86.avx512.pmul.dq.512" => "__builtin_ia32_pmuldq512_mask", + "llvm.x86.avx512.pmulu.dq.512" => "__builtin_ia32_pmuludq512_mask", + "llvm.x86.avx512.mask.pmaxs.q.256" => "__builtin_ia32_pmaxsq256_mask", + "llvm.x86.avx512.mask.pmaxs.q.128" => "__builtin_ia32_pmaxsq128_mask", + "llvm.x86.avx512.max.ps.512" => "__builtin_ia32_maxps512_mask", + "llvm.x86.avx512.max.pd.512" => "__builtin_ia32_maxpd512_mask", + "llvm.x86.avx512.mask.pmaxu.q.256" => "__builtin_ia32_pmaxuq256_mask", + "llvm.x86.avx512.mask.pmaxu.q.128" => "__builtin_ia32_pmaxuq128_mask", + "llvm.x86.avx512.mask.pmins.q.256" => "__builtin_ia32_pminsq256_mask", + "llvm.x86.avx512.mask.pmins.q.128" => "__builtin_ia32_pminsq128_mask", + "llvm.x86.avx512.min.ps.512" => "__builtin_ia32_minps512_mask", + "llvm.x86.avx512.min.pd.512" => "__builtin_ia32_minpd512_mask", + "llvm.x86.avx512.mask.pminu.q.256" => "__builtin_ia32_pminuq256_mask", + "llvm.x86.avx512.mask.pminu.q.128" => "__builtin_ia32_pminuq128_mask", + "llvm.fma.v16f32" => "__builtin_ia32_vfmaddps512_mask", + "llvm.fma.v8f64" => "__builtin_ia32_vfmaddpd512_mask", + "llvm.x86.avx512.vfmaddsub.ps.512" => "__builtin_ia32_vfmaddsubps512_mask", + "llvm.x86.avx512.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_mask", + "llvm.x86.avx512.pternlog.d.512" => "__builtin_ia32_pternlogd512_mask", + "llvm.x86.avx512.pternlog.d.256" => "__builtin_ia32_pternlogd256_mask", + "llvm.x86.avx512.pternlog.d.128" => "__builtin_ia32_pternlogd128_mask", + "llvm.x86.avx512.pternlog.q.512" => "__builtin_ia32_pternlogq512_mask", + "llvm.x86.avx512.pternlog.q.256" => "__builtin_ia32_pternlogq256_mask", + "llvm.x86.avx512.pternlog.q.128" => "__builtin_ia32_pternlogq128_mask", + "llvm.x86.avx512.add.ps.512" => "__builtin_ia32_addps512_mask", + "llvm.x86.avx512.add.pd.512" => "__builtin_ia32_addpd512_mask", + "llvm.x86.avx512.sub.ps.512" => "__builtin_ia32_subps512_mask", + "llvm.x86.avx512.sub.pd.512" => "__builtin_ia32_subpd512_mask", + "llvm.x86.avx512.mul.ps.512" => "__builtin_ia32_mulps512_mask", + "llvm.x86.avx512.mul.pd.512" => "__builtin_ia32_mulpd512_mask", + "llvm.x86.avx512.div.ps.512" => "__builtin_ia32_divps512_mask", + "llvm.x86.avx512.div.pd.512" => "__builtin_ia32_divpd512_mask", + "llvm.x86.avx512.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_mask", + "llvm.x86.avx512.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_mask", + + // The above doc points to unknown builtins for the following, so override them: + "llvm.x86.avx2.gather.d.d" => "__builtin_ia32_gathersiv4si", + "llvm.x86.avx2.gather.d.d.256" => "__builtin_ia32_gathersiv8si", + "llvm.x86.avx2.gather.d.ps" => "__builtin_ia32_gathersiv4sf", + "llvm.x86.avx2.gather.d.ps.256" => "__builtin_ia32_gathersiv8sf", + "llvm.x86.avx2.gather.d.q" => "__builtin_ia32_gathersiv2di", + "llvm.x86.avx2.gather.d.q.256" => "__builtin_ia32_gathersiv4di", + "llvm.x86.avx2.gather.d.pd" => "__builtin_ia32_gathersiv2df", + "llvm.x86.avx2.gather.d.pd.256" => "__builtin_ia32_gathersiv4df", + "llvm.x86.avx2.gather.q.d" => "__builtin_ia32_gatherdiv4si", + "llvm.x86.avx2.gather.q.d.256" => "__builtin_ia32_gatherdiv4si256", + "llvm.x86.avx2.gather.q.ps" => "__builtin_ia32_gatherdiv4sf", + "llvm.x86.avx2.gather.q.ps.256" => "__builtin_ia32_gatherdiv4sf256", + "llvm.x86.avx2.gather.q.q" => "__builtin_ia32_gatherdiv2di", + "llvm.x86.avx2.gather.q.q.256" => "__builtin_ia32_gatherdiv4di", + "llvm.x86.avx2.gather.q.pd" => "__builtin_ia32_gatherdiv2df", + "llvm.x86.avx2.gather.q.pd.256" => "__builtin_ia32_gatherdiv4df", + "" => "", + // NOTE: this file is generated by https://github.com/GuillaumeGomez/llvmint/blob/master/generate_list.py + _ => include!("archs.rs"), + }; + + let func = cx.context.get_target_builtin_function(gcc_name); + cx.functions.borrow_mut().insert(gcc_name.to_string(), func); + func +} diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs new file mode 100644 index 000000000..5fbdedac0 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs @@ -0,0 +1,1134 @@ +pub mod llvm; +mod simd; + +use gccjit::{ComparisonOp, Function, RValue, ToRValue, Type, UnaryOp, FunctionType}; +use rustc_codegen_ssa::MemFlags; +use rustc_codegen_ssa::base::wants_msvc_seh; +use rustc_codegen_ssa::common::{IntPredicate, span_invalid_monomorphization_error}; +use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue}; +use rustc_codegen_ssa::mir::place::PlaceRef; +use rustc_codegen_ssa::traits::{ArgAbiMethods, BaseTypeMethods, BuilderMethods, ConstMethods, IntrinsicCallMethods}; +use rustc_middle::bug; +use rustc_middle::ty::{self, Instance, Ty}; +use rustc_middle::ty::layout::LayoutOf; +use rustc_span::{Span, Symbol, symbol::kw, sym}; +use rustc_target::abi::HasDataLayout; +use rustc_target::abi::call::{ArgAbi, FnAbi, PassMode}; +use rustc_target::spec::PanicStrategy; + +use crate::abi::GccType; +use crate::builder::Builder; +use crate::common::{SignType, TypeReflection}; +use crate::context::CodegenCx; +use crate::type_of::LayoutGccExt; +use crate::intrinsic::simd::generic_simd_intrinsic; + +fn get_simple_intrinsic<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, name: Symbol) -> Option<Function<'gcc>> { + let gcc_name = match name { + sym::sqrtf32 => "sqrtf", + sym::sqrtf64 => "sqrt", + sym::powif32 => "__builtin_powif", + sym::powif64 => "__builtin_powi", + sym::sinf32 => "sinf", + sym::sinf64 => "sin", + sym::cosf32 => "cosf", + sym::cosf64 => "cos", + sym::powf32 => "powf", + sym::powf64 => "pow", + sym::expf32 => "expf", + sym::expf64 => "exp", + sym::exp2f32 => "exp2f", + sym::exp2f64 => "exp2", + sym::logf32 => "logf", + sym::logf64 => "log", + sym::log10f32 => "log10f", + sym::log10f64 => "log10", + sym::log2f32 => "log2f", + sym::log2f64 => "log2", + sym::fmaf32 => "fmaf", + sym::fmaf64 => "fma", + sym::fabsf32 => "fabsf", + sym::fabsf64 => "fabs", + sym::minnumf32 => "fminf", + sym::minnumf64 => "fmin", + sym::maxnumf32 => "fmaxf", + sym::maxnumf64 => "fmax", + sym::copysignf32 => "copysignf", + sym::copysignf64 => "copysign", + sym::floorf32 => "floorf", + sym::floorf64 => "floor", + sym::ceilf32 => "ceilf", + sym::ceilf64 => "ceil", + sym::truncf32 => "truncf", + sym::truncf64 => "trunc", + sym::rintf32 => "rintf", + sym::rintf64 => "rint", + sym::nearbyintf32 => "nearbyintf", + sym::nearbyintf64 => "nearbyint", + sym::roundf32 => "roundf", + sym::roundf64 => "round", + sym::abort => "abort", + _ => return None, + }; + Some(cx.context.get_builtin_function(&gcc_name)) +} + +impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { + fn codegen_intrinsic_call(&mut self, instance: Instance<'tcx>, fn_abi: &FnAbi<'tcx, Ty<'tcx>>, args: &[OperandRef<'tcx, RValue<'gcc>>], llresult: RValue<'gcc>, span: Span) { + let tcx = self.tcx; + let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all()); + + let (def_id, substs) = match *callee_ty.kind() { + ty::FnDef(def_id, substs) => (def_id, substs), + _ => bug!("expected fn item type, found {}", callee_ty), + }; + + let sig = callee_ty.fn_sig(tcx); + let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig); + let arg_tys = sig.inputs(); + let ret_ty = sig.output(); + let name = tcx.item_name(def_id); + let name_str = name.as_str(); + + let llret_ty = self.layout_of(ret_ty).gcc_type(self, true); + let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout); + + let simple = get_simple_intrinsic(self, name); + let llval = + match name { + _ if simple.is_some() => { + // FIXME(antoyo): remove this cast when the API supports function. + let func = unsafe { std::mem::transmute(simple.expect("simple")) }; + self.call(self.type_void(), func, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None) + }, + sym::likely => { + self.expect(args[0].immediate(), true) + } + sym::unlikely => { + self.expect(args[0].immediate(), false) + } + kw::Try => { + try_intrinsic( + self, + args[0].immediate(), + args[1].immediate(), + args[2].immediate(), + llresult, + ); + return; + } + sym::breakpoint => { + unimplemented!(); + } + sym::va_copy => { + unimplemented!(); + } + sym::va_arg => { + unimplemented!(); + } + + sym::volatile_load | sym::unaligned_volatile_load => { + let tp_ty = substs.type_at(0); + let mut ptr = args[0].immediate(); + if let PassMode::Cast(ty) = fn_abi.ret.mode { + ptr = self.pointercast(ptr, self.type_ptr_to(ty.gcc_type(self))); + } + let load = self.volatile_load(ptr.get_type(), ptr); + // TODO(antoyo): set alignment. + self.to_immediate(load, self.layout_of(tp_ty)) + } + sym::volatile_store => { + let dst = args[0].deref(self.cx()); + args[1].val.volatile_store(self, dst); + return; + } + sym::unaligned_volatile_store => { + let dst = args[0].deref(self.cx()); + args[1].val.unaligned_volatile_store(self, dst); + return; + } + sym::prefetch_read_data + | sym::prefetch_write_data + | sym::prefetch_read_instruction + | sym::prefetch_write_instruction => { + unimplemented!(); + } + sym::ctlz + | sym::ctlz_nonzero + | sym::cttz + | sym::cttz_nonzero + | sym::ctpop + | sym::bswap + | sym::bitreverse + | sym::rotate_left + | sym::rotate_right + | sym::saturating_add + | sym::saturating_sub => { + let ty = arg_tys[0]; + match int_type_width_signed(ty, self) { + Some((width, signed)) => match name { + sym::ctlz | sym::cttz => { + let func = self.current_func.borrow().expect("func"); + let then_block = func.new_block("then"); + let else_block = func.new_block("else"); + let after_block = func.new_block("after"); + + let arg = args[0].immediate(); + let result = func.new_local(None, arg.get_type(), "zeros"); + let zero = self.cx.gcc_zero(arg.get_type()); + let cond = self.gcc_icmp(IntPredicate::IntEQ, arg, zero); + self.llbb().end_with_conditional(None, cond, then_block, else_block); + + let zero_result = self.cx.gcc_uint(arg.get_type(), width); + then_block.add_assignment(None, result, zero_result); + then_block.end_with_jump(None, after_block); + + // NOTE: since jumps were added in a place + // count_leading_zeroes() does not expect, the current block + // in the state need to be updated. + self.switch_to_block(else_block); + + let zeros = + match name { + sym::ctlz => self.count_leading_zeroes(width, arg), + sym::cttz => self.count_trailing_zeroes(width, arg), + _ => unreachable!(), + }; + self.llbb().add_assignment(None, result, zeros); + self.llbb().end_with_jump(None, after_block); + + // NOTE: since jumps were added in a place rustc does not + // expect, the current block in the state need to be updated. + self.switch_to_block(after_block); + + result.to_rvalue() + } + sym::ctlz_nonzero => { + self.count_leading_zeroes(width, args[0].immediate()) + }, + sym::cttz_nonzero => { + self.count_trailing_zeroes(width, args[0].immediate()) + } + sym::ctpop => self.pop_count(args[0].immediate()), + sym::bswap => { + if width == 8 { + args[0].immediate() // byte swap a u8/i8 is just a no-op + } + else { + self.gcc_bswap(args[0].immediate(), width) + } + }, + sym::bitreverse => self.bit_reverse(width, args[0].immediate()), + sym::rotate_left | sym::rotate_right => { + // TODO(antoyo): implement using algorithm from: + // https://blog.regehr.org/archives/1063 + // for other platforms. + let is_left = name == sym::rotate_left; + let val = args[0].immediate(); + let raw_shift = args[1].immediate(); + if is_left { + self.rotate_left(val, raw_shift, width) + } + else { + self.rotate_right(val, raw_shift, width) + } + }, + sym::saturating_add => { + self.saturating_add(args[0].immediate(), args[1].immediate(), signed, width) + }, + sym::saturating_sub => { + self.saturating_sub(args[0].immediate(), args[1].immediate(), signed, width) + }, + _ => bug!(), + }, + None => { + span_invalid_monomorphization_error( + tcx.sess, + span, + &format!( + "invalid monomorphization of `{}` intrinsic: \ + expected basic integer type, found `{}`", + name, ty + ), + ); + return; + } + } + } + + sym::raw_eq => { + use rustc_target::abi::Abi::*; + let tp_ty = substs.type_at(0); + let layout = self.layout_of(tp_ty).layout; + let _use_integer_compare = match layout.abi() { + Scalar(_) | ScalarPair(_, _) => true, + Uninhabited | Vector { .. } => false, + Aggregate { .. } => { + // For rusty ABIs, small aggregates are actually passed + // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`), + // so we re-use that same threshold here. + layout.size() <= self.data_layout().pointer_size * 2 + } + }; + + let a = args[0].immediate(); + let b = args[1].immediate(); + if layout.size().bytes() == 0 { + self.const_bool(true) + } + /*else if use_integer_compare { + let integer_ty = self.type_ix(layout.size.bits()); // FIXME(antoyo): LLVM creates an integer of 96 bits for [i32; 3], but gcc doesn't support this, so it creates an integer of 128 bits. + let ptr_ty = self.type_ptr_to(integer_ty); + let a_ptr = self.bitcast(a, ptr_ty); + let a_val = self.load(integer_ty, a_ptr, layout.align.abi); + let b_ptr = self.bitcast(b, ptr_ty); + let b_val = self.load(integer_ty, b_ptr, layout.align.abi); + self.icmp(IntPredicate::IntEQ, a_val, b_val) + }*/ + else { + let void_ptr_type = self.context.new_type::<*const ()>(); + let a_ptr = self.bitcast(a, void_ptr_type); + let b_ptr = self.bitcast(b, void_ptr_type); + let n = self.context.new_cast(None, self.const_usize(layout.size().bytes()), self.sizet_type); + let builtin = self.context.get_builtin_function("memcmp"); + let cmp = self.context.new_call(None, builtin, &[a_ptr, b_ptr, n]); + self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0)) + } + } + + sym::black_box => { + args[0].val.store(self, result); + + let block = self.llbb(); + let extended_asm = block.add_extended_asm(None, ""); + extended_asm.add_input_operand(None, "r", result.llval); + extended_asm.add_clobber("memory"); + extended_asm.set_volatile_flag(true); + + // We have copied the value to `result` already. + return; + } + + _ if name_str.starts_with("simd_") => { + match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) { + Ok(llval) => llval, + Err(()) => return, + } + } + + _ => bug!("unknown intrinsic '{}'", name), + }; + + if !fn_abi.ret.is_ignore() { + if let PassMode::Cast(ty) = fn_abi.ret.mode { + let ptr_llty = self.type_ptr_to(ty.gcc_type(self)); + let ptr = self.pointercast(result.llval, ptr_llty); + self.store(llval, ptr, result.align); + } + else { + OperandRef::from_immediate_or_packed_pair(self, llval, result.layout) + .val + .store(self, result); + } + } + } + + fn abort(&mut self) { + let func = self.context.get_builtin_function("abort"); + let func: RValue<'gcc> = unsafe { std::mem::transmute(func) }; + self.call(self.type_void(), func, &[], None); + } + + fn assume(&mut self, value: Self::Value) { + // TODO(antoyo): switch to assume when it exists. + // Or use something like this: + // #define __assume(cond) do { if (!(cond)) __builtin_unreachable(); } while (0) + self.expect(value, true); + } + + fn expect(&mut self, cond: Self::Value, _expected: bool) -> Self::Value { + // TODO(antoyo) + cond + } + + fn type_test(&mut self, _pointer: Self::Value, _typeid: Self::Value) -> Self::Value { + // Unsupported. + self.context.new_rvalue_from_int(self.int_type, 0) + } + + fn type_checked_load( + &mut self, + _llvtable: Self::Value, + _vtable_byte_offset: u64, + _typeid: Self::Value, + ) -> Self::Value { + // Unsupported. + self.context.new_rvalue_from_int(self.int_type, 0) + } + + fn va_start(&mut self, _va_list: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } + + fn va_end(&mut self, _va_list: RValue<'gcc>) -> RValue<'gcc> { + unimplemented!(); + } +} + +impl<'a, 'gcc, 'tcx> ArgAbiMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { + fn store_fn_arg(&mut self, arg_abi: &ArgAbi<'tcx, Ty<'tcx>>, idx: &mut usize, dst: PlaceRef<'tcx, Self::Value>) { + arg_abi.store_fn_arg(self, idx, dst) + } + + fn store_arg(&mut self, arg_abi: &ArgAbi<'tcx, Ty<'tcx>>, val: RValue<'gcc>, dst: PlaceRef<'tcx, RValue<'gcc>>) { + arg_abi.store(self, val, dst) + } + + fn arg_memory_ty(&self, arg_abi: &ArgAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> { + arg_abi.memory_ty(self) + } +} + +pub trait ArgAbiExt<'gcc, 'tcx> { + fn memory_ty(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; + fn store(&self, bx: &mut Builder<'_, 'gcc, 'tcx>, val: RValue<'gcc>, dst: PlaceRef<'tcx, RValue<'gcc>>); + fn store_fn_arg(&self, bx: &mut Builder<'_, 'gcc, 'tcx>, idx: &mut usize, dst: PlaceRef<'tcx, RValue<'gcc>>); +} + +impl<'gcc, 'tcx> ArgAbiExt<'gcc, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> { + /// Gets the LLVM type for a place of the original Rust type of + /// this argument/return, i.e., the result of `type_of::type_of`. + fn memory_ty(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> { + self.layout.gcc_type(cx, true) + } + + /// Stores a direct/indirect value described by this ArgAbi into a + /// place for the original Rust type of this argument/return. + /// Can be used for both storing formal arguments into Rust variables + /// or results of call/invoke instructions into their destinations. + fn store(&self, bx: &mut Builder<'_, 'gcc, 'tcx>, val: RValue<'gcc>, dst: PlaceRef<'tcx, RValue<'gcc>>) { + if self.is_ignore() { + return; + } + if self.is_sized_indirect() { + OperandValue::Ref(val, None, self.layout.align.abi).store(bx, dst) + } + else if self.is_unsized_indirect() { + bug!("unsized `ArgAbi` must be handled through `store_fn_arg`"); + } + else if let PassMode::Cast(cast) = self.mode { + // FIXME(eddyb): Figure out when the simpler Store is safe, clang + // uses it for i16 -> {i8, i8}, but not for i24 -> {i8, i8, i8}. + let can_store_through_cast_ptr = false; + if can_store_through_cast_ptr { + let cast_ptr_llty = bx.type_ptr_to(cast.gcc_type(bx)); + let cast_dst = bx.pointercast(dst.llval, cast_ptr_llty); + bx.store(val, cast_dst, self.layout.align.abi); + } + else { + // The actual return type is a struct, but the ABI + // adaptation code has cast it into some scalar type. The + // code that follows is the only reliable way I have + // found to do a transform like i64 -> {i32,i32}. + // Basically we dump the data onto the stack then memcpy it. + // + // Other approaches I tried: + // - Casting rust ret pointer to the foreign type and using Store + // is (a) unsafe if size of foreign type > size of rust type and + // (b) runs afoul of strict aliasing rules, yielding invalid + // assembly under -O (specifically, the store gets removed). + // - Truncating foreign type to correct integral type and then + // bitcasting to the struct type yields invalid cast errors. + + // We instead thus allocate some scratch space... + let scratch_size = cast.size(bx); + let scratch_align = cast.align(bx); + let llscratch = bx.alloca(cast.gcc_type(bx), scratch_align); + bx.lifetime_start(llscratch, scratch_size); + + // ... where we first store the value... + bx.store(val, llscratch, scratch_align); + + // ... and then memcpy it to the intended destination. + bx.memcpy( + dst.llval, + self.layout.align.abi, + llscratch, + scratch_align, + bx.const_usize(self.layout.size.bytes()), + MemFlags::empty(), + ); + + bx.lifetime_end(llscratch, scratch_size); + } + } + else { + OperandValue::Immediate(val).store(bx, dst); + } + } + + fn store_fn_arg<'a>(&self, bx: &mut Builder<'a, 'gcc, 'tcx>, idx: &mut usize, dst: PlaceRef<'tcx, RValue<'gcc>>) { + let mut next = || { + let val = bx.current_func().get_param(*idx as i32); + *idx += 1; + val.to_rvalue() + }; + match self.mode { + PassMode::Ignore => {}, + PassMode::Pair(..) => { + OperandValue::Pair(next(), next()).store(bx, dst); + }, + PassMode::Indirect { extra_attrs: Some(_), .. } => { + OperandValue::Ref(next(), Some(next()), self.layout.align.abi).store(bx, dst); + }, + PassMode::Direct(_) | PassMode::Indirect { extra_attrs: None, .. } | PassMode::Cast(_) => { + let next_arg = next(); + self.store(bx, next_arg, dst); + }, + } + } +} + +fn int_type_width_signed<'gcc, 'tcx>(ty: Ty<'tcx>, cx: &CodegenCx<'gcc, 'tcx>) -> Option<(u64, bool)> { + match ty.kind() { + ty::Int(t) => Some(( + match t { + rustc_middle::ty::IntTy::Isize => u64::from(cx.tcx.sess.target.pointer_width), + rustc_middle::ty::IntTy::I8 => 8, + rustc_middle::ty::IntTy::I16 => 16, + rustc_middle::ty::IntTy::I32 => 32, + rustc_middle::ty::IntTy::I64 => 64, + rustc_middle::ty::IntTy::I128 => 128, + }, + true, + )), + ty::Uint(t) => Some(( + match t { + rustc_middle::ty::UintTy::Usize => u64::from(cx.tcx.sess.target.pointer_width), + rustc_middle::ty::UintTy::U8 => 8, + rustc_middle::ty::UintTy::U16 => 16, + rustc_middle::ty::UintTy::U32 => 32, + rustc_middle::ty::UintTy::U64 => 64, + rustc_middle::ty::UintTy::U128 => 128, + }, + false, + )), + _ => None, + } +} + +impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { + fn bit_reverse(&mut self, width: u64, value: RValue<'gcc>) -> RValue<'gcc> { + let result_type = value.get_type(); + let typ = result_type.to_unsigned(self.cx); + + let value = + if result_type.is_signed(self.cx) { + self.gcc_int_cast(value, typ) + } + else { + value + }; + + let context = &self.cx.context; + let result = + match width { + 8 => { + // First step. + let left = self.and(value, context.new_rvalue_from_int(typ, 0xF0)); + let left = self.lshr(left, context.new_rvalue_from_int(typ, 4)); + let right = self.and(value, context.new_rvalue_from_int(typ, 0x0F)); + let right = self.shl(right, context.new_rvalue_from_int(typ, 4)); + let step1 = self.or(left, right); + + // Second step. + let left = self.and(step1, context.new_rvalue_from_int(typ, 0xCC)); + let left = self.lshr(left, context.new_rvalue_from_int(typ, 2)); + let right = self.and(step1, context.new_rvalue_from_int(typ, 0x33)); + let right = self.shl(right, context.new_rvalue_from_int(typ, 2)); + let step2 = self.or(left, right); + + // Third step. + let left = self.and(step2, context.new_rvalue_from_int(typ, 0xAA)); + let left = self.lshr(left, context.new_rvalue_from_int(typ, 1)); + let right = self.and(step2, context.new_rvalue_from_int(typ, 0x55)); + let right = self.shl(right, context.new_rvalue_from_int(typ, 1)); + let step3 = self.or(left, right); + + step3 + }, + 16 => { + // First step. + let left = self.and(value, context.new_rvalue_from_int(typ, 0x5555)); + let left = self.shl(left, context.new_rvalue_from_int(typ, 1)); + let right = self.and(value, context.new_rvalue_from_int(typ, 0xAAAA)); + let right = self.lshr(right, context.new_rvalue_from_int(typ, 1)); + let step1 = self.or(left, right); + + // Second step. + let left = self.and(step1, context.new_rvalue_from_int(typ, 0x3333)); + let left = self.shl(left, context.new_rvalue_from_int(typ, 2)); + let right = self.and(step1, context.new_rvalue_from_int(typ, 0xCCCC)); + let right = self.lshr(right, context.new_rvalue_from_int(typ, 2)); + let step2 = self.or(left, right); + + // Third step. + let left = self.and(step2, context.new_rvalue_from_int(typ, 0x0F0F)); + let left = self.shl(left, context.new_rvalue_from_int(typ, 4)); + let right = self.and(step2, context.new_rvalue_from_int(typ, 0xF0F0)); + let right = self.lshr(right, context.new_rvalue_from_int(typ, 4)); + let step3 = self.or(left, right); + + // Fourth step. + let left = self.and(step3, context.new_rvalue_from_int(typ, 0x00FF)); + let left = self.shl(left, context.new_rvalue_from_int(typ, 8)); + let right = self.and(step3, context.new_rvalue_from_int(typ, 0xFF00)); + let right = self.lshr(right, context.new_rvalue_from_int(typ, 8)); + let step4 = self.or(left, right); + + step4 + }, + 32 => { + // TODO(antoyo): Refactor with other implementations. + // First step. + let left = self.and(value, context.new_rvalue_from_long(typ, 0x55555555)); + let left = self.shl(left, context.new_rvalue_from_long(typ, 1)); + let right = self.and(value, context.new_rvalue_from_long(typ, 0xAAAAAAAA)); + let right = self.lshr(right, context.new_rvalue_from_long(typ, 1)); + let step1 = self.or(left, right); + + // Second step. + let left = self.and(step1, context.new_rvalue_from_long(typ, 0x33333333)); + let left = self.shl(left, context.new_rvalue_from_long(typ, 2)); + let right = self.and(step1, context.new_rvalue_from_long(typ, 0xCCCCCCCC)); + let right = self.lshr(right, context.new_rvalue_from_long(typ, 2)); + let step2 = self.or(left, right); + + // Third step. + let left = self.and(step2, context.new_rvalue_from_long(typ, 0x0F0F0F0F)); + let left = self.shl(left, context.new_rvalue_from_long(typ, 4)); + let right = self.and(step2, context.new_rvalue_from_long(typ, 0xF0F0F0F0)); + let right = self.lshr(right, context.new_rvalue_from_long(typ, 4)); + let step3 = self.or(left, right); + + // Fourth step. + let left = self.and(step3, context.new_rvalue_from_long(typ, 0x00FF00FF)); + let left = self.shl(left, context.new_rvalue_from_long(typ, 8)); + let right = self.and(step3, context.new_rvalue_from_long(typ, 0xFF00FF00)); + let right = self.lshr(right, context.new_rvalue_from_long(typ, 8)); + let step4 = self.or(left, right); + + // Fifth step. + let left = self.and(step4, context.new_rvalue_from_long(typ, 0x0000FFFF)); + let left = self.shl(left, context.new_rvalue_from_long(typ, 16)); + let right = self.and(step4, context.new_rvalue_from_long(typ, 0xFFFF0000)); + let right = self.lshr(right, context.new_rvalue_from_long(typ, 16)); + let step5 = self.or(left, right); + + step5 + }, + 64 => { + // First step. + let left = self.shl(value, context.new_rvalue_from_long(typ, 32)); + let right = self.lshr(value, context.new_rvalue_from_long(typ, 32)); + let step1 = self.or(left, right); + + // Second step. + let left = self.and(step1, context.new_rvalue_from_long(typ, 0x0001FFFF0001FFFF)); + let left = self.shl(left, context.new_rvalue_from_long(typ, 15)); + let right = self.and(step1, context.new_rvalue_from_long(typ, 0xFFFE0000FFFE0000u64 as i64)); // TODO(antoyo): transmute the number instead? + let right = self.lshr(right, context.new_rvalue_from_long(typ, 17)); + let step2 = self.or(left, right); + + // Third step. + let left = self.lshr(step2, context.new_rvalue_from_long(typ, 10)); + let left = self.xor(step2, left); + let temp = self.and(left, context.new_rvalue_from_long(typ, 0x003F801F003F801F)); + + let left = self.shl(temp, context.new_rvalue_from_long(typ, 10)); + let left = self.or(temp, left); + let step3 = self.xor(left, step2); + + // Fourth step. + let left = self.lshr(step3, context.new_rvalue_from_long(typ, 4)); + let left = self.xor(step3, left); + let temp = self.and(left, context.new_rvalue_from_long(typ, 0x0E0384210E038421)); + + let left = self.shl(temp, context.new_rvalue_from_long(typ, 4)); + let left = self.or(temp, left); + let step4 = self.xor(left, step3); + + // Fifth step. + let left = self.lshr(step4, context.new_rvalue_from_long(typ, 2)); + let left = self.xor(step4, left); + let temp = self.and(left, context.new_rvalue_from_long(typ, 0x2248884222488842)); + + let left = self.shl(temp, context.new_rvalue_from_long(typ, 2)); + let left = self.or(temp, left); + let step5 = self.xor(left, step4); + + step5 + }, + 128 => { + // TODO(antoyo): find a more efficient implementation? + let sixty_four = self.gcc_int(typ, 64); + let right_shift = self.gcc_lshr(value, sixty_four); + let high = self.gcc_int_cast(right_shift, self.u64_type); + let low = self.gcc_int_cast(value, self.u64_type); + + let reversed_high = self.bit_reverse(64, high); + let reversed_low = self.bit_reverse(64, low); + + let new_low = self.gcc_int_cast(reversed_high, typ); + let new_high = self.shl(self.gcc_int_cast(reversed_low, typ), sixty_four); + + self.gcc_or(new_low, new_high) + }, + _ => { + panic!("cannot bit reverse with width = {}", width); + }, + }; + + self.gcc_int_cast(result, result_type) + } + + fn count_leading_zeroes(&mut self, width: u64, arg: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): use width? + let arg_type = arg.get_type(); + let count_leading_zeroes = + // TODO(antoyo): write a new function Type::is_compatible_with(&Type) and use it here + // instead of using is_uint(). + if arg_type.is_uint(&self.cx) { + "__builtin_clz" + } + else if arg_type.is_ulong(&self.cx) { + "__builtin_clzl" + } + else if arg_type.is_ulonglong(&self.cx) { + "__builtin_clzll" + } + else if width == 128 { + // Algorithm from: https://stackoverflow.com/a/28433850/389119 + let array_type = self.context.new_array_type(None, arg_type, 3); + let result = self.current_func() + .new_local(None, array_type, "count_loading_zeroes_results"); + + let sixty_four = self.const_uint(arg_type, 64); + let shift = self.lshr(arg, sixty_four); + let high = self.gcc_int_cast(shift, self.u64_type); + let low = self.gcc_int_cast(arg, self.u64_type); + + let zero = self.context.new_rvalue_zero(self.usize_type); + let one = self.context.new_rvalue_one(self.usize_type); + let two = self.context.new_rvalue_from_long(self.usize_type, 2); + + let clzll = self.context.get_builtin_function("__builtin_clzll"); + + let first_elem = self.context.new_array_access(None, result, zero); + let first_value = self.gcc_int_cast(self.context.new_call(None, clzll, &[high]), arg_type); + self.llbb() + .add_assignment(None, first_elem, first_value); + + let second_elem = self.context.new_array_access(None, result, one); + let cast = self.gcc_int_cast(self.context.new_call(None, clzll, &[low]), arg_type); + let second_value = self.add(cast, sixty_four); + self.llbb() + .add_assignment(None, second_elem, second_value); + + let third_elem = self.context.new_array_access(None, result, two); + let third_value = self.const_uint(arg_type, 128); + self.llbb() + .add_assignment(None, third_elem, third_value); + + let not_high = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, high); + let not_low = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, low); + let not_low_and_not_high = not_low & not_high; + let index = not_high + not_low_and_not_high; + // NOTE: the following cast is necessary to avoid a GIMPLE verification failure in + // gcc. + // TODO(antoyo): do the correct verification in libgccjit to avoid an error at the + // compilation stage. + let index = self.context.new_cast(None, index, self.i32_type); + + let res = self.context.new_array_access(None, result, index); + + return self.gcc_int_cast(res.to_rvalue(), arg_type); + } + else { + let count_leading_zeroes = self.context.get_builtin_function("__builtin_clzll"); + let arg = self.context.new_cast(None, arg, self.ulonglong_type); + let diff = self.ulonglong_type.get_size() as i64 - arg_type.get_size() as i64; + let diff = self.context.new_rvalue_from_long(self.int_type, diff * 8); + let res = self.context.new_call(None, count_leading_zeroes, &[arg]) - diff; + return self.context.new_cast(None, res, arg_type); + }; + let count_leading_zeroes = self.context.get_builtin_function(count_leading_zeroes); + let res = self.context.new_call(None, count_leading_zeroes, &[arg]); + self.context.new_cast(None, res, arg_type) + } + + fn count_trailing_zeroes(&mut self, _width: u64, arg: RValue<'gcc>) -> RValue<'gcc> { + let result_type = arg.get_type(); + let arg = + if result_type.is_signed(self.cx) { + let new_type = result_type.to_unsigned(self.cx); + self.gcc_int_cast(arg, new_type) + } + else { + arg + }; + let arg_type = arg.get_type(); + let (count_trailing_zeroes, expected_type) = + // TODO(antoyo): write a new function Type::is_compatible_with(&Type) and use it here + // instead of using is_uint(). + if arg_type.is_uchar(&self.cx) || arg_type.is_ushort(&self.cx) || arg_type.is_uint(&self.cx) { + // NOTE: we don't need to & 0xFF for uchar because the result is undefined on zero. + ("__builtin_ctz", self.cx.uint_type) + } + else if arg_type.is_ulong(&self.cx) { + ("__builtin_ctzl", self.cx.ulong_type) + } + else if arg_type.is_ulonglong(&self.cx) { + ("__builtin_ctzll", self.cx.ulonglong_type) + } + else if arg_type.is_u128(&self.cx) { + // Adapted from the algorithm to count leading zeroes from: https://stackoverflow.com/a/28433850/389119 + let array_type = self.context.new_array_type(None, arg_type, 3); + let result = self.current_func() + .new_local(None, array_type, "count_loading_zeroes_results"); + + let sixty_four = self.gcc_int(arg_type, 64); + let shift = self.gcc_lshr(arg, sixty_four); + let high = self.gcc_int_cast(shift, self.u64_type); + let low = self.gcc_int_cast(arg, self.u64_type); + + let zero = self.context.new_rvalue_zero(self.usize_type); + let one = self.context.new_rvalue_one(self.usize_type); + let two = self.context.new_rvalue_from_long(self.usize_type, 2); + + let ctzll = self.context.get_builtin_function("__builtin_ctzll"); + + let first_elem = self.context.new_array_access(None, result, zero); + let first_value = self.gcc_int_cast(self.context.new_call(None, ctzll, &[low]), arg_type); + self.llbb() + .add_assignment(None, first_elem, first_value); + + let second_elem = self.context.new_array_access(None, result, one); + let second_value = self.gcc_add(self.gcc_int_cast(self.context.new_call(None, ctzll, &[high]), arg_type), sixty_four); + self.llbb() + .add_assignment(None, second_elem, second_value); + + let third_elem = self.context.new_array_access(None, result, two); + let third_value = self.gcc_int(arg_type, 128); + self.llbb() + .add_assignment(None, third_elem, third_value); + + let not_low = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, low); + let not_high = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, high); + let not_low_and_not_high = not_low & not_high; + let index = not_low + not_low_and_not_high; + // NOTE: the following cast is necessary to avoid a GIMPLE verification failure in + // gcc. + // TODO(antoyo): do the correct verification in libgccjit to avoid an error at the + // compilation stage. + let index = self.context.new_cast(None, index, self.i32_type); + + let res = self.context.new_array_access(None, result, index); + + return self.gcc_int_cast(res.to_rvalue(), result_type); + } + else { + let count_trailing_zeroes = self.context.get_builtin_function("__builtin_ctzll"); + let arg_size = arg_type.get_size(); + let casted_arg = self.context.new_cast(None, arg, self.ulonglong_type); + let byte_diff = self.ulonglong_type.get_size() as i64 - arg_size as i64; + let diff = self.context.new_rvalue_from_long(self.int_type, byte_diff * 8); + let mask = self.context.new_rvalue_from_long(arg_type, -1); // To get the value with all bits set. + let masked = mask & self.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, arg); + let cond = self.context.new_comparison(None, ComparisonOp::Equals, masked, mask); + let diff = diff * self.context.new_cast(None, cond, self.int_type); + let res = self.context.new_call(None, count_trailing_zeroes, &[casted_arg]) - diff; + return self.context.new_cast(None, res, result_type); + }; + let count_trailing_zeroes = self.context.get_builtin_function(count_trailing_zeroes); + let arg = + if arg_type != expected_type { + self.context.new_cast(None, arg, expected_type) + } + else { + arg + }; + let res = self.context.new_call(None, count_trailing_zeroes, &[arg]); + self.context.new_cast(None, res, result_type) + } + + fn pop_count(&mut self, value: RValue<'gcc>) -> RValue<'gcc> { + // TODO(antoyo): use the optimized version with fewer operations. + let result_type = value.get_type(); + let value_type = result_type.to_unsigned(self.cx); + + let value = + if result_type.is_signed(self.cx) { + self.gcc_int_cast(value, value_type) + } + else { + value + }; + + if value_type.is_u128(&self.cx) { + // TODO(antoyo): implement in the normal algorithm below to have a more efficient + // implementation (that does not require a call to __popcountdi2). + let popcount = self.context.get_builtin_function("__builtin_popcountll"); + let sixty_four = self.gcc_int(value_type, 64); + let right_shift = self.gcc_lshr(value, sixty_four); + let high = self.gcc_int_cast(right_shift, self.cx.ulonglong_type); + let high = self.context.new_call(None, popcount, &[high]); + let low = self.gcc_int_cast(value, self.cx.ulonglong_type); + let low = self.context.new_call(None, popcount, &[low]); + let res = high + low; + return self.gcc_int_cast(res, result_type); + } + + // First step. + let mask = self.context.new_rvalue_from_long(value_type, 0x5555555555555555); + let left = value & mask; + let shifted = value >> self.context.new_rvalue_from_int(value_type, 1); + let right = shifted & mask; + let value = left + right; + + // Second step. + let mask = self.context.new_rvalue_from_long(value_type, 0x3333333333333333); + let left = value & mask; + let shifted = value >> self.context.new_rvalue_from_int(value_type, 2); + let right = shifted & mask; + let value = left + right; + + // Third step. + let mask = self.context.new_rvalue_from_long(value_type, 0x0F0F0F0F0F0F0F0F); + let left = value & mask; + let shifted = value >> self.context.new_rvalue_from_int(value_type, 4); + let right = shifted & mask; + let value = left + right; + + if value_type.is_u8(&self.cx) { + return self.context.new_cast(None, value, result_type); + } + + // Fourth step. + let mask = self.context.new_rvalue_from_long(value_type, 0x00FF00FF00FF00FF); + let left = value & mask; + let shifted = value >> self.context.new_rvalue_from_int(value_type, 8); + let right = shifted & mask; + let value = left + right; + + if value_type.is_u16(&self.cx) { + return self.context.new_cast(None, value, result_type); + } + + // Fifth step. + let mask = self.context.new_rvalue_from_long(value_type, 0x0000FFFF0000FFFF); + let left = value & mask; + let shifted = value >> self.context.new_rvalue_from_int(value_type, 16); + let right = shifted & mask; + let value = left + right; + + if value_type.is_u32(&self.cx) { + return self.context.new_cast(None, value, result_type); + } + + // Sixth step. + let mask = self.context.new_rvalue_from_long(value_type, 0x00000000FFFFFFFF); + let left = value & mask; + let shifted = value >> self.context.new_rvalue_from_int(value_type, 32); + let right = shifted & mask; + let value = left + right; + + self.context.new_cast(None, value, result_type) + } + + // Algorithm from: https://blog.regehr.org/archives/1063 + fn rotate_left(&mut self, value: RValue<'gcc>, shift: RValue<'gcc>, width: u64) -> RValue<'gcc> { + let max = self.const_uint(shift.get_type(), width); + let shift = self.urem(shift, max); + let lhs = self.shl(value, shift); + let result_neg = self.neg(shift); + let result_and = + self.and( + result_neg, + self.const_uint(shift.get_type(), width - 1), + ); + let rhs = self.lshr(value, result_and); + self.or(lhs, rhs) + } + + // Algorithm from: https://blog.regehr.org/archives/1063 + fn rotate_right(&mut self, value: RValue<'gcc>, shift: RValue<'gcc>, width: u64) -> RValue<'gcc> { + let max = self.const_uint(shift.get_type(), width); + let shift = self.urem(shift, max); + let lhs = self.lshr(value, shift); + let result_neg = self.neg(shift); + let result_and = + self.and( + result_neg, + self.const_uint(shift.get_type(), width - 1), + ); + let rhs = self.shl(value, result_and); + self.or(lhs, rhs) + } + + fn saturating_add(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> { + let result_type = lhs.get_type(); + if signed { + // Based on algorithm from: https://stackoverflow.com/a/56531252/389119 + let func = self.current_func.borrow().expect("func"); + let res = func.new_local(None, result_type, "saturating_sum"); + let supports_native_type = self.is_native_int_type(result_type); + let overflow = + if supports_native_type { + let func_name = + match width { + 8 => "__builtin_add_overflow", + 16 => "__builtin_add_overflow", + 32 => "__builtin_sadd_overflow", + 64 => "__builtin_saddll_overflow", + 128 => "__builtin_add_overflow", + _ => unreachable!(), + }; + let overflow_func = self.context.get_builtin_function(func_name); + self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None) + } + else { + let func_name = + match width { + 128 => "__rust_i128_addo", + _ => unreachable!(), + }; + let param_a = self.context.new_parameter(None, result_type, "a"); + let param_b = self.context.new_parameter(None, result_type, "b"); + let result_field = self.context.new_field(None, result_type, "result"); + let overflow_field = self.context.new_field(None, self.bool_type, "overflow"); + let return_type = self.context.new_struct_type(None, "result_overflow", &[result_field, overflow_field]); + let func = self.context.new_function(None, FunctionType::Extern, return_type.as_type(), &[param_a, param_b], func_name, false); + let result = self.context.new_call(None, func, &[lhs, rhs]); + let overflow = result.access_field(None, overflow_field); + let int_result = result.access_field(None, result_field); + self.llbb().add_assignment(None, res, int_result); + overflow + }; + + let then_block = func.new_block("then"); + let after_block = func.new_block("after"); + + // Return `result_type`'s maximum or minimum value on overflow + // NOTE: convert the type to unsigned to have an unsigned shift. + let unsigned_type = result_type.to_unsigned(&self.cx); + let shifted = self.gcc_lshr(self.gcc_int_cast(lhs, unsigned_type), self.gcc_int(unsigned_type, width as i64 - 1)); + let uint_max = self.gcc_not(self.gcc_int(unsigned_type, 0)); + let int_max = self.gcc_lshr(uint_max, self.gcc_int(unsigned_type, 1)); + then_block.add_assignment(None, res, self.gcc_int_cast(self.gcc_add(shifted, int_max), result_type)); + then_block.end_with_jump(None, after_block); + + self.llbb().end_with_conditional(None, overflow, then_block, after_block); + + // NOTE: since jumps were added in a place rustc does not + // expect, the current block in the state need to be updated. + self.switch_to_block(after_block); + + res.to_rvalue() + } + else { + // Algorithm from: http://locklessinc.com/articles/sat_arithmetic/ + let res = self.gcc_add(lhs, rhs); + let cond = self.gcc_icmp(IntPredicate::IntULT, res, lhs); + let value = self.gcc_neg(self.gcc_int_cast(cond, result_type)); + self.gcc_or(res, value) + } + } + + // Algorithm from: https://locklessinc.com/articles/sat_arithmetic/ + fn saturating_sub(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> { + let result_type = lhs.get_type(); + if signed { + // Based on algorithm from: https://stackoverflow.com/a/56531252/389119 + let func = self.current_func.borrow().expect("func"); + let res = func.new_local(None, result_type, "saturating_diff"); + let supports_native_type = self.is_native_int_type(result_type); + let overflow = + if supports_native_type { + let func_name = + match width { + 8 => "__builtin_sub_overflow", + 16 => "__builtin_sub_overflow", + 32 => "__builtin_ssub_overflow", + 64 => "__builtin_ssubll_overflow", + 128 => "__builtin_sub_overflow", + _ => unreachable!(), + }; + let overflow_func = self.context.get_builtin_function(func_name); + self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None) + } + else { + let func_name = + match width { + 128 => "__rust_i128_subo", + _ => unreachable!(), + }; + let param_a = self.context.new_parameter(None, result_type, "a"); + let param_b = self.context.new_parameter(None, result_type, "b"); + let result_field = self.context.new_field(None, result_type, "result"); + let overflow_field = self.context.new_field(None, self.bool_type, "overflow"); + let return_type = self.context.new_struct_type(None, "result_overflow", &[result_field, overflow_field]); + let func = self.context.new_function(None, FunctionType::Extern, return_type.as_type(), &[param_a, param_b], func_name, false); + let result = self.context.new_call(None, func, &[lhs, rhs]); + let overflow = result.access_field(None, overflow_field); + let int_result = result.access_field(None, result_field); + self.llbb().add_assignment(None, res, int_result); + overflow + }; + + let then_block = func.new_block("then"); + let after_block = func.new_block("after"); + + // Return `result_type`'s maximum or minimum value on overflow + // NOTE: convert the type to unsigned to have an unsigned shift. + let unsigned_type = result_type.to_unsigned(&self.cx); + let shifted = self.gcc_lshr(self.gcc_int_cast(lhs, unsigned_type), self.gcc_int(unsigned_type, width as i64 - 1)); + let uint_max = self.gcc_not(self.gcc_int(unsigned_type, 0)); + let int_max = self.gcc_lshr(uint_max, self.gcc_int(unsigned_type, 1)); + then_block.add_assignment(None, res, self.gcc_int_cast(self.gcc_add(shifted, int_max), result_type)); + then_block.end_with_jump(None, after_block); + + self.llbb().end_with_conditional(None, overflow, then_block, after_block); + + // NOTE: since jumps were added in a place rustc does not + // expect, the current block in the state need to be updated. + self.switch_to_block(after_block); + + res.to_rvalue() + } + else { + let res = self.gcc_sub(lhs, rhs); + let comparison = self.gcc_icmp(IntPredicate::IntULE, res, lhs); + let value = self.gcc_neg(self.gcc_int_cast(comparison, result_type)); + self.gcc_and(res, value) + } + } +} + +fn try_intrinsic<'gcc, 'tcx>(bx: &mut Builder<'_, 'gcc, 'tcx>, try_func: RValue<'gcc>, data: RValue<'gcc>, _catch_func: RValue<'gcc>, dest: RValue<'gcc>) { + // NOTE: the `|| true` here is to use the panic=abort strategy with panic=unwind too + if bx.sess().panic_strategy() == PanicStrategy::Abort || true { + // TODO(bjorn3): Properly implement unwinding and remove the `|| true` once this is done. + bx.call(bx.type_void(), try_func, &[data], None); + // Return 0 unconditionally from the intrinsic call; + // we can never unwind. + let ret_align = bx.tcx.data_layout.i32_align.abi; + bx.store(bx.const_i32(0), dest, ret_align); + } + else if wants_msvc_seh(bx.sess()) { + unimplemented!(); + } + else { + unimplemented!(); + } +} diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs b/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs new file mode 100644 index 000000000..2401f3350 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs @@ -0,0 +1,751 @@ +use std::cmp::Ordering; + +use gccjit::{BinaryOp, RValue, Type, ToRValue}; +use rustc_codegen_ssa::base::compare_simd_types; +use rustc_codegen_ssa::common::{TypeKind, span_invalid_monomorphization_error}; +use rustc_codegen_ssa::mir::operand::OperandRef; +use rustc_codegen_ssa::mir::place::PlaceRef; +use rustc_codegen_ssa::traits::{BaseTypeMethods, BuilderMethods}; +use rustc_hir as hir; +use rustc_middle::span_bug; +use rustc_middle::ty::layout::HasTyCtxt; +use rustc_middle::ty::{self, Ty}; +use rustc_span::{Span, Symbol, sym}; +use rustc_target::abi::Align; + +use crate::builder::Builder; +use crate::intrinsic; + +pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>, name: Symbol, callee_ty: Ty<'tcx>, args: &[OperandRef<'tcx, RValue<'gcc>>], ret_ty: Ty<'tcx>, llret_ty: Type<'gcc>, span: Span) -> Result<RValue<'gcc>, ()> { + // macros for error handling: + #[allow(unused_macro_rules)] + macro_rules! emit_error { + ($msg: tt) => { + emit_error!($msg, ) + }; + ($msg: tt, $($fmt: tt)*) => { + span_invalid_monomorphization_error( + bx.sess(), span, + &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg), + name, $($fmt)*)); + } + } + + macro_rules! return_error { + ($($fmt: tt)*) => { + { + emit_error!($($fmt)*); + return Err(()); + } + } + } + + macro_rules! require { + ($cond: expr, $($fmt: tt)*) => { + if !$cond { + return_error!($($fmt)*); + } + }; + } + + macro_rules! require_simd { + ($ty: expr, $position: expr) => { + require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty) + }; + } + + let tcx = bx.tcx(); + let sig = + tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx)); + let arg_tys = sig.inputs(); + + if name == sym::simd_select_bitmask { + require_simd!(arg_tys[1], "argument"); + let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx()); + + let expected_int_bits = (len.max(8) - 1).next_power_of_two(); + let expected_bytes = len / 8 + ((len % 8 > 0) as u64); + + let mask_ty = arg_tys[0]; + let mut mask = match mask_ty.kind() { + ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(), + ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(), + ty::Array(elem, len) + if matches!(elem.kind(), ty::Uint(ty::UintTy::U8)) + && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all()) + == Some(expected_bytes) => + { + let place = PlaceRef::alloca(bx, args[0].layout); + args[0].val.store(bx, place); + let int_ty = bx.type_ix(expected_bytes * 8); + let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty)); + bx.load(int_ty, ptr, Align::ONE) + } + _ => return_error!( + "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`", + mask_ty, + expected_int_bits, + expected_bytes + ), + }; + + let arg1 = args[1].immediate(); + let arg1_type = arg1.get_type(); + let arg1_vector_type = arg1_type.unqualified().dyncast_vector().expect("vector type"); + let arg1_element_type = arg1_vector_type.get_element_type(); + + let mut elements = vec![]; + let one = bx.context.new_rvalue_one(mask.get_type()); + for _ in 0..len { + let element = bx.context.new_cast(None, mask & one, arg1_element_type); + elements.push(element); + mask = mask >> one; + } + let vector_mask = bx.context.new_rvalue_from_vector(None, arg1_type, &elements); + + return Ok(bx.vector_select(vector_mask, arg1, args[2].immediate())); + } + + // every intrinsic below takes a SIMD vector as its first argument + require_simd!(arg_tys[0], "input"); + let in_ty = arg_tys[0]; + + let comparison = match name { + sym::simd_eq => Some(hir::BinOpKind::Eq), + sym::simd_ne => Some(hir::BinOpKind::Ne), + sym::simd_lt => Some(hir::BinOpKind::Lt), + sym::simd_le => Some(hir::BinOpKind::Le), + sym::simd_gt => Some(hir::BinOpKind::Gt), + sym::simd_ge => Some(hir::BinOpKind::Ge), + _ => None, + }; + + let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx()); + if let Some(cmp_op) = comparison { + require_simd!(ret_ty, "return"); + + let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx()); + require!( + in_len == out_len, + "expected return type with length {} (same as input type `{}`), \ + found `{}` with length {}", + in_len, + in_ty, + ret_ty, + out_len + ); + require!( + bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer, + "expected return type with integer elements, found `{}` with non-integer `{}`", + ret_ty, + out_ty + ); + + return Ok(compare_simd_types( + bx, + args[0].immediate(), + args[1].immediate(), + in_elem, + llret_ty, + cmp_op, + )); + } + + if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") { + let n: u64 = + if stripped.is_empty() { + // Make sure this is actually an array, since typeck only checks the length-suffixed + // version of this intrinsic. + match args[2].layout.ty.kind() { + ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => { + len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| { + span_bug!(span, "could not evaluate shuffle index array length") + }) + } + _ => return_error!( + "simd_shuffle index must be an array of `u32`, got `{}`", + args[2].layout.ty + ), + } + } + else { + stripped.parse().unwrap_or_else(|_| { + span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?") + }) + }; + + require_simd!(ret_ty, "return"); + + let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx()); + require!( + out_len == n, + "expected return type of length {}, found `{}` with length {}", + n, + ret_ty, + out_len + ); + require!( + in_elem == out_ty, + "expected return element type `{}` (element of input `{}`), \ + found `{}` with element type `{}`", + in_elem, + in_ty, + ret_ty, + out_ty + ); + + let vector = args[2].immediate(); + + return Ok(bx.shuffle_vector( + args[0].immediate(), + args[1].immediate(), + vector, + )); + } + + #[cfg(feature="master")] + if name == sym::simd_insert { + require!( + in_elem == arg_tys[2], + "expected inserted type `{}` (element of input `{}`), found `{}`", + in_elem, + in_ty, + arg_tys[2] + ); + let vector = args[0].immediate(); + let index = args[1].immediate(); + let value = args[2].immediate(); + // TODO(antoyo): use a recursive unqualified() here. + let vector_type = vector.get_type().unqualified().dyncast_vector().expect("vector type"); + let element_type = vector_type.get_element_type(); + // NOTE: we cannot cast to an array and assign to its element here because the value might + // not be an l-value. So, call a builtin to set the element. + // TODO(antoyo): perhaps we could create a new vector or maybe there's a GIMPLE instruction for that? + // TODO(antoyo): don't use target specific builtins here. + let func_name = + match in_len { + 2 => { + if element_type == bx.i64_type { + "__builtin_ia32_vec_set_v2di" + } + else { + unimplemented!(); + } + }, + 4 => { + if element_type == bx.i32_type { + "__builtin_ia32_vec_set_v4si" + } + else { + unimplemented!(); + } + }, + 8 => { + if element_type == bx.i16_type { + "__builtin_ia32_vec_set_v8hi" + } + else { + unimplemented!(); + } + }, + _ => unimplemented!("Len: {}", in_len), + }; + let builtin = bx.context.get_target_builtin_function(func_name); + let param1_type = builtin.get_param(0).to_rvalue().get_type(); + // TODO(antoyo): perhaps use __builtin_convertvector for vector casting. + let vector = bx.cx.bitcast_if_needed(vector, param1_type); + let result = bx.context.new_call(None, builtin, &[vector, value, bx.context.new_cast(None, index, bx.int_type)]); + // TODO(antoyo): perhaps use __builtin_convertvector for vector casting. + return Ok(bx.context.new_bitcast(None, result, vector.get_type())); + } + + #[cfg(feature="master")] + if name == sym::simd_extract { + require!( + ret_ty == in_elem, + "expected return type `{}` (element of input `{}`), found `{}`", + in_elem, + in_ty, + ret_ty + ); + let vector = args[0].immediate(); + return Ok(bx.context.new_vector_access(None, vector, args[1].immediate()).to_rvalue()); + } + + if name == sym::simd_select { + let m_elem_ty = in_elem; + let m_len = in_len; + require_simd!(arg_tys[1], "argument"); + let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx()); + require!( + m_len == v_len, + "mismatched lengths: mask length `{}` != other vector length `{}`", + m_len, + v_len + ); + match m_elem_ty.kind() { + ty::Int(_) => {} + _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty), + } + return Ok(bx.vector_select(args[0].immediate(), args[1].immediate(), args[2].immediate())); + } + + if name == sym::simd_cast { + require_simd!(ret_ty, "return"); + let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx()); + require!( + in_len == out_len, + "expected return type with length {} (same as input type `{}`), \ + found `{}` with length {}", + in_len, + in_ty, + ret_ty, + out_len + ); + // casting cares about nominal type, not just structural type + if in_elem == out_elem { + return Ok(args[0].immediate()); + } + + enum Style { + Float, + Int(/* is signed? */ bool), + Unsupported, + } + + let (in_style, in_width) = match in_elem.kind() { + // vectors of pointer-sized integers should've been + // disallowed before here, so this unwrap is safe. + ty::Int(i) => ( + Style::Int(true), + i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(), + ), + ty::Uint(u) => ( + Style::Int(false), + u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(), + ), + ty::Float(f) => (Style::Float, f.bit_width()), + _ => (Style::Unsupported, 0), + }; + let (out_style, out_width) = match out_elem.kind() { + ty::Int(i) => ( + Style::Int(true), + i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(), + ), + ty::Uint(u) => ( + Style::Int(false), + u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(), + ), + ty::Float(f) => (Style::Float, f.bit_width()), + _ => (Style::Unsupported, 0), + }; + + let extend = |in_type, out_type| { + let vector_type = bx.context.new_vector_type(out_type, 8); + let vector = args[0].immediate(); + let array_type = bx.context.new_array_type(None, in_type, 8); + // TODO(antoyo): switch to using new_vector_access or __builtin_convertvector for vector casting. + let array = bx.context.new_bitcast(None, vector, array_type); + + let cast_vec_element = |index| { + let index = bx.context.new_rvalue_from_int(bx.int_type, index); + bx.context.new_cast(None, bx.context.new_array_access(None, array, index).to_rvalue(), out_type) + }; + + bx.context.new_rvalue_from_vector(None, vector_type, &[ + cast_vec_element(0), + cast_vec_element(1), + cast_vec_element(2), + cast_vec_element(3), + cast_vec_element(4), + cast_vec_element(5), + cast_vec_element(6), + cast_vec_element(7), + ]) + }; + + match (in_style, out_style) { + (Style::Int(in_is_signed), Style::Int(_)) => { + return Ok(match in_width.cmp(&out_width) { + Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty), + Ordering::Equal => args[0].immediate(), + Ordering::Less => { + if in_is_signed { + match (in_width, out_width) { + // FIXME(antoyo): the function _mm_cvtepi8_epi16 should directly + // call an intrinsic equivalent to __builtin_ia32_pmovsxbw128 so that + // we can generate a call to it. + (8, 16) => extend(bx.i8_type, bx.i16_type), + (8, 32) => extend(bx.i8_type, bx.i32_type), + (8, 64) => extend(bx.i8_type, bx.i64_type), + (16, 32) => extend(bx.i16_type, bx.i32_type), + (32, 64) => extend(bx.i32_type, bx.i64_type), + (16, 64) => extend(bx.i16_type, bx.i64_type), + _ => unimplemented!("in: {}, out: {}", in_width, out_width), + } + } else { + match (in_width, out_width) { + (8, 16) => extend(bx.u8_type, bx.u16_type), + (8, 32) => extend(bx.u8_type, bx.u32_type), + (8, 64) => extend(bx.u8_type, bx.u64_type), + (16, 32) => extend(bx.u16_type, bx.u32_type), + (16, 64) => extend(bx.u16_type, bx.u64_type), + (32, 64) => extend(bx.u32_type, bx.u64_type), + _ => unimplemented!("in: {}, out: {}", in_width, out_width), + } + } + } + }); + } + (Style::Int(_), Style::Float) => { + // TODO: add support for internal functions in libgccjit to get access to IFN_VEC_CONVERT which is + // doing like __builtin_convertvector? + // Or maybe provide convert_vector as an API since it might not easy to get the + // types of internal functions. + unimplemented!(); + } + (Style::Float, Style::Int(_)) => { + unimplemented!(); + } + (Style::Float, Style::Float) => { + unimplemented!(); + } + _ => { /* Unsupported. Fallthrough. */ } + } + require!( + false, + "unsupported cast from `{}` with element `{}` to `{}` with element `{}`", + in_ty, + in_elem, + ret_ty, + out_elem + ); + } + + macro_rules! arith_binary { + ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => { + $(if name == sym::$name { + match in_elem.kind() { + $($(ty::$p(_))|* => { + return Ok(bx.$call(args[0].immediate(), args[1].immediate())) + })* + _ => {}, + } + require!(false, + "unsupported operation on `{}` with element `{}`", + in_ty, + in_elem) + })* + } + } + + fn simd_simple_float_intrinsic<'gcc, 'tcx>( + name: Symbol, + in_elem: Ty<'_>, + in_ty: Ty<'_>, + in_len: u64, + bx: &mut Builder<'_, 'gcc, 'tcx>, + span: Span, + args: &[OperandRef<'tcx, RValue<'gcc>>], + ) -> Result<RValue<'gcc>, ()> { + macro_rules! emit_error { + ($msg: tt, $($fmt: tt)*) => { + span_invalid_monomorphization_error( + bx.sess(), span, + &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg), + name, $($fmt)*)); + } + } + macro_rules! return_error { + ($($fmt: tt)*) => { + { + emit_error!($($fmt)*); + return Err(()); + } + } + } + + let (elem_ty_str, elem_ty) = + if let ty::Float(f) = in_elem.kind() { + let elem_ty = bx.cx.type_float_from_ty(*f); + match f.bit_width() { + 32 => ("f32", elem_ty), + 64 => ("f64", elem_ty), + _ => { + return_error!( + "unsupported element type `{}` of floating-point vector `{}`", + f.name_str(), + in_ty + ); + } + } + } + else { + return_error!("`{}` is not a floating-point type", in_ty); + }; + + let vec_ty = bx.cx.type_vector(elem_ty, in_len); + + let (intr_name, fn_ty) = + match name { + sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)), // TODO(antoyo): pand with 170141183420855150465331762880109871103 + sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)), + sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)), + sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)), + sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)), + sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)), + _ => return_error!("unrecognized intrinsic `{}`", name), + }; + let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str); + let function = intrinsic::llvm::intrinsic(llvm_name, &bx.cx); + let function: RValue<'gcc> = unsafe { std::mem::transmute(function) }; + let c = bx.call(fn_ty, function, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None); + Ok(c) + } + + if std::matches!( + name, + sym::simd_ceil + | sym::simd_fabs + | sym::simd_fcos + | sym::simd_fexp2 + | sym::simd_fexp + | sym::simd_flog10 + | sym::simd_flog2 + | sym::simd_flog + | sym::simd_floor + | sym::simd_fma + | sym::simd_fpow + | sym::simd_fpowi + | sym::simd_fsin + | sym::simd_fsqrt + | sym::simd_round + | sym::simd_trunc + ) { + return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args); + } + + arith_binary! { + simd_add: Uint, Int => add, Float => fadd; + simd_sub: Uint, Int => sub, Float => fsub; + simd_mul: Uint, Int => mul, Float => fmul; + simd_div: Uint => udiv, Int => sdiv, Float => fdiv; + simd_rem: Uint => urem, Int => srem, Float => frem; + simd_shl: Uint, Int => shl; + simd_shr: Uint => lshr, Int => ashr; + simd_and: Uint, Int => and; + simd_or: Uint, Int => or; // FIXME(antoyo): calling `or` might not work on vectors. + simd_xor: Uint, Int => xor; + } + + macro_rules! arith_unary { + ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => { + $(if name == sym::$name { + match in_elem.kind() { + $($(ty::$p(_))|* => { + return Ok(bx.$call(args[0].immediate())) + })* + _ => {}, + } + require!(false, + "unsupported operation on `{}` with element `{}`", + in_ty, + in_elem) + })* + } + } + + arith_unary! { + simd_neg: Int => neg, Float => fneg; + } + + #[cfg(feature="master")] + if name == sym::simd_saturating_add || name == sym::simd_saturating_sub { + let lhs = args[0].immediate(); + let rhs = args[1].immediate(); + let is_add = name == sym::simd_saturating_add; + let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _; + let (signed, elem_width, elem_ty) = match *in_elem.kind() { + ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)), + ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)), + _ => { + return_error!( + "expected element type `{}` of vector type `{}` \ + to be a signed or unsigned integer type", + arg_tys[0].simd_size_and_type(bx.tcx()).1, + arg_tys[0] + ); + } + }; + let builtin_name = + match (signed, is_add, in_len, elem_width) { + (true, true, 32, 8) => "__builtin_ia32_paddsb256", // TODO(antoyo): cast arguments to unsigned. + (false, true, 32, 8) => "__builtin_ia32_paddusb256", + (true, true, 16, 16) => "__builtin_ia32_paddsw256", + (false, true, 16, 16) => "__builtin_ia32_paddusw256", + (true, false, 16, 16) => "__builtin_ia32_psubsw256", + (false, false, 16, 16) => "__builtin_ia32_psubusw256", + (true, false, 32, 8) => "__builtin_ia32_psubsb256", + (false, false, 32, 8) => "__builtin_ia32_psubusb256", + _ => unimplemented!("signed: {}, is_add: {}, in_len: {}, elem_width: {}", signed, is_add, in_len, elem_width), + }; + let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64); + + let func = bx.context.get_target_builtin_function(builtin_name); + let param1_type = func.get_param(0).to_rvalue().get_type(); + let param2_type = func.get_param(1).to_rvalue().get_type(); + let lhs = bx.cx.bitcast_if_needed(lhs, param1_type); + let rhs = bx.cx.bitcast_if_needed(rhs, param2_type); + let result = bx.context.new_call(None, func, &[lhs, rhs]); + // TODO(antoyo): perhaps use __builtin_convertvector for vector casting. + return Ok(bx.context.new_bitcast(None, result, vec_ty)); + } + + macro_rules! arith_red { + ($name:ident : $vec_op:expr, $float_reduce:ident, $ordered:expr, $op:ident, + $identity:expr) => { + if name == sym::$name { + require!( + ret_ty == in_elem, + "expected return type `{}` (element of input `{}`), found `{}`", + in_elem, + in_ty, + ret_ty + ); + return match in_elem.kind() { + ty::Int(_) | ty::Uint(_) => { + let r = bx.vector_reduce_op(args[0].immediate(), $vec_op); + if $ordered { + // if overflow occurs, the result is the + // mathematical result modulo 2^n: + Ok(bx.$op(args[1].immediate(), r)) + } + else { + Ok(bx.vector_reduce_op(args[0].immediate(), $vec_op)) + } + } + ty::Float(_) => { + if $ordered { + // ordered arithmetic reductions take an accumulator + let acc = args[1].immediate(); + Ok(bx.$float_reduce(acc, args[0].immediate())) + } + else { + Ok(bx.vector_reduce_op(args[0].immediate(), $vec_op)) + } + } + _ => return_error!( + "unsupported {} from `{}` with element `{}` to `{}`", + sym::$name, + in_ty, + in_elem, + ret_ty + ), + }; + } + }; + } + + arith_red!( + simd_reduce_add_unordered: BinaryOp::Plus, + vector_reduce_fadd_fast, + false, + add, + 0.0 // TODO: Use this argument. + ); + arith_red!( + simd_reduce_mul_unordered: BinaryOp::Mult, + vector_reduce_fmul_fast, + false, + mul, + 1.0 + ); + + macro_rules! minmax_red { + ($name:ident: $reduction:ident) => { + if name == sym::$name { + require!( + ret_ty == in_elem, + "expected return type `{}` (element of input `{}`), found `{}`", + in_elem, + in_ty, + ret_ty + ); + return match in_elem.kind() { + ty::Int(_) | ty::Uint(_) | ty::Float(_) => Ok(bx.$reduction(args[0].immediate())), + _ => return_error!( + "unsupported {} from `{}` with element `{}` to `{}`", + sym::$name, + in_ty, + in_elem, + ret_ty + ), + }; + } + }; + } + + minmax_red!(simd_reduce_min: vector_reduce_min); + minmax_red!(simd_reduce_max: vector_reduce_max); + + macro_rules! bitwise_red { + ($name:ident : $op:expr, $boolean:expr) => { + if name == sym::$name { + let input = if !$boolean { + require!( + ret_ty == in_elem, + "expected return type `{}` (element of input `{}`), found `{}`", + in_elem, + in_ty, + ret_ty + ); + args[0].immediate() + } else { + match in_elem.kind() { + ty::Int(_) | ty::Uint(_) => {} + _ => return_error!( + "unsupported {} from `{}` with element `{}` to `{}`", + sym::$name, + in_ty, + in_elem, + ret_ty + ), + } + + // boolean reductions operate on vectors of i1s: + let i1 = bx.type_i1(); + let i1xn = bx.type_vector(i1, in_len as u64); + bx.trunc(args[0].immediate(), i1xn) + }; + return match in_elem.kind() { + ty::Int(_) | ty::Uint(_) => { + let r = bx.vector_reduce_op(input, $op); + Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) }) + } + _ => return_error!( + "unsupported {} from `{}` with element `{}` to `{}`", + sym::$name, + in_ty, + in_elem, + ret_ty + ), + }; + } + }; + } + + bitwise_red!(simd_reduce_and: BinaryOp::BitwiseAnd, false); + bitwise_red!(simd_reduce_or: BinaryOp::BitwiseOr, false); + + unimplemented!("simd {}", name); +} diff --git a/compiler/rustc_codegen_gcc/src/lib.rs b/compiler/rustc_codegen_gcc/src/lib.rs new file mode 100644 index 000000000..8a206c036 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/lib.rs @@ -0,0 +1,331 @@ +/* + * TODO(antoyo): implement equality in libgccjit based on https://zpz.github.io/blog/overloading-equality-operator-in-cpp-class-hierarchy/ (for type equality?) + * TODO(antoyo): support #[inline] attributes. + * TODO(antoyo): support LTO (gcc's equivalent to Thin LTO is enabled by -fwhopr: https://stackoverflow.com/questions/64954525/does-gcc-have-thin-lto). + * + * TODO(antoyo): remove the patches. + */ + +#![feature( + rustc_private, + decl_macro, + associated_type_bounds, + never_type, + trusted_len, + hash_raw_entry +)] +#![allow(broken_intra_doc_links)] +#![recursion_limit="256"] +#![warn(rust_2018_idioms)] +#![warn(unused_lifetimes)] + +extern crate rustc_ast; +extern crate rustc_codegen_ssa; +extern crate rustc_data_structures; +extern crate rustc_errors; +extern crate rustc_hir; +extern crate rustc_metadata; +extern crate rustc_middle; +extern crate rustc_session; +extern crate rustc_span; +extern crate rustc_target; +extern crate tempfile; + +// This prevents duplicating functions and statics that are already part of the host rustc process. +#[allow(unused_extern_crates)] +extern crate rustc_driver; + +mod abi; +mod allocator; +mod archive; +mod asm; +mod back; +mod base; +mod builder; +mod callee; +mod common; +mod consts; +mod context; +mod coverageinfo; +mod debuginfo; +mod declare; +mod int; +mod intrinsic; +mod mono_item; +mod type_; +mod type_of; + +use std::any::Any; +use std::sync::{Arc, Mutex}; + +use gccjit::{Context, OptimizationLevel, CType}; +use rustc_ast::expand::allocator::AllocatorKind; +use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen}; +use rustc_codegen_ssa::base::codegen_crate; +use rustc_codegen_ssa::back::write::{CodegenContext, FatLTOInput, ModuleConfig, TargetMachineFactoryFn}; +use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule}; +use rustc_codegen_ssa::target_features::supported_target_features; +use rustc_codegen_ssa::traits::{CodegenBackend, ExtraBackendMethods, ModuleBufferMethods, ThinBufferMethods, WriteBackendMethods}; +use rustc_data_structures::fx::FxHashMap; +use rustc_errors::{ErrorGuaranteed, Handler}; +use rustc_metadata::EncodedMetadata; +use rustc_middle::dep_graph::{WorkProduct, WorkProductId}; +use rustc_middle::ty::TyCtxt; +use rustc_middle::ty::query::Providers; +use rustc_session::config::{Lto, OptLevel, OutputFilenames}; +use rustc_session::Session; +use rustc_span::Symbol; +use rustc_span::fatal_error::FatalError; +use tempfile::TempDir; + +pub struct PrintOnPanic<F: Fn() -> String>(pub F); + +impl<F: Fn() -> String> Drop for PrintOnPanic<F> { + fn drop(&mut self) { + if ::std::thread::panicking() { + println!("{}", (self.0)()); + } + } +} + +#[derive(Clone)] +pub struct GccCodegenBackend { + supports_128bit_integers: Arc<Mutex<bool>>, +} + +impl CodegenBackend for GccCodegenBackend { + fn init(&self, sess: &Session) { + if sess.lto() != Lto::No { + sess.warn("LTO is not supported. You may get a linker error."); + } + + let temp_dir = TempDir::new().expect("cannot create temporary directory"); + let temp_file = temp_dir.into_path().join("result.asm"); + let check_context = Context::default(); + check_context.set_print_errors_to_stderr(false); + let _int128_ty = check_context.new_c_type(CType::UInt128t); + // NOTE: we cannot just call compile() as this would require other files than libgccjit.so. + check_context.compile_to_file(gccjit::OutputKind::Assembler, temp_file.to_str().expect("path to str")); + *self.supports_128bit_integers.lock().expect("lock") = check_context.get_last_error() == Ok(None); + } + + fn provide(&self, providers: &mut Providers) { + // FIXME(antoyo) compute list of enabled features from cli flags + providers.global_backend_features = |_tcx, ()| vec![]; + } + + fn codegen_crate<'tcx>(&self, tcx: TyCtxt<'tcx>, metadata: EncodedMetadata, need_metadata_module: bool) -> Box<dyn Any> { + let target_cpu = target_cpu(tcx.sess); + let res = codegen_crate(self.clone(), tcx, target_cpu.to_string(), metadata, need_metadata_module); + + Box::new(res) + } + + fn join_codegen(&self, ongoing_codegen: Box<dyn Any>, sess: &Session, _outputs: &OutputFilenames) -> Result<(CodegenResults, FxHashMap<WorkProductId, WorkProduct>), ErrorGuaranteed> { + let (codegen_results, work_products) = ongoing_codegen + .downcast::<rustc_codegen_ssa::back::write::OngoingCodegen<GccCodegenBackend>>() + .expect("Expected GccCodegenBackend's OngoingCodegen, found Box<Any>") + .join(sess); + + Ok((codegen_results, work_products)) + } + + fn link(&self, sess: &Session, codegen_results: CodegenResults, outputs: &OutputFilenames) -> Result<(), ErrorGuaranteed> { + use rustc_codegen_ssa::back::link::link_binary; + + link_binary( + sess, + &crate::archive::ArArchiveBuilderBuilder, + &codegen_results, + outputs, + ) + } + + fn target_features(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> { + target_features(sess, allow_unstable) + } +} + +impl ExtraBackendMethods for GccCodegenBackend { + fn codegen_allocator<'tcx>(&self, tcx: TyCtxt<'tcx>, module_name: &str, kind: AllocatorKind, has_alloc_error_handler: bool) -> Self::Module { + let mut mods = GccContext { + context: Context::default(), + }; + unsafe { allocator::codegen(tcx, &mut mods, module_name, kind, has_alloc_error_handler); } + mods + } + + fn compile_codegen_unit<'tcx>(&self, tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (ModuleCodegen<Self::Module>, u64) { + base::compile_codegen_unit(tcx, cgu_name, *self.supports_128bit_integers.lock().expect("lock")) + } + + fn target_machine_factory(&self, _sess: &Session, _opt_level: OptLevel, _features: &[String]) -> TargetMachineFactoryFn<Self> { + // TODO(antoyo): set opt level. + Arc::new(|_| { + Ok(()) + }) + } + + fn target_cpu<'b>(&self, _sess: &'b Session) -> &'b str { + unimplemented!(); + } + + fn tune_cpu<'b>(&self, _sess: &'b Session) -> Option<&'b str> { + None + // TODO(antoyo) + } +} + +pub struct ModuleBuffer; + +impl ModuleBufferMethods for ModuleBuffer { + fn data(&self) -> &[u8] { + unimplemented!(); + } +} + +pub struct ThinBuffer; + +impl ThinBufferMethods for ThinBuffer { + fn data(&self) -> &[u8] { + unimplemented!(); + } +} + +pub struct GccContext { + context: Context<'static>, +} + +unsafe impl Send for GccContext {} +// FIXME(antoyo): that shouldn't be Sync. Parallel compilation is currently disabled with "-Zno-parallel-llvm". Try to disable it here. +unsafe impl Sync for GccContext {} + +impl WriteBackendMethods for GccCodegenBackend { + type Module = GccContext; + type TargetMachine = (); + type ModuleBuffer = ModuleBuffer; + type Context = (); + type ThinData = (); + type ThinBuffer = ThinBuffer; + + fn run_fat_lto(_cgcx: &CodegenContext<Self>, mut modules: Vec<FatLTOInput<Self>>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> { + // TODO(antoyo): implement LTO by sending -flto to libgccjit and adding the appropriate gcc linker plugins. + // NOTE: implemented elsewhere. + // TODO(antoyo): what is implemented elsewhere ^ ? + let module = + match modules.remove(0) { + FatLTOInput::InMemory(module) => module, + FatLTOInput::Serialized { .. } => { + unimplemented!(); + } + }; + Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: vec![] }) + } + + fn run_thin_lto(_cgcx: &CodegenContext<Self>, _modules: Vec<(String, Self::ThinBuffer)>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<(Vec<LtoModuleCodegen<Self>>, Vec<WorkProduct>), FatalError> { + unimplemented!(); + } + + fn print_pass_timings(&self) { + unimplemented!(); + } + + unsafe fn optimize(_cgcx: &CodegenContext<Self>, _diag_handler: &Handler, module: &ModuleCodegen<Self::Module>, config: &ModuleConfig) -> Result<(), FatalError> { + module.module_llvm.context.set_optimization_level(to_gcc_opt_level(config.opt_level)); + Ok(()) + } + + fn optimize_fat(_cgcx: &CodegenContext<Self>, _module: &mut ModuleCodegen<Self::Module>) -> Result<(), FatalError> { + // TODO(antoyo) + Ok(()) + } + + unsafe fn optimize_thin(_cgcx: &CodegenContext<Self>, _thin: ThinModule<Self>) -> Result<ModuleCodegen<Self::Module>, FatalError> { + unimplemented!(); + } + + unsafe fn codegen(cgcx: &CodegenContext<Self>, diag_handler: &Handler, module: ModuleCodegen<Self::Module>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> { + back::write::codegen(cgcx, diag_handler, module, config) + } + + fn prepare_thin(_module: ModuleCodegen<Self::Module>) -> (String, Self::ThinBuffer) { + unimplemented!(); + } + + fn serialize_module(_module: ModuleCodegen<Self::Module>) -> (String, Self::ModuleBuffer) { + unimplemented!(); + } + + fn run_link(cgcx: &CodegenContext<Self>, diag_handler: &Handler, modules: Vec<ModuleCodegen<Self::Module>>) -> Result<ModuleCodegen<Self::Module>, FatalError> { + back::write::link(cgcx, diag_handler, modules) + } +} + +/// This is the entrypoint for a hot plugged rustc_codegen_gccjit +#[no_mangle] +pub fn __rustc_codegen_backend() -> Box<dyn CodegenBackend> { + Box::new(GccCodegenBackend { + supports_128bit_integers: Arc::new(Mutex::new(false)), + }) +} + +fn to_gcc_opt_level(optlevel: Option<OptLevel>) -> OptimizationLevel { + match optlevel { + None => OptimizationLevel::None, + Some(level) => { + match level { + OptLevel::No => OptimizationLevel::None, + OptLevel::Less => OptimizationLevel::Limited, + OptLevel::Default => OptimizationLevel::Standard, + OptLevel::Aggressive => OptimizationLevel::Aggressive, + OptLevel::Size | OptLevel::SizeMin => OptimizationLevel::Limited, + } + }, + } +} + +fn handle_native(name: &str) -> &str { + if name != "native" { + return name; + } + + unimplemented!(); +} + +pub fn target_cpu(sess: &Session) -> &str { + match sess.opts.cg.target_cpu { + Some(ref name) => handle_native(name), + None => handle_native(sess.target.cpu.as_ref()), + } +} + +pub fn target_features(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { + supported_target_features(sess) + .iter() + .filter_map( + |&(feature, gate)| { + if sess.is_nightly_build() || allow_unstable || gate.is_none() { Some(feature) } else { None } + }, + ) + .filter(|_feature| { + // TODO(antoyo): implement a way to get enabled feature in libgccjit. + // Probably using the equivalent of __builtin_cpu_supports. + #[cfg(feature="master")] + { + _feature.contains("sse") || _feature.contains("avx") + } + #[cfg(not(feature="master"))] + { + false + } + /* + adx, aes, avx, avx2, avx512bf16, avx512bitalg, avx512bw, avx512cd, avx512dq, avx512er, avx512f, avx512gfni, + avx512ifma, avx512pf, avx512vaes, avx512vbmi, avx512vbmi2, avx512vl, avx512vnni, avx512vp2intersect, avx512vpclmulqdq, + avx512vpopcntdq, bmi1, bmi2, cmpxchg16b, ermsb, f16c, fma, fxsr, lzcnt, movbe, pclmulqdq, popcnt, rdrand, rdseed, rtm, + sha, sse, sse2, sse3, sse4.1, sse4.2, sse4a, ssse3, tbm, xsave, xsavec, xsaveopt, xsaves + */ + //false + }) + .map(|feature| Symbol::intern(feature)) + .collect() +} diff --git a/compiler/rustc_codegen_gcc/src/mono_item.rs b/compiler/rustc_codegen_gcc/src/mono_item.rs new file mode 100644 index 000000000..9468a1ef4 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/mono_item.rs @@ -0,0 +1,38 @@ +use rustc_codegen_ssa::traits::PreDefineMethods; +use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrFlags; +use rustc_middle::mir::mono::{Linkage, Visibility}; +use rustc_middle::ty::{self, Instance, TypeVisitable}; +use rustc_middle::ty::layout::{FnAbiOf, LayoutOf}; +use rustc_span::def_id::DefId; + +use crate::base; +use crate::context::CodegenCx; +use crate::type_of::LayoutGccExt; + +impl<'gcc, 'tcx> PreDefineMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn predefine_static(&self, def_id: DefId, _linkage: Linkage, _visibility: Visibility, symbol_name: &str) { + let attrs = self.tcx.codegen_fn_attrs(def_id); + let instance = Instance::mono(self.tcx, def_id); + let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all()); + let gcc_type = self.layout_of(ty).gcc_type(self, true); + + let is_tls = attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL); + let global = self.define_global(symbol_name, gcc_type, is_tls, attrs.link_section); + + // TODO(antoyo): set linkage and visibility. + self.instances.borrow_mut().insert(instance, global); + } + + fn predefine_fn(&self, instance: Instance<'tcx>, linkage: Linkage, _visibility: Visibility, symbol_name: &str) { + assert!(!instance.substs.needs_infer()); + + let fn_abi = self.fn_abi_of_instance(instance, ty::List::empty()); + self.linkage.set(base::linkage_to_gcc(linkage)); + let _decl = self.declare_fn(symbol_name, &fn_abi); + //let attrs = self.tcx.codegen_fn_attrs(instance.def_id()); + + // TODO(antoyo): call set_link_section() to allow initializing argc/argv. + // TODO(antoyo): set unique comdat. + // TODO(antoyo): use inline attribute from there in linkage.set() above. + } +} diff --git a/compiler/rustc_codegen_gcc/src/type_.rs b/compiler/rustc_codegen_gcc/src/type_.rs new file mode 100644 index 000000000..68bdb8d4e --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/type_.rs @@ -0,0 +1,303 @@ +use std::convert::TryInto; + +use gccjit::{RValue, Struct, Type}; +use rustc_codegen_ssa::traits::{BaseTypeMethods, DerivedTypeMethods, TypeMembershipMethods}; +use rustc_codegen_ssa::common::TypeKind; +use rustc_middle::{bug, ty}; +use rustc_middle::ty::layout::TyAndLayout; +use rustc_target::abi::{AddressSpace, Align, Integer, Size}; + +use crate::common::TypeReflection; +use crate::context::CodegenCx; +use crate::type_of::LayoutGccExt; + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn type_ix(&self, num_bits: u64) -> Type<'gcc> { + // gcc only supports 1, 2, 4 or 8-byte integers. + // FIXME(antoyo): this is misleading to use the next power of two as rustc_codegen_ssa + // sometimes use 96-bit numbers and the following code will give an integer of a different + // size. + let bytes = (num_bits / 8).next_power_of_two() as i32; + match bytes { + 1 => self.i8_type, + 2 => self.i16_type, + 4 => self.i32_type, + 8 => self.i64_type, + 16 => self.i128_type, + _ => panic!("unexpected num_bits: {}", num_bits), + } + } + + pub fn type_void(&self) -> Type<'gcc> { + self.context.new_type::<()>() + } + + pub fn type_size_t(&self) -> Type<'gcc> { + self.context.new_type::<usize>() + } + + pub fn type_u8(&self) -> Type<'gcc> { + self.u8_type + } + + pub fn type_u16(&self) -> Type<'gcc> { + self.u16_type + } + + pub fn type_u32(&self) -> Type<'gcc> { + self.u32_type + } + + pub fn type_u64(&self) -> Type<'gcc> { + self.u64_type + } + + pub fn type_u128(&self) -> Type<'gcc> { + self.u128_type + } + + pub fn type_pointee_for_align(&self, align: Align) -> Type<'gcc> { + // FIXME(eddyb) We could find a better approximation if ity.align < align. + let ity = Integer::approximate_align(self, align); + self.type_from_integer(ity) + } + + pub fn type_vector(&self, ty: Type<'gcc>, len: u64) -> Type<'gcc> { + self.context.new_vector_type(ty, len) + } + + pub fn type_float_from_ty(&self, t: ty::FloatTy) -> Type<'gcc> { + match t { + ty::FloatTy::F32 => self.type_f32(), + ty::FloatTy::F64 => self.type_f64(), + } + } +} + +impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn type_i1(&self) -> Type<'gcc> { + self.bool_type + } + + fn type_i8(&self) -> Type<'gcc> { + self.i8_type + } + + fn type_i16(&self) -> Type<'gcc> { + self.i16_type + } + + fn type_i32(&self) -> Type<'gcc> { + self.i32_type + } + + fn type_i64(&self) -> Type<'gcc> { + self.i64_type + } + + fn type_i128(&self) -> Type<'gcc> { + self.i128_type + } + + fn type_isize(&self) -> Type<'gcc> { + self.isize_type + } + + fn type_f32(&self) -> Type<'gcc> { + self.context.new_type::<f32>() + } + + fn type_f64(&self) -> Type<'gcc> { + self.context.new_type::<f64>() + } + + fn type_func(&self, params: &[Type<'gcc>], return_type: Type<'gcc>) -> Type<'gcc> { + self.context.new_function_pointer_type(None, return_type, params, false) + } + + fn type_struct(&self, fields: &[Type<'gcc>], packed: bool) -> Type<'gcc> { + let types = fields.to_vec(); + if let Some(typ) = self.struct_types.borrow().get(fields) { + return typ.clone(); + } + let fields: Vec<_> = fields.iter().enumerate() + .map(|(index, field)| self.context.new_field(None, *field, &format!("field{}_TODO", index))) + .collect(); + let typ = self.context.new_struct_type(None, "struct", &fields).as_type(); + if packed { + #[cfg(feature="master")] + typ.set_packed(); + } + self.struct_types.borrow_mut().insert(types, typ); + typ + } + + fn type_kind(&self, typ: Type<'gcc>) -> TypeKind { + if self.is_int_type_or_bool(typ) { + TypeKind::Integer + } + else if typ.is_compatible_with(self.float_type) { + TypeKind::Float + } + else if typ.is_compatible_with(self.double_type) { + TypeKind::Double + } + else if typ.is_vector() { + TypeKind::Vector + } + else { + // TODO(antoyo): support other types. + TypeKind::Void + } + } + + fn type_ptr_to(&self, ty: Type<'gcc>) -> Type<'gcc> { + ty.make_pointer() + } + + fn type_ptr_to_ext(&self, ty: Type<'gcc>, _address_space: AddressSpace) -> Type<'gcc> { + // TODO(antoyo): use address_space, perhaps with TYPE_ADDR_SPACE? + ty.make_pointer() + } + + fn element_type(&self, ty: Type<'gcc>) -> Type<'gcc> { + if let Some(typ) = ty.dyncast_array() { + typ + } + else if let Some(vector_type) = ty.dyncast_vector() { + vector_type.get_element_type() + } + else if let Some(typ) = ty.get_pointee() { + typ + } + else { + unreachable!() + } + } + + fn vector_length(&self, _ty: Type<'gcc>) -> usize { + unimplemented!(); + } + + fn float_width(&self, typ: Type<'gcc>) -> usize { + let f32 = self.context.new_type::<f32>(); + let f64 = self.context.new_type::<f64>(); + if typ.is_compatible_with(f32) { + 32 + } + else if typ.is_compatible_with(f64) { + 64 + } + else { + panic!("Cannot get width of float type {:?}", typ); + } + // TODO(antoyo): support other sizes. + } + + fn int_width(&self, typ: Type<'gcc>) -> u64 { + self.gcc_int_width(typ) + } + + fn val_ty(&self, value: RValue<'gcc>) -> Type<'gcc> { + value.get_type() + } +} + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn type_padding_filler(&self, size: Size, align: Align) -> Type<'gcc> { + let unit = Integer::approximate_align(self, align); + let size = size.bytes(); + let unit_size = unit.size().bytes(); + assert_eq!(size % unit_size, 0); + self.type_array(self.type_from_integer(unit), size / unit_size) + } + + pub fn set_struct_body(&self, typ: Struct<'gcc>, fields: &[Type<'gcc>], packed: bool) { + let fields: Vec<_> = fields.iter().enumerate() + .map(|(index, field)| self.context.new_field(None, *field, &format!("field_{}", index))) + .collect(); + typ.set_fields(None, &fields); + if packed { + #[cfg(feature="master")] + typ.as_type().set_packed(); + } + } + + pub fn type_named_struct(&self, name: &str) -> Struct<'gcc> { + self.context.new_opaque_struct_type(None, name) + } + + pub fn type_array(&self, ty: Type<'gcc>, mut len: u64) -> Type<'gcc> { + if let Some(struct_type) = ty.is_struct() { + if struct_type.get_field_count() == 0 { + // NOTE: since gccjit only supports i32 for the array size and libcore's tests uses a + // size of usize::MAX in test_binary_search, we workaround this by setting the size to + // zero for ZSTs. + // FIXME(antoyo): fix gccjit API. + len = 0; + } + } + + // NOTE: see note above. Some other test uses usize::MAX. + if len == u64::MAX { + len = 0; + } + + let len: i32 = len.try_into().expect("array len"); + + self.context.new_array_type(None, ty, len) + } + + pub fn type_bool(&self) -> Type<'gcc> { + self.context.new_type::<bool>() + } +} + +pub fn struct_fields<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout<'tcx>) -> (Vec<Type<'gcc>>, bool) { + let field_count = layout.fields.count(); + + let mut packed = false; + let mut offset = Size::ZERO; + let mut prev_effective_align = layout.align.abi; + let mut result: Vec<_> = Vec::with_capacity(1 + field_count * 2); + for i in layout.fields.index_by_increasing_offset() { + let target_offset = layout.fields.offset(i as usize); + let field = layout.field(cx, i); + let effective_field_align = + layout.align.abi.min(field.align.abi).restrict_for_offset(target_offset); + packed |= effective_field_align < field.align.abi; + + assert!(target_offset >= offset); + let padding = target_offset - offset; + let padding_align = prev_effective_align.min(effective_field_align); + assert_eq!(offset.align_to(padding_align) + padding, target_offset); + result.push(cx.type_padding_filler(padding, padding_align)); + + result.push(field.gcc_type(cx, !field.ty.is_any_ptr())); // FIXME(antoyo): might need to check if the type is inside another, like Box<Type>. + offset = target_offset + field.size; + prev_effective_align = effective_field_align; + } + if !layout.is_unsized() && field_count > 0 { + if offset > layout.size { + bug!("layout: {:#?} stride: {:?} offset: {:?}", layout, layout.size, offset); + } + let padding = layout.size - offset; + let padding_align = prev_effective_align; + assert_eq!(offset.align_to(padding_align) + padding, layout.size); + result.push(cx.type_padding_filler(padding, padding_align)); + assert_eq!(result.len(), 1 + field_count * 2); + } + + (result, packed) +} + +impl<'gcc, 'tcx> TypeMembershipMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn set_type_metadata(&self, _function: RValue<'gcc>, _typeid: String) { + // Unsupported. + } + + fn typeid_metadata(&self, _typeid: String) -> RValue<'gcc> { + // Unsupported. + self.context.new_rvalue_from_int(self.int_type, 0) + } +} diff --git a/compiler/rustc_codegen_gcc/src/type_of.rs b/compiler/rustc_codegen_gcc/src/type_of.rs new file mode 100644 index 000000000..524d10fb5 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/type_of.rs @@ -0,0 +1,385 @@ +use std::fmt::Write; + +use gccjit::{Struct, Type}; +use crate::rustc_codegen_ssa::traits::{BaseTypeMethods, DerivedTypeMethods, LayoutTypeMethods}; +use rustc_middle::bug; +use rustc_middle::ty::{self, Ty, TypeVisitable}; +use rustc_middle::ty::layout::{FnAbiOf, LayoutOf, TyAndLayout}; +use rustc_middle::ty::print::with_no_trimmed_paths; +use rustc_target::abi::{self, Abi, F32, F64, FieldsShape, Int, Integer, Pointer, PointeeInfo, Size, TyAbiInterface, Variants}; +use rustc_target::abi::call::{CastTarget, FnAbi, Reg}; + +use crate::abi::{FnAbiGccExt, GccType}; +use crate::context::CodegenCx; +use crate::type_::struct_fields; + +impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + fn type_from_unsigned_integer(&self, i: Integer) -> Type<'gcc> { + use Integer::*; + match i { + I8 => self.type_u8(), + I16 => self.type_u16(), + I32 => self.type_u32(), + I64 => self.type_u64(), + I128 => self.type_u128(), + } + } + + #[cfg(feature="master")] + pub fn type_int_from_ty(&self, t: ty::IntTy) -> Type<'gcc> { + match t { + ty::IntTy::Isize => self.type_isize(), + ty::IntTy::I8 => self.type_i8(), + ty::IntTy::I16 => self.type_i16(), + ty::IntTy::I32 => self.type_i32(), + ty::IntTy::I64 => self.type_i64(), + ty::IntTy::I128 => self.type_i128(), + } + } + + #[cfg(feature="master")] + pub fn type_uint_from_ty(&self, t: ty::UintTy) -> Type<'gcc> { + match t { + ty::UintTy::Usize => self.type_isize(), + ty::UintTy::U8 => self.type_i8(), + ty::UintTy::U16 => self.type_i16(), + ty::UintTy::U32 => self.type_i32(), + ty::UintTy::U64 => self.type_i64(), + ty::UintTy::U128 => self.type_i128(), + } + } +} + +pub fn uncached_gcc_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout<'tcx>, defer: &mut Option<(Struct<'gcc>, TyAndLayout<'tcx>)>) -> Type<'gcc> { + match layout.abi { + Abi::Scalar(_) => bug!("handled elsewhere"), + Abi::Vector { ref element, count } => { + let element = layout.scalar_gcc_type_at(cx, element, Size::ZERO); + return cx.context.new_vector_type(element, count); + }, + Abi::ScalarPair(..) => { + return cx.type_struct( + &[ + layout.scalar_pair_element_gcc_type(cx, 0, false), + layout.scalar_pair_element_gcc_type(cx, 1, false), + ], + false, + ); + } + Abi::Uninhabited | Abi::Aggregate { .. } => {} + } + + let name = match layout.ty.kind() { + // FIXME(eddyb) producing readable type names for trait objects can result + // in problematically distinct types due to HRTB and subtyping (see #47638). + // ty::Dynamic(..) | + ty::Adt(..) | ty::Closure(..) | ty::Foreign(..) | ty::Generator(..) | ty::Str + if !cx.sess().fewer_names() => + { + let mut name = with_no_trimmed_paths!(layout.ty.to_string()); + if let (&ty::Adt(def, _), &Variants::Single { index }) = + (layout.ty.kind(), &layout.variants) + { + if def.is_enum() && !def.variants().is_empty() { + write!(&mut name, "::{}", def.variant(index).name).unwrap(); + } + } + if let (&ty::Generator(_, _, _), &Variants::Single { index }) = + (layout.ty.kind(), &layout.variants) + { + write!(&mut name, "::{}", ty::GeneratorSubsts::variant_name(index)).unwrap(); + } + Some(name) + } + ty::Adt(..) => { + // If `Some` is returned then a named struct is created in LLVM. Name collisions are + // avoided by LLVM (with increasing suffixes). If rustc doesn't generate names then that + // can improve perf. + // FIXME(antoyo): I don't think that's true for libgccjit. + Some(String::new()) + } + _ => None, + }; + + match layout.fields { + FieldsShape::Primitive | FieldsShape::Union(_) => { + let fill = cx.type_padding_filler(layout.size, layout.align.abi); + let packed = false; + match name { + None => cx.type_struct(&[fill], packed), + Some(ref name) => { + let gcc_type = cx.type_named_struct(name); + cx.set_struct_body(gcc_type, &[fill], packed); + gcc_type.as_type() + }, + } + } + FieldsShape::Array { count, .. } => cx.type_array(layout.field(cx, 0).gcc_type(cx, true), count), + FieldsShape::Arbitrary { .. } => + match name { + None => { + let (gcc_fields, packed) = struct_fields(cx, layout); + cx.type_struct(&gcc_fields, packed) + }, + Some(ref name) => { + let gcc_type = cx.type_named_struct(name); + *defer = Some((gcc_type, layout)); + gcc_type.as_type() + }, + }, + } +} + +pub trait LayoutGccExt<'tcx> { + fn is_gcc_immediate(&self) -> bool; + fn is_gcc_scalar_pair(&self) -> bool; + fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, set_fields: bool) -> Type<'gcc>; + fn immediate_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; + fn scalar_gcc_type_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, scalar: &abi::Scalar, offset: Size) -> Type<'gcc>; + fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize, immediate: bool) -> Type<'gcc>; + fn gcc_field_index(&self, index: usize) -> u64; + fn pointee_info_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, offset: Size) -> Option<PointeeInfo>; +} + +impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> { + fn is_gcc_immediate(&self) -> bool { + match self.abi { + Abi::Scalar(_) | Abi::Vector { .. } => true, + Abi::ScalarPair(..) => false, + Abi::Uninhabited | Abi::Aggregate { .. } => self.is_zst(), + } + } + + fn is_gcc_scalar_pair(&self) -> bool { + match self.abi { + Abi::ScalarPair(..) => true, + Abi::Uninhabited | Abi::Scalar(_) | Abi::Vector { .. } | Abi::Aggregate { .. } => false, + } + } + + /// Gets the GCC type corresponding to a Rust type, i.e., `rustc_middle::ty::Ty`. + /// The pointee type of the pointer in `PlaceRef` is always this type. + /// For sized types, it is also the right LLVM type for an `alloca` + /// containing a value of that type, and most immediates (except `bool`). + /// Unsized types, however, are represented by a "minimal unit", e.g. + /// `[T]` becomes `T`, while `str` and `Trait` turn into `i8` - this + /// is useful for indexing slices, as `&[T]`'s data pointer is `T*`. + /// If the type is an unsized struct, the regular layout is generated, + /// with the inner-most trailing unsized field using the "minimal unit" + /// of that field's type - this is useful for taking the address of + /// that field and ensuring the struct has the right alignment. + //TODO(antoyo): do we still need the set_fields parameter? + fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, set_fields: bool) -> Type<'gcc> { + if let Abi::Scalar(ref scalar) = self.abi { + // Use a different cache for scalars because pointers to DSTs + // can be either fat or thin (data pointers of fat pointers). + if let Some(&ty) = cx.scalar_types.borrow().get(&self.ty) { + return ty; + } + let ty = + match *self.ty.kind() { + ty::Ref(_, ty, _) | ty::RawPtr(ty::TypeAndMut { ty, .. }) => { + cx.type_ptr_to(cx.layout_of(ty).gcc_type(cx, set_fields)) + } + ty::Adt(def, _) if def.is_box() => { + cx.type_ptr_to(cx.layout_of(self.ty.boxed_ty()).gcc_type(cx, true)) + } + ty::FnPtr(sig) => cx.fn_ptr_backend_type(&cx.fn_abi_of_fn_ptr(sig, ty::List::empty())), + _ => self.scalar_gcc_type_at(cx, scalar, Size::ZERO), + }; + cx.scalar_types.borrow_mut().insert(self.ty, ty); + return ty; + } + + // Check the cache. + let variant_index = + match self.variants { + Variants::Single { index } => Some(index), + _ => None, + }; + let cached_type = cx.types.borrow().get(&(self.ty, variant_index)).cloned(); + if let Some(ty) = cached_type { + let type_to_set_fields = cx.types_with_fields_to_set.borrow_mut().remove(&ty); + if let Some((struct_type, layout)) = type_to_set_fields { + // Since we might be trying to generate a type containing another type which is not + // completely generated yet, we deferred setting the fields until now. + let (fields, packed) = struct_fields(cx, layout); + cx.set_struct_body(struct_type, &fields, packed); + } + return ty; + } + + assert!(!self.ty.has_escaping_bound_vars(), "{:?} has escaping bound vars", self.ty); + + // Make sure lifetimes are erased, to avoid generating distinct LLVM + // types for Rust types that only differ in the choice of lifetimes. + let normal_ty = cx.tcx.erase_regions(self.ty); + + let mut defer = None; + let ty = + if self.ty != normal_ty { + let mut layout = cx.layout_of(normal_ty); + if let Some(v) = variant_index { + layout = layout.for_variant(cx, v); + } + layout.gcc_type(cx, true) + } + else { + uncached_gcc_type(cx, *self, &mut defer) + }; + + cx.types.borrow_mut().insert((self.ty, variant_index), ty); + + if let Some((ty, layout)) = defer { + let (fields, packed) = struct_fields(cx, layout); + cx.set_struct_body(ty, &fields, packed); + } + + ty + } + + fn immediate_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> { + if let Abi::Scalar(ref scalar) = self.abi { + if scalar.is_bool() { + return cx.type_i1(); + } + } + self.gcc_type(cx, true) + } + + fn scalar_gcc_type_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, scalar: &abi::Scalar, offset: Size) -> Type<'gcc> { + match scalar.primitive() { + Int(i, true) => cx.type_from_integer(i), + Int(i, false) => cx.type_from_unsigned_integer(i), + F32 => cx.type_f32(), + F64 => cx.type_f64(), + Pointer => { + // If we know the alignment, pick something better than i8. + let pointee = + if let Some(pointee) = self.pointee_info_at(cx, offset) { + cx.type_pointee_for_align(pointee.align) + } + else { + cx.type_i8() + }; + cx.type_ptr_to(pointee) + } + } + } + + fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize, immediate: bool) -> Type<'gcc> { + // TODO(antoyo): remove llvm hack: + // HACK(eddyb) special-case fat pointers until LLVM removes + // pointee types, to avoid bitcasting every `OperandRef::deref`. + match self.ty.kind() { + ty::Ref(..) | ty::RawPtr(_) => { + return self.field(cx, index).gcc_type(cx, true); + } + // only wide pointer boxes are handled as pointers + // thin pointer boxes with scalar allocators are handled by the general logic below + ty::Adt(def, substs) if def.is_box() && cx.layout_of(substs.type_at(1)).is_zst() => { + let ptr_ty = cx.tcx.mk_mut_ptr(self.ty.boxed_ty()); + return cx.layout_of(ptr_ty).scalar_pair_element_gcc_type(cx, index, immediate); + } + _ => {} + } + + let (a, b) = match self.abi { + Abi::ScalarPair(ref a, ref b) => (a, b), + _ => bug!("TyAndLayout::scalar_pair_element_llty({:?}): not applicable", self), + }; + let scalar = [a, b][index]; + + // Make sure to return the same type `immediate_gcc_type` would when + // dealing with an immediate pair. This means that `(bool, bool)` is + // effectively represented as `{i8, i8}` in memory and two `i1`s as an + // immediate, just like `bool` is typically `i8` in memory and only `i1` + // when immediate. We need to load/store `bool` as `i8` to avoid + // crippling LLVM optimizations or triggering other LLVM bugs with `i1`. + // TODO(antoyo): this bugs certainly don't happen in this case since the bool type is used instead of i1. + if scalar.is_bool() { + return cx.type_i1(); + } + + let offset = + if index == 0 { + Size::ZERO + } + else { + a.size(cx).align_to(b.align(cx).abi) + }; + self.scalar_gcc_type_at(cx, scalar, offset) + } + + fn gcc_field_index(&self, index: usize) -> u64 { + match self.abi { + Abi::Scalar(_) | Abi::ScalarPair(..) => { + bug!("TyAndLayout::gcc_field_index({:?}): not applicable", self) + } + _ => {} + } + match self.fields { + FieldsShape::Primitive | FieldsShape::Union(_) => { + bug!("TyAndLayout::gcc_field_index({:?}): not applicable", self) + } + + FieldsShape::Array { .. } => index as u64, + + FieldsShape::Arbitrary { .. } => 1 + (self.fields.memory_index(index) as u64) * 2, + } + } + + fn pointee_info_at<'a>(&self, cx: &CodegenCx<'a, 'tcx>, offset: Size) -> Option<PointeeInfo> { + if let Some(&pointee) = cx.pointee_infos.borrow().get(&(self.ty, offset)) { + return pointee; + } + + let result = Ty::ty_and_layout_pointee_info_at(*self, cx, offset); + + cx.pointee_infos.borrow_mut().insert((self.ty, offset), result); + result + } +} + +impl<'gcc, 'tcx> LayoutTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> { + fn backend_type(&self, layout: TyAndLayout<'tcx>) -> Type<'gcc> { + layout.gcc_type(self, true) + } + + fn immediate_backend_type(&self, layout: TyAndLayout<'tcx>) -> Type<'gcc> { + layout.immediate_gcc_type(self) + } + + fn is_backend_immediate(&self, layout: TyAndLayout<'tcx>) -> bool { + layout.is_gcc_immediate() + } + + fn is_backend_scalar_pair(&self, layout: TyAndLayout<'tcx>) -> bool { + layout.is_gcc_scalar_pair() + } + + fn backend_field_index(&self, layout: TyAndLayout<'tcx>, index: usize) -> u64 { + layout.gcc_field_index(index) + } + + fn scalar_pair_element_backend_type(&self, layout: TyAndLayout<'tcx>, index: usize, immediate: bool) -> Type<'gcc> { + layout.scalar_pair_element_gcc_type(self, index, immediate) + } + + fn cast_backend_type(&self, ty: &CastTarget) -> Type<'gcc> { + ty.gcc_type(self) + } + + fn fn_ptr_backend_type(&self, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> { + fn_abi.ptr_to_gcc_type(self) + } + + fn reg_backend_type(&self, _ty: &Reg) -> Type<'gcc> { + unimplemented!(); + } + + fn fn_decl_backend_type(&self, _fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> { + // FIXME(antoyo): return correct type. + self.type_void() + } +} |