diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-17 12:11:38 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-17 12:13:23 +0000 |
commit | 20431706a863f92cb37dc512fef6e48d192aaf2c (patch) | |
tree | 2867f13f5fd5437ba628c67d7f87309ccadcd286 /compiler/rustc_codegen_cranelift/src/intrinsics | |
parent | Releasing progress-linux version 1.65.0+dfsg1-2~progress7.99u1. (diff) | |
download | rustc-20431706a863f92cb37dc512fef6e48d192aaf2c.tar.xz rustc-20431706a863f92cb37dc512fef6e48d192aaf2c.zip |
Merging upstream version 1.66.0+dfsg1.
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to '')
3 files changed, 158 insertions, 39 deletions
diff --git a/compiler/rustc_codegen_cranelift/src/intrinsics/llvm.rs b/compiler/rustc_codegen_cranelift/src/intrinsics/llvm.rs index a799dca93..783d426c3 100644 --- a/compiler/rustc_codegen_cranelift/src/intrinsics/llvm.rs +++ b/compiler/rustc_codegen_cranelift/src/intrinsics/llvm.rs @@ -14,6 +14,10 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>( target: Option<BasicBlock>, ) { match intrinsic { + "llvm.x86.sse2.pause" | "llvm.aarch64.isb" => { + // Spin loop hint + } + // Used by `_mm_movemask_epi8` and `_mm256_movemask_epi8` "llvm.x86.sse2.pmovmskb.128" | "llvm.x86.avx2.pmovmskb" | "llvm.x86.sse2.movmsk.pd" => { intrinsic_args!(fx, args => (a); intrinsic); @@ -25,8 +29,7 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>( let mut res = fx.bcx.ins().iconst(types::I32, 0); for lane in (0..lane_count).rev() { - let a_lane = - a.value_field(fx, mir::Field::new(lane.try_into().unwrap())).load_scalar(fx); + let a_lane = a.value_lane(fx, lane).load_scalar(fx); // cast float to int let a_lane = match lane_ty { diff --git a/compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs b/compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs index 2e4ca594f..0302b843a 100644 --- a/compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs +++ b/compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs @@ -84,6 +84,30 @@ fn simd_for_each_lane<'tcx>( } } +fn simd_pair_for_each_lane_typed<'tcx>( + fx: &mut FunctionCx<'_, '_, 'tcx>, + x: CValue<'tcx>, + y: CValue<'tcx>, + ret: CPlace<'tcx>, + f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, CValue<'tcx>, CValue<'tcx>) -> CValue<'tcx>, +) { + assert_eq!(x.layout(), y.layout()); + let layout = x.layout(); + + let (lane_count, _lane_ty) = layout.ty.simd_size_and_type(fx.tcx); + let (ret_lane_count, _ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx); + assert_eq!(lane_count, ret_lane_count); + + for lane_idx in 0..lane_count { + let x_lane = x.value_lane(fx, lane_idx); + let y_lane = y.value_lane(fx, lane_idx); + + let res_lane = f(fx, x_lane, y_lane); + + ret.place_lane(fx, lane_idx).write_cvalue(fx, res_lane); + } +} + fn simd_pair_for_each_lane<'tcx>( fx: &mut FunctionCx<'_, '_, 'tcx>, x: CValue<'tcx>, @@ -504,37 +528,7 @@ fn codegen_regular_intrinsic_call<'tcx>( _ => unreachable!(), }; - let signed = type_sign(lhs.layout().ty); - - let checked_res = crate::num::codegen_checked_int_binop(fx, bin_op, lhs, rhs); - - let (val, has_overflow) = checked_res.load_scalar_pair(fx); - let clif_ty = fx.clif_type(lhs.layout().ty).unwrap(); - - let (min, max) = type_min_max_value(&mut fx.bcx, clif_ty, signed); - - let val = match (intrinsic, signed) { - (sym::saturating_add, false) => fx.bcx.ins().select(has_overflow, max, val), - (sym::saturating_sub, false) => fx.bcx.ins().select(has_overflow, min, val), - (sym::saturating_add, true) => { - let rhs = rhs.load_scalar(fx); - let rhs_ge_zero = - fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThanOrEqual, rhs, 0); - let sat_val = fx.bcx.ins().select(rhs_ge_zero, max, min); - fx.bcx.ins().select(has_overflow, sat_val, val) - } - (sym::saturating_sub, true) => { - let rhs = rhs.load_scalar(fx); - let rhs_ge_zero = - fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThanOrEqual, rhs, 0); - let sat_val = fx.bcx.ins().select(rhs_ge_zero, min, max); - fx.bcx.ins().select(has_overflow, sat_val, val) - } - _ => unreachable!(), - }; - - let res = CValue::by_val(val, lhs.layout()); - + let res = crate::num::codegen_saturating_int_binop(fx, bin_op, lhs, rhs); ret.write_cvalue(fx, res); } sym::rotate_left => { @@ -819,8 +813,8 @@ fn codegen_regular_intrinsic_call<'tcx>( sym::ptr_guaranteed_cmp => { intrinsic_args!(fx, args => (a, b); intrinsic); - let val = crate::num::codegen_ptr_binop(fx, BinOp::Eq, a, b); - ret.write_cvalue(fx, val); + let val = crate::num::codegen_ptr_binop(fx, BinOp::Eq, a, b).load_scalar(fx); + ret.write_cvalue(fx, CValue::by_val(val, fx.layout_of(fx.tcx.types.u8))); } sym::caller_location => { @@ -1206,7 +1200,7 @@ fn codegen_regular_intrinsic_call<'tcx>( // FIXME once unwinding is supported, change this to actually catch panics let f_sig = fx.bcx.func.import_signature(Signature { call_conv: fx.target_config.default_call_conv, - params: vec![AbiParam::new(fx.bcx.func.dfg.value_type(data))], + params: vec![AbiParam::new(pointer_ty(fx.tcx))], returns: vec![], }); diff --git a/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs b/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs index 1f358b1bb..51fce8c85 100644 --- a/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs +++ b/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs @@ -2,6 +2,7 @@ use rustc_middle::ty::subst::SubstsRef; use rustc_span::Symbol; +use rustc_target::abi::Endian; use super::*; use crate::prelude::*; @@ -26,7 +27,7 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>( span: Span, ) { match intrinsic { - sym::simd_cast => { + sym::simd_as | sym::simd_cast => { intrinsic_args!(fx, args => (a); intrinsic); if !a.layout().ty.is_simd() { @@ -162,6 +163,7 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>( } } } else { + // FIXME remove this case intrinsic.as_str()["simd_shuffle".len()..].parse().unwrap() }; @@ -650,8 +652,128 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>( } } - // simd_saturating_* - // simd_bitmask + sym::simd_select_bitmask => { + intrinsic_args!(fx, args => (m, a, b); intrinsic); + + if !a.layout().ty.is_simd() { + report_simd_type_validation_error(fx, intrinsic, span, a.layout().ty); + return; + } + assert_eq!(a.layout(), b.layout()); + + let (lane_count, lane_ty) = a.layout().ty.simd_size_and_type(fx.tcx); + let lane_layout = fx.layout_of(lane_ty); + + let m = m.load_scalar(fx); + + for lane in 0..lane_count { + let m_lane = fx.bcx.ins().ushr_imm(m, u64::from(lane) as i64); + let m_lane = fx.bcx.ins().band_imm(m_lane, 1); + let a_lane = a.value_lane(fx, lane).load_scalar(fx); + let b_lane = b.value_lane(fx, lane).load_scalar(fx); + + let m_lane = fx.bcx.ins().icmp_imm(IntCC::Equal, m_lane, 0); + let res_lane = + CValue::by_val(fx.bcx.ins().select(m_lane, b_lane, a_lane), lane_layout); + + ret.place_lane(fx, lane).write_cvalue(fx, res_lane); + } + } + + sym::simd_bitmask => { + intrinsic_args!(fx, args => (a); intrinsic); + + let (lane_count, lane_ty) = a.layout().ty.simd_size_and_type(fx.tcx); + let lane_clif_ty = fx.clif_type(lane_ty).unwrap(); + + // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a + // vector mask and returns the most significant bit (MSB) of each lane in the form + // of either: + // * an unsigned integer + // * an array of `u8` + // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits. + // + // The bit order of the result depends on the byte endianness, LSB-first for little + // endian and MSB-first for big endian. + let expected_int_bits = lane_count.max(8); + let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64); + + match lane_ty.kind() { + ty::Int(_) | ty::Uint(_) => {} + _ => { + fx.tcx.sess.span_fatal( + span, + &format!( + "invalid monomorphization of `simd_bitmask` intrinsic: \ + vector argument `{}`'s element type `{}`, expected integer element \ + type", + a.layout().ty, + lane_ty + ), + ); + } + } + + let res_type = + Type::int_with_byte_size(u16::try_from(expected_bytes).unwrap()).unwrap(); + let mut res = fx.bcx.ins().iconst(res_type, 0); + + let lanes = match fx.tcx.sess.target.endian { + Endian::Big => Box::new(0..lane_count) as Box<dyn Iterator<Item = u64>>, + Endian::Little => Box::new((0..lane_count).rev()) as Box<dyn Iterator<Item = u64>>, + }; + for lane in lanes { + let a_lane = a.value_lane(fx, lane).load_scalar(fx); + + // extract sign bit of an int + let a_lane_sign = fx.bcx.ins().ushr_imm(a_lane, i64::from(lane_clif_ty.bits() - 1)); + + // shift sign bit into result + let a_lane_sign = clif_intcast(fx, a_lane_sign, res_type, false); + res = fx.bcx.ins().ishl_imm(res, 1); + res = fx.bcx.ins().bor(res, a_lane_sign); + } + + match ret.layout().ty.kind() { + ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {} + ty::Array(elem, len) + if matches!(elem.kind(), ty::Uint(ty::UintTy::U8)) + && len.try_eval_usize(fx.tcx, ty::ParamEnv::reveal_all()) + == Some(expected_bytes) => {} + _ => { + fx.tcx.sess.span_fatal( + span, + &format!( + "invalid monomorphization of `simd_bitmask` intrinsic: \ + cannot return `{}`, expected `u{}` or `[u8; {}]`", + ret.layout().ty, + expected_int_bits, + expected_bytes + ), + ); + } + } + + let res = CValue::by_val(res, ret.layout()); + ret.write_cvalue(fx, res); + } + + sym::simd_saturating_add | sym::simd_saturating_sub => { + intrinsic_args!(fx, args => (x, y); intrinsic); + + let bin_op = match intrinsic { + sym::simd_saturating_add => BinOp::Add, + sym::simd_saturating_sub => BinOp::Sub, + _ => unreachable!(), + }; + + // FIXME use vector instructions when possible + simd_pair_for_each_lane_typed(fx, x, y, ret, &|fx, x_lane, y_lane| { + crate::num::codegen_saturating_int_binop(fx, bin_op, x_lane, y_lane) + }); + } + + // simd_arith_offset // simd_scatter // simd_gather _ => { |