diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-28 14:29:10 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-28 14:29:10 +0000 |
commit | 2aa4a82499d4becd2284cdb482213d541b8804dd (patch) | |
tree | b80bf8bf13c3766139fbacc530efd0dd9d54394c /third_party/rust/cranelift-codegen/src/souper_harvest.rs | |
parent | Initial commit. (diff) | |
download | firefox-upstream.tar.xz firefox-upstream.zip |
Adding upstream version 86.0.1.upstream/86.0.1upstream
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'third_party/rust/cranelift-codegen/src/souper_harvest.rs')
-rw-r--r-- | third_party/rust/cranelift-codegen/src/souper_harvest.rs | 527 |
1 files changed, 527 insertions, 0 deletions
diff --git a/third_party/rust/cranelift-codegen/src/souper_harvest.rs b/third_party/rust/cranelift-codegen/src/souper_harvest.rs new file mode 100644 index 0000000000..fcf53b0ed5 --- /dev/null +++ b/third_party/rust/cranelift-codegen/src/souper_harvest.rs @@ -0,0 +1,527 @@ +//! Harvest left-hand side superoptimization candidates. +//! +//! Given a clif function, harvest all its integer subexpressions, so that they +//! can be fed into [Souper](https://github.com/google/souper) as candidates for +//! superoptimization. For some of these candidates, Souper will successfully +//! synthesize a right-hand side that is equivalent but has lower cost than the +//! left-hand side. Then, we can combine these left- and right-hand sides into a +//! complete optimization, and add it to our peephole passes. +//! +//! To harvest the expression that produced a given value `x`, we do a +//! post-order traversal of the dataflow graph starting from `x`. As we do this +//! traversal, we maintain a map from clif values to their translated Souper +//! values. We stop traversing when we reach anything that can't be translated +//! into Souper IR: a memory load, a float-to-int conversion, a block parameter, +//! etc. For values produced by these instructions, we create a Souper `var`, +//! which is an input variable to the optimization. For instructions that have a +//! direct mapping into Souper IR, we get the Souper version of each of its +//! operands and then create the Souper version of the instruction itself. It +//! should now be clear why we do a post-order traversal: we need an +//! instruction's translated operands in order to translate the instruction +//! itself. Once this instruction is translated, we update the clif-to-souper +//! map with this new translation so that any other instruction that uses this +//! result as an operand has access to the translated value. When the traversal +//! is complete we return the translation of `x` as the root of left-hand side +//! candidate. + +use crate::ir; +use souper_ir::ast; +use std::collections::{HashMap, HashSet}; +use std::string::String; +use std::sync::mpsc; +use std::vec::Vec; + +/// Harvest Souper left-hand side candidates from the given function. +/// +/// Candidates are reported through the given MPSC sender. +pub fn do_souper_harvest(func: &ir::Function, out: &mut mpsc::Sender<String>) { + let mut allocs = Allocs::default(); + + // Iterate over each instruction in each block and try and harvest a + // left-hand side from its result. + for block in func.layout.blocks() { + let mut option_inst = func.layout.first_inst(block); + while let Some(inst) = option_inst { + let results = func.dfg.inst_results(inst); + if results.len() == 1 { + let val = results[0]; + let ty = func.dfg.value_type(val); + if ty.is_int() && ty.lane_count() == 1 { + harvest_candidate_lhs(&mut allocs, func, val, out); + } + } + option_inst = func.layout.next_inst(inst); + } + } +} + +/// Allocations that we reuse across many LHS candidate harvests. +#[derive(Default)] +struct Allocs { + /// A map from cranelift IR to souper IR for values that we've already + /// translated into souper IR. + ir_to_souper_val: HashMap<ir::Value, ast::ValueId>, + + /// Stack of to-visit and to-trace values for the post-order DFS. + dfs_stack: Vec<StackEntry>, + + /// Set of values we've already seen in our post-order DFS. + dfs_seen: HashSet<ir::Value>, +} + +impl Allocs { + /// Reset the collections to their empty state (without deallocating their + /// backing data). + fn reset(&mut self) { + self.ir_to_souper_val.clear(); + self.dfs_stack.clear(); + self.dfs_seen.clear(); + } +} + +/// Harvest a candidate LHS for `val` from the dataflow graph. +fn harvest_candidate_lhs( + allocs: &mut Allocs, + func: &ir::Function, + val: ir::Value, + out: &mut mpsc::Sender<String>, +) { + allocs.reset(); + let mut lhs = ast::LeftHandSideBuilder::default(); + let mut non_var_count = 0; + + // Should we keep tracing through the given `val`? Only if it is defined + // by an instruction that we can translate to Souper IR. + let should_trace = |val| match func.dfg.value_def(val) { + ir::ValueDef::Result(inst, 0) => match func.dfg[inst].opcode() { + ir::Opcode::Iadd + | ir::Opcode::IaddImm + | ir::Opcode::IrsubImm + | ir::Opcode::Imul + | ir::Opcode::ImulImm + | ir::Opcode::Udiv + | ir::Opcode::UdivImm + | ir::Opcode::Sdiv + | ir::Opcode::SdivImm + | ir::Opcode::Urem + | ir::Opcode::UremImm + | ir::Opcode::Srem + | ir::Opcode::SremImm + | ir::Opcode::Band + | ir::Opcode::BandImm + | ir::Opcode::Bor + | ir::Opcode::BorImm + | ir::Opcode::Bxor + | ir::Opcode::BxorImm + | ir::Opcode::Ishl + | ir::Opcode::IshlImm + | ir::Opcode::Sshr + | ir::Opcode::SshrImm + | ir::Opcode::Ushr + | ir::Opcode::UshrImm + | ir::Opcode::Select + | ir::Opcode::Uextend + | ir::Opcode::Sextend + | ir::Opcode::Trunc + | ir::Opcode::Icmp + | ir::Opcode::Popcnt + | ir::Opcode::Bitrev + | ir::Opcode::Clz + | ir::Opcode::Ctz + // TODO: ir::Opcode::IaddCarry + // TODO: ir::Opcode::IaddCout + | ir::Opcode::SaddSat + | ir::Opcode::SsubSat + | ir::Opcode::UsubSat => true, + _ => false, + }, + _ => false, + }; + + post_order_dfs(allocs, &func.dfg, val, should_trace, |allocs, val| { + let souper_assignment_rhs = match func.dfg.value_def(val) { + ir::ValueDef::Result(inst, 0) => { + let args = func.dfg.inst_args(inst); + let arg = |allocs: &mut Allocs, n| allocs.ir_to_souper_val[&args[n]].into(); + + match (func.dfg[inst].opcode(), &func.dfg[inst]) { + (ir::Opcode::Iadd, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Add { a, b }.into() + } + (ir::Opcode::IaddImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Add { a, b }.into() + } + (ir::Opcode::IrsubImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let b = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let a = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Sub { a, b }.into() + } + (ir::Opcode::Imul, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Mul { a, b }.into() + } + (ir::Opcode::ImulImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Mul { a, b }.into() + } + (ir::Opcode::Udiv, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Udiv { a, b }.into() + } + (ir::Opcode::UdivImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Udiv { a, b }.into() + } + (ir::Opcode::Sdiv, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Sdiv { a, b }.into() + } + (ir::Opcode::SdivImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Sdiv { a, b }.into() + } + (ir::Opcode::Urem, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Urem { a, b }.into() + } + (ir::Opcode::UremImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Urem { a, b }.into() + } + (ir::Opcode::Srem, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Srem { a, b }.into() + } + (ir::Opcode::SremImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Srem { a, b }.into() + } + (ir::Opcode::Band, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::And { a, b }.into() + } + (ir::Opcode::BandImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::And { a, b }.into() + } + (ir::Opcode::Bor, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Or { a, b }.into() + } + (ir::Opcode::BorImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Or { a, b }.into() + } + (ir::Opcode::Bxor, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Xor { a, b }.into() + } + (ir::Opcode::BxorImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Xor { a, b }.into() + } + (ir::Opcode::Ishl, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Shl { a, b }.into() + } + (ir::Opcode::IshlImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Shl { a, b }.into() + } + (ir::Opcode::Sshr, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Ashr { a, b }.into() + } + (ir::Opcode::SshrImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Ashr { a, b }.into() + } + (ir::Opcode::Ushr, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::Lshr { a, b }.into() + } + (ir::Opcode::UshrImm, ir::InstructionData::BinaryImm64 { imm, .. }) => { + let a = arg(allocs, 0); + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + let b = ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into(); + ast::Instruction::Lshr { a, b }.into() + } + (ir::Opcode::Select, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + let c = arg(allocs, 2); + ast::Instruction::Select { a, b, c }.into() + } + (ir::Opcode::Uextend, _) => { + let a = arg(allocs, 0); + ast::Instruction::Zext { a }.into() + } + (ir::Opcode::Sextend, _) => { + let a = arg(allocs, 0); + ast::Instruction::Sext { a }.into() + } + (ir::Opcode::Trunc, _) => { + let a = arg(allocs, 0); + ast::Instruction::Trunc { a }.into() + } + (ir::Opcode::Icmp, ir::InstructionData::IntCompare { cond, .. }) + | (ir::Opcode::IcmpImm, ir::InstructionData::IntCompare { cond, .. }) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + match cond { + ir::condcodes::IntCC::Equal => ast::Instruction::Eq { a, b }.into(), + ir::condcodes::IntCC::NotEqual => ast::Instruction::Ne { a, b }.into(), + ir::condcodes::IntCC::UnsignedLessThan => { + ast::Instruction::Ult { a, b }.into() + } + ir::condcodes::IntCC::SignedLessThan => { + ast::Instruction::Slt { a, b }.into() + } + ir::condcodes::IntCC::UnsignedLessThanOrEqual => { + ast::Instruction::Sle { a, b }.into() + } + ir::condcodes::IntCC::SignedLessThanOrEqual => { + ast::Instruction::Sle { a, b }.into() + } + _ => ast::AssignmentRhs::Var, + } + } + (ir::Opcode::Popcnt, _) => { + let a = arg(allocs, 0); + ast::Instruction::Ctpop { a }.into() + } + (ir::Opcode::Bitrev, _) => { + let a = arg(allocs, 0); + ast::Instruction::BitReverse { a }.into() + } + (ir::Opcode::Clz, _) => { + let a = arg(allocs, 0); + ast::Instruction::Ctlz { a }.into() + } + (ir::Opcode::Ctz, _) => { + let a = arg(allocs, 0); + ast::Instruction::Cttz { a }.into() + } + // TODO: ir::Opcode::IaddCarry + // TODO: ir::Opcode::IaddCout + (ir::Opcode::SaddSat, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::SaddSat { a, b }.into() + } + (ir::Opcode::SsubSat, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::SsubSat { a, b }.into() + } + (ir::Opcode::UsubSat, _) => { + let a = arg(allocs, 0); + let b = arg(allocs, 1); + ast::Instruction::UsubSat { a, b }.into() + } + (ir::Opcode::Iconst, ir::InstructionData::UnaryImm { imm, .. }) => { + let value: i64 = (*imm).into(); + let value: i128 = value.into(); + ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into() + } + (ir::Opcode::Bconst, ir::InstructionData::UnaryBool { imm, .. }) => { + let value = *imm as i128; + ast::Constant { + value, + r#type: souper_type_of(&func.dfg, val), + } + .into() + } + _ => ast::AssignmentRhs::Var, + } + } + _ => ast::AssignmentRhs::Var, + }; + + non_var_count += match souper_assignment_rhs { + ast::AssignmentRhs::Var => 0, + _ => 1, + }; + let souper_ty = souper_type_of(&func.dfg, val); + let souper_val = lhs.assignment(None, souper_ty, souper_assignment_rhs, vec![]); + let old_value = allocs.ir_to_souper_val.insert(val, souper_val); + assert!(old_value.is_none()); + }); + + // We end up harvesting a lot of candidates like: + // + // %0:i32 = var + // infer %0 + // + // and + // + // %0:i32 = var + // %1:i32 = var + // %2:i32 = add %0, %1 + // + // Both of these are useless. Only actually harvest the candidate if there + // are at least two actual operations. + if non_var_count >= 2 { + let lhs = lhs.finish(allocs.ir_to_souper_val[&val], None); + out.send(format!( + ";; Harvested from `{}` in `{}`\n{}\n", + val, func.name, lhs + )) + .unwrap(); + } +} + +fn souper_type_of(dfg: &ir::DataFlowGraph, val: ir::Value) -> Option<ast::Type> { + let ty = dfg.value_type(val); + assert!(ty.is_int() || ty.is_bool()); + assert_eq!(ty.lane_count(), 1); + Some(ast::Type { width: ty.bits() }) +} + +#[derive(Debug)] +enum StackEntry { + Visit(ir::Value), + Trace(ir::Value), +} + +fn post_order_dfs( + allocs: &mut Allocs, + dfg: &ir::DataFlowGraph, + val: ir::Value, + should_trace: impl Fn(ir::Value) -> bool, + mut visit: impl FnMut(&mut Allocs, ir::Value), +) { + allocs.dfs_stack.push(StackEntry::Trace(val)); + + while let Some(entry) = allocs.dfs_stack.pop() { + match entry { + StackEntry::Visit(val) => { + let is_new = allocs.dfs_seen.insert(val); + if is_new { + visit(allocs, val); + } + } + StackEntry::Trace(val) => { + if allocs.dfs_seen.contains(&val) { + continue; + } + + allocs.dfs_stack.push(StackEntry::Visit(val)); + if should_trace(val) { + if let ir::ValueDef::Result(inst, 0) = dfg.value_def(val) { + let args = dfg.inst_args(inst); + for v in args.iter().rev().copied() { + allocs.dfs_stack.push(StackEntry::Trace(v)); + } + } + } + } + } + } +} |