T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

machinst x64: use zero-latency move instructions for f32/f64;

Browse Source 
As found by @julian-seward1, movss/movsd aren't included in the
zero-latency move instructions section of the Intel optimization manual.
Use MOVAPS instead for those moves.
This commit is contained in:
Benjamin Bouvier
2020-10-06 18:30:14 +02:00
parent fc430eef76
commit 84ac3feef8
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
cranelift/codegen/src/isa/x64/inst/mod.rs
View File

@@ -2401,10 +2401,12 @@ impl MachInst for Inst {
match rc_dst { match rc_dst {
RegClass::I64 => Inst::mov_r_r(true, src_reg, dst_reg), RegClass::I64 => Inst::mov_r_r(true, src_reg, dst_reg),
RegClass::V128 => { RegClass::V128 => {
// The Intel optimization manual, in "3.5.1.13 Zero-Latency MOV Instructions",
// doesn't include MOVSS/MOVSD as instructions with zero-latency. Use movaps for
// those, which may write more lanes that we need, but are specified to have
// zero-latency.
let opcode = match ty { let opcode = match ty {
types::F32 => SseOpcode::Movss, types::F32 | types::F64 | types::F32X4 => SseOpcode::Movaps,
types::F64 => SseOpcode::Movsd,
types::F32X4 => SseOpcode::Movaps,
types::F64X2 => SseOpcode::Movapd, types::F64X2 => SseOpcode::Movapd,
_ if ty.is_vector() && ty.bits() == 128 => SseOpcode::Movdqa, _ if ty.is_vector() && ty.bits() == 128 => SseOpcode::Movdqa,
_ => unimplemented!("unable to move type: {}", ty), _ => unimplemented!("unable to move type: {}", ty),