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

x64: Add AVX support for some more float-related instructions (#6092)

Browse Source 
* x64: Add AVX encodings of `vcvt{ss2sd,sd2ss}`

Additionally update the instruction helpers to take an `XmmMem` argument
to allow load sinking into the instruction.

* x64: Add AVX encoding of `sqrts{s,d}`

* x64: Add AVX support for `rounds{s,d}`
This commit is contained in:
Alex Crichton
2023-03-29 13:09:49 -05:00
committed by GitHub
parent afb417920d
commit 0b0ac3ff73
8 changed files with 543 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
cranelift/codegen/src/isa/x64/inst.isle
View File

@@ -1299,6 +1299,12 @@
Vpmovmskb
Vcvtsi2ss
Vcvtsi2sd
Vcvtss2sd
Vcvtsd2ss
Vsqrtss
Vsqrtsd
Vroundss
Vroundsd
))
(type Avx512Opcode extern
@@ -3348,11 +3354,17 @@
(decl x64_roundss (XmmMem RoundImm) Xmm)
(rule (x64_roundss src1 round)
(xmm_unary_rm_r_imm (SseOpcode.Roundss) src1 (encode_round_imm round)))
(rule 1 (x64_roundss src1 round)
(if-let $true (use_avx_simd))
(xmm_unary_rm_r_imm_vex (AvxOpcode.Vroundss) src1 (encode_round_imm round)))
;; Helper for creating `roundsd` instructions.
(decl x64_roundsd (XmmMem RoundImm) Xmm)
(rule (x64_roundsd src1 round)
(xmm_unary_rm_r_imm (SseOpcode.Roundsd) src1 (encode_round_imm round)))
(rule 1 (x64_roundsd src1 round)
(if-let $true (use_avx_simd))
(xmm_unary_rm_r_imm_vex (AvxOpcode.Vroundsd) src1 (encode_round_imm round)))
;; Helper for creating `roundps` instructions.
(decl x64_roundps (XmmMem RoundImm) Xmm)
@@ -3985,10 +3997,16 @@
;; Helper for creating `sqrtss` instructions.
(decl x64_sqrtss (XmmMem) Xmm)
(rule (x64_sqrtss x) (xmm_unary_rm_r_unaligned (SseOpcode.Sqrtss) x))
(rule 1 (x64_sqrtss x)
(if-let $true (use_avx_simd))
(xmm_unary_rm_r_vex (AvxOpcode.Vsqrtss) x))
;; Helper for creating `sqrtsd` instructions.
(decl x64_sqrtsd (XmmMem) Xmm)
(rule (x64_sqrtsd x) (xmm_unary_rm_r_unaligned (SseOpcode.Sqrtsd) x))
(rule 1 (x64_sqrtsd x)
(if-let $true (use_avx_simd))
(xmm_unary_rm_r_vex (AvxOpcode.Vsqrtsd) x))
;; Helper for creating `sqrtps` instructions.
(decl x64_sqrtps (XmmMem) Xmm)
@@ -4005,12 +4023,18 @@
(xmm_unary_rm_r_vex (AvxOpcode.Vsqrtpd) x))
;; Helper for creating `cvtss2sd` instructions.
(decl x64_cvtss2sd (Xmm) Xmm)
(rule (x64_cvtss2sd x) (xmm_unary_rm_r (SseOpcode.Cvtss2sd) x))
(decl x64_cvtss2sd (XmmMem) Xmm)
(rule (x64_cvtss2sd x) (xmm_unary_rm_r_unaligned (SseOpcode.Cvtss2sd) x))
(rule 1 (x64_cvtss2sd x)
(if-let $true (use_avx_simd))
(xmm_unary_rm_r_vex (AvxOpcode.Vcvtss2sd) x))
;; Helper for creating `cvtsd2ss` instructions.
(decl x64_cvtsd2ss (Xmm) Xmm)
(rule (x64_cvtsd2ss x) (xmm_unary_rm_r (SseOpcode.Cvtsd2ss) x))
(decl x64_cvtsd2ss (XmmMem) Xmm)
(rule (x64_cvtsd2ss x) (xmm_unary_rm_r_unaligned (SseOpcode.Cvtsd2ss) x))
(rule 1 (x64_cvtsd2ss x)
(if-let $true (use_avx_simd))
(xmm_unary_rm_r_vex (AvxOpcode.Vcvtsd2ss) x))
;; Helper for creating `cvtdq2ps` instructions.
(decl x64_cvtdq2ps (XmmMem) Xmm)

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

@@ -1722,7 +1722,13 @@ impl AvxOpcode {
| AvxOpcode::Vmovmskpd
| AvxOpcode::Vpmovmskb
| AvxOpcode::Vcvtsi2ss
| AvxOpcode::Vcvtsi2sd => {
| AvxOpcode::Vcvtsi2sd
| AvxOpcode::Vcvtss2sd
| AvxOpcode::Vcvtsd2ss
| AvxOpcode::Vsqrtss
| AvxOpcode::Vsqrtsd
| AvxOpcode::Vroundss
| AvxOpcode::Vroundsd => {
smallvec![InstructionSet::AVX]
}

42
cranelift/codegen/src/isa/x64/inst/emit.rs
View File

@@ -2405,17 +2405,36 @@ pub(crate) fn emit(
AvxOpcode::Vbroadcastss => (LegacyPrefixes::_66, OpcodeMap::_0F38, 0x18),
AvxOpcode::Vmovddup => (LegacyPrefixes::_F2, OpcodeMap::_0F, 0x12),
AvxOpcode::Vcvtss2sd => (LegacyPrefixes::_F3, OpcodeMap::_0F, 0x5A),
AvxOpcode::Vcvtsd2ss => (LegacyPrefixes::_F2, OpcodeMap::_0F, 0x5A),
AvxOpcode::Vsqrtss => (LegacyPrefixes::_F3, OpcodeMap::_0F, 0x51),
AvxOpcode::Vsqrtsd => (LegacyPrefixes::_F2, OpcodeMap::_0F, 0x51),
_ => panic!("unexpected rmr_imm_vex opcode {op:?}"),
};
VexInstruction::new()
let vex = VexInstruction::new()
.length(VexVectorLength::V128)
.prefix(prefix)
.map(map)
.opcode(opcode)
.reg(dst.to_real_reg().unwrap().hw_enc())
.rm(src)
.encode(sink);
.rm(src);
// These opcodes take a second operand through `vvvv` which copies
// the upper bits into the destination register. That's not
// reflected in the CLIF instruction, however, since the SSE version
// doesn't have this functionality. Instead just copy whatever
// happens to already be in the destination, which at least is what
// LLVM seems to do.
let vex = match op {
AvxOpcode::Vcvtss2sd
| AvxOpcode::Vcvtsd2ss
| AvxOpcode::Vsqrtss
| AvxOpcode::Vsqrtsd => vex.vvvv(dst.to_real_reg().unwrap().hw_enc()),
_ => vex,
};
vex.encode(sink);
}
Inst::XmmUnaryRmRImmVex { op, src, dst, imm } => {
@@ -2433,18 +2452,29 @@ pub(crate) fn emit(
AvxOpcode::Vpshuflw => (LegacyPrefixes::_F2, OpcodeMap::_0F, 0x70),
AvxOpcode::Vpshufhw => (LegacyPrefixes::_F3, OpcodeMap::_0F, 0x70),
AvxOpcode::Vpshufd => (LegacyPrefixes::_66, OpcodeMap::_0F, 0x70),
AvxOpcode::Vroundss => (LegacyPrefixes::_66, OpcodeMap::_0F3A, 0x0A),
AvxOpcode::Vroundsd => (LegacyPrefixes::_66, OpcodeMap::_0F3A, 0x0B),
_ => panic!("unexpected rmr_imm_vex opcode {op:?}"),
};
VexInstruction::new()
let vex = VexInstruction::new()
.length(VexVectorLength::V128)
.prefix(prefix)
.map(map)
.opcode(opcode)
.reg(dst.to_real_reg().unwrap().hw_enc())
.rm(src)
.imm(*imm)
.encode(sink);
.imm(*imm);
// See comments in similar block above in `XmmUnaryRmRVex` for what
// this is doing.
let vex = match op {
AvxOpcode::Vroundss | AvxOpcode::Vroundsd => {
vex.vvvv(dst.to_real_reg().unwrap().hw_enc())
}
_ => vex,
};
vex.encode(sink);
}
Inst::XmmMovRMVex { op, src, dst } => {