cmd/compile: make 386 float-to-int conversions match amd64

Change-Id: Iff13b4471f94a6a91d8b159603a9338cb9c89747
Reviewed-on: https://go-review.googlesource.com/c/go/+/708295
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
This commit is contained in:
David Chase 2025-09-30 16:27:32 -04:00
parent 0e466a8d1d
commit 78d75b3799
9 changed files with 182 additions and 23 deletions

View file

@ -88,8 +88,10 @@
(Cvt32to32F ...) => (CVTSL2SS ...)
(Cvt32to64F ...) => (CVTSL2SD ...)
(Cvt32Fto32 ...) => (CVTTSS2SL ...)
(Cvt64Fto32 ...) => (CVTTSD2SL ...)
(Cvt32Fto32 <t> x) && base.ConvertHash.MatchPos(v.Pos, nil) => (XORL <t> y (SARLconst <t> [31] (ANDL <t> y:(CVTTSS2SL <t> x) (NOTL <typ.Int32> (MOVLf2i x)))))
(Cvt64Fto32 <t> x) && base.ConvertHash.MatchPos(v.Pos, nil) => (XORL <t> y (SARLconst <t> [31] (ANDL <t> y:(CVTTSD2SL <t> x) (NOTL <typ.Int32> (MOVLf2i (CVTSD2SS <typ.Float32> x))))))
(Cvt32Fto32 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSS2SL <t> x)
(Cvt64Fto32 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSD2SL <t> x)
(Cvt32Fto64F ...) => (CVTSS2SD ...)
(Cvt64Fto32F ...) => (CVTSD2SS ...)

View file

@ -342,6 +342,8 @@ func init() {
{name: "MOVWLSX", argLength: 1, reg: gp11, asm: "MOVWLSX"}, // sign extend arg0 from int16 to int32
{name: "MOVWLZX", argLength: 1, reg: gp11, asm: "MOVWLZX"}, // zero extend arg0 from int16 to int32
{name: "MOVLf2i", argLength: 1, reg: fpgp, typ: "UInt32"}, // move 32 bits from float to int reg, zero extend
{name: "MOVLconst", reg: gp01, asm: "MOVL", typ: "UInt32", aux: "Int32", rematerializeable: true}, // 32 low bits of auxint
{name: "CVTTSD2SL", argLength: 1, reg: fpgp, asm: "CVTTSD2SL"}, // convert float64 to int32

View file

@ -499,6 +499,7 @@ const (
Op386MOVBLZX
Op386MOVWLSX
Op386MOVWLZX
Op386MOVLf2i
Op386MOVLconst
Op386CVTTSD2SL
Op386CVTTSS2SL
@ -5601,6 +5602,18 @@ var opcodeTable = [...]opInfo{
},
},
},
{
name: "MOVLf2i",
argLen: 1,
reg: regInfo{
inputs: []inputInfo{
{0, 65280}, // X0 X1 X2 X3 X4 X5 X6 X7
},
outputs: []outputInfo{
{0, 239}, // AX CX DX BX BP SI DI
},
},
},
{
name: "MOVLconst",
auxType: auxInt32,

View file

@ -3,6 +3,7 @@
package ssa
import "math"
import "cmd/compile/internal/base"
import "cmd/compile/internal/types"
func rewriteValue386(v *Value) bool {
@ -340,8 +341,7 @@ func rewriteValue386(v *Value) bool {
v.Op = Op386BSFL
return true
case OpCvt32Fto32:
v.Op = Op386CVTTSS2SL
return true
return rewriteValue386_OpCvt32Fto32(v)
case OpCvt32Fto64F:
v.Op = Op386CVTSS2SD
return true
@ -352,8 +352,7 @@ func rewriteValue386(v *Value) bool {
v.Op = Op386CVTSL2SD
return true
case OpCvt64Fto32:
v.Op = Op386CVTTSD2SL
return true
return rewriteValue386_OpCvt64Fto32(v)
case OpCvt64Fto32F:
v.Op = Op386CVTSD2SS
return true
@ -7964,6 +7963,98 @@ func rewriteValue386_OpCtz8(v *Value) bool {
return true
}
}
func rewriteValue386_OpCvt32Fto32(v *Value) bool {
v_0 := v.Args[0]
b := v.Block
typ := &b.Func.Config.Types
// match: (Cvt32Fto32 <t> x)
// cond: base.ConvertHash.MatchPos(v.Pos, nil)
// result: (XORL <t> y (SARLconst <t> [31] (ANDL <t> y:(CVTTSS2SL <t> x) (NOTL <typ.Int32> (MOVLf2i x)))))
for {
t := v.Type
x := v_0
if !(base.ConvertHash.MatchPos(v.Pos, nil)) {
break
}
v.reset(Op386XORL)
v.Type = t
v0 := b.NewValue0(v.Pos, Op386SARLconst, t)
v0.AuxInt = int32ToAuxInt(31)
v1 := b.NewValue0(v.Pos, Op386ANDL, t)
y := b.NewValue0(v.Pos, Op386CVTTSS2SL, t)
y.AddArg(x)
v3 := b.NewValue0(v.Pos, Op386NOTL, typ.Int32)
v4 := b.NewValue0(v.Pos, Op386MOVLf2i, typ.UInt32)
v4.AddArg(x)
v3.AddArg(v4)
v1.AddArg2(y, v3)
v0.AddArg(v1)
v.AddArg2(y, v0)
return true
}
// match: (Cvt32Fto32 <t> x)
// cond: !base.ConvertHash.MatchPos(v.Pos, nil)
// result: (CVTTSS2SL <t> x)
for {
t := v.Type
x := v_0
if !(!base.ConvertHash.MatchPos(v.Pos, nil)) {
break
}
v.reset(Op386CVTTSS2SL)
v.Type = t
v.AddArg(x)
return true
}
return false
}
func rewriteValue386_OpCvt64Fto32(v *Value) bool {
v_0 := v.Args[0]
b := v.Block
typ := &b.Func.Config.Types
// match: (Cvt64Fto32 <t> x)
// cond: base.ConvertHash.MatchPos(v.Pos, nil)
// result: (XORL <t> y (SARLconst <t> [31] (ANDL <t> y:(CVTTSD2SL <t> x) (NOTL <typ.Int32> (MOVLf2i (CVTSD2SS <typ.Float32> x))))))
for {
t := v.Type
x := v_0
if !(base.ConvertHash.MatchPos(v.Pos, nil)) {
break
}
v.reset(Op386XORL)
v.Type = t
v0 := b.NewValue0(v.Pos, Op386SARLconst, t)
v0.AuxInt = int32ToAuxInt(31)
v1 := b.NewValue0(v.Pos, Op386ANDL, t)
y := b.NewValue0(v.Pos, Op386CVTTSD2SL, t)
y.AddArg(x)
v3 := b.NewValue0(v.Pos, Op386NOTL, typ.Int32)
v4 := b.NewValue0(v.Pos, Op386MOVLf2i, typ.UInt32)
v5 := b.NewValue0(v.Pos, Op386CVTSD2SS, typ.Float32)
v5.AddArg(x)
v4.AddArg(v5)
v3.AddArg(v4)
v1.AddArg2(y, v3)
v0.AddArg(v1)
v.AddArg2(y, v0)
return true
}
// match: (Cvt64Fto32 <t> x)
// cond: !base.ConvertHash.MatchPos(v.Pos, nil)
// result: (CVTTSD2SL <t> x)
for {
t := v.Type
x := v_0
if !(!base.ConvertHash.MatchPos(v.Pos, nil)) {
break
}
v.reset(Op386CVTTSD2SL)
v.Type = t
v.AddArg(x)
return true
}
return false
}
func rewriteValue386_OpDiv8(v *Value) bool {
v_1 := v.Args[1]
v_0 := v.Args[0]

View file

@ -538,6 +538,13 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
p.To.Type = obj.TYPE_MEM
p.To.Reg = v.Args[0].Reg()
ssagen.AddAux(&p.To, v)
case ssa.Op386MOVLf2i:
var p *obj.Prog
p = s.Prog(x86.AMOVL)
p.From.Type = obj.TYPE_REG
p.From.Reg = v.Args[0].Reg()
p.To.Type = obj.TYPE_REG
p.To.Reg = v.Reg()
case ssa.Op386ADDLconstmodify:
sc := v.AuxValAndOff()
val := sc.Val()

View file

@ -1407,16 +1407,6 @@ TEXT runtime·uint32tofloat64(SB),NOSPLIT,$8-12
FMOVDP F0, ret+4(FP)
RET
TEXT runtime·float64touint32(SB),NOSPLIT,$12-12
FMOVD a+0(FP), F0
FSTCW 0(SP)
FLDCW runtime·controlWord64trunc(SB)
FMOVVP F0, 4(SP)
FLDCW 0(SP)
MOVL 4(SP), AX
MOVL AX, ret+8(FP)
RET
// gcWriteBarrier informs the GC about heap pointer writes.
//
// gcWriteBarrier returns space in a write barrier buffer which

View file

@ -6,7 +6,6 @@ package runtime
import "unsafe"
func float64touint32(a float64) uint32
func uint32tofloat64(a uint32) float64
// stackcheck checks that SP is in range [g->stack.lo, g->stack.hi).

View file

@ -40,10 +40,17 @@ func float64toint64(d float64) (y uint64) {
}
func float64touint64(d float64) (y uint64) {
_d2v(&y, d)
_d2vu(&y, d)
return
}
func float64touint32(a float64) uint32 {
if a >= 0xffffffff {
return 0xffffffff
}
return uint32(float64touint64(a))
}
func int64tofloat64(y int64) float64 {
if y < 0 {
return -uint64tofloat64(-uint64(y))
@ -117,12 +124,16 @@ func _d2v(y *uint64, d float64) {
} else {
/* v = (hi||lo) << -sh */
sh := uint32(-sh)
if sh <= 11 {
if sh <= 10 {
ylo = xlo << sh
yhi = xhi<<sh | xlo>>(32-sh)
} else {
/* overflow */
yhi = uint32(d) /* causes something awful */
if x&sign64 != 0 {
*y = 0x8000000000000000
} else {
*y = 0x7fffffffffffffff
}
return
}
}
if x&sign64 != 0 {
@ -136,6 +147,50 @@ func _d2v(y *uint64, d float64) {
*y = uint64(yhi)<<32 | uint64(ylo)
}
func _d2vu(y *uint64, d float64) {
x := *(*uint64)(unsafe.Pointer(&d))
if x&sign64 != 0 {
*y = 0
return
}
xhi := uint32(x>>32)&0xfffff | 0x100000
xlo := uint32(x)
sh := 1075 - int32(uint32(x>>52)&0x7ff)
var ylo, yhi uint32
if sh >= 0 {
sh := uint32(sh)
/* v = (hi||lo) >> sh */
if sh < 32 {
if sh == 0 {
ylo = xlo
yhi = xhi
} else {
ylo = xlo>>sh | xhi<<(32-sh)
yhi = xhi >> sh
}
} else {
if sh == 32 {
ylo = xhi
} else if sh < 64 {
ylo = xhi >> (sh - 32)
}
}
} else {
/* v = (hi||lo) << -sh */
sh := uint32(-sh)
if sh <= 11 {
ylo = xlo << sh
yhi = xhi<<sh | xlo>>(32-sh)
} else {
/* overflow */
*y = 0xffffffffffffffff
return
}
}
*y = uint64(yhi)<<32 | uint64(ylo)
}
func uint64div(n, d uint64) uint64 {
// Check for 32 bit operands
if uint32(n>>32) == 0 && uint32(d>>32) == 0 {

View file

@ -4,9 +4,9 @@
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
//go:build !wasm && !386 && !arm && !mips
//go:build !wasm
// TODO fix this to work for wasm and 32-bit architectures.
// TODO fix this to work for wasm
// Doing more than this, however, expands the change.
package main