languages/go
languages/go
1
0
Fork 0
mirror of https://github.com/golang/go synced 2024-09-05 07:54:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

[dev.regabi] cmd/compile: make ordering for InvertFlags more stable

Browse source 
Current many architectures use a rule along the lines of

// Canonicalize the order of arguments to comparisons - helps with CSE.
((CMP|CMPW) x y) && x.ID > y.ID => (InvertFlags ((CMP|CMPW) y x))

to normalize comparisons as much as possible for CSE.  Replace the
ID comparison with something less variable across compiler changes.
This helps avoid spurious failures in some of the codegen-comparison
tests (though the current choice of comparison is sensitive to Op
ordering).

Two tests changed to accommodate modified instruction choice.

Change-Id: Ib35f450bd2bae9d4f9f7838ceaf7ec682bcf1e1a
Reviewed-on: https://go-review.googlesource.com/c/go/+/280155
Trust: David Chase <drchase@google.com>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
This commit is contained in:
David Chase 2020-12-30 12:05:57 -05:00
parent d9acf6f3a3
commit 9a19481acb
15 changed files with 59 additions and 47 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/cmd/compile/internal/ssa/gen/386.rules
View file

@ -475,7 +475,7 @@
(CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
// Canonicalize the order of arguments to comparisons - helps with CSE.
(CMP(L|W|B) x y) && x.ID > y.ID => (InvertFlags (CMP(L|W|B) y x))
(CMP(L|W|B) x y) && canonLessThan(x,y) => (InvertFlags (CMP(L|W|B) y x))
// strength reduction
// Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:

2
src/cmd/compile/internal/ssa/gen/AMD64.rules
View file

@ -916,7 +916,7 @@
(CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
// Canonicalize the order of arguments to comparisons - helps with CSE.
(CMP(Q|L|W|B) x y) && x.ID > y.ID => (InvertFlags (CMP(Q|L|W|B) y x))
(CMP(Q|L|W|B) x y) && canonLessThan(x,y) => (InvertFlags (CMP(Q|L|W|B) y x))
// Using MOVZX instead of AND is cheaper.
(AND(Q|L)const [ 0xFF] x) => (MOVBQZX x)

2
src/cmd/compile/internal/ssa/gen/ARM.rules
View file

@ -507,7 +507,7 @@
(TEQ x (MOVWconst [c])) => (TEQconst [c] x)
// Canonicalize the order of arguments to comparisons - helps with CSE.
(CMP x y) && x.ID > y.ID => (InvertFlags (CMP y x))
(CMP x y) && canonLessThan(x,y) => (InvertFlags (CMP y x))
// don't extend after proper load
// MOVWreg instruction is not emitted if src and dst registers are same, but it ensures the type.

2
src/cmd/compile/internal/ssa/gen/ARM64.rules
View file

@ -1151,7 +1151,7 @@
(CMPW (MOVDconst [c]) x) => (InvertFlags (CMPWconst [int32(c)] x))
// Canonicalize the order of arguments to comparisons - helps with CSE.
((CMP|CMPW) x y) && x.ID > y.ID => (InvertFlags ((CMP|CMPW) y x))
((CMP|CMPW) x y) && canonLessThan(x,y) => (InvertFlags ((CMP|CMPW) y x))
// mul-neg => mneg
(NEG (MUL x y)) => (MNEG x y)

2
src/cmd/compile/internal/ssa/gen/PPC64.rules
View file

@ -1088,7 +1088,7 @@
(CMPWU (MOVDconst [c]) y) && isU16Bit(c) => (InvertFlags (CMPWUconst y [int32(c)]))
// Canonicalize the order of arguments to comparisons - helps with CSE.
((CMP|CMPW|CMPU|CMPWU) x y) && x.ID > y.ID => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x))
((CMP|CMPW|CMPU|CMPWU) x y) && canonLessThan(x,y) => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x))
// ISEL auxInt values 0=LT 1=GT 2=EQ arg2 ? arg0 : arg1
// ISEL auxInt values 4=GE 5=LE 6=NE arg2 ? arg1 : arg0

2
src/cmd/compile/internal/ssa/gen/S390X.rules
View file

@ -785,7 +785,7 @@
=> (RISBGZ x {s390x.NewRotateParams(r.Start, r.Start, -r.Start&63)})
// Canonicalize the order of arguments to comparisons - helps with CSE.
((CMP|CMPW|CMPU|CMPWU) x y) && x.ID > y.ID => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x))
((CMP|CMPW|CMPU|CMPWU) x y) && canonLessThan(x,y) => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x))
// Use sign/zero extend instead of RISBGZ.
(RISBGZ x {r}) && r == s390x.NewRotateParams(56, 63, 0) => (MOVBZreg x)

12
src/cmd/compile/internal/ssa/rewrite.go
View file

@ -521,6 +521,18 @@ func shiftIsBounded(v *Value) bool {
return v.AuxInt != 0
}
// canonLessThan returns whether x is "ordered" less than y, for purposes of normalizing
// generated code as much as possible.
func canonLessThan(x, y *Value) bool {
if x.Op != y.Op {
return x.Op < y.Op
}
if !x.Pos.SameFileAndLine(y.Pos) {
return x.Pos.Before(y.Pos)
}
return x.ID < y.ID
}
// truncate64Fto32F converts a float64 value to a float32 preserving the bit pattern
// of the mantissa. It will panic if the truncation results in lost information.
func truncate64Fto32F(f float64) float32 {

12
src/cmd/compile/internal/ssa/rewrite386.go
View file

@ -1785,12 +1785,12 @@ func rewriteValue386_Op386CMPB(v *Value) bool {
return true
}
// match: (CMPB x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPB y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(Op386InvertFlags)
@ -2078,12 +2078,12 @@ func rewriteValue386_Op386CMPL(v *Value) bool {
return true
}
// match: (CMPL x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPL y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(Op386InvertFlags)
@ -2386,12 +2386,12 @@ func rewriteValue386_Op386CMPW(v *Value) bool {
return true
}
// match: (CMPW x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPW y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(Op386InvertFlags)

16
src/cmd/compile/internal/ssa/rewriteAMD64.go
View file

@ -6749,12 +6749,12 @@ func rewriteValueAMD64_OpAMD64CMPB(v *Value) bool {
return true
}
// match: (CMPB x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPB y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpAMD64InvertFlags)
@ -7135,12 +7135,12 @@ func rewriteValueAMD64_OpAMD64CMPL(v *Value) bool {
return true
}
// match: (CMPL x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPL y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpAMD64InvertFlags)
@ -7544,12 +7544,12 @@ func rewriteValueAMD64_OpAMD64CMPQ(v *Value) bool {
return true
}
// match: (CMPQ x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPQ y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpAMD64InvertFlags)
@ -8106,12 +8106,12 @@ func rewriteValueAMD64_OpAMD64CMPW(v *Value) bool {
return true
}
// match: (CMPW x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPW y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpAMD64InvertFlags)

4
src/cmd/compile/internal/ssa/rewriteARM.go
View file

@ -3728,12 +3728,12 @@ func rewriteValueARM_OpARMCMP(v *Value) bool {
return true
}
// match: (CMP x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMP y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpARMInvertFlags)

8
src/cmd/compile/internal/ssa/rewriteARM64.go
View file

@ -2772,12 +2772,12 @@ func rewriteValueARM64_OpARM64CMP(v *Value) bool {
return true
}
// match: (CMP x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMP y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpARM64InvertFlags)
@ -2941,12 +2941,12 @@ func rewriteValueARM64_OpARM64CMPW(v *Value) bool {
return true
}
// match: (CMPW x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPW y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpARM64InvertFlags)

16
src/cmd/compile/internal/ssa/rewritePPC64.go
View file

@ -4777,12 +4777,12 @@ func rewriteValuePPC64_OpPPC64CMP(v *Value) bool {
return true
}
// match: (CMP x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMP y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpPPC64InvertFlags)
@ -4834,12 +4834,12 @@ func rewriteValuePPC64_OpPPC64CMPU(v *Value) bool {
return true
}
// match: (CMPU x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPU y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpPPC64InvertFlags)
@ -4964,12 +4964,12 @@ func rewriteValuePPC64_OpPPC64CMPW(v *Value) bool {
return true
}
// match: (CMPW x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPW y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpPPC64InvertFlags)
@ -5045,12 +5045,12 @@ func rewriteValuePPC64_OpPPC64CMPWU(v *Value) bool {
return true
}
// match: (CMPWU x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPWU y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpPPC64InvertFlags)

16
src/cmd/compile/internal/ssa/rewriteS390X.go
View file

@ -6332,12 +6332,12 @@ func rewriteValueS390X_OpS390XCMP(v *Value) bool {
return true
}
// match: (CMP x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMP y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpS390XInvertFlags)
@ -6389,12 +6389,12 @@ func rewriteValueS390X_OpS390XCMPU(v *Value) bool {
return true
}
// match: (CMPU x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPU y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpS390XInvertFlags)
@ -6624,12 +6624,12 @@ func rewriteValueS390X_OpS390XCMPW(v *Value) bool {
return true
}
// match: (CMPW x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPW y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpS390XInvertFlags)
@ -6721,12 +6721,12 @@ func rewriteValueS390X_OpS390XCMPWU(v *Value) bool {
return true
}
// match: (CMPWU x y)
// cond: x.ID > y.ID
// cond: canonLessThan(x,y)
// result: (InvertFlags (CMPWU y x))
for {
x := v_0
y := v_1
if !(x.ID > y.ID) {
if !(canonLessThan(x, y)) {
break
}
v.reset(OpS390XInvertFlags)

6
test/codegen/condmove.go
View file

@ -31,7 +31,7 @@ func cmovuintptr(x, y uintptr) uintptr {
if x < y {
x = -y
}
// amd64:"CMOVQCS"
// amd64:"CMOVQ(HI|CS)"
// arm64:"CSEL\t(LO|HI)"
// wasm:"Select"
return x
@ -41,7 +41,7 @@ func cmov32bit(x, y uint32) uint32 {
if x < y {
x = -y
}
// amd64:"CMOVLCS"
// amd64:"CMOVL(HI|CS)"
// arm64:"CSEL\t(LO|HI)"
// wasm:"Select"
return x
@ -51,7 +51,7 @@ func cmov16bit(x, y uint16) uint16 {
if x < y {
x = -y
}
// amd64:"CMOVWCS"
// amd64:"CMOVW(HI|CS)"
// arm64:"CSEL\t(LO|HI)"
// wasm:"Select"
return x

4
test/codegen/spectre.go
View file

@ -13,12 +13,12 @@ func IndexArray(x *[10]int, i int) int {
}
func IndexString(x string, i int) byte {
// amd64:`CMOVQCC`
// amd64:`CMOVQLS`
return x[i]
}
func IndexSlice(x []float64, i int) float64 {
// amd64:`CMOVQCC`
// amd64:`CMOVQLS`
return x[i]
}