cmd/compile/internal/amd64: break dependency for CVTS[LQ]2S[DS] (8589049d) · Commits · Kirill Smelkov / go

Commit 8589049d authored Oct 19, 2016 by

Ilya Tocar

cmd/compile/internal/amd64: break dependency for CVTS[LQ]2S[DS]

CVTSL2SS, CVTSQ2SS, CVTSL2SD, CVTSQ2SD preserve upper part of xmm register,
introducing false dependency on a previous value.
Break it by xoring destination with itself.
Increases size of go executable by 320 bytes, but shows nice improvement on go1.
Also fixes performance degradation introduced by 1.7.

name old time/op new time/op delta
BinaryTree17-4 2.20s ± 1% 2.19s ± 0% -0.36% (p=0.000 n=18+16)
Fannkuch11-4 2.44s ± 1% 2.45s ± 2% +0.47% (p=0.030 n=20+20)
FmtFprintfEmpty-4 40.9ns ± 7% 40.5ns ± 1% ~ (p=0.531 n=20+16)
FmtFprintfString-4 111ns ± 2% 111ns ± 1% ~ (p=0.510 n=18+19)
FmtFprintfInt-4 98.3ns ± 3% 99.3ns ± 1% +1.01% (p=0.003 n=20+18)
FmtFprintfIntInt-4 148ns ± 3% 147ns ± 1% ~ (p=0.919 n=20+17)
FmtFprintfPrefixedInt-4 149ns ± 1% 152ns ± 0% +1.73% (p=0.000 n=19+17)
FmtFprintfFloat-4 231ns ± 0% 231ns ± 1% ~ (p=0.678 n=18+19)
FmtManyArgs-4 667ns ± 1% 672ns ± 1% +0.73% (p=0.005 n=20+20)
GobDecode-4 5.60ms ± 0% 5.61ms ± 0% +0.24% (p=0.000 n=20+20)
GobEncode-4 4.74ms ± 0% 4.73ms ± 1% -0.20% (p=0.002 n=20+20)
Gzip-4 199ms ± 0% 199ms ± 1% +0.35% (p=0.000 n=19+20)
Gunzip-4 31.8ms ± 1% 31.5ms ± 1% -0.89% (p=0.000 n=20+20)
HTTPClientServer-4 38.1µs ± 1% 38.0µs ± 1% ~ (p=0.117 n=19+18)
JSONEncode-4 14.2ms ± 1% 13.4ms ± 0% -5.73% (p=0.000 n=20+20)
JSONDecode-4 42.7ms ± 0% 42.7ms ± 1% +0.18% (p=0.019 n=18+19)
Mandelbrot200-4 3.26ms ± 0% 2.99ms ± 0% -8.38% (p=0.000 n=19+19)
GoParse-4 2.76ms ± 1% 2.76ms ± 1% ~ (p=0.583 n=20+20)
RegexpMatchEasy0_32-4 69.5ns ± 0% 69.6ns ± 0% +0.10% (p=0.017 n=16+17)
RegexpMatchEasy0_1K-4 703ns ± 0% 708ns ± 3% +0.65% (p=0.000 n=17+18)
RegexpMatchEasy1_32-4 68.2ns ± 1% 68.2ns ± 2% ~ (p=0.094 n=18+20)
RegexpMatchEasy1_1K-4 288ns ± 1% 288ns ± 0% ~ (p=0.403 n=17+18)
RegexpMatchMedium_32-4 104ns ± 2% 103ns ± 1% ~ (p=0.110 n=20+16)
RegexpMatchMedium_1K-4 31.7µs ± 3% 31.7µs ± 3% ~ (p=0.091 n=19+20)
RegexpMatchHard_32-4 1.59µs ± 2% 1.58µs ± 2% ~ (p=0.083 n=20+20)
RegexpMatchHard_1K-4 48.1µs ± 3% 47.9µs ± 2% ~ (p=0.461 n=20+19)
Revcomp-4 344ms ± 0% 345ms ± 0% +0.08% (p=0.009 n=18+17)
Template-4 44.8ms ± 1% 44.7ms ± 1% ~ (p=0.277 n=20+20)
TimeParse-4 258ns ± 0% 258ns ± 0% ~ (all samples are equal)
TimeFormat-4 275ns ± 0% 273ns ± 0% -0.64% (p=0.000 n=20+18)

name old speed new speed delta
GobDecode-4 137MB/s ± 0% 137MB/s ± 0% -0.24% (p=0.000 n=20+20)
GobEncode-4 162MB/s ± 0% 162MB/s ± 0% +0.20% (p=0.002 n=20+20)
Gzip-4 97.6MB/s ± 0% 97.3MB/s ± 1% -0.35% (p=0.000 n=19+20)
Gunzip-4 610MB/s ± 1% 615MB/s ± 1% +0.89% (p=0.000 n=20+20)
JSONEncode-4 136MB/s ± 1% 145MB/s ± 0% +6.08% (p=0.000 n=20+20)
JSONDecode-4 45.5MB/s ± 0% 45.4MB/s ± 1% -0.17% (p=0.017 n=18+19)
GoParse-4 21.0MB/s ± 1% 21.0MB/s ± 1% ~ (p=0.578 n=20+20)
RegexpMatchEasy0_32-4 460MB/s ± 0% 460MB/s ± 0% -0.09% (p=0.031 n=16+17)
RegexpMatchEasy0_1K-4 1.46GB/s ± 0% 1.45GB/s ± 3% -0.64% (p=0.000 n=17+18)
RegexpMatchEasy1_32-4 469MB/s ± 0% 469MB/s ± 2% +0.06% (p=0.043 n=18+20)
RegexpMatchEasy1_1K-4 3.55GB/s ± 1% 3.55GB/s ± 0% ~ (p=0.057 n=17+18)
RegexpMatchMedium_32-4 9.61MB/s ± 2% 9.64MB/s ± 2% ~ (p=0.856 n=20+20)
RegexpMatchMedium_1K-4 32.3MB/s ± 3% 32.3MB/s ± 3% ~ (p=0.085 n=19+20)
RegexpMatchHard_32-4 20.1MB/s ± 2% 20.2MB/s ± 2% ~ (p=0.086 n=20+20)
RegexpMatchHard_1K-4 21.3MB/s ± 3% 21.4MB/s ± 2% ~ (p=0.578 n=20+20)
Revcomp-4 738MB/s ± 0% 737MB/s ± 0% -0.08% (p=0.009 n=18+17)
Template-4 43.3MB/s ± 1% 43.4MB/s ± 1% ~ (p=0.274 n=20+20)

Fixes #16982

Change-Id: If574d66f39f4183a9b1d5ffff0339909cc73f59d
Reviewed-on: https://go-review.googlesource.com/31490
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

parent 5ce715fd

Hide whitespace changes

Inline Side-by-side

View file @ 8589049d

@@ -637,10 +637,14 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p.To.Index = i
 		gc.AddAux2(&p.To, v, sc.Off())
 	case ssa.OpAMD64MOVLQSX, ssa.OpAMD64MOVWQSX, ssa.OpAMD64MOVBQSX, ssa.OpAMD64MOVLQZX, ssa.OpAMD64MOVWQZX, ssa.OpAMD64MOVBQZX,
 		ssa.OpAMD64CVTSL2SS, ssa.OpAMD64CVTSL2SD, ssa.OpAMD64CVTSQ2SS, ssa.OpAMD64CVTSQ2SD,
 		ssa.OpAMD64CVTTSS2SL, ssa.OpAMD64CVTTSD2SL, ssa.OpAMD64CVTTSS2SQ, ssa.OpAMD64CVTTSD2SQ,
 		ssa.OpAMD64CVTSS2SD, ssa.OpAMD64CVTSD2SS:
 		opregreg(v.Op.Asm(), v.Reg(), v.Args[0].Reg())
 	case ssa.OpAMD64CVTSL2SD, ssa.OpAMD64CVTSQ2SD, ssa.OpAMD64CVTSQ2SS, ssa.OpAMD64CVTSL2SS:
 		r := v.Reg()
 		// Break false dependency on destination register.
 		opregreg(x86.AXORPS, r, r)
 		opregreg(v.Op.Asm(), r, v.Args[0].Reg())
 	case ssa.OpAMD64DUFFZERO:
 		off := duffStart(v.AuxInt)
 		adj := duffAdj(v.AuxInt)
-...

Please register or to comment