mirror of
https://github.com/golang/go
synced 2024-11-17 13:14:56 -07:00
cmd/compile/internal/ssa: tweak branchelim cost model on amd64
Currently branchelim is too aggressive in converting branches to conditinal movs. On most x86 cpus resulting cmov* are more expensive than most simple instructions, because they have a latency of 2, instead of 1, So by teaching branchelim to limit number of CondSelects and consider possible need to recalculate flags, we can archive huge speed-ups (fix big regressions). In package strings: ToUpper/#00-6 10.9ns ± 1% 11.8ns ± 1% +8.29% (p=0.000 n=10+10) ToUpper/ONLYUPPER-6 27.9ns ± 0% 27.8ns ± 1% ~ (p=0.106 n=9+10) ToUpper/abc-6 90.3ns ± 2% 90.3ns ± 1% ~ (p=0.956 n=10+10) ToUpper/AbC123-6 110ns ± 1% 113ns ± 2% +3.19% (p=0.000 n=10+10) ToUpper/azAZ09_-6 109ns ± 2% 110ns ± 1% ~ (p=0.174 n=10+10) ToUpper/longStrinGwitHmixofsmaLLandcAps-6 228ns ± 1% 233ns ± 2% +2.11% (p=0.000 n=10+10) ToUpper/longɐstringɐwithɐnonasciiⱯchars-6 907ns ± 1% 709ns ± 2% -21.85% (p=0.000 n=10+10) ToUpper/ɐɐɐɐɐ-6 793ns ± 2% 562ns ± 2% -29.06% (p=0.000 n=10+10) In fmt: SprintfQuoteString-6 272ns ± 2% 195ns ± 4% -28.39% (p=0.000 n=10+10) And in archive/zip: CompressedZipGarbage-6 4.00ms ± 0% 4.03ms ± 0% +0.71% (p=0.000 n=10+10) Zip64Test-6 27.5ms ± 1% 24.2ms ± 0% -12.01% (p=0.000 n=10+10) Zip64TestSizes/4096-6 10.4µs ±12% 10.7µs ± 8% ~ (p=0.068 n=10+8) Zip64TestSizes/1048576-6 79.0µs ± 3% 70.2µs ± 2% -11.14% (p=0.000 n=10+10) Zip64TestSizes/67108864-6 4.64ms ± 1% 4.11ms ± 1% -11.43% (p=0.000 n=10+10) As far as I can tell, no cases with significant gain from cmov have regressed. On go1 it looks like most changes are unrelated, but I've verified that TimeFormat really switched from cmov to branch in a hot spot. Fill results below: name old time/op new time/op delta BinaryTree17-6 4.42s ± 1% 4.44s ± 1% ~ (p=0.075 n=10+10) Fannkuch11-6 4.23s ± 0% 4.18s ± 0% -1.16% (p=0.000 n=8+10) FmtFprintfEmpty-6 67.5ns ± 2% 67.5ns ± 0% ~ (p=0.950 n=10+7) FmtFprintfString-6 117ns ± 2% 119ns ± 1% +1.07% (p=0.045 n=9+10) FmtFprintfInt-6 122ns ± 0% 123ns ± 2% ~ (p=0.825 n=8+10) FmtFprintfIntInt-6 188ns ± 1% 187ns ± 1% -0.85% (p=0.001 n=10+10) FmtFprintfPrefixedInt-6 223ns ± 1% 226ns ± 1% +1.40% (p=0.000 n=9+10) FmtFprintfFloat-6 380ns ± 1% 379ns ± 0% ~ (p=0.350 n=9+7) FmtManyArgs-6 784ns ± 0% 790ns ± 1% +0.81% (p=0.000 n=10+9) GobDecode-6 10.7ms ± 1% 10.8ms ± 0% +0.68% (p=0.000 n=10+10) GobEncode-6 8.95ms ± 0% 8.94ms ± 0% ~ (p=0.436 n=10+10) Gzip-6 378ms ± 0% 378ms ± 0% ~ (p=0.696 n=8+10) Gunzip-6 60.5ms ± 0% 60.9ms ± 0% +0.73% (p=0.000 n=8+10) HTTPClientServer-6 109µs ± 3% 111µs ± 2% +2.53% (p=0.000 n=10+10) JSONEncode-6 20.2ms ± 0% 20.2ms ± 0% ~ (p=0.382 n=8+8) JSONDecode-6 85.9ms ± 1% 84.5ms ± 0% -1.59% (p=0.000 n=10+8) Mandelbrot200-6 6.89ms ± 0% 6.85ms ± 1% -0.49% (p=0.000 n=9+10) GoParse-6 5.49ms ± 0% 5.40ms ± 0% -1.63% (p=0.000 n=8+10) RegexpMatchEasy0_32-6 126ns ± 1% 129ns ± 1% +2.14% (p=0.000 n=10+10) RegexpMatchEasy0_1K-6 320ns ± 1% 317ns ± 2% ~ (p=0.089 n=10+10) RegexpMatchEasy1_32-6 119ns ± 2% 121ns ± 4% ~ (p=0.591 n=10+10) RegexpMatchEasy1_1K-6 544ns ± 1% 541ns ± 0% -0.67% (p=0.020 n=8+8) RegexpMatchMedium_32-6 184ns ± 1% 184ns ± 1% ~ (p=0.360 n=10+10) RegexpMatchMedium_1K-6 57.7µs ± 2% 58.3µs ± 1% +1.12% (p=0.022 n=10+10) RegexpMatchHard_32-6 2.72µs ± 5% 2.70µs ± 0% ~ (p=0.166 n=10+8) RegexpMatchHard_1K-6 80.2µs ± 0% 81.0µs ± 0% +1.01% (p=0.000 n=8+10) Revcomp-6 607ms ± 0% 601ms ± 2% -1.00% (p=0.006 n=8+10) Template-6 93.1ms ± 1% 92.6ms ± 0% -0.56% (p=0.000 n=9+9) TimeParse-6 472ns ± 0% 470ns ± 0% -0.28% (p=0.001 n=9+9) TimeFormat-6 546ns ± 0% 511ns ± 0% -6.41% (p=0.001 n=7+7) [Geo mean] 76.4µs 76.3µs -0.12% name old speed new speed delta GobDecode-6 71.5MB/s ± 1% 71.1MB/s ± 0% -0.68% (p=0.000 n=10+10) GobEncode-6 85.8MB/s ± 0% 85.8MB/s ± 0% ~ (p=0.425 n=10+10) Gzip-6 51.3MB/s ± 0% 51.3MB/s ± 0% ~ (p=0.680 n=8+10) Gunzip-6 321MB/s ± 0% 318MB/s ± 0% -0.73% (p=0.000 n=8+10) JSONEncode-6 95.9MB/s ± 0% 96.0MB/s ± 0% ~ (p=0.367 n=8+8) JSONDecode-6 22.6MB/s ± 1% 22.9MB/s ± 0% +1.62% (p=0.000 n=10+8) GoParse-6 10.6MB/s ± 0% 10.7MB/s ± 0% +1.64% (p=0.000 n=8+10) RegexpMatchEasy0_32-6 252MB/s ± 1% 247MB/s ± 1% -2.22% (p=0.000 n=10+10) RegexpMatchEasy0_1K-6 3.19GB/s ± 1% 3.22GB/s ± 2% ~ (p=0.105 n=10+10) RegexpMatchEasy1_32-6 267MB/s ± 2% 264MB/s ± 4% ~ (p=0.481 n=10+10) RegexpMatchEasy1_1K-6 1.88GB/s ± 1% 1.89GB/s ± 0% +0.62% (p=0.038 n=8+8) RegexpMatchMedium_32-6 5.41MB/s ± 2% 5.43MB/s ± 0% ~ (p=0.339 n=10+8) RegexpMatchMedium_1K-6 17.8MB/s ± 1% 17.6MB/s ± 1% -1.12% (p=0.029 n=10+10) RegexpMatchHard_32-6 11.8MB/s ± 5% 11.8MB/s ± 0% ~ (p=0.163 n=10+8) RegexpMatchHard_1K-6 12.8MB/s ± 0% 12.6MB/s ± 0% -1.06% (p=0.000 n=7+10) Revcomp-6 419MB/s ± 0% 423MB/s ± 2% +1.02% (p=0.006 n=8+10) Template-6 20.9MB/s ± 0% 21.0MB/s ± 0% +0.53% (p=0.000 n=8+8) [Geo mean] 77.0MB/s 77.0MB/s +0.05% diff --git a/src/cmd/compile/internal/ssa/branchelim.go b/src/cmd/compile/internal/ssa/branchelim.go Change-Id: Ibdffa9ea9b4c72668617ce3202ec4a83a1cd59be Reviewed-on: https://go-review.googlesource.com/107936 Run-TryBot: Ilya Tocar <ilya.tocar@intel.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Keith Randall <khr@golang.org>
This commit is contained in:
Ilya Tocar
parent
01a7487534
commit
25813f9f91
@ -193,6 +193,11 @@ func elimIfElse(f *Func, b *Block) bool {
|
||||
return false
|
||||
}
|
||||
|
||||
// Don't generate CondSelects if branch is cheaper.
|
||||
if !shouldElimIfElse(no, yes, post, f.Config.arch) {
|
||||
return false
|
||||
}
|
||||
|
||||
// now we're committed: rewrite each Phi as a CondSelect
|
||||
swap := post.Preds[0].Block() != b.Succs[0].Block()
|
||||
for _, v := range post.Values {
|
||||
@ -238,6 +243,39 @@ func elimIfElse(f *Func, b *Block) bool {
|
||||
return true
|
||||
}
|
||||
|
||||
// shouldElimIfElse reports whether estimated cost of eliminating branch
|
||||
// is lower than threshold.
|
||||
func shouldElimIfElse(no, yes, post *Block, arch string) bool {
|
||||
switch arch {
|
||||
default:
|
||||
return true
|
||||
case "amd64":
|
||||
const maxcost = 2
|
||||
phi := 0
|
||||
other := 0
|
||||
for _, v := range post.Values {
|
||||
if v.Op == OpPhi {
|
||||
// Each phi results in CondSelect, which lowers into CMOV,
|
||||
// CMOV has latency >1 on most CPUs.
|
||||
phi++
|
||||
}
|
||||
for _, x := range v.Args {
|
||||
if x.Block == no || x.Block == yes {
|
||||
other++
|
||||
}
|
||||
}
|
||||
}
|
||||
cost := phi * 1
|
||||
if phi > 1 {
|
||||
// If we have more than 1 phi and some values in post have args
|
||||
// in yes or no blocks, we may have to recalucalte condition, because
|
||||
// those args may clobber flags. For now assume that all operations clobber flags.
|
||||
cost += other * 1
|
||||
}
|
||||
return cost < maxcost
|
||||
}
|
||||
}
|
||||
|
||||
func allTrivial(b *Block) bool {
|
||||
// don't fuse memory ops, Phi ops, divides (can panic),
|
||||
// or anything else with side-effects
|
||||
|
||||
Loading…
Reference in New Issue
Block a user