cmd/compile: hoist some loop invariants #59194

y1yang0 · 2023-03-23T09:25:49Z

This patch tries to conservatively hoist profitable loop invariant to block that dominates the entire loop. Value is considered as loop invariant if all its inputs are defined outside the loop or all its inputs are loop invariants. Since loop invariant will immediately moved to dominator block of loop, the first rule actually already implies the second rule.

The following limitations make this optimization tend to be conservative, it could be more aggressive in the future.

Only short loop(8 blocks) is considered an optimization candidate, this slightly reduces register pressure.
Give up further lifting on blocks when they contain Calls and Load/Store(This is too conservative) because they kill all/ part of memory and are too complicated.
Constant values and values from unlikely branch is not considered as a candidate. The former is obvious, but the latter is debatable, because it seems that likelyadjust will further optimize the likeness, and if a certain value is located in a deep inner loop, it may not be restricted by likeness, allowing it to be an optimization candidate

The ad-hoc and go1 benchmark show slight improvements:
darwin + amd64 +cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
│ old.log │ new.log │
│ sec/op │ sec/op vs base │
BinaryTree17-12 2.306 ± 10% 2.313 ± 6% ~ (p=0.912 n=10)
Fannkuch11-12 2.005 ± 7% 2.002 ± 2% ~ (p=0.853 n=10)
FmtFprintfEmpty-12 23.89n ± 10% 24.52n ± 9% ~ (p=0.393 n=10)
FmtFprintfString-12 40.36n ± 11% 43.68n ± 27% +8.25% (p=0.035 n=10)
FmtFprintfInt-12 49.09n ± 10% 49.88n ± 7% ~ (p=0.796 n=10)
FmtFprintfIntInt-12 81.09n ± 32% 78.13n ± 5% ~ (p=0.052 n=10)
FmtFprintfPrefixedInt-12 85.78n ± 3% 86.34n ± 1% ~ (p=0.955 n=10)
FmtFprintfFloat-12 132.9n ± 7% 128.0n ± 3% -3.69% (p=0.045 n=10)
FmtManyArgs-12 320.7n ± 33% 342.6n ± 19% +6.84% (p=0.042 n=10)
GobDecode-12 4.052m ± 13% 3.984m ± 8% ~ (p=0.579 n=10)
GobEncode-12 2.831m ± 20% 2.691m ± 6% -4.95% (p=0.043 n=10)
Gzip-12 183.5m ± 1% 190.1m ± 6% ~ (p=0.063 n=10)
Gunzip-12 21.17m ± 13% 21.22m ± 17% ~ (p=1.000 n=10)
HTTPClientServer-12 91.00µ ± 22% 78.75µ ± 1% -13.47% (p=0.000 n=10)
JSONEncode-12 5.527m ± 1% 5.575m ± 5% ~ (p=0.631 n=10)
JSONDecode-12 26.01m ± 1% 26.02m ± 5% ~ (p=0.684 n=10)
Mandelbrot200-12 3.188m ± 9% 3.333m ± 4% ~ (p=0.218 n=10)
GoParse-12 2.276m ± 10% 2.281m ± 5% ~ (p=0.739 n=10)
RegexpMatchEasy0_32-12 42.69n ± 7% 39.21n ± 5% -8.15% (p=0.009 n=10)
RegexpMatchEasy0_1K-12 128.5n ± 4% 127.9n ± 12% ~ (p=0.926 n=10)
RegexpMatchEasy1_32-12 36.29n ± 6% 36.19n ± 2% ~ (p=0.724 n=10)
RegexpMatchEasy1_1K-12 195.8n ± 16% 191.0n ± 10% ~ (p=0.739 n=10)
RegexpMatchMedium_32-12 600.1n ± 2% 629.7n ± 1% +4.93% (p=0.002 n=10)
RegexpMatchMedium_1K-12 20.81µ ± 3% 20.74µ ± 8% ~ (p=0.971 n=10)
RegexpMatchHard_32-12 915.1n ± 2% 910.1n ± 4% ~ (p=1.000 n=10)
RegexpMatchHard_1K-12 28.35µ ± 11% 27.26µ ± 3% -3.83% (p=0.003 n=10)
Revcomp-12 290.9m ± 15% 279.2m ± 11% ~ (p=0.353 n=10)
Template-12 36.29m ± 16% 34.43m ± 2% -5.13% (p=0.043 n=10)
TimeParse-12 199.4n ± 40% 187.4n ± 42% ~ (p=0.579 n=10)
TimeFormat-12 168.2n ± 10% 178.6n ± 42% ~ (p=0.289 n=10)
geomean 31.84µ 31.62µ -0.70%
linux + amd64 + cpu: Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz
│ old.log │ new.log │
│ sec/op │ sec/op vs base │
BinaryTree17-96 5.097 ± 3% 5.159 ± 2% ~ (p=0.315 n=10)
Fannkuch11-96 3.189 ± 0% 3.191 ± 1% +0.07% (p=0.023 n=10)
FmtFprintfEmpty-96 38.04n ± 2% 37.44n ± 1% -1.58% (p=0.003 n=10)
FmtFprintfString-96 66.65n ± 4% 63.49n ± 1% -4.74% (p=0.000 n=10)
FmtFprintfInt-96 81.01n ± 2% 80.96n ± 1% ~ (p=0.493 n=10)
FmtFprintfIntInt-96 118.7n ± 1% 119.0n ± 2% ~ (p=0.614 n=10)
FmtFprintfPrefixedInt-96 144.0n ± 1% 142.6n ± 1% ~ (p=0.169 n=10)
FmtFprintfFloat-96 203.9n ± 1% 201.0n ± 1% -1.42% (p=0.004 n=10)
FmtManyArgs-96 505.2n ± 1% 506.9n ± 1% ~ (p=0.101 n=10)
GobDecode-96 6.602m ± 3% 6.675m ± 3% ~ (p=0.631 n=10)
GobEncode-96 5.847m ± 3% 5.808m ± 3% ~ (p=0.796 n=10)
Gzip-96 284.2m ± 1% 284.8m ± 4% ~ (p=0.247 n=10)
Gunzip-96 33.25m ± 2% 33.28m ± 1% ~ (p=0.971 n=10)
HTTPClientServer-96 64.59µ ± 1% 64.10µ ± 3% -0.76% (p=0.023 n=10)
JSONEncode-96 9.262m ± 4% 9.617m ± 3% +3.83% (p=0.043 n=10)
JSONDecode-96 41.17m ± 1% 41.17m ± 1% ~ (p=0.631 n=10)
Mandelbrot200-96 5.063m ± 1% 5.059m ± 0% ~ (p=0.280 n=10)
GoParse-96 4.374m ± 3% 4.284m ± 1% ~ (p=0.052 n=10)
RegexpMatchEasy0_32-96 61.80n ± 2% 61.31n ± 2% ~ (p=0.363 n=10)
RegexpMatchEasy0_1K-96 187.1n ± 1% 189.9n ± 2% ~ (p=0.065 n=10)
RegexpMatchEasy1_32-96 55.92n ± 3% 55.05n ± 2% -1.54% (p=0.023 n=10)
RegexpMatchEasy1_1K-96 295.4n ± 1% 294.6n ± 1% ~ (p=0.160 n=10)
RegexpMatchMedium_32-96 1.023µ ± 1% 1.016µ ± 1% ~ (p=0.091 n=10)
RegexpMatchMedium_1K-96 31.80µ ± 1% 31.67µ ± 1% ~ (p=0.190 n=10)
RegexpMatchHard_32-96 1.450µ ± 1% 1.450µ ± 1% ~ (p=0.925 n=10)
RegexpMatchHard_1K-96 44.25µ ± 4% 43.88µ ± 1% ~ (p=0.190 n=10)
Revcomp-96 500.1m ± 7% 464.4m ± 10% ~ (p=0.315 n=10)
Template-96 58.88m ± 4% 58.86m ± 6% ~ (p=0.853 n=10)
TimeParse-96 252.0n ± 5% 251.9n ± 1% ~ (p=0.537 n=10)
TimeFormat-96 272.0n ± 2% 274.8n ± 2% ~ (p=0.147 n=10)
geomean 50.02µ 49.77µ -0.50%

Updates #15808

Conservatively hoist some loop invariants outside the loop. Updates golang#15808

gopherbot · 2023-03-23T09:29:52Z

This PR (HEAD: 0672dff) has been imported to Gerrit for code review.

Please visit https://go-review.googlesource.com/c/go/+/478815 to see it.

Tip: You can toggle comments from me using the comments slash command (e.g. /comments off)
See the Wiki page for more info

gopherbot · 2023-03-24T09:02:44Z

Message from Iskander Sharipov:

Patch Set 6:

(1 comment)