cmd/compile: move value around before kick it out of register
When allocating registers, before kicking out the existing value, copy it to a spare register if there is one. So later use of this value can be found in register instead of reload from spill. This is very helpful for instructions of which the input and/or output can only be in specific registers, e.g. DIV on x86, MUL/DIV on MIPS. May also be helpful in general. For "go build -a cmd/go" on AMD64, reduce "spilled value remains" by 1% (not including args, which almost certainly remain). For the code in issue #16061 on AMD64: MaxRem-12 111µs ± 1% 94µs ± 0% -15.38% (p=0.008 n=5+5) Go1 benchmark on AMD64: BinaryTree17-12 2.32s ± 2% 2.30s ± 1% ~ (p=0.421 n=5+5) Fannkuch11-12 2.52s ± 0% 2.44s ± 0% -3.44% (p=0.008 n=5+5) FmtFprintfEmpty-12 39.9ns ± 3% 39.8ns ± 0% ~ (p=0.635 n=5+4) FmtFprintfString-12 114ns ± 1% 113ns ± 1% ~ (p=0.905 n=5+5) FmtFprintfInt-12 102ns ± 6% 98ns ± 1% ~ (p=0.087 n=5+5) FmtFprintfIntInt-12 146ns ± 5% 147ns ± 1% ~ (p=0.238 n=5+5) FmtFprintfPrefixedInt-12 155ns ± 2% 151ns ± 1% -2.58% (p=0.008 n=5+5) FmtFprintfFloat-12 231ns ± 1% 232ns ± 1% ~ (p=0.286 n=5+5) FmtManyArgs-12 657ns ± 1% 649ns ± 0% -1.31% (p=0.008 n=5+5) GobDecode-12 6.35ms ± 0% 6.29ms ± 1% ~ (p=0.056 n=5+5) GobEncode-12 5.38ms ± 1% 5.45ms ± 1% ~ (p=0.056 n=5+5) Gzip-12 209ms ± 0% 209ms ± 1% ~ (p=0.690 n=5+5) Gunzip-12 31.2ms ± 1% 31.1ms ± 1% ~ (p=0.548 n=5+5) HTTPClientServer-12 123µs ± 4% 130µs ± 8% ~ (p=0.151 n=5+5) JSONEncode-12 14.0ms ± 1% 14.0ms ± 1% ~ (p=0.421 n=5+5) JSONDecode-12 41.2ms ± 1% 41.1ms ± 2% ~ (p=0.421 n=5+5) Mandelbrot200-12 3.96ms ± 1% 3.98ms ± 0% ~ (p=0.421 n=5+5) GoParse-12 2.88ms ± 1% 2.88ms ± 1% ~ (p=0.841 n=5+5) RegexpMatchEasy0_32-12 68.0ns ± 3% 66.6ns ± 1% -2.00% (p=0.024 n=5+5) RegexpMatchEasy0_1K-12 728ns ± 8% 682ns ± 1% -6.26% (p=0.008 n=5+5) RegexpMatchEasy1_32-12 66.8ns ± 2% 66.0ns ± 1% ~ (p=0.302 n=5+5) RegexpMatchEasy1_1K-12 291ns ± 2% 288ns ± 1% ~ (p=0.111 n=5+5) RegexpMatchMedium_32-12 103ns ± 2% 100ns ± 0% -2.53% (p=0.016 n=5+4) RegexpMatchMedium_1K-12 31.9µs ± 1% 31.3µs ± 0% -1.75% (p=0.008 n=5+5) RegexpMatchHard_32-12 1.59µs ± 2% 1.59µs ± 1% ~ (p=0.548 n=5+5) RegexpMatchHard_1K-12 48.3µs ± 2% 47.7µs ± 1% ~ (p=0.222 n=5+5) Revcomp-12 340ms ± 1% 338ms ± 1% ~ (p=0.421 n=5+5) Template-12 46.3ms ± 1% 46.5ms ± 1% ~ (p=0.690 n=5+5) TimeParse-12 252ns ± 1% 247ns ± 0% -1.91% (p=0.000 n=5+4) TimeFormat-12 277ns ± 1% 267ns ± 0% -3.82% (p=0.008 n=5+5) [Geo mean] 48.8µs 48.3µs -0.93% It has very little effect on binary size and compiler speed. compilebench: Template 230ms ±10% 231ms ± 8% ~ (p=0.546 n=9+9) Unicode 123ms ± 6% 124ms ± 9% ~ (p=0.481 n=10+10) GoTypes 742ms ± 6% 755ms ± 3% ~ (p=0.123 n=10+10) Compiler 3.10s ± 3% 3.08s ± 1% ~ (p=0.631 n=10+10) Fixes #16061. Change-Id: Id99cdc7a182ee10a704fa0f04e8e0d0809b2ac56 Reviewed-on: https://go-review.googlesource.com/29732 Run-TryBot: Cherry Zhang <cherryyz@google.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
Showing
Please register or sign in to comment