runtime: disable prefetching on 386 (0adf6dce) · Commits · go / golang

Commit 0adf6dce authored Oct 17, 2015 by

Shenghou Ma Committed by Minux Ma Nov 13, 2015

runtime: disable prefetching on 386

It doesn't seem to help on modern processors and it makes Go impossible to run
on Pentium MMX (which is the documented minimum hardware requirement.)

Old is with prefetch, new is w/o. Both are compiled with GO386=sse2.
Benchmarking is done on Intel(R) Core(TM) i5-3570K CPU @ 3.40GHz.

name old time/op new time/op delta
BinaryTree17-4 2.89s ± 2% 2.87s ± 0% ~ (p=0.061 n=11+10)
Fannkuch11-4 3.65s ± 0% 3.65s ± 0% ~ (p=0.365 n=11+11)
FmtFprintfEmpty-4 52.1ns ± 0% 52.1ns ± 0% ~ (p=0.065 n=10+9)
FmtFprintfString-4 168ns ± 0% 167ns ± 0% -0.48% (p=0.000 n=8+10)
FmtFprintfInt-4 167ns ± 0% 167ns ± 1% ~ (p=0.591 n=9+10)
FmtFprintfIntInt-4 295ns ± 0% 292ns ± 0% -0.99% (p=0.000 n=9+10)
FmtFprintfPrefixedInt-4 327ns ± 0% 326ns ± 0% -0.24% (p=0.007 n=10+10)
FmtFprintfFloat-4 431ns ± 0% 431ns ± 0% -0.07% (p=0.000 n=10+11)
FmtManyArgs-4 1.13µs ± 0% 1.13µs ± 0% -0.37% (p=0.009 n=11+11)
GobDecode-4 9.36ms ± 1% 9.33ms ± 0% -0.31% (p=0.006 n=11+10)
GobEncode-4 7.38ms ± 1% 7.38ms ± 1% ~ (p=0.797 n=11+11)
Gzip-4 394ms ± 0% 395ms ± 1% ~ (p=0.519 n=11+11)
Gunzip-4 65.4ms ± 0% 65.4ms ± 0% ~ (p=0.739 n=10+10)
HTTPClientServer-4 52.4µs ± 1% 52.5µs ± 1% ~ (p=0.748 n=11+11)
JSONEncode-4 19.0ms ± 0% 19.0ms ± 0% ~ (p=0.780 n=9+10)
JSONDecode-4 59.6ms ± 0% 59.6ms ± 0% ~ (p=0.720 n=9+10)
Mandelbrot200-4 4.09ms ± 0% 4.09ms ± 0% ~ (p=0.295 n=11+9)
GoParse-4 3.45ms ± 1% 3.43ms ± 1% -0.35% (p=0.040 n=11+11)
RegexpMatchEasy0_32-4 101ns ± 1% 101ns ± 1% ~ (p=1.000 n=11+11)
RegexpMatchEasy0_1K-4 796ns ± 0% 796ns ± 0% ~ (p=0.954 n=10+8)
RegexpMatchEasy1_32-4 110ns ± 0% 110ns ± 1% ~ (p=0.289 n=9+11)
RegexpMatchEasy1_1K-4 991ns ± 0% 991ns ± 0% ~ (p=0.784 n=10+8)
RegexpMatchMedium_32-4 131ns ± 0% 130ns ± 0% -0.42% (p=0.004 n=11+9)
RegexpMatchMedium_1K-4 41.9µs ± 1% 41.6µs ± 0% ~ (p=0.067 n=11+9)
RegexpMatchHard_32-4 2.34µs ± 0% 2.34µs ± 0% ~ (p=0.208 n=11+11)
RegexpMatchHard_1K-4 70.9µs ± 0% 71.0µs ± 0% ~ (p=0.968 n=9+10)
Revcomp-4 819ms ± 0% 818ms ± 0% ~ (p=0.251 n=10+11)
Template-4 73.9ms ± 0% 73.8ms ± 0% -0.25% (p=0.013 n=10+11)
TimeParse-4 414ns ± 0% 414ns ± 0% ~ (p=0.809 n=11+10)
TimeFormat-4 485ns ± 0% 485ns ± 0% ~ (p=0.404 n=11+7)

name old speed new speed delta
GobDecode-4 82.0MB/s ± 1% 82.3MB/s ± 0% +0.31% (p=0.007 n=11+10)
GobEncode-4 104MB/s ± 1% 104MB/s ± 1% ~ (p=0.797 n=11+11)
Gzip-4 49.2MB/s ± 0% 49.1MB/s ± 1% ~ (p=0.507 n=11+11)
Gunzip-4 297MB/s ± 0% 297MB/s ± 0% ~ (p=0.670 n=10+10)
JSONEncode-4 102MB/s ± 0% 102MB/s ± 0% ~ (p=0.794 n=9+10)
JSONDecode-4 32.6MB/s ± 0% 32.6MB/s ± 0% ~ (p=0.334 n=9+9)
GoParse-4 16.8MB/s ± 1% 16.9MB/s ± 1% ~ (p=0.052 n=11+11)
RegexpMatchEasy0_32-4 314MB/s ± 0% 314MB/s ± 1% ~ (p=0.618 n=11+11)
RegexpMatchEasy0_1K-4 1.29GB/s ± 0% 1.29GB/s ± 0% ~ (p=0.315 n=10+10)
RegexpMatchEasy1_32-4 290MB/s ± 1% 290MB/s ± 1% ~ (p=0.667 n=10+11)
RegexpMatchEasy1_1K-4 1.03GB/s ± 0% 1.03GB/s ± 0% ~ (p=0.829 n=10+8)
RegexpMatchMedium_32-4 7.63MB/s ± 0% 7.65MB/s ± 0% ~ (p=0.142 n=11+11)
RegexpMatchMedium_1K-4 24.4MB/s ± 1% 24.6MB/s ± 0% ~ (p=0.063 n=11+9)
RegexpMatchHard_32-4 13.7MB/s ± 0% 13.7MB/s ± 0% ~ (p=0.302 n=11+11)
RegexpMatchHard_1K-4 14.4MB/s ± 0% 14.4MB/s ± 0% ~ (p=0.784 n=9+10)
Revcomp-4 310MB/s ± 0% 311MB/s ± 0% ~ (p=0.243 n=10+11)
Template-4 26.2MB/s ± 0% 26.3MB/s ± 0% +0.24% (p=0.009 n=10+11)

Update #12970.

Change-Id: Id185080687a60c229a5cb2e5220e7ca1b53910e2
Reviewed-on: https://go-review.googlesource.com/15999Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>

parent 85ca578d

Hide whitespace changes

Inline Side-by-side

View file @ 0adf6dce

@@ -1530,23 +1530,15 @@ TEXT runtime·goexit(SB),NOSPLIT,$0-0
 	// traceback from goexit1 must hit code range of goexit
 	BYTE	$0x90	// NOP
 // Prefetching doesn't seem to help.
 TEXT runtime·prefetcht0(SB),NOSPLIT,$0-4
 	MOVL	addr+0(FP), AX
 	PREFETCHT0	(AX)
 	RET
 TEXT runtime·prefetcht1(SB),NOSPLIT,$0-4
 	MOVL	addr+0(FP), AX
 	PREFETCHT1	(AX)
 	RET
 TEXT runtime·prefetcht2(SB),NOSPLIT,$0-4
 	MOVL	addr+0(FP), AX
 	PREFETCHT2	(AX)
 	RET
 TEXT runtime·prefetchnta(SB),NOSPLIT,$0-4
 	MOVL	addr+0(FP), AX
 	PREFETCHNTA	(AX)
 	RET

Please register or to comment