mirror
/
go
mirror of https://github.com/golang/go.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

internal/bytealg, runtime: align some loong64 asm loops to 16-byte boundaries 

The LA464 micro-architecture is very sensitive to alignment of loops,
so the final performance of linked binaries can vary wildly due to
uncontrolled alignment of certain performance-critical loops. Now that
PCALIGN is available on loong64, let's make use of it and manually align
some assembly loops. The functions are identified based on perf records
of some easily regressed go1 benchmark cases (e.g. FmtFprintfPrefixedInt,
RegexpMatchEasy0_1K and Revcomp are particularly sensitive; even those
optimizations purely reducing dynamic instruction counts can regress
those cases by 6~12%, making the numbers almost useless).

Benchmark results on Loongson 3A5000 (which is an LA464 implementation):

goos: linux
goarch: loong64
pkg: test/bench/go1
                      │  CL 416154  │               this CL               │
                      │   sec/op    │   sec/op     vs base                │
BinaryTree17             14.10 ± 1%    14.10 ± 1%        ~ (p=1.000 n=10)
Fannkuch11               3.672 ± 0%    3.579 ± 0%   -2.53% (p=0.000 n=10)
FmtFprintfEmpty         94.72n ± 0%   94.73n ± 0%   +0.01% (p=0.000 n=10)
FmtFprintfString        149.9n ± 0%   151.9n ± 0%   +1.33% (p=0.000 n=10)
FmtFprintfInt           154.1n ± 0%   158.3n ± 0%   +2.73% (p=0.000 n=10)
FmtFprintfIntInt        236.2n ± 0%   241.4n ± 0%   +2.20% (p=0.000 n=10)
FmtFprintfPrefixedInt   314.2n ± 0%   320.2n ± 0%   +1.91% (p=0.000 n=10)
FmtFprintfFloat         405.0n ± 0%   414.3n ± 0%   +2.30% (p=0.000 n=10)
FmtManyArgs             933.6n ± 0%   949.9n ± 0%   +1.75% (p=0.000 n=10)
GobDecode               15.51m ± 1%   15.24m ± 0%   -1.77% (p=0.000 n=10)
GobEncode               18.42m ± 4%   18.10m ± 2%        ~ (p=0.631 n=10)
Gzip                    423.6m ± 0%   429.9m ± 0%   +1.49% (p=0.000 n=10)
Gunzip                  88.75m ± 0%   88.31m ± 0%   -0.50% (p=0.000 n=10)
HTTPClientServer        85.44µ ± 0%   85.71µ ± 0%   +0.31% (p=0.035 n=10)
JSONEncode              18.65m ± 0%   19.74m ± 0%   +5.81% (p=0.000 n=10)
JSONDecode              77.75m ± 0%   78.60m ± 1%   +1.09% (p=0.000 n=10)
Mandelbrot200           7.214m ± 0%   7.208m ± 0%        ~ (p=0.481 n=10)
GoParse                 7.616m ± 2%   7.616m ± 1%        ~ (p=0.739 n=10)
RegexpMatchEasy0_32     142.9n ± 0%   133.0n ± 0%   -6.93% (p=0.000 n=10)
RegexpMatchEasy0_1K     1.535µ ± 0%   1.362µ ± 0%  -11.27% (p=0.000 n=10)
RegexpMatchEasy1_32     161.8n ± 0%   161.8n ± 0%        ~ (p=0.628 n=10)
RegexpMatchEasy1_1K     1.635µ ± 0%   1.497µ ± 0%   -8.41% (p=0.000 n=10)
RegexpMatchMedium_32    1.429µ ± 0%   1.420µ ± 0%   -0.63% (p=0.000 n=10)
RegexpMatchMedium_1K    41.86µ ± 0%   42.25µ ± 0%   +0.93% (p=0.000 n=10)
RegexpMatchHard_32      2.144µ ± 0%   2.108µ ± 0%   -1.68% (p=0.000 n=10)
RegexpMatchHard_1K      63.83µ ± 0%   62.65µ ± 0%   -1.86% (p=0.000 n=10)
Revcomp                  1.337 ± 0%    1.192 ± 0%  -10.89% (p=0.000 n=10)
Template                116.4m ± 1%   115.6m ± 2%        ~ (p=0.579 n=10)
TimeParse               421.4n ± 2%   418.1n ± 1%   -0.78% (p=0.001 n=10)
TimeFormat              515.1n ± 0%   517.9n ± 0%   +0.54% (p=0.001 n=10)
geomean                 104.5µ        103.5µ        -0.99%

                     │  CL 416154   │               this CL                │
                     │     B/s      │     B/s       vs base                │
GobDecode              47.19Mi ± 1%   48.04Mi ± 0%   +1.80% (p=0.000 n=10)
GobEncode              39.73Mi ± 4%   40.44Mi ± 2%        ~ (p=0.631 n=10)
Gzip                   43.68Mi ± 0%   43.04Mi ± 0%   -1.47% (p=0.000 n=10)
Gunzip                 208.5Mi ± 0%   209.6Mi ± 0%   +0.50% (p=0.000 n=10)
JSONEncode             99.21Mi ± 0%   93.76Mi ± 0%   -5.49% (p=0.000 n=10)
JSONDecode             23.80Mi ± 0%   23.55Mi ± 1%   -1.08% (p=0.000 n=10)
GoParse                7.253Mi ± 2%   7.253Mi ± 1%        ~ (p=0.810 n=10)
RegexpMatchEasy0_32    213.6Mi ± 0%   229.4Mi ± 0%   +7.41% (p=0.000 n=10)
RegexpMatchEasy0_1K    636.3Mi ± 0%   717.3Mi ± 0%  +12.73% (p=0.000 n=10)
RegexpMatchEasy1_32    188.6Mi ± 0%   188.6Mi ± 0%        ~ (p=0.810 n=10)
RegexpMatchEasy1_1K    597.4Mi ± 0%   652.2Mi ± 0%   +9.17% (p=0.000 n=10)
RegexpMatchMedium_32   21.35Mi ± 0%   21.49Mi ± 0%   +0.63% (p=0.000 n=10)
RegexpMatchMedium_1K   23.33Mi ± 0%   23.11Mi ± 0%   -0.94% (p=0.000 n=10)
RegexpMatchHard_32     14.24Mi ± 0%   14.48Mi ± 0%   +1.67% (p=0.000 n=10)
RegexpMatchHard_1K     15.30Mi ± 0%   15.59Mi ± 0%   +1.93% (p=0.000 n=10)
Revcomp                181.3Mi ± 0%   203.4Mi ± 0%  +12.21% (p=0.000 n=10)
Template               15.89Mi ± 1%   16.00Mi ± 2%        ~ (p=0.542 n=10)
geomean                59.33Mi        60.72Mi        +2.33%

Change-Id: I9ac28d936e03d21c46bb19fa100018f61ace6b42
Reviewed-on: https://go-review.googlesource.com/c/go/+/479816
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Auto-Submit: Ian Lance Taylor <iant@google.com>
Run-TryBot: WANG Xuerui <git@xen0n.name>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Ian Lance Taylor <iant@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
This commit is contained in:
WANG Xuerui 2023-03-28 19:58:17 +08:00 committed by Gopher Robot
parent 47b22b6548
commit 157999f512
5 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
src/internal/bytealg/compare_loong64.s
View File

@ -48,6 +48,7 @@ entry:
AND $7, R15
BNE R0, R15, byte_loop
PCALIGN $16
chunk16_loop:
BEQ R0, R14, byte_loop
MOVV (R6), R8

1
src/internal/bytealg/equal_loong64.s
View File

@ -14,6 +14,7 @@ TEXT runtime·memequal(SB),NOSPLIT|NOFRAME,$0-25
BEQ R4, R5, eq
MOVV size+16(FP), R6
ADDV R4, R6, R7
PCALIGN $16
loop:
BNE R4, R7, test
MOVV $1, R4

2
src/internal/bytealg/indexbyte_loong64.s
View File

@ -13,6 +13,7 @@ TEXT ·IndexByte(SB),NOSPLIT,$0-40
ADDV R4, R5 // end
ADDV $-1, R4
PCALIGN $16
loop:
ADDV $1, R4
BEQ R4, R5, notfound
@ -36,6 +37,7 @@ TEXT ·IndexByteString(SB),NOSPLIT,$0-32
ADDV R4, R5 // end
ADDV $-1, R4
PCALIGN $16
loop:
ADDV $1, R4
BEQ R4, R5, notfound

1
src/runtime/memclr_loong64.s
View File

@ -26,6 +26,7 @@ words:
// do 8 bytes at a time if there is room
ADDV $-7, R4, R7
PCALIGN $16
SGTU R7, R6, R8
BEQ R8, out
MOVV R0, (R6)

2
src/runtime/memmove_loong64.s
View File

@ -42,6 +42,7 @@ words:
// do 8 bytes at a time if there is room
ADDV $-7, R9, R6 // R6 is end pointer-7
PCALIGN $16
SGTU R6, R4, R8
BEQ R8, out
MOVV (R5), R7
@ -86,6 +87,7 @@ words1:
// do 8 bytes at a time if there is room
ADDV $7, R4, R6 // R6 is start pointer+7
PCALIGN $16
SGTU R9, R6, R8
BEQ R8, out1
ADDV $-8, R5