|
Both Keith's https://go-review.googlesource.com/c/41612/ and
and Ben's https://go-review.googlesource.com/c/41679/ optimized ARM's
constant pool. But neither was complete.
First, BIC was forgotten.
1. "BIC $0xff00ff00, Reg" can be optimized to
"BIC $0xff000000, Reg
BIC $0x0000ff00, Reg"
2. "BIC $0xffff00ff, Reg" can be optimized to
"AND $0x0000ff00, Reg"
3. "AND $0xffff00ff, Reg" can be optimized to
"BIC $0x0000ff00, Reg"
Second, break a non-ARMImmRot to the subtraction of two ARMImmRots was
left as TODO.
1. "ADD $0x00fffff0, Reg" can be optimized to
"ADD $0x01000000, Reg
SUB $0x00000010, Reg"
2. "SUB $0x00fffff0, Reg" can be optimized to
"SUB $0x01000000, Reg
ADD $0x00000010, Reg"
This patch fixes them and issue #19844.
The go1 benchmark shows improvements.
name old time/op new time/op delta
BinaryTree17-4 41.4s ± 1% 41.7s ± 1% +0.54% (p=0.000 n=50+49)
Fannkuch11-4 24.7s ± 1% 25.1s ± 0% +1.70% (p=0.000 n=50+49)
FmtFprintfEmpty-4 853ns ± 1% 852ns ± 1% ~ (p=0.833 n=50+50)
FmtFprintfString-4 1.33µs ± 1% 1.33µs ± 1% ~ (p=0.163 n=50+50)
FmtFprintfInt-4 1.40µs ± 1% 1.40µs ± 0% ~ (p=0.293 n=50+35)
FmtFprintfIntInt-4 2.09µs ± 1% 2.08µs ± 1% -0.39% (p=0.000 n=50+49)
FmtFprintfPrefixedInt-4 2.43µs ± 1% 2.43µs ± 1% ~ (p=0.552 n=50+50)
FmtFprintfFloat-4 4.57µs ± 1% 4.42µs ± 1% -3.18% (p=0.000 n=50+50)
FmtManyArgs-4 8.62µs ± 1% 8.52µs ± 0% -1.08% (p=0.000 n=50+50)
GobDecode-4 101ms ± 1% 101ms ± 2% +0.45% (p=0.001 n=49+49)
GobEncode-4 90.7ms ± 1% 91.1ms ± 2% +0.51% (p=0.001 n=50+50)
Gzip-4 4.23s ± 1% 4.21s ± 1% -0.62% (p=0.000 n=50+50)
Gunzip-4 623ms ± 1% 619ms ± 0% -0.63% (p=0.000 n=50+42)
HTTPClientServer-4 721µs ± 5% 683µs ± 3% -5.25% (p=0.000 n=50+47)
JSONEncode-4 251ms ± 1% 253ms ± 1% +0.54% (p=0.000 n=49+50)
JSONDecode-4 941ms ± 1% 944ms ± 1% +0.30% (p=0.001 n=49+50)
Mandelbrot200-4 49.3ms ± 1% 49.3ms ± 0% ~ (p=0.918 n=50+48)
GoParse-4 47.1ms ± 1% 47.2ms ± 1% +0.18% (p=0.025 n=50+50)
RegexpMatchEasy0_32-4 1.23µs ± 1% 1.24µs ± 1% +0.30% (p=0.000 n=49+50)
RegexpMatchEasy0_1K-4 7.74µs ± 7% 7.76µs ± 5% ~ (p=0.888 n=50+50)
RegexpMatchEasy1_32-4 1.32µs ± 1% 1.32µs ± 1% +0.23% (p=0.003 n=50+50)
RegexpMatchEasy1_1K-4 10.6µs ± 2% 10.5µs ± 3% -1.29% (p=0.000 n=49+50)
RegexpMatchMedium_32-4 2.19µs ± 1% 2.10µs ± 1% -3.79% (p=0.000 n=49+49)
RegexpMatchMedium_1K-4 544µs ± 0% 545µs ± 0% ~ (p=0.123 n=41+50)
RegexpMatchHard_32-4 28.8µs ± 0% 28.8µs ± 1% ~ (p=0.580 n=46+50)
RegexpMatchHard_1K-4 863µs ± 1% 865µs ± 1% +0.31% (p=0.027 n=47+50)
Revcomp-4 82.2ms ± 2% 82.3ms ± 2% ~ (p=0.894 n=48+49)
Template-4 1.06s ± 1% 1.04s ± 1% -1.18% (p=0.000 n=50+49)
TimeParse-4 7.25µs ± 1% 7.35µs ± 0% +1.48% (p=0.000 n=50+50)
TimeFormat-4 13.3µs ± 1% 13.2µs ± 1% -0.13% (p=0.007 n=50+50)
[Geo mean] 736µs 733µs -0.37%
name old speed new speed delta
GobDecode-4 7.60MB/s ± 1% 7.56MB/s ± 2% -0.46% (p=0.001 n=49+49)
GobEncode-4 8.47MB/s ± 1% 8.42MB/s ± 2% -0.50% (p=0.001 n=50+50)
Gzip-4 4.58MB/s ± 1% 4.61MB/s ± 1% +0.59% (p=0.000 n=50+50)
Gunzip-4 31.2MB/s ± 1% 31.4MB/s ± 0% +0.63% (p=0.000 n=50+42)
JSONEncode-4 7.73MB/s ± 1% 7.69MB/s ± 1% -0.53% (p=0.000 n=49+50)
JSONDecode-4 2.06MB/s ± 1% 2.06MB/s ± 1% ~ (p=0.052 n=44+50)
GoParse-4 1.23MB/s ± 0% 1.23MB/s ± 2% ~ (p=0.526 n=26+50)
RegexpMatchEasy0_32-4 25.9MB/s ± 1% 25.9MB/s ± 1% -0.30% (p=0.000 n=49+50)
RegexpMatchEasy0_1K-4 132MB/s ± 7% 132MB/s ± 6% ~ (p=0.885 n=50+50)
RegexpMatchEasy1_32-4 24.2MB/s ± 1% 24.1MB/s ± 1% -0.22% (p=0.003 n=50+50)
RegexpMatchEasy1_1K-4 96.4MB/s ± 2% 97.8MB/s ± 3% +1.36% (p=0.000 n=50+50)
RegexpMatchMedium_32-4 460kB/s ± 0% 476kB/s ± 1% +3.43% (p=0.000 n=49+50)
RegexpMatchMedium_1K-4 1.88MB/s ± 0% 1.88MB/s ± 0% ~ (all equal)
RegexpMatchHard_32-4 1.11MB/s ± 0% 1.11MB/s ± 1% +0.34% (p=0.000 n=45+50)
RegexpMatchHard_1K-4 1.19MB/s ± 1% 1.18MB/s ± 1% -0.34% (p=0.033 n=50+50)
Revcomp-4 30.9MB/s ± 2% 30.9MB/s ± 2% ~ (p=0.894 n=48+49)
Template-4 1.84MB/s ± 1% 1.86MB/s ± 2% +1.19% (p=0.000 n=48+50)
[Geo mean] 6.63MB/s 6.65MB/s +0.26%
Fixes #19844.
Change-Id: I5ad16cc0b29267bb4579aca3dcc10a0b8ade1aa4
Reviewed-on: https://go-review.googlesource.com/42430
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
|
|
When a constant doesn't fit in a single instruction, use two
paired instructions instead of the constant pool. For example
ADD $0xaa00bb, R0, R1
Used to rewrite to:
MOV ?(IP), R11
ADD R11, R0, R1
Instead, do:
ADD $0xaa0000, R0, R1
ADD $0xbb, R1, R1
Same number of instructions.
Good:
4 less bytes (no constant pool entry)
One less load.
Bad:
Critical path is one instruction longer.
It's probably worth it to avoid the loads, they are expensive.
Dave Cheney got us some performance numbers: https://perf.golang.org/search?q=upload:20170426.1
TL;DR mean 1.37% improvement.
Change-Id: Ib206836161fdc94a3962db6f9caa635c87d57cf1
Reviewed-on: https://go-review.googlesource.com/41612
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
|
