runtime/internal/sys: implement Ctz and Bswap in assembly for 386
Ctz is a hot-spot in the Go 1.7 memory manager. In SSA it's
implemented as an intrinsic that compiles to a few instructions, but
on the old backend (all architectures other than amd64), it's
implemented as a fairly complex Go function. As a result, switching to
bitmap-based allocation was a significant hit to allocation-heavy
workloads like BinaryTree17 on non-SSA platforms.
For unknown reasons, this hit 386 particularly hard. We can regain a
lot of the lost performance by implementing Ctz in assembly on the
386. This isn't as good as an intrinsic, since it still generates a
function call and prevents useful inlining, but it's much better than
the pure Go implementation:
name old time/op new time/op delta
BinaryTree17-12 3.59s ± 1% 3.06s ± 1% -14.74% (p=0.000 n=19+20)
Fannkuch11-12 3.72s ± 1% 3.64s ± 1% -2.09% (p=0.000 n=17+19)
FmtFprintfEmpty-12 52.3ns ± 3% 52.3ns ± 3% ~ (p=0.829 n=20+19)
FmtFprintfString-12 156ns ± 1% 148ns ± 3% -5.20% (p=0.000 n=18+19)
FmtFprintfInt-12 137ns ± 1% 136ns ± 1% -0.56% (p=0.000 n=19+13)
FmtFprintfIntInt-12 227ns ± 2% 225ns ± 2% -0.93% (p=0.000 n=19+17)
FmtFprintfPrefixedInt-12 210ns ± 1% 208ns ± 1% -0.91% (p=0.000 n=19+17)
FmtFprintfFloat-12 375ns ± 1% 371ns ± 1% -1.06% (p=0.000 n=19+18)
FmtManyArgs-12 995ns ± 2% 978ns ± 1% -1.63% (p=0.000 n=17+17)
GobDecode-12 9.33ms ± 1% 9.19ms ± 0% -1.59% (p=0.000 n=20+17)
GobEncode-12 7.73ms ± 1% 7.73ms ± 1% ~ (p=0.771 n=19+20)
Gzip-12 375ms ± 1% 374ms ± 1% ~ (p=0.141 n=20+18)
Gunzip-12 61.8ms ± 1% 61.8ms ± 1% ~ (p=0.602 n=20+20)
HTTPClientServer-12 87.7µs ± 2% 86.9µs ± 3% -0.87% (p=0.024 n=19+20)
JSONEncode-12 20.2ms ± 1% 20.4ms ± 0% +0.53% (p=0.000 n=18+19)
JSONDecode-12 65.3ms ± 0% 65.4ms ± 1% ~ (p=0.385 n=16+19)
Mandelbrot200-12 4.11ms ± 1% 4.12ms ± 0% +0.29% (p=0.020 n=19+19)
GoParse-12 3.75ms ± 1% 3.61ms ± 2% -3.90% (p=0.000 n=20+20)
RegexpMatchEasy0_32-12 104ns ± 0% 103ns ± 0% -0.96% (p=0.000 n=13+16)
RegexpMatchEasy0_1K-12 805ns ± 1% 803ns ± 1% ~ (p=0.189 n=18+18)
RegexpMatchEasy1_32-12 111ns ± 0% 111ns ± 3% ~ (p=1.000 n=14+19)
RegexpMatchEasy1_1K-12 1.00µs ± 1% 1.00µs ± 1% +0.50% (p=0.003 n=19+19)
RegexpMatchMedium_32-12 133ns ± 2% 133ns ± 2% ~ (p=0.218 n=20+20)
RegexpMatchMedium_1K-12 41.2µs ± 1% 42.2µs ± 1% +2.52% (p=0.000 n=18+16)
RegexpMatchHard_32-12 2.35µs ± 1% 2.38µs ± 1% +1.53% (p=0.000 n=18+18)
RegexpMatchHard_1K-12 70.9µs ± 2% 72.0µs ± 1% +1.42% (p=0.000 n=19+17)
Revcomp-12 1.06s ± 0% 1.05s ± 0% -1.36% (p=0.000 n=20+18)
Template-12 86.2ms ± 1% 84.6ms ± 0% -1.89% (p=0.000 n=20+18)
TimeParse-12 425ns ± 2% 428ns ± 1% +0.77% (p=0.000 n=18+19)
TimeFormat-12 517ns ± 1% 519ns ± 1% +0.43% (p=0.001 n=20+19)
[Geo mean] 74.3µs 73.5µs -1.05%
Prior to this commit, BinaryTree17-12 on 386 was 33% slower than at
the go1.6 tag. With this commit, it's 13% slower.
On arm and arm64, BinaryTree17-12 is only ~5% slower than it was at
go1.6. It may be worth implementing Ctz for them as well.
I consider this change low risk, since the functions it replaces are
simple, very well specified, and well tested.
For #16117.
Change-Id: Ic39d851d5aca91330134596effd2dab9689ba066
Reviewed-on: https://go-review.googlesource.com/24640
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>