diff options
author | Filippo Valsorda <filippo@golang.org> | 2022-12-02 17:58:00 +0100 |
---|---|---|
committer | Gopher Robot <gobot@golang.org> | 2022-12-02 17:52:41 +0000 |
commit | eb122456455469b512c4a1995ab8d0f97b407ba0 (patch) | |
tree | b59bb354c574931e963f3495e7a07790514d9815 | |
parent | 6ec6492f5599f9fa811258ddec5cdf3dced6af82 (diff) | |
download | go-eb122456455469b512c4a1995ab8d0f97b407ba0.tar.gz go-eb122456455469b512c4a1995ab8d0f97b407ba0.zip |
math/big: fix BitLen performance regression
CL 450055 replaced BitLen with a slower constant-time implementation,
which caused a performance regression in some ecosystem benchmarks.
https://perf.golang.org/search?q=upload%3A20221130.13+pkg%3Agithub.com%2Fericlagergren%2Fdecimal%2Fbenchmarks
Current tip vs this CL
name old time/op new time/op delta
Pi/foo=ericlagergren_(Go)/prec=100-4 151µs ± 0% 129µs ± 0% -14.89% (p=0.000 n=10+9)
Pi/foo=ericlagergren_(GDA)/prec=100-4 305µs ± 0% 269µs ± 1% -11.88% (p=0.000 n=9+10)
Pi/foo=cockroachdb/apd/prec=100-4 5.74ms ± 0% 5.33ms ± 0% -7.02% (p=0.000 n=9+10)
Pi/foo=shopspring/prec=100-4 265µs ±16% 268µs ±11% ~ (p=0.796 n=10+10)
Pi/foo=apmckinlay/prec=100-4 3.10µs ± 0% 3.08µs ± 0% -0.60% (p=0.000 n=8+10)
Pi/foo=go-inf/prec=100-4 132µs ± 9% 137µs ± 9% ~ (p=0.182 n=10+9)
Pi/foo=float64/prec=100-4 4.97µs ± 0% 4.98µs ± 0% ~ (p=0.196 n=10+10)
CL 450055's parent vs this CL
name old time/op new time/op delta
Pi/foo=ericlagergren_(Go)/prec=100-4 129µs ± 1% 129µs ± 0% ~ (p=0.182 n=10+9)
Pi/foo=ericlagergren_(GDA)/prec=100-4 267µs ± 1% 269µs ± 1% +0.93% (p=0.001 n=9+10)
Pi/foo=shopspring/prec=100-4 252µs ± 9% 268µs ±11% ~ (p=0.052 n=10+10)
Pi/foo=apmckinlay/prec=100-4 3.10µs ± 1% 3.08µs ± 0% -0.66% (p=0.000 n=9+10)
Pi/foo=go-inf/prec=100-4 135µs ± 6% 137µs ± 9% ~ (p=0.605 n=9+9)
Pi/foo=float64/prec=100-4 4.97µs ± 0% 4.98µs ± 0% +0.23% (p=0.005 n=8+10)
goos: linux
goarch: amd64
pkg: github.com/ericlagergren/decimal_benchmarks
cpu: Intel(R) Core(TM) i5-7400 CPU @ 3.00GHz
Fixes #57014
Change-Id: I08478bea122212320a592ad2652e33077807de09
Reviewed-on: https://go-review.googlesource.com/c/go/+/454617
Reviewed-by: Roland Shoemaker <roland@golang.org>
Run-TryBot: Filippo Valsorda <filippo@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Auto-Submit: Filippo Valsorda <filippo@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
-rw-r--r-- | src/math/big/nat.go | 19 |
1 files changed, 11 insertions, 8 deletions
diff --git a/src/math/big/nat.go b/src/math/big/nat.go index 4166a90ac0..90ce6d19c4 100644 --- a/src/math/big/nat.go +++ b/src/math/big/nat.go @@ -664,15 +664,18 @@ func (x nat) bitLen() int { // This function is used in cryptographic operations. It must not leak // anything but the Int's sign and bit size through side-channels. Any // changes must be reviewed by a security expert. - // - // In particular, bits.Len and bits.LeadingZeros use a lookup table for the - // low-order bits on some architectures. if i := len(x) - 1; i >= 0 { - l := i * _W - for top := x[i]; top != 0; top >>= 1 { - l++ - } - return l + // bits.Len uses a lookup table for the low-order bits on some + // architectures. Neutralize any input-dependent behavior by setting all + // bits after the first one bit. + top := uint(x[i]) + top |= top >> 1 + top |= top >> 2 + top |= top >> 4 + top |= top >> 8 + top |= top >> 16 + top |= top >> 16 >> 16 // ">> 32" doesn't compile on 32-bit architectures + return i*_W + bits.Len(top) } return 0 } |