Abs is an intrinsic (or a relatively cheap operation) on most
architectures. Using it in IsInf typically saves a branch when
`sign` is 0 (note the `sign` variable is typically a constant).
This change doesn't make a huge difference on amd64 (these
benchmarks are fairly noisy too) but removing the branch will
allow rewrite rules to detect and optimize infinity checks on
other architectures. For example, riscv64 can check for
infinities with the FCLASSD instruction and s390x can use the
TCDB instruction.
goos: linux
goarch: amd64
pkg: math
cpu: 12th Gen Intel(R) Core(TM) i7-12700T
│ sec/op │ sec/op vs base │
Acos 4.317n ± 1% 4.321n ± 0% ~ (p=0.466 n=10)
Acosh 8.857n ± 1% 8.411n ± 2% -5.05% (p=0.001 n=10)
Asin 4.260n ± 1% 4.204n ± 6% -1.31% (p=0.021 n=10)
Asinh 10.63n ± 2% 10.37n ± 0% -2.49% (p=0.000 n=10)
Atan 2.493n ± 1% 2.368n ± 0% -5.01% (p=0.000 n=10)
Atanh 8.820n ± 4% 8.770n ± 2% ~ (p=0.579 n=10)
Atan2 4.212n ± 1% 4.066n ± 11% -3.45% (p=0.023 n=10)
Cbrt 4.859n ± 0% 4.845n ± 0% -0.29% (p=0.000 n=10)
Ceil 0.3877n ± 3% 0.2514n ± 0% -35.17% (p=0.000 n=10)
Copysign 0.3479n ± 2% 0.4179n ± 0% +20.14% (p=0.000 n=10)
Cos 4.734n ± 2% 4.486n ± 0% -5.26% (p=0.000 n=10)
Cosh 5.244n ± 0% 5.071n ± 0% -3.29% (p=0.000 n=10)
Erf 2.975n ± 1% 2.788n ± 0% -6.29% (p=0.000 n=10)
Erfc 3.259n ± 1% 3.121n ± 0% -4.23% (p=0.000 n=10)
Erfinv 4.015n ± 1% 3.904n ± 0% -2.76% (p=0.000 n=10)
Erfcinv 4.166n ± 1% 4.039n ± 0% -3.04% (p=0.000 n=10)
Exp 3.567n ± 1% 3.429n ± 0% -3.87% (p=0.000 n=10)
ExpGo 9.173n ± 1% 8.368n ± 2% -8.78% (p=0.000 n=10)
Expm1 4.466n ± 3% 4.419n ± 0% -1.05% (p=0.000 n=10)
Exp2 8.328n ± 0% 8.046n ± 0% -3.39% (p=0.000 n=10)
Exp2Go 8.796n ± 5% 8.237n ± 2% -6.36% (p=0.000 n=10)
Abs 0.2400n ± 2% 0.2144n ± 0% -10.71% (p=0.000 n=10)
Dim 0.4077n ± 3% 0.3795n ± 1% -6.91% (p=0.000 n=10)
Floor 0.3616n ± 2% 0.2528n ± 3% -30.10% (p=0.000 n=10)
Max 1.401n ± 1% 1.344n ± 1% -4.14% (p=0.000 n=10)
Min 1.391n ± 1% 1.345n ± 1% -3.27% (p=0.000 n=10)
Mod 15.45n ± 1% 15.62n ± 2% ~ (p=0.066 n=10)
Frexp 1.838n ± 2% 1.605n ± 1% -12.70% (p=0.000 n=10)
Gamma 4.465n ± 1% 4.458n ± 1% ~ (p=0.256 n=10)
Hypot 2.237n ± 1% 2.208n ± 0% -1.32% (p=0.000 n=10)
HypotGo 2.610n ± 3% 2.663n ± 5% ~ (p=0.280 n=10)
Ilogb 1.793n ± 1% 1.566n ± 1% -12.66% (p=0.000 n=10)
J0 22.11n ± 1% 21.45n ± 1% -2.99% (p=0.000 n=10)
J1 21.71n ± 1% 21.38n ± 1% -1.54% (p=0.000 n=10)
Jn 46.43n ± 1% 45.83n ± 1% -1.30% (p=0.001 n=10)
Ldexp 2.360n ± 1% 2.111n ± 1% -10.51% (p=0.000 n=10)
Lgamma 4.728n ± 1% 4.850n ± 2% +2.59% (p=0.000 n=10)
Log 4.304n ± 2% 4.228n ± 1% -1.78% (p=0.000 n=10)
Logb 1.833n ± 2% 1.635n ± 2% -10.80% (p=0.000 n=10)
Log1p 5.262n ± 2% 5.173n ± 2% -1.69% (p=0.001 n=10)
Log10 4.534n ± 1% 4.474n ± 1% -1.33% (p=0.024 n=10)
Log2 2.510n ± 2% 2.246n ± 2% -10.48% (p=0.000 n=10)
Modf 1.712n ± 3% 1.700n ± 1% ~ (p=0.055 n=10)
Nextafter32 2.190n ± 3% 2.187n ± 0% ~ (p=0.266 n=10)
Nextafter64 2.184n ± 0% 2.183n ± 0% -0.05% (p=0.017 n=10)
PowInt 11.45n ± 7% 11.32n ± 9% ~ (p=0.137 n=10)
PowFrac 27.46n ± 3% 27.04n ± 1% -1.55% (p=0.001 n=10)
Pow10Pos 0.5367n ± 3% 0.5466n ± 2% +1.84% (p=0.009 n=10)
Pow10Neg 0.8939n ± 1% 0.8720n ± 2% -2.45% (p=0.000 n=10)
Round 1.218n ± 1% 1.198n ± 1% -1.56% (p=0.005 n=10)
RoundToEven 1.711n ± 0% 1.710n ± 0% ~ (p=0.464 n=10)
Remainder 12.87n ± 10% 13.79n ± 14% +7.11% (p=0.027 n=10)
Signbit 0.4072n ± 2% 0.3839n ± 2% -5.71% (p=0.000 n=10)
Sin 4.102n ± 1% 4.058n ± 3% ~ (p=0.138 n=10)
Sincos 5.837n ± 1% 5.715n ± 2% -2.10% (p=0.000 n=10)
Sinh 5.622n ± 1% 5.567n ± 2% -0.96% (p=0.006 n=10)
SqrtIndirect 0.4284n ± 0% 0.4279n ± 0% ~ (p=0.084 n=10)
SqrtLatency 2.779n ± 0% 2.777n ± 0% ~ (p=0.089 n=10)
SqrtIndirectLatency 2.777n ± 0% 2.778n ± 0% ~ (p=0.305 n=10)
SqrtGoLatency 24.00n ± 0% 24.51n ± 0% +2.12% (p=0.000 n=10)
SqrtPrime 673.0n ± 0% 673.0n ± 0% ~ (p=0.574 n=10)
Tan 4.111n ± 4% 4.123n ± 5% ~ (p=0.424 n=10)
Tanh 5.787n ± 1% 5.723n ± 1% -1.11% (p=0.010 n=10)
Trunc 0.3441n ± 3% 0.2596n ± 2% -24.56% (p=0.000 n=10)
Y0 21.63n ± 2% 21.07n ± 2% -2.61% (p=0.001 n=10)
Y1 21.42n ± 1% 20.93n ± 3% -2.29% (p=0.041 n=10)
Yn 45.78n ± 1% 45.83n ± 1% ~ (p=0.671 n=10)
Float64bits 0.2187n ± 2% 0.2199n ± 2% ~ (p=0.138 n=10)
Float64frombits 0.2198n ± 1% 0.2199n ± 1% ~ (p=0.956 n=10)
Float32bits 0.2237n ± 2% 0.2213n ± 1% ~ (p=0.060 n=10)
Float32frombits 0.2251n ± 1% 0.2219n ± 2% -1.42% (p=0.000 n=10)
FMA 0.8557n ± 1% 0.8555n ± 0% ~ (p=0.286 n=10)
geomean 3.186n 3.070n -3.61%
Change-Id: I4814bb1e3d9d20e9d8cd7689e8d5383e36b00331
Reviewed-on: https://go-review.googlesource.com/c/go/+/694955
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Sean Liao <sean@liao.dev>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>
// To avoid the floating-point hardware, could use:
// x := Float64bits(f);
// return sign >= 0 && x == uvinf || sign <= 0 && x == uvneginf;
- return sign >= 0 && f > MaxFloat64 || sign <= 0 && f < -MaxFloat64
+ if sign == 0 {
+ f = Abs(f)
+ } else if sign < 0 {
+ f = -f
+ }
+ return f > MaxFloat64
}
// normalize returns a normal number y and exponent exp