aboutsummaryrefslogtreecommitdiff
path: root/src/simd
diff options
context:
space:
mode:
authorCherry Mui <cherryyz@google.com>2026-01-02 14:02:07 -0500
committerCherry Mui <cherryyz@google.com>2026-01-02 12:16:34 -0800
commite84983fa40a6e97d3e169f1f3549af889b1b1f22 (patch)
tree782393e6b94f39a862623149faac9bad11fabaf1 /src/simd
parent8244b8567704739d9d6fa69a0f4b50b3203d6504 (diff)
downloadgo-e84983fa40a6e97d3e169f1f3549af889b1b1f22.tar.xz
cmd/compile: optimize SIMD IsNaN.Or(IsNaN)
IsNaN's underlying instruction, VCMPPS (or VCMPPD), takes two inputs, and computes either of them is NaN. Optimize the Or pattern to generate two-operand form. This implements the optimization mentioned in CL 733660. Change-Id: I13943b377ee384864c913eed320763f333a03e41 Reviewed-on: https://go-review.googlesource.com/c/go/+/733680 Reviewed-by: David Chase <drchase@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Diffstat (limited to 'src/simd')
-rw-r--r--src/simd/archsimd/internal/simd_test/compare_test.go34
1 files changed, 34 insertions, 0 deletions
diff --git a/src/simd/archsimd/internal/simd_test/compare_test.go b/src/simd/archsimd/internal/simd_test/compare_test.go
index e678676be0..ea8514ac93 100644
--- a/src/simd/archsimd/internal/simd_test/compare_test.go
+++ b/src/simd/archsimd/internal/simd_test/compare_test.go
@@ -309,4 +309,38 @@ func TestIsNaN(t *testing.T) {
testFloat32x16UnaryCompare(t, archsimd.Float32x16.IsNaN, isNaNSlice[float32])
testFloat64x8UnaryCompare(t, archsimd.Float64x8.IsNaN, isNaNSlice[float64])
}
+
+ // Test x.IsNaN().Or(y.IsNaN()), which is optimized to VCMPP(S|D) $3, x, y.
+ want32 := mapCompare(func(x, y float32) bool { return x != x || y != y })
+ want64 := mapCompare(func(x, y float64) bool { return x != x || y != y })
+ testFloat32x4Compare(t,
+ func(x, y archsimd.Float32x4) archsimd.Mask32x4 {
+ return x.IsNaN().Or(y.IsNaN())
+ }, want32)
+ testFloat64x2Compare(t,
+ func(x, y archsimd.Float64x2) archsimd.Mask64x2 {
+ return x.IsNaN().Or(y.IsNaN())
+ }, want64)
+
+ if archsimd.X86.AVX2() {
+ testFloat32x8Compare(t,
+ func(x, y archsimd.Float32x8) archsimd.Mask32x8 {
+ return x.IsNaN().Or(y.IsNaN())
+ }, want32)
+ testFloat64x4Compare(t,
+ func(x, y archsimd.Float64x4) archsimd.Mask64x4 {
+ return x.IsNaN().Or(y.IsNaN())
+ }, want64)
+ }
+
+ if archsimd.X86.AVX512() {
+ testFloat32x16Compare(t,
+ func(x, y archsimd.Float32x16) archsimd.Mask32x16 {
+ return x.IsNaN().Or(y.IsNaN())
+ }, want32)
+ testFloat64x8Compare(t,
+ func(x, y archsimd.Float64x8) archsimd.Mask64x8 {
+ return x.IsNaN().Or(y.IsNaN())
+ }, want64)
+ }
}