From 8ccd6c20347dfb6095a572ec3dc43f19c60f622c Mon Sep 17 00:00:00 2001
From: David Chase <drchase@google.com>
Date: Mon, 18 Aug 2025 15:04:45 -0400
Subject: [PATCH] [dev.simd] simd, cmd/compile: mark BLEND instructions as
 not-zero-mask

Change-Id: Ida9f29423d62a25be41dcf637ffb9275b7cae642
Reviewed-on: https://go-review.googlesource.com/c/go/+/697055
Reviewed-by: Cherry Mui <cherryyz@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
---
 src/cmd/compile/internal/amd64/simdssa.go |  4 ----
 src/simd/_gen/simdgen/ops/Moves/go.yaml   |  2 ++
 src/simd/simd_test.go                     | 22 ++++++++++++++++++++++
 3 files changed, 24 insertions(+), 4 deletions(-)

diff --git a/src/cmd/compile/internal/amd64/simdssa.go b/src/cmd/compile/internal/amd64/simdssa.go
index 466e6c9cc7..1ab4c88cba 100644
--- a/src/cmd/compile/internal/amd64/simdssa.go
+++ b/src/cmd/compile/internal/amd64/simdssa.go
@@ -1654,10 +1654,6 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VPXORQMasked128,
 		ssa.OpAMD64VPXORQMasked256,
 		ssa.OpAMD64VPXORQMasked512,
-		ssa.OpAMD64VPBLENDMBMasked512,
-		ssa.OpAMD64VPBLENDMWMasked512,
-		ssa.OpAMD64VPBLENDMDMasked512,
-		ssa.OpAMD64VPBLENDMQMasked512,
 		ssa.OpAMD64VPSLLWMasked128const,
 		ssa.OpAMD64VPSLLWMasked256const,
 		ssa.OpAMD64VPSLLWMasked512const,
diff --git a/src/simd/_gen/simdgen/ops/Moves/go.yaml b/src/simd/_gen/simdgen/ops/Moves/go.yaml
index 0e5997deeb..d4d1b4b9bd 100644
--- a/src/simd/_gen/simdgen/ops/Moves/go.yaml
+++ b/src/simd/_gen/simdgen/ops/Moves/go.yaml
@@ -253,6 +253,7 @@
 # That means the signature is wrong.
 - go: blend
   asm: VPBLENDVB
+  zeroing: false
   in:
   - &v
     go: $t
@@ -269,6 +270,7 @@
 # For AVX512
 - go: blend
   asm: VPBLENDM[BWDQ]
+  zeroing: false
   in:
   - &v
     go: $t
diff --git a/src/simd/simd_test.go b/src/simd/simd_test.go
index 831dc4f268..ce982409ea 100644
--- a/src/simd/simd_test.go
+++ b/src/simd/simd_test.go
@@ -397,6 +397,28 @@ func TestMergeFloat(t *testing.T) {
 	checkSlices[float64](t, s, []float64{4, 2, 3, 4})
 }
 
+func TestMergeFloat512(t *testing.T) {
+	if !simd.HasAVX512() {
+		t.Skip("Test requires HasAVX512, not available on this hardware")
+		return
+	}
+	a := simd.LoadFloat64x8Slice([]float64{1, 2, 3, 4, 5, 6, 7, 8})
+	b := simd.LoadFloat64x8Slice([]float64{8, 7, 6, 5, 4, 2, 3, 1})
+	g := a.Greater(b)
+	k := make([]int64, 8, 8)
+	g.AsInt64x8().StoreSlice(k)
+	checkSlices[int64](t, k, []int64{0, 0, 0, 0, -1, -1, -1, -1})
+	c := a.Merge(b, g)
+	d := a.Masked(g)
+
+	s := make([]float64, 8, 8)
+	c.StoreSlice(s)
+	checkSlices[float64](t, s, []float64{8, 7, 6, 5, 5, 6, 7, 8})
+
+	d.StoreSlice(s)
+	checkSlices[float64](t, s, []float64{0, 0, 0, 0, 5, 6, 7, 8})
+}
+
 var ro uint8 = 2
 
 func TestRotateAllVariable(t *testing.T) {
-- 
2.52.0