[X86] lowerVECTOR_SHUFFLE - canonicalize zeros/ones/fp splat constants to ensure no undefs (#141214)

RKSimon · web-flow · commit 6e574a4fa332 · 2025-05-23T11:02:46.000+01:00
Make it easier for splat/element-equivalent detection by ensuring
constant splats contain no undefs.

Integer constants are limited to rematerializable zeros/ones values to
avoid unnecessary scalar_to_vector(int) -&gt; load conversions - we can
relax this later if useful
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -18322,6 +18322,25 @@ static SDValue lowerVECTOR_SHUFFLE(SDValue Op, const X86Subtarget &Subtarget,
          "canonicalizeShuffleMaskWithHorizOp "
          "shouldn't alter the shuffle mask size");
 
+  // Canonicalize zeros/ones/fp splat constants to ensure no undefs.
+  // These will be materialized uniformly anyway, so make splat matching easier.
+  // TODO: Allow all int constants?
+  auto CanonicalizeConstant = [VT, &DL, &DAG](SDValue V) {
+    if (auto *BV = dyn_cast<BuildVectorSDNode>(V)) {
+      BitVector Undefs;
+      if (SDValue Splat = BV->getSplatValue(&Undefs)) {
+        if (Undefs.any() &&
+            (isNullConstant(Splat) || isAllOnesConstant(Splat) ||
+             isa<ConstantFPSDNode>(Splat))) {
+          V = DAG.getBitcast(VT, DAG.getSplat(BV->getValueType(0), DL, Splat));
+        }
+      }
+    }
+    return V;
+  };
+  V1 = CanonicalizeConstant(V1);
+  V2 = CanonicalizeConstant(V2);
+
   // Commute the shuffle if it will improve canonicalization.
   if (canonicalizeShuffleMaskWithCommute(Mask)) {
     ShuffleVectorSDNode::commuteMask(Mask);
diff --git a/llvm/test/CodeGen/X86/pr34592.ll b/llvm/test/CodeGen/X86/pr34592.ll
@@ -24,12 +24,12 @@ define <16 x i64> @pluto(<16 x i64> %arg, <16 x i64> %arg1, <16 x i64> %arg2, <1
 ; CHECK-O0-NEXT:    vmovaps 48(%rbp), %ymm11
 ; CHECK-O0-NEXT:    vmovaps 16(%rbp), %ymm11
 ; CHECK-O0-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]
+; CHECK-O0-NEXT:    vxorps %xmm3, %xmm3, %xmm3
 ; CHECK-O0-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
 ; CHECK-O0-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]
 ; CHECK-O0-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,1,2,1]
 ; CHECK-O0-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3,4,5],ymm0[6,7]
 ; CHECK-O0-NEXT:    vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm6[0,1]
-; CHECK-O0-NEXT:    vxorps %xmm3, %xmm3, %xmm3
 ; CHECK-O0-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7]
 ; CHECK-O0-NEXT:    vmovaps %xmm1, %xmm3
 ; CHECK-O0-NEXT:    vmovaps %xmm7, %xmm1
@@ -55,12 +55,12 @@ define <16 x i64> @pluto(<16 x i64> %arg, <16 x i64> %arg1, <16 x i64> %arg2, <1
 ; CHECK-O3-NEXT:    vmovdqa 208(%rbp), %ymm3
 ; CHECK-O3-NEXT:    vmovdqa 144(%rbp), %ymm0
 ; CHECK-O3-NEXT:    vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3,4,5],ymm2[6,7]
+; CHECK-O3-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; CHECK-O3-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
 ; CHECK-O3-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; CHECK-O3-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,1,2,1]
 ; CHECK-O3-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7]
 ; CHECK-O3-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm7[2,3],ymm6[0,1]
-; CHECK-O3-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; CHECK-O3-NEXT:    vpblendd {{.*#+}} ymm2 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
 ; CHECK-O3-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm7[0],ymm5[0],ymm7[2],ymm5[2]
 ; CHECK-O3-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
diff --git a/llvm/test/CodeGen/X86/pr38639.ll b/llvm/test/CodeGen/X86/pr38639.ll
@@ -6,11 +6,8 @@ define <8 x double> @test(<4 x double> %a, <4 x double> %b) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm1 = [8.2071743224100002E-1,8.2071743224100002E-1,8.2071743224100002E-1,8.2071743224100002E-1]
 ; CHECK-NEXT:    vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3],ymm1[4,5,6,7]
-; CHECK-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
-; CHECK-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
-; CHECK-NEXT:    vmovddup {{.*#+}} xmm2 = [8.2071743224100002E-1,8.2071743224100002E-1]
-; CHECK-NEXT:    # xmm2 = mem[0,0]
-; CHECK-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; CHECK-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; CHECK-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
 ; CHECK-NEXT:    retq
   %1 = shufflevector <4 x double> %a, <4 x double> <double undef, double 0x3FEA435134576E1C, double 0x3FEA435134576E1C, double 0x3FEA435134576E1C>, <8 x i32> <i32 6, i32 5, i32 2, i32 3, i32 5, i32 1, i32 3, i32 7>
   ret <8 x double> %1
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-128-v4.ll b/llvm/test/CodeGen/X86/vector-shuffle-128-v4.ll
@@ -2402,7 +2402,7 @@ define <4 x float> @shuffle_mem_pmovzx_v4f32(ptr %p0, ptr %p1) {
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
 ; AVX1-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX1-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; AVX1-NEXT:    vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2]
 ; AVX1-NEXT:    vmovaps %xmm1, (%rsi)
 ; AVX1-NEXT:    retq
@@ -2411,7 +2411,7 @@ define <4 x float> @shuffle_mem_pmovzx_v4f32(ptr %p0, ptr %p1) {
 ; AVX2OR512VL:       # %bb.0:
 ; AVX2OR512VL-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
 ; AVX2OR512VL-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX2OR512VL-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX2OR512VL-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; AVX2OR512VL-NEXT:    vbroadcastss %xmm0, %xmm0
 ; AVX2OR512VL-NEXT:    vmovaps %xmm1, (%rsi)
 ; AVX2OR512VL-NEXT:    retq
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll b/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
@@ -640,8 +640,7 @@ define <32 x float> @PR47534(<8 x float> %tmp) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; CHECK-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [7,25,26,27,7,29,30,31,7,25,26,27,7,29,30,31]
-; CHECK-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
+; CHECK-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [7,17,18,19,7,21,22,23,0,25,26,27,0,29,30,31]
 ; CHECK-NEXT:    vpermi2ps %zmm2, %zmm0, %zmm1
 ; CHECK-NEXT:    ret{{[l|q]}}
   %tmp1 = shufflevector <8 x float> %tmp, <8 x float> undef, <32 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
