[llvm] [X86][AVX512] Better lowering for `_mm512_maskz_shuffle_epi32` (PR #121147)

[Simon Pilgrim via llvm-commits]

================
@@ -17172,6 +17172,58 @@ static SDValue lowerV8I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
   return lowerShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, Subtarget, DAG);
 }
 
+static SDValue lowerShuffleAsVSELECT(const SDLoc &DL,
+                                     ArrayRef<int> RepeatedMask, SDValue V1,
+                                     SDValue V2, SelectionDAG &DAG) {
+  if (V1.getOpcode() != ISD::BUILD_VECTOR &&
+      V2.getOpcode() != ISD::BUILD_VECTOR)
+    return SDValue();
+  SDValue BuildVector;
+  if (V1.getOpcode() == ISD::BUILD_VECTOR) {
+    BuildVector = V1;
+    if (V2.getOpcode() != ISD::BITCAST)
+      return SDValue();
+  } else {
+    BuildVector = V2;
+    if (V1.getOpcode() != ISD::BITCAST)
+      return SDValue();
+  }
+  if (!ISD::isBuildVectorAllZeros(BuildVector.getNode()))
+    return SDValue();
+  APInt DestMask(16, 0);
+  for (unsigned i = 0; i < 16; ++i) {
+    SDValue Op = BuildVector->getOperand(i);
+    if (Op.isUndef())
+      DestMask.setBit(i);
+  }
+  if (DestMask.isZero())
+    return SDValue();
+
+  unsigned Imm8 = 0;
+  for (unsigned i = 0; i < 4; ++i) {
+    if (V1.getOpcode() != ISD::BUILD_VECTOR) {
+      if (RepeatedMask[i] >= 4) {
+        continue;
+      }
+    } else if (RepeatedMask[i] < 4) {
+      continue;
+    }
+    Imm8 += (RepeatedMask[i] % 4) << (2 * i);
+  }
----------------
RKSimon wrote:

It'd be better if you instead created a new shuffle mask instead of lowering to PSHUFD directly, so you lower to a vselect(shuffle(),zero) pattern - will make it easier to reuse in other shuffles.

https://github.com/llvm/llvm-project/pull/121147