[llvm] [X86][AVX] Fix handling of out-of-bounds shift amounts in AVX2 vector logical shift nodes #83840 (PR #86922)

[Simon Pilgrim via llvm-commits]

================
@@ -47701,12 +47725,31 @@ static SDValue combineShiftToPMULH(SDNode *N, SelectionDAG &DAG,
   return DAG.getNode(ExtOpc, DL, VT, Mulh);
 }
 
-static SDValue combineShiftLeft(SDNode *N, SelectionDAG &DAG) {
+static SDValue combineShiftLeft(SDNode *N, SelectionDAG &DAG,
+                                const X86Subtarget &Subtarget) {
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
   ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
   EVT VT = N0.getValueType();
 
+  // Exploits AVX2 VSHLV/VSRLV instructions for efficient unsigned vector shifts
+  // with out-of-bounds clamping.
+  if (N0.getOpcode() == ISD::VSELECT &&
+      supportedVectorVarShift(VT, Subtarget, ISD::SHL)) {
+    SDValue Cond = N0.getOperand(0);
+    SDValue N00 = N0.getOperand(1);
+    SDValue N01 = N0.getOperand(2);
+    APInt SV;
+    if (Cond.getOpcode() == ISD::SETCC && Cond.getOperand(0) == N1 &&
----------------
RKSimon wrote:

Add a description comment:
```
// fold shl(select(icmp_ult(amt,BW),x,0),amt) -> avx2 psllv(x,amt)
```

https://github.com/llvm/llvm-project/pull/86922