[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]

================
@@ -45950,6 +45950,30 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // Exploits AVX2 VSHLV/VSRLV instructions for efficient unsigned vector shifts
+  // with out-of-bounds clamping.
+
+  // Unlike general shift instructions (SHL/SRL), AVX2's VSHLV/VSRLV handle
+  // shift amounts exceeding the element bitwidth. VSHLV/VSRLV clamps the amount
+  // to bitwidth-1 for unsigned shifts, effectively performing a maximum left
+  // shift of bitwidth-1 positions. and returns zero for unsigned right shifts
+  // exceeding bitwidth-1.
+  if (N->getOpcode() == ISD::VSELECT &&
+      (LHS.getOpcode() == ISD::SRL || LHS.getOpcode() == ISD::SHL) &&
+      supportedVectorVarShift(VT, Subtarget, LHS.getOpcode())) {
+    APInt SV;
+    if (Cond.getOpcode() == ISD::SETCC &&
----------------
RKSimon wrote:

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

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