[llvm] [InstCombine] Fold icmp (zext X), (and (trunc Y), Mask) -> icmp X, trunc Y (PR #221189)

[armandeep singh via llvm-commits]

================
@@ -6626,12 +6648,42 @@ Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
     return new ICmpInst(ICmp.getPredicate(),
                         SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(0),
                         SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(1));
+  // Catch the mirrored operand order: icmp Pred (and (trunc Y), Mask), (zext
+  // X). Canonicalization (like ugt -> ult) often swaps operands, pushing the
+  // zext to Operand 1. We need to handle this to avoid missing optimizations.
+  {
+    Value *X, *Y;
+    const APInt *Mask;
 
+    // Ensure we have an unsigned compare with a zext on the right side.
+    // Then look for our one use AND with a mask on the left.
+    if (match(ICmp.getOperand(1), m_ZExt(m_Value(X))) && !ICmp.isSigned() &&
+        match(ICmp.getOperand(0),
+              m_OneUse(m_c_And(m_Trunc(m_Value(Y)), m_APInt(Mask))))) {
+
+      Type *SmallType = X->getType();
+      unsigned SmallWidth = SmallType->getScalarSizeInBits();
+
+      // If the mask exactly covers the bits of the narrower type, the higher
+      // bits are already guaranteed to be zero. We can bypass the extension and
+      // compare the truncated values directly.
+      if (*Mask == APInt::getLowBitsSet(Mask->getBitWidth(), SmallWidth)) {
----------------
armandeep23947 wrote:

Fixed.

https://github.com/llvm/llvm-project/pull/221189