[llvm] [GlobalISel] [AArch64] Skip redundant shift amount masking during isel (PR #223136)

[David Green via llvm-commits]

================
@@ -766,6 +769,30 @@ bool AArch64DAGToDAGISel::SelectInlineAsmMemoryOperand(
   return true;
 }
 
+template <unsigned ShiftWidth>
+bool AArch64DAGToDAGISel::SelectShiftMask(SDValue N, SDValue &ShAmt) {
+  // AArch64 shift instructions only use the low log2(ShiftWidth) bits of the
+  // shift amount. If the shift amount has a redundant AND mask that exactly
+  // covers a narrow type, we can remove it.
+  // Note: do not strip ZERO_EXTEND/ANY_EXTEND here as that can cause
+  // register class mismatches (e.g. returning a 32-bit value for an i64 shift).
+  // Do not match ZERO_EXTEND or ANY_EXTEND — these are handled by existing
+  // patterns in the Shift multiclass (e.g. zext GPR32 → SUBREG_TO_REG).
+  if (N.getOpcode() == ISD::ZERO_EXTEND || N.getOpcode() == ISD::ANY_EXTEND)
+    return false;
----------------
davemgreen wrote:

Checking for one pattern in another isn't the best. I think returning false below would help

https://github.com/llvm/llvm-project/pull/223136