[llvm] [CodeGen] Optimize ctpop expansion using computeKnownBits (PR #221671)

[via llvm-commits]

================
@@ -10811,6 +10798,25 @@ SDValue TargetLowering::expandCTPOP(SDNode *Node, SelectionDAG &DAG) const {
   EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
   SDValue Op = Node->getOperand(0);
   unsigned Len = VT.getScalarSizeInBits();
+
+  // Compute effective bit width from known bits, allowing us to shift the
+  // active bits down if necessary to fit into smaller specialized expansions.
+  KnownBits Known = DAG.computeKnownBits(Op);
+  unsigned LZ = Known.countMinLeadingZeros();
+  unsigned TZ = Known.countMinTrailingZeros();
+  unsigned ShiftedActiveBits = Known.getBitWidth() - (LZ + TZ);
+
+  // If the active bits are not at the low end, shift them down
+  if (ShiftedActiveBits < Len && TZ > 0) {
+    Op = DAG.getNode(ISD::SRL, dl, VT, Op,
+                     DAG.getShiftAmountConstant(TZ, VT, dl));
+  }
----------------
milkHongYe wrote:

Done

https://github.com/llvm/llvm-project/pull/221671