[llvm] [MIPS]Remove unnecessary SLL instructions on MIPS64el (PR #109386)

via llvm-commits llvm-commits at lists.llvm.org
Tue Apr 29 19:49:35 PDT 2025


================
@@ -1221,6 +1222,37 @@ static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
                      DAG.getConstant(SMSize, DL, MVT::i32));
 }
 
+static SDValue performSignExtendCombine(SDNode *N, SelectionDAG &DAG,
+                                        TargetLowering::DAGCombinerInfo &DCI,
+                                        const MipsSubtarget &Subtarget) {
+  if (DCI.Level != AfterLegalizeDAG || !Subtarget.isGP64bit()) {
+    return SDValue();
+  }
+
+  SDValue N0 = N->getOperand(0);
+  EVT VT = N->getValueType(0);
+
+  // Pattern match XOR.
+  // $dst = sign_extend (xor (trunc $src, i32), imm)
+  // => $dst = xor (signext_inreg $src, i32), imm
+  if (N0.getOpcode() == ISD::XOR &&
+      N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
+      N0.getOperand(1).getOpcode() == ISD::Constant) {
+    SDValue TruncateSource = N0.getOperand(0).getOperand(0);
+    auto *ConstantOperand = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+    if (VT == MVT::i64 && VT == TruncateSource.getValueType()) {
----------------
yingopq wrote:

Do we have to put a limit here?
If the input parameter is i16, can it also be adapted to this optimization?

https://github.com/llvm/llvm-project/pull/109386


More information about the llvm-commits mailing list