[llvm] [AArch64] Remove EXT instr before UZP when extracting elements from vector (PR #91328)

[via llvm-commits]

================
@@ -21448,6 +21448,29 @@ static SDValue performUzpCombine(SDNode *N, SelectionDAG &DAG,
   SDValue Op1 = N->getOperand(1);
   EVT ResVT = N->getValueType(0);
 
+  // uzp(extract_lo(x), extract_hi(x)) -> extract_lo(uzp x, x)
+  if (Op0.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+      Op1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+      Op0.getOperand(0) == Op1.getOperand(0)) {
+
+    SDValue SourceVec = Op0.getOperand(0);
+    uint64_t ExtIdx0 = Op0.getConstantOperandVal(1);
+    uint64_t ExtIdx1 = Op1.getConstantOperandVal(1);
+    uint64_t NumElements = SourceVec.getValueType().getVectorMinNumElements();
+    if (ExtIdx0 == 0 && ExtIdx1 == NumElements / 2) {
+      EVT OpVT = Op0.getOperand(1).getValueType();
+      EVT WidenedResVT = ResVT.getDoubleNumVectorElementsVT(*DAG.getContext());
+      SDValue uzp2 =
----------------
Lukacma wrote:

It certainly produces different code. But based on what I have seen from current testcases, it doesn't seem to degrade performance. The only thing is that sometimes compiler now prefers using XTN instruction instead of UZP1 for extracting even elements sometimes, but based on software optimizations guides for V3 and N3 core, both instructions have same latency and throughput. So unless smth changes in the future this should be fine.

I changed code to use UNDEF now. Thanks for the suggestions ! 


https://github.com/llvm/llvm-project/pull/91328