[llvm-branch-commits] [llvm] [LoongArch] Select V{ADD, SUB}I for operations with negative splat immediates (PR #191966)

[via llvm-branch-commits]

================
@@ -425,6 +425,27 @@ bool LoongArchDAGToDAGISel::selectVSplatImm(SDValue N, SDValue &SplatVal) {
   return false;
 }
 
+template <unsigned ImmBitSize>
+bool LoongArchDAGToDAGISel::selectVSplatImmNeg(SDValue N,
+                                               SDValue &SplatVal) const {
+  APInt ImmValue;
+  EVT EltTy = N->getValueType(0).getVectorElementType();
+
+  if (N->getOpcode() == ISD::BITCAST)
+    N = N->getOperand(0);
+
+  if (selectVSplat(N.getNode(), ImmValue, EltTy.getSizeInBits()) &&
+      ImmValue.getBitWidth() == EltTy.getSizeInBits()) {
+    if ((-ImmValue).isIntN(ImmBitSize)) {
+      SplatVal = CurDAG->getTargetConstant(-ImmValue.getSExtValue(), SDLoc(N),
----------------
wangleiat wrote:

LGTM. ImmBitSize is unlikely to be 64, but using (-ImmValue) would make this more robust. This is not blocking.

https://github.com/llvm/llvm-project/pull/191966