[llvm] [X86] combineINSERT_SUBVECTOR - fold insert_subvector(base,extract_subvector(broadcast)) -> blend shuffle(base,broadcast) (REAPPLIED) (PR #133724)

[Phoebe Wang via llvm-commits]

================
@@ -58896,6 +58896,30 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
         Mask[i + IdxVal] = i + ExtIdxVal + VecNumElts;
       return DAG.getVectorShuffle(OpVT, dl, Vec, ExtSrc, Mask);
     }
+    // If we're broadcasting, see if we can use a blend instead of
+    // extract/insert pair. Ensure that the subvector is aligned with the
+    // insertion/extractions.
+    if ((ExtIdxVal % SubVecNumElts) == 0 && (IdxVal % SubVecNumElts) == 0 &&
+        (ExtSrc.getOpcode() == X86ISD::VBROADCAST ||
+         ExtSrc.getOpcode() == X86ISD::VBROADCAST_LOAD ||
+         (ExtSrc.getOpcode() == X86ISD::SUBV_BROADCAST_LOAD &&
----------------
phoebewang wrote:

Do we still need to limit to broadcast now?

https://github.com/llvm/llvm-project/pull/133724