[llvm] [X86] Optimize v4i16->v4i8 truncating stores via v4i32 widening (PR #186676)

[Simon Pilgrim via llvm-commits]

================
@@ -54294,6 +54294,20 @@ static SDValue combineStore(SDNode *N, SelectionDAG &DAG,
   EVT VT = StoredVal.getValueType();
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
+  // Pattern: store(trunc(load v4i16) to v4i8)
+  if (!St->isTruncatingStore() && VT == MVT::v4i8 && Subtarget.hasAVX512() &&
+      StoredVal.getOpcode() == ISD::TRUNCATE &&
----------------
RKSimon wrote:

Can this be done more tidily as a sd_match pattern?

https://github.com/llvm/llvm-project/pull/186676