[CodeGen][SVE] Avoid scalarizing zero splat stores on scalable vectors.

Summary: Implemented in replaceZeroVectorStore(). Fixes several warnings in AArch64 SVE unit tests.

Reviewers: sdesmalen, kmclaughlin, dancgr, efriedma, each, andwar, rengolin

Reviewed By: sdesmalen

Subscribers: tschuett, kristof.beyls, hiraditya, rkruppe, psnobl, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D80824

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index a3fec1d..8671209 100644 (file)
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -12069,6 +12069,10 @@ static SDValue replaceZeroVectorStore(SelectionDAG &DAG, StoreSDNode &St) {
   SDValue StVal = St.getValue();
   EVT VT = StVal.getValueType();
 
+  // Avoid scalarizing zero splat stores for scalable vectors.
+  if (VT.isScalableVector())
+    return SDValue();
+
   // It is beneficial to scalarize a zero splat store for 2 or 3 i64 elements or
   // 2, 3 or 4 i32 elements.
   int NumVecElts = VT.getVectorNumElements();