[Mlir-commits] [mlir] [mlir][vectorize] Support affine.apply in SuperVectorize (PR #77968)

[Hsiangkai Wang]

================
@@ -1184,10 +1202,15 @@ static Operation *vectorizeAffineLoad(AffineLoadOp loadOp,
   SmallVector<Value, 8> indices;
   indices.reserve(memRefType.getRank());
   if (loadOp.getAffineMap() !=
-      state.builder.getMultiDimIdentityMap(memRefType.getRank()))
+      state.builder.getMultiDimIdentityMap(memRefType.getRank())) {
+    // Check the operand in loadOp affine map does not come from AffineApplyOp.
+    for (auto op : mapOperands) {
+      if (op.getDefiningOp<AffineApplyOp>())
+        return nullptr;
+    }
     computeMemoryOpIndices(loadOp, loadOp.getAffineMap(), mapOperands, state,
                            indices);
-  else
+  } else
----------------
Hsiangkai wrote:

Done

https://github.com/llvm/llvm-project/pull/77968