[Mlir-commits] [mlir] [MLIR][MemRef] Fix AllocOp/AllocaOp flattening domination violation (PR #188980)

[Hocky Yudhiono]

================
@@ -230,19 +210,98 @@ static LogicalResult canBeFlattened(T op, PatternRewriter &rewriter) {
       .Default([&](auto op) { return success(); });
 }
 
+// Pattern for memref::AllocOp and memref::AllocaOp.
+//
+// The "source" memref for these ops IS the op's own result, so the generic
+// MemRefRewritePattern cannot be used: getFlattenMemrefAndOffset would insert
+// ExtractStridedMetadataOp and ReinterpretCastOp that use op.result BEFORE op
+// in the block. After replaceOpWithNewOp the original result is RAUW'd to the
+// new ReinterpretCastOp, leaving the earlier ops with forward references
+// (domination violations) caught by MLIR_ENABLE_EXPENSIVE_PATTERN_API_CHECKS.
+//
+// Instead, sizes and strides are computed from the op's operands and type
+// (which all dominate the op), avoiding any reference to op.result until the
+// final replaceOpWithNewOp inside castAllocResult.
+template <typename T>
+struct AllocLikeFlattenPattern : public OpRewritePattern<T> {
+  using OpRewritePattern<T>::OpRewritePattern;
+  LogicalResult matchAndRewrite(T op,
----------------
hockyy wrote:

```suggestion
template <typename AllocLikeOp>
struct AllocLikeFlattenPattern : public OpRewritePattern<AllocLikeOp> {
  using OpRewritePattern<AllocLikeOp>::OpRewritePattern;
  LogicalResult matchAndRewrite(AllocLikeOp op,
```

https://github.com/llvm/llvm-project/pull/188980