[mlir] [mlir][Vector] Tighten up application conditions in TransferReadAfter… (PR #143869)

[Andrzej Warzyński]

================
@@ -4668,12 +4668,15 @@ struct TransferReadAfterWriteToBroadcast
 
   LogicalResult matchAndRewrite(TransferReadOp readOp,
                                 PatternRewriter &rewriter) const override {
-    if (readOp.hasOutOfBoundsDim() ||
-        !llvm::isa<RankedTensorType>(readOp.getShapedType()))
-      return failure();
     auto defWrite = readOp.getBase().getDefiningOp<vector::TransferWriteOp>();
     if (!defWrite)
       return failure();
+    // Bail if we need an alias analysis.
+    if (!readOp.hasPureTensorSemantics() || !defWrite.hasPureTensorSemantics())
+      return failure();
+    // Bail if we need a bounds analysis.
+    if (readOp.hasOutOfBoundsDim() || defWrite.hasOutOfBoundsDim())
+      return failure();
----------------
banach-space wrote:

Looks like we are missing negative tests for these cases as well? And masking?

https://github.com/llvm/llvm-project/pull/143869