[Mlir-commits] [mlir] [mlir][scf] Rewrite vector.transfer_read/write after peeling (PR #88684)

[Matthias Springer]

================
@@ -166,6 +167,52 @@ static LogicalResult peelForLoop(RewriterBase &b, ForOp forOp,
   return success();
 }
 
+static void rewriteVectorReadWriteToLoadStore(RewriterBase &b, Operation *op) {
+  b.setInsertionPoint(op);
+  if (auto write = dyn_cast<vector::TransferWriteOp>(op)) {
+    b.replaceOpWithNewOp<vector::StoreOp>(
----------------
matthias-springer wrote:

How about simply setting the `in_bounds` attribute to "true"? I think we already have canonicalization/lowering patterns that would then lower the op to `vector.load/store`.


https://github.com/llvm/llvm-project/pull/88684