transformations: (csl-stencil-bufferize) AccessOp to read to_tensor's…

… underlying memref (#3663)

`csl_stencil.AccessOp`s can read directly from memrefs underlying any
`bufferization.to_tensor` passed into them.

Currently, in the presence of several buffers (ie, stencil fields
translated to memrefs at this stage), all but the main buffer will have
intermediate `bufferization.to_tensor` ops, which creates problems in
lowering. The AccessOps can instead work directly off of the underlying
memrefs.

Co-authored-by: n-io <n-io@users.noreply.github.com>

xdsl/transforms/csl_stencil_bufferize.py

@@ -289,10 +289,18 @@ def match_and_rewrite(self, op: csl_stencil.AccessOp, rewriter: PatternRewriter,
             rewriter.replace_matched_op(to_tensor_op(op.op))
             return
 
+        # accesses to buffers passed in additional args can read directly from memref underlying `to_tensor`
+        source = (
+            op.op.op.memref
+            if isinstance(op.op, OpResult)
+            and isinstance(op.op.op, bufferization.ToTensorOp)
+            else op.op
+        )
+
         rewriter.replace_matched_op(
             [
                 access := csl_stencil.AccessOp(
-                    op.op,
+                    source,
                     op.offset,
                     r_type,
                     op.offset_mapping,