Since the load is inserted in place of the reshape (so the shape tensor
cannot be overwritten) the index has to dominate the reshape. With a
dynamic index there I don't see a way of enforcing that. Constants are
moved to the top so it's safe to do this transformation for them.
Details
Details
Diff Detail
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
mlir/test/Dialect/Standard/canonicalize.mlir | ||
---|---|---|
147 | maybe insert a couple of CHECK-NOTs to make sure that it actually folded? memref.dim alone would pass with the original IR. |
maybe insert a couple of CHECK-NOTs to make sure that it actually folded? memref.dim alone would pass with the original IR.