[Mlir-commits] [mlir] Fold `memref.dim` into `memref.expand_shape` (PR #88423)

[Benoit Jacob]

bjacob wrote:

I can see how that makes sense. It means that for each such op, we just teach the compiler about computing its result shapes (by implementing `reifyResultShapes`) and it figures out how to do things such as folding `dim` ops, without the need for handwritten pattern rewriter code.

I should have the patch for `memref.expand_shape` today. But it seems that this treatment will be needed in more places, and I'm not sure I can commit to take care of them.

https://github.com/llvm/llvm-project/pull/88423