[Mlir-commits] [mlir] [mlir][memref] canonicalization for erasing copying subview to identical subview (PR #125852)

[Frank Schlimbach]

fschlimb wrote:

> > I guess the question is, why can't you run canonicalize->cse->canonicalize to solve this problem?
> 
> This is really expensive though: so rather than a "can" I see it as whether it is desirable. Beyond the "purity" of the design, this gets into practical territory: the tradeoff for the canonicalization to be useful enough can be driven by actual patterns of IR observed. 

I second @joker-eph. The pattern covered here I actually do see, in particular MLIR's tiling machinery (like `transform.structured.tile_using_for*` ) creates such. Should other op combinations get observed they can still be added.


https://github.com/llvm/llvm-project/pull/125852