[Mlir-commits] [mlir] [mlir][vector] Add `SwapShapeCastOfTranspose` canonicalizer pattern (PR #100933)

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

My intiution here is that such a pattern can be a canonicalization. If I had to give a reasoning here, it is effectively bubbling up the shape cast operation, and that could be applied repeatedly as a fixed point to move the shape casts up as much as possible which I think is good for the program as a whole. But for consistency, I would also see if there are a group of patterns that do the same thing and create a fixed point outside of canonicalizations to apply the fixed point where necessary (basically everything in the compiler stack should be deliberate and whoever is putting things together must know what they are doing when they add something to a pass pipeline. The  kitchen sink approach to canonicalizer is a problem in general). But I wont really push back against this being a canonicalizer pattern.

https://github.com/llvm/llvm-project/pull/100933