[llvm] [Passes] Always run LowerMatrixIntrinsics, request EarlyCSE on demand. (PR #118507)

[Florian Hahn via llvm-commits]

fhahn wrote:

> why does lower matrix intrinsics run later in the pipeline rather than earlier? if run earlier we wouldn't need the conditional EarlyCSE run later?

The main motivation for the current position is that it is likely more beneficial to retain the more high-level intrinsics during scalar simplifications (e.g. we want to fold `@llvm.matrix.multiply` / `@llvm.matrix.transpose` instead of their expanded forms which would be much more difficult to match) and also keep the IR more compact early on (expand the intrinsics can add a significant amount of IR instructions).

https://github.com/llvm/llvm-project/pull/118507