[llvm] AMDGPU: VALU data fast-forwarding needs no s_delay (PR #205481)

[Jay Foad via llvm-commits]

jayfoad wrote:

Sorry for the long delays on this. Overall I think it is good work. I do still have some concern about the amount of stuff that is going into a single PR:

1. Skipping unnecessary delays for fast forward cases
2. Handling VOPD encoding, which seems to add a lot of complexity
3. Handling implicit uses and defs, which causes a lot of lit test churn

Do you think it would be possible to split this into separate PRs? Or would that cause code quality regressions in the intermediate steps?

https://github.com/llvm/llvm-project/pull/205481