[llvm] [MachinePipeliner] Fix incorrect handlings of unpipelineable insts (PR #126057)

[via llvm-commits]

dpenry wrote:

There's a bit of confusion of terminology around dependence chain, but what is going on here does make some sense to me now.  SU(16) causes its two predecessors (0,1) to be unpipelineable.  SU(0) is causing SU(15) to be unpipelineable because of traversing the anti-dependence backwards and then SU(14) is made unpipelineable by SU(15).

So, I think what's happened here is that the changes to the data dependence handling has probably confused the issue, because executing in SU order *should* be correct -- no actual dependence can be violated, given that the original basic block was in SU order and the transitive backwards closure of a non-pipelineable instruction is considered non-pipelineable. Is it possible that there was code removed from normalizeNonPipelinedInstructions which formerly ignored the artificial dependence when computing SU(14)'s scheduling time?  Was the anti-dependence traversed backwards previously?  Is the artificial dependence correct?

https://github.com/llvm/llvm-project/pull/126057