[PATCH] D34144: [CodeGen] Add generic MacroFusion pass.

[Evandro Menezes via Phabricator via llvm-commits]

evandro added a comment.

In https://reviews.llvm.org/D34144#781159, @fhahn wrote:

> - I've simplified `MacroFusion::scheduleAdjacentImpl` to only look at predecessors. I think that shouldn't make a difference as the same pairs should be found, just the order they are discovered may be different. @evandro was there a reason why the successors were considered for SUs in a block?


If memory serves, I came across some anecdotal instances when not all instructions were visited otherwise.  The reason, methinks, is that some instructions are marked in a special manner, like stores and prefetches.  So, I had to start with at the successors of the `EntrySU` and then again at the predecessors of `ExitSU` to make sure that all instructions in a BB would be visited.


https://reviews.llvm.org/D34144