[PATCH] D65280: Add a pass to lower is.constant and objectsize intrinsics

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc added a comment.

Thanks to Joerg for some useful discussion on IRC -- there was a concern I hadn't thought about that is exactly right: we somewhat want this pass to minimally disrupt things but also to be reasonably self contained.

Based on the discussion, I think the patch can still be simplified a bit though (although not as much as I had originally suggested). Notably, the additional API for inst simplify makes a lot more sense to me now.

I'd suggest thinking about this in three phases.

First, simplify the instructions to constants. Here, you would keep the new API to extract a list of potentially relevant instructions. After each simplification, scan this and append any branches or switches to a list of terminators needing an update. (May want it to be a set vector so you don't duplicate.) But don't actually update the CFG in any way during this phase.

Second, rewrite the branches and switches collected in the first phase, collecting basic blocks that might be made unreachable in another `SetVector`. This just mutates the CFG but doesn't delete any blocks and so doesn't make any of the instructions invalid to visit. You should be able to collect a batch of domtree updates during this phase as well without applying them eagerly.

Third, apply the domtree updates (so you have precise reachability) and remove any blocks from the list that are still reachable. Hand the remaining blocks to the `DeleteBasicBlocks` utility to completely remove them. This will remove the implicit dependence between this pass and the later codegen passes which I think is much cleaner.

I think that will still remove the need to iterate, and will enhance this to actually fully remove the blocks orphaned by the folding of these branches (but *only* those blocks). Thoughts?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D65280/new/

https://reviews.llvm.org/D65280