[draft][flang] Undo the effects of CSE for hlfir.exactly_once. #140190
Conversation
CSE may delete operations from hlfir.exactly_once and reuse the equivalent results from the parent region(s), e.g. from the parent hlfir.region_assign. This makes it problematic to clone hlfir.exactly_once before the top-level hlfir.where. This patch adds a "canonicalizer" that pulls in such operations back into hlfir.exactly_once.
|
Hi Jean, I want to share what I have so far. Let me know if you see any problems with this approach. It works on a couple of tests. Even one from our LIT tests fails at |
There was a problem hiding this comment.
Thanks Slava, this looks great!
I also thought about what you said about having exactly_once be a side region, and I think that may be the cleanest solution if what you have here were to hit issues with operations with memory effect being moved (which I do not think is possible given the current CSE works because the dominance is clear in the IR syntax, and that is all it needs for operation that will behave the same if re-evaluated with the same inputs, but any pass moving memory operations would need to understand the control path between the two operations to check for modifications, and I do not think it is possible for MLIR to do such analysis without more info on the control flow behavior of the where/forall/region_assign/exactly_once operations).
CSE may delete operations from hlfir.exactly_once and reuse
the equivalent results from the parent region(s), e.g. from the parent
hlfir.region_assign. This makes it problematic to clone
hlfir.exactly_once before the top-level hlfir.where.
This patch adds a "canonicalizer" that pulls in such operations
back into hlfir.exactly_once.