Don't look at past partitions when doing inference #59613
Conversation
|
Yeah, this is probably the wrong fix. Either codegen needs to handle whatever partition structure is causing the problem, or inference can truncate the range in a particular scenario, but needs to be more precise than this. |
|
It seems quite tricky for codegen to handle this if inference doesn't bound the world age to it's primary world. Because what is happening here is a global that has two partition, one that's visible in the current/world age of the method, and another that isn't. But those partitions have the same RTE (there's a comment there that guard should infer to Union{} but it doesn't), and with the same RTE we merge the partitions expanding the legal world age to the guard validity |
|
You can extend the rte struct to flag this situation if codegen, can't handle it (although I could of course emit a world age check) but this fix is too aggressive. |
| this_rte = query(interp, leaf_binding, leaf_partition) | ||
| if @isdefined(rte) | ||
| if this_rte === rte | ||
| if this_rte === rte && !(is_some_guard(kind) ⊻ is_some_guard(binding_kind(leaf_partition))) # Don't merge from no guard to guard |
There was a problem hiding this comment.
The same check should apply to backdated const, which needs to be treated uniformly with guard.
There was a problem hiding this comment.
should is_some_guard be true for it?
There was a problem hiding this comment.
No, it's not a guard, but there should an is_backdated, which #59735 does add.
We were looking at partitions that could never be reached here, causing odd codegen behaviour.
Fixes some part of #42559