fix some cases where we incorrectly removed global var accesses #16106
Conversation
- if a global var might be undefined, consider it not `effect_free` since accessing it can throw an error. - typeasserts that don't fire are removed by inlining, so don't consider typeassert `effect_free` since any remaining calls might throw.
|
iirc, this will cause performance regressions due to inhibiting the later optimization passes (such as tuple destructuring) |
|
Which part --- the global var part or the typeassert part? I wouldn't think accessing possibly-undefined global variables was that important a case. |
|
the typeassert. i don't know of any cases where the globals access could cause issues either |
|
My thinking was that since we remove typeasserts that won't throw, any that remain indeed have effects that need to be preserved. Am I missing a case? |
|
That would make sense (I forgot we have a custom inliner for this). I would have to look back through blame to check, but I think it was the |
|
I'll check that. |
|
Will this fix the original issue in #9765? (maybe not the |
|
Yes it fixes the original issue there, but not all the cases mentioned in the later comments. So far I don't really see any regressions here. |
Possible fix for the issue identified here: #16090 (comment)