Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Fix crash in daemon mode on new import cycle (#14508)
Fixes #14329 This fixes the second crash reported in the issue (other one is already fixed). This one is tricky, it looks like it happens only when we bring in a new import cycle in an incremental update with `--follow-import=normal`. To explain the reason, a little reminder of how semantic analyzer passes work: * Originally, we recorded progress automatically when some new symbol was resolved and added to symbol tables. * With implementation of recursive types, this mechanism was insufficient, as recursive types require modifying some symbols _in place_, this is why `force_progress` flag was added to `defer()`. * I was only careful with this flag for recursive type aliases (that were implemented first), for other things (like recursive TypedDicts, etc) I just always passed `force_progress=True` (without checking if we actually resolved some placeholder types). * The reasoning for that is if we ever add `becomes_typeinfo=True`, there is no way this symbol will later be unresolved (otherwise how would we know this is something that is a type). * It turns out this reasoning doesn't work in some edge cases in daemon mode, we do put some placeholders with `becomes_typeinfo=True` for symbols imported from modules that were not yet processed, thus causing a crash (see test cases). * There were two options to fix this: one is to stop creating placeholders with `becomes_typeinfo=True` for unimported symbols in daemon mode, other one is to always carefully check if in-place update of a symbol actually resulted in progress. * Ultimately I decided that the first way is too fragile (and I don't understand how import following works for daemon anyway), and the second way is something that is technically correct anyway, so here is this PR I didn't add test cases for each of the crash scenarios, since they are all very similar. I only added two that I encountered "in the wild", upper bound and tuple base caused actual crash in `trio` stubs, plus also randomly a test for a TypedDict crash. _EDIT:_ and one more thing, the "cannot resolve name" error should never appear in normal mode, only in daemon update (see reasoning above), so I don't make those error messages detailed, just add some minimal info if we will need to debug them.
- Loading branch information