Preserve types during empty container assignment #58911
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
This was referenced May 25, 2021
Closed
facebook-github-bot
added
the
oncall: jit
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit 7a9b4f4 (more details on the Dr. CI page):
🕵️ 4 new failures recognized by patternsThe following CI failures do not appear to be due to upstream breakages:
|
[ghstack-poisoned]
[ghstack-poisoned]
[ghstack-poisoned]
[ghstack-poisoned]
|
deferring to @gmagogsfm unless there is a specific thing u want comment on |
eellison
reviewed
Aug 4, 2021
| @@ -1127,15 +1127,6 @@ void AliasDb::makePointerTo(const Value* from, const Value* to) { | |||
| // immutable. `Any` is mutable but can point to an immutable type | |||
| // through refinement | |||
| if (isMutableTypeInternal(from) != isMutableTypeInternal(to)) { | |||
There was a problem hiding this comment.
didn't look at any of the other code in here, but we should still be making a pointer if there is any type overlap between the sets. please tag me for changes related to alias analysis. this should actually have been changed as part of the Union changes... maybe you could factor out the code from refinement into a "has type overlap" function or something similar ? it would be nice if we could reuse that logic to peephole optimize away prim::isinstance
There was a problem hiding this comment.
We're still making a pointer if there's type overlap; the only thing I deleted was the check leading up to the TORCH_INTERNAL_ASSERT on line 1137. The reason we need to do this is because, if we don't, the two branches for the unchecked_cast will falsely throw an error.
Example traceback from the TestUnion.test_union_with_dictliteral test:
Traceback (most recent call last):
File "/data/users/ansley/pytorch/test/jit/test_union.py", line 705, in test_union_with_dictliteral
self.checkScript(fn, ())
File "/home/ansley/local/miniconda3/envs/pytorch/lib/python3.8/site-packages/torch/testing/_internal/jit_utils.py", line 454, in checkScript
self.checkScript(
File "/home/ansley/local/miniconda3/envs/pytorch/lib/python3.8/site-packages/torch/testing/_internal/jit_utils.py", line 485, in checkScript
script_outputs = scripted_fn(*recording_inputs)
File "/home/ansley/local/miniconda3/envs/pytorch/lib/python3.8/site-packages/torch/testing/_internal/common_utils.py", line 139, in prof_func_call
return prof_callable(func_call, *args, **kwargs)
File "/home/ansley/local/miniconda3/envs/pytorch/lib/python3.8/site-packages/torch/testing/_internal/common_utils.py", line 136, in prof_callable
return callable(*args, **kwargs)
RuntimeError: expected_kindINTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/alias_analysis.cpp":1138, please report a bug to PyTorch. intDict(str, Tensor)
There was a problem hiding this comment.
(isMutableTypeInternal(from) != isMutableTypeInternal(to)) we're checking equality not overlap here
There was a problem hiding this comment.
Sorry for the wording choice, I did mean equality
There was a problem hiding this comment.
i think we want to be checking overlap, e.g. Optional[List[int]] should have a pointer to List[int]. If the sometimes/never/always subtyping logic from ir_emitter was factored out we could just check that it's not never, and then make a pointer
[ghstack-poisoned]
gmagogsfm
requested changes
Aug 12, 2021
| union_type_hint->containedTypes().end(), | ||
| std::back_inserter(list_types), | ||
| [&](TypePtr type_ptr) { | ||
| return type_ptr->isSubtypeOf(AnyListType::get()); |
There was a problem hiding this comment.
Check for type_ptr->kind == ListType::Kind
| Node* result = graph->insertNode(graph->createList(elem_type, values)); | ||
| if (annotated_union_type) { | ||
| Node* n = graph->insertNode( | ||
| graph->create(prim::unchecked_cast, {result->output()})); |
There was a problem hiding this comment.
Discussed offline:
This node is inserted because interpreter isn't happy about the type mismatch in assignment.
TODO: Find out why interpreter isn't happy about it
TODO: Find out performance implications (if any) about inserting this unchecked_cast node into graph.
| c10::optional<TypePtr> unified = unifyTypeList( | ||
| types, nowhere, /*default_to_union=*/true, element_type_hint); | ||
| types, nowhere, /*default_to_union=*/true, known_elem_type); | ||
|
|
||
| if (!type_hint && *unified == AnyType::get()) { |
There was a problem hiding this comment.
In theory, this should never trigger when default_to_union=True.
| // We don't want to use `elem_type` as the final argument to | ||
| // `unifyTypeList` because there's a chance that `elem_type` is | ||
| // the Tensor default | ||
| auto known_elem_type = |
There was a problem hiding this comment.
Suggested change
| auto known_elem_type = | |
| auto elem_type_hint = |
| @@ -3789,13 +3837,48 @@ struct to_ir { | |||
| // `Any` as a catch-all supertype). Assume `[]` is `List[Tensor]` | |||
| TypePtr elem_type = TensorType::get(); | |||
There was a problem hiding this comment.
Suggested change
| TypePtr elem_type = TensorType::get(); | |
| TypePtr inferred_elem_type = TensorType::get(); |
test/jit/test_union.py
Outdated
| x["bar"] = torch.tensor(3) | ||
| return x | ||
|
|
||
| self.checkScript(fn, ()) |
There was a problem hiding this comment.
Nit:
Use FileCheck on generated IR to make sure right nodes/types are inserted into graph
[ghstack-poisoned]
[ghstack-poisoned]
Differential Revision: [D30651066](https://our.internmc.facebook.com/intern/diff/D30651066) [ghstack-poisoned]
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
gmagogsfm
approved these changes
Sep 7, 2021
| // If the type hint was not `List[T]`, throw an error | ||
| throw ErrorReport(ll) << "Expected a List type hint but instead got " | ||
| << type_hint->repr_str(); | ||
| // If necessary/possible, make `type_hint` a ListType |
There was a problem hiding this comment.
The added logic is complex enough to warrant a helper function I think. What do you think?
Differential Revision: [D30785623](https://our.internmc.facebook.com/intern/diff/D30785623) [ghstack-poisoned]
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D30785623](https://our.internmc.facebook.com/intern/diff/D30785623) [ghstack-poisoned]
Differential Revision: [D30785623](https://our.internmc.facebook.com/intern/diff/D30785623) [ghstack-poisoned]
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
2 similar comments
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D30785623](https://our.internmc.facebook.com/intern/diff/D30785623) [ghstack-poisoned]
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D30785623](https://our.internmc.facebook.com/intern/diff/D30785623) [ghstack-poisoned]
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
4 similar comments
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
|
@ansley has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack:
Differential Revision: D30785623