fix tests and replace exec() in compute/_jit.py

[water0615]

Description

closes

fix tests and replace exec() in compute/_jit.py

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[shwina]

Thanks for the PR, @water0615!

By its nature, cuda.compute accepts arbitrary user callables and should be run in trusted environments.

We don't want to be too restrictive in terms of the contents of the user-provided callables. For instance, the __globals__ dict could contain e.g., other arbitrarily-named device functions. Thus, I think the restrictions in _validate_source_safety are a bit too strict and may prevent legitimate use-cases.

That being said, there are improvements we can make here in terms of validating, e.g., struct members to ensure they are identifiers. For example, I think the following patch would take us quite far in terms of better error reporting if someone passes something other than a identifier for the struct member name:

@@ -55,6 +55,13 @@ def gpu_struct(
     # At this point, field_dict must be a dict
     assert isinstance(field_dict, dict)

+    # Validate field names are valid Python identifiers before use in exec()
+    for key in field_dict:
+        if not isinstance(key, str) or not key.isidentifier():
+            raise ValueError(
+                f"gpu_struct field name {key!r} is not a valid Python identifier"
+            )
+
     # Normalize fields for storage on the struct class
     field_spec = {}
     for key, val in field_dict.items():

Do you think we could start there (and possibly some unit tests for the above)?

[shwina]

I believe numba's type system will need to see the attribute access rather than access via getattr; so this may not work as intended.

Instead, we could assert that all field names are valid Python identifiers via str.isidentifier().

[shwina]

I'm going to go ahead and close this PR, as #8235 addresses the issue without restricting the contents of user-provided callables.