[
https://issues.apache.org/jira/browse/ARROW-16013?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Tobias Zagorni updated ARROW-16013:
-----------------------------------
Description:
The calculation of the stride value passed to {{CopyStridedNatural}} by
{{NumPyStridedConverter::Visit}} divides the stride value by sizeof(T), which
is unsigned, without an appropriate type case. This causes
{{CopyStridedNatural}} to access bogus indices.
This is triggered by the existing test method {{test_numpy_to_pyarrow}} in
{{test_array.py}} (at {{{}pa.array(np_arr[case]){}}}). When using ubsan catches
a signed integer overflow, otherwise it apparently does what we want, but is
still undefined behavoir.
Also the existing behavoir would probably not work for data types that are not
a power of 2, if such would exist (numpy docs mention a float96 can exist on
some platforms, but I don't have them on x86_64)
was:
The calculation of the stride value passed to {{CopyStridedNatural}} by
{{NumPyStridedConverter::Visit}} divides the stride value by sizeof(T), which
is unsigned, without an appropriate type case. This causes
{{CopyStridedNatural}} to access bogus indices.
This is triggered by the existing test method {{test_numpy_to_pyarrow}} in
{{test_array.py}} (at {{{}pa.array(np_arr[case]){}}}). When using ubsan catches
a signed integer overflow, otherwise it apparently does what we want, but is
still undefined behavoir
> [C++][Python] Signed overflow when using negative stride in
> NumPyStridedConverter
> ---------------------------------------------------------------------------------
>
> Key: ARROW-16013
> URL: https://issues.apache.org/jira/browse/ARROW-16013
> Project: Apache Arrow
> Issue Type: Bug
> Components: C++, Python
> Reporter: Tobias Zagorni
> Assignee: Tobias Zagorni
> Priority: Minor
>
> The calculation of the stride value passed to {{CopyStridedNatural}} by
> {{NumPyStridedConverter::Visit}} divides the stride value by sizeof(T), which
> is unsigned, without an appropriate type case. This causes
> {{CopyStridedNatural}} to access bogus indices.
> This is triggered by the existing test method {{test_numpy_to_pyarrow}} in
> {{test_array.py}} (at {{{}pa.array(np_arr[case]){}}}). When using ubsan
> catches a signed integer overflow, otherwise it apparently does what we want,
> but is still undefined behavoir.
> Also the existing behavoir would probably not work for data types that are
> not a power of 2, if such would exist (numpy docs mention a float96 can exist
> on some platforms, but I don't have them on x86_64)
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
