[BUG] A data buffer is cast into a type with a different layout

[vuule]

A device array containing objects of type thrust::pair<const char*, size_type> is cast to nvstrdesc_s using reinterpret_cast. These two types are not guaranteed to have the same memory layout.
The code works, but it's not guaranteed to work correctly on all compilers/platforms.

Observed the issue in ORC and Parquet readers.

[harrism]

@vuule can you add a link to the problem code?

[vuule]

in column_buffer:

cudf/cpp/src/io/utilities/column_buffer.hpp

Line 65 in b773c9e

using str_pair = thrust::pair<const char*, size_type>;

cudf/cpp/src/io/utilities/column_buffer.hpp

Line 94 in b773c9e

rmm::device_vector<str_pair> _strings;

cudf/cpp/src/io/utilities/column_buffer.hpp

Line 82 in b773c9e

auto data() { return _strings.size() ? _strings.data().get() : _data.data(); }

In the Parquet reader:

cudf/cpp/src/io/parquet/reader_impl.cu

Line 673 in b773c9e

chunks[c].column_data_base = out_buffers[chunk_col_map[c]].data();

and then in page decode there are a few lines like this one:

cudf/cpp/src/io/parquet/page_dict.cu

Line 203 in ef1629c

const char *ptr = reinterpret_cast<const nvstrdesc_s *>(s->col.column_data_base)[row].ptr;

[devavret]

Can this be remedied by the suggestion in #5682 . Essentially doing away with nvstrdesc

[vuule]

Thanks @devavret . Assigned to you since the other issue is already on you.

[github-actions]

This issue has been marked rotten due to no recent activity in the past 90d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[vuule]

Issue still relevant.

[vyasr]

This was resolved in #7841