Jörn Horstmann created ARROW-10990:
--------------------------------------
Summary: [Rust]: SIMD implementation of compare kernels reads out
of bounds
Key: ARROW-10990
URL: https://issues.apache.org/jira/browse/ARROW-10990
Project: Apache Arrow
Issue Type: Bug
Components: Rust
Reporter: Jörn Horstmann
The simd compare kernels use the following pattern to handle the remainder that
is not a multiple of the number of vector lanes:
{code}
if rem > 0 {
let simd_left = T::load(left.value_slice(len - rem, lanes));
let simd_right = T::load(right.value_slice(len - rem, lanes));
let simd_result = op(simd_left, simd_right);
let rem_buffer_size = (rem as f32 / 8f32).ceil() as usize;
T::bitmask(&simd_result, |b| {
result.extend_from_slice(&b[0..rem_buffer_size]);
});
}
{code}
While this avoids writing into result out of bounds, it still reads from the
{{left}} and {{right}} arrays at out of bounds indices and valgrind complains
about that. I propose to rewrite the logic to use chunked iteration, with a
scalar loop for the remainder, similar to the change for arithmetic kernels in
ARROW-10914.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
