[PATCH v1] x86: memcmp-avx2-movbe.S and memcmp-evex-movbe.S fix overflow bug.

[Noah Goldstein]

On Mon, Jun 7, 2021 at 1:45 PM Paul Eggert <eggert@cs.ucla.edu> wrote:

> On 6/7/21 10:28 AM, Noah Goldstein via Libc-alpha wrote:
> >
> > Is it actually UB? The caller is not causing overflow. The implementation
> > method is. It is possible to implement without overflow.
>
> The C Standard says that unless otherwise specified, when you pass an
> array (pointer + size) to a standard function, all the addresses in the
> array must be valid. It's valid if (say) memcmp is multithreaded and
> compares the first halves of the two arrays in parallel with comparing
> the second halves.
>

Does this mean there is an issue with functions like wcsnlen?

They do say that they only need to be able to access memory up
to first null terminator but currently the x86_64 wcsnlen-avx2.S
implementation will multiple length by 4 which could cause overflow.

For example with a string length 1000 and maxlen passed as
2^62 + 1 the return will be 1.

Is that an issue? It's a pretty simple fix although I think this idiom
of multiply length by 4 to get byte count is in a lot of similarly specified
files.


>
> If I understand things correctly this patch isn't fixing a conformance
> bug; it's merely a QoI issue, where by "quality" one means "I want this
> particular undefined behavior to cause a core dump".
>