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

[Noah Goldstein]

On Mon, Jun 7, 2021 at 10:21 AM Siddhesh Poyarekar <siddhesh@gotplt.org>
wrote:

> On 6/7/21 2:00 PM, Noah Goldstein via Libc-alpha wrote:
> > Fix bugs introducted in commits:
> >
> > author        Noah Goldstein <goldstein.w.n@gmail.com>
> > Mon, 17 May 2021 17:57:24 +0000 (13:57 -0400)
> > commit        4ad473e97acdc5f6d811755b67c09f2128a644ce
> >
> > And
> >
> > author        Noah Goldstein <goldstein.w.n@gmail.com>
> > Mon, 17 May 2021 17:56:52 +0000 (13:56 -0400)
> > commit        16d12015c57701b08d7bbed6ec536641bcafb428
> >
> > Which added a bug which would cause pointer + length overflow to lead
> > to an early return as opposed to a Segmentation Fault.
>
> If we end up making this change, IMO it should come with an explicit
> note that this behaviour is not guaranteed for invalid inputs in other
> implementations of memcmp or for that matter, in future versions of this
> memcmp.
>
> An input that causes pointer + length overflow is undefined behaviour
> and IMO we shouldn't try to define it for glibc.


Is it actually UB? The caller is not causing overflow. The implementation
method is. It is possible to implement without overflow.


> This change may not
> have a noticeable performance impact


Minor impact actually for memcmp.


> but future requests to guarantee
> this behaviour may not necessarily be this straightforward and IMO we
> should not bind ourselves to it.


Agreed.


>
> Siddhesh
>