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

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Wed, Jun 9, 2021 at 2:48 AM Siddhesh Poyarekar <siddhesh@gotplt.org>
wrote:

> On 6/9/21 12:02 PM, Noah Goldstein wrote:
> > Wait are you saying that the return value overflowing causes UB or
> > that the act of passing a maxlen where s + maxlen * sizeof(wchar_t) is
> > outside range object is UB?
>
> Not specifically the return value, but more broadly, causing wcsnlen to
> invoke undefined behaviour, which could include the former.
>
> > If the former then why is the follow okay:
> >
> > Previous example with a string whose length is 1000
> > but because wcslen is passed maxlen where maxlen * sizeof(wchar_t)
> > overflows and leads to a result less than 1000 the implementation of
> > wcslen in wcsnlen-avx2.S will return a length less than 1000.
> >
> >
> > If the latter then:
> >
> > For test-wcsnlen which redirects to test-strnlen
> > If the UB is when is s+maxlen*sizeof(wchar_t) is outside object bound
> > then s + SIZE_MAX + sizeof(wchar_t) surely is.
> >
> > Although even then test-strnlen s + SIZE_MAX will also overflow if s non
> > null.
>
> That is a good catch; I did not notice that.  That should be SIZE_MAX /
> sizeof (CHAR).
>

Do we want to support s + maxlen + sizeof(CHAR) overflowing? If not we can
speed up the AVX2/EVEX implementation of strnlen/wcsnlen/memchr/wmemchr.

(Assuming memchr is of same standard where s + n + sizeof(CHAR)
can't overflow. If not it has this bug).


> Thanks,
> Siddhesh
>