[compiler-rt] r215295 - builtins: correct __umodsi3, __udivsi3 on ARM

[Renato Golin]

So, we have look and we believe that this time, it's correct. Saleem's
patch does indeed fixes the two blatant bugs that we introduced by
copy&paste.

Jim/Bill, we also believe this fix needs to be in 3.5, as things will
be completely broken if not. Can I merge it back?

cheers,
--renato

On 19 August 2014 23:04, Renato Golin <renato.golin at linaro.org> wrote:
> On 19 August 2014 22:55, Anton Korobeynikov <anton at korobeynikov.info> wrote:
>> Did some git / svn archaeology: support for cores with division
>> instructions was added in r182665 / a09d09d29e
>>
>> And the code there (see e.g.
>> http://llvm.org/klaus/compiler-rt/commit/a09d09d29e250e905bdfaf819979b9c3e9adc047/)
>> is precisely the same as in Saleem's latest commit.
>>
>> Looks like in r200001 umodsi3.S was completely replaced (!) by
>> udivmodsi4.S making it broken.
>>
>> CC'ing Joerg as he was the author of that commit.
>
> I think we need to look at this one carefully...
>
> Joerg may correct me, but I believe the intention was to implement a
> divmod, that happened to be in a badly named file (umod) and with a
> copy&paste mistake on the function name. So that one should actually
> be the divmod function, not the mod, and the latest Saleem's patch
> actually finished killing the original implementation at all, meaning
> that it's still divmod for armv4 but not for the rest, but being used
> as div/mod.
>
> I can't believe we used to pass all tests with compiler-rt on ARM
> about two months ago with that all messed up!
>
> This is the review in Phab:
>
> http://reviews.llvm.org/D2595
>
> cheers,
> --renato