Re: [PATCH] middle-end: fix de-optimizations with bitclear patterns on signed values

[Jonathan Wakely <jwakely@redhat.com>]

On Tue, 26 Oct 2021 at 14:36, Jakub Jelinek <jakub@redhat.com> wrote:

> On Tue, Oct 26, 2021 at 03:21:55PM +0200, Richard Biener wrote:
> > On Tue, 26 Oct 2021, Jakub Jelinek wrote:
> >
> > > On Tue, Oct 26, 2021 at 03:13:29PM +0200, Richard Biener wrote:
> > > > try
> > > >   auto c = ...;
> > > >   signed char c2 = c;
> > > >   return c2 >= ...
> > > > then
> > >
> > > That won't work, at least when using <compare>, which is what we with
> the
> > > optimization want to deal with primarily.
> > > Because std::partial_ordering etc. aren't implicitly nor explicitly
> > > convertible to int or signed char etc.
> > > Sure, one could in the testcase define its own std::strong_ordering
> etc.
> > > and define a conversion operator for it...
> >
> > So how do we end up with the signed char case in the first place?
> > Is the frontend using a type that's target dependent?
>
> <compare> uses explicitly signed char:
> namespace std
> {
>   // [cmp.categories], comparison category types
>   namespace __cmp_cat
>   {
>     using type = signed char;
>     enum class _Ord : type { equivalent = 0, less = -1, greater = 1 };
>     enum class _Ncmp : type { _Unordered = 2 };
> ...
> and __cmp_cat::type is what is used as type of _M_value of std::*_ordering
> -fsigned-char vs. -funsigned-char make no difference on the testcases on
> x86, but as mentioned in
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94589#c24
> some target decisions like load_extend_op uses in fold-const.c can affect
> it.  See https://gcc.gnu.org/pipermail/gcc-patches/2021-May/570714.html
>
>

We can change __cmp_cat::type if that would result in better code. I picked
signed char because we only need two bits, and preferably have a signed
type as it simplifies some things. Would int make more sense? Or int:2 ?