RFC [1/2] divmod transform

[Richard Biener]

On Tue, 31 May 2016, Prathamesh Kulkarni wrote:

> On 30 May 2016 at 13:15, Richard Biener <rguenther@suse.de> wrote:
> > On Mon, 30 May 2016, Prathamesh Kulkarni wrote:
> >
> >> The attached patch ICE's during bootstrap for x86_64, and is reproducible with
> >> following case with -m32 -O2:
> >>
> >> typedef long long type;
> >>
> >> type f(type x, type y)
> >> {
> >>   type q = x / y;
> >>   type r = x % y;
> >>   return q + r;
> >> }
> >>
> >> The ICE happens because the test-case hits
> >> gcc_assert (unsignedp);
> >> in default_expand_divmod_libfunc ().
> >
> > That's of course your function (and ICE).
> >
> >> Surprisingly, optab_libfunc (sdivmod_optab, DImode) returns optab_libfunc
> >> with name "__divmoddi4" although __divmoddi4() is nowhere defined in
> >> libgcc for x86.
> >> (I verified that by forcing the patch to generate call to __divmoddi4,
> >> which results in undefined reference to __divmoddi4).
> >>
> >> This happens because in optabs.def we have:
> >> OPTAB_NL(sdivmod_optab, "divmod$a4", UNKNOWN, "divmod", '4', gen_int_libfunc)
> >>
> >> and gen_int_libfunc generates "__divmoddi4" on first call to optab_libfunc
> >> and sets optab_libfunc (sdivmod_optab, DImode) to "__divmoddi4".
> >> I wonder if we should remove gen_int_libfunc entry in optabs.def for
> >> sdivmod_optab ?
> >
> > Hum, not sure - you might want to look at expand_divmod (though that
> > always just computes one part of the result in the end).
> As a workaround, would it be OK to check if libfunc is __udivmoddi4
> if expand_divmod_libfunc is default, as in attached patch ?

Humm.  I suppose it is because until now we never expanded divmod
directly but only the modulo and division libfuncs are all 
implemented in terms of udivmoddi4.

This means that the signed divmod libfunc expander needs to use
that as well - not sure if that's efficiently possible though.
Which would mean to simply never consider divmoddi4 as available.

Richard.

> This prevents ICE for the above test-case.
> Bootstrap+test on x86_64 in progress.
> 
> Thanks,
> Prathamesh
> >
> > Joseph - do you know sth about why there's not a full set of divmod
> > libfuncs in libgcc?
> >
> > Thanks,
> > Richard.
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)