On Mon, Feb 11, 2019 at 07:17:16AM -0600, Bill Schmidt wrote:
> At -O0 (if I hand-inline everything myself to avoid errors), we scalarize
> the modulo/masking operation into a rldicl for each doubleword. I really
> don't see any reason to change the code.
So what does this look like at expand (at -O0)? Is it something that
is done at gimple level, is it expand itself, is it some target thing?
> >> For -mcpu=power9, we get close, but have some bad register allocation and
> >> an unnecessary extend:
> >>
> >> xxspltib 0,4 <- why not just xxspltib 32,4?
> >> xxlor 32,0,0 <- wasted copy
> > Yeah, huh. Where does that come from... I blame splitters after reload.
>
> This only happens at -O2 and up, FWIW. At -O1 we allocate the registers
> reasonably.
Heh.
> >> Weird. I just tried adding -mvsx
> > Does it _need_ VSX anyway? Are these builtins defined without it, too?
>
> Yes (vector long long / V2DImode requires VSX).
So something like
/* { dg-do run } */
/* { dg-require-effective-target vsx_hw } */
/* { dg-options "-mvsx -O2" } */
then?
> I pointed to the bugzilla in another reply -- which was "resolved" with a
> hack.
> I consider it still broken this way...
Reopen that PR?
> I tested a revised version of the patch overnight and will submit shortly.
Thanks.
Segher
