This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: RFC: generic vectors and GPR/simd duality
On Wed, May 12, 2004 at 11:54:39AM -0400, Aldy Hernandez wrote:
> On Wed, May 12, 2004 at 11:49:49AM -0400, Daniel Jacobowitz wrote:
> > On Wed, May 12, 2004 at 10:58:00AM -0400, Aldy Hernandez wrote:
> > > mode_multiple from V2SI to V4SI is 2, whereas nregs_multiple is 4.
> > > GPR #26 can hold a V2SI in just one register, but needs 4 for a V4SI.
> > > In other architectures, a given SIMD register can hold only one type
> > > of simd, whereas the e500 can hold any (albeit in subsequent registers
> > > for anything not V2SI). Confusion.
> >
> > Why does V4SI take four GPRs, that's the question. If you treat them
> > as V2SI registers GCC ought to be able to synthesize the V4SI as two
> > V2SIs. It might be as simple as "fixing" HARD_REGNO_NREGS... probably
> > not, though.
>
> Shudder. I can almost guarantee you that that's going to create all
> sorts of havoc. I'm willing to give it a whirl, but watch the pain...
My brain hurts.
When I change HARD_REGNO_NREGS, reload does some gymnastics of mode
changing, ending up with:
(subreg:SI (reg:V4SI 11 11) 0)
Now, SI is 32 bits. V4SI is 128 bits. But HARD_REGNO_NREGS now
returns 2 for a V4SI. You can imagine the confusion wrought by all
this.
The more I think about it, the more I want to remove the creation of
the subregs in the first place.
Aldy