This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH], PowerPC: Allow DImode in Altivec registers
- From: Michael Meissner <meissner at linux dot vnet dot ibm dot com>
- To: Segher Boessenkool <segher at kernel dot crashing dot org>
- Cc: Michael Meissner <meissner at linux dot vnet dot ibm dot com>, gcc-patches at gcc dot gnu dot org, David Edelsohn <dje dot gcc at gmail dot com>, Bill Schmidt <wschmidt at linux dot vnet dot ibm dot com>
- Date: Wed, 15 Jun 2016 17:12:44 -0400
- Subject: Re: [PATCH], PowerPC: Allow DImode in Altivec registers
- Authentication-results: sourceware.org; auth=none
- References: <20160613172816 dot GA23271 at ibm-tiger dot the-meissners dot org> <20160613182941 dot GA3747 at ibm-tiger dot the-meissners dot org> <20160614225346 dot GD13674 at gate dot crashing dot org> <20160615182441 dot GA9771 at ibm-tiger dot the-meissners dot org> <20160615195120 dot GA16348 at gate dot crashing dot org>
On Wed, Jun 15, 2016 at 02:51:20PM -0500, Segher Boessenkool wrote:
> On Wed, Jun 15, 2016 at 02:24:41PM -0400, Michael Meissner wrote:
> > > > ; Some DImode loads are best done as a load of -1 followed by a mask
> > > > ; instruction.
> > > > (define_split
> > > > - [(set (match_operand:DI 0 "gpc_reg_operand")
> > > > + [(set (match_operand:DI 0 "int_reg_operand_not_pseudo")
> > >
> > > Not sure what this is for... If you want to say this split is only to
> > > be done after RA, just say that explicitly in the split condition (i.e.
> > > "reload_completed"). Or does this mean something else?
> >
> > This is so that constants being loaded into the vector registers aren't split
> > (they are handled via different define_splits). Previously, the only constant
> > that was loaded in vector registers was 0.
> >
> > The int_reg_operand_not_pseudo allows the split to take place if it has already
> > gotten hard registers before register allocation.
>
> When does that happen?
Using arguments, function returns, and of course explicit registers, but I
agree it is fairly low.
> > It could have been the
> > normal int_reg_operand and then use a reload_completed check.
>
> That is preferred if it makes no difference (otherwise, bebfore you know
> it we'll have twice as many predicates).
We already had the predicate for another use, so I wasn't adding a new one.
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meissner@linux.vnet.ibm.com, phone: +1 (978) 899-4797