On Thu, 19 Aug 2021, Richard Sandiford wrote:
> Richard Biener writes:
> > On Wed, 18 Aug 2021, Richard Sandiford wrote:
> >> I think it would be OK/sensible to use the larger of the index or
> >> result vectors to determine the mask, if that helps. There just
> >> wasn't any need to make a dist
Richard Biener writes:
> On Wed, 18 Aug 2021, Richard Sandiford wrote:
>> I think it would be OK/sensible to use the larger of the index or
>> result vectors to determine the mask, if that helps. There just
>> wasn't any need to make a distinction for SVE, since there the
>> mask type is determin
On Wed, 18 Aug 2021, Richard Sandiford wrote:
> Richard Biener writes:
> > On Wed, 18 Aug 2021, Hongtao Liu wrote:
> >
> >> On Wed, Aug 18, 2021 at 7:37 PM Hongtao Liu wrote:
> >> >
> >> > On Wed, Aug 18, 2021 at 7:30 PM Hongtao Liu wrote:
> >> > >
> >> > > On Wed, Aug 18, 2021 at 6:28 PM Richa
Richard Biener writes:
> On Wed, 18 Aug 2021, Hongtao Liu wrote:
>
>> On Wed, Aug 18, 2021 at 7:37 PM Hongtao Liu wrote:
>> >
>> > On Wed, Aug 18, 2021 at 7:30 PM Hongtao Liu wrote:
>> > >
>> > > On Wed, Aug 18, 2021 at 6:28 PM Richard Biener wrote:
>> > > >
>> > > > On Wed, 18 Aug 2021, Richar
On Wed, 18 Aug 2021, Hongtao Liu wrote:
> On Wed, Aug 18, 2021 at 7:37 PM Hongtao Liu wrote:
> >
> > On Wed, Aug 18, 2021 at 7:30 PM Hongtao Liu wrote:
> > >
> > > On Wed, Aug 18, 2021 at 6:28 PM Richard Biener wrote:
> > > >
> > > > On Wed, 18 Aug 2021, Richard Biener wrote:
> > > >
> > > > >
target = gen_reg_rtx (V4SFmode);
> > emit_insn (gen_vec_extract_lo_v8sf (target, subtarget));
> > ---cut from i386-expand.c---
> >
> > and do ugly things when expand builtin
> >
> > > From the ISA docs I conclude that vgatherqps is not supported for
t; >(V2DF "V4SI") (V4DF "V4SI") (V8DF "V8SI")
> > (V4SI "V4SI") (V8SI "V8SI") (V16SI "V16SI")
> >(V4SF "V4SI") (V8SF "V8SI") (V16SF
F "V8DI") (V16SF "V8DI")]
Hum, it seems to be more complicated (or simpler). The gatherq
instructions will always fill only half of the SFmode destination
while the gatherd will only use half of the index vector for a DFmode
destination.
So we support mask_gather_load
gatherdi"
> [(parallel [(set (match_operand:VI48F 0 "register_operand")
> (unspec:VI48F
>
> "supports" V16SI and V16SF gathers, leading to broken patterns?
>
> I can test if anything goes wrong fixing VEC_GATHER_IDXDI to
>
> (define
nce
> > AVX doesn't have scatter).
> >
> > Thanks,
> > Richard.
> >
> > From 706d3ac96b384eaad249cc83ec542ec643e21a4c Mon Sep 17 00:00:00 2001
> > From: Richard Biener
> > Date: Tue, 17 Aug 2021 15:14:38 +0200
> > Subject: [PATCH] move x86 to use gather/scatter internal
ster_operand")
(unspec:VI48F
"supports" V16SI and V16SF gathers, leading to broken patterns?
I can test if anything goes wrong fixing VEC_GATHER_IDXDI to
(define_mode_attr VEC_GATHER_IDXDI
[(V2DI "V2DI") (V4DI "V4DI") (V8
43e21a4c Mon Sep 17 00:00:00 2001
> From: Richard Biener
> Date: Tue, 17 Aug 2021 15:14:38 +0200
> Subject: [PATCH] move x86 to use gather/scatter internal functions
> To: gcc-patches@gcc.gnu.org
>
> This moves the x86 backend to use standard pattern names for
> mask_gather
below? If not I'll
do mask_scatter_store accordingly (that will be simpler since
AVX doesn't have scatter).
Thanks,
Richard.
>From 706d3ac96b384eaad249cc83ec542ec643e21a4c Mon Sep 17 00:00:00 2001
From: Richard Biener
Date: Tue, 17 Aug 2021 15:14:38 +0200
Subject: [PATCH] move x86 to
On Wed, 18 Aug 2021, Hongtao Liu wrote:
> On Wed, Aug 18, 2021 at 4:32 PM Richard Biener wrote:
> >
> > On Wed, 18 Aug 2021, Hongtao Liu wrote:
> >
> > > On Wed, Aug 18, 2021 at 11:24 AM Hongtao Liu wrote:
> > > >
> > > > On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
> > > > w
On Wed, Aug 18, 2021 at 4:32 PM Richard Biener wrote:
>
> On Wed, 18 Aug 2021, Hongtao Liu wrote:
>
> > On Wed, Aug 18, 2021 at 11:24 AM Hongtao Liu wrote:
> > >
> > > On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
> > > wrote:
> > > >
> > > > On Tue, Aug 17, 2021 at 3:29 PM Ric
On Wed, 18 Aug 2021, Hongtao Liu wrote:
> On Wed, Aug 18, 2021 at 11:24 AM Hongtao Liu wrote:
> >
> > On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
> > wrote:
> > >
> > > On Tue, Aug 17, 2021 at 3:29 PM Richard Biener via Gcc-patches
> > > wrote:
> > > >
> > > > This is an att
On Wed, Aug 18, 2021 at 11:24 AM Hongtao Liu wrote:
>
> On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
> wrote:
> >
> > On Tue, Aug 17, 2021 at 3:29 PM Richard Biener via Gcc-patches
> > wrote:
> > >
> > > This is an attempt to start moving the x86 backend to use
> > > standard
On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
wrote:
>
> On Tue, Aug 17, 2021 at 3:29 PM Richard Biener via Gcc-patches
> wrote:
> >
> > This is an attempt to start moving the x86 backend to use
> > standard pattern names for [mask_]gather_load and [mask_]scatter_store
> > rathe
On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
wrote:
>
> On Tue, Aug 17, 2021 at 3:29 PM Richard Biener via Gcc-patches
> wrote:
> >
> > This is an attempt to start moving the x86 backend to use
> > standard pattern names for [mask_]gather_load and [mask_]scatter_store
> > rathe
On Tue, Aug 17, 2021 at 3:29 PM Richard Biener via Gcc-patches
wrote:
>
> This is an attempt to start moving the x86 backend to use
> standard pattern names for [mask_]gather_load and [mask_]scatter_store
> rather than using the builtin_{gather,scatter} target hooks.
>
> I've started with AVX2 gat
This is an attempt to start moving the x86 backend to use
standard pattern names for [mask_]gather_load and [mask_]scatter_store
rather than using the builtin_{gather,scatter} target hooks.
I've started with AVX2 gathers and given x86 only supports masked
gather I only implemented mask_gather_load
21 matches
Mail list logo
