[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183 rsandifo at gcc dot gnu.org changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #6 from rsandifo at gcc dot gnu.org --- Fixed.
[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183 --- Comment #5 from CVS Commits --- The master branch has been updated by Prathamesh Kulkarni : https://gcc.gnu.org/g:6b4c18b98127087d7f14062b81bc678f0589cd36 commit r12-4493-g6b4c18b98127087d7f14062b81bc678f0589cd36 Author: prathamesh.kulkarni Date: Tue Oct 19 13:51:51 2021 +0530 [sve] PR93183 - Add support for conditional neg. gcc/testsuite/ChangeLog: PR target/93183 * gcc.target/aarch64/sve/pr93183.c: Remove -mcpu=generic+sve from dg-options.
[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183 --- Comment #4 from CVS Commits --- The master branch has been updated by Prathamesh Kulkarni : https://gcc.gnu.org/g:20dcda98ed376cb61c74b2c71656f99c671ec9ce commit r12-4470-g20dcda98ed376cb61c74b2c71656f99c671ec9ce Author: prathamesh.kulkarni Date: Mon Oct 18 15:44:06 2021 +0530 [sve] PR93183 - Add support for conditional neg. gcc/ChangeLog: PR target/93183 * gimple-match-head.c (try_conditional_simplification): Add case for single operand. * internal-fn.def: Add entry for COND_NEG internal function. * internal-fn.c (FOR_EACH_CODE_MAPPING): Add entry for NEGATE_EXPR, COND_NEG mapping. * optabs.def: Add entry for cond_neg_optab. * match.pd (UNCOND_UNARY, COND_UNARY): New operator lists. (vec_cond COND (foo A) B) -> (IFN_COND_FOO COND A B): New pattern. (vec_cond COND B (foo A)) -> (IFN_COND_FOO ~COND A B): Likewise. gcc/testsuite/ChangeLog: PR target/93183 * gcc.target/aarch64/sve/cond_unary_4.c: Adjust. * gcc.target/aarch64/sve/pr93183.c: New test.
[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183
--- Comment #3 from prathamesh3492 at gcc dot gnu.org ---
(In reply to rsand...@gcc.gnu.org from comment #2)
> (In reply to Andrew Pinski from comment #1)
> > We get:
> > .L3:
> > ld1bz0.b, p0/z, [x1, x3]
> > movprfx z2, z0
> > and z2.b, z2.b, #0xc0
> > neg z1.b, p1/m, z0.b ;;; THIS
> > cmpeq p2.b, p1/z, z2.b, #0
> > sel z0.b, p2, z0.b, z1.b AND THIS
> > st1bz0.b, p0, [x0, x3]
> > incbx3
> > whilelo p0.b, w3, w2
> > b.any .L3
> >
> > The two instructions marked should be combined.
>
> The problem is that it isn't a straight combination of the
> NEG and SEL, because the condition is the inverse of the one
> that we want for predication.
IIUC, sel is redundant and we could generate following code
instead for the inner loop ?
ld1bz0.b, p0/z, [x1, x2]
movprfx z2, z0
and z2.b, z2.b, #0xc0
cmpne p2.b, p1/z, z2.b, #0
neg z0.b, p2/m, z0.b
st1bz0.b, p0, [x0, x3]
incbx3
whilelo p0.b, w3, w2
b.any .L3
>
> This is one of the things that the IFN_COND_* functions were
> designed to fix. We should probably add unary versions of those.
The input to isel pass is:
vect__3.11_39 = .MASK_LOAD (_22, 8B, loop_mask_38);
vect_t1_15.12_41 = vect__3.11_39 & { 192, ... };
vect_t_12.13_42 = -vect__3.11_39;
_44 = vect_t1_15.12_41 == { 0, ... };
vect_iftmp.14_45 = VEC_COND_EXPR <_44, vect__3.11_39, vect_t_12.13_42>;
where vect__3.11_39 and vect_t_12.13_42 are negatives of each other.
I suppose in isel pass if we come across vec_cond_expr of the form:
op2 = vec_cond_expr
then we could lower it to a new internal function say IFN_COND_NEG.
IFN_COND_NEG could use a new optab say cond_neg_optab to expand it to:
movprfx op2, op1
set predicate according to inverted cond
op2 = predicate/m neg op2
Does this look reasonable ?
Thanks,
Prathamesh
[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183 Richard Biener changed: What|Removed |Added Status|NEW |ASSIGNED
[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183 rsandifo at gcc dot gnu.org changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2020-01-07 CC||rsandifo at gcc dot gnu.org Ever confirmed|0 |1 --- Comment #2 from rsandifo at gcc dot gnu.org --- (In reply to Andrew Pinski from comment #1) > We get: > .L3: > ld1bz0.b, p0/z, [x1, x3] > movprfx z2, z0 > and z2.b, z2.b, #0xc0 > neg z1.b, p1/m, z0.b ;;; THIS > cmpeq p2.b, p1/z, z2.b, #0 > sel z0.b, p2, z0.b, z1.b AND THIS > st1bz0.b, p0, [x0, x3] > incbx3 > whilelo p0.b, w3, w2 > b.any .L3 > > The two instructions marked should be combined. The problem is that it isn't a straight combination of the NEG and SEL, because the condition is the inverse of the one that we want for predication. This is one of the things that the IFN_COND_* functions were designed to fix. We should probably add unary versions of those.
[Bug target/93183] SVE does not use neg as conditional
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93183
Andrew Pinski changed:
What|Removed |Added
Blocks||26163, 53947
Summary|SVE only makes the last |SVE does not use neg as
|thing conditional with |conditional
|ifconversion|
--- Comment #1 from Andrew Pinski ---
Actually it looks more like conditional neg does not work.
Take:
typedef unsigned char uint8_t;
static inline uint8_t x264_clip_uint8( uint8_t x )
{
uint8_t t = -x;
uint8_t t1 = x&(~63);
return t1 ? t : x;
}
void mc_weight( uint8_t *__restrict dst, uint8_t *__restrict src, int n)
{
for( int x = 0; x < n*16; x++ )
dst[x] = x264_clip_uint8(src[x]);
}
--- CUT ---
We get:
.L3:
ld1bz0.b, p0/z, [x1, x3]
movprfx z2, z0
and z2.b, z2.b, #0xc0
neg z1.b, p1/m, z0.b ;;; THIS
cmpeq p2.b, p1/z, z2.b, #0
sel z0.b, p2, z0.b, z1.b AND THIS
st1bz0.b, p0, [x0, x3]
incbx3
whilelo p0.b, w3, w2
b.any .L3
The two instructions marked should be combined.
NOTE this does show up in real code in SPEC CPU, see PR 92492.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26163
[Bug 26163] [meta-bug] missed optimization in SPEC (2k17, 2k and 2k6 and 95)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
