[Bug middle-end/88487] union prevents autovectorization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487
--- Comment #6 from Daniel Fruzynski ---
Not good. Fortunately I found workaround. This is probably the best what one
can get:
[code]
#include
#include
template
struct TypeHelper
{
constexpr unsigned offset();
operator Type&()
{
uint8_t*__restrict p = (uint8_t*__restrict)this - offset();
Type*__restrict pt = (Type*__restrict)p;
return *pt;
}
};
struct S
{
struct Union
{
void*__restrict*__restrict ptr;
TypeHelper d;
} u;
};
template<>
constexpr unsigned TypeHelper::offset()
{
return offsetof(S::Union, d) - offsetof(S::Union, ptr);
}
void test(S* __restrict s1, S* __restrict s2)
{
for (int n = 0; n < 2; ++n)
{
s1->u.d[n][0] = s2->u.d[n][0];
s1->u.d[n][1] = s2->u.d[n][1];
}
}
[/code]
[Bug middle-end/88487] union prevents autovectorization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487 --- Comment #5 from rguenther at suse dot de --- On Fri, 14 Dec 2018, bugzi...@poradnik-webmastera.com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487 > > --- Comment #4 from Daniel Fruzynski --- > OK, I see. Is there any workaround for this? Only not using a union ... > I tried to assign pointer to local > variable directly and with intermediate casting via void*, but it did not > help. > Casting S1* to S2* also does not work. Yes, that doesn't work by design.
[Bug middle-end/88487] union prevents autovectorization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487 --- Comment #4 from Daniel Fruzynski --- OK, I see. Is there any workaround for this? I tried to assign pointer to local variable directly and with intermediate casting via void*, but it did not help. Casting S1* to S2* also does not work.
[Bug middle-end/88487] union prevents autovectorization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487
Richard Biener changed:
What|Removed |Added
Keywords||missed-optimization
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed||2018-12-14
Blocks||49774
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Ever confirmed|0 |1
--- Comment #3 from Richard Biener ---
t.c:22:1: note: can't determine dependence between MEM[(double *)_28 clique 1
base 0] and MEM[(double *)_25 + 8B clique 1 base 0]
t.c:22:1: note: removing SLP instance operations starting from: MEM[(double
*)_28 clique 1 base 0] = _29;
t.c:22:1: note: can't determine dependence between MEM[(double *)_43 clique 1
base 0] and MEM[(double *)_40 + 8B clique 1 base 0]
t.c:22:1: note: removing SLP instance operations starting from: MEM[(double
*)_43 clique 1 base 0] = _44;
# PT = nonlocal escaped null
_25 = MEM[(double * restrict *)_21 clique 1 base 0];
# PT = nonlocal escaped null
_28 = MEM[(double * restrict *)_26 clique 1 base 0];
..
MEM[(double *)_28 clique 1 base 0] = _29;
_31 = MEM[(double *)_25 + 8B clique 1 base 0];
while w/o unions we have
# PT = null { D.2686 } (nonlocal, restrict)
_25 = MEM[(double * restrict *)_21 clique 1 base 1];
..
_29 = MEM[(double *)_25 clique 1 base 2];
that is, the indirect loads from non-union members produce restricted
pointers while those from union members not.
The reason for this is that points-to analysis doesn't handle unions
in field-sensitive analysis and thus the restrict code doesn't apply.
This can probably be fixed in a reasonable manner in
push_fields_onto_fieldstack
by initializing only_restrict_pointers appropriately for the UNION case.
Not really my top-priority though.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=49774
[Bug 49774] [meta-bug] restrict qualification aliasing issues
[Bug middle-end/88487] union prevents autovectorization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487 --- Comment #2 from Daniel Fruzynski --- I spotted that test3 in previous comment uses structure S2 which does not have union inside. When I changes it to use S1, I got non-vectorized code. So this workaround does not work.
[Bug middle-end/88487] union prevents autovectorization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88487
--- Comment #1 from Daniel Fruzynski ---
Update: when pointers to data are copied to local variables like below,
autovectorization starts working again.
[code]
void test3(S2* __restrict__ s1, S2* __restrict__ s2)
{
double* __restrict__ * __restrict__ d1 = s1->d;
double* __restrict__ * __restrict__ d2 = s2->d;
for (int n = 0; n < 2; ++n)
{
d1[n][0] = d2[n][0];
d1[n][1] = d2[n][1];
}
}
[/code]
