https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125941
Tamar Christina <tnfchris at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |tnfchris at gcc dot gnu.org
--- Comment #1 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
Which particular part are you referring to?
the testcase is
#include <arm_sve.h>
#define VECT_CSTN { -1, t (), 0, -1, 0, f (), 0, 0, 0, -1, 0, -1, 0, -1, 0, -1
}
int __attribute__ ((noipa))
t ()
{
return -1;
}
int __attribute__ ((noipa))
f ()
{
return 0;
}
svbool_t __attribute__ ((noipa))
func_init4 ()
{
svbool_t temp = VECT_CSTN;
return temp;
}
so it's trying to create a boolean vector from a data-vector. Without SVE2p1
the only
way to create this is through memory or another operation like compare. with
SVE2p1 it can use a PMOV.
But if you're referring to the register preservations:
str z8, [sp, #2, mul vl]
str z9, [sp, #3, mul vl]
str z10, [sp, #4, mul vl]
str z11, [sp, #5, mul vl]
str p15, [sp, #11, mul vl]
str p5, [sp, #1, mul vl]
str p6, [sp, #2, mul vl]
str p7, [sp, #3, mul vl]
str p8, [sp, #4, mul vl]
str p9, [sp, #5, mul vl]
str p10, [sp, #6, mul vl]
str p11, [sp, #7, mul vl]
str p12, [sp, #8, mul vl]
str p13, [sp, #9, mul vl]
str p14, [sp, #10, mul vl]
str z12, [sp, #6, mul vl]
str z13, [sp, #7, mul vl]
str z14, [sp, #8, mul vl]
str z15, [sp, #9, mul vl]
str z16, [sp, #10, mul vl]
str z17, [sp, #11, mul vl]
str z18, [sp, #12, mul vl]
str z19, [sp, #13, mul vl]
str z20, [sp, #14, mul vl]
str z21, [sp, #15, mul vl]
str p4, [sp]
str z22, [sp, #16, mul vl]
str z23, [sp, #17, mul vl]
This has to do with the PCS of the functions
func_init4 returns an svbool_t so it's calling convention is aarch64_pcs_sve.
but t and f do not use an SVE type, so their calling convention is the base
pcs.
that means that to call t and f it has to preserve the state of the sve pcs
and restore it. This is what's causing the saves. If you make t and f use the
sve PCS you'll notice they go away
https://godbolt.org/z/G8xcaxrjs
The extra compares is because the element types of an svbool_t are booleans so
they need to be converted to a C boolean first.
So if the ticket is about the spills then that's not an issue. if it's about
adding
PMOV support, then yes :)
