https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114339
--- Comment #6 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
vectorizer generates:
mask_patt_21.19_58 = vect_perm_even_49 >= vect_cst__57;
mask_patt_21.19_59 = vect_perm_even_55 >= vect_cst__57;
vexit_reduc_63 = mask_patt_21.19_58 | mask_patt_21.19_59;
if (vexit_reduc_63 != { 0, 0 })
goto <bb 14>; [20.00%]
else
goto <bb 5>; [80.00%]
This is changed at loopdone into:
delays[3].nonprimary_delay = 129600;
vect_cst__57 = {tdiff_6, tdiff_6};
mask_patt_21.19_58 = vect_cst__57 <= { 0, 0 };
mask_patt_21.19_59 = vect_cst__57 <= { 0, 0x7FFFFFFFFFFFFFFF };
vexit_reduc_63 = mask_patt_21.19_58 | mask_patt_21.19_59;
if (vexit_reduc_63 != { 0, 0 })
goto <bb 3>; [20.00%]
else
goto <bb 7>; [80.00%]
or in other words, if there's any value where the compare succeeds, find it and
return.
This looks correct to me.
It could be that my AVX is rusty but, this generates:
vmovdqa 0xf9c(%rip),%xmm1 # 0x402010
mov $0x1,%eax
vmovq %rcx,%xmm3
vmovdqa %xmm0,(%rsp)
vpunpcklqdq %xmm3,%xmm3,%xmm2
vmovdqa %xmm0,0x10(%rsp)
vpcmpgtq %xmm2,%xmm1,%xmm1#
vmovdqa %xmm0,0x20(%rsp)
vmovq %rax,%xmm0
vpunpcklqdq %xmm0,%xmm0,%xmm0
movl $0x1fa40,0x38(%rsp)
vpcmpgtq %xmm2,%xmm0,%xmm0
vpor %xmm1,%xmm0,%xmm0
vptest %xmm0,%xmm0
which looks off, particularly for the second compare it look like it doesn't do
a load but instead just duplicates the constant 1.
gdb seems to confirm this. At the first compare:
(gdb) p $xmm2.v2_int64
$4 = {10412095, 10412095}
(gdb) p $xmm0.v2_int64
$5 = {0, 0}
which is what's expected, but at the second compare:
(gdb) p $xmm2.v2_int64
$7 = {10412095, 10412095}
(gdb) p $xmm0.v2_int64
$6 = {1, 1}
at the second it's comparing {1, 1} instead of {0, 0x7FFFFFFFFFFFFFFF}.
on AArch64 where it doesn't fail the comparison is:
movi v29.4s, 0
add x1, sp, 16
ldr x5, [x0, 8]
mov w0, 64064
movk w0, 0x1, lsl 16
add x3, sp, 48
str q29, [sp, 64]
mov x2, 57407
mov x4, 9223372036854775807
str x4, [sp, 64]
movk x2, 0x9e, lsl 16
str w0, [sp, 72]
sub x2, x2, x5
stp q29, q29, [x1]
dup v27.2d, x2
ld2 {v30.2d - v31.2d}, [x1]
str q29, [sp, 48]
ld2 {v28.2d - v29.2d}, [x3]
cmge v30.2d, v30.2d, v27.2d
cmge v28.2d, v28.2d, v27.2d
orr v30.16b, v30.16b, v28.16b
umaxp v30.4s, v30.4s, v30.4s
fmov x0, d30
cbnz x0, .L12
which has v30.2d being {0, 0} and v28.2d being {0, 0x7FFFFFFFFFFFFFFF} as
expected...
On AArch64 we don't inline the constants because whatever is propagating the
constants can't understand the LOAD_LANES:
mask_patt_19.21_50 = vect__2.16_44 >= vect_cst__49;
mask_patt_19.21_51 = vect__2.19_47 >= vect_cst__49;
vexit_reduc_55 = mask_patt_19.21_50 | mask_patt_19.21_51;
if (vexit_reduc_55 != { 0, 0 })
goto <bb 3>; [20.00%]
else
goto <bb 7>; [80.00%]
so could this be another expansion bug?
Note that a simpler reproducer is this:
---
long tdiff = 10412095;
int main() {
struct {
long maximum;
int nonprimary_delay;
} delays[] = {{}, {}, {}, {9223372036854775807, 36 * 60 * 60}};
for (unsigned i = 0; i < sizeof(delays) / sizeof(delays[0]); ++i)
if (tdiff <= delays[i].maximum)
return delays[i].nonprimary_delay;
__builtin_abort();
}
---
the key point is that we're not allowed to constprop tdiff at GIMPLE. If we do,
e.g:
int main() {
struct {
long maximum;
int nonprimary_delay;
} delays[] = {{}, {}, {}, {9223372036854775807, 36 * 60 * 60}};
long tdiff = 10412095;
for (unsigned i = 0; i < sizeof(delays) / sizeof(delays[0]); ++i)
if (tdiff <= delays[i].maximum)
return delays[i].nonprimary_delay;
__builtin_abort();
}
then after vectorization the const prop the entire expression is evaluated at
GIMPLE and it gets the right result.
This makes me believe it's a target expansion bug.