https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126770
Bug ID: 126770
Summary: inefficient reduction of comparison sum
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Target: x86_64-*-*
For the tail reduction in PR126028,
bool test0 = (r2_0 < groupplcutoff2);
bool test1 = (r2_1 < groupplcutoff2);
hu += test0 + test1;
we generate awful code with -O3 -march=znver5 which can be seen with the
following reduced testcase (which needs -fno-vect-cost-model):
int foo (double g, double *r, int n)
{
int hu = 0;
bool test0 = r[0] < g;
bool test1 = r[1] < g;
hu += test0 + test1;
return hu;
}
we generate
foo:
.LFB0:
.cfi_startproc
vmovupd (%rdi), %xmm1
vmovddup %xmm0, %xmm0
movl $1, %eax
vcmppd $1, %xmm0, %xmm1, %k1
vpbroadcastq %rax, %xmm0{%k1}{z}
vpmovqd %xmm0, %xmm0
vmovd %xmm0, %eax
vpextrd $1, %xmm0, %edx
addl %edx, %eax
which is in .optimized
_17 = {g_8(D), g_8(D)};
vect__1.5_16 = MEM <vector(2) double> [(double *)r_7(D)];
mask_test0_9.6_18 = vect__1.5_16 < _17;
vect_patt_11.7_19 = .VCOND_MASK (mask_test0_9.6_18, { 1, 1 }, { 0, 0 });
vect_patt_12.8_20 = (vector(2) int) vect_patt_11.7_19;
_22 = BIT_FIELD_REF <vect_patt_12.8_20, 32, 0>;
_23 = BIT_FIELD_REF <vect_patt_12.8_20, 32, 32>;
_24 = _22 + _23;
on a high-level the reduction of the mask as 1/0 could be matched as
population count. But on a low-level STV could have figured that
doing the addl in SSE and the SSE->GPR move only once should be prefered
and similar the movl $1,%eax; vpbroadcastq %rax, %xmm0{%k1}{z}
sequence looks like there must be a more optimal form, not to name
the vpmovqd.
I realize most of this should probably happen on the GIMPLE level.
The first mistake is from the vectorizer and the bool pattern recognition of
_1 = *r_7(D);
test0_9 = _1 < g_8(D);
_2 = MEM[(double *)r_7(D) + 8B];
test1_10 = _2 < g_8(D);
_3 = (int) test0_9;
_4 = (int) test1_10;
where it choses a COND_EXPR with a unsigned long. I'll open a separate PR
for that.
