https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116582
--- Comment #6 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Here is a variant of benchmark that needs masking
#include <stdlib.h>
#define M 1024*1024
T a[M], b[M];
int indices[M];
char c[M];
__attribute__ ((noipa))
void
test ()
{
for (int i = 0; i < 1024* 16; i++)
if (c[i])
a[i] += b[indices[i]];
}
int
main()
{
for (int i = 0 ; i < M; i++)
{
indices[i] = rand () % M;
c[i] = rand () % 2;
}
for (int i = 0 ; i < 10000; i++)
test ();
return 0;
}
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
281.03 msec task-clock:u # 0.999 CPUs
utilized ( +- 0.62% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
659 page-faults:u # 2.345 K/sec
( +- 0.06% )
1,156,011,975 cycles:u # 4.113 GHz
( +- 0.65% )
757,216,769 stalled-cycles-frontend:u # 65.50% frontend
cycles idle ( +- 1.59% )
1,292,982,312 instructions:u # 1.12 insn per
cycle
# 0.59 stalled cycles per
insn ( +- 0.00% )
360,669,069 branches:u # 1.283 G/sec
( +- 0.00% )
118,731 branch-misses:u # 0.03% of all
branches ( +- 8.51% )
0.28126 +- 0.00173 seconds time elapsed ( +- 0.62% )
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat -r 10 ./a.out
401241: 62 f2 7d 4d 92 1c 8d vgatherdps
0x904080(,%zmm1,4),%zmm3{%k5}
40125b: 62 f2 7d 4e 92 14 8d vgatherdps
0x904080(,%zmm1,4),%zmm2{%k6}
40126a: 62 f2 7d 4f 92 0c a5 vgatherdps
0x904080(,%zmm4,4),%zmm1{%k7}
401280: 62 f2 7d 4d 92 2c a5 vgatherdps
0x904080(,%zmm4,4),%zmm5{%k5}
Performance counter stats for './a.out' (10 runs):
266.73 msec task-clock:u # 0.999 CPUs
utilized ( +- 4.31% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
659 page-faults:u # 2.471 K/sec
( +- 0.05% )
1,097,343,324 cycles:u # 4.114 GHz
( +- 4.33% )
4,009,606 stalled-cycles-frontend:u # 0.37% frontend
cycles idle ( +- 6.91% )
241,592,306 instructions:u # 0.22 insn per
cycle
# 0.02 stalled cycles per
insn ( +- 0.00% )
35,549,063 branches:u # 133.279 M/sec
( +- 0.00% )
92,191 branch-misses:u # 0.26% of all
branches ( +- 0.06% )
0.2670 +- 0.0115 seconds time elapsed ( +- 4.30% )
so the difference in number of cycles is quite small while frontend works much
harder without gether. If c array is ocnstnat 1:
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
520.92 msec task-clock:u # 1.000 CPUs
utilized ( +- 5.29% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
659 page-faults:u # 1.265 K/sec
( +- 0.04% )
2,142,512,947 cycles:u # 4.113 GHz
( +- 5.31% )
137,707,449 stalled-cycles-frontend:u # 6.43% frontend
cycles idle ( +- 94.67% )
1,553,801,640 instructions:u # 0.73 insn per
cycle
# 0.09 stalled cycles per
insn ( +- 0.00% )
344,940,506 branches:u # 662.177 M/sec
( +- 0.00% )
58,418 branch-misses:u # 0.02% of all
branches ( +- 0.07% )
0.5212 +- 0.0276 seconds time elapsed ( +- 5.29% )
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat -r 10 ./a.out
401231: 62 f2 7d 4d 92 1c 8d vgatherdps
0x904080(,%zmm1,4),%zmm3{%k5}
40124b: 62 f2 7d 4e 92 14 8d vgatherdps
0x904080(,%zmm1,4),%zmm2{%k6}
40125a: 62 f2 7d 4f 92 0c a5 vgatherdps
0x904080(,%zmm4,4),%zmm1{%k7}
401270: 62 f2 7d 4d 92 2c a5 vgatherdps
0x904080(,%zmm4,4),%zmm5{%k5}
Performance counter stats for './a.out' (10 runs):
545.45 msec task-clock:u # 1.000 CPUs
utilized ( +- 5.28% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
658 page-faults:u # 1.206 K/sec
( +- 0.05% )
2,255,436,021 cycles:u # 4.135 GHz
( +- 5.30% )
6,191,015 stalled-cycles-frontend:u # 0.27% frontend
cycles idle ( +- 7.58% )
171,371,626 instructions:u # 0.08 insn per
cycle
# 0.04 stalled cycles per
insn ( +- 0.00% )
19,820,508 branches:u # 36.338 M/sec
( +- 0.00% )
58,434 branch-misses:u # 0.29% of all
branches ( +- 0.06% )
0.5457 +- 0.0288 seconds time elapsed ( +- 5.28% )
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
690.27 msec task-clock:u # 1.000 CPUs
utilized ( +- 3.07% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
659 page-faults:u # 954.698 /sec
( +- 0.04% )
2,862,994,747 cycles:u # 4.148 GHz
( +- 3.08% )
60,389,319 stalled-cycles-frontend:u # 2.11% frontend
cycles idle ( +- 92.61% )
1,477,776,699 instructions:u # 0.52 insn per
cycle
# 0.04 stalled cycles per
insn ( +- 0.00% )
328,232,959 branches:u # 475.514 M/sec
( +- 0.00% )
24,844 branch-misses:u # 0.01% of all
branches ( +- 1.40% )
0.6905 +- 0.0212 seconds time elapsed ( +- 3.07% )
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat -r 10 ./a.out
401271: 62 f2 7d 4d 92 1c 8d vgatherdps
0x904080(,%zmm1,4),%zmm3{%k5}
40128b: 62 f2 7d 4e 92 14 8d vgatherdps
0x904080(,%zmm1,4),%zmm2{%k6}
40129a: 62 f2 7d 4f 92 0c a5 vgatherdps
0x904080(,%zmm4,4),%zmm1{%k7}
4012b0: 62 f2 7d 4d 92 2c a5 vgatherdps
0x904080(,%zmm4,4),%zmm5{%k5}
Performance counter stats for './a.out' (10 runs):
849.59 msec task-clock:u # 1.000 CPUs
utilized ( +- 0.02% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
659 page-faults:u # 775.667 /sec
( +- 0.03% )
3,514,169,631 cycles:u # 4.136 GHz
( +- 0.01% )
20,390,142 stalled-cycles-frontend:u # 0.58% frontend
cycles idle ( +- 1.47% )
95,346,747 instructions:u # 0.03 insn per
cycle
# 0.21 stalled cycles per
insn ( +- 0.00% )
3,113,006 branches:u # 3.664 M/sec
( +- 0.00% )
24,535 branch-misses:u # 0.79% of all
branches ( +- 0.11% )
0.849857 +- 0.000176 seconds time elapsed ( +- 0.02% )
What confuses me is the branch mispredict count. Non-gather loop is:
0.60 │ ↓ je 6c
│ movslq 0x504080(,%rax,4),%rdx
1.19 │ vmovss 0x904080(,%rdx,4),%xmm0
84.09 │ vaddss 0xd04080(,%rax,4),%xmm0,%xmm0
8.77 │ vmovss %xmm0,0xd04080(,%rax,4)
1.62 │6c: inc %rax
3.03 │ cmp $0x4000,%rax
While gathering loop is:
│ 10:┌─→vmovdqa64 0x404080(%rax),%zmm1
│ │ vxorps %xmm3,%xmm3,%xmm3
│ │ vmovdqa32 0x504100(,%rax,4),%zmm4
│ │ vxorps %xmm2,%xmm2,%xmm2
0.15 │ │ vxorps %xmm5,%xmm5,%xmm5
│ │ vpcmpneqb %zmm6,%zmm1,%k1
│ │ vmovdqa32 0x504080(,%rax,4),%zmm1
0.07 │ │ kshiftrd $0x10,%k1,%k4
│ │ kmovw %k1,%k5
│ │ kshiftrq $0x20,%k1,%k2
│ │ kmovw %k4,%k6
│ │ kshiftrd $0x10,%k2,%k3
0.15 │ │ kmovw %k2,%k7
│ │ vgatherdps 0x904080(,%zmm1,4),%zmm3{%k5}
25.63 │ │ vmovdqa32 0x5040c0(,%rax,4),%zmm1
│ │ kmovw %k3,%k5
│ │ vgatherdps 0x904080(,%zmm1,4),%zmm2{%k6}
26.30 │ │ vxorps %xmm1,%xmm1,%xmm1
│ │ vgatherdps 0x904080(,%zmm4,4),%zmm1{%k7}
27.53 │ │ vmovdqa32 0x504140(,%rax,4),%zmm4
│ │ vgatherdps 0x904080(,%zmm4,4),%zmm5{%k5}
19.08 │ │ vmovaps 0xd04080(,%rax,4),%zmm4
│ │ vaddps %zmm4,%zmm3,%zmm4{%k1}
│ │ vmovaps 0xd040c0(,%rax,4),%zmm3
0.51 │ │ vmovaps %zmm4,0xd04080(,%rax,4)
│ │ vaddps %zmm3,%zmm2,%zmm3{%k4}
│ │ vmovaps 0xd04100(,%rax,4),%zmm2
│ │ vmovaps %zmm3,0xd040c0(,%rax,4)
│ │ vaddps %zmm2,%zmm1,%zmm2{%k2}
│ │ vmovaps 0xd04140(,%rax,4),%zmm1
0.07 │ │ vmovaps %zmm2,0xd04100(,%rax,4)
0.07 │ │ vaddps %zmm1,%zmm5,%zmm1{%k3}
0.44 │ │ vmovaps %zmm1,0xd04140(,%rax,4)
│ │ add $0x40,%rax
│ ├──cmp $0x4000,%rax
│ └──jne 10
0.70 │ ↑ jne 40
How it is possible that there is no increase in mispredicts?
Curiously enabling scatter disables vectorization again:
jh@shroud:~> ~/trunk-install-znver5/bin/g++ -DT=float -march=native cnd.c
-Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details
-mtune-ctrl=use_scatter_4parts -mtune-ctrl=use_scatter_8parts
-mtune-ctrl=use_scatter_2parts ; objdump -d a.out | grep gather ; perf stat -r
10 ./a.out
Performance counter stats for './a.out' (10 runs):
475.05 msec task-clock:u # 1.000 CPUs
utilized ( +- 0.05% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
660 page-faults:u # 1.389 K/sec
( +- 0.05% )
1,968,735,770 cycles:u # 4.144 GHz
( +- 0.01% )
121,150,048 stalled-cycles-frontend:u # 6.15% frontend
cycles idle ( +- 0.36% )
1,594,448,124 instructions:u # 0.81 insn per
cycle
# 0.08 stalled cycles per
insn ( +- 0.00% )
360,669,133 branches:u # 759.224 M/sec
( +- 0.00% )
89,997 branch-misses:u # 0.02% of all
branches ( +- 0.54% )
0.475287 +- 0.000236 seconds time elapsed ( +- 0.05% )
We seem to give up on:
cnd.c:11:21: note: vect_is_simple_use: vectype vector(16)
<signed-boolean:1>
cnd.c:11:21: note: vect_is_simple_use: operand indices[i_14], type of def:
internal
cnd.c:11:21: note: vect_is_simple_use: vectype vector(16) int
cnd.c:11:21: missed: unsupported masked emulated gather.
cnd.c:9:1: missed: not vectorized: relevant stmt not supported: patt_31 =
.MASK_LOAD (_30, 32B, patt_12);
cnd.c:11:21: missed: bad operation or unsupported loop bound.
