https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85833
Bug ID: 85833
Summary: [AVX512] use mask registers instructions instead of
scalar code
Product: gcc
Version: 7.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: wojciech_mula at poczta dot onet.pl
Target Milestone: ---
There is a simple function, which checks if there is any non-zero element
in a vector:
---ktest.c---
#include <immintrin.h>
int anynonzero_epi32(__m512i x) {
const __m512i zero = _mm512_setzero_si512();
const __mmask16 mask = _mm512_cmpneq_epi32_mask(x, zero);
return mask != 0;
}
---eof---
$ gcc --version
gcc (Debian 7.3.0-16) 7.3.0
$ gcc -O2 -S -mavx512f ktest.c && cat ktest.s
anynonzero_epi32:
vpxord %zmm1, %zmm1, %zmm1
vpcmpd $4, %zmm1, %zmm0, %k1
kmovw %k1, %eax # <<< HERE
testw %ax, %ax #
setne %al
movzbl %al, %eax
vzeroupper
ret
The problem is that GCC copies content of the mask register k1 into
GPR (using KMOV instruction), and then perform test. AVX512F has got
instruction KTEST kx, ky which sets ZF and CF:
ZF = (kx AND ky) == 0
CF = (kx AND NOT ky) == 0
In this case we might use KTEST k1, k1 to set ZF when k1 == 0.
The procedure might be then compiled as:
anynonzero_epi32:
vpxord %zmm1, %zmm1, %zmm1
vpcmpd $4, %zmm1, %zmm0, %k1
xor %eax, %eax #
ktestw %k1, %k1 #
setne %al #
vzeroupper
ret
