https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70059
Bug ID: 70059
Summary: Invalid codegen on AVX-512 when using
_mm512_inserti64x4(x, y, 0)
Product: gcc
Version: 5.3.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: povilas at radix dot lt
Target Milestone: ---
Use of _mm512_inserti64x4 in certain scenario outlined below fails on GCC
5.3.1.
$ g++-5 --version
g++-5 (Ubuntu 5.3.1-9ubuntu3) 5.3.1 20160222
$ cat main.cc
#include "immintrin.h"
__m512i run(__m256i a, __m256i b)
{
__m512i r = _mm512_undefined_si512();
r = _mm512_inserti64x4(r, a, 0);
r = _mm512_inserti64x4(r, b, 1);
return r;
}
$ g++-5 -O1 -c main.cc --save-temps -o main.o -mavx512f
The following assembly is generated:
_Z3runDv4_xS_:
vinserti64x4 $0x1, %ymm0, %zmm1, %zmm0
ret
As you can see, the first argument is inserted into the upper half of zmm0
register and the second into the lower. The intention is the other way round.
The problem can be worked around:
$ cat main.cc
#include "immintrin.h"
__m512i run(__m256i a, __m256i b)
{
__m512i r;
r = _mm512_castsi256_si512(a);
r = _mm512_inserti64x4(r, b, 1);
return r;
}
$ g++-5 -O1 -c main.cc --save-temps -o main.o -mavx512f
The resulting assembly is correct:
_Z3runDv4_xS_:
vinserti64x4 $0x1, %ymm1, %zmm0, %zmm0
ret
Regards,
Povilas
