CODEWITHSUNDEEP
c++gccassemblyavx2
user3401776
user3401776

Reputation: 23

Assemble Error for AVX2

I've tried to compile a AVX2 program with gcc(g++). But it didn't work right.

#include<immintrin.h>
....
__m256i _vector256  = _mm256_loadu_si256((__m256i*)pin);
__m256i _vectorMask = _mm256_loadu_si256((__m256i*)mask_hbits); 
_vector256 = _mm256_slli_epi32 (_vector256, AVX_LOGDESC); // AVX_LOGDESC == 4
__m256i _vectorRes = _mm256_and_si256(_vector256, _vectorMask);
....

My compile command is (as one line in Makefile):

g++ avx_shift.cpp -c -mavx2

I also tried this one:

g++ avx_shift.cpp -c -march=native

But the error messages were same:

/tmp/ccSFs6U0.s: Assembler messages:
/tmp/ccSFs6U0.s:1083: Error: suffix or operands invalid for `vpslld'
/tmp/ccSFs6U0.s:1091: Error: suffix or operands invalid for `vpand'
....

The assembly code (added the line number):

vmovdqa 988(%rsp), %ymm0    # __A, D.48219
movl    984(%rsp), %eax # __B, tmp205
movl    %eax, -120(%rsp)    # tmp205, %sfp
vmovd   -120(%rsp), %xmm6   # %sfp, tmp205
vpslld  %xmm6, %ymm0, %ymm0 # tmp205, D.48219, D.48220      // 1083
....
vmovdqa 668(%rsp), %ymm0    # __B, tmp220
vmovdqa 700(%rsp), %ymm1    # __A, tmp221
vpand   %ymm0, %ymm1, %ymm0 # tmp220, tmp221, D.48215       // 1091
....

My gcc(g++) version is 4.8.0. My OS version is Centos6.5 X86_64. I've confirmed that my CPU supports AVX2 instructions. So who can help me find the errors?

Upvotes: 1

Views: 1713

Answers (1)

Jester
Jester

Reputation: 58822

The two source lines you pasted shouldn't, and on my machine don't, generate any vpslld or vpand instructions. Use the -S -g -fverbose-asm switches to ask for assembly source and try to find the matching source lines.

Upvotes: 3

Related Questions