Fix the seg fault issue (108) due to mavx CFLAG

[mamthan]

We are able to reproduce the issue #108 with the following build command:

cmake -DCMAKE_C_FLAGS="-mavx" ..
make

Encode command:

./xeve_app -i <input.y4m> -o <output.evc>

Error message:

XEVE: eXtra-fast Essential Video Encoder
Segmentation fault

On passing "-mavx2" as C_Flag (cmake -DCMAKE_C_FLAGS="-mavx" ..), it sets the flag for all files in the project, including the source C files. So, the compiler is creating optimizations which are not handled properly resulting in a segmentation fault.

This PR fixes the issue by updating the CmakeLists.txt files.

[kpchoi]

Did you check why the 'mavx2' option generate segfault exactly?
Theoretically, it should not have problem even though we set 'mavx2' option to c-code.
There could be some buggy code for AVX2 optimization generated by c-compiler?
If we could find the weak code to avx2 optimization, it would be better to change the code itself rather than adding 'mnoavx2' cflag option.

[dariusz-f]

Output form gdb:

XEVE: eXtra-fast Essential Video Encoder

Program received signal SIGSEGV, Segmentation fault.
0x00005555555a3732 in xeve_pintra_create ()
(gdb) bt
#0  0x00005555555a3732 in xeve_pintra_create ()
#1  0x0000555555570e8a in xeve_platform_init ()
#2  0x00005555555ea4f5 in xevem_platform_init ()
#3  0x0000555555566658 in xeve_create ()
#4  0x000055555555cc83 in main ()

In debug build and under valgrind it works correctly in my case.

[kpchoi]

@neergil ,
Could check this PR?
It is not recommened to change cflag if it is not issue relating to compile and build issue.

[kpchoi]

@neergil

I don't understand this code fully, because the code change is weird.
It is required to cross-check the code changes.

[kpchoi]

We looked into this further and found the root cause: XEVE_CTX requires 32-byte alignment (XEVE_PINTRA/XEVE_PINTER contain ALIGNED_32 members) but it is allocated with plain malloc(), which only guarantees 16 bytes. With -mavx the compiler trusts the declared alignment and emits aligned vmovdqa stores for adjacent member assignments, hence the crash in xeve_pintra_create().

Reordering the struct members only prevents the vectorization at this particular spot, so the problem could reappear with a different compiler or flags. #162 fixes the allocation itself instead. Could you check whether it resolves the issue on your side as well?