PARSEC x264 基准测试中的奇怪行为

Question

I am running the PARSEC x264 benchmark and it runs fine until the close down when it crashes. Unfortunately the suite seems to be unmaintained or on low maintenance - this problem (the crash, not the specifics) was posted on the mailing list nearly a year ago but no replies.

I have isolated the problem to part of a particular function (see below) using good old fashioned debugging by printf:

void x264_cqm_delete( x264_t *h )
{
     int i, j;
     printf("IN BAD CODE\n");
     for( i = 0; i < 6; i++ )
     {
         for( j = 0; j < i; j++ )
             if( h->quant4_mf[i] == h->quant4_mf[j] )
                 break;
         if( j == i )
         {
            printf("DELETE CODE with i = %i and j = %i\n", i, j);
            x264_free( h->  quant4_mf[i] );
            x264_free( h->dequant4_mf[i] );
            x264_free( h->unquant4_mf[i] );
        }
        for( j = 0; j < i; j++ )
            if( h->quant4_bias[i] == h->quant4_bias[j] )
                break;
        if( j == i )
            x264_free( h->quant4_bias[i] );
}

}

This generates the output...

IN BAD CODE
DELETE CODE with i = 0 and j = 0
DELETE CODE with i = 4 and j = 4
DELETE CODE with i = 5 and j = 4

What is going on?

Answer 1

From examining the canonical code in the x264 library itself it seems clear that the code is broken and the reason I am seeing these issues is that the code is trashing the heap.

In fact the outer loop's limit should be 4 - it should match the number used in the name of the various arrays - that is the pattern used in the x264 library itself - and the quant4_mf array has only 4 elements in the first dimension. Patching the code in this way sees the compiled binary run without issue.

Answer 2

Just to make it work, you can change compiler optimization from O3 to O2: https://github.com/cirosantilli/parsec-benchmark/issues/3 Tested with Ubuntu-18.04 gcc 7.5.0