Change free_sh_audio() to take demuxer and stream id as parameters (same as new_sh_audio()) instead of sh_audio_t *, use those to remove the pointer from demuxer->a_streams[] before freeing it. Some demuxers use free_sh_audio() to undo the creation of an already-allocated audio stream in case of error. These uses were unsafe since free_sh_audio() freed the data structure but left the pointer in demuxer->a_streams[], leading to double free later in free_demuxer() (and perhaps use of the freed stream before that, I didn't check).