-
Martin Storsjö authored
REBASE_PICTURE (more specifically, this half of it) takes a Picture pointer that points into one larger struct, finds the offset of that Picture within the struct and finds the corresponding field within another instance of a similar struct. The pointer difference "pic - (Picture*)old_ctx" is a value given in sizeof(Picture) units, and when applied back on (Picture*)new_ctx gets multiplied back with sizeof(Picture). Many compilers seem to optimize out this division/multiplication, but not all do. GCC 4.2 on OS X doesn't seem to remove the division/multiplication, therefore the new pointer didn't turn out to point to exactly the right place in the new struct since it only had sizeof(Picture) granularity (and the Picture is not aligned on a sizeof(Picture) boundary within the encompassing struct). This bug has been present before 47318953 as well - with H264, pointers to h->ref_list[0][0] pointed to 88 bytes before h->ref_list[0][0] after the rebase. After shrinking Picture, the difference ended up even larger, making writes via such a Picture pointer overwrite other fields at random in H264Context, ending up in crashes later. This fixes H264 multithreaded decoding on OS X with GCC 4.2. Signed-off-by:
Martin Storsjö <martin@martin.st>
a65f965c