'Does not work' isn't very specific. 

As far as I know, there's no fundamental reason why this shouldn't work fine on
MinGW - so if it doesn't, that's likely to be a bug.

So, what is the actual bug you indirectly refer to here?

Sorry, for the general title. I did not known much more that before. 

Now it seems I have found the problem:
gst_mem_chunk_alloc ends in infinite cycle, since
gst_trash_stack_pop always returns NULL.

(The function gst_mem_chunk_alloc is trying to allocate new chunks and push them
to TrashStack until gst_trash_stack_pop returns non-NULL)

When I try to add debug output to gst_mem_chunk_alloc:
    210 gpointer
    211 gst_mem_chunk_alloc (GstMemChunk * mem_chunk)
    212 {
    213   GstMemChunkElement *chunk;
    214 
    215   g_return_val_if_fail (mem_chunk != NULL, NULL);
    216 
    217 again:
    218   printf("HERE\n");
    219   chunk = gst_trash_stack_pop (&mem_chunk->stack);
it prints infinite number of HERE messages. 

However, if I try the following debug prints
    210 gpointer
    211 gst_mem_chunk_alloc (GstMemChunk * mem_chunk)
    212 {
    213   GstMemChunkElement *chunk;
    214 
    215   g_return_val_if_fail (mem_chunk != NULL, NULL);
    216 
    217 again:
    218   printf("HERE %d\n", mem_chunk->stack.count);
    219   chunk = gst_trash_stack_pop (&mem_chunk->stack);
    220   printf(" ----- %p\n", chunk);
it works! (Note that it does not work without the second printf).

So it seems that it is some optimization problem in GCC (probably related to the
use of inline assmebler). 

This code no longer exists in the GStreamer repository.