Shouldn't this specify the deallocator too while we're at it changing global headers? It could perhaps find mismatches, or perhaps optimize even more combined malloc / free?

yes it could (even if the gcc version supporting it is newer)

could also be done for msvcrt fopen/fclose

from the initial patch, I wonder if adding the attribute to inline functions in heap.h is needed (as called function are already marked)

I've been using this patch (generates one potential mismatch to be fixed)

A+