Fixes Heap Allocations for Large alignMask #238
Conversation
Uses the aligned_alloc C function from the stdlib instead of malloc so that incase the alignMask is larger than what the system supports we do not get unnecessary crashes.
|
The alignMask used by swift::slowAlloc() and the alignment used by posix_memalign() are not compatible. You need to convert them. (I think alignment = alignMask+1 works.) Calling posix_memalign() for every allocation may be slow. You should measure performance. You may need to call malloc directly when alignMask is within the system's guaranteed alignment. |
|
I could not find anything that showed that swift::slowAlloc() requires an alignment mask + 1 maybe I am missing something, but that should be an easy enough fix. However, we should make sure this is true because posix_memalign requires the alignMask to be a power of 2, or else will fail. You are right, posix_memalign incurs a significant overhead in order to align memory, so we can use it only if the alignMask exceeds the system's guaranteed alignment. However, after a fair amount of research I've found that on Darwin, malloc promises a 16 byte alignment whereas on other posix systems the glibc only promises 8 byte alignment. There probably is a clean way to work around the platform dependent system size, any suggestions? |
|
The compiler could emit more specific function calls when it knows the alignment needed for an allocation, and only fall back to |
Fixes a segmentation fault where the IRGen module would crash when a generic (empty) struct was part of the source code because it tried to call std::next on an end iterator.
Changes needed to build/test on Linux.
Update Serpent hash to 4.2
Uses the posix_memalign C function from the stdlib instead of malloc so that in case the alignMask is larger than what the system (or malloc) supports (on OS X malloc aligns memory on 16 byte boundaries so for larger values) we do not get unnecessary crashes.