Make malloc(0) and posix_memalign(0) return a non-null pointer

[RalfJung]

Currently we just return a null pointer, but we're also allowed to return a unique non-zero address. That would have the benefit that we can detect leaks (not calling free) or double-free.

[tiif]

This is interesting. I've had this question on my mind for quite a while, how will malloc(0) cause a memory leak if the memory allocated is 0 byte?

Besides, the return value of malloc(0) seems to be implementation-defined¹, and may return null or a valid pointer. If miri only return a non-zero address, and the user assumed malloc(0) return null (hence they don't free it), would it be considered a false positive?

Footnotes

1. https://stackoverflow.com/questions/5372591/why-is-the-return-value-of-malloc0-implementation-defined ↩

[RalfJung]

This is interesting. I've had this question on my mind for quite a while, how will malloc(0) cause a memory leak if the memory allocated is 0 byte?

There can be metadata associated with this allocation, managed by the allocator, that would be leaked.

Besides, the return value of malloc(0) seems to be implementation-defined1, and may return null or a valid pointer. If miri only return a non-zero address, and the user assumed malloc(0) return null (hence they don't free it), would it be considered a false positive?

Implementation-defined means that the implementation gets to decide. If the user assumes that malloc(0) returns null, then they are making a false assumption, since it is not necessarily true that the implementation will return null.

[tiif]

I wonder if we should do the same for posix_memalign? It is still returning a null pointer currently.

If size is 0, then the value placed in *memptr is either NULL or a unique pointer value.
https://man7.org/linux/man-pages/man3/posix_memalign.3.html

[RalfJung]

Yeah, posix_memalign shod also return non-null on size 0. (But that can be a separate PR.) One of the key parts of the PR will be two tests to ensure we detect double free and memory leaks of malloc(0).

[RalfJung]

#3580 implements this for malloc; posix_memalign could get the same treatment as discussed above.