Zero-size SYCLMalloc returns non-unique addresses

[codeplaysoftware]

Zero-size SYCLMalloc returns non-unique addresses

[Ruyk]

What would be the expected behaviour? According to the POSIX and C standards, malloc of size 0 behavior:

 If the size of the space requested is 0, the behavior is implementation-defined: either a null pointer shall be returned, or the behavior shall be as if the size were some non-zero value, except that the behavior is undefined if the returned pointer is used to access an object

Which is technically what is happening. I am happy to return a nullptr if that makes detecting errors easier...

[DuncanMcBain]

I believe either one of TensorFlow or Eigen relies on size-zero allocations being allowed. In that case, a nullptr would be a problem. However, I think the current behaviour isn't consistent with the second bullet point - shouldn't malloc return unique pointers? Moreover, this means that the PointerMapper will misbehave (what happens when you try to free once, or three times?)

[r-potter]

I agree with @DuncanMcBain . Either null or the equivalent of a non-zero allocation would be fine (although I believe there is some Eigen or Tensorflow requirement that prefers non-null). What's problematic is returning the same non-null address. There is actually a much more problematic case:

auto a = SYCLMalloc(0, pMap); // Returns 0x1000
auto b = SYCLMalloc(4096, pMap); // Returns 0x1000 + 0
if (a) {
    SYCLFree(a, pMap); // b gets accidentally freed because it shares an address.
}

This is why POSIX says the equivalent of a non-zero allocation. It ensures allocations get unique addresses, and that free can determine what to actually release without aliasing other allocations.

[Ruyk]

That makes sense to me. I didn't realize at first about the potential double free issue.

Thanks for the pull request from @Rbiessy, this is sorted out now. If you are happy with the solution I think we can close it.

[r-potter]

LGTM

[DuncanMcBain]

Thanks, all!