IssueTracker

Subtracting an unsigned 32-bit integer from a 64-bit pointer in Reactor requires to first zero-extend it. Currently LLVM implicitly sign-extends the index in a GEP operation.

This could be implemented elegantly by defining an IntPtr (and UIntPtr) type, which corresponds to Int or Long (UInt or ULong) on 32-bit or 64-bit platforms respectively. This matches (u)intptr_t in C++.

https://chromium-review.googlesource.com/433365 implements an optimization related to this, but it's actually slightly incorrect. We should only optimize pointer arithmetic where the index first gets zero-extended, then shifted/multiplied. Currently we do it the other way around, which in theory could lose the high index bits before the extension. Since this overflow is extremely unlikely to be intentional, it should be safe for now.

Likewise constant unsigned 32-bit array indexes and pointer offsets are not handled entirely correctly at the moment.