IssueTracker

LLVM's bitcast instruction is specified to only support casting pointers to/from other pointers. One must use ptrtoint and inttoptr to cast them to/from integers. In practice we've found bitcast to support these casts as well, but this might be highly target-specific and should be avoided. That said, it's worth noting that C++'s reinterpret_cast<> has this flexibility as well, and Reactor's As<> currently has no explicit restriction other than the sizes must match.

That said, we had to add support for casting 'emulated' vectors and casting scalars to vectors.

For SPIR-V's OpBitcast instruction we need support for casting between 64-bit pointers and int2 vectors. This happens to already work for the LLVM backend, but Subzero only supports casting between vectors and 32-bit values.

We should check support for all combinations for both backends, add unit tests, comply with LLVM's restrictions, and where necessary optimizing things. For example Subzero could use movq for 64-bit values, similar to its use of movd for 32-bit values, instead of storing to memory and loading from it again.