[llvm-dev] Incompatible x86 calling convention for __m64 parameters and returns.

[Brad Conroy via llvm-dev]

On x86 (32 bit) GCC passes mmx parameters in mm0, mm1 and mm2 and returns in mm0.
According to comments in X86CallingConv.td, llvm supposedly does the same:

>  // MMX vector types are always returned in MM0. If the target doesn't have
>  // MM0, it doesn't support these vector types.
>  CCIfType<[x86mmx], CCAssignToReg<[MM0]>>,

and

>  // The first 3 __m64 vector arguments are passed in mmx registers if the
>  // call is not a vararg call.
>  CCIfNotVarArg<CCIfType<[x86mmx],
>                CCAssignToReg<[MM0, MM1, MM2]>>>,

However in practice, llvm is actually passing parameters on the stack and returning them
in the eax:edx register pair as if they were long long or unsigned long long.

This goes back to at least clang-3.0, but I couldn't find a bug report on it... maybe because
mmx intrinsics are rarely used, so it may have no real world implications.

R,
Brad Conroy