Non-native can be a bit harder for constant folding (you have to emulate the target’s behavior for floating point, for example), but I think that mostly is a thing of the past because most architectures use the same types.
What can make a difference is the architecture. Examples:
- Register assignment is easier on orthogonal architectures.
- A compiler doesn’t need to spend time looking for auto-vectorization opportunities if the target architecture doesn’t have vector instructions.
Probably more importantly, there can be a difference in how much effort the compiler makes for finding good code. Typically, newer compilers start out with worse code generation that is faster to generate (make it work first, then make it produce good code)
I wouldn’t know whether any of these are an issue in this case.
What can make a difference is the architecture. Examples:
- Register assignment is easier on orthogonal architectures.
- A compiler doesn’t need to spend time looking for auto-vectorization opportunities if the target architecture doesn’t have vector instructions.
Probably more importantly, there can be a difference in how much effort the compiler makes for finding good code. Typically, newer compilers start out with worse code generation that is faster to generate (make it work first, then make it produce good code)
I wouldn’t know whether any of these are an issue in this case.