Add CUDA support for MSVC

[trolleyman]

Tested building https://github.com/trolleyman/cuda-macros with this (the cuda-macros-test crate) and it builds & links correctly. Haven't had a chance to test that this runs yet, but will in the morning.

I wasn't sure that this way is the most elegant, but this seemed like the way that did the least amount of changes. I am also not sure that this is the correct way of doing this, especially regarding cross-compiling, but it gets it up and running at least.

To test you can do a cargo build in the root of the repo linked above. The build stuff is a bit hacky, but essentially it generates the CUDA function below & calls it.

extern "C" __global__ void hello(int32_t* x, int32_t y) {
    printf("Hello from block %d, thread %d (y=%d)\n", blockIdx.x, threadIdx.x, y);
    *x = 2;
}

extern "C" unsafe fn hello(x: *mut i32, y: i32);

[alexcrichton]

Thanks! It looks like the main difference here is /X vs -X but I think that MSVC considers them equivalent for C compilation, could the MSVC bits just be updated to use the dash instead?

[trolleyman]

Oh cool, didn't know that. Does that mean we can remove all of those *_flag functions from the compiler family struct & just use raw string literals everywhere?

[alexcrichton]

Perhaps yeah! I forget which target those were added for, so I'd just follow the various logic internally to see if it's necessary to keep it or not.

[trolleyman]

Alright, there we go. I also tested cuda-macros & it compiles & runs a simple test program. Should be ready to merge

[trolleyman]

Ah. That commit broke it

[trolleyman]

There we go

[alexcrichton]

Are changes like this necessary? This PR feels a bit overreaching in terms of changing everything it finds?

[trolleyman]

I wasn't too sure about this define, so I added it when compiling with nvcc as well. We can just simplify it to a -D option instead and it will define it for both nvcc & non-nvcc versions?

[alexcrichton]

I'm not really sure what's going on here, how come the flag is pushed twice?

[trolleyman]

I wasn't too sure about if the nvcc compiler passed through the -m32 and -m64 options, so I added it for both. I think it does though, so I'll modify these lines to just add the options onto the command args directly, as both msvc and nvcc support these options.

Also, for context, the nvcc compiler only supports 32/64 bit cross-compilation, and it's phasing out 32 bit compilation altogether as well it seems. I'm not sure what to do about cross-compilation & CUDA btw -- should it be better supported/disallowed entirely?

[alexcrichton]

Hm I was sort of hoping that everything that was /foo could be changed to -foo, but this seems like it's sort of just a smattering of some one way and some another way...

[trolleyman]

Ah, ok - should I try and remove all of the /opt options?

[alexcrichton]

I think that should work yeah, and I think it's probably better than sprinkling if cuda everywhere

[trolleyman]

Ok, removed all /- options everywhere, and simplified command_add_output_file. I've tested that -Foout.obj is the same as -Fo out.obj when compiling assembly.

Removed the edits that are mentioned in #426 (comment) as cross-compilation is not supported with nvcc, so it's probably better not to even pretend that we support it.

Simplified the change in #426 (comment)

[alexcrichton]

Thanks @trolleyman!