[CUDA] Replace 'nv_weak' attributes in CUDA headers with 'weak'.

[tra]

Summary

An alternative to implementing nv_weak attribute (D47201).

The patch should make runtime sub functions to have .weak attribute in PTX and that should avoid GPU-side linking errors.

[jlebar]

What could possibly go wrong.

[tra]

I'll wait to see if that fixes @Hahnfeld's problem.

AFAICT, nv_weak is not used for anything interesting other than the device-side runtime stubs that return errors (apparently, until they are linked with the proper device runtime which would have strong version of the symbols.

I've tested with 8.0, 9.0 and 9.2 with -fcuda-rdc and all appear to produce .weak variants of those stubs in PTX.
One thing that I've noticed that we do differently is that those stubs being present in clang-generated PTX regardless of whether -fcuda-rdc is in effect. Well, there's only so much we can do with preprocessor magic. Other than wasted space, it's benign AFAICT. If it turns out that it's problematic after all, we should be able to disable them away with attribute((if_enabled...)).

[Hahnfeld]

Works for me, I'll abandon D47201 if you prefer this.

[aaron.ballman]

LGTM as well, for what little it's worth. :-)

lib/Headers/__clang_cuda_runtime_wrapper.h

 #define __CUDA_LIBDEVICE__
 #endif
 // Disables definitions of device-side runtime support stubs in
 // cuda_device_runtime_api.h
 #include "driver_types.h"
 #include "host_config.h"
 #include "host_defines.h"
 
+// Temporarily replace "nv_weak" with weak, so __attribute__((nv_weak)) in
+// cuda_device_runtime_api.h ends up being __attribute__((weak)) which is the
+// functional equivalent of what we need.
+#pragma push_macro("nv_weak")
+#define nv_weak weak
 #undef __CUDABE__
 #undef __CUDA_LIBDEVICE__
 #define __CUDACC__
 #include "cuda_runtime.h"
 
+#pragma pop_macro("nv_weak")
 #undef __CUDACC__
 #define __CUDABE__
 
 // CUDA headers use __nvvm_memcpy and __nvvm_memset which Clang does
 // not have at the moment. Emulate them with a builtin memcpy/memset.
 #define __nvvm_memcpy(s, d, n, a) __builtin_memcpy(s, d, n)
 #define __nvvm_memset(d, c, n, a) __builtin_memset(d, c, n)