[CUDA] Add conversion operators for threadIdx, blockIdx, gridDim, and blockDim to uint3 and dim3.

[jlebar]

Summary

This lets you write, e.g.

uint3 a = threadIdx;
uint3 b = blockIdx;
dim3 c = gridDim;
dim3 d = blockDim;

which is legal in nvcc, but was not legal in clang.

The fact that e.g. the type of threadIdx is not actually uint3 is still
observable, but now you have to try to observe it.

[tra]

lib/Headers/cuda_builtin_vars.h

72

Considering that built-in variables are never instantiated, I wonder how it's going to work as the operator will presumably need 'this' pointing *somewhere*, even if we don't use it. Unused 'this' would probably get optimized away with optimizations on, but -O0 may cause problems.

[jlebar]

lib/Headers/cuda_builtin_vars.h

72

This is interesting. In the ptx, threadIdx actually gets instantiated, as a non-weak global:

.global .align 1 .b8 threadIdx[1];

Then we take the address of this thing.

At -O2, we don't emit a threadIdx global at all.

I think this is basically fine. It's actually not right to change extern to static in the decl, because then we try to construct a __cuda_builtin_threadIdx_t, and the default constructor is deleted. :)

[tra]

OK.

lib/Headers/__clang_cuda_runtime_wrapper.h

 // We also need device-side std::malloc and std::free.
 namespace std {
 __device__ static inline void free(void *__ptr) { ::free(__ptr); }
 __device__ static inline void *malloc(size_t __size) {
   return ::malloc(__size);
 }
 } // namespace std
 
+// Out-of-line implementations from cuda_builtin_vars.h.  These need to come
+// after we've pulled in the definition of uint3 and dim3.
+
+__device__ inline __cuda_builtin_threadIdx_t::operator uint3() const {
+  uint3 ret;
+  ret.x = x;
+  ret.y = y;
+  ret.z = z;
+  return ret;
+}
+
+__device__ inline __cuda_builtin_blockIdx_t::operator uint3() const {
+  uint3 ret;
+  ret.x = x;
+  ret.y = y;
+  ret.z = z;
+  return ret;
+}
+
+__device__ inline __cuda_builtin_blockDim_t::operator dim3() const {
+  return dim3(x, y, z);
+}
+
+__device__ inline __cuda_builtin_gridDim_t::operator dim3() const {
+  return dim3(x, y, z);
+}
+
 #include <__clang_cuda_cmath.h>
 
 #endif // __CUDA__
 #endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__

lib/Headers/cuda_builtin_vars.h

  * THE SOFTWARE.
  *
  *===-----------------------------------------------------------------------===
  */
 
 #ifndef __CUDA_BUILTIN_VARS_H
 #define __CUDA_BUILTIN_VARS_H
 
+// Forward declares from vector_types.h.
+struct uint3;
+struct dim3;
+
 // The file implements built-in CUDA variables using __declspec(property).
 // https://msdn.microsoft.com/en-us/library/yhfk0thd.aspx
 // All read accesses of built-in variable fields get converted into calls to a
-// getter function which in turn would call appropriate builtin to fetch the
+// getter function which in turn calls the appropriate builtin to fetch the
 // value.
 //
 // Example:
 //    int x = threadIdx.x;
 // IR output:
 //  %0 = call i32 @llvm.ptx.read.tid.x() #3
 // PTX output:
 //  mov.u32     %r2, %tid.x;
#define __CUDA_DISALLOW_BUILTINVAR_ACCESS(TypeName)                            \
   __attribute__((device)) TypeName(const TypeName &) __DELETE;                 \
   __attribute__((device)) void operator=(const TypeName &) const __DELETE;     \
   __attribute__((device)) TypeName *operator&() const __DELETE
 
 struct __cuda_builtin_threadIdx_t {
   __CUDA_DEVICE_BUILTIN(x,__builtin_ptx_read_tid_x());
   __CUDA_DEVICE_BUILTIN(y,__builtin_ptx_read_tid_y());
   __CUDA_DEVICE_BUILTIN(z,__builtin_ptx_read_tid_z());
+  // threadIdx should be convertible to uint3 (in fact in nvcc, it *is* a
+  // uint3).  This function is defined after we pull in vector_types.h.
+  __attribute__((device)) operator uint3() const;
 private:
   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_threadIdx_t);
 };
 
 struct __cuda_builtin_blockIdx_t {
   __CUDA_DEVICE_BUILTIN(x,__builtin_ptx_read_ctaid_x());
   __CUDA_DEVICE_BUILTIN(y,__builtin_ptx_read_ctaid_y());
   __CUDA_DEVICE_BUILTIN(z,__builtin_ptx_read_ctaid_z());
+  // blockIdx should be convertible to uint3 (in fact in nvcc, it *is* a
+  // uint3).  This function is defined after we pull in vector_types.h.
+  __attribute__((device)) operator uint3() const;
 private:
   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_blockIdx_t);
 };
 
 struct __cuda_builtin_blockDim_t {
   __CUDA_DEVICE_BUILTIN(x,__builtin_ptx_read_ntid_x());
   __CUDA_DEVICE_BUILTIN(y,__builtin_ptx_read_ntid_y());
   __CUDA_DEVICE_BUILTIN(z,__builtin_ptx_read_ntid_z());
+  // blockDim should be convertible to dim3 (in fact in nvcc, it *is* a
+  // dim3).  This function is defined after we pull in vector_types.h.
+  __attribute__((device)) operator dim3() const;
 private:
   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_blockDim_t);
 };
 
 struct __cuda_builtin_gridDim_t {
   __CUDA_DEVICE_BUILTIN(x,__builtin_ptx_read_nctaid_x());
   __CUDA_DEVICE_BUILTIN(y,__builtin_ptx_read_nctaid_y());
   __CUDA_DEVICE_BUILTIN(z,__builtin_ptx_read_nctaid_z());
+  // gridDim should be convertible to dim3 (in fact in nvcc, it *is* a
+  // dim3).  This function is defined after we pull in vector_types.h.
+  __attribute__((device)) operator dim3() const;
 private:
   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_gridDim_t);
 };
 
 #define __CUDA_BUILTIN_VAR                                                     \
   extern const __attribute__((device)) __attribute__((weak))
 __CUDA_BUILTIN_VAR __cuda_builtin_threadIdx_t threadIdx;
 __CUDA_BUILTIN_VAR __cuda_builtin_blockIdx_t blockIdx;