[Polly][GPGPU] Fixed undefined reference for CUDA's managed memory in Runtime library.

[PhilippSchaad]

Summary

References to polly_freeManaged and polly_mallocManaged resulted in undefined references to the GPUJIT Runtime library. The API was broken and has now been generalized again.

[bollu]

LGTM, thanks for generalising it :) I'm sorry, I'll keep the mixed CUDA/OpenCL thing in mind next time I mess with GPUJIt.

[PhilippSchaad]

Not a problem, thanks for reacting so quickly! :-)

[PhilippSchaad]

Was committed with https://llvm.org/svn/llvm-project/polly/trunk@311848 91177308-0d34-0410-b5e6-96231b3b80d8

tools/GPURuntime/GPUJIT.c

 int isManagedPtr(void *mem) {
   for (unsigned long long i = 0; i < g_nmanagedptrs; i++) {
     if (g_managedptrs[i] == mem)
       return 1;
   }
   return 0;
 }
 
-void polly_freeManaged(void *mem) {
+void freeManagedCUDA(void *mem) {
   dump_function();
 
   // In a real-world program this was used (COSMO), there were more `free`
   // calls in the original source than `malloc` calls. Hence, replacing all
   // `free`s with `cudaFree` does not work, since we would try to free
   // 'illegal' memory.
   // As a quick fix, we keep a free list and check if `mem` is a managed memory
   // pointer. If it is, we call `cudaFree`.
   // If not, we pass it along to the underlying allocator.
   // This is a hack, and can be removed if the underlying issue is fixed.
   if (isManagedPtr(mem)) {
     if (CuMemFreeFcnPtr((size_t)mem) != CUDA_SUCCESS) {
       fprintf(stderr, "cudaFree failed.\n");
       exit(-1);
     }
     return;
   } else {
     free(mem);
   }
 }
 
-void *polly_mallocManaged(size_t size) {
+void *mallocManagedCUDA(size_t size) {
   // Note: [Size 0 allocations]
   // Sometimes, some runtime computation of size could create a size of 0
   // for an allocation. In these cases, we do not wish to fail.
   // The CUDA API fails on size 0 allocations.
   // So, we allocate size a minimum of size 1.
   if (!size && DebugMode)
     fprintf(stderr, "cudaMallocManaged called with size 0. "
                     "Promoting to size 1");
case RUNTIME_CL:
     freeContextCL(Context);
     break;
 #endif /* HAS_LIBOPENCL */
   default:
     err_runtime();
   }
 }
 
+void polly_freeManaged(void *mem) {
+  dump_function();
+
+#ifdef HAS_LIBCUDART
+  freeManagedCUDA(mem);
+#else
+  fprintf(stderr, "No CUDA Runtime. Managed memory only supported by CUDA\n");
+  exit(-1);
+#endif
+}
+
+void *polly_mallocManaged(size_t size) {
+  dump_function();
+
+#ifdef HAS_LIBCUDART
+  return mallocManagedCUDA(size);
+#else
+  fprintf(stderr, "No CUDA Runtime. Managed memory only supported by CUDA\n");
+  exit(-1);
+#endif
+}
+
 /* Initialize GPUJIT with CUDA as runtime library. */
 PollyGPUContext *polly_initContextCUDA() {
 #ifdef HAS_LIBCUDART
   Runtime = RUNTIME_CUDA;
   return polly_initContext();
 #else
   fprintf(stderr, "GPU Runtime was built without CUDA support.\n");
   exit(-1);