[llvm] [offload] Fix CUDA args size by subtracting tail padding (PR #172249)

[Kevin Sala Penades via llvm-commits]

================
@@ -173,11 +174,29 @@ struct CUDAKernelTy : public GenericKernelTy {
   }
 
 private:
+  /// Initialize the size of the arguments.
+  Error initArgsSize() {
+    CUresult Res;
+    size_t ArgOffset, ArgSize;
+    size_t Arg = 0;
+
+    // Find the last argument to know the total size of the arguments.
+    while ((Res = cuFuncGetParamInfo(Func, Arg++, &ArgOffset, &ArgSize)) == CUDA_SUCCESS)
+      ArgsSize = ArgOffset + ArgSize;
+
+    if (Res != CUDA_ERROR_INVALID_VALUE)
----------------
kevinsala wrote:

`ArgsSize` initialized to zero in the declaration. But I will move the initialization here to be clearer.

https://github.com/llvm/llvm-project/pull/172249