[llvm] [Offload] Add `--kernel <name>` command to `llvm-gpu-loader` (PR #213738)

[Joseph Huber via llvm-commits]

================
@@ -263,21 +273,31 @@ int main(int argc, const char **argv, const char **envp) {
   OFFLOAD_ERR(olMemAlloc(Device, OL_ALLOC_TYPE_DEVICE, sizeof(int), &DevRet));
   OFFLOAD_ERR(olMemcpy(Queue, DevRet, Device, &Zero, Host, sizeof(int)));
 
+  // The '_begin' and '_end' kernels perform libc startup and teardown. Global
+  // constructors and destructors are handled automatically by the runtime.
   ol_kernel_launch_size_args_t BeginLaunch{1, {1, 1, 1}, {1, 1, 1}, 0};
-  launchKernel(Queue, Device, Program, "_begin", BeginLaunch, DevArgc, DevArgv,
-               DevEnvp);
+  if (hasKernel(Program, "_begin"))
+    launchKernel(Queue, Device, Program, "_begin", BeginLaunch, DevArgc,
+                 DevArgv, DevEnvp);
   OFFLOAD_ERR(olSyncQueue(Queue));
 
   uint32_t Dims = (BlocksZ > 1) ? 3 : (BlocksY > 1) ? 2 : 1;
   ol_kernel_launch_size_args_t StartLaunch{Dims,
                                            {BlocksX, BlocksY, BlocksZ},
                                            {ThreadsX, ThreadsY, ThreadsZ},
                                            /*SharedMemBytes=*/0};
-  launchKernel(Queue, Device, Program, "_start", StartLaunch, DevArgc, DevArgv,
-               DevEnvp, DevRet);
+  if (Kernels.empty()) {
+    launchKernel(Queue, Device, Program, "_start", StartLaunch, DevArgc,
+                 DevArgv, DevEnvp, DevRet);
+  } else {
+    // Launch the user-specified kernels in order. These must take no arguments.
+    for (const std::string &Kernel : Kernels)
+      launchKernel(Queue, Device, Program, Kernel.c_str(), StartLaunch);
+  }
 
   ol_kernel_launch_size_args_t EndLaunch{1, {1, 1, 1}, {1, 1, 1}, 0};
-  launchKernel(Queue, Device, Program, "_end", EndLaunch);
+  if (hasKernel(Program, "_end"))
+    launchKernel(Queue, Device, Program, "_end", EndLaunch);
 
   int Ret;
----------------
jhuber6 wrote:

Not easily, there's nowhere in tree that it would fit so I'm mostly just hoping that it's simple enough, though I'm hopefully able to use this in the future for some compiler-rt code.

https://github.com/llvm/llvm-project/pull/213738