[clang] [llvm] [NVPTX] Add errors for incorrect CUDA addrpaces (PR #138706)

[Lewis Crawford via llvm-commits]

================
@@ -1399,19 +1399,27 @@ void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
       if (PTy) {
         O << "\t.param .u" << PTySizeInBits << " .ptr";
 
+        bool IsCUDA = static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() ==
+                      NVPTX::CUDA;
         switch (PTy->getAddressSpace()) {
         default:
           break;
         case ADDRESS_SPACE_GLOBAL:
           O << " .global";
           break;
         case ADDRESS_SPACE_SHARED:
+          if (IsCUDA)
+            report_fatal_error(".shared ptr kernel args unsupported in CUDA.");
           O << " .shared";
           break;
         case ADDRESS_SPACE_CONST:
+          if (IsCUDA)
+            report_fatal_error(".const ptr kernel args unsupported in CUDA.");
----------------
LewisCrawford wrote:

I've updated the patch to fix the `CodeGenCUDA/memcpy-libcall.cu` test, which shows that even the clang frontend with a `-cuda` triple can still generate invalid PTX which will crash when executed in a CUDA context. This will cause a confusing `CUDA_ERROR_INVALID_IMAGE` error message at run-time (similar to the one experienced here https://github.com/llvm/llvm-project/pull/114874#issuecomment-2658056661 ), rather than failing to compile and giving the more explicit error message I've introduced in this patch.

Unfortunately, this issue about CUDA only supporting .global and .generic pointers in kernel args, is one of several cases that are missing in the PTX documentation, where the PTX is technically legal according to the spec, but is invalid when used within the CUDA API (some texture-related queries is another case of this incompatibilty). We plan on adding public documentation about this soon though.

In the mean-time, it seems better to me to have some kind of readable error message about this to help implementers of various front-ends, even if the docs are currently missing it, since the `CUDA_ERROR_INVALID_IMAGE` caused when running this does not provide any useful information about the problem.

https://github.com/llvm/llvm-project/pull/138706