[Cuda] Handle -fcuda-short-ptr even with -nocudalib (#111682)

When passed -nocudalib/-nogpulib, Cuda's argument handling would bail
out before handling -fcuda-short-ptr, meaning the frontend and backend
data layouts would mismatch.

Author: frasercrmck

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -848,6 +848,10 @@ void CudaToolChain::addClangTargetOptions(
   if (CudaInstallation.version() >= CudaVersion::CUDA_90)
     CC1Args.push_back("-fcuda-allow-variadic-functions");
 
+  if (DriverArgs.hasFlag(options::OPT_fcuda_short_ptr,
+                         options::OPT_fno_cuda_short_ptr, false))
+    CC1Args.append({"-mllvm", "--nvptx-short-ptr"});
+
   if (DriverArgs.hasArg(options::OPT_nogpulib))
     return;
 
@@ -873,10 +877,6 @@ void CudaToolChain::addClangTargetOptions(
 
   clang::CudaVersion CudaInstallationVersion = CudaInstallation.version();
 
-  if (DriverArgs.hasFlag(options::OPT_fcuda_short_ptr,
-                         options::OPT_fno_cuda_short_ptr, false))
-    CC1Args.append({"-mllvm", "--nvptx-short-ptr"});
-
   if (CudaInstallationVersion >= CudaVersion::UNKNOWN)
     CC1Args.push_back(
         DriverArgs.MakeArgString(Twine("-target-sdk-version=") +

clang/test/Driver/cuda-short-ptr.cu

@@ -0,0 +1,6 @@
+// Checks that cuda compilation does the right thing when passed -fcuda-short-ptr
+
+// RUN: %clang -### --target=x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=sm_20 -fcuda-short-ptr -nocudainc -nocudalib --cuda-path=%S/Inputs/CUDA/usr/local/cuda %s 2>&1 | FileCheck %s
+
+// CHECK: "-mllvm" "--nvptx-short-ptr"
+// CHECK-SAME: "-fcuda-short-ptr"