[flang-rt] Fixed warnings and miscompilations in CUDA build. (#134470)

* DescribeIEEESignaledExceptions() is unused on the device - warning.
* StopStatementText() could return while marked noreturn - warning.
* Including cuda/std/complex only in the device compilation
  may cause nvcc to try to register variables in `cuda` namespace,
  while they are not defined in the host compilation - error.
  I decided to include cuda/std/complex always under RT_USE_LIBCUDACXX.

flang-rt/include/flang-rt/runtime/terminator.h

@@ -115,6 +115,21 @@ private:
 RT_API_ATTRS void NotifyOtherImagesOfNormalEnd();
 RT_API_ATTRS void NotifyOtherImagesOfFailImageStatement();
 RT_API_ATTRS void NotifyOtherImagesOfErrorTermination();
+
+#if defined(RT_DEVICE_COMPILATION)
+/// Trap the execution on the device.
+[[noreturn]] void RT_API_ATTRS DeviceTrap() {
+#if defined(__CUDACC__)
+  // NVCC supports __trap().
+  __trap();
+#elif defined(__clang__)
+  // Clang supports __builtin_trap().
+  __builtin_trap();
+#else
+#error "unsupported compiler"
+#endif
+}
+#endif
 } // namespace Fortran::runtime
 
 namespace Fortran::runtime::io {

flang-rt/lib/runtime/stop.cpp

@@ -23,7 +23,7 @@
 
 extern "C" {
 
-static void DescribeIEEESignaledExceptions() {
+[[maybe_unused]] static void DescribeIEEESignaledExceptions() {
 #ifdef fetestexcept // a macro in some environments; omit std::
   auto excepts{fetestexcept(FE_ALL_EXCEPT)};
 #else
@@ -82,15 +82,7 @@ static void CloseAllExternalUnits(const char *why) {
     }
     std::printf("\n");
   }
-#if defined(__CUDACC__)
+  Fortran::runtime::DeviceTrap();
-  // NVCC supports __trap().
-  __trap();
-#elif defined(__clang__)
-  // Clang supports __builtin_trap().
-  __builtin_trap();
-#else
-#error "unsupported compiler"
-#endif
 #else
   CloseAllExternalUnits("STOP statement");
   if (Fortran::runtime::executionEnvironment.noStopMessage && code == 0) {
@@ -119,17 +111,7 @@ static void CloseAllExternalUnits(const char *why) {
           "Fortran %s: %s\n", isErrorStop ? "ERROR STOP" : "STOP", code);
     }
   }
-  if (isErrorStop) {
+  Fortran::runtime::DeviceTrap();
-#if defined(__CUDACC__)
-    // NVCC supports __trap().
-    __trap();
-#elif defined(__clang__)
-    // Clang supports __builtin_trap().
-    __builtin_trap();
-#else
-#error "unsupported compiler"
-#endif
-  }
 #else
   CloseAllExternalUnits("STOP statement");
   if (!quiet) {

flang-rt/lib/runtime/terminator.cpp

@@ -75,15 +75,7 @@ RT_API_ATTRS void Terminator::CrashHeader() const {
 #endif
   NotifyOtherImagesOfErrorTermination();
 #if defined(RT_DEVICE_COMPILATION)
-#if defined(__CUDACC__)
+  DeviceTrap();
-  // NVCC supports __trap().
-  __trap();
-#elif defined(__clang__)
-  // Clang supports __builtin_trap().
-  __builtin_trap();
-#else
-#error "unsupported compiler"
-#endif
 #else
   std::abort();
 #endif

flang/include/flang/Common/enum-class.h

@@ -17,7 +17,6 @@
 #ifndef FORTRAN_COMMON_ENUM_CLASS_H_
 #define FORTRAN_COMMON_ENUM_CLASS_H_
 
-#include "flang/Common/variant.h"
 #include <array>
 #include <string>
 

flang/include/flang/Runtime/complex.h

@@ -16,8 +16,11 @@
 
 #include "flang/Common/api-attrs.h"
 
-#if RT_USE_LIBCUDACXX && defined(RT_DEVICE_COMPILATION)
+#if RT_USE_LIBCUDACXX
 #include <cuda/std/complex>
+#endif
+
+#if RT_USE_LIBCUDACXX && defined(RT_DEVICE_COMPILATION)
 namespace Fortran::runtime::rtcmplx {
 using cuda::std::complex;
 using cuda::std::conj;