CUDA 11.2+ diagnostic-error flags are appended to the wrong variable

[WeiqunZhang]

• Type: correctness
• Severity: medium
• Component: CMake CUDA backend
• Location: Tools/CMake/AMReXParallelBackends.cmake:199

Problem

The CUDA backend accumulates compiler flags in _cuda_flags and later applies
that list to every amrex_${D}d target. The CUDA 11.2+ branch intended to make
device-variable misuse a hard error appends to _cuda_flag instead:

if (CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11.2)
   list(APPEND _cuda_flag --display-error-number "SHELL:--diag-error 20092")
endif ()

The target application step uses _cuda_flags, so these diagnostics are never
actually enabled.

Impact

• CUDA 11.2+ builds silently miss the extra diagnostic enforcement that this
  branch is meant to provide.
• Misuse that should be diagnosed as an error can pass through this configuration
  path without the intended nvcc flags.

Suggested patch

--- a/Tools/CMake/AMReXParallelBackends.cmake
+++ b/Tools/CMake/AMReXParallelBackends.cmake
@@
-   if (CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11.2)
-      list(APPEND _cuda_flag --display-error-number "SHELL:--diag-error 20092")
+   if (CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11.2)
+      list(APPEND _cuda_flags --display-error-number "SHELL:--diag-error 20092")
    endif ()

Prepared by Codex