Merge topic 'cuda_host_compiler_fail'

01428c5560 CUDA: Fail fast if CMAKE_CUDA_ARCHITECTURES doesn't work during detection
9f81aa0f69 CUDA: Fail if compiler detection using the host compiler fails

Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !5155

3 files changed, 35 insertions, 21 deletions

diff --git a/Help/release/dev/cuda-fail-fast.rst b/Help/release/dev/cuda-fail-fast.rst
new file mode 100644
index 0000000..d857eb2
--- /dev/null
+++ b/Help/release/dev/cuda-fail-fast.rst
@@ -0,0 +1,6 @@
+cuda-fail-fast
+--------------
+
+* If ``CUDA`` compiler detection fails with user-specified
+  :variable:`CMAKE_CUDA_ARCHITECTURES` or :variable:`CMAKE_CUDA_HOST_COMPILER`
+  an error is raised.
diff --git a/Modules/CMakeDetermineCUDACompiler.cmake b/Modules/CMakeDetermineCUDACompiler.cmake
index daca382..6bdd06b 100644
--- a/Modules/CMakeDetermineCUDACompiler.cmake
+++ b/Modules/CMakeDetermineCUDACompiler.cmake
@@ -189,6 +189,10 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
     set(nvcc_test_flags "--keep --keep-dir tmp")
     if(CMAKE_CUDA_HOST_COMPILER)
       string(APPEND nvcc_test_flags " -ccbin=\"${CMAKE_CUDA_HOST_COMPILER}\"")
+
+      # If the user has specified a host compiler we should fail instead of trying without.
+      # Succeeding detection without may result in confusing errors later on, see #21076.
+      set(CMAKE_CUDA_COMPILER_ID_REQUIRE_SUCCESS ON)
     endif()
   elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
     if(WIN32)
@@ -202,34 +206,35 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
     endif()
   endif()
 
-  # First try with the user-specified architectures.
+  # Append user-specified architectures.
   if(CMAKE_CUDA_ARCHITECTURES)
-    set(clang_archs "${clang_test_flags}")
-    set(nvcc_archs "${nvcc_test_flags}")
-
     foreach(arch ${CMAKE_CUDA_ARCHITECTURES})
       # Strip specifiers as PTX vs binary doesn't matter.
       string(REGEX MATCH "[0-9]+" arch_name "${arch}")
-      string(APPEND clang_archs " --cuda-gpu-arch=sm_${arch_name}")
-      string(APPEND nvcc_archs " -gencode=arch=compute_${arch_name},code=sm_${arch_name}")
+      string(APPEND clang_test_flags " --cuda-gpu-arch=sm_${arch_name}")
+      string(APPEND nvcc_test_flags " -gencode=arch=compute_${arch_name},code=sm_${arch_name}")
       list(APPEND tested_architectures "${arch_name}")
     endforeach()
 
-    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_archs}")
-    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${nvcc_archs}")
+    # If the user has specified architectures we'll want to fail during compiler detection if they don't work.
+    set(CMAKE_CUDA_COMPILER_ID_REQUIRE_SUCCESS ON)
   endif()
 
-  # Fallback default NVCC flags.
-  list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST ${nvcc_test_flags})
-
-  # Clang doesn't automatically select an architecture supported by the SDK.
-  # Try in reverse order of deprecation with the most recent at front (i.e. the most likely to work for new setups).
-  foreach(arch "20" "30" "52")
-    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags} --cuda-gpu-arch=sm_${arch}")
-  endforeach()
+  if(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
+    if(NOT CMAKE_CUDA_ARCHITECTURES)
+      # Clang doesn't automatically select an architecture supported by the SDK.
+      # Try in reverse order of deprecation with the most recent at front (i.e. the most likely to work for new setups).
+      foreach(arch "20" "30" "52")
+        list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags} --cuda-gpu-arch=sm_${arch}")
+      endforeach()
+    endif()
 
-  # Finally also try the default.
-  list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags}")
+    # If the user specified CMAKE_CUDA_ARCHITECTURES this will include all the architecture flags.
+    # Otherwise this won't include any architecture flags and we'll fallback to Clang's defaults.
+    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags}")
+  elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
+    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${nvcc_test_flags}")
+  endif()
 
   # We perform compiler identification for a second time to extract implicit linking info and host compiler for NVCC.
   # We also use it to verify that CMAKE_CUDA_ARCHITECTURES and additionally on Clang that CUDA toolkit path works.
diff --git a/Modules/CMakeDetermineCompilerId.cmake b/Modules/CMakeDetermineCompilerId.cmake
index 8724ed8..688e1d8 100644
--- a/Modules/CMakeDetermineCompilerId.cmake
+++ b/Modules/CMakeDetermineCompilerId.cmake
@@ -628,9 +628,12 @@ ${CMAKE_${lang}_COMPILER_ID_OUTPUT}
 
 ")
     file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeError.log "${MSG}")
-    #if(NOT CMAKE_${lang}_COMPILER_ID_ALLOW_FAIL)
-    #  message(FATAL_ERROR "${MSG}")
-    #endif()
+
+    # Some languages may know the correct/desired set of flags and want to fail right away if they don't work.
+    # This is currently only used by CUDA.
+    if(CMAKE_${lang}_COMPILER_ID_REQUIRE_SUCCESS)
+      message(FATAL_ERROR "${MSG}")
+    endif()
 
     # No output files should be inspected.
     set(COMPILER_${lang}_PRODUCED_FILES)