Improve CUDA/LLVM compatibility

src/cmake/cuda_compat/crt/math_functions.hpp

@@ -0,0 +1,21 @@
+// Copyright Contributors to the Open Shading Language project.
+// SPDX-License-Identifier: BSD-3-Clause
+
+// CUDA 13.2 added _NV_RSQRT_SPECIFIER to math_functions.hpp, but older Clang
+// CUDA wrappers include that file without defining the macro first. Match the
+// definition used by newer Clang wrappers, including glibc 2.42's exception
+// specifier.
+// Upstream fix: https://github.com/llvm/llvm-project/pull/185701
+#ifndef _NV_RSQRT_SPECIFIER
+#    if defined(__GNUC__) && defined(__GLIBC_PREREQ)
+#        if __GLIBC_PREREQ(2, 42)
+#            define _NV_RSQRT_SPECIFIER noexcept(true)
+#        endif
+#    endif
+#    ifndef _NV_RSQRT_SPECIFIER
+#        define _NV_RSQRT_SPECIFIER
+#    endif
+#endif
+
+// Clang's CUDA wrapper intentionally includes this header more than once.
+#include_next <crt/math_functions.hpp>

src/cmake/cuda_compat/cuda_stdlib.h

@@ -0,0 +1,14 @@
+// Copyright Contributors to the Open Shading Language project.
+// SPDX-License-Identifier: BSD-3-Clause
+
+#pragma once
+
+// Preload these headers before compiling OSL device code. CUDA's __noinline__
+// macro conflicts with attribute names in libstdc++; hide it during the includes
+// and restore it for device code. The headers' include guards prevent reprocessing.
+// Related upstream fix: https://github.com/llvm/llvm-project/pull/66138
+#pragma push_macro("__noinline__")
+#undef __noinline__
+#include <memory>
+#include <string>
+#pragma pop_macro("__noinline__")

src/cmake/cuda_compat/texture_fetch_functions.h

@@ -0,0 +1,7 @@
+// Copyright Contributors to the Open Shading Language project.
+// SPDX-License-Identifier: BSD-3-Clause
+
+// CUDA 13 removed this header, but older Clang CUDA wrappers still include it.
+// OSL does not use the legacy texture-reference API, so an empty stub is
+// sufficient for CUDA bitcode generation.
+#pragma once

src/cmake/cuda_macros.cmake

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # https://github.com/AcademySoftwareFoundation/OpenShadingLanguage
 
-set ($OSL_EXTRA_NVCC_ARGS "" CACHE STRING "Custom args passed to nvcc when compiling CUDA code")
+set (OSL_EXTRA_NVCC_ARGS "" CACHE STRING "Custom args passed to nvcc when compiling CUDA code")
 set (CUDA_OPT_FLAG_NVCC "-O3" CACHE STRING "The optimization level to use when compiling CUDA/C++ files with nvcc")
 set (CUDA_OPT_FLAG_CLANG "-O3" CACHE STRING "The optimization level to use when compiling CUDA/C++ files with clang")
 
@@ -148,6 +148,22 @@ function ( MAKE_CUDA_BITCODE src suffix generated_bc extra_clang_args )
         set (CUDA_TEXREF_FIX "-D__CLANG_CUDA_TEXTURE_INTRINSICS_H__")
     endif()
 
+    if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "13.0")
+        # CUDA 13 removed texture_fetch_functions.h, but older Clang CUDA
+        # wrappers still include it. Search OSL's compatibility headers first.
+        set (CUDA_COMPAT_INCLUDE
+             "-I${PROJECT_SOURCE_DIR}/src/cmake/cuda_compat")
+        list (APPEND exec_headers
+              "${PROJECT_SOURCE_DIR}/src/cmake/cuda_compat/crt/math_functions.hpp")
+    endif ()
+
+    # Load libstdc++ before OSL headers while CUDA's __noinline__ macro is hidden.
+    if (NOT WIN32 AND LLVM_VERSION VERSION_LESS 18.0)
+        set (CUDA_STDLIB_HEADER "${PROJECT_SOURCE_DIR}/src/cmake/cuda_compat/cuda_stdlib.h")
+        set (CLANG_CUDA_COMPAT_FLAGS -include "${CUDA_STDLIB_HEADER}")
+        list (APPEND exec_headers "${CUDA_STDLIB_HEADER}")
+    endif ()
+
     list (TRANSFORM IMATH_INCLUDES PREPEND -I
           OUTPUT_VARIABLE ALL_IMATH_INCLUDES)
     list (TRANSFORM OPENEXR_INCLUDES PREPEND -I
@@ -157,6 +173,8 @@ function ( MAKE_CUDA_BITCODE src suffix generated_bc extra_clang_args )
 
     add_custom_command (OUTPUT ${bc_cuda}
         COMMAND ${LLVM_BC_GENERATOR}
+            ${CUDA_COMPAT_INCLUDE}
+            ${CLANG_CUDA_COMPAT_FLAGS}
             "-I${OPTIX_INCLUDES}"
             "-I${CUDA_INCLUDES}"
             "-I${CMAKE_CURRENT_SOURCE_DIR}"

src/cmake/externalpackages.cmake

@@ -141,16 +141,6 @@ if (OSL_USE_OPTIX)
             message (FATAL_ERROR "NVPTX target is not available in the provided LLVM build")
         endif()
 
-        # CUDA 13.2+ requires LLVM 22.1.2+ for the _NV_RSQRT_SPECIFIER fix.
-        # See: https://github.com/AcademySoftwareFoundation/OpenShadingLanguage/issues/2129
-        if (CUDA_VERSION VERSION_GREATER_EQUAL "13.2" AND LLVM_VERSION VERSION_LESS "22.1.2")
-            message (FATAL_ERROR
-                "${ColorRed}CUDA ${CUDA_VERSION} requires LLVM 22.1.2 or newer "
-                "(you have LLVM ${LLVM_VERSION}). CUDA 13.2+ needs the "
-                "_NV_RSQRT_SPECIFIER fix from llvm-project commit c1c833734cf0, "
-                "first shipped in LLVM 22.1.2. Either upgrade LLVM or downgrade CUDA.${ColorReset}")
-        endif ()
-
         set (CUDA_LIB_FLAGS "--cuda-path=${CUDA_TOOLKIT_ROOT_DIR}")
 
         # If the user wants, try to use static libs here to putting static lib