LLVM 16 and 17 support (#1730)

* Bump minimum supported LLVM version to 17 * Rename `makeArrayRef` to `toArrayRef` to avoid `llvm::makeArrayRef` deprecation warnings. * Minor required changes related to `llvm::Align`. * Disable `initializeInstrumentation` and `createPruneEHPass` which were remove and appear to serve no purpose now. * Add assert when using default optimization levels with legacy pass manager, which is no longer supported. This depends on both opaque pointers and new pass manager support. Signed-off-by: Brecht Van Lommel <brecht@blender.org> * Keep CMAKE_CXX_STANDARD warning for LLVM 16 * Enable LLVM 17 too, adjust for changes in new pass manager PR * Fix OptiX data layout mismatch with LLVM 16+ i128:128 was already in the NVPTX backend since LLVM 6, the docs this was taken from seem outdated. Unclear it started failing now, but the data layout seems clearly incomplete without this. * Modify CI to run the bleeding edge test on LLVM 16 * Manually make global variable names valid for PTX * Fix build with older LLVM --------- Signed-off-by: Brecht Van Lommel <brecht@blender.org> Co-authored-by: Larry Gritz <lg@larrygritz.com> Co-authored-by: Tim Grant <tgrant@nvidia.com>
AcademySoftwareFoundation · Oct 4, 2023 · cdfec4f · cdfec4f
1 parent 3567aac
commit cdfec4f
Show file tree

Hide file tree

Showing 9 changed files with 148 additions and 115 deletions.
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -334,7 +334,7 @@ jobs:
             setenvs: export LLVM_VERSION=10.0.0
                             OPENIMAGEIO_CMAKE_FLAGS="-DBUILD_FMT_VERSION=7.0.1"
                             PUGIXML_VERSION=v1.10
-          - desc: latest releases gcc11/C++17 llvm13 boost1.71 exr3.2 py3.9 avx2 batch-b16avx512
+          - desc: latest releases gcc11/C++17 llvm15 boost1.71 exr3.2 py3.9 avx2 batch-b16avx512
             nametag: linux-latest-releases
             os: ubuntu-22.04
             cxx_compiler: g++-11
@@ -346,11 +346,11 @@ jobs:
             python_ver: "3.10"
             simd: avx2,f16c
             batched: b8_AVX2,b8_AVX512,b16_AVX512
-            setenvs: export LLVM_VERSION=14.0.0
+            setenvs: export LLVM_VERSION=15.0.6
                             LLVM_DISTRO_NAME=ubuntu-18.04
                             OPENCOLORIO_VERSION=v2.2.0
                             PUGIXML_VERSION=v1.13
-          - desc: bleeding edge gcc12/C++17 llvm14 oiio/ocio/exr/pybind-master boost1.71 py3.10 avx2 batch-b16avx512
+          - desc: bleeding edge gcc12/C++17 llvm16 oiio/ocio/exr/pybind-master boost1.71 py3.10 avx2 batch-b16avx512
             nametag: linux-bleeding-edge
             os: ubuntu-22.04
             cxx_compiler: g++-12
@@ -362,8 +362,8 @@ jobs:
             python_ver: "3.10"
             simd: avx2,f16c
             batched: b8_AVX2,b8_AVX512,b16_AVX512
-            setenvs: export LLVM_VERSION=15.0.6
-                            LLVM_DISTRO_NAME=ubuntu-18.04
+            setenvs: export LLVM_VERSION=16.0.4
+                            LLVM_DISTRO_NAME=ubuntu-22.04
                             OPENCOLORIO_VERSION=main
                             PUGIXML_VERSION=master
           - desc: clang14/C++17 llvm14 boost1.71 exr3.1 py3.8 avx2 batch-b16avx512

diff --git a/src/cmake/cuda_macros.cmake b/src/cmake/cuda_macros.cmake
@@ -207,8 +207,8 @@ function ( CUDA_SHADEOPS_COMPILE prefix output_bc output_ptx input_srcs headers
     endforeach ()
 
     if (LLVM_NEW_PASS_MANAGER)
-      # --nvptx-assign-valid-global-names does not appaer to be supported
-      # by either opt or llc with the new pass manager.
+      # There is no --nvptx-assign-valid-global-names flag for the new
+      # pass manager, but it appears to run this pass by default.
       string(REPLACE "-O" "O" opt_tool_flags ${CUDA_OPT_FLAG_CLANG})
       set (opt_tool_flags -passes="default<${opt_tool_flags}>")
     else()

diff --git a/src/cmake/externalpackages.cmake b/src/cmake/externalpackages.cmake
@@ -117,7 +117,7 @@ checked_find_package (pugixml REQUIRED
 # LLVM library setup
 checked_find_package (LLVM REQUIRED
                       VERSION_MIN 9.0
-                      VERSION_MAX 15.9
+                      VERSION_MAX 17.9
                       PRINT LLVM_SYSTEM_LIBRARIES CLANG_LIBRARIES)
 # ensure include directory is added (in case of non-standard locations
 include_directories (BEFORE SYSTEM "${LLVM_INCLUDES}")
@@ -148,7 +148,6 @@ if (LLVM_VERSION VERSION_GREATER_EQUAL 15.0
          "15 or higher, or you may get build errors.${ColorReset}\n")
 endif ()
 if (LLVM_VERSION VERSION_GREATER_EQUAL 16.0)
-    message (ERROR "${ColorYellow}OSL is not yet compatible with LLVM 16.${ColorReset}\n")
     if (CMAKE_CXX_STANDARD VERSION_LESS 17)
         message (WARNING "${ColorYellow}LLVM 16+ requires C++17 or higher. "
             "Please set CMAKE_CXX_STANDARD to 17 or higher.${ColorReset}\n")

diff --git a/src/cmake/modules/FindLLVM.cmake b/src/cmake/modules/FindLLVM.cmake
@@ -126,8 +126,9 @@ if (LLVM_VERSION VERSION_GREATER_EQUAL 9.0 AND (LLVM_SHARED_MODE STREQUAL "share
 endif ()
 
 foreach (COMPONENT clangFrontend clangDriver clangSerialization
-                   clangParse clangSema clangAnalysis clangAST clangBasic
-                   clangEdit clangLex clangSupport)
+                   clangParse clangSema clangAnalysis clangAST
+                   clangASTMatchers clangBasic clangEdit clangLex
+                   clangSupport)
     find_library ( _CLANG_${COMPONENT}_LIBRARY
                   NAMES ${COMPONENT}
                   PATHS ${LLVM_LIB_DIR})

diff --git a/src/include/OSL/oslexec.h b/src/include/OSL/oslexec.h
@@ -51,7 +51,8 @@ class ShadingSystemImpl;
 }
 
 #ifdef __CUDA_ARCH__
-#    define STRING_PARAMS(x) UStringHash::Hash(__OSL_STRINGIFY(x))
+#    define STRING_PARAMS(x) \
+        UStringHash::HashConstEval<UStringHash::Hash(__OSL_STRINGIFY(x))>
 #else
 #    define STRING_PARAMS(x) StringParams::x
 #endif

diff --git a/src/liboslexec/llvm_instance.cpp b/src/liboslexec/llvm_instance.cpp
@@ -1751,7 +1751,7 @@ BackendLLVM::run()
             // The target triple and data layout used here are those specified
             // for NVPTX (https://www.llvm.org/docs/NVPTXUsage.html#triples).
             ll.module()->setDataLayout(
-                "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
+                "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
             ll.module()->setTargetTriple("nvptx64-nvidia-cuda");
         }
 #    endif
@@ -1816,7 +1816,7 @@ BackendLLVM::run()
                     err);
 
             shadeops_module->setDataLayout(
-                "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
+                "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
             shadeops_module->setTargetTriple("nvptx64-nvidia-cuda");
 
             std::unique_ptr<llvm::Module> shadeops_ptr(shadeops_module);
@@ -1843,7 +1843,7 @@ BackendLLVM::run()
                         err);
 
                 rend_lib_module->setDataLayout(
-                    "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
+                    "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
                 rend_lib_module->setTargetTriple("nvptx64-nvidia-cuda");
 
                 for (llvm::Function& fn : *rend_lib_module) {
@@ -1863,7 +1863,7 @@ BackendLLVM::run()
             // See: https://llvm.org/docs/NVPTXUsage.html
             ll.module()->setTargetTriple("nvptx64-nvidia-cuda");
             ll.module()->setDataLayout(
-                "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
+                "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64");
 
             // Tag each function as an OSL library function to help with
             // inlining and optimization after codegen.
@@ -1877,6 +1877,16 @@ BackendLLVM::run()
                 global.setLinkage(llvm::GlobalValue::ExternalLinkage);
                 global.setExternallyInitialized(true);
                 global.setInitializer(nullptr);
+                // Replace characters not supported in ptx, matching the LLVM
+                // NVPTXAssignValidGlobalNames pass.
+                string_view global_name(global.getName().data(),
+                                        global.getName().size());
+                if (Strutil::contains_any_char(global_name, ".@")) {
+                    std::string valid_name = global_name;
+                    valid_name = Strutil::replace(valid_name, ".", "_$_", true);
+                    valid_name = Strutil::replace(valid_name, "@", "_$_", true);
+                    global.setName(valid_name);
+                }
             }
         }
         OSL_ASSERT(ll.module());