Use opaque pointers

docker/Dockerfile

@@ -31,15 +31,15 @@ RUN apt-get update && apt-get install -y \
 # LLVM
 RUN wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key | apt-key add - && \
     apt-add-repository \
-        'deb http://apt.llvm.org/jammy/ llvm-toolchain-jammy main' && \
+        'deb http://apt.llvm.org/jammy/ llvm-toolchain-jammy-16 main' && \
     apt-get update && apt-get install -y \
-        llvm-15 \
-        llvm-15-dev \
-        clang-15 \
-        libclang-15-dev \
+        llvm-16 \
+        llvm-16-dev \
+        clang-16 \
+        libclang-16-dev \
         --
 
-ENV PATH=/usr/lib/llvm-15/bin${PATH:+:${PATH}}
+ENV PATH=/usr/lib/llvm-16/bin${PATH:+:${PATH}}
 
 # Dependencies
 RUN apt-get update && apt-get install -y \

docker/Dockerfile.l0

@@ -12,7 +12,7 @@ RUN mkdir /intel-gpu-drivers && cd /intel-gpu-drivers && \
 
 RUN cd /intel-gpu-drivers && dpkg -i *.deb
 
-RUN git clone -b llvm_release_150 https://github.com/KhronosGroup/SPIRV-LLVM-Translator.git && \
+RUN git clone -b llvm_release_160 https://github.com/KhronosGroup/SPIRV-LLVM-Translator.git && \
     mkdir SPIRV-LLVM-Translator/build && cd SPIRV-LLVM-Translator/build && \
     cmake -Wno-dev .. && make llvm-spirv -j`nproc` && make install
 

omniscidb/QueryEngine/ArrayIR.cpp

@@ -142,9 +142,8 @@ std::vector<llvm::Value*> CodeGenerator::codegenArrayExpr(
         codegen_traits_desc.local_addr_space_) {
       allocated_target_buffer = ir_builder.CreateAddrSpaceCast(
           allocated_target_buffer,
-          llvm::PointerType::get(
-              allocated_target_buffer->getType()->getPointerElementType(),
-              co.codegen_traits_desc.local_addr_space_),
+          llvm::PointerType::get(cgen_state_->context_,
+                                 co.codegen_traits_desc.local_addr_space_),
           "allocated.target.buffer.cast");
     }
   } else {

omniscidb/QueryEngine/CMakeLists.txt

@@ -189,11 +189,6 @@ set(hdk_default_runtime_functions_module_dependencies
     GroupByRuntime.cpp
     TopKRuntime.cpp)
 
-list(APPEND DISABLE_OPAQUE_POINTERS_CMD "")
-if(${LLVM_PACKAGE_VERSION} VERSION_GREATER_EQUAL "15")
-    list(APPEND DISABLE_OPAQUE_POINTERS_CMD "-Xclang" "-no-opaque-pointers")
-endif()
-
 # Adds a custom command for producing llvm bitcode modules out of a .cpp source.
 # Any additional arguments are treated as clang args and bypassed as-is.
 function(precompile_llvm_module SOURCE_FILE SUFFIX)
@@ -204,7 +199,7 @@ function(precompile_llvm_module SOURCE_FILE SUFFIX)
         DEPENDS ${hdk_default_runtime_functions_module_dependencies} ${SOURCE_FILE}
         OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${result_module_name}
         COMMAND ${llvm_clangpp_cmd}
-        ARGS -std=c++17 ${RT_OPT_FLAGS} ${ARGN} -c -emit-llvm ${DISABLE_OPAQUE_POINTERS_CMD}
+        ARGS -std=c++17 ${RT_OPT_FLAGS} ${ARGN} -c -emit-llvm
         ${CLANG_SDK_INC} ${CLANG_CRT_INC} ${MAPD_DEFINITIONS} -DEXECUTE_INCLUDE
         -o ${CMAKE_CURRENT_BINARY_DIR}/${result_module_name}
         -I ${CMAKE_CURRENT_SOURCE_DIR}/../
@@ -281,7 +276,7 @@ add_custom_command(
         DEPENDS ExtensionFunctions.hpp ExtensionFunctionsArray.hpp ExtensionFunctionsTesting.hpp
         OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ExtensionFunctions.ast.raw
         COMMAND ${llvm_clangpp_cmd}
-        ARGS -DNO_BOOST -std=c++17 -fsyntax-only -Xclang -ast-dump -fno-diagnostics-color -Wno-return-type-c-linkage ${DISABLE_OPAQUE_POINTERS_CMD}
+        ARGS -DNO_BOOST -std=c++17 -fsyntax-only -Xclang -ast-dump -fno-diagnostics-color -Wno-return-type-c-linkage
         -I ${CMAKE_CURRENT_SOURCE_DIR}/../ ${CMAKE_CURRENT_SOURCE_DIR}/ExtensionFunctions.hpp > ${CMAKE_CURRENT_BINARY_DIR}/ExtensionFunctions.ast.raw)
 
 add_custom_command(

omniscidb/QueryEngine/CgenState.h

@@ -214,14 +214,26 @@ struct CgenState {
   llvm::Value* emitExternalCall(
       const std::string& fname,
       llvm::Type* ret_type,
-      const std::vector<llvm::Value*> args,
+      const std::vector<llvm::Value*> args_in,
       const std::vector<llvm::Attribute::AttrKind>& fnattrs = {},
       const bool has_struct_return = false) {
+    auto args = args_in;
     std::vector<llvm::Type*> arg_types;
-    for (const auto arg : args) {
+    std::set<size_t> struct_args;
+    for (size_t i = 0; i < args.size(); i++) {
+      const auto arg = args[i];
       CHECK(arg);
-      arg_types.push_back(arg->getType());
+      if (arg->getType()->isStructTy()) {
+        CHECK(struct_args.insert(i).second);
+        auto arg_as_load = llvm::dyn_cast<llvm::LoadInst>(arg);
+        CHECK(arg_as_load);
+        arg_types.push_back(arg_as_load->getPointerOperand()->getType());
+        args[i] = arg_as_load->getPointerOperand();
+      } else {
+        arg_types.push_back(arg->getType());
+      }
     }
+    // TODO(adb): check ret type
     auto func_ty = llvm::FunctionType::get(ret_type, arg_types, false);
     llvm::AttributeList attrs;
     if (!fnattrs.empty()) {
@@ -253,27 +265,26 @@ struct CgenState {
     CHECK(func_type);
     if (has_struct_return) {
       const auto arg_ti = func_type->getParamType(0);
-      CHECK(arg_ti->isPointerTy() && arg_ti->getPointerElementType()->isStructTy());
+      CHECK(arg_ti->isPointerTy());
       auto attr_list = func->getAttributes();
 #if LLVM_VERSION_MAJOR > 13
       llvm::AttrBuilder arr_arg_builder(context_, attr_list.getParamAttrs(0));
 #else
       llvm::AttrBuilder arr_arg_builder(attr_list.getParamAttributes(0));
 #endif
-      arr_arg_builder.addAttribute(llvm::Attribute::StructRet);
+      arr_arg_builder.addStructRetAttr(args_in[0]->getType());
       func->addParamAttrs(0, arr_arg_builder);
     }
     const size_t arg_start = has_struct_return ? 1 : 0;
     for (size_t i = arg_start; i < func->arg_size(); i++) {
-      const auto arg_ti = func_type->getParamType(i);
-      if (arg_ti->isPointerTy() && arg_ti->getPointerElementType()->isStructTy()) {
+      if (struct_args.count(i) > 0) {
         auto attr_list = func->getAttributes();
 #if LLVM_VERSION_MAJOR > 13
         llvm::AttrBuilder arr_arg_builder(context_, attr_list.getParamAttrs(i));
 #else
         llvm::AttrBuilder arr_arg_builder(attr_list.getParamAttributes(i));
 #endif
-        arr_arg_builder.addByValAttr(arg_ti->getPointerElementType());
+        arr_arg_builder.addByValAttr(args_in[i]->getType());
         func->addParamAttrs(i, arr_arg_builder);
       }
     }

omniscidb/QueryEngine/ColumnIR.cpp

@@ -273,7 +273,7 @@ llvm::Value* CodeGenerator::codegenRowId(const hdk::ir::ColumnVar* col_var,
   } else if (col_var->rteIdx() > 0) {
     auto frag_off_ptr = get_arg_by_name(cgen_state_->row_func_, "frag_row_off");
     auto input_off_ptr = cgen_state_->ir_builder_.CreateGEP(
-        frag_off_ptr->getType()->getScalarType()->getPointerElementType(),
+        get_int_type(64, cgen_state_->context_),
         frag_off_ptr,
         cgen_state_->llInt(int32_t(col_var->rteIdx())));
     auto rowid_offset_lv = cgen_state_->ir_builder_.CreateLoad(

omniscidb/QueryEngine/Compiler/Backend.cpp

@@ -27,6 +27,7 @@
 #include <llvm/IR/LegacyPassManager.h>
 #include <llvm/IR/Verifier.h>
 #include <llvm/IRReader/IRReader.h>
+#include <llvm/Passes/PassBuilder.h>
 #include <llvm/Support/TargetSelect.h>
 #include <llvm/Transforms/IPO/PassManagerBuilder.h>
 
@@ -592,7 +593,6 @@ auto insert_emtpy_abort_replacement(llvm::Module* m) {
 
 void CUDABackend::linkModuleWithLibdevice(const std::unique_ptr<llvm::Module>& ext,
                                           llvm::Module& llvm_module,
-                                          llvm::PassManagerBuilder& pass_manager_builder,
                                           const GPUTarget& gpu_target,
                                           llvm::TargetMachine* nvptx_target_machine) {
 #ifdef HAVE_CUDA
@@ -641,17 +641,13 @@ void CUDABackend::linkModuleWithLibdevice(const std::unique_ptr<llvm::Module>& e
     fn.addFnAttr("nvptx-f32ftz", "true");
   }
 
-  // add nvvm reflect pass replacing any NVVM conditionals with constants
-  nvptx_target_machine->adjustPassManager(pass_manager_builder);
-  llvm::legacy::FunctionPassManager FPM(&llvm_module);
-  pass_manager_builder.populateFunctionPassManager(FPM);
+  llvm::PassBuilder pass_builder(nvptx_target_machine);
+  llvm::ModuleAnalysisManager module_analysis_manager;
+  pass_builder.registerModuleAnalyses(module_analysis_manager);
+  llvm::ModulePassManager module_pass_manager;
+  CHECK(!pass_builder.parsePassPipeline(module_pass_manager, "NVVMReflect"));
 
-  // Run the NVVMReflectPass here rather than inside optimize_ir
-  FPM.doInitialization();
-  for (auto& F : llvm_module) {
-    FPM.run(F);
-  }
-  FPM.doFinalization();
+  module_pass_manager.run(llvm_module, module_analysis_manager);
 #endif
 }
 
@@ -688,18 +684,12 @@ std::shared_ptr<CudaCompilationContext> CUDABackend::generateNativeGPUCode(
     `gpu_target.cgen_state->module_` appears to be the same as `llvm_module`
    */
   CHECK(gpu_target.cgen_state->module_ == llvm_module);
-  llvm::PassManagerBuilder pass_manager_builder = llvm::PassManagerBuilder();
-
-  pass_manager_builder.OptLevel = 0;
-  llvm::legacy::PassManager module_pass_manager;
-  pass_manager_builder.populateModulePassManager(module_pass_manager);
 
   bool requires_libdevice = check_module_requires_libdevice(llvm_module);
 
   if (requires_libdevice) {
     linkModuleWithLibdevice(exts.at(ExtModuleKinds::rt_libdevice_module),
                             *llvm_module,
-                            pass_manager_builder,
                             gpu_target,
                             nvptx_target_machine);
   }

omniscidb/QueryEngine/Compiler/Backend.h

@@ -87,12 +87,18 @@ class CodegenTraits {
   llvm::PointerType* localPointerType(llvm::Type* ElementType) const {
     return llvm::PointerType::get(ElementType, local_addr_space_);
   }
+  llvm::PointerType* localOpaquePtr(llvm::LLVMContext& ctx) const {
+    return llvm::PointerType::get(ctx, local_addr_space_);
+  }
   llvm::PointerType* smemPointerType(llvm::Type* ElementType) const {
     return llvm::PointerType::get(ElementType, smem_addr_space_);
   }
   llvm::PointerType* globalPointerType(llvm::Type* ElementType) const {
     return llvm::PointerType::get(ElementType, global_addr_space_);
   }
+  llvm::PointerType* globalOpaquePtr(llvm::LLVMContext& ctx) const {
+    return llvm::PointerType::get(ctx, global_addr_space_);
+  }
   llvm::CallingConv::ID callingConv() const { return conv_; }
   llvm::StringRef dataLayout() const {
     return llvm::StringRef(
@@ -167,7 +173,6 @@ class CUDABackend : public Backend {
 
   static void linkModuleWithLibdevice(const std::unique_ptr<llvm::Module>& ext,
                                       llvm::Module& module,
-                                      llvm::PassManagerBuilder& pass_manager_builder,
                                       const GPUTarget& gpu_target,
                                       llvm::TargetMachine* nvptx_target_machine);
 

omniscidb/QueryEngine/Compiler/HelperFunctions.cpp

@@ -55,7 +55,10 @@ namespace compiler {
 std::string mangle_spirv_builtin(const llvm::Function& func) {
   CHECK(func.getName().startswith("__spirv_")) << func.getName().str();
   std::string new_name;
-#if LLVM_VERSION_MAJOR > 14
+#if LLVM_VERSION_MAJOR > 15
+  llvm::mangleOpenClBuiltin(func.getName().str(), func.getArg(0)->getType(), new_name);
+
+#elif LLVM_VERSION_MAJOR > 14
   mangleOpenClBuiltin(
       func.getName().str(), func.getArg(0)->getType(), /*pointer_types=*/{}, new_name);
 #else