NativeCodegen.cpp File Reference

#include "QueryEngine/Execute.h"
#include <llvm/Analysis/ScopedNoAliasAA.h>
#include <llvm/Analysis/TypeBasedAliasAnalysis.h>
#include <llvm/Bitcode/BitcodeReader.h>
#include <llvm/Bitcode/BitcodeWriter.h>
#include <llvm/ExecutionEngine/MCJIT.h>
#include <llvm/IR/Attributes.h>
#include <llvm/IR/GlobalValue.h>
#include <llvm/IR/InstIterator.h>
#include <llvm/IR/IntrinsicInst.h>
#include <llvm/IR/Intrinsics.h>
#include <llvm/IR/LegacyPassManager.h>
#include <llvm/IR/Verifier.h>
#include <llvm/IRReader/IRReader.h>
#include <llvm/Support/TargetRegistry.h>
#include <llvm/Support/Casting.h>
#include <llvm/Support/FileSystem.h>
#include <llvm/Support/FormattedStream.h>
#include <llvm/Support/MemoryBuffer.h>
#include <llvm/Support/SourceMgr.h>
#include <llvm/Support/TargetSelect.h>
#include <llvm/Support/raw_os_ostream.h>
#include <llvm/Support/raw_ostream.h>
#include <llvm/Transforms/IPO.h>
#include <llvm/Transforms/IPO/AlwaysInliner.h>
#include <llvm/Transforms/IPO/InferFunctionAttrs.h>
#include <llvm/Transforms/IPO/PassManagerBuilder.h>
#include <llvm/Transforms/InstCombine/InstCombine.h>
#include <llvm/Transforms/Instrumentation.h>
#include <llvm/Transforms/Scalar.h>
#include <llvm/Transforms/Scalar/GVN.h>
#include <llvm/Transforms/Scalar/InstSimplifyPass.h>
#include <llvm/Transforms/Utils.h>
#include <llvm/Transforms/Utils/BasicBlockUtils.h>
#include <llvm/Transforms/Utils/Cloning.h>
#include "CudaMgr/CudaMgr.h"
#include "QueryEngine/CodeGenerator.h"
#include "QueryEngine/CodegenHelper.h"
#include "QueryEngine/ExtensionFunctionsWhitelist.h"
#include "QueryEngine/GpuSharedMemoryUtils.h"
#include "QueryEngine/LLVMFunctionAttributesUtil.h"
#include "QueryEngine/Optimization/AnnotateInternalFunctionsPass.h"
#include "QueryEngine/OutputBufferInitialization.h"
#include "QueryEngine/QueryEngine.h"
#include "QueryEngine/QueryTemplateGenerator.h"
#include "QueryEngine/UsedColumnsVisitor.h"
#include "Shared/InlineNullValues.h"
#include "Shared/MathUtils.h"
#include "StreamingTopN.h"

Include dependency graph for NativeCodegen.cpp:

Go to the source code of this file.

Classes
class	anonymous_namespace{NativeCodegen.cpp}::CaseExprDetector

Namespaces
	anonymous_namespace{NativeCodegen.cpp}

Macros
#define	SHOW_DEFINED(MODULE)
#define	SHOW_FUNCTIONS(MODULE)

Functions
void	anonymous_namespace{NativeCodegen.cpp}::throw_parseIR_error (const llvm::SMDiagnostic &parse_error, std::string src="", const bool is_gpu=false)
template<typename T = void>
void	anonymous_namespace{NativeCodegen.cpp}::show_defined (llvm::Module &llvm_module)
template<typename T = void>
void	anonymous_namespace{NativeCodegen.cpp}::show_defined (llvm::Module *llvm_module)
template<typename T = void>
void	anonymous_namespace{NativeCodegen.cpp}::show_defined (std::unique_ptr< llvm::Module > &llvm_module)
template<typename T = void>
void	anonymous_namespace{NativeCodegen.cpp}::scan_function_calls (llvm::Function &F, std::unordered_set< std::string > &defined, std::unordered_set< std::string > &undefined, const std::unordered_set< std::string > &ignored)
template<typename T = void>
void	anonymous_namespace{NativeCodegen.cpp}::scan_function_calls (llvm::Module &llvm_module, std::unordered_set< std::string > &defined, std::unordered_set< std::string > &undefined, const std::unordered_set< std::string > &ignored)
template<typename T = void>
std::tuple< std::unordered_set < std::string > , std::unordered_set < std::string > >	anonymous_namespace{NativeCodegen.cpp}::scan_function_calls (llvm::Module &llvm_module, const std::unordered_set< std::string > &ignored={})
void	anonymous_namespace{NativeCodegen.cpp}::eliminate_dead_self_recursive_funcs (llvm::Module &M, const std::unordered_set< llvm::Function * > &live_funcs)
void	anonymous_namespace{NativeCodegen.cpp}::optimize_ir (llvm::Function *query_func, llvm::Module *llvm_module, llvm::legacy::PassManager &pass_manager, const std::unordered_set< llvm::Function * > &live_funcs, const bool is_gpu_smem_used, const CompilationOptions &co)
void	verify_function_ir (const llvm::Function *func)
std::string	anonymous_namespace{NativeCodegen.cpp}::assemblyForCPU (ExecutionEngineWrapper &execution_engine, llvm::Module *llvm_module)
ExecutionEngineWrapper	anonymous_namespace{NativeCodegen.cpp}::create_execution_engine (llvm::Module *llvm_module, llvm::EngineBuilder &eb, const CompilationOptions &co)
std::string	anonymous_namespace{NativeCodegen.cpp}::cpp_to_llvm_name (const std::string &s)
std::string	anonymous_namespace{NativeCodegen.cpp}::gen_array_any_all_sigs ()
std::string	anonymous_namespace{NativeCodegen.cpp}::gen_translate_null_key_sigs ()
llvm::StringRef	get_gpu_target_triple_string ()
llvm::StringRef	get_gpu_data_layout ()
std::map< std::string, std::string >	get_device_parameters (bool cpu_only)
std::unique_ptr< llvm::Module >	read_llvm_module_from_bc_file (const std::string &bc_filename, llvm::LLVMContext &context)
std::unique_ptr< llvm::Module >	read_llvm_module_from_ir_file (const std::string &udf_ir_filename, llvm::LLVMContext &ctx, bool is_gpu=false)
std::unique_ptr< llvm::Module >	read_llvm_module_from_ir_string (const std::string &udf_ir_string, llvm::LLVMContext &ctx, bool is_gpu=false)
void	anonymous_namespace{NativeCodegen.cpp}::bind_pos_placeholders (const std::string &pos_fn_name, const bool use_resume_param, llvm::Function *query_func, llvm::Module *llvm_module)
void	anonymous_namespace{NativeCodegen.cpp}::set_row_func_argnames (llvm::Function *row_func, const size_t in_col_count, const size_t agg_col_count, const bool hoist_literals)
llvm::Function *	anonymous_namespace{NativeCodegen.cpp}::create_row_function (const size_t in_col_count, const size_t agg_col_count, const bool hoist_literals, llvm::Module *llvm_module, llvm::LLVMContext &context)
void	anonymous_namespace{NativeCodegen.cpp}::bind_query (llvm::Function *query_func, const std::string &query_fname, llvm::Function *multifrag_query_func, llvm::Module *llvm_module)
std::vector< std::string >	anonymous_namespace{NativeCodegen.cpp}::get_agg_fnames (const std::vector< Analyzer::Expr * > &target_exprs, const bool is_group_by)
template<typename InstType >
llvm::Value *	anonymous_namespace{NativeCodegen.cpp}::find_variable_in_basic_block (llvm::Function *func, std::string bb_name, std::string variable_name)
size_t	anonymous_namespace{NativeCodegen.cpp}::get_shared_memory_size (const bool shared_mem_used, const QueryMemoryDescriptor *query_mem_desc_ptr)
bool	anonymous_namespace{NativeCodegen.cpp}::has_count_expr (RelAlgExecutionUnit const &ra_exe_unit)
bool	anonymous_namespace{NativeCodegen.cpp}::has_case_expr_within_groupby_expr (RelAlgExecutionUnit const &ra_exe_unit)
bool	anonymous_namespace{NativeCodegen.cpp}::is_gpu_shared_mem_supported (const QueryMemoryDescriptor *query_mem_desc_ptr, const RelAlgExecutionUnit &ra_exe_unit, const CudaMgr_Namespace::CudaMgr *cuda_mgr, const ExecutorDeviceType device_type, const unsigned cuda_blocksize, const unsigned num_blocks_per_mp)
std::string	anonymous_namespace{NativeCodegen.cpp}::serialize_llvm_metadata_footnotes (llvm::Function *query_func, CgenState *cgen_state)
std::vector< llvm::Value * >	generate_column_heads_load (const int num_columns, llvm::Value *byte_stream_arg, llvm::IRBuilder<> &ir_builder, llvm::LLVMContext &ctx)

Variables
float	g_fraction_code_cache_to_evict = 0.2
const std::string	anonymous_namespace{NativeCodegen.cpp}::cuda_rt_decls
constexpr std::array < std::string_view, 18 >	TARGET_RUNTIME_FUNCTIONS_FOR_MODULE_CLONING

Macro Definition Documentation

#define SHOW_DEFINED

(

MODULE

)

Value:

{                                                                  \
    std::cout << __func__ << "#" << __LINE__ << ": " #MODULE << " "; \
    ::show_defined(MODULE);                                          \
  }

Definition at line 141 of file NativeCodegen.cpp.

#define SHOW_FUNCTIONS

(

MODULE

)

Value:

{                                                                  \
    std::cout << __func__ << "#" << __LINE__ << ": " #MODULE << " "; \
    ::show_functions(MODULE);                                        \
  }

Definition at line 147 of file NativeCodegen.cpp.

Function Documentation

std::vector<llvm::Value*> generate_column_heads_load	(	const int	num_columns,
		llvm::Value *	byte_stream_arg,
		llvm::IRBuilder<> &	ir_builder,
		llvm::LLVMContext &	ctx
	)

Loads individual columns from a single, packed pointers buffer (the byte stream arg)

Definition at line 3454 of file NativeCodegen.cpp.

Referenced by TableFunctionCompilationContext::generateEntryPoint().

                                                                           {
  CHECK(byte_stream_arg);
  const auto max_col_local_id = num_columns - 1;

  std::vector<llvm::Value*> col_heads;
  for (int col_id = 0; col_id <= max_col_local_id; ++col_id) {
    auto* gep = ir_builder.CreateGEP(
        byte_stream_arg->getType()->getScalarType()->getPointerElementType(),
        byte_stream_arg,
        llvm::ConstantInt::get(llvm::Type::getInt32Ty(ctx), col_id));
    auto* load_gep = ir_builder.CreateLoad(gep->getType()->getPointerElementType(), gep);
    load_gep->setName(byte_stream_arg->getName() + "_" + std::to_string(col_id) + "_ptr");
    col_heads.emplace_back(load_gep);
  }
  return col_heads;
}

Here is the caller graph for this function:

std::map<std::string, std::string> get_device_parameters

(

bool

cpu_only

)

Definition at line 968 of file NativeCodegen.cpp.

                                                                  {
  std::map<std::string, std::string> result;

  result.insert(std::make_pair("cpu_name", llvm::sys::getHostCPUName()));
  result.insert(std::make_pair("cpu_triple", llvm::sys::getProcessTriple()));
  result.insert(
      std::make_pair("cpu_cores", std::to_string(llvm::sys::getHostNumPhysicalCores())));
  result.insert(std::make_pair("cpu_threads", std::to_string(cpu_threads())));

  // https://en.cppreference.com/w/cpp/language/types
  std::string sizeof_types;
  sizeof_types += "bool:" + std::to_string(sizeof(bool)) + ";";
  sizeof_types += "size_t:" + std::to_string(sizeof(size_t)) + ";";
  sizeof_types += "ssize_t:" + std::to_string(sizeof(ssize_t)) + ";";
  sizeof_types += "char:" + std::to_string(sizeof(char)) + ";";
  sizeof_types += "uchar:" + std::to_string(sizeof(unsigned char)) + ";";
  sizeof_types += "short:" + std::to_string(sizeof(short)) + ";";
  sizeof_types += "ushort:" + std::to_string(sizeof(unsigned short int)) + ";";
  sizeof_types += "int:" + std::to_string(sizeof(int)) + ";";
  sizeof_types += "uint:" + std::to_string(sizeof(unsigned int)) + ";";
  sizeof_types += "long:" + std::to_string(sizeof(long int)) + ";";
  sizeof_types += "ulong:" + std::to_string(sizeof(unsigned long int)) + ";";
  sizeof_types += "longlong:" + std::to_string(sizeof(long long int)) + ";";
  sizeof_types += "ulonglong:" + std::to_string(sizeof(unsigned long long int)) + ";";
  sizeof_types += "float:" + std::to_string(sizeof(float)) + ";";
  sizeof_types += "double:" + std::to_string(sizeof(double)) + ";";
  sizeof_types += "longdouble:" + std::to_string(sizeof(long double)) + ";";
  sizeof_types += "voidptr:" + std::to_string(sizeof(void*)) + ";";

  result.insert(std::make_pair("type_sizeof", sizeof_types));

  std::string null_values;
  null_values += "boolean1:" + std::to_string(serialized_null_value<bool>()) + ";";
  null_values += "boolean8:" + std::to_string(serialized_null_value<int8_t>()) + ";";
  null_values += "int8:" + std::to_string(serialized_null_value<int8_t>()) + ";";
  null_values += "int16:" + std::to_string(serialized_null_value<int16_t>()) + ";";
  null_values += "int32:" + std::to_string(serialized_null_value<int32_t>()) + ";";
  null_values += "int64:" + std::to_string(serialized_null_value<int64_t>()) + ";";
  null_values += "uint8:" + std::to_string(serialized_null_value<uint8_t>()) + ";";
  null_values += "uint16:" + std::to_string(serialized_null_value<uint16_t>()) + ";";
  null_values += "uint32:" + std::to_string(serialized_null_value<uint32_t>()) + ";";
  null_values += "uint64:" + std::to_string(serialized_null_value<uint64_t>()) + ";";
  null_values += "float32:" + std::to_string(serialized_null_value<float>()) + ";";
  null_values += "float64:" + std::to_string(serialized_null_value<double>()) + ";";
  null_values +=
      "Array<boolean8>:" + std::to_string(serialized_null_value<int8_t, true>()) + ";";
  null_values +=
      "Array<int8>:" + std::to_string(serialized_null_value<int8_t, true>()) + ";";
  null_values +=
      "Array<int16>:" + std::to_string(serialized_null_value<int16_t, true>()) + ";";
  null_values +=
      "Array<int32>:" + std::to_string(serialized_null_value<int32_t, true>()) + ";";
  null_values +=
      "Array<int64>:" + std::to_string(serialized_null_value<int64_t, true>()) + ";";
  null_values +=
      "Array<float32>:" + std::to_string(serialized_null_value<float, true>()) + ";";
  null_values +=
      "Array<float64>:" + std::to_string(serialized_null_value<double, true>()) + ";";

  result.insert(std::make_pair("null_values", null_values));

  llvm::StringMap<bool> cpu_features;
  if (llvm::sys::getHostCPUFeatures(cpu_features)) {
    std::string features_str = "";
    for (auto it = cpu_features.begin(); it != cpu_features.end(); ++it) {
      features_str += (it->getValue() ? " +" : " -");
      features_str += it->getKey().str();
    }
    result.insert(std::make_pair("cpu_features", features_str));
  }

  result.insert(std::make_pair("llvm_version",
                               std::to_string(LLVM_VERSION_MAJOR) + "." +
                                   std::to_string(LLVM_VERSION_MINOR) + "." +
                                   std::to_string(LLVM_VERSION_PATCH)));

#ifdef HAVE_CUDA
  if (!cpu_only) {
    int device_count = 0;
    checkCudaErrors(cuDeviceGetCount(&device_count));
    if (device_count) {
      CUdevice device{};
      char device_name[256];
      int major = 0, minor = 0;
      int driver_version;
      checkCudaErrors(cuDeviceGet(&device, 0));  // assuming homogeneous multi-GPU system
      checkCudaErrors(cuDeviceGetName(device_name, 256, device));
      checkCudaErrors(cuDeviceGetAttribute(
          &major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device));
      checkCudaErrors(cuDeviceGetAttribute(
          &minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device));
      checkCudaErrors(cuDriverGetVersion(&driver_version));

      result.insert(std::make_pair("gpu_name", device_name));
      result.insert(std::make_pair("gpu_count", std::to_string(device_count)));
      result.insert(std::make_pair("gpu_compute_capability",
                                   std::to_string(major) + "." + std::to_string(minor)));
      result.insert(std::make_pair("gpu_triple", get_gpu_target_triple_string()));
      result.insert(std::make_pair("gpu_datalayout", get_gpu_data_layout()));
      result.insert(std::make_pair("gpu_driver",
                                   "CUDA " + std::to_string(driver_version / 1000) + "." +
                                       std::to_string((driver_version % 1000) / 10)));

      auto rt_libdevice_path = get_cuda_libdevice_dir() + "/libdevice.10.bc";
      result.insert(
          std::make_pair("gpu_has_libdevice",
                         std::to_string(boost::filesystem::exists(rt_libdevice_path))));
    }
  }
#endif

  return result;
}

llvm::StringRef get_gpu_data_layout

(

)

Definition at line 960 of file NativeCodegen.cpp.

                                    {
  return llvm::StringRef(
      "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");
}

llvm::StringRef get_gpu_target_triple_string

(

)

Definition at line 956 of file NativeCodegen.cpp.

References DEBUG_TIMER.

                                             {
  return llvm::StringRef("nvptx64-nvidia-cuda");
}

std::unique_ptr<llvm::Module> read_llvm_module_from_bc_file	(	const std::string &	bc_filename,
		llvm::LLVMContext &	context
	)

Definition at line 1587 of file NativeCodegen.cpp.

Referenced by Executor::update_extension_modules().

                              {
  llvm::SMDiagnostic err;

  auto buffer_or_error = llvm::MemoryBuffer::getFile(bc_filename);
  CHECK(!buffer_or_error.getError()) << "bc_filename=" << bc_filename;
  llvm::MemoryBuffer* buffer = buffer_or_error.get().get();

  auto owner = llvm::parseBitcodeFile(buffer->getMemBufferRef(), context);
  CHECK(!owner.takeError());
  CHECK(owner->get());
  return std::move(owner.get());
}

Here is the caller graph for this function:

std::unique_ptr<llvm::Module> read_llvm_module_from_ir_file	(	const std::string &	udf_ir_filename,
		llvm::LLVMContext &	ctx,
		bool	is_gpu = false
	)

Definition at line 1602 of file NativeCodegen.cpp.

Referenced by Executor::update_extension_modules().

                         {
  llvm::SMDiagnostic parse_error;

  llvm::StringRef file_name_arg(udf_ir_filename);

  auto owner = llvm::parseIRFile(file_name_arg, parse_error, ctx);
  if (!owner) {
    throw_parseIR_error(parse_error, udf_ir_filename, is_gpu);
  }

  if (is_gpu) {
    llvm::Triple gpu_triple(owner->getTargetTriple());
    if (!gpu_triple.isNVPTX()) {
      LOG(WARNING)
          << "Expected triple nvptx64-nvidia-cuda for NVVM IR of loadtime UDFs but got "
          << gpu_triple.str() << ". Disabling the NVVM IR module.";
      return std::unique_ptr<llvm::Module>();
    }
  }
  return owner;
}

Here is the caller graph for this function:

std::unique_ptr<llvm::Module> read_llvm_module_from_ir_string	(	const std::string &	udf_ir_string,
		llvm::LLVMContext &	ctx,
		bool	is_gpu = false
	)

Definition at line 1627 of file NativeCodegen.cpp.

Referenced by Executor::update_extension_modules().

                         {
  llvm::SMDiagnostic parse_error;

  auto buf = std::make_unique<llvm::MemoryBufferRef>(udf_ir_string,
                                                     "Runtime UDF/UDTF LLVM/NVVM IR");

  auto owner = llvm::parseIR(*buf, parse_error, ctx);
  if (!owner) {
    LOG(IR) << "read_llvm_module_from_ir_string:\n"
            << udf_ir_string << "\nEnd of LLVM/NVVM IR";
    throw_parseIR_error(parse_error, "", /* is_gpu= */ is_gpu);
  }

  if (is_gpu) {
    llvm::Triple gpu_triple(owner->getTargetTriple());
    if (!gpu_triple.isNVPTX()) {
      LOG(IR) << "read_llvm_module_from_ir_string:\n"
              << udf_ir_string << "\nEnd of NNVM IR";
      LOG(WARNING) << "Expected triple nvptx64-nvidia-cuda for NVVM IR but got "
                   << gpu_triple.str()
                   << ". Executing runtime UDF/UDTFs on GPU will be disabled.";
      return std::unique_ptr<llvm::Module>();
      ;
    }
  }
  return owner;
}

Here is the caller graph for this function:

void verify_function_ir

(

const llvm::Function *

func

)

Definition at line 386 of file NativeCodegen.cpp.

References logger::FATAL, and LOG.

Referenced by GpuSharedMemCodeBuilder::codegen(), spatial_type::Transform::codegen(), anonymous_namespace{JoinLoopTest.cpp}::create_loop_test_function(), TableFunctionCompilationContext::generateEntryPoint(), StubGenerator::generateStub(), and translate_function().

                                                  {
  std::stringstream err_ss;
  llvm::raw_os_ostream err_os(err_ss);
  err_os << "\n-----\n";
  if (llvm::verifyFunction(*func, &err_os)) {
    err_os << "\n-----\n";
    func->print(err_os, nullptr);
    err_os << "\n-----\n";
    LOG(FATAL) << err_ss.str();
  }
}

Here is the caller graph for this function:

Variable Documentation

float g_fraction_code_cache_to_evict = 0.2

Definition at line 81 of file NativeCodegen.cpp.

Referenced by CommandLineOptions::fillDeveloperOptions().

constexpr std::array<std::string_view, 18> TARGET_RUNTIME_FUNCTIONS_FOR_MODULE_CLONING

Initial value:

{"query_stub_hoisted_literals",
     "multifrag_query_hoisted_literals",
     "query_stub",
     "multifrag_query",
     "fixed_width_int_decode",
     "fixed_width_unsigned_decode",
     "diff_fixed_width_int_decode",
     "fixed_width_double_decode",
     "fixed_width_float_decode",
     "fixed_width_small_date_decode",
     "record_error_code",
     "get_error_code",
     "pos_start_impl",
     "pos_step_impl",
     "group_buff_idx_impl",
     "init_shared_mem",
     "init_shared_mem_nop",
     "write_back_nop"}}

Definition at line 1560 of file NativeCodegen.cpp.