blob: 44f23c5abb07a7c985cc1b204b7ac4b28104349a [file] [log] [blame]
/*
* Copyright 2010-2012, The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "Assert.h"
#include "Log.h"
#include "RSTransforms.h"
#include "RSUtils.h"
#include "rsDefines.h"
#include "bcc/Compiler.h"
#include "bcc/CompilerConfig.h"
#include "bcc/Config.h"
#include "bcc/Script.h"
#include "bcc/Source.h"
#include "bcinfo/MetadataExtractor.h"
#include <llvm/Analysis/Passes.h>
#include <llvm/Analysis/TargetTransformInfo.h>
#include <llvm/CodeGen/RegAllocRegistry.h>
#include <llvm/IR/LegacyPassManager.h>
#include <llvm/IR/Module.h>
#include <llvm/Support/TargetRegistry.h>
#include <llvm/Support/raw_ostream.h>
#include <llvm/IR/DataLayout.h>
#include <llvm/Target/TargetSubtargetInfo.h>
#include <llvm/Target/TargetMachine.h>
#include <llvm/Transforms/IPO.h>
#include <llvm/Transforms/IPO/PassManagerBuilder.h>
#include <llvm/Transforms/Scalar.h>
#include <llvm/Transforms/Vectorize.h>
#include <string>
#include <set>
namespace {
// Name of metadata node where list of exported types resides
// (should be synced with slang_rs_metadata.h)
static const llvm::StringRef ExportedTypeMetadataName = "#rs_export_type";
// Every exported struct type must have the same layout according to
// the Module's DataLayout that it does according to the
// TargetMachine's DataLayout -- that is, the front end (represented
// by Module) and back end (represented by TargetMachine) must agree.
bool validateLayoutOfExportedTypes(const llvm::Module &module,
const llvm::DataLayout &moduleDataLayout,
const llvm::DataLayout &targetDataLayout) {
if (moduleDataLayout == targetDataLayout)
return true;
const llvm::NamedMDNode *const exportedTypesMD =
module.getNamedMetadata(ExportedTypeMetadataName);
if (!exportedTypesMD)
return true;
bool allOk = true;
for (const llvm::MDNode *const exportedTypeMD : exportedTypesMD->operands()) {
bccAssert(exportedTypeMD->getNumOperands() == 1);
// The name of the type in LLVM is the name of the type in the
// metadata with "struct." prepended.
std::string exportedTypeName =
"struct." +
llvm::cast<llvm::MDString>(exportedTypeMD->getOperand(0))->getString().str();
llvm::StructType *const exportedType = module.getTypeByName(exportedTypeName);
if (!exportedType) {
// presumably this means the type got optimized away
continue;
}
const llvm::StructLayout *const moduleStructLayout = moduleDataLayout.getStructLayout(exportedType);
const llvm::StructLayout *const targetStructLayout = targetDataLayout.getStructLayout(exportedType);
if (moduleStructLayout->getSizeInBits() != targetStructLayout->getSizeInBits()) {
ALOGE("%s: getSizeInBits() does not match (%u, %u)", exportedTypeName.c_str(),
unsigned(moduleStructLayout->getSizeInBits()), unsigned(targetStructLayout->getSizeInBits()));
allOk = false;
}
// We deliberately do not check alignment of the struct as a whole -- the explicit padding
// from slang doesn't force the alignment.
for (unsigned elementCount = exportedType->getNumElements(), elementIdx = 0;
elementIdx < elementCount; ++elementIdx) {
if (moduleStructLayout->getElementOffsetInBits(elementIdx) !=
targetStructLayout->getElementOffsetInBits(elementIdx)) {
ALOGE("%s: getElementOffsetInBits(%u) does not match (%u, %u)",
exportedTypeName.c_str(), elementIdx,
unsigned(moduleStructLayout->getElementOffsetInBits(elementIdx)),
unsigned(targetStructLayout->getElementOffsetInBits(elementIdx)));
allOk = false;
}
}
}
return allOk;
}
} // end unnamed namespace
using namespace bcc;
const char *Compiler::GetErrorString(enum ErrorCode pErrCode) {
switch (pErrCode) {
case kSuccess:
return "Successfully compiled.";
case kInvalidConfigNoTarget:
return "Invalid compiler config supplied (getTarget() returns nullptr.) "
"(missing call to CompilerConfig::initialize()?)";
case kErrCreateTargetMachine:
return "Failed to create llvm::TargetMachine.";
case kErrSwitchTargetMachine:
return "Failed to switch llvm::TargetMachine.";
case kErrNoTargetMachine:
return "Failed to compile the script since there's no available "
"TargetMachine. (missing call to Compiler::config()?)";
case kErrMaterialization:
return "Failed to materialize the module.";
case kErrInvalidOutputFileState:
return "Supplied output file was invalid (in the error state.)";
case kErrPrepareOutput:
return "Failed to prepare file for output.";
case kPrepareCodeGenPass:
return "Failed to construct pass list for code-generation.";
case kErrCustomPasses:
return "Error occurred while adding custom passes.";
case kErrInvalidSource:
return "Error loading input bitcode";
case kIllegalGlobalFunction:
return "Use of undefined external function";
case kErrInvalidTargetMachine:
return "Invalid/unexpected llvm::TargetMachine.";
case kErrInvalidLayout:
return "Invalid layout (RenderScript ABI and native ABI are incompatible)";
}
// This assert should never be reached as the compiler verifies that the
// above switch coveres all enum values.
bccAssert(false && "Unknown error code encountered");
return "";
}
//===----------------------------------------------------------------------===//
// Instance Methods
//===----------------------------------------------------------------------===//
Compiler::Compiler() : mTarget(nullptr), mEnableOpt(true) {
return;
}
Compiler::Compiler(const CompilerConfig &pConfig) : mTarget(nullptr),
mEnableOpt(true) {
const std::string &triple = pConfig.getTriple();
enum ErrorCode err = config(pConfig);
if (err != kSuccess) {
ALOGE("%s (%s, features: %s)", GetErrorString(err),
triple.c_str(), pConfig.getFeatureString().c_str());
return;
}
return;
}
enum Compiler::ErrorCode Compiler::config(const CompilerConfig &pConfig) {
if (pConfig.getTarget() == nullptr) {
return kInvalidConfigNoTarget;
}
llvm::TargetMachine *new_target =
(pConfig.getTarget())->createTargetMachine(pConfig.getTriple(),
pConfig.getCPU(),
pConfig.getFeatureString(),
pConfig.getTargetOptions(),
pConfig.getRelocationModel(),
pConfig.getCodeModel(),
pConfig.getOptimizationLevel());
if (new_target == nullptr) {
return ((mTarget != nullptr) ? kErrSwitchTargetMachine :
kErrCreateTargetMachine);
}
// Replace the old TargetMachine.
delete mTarget;
mTarget = new_target;
// Adjust register allocation policy according to the optimization level.
// createFastRegisterAllocator: fast but bad quality
// createLinearScanRegisterAllocator: not so fast but good quality
if ((pConfig.getOptimizationLevel() == llvm::CodeGenOpt::None)) {
llvm::RegisterRegAlloc::setDefault(llvm::createFastRegisterAllocator);
} else {
llvm::RegisterRegAlloc::setDefault(llvm::createGreedyRegisterAllocator);
}
return kSuccess;
}
Compiler::~Compiler() {
delete mTarget;
}
// This function has complete responsibility for creating and executing the
// exact list of compiler passes.
enum Compiler::ErrorCode Compiler::runPasses(Script &script,
llvm::raw_pwrite_stream &pResult) {
// Pass manager for link-time optimization
llvm::legacy::PassManager transformPasses;
// Empty MCContext.
llvm::MCContext *mc_context = nullptr;
transformPasses.add(
createTargetTransformInfoWrapperPass(mTarget->getTargetIRAnalysis()));
// Add some initial custom passes.
addInvokeHelperPass(transformPasses);
addExpandKernelPass(transformPasses);
addDebugInfoPass(script, transformPasses);
addInvariantPass(transformPasses);
if (mTarget->getOptLevel() != llvm::CodeGenOpt::None) {
if (!addInternalizeSymbolsPass(script, transformPasses))
return kErrCustomPasses;
}
addGlobalInfoPass(script, transformPasses);
if (mTarget->getOptLevel() == llvm::CodeGenOpt::None) {
transformPasses.add(llvm::createGlobalOptimizerPass());
transformPasses.add(llvm::createConstantMergePass());
} else {
// FIXME: Figure out which passes should be executed.
llvm::PassManagerBuilder Builder;
Builder.Inliner = llvm::createFunctionInliningPass();
Builder.populateLTOPassManager(transformPasses);
/* FIXME: Reenable autovectorization after rebase.
bug 19324423
// Add vectorization passes after LTO passes are in
// additional flag: -unroll-runtime
transformPasses.add(llvm::createLoopUnrollPass(-1, 16, 0, 1));
// Need to pass appropriate flags here: -scalarize-load-store
transformPasses.add(llvm::createScalarizerPass());
transformPasses.add(llvm::createCFGSimplificationPass());
transformPasses.add(llvm::createScopedNoAliasAAPass());
transformPasses.add(llvm::createScalarEvolutionAliasAnalysisPass());
// additional flags: -slp-vectorize-hor -slp-vectorize-hor-store (unnecessary?)
transformPasses.add(llvm::createSLPVectorizerPass());
transformPasses.add(llvm::createDeadCodeEliminationPass());
transformPasses.add(llvm::createInstructionCombiningPass());
*/
}
// These passes have to come after LTO, since we don't want to examine
// functions that are never actually called.
if (llvm::Triple(getTargetMachine().getTargetTriple()).getArch() == llvm::Triple::x86_64 ||
llvm::Triple(getTargetMachine().getTargetTriple()).getArch() == llvm::Triple::mips64el)
transformPasses.add(createRSX86_64CallConvPass()); // Add pass to correct calling convention for X86-64 and mips64.
transformPasses.add(createRSIsThreadablePass()); // Add pass to mark script as threadable.
// RSEmbedInfoPass needs to come after we have scanned for non-threadable
// functions.
if (script.getEmbedInfo())
transformPasses.add(createRSEmbedInfoPass());
// Execute the passes.
transformPasses.run(script.getSource().getModule());
// Run backend separately to avoid interference between debug metadata
// generation and backend initialization.
llvm::legacy::PassManager codeGenPasses;
// Add passes to the pass manager to emit machine code through MC layer.
if (mTarget->addPassesToEmitMC(codeGenPasses, mc_context, pResult,
/* DisableVerify */false)) {
return kPrepareCodeGenPass;
}
// Execute the passes.
codeGenPasses.run(script.getSource().getModule());
return kSuccess;
}
enum Compiler::ErrorCode Compiler::compile(Script &script,
llvm::raw_pwrite_stream &pResult,
llvm::raw_ostream *IRStream) {
llvm::Module &module = script.getSource().getModule();
enum ErrorCode err;
if (mTarget == nullptr) {
return kErrNoTargetMachine;
}
const std::string &triple = module.getTargetTriple();
const llvm::DataLayout dl = getTargetMachine().createDataLayout();
unsigned int pointerSize = dl.getPointerSizeInBits();
if (triple == "armv7-none-linux-gnueabi") {
if (pointerSize != 32) {
return kErrInvalidSource;
}
} else if (triple == "aarch64-none-linux-gnueabi") {
if (pointerSize != 64) {
return kErrInvalidSource;
}
} else {
return kErrInvalidSource;
}
if (script.isStructExplicitlyPaddedBySlang()) {
if (!validateLayoutOfExportedTypes(module, module.getDataLayout(), dl))
return kErrInvalidLayout;
} else {
if (getTargetMachine().getTargetTriple().getArch() == llvm::Triple::x86) {
// Detect and fail if TargetMachine datalayout is different than what we
// expect. This is to detect changes in default target layout for x86 and
// update X86_CUSTOM_DL_STRING in include/bcc/Config/Config.h appropriately.
if (dl.getStringRepresentation().compare(X86_DEFAULT_DL_STRING) != 0) {
return kErrInvalidTargetMachine;
}
}
}
// Sanitize module's target information.
module.setTargetTriple(getTargetMachine().getTargetTriple().str());
module.setDataLayout(getTargetMachine().createDataLayout());
// Materialize the bitcode module.
if (module.getMaterializer() != nullptr) {
// A module with non-null materializer means that it is a lazy-load module.
// Materialize it now. This function returns false when the materialization
// is successful.
std::error_code ec = module.materializeAll();
if (ec) {
ALOGE("Failed to materialize the module `%s'! (%s)",
module.getModuleIdentifier().c_str(), ec.message().c_str());
return kErrMaterialization;
}
}
if ((err = runPasses(script, pResult)) != kSuccess) {
return err;
}
if (IRStream) {
*IRStream << module;
}
return kSuccess;
}
bool Compiler::addInternalizeSymbolsPass(Script &script, llvm::legacy::PassManager &pPM) {
// Add a pass to internalize the symbols that don't need to have global
// visibility.
llvm::Module &module = script.getSource().getModule();
bcinfo::MetadataExtractor me(&module);
if (!me.extract()) {
bccAssert(false && "Could not extract metadata for module!");
return false;
}
// Set of symbols that should not be internalized.
std::set<std::string> export_symbols;
const char *sf[] = {
kRoot, // Graphics drawing function or compute kernel.
kInit, // Initialization routine called implicitly on startup.
kRsDtor, // Static global destructor for a script instance.
kRsInfo, // Variable containing string of RS metadata info.
kRsGlobalEntries, // Optional number of global variables.
kRsGlobalNames, // Optional global variable name info.
kRsGlobalAddresses, // Optional global variable address info.
kRsGlobalSizes, // Optional global variable size info.
kRsGlobalProperties, // Optional global variable properties.
nullptr // Must be nullptr-terminated.
};
const char **special_functions = sf;
// Special RS functions should always be global symbols.
while (*special_functions != nullptr) {
export_symbols.insert(*special_functions);
special_functions++;
}
// Visibility of symbols appeared in rs_export_var and rs_export_func should
// also be preserved.
size_t exportVarCount = me.getExportVarCount();
size_t exportFuncCount = me.getExportFuncCount();
size_t exportForEachCount = me.getExportForEachSignatureCount();
size_t exportReduceCount = me.getExportReduceCount();
const char **exportVarNameList = me.getExportVarNameList();
const char **exportFuncNameList = me.getExportFuncNameList();
const char **exportForEachNameList = me.getExportForEachNameList();
const bcinfo::MetadataExtractor::Reduce *exportReduceList = me.getExportReduceList();
size_t i;
for (i = 0; i < exportVarCount; ++i) {
export_symbols.insert(exportVarNameList[i]);
}
for (i = 0; i < exportFuncCount; ++i) {
export_symbols.insert(exportFuncNameList[i]);
}
// Expanded foreach functions should not be internalized; nor should
// general reduction initializer, combiner, and outconverter
// functions. keep_funcs keeps the names of these functions around
// until createInternalizePass() is finished making its own copy of
// the visible symbols.
std::vector<std::string> keep_funcs;
keep_funcs.reserve(exportForEachCount + exportReduceCount*4);
for (i = 0; i < exportForEachCount; ++i) {
keep_funcs.push_back(std::string(exportForEachNameList[i]) + ".expand");
}
auto keepFuncsPushBackIfPresent = [&keep_funcs](const char *Name) {
if (Name) keep_funcs.push_back(Name);
};
for (i = 0; i < exportReduceCount; ++i) {
keep_funcs.push_back(std::string(exportReduceList[i].mAccumulatorName) + ".expand");
keepFuncsPushBackIfPresent(exportReduceList[i].mInitializerName);
if (exportReduceList[i].mCombinerName != nullptr) {
keep_funcs.push_back(exportReduceList[i].mCombinerName);
} else {
keep_funcs.push_back(nameReduceCombinerFromAccumulator(exportReduceList[i].mAccumulatorName));
}
keepFuncsPushBackIfPresent(exportReduceList[i].mOutConverterName);
}
for (auto &symbol_name : keep_funcs) {
export_symbols.insert(symbol_name);
}
auto IsExportedSymbol = [=](const llvm::GlobalValue &GV) {
return export_symbols.count(GV.getName()) > 0;
};
pPM.add(llvm::createInternalizePass(IsExportedSymbol));
return true;
}
void Compiler::addInvokeHelperPass(llvm::legacy::PassManager &pPM) {
llvm::Triple arch(getTargetMachine().getTargetTriple());
if (arch.isArch64Bit()) {
pPM.add(createRSInvokeHelperPass());
}
}
void Compiler::addDebugInfoPass(Script &script, llvm::legacy::PassManager &pPM) {
if (script.getSource().getDebugInfoEnabled())
pPM.add(createRSAddDebugInfoPass());
}
void Compiler::addExpandKernelPass(llvm::legacy::PassManager &pPM) {
// Expand ForEach and reduce on CPU path to reduce launch overhead.
bool pEnableStepOpt = true;
pPM.add(createRSKernelExpandPass(pEnableStepOpt));
}
void Compiler::addGlobalInfoPass(Script &script, llvm::legacy::PassManager &pPM) {
// Add additional information about RS global variables inside the Module.
if (script.getEmbedGlobalInfo()) {
pPM.add(createRSGlobalInfoPass(script.getEmbedGlobalInfoSkipConstant()));
}
}
void Compiler::addInvariantPass(llvm::legacy::PassManager &pPM) {
// Mark Loads from RsExpandKernelDriverInfo as "load.invariant".
// Should run after ExpandForEach and before inlining.
pPM.add(createRSInvariantPass());
}
enum Compiler::ErrorCode Compiler::screenGlobalFunctions(Script &script) {
llvm::Module &module = script.getSource().getModule();
// Materialize the bitcode module in case this is a lazy-load module. Do not
// clear the materializer by calling materializeAllPermanently since the
// runtime library has not been merged into the module yet.
if (module.getMaterializer() != nullptr) {
std::error_code ec = module.materializeAll();
if (ec) {
ALOGE("Failed to materialize module `%s' when screening globals! (%s)",
module.getModuleIdentifier().c_str(), ec.message().c_str());
return kErrMaterialization;
}
}
// Add pass to check for illegal function calls.
llvm::legacy::PassManager pPM;
pPM.add(createRSScreenFunctionsPass());
pPM.run(module);
return kSuccess;
}
void Compiler::translateGEPs(Script &script) {
llvm::legacy::PassManager pPM;
pPM.add(createRSX86TranslateGEPPass());
// Materialization done in screenGlobalFunctions above.
pPM.run(script.getSource().getModule());
}