/// This routine adds optimization passes based on selected optimization level,
/// OptLevel.
///
/// OptLevel - Optimization Level
static void AddOptimizationPasses(PassManagerBase &MPM,FunctionPassManager &FPM,
unsigned OptLevel, unsigned SizeLevel) {
FPM.add(createVerifierPass()); // Verify that input is correct
MPM.add(createDebugInfoVerifierPass()); // Verify that debug info is correct
PassManagerBuilder Builder;
Builder.OptLevel = OptLevel;
Builder.SizeLevel = SizeLevel;
if (DisableInline) {
// No inlining pass
} else if (OptLevel > 1) {
Builder.Inliner = createFunctionInliningPass(OptLevel, SizeLevel);
} else {
Builder.Inliner = createAlwaysInlinerPass();
}
Builder.DisableUnitAtATime = !UnitAtATime;
Builder.DisableUnrollLoops = (DisableLoopUnrolling.getNumOccurrences() > 0) ?
DisableLoopUnrolling : OptLevel == 0;
// This is final, unless there is a #pragma vectorize enable
if (DisableLoopVectorization)
Builder.LoopVectorize = false;
// If option wasn't forced via cmd line (-vectorize-loops, -loop-vectorize)
else if (!Builder.LoopVectorize)
Builder.LoopVectorize = OptLevel > 1 && SizeLevel < 2;
// When #pragma vectorize is on for SLP, do the same as above
Builder.SLPVectorize =
DisableSLPVectorization ? false : OptLevel > 1 && SizeLevel < 2;
Builder.populateFunctionPassManager(FPM);
Builder.populateModulePassManager(MPM);
}
/// AddOptimizationPasses - This routine adds optimization passes
/// based on selected optimization level, OptLevel. This routine
/// duplicates llvm-gcc behaviour.
///
/// OptLevel - Optimization Level
static void AddOptimizationPasses(PassManagerBase &MPM,FunctionPassManager &FPM,
unsigned OptLevel, unsigned SizeLevel) {
FPM.add(createVerifierPass()); // Verify that input is correct
PassManagerBuilder Builder;
Builder.OptLevel = OptLevel;
Builder.SizeLevel = SizeLevel;
if (DisableInline) {
// No inlining pass
} else if (OptLevel > 1) {
unsigned Threshold = 225;
if (SizeLevel == 1) // -Os
Threshold = 75;
else if (SizeLevel == 2) // -Oz
Threshold = 25;
if (OptLevel > 2)
Threshold = 275;
Builder.Inliner = createFunctionInliningPass(Threshold);
} else {
Builder.Inliner = createAlwaysInlinerPass();
}
Builder.DisableUnitAtATime = !UnitAtATime;
Builder.DisableUnrollLoops = OptLevel == 0;
Builder.DisableSimplifyLibCalls = DisableSimplifyLibCalls;
Builder.populateFunctionPassManager(FPM);
Builder.populateModulePassManager(MPM);
}
bool LLVMOptimizeModule(LLVMModuleRef Mod) {
Module* M = unwrap(Mod);
// Create a PassManager to hold and optimize the collection of passes we are
// about to build.
PassManager Passes;
// Add an appropriate TargetLibraryInfo pass for the module's triple.
TargetLibraryInfo *TLI = new TargetLibraryInfo(Triple(M->getTargetTriple()));
// Add an appropriate DataLayout instance for this module.
// const std::string &ModuleDataLayout = M->getDataLayout();
// if (!ModuleDataLayout.empty()) {
// DataLayout *TD = NULL; // new DataLayout(ModuleDataLayout);
// Passes.add(TD);
// }
Passes.add(createVerifierPass()); // Verify that input is correct
// -std-compile-opts adds the same module passes as -O3.
PassManagerBuilder Builder;
Builder.Inliner = createFunctionInliningPass();
Builder.OptLevel = 3;
Builder.populateModulePassManager(Passes);
// Now that we have all of the passes ready, run them.
bool change = Passes.run(*M);
return change;
}
void
LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
LLVMPassManagerRef PM) {
PassManagerBuilder *Builder = unwrap(PMB);
legacy::PassManagerBase *MPM = unwrap(PM);
Builder->populateModulePassManager(*MPM);
}
Module * llvmutil_extractmodule(Module * OrigMod, TargetMachine * TM, std::vector<llvm::GlobalValue*> * livevalues, std::vector<std::string> * symbolnames, bool internalizeandoptimize) {
assert(symbolnames == NULL || livevalues->size() == symbolnames->size());
ValueToValueMapTy VMap;
#if LLVM_VERSION >= 34
Module * M = llvmutil_extractmodulewithproperties(OrigMod->getModuleIdentifier(), OrigMod, (llvm::GlobalValue **)&(*livevalues)[0], livevalues->size(), AlwaysCopy, NULL, VMap);
#else
Module * M = CloneModule(OrigMod, VMap);
internalizeandoptimize = true; //we need to do this regardless of the input because it is the only way we can extract just the needed functions from the module
#endif
//rename values to symbolsnames
std::vector<const char *> names;
for(size_t i = 0; i < livevalues->size(); i++) {
GlobalValue * fn = cast<GlobalValue>(VMap[(*livevalues)[i]]);
const std::string & name = (*symbolnames)[i];
GlobalValue * gv = M->getNamedValue(name);
if(gv) { //if there is already a symbol with this name, rename it
gv->setName(Twine((*symbolnames)[i],"_renamed"));
}
fn->setName(name); //and set our function to this name
assert(fn->getName() == name);
names.push_back(name.c_str()); //internalize pass has weird interface, so we need to copy the names here
}
if (!internalizeandoptimize)
return M;
//at this point we run optimizations on the module
//first internalize all functions not mentioned in "names" using an internalize pass and then perform
//standard optimizations
PassManager MPM;
llvmutil_addtargetspecificpasses(&MPM, TM);
MPM.add(createVerifierPass()); //make sure we haven't messed stuff up yet
MPM.add(createInternalizePass(names));
MPM.add(createGlobalDCEPass()); //run this early since anything not in the table of exported functions is still in this module
//this will remove dead functions
PassManagerBuilder PMB;
PMB.OptLevel = 3;
PMB.SizeLevel = 0;
#if LLVM_VERSION >= 35
PMB.LoopVectorize = true;
PMB.SLPVectorize = true;
#endif
PMB.populateModulePassManager(MPM);
MPM.run(*M);
return M;
}
void LLVMZigOptimizeModule(LLVMTargetMachineRef targ_machine_ref, LLVMModuleRef module_ref) {
TargetMachine* target_machine = reinterpret_cast<TargetMachine*>(targ_machine_ref);
Module* module = unwrap(module_ref);
TargetLibraryInfoImpl tlii(Triple(module->getTargetTriple()));
PassManagerBuilder *PMBuilder = new PassManagerBuilder();
PMBuilder->OptLevel = target_machine->getOptLevel();
PMBuilder->SizeLevel = 0;
PMBuilder->BBVectorize = true;
PMBuilder->SLPVectorize = true;
PMBuilder->LoopVectorize = true;
PMBuilder->DisableUnitAtATime = false;
PMBuilder->DisableUnrollLoops = false;
PMBuilder->MergeFunctions = true;
PMBuilder->PrepareForLTO = true;
PMBuilder->RerollLoops = true;
PMBuilder->addExtension(PassManagerBuilder::EP_EarlyAsPossible, addAddDiscriminatorsPass);
PMBuilder->LibraryInfo = &tlii;
PMBuilder->Inliner = createFunctionInliningPass(PMBuilder->OptLevel, PMBuilder->SizeLevel);
// Set up the per-function pass manager.
legacy::FunctionPassManager *FPM = new legacy::FunctionPassManager(module);
FPM->add(createTargetTransformInfoWrapperPass(target_machine->getTargetIRAnalysis()));
#ifndef NDEBUG
bool verify_module = true;
#else
bool verify_module = false;
#endif
if (verify_module) {
FPM->add(createVerifierPass());
}
PMBuilder->populateFunctionPassManager(*FPM);
// Set up the per-module pass manager.
legacy::PassManager *MPM = new legacy::PassManager();
MPM->add(createTargetTransformInfoWrapperPass(target_machine->getTargetIRAnalysis()));
PMBuilder->populateModulePassManager(*MPM);
// run per function optimization passes
FPM->doInitialization();
for (Function &F : *module)
if (!F.isDeclaration())
FPM->run(F);
FPM->doFinalization();
// run per module optimization passes
MPM->run(*module);
}
Function &OptimizeForRuntime(Function &F) {
#ifdef DEBUG
static PassManagerBuilder Builder = getDebugBuilder();
#else
static PassManagerBuilder Builder = getBuilder();
#endif
Module *M = F.getParent();
opt::GenerateOutput = true;
polly::opt::PollyParallel = true;
FunctionPassManager PM = FunctionPassManager(M);
Builder.populateFunctionPassManager(PM);
PM.doInitialization();
PM.run(F);
PM.doFinalization();
if (opt::havePapi()) {
PassManager MPM;
Builder.populateModulePassManager(MPM);
MPM.add(polli::createTraceMarkerPass());
MPM.run(*M);
}
if (opt::haveLikwid()) {
PassManager MPM;
Builder.populateModulePassManager(MPM);
MPM.add(polli::createLikwidMarkerPass());
MPM.run(*M);
}
DEBUG(
StoreModule(*M, M->getModuleIdentifier() + ".after.polly.ll")
);
opt::GenerateOutput = false;
return F;
}
Module * llvmutil_extractmodule(Module * OrigMod, TargetMachine * TM, std::vector<Function*> * livefns, std::vector<std::string> * symbolnames) {
assert(symbolnames == NULL || livefns->size() == symbolnames->size());
ValueToValueMapTy VMap;
Module * M = CloneModule(OrigMod, VMap);
PassManager * MPM = new PassManager();
llvmutil_addtargetspecificpasses(MPM, TM);
std::vector<const char *> names;
for(size_t i = 0; i < livefns->size(); i++) {
Function * fn = cast<Function>(VMap[(*livefns)[i]]);
const char * name;
if(symbolnames) {
GlobalAlias * ga = new GlobalAlias(fn->getType(), Function::ExternalLinkage, (*symbolnames)[i], fn, M);
name = copyName(ga->getName());
} else {
name = copyName(fn->getName());
}
names.push_back(name); //internalize pass has weird interface, so we need to copy the names here
}
//at this point we run optimizations on the module
//first internalize all functions not mentioned in "names" using an internalize pass and then perform
//standard optimizations
MPM->add(createVerifierPass()); //make sure we haven't messed stuff up yet
MPM->add(createInternalizePass(names));
MPM->add(createGlobalDCEPass()); //run this early since anything not in the table of exported functions is still in this module
//this will remove dead functions
//clean up the name list
for(size_t i = 0; i < names.size(); i++) {
free((char*)names[i]);
names[i] = NULL;
}
PassManagerBuilder PMB;
PMB.OptLevel = 3;
PMB.DisableUnrollLoops = true;
PMB.populateModulePassManager(*MPM);
//PMB.populateLTOPassManager(*MPM, false, false); //no need to re-internalize, we already did it
MPM->run(*M);
delete MPM;
MPM = NULL;
return M;
}
void llvmutil_addoptimizationpasses(PassManagerBase * fpm) {
PassManagerBuilder PMB;
PMB.OptLevel = 3;
PMB.SizeLevel = 0;
PMB.DisableUnitAtATime = true;
#if LLVM_VERSION <= 34 && LLVM_VERSION >= 32
PMB.LoopVectorize = false;
#elif LLVM_VERSION >= 35
PMB.LoopVectorize = true;
PMB.SLPVectorize = true;
#endif
PassManagerWrapper W(fpm);
PMB.populateModulePassManager(W);
}
static void AddStandardCompilePasses(PassManagerBase &PM) {
PM.add(createVerifierPass()); // Verify that input is correct
// If the -strip-debug command line option was specified, do it.
if (StripDebug)
addPass(PM, createStripSymbolsPass(true));
if (DisableOptimizations) return;
// -std-compile-opts adds the same module passes as -O3.
PassManagerBuilder Builder;
if (!DisableInline)
Builder.Inliner = createFunctionInliningPass();
Builder.OptLevel = 3;
Builder.populateModulePassManager(PM);
}
/**
* Adds a set of optimization passes to the given module/function pass
* managers based on the given optimization and size reduction levels.
*
* The selection mirrors Clang behavior and is based on LLVM's
* PassManagerBuilder.
*/
static void addOptimizationPasses(PassManagerBase &mpm, FunctionPassManager &fpm,
unsigned optLevel, unsigned sizeLevel) {
fpm.add(createVerifierPass()); // Verify that input is correct
PassManagerBuilder builder;
builder.OptLevel = optLevel;
builder.SizeLevel = sizeLevel;
if (willInline()) {
unsigned threshold = 225;
if (sizeLevel == 1) // -Os
threshold = 75;
else if (sizeLevel == 2) // -Oz
threshold = 25;
if (optLevel > 2)
threshold = 275;
builder.Inliner = createFunctionInliningPass(threshold);
} else {
builder.Inliner = createAlwaysInlinerPass();
}
builder.DisableSimplifyLibCalls = disableSimplifyLibCalls;
builder.DisableUnitAtATime = !unitAtATime;
builder.DisableUnrollLoops = optLevel == 0;
/* builder.Vectorize is set in ctor from command line switch */
if (!disableLangSpecificPasses) {
if (!disableSimplifyDruntimeCalls)
builder.addExtension(PassManagerBuilder::EP_LoopOptimizerEnd, addSimplifyDRuntimeCallsPass);
#if USE_METADATA
if (!disableGCToStack)
builder.addExtension(PassManagerBuilder::EP_LoopOptimizerEnd, addGarbageCollect2StackPass);
#endif // USE_METADATA
}
#if LDC_LLVM_VER >= 301
// EP_OptimizerLast does not exist in LLVM 3.0, add it manually below.
builder.addExtension(PassManagerBuilder::EP_OptimizerLast, addStripExternalsPass);
#endif
builder.populateFunctionPassManager(fpm);
builder.populateModulePassManager(mpm);
#if LDC_LLVM_VER < 301
addStripExternalsPass(builder, mpm);
#endif
}
/// This routine adds optimization passes based on selected optimization level,
/// OptLevel.
///
/// OptLevel - Optimization Level
static void AddOptimizationPasses(legacy::PassManagerBase &MPM,
legacy::FunctionPassManager &FPM,
TargetMachine *TM, unsigned OptLevel,
unsigned SizeLevel) {
if (!NoVerify || VerifyEach)
FPM.add(createVerifierPass()); // Verify that input is correct
PassManagerBuilder Builder;
Builder.OptLevel = OptLevel;
Builder.SizeLevel = SizeLevel;
if (DisableInline) {
// No inlining pass
} else if (OptLevel > 1) {
Builder.Inliner = createFunctionInliningPass(OptLevel, SizeLevel);
} else {
Builder.Inliner = createAlwaysInlinerLegacyPass();
}
Builder.DisableUnitAtATime = !UnitAtATime;
Builder.DisableUnrollLoops = (DisableLoopUnrolling.getNumOccurrences() > 0) ?
DisableLoopUnrolling : OptLevel == 0;
// This is final, unless there is a #pragma vectorize enable
if (DisableLoopVectorization)
Builder.LoopVectorize = false;
// If option wasn't forced via cmd line (-vectorize-loops, -loop-vectorize)
else if (!Builder.LoopVectorize)
Builder.LoopVectorize = OptLevel > 1 && SizeLevel < 2;
// When #pragma vectorize is on for SLP, do the same as above
Builder.SLPVectorize =
DisableSLPVectorization ? false : OptLevel > 1 && SizeLevel < 2;
// Add target-specific passes that need to run as early as possible.
if (TM)
Builder.addExtension(
PassManagerBuilder::EP_EarlyAsPossible,
[&](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
TM->addEarlyAsPossiblePasses(PM);
});
if (Coroutines)
addCoroutinePassesToExtensionPoints(Builder);
Builder.populateFunctionPassManager(FPM);
Builder.populateModulePassManager(MPM);
}
void llvmutil_optimizemodule(Module * M, TargetMachine * TM) {
PassManager MPM;
llvmutil_addtargetspecificpasses(&MPM, TM);
MPM.add(createVerifierPass()); //make sure we haven't messed stuff up yet
MPM.add(createGlobalDCEPass()); //run this early since anything not in the table of exported functions is still in this module
//this will remove dead functions
PassManagerBuilder PMB;
PMB.OptLevel = 3;
PMB.SizeLevel = 0;
PMB.Inliner = createFunctionInliningPass(PMB.OptLevel, 0);
#if LLVM_VERSION >= 35
PMB.LoopVectorize = true;
PMB.SLPVectorize = true;
#endif
PMB.populateModulePassManager(MPM);
MPM.run(*M);
}
/// AddOptimizationPasses - This routine adds optimization passes
/// based on selected optimization level, OptLevel. This routine
/// duplicates llvm-gcc behaviour.
///
/// OptLevel - Optimization Level
static void AddOptimizationPasses(PassManagerBase &MPM,FunctionPassManager &FPM,
unsigned OptLevel) {
PassManagerBuilder Builder;
Builder.OptLevel = OptLevel;
if (DisableInline) {
// No inlining pass
} else if (OptLevel > 1) {
unsigned Threshold = 225;
if (OptLevel > 2)
Threshold = 275;
Builder.Inliner = createFunctionInliningPass(Threshold);
} else {
Builder.Inliner = createAlwaysInlinerPass();
}
Builder.DisableUnitAtATime = !UnitAtATime;
Builder.DisableUnrollLoops = OptLevel == 0;
Builder.DisableSimplifyLibCalls = DisableSimplifyLibCalls;
Builder.populateFunctionPassManager(FPM);
Builder.populateModulePassManager(MPM);
}
void EmitAssemblyHelper::CreatePasses() {
unsigned OptLevel = CodeGenOpts.OptimizationLevel;
CodeGenOptions::InliningMethod Inlining = CodeGenOpts.Inlining;
// Handle disabling of LLVM optimization, where we want to preserve the
// internal module before any optimization.
if (CodeGenOpts.DisableLLVMOpts) {
OptLevel = 0;
Inlining = CodeGenOpts.NoInlining;
}
PassManagerBuilder PMBuilder;
PMBuilder.OptLevel = OptLevel;
PMBuilder.SizeLevel = CodeGenOpts.OptimizeSize;
PMBuilder.DisableSimplifyLibCalls = !CodeGenOpts.SimplifyLibCalls;
PMBuilder.DisableUnitAtATime = !CodeGenOpts.UnitAtATime;
PMBuilder.DisableUnrollLoops = !CodeGenOpts.UnrollLoops;
// In ObjC ARC mode, add the main ARC optimization passes.
if (LangOpts.ObjCAutoRefCount) {
PMBuilder.addExtension(PassManagerBuilder::EP_EarlyAsPossible,
addObjCARCExpandPass);
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addObjCARCOptPass);
}
if (LangOpts.AddressSanitizer) {
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addAddressSanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addAddressSanitizerPass);
}
// Figure out TargetLibraryInfo.
Triple TargetTriple(TheModule->getTargetTriple());
PMBuilder.LibraryInfo = new TargetLibraryInfo(TargetTriple);
if (!CodeGenOpts.SimplifyLibCalls)
PMBuilder.LibraryInfo->disableAllFunctions();
switch (Inlining) {
case CodeGenOptions::NoInlining: break;
case CodeGenOptions::NormalInlining: {
// FIXME: Derive these constants in a principled fashion.
unsigned Threshold = 225;
if (CodeGenOpts.OptimizeSize == 1) // -Os
Threshold = 75;
else if (CodeGenOpts.OptimizeSize == 2) // -Oz
Threshold = 25;
else if (OptLevel > 2)
Threshold = 275;
PMBuilder.Inliner = createFunctionInliningPass(Threshold);
break;
}
case CodeGenOptions::OnlyAlwaysInlining:
// Respect always_inline.
PMBuilder.Inliner = createAlwaysInlinerPass();
break;
}
// Set up the per-function pass manager.
FunctionPassManager *FPM = getPerFunctionPasses();
if (CodeGenOpts.VerifyModule)
FPM->add(createVerifierPass());
PMBuilder.populateFunctionPassManager(*FPM);
// Set up the per-module pass manager.
PassManager *MPM = getPerModulePasses();
if (CodeGenOpts.EmitGcovArcs || CodeGenOpts.EmitGcovNotes) {
MPM->add(createGCOVProfilerPass(CodeGenOpts.EmitGcovNotes,
CodeGenOpts.EmitGcovArcs,
TargetTriple.isMacOSX()));
if (!CodeGenOpts.DebugInfo)
MPM->add(createStripSymbolsPass(true));
}
PMBuilder.populateModulePassManager(*MPM);
}
// This is the method invoked by the EE to Jit code.
CorJitResult LLILCJit::compileMethod(ICorJitInfo *JitInfo,
CORINFO_METHOD_INFO *MethodInfo,
UINT Flags, BYTE **NativeEntry,
ULONG *NativeSizeOfCode) {
// Bail if input is malformed
if (nullptr == JitInfo || nullptr == MethodInfo || nullptr == NativeEntry ||
nullptr == NativeSizeOfCode) {
return CORJIT_INTERNALERROR;
}
// Prep main outputs
*NativeEntry = nullptr;
*NativeSizeOfCode = 0;
// Set up state for this thread (if necessary)
LLILCJitPerThreadState *PerThreadState = State.get();
if (PerThreadState == nullptr) {
PerThreadState = new LLILCJitPerThreadState();
State.set(PerThreadState);
}
// Set up context for this Jit request
LLILCJitContext Context = LLILCJitContext(PerThreadState);
// Fill in context information from the CLR
Context.JitInfo = JitInfo;
Context.MethodInfo = MethodInfo;
Context.Flags = Flags;
JitInfo->getEEInfo(&Context.EEInfo);
// Fill in context information from LLVM
Context.LLVMContext = &PerThreadState->LLVMContext;
std::unique_ptr<Module> M = Context.getModuleForMethod(MethodInfo);
Context.CurrentModule = M.get();
Context.CurrentModule->setTargetTriple(LLILC_TARGET_TRIPLE);
Context.MethodName = Context.CurrentModule->getModuleIdentifier();
Context.TheABIInfo = ABIInfo::get(*Context.CurrentModule);
// Initialize per invocation JIT options. This should be done after the
// rest of the Context is filled out as it has dependencies on JitInfo,
// Flags and MethodInfo.
JitOptions JitOptions(Context);
Context.Options = &JitOptions;
EngineBuilder Builder(std::move(M));
std::string ErrStr;
Builder.setErrorStr(&ErrStr);
std::unique_ptr<RTDyldMemoryManager> MM(new EEMemoryManager(&Context));
Builder.setMCJITMemoryManager(std::move(MM));
TargetOptions Options;
if (Context.Options->OptLevel != OptLevel::DEBUG_CODE) {
Builder.setOptLevel(CodeGenOpt::Level::Default);
} else {
Builder.setOptLevel(CodeGenOpt::Level::None);
Options.NoFramePointerElim = true;
}
Builder.setTargetOptions(Options);
ExecutionEngine *NewEngine = Builder.create();
if (!NewEngine) {
errs() << "Could not create ExecutionEngine: " << ErrStr << "\n";
return CORJIT_INTERNALERROR;
}
// Don't allow the EE to search for external symbols.
NewEngine->DisableSymbolSearching();
Context.EE = NewEngine;
// Now jit the method.
CorJitResult Result = CORJIT_INTERNALERROR;
if (Context.Options->DumpLevel == DumpLevel::VERBOSE) {
Context.outputDebugMethodName();
}
bool HasMethod = this->readMethod(&Context);
if (HasMethod) {
if (Context.Options->DoInsertStatepoints) {
// If using Precise GC, run the GC-Safepoint insertion
// and lowering passes before generating code.
legacy::PassManager Passes;
Passes.add(createPlaceSafepointsPass());
PassManagerBuilder PMBuilder;
PMBuilder.OptLevel = 0; // Set optimization level to -O0
PMBuilder.SizeLevel = 0; // so that no additional phases are run.
PMBuilder.populateModulePassManager(Passes);
Passes.add(createRewriteStatepointsForGCPass(false));
Passes.run(*Context.CurrentModule);
}
Context.EE->generateCodeForModule(Context.CurrentModule);
// You need to pick up the COFFDyld changes from the MS branch of LLVM
// or this will fail with an "Incompatible object format!" error
// from LLVM's dynamic loader.
uint64_t FunctionAddress =
Context.EE->getFunctionAddress(Context.MethodName);
*NativeEntry = (BYTE *)FunctionAddress;
// TODO: ColdCodeSize, or separated code, is not enabled or included.
*NativeSizeOfCode = Context.HotCodeSize + Context.ReadOnlyDataSize;
// This is a stop-gap point to issue a default stub of GC info. This lets
// the CLR consume our methods cleanly. (and the ETW tracing still works)
// Down the road this will be superseded by a CLR specific
// GCMetadataPrinter instance or similar.
this->outputGCInfo(&Context);
// Dump out any enabled timing info.
TimerGroup::printAll(errs());
// Tell the CLR that we've successfully generated code for this method.
Result = CORJIT_OK;
}
// Clean up a bit
delete Context.EE;
Context.EE = nullptr;
delete Context.TheABIInfo;
Context.TheABIInfo = nullptr;
return Result;
}
int main(int argc, char **argv) {
#ifndef DEBUG_BUGPOINT
llvm::sys::PrintStackTraceOnErrorSignal();
llvm::PrettyStackTraceProgram X(argc, argv);
llvm_shutdown_obj Y; // Call llvm_shutdown() on exit.
#endif
// Initialize passes
PassRegistry &Registry = *PassRegistry::getPassRegistry();
initializeCore(Registry);
initializeScalarOpts(Registry);
initializeObjCARCOpts(Registry);
initializeVectorization(Registry);
initializeIPO(Registry);
initializeAnalysis(Registry);
initializeIPA(Registry);
initializeTransformUtils(Registry);
initializeInstCombine(Registry);
initializeInstrumentation(Registry);
initializeTarget(Registry);
#ifdef LINK_POLLY_INTO_TOOLS
polly::initializePollyPasses(Registry);
#endif
// @LOCALMOD-BEGIN
initializeAddPNaClExternalDeclsPass(Registry);
initializeAllocateDataSegmentPass(Registry);
initializeBackendCanonicalizePass(Registry);
initializeCanonicalizeMemIntrinsicsPass(Registry);
initializeCleanupUsedGlobalsMetadataPass(Registry);
initializeConstantInsertExtractElementIndexPass(Registry);
initializeExpandAllocasPass(Registry);
initializeExpandArithWithOverflowPass(Registry);
initializeExpandByValPass(Registry);
initializeExpandConstantExprPass(Registry);
initializeExpandCtorsPass(Registry);
initializeExpandGetElementPtrPass(Registry);
initializeExpandIndirectBrPass(Registry);
initializeExpandLargeIntegersPass(Registry);
initializeExpandShuffleVectorPass(Registry);
initializeExpandSmallArgumentsPass(Registry);
initializeExpandStructRegsPass(Registry);
initializeExpandTlsConstantExprPass(Registry);
initializeExpandTlsPass(Registry);
initializeExpandVarArgsPass(Registry);
initializeFixVectorLoadStoreAlignmentPass(Registry);
initializeFlattenGlobalsPass(Registry);
initializeGlobalCleanupPass(Registry);
initializeGlobalizeConstantVectorsPass(Registry);
initializeInsertDivideCheckPass(Registry);
initializeInternalizeUsedGlobalsPass(Registry);
initializeNormalizeAlignmentPass(Registry);
initializePNaClABIVerifyFunctionsPass(Registry);
initializePNaClABIVerifyModulePass(Registry);
initializePNaClSjLjEHPass(Registry);
initializePromoteI1OpsPass(Registry);
initializePromoteIntegersPass(Registry);
initializeRemoveAsmMemoryPass(Registry);
initializeRenameEntryPointPass(Registry);
initializeReplacePtrsWithIntsPass(Registry);
initializeResolveAliasesPass(Registry);
initializeResolvePNaClIntrinsicsPass(Registry);
initializeRewriteAtomicsPass(Registry);
initializeRewriteLLVMIntrinsicsPass(Registry);
initializeRewritePNaClLibraryCallsPass(Registry);
initializeSandboxIndirectCallsPass(Registry);
initializeSandboxMemoryAccessesPass(Registry);
initializeSimplifyAllocasPass(Registry);
initializeSimplifyStructRegSignaturesPass(Registry);
initializeStripAttributesPass(Registry);
initializeStripMetadataPass(Registry);
initializeStripModuleFlagsPass(Registry);
initializeStripTlsPass(Registry);
initializeSubstituteUndefsPass(Registry);
// Emscripten passes:
initializeExpandI64Pass(Registry);
initializeExpandInsertExtractElementPass(Registry);
initializeLowerEmAsyncifyPass(Registry);
initializeLowerEmExceptionsPass(Registry);
initializeLowerEmSetjmpPass(Registry);
initializeNoExitRuntimePass(Registry);
// Emscripten passes end.
// @LOCALMOD-END
cl::ParseCommandLineOptions(argc, argv,
"LLVM automatic testcase reducer. See\nhttp://"
"llvm.org/cmds/bugpoint.html"
" for more information.\n");
#ifndef DEBUG_BUGPOINT
sys::SetInterruptFunction(BugpointInterruptFunction);
#endif
LLVMContext& Context = getGlobalContext();
// If we have an override, set it and then track the triple we want Modules
// to use.
if (!OverrideTriple.empty()) {
TargetTriple.setTriple(Triple::normalize(OverrideTriple));
outs() << "Override triple set to '" << TargetTriple.getTriple() << "'\n";
}
if (MemoryLimit < 0) {
// Set the default MemoryLimit. Be sure to update the flag's description if
// you change this.
if (sys::RunningOnValgrind() || UseValgrind)
MemoryLimit = 800;
else
MemoryLimit = 300;
}
BugDriver D(argv[0], FindBugs, TimeoutValue, MemoryLimit,
UseValgrind, Context);
if (D.addSources(InputFilenames)) return 1;
AddToDriver PM(D);
if (StandardLinkOpts) {
PassManagerBuilder Builder;
Builder.Inliner = createFunctionInliningPass();
Builder.populateLTOPassManager(PM);
}
if (OptLevelO1 || OptLevelO2 || OptLevelO3) {
PassManagerBuilder Builder;
if (OptLevelO1)
Builder.Inliner = createAlwaysInlinerPass();
else if (OptLevelO2)
Builder.Inliner = createFunctionInliningPass(225);
else
Builder.Inliner = createFunctionInliningPass(275);
// Note that although clang/llvm-gcc use two separate passmanagers
// here, it shouldn't normally make a difference.
Builder.populateFunctionPassManager(PM);
Builder.populateModulePassManager(PM);
}
for (std::vector<const PassInfo*>::iterator I = PassList.begin(),
E = PassList.end();
I != E; ++I) {
const PassInfo* PI = *I;
D.addPass(PI->getPassArgument());
}
// Bugpoint has the ability of generating a plethora of core files, so to
// avoid filling up the disk, we prevent it
#ifndef DEBUG_BUGPOINT
sys::Process::PreventCoreFiles();
#endif
std::string Error;
bool Failure = D.run(Error);
if (!Error.empty()) {
errs() << Error;
return 1;
}
return Failure;
}
int main(int argc, char* argv[])
{
if(argc < 2)
{
cerr << "Usage: " << argv[0] << " bf_file" << endl;
return -1;
}
ifstream sourceFile(argv[1]);
string line, source;
while(getline(sourceFile, line)) source += line;
// Setup a module and engine for JIT-ing
std::string error;
InitializeNativeTarget();
Module* module = new Module("bfcode", getGlobalContext());
InitializeNativeTarget();
LLVMLinkInJIT();
ExecutionEngine *engine = EngineBuilder(module)
.setErrorStr(&error)
.setOptLevel(CodeGenOpt::Aggressive)
.create();
if(!engine)
{
cout << "No engine created: " << error << endl;
return -1;
}
module->setDataLayout(engine->getTargetData()->getStringRepresentation());
// Compile the BF to IR
cout << "Parsing… " << flush;
Function* func = makeFunc(module, source.c_str());
cout << "done" << endl;
{
ofstream dst("out.ll");
raw_os_ostream rawdst(dst);
rawdst << *module;
}
// Run optimization passes
cout << "Optimizing… " << flush;
PassManagerBuilder PMBuilder;
FunctionPassManager pm(module);
PMBuilder.populateFunctionPassManager(pm);
pm.add(new TargetData(*(engine->getTargetData())));
pm.add(createVerifierPass());
// Eliminate simple loops such as [>>++<<-]
pm.add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
pm.add(createLICMPass()); // Hoist loop invariants
pm.add(createPromoteMemoryToRegisterPass());
pm.add(createIndVarSimplifyPass()); // Canonicalize indvars
pm.add(createLoopDeletionPass()); // Delete dead loops
pm.add(createConstantPropagationPass()); // Propagate constants
pm.add(new CondProp); // Propagate conditionals
// Simplify code
for(int repeat=0; repeat < 3; repeat++)
{
pm.add(createPromoteMemoryToRegisterPass());
pm.add(createGVNPass()); // Remove redundancies
pm.add(createSCCPPass()); // Constant prop with SCCP
pm.add(createLoopDeletionPass());
pm.add(createLoopUnrollPass());
pm.add(createCFGSimplificationPass()); // Merge & remove BBs
pm.add(createInstructionCombiningPass());
pm.add(createConstantPropagationPass()); // Propagate constants
pm.add(createAggressiveDCEPass()); // Delete dead instructions
pm.add(createCFGSimplificationPass()); // Merge & remove BBs
pm.add(createDeadStoreEliminationPass()); // Delete dead stores
pm.add(createMemCpyOptPass()); // Combine multiple stores into memset's
//pm.add(new PutCharAggregatePass);
}
pm.add(createPromoteMemoryToRegisterPass());
// Process
foreach (Function& f, *module)
if (!f.isDeclaration)
pm.run(f);
PassManager pmm;
PMBuilder.populateModulePassManager(pmm);
pmm.add(createConstantMergePass());
pmm.add(createGlobalOptimizerPass());
pmm.add(createGlobalDCEPass());
pmm.add(createIPConstantPropagationPass());
pmm.run(*module);
foreach (Function& f, *module)
if (!f.isDeclaration)
pm.run(f);
pmm.run(*module);
cout << "done" << endl;
{
ofstream dst("optout.ll");
raw_os_ostream rawdst(dst);
rawdst << *module;
}
// Compile …
cout << "Compiling…" << flush;
int (*bf)() = (int (*)())engine->getPointerToFunction(func);
cout << " done" << endl;
// … and run!
return bf();
}
vector<char> CodeGen_PTX_Dev::compile_to_src() {
#ifdef WITH_PTX
debug(2) << "In CodeGen_PTX_Dev::compile_to_src";
// DISABLED - hooked in here to force PrintBeforeAll option - seems to be the only way?
/*char* argv[] = { "llc", "-print-before-all" };*/
/*int argc = sizeof(argv)/sizeof(char*);*/
/*cl::ParseCommandLineOptions(argc, argv, "Halide PTX internal compiler\n");*/
llvm::Triple triple(module->getTargetTriple());
// Allocate target machine
std::string err_str;
const llvm::Target *target = TargetRegistry::lookupTarget(triple.str(), err_str);
internal_assert(target) << err_str << "\n";
TargetOptions options;
#if LLVM_VERSION < 50
options.LessPreciseFPMADOption = true;
#endif
options.PrintMachineCode = false;
options.AllowFPOpFusion = FPOpFusion::Fast;
options.UnsafeFPMath = true;
options.NoInfsFPMath = true;
options.NoNaNsFPMath = true;
options.HonorSignDependentRoundingFPMathOption = false;
options.NoZerosInBSS = false;
options.GuaranteedTailCallOpt = false;
options.StackAlignmentOverride = 0;
std::unique_ptr<TargetMachine>
target_machine(target->createTargetMachine(triple.str(),
mcpu(), mattrs(), options,
llvm::Reloc::PIC_,
#if LLVM_VERSION < 60
llvm::CodeModel::Default,
#else
llvm::CodeModel::Small,
#endif
CodeGenOpt::Aggressive));
internal_assert(target_machine.get()) << "Could not allocate target machine!";
#if LLVM_VERSION >= 60
module->setDataLayout(target_machine->createDataLayout());
#endif
// Set up passes
llvm::SmallString<8> outstr;
raw_svector_ostream ostream(outstr);
ostream.SetUnbuffered();
legacy::FunctionPassManager function_pass_manager(module.get());
legacy::PassManager module_pass_manager;
module_pass_manager.add(createTargetTransformInfoWrapperPass(target_machine->getTargetIRAnalysis()));
function_pass_manager.add(createTargetTransformInfoWrapperPass(target_machine->getTargetIRAnalysis()));
// NVidia's libdevice library uses a __nvvm_reflect to choose
// how to handle denormalized numbers. (The pass replaces calls
// to __nvvm_reflect with a constant via a map lookup. The inliner
// pass then resolves these situations to fast code, often a single
// instruction per decision point.)
//
// The default is (more) IEEE like handling. FTZ mode flushes them
// to zero. (This may only apply to single-precision.)
//
// The libdevice documentation covers other options for math accuracy
// such as replacing division with multiply by the reciprocal and
// use of fused-multiply-add, but they do not seem to be controlled
// by this __nvvvm_reflect mechanism and may be flags to earlier compiler
// passes.
#define kDefaultDenorms 0
#define kFTZDenorms 1
#if LLVM_VERSION <= 40
StringMap<int> reflect_mapping;
reflect_mapping[StringRef("__CUDA_FTZ")] = kFTZDenorms;
module_pass_manager.add(createNVVMReflectPass(reflect_mapping));
#else
// Insert a module flag for the FTZ handling.
module->addModuleFlag(llvm::Module::Override, "nvvm-reflect-ftz",
kFTZDenorms);
if (kFTZDenorms) {
for (llvm::Function &fn : *module) {
fn.addFnAttr("nvptx-f32ftz", "true");
}
}
#endif
PassManagerBuilder b;
b.OptLevel = 3;
#if LLVM_VERSION >= 50
b.Inliner = createFunctionInliningPass(b.OptLevel, 0, false);
#else
b.Inliner = createFunctionInliningPass(b.OptLevel, 0);
#endif
b.LoopVectorize = true;
b.SLPVectorize = true;
#if LLVM_VERSION > 40
target_machine->adjustPassManager(b);
#endif
b.populateFunctionPassManager(function_pass_manager);
b.populateModulePassManager(module_pass_manager);
// Override default to generate verbose assembly.
target_machine->Options.MCOptions.AsmVerbose = true;
// Output string stream
// Ask the target to add backend passes as necessary.
bool fail = target_machine->addPassesToEmitFile(module_pass_manager, ostream,
TargetMachine::CGFT_AssemblyFile,
true);
if (fail) {
internal_error << "Failed to set up passes to emit PTX source\n";
}
// Run optimization passes
function_pass_manager.doInitialization();
for (llvm::Module::iterator i = module->begin(); i != module->end(); i++) {
function_pass_manager.run(*i);
}
function_pass_manager.doFinalization();
module_pass_manager.run(*module);
if (debug::debug_level() >= 2) {
dump();
}
debug(2) << "Done with CodeGen_PTX_Dev::compile_to_src";
debug(1) << "PTX kernel:\n" << outstr.c_str() << "\n";
vector<char> buffer(outstr.begin(), outstr.end());
buffer.push_back(0);
return buffer;
#else // WITH_PTX
return vector<char>();
#endif
}
void CompileContext::CreateEE()
{
string errMsg;
if (verifyModule(*module, &raw_string_ostream(errMsg))) {
Alert((char*)errMsg.c_str());
}
EE.reset(EngineBuilder(module)
.setEngineKind(EngineKind::JIT)
.setUseMCJIT(true)
.setOptLevel(CodeGenOpt::Aggressive)
.create());
EE->InstallLazyFunctionCreator(HspLazyFunctionCreator);
#define REGISTER_RT_(t, name, func) \
do {\
Function *KnownFunction = Function::Create(\
TypeBuilder<t, false>::get(context),\
GlobalValue::ExternalLinkage, name,\
module);\
EE->addGlobalMapping(KnownFunction, (void*)(intptr_t)func);\
} while (false);
#define REGISTER_RT(t, func) REGISTER_RT_(t, #func, func)
REGISTER_RT(void(int, int), Prgcmd);
REGISTER_RT(void(int, int), Modcmd);
REGISTER_RT(void(void*, int, int), VarSet);
REGISTER_RT(void(void*, int), VarSetIndex1);
REGISTER_RT(void(void*, int, int), VarSetIndex2);
REGISTER_RT(void(void*, int, int), VarSetIndex1i);
REGISTER_RT(void(void*, int, int, int), VarSetIndex2i);
REGISTER_RT(void(void*, double, int), VarSetIndex1d);
REGISTER_RT(void(void*, double, int, int), VarSetIndex2d);
REGISTER_RT(void(int), StackPushi);
REGISTER_RT_(void(int), "PushInt", StackPushi);
REGISTER_RT(void(int), StackPushd);
REGISTER_RT_(void(double), "PushDouble", StackPushd);
REGISTER_RT(void(int), StackPushl);
REGISTER_RT_(void(int), "PushLabel", StackPushl);
REGISTER_RT(void(char*), PushStr);
REGISTER_RT(void(void*, int), PushVar);
REGISTER_RT(void(void*, int), PushVAP);
REGISTER_RT(void(), PushDefault);
REGISTER_RT(void(), PushFuncEnd);
REGISTER_RT(void(int), PushFuncPrm1);
REGISTER_RT(void(int), PushFuncPrmI);
REGISTER_RT(void(int), PushFuncPrmD);
REGISTER_RT(void(int, int), PushFuncPrm);
REGISTER_RT(void(int, int), PushFuncPAP);
REGISTER_RT(void*(int), FuncPrm);
REGISTER_RT(void*(int), LocalPrm);
REGISTER_RT(int(int), FuncPrmI);
REGISTER_RT(double(int), FuncPrmD);
REGISTER_RT(void(), CalcAddI);
REGISTER_RT(void(), CalcSubI);
REGISTER_RT(void(), CalcMulI);
REGISTER_RT(void(), CalcDivI);
REGISTER_RT(void(), CalcModI);
REGISTER_RT(void(), CalcAndI);
REGISTER_RT(void(), CalcOrI);
REGISTER_RT(void(), CalcXorI);
REGISTER_RT(void(), CalcEqI);
REGISTER_RT(void(), CalcNeI);
REGISTER_RT(void(), CalcGtI);
REGISTER_RT(void(), CalcLtI);
REGISTER_RT(void(), CalcGtEqI);
REGISTER_RT(void(), CalcLtEqI);
REGISTER_RT(void(), CalcRrI);
REGISTER_RT(void(), CalcLrI);
REGISTER_RT(void(int, int), PushIntfunc);
REGISTER_RT(void(void*, int, int), VarCalc);
REGISTER_RT(void(void*, int), VarInc);
REGISTER_RT(void(int), TaskSwitch);
REGISTER_RT(char(), HspIf);
REGISTER_RT(void(int, int), PushSysvar);
REGISTER_RT(void(int, int), PushExtvar);
REGISTER_RT(void(int, int), PushModcmd);
REGISTER_RT(void(int, int), Extcmd);
REGISTER_RT(void(int, int), Intcmd);
REGISTER_RT(void(int, int), PushDllfunc);
REGISTER_RT(void(int, int), PushDllctrl);
REGISTER_RT(int(), GetTaskID);
//REGISTER_RT(int(Hsp3r*, int, int), Hsp3rReset);
REGISTER_RT(void(void*, int), HspVarCoreArray2);
REGISTER_RT(double(int, int), CallDoubleIntfunc);
REGISTER_RT(int(int, int), CallIntIntfunc);
REGISTER_RT(double(int, int), CallDoubleSysvar);
REGISTER_RT(int(int, int), CallIntSysvar);
REGISTER_RT(int(), PopInt);
REGISTER_RT(double(), PopDouble);
//REGISTER_RT(char(), HspIf);
//REGISTER_RT(void(int, int), PushSysvar);
#undef REGISTER_RT
Passes.reset(new PassManager());
FPM.reset(new FunctionPassManager(module));
Passes->add(new DataLayoutPass(module));
Passes->add(createVerifierPass());
PassManagerBuilder Builder;
Builder.OptLevel = 3;
Builder.SizeLevel = 0;
Builder.Inliner = createFunctionInliningPass(Builder.OptLevel, Builder.SizeLevel);
Builder.populateModulePassManager(*Passes);
Builder.populateLTOPassManager(*Passes, false, true);
Builder.populateFunctionPassManager(*FPM);
Builder.populateModulePassManager(*Passes);
FPM->doInitialization();
for (auto& f : *module) {
// Run the FPM on this function
FPM->run(f);
}
return;
}
int compile(list<string> args, list<string> kgen_args,
string merge, list<string> merge_args,
string input, string output, int arch,
string host_compiler, string fileprefix)
{
//
// The LLVM compiler to emit IR.
//
const char* llvm_compiler = "kernelgen-gfortran";
//
// Interpret kernelgen compile options.
//
for (list<string>::iterator iarg = kgen_args.begin(),
iearg = kgen_args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strncmp(arg, "-Wk,--llvm-compiler=", 20))
llvm_compiler = arg + 20;
}
//
// Generate temporary output file.
// Check if output file is specified in the command line.
// Replace or add output to the temporary file.
//
cfiledesc tmp_output = cfiledesc::mktemp(fileprefix);
bool output_specified = false;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strcmp(arg, "-o"))
{
iarg++;
*iarg = tmp_output.getFilename();
output_specified = true;
break;
}
}
if (!output_specified)
{
args.push_back("-o");
args.push_back(tmp_output.getFilename());
}
//
// 1) Compile source code using regular host compiler.
//
{
if (verbose)
{
cout << host_compiler;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
cout << " " << *iarg;
cout << endl;
}
int status = execute(host_compiler, args, "", NULL, NULL);
if (status) return status;
}
//
// 2) Emit LLVM IR.
//
string out = "";
{
list<string> emit_ir_args;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strcmp(arg, "-c") || !strcmp(arg, "-o"))
{
iarg++;
continue;
}
if (!strcmp(arg, "-g"))
{
continue;
}
emit_ir_args.push_back(*iarg);
}
emit_ir_args.push_back("-fplugin=/opt/kernelgen/lib/dragonegg.so");
emit_ir_args.push_back("-fplugin-arg-dragonegg-emit-ir");
emit_ir_args.push_back("-S");
emit_ir_args.push_back(input);
emit_ir_args.push_back("-o");
emit_ir_args.push_back("-");
if (verbose)
{
cout << llvm_compiler;
for (list<string>::iterator iarg = emit_ir_args.begin(),
iearg = emit_ir_args.end(); iarg != iearg; iarg++)
cout << " " << *iarg;
cout << endl;
}
int status = execute(llvm_compiler, emit_ir_args, "", &out, NULL);
if (status) return status;
}
//
// 3) Record existing module functions.
//
LLVMContext &context = getGlobalContext();
SMDiagnostic diag;
MemoryBuffer* buffer1 = MemoryBuffer::getMemBuffer(out);
auto_ptr<Module> m1;
m1.reset(ParseIR(buffer1, diag, context));
//m1.get()->dump();
//
// 4) Inline calls and extract loops into new functions.
//
MemoryBuffer* buffer2 = MemoryBuffer::getMemBuffer(out);
auto_ptr<Module> m2;
m2.reset(ParseIR(buffer2, diag, context));
{
PassManager manager;
manager.add(createInstructionCombiningPass());
manager.run(*m2.get());
}
std::vector<CallInst *> LoopFuctionCalls;
{
PassManager manager;
manager.add(createBranchedLoopExtractorPass(LoopFuctionCalls));
manager.run(*m2.get());
}
//m2.get()->dump();
//
// 5) Replace call to loop functions with call to launcher.
// Append "always inline" attribute to all other functions.
//
Type* int32Ty = Type::getInt32Ty(context);
Function* launch = Function::Create(
TypeBuilder<types::i<32>(types::i<8>*, types::i<64>, types::i<32>*), true>::get(context),
GlobalValue::ExternalLinkage, "kernelgen_launch", m2.get());
for (Module::iterator f1 = m2.get()->begin(), fe1 = m2.get()->end(); f1 != fe1; f1++)
{
Function* func = f1;
if (func->isDeclaration()) continue;
// Search for the current function in original module
// functions list.
// If function is not in list of original module, then
// it is generated by the loop extractor.
// Append "always inline" attribute to all other functions.
if (m1.get()->getFunction(func->getName()))
{
const AttrListPtr attr = func->getAttributes();
const AttrListPtr attr_new = attr.addAttr(~0U, Attribute::AlwaysInline);
func->setAttributes(attr_new);
continue;
}
// Each such function must be extracted to the
// standalone module and packed into resulting
// object file data section.
if (verbose)
cout << "Preparing loop function " << func->getName().data() <<
" ..." << endl;
// Reset to default visibility.
func->setVisibility(GlobalValue::DefaultVisibility);
// Reset to default linkage.
func->setLinkage(GlobalValue::ExternalLinkage);
// Replace call to this function in module with call to launcher.
bool found = false;
for (Module::iterator f2 = m2->begin(), fe2 = m2->end(); (f2 != fe2) && !found; f2++)
for (Function::iterator bb = f2->begin(); (bb != f2->end()) && !found; bb++)
for (BasicBlock::iterator i = bb->begin(); i != bb->end(); i++)
{
// Check if instruction in focus is a call.
CallInst* call = dyn_cast<CallInst>(cast<Value>(i));
if (!call) continue;
// Check if function is called (needs -instcombine pass).
Function* callee = call->getCalledFunction();
if (!callee) continue;
if (callee->isDeclaration()) continue;
if (callee->getName() != func->getName()) continue;
// Create a constant array holding original called
// function name.
Constant* name = ConstantArray::get(
context, callee->getName(), true);
// Create and initialize the memory buffer for name.
ArrayType* nameTy = cast<ArrayType>(name->getType());
AllocaInst* nameAlloc = new AllocaInst(nameTy, "", call);
StoreInst* nameInit = new StoreInst(name, nameAlloc, "", call);
Value* Idx[2];
Idx[0] = Constant::getNullValue(Type::getInt32Ty(context));
Idx[1] = ConstantInt::get(Type::getInt32Ty(context), 0);
GetElementPtrInst* namePtr = GetElementPtrInst::Create(nameAlloc, Idx, "", call);
// Add pointer to the original function string name.
SmallVector<Value*, 16> call_args;
call_args.push_back(namePtr);
// Add size of the aggregated arguments structure.
{
BitCastInst* BC = new BitCastInst(
call->getArgOperand(0), Type::getInt64PtrTy(context),
"", call);
LoadInst* LI = new LoadInst(BC, "", call);
call_args.push_back(LI);
}
// Add original aggregated structure argument.
call_args.push_back(call->getArgOperand(0));
// Create new function call with new call arguments
// and copy old call properties.
CallInst* newcall = CallInst::Create(launch, call_args, "", call);
//newcall->takeName(call);
newcall->setCallingConv(call->getCallingConv());
newcall->setAttributes(call->getAttributes());
newcall->setDebugLoc(call->getDebugLoc());
// Replace old call with new one.
call->replaceAllUsesWith(newcall);
call->eraseFromParent();
found = true;
break;
}
}
//m2.get()->dump();
//
// 6) Apply optimization passes to the resulting common
// module.
//
{
PassManager manager;
manager.add(createLowerSetJmpPass());
PassManagerBuilder builder;
builder.Inliner = createFunctionInliningPass();
builder.OptLevel = 3;
builder.DisableSimplifyLibCalls = true;
builder.populateModulePassManager(manager);
manager.run(*m2.get());
}
//m2.get()->dump();
//
// 7) Embed the resulting module into object file.
//
{
string ir_string;
raw_string_ostream ir(ir_string);
ir << (*m2.get());
celf e(tmp_output.getFilename(), output);
e.getSection(".data")->addSymbol(
"__kernelgen_" + string(input),
ir_string.c_str(), ir_string.size() + 1);
}
return 0;
}
void swift::performLLVMOptimizations(IRGenOptions &Opts, llvm::Module *Module,
llvm::TargetMachine *TargetMachine) {
SharedTimer timer("LLVM optimization");
// Set up a pipeline.
PassManagerBuilder PMBuilder;
if (Opts.Optimize && !Opts.DisableLLVMOptzns) {
PMBuilder.OptLevel = 3;
PMBuilder.Inliner = llvm::createFunctionInliningPass(200);
PMBuilder.SLPVectorize = true;
PMBuilder.LoopVectorize = true;
PMBuilder.MergeFunctions = true;
} else {
PMBuilder.OptLevel = 0;
if (!Opts.DisableLLVMOptzns)
PMBuilder.Inliner =
llvm::createAlwaysInlinerPass(/*insertlifetime*/false);
}
PMBuilder.addExtension(PassManagerBuilder::EP_ModuleOptimizerEarly,
addSwiftStackPromotionPass);
// If the optimizer is enabled, we run the ARCOpt pass in the scalar optimizer
// and the Contract pass as late as possible.
if (!Opts.DisableLLVMARCOpts) {
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addSwiftARCOptPass);
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addSwiftContractPass);
}
if (Opts.Sanitize == SanitizerKind::Address) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addAddressSanitizerPasses);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addAddressSanitizerPasses);
}
if (Opts.Sanitize == SanitizerKind::Thread) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addThreadSanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addThreadSanitizerPass);
}
// Configure the function passes.
legacy::FunctionPassManager FunctionPasses(Module);
FunctionPasses.add(createTargetTransformInfoWrapperPass(
TargetMachine->getTargetIRAnalysis()));
if (Opts.Verify)
FunctionPasses.add(createVerifierPass());
PMBuilder.populateFunctionPassManager(FunctionPasses);
// The PMBuilder only knows about LLVM AA passes. We should explicitly add
// the swift AA pass after the other ones.
if (!Opts.DisableLLVMARCOpts) {
FunctionPasses.add(createSwiftAAWrapperPass());
FunctionPasses.add(createExternalAAWrapperPass([](Pass &P, Function &,
AAResults &AAR) {
if (auto *WrapperPass = P.getAnalysisIfAvailable<SwiftAAWrapperPass>())
AAR.addAAResult(WrapperPass->getResult());
}));
}
// Run the function passes.
FunctionPasses.doInitialization();
for (auto I = Module->begin(), E = Module->end(); I != E; ++I)
if (!I->isDeclaration())
FunctionPasses.run(*I);
FunctionPasses.doFinalization();
// Configure the module passes.
legacy::PassManager ModulePasses;
ModulePasses.add(createTargetTransformInfoWrapperPass(
TargetMachine->getTargetIRAnalysis()));
PMBuilder.populateModulePassManager(ModulePasses);
// The PMBuilder only knows about LLVM AA passes. We should explicitly add
// the swift AA pass after the other ones.
if (!Opts.DisableLLVMARCOpts) {
ModulePasses.add(createSwiftAAWrapperPass());
ModulePasses.add(createExternalAAWrapperPass([](Pass &P, Function &,
AAResults &AAR) {
if (auto *WrapperPass = P.getAnalysisIfAvailable<SwiftAAWrapperPass>())
AAR.addAAResult(WrapperPass->getResult());
}));
}
// If we're generating a profile, add the lowering pass now.
if (Opts.GenerateProfile)
ModulePasses.add(createInstrProfilingPass());
if (Opts.Verify)
ModulePasses.add(createVerifierPass());
if (Opts.PrintInlineTree)
ModulePasses.add(createInlineTreePrinterPass());
// Do it.
ModulePasses.run(*Module);
}
int main(int argc, char **argv) {
#ifndef DEBUG_BUGPOINT
llvm::sys::PrintStackTraceOnErrorSignal();
llvm::PrettyStackTraceProgram X(argc, argv);
llvm_shutdown_obj Y; // Call llvm_shutdown() on exit.
#endif
// Initialize passes
PassRegistry &Registry = *PassRegistry::getPassRegistry();
initializeCore(Registry);
initializeScalarOpts(Registry);
initializeObjCARCOpts(Registry);
initializeVectorization(Registry);
initializeIPO(Registry);
initializeAnalysis(Registry);
initializeIPA(Registry);
initializeTransformUtils(Registry);
initializeInstCombine(Registry);
initializeInstrumentation(Registry);
initializeTarget(Registry);
#ifdef LINK_POLLY_INTO_TOOLS
polly::initializePollyPasses(Registry);
#endif
cl::ParseCommandLineOptions(argc, argv,
"LLVM automatic testcase reducer. See\nhttp://"
"llvm.org/cmds/bugpoint.html"
" for more information.\n");
#ifndef DEBUG_BUGPOINT
sys::SetInterruptFunction(BugpointInterruptFunction);
#endif
LLVMContext& Context = getGlobalContext();
// If we have an override, set it and then track the triple we want Modules
// to use.
if (!OverrideTriple.empty()) {
TargetTriple.setTriple(Triple::normalize(OverrideTriple));
outs() << "Override triple set to '" << TargetTriple.getTriple() << "'\n";
}
if (MemoryLimit < 0) {
// Set the default MemoryLimit. Be sure to update the flag's description if
// you change this.
if (sys::RunningOnValgrind() || UseValgrind)
MemoryLimit = 800;
else
MemoryLimit = 400;
}
BugDriver D(argv[0], FindBugs, TimeoutValue, MemoryLimit,
UseValgrind, Context);
if (D.addSources(InputFilenames)) return 1;
AddToDriver PM(D);
if (StandardLinkOpts) {
PassManagerBuilder Builder;
Builder.Inliner = createFunctionInliningPass();
Builder.populateLTOPassManager(PM);
}
if (OptLevelO1 || OptLevelO2 || OptLevelO3) {
PassManagerBuilder Builder;
if (OptLevelO1)
Builder.Inliner = createAlwaysInlinerPass();
else if (OptLevelO2)
Builder.Inliner = createFunctionInliningPass(225);
else
Builder.Inliner = createFunctionInliningPass(275);
// Note that although clang/llvm-gcc use two separate passmanagers
// here, it shouldn't normally make a difference.
Builder.populateFunctionPassManager(PM);
Builder.populateModulePassManager(PM);
}
for (std::vector<const PassInfo*>::iterator I = PassList.begin(),
E = PassList.end();
I != E; ++I) {
const PassInfo* PI = *I;
D.addPass(PI->getPassArgument());
}
// Bugpoint has the ability of generating a plethora of core files, so to
// avoid filling up the disk, we prevent it
#ifndef DEBUG_BUGPOINT
sys::Process::PreventCoreFiles();
#endif
std::string Error;
bool Failure = D.run(Error);
if (!Error.empty()) {
errs() << Error;
return 1;
}
return Failure;
}
// main - Entry point for the sync compiler.
//
int main(int argc, char **argv) {
sys::PrintStackTraceOnErrorSignal();
PrettyStackTraceProgram X(argc, argv);
// Enable debug stream buffering.
EnableDebugBuffering = true;
LLVMContext &Context = getGlobalContext();
llvm_shutdown_obj Y; // Call llvm_shutdown() on exit.
// Initialize target first, so that --version shows registered targets.
LLVMInitializeVerilogBackendTarget();
LLVMInitializeVerilogBackendTargetInfo();
LLVMInitializeVerilogBackendTargetMC();
cl::ParseCommandLineOptions(argc, argv, "llvm system compiler\n");
SMDiagnostic Err;
LuaScript *S = &scriptEngin();
S->init();
// Run the lua script.
if (!S->runScriptFile(InputFilename, Err)){
Err.print(argv[0], errs());
return 1;
}
S->updateStatus();
// Load the module to be compiled...
std::auto_ptr<Module> M;
M.reset(ParseIRFile(S->getValue<std::string>("InputFile"),
Err, Context));
if (M.get() == 0) {
Err.print(argv[0], errs());
return 1;
}
Module &mod = *M.get();
// TODO: Build the right triple.
Triple TheTriple(mod.getTargetTriple());
TargetOptions TO;
std::auto_ptr<TargetMachine>
target(TheVBackendTarget.createTargetMachine(TheTriple.getTriple(), "",
S->getDataLayout(), TO));
// Build up all of the passes that we want to do to the module.
PassManagerBuilder Builder;
Builder.DisableUnrollLoops = true;
Builder.LibraryInfo = new TargetLibraryInfo();
Builder.LibraryInfo->disableAllFunctions();
Builder.OptLevel = 3;
Builder.SizeLevel = 2;
Builder.DisableSimplifyLibCalls = true;
Builder.Inliner = createHLSInlinerPass();
Builder.addExtension(PassManagerBuilder::EP_LoopOptimizerEnd,
LoopOptimizerEndExtensionFn);
PassManager Passes;
Passes.add(new TargetData(*target->getTargetData()));
// Add the immutable target-specific alias analysis ahead of all others AAs.
Passes.add(createVAliasAnalysisPass(target->getIntrinsicInfo()));
Passes.add(createVerifierPass());
// This is the final bitcode, internalize it to expose more optimization
// opportunities. Note that we should internalize it before SW/HW partition,
// otherwise we may lost some information that help the later internalize.
Passes.add(createInternalizePass(true));
// Perform Software/Hardware partition.
Passes.add(createFunctionFilterPass(S->getOutputStream("SoftwareIROutput")));
Passes.add(createGlobalDCEPass());
// Optimize the hardware part.
//Builder.populateFunctionPassManager(*FPasses);
Builder.populateModulePassManager(Passes);
Builder.populateLTOPassManager(Passes,
/*Internalize*/true,
/*RunInliner*/true);
//PM.add(createPrintModulePass(&dbgs()));
// We do not use the stream that passing into addPassesToEmitFile.
formatted_raw_ostream formatted_nulls(nulls());
// Ask the target to add backend passes as necessary.
target->addPassesToEmitFile(Passes, formatted_nulls,
TargetMachine::CGFT_Null,
false/*NoVerify*/);
// Analyse the slack between registers.
Passes.add(createCombPathDelayAnalysisPass());
Passes.add(createVerilogASTWriterPass(S->getOutputStream("RTLOutput")));
// Run some scripting passes.
for (LuaScript::scriptpass_it I = S->passes_begin(), E = S->passes_end();
I != E; ++I) {
const luabind::object &o = *I;
Pass *P = createScriptingPass(
luabind::object_cast<std::string>(I.key()).c_str(),
luabind::object_cast<std::string>(o["FunctionScript"]).c_str(),
luabind::object_cast<std::string>(o["GlobalScript"]).c_str());
Passes.add(P);
}
// Run the passes.
Passes.run(mod);
// If no error occur, keep the files.
S->keepAllFiles();
return 0;
}
static void addOptimizationPasses(PassManagerBase &mpm, FunctionPassManager &fpm,
#endif
unsigned optLevel, unsigned sizeLevel) {
fpm.add(createVerifierPass()); // Verify that input is correct
PassManagerBuilder builder;
builder.OptLevel = optLevel;
builder.SizeLevel = sizeLevel;
if (willInline()) {
unsigned threshold = 225;
if (sizeLevel == 1) // -Os
threshold = 75;
else if (sizeLevel == 2) // -Oz
threshold = 25;
if (optLevel > 2)
threshold = 275;
builder.Inliner = createFunctionInliningPass(threshold);
} else {
builder.Inliner = createAlwaysInlinerPass();
}
#if LDC_LLVM_VER < 304
builder.DisableSimplifyLibCalls = disableSimplifyLibCalls;
#endif
builder.DisableUnitAtATime = !unitAtATime;
builder.DisableUnrollLoops = optLevel == 0;
#if LDC_LLVM_VER >= 304
builder.DisableUnrollLoops = (disableLoopUnrolling.getNumOccurrences() > 0) ?
disableLoopUnrolling : optLevel == 0;
// This is final, unless there is a #pragma vectorize enable
if (disableLoopVectorization)
builder.LoopVectorize = false;
// If option wasn't forced via cmd line (-vectorize-loops, -loop-vectorize)
else if (!builder.LoopVectorize)
builder.LoopVectorize = optLevel > 1 && sizeLevel < 2;
// When #pragma vectorize is on for SLP, do the same as above
builder.SLPVectorize =
disableSLPVectorization ? false : optLevel > 1 && sizeLevel < 2;
#else
/* builder.Vectorize is set in ctor from command line switch */
#endif
#if LDC_LLVM_VER >= 303
if (opts::sanitize == opts::AddressSanitizer) {
builder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addAddressSanitizerPasses);
builder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addAddressSanitizerPasses);
}
if (opts::sanitize == opts::MemorySanitizer) {
builder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addMemorySanitizerPass);
builder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addMemorySanitizerPass);
}
if (opts::sanitize == opts::ThreadSanitizer) {
builder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addThreadSanitizerPass);
builder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addThreadSanitizerPass);
}
#endif
if (!disableLangSpecificPasses) {
if (!disableSimplifyDruntimeCalls)
builder.addExtension(PassManagerBuilder::EP_LoopOptimizerEnd, addSimplifyDRuntimeCallsPass);
if (!disableGCToStack)
builder.addExtension(PassManagerBuilder::EP_LoopOptimizerEnd, addGarbageCollect2StackPass);
}
// EP_OptimizerLast does not exist in LLVM 3.0, add it manually below.
builder.addExtension(PassManagerBuilder::EP_OptimizerLast, addStripExternalsPass);
builder.populateFunctionPassManager(fpm);
builder.populateModulePassManager(mpm);
}
int main(int argc, char **argv) {
#ifndef DEBUG_BUGPOINT
InitLLVM X(argc, argv);
#endif
// Initialize passes
PassRegistry &Registry = *PassRegistry::getPassRegistry();
initializeCore(Registry);
initializeScalarOpts(Registry);
initializeObjCARCOpts(Registry);
initializeVectorization(Registry);
initializeIPO(Registry);
initializeAnalysis(Registry);
initializeTransformUtils(Registry);
initializeInstCombine(Registry);
initializeAggressiveInstCombine(Registry);
initializeInstrumentation(Registry);
initializeTarget(Registry);
#ifdef LINK_POLLY_INTO_TOOLS
polly::initializePollyPasses(Registry);
#endif
if (std::getenv("bar") == (char*) -1) {
InitializeAllTargets();
InitializeAllTargetMCs();
InitializeAllAsmPrinters();
InitializeAllAsmParsers();
}
cl::ParseCommandLineOptions(argc, argv,
"LLVM automatic testcase reducer. See\nhttp://"
"llvm.org/cmds/bugpoint.html"
" for more information.\n");
#ifndef DEBUG_BUGPOINT
sys::SetInterruptFunction(BugpointInterruptFunction);
#endif
LLVMContext Context;
// If we have an override, set it and then track the triple we want Modules
// to use.
if (!OverrideTriple.empty()) {
TargetTriple.setTriple(Triple::normalize(OverrideTriple));
outs() << "Override triple set to '" << TargetTriple.getTriple() << "'\n";
}
if (MemoryLimit < 0) {
// Set the default MemoryLimit. Be sure to update the flag's description if
// you change this.
if (sys::RunningOnValgrind() || UseValgrind)
MemoryLimit = 800;
else
MemoryLimit = 400;
#if (LLVM_ADDRESS_SANITIZER_BUILD || LLVM_MEMORY_SANITIZER_BUILD || \
LLVM_THREAD_SANITIZER_BUILD)
// Starting from kernel 4.9 memory allocated with mmap is counted against
// RLIMIT_DATA. Sanitizers need to allocate tens of terabytes for shadow.
MemoryLimit = 0;
#endif
}
BugDriver D(argv[0], FindBugs, TimeoutValue, MemoryLimit, UseValgrind,
Context);
if (D.addSources(InputFilenames))
return 1;
AddToDriver PM(D);
if (StandardLinkOpts) {
PassManagerBuilder Builder;
Builder.Inliner = createFunctionInliningPass();
Builder.populateLTOPassManager(PM);
}
if (OptLevelO1 || OptLevelO2 || OptLevelO3) {
PassManagerBuilder Builder;
if (OptLevelO1)
Builder.Inliner = createAlwaysInlinerLegacyPass();
else if (OptLevelOs || OptLevelO2)
Builder.Inliner = createFunctionInliningPass(
2, OptLevelOs ? 1 : 0, false);
else
Builder.Inliner = createFunctionInliningPass(275);
Builder.populateFunctionPassManager(PM);
Builder.populateModulePassManager(PM);
}
for (const PassInfo *PI : PassList)
D.addPass(PI->getPassArgument());
// Bugpoint has the ability of generating a plethora of core files, so to
// avoid filling up the disk, we prevent it
#ifndef DEBUG_BUGPOINT
sys::Process::PreventCoreFiles();
#endif
if (Error E = D.run()) {
errs() << toString(std::move(E));
return 1;
}
return 0;
}
static bool ParseFile(const std::string &Filename,
const std::vector<std::string> &IncludeDirs,
SmallVectorImpl<std::string> &OutputFiles) {
ErrorOr<std::unique_ptr<MemoryBuffer>> MBOrErr =
MemoryBuffer::getFileOrSTDIN(Filename);
if (std::error_code EC = MBOrErr.getError()) {
llvm::errs() << "Could not open input file '" << Filename << "': "
<< EC.message() <<"\n";
return true;
}
std::unique_ptr<llvm::MemoryBuffer> MB = std::move(MBOrErr.get());
// Record the location of the include directory so that the lexer can find it
// later.
SourceMgr SrcMgr;
SrcMgr.setIncludeDirs(IncludeDirs);
// Tell SrcMgr about this buffer, which is what Parser will pick up.
SrcMgr.AddNewSourceBuffer(std::move(MB), llvm::SMLoc());
LangOptions Opts;
Opts.DefaultReal8 = DefaultReal8;
Opts.DefaultDouble8 = DefaultDouble8;
Opts.DefaultInt8 = DefaultInt8;
Opts.ReturnComments = ReturnComments;
if(!FreeForm && !FixedForm) {
llvm::StringRef Ext = llvm::sys::path::extension(Filename);
if(Ext.equals_lower(".f")) {
Opts.FixedForm = 1;
Opts.FreeForm = 0;
}
} else if(FixedForm) {
Opts.FixedForm = 1;
Opts.FreeForm = 0;
}
TextDiagnosticPrinter TDP(SrcMgr);
DiagnosticsEngine Diag(new DiagnosticIDs,&SrcMgr, &TDP, false);
// Chain in -verify checker, if requested.
if(RunVerifier)
Diag.setClient(new VerifyDiagnosticConsumer(Diag));
ASTContext Context(SrcMgr, Opts);
Sema SA(Context, Diag);
Parser P(SrcMgr, Opts, Diag, SA);
Diag.getClient()->BeginSourceFile(Opts, &P.getLexer());
P.ParseProgramUnits();
Diag.getClient()->EndSourceFile();
// Dump
if(PrintAST || DumpAST) {
auto Dumper = CreateASTDumper("");
Dumper->HandleTranslationUnit(Context);
delete Dumper;
}
// Emit
if(!SyntaxOnly && !Diag.hadErrors()) {
flang::TargetOptions TargetOptions;
TargetOptions.Triple = TargetTriple.empty()? llvm::sys::getDefaultTargetTriple() :
TargetTriple;
TargetOptions.CPU = llvm::sys::getHostCPUName();
auto CG = CreateLLVMCodeGen(Diag, Filename == ""? std::string("module") : Filename,
CodeGenOptions(), TargetOptions, llvm::getGlobalContext());
CG->Initialize(Context);
CG->HandleTranslationUnit(Context);
BackendAction BA = Backend_EmitObj;
if(EmitASM) BA = Backend_EmitAssembly;
if(EmitLLVM) BA = Backend_EmitLL;
const llvm::Target *TheTarget = 0;
std::string Err;
TheTarget = llvm::TargetRegistry::lookupTarget(TargetOptions.Triple, Err);
CodeGenOpt::Level TMOptLevel = CodeGenOpt::Default;
switch(OptLevel) {
case 0: TMOptLevel = CodeGenOpt::None; break;
case 3: TMOptLevel = CodeGenOpt::Aggressive; break;
}
llvm::TargetOptions Options;
auto TM = TheTarget->createTargetMachine(TargetOptions.Triple, TargetOptions.CPU, "", Options,
Reloc::Default, CodeModel::Default,
TMOptLevel);
if(!(EmitLLVM && OptLevel == 0)) {
auto TheModule = CG->GetModule();
auto PM = new llvm::legacy::PassManager();
//llvm::legacy::FunctionPassManager *FPM = new llvm::legacy::FunctionPassManager(TheModule);
//FPM->add(new DataLayoutPass());
//PM->add(new llvm::DataLayoutPass());
//TM->addAnalysisPasses(*PM);
PM->add(createPromoteMemoryToRegisterPass());
PassManagerBuilder PMBuilder;
PMBuilder.OptLevel = OptLevel;
PMBuilder.SizeLevel = 0;
PMBuilder.LoopVectorize = true;
PMBuilder.SLPVectorize = true;
unsigned Threshold = 225;
if (OptLevel > 2)
Threshold = 275;
PMBuilder.Inliner = createFunctionInliningPass(Threshold);
PMBuilder.populateModulePassManager(*PM);
//llvm::legacy::PassManager *MPM = new llvm::legacy::PassManager();
//PMBuilder.populateModulePassManager(*MPM);
PM->run(*TheModule);
//MPM->run(*TheModule);
delete PM;
//delete MPM;
}
if(Interpret) {
//const char *Env[] = { "", nullptr };
//Execute(CG->ReleaseModule(), Env);
} else if(OutputFile == "-"){
// FIXME: outputting to stdout is broken
//EmitFile(llvm::outs(), CG->GetModule(), TM, BA);
OutputFiles.push_back(GetOutputName("stdout",BA));
EmitOutputFile(OutputFiles.back(), CG->GetModule(), TM, BA);
}else {
OutputFiles.push_back(GetOutputName(Filename, BA));
EmitOutputFile(OutputFiles.back(), CG->GetModule(), TM, BA);
}
delete CG;
}
return Diag.hadErrors();
}
void EmitAssemblyHelper::CreatePasses() {
unsigned OptLevel = CodeGenOpts.OptimizationLevel;
CodeGenOptions::InliningMethod Inlining = CodeGenOpts.Inlining;
// Handle disabling of LLVM optimization, where we want to preserve the
// internal module before any optimization.
if (CodeGenOpts.DisableLLVMOpts) {
OptLevel = 0;
Inlining = CodeGenOpts.NoInlining;
}
PassManagerBuilder PMBuilder;
PMBuilder.OptLevel = OptLevel;
PMBuilder.SizeLevel = CodeGenOpts.OptimizeSize;
PMBuilder.DisableSimplifyLibCalls = !CodeGenOpts.SimplifyLibCalls;
PMBuilder.DisableUnitAtATime = !CodeGenOpts.UnitAtATime;
PMBuilder.DisableUnrollLoops = !CodeGenOpts.UnrollLoops;
// In ObjC ARC mode, add the main ARC optimization passes.
if (LangOpts.ObjCAutoRefCount) {
PMBuilder.addExtension(PassManagerBuilder::EP_EarlyAsPossible,
addObjCARCExpandPass);
PMBuilder.addExtension(PassManagerBuilder::EP_ModuleOptimizerEarly,
addObjCARCAPElimPass);
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addObjCARCOptPass);
}
if (CodeGenOpts.BoundsChecking > 0) {
BoundsChecking = CodeGenOpts.BoundsChecking;
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addBoundsCheckingPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addBoundsCheckingPass);
}
if (LangOpts.AddressSanitizer) {
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addAddressSanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addAddressSanitizerPass);
}
if (LangOpts.ThreadSanitizer) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addThreadSanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addThreadSanitizerPass);
}
// Figure out TargetLibraryInfo.
Triple TargetTriple(TheModule->getTargetTriple());
PMBuilder.LibraryInfo = new TargetLibraryInfo(TargetTriple);
if (!CodeGenOpts.SimplifyLibCalls)
PMBuilder.LibraryInfo->disableAllFunctions();
switch (Inlining) {
case CodeGenOptions::NoInlining: break;
case CodeGenOptions::NormalInlining: {
// FIXME: Derive these constants in a principled fashion.
unsigned Threshold = 225;
if (CodeGenOpts.OptimizeSize == 1) // -Os
Threshold = 75;
else if (CodeGenOpts.OptimizeSize == 2) // -Oz
Threshold = 25;
else if (OptLevel > 2)
Threshold = 275;
PMBuilder.Inliner = createFunctionInliningPass(Threshold);
break;
}
case CodeGenOptions::OnlyAlwaysInlining:
// Respect always_inline.
if (OptLevel == 0)
// Do not insert lifetime intrinsics at -O0.
PMBuilder.Inliner = createAlwaysInlinerPass(false);
else
PMBuilder.Inliner = createAlwaysInlinerPass();
break;
}
// Set up the per-function pass manager.
FunctionPassManager *FPM = getPerFunctionPasses();
if (CodeGenOpts.VerifyModule)
FPM->add(createVerifierPass());
PMBuilder.populateFunctionPassManager(*FPM);
// Set up the per-module pass manager.
PassManager *MPM = getPerModulePasses();
if (CodeGenOpts.EmitGcovArcs || CodeGenOpts.EmitGcovNotes) {
MPM->add(createGCOVProfilerPass(CodeGenOpts.EmitGcovNotes,
CodeGenOpts.EmitGcovArcs,
TargetTriple.isMacOSX()));
if (CodeGenOpts.DebugInfo == CodeGenOptions::NoDebugInfo)
MPM->add(createStripSymbolsPass(true));
}
// Add the memory safety passes for control-flow integrity
if (CodeGenOpts.MemSafety) {
// Make sure everything that can be in an LLVM register is.
MPM->add (createPromoteMemoryToRegisterPass());
MPM->add (createUnifyFunctionExitNodesPass());
MPM->add (new CFIChecks());
}
PMBuilder.populateModulePassManager(*MPM);
if (CodeGenOpts.SoftBound) {
// Make sure SoftBound+CETS is run after optimization with atleast mem2reg run
MPM->add(new DominatorTree());
MPM->add(new DominanceFrontier());
MPM->add(new LoopInfo());
MPM->add(new InitializeSoftBound());
MPM->add(new SoftBoundCETSPass());
}
// Add the memory safety passes
if (CodeGenOpts.MemSafety) {
MPM->add (createCommonMSCInfoPass());
MPM->add (createSAFECodeMSCInfoPass());
// C standard library / format string function transforms
if (!CodeGenOpts.BaggyBounds) {
MPM->add (new StringTransform());
MPM->add (new FormatStringTransform());
MPM->add (new RegisterVarargCallSites());
MPM->add (new LoggingFunctions());
}
MPM->add (new InitAllocas());
MPM->add (createRegisterGlobalsPass(/*RegUncertain=*/true,
/*MakeInternal=*/false));
MPM->add (createRemoveUnsuitableGlobalRegistrationsPass());
MPM->add (new RegisterMainArgs());
MPM->add (createInstrumentFreeCallsPass());
MPM->add (new RegisterCustomizedAllocation());
MPM->add (new LoopInfo ());
MPM->add (new DominatorTree ());
MPM->add (createRegisterStackPoolsPass(/*RegByval=*/true));
MPM->add (createUnregisterStackPoolsPass());
MPM->add (createSpecializeCMSCallsPass());
MPM->add (new RegisterRuntimeInitializer(CodeGenOpts.MemSafetyLogFile.c_str()));
MPM->add (new DebugInstrument());
MPM->add (createInstrumentMemoryAccessesPass());
MPM->add (createInstrumentGEPsPass());
MPM->add (createSpecializeCMSCallsPass());
MPM->add (new ScalarEvolution());
MPM->add (new ArrayBoundsCheckLocal());
MPM->add (createOptimizeGEPChecksPass());
MPM->add (createExactCheckOptPass());
MPM->add (new ScalarEvolution());
MPM->add (createLocalArrayBoundsAnalysisPass());
MPM->add (createOptimizeFastMemoryChecksPass());
MPM->add (createOptimizeIdenticalLSChecksPass());
MPM->add (new DominatorTree());
MPM->add (new ScalarEvolution());
MPM->add (createOptimizeImpliedFastLSChecksPass());
MPM->add (new OptimizeChecks());
MPM->add (createOptimizeMemoryRegistrationsPass(/*AllowFastChecks=*/true));
if (CodeGenOpts.MemSafeTerminate) {
MPM->add (llvm::createSCTerminatePass ());
}
}
if (CodeGenOpts.BaggyBounds) {
MPM->add (new InsertBaggyBoundsChecks());
}
//
// Rerun the LLVM optimizations again.
//
PMBuilder.populateModulePassManager(*MPM);
// For SAFECode, do the debug instrumentation and OOB rewriting after
// all optimization is done.
if (CodeGenOpts.MemSafety) {
MPM->add (new DebugInstrument());
MPM->add (new RewriteOOB());
}
}
