static void AddStandardCompilePasses(PassManager &PM) {
PM.add(createVerifierPass()); // Verify that input is correct
#if LLVM_VERSION_CODE < LLVM_VERSION(3, 0)
addPass(PM, createLowerSetJmpPass()); // Lower llvm.setjmp/.longjmp
#endif
// If the -strip-debug command line option was specified, do it.
if (StripDebug)
addPass(PM, createStripSymbolsPass(true));
if (DisableOptimizations) return;
addPass(PM, createCFGSimplificationPass()); // Clean up disgusting code
addPass(PM, createPromoteMemoryToRegisterPass());// Kill useless allocas
addPass(PM, createGlobalOptimizerPass()); // Optimize out global vars
addPass(PM, createGlobalDCEPass()); // Remove unused fns and globs
addPass(PM, createIPConstantPropagationPass());// IP Constant Propagation
addPass(PM, createDeadArgEliminationPass()); // Dead argument elimination
addPass(PM, createInstructionCombiningPass()); // Clean up after IPCP & DAE
addPass(PM, createCFGSimplificationPass()); // Clean up after IPCP & DAE
addPass(PM, createPruneEHPass()); // Remove dead EH info
addPass(PM, createFunctionAttrsPass()); // Deduce function attrs
if (!DisableInline)
addPass(PM, createFunctionInliningPass()); // Inline small functions
addPass(PM, createArgumentPromotionPass()); // Scalarize uninlined fn args
#if LLVM_VERSION_CODE < LLVM_VERSION(3, 4)
addPass(PM, createSimplifyLibCallsPass()); // Library Call Optimizations
#endif
addPass(PM, createInstructionCombiningPass()); // Cleanup for scalarrepl.
addPass(PM, createJumpThreadingPass()); // Thread jumps.
addPass(PM, createCFGSimplificationPass()); // Merge & remove BBs
addPass(PM, createScalarReplAggregatesPass()); // Break up aggregate allocas
addPass(PM, createInstructionCombiningPass()); // Combine silly seq's
addPass(PM, createTailCallEliminationPass()); // Eliminate tail calls
addPass(PM, createCFGSimplificationPass()); // Merge & remove BBs
addPass(PM, createReassociatePass()); // Reassociate expressions
addPass(PM, createLoopRotatePass());
addPass(PM, createLICMPass()); // Hoist loop invariants
addPass(PM, createLoopUnswitchPass()); // Unswitch loops.
// FIXME : Removing instcombine causes nestedloop regression.
addPass(PM, createInstructionCombiningPass());
addPass(PM, createIndVarSimplifyPass()); // Canonicalize indvars
addPass(PM, createLoopDeletionPass()); // Delete dead loops
addPass(PM, createLoopUnrollPass()); // Unroll small loops
addPass(PM, createInstructionCombiningPass()); // Clean up after the unroller
addPass(PM, createGVNPass()); // Remove redundancies
addPass(PM, createMemCpyOptPass()); // Remove memcpy / form memset
addPass(PM, createSCCPPass()); // Constant prop with SCCP
// Run instcombine after redundancy elimination to exploit opportunities
// opened up by them.
addPass(PM, createInstructionCombiningPass());
addPass(PM, createDeadStoreEliminationPass()); // Delete dead stores
addPass(PM, createAggressiveDCEPass()); // Delete dead instructions
addPass(PM, createCFGSimplificationPass()); // Merge & remove BBs
addPass(PM, createStripDeadPrototypesPass()); // Get rid of dead prototypes
#if LLVM_VERSION_CODE < LLVM_VERSION(3, 0)
addPass(PM, createDeadTypeEliminationPass()); // Eliminate dead types
#endif
addPass(PM, createConstantMergePass()); // Merge dup global constants
}
/// Optimize - Perform link time optimizations. This will run the scalar
/// optimizations, any loaded plugin-optimization modules, and then the
/// inter-procedural optimizations if applicable.
void Optimize(Module *M, const std::string &EntryPoint) {
// Instantiate the pass manager to organize the passes.
PassManager Passes;
// If we're verifying, start off with a verification pass.
if (VerifyEach)
Passes.add(createVerifierPass());
#if LLVM_VERSION_CODE <= LLVM_VERSION(3, 1)
// Add an appropriate TargetData instance for this module...
addPass(Passes, new TargetData(M));
#elif LLVM_VERSION_CODE < LLVM_VERSION(3, 5)
// Add an appropriate DataLayout instance for this module...
addPass(Passes, new DataLayout(M));
#else
// Add an appropriate DataLayout instance for this module...
addPass(Passes, new DataLayoutPass(M));
#endif
// DWD - Run the opt standard pass list as well.
AddStandardCompilePasses(Passes);
if (!DisableOptimizations) {
// Now that composite has been compiled, scan through the module, looking
// for a main function. If main is defined, mark all other functions
// internal.
if (!DisableInternalize) {
#if LLVM_VERSION_CODE >= LLVM_VERSION(3, 2)
ModulePass *pass = createInternalizePass(
std::vector<const char *>(1, EntryPoint.c_str()));
#else
ModulePass *pass = createInternalizePass(true);
#endif
addPass(Passes, pass);
}
// Propagate constants at call sites into the functions they call. This
// opens opportunities for globalopt (and inlining) by substituting function
// pointers passed as arguments to direct uses of functions.
addPass(Passes, createIPSCCPPass());
// Now that we internalized some globals, see if we can hack on them!
addPass(Passes, createGlobalOptimizerPass());
// Linking modules together can lead to duplicated global constants, only
// keep one copy of each constant...
addPass(Passes, createConstantMergePass());
// Remove unused arguments from functions...
addPass(Passes, createDeadArgEliminationPass());
// Reduce the code after globalopt and ipsccp. Both can open up significant
// simplification opportunities, and both can propagate functions through
// function pointers. When this happens, we often have to resolve varargs
// calls, etc, so let instcombine do this.
addPass(Passes, createInstructionCombiningPass());
if (!DisableInline)
addPass(Passes, createFunctionInliningPass()); // Inline small functions
addPass(Passes, createPruneEHPass()); // Remove dead EH info
addPass(Passes, createGlobalOptimizerPass()); // Optimize globals again.
addPass(Passes, createGlobalDCEPass()); // Remove dead functions
// If we didn't decide to inline a function, check to see if we can
// transform it to pass arguments by value instead of by reference.
addPass(Passes, createArgumentPromotionPass());
// The IPO passes may leave cruft around. Clean up after them.
addPass(Passes, createInstructionCombiningPass());
addPass(Passes, createJumpThreadingPass()); // Thread jumps.
addPass(Passes, createScalarReplAggregatesPass()); // Break up allocas
// Run a few AA driven optimizations here and now, to cleanup the code.
addPass(Passes, createFunctionAttrsPass()); // Add nocapture
addPass(Passes, createGlobalsModRefPass()); // IP alias analysis
addPass(Passes, createLICMPass()); // Hoist loop invariants
addPass(Passes, createGVNPass()); // Remove redundancies
addPass(Passes, createMemCpyOptPass()); // Remove dead memcpy's
addPass(Passes, createDeadStoreEliminationPass()); // Nuke dead stores
// Cleanup and simplify the code after the scalar optimizations.
addPass(Passes, createInstructionCombiningPass());
addPass(Passes, createJumpThreadingPass()); // Thread jumps.
addPass(Passes, createPromoteMemoryToRegisterPass()); // Cleanup jumpthread.
// Delete basic blocks, which optimization passes may have killed...
addPass(Passes, createCFGSimplificationPass());
// Now that we have optimized the program, discard unreachable functions...
addPass(Passes, createGlobalDCEPass());
}
// If the -s or -S command line options were specified, strip the symbols out
// of the resulting program to make it smaller. -s and -S are GNU ld options
// that we are supporting; they alias -strip-all and -strip-debug.
if (Strip || StripDebug)
addPass(Passes, createStripSymbolsPass(StripDebug && !Strip));
#if 0
// Create a new optimization pass for each one specified on the command line
std::auto_ptr<TargetMachine> target;
for (unsigned i = 0; i < OptimizationList.size(); ++i) {
const PassInfo *Opt = OptimizationList[i];
if (Opt->getNormalCtor())
addPass(Passes, Opt->getNormalCtor()());
else
llvm::errs() << "llvm-ld: cannot create pass: "******"\n";
}
#endif
// The user's passes may leave cruft around. Clean up after them them but
// only if we haven't got DisableOptimizations set
if (!DisableOptimizations) {
addPass(Passes, createInstructionCombiningPass());
addPass(Passes, createCFGSimplificationPass());
addPass(Passes, createAggressiveDCEPass());
addPass(Passes, createGlobalDCEPass());
}
// Make sure everything is still good.
if (!DontVerify)
Passes.add(createVerifierPass());
// Run our queue of passes all at once now, efficiently.
Passes.run(*M);
}
void BackendPass::CreatePasses(const clang::LangOptions& LangOpts,
const clang::CodeGenOptions& CodeGenOpts) {
// See clang/lib/CodeGen/BackendUtil.cpp EmitAssemblyHelper::CreatePasses()
unsigned OptLevel = CodeGenOpts.OptimizationLevel;
CodeGenOptions::InliningMethod Inlining = CodeGenOpts.getInlining();
// Handle disabling of LLVM optimization, where we want to preserve the
// internal module before any optimization.
if (CodeGenOpts.DisableLLVMOpts) {
OptLevel = 0;
Inlining = CodeGenOpts.NoInlining;
}
PassManagerBuilderWithOpts PMBuilder(LangOpts, CodeGenOpts);
PMBuilder.OptLevel = OptLevel;
PMBuilder.SizeLevel = CodeGenOpts.OptimizeSize;
PMBuilder.BBVectorize = CodeGenOpts.VectorizeBB;
PMBuilder.SLPVectorize = CodeGenOpts.VectorizeSLP;
PMBuilder.LoopVectorize = CodeGenOpts.VectorizeLoop;
PMBuilder.DisableUnitAtATime = !CodeGenOpts.UnitAtATime;
PMBuilder.DisableUnrollLoops = !CodeGenOpts.UnrollLoops;
PMBuilder.RerollLoops = CodeGenOpts.RerollLoops;
#if 0
if (!CodeGenOpts.SampleProfileFile.empty())
PMBuilder.addExtension(PassManagerBuilder::EP_EarlyAsPossible,
addSampleProfileLoaderPass);
#endif
// 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 (LangOpts.Sanitize.LocalBounds) {
PMBuilder.addExtension(PassManagerBuilder::EP_ScalarOptimizerLate,
addBoundsCheckingPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addBoundsCheckingPass);
}
if (LangOpts.Sanitize.Address) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addAddressSanitizerPasses);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addAddressSanitizerPasses);
}
if (LangOpts.Sanitize.Memory) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addMemorySanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addMemorySanitizerPass);
}
if (LangOpts.Sanitize.Thread) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addThreadSanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addThreadSanitizerPass);
}
if (LangOpts.Sanitize.DataFlow) {
PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
addDataFlowSanitizerPass);
PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
addDataFlowSanitizerPass);
}
// Figure out TargetLibraryInfo.
Triple TargetTriple(m_Module->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;
// Creates a SimpleInliner that requests InsertLifetime.
PMBuilder.Inliner
= new InlinerKeepDeadFunc(createFunctionInliningPass(Threshold));
break;
}
case CodeGenOptions::OnlyAlwaysInlining:
// Respect always_inline.
if (OptLevel == 0)
// Do not insert lifetime intrinsics at -O0.
PMBuilder.Inliner
= new InlinerKeepDeadFunc(createAlwaysInlinerPass(false));
else
PMBuilder.Inliner
= new InlinerKeepDeadFunc(createAlwaysInlinerPass());
break;
}
// Set up the per-function pass manager.
FunctionPassManager *FPM = getPerFunctionPasses();
if (CodeGenOpts.VerifyModule)
FPM->add(createVerifierPass());
PMBuilder.populateFunctionPassManager(*FPM);
// The Inliner is a module pass; register it.
PMBuilder.populateModulePassManager(*getPerModulePasses());
}
/// adopted from: llvm-2.9/include/llvm/Support/StandardPasses.h
void optimizeFunction(Function* f, const bool disableLICM, const bool disableLoopRotate) {
assert (f);
assert (f->getParent());
Module* mod = f->getParent();
TargetData* targetData = new TargetData(mod);
const unsigned OptimizationLevel = 3;
const bool OptimizeSize = false;
const bool UnitAtATime = true;
const bool UnrollLoops = true;
const bool SimplifyLibCalls = true;
const bool HaveExceptions = false;
Pass* InliningPass = createFunctionInliningPass(275);
//PassManager Passes;
FunctionPassManager Passes(mod);
Passes.add(targetData);
//
// custom
//
Passes.add(createScalarReplAggregatesPass(-1, false));
//
// createStandardFunctionPasses
//
Passes.add(createCFGSimplificationPass());
Passes.add(createPromoteMemoryToRegisterPass());
Passes.add(createInstructionCombiningPass());
// Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
// BasicAliasAnalysis wins if they disagree. This is intended to help
// support "obvious" type-punning idioms.
Passes.add(createTypeBasedAliasAnalysisPass());
Passes.add(createBasicAliasAnalysisPass());
// Start of function pass.
// Break up aggregate allocas, using SSAUpdater.
Passes.add(createScalarReplAggregatesPass(-1, false));
Passes.add(createEarlyCSEPass()); // Catch trivial redundancies
if (SimplifyLibCalls)
Passes.add(createSimplifyLibCallsPass()); // Library Call Optimizations
Passes.add(createJumpThreadingPass()); // Thread jumps.
Passes.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
Passes.add(createCFGSimplificationPass()); // Merge & remove BBs
Passes.add(createInstructionCombiningPass()); // Combine silly seq's
Passes.add(createTailCallEliminationPass()); // Eliminate tail calls
Passes.add(createCFGSimplificationPass()); // Merge & remove BBs
Passes.add(createReassociatePass()); // Reassociate expressions
if (!disableLoopRotate) Passes.add(createLoopRotatePass()); // Rotate Loop // makes packetized Mandelbrot fail
if (!disableLICM) Passes.add(createLICMPass()); // Hoist loop invariants // makes scalar driver crash after optimization
//Passes.add(createLoopUnswitchPass(OptimizeSize || OptimizationLevel < 3)); // breaks DCT with UNIFORM_ANALYSIS=0
Passes.add(createInstructionCombiningPass());
Passes.add(createIndVarSimplifyPass()); // Canonicalize indvars
Passes.add(createLoopIdiomPass()); // Recognize idioms like memset.
Passes.add(createLoopDeletionPass()); // Delete dead loops
if (UnrollLoops)
Passes.add(createLoopUnrollPass()); // Unroll small loops
Passes.add(createInstructionCombiningPass()); // Clean up after the unroller
if (OptimizationLevel > 1)
Passes.add(createGVNPass()); // Remove redundancies
Passes.add(createMemCpyOptPass()); // Remove memcpy / form memset
Passes.add(createSCCPPass()); // Constant prop with SCCP
// Run instcombine after redundancy elimination to exploit opportunities
// opened up by them.
Passes.add(createInstructionCombiningPass());
Passes.add(createJumpThreadingPass()); // Thread jumps
Passes.add(createCorrelatedValuePropagationPass());
Passes.add(createDeadStoreEliminationPass()); // Delete dead stores
Passes.add(createAggressiveDCEPass()); // Delete dead instructions
Passes.add(createCFGSimplificationPass()); // Merge & remove BBs
WFVOPENCL_DEBUG( Passes.add(createVerifierPass()); );
static void createOptimizationPasses() {
// Create and set up the per-function pass manager.
// FIXME: Move the code generator to be function-at-a-time.
PerFunctionPasses =
new FunctionPassManager(new ExistingModuleProvider(TheModule));
PerFunctionPasses->add(new TargetData(*TheTarget->getTargetData()));
// In -O0 if checking is disabled, we don't even have per-function passes.
bool HasPerFunctionPasses = false;
#ifdef ENABLE_CHECKING
PerFunctionPasses->add(createVerifierPass());
HasPerFunctionPasses = true;
#endif
if (optimize > 0 && !DisableLLVMOptimizations) {
HasPerFunctionPasses = true;
PerFunctionPasses->add(createCFGSimplificationPass());
if (optimize == 1)
PerFunctionPasses->add(createPromoteMemoryToRegisterPass());
else
PerFunctionPasses->add(createScalarReplAggregatesPass());
PerFunctionPasses->add(createInstructionCombiningPass());
// PerFunctionPasses->add(createCFGSimplificationPass());
}
// FIXME: AT -O0/O1, we should stream out functions at a time.
PerModulePasses = new PassManager();
PerModulePasses->add(new TargetData(*TheTarget->getTargetData()));
bool HasPerModulePasses = false;
if (optimize > 0 && !DisableLLVMOptimizations) {
HasPerModulePasses = true;
PassManager *PM = PerModulePasses;
if (flag_unit_at_a_time)
PM->add(createRaiseAllocationsPass()); // call %malloc -> malloc inst
PM->add(createCFGSimplificationPass()); // Clean up disgusting code
PM->add(createPromoteMemoryToRegisterPass()); // Kill useless allocas
if (flag_unit_at_a_time) {
PM->add(createGlobalOptimizerPass()); // Optimize out global vars
PM->add(createGlobalDCEPass()); // Remove unused fns and globs
PM->add(createIPConstantPropagationPass()); // IP Constant Propagation
PM->add(createDeadArgEliminationPass()); // Dead argument elimination
}
PM->add(createInstructionCombiningPass()); // Clean up after IPCP & DAE
// DISABLE PREDSIMPLIFY UNTIL PR967 is fixed.
//PM->add(createPredicateSimplifierPass()); // Canonicalize registers
PM->add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
if (flag_unit_at_a_time)
PM->add(createPruneEHPass()); // Remove dead EH info
if (optimize > 1) {
if (flag_inline_trees > 1) // respect -fno-inline-functions
PM->add(createFunctionInliningPass()); // Inline small functions
if (flag_unit_at_a_time && !lang_hooks.flag_no_builtin())
PM->add(createSimplifyLibCallsPass()); // Library Call Optimizations
if (optimize > 2)
PM->add(createArgumentPromotionPass()); // Scalarize uninlined fn args
}
PM->add(createTailDuplicationPass()); // Simplify cfg by copying code
PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
PM->add(createCFGSimplificationPass()); // Merge & remove BBs
PM->add(createScalarReplAggregatesPass()); // Break up aggregate allocas
PM->add(createInstructionCombiningPass()); // Combine silly seq's
PM->add(createCondPropagationPass()); // Propagate conditionals
PM->add(createTailCallEliminationPass()); // Eliminate tail calls
PM->add(createCFGSimplificationPass()); // Merge & remove BBs
PM->add(createReassociatePass()); // Reassociate expressions
PM->add(createLoopRotatePass()); // Rotate Loop
PM->add(createLICMPass()); // Hoist loop invariants
PM->add(createLoopUnswitchPass(optimize_size ? true : false));
PM->add(createInstructionCombiningPass()); // Clean up after LICM/reassoc
PM->add(createIndVarSimplifyPass()); // Canonicalize indvars
if (flag_unroll_loops)
PM->add(createLoopUnrollPass()); // Unroll small loops
PM->add(createInstructionCombiningPass()); // Clean up after the unroller
PM->add(createLoadValueNumberingPass()); // GVN for load instructions
PM->add(createGCSEPass()); // Remove common subexprs
PM->add(createSCCPPass()); // Constant prop with SCCP
// Run instcombine after redundancy elimination to exploit opportunities
// opened up by them.
PM->add(createInstructionCombiningPass());
PM->add(createCondPropagationPass()); // Propagate conditionals
PM->add(createDeadStoreEliminationPass()); // Delete dead stores
PM->add(createAggressiveDCEPass()); // SSA based 'Aggressive DCE'
PM->add(createCFGSimplificationPass()); // Merge & remove BBs
if (optimize > 1 && flag_unit_at_a_time)
PM->add(createConstantMergePass()); // Merge dup global constants
PM->add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
}
if (emit_llvm_bc) {
// Emit an LLVM .bc file to the output. This is used when passed
// -emit-llvm -c to the GCC driver.
PerModulePasses->add(CreateBitcodeWriterPass(*AsmOutStream));
// Disable emission of .ident into the output file... which is completely
// wrong for llvm/.bc emission cases.
flag_no_ident = 1;
HasPerModulePasses = true;
} else if (emit_llvm) {
// Emit an LLVM .ll file to the output. This is used when passed
// -emit-llvm -S to the GCC driver.
PerModulePasses->add(new PrintModulePass(AsmOutFile));
// Disable emission of .ident into the output file... which is completely
// wrong for llvm/.bc emission cases.
flag_no_ident = 1;
HasPerModulePasses = true;
} else {
FunctionPassManager *PM;
// If there are passes we have to run on the entire module, we do codegen
// as a separate "pass" after that happens.
// FIXME: This is disabled right now until bugs can be worked out. Reenable
// this for fast -O0 compiles!
if (HasPerModulePasses || 1) {
CodeGenPasses = PM =
new FunctionPassManager(new ExistingModuleProvider(TheModule));
PM->add(new TargetData(*TheTarget->getTargetData()));
} else {
// If there are no module-level passes that have to be run, we codegen as
// each function is parsed.
PM = PerFunctionPasses;
HasPerFunctionPasses = true;
}
// Normal mode, emit a .s file by running the code generator.
switch (TheTarget->addPassesToEmitFile(*PM, *AsmOutStream,
TargetMachine::AssemblyFile,
/*FAST*/optimize == 0)) {
default:
case FileModel::Error:
cerr << "Error interfacing to target machine!\n";
exit(1);
case FileModel::AsmFile:
break;
}
if (TheTarget->addPassesToEmitFileFinish(*PM, 0, /*Fast*/optimize == 0)) {
cerr << "Error interfacing to target machine!\n";
exit(1);
}
}
if (HasPerFunctionPasses) {
PerFunctionPasses->doInitialization();
} else {
delete PerFunctionPasses;
PerFunctionPasses = 0;
}
if (!HasPerModulePasses) {
delete PerModulePasses;
PerModulePasses = 0;
}
}
