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
}
/// initialize_module - perform setup of the LLVM code generation system.
void BrainFTraceRecorder::initialize_module() {
LLVMContext &Context = module->getContext();
// Initialize the code generator, and enable aggressive code generation.
InitializeNativeTarget();
EngineBuilder builder(module);
builder.setOptLevel(CodeGenOpt::Aggressive);
EE = builder.create();
// Create a FunctionPassManager to handle running optimization passes
// on our generated code. Setup a basic suite of optimizations for it.
FPM = new llvm::FunctionPassManager(module);
FPM->add(createInstructionCombiningPass());
FPM->add(createCFGSimplificationPass());
FPM->add(createScalarReplAggregatesPass());
FPM->add(createSimplifyLibCallsPass());
FPM->add(createInstructionCombiningPass());
FPM->add(createJumpThreadingPass());
FPM->add(createCFGSimplificationPass());
FPM->add(createInstructionCombiningPass());
FPM->add(createCFGSimplificationPass());
FPM->add(createReassociatePass());
FPM->add(createLoopRotatePass());
FPM->add(createLICMPass());
FPM->add(createLoopUnswitchPass(false));
FPM->add(createInstructionCombiningPass());
FPM->add(createIndVarSimplifyPass());
FPM->add(createLoopDeletionPass());
FPM->add(createLoopUnrollPass());
FPM->add(createInstructionCombiningPass());
FPM->add(createGVNPass());
FPM->add(createSCCPPass());
FPM->add(createInstructionCombiningPass());
FPM->add(createJumpThreadingPass());
FPM->add(createDeadStoreEliminationPass());
FPM->add(createAggressiveDCEPass());
FPM->add(createCFGSimplificationPass());
// Cache the LLVM type signature of an opcode function
int_type = sizeof(size_t) == 4 ?
IntegerType::getInt32Ty(Context) :
IntegerType::getInt64Ty(Context);
const Type *data_type =
PointerType::getUnqual(IntegerType::getInt8Ty(Context));
std::vector<const Type*> args;
args.push_back(int_type);
args.push_back(data_type);
op_type =
FunctionType::get(Type::getVoidTy(Context), args, false);
// Setup a global variable in the LLVM module to represent the bytecode
// array. Bind it to the actual bytecode array at JIT time.
const Type *bytecode_type = PointerType::getUnqual(op_type);
bytecode_array = cast<GlobalValue>(module->
getOrInsertGlobal("BytecodeArray", bytecode_type));
EE->addGlobalMapping(bytecode_array, BytecodeArray);
// Setup a similar mapping for the global mode flag.
const IntegerType *flag_type = IntegerType::get(Context, 8);
mode_flag =
cast<GlobalValue>(module->getOrInsertGlobal("mode", flag_type));
EE->addGlobalMapping(mode_flag, &mode);
// Setup a similar mapping for the global extension root flag.
ext_root =
cast<GlobalValue>(module->getOrInsertGlobal("ext_root", int_type));
EE->addGlobalMapping(ext_root, &extension_root);
// Setup a similar mapping for the global extension leaf flag.
ext_leaf =
cast<GlobalValue>(module->getOrInsertGlobal("ext_leaf", int_type));
EE->addGlobalMapping(ext_leaf, &extension_leaf);
// Cache LLVM declarations for putchar() and getchar().
const Type *int_type = sizeof(int) == 4 ? IntegerType::getInt32Ty(Context)
: IntegerType::getInt64Ty(Context);
putchar_func =
module->getOrInsertFunction("putchar", int_type, int_type, NULL);
getchar_func = module->getOrInsertFunction("getchar", int_type, NULL);
}
/// 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;
}
}
