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
}
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;
}
}
/// 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()); );
Function * futamurize( const Function * orig_func, DenseMap<const Value*, Value*> &argmap, std::set<const unsigned char *> &constant_addresses_set )
{
LLVMContext &context = getGlobalContext();
// Make a copy of the function, removing constant arguments
Function * specialized_func = CloneFunction( orig_func, argmap );
specialized_func->setName( orig_func->getNameStr() + "_1" );
// add it to our module
LLVM_Module->getFunctionList().push_back( specialized_func );
printf("\nspecialized_func = %p <%s>\n", specialized_func, specialized_func->getName().data());
//~ specialized_func->dump();
// Optimize it
FunctionPassManager PM( LLVM_Module );
createStandardFunctionPasses( &PM, 3 );
PM.add(createScalarReplAggregatesPass()); // Break up aggregate allocas
PM.add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
PM.add(createJumpThreadingPass()); // Thread jumps.
PM.add(createCFGSimplificationPass()); // Merge & remove BBs
PM.add(createInstructionCombiningPass()); // Combine silly seq's
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( false ));
PM.add(createInstructionCombiningPass());
PM.add(createIndVarSimplifyPass()); // Canonicalize indvars
PM.add(createLoopDeletionPass()); // Delete dead loops
PM.add(createLoopUnroll2Pass()); // Unroll small loops
PM.add(createInstructionCombiningPass()); // Clean up after the unroller
PM.add(createGVNPass()); // Remove redundancies
PM.add(createMemCpyOptPass()); // Remove memcpy / form memset
PM.add(createSCCPPass()); // Constant prop with SCCP
PM.add(createPromoteMemoryToRegisterPass());
PM.add(createConstantPropagationPass());
PM.add(createDeadStoreEliminationPass());
PM.add(createAggressiveDCEPass());
PM.add(new MemoryDependenceAnalysis());
//~ PM.add(createAAEvalPass());
const PassInfo * pinfo = Pass::lookupPassInfo( "print-alias-sets" );
if( !pinfo ) { printf( "print-alias-sets not found\n" ); exit(-1); }
PM.add( pinfo->createPass() );
FunctionPassManager PM_Inline( LLVM_Module );
PM_Inline.add(createSingleFunctionInliningPass());
bool Changed = false;
int iterations = 2;
int inline_iterations = 6;
do
{
Changed = false;
// first do some optimizations
PM.doInitialization();
PM.run( *specialized_func );
PM.doFinalization();
// Load from Constant Memory detection
const TargetData *TD = LLVM_EE->getTargetData();
for (inst_iterator I = inst_begin(specialized_func), E = inst_end(specialized_func); I != E; ++I)
{
Instruction * inst = (Instruction *) &*I;
// get all Load instructions
LoadInst * load = dyn_cast<LoadInst>( inst );
if( !load ) continue;
if( load->isVolatile() ) continue;
if (load->use_empty()) continue; // Don't muck with dead instructions...
// get the address loaded by load instruction
Value *ptr_value = load->getPointerOperand();
// we're only interested in constant addresses
ConstantExpr * ptr_constant_expr = dyn_cast<ConstantExpr>( ptr_value );
if( !ptr_constant_expr ) continue;
ptr_constant_expr->dump();
// compute real address of constant pointer expression
Constant * ptr_constant = ConstantFoldConstantExpression( ptr_constant_expr, TD );
if( !ptr_constant ) continue;
ptr_constant->dump();
// convert to int constant
ConstantInt *int_constant = dyn_cast<ConstantInt>( ConstantExpr::getPtrToInt( ptr_constant, Type::getInt64Ty( context )));
if( !int_constant ) continue;
int_constant->dump();
// get data size
int data_length = TD->getTypeAllocSize( load->getType() );
ptr_value->getType()->dump();
// get real address (at last !)
const unsigned char * c_ptr = (const unsigned char *) int_constant->getLimitedValue();
printf( "%ld %d %d\n", c_ptr, constant_addresses_set.count( c_ptr ), data_length );
// check what's in this address
int isconst = 1;
for( int offset=0; offset<data_length; offset++ )
isconst &= constant_addresses_set.count( c_ptr + offset );
if( !isconst ) continue;
printf( "It is constant.\n" );
// make a LLVM const with the data
Constant *new_constant = NULL;
switch( data_length )
{
case 1: new_constant = ConstantInt::get( Type::getInt8Ty( context ), *(uint8_t*)c_ptr, false /* signed */ ); break;
case 2: new_constant = ConstantInt::get( Type::getInt16Ty( context ), *(uint16_t*)c_ptr, false /* signed */ ); break;
case 4: new_constant = ConstantInt::get( Type::getInt32Ty( context ), *(uint32_t*)c_ptr, false /* signed */ ); break;
case 8: new_constant = ConstantInt::get( Type::getInt64Ty( context ), *(uint64_t*)c_ptr, false /* signed */ ); break;
default:
{
StringRef const_data ( (const char *) c_ptr, data_length );
new_constant = ConstantArray::get( context, const_data, false /* dont add terminating null */ );
}
}
if( !new_constant ) continue;
new_constant->dump();
//~ // get the type that is loaded
const Type *Ty = load->getType();
// do we need a cast ?
if( load->getType() != new_constant->getType() )
{
new_constant = ConstantExpr::getBitCast( new_constant, Ty );
new_constant->dump();
}
// zap the load and replace with constant address
load->replaceAllUsesWith( new_constant );
printf( "\nREPLACED :...\n" );
load->dump();
new_constant->dump();
Changed = true;
}
if( Changed )
continue; // re-optimize and do another pass of constant load elimination
// if we can't do anything else, do an inlining pass
if( inline_iterations > 0 )
{
inline_iterations --;
PM_Inline.doInitialization();
Changed |= PM_Inline.run( *specialized_func );
PM_Inline.doFinalization();
//~ for( int i=0; i<3; i++ )
{
PM.doInitialization();
Changed |= PM.run( *specialized_func );
PM.doFinalization();
}
}
if( iterations>0 && !Changed )
iterations--;
} while( Changed || iterations>0 );
return specialized_func;
}