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 addOptimizationPasses(T *PM)
{
#ifdef JL_DEBUG_BUILD
PM->add(createVerifierPass());
#endif
#ifdef __has_feature
# if __has_feature(address_sanitizer)
# if defined(LLVM37) && !defined(LLVM38)
// LLVM 3.7 BUG: ASAN pass doesn't properly initialize its dependencies
initializeTargetLibraryInfoWrapperPassPass(*PassRegistry::getPassRegistry());
# endif
PM->add(createAddressSanitizerFunctionPass());
# endif
# if __has_feature(memory_sanitizer)
PM->add(llvm::createMemorySanitizerPass(true));
# endif
#endif
if (jl_options.opt_level <= 1) {
return;
}
#ifdef LLVM37
PM->add(createTargetTransformInfoWrapperPass(jl_TargetMachine->getTargetIRAnalysis()));
#else
jl_TargetMachine->addAnalysisPasses(*PM);
#endif
#ifdef LLVM38
PM->add(createTypeBasedAAWrapperPass());
#else
PM->add(createTypeBasedAliasAnalysisPass());
#endif
if (jl_options.opt_level >= 3) {
#ifdef LLVM38
PM->add(createBasicAAWrapperPass());
#else
PM->add(createBasicAliasAnalysisPass());
#endif
}
// list of passes from vmkit
PM->add(createCFGSimplificationPass()); // Clean up disgusting code
PM->add(createPromoteMemoryToRegisterPass());// Kill useless allocas
#ifndef INSTCOMBINE_BUG
PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
#endif
PM->add(createSROAPass()); // Break up aggregate allocas
#ifndef INSTCOMBINE_BUG
PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
#endif
PM->add(createJumpThreadingPass()); // Thread jumps.
// NOTE: CFG simp passes after this point seem to hurt native codegen.
// See issue #6112. Should be re-evaluated when we switch to MCJIT.
//PM->add(createCFGSimplificationPass()); // Merge & remove BBs
#ifndef INSTCOMBINE_BUG
PM->add(createInstructionCombiningPass()); // Combine silly seq's
#endif
//PM->add(createCFGSimplificationPass()); // Merge & remove BBs
PM->add(createReassociatePass()); // Reassociate expressions
// this has the potential to make some things a bit slower
//PM->add(createBBVectorizePass());
PM->add(createEarlyCSEPass()); //// ****
PM->add(createLoopIdiomPass()); //// ****
PM->add(createLoopRotatePass()); // Rotate loops.
// LoopRotate strips metadata from terminator, so run LowerSIMD afterwards
PM->add(createLowerSimdLoopPass()); // Annotate loop marked with "simdloop" as LLVM parallel loop
PM->add(createLICMPass()); // Hoist loop invariants
PM->add(createLoopUnswitchPass()); // Unswitch loops.
// Subsequent passes not stripping metadata from terminator
#ifndef INSTCOMBINE_BUG
PM->add(createInstructionCombiningPass());
#endif
PM->add(createIndVarSimplifyPass()); // Canonicalize indvars
PM->add(createLoopDeletionPass()); // Delete dead loops
#if defined(LLVM35)
PM->add(createSimpleLoopUnrollPass()); // Unroll small loops
#else
PM->add(createLoopUnrollPass()); // Unroll small loops
#endif
#if !defined(LLVM35) && !defined(INSTCOMBINE_BUG)
PM->add(createLoopVectorizePass()); // Vectorize loops
#endif
//PM->add(createLoopStrengthReducePass()); // (jwb added)
#ifndef INSTCOMBINE_BUG
PM->add(createInstructionCombiningPass()); // Clean up after the unroller
#endif
PM->add(createGVNPass()); // Remove redundancies
//PM->add(createMemCpyOptPass()); // Remove memcpy / form memset
PM->add(createSCCPPass()); // Constant prop with SCCP
// Run instcombine after redundancy elimination to exploit opportunities
// opened up by them.
PM->add(createSinkingPass()); ////////////// ****
PM->add(createInstructionSimplifierPass());///////// ****
#ifndef INSTCOMBINE_BUG
PM->add(createInstructionCombiningPass());
#endif
PM->add(createJumpThreadingPass()); // Thread jumps
PM->add(createDeadStoreEliminationPass()); // Delete dead stores
#if !defined(INSTCOMBINE_BUG)
if (jl_options.opt_level >= 3) {
#ifdef LLVM39
initializeDemandedBitsPass(*PassRegistry::getPassRegistry());
#endif
PM->add(createSLPVectorizerPass()); // Vectorize straight-line code
}
#endif
PM->add(createAggressiveDCEPass()); // Delete dead instructions
#if !defined(INSTCOMBINE_BUG)
if (jl_options.opt_level >= 3)
PM->add(createInstructionCombiningPass()); // Clean up after SLP loop vectorizer
#endif
#if defined(LLVM35)
PM->add(createLoopVectorizePass()); // Vectorize loops
PM->add(createInstructionCombiningPass()); // Clean up after loop vectorizer
#endif
//PM->add(createCFGSimplificationPass()); // Merge & remove BBs
}
/// 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()); );
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;
}
