Context::Context(LLVMState* ls)
: ls_(ls)
, root_info_(0)
, inlined_block_(false)
, inline_depth_(0)
, rds_(new jit::RuntimeDataHolder)
, function_(0)
, state_(0)
, out_args_(0)
, counter_(0)
{
VoidTy = Type::getVoidTy(ctx_);
Int1Ty = Type::getInt1Ty(ctx_);
Int8Ty = Type::getInt8Ty(ctx_);
Int16Ty = Type::getInt16Ty(ctx_);
Int32Ty = Type::getInt32Ty(ctx_);
Int64Ty = Type::getInt64Ty(ctx_);
#ifdef IS_X8664
IntPtrTy = Int64Ty;
#else
IntPtrTy = Int32Ty;
#endif
VoidPtrTy = llvm::PointerType::getUnqual(Int8Ty);
FloatTy = Type::getFloatTy(ctx_);
DoubleTy = Type::getDoubleTy(ctx_);
Int8PtrTy = llvm::PointerType::getUnqual(Int8Ty);
Zero = llvm::ConstantInt::get(Int32Ty, 0);
One = llvm::ConstantInt::get(Int32Ty, 1);
module_ = new llvm::Module("rubinius", ctx_);
autogen_types::makeLLVMModuleContents(module_);
llvm::EngineBuilder factory(module_);
std::string error;
factory.setAllocateGVsWithCode(false);
memory_ = new jit::RubiniusRequestJITMemoryManager(ls->memory());
factory.setJITMemoryManager(memory_);
factory.setEngineKind(EngineKind::JIT);
factory.setErrorStr(&error);
#if RBX_LLVM_API_VER > 300
llvm::TargetOptions opts;
opts.NoFramePointerElim = true;
opts.NoFramePointerElimNonLeaf = true;
opts.JITEmitDebugInfo = true;
factory.setTargetOptions(opts);
#endif
factory.setMCPU(ls_->cpu());
engine_ = factory.create();
if(!engine_) {
std::cerr << "Error setting up LLVM Execution Engine: "<< error << std::endl;
rubinius::bug("error configuring LLVM");
}
if(ls_->jit_event_listener()) {
engine_->RegisterJITEventListener(ls_->jit_event_listener());
}
builder_ = new llvm::PassManagerBuilder();
builder_->OptLevel = 2;
passes_ = new llvm::FunctionPassManager(module_);
#if RBX_LLVM_API_VER >= 302
module_->setDataLayout(engine_->getDataLayout()->getStringRepresentation());
passes_->add(new llvm::DataLayout(*engine_->getDataLayout()));
#else
module_->setDataLayout(engine_->getTargetData()->getStringRepresentation());
passes_->add(new llvm::TargetData(*engine_->getTargetData()));
#endif
builder_->populateFunctionPassManager(*passes_);
// Eliminate unnecessary alloca.
passes_->add(createPromoteMemoryToRegisterPass());
// Do simple "peephole" optimizations and bit-twiddling optzns.
passes_->add(createInstructionCombiningPass());
// Reassociate expressions.
passes_->add(createReassociatePass());
// Eliminate Common SubExpressions.
passes_->add(createGVNPass());
passes_->add(createDeadStoreEliminationPass());
passes_->add(createInstructionCombiningPass());
// Simplify the control flow graph (deleting unreachable blocks, etc).
passes_->add(createCFGSimplificationPass());
passes_->add(create_rubinius_alias_analysis());
passes_->add(createGVNPass());
// passes_->add(createCFGSimplificationPass());
passes_->add(createDeadStoreEliminationPass());
// passes_->add(createVerifierPass());
passes_->add(createScalarReplAggregatesPass());
passes_->add(create_overflow_folding_pass());
passes_->add(create_guard_eliminator_pass());
passes_->add(createCFGSimplificationPass());
passes_->add(createInstructionCombiningPass());
passes_->add(createScalarReplAggregatesPass());
passes_->add(createDeadStoreEliminationPass());
passes_->add(createCFGSimplificationPass());
passes_->add(createInstructionCombiningPass());
passes_->doInitialization();
ObjTy = ptr_type("Object");
profiling_ = new GlobalVariable(
*module_, Int1Ty, false,
GlobalVariable::ExternalLinkage,
0, "profiling_flag");
metadata_id_ = ctx_.getMDKindID("rbx-classid");
}
/// 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) {
// 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());
// Add an appropriate TargetData instance for this module...
#if LLVM_VERSION_CODE >= LLVM_VERSION(3, 1)
addPass(Passes, new DataLayout(M));
#else
addPass(Passes, new TargetData(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)
addPass(Passes, createInternalizePass());
#else
addPass(Passes, createInternalizePass(true));
#endif
// 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
std::cerr << "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);
}
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;
#if LLVM_VERSION_CODE < LLVM_VERSION(2, 7)
addPass(PM, createRaiseAllocationsPass()); // call %malloc -> malloc inst
#endif
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
addPass(PM, createSimplifyLibCallsPass()); // Library Call Optimizations
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
#if LLVM_VERSION_CODE < LLVM_VERSION(2, 7)
addPass(PM, createCondPropagationPass()); // Propagate conditionals
#endif
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.
#if LLVM_VERSION_CODE < LLVM_VERSION(2, 9)
addPass(PM, createLoopIndexSplitPass()); // Index split loops.
#endif
// 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());
#if LLVM_VERSION_CODE < LLVM_VERSION(2, 7)
addPass(PM, createCondPropagationPass()); // Propagate conditionals
#endif
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
}
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;
}
