Function* Create(const STREAMOUT_COMPILE_STATE& state)
{
static std::size_t soNum = 0;
std::stringstream fnName("SOShader", std::ios_base::in | std::ios_base::out | std::ios_base::ate);
fnName << soNum++;
// SO function signature
// typedef void(__cdecl *PFN_SO_FUNC)(SWR_STREAMOUT_CONTEXT*)
std::vector<Type*> args{
PointerType::get(Gen_SWR_STREAMOUT_CONTEXT(JM()), 0), // SWR_STREAMOUT_CONTEXT*
};
FunctionType* fTy = FunctionType::get(IRB()->getVoidTy(), args, false);
Function* soFunc = Function::Create(fTy, GlobalValue::ExternalLinkage, fnName.str(), JM()->mpCurrentModule);
// create return basic block
BasicBlock* entry = BasicBlock::Create(JM()->mContext, "entry", soFunc);
BasicBlock* returnBB = BasicBlock::Create(JM()->mContext, "return", soFunc);
IRB()->SetInsertPoint(entry);
// arguments
auto argitr = soFunc->getArgumentList().begin();
Value* pSoCtx = &*argitr++;
pSoCtx->setName("pSoCtx");
const STREAMOUT_STREAM& streamState = state.stream;
buildStream(state, streamState, pSoCtx, returnBB, soFunc);
BR(returnBB);
IRB()->SetInsertPoint(returnBB);
RET_VOID();
JitManager::DumpToFile(soFunc, "SoFunc");
::FunctionPassManager passes(JM()->mpCurrentModule);
passes.add(createBreakCriticalEdgesPass());
passes.add(createCFGSimplificationPass());
passes.add(createEarlyCSEPass());
passes.add(createPromoteMemoryToRegisterPass());
passes.add(createCFGSimplificationPass());
passes.add(createEarlyCSEPass());
passes.add(createInstructionCombiningPass());
passes.add(createInstructionSimplifierPass());
passes.add(createConstantPropagationPass());
passes.add(createSCCPPass());
passes.add(createAggressiveDCEPass());
passes.run(*soFunc);
JitManager::DumpToFile(soFunc, "SoFunc_optimized");
return soFunc;
}
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
}
/// 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()); );