1 //===------ LoopGenerators.cpp - IR helper to create loops ---------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains functions to create scalar and parallel loops as LLVM-IR. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "polly/CodeGen/LoopGenerators.h" 15 #include "polly/ScopDetection.h" 16 #include "llvm/Analysis/LoopInfo.h" 17 #include "llvm/IR/DataLayout.h" 18 #include "llvm/IR/Dominators.h" 19 #include "llvm/IR/Module.h" 20 #include "llvm/Support/CommandLine.h" 21 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 22 23 using namespace llvm; 24 using namespace polly; 25 26 static cl::opt<int> 27 PollyNumThreads("polly-num-threads", 28 cl::desc("Number of threads to use (0 = auto)"), cl::Hidden, 29 cl::init(0)); 30 31 // We generate a loop of either of the following structures: 32 // 33 // BeforeBB BeforeBB 34 // | | 35 // v v 36 // GuardBB PreHeaderBB 37 // / | | _____ 38 // __ PreHeaderBB | v \/ | 39 // / \ / | HeaderBB latch 40 // latch HeaderBB | |\ | 41 // \ / \ / | \------/ 42 // < \ / | 43 // \ / v 44 // ExitBB ExitBB 45 // 46 // depending on whether or not we know that it is executed at least once. If 47 // not, GuardBB checks if the loop is executed at least once. If this is the 48 // case we branch to PreHeaderBB and subsequently to the HeaderBB, which 49 // contains the loop iv 'polly.indvar', the incremented loop iv 50 // 'polly.indvar_next' as well as the condition to check if we execute another 51 // iteration of the loop. After the loop has finished, we branch to ExitBB. 52 Value *polly::createLoop(Value *LB, Value *UB, Value *Stride, 53 PollyIRBuilder &Builder, Pass *P, LoopInfo &LI, 54 DominatorTree &DT, BasicBlock *&ExitBB, 55 ICmpInst::Predicate Predicate, 56 ScopAnnotator *Annotator, bool Parallel, 57 bool UseGuard) { 58 Function *F = Builder.GetInsertBlock()->getParent(); 59 LLVMContext &Context = F->getContext(); 60 61 assert(LB->getType() == UB->getType() && "Types of loop bounds do not match"); 62 IntegerType *LoopIVType = dyn_cast<IntegerType>(UB->getType()); 63 assert(LoopIVType && "UB is not integer?"); 64 assert((LoopIVType == LB->getType() && LoopIVType == Stride->getType()) && 65 "LB, UB and Stride should have equal types."); 66 67 BasicBlock *BeforeBB = Builder.GetInsertBlock(); 68 BasicBlock *GuardBB = 69 UseGuard ? BasicBlock::Create(Context, "polly.loop_if", F) : nullptr; 70 BasicBlock *HeaderBB = BasicBlock::Create(Context, "polly.loop_header", F); 71 BasicBlock *PreHeaderBB = 72 BasicBlock::Create(Context, "polly.loop_preheader", F); 73 74 // Update LoopInfo 75 Loop *OuterLoop = LI.getLoopFor(BeforeBB); 76 Loop *NewLoop = new Loop(); 77 78 if (OuterLoop) 79 OuterLoop->addChildLoop(NewLoop); 80 else 81 LI.addTopLevelLoop(NewLoop); 82 83 if (OuterLoop) { 84 if (GuardBB) 85 OuterLoop->addBasicBlockToLoop(GuardBB, LI); 86 OuterLoop->addBasicBlockToLoop(PreHeaderBB, LI); 87 } 88 89 NewLoop->addBasicBlockToLoop(HeaderBB, LI); 90 91 // Notify the annotator (if present) that we have a new loop, but only 92 // after the header block is set. 93 if (Annotator) 94 Annotator->pushLoop(NewLoop, Parallel); 95 96 // ExitBB 97 ExitBB = SplitBlock(BeforeBB, &*Builder.GetInsertPoint(), &DT, &LI); 98 ExitBB->setName("polly.loop_exit"); 99 100 // BeforeBB 101 if (GuardBB) { 102 BeforeBB->getTerminator()->setSuccessor(0, GuardBB); 103 DT.addNewBlock(GuardBB, BeforeBB); 104 105 // GuardBB 106 Builder.SetInsertPoint(GuardBB); 107 Value *LoopGuard; 108 LoopGuard = Builder.CreateICmp(Predicate, LB, UB); 109 LoopGuard->setName("polly.loop_guard"); 110 Builder.CreateCondBr(LoopGuard, PreHeaderBB, ExitBB); 111 DT.addNewBlock(PreHeaderBB, GuardBB); 112 } else { 113 BeforeBB->getTerminator()->setSuccessor(0, PreHeaderBB); 114 DT.addNewBlock(PreHeaderBB, BeforeBB); 115 } 116 117 // PreHeaderBB 118 Builder.SetInsertPoint(PreHeaderBB); 119 Builder.CreateBr(HeaderBB); 120 121 // HeaderBB 122 DT.addNewBlock(HeaderBB, PreHeaderBB); 123 Builder.SetInsertPoint(HeaderBB); 124 PHINode *IV = Builder.CreatePHI(LoopIVType, 2, "polly.indvar"); 125 IV->addIncoming(LB, PreHeaderBB); 126 Value *IncrementedIV = Builder.CreateNSWAdd(IV, Stride, "polly.indvar_next"); 127 Value *LoopCondition; 128 UB = Builder.CreateSub(UB, Stride, "polly.adjust_ub"); 129 LoopCondition = Builder.CreateICmp(Predicate, IV, UB); 130 LoopCondition->setName("polly.loop_cond"); 131 132 // Create the loop latch and annotate it as such. 133 BranchInst *B = Builder.CreateCondBr(LoopCondition, HeaderBB, ExitBB); 134 if (Annotator) 135 Annotator->annotateLoopLatch(B, NewLoop, Parallel); 136 137 IV->addIncoming(IncrementedIV, HeaderBB); 138 if (GuardBB) 139 DT.changeImmediateDominator(ExitBB, GuardBB); 140 else 141 DT.changeImmediateDominator(ExitBB, HeaderBB); 142 143 // The loop body should be added here. 144 Builder.SetInsertPoint(HeaderBB->getFirstNonPHI()); 145 return IV; 146 } 147 148 Value *ParallelLoopGenerator::createParallelLoop( 149 Value *LB, Value *UB, Value *Stride, SetVector<Value *> &UsedValues, 150 ValueMapT &Map, BasicBlock::iterator *LoopBody) { 151 152 // Adjust the types to match the GOMP API. 153 LB = Builder.CreateSExt(LB, LongType); 154 UB = Builder.CreateSExt(UB, LongType); 155 Stride = Builder.CreateSExt(Stride, LongType); 156 157 Function *SubFn; 158 159 AllocaInst *Struct = storeValuesIntoStruct(UsedValues); 160 BasicBlock::iterator BeforeLoop = Builder.GetInsertPoint(); 161 Value *IV = createSubFn(Stride, Struct, UsedValues, Map, &SubFn); 162 *LoopBody = Builder.GetInsertPoint(); 163 Builder.SetInsertPoint(&*BeforeLoop); 164 165 Value *SubFnParam = Builder.CreateBitCast(Struct, Builder.getInt8PtrTy(), 166 "polly.par.userContext"); 167 168 // Add one as the upper bound provided by openmp is a < comparison 169 // whereas the codegenForSequential function creates a <= comparison. 170 UB = Builder.CreateAdd(UB, ConstantInt::get(LongType, 1)); 171 172 // Tell the runtime we start a parallel loop 173 createCallSpawnThreads(SubFn, SubFnParam, LB, UB, Stride); 174 Builder.CreateCall(SubFn, SubFnParam); 175 createCallJoinThreads(); 176 177 // Mark the end of the lifetime for the parameter struct. 178 Type *Ty = Struct->getType(); 179 ConstantInt *SizeOf = Builder.getInt64(DL.getTypeAllocSize(Ty)); 180 Builder.CreateLifetimeEnd(Struct, SizeOf); 181 182 return IV; 183 } 184 185 void ParallelLoopGenerator::createCallSpawnThreads(Value *SubFn, 186 Value *SubFnParam, Value *LB, 187 Value *UB, Value *Stride) { 188 const std::string Name = "GOMP_parallel_loop_runtime_start"; 189 190 Function *F = M->getFunction(Name); 191 192 // If F is not available, declare it. 193 if (!F) { 194 GlobalValue::LinkageTypes Linkage = Function::ExternalLinkage; 195 196 Type *Params[] = {PointerType::getUnqual(FunctionType::get( 197 Builder.getVoidTy(), Builder.getInt8PtrTy(), false)), 198 Builder.getInt8PtrTy(), Builder.getInt32Ty(), LongType, 199 LongType, LongType}; 200 201 FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), Params, false); 202 F = Function::Create(Ty, Linkage, Name, M); 203 } 204 205 Value *NumberOfThreads = Builder.getInt32(PollyNumThreads); 206 Value *Args[] = {SubFn, SubFnParam, NumberOfThreads, LB, UB, Stride}; 207 208 Builder.CreateCall(F, Args); 209 } 210 211 Value *ParallelLoopGenerator::createCallGetWorkItem(Value *LBPtr, 212 Value *UBPtr) { 213 const std::string Name = "GOMP_loop_runtime_next"; 214 215 Function *F = M->getFunction(Name); 216 217 // If F is not available, declare it. 218 if (!F) { 219 GlobalValue::LinkageTypes Linkage = Function::ExternalLinkage; 220 Type *Params[] = {LongType->getPointerTo(), LongType->getPointerTo()}; 221 FunctionType *Ty = FunctionType::get(Builder.getInt8Ty(), Params, false); 222 F = Function::Create(Ty, Linkage, Name, M); 223 } 224 225 Value *Args[] = {LBPtr, UBPtr}; 226 Value *Return = Builder.CreateCall(F, Args); 227 Return = Builder.CreateICmpNE( 228 Return, Builder.CreateZExt(Builder.getFalse(), Return->getType())); 229 return Return; 230 } 231 232 void ParallelLoopGenerator::createCallJoinThreads() { 233 const std::string Name = "GOMP_parallel_end"; 234 235 Function *F = M->getFunction(Name); 236 237 // If F is not available, declare it. 238 if (!F) { 239 GlobalValue::LinkageTypes Linkage = Function::ExternalLinkage; 240 241 FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), false); 242 F = Function::Create(Ty, Linkage, Name, M); 243 } 244 245 Builder.CreateCall(F, {}); 246 } 247 248 void ParallelLoopGenerator::createCallCleanupThread() { 249 const std::string Name = "GOMP_loop_end_nowait"; 250 251 Function *F = M->getFunction(Name); 252 253 // If F is not available, declare it. 254 if (!F) { 255 GlobalValue::LinkageTypes Linkage = Function::ExternalLinkage; 256 257 FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), false); 258 F = Function::Create(Ty, Linkage, Name, M); 259 } 260 261 Builder.CreateCall(F, {}); 262 } 263 264 Function *ParallelLoopGenerator::createSubFnDefinition() { 265 Function *F = Builder.GetInsertBlock()->getParent(); 266 std::vector<Type *> Arguments(1, Builder.getInt8PtrTy()); 267 FunctionType *FT = FunctionType::get(Builder.getVoidTy(), Arguments, false); 268 Function *SubFn = Function::Create(FT, Function::InternalLinkage, 269 F->getName() + "_polly_subfn", M); 270 271 // Certain backends (e.g., NVPTX) do not support '.'s in function names. 272 // Hence, we ensure that all '.'s are replaced by '_'s. 273 std::string FunctionName = SubFn->getName(); 274 std::replace(FunctionName.begin(), FunctionName.end(), '.', '_'); 275 SubFn->setName(FunctionName); 276 277 // Do not run any polly pass on the new function. 278 SubFn->addFnAttr(PollySkipFnAttr); 279 280 Function::arg_iterator AI = SubFn->arg_begin(); 281 AI->setName("polly.par.userContext"); 282 283 return SubFn; 284 } 285 286 AllocaInst * 287 ParallelLoopGenerator::storeValuesIntoStruct(SetVector<Value *> &Values) { 288 SmallVector<Type *, 8> Members; 289 290 for (Value *V : Values) 291 Members.push_back(V->getType()); 292 293 // We do not want to allocate the alloca inside any loop, thus we allocate it 294 // in the entry block of the function and use annotations to denote the actual 295 // live span (similar to clang). 296 BasicBlock &EntryBB = Builder.GetInsertBlock()->getParent()->getEntryBlock(); 297 Instruction *IP = &*EntryBB.getFirstInsertionPt(); 298 StructType *Ty = StructType::get(Builder.getContext(), Members); 299 AllocaInst *Struct = new AllocaInst(Ty, 0, "polly.par.userContext", IP); 300 301 // Mark the start of the lifetime for the parameter struct. 302 ConstantInt *SizeOf = Builder.getInt64(DL.getTypeAllocSize(Ty)); 303 Builder.CreateLifetimeStart(Struct, SizeOf); 304 305 for (unsigned i = 0; i < Values.size(); i++) { 306 Value *Address = Builder.CreateStructGEP(Ty, Struct, i); 307 Address->setName("polly.subfn.storeaddr." + Values[i]->getName()); 308 Builder.CreateStore(Values[i], Address); 309 } 310 311 return Struct; 312 } 313 314 void ParallelLoopGenerator::extractValuesFromStruct( 315 SetVector<Value *> OldValues, Type *Ty, Value *Struct, ValueMapT &Map) { 316 for (unsigned i = 0; i < OldValues.size(); i++) { 317 Value *Address = Builder.CreateStructGEP(Ty, Struct, i); 318 Value *NewValue = Builder.CreateLoad(Address); 319 NewValue->setName("polly.subfunc.arg." + OldValues[i]->getName()); 320 Map[OldValues[i]] = NewValue; 321 } 322 } 323 324 Value *ParallelLoopGenerator::createSubFn(Value *Stride, AllocaInst *StructData, 325 SetVector<Value *> Data, 326 ValueMapT &Map, Function **SubFnPtr) { 327 BasicBlock *PrevBB, *HeaderBB, *ExitBB, *CheckNextBB, *PreHeaderBB, *AfterBB; 328 Value *LBPtr, *UBPtr, *UserContext, *Ret1, *HasNextSchedule, *LB, *UB, *IV; 329 Function *SubFn = createSubFnDefinition(); 330 LLVMContext &Context = SubFn->getContext(); 331 332 // Store the previous basic block. 333 PrevBB = Builder.GetInsertBlock(); 334 335 // Create basic blocks. 336 HeaderBB = BasicBlock::Create(Context, "polly.par.setup", SubFn); 337 ExitBB = BasicBlock::Create(Context, "polly.par.exit", SubFn); 338 CheckNextBB = BasicBlock::Create(Context, "polly.par.checkNext", SubFn); 339 PreHeaderBB = BasicBlock::Create(Context, "polly.par.loadIVBounds", SubFn); 340 341 DT.addNewBlock(HeaderBB, PrevBB); 342 DT.addNewBlock(ExitBB, HeaderBB); 343 DT.addNewBlock(CheckNextBB, HeaderBB); 344 DT.addNewBlock(PreHeaderBB, HeaderBB); 345 346 // Fill up basic block HeaderBB. 347 Builder.SetInsertPoint(HeaderBB); 348 LBPtr = Builder.CreateAlloca(LongType, 0, "polly.par.LBPtr"); 349 UBPtr = Builder.CreateAlloca(LongType, 0, "polly.par.UBPtr"); 350 UserContext = Builder.CreateBitCast( 351 &*SubFn->arg_begin(), StructData->getType(), "polly.par.userContext"); 352 353 extractValuesFromStruct(Data, StructData->getAllocatedType(), UserContext, 354 Map); 355 Builder.CreateBr(CheckNextBB); 356 357 // Add code to check if another set of iterations will be executed. 358 Builder.SetInsertPoint(CheckNextBB); 359 Ret1 = createCallGetWorkItem(LBPtr, UBPtr); 360 HasNextSchedule = Builder.CreateTrunc(Ret1, Builder.getInt1Ty(), 361 "polly.par.hasNextScheduleBlock"); 362 Builder.CreateCondBr(HasNextSchedule, PreHeaderBB, ExitBB); 363 364 // Add code to load the iv bounds for this set of iterations. 365 Builder.SetInsertPoint(PreHeaderBB); 366 LB = Builder.CreateLoad(LBPtr, "polly.par.LB"); 367 UB = Builder.CreateLoad(UBPtr, "polly.par.UB"); 368 369 // Subtract one as the upper bound provided by openmp is a < comparison 370 // whereas the codegenForSequential function creates a <= comparison. 371 UB = Builder.CreateSub(UB, ConstantInt::get(LongType, 1), 372 "polly.par.UBAdjusted"); 373 374 Builder.CreateBr(CheckNextBB); 375 Builder.SetInsertPoint(&*--Builder.GetInsertPoint()); 376 IV = createLoop(LB, UB, Stride, Builder, P, LI, DT, AfterBB, 377 ICmpInst::ICMP_SLE, nullptr, true, /* UseGuard */ false); 378 379 BasicBlock::iterator LoopBody = Builder.GetInsertPoint(); 380 381 // Add code to terminate this subfunction. 382 Builder.SetInsertPoint(ExitBB); 383 createCallCleanupThread(); 384 Builder.CreateRetVoid(); 385 386 Builder.SetInsertPoint(&*LoopBody); 387 *SubFnPtr = SubFn; 388 389 return IV; 390 } 391