1 //===- OpenMPIRBuilder.cpp - Builder for LLVM-IR for OpenMP directives ----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 ///
10 /// This file implements the OpenMPIRBuilder class, which is used as a
11 /// convenient way to create LLVM instructions for OpenMP directives.
12 ///
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
16 #include "llvm/ADT/StringRef.h"
17 #include "llvm/ADT/Triple.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CodeMetrics.h"
20 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
21 #include "llvm/Analysis/ScalarEvolution.h"
22 #include "llvm/Analysis/TargetLibraryInfo.h"
23 #include "llvm/IR/CFG.h"
24 #include "llvm/IR/DebugInfo.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/MDBuilder.h"
27 #include "llvm/IR/PassManager.h"
28 #include "llvm/IR/Value.h"
29 #include "llvm/MC/TargetRegistry.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Error.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
35 #include "llvm/Transforms/Utils/CodeExtractor.h"
36 #include "llvm/Transforms/Utils/LoopPeel.h"
37 #include "llvm/Transforms/Utils/ModuleUtils.h"
38 #include "llvm/Transforms/Utils/UnrollLoop.h"
39
40 #include <sstream>
41
42 #define DEBUG_TYPE "openmp-ir-builder"
43
44 using namespace llvm;
45 using namespace omp;
46
47 static cl::opt<bool>
48 OptimisticAttributes("openmp-ir-builder-optimistic-attributes", cl::Hidden,
49 cl::desc("Use optimistic attributes describing "
50 "'as-if' properties of runtime calls."),
51 cl::init(false));
52
53 static cl::opt<double> UnrollThresholdFactor(
54 "openmp-ir-builder-unroll-threshold-factor", cl::Hidden,
55 cl::desc("Factor for the unroll threshold to account for code "
56 "simplifications still taking place"),
57 cl::init(1.5));
58
addAttributes(omp::RuntimeFunction FnID,Function & Fn)59 void OpenMPIRBuilder::addAttributes(omp::RuntimeFunction FnID, Function &Fn) {
60 LLVMContext &Ctx = Fn.getContext();
61
62 // Get the function's current attributes.
63 auto Attrs = Fn.getAttributes();
64 auto FnAttrs = Attrs.getFnAttrs();
65 auto RetAttrs = Attrs.getRetAttrs();
66 SmallVector<AttributeSet, 4> ArgAttrs;
67 for (size_t ArgNo = 0; ArgNo < Fn.arg_size(); ++ArgNo)
68 ArgAttrs.emplace_back(Attrs.getParamAttrs(ArgNo));
69
70 #define OMP_ATTRS_SET(VarName, AttrSet) AttributeSet VarName = AttrSet;
71 #include "llvm/Frontend/OpenMP/OMPKinds.def"
72
73 // Add attributes to the function declaration.
74 switch (FnID) {
75 #define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets) \
76 case Enum: \
77 FnAttrs = FnAttrs.addAttributes(Ctx, FnAttrSet); \
78 RetAttrs = RetAttrs.addAttributes(Ctx, RetAttrSet); \
79 for (size_t ArgNo = 0; ArgNo < ArgAttrSets.size(); ++ArgNo) \
80 ArgAttrs[ArgNo] = \
81 ArgAttrs[ArgNo].addAttributes(Ctx, ArgAttrSets[ArgNo]); \
82 Fn.setAttributes(AttributeList::get(Ctx, FnAttrs, RetAttrs, ArgAttrs)); \
83 break;
84 #include "llvm/Frontend/OpenMP/OMPKinds.def"
85 default:
86 // Attributes are optional.
87 break;
88 }
89 }
90
91 FunctionCallee
getOrCreateRuntimeFunction(Module & M,RuntimeFunction FnID)92 OpenMPIRBuilder::getOrCreateRuntimeFunction(Module &M, RuntimeFunction FnID) {
93 FunctionType *FnTy = nullptr;
94 Function *Fn = nullptr;
95
96 // Try to find the declation in the module first.
97 switch (FnID) {
98 #define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...) \
99 case Enum: \
100 FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__}, \
101 IsVarArg); \
102 Fn = M.getFunction(Str); \
103 break;
104 #include "llvm/Frontend/OpenMP/OMPKinds.def"
105 }
106
107 if (!Fn) {
108 // Create a new declaration if we need one.
109 switch (FnID) {
110 #define OMP_RTL(Enum, Str, ...) \
111 case Enum: \
112 Fn = Function::Create(FnTy, GlobalValue::ExternalLinkage, Str, M); \
113 break;
114 #include "llvm/Frontend/OpenMP/OMPKinds.def"
115 }
116
117 // Add information if the runtime function takes a callback function
118 if (FnID == OMPRTL___kmpc_fork_call || FnID == OMPRTL___kmpc_fork_teams) {
119 if (!Fn->hasMetadata(LLVMContext::MD_callback)) {
120 LLVMContext &Ctx = Fn->getContext();
121 MDBuilder MDB(Ctx);
122 // Annotate the callback behavior of the runtime function:
123 // - The callback callee is argument number 2 (microtask).
124 // - The first two arguments of the callback callee are unknown (-1).
125 // - All variadic arguments to the runtime function are passed to the
126 // callback callee.
127 Fn->addMetadata(
128 LLVMContext::MD_callback,
129 *MDNode::get(Ctx, {MDB.createCallbackEncoding(
130 2, {-1, -1}, /* VarArgsArePassed */ true)}));
131 }
132 }
133
134 LLVM_DEBUG(dbgs() << "Created OpenMP runtime function " << Fn->getName()
135 << " with type " << *Fn->getFunctionType() << "\n");
136 addAttributes(FnID, *Fn);
137
138 } else {
139 LLVM_DEBUG(dbgs() << "Found OpenMP runtime function " << Fn->getName()
140 << " with type " << *Fn->getFunctionType() << "\n");
141 }
142
143 assert(Fn && "Failed to create OpenMP runtime function");
144
145 // Cast the function to the expected type if necessary
146 Constant *C = ConstantExpr::getBitCast(Fn, FnTy->getPointerTo());
147 return {FnTy, C};
148 }
149
getOrCreateRuntimeFunctionPtr(RuntimeFunction FnID)150 Function *OpenMPIRBuilder::getOrCreateRuntimeFunctionPtr(RuntimeFunction FnID) {
151 FunctionCallee RTLFn = getOrCreateRuntimeFunction(M, FnID);
152 auto *Fn = dyn_cast<llvm::Function>(RTLFn.getCallee());
153 assert(Fn && "Failed to create OpenMP runtime function pointer");
154 return Fn;
155 }
156
initialize()157 void OpenMPIRBuilder::initialize() { initializeTypes(M); }
158
finalize(Function * Fn,bool AllowExtractorSinking)159 void OpenMPIRBuilder::finalize(Function *Fn, bool AllowExtractorSinking) {
160 SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
161 SmallVector<BasicBlock *, 32> Blocks;
162 SmallVector<OutlineInfo, 16> DeferredOutlines;
163 for (OutlineInfo &OI : OutlineInfos) {
164 // Skip functions that have not finalized yet; may happen with nested
165 // function generation.
166 if (Fn && OI.getFunction() != Fn) {
167 DeferredOutlines.push_back(OI);
168 continue;
169 }
170
171 ParallelRegionBlockSet.clear();
172 Blocks.clear();
173 OI.collectBlocks(ParallelRegionBlockSet, Blocks);
174
175 Function *OuterFn = OI.getFunction();
176 CodeExtractorAnalysisCache CEAC(*OuterFn);
177 CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr,
178 /* AggregateArgs */ false,
179 /* BlockFrequencyInfo */ nullptr,
180 /* BranchProbabilityInfo */ nullptr,
181 /* AssumptionCache */ nullptr,
182 /* AllowVarArgs */ true,
183 /* AllowAlloca */ true,
184 /* Suffix */ ".omp_par");
185
186 LLVM_DEBUG(dbgs() << "Before outlining: " << *OuterFn << "\n");
187 LLVM_DEBUG(dbgs() << "Entry " << OI.EntryBB->getName()
188 << " Exit: " << OI.ExitBB->getName() << "\n");
189 assert(Extractor.isEligible() &&
190 "Expected OpenMP outlining to be possible!");
191
192 Function *OutlinedFn = Extractor.extractCodeRegion(CEAC);
193
194 LLVM_DEBUG(dbgs() << "After outlining: " << *OuterFn << "\n");
195 LLVM_DEBUG(dbgs() << " Outlined function: " << *OutlinedFn << "\n");
196 assert(OutlinedFn->getReturnType()->isVoidTy() &&
197 "OpenMP outlined functions should not return a value!");
198
199 // For compability with the clang CG we move the outlined function after the
200 // one with the parallel region.
201 OutlinedFn->removeFromParent();
202 M.getFunctionList().insertAfter(OuterFn->getIterator(), OutlinedFn);
203
204 // Remove the artificial entry introduced by the extractor right away, we
205 // made our own entry block after all.
206 {
207 BasicBlock &ArtificialEntry = OutlinedFn->getEntryBlock();
208 assert(ArtificialEntry.getUniqueSuccessor() == OI.EntryBB);
209 assert(OI.EntryBB->getUniquePredecessor() == &ArtificialEntry);
210 if (AllowExtractorSinking) {
211 // Move instructions from the to-be-deleted ArtificialEntry to the entry
212 // basic block of the parallel region. CodeExtractor may have sunk
213 // allocas/bitcasts for values that are solely used in the outlined
214 // region and do not escape.
215 assert(!ArtificialEntry.empty() &&
216 "Expected instructions to sink in the outlined region");
217 for (BasicBlock::iterator It = ArtificialEntry.begin(),
218 End = ArtificialEntry.end();
219 It != End;) {
220 Instruction &I = *It;
221 It++;
222
223 if (I.isTerminator())
224 continue;
225
226 I.moveBefore(*OI.EntryBB, OI.EntryBB->getFirstInsertionPt());
227 }
228 }
229 OI.EntryBB->moveBefore(&ArtificialEntry);
230 ArtificialEntry.eraseFromParent();
231 }
232 assert(&OutlinedFn->getEntryBlock() == OI.EntryBB);
233 assert(OutlinedFn && OutlinedFn->getNumUses() == 1);
234
235 // Run a user callback, e.g. to add attributes.
236 if (OI.PostOutlineCB)
237 OI.PostOutlineCB(*OutlinedFn);
238 }
239
240 // Remove work items that have been completed.
241 OutlineInfos = std::move(DeferredOutlines);
242 }
243
~OpenMPIRBuilder()244 OpenMPIRBuilder::~OpenMPIRBuilder() {
245 assert(OutlineInfos.empty() && "There must be no outstanding outlinings");
246 }
247
createGlobalFlag(unsigned Value,StringRef Name)248 GlobalValue *OpenMPIRBuilder::createGlobalFlag(unsigned Value, StringRef Name) {
249 IntegerType *I32Ty = Type::getInt32Ty(M.getContext());
250 auto *GV =
251 new GlobalVariable(M, I32Ty,
252 /* isConstant = */ true, GlobalValue::WeakODRLinkage,
253 ConstantInt::get(I32Ty, Value), Name);
254
255 return GV;
256 }
257
getOrCreateIdent(Constant * SrcLocStr,IdentFlag LocFlags,unsigned Reserve2Flags)258 Value *OpenMPIRBuilder::getOrCreateIdent(Constant *SrcLocStr,
259 IdentFlag LocFlags,
260 unsigned Reserve2Flags) {
261 // Enable "C-mode".
262 LocFlags |= OMP_IDENT_FLAG_KMPC;
263
264 Value *&Ident =
265 IdentMap[{SrcLocStr, uint64_t(LocFlags) << 31 | Reserve2Flags}];
266 if (!Ident) {
267 Constant *I32Null = ConstantInt::getNullValue(Int32);
268 Constant *IdentData[] = {
269 I32Null, ConstantInt::get(Int32, uint32_t(LocFlags)),
270 ConstantInt::get(Int32, Reserve2Flags), I32Null, SrcLocStr};
271 Constant *Initializer =
272 ConstantStruct::get(OpenMPIRBuilder::Ident, IdentData);
273
274 // Look for existing encoding of the location + flags, not needed but
275 // minimizes the difference to the existing solution while we transition.
276 for (GlobalVariable &GV : M.getGlobalList())
277 if (GV.getValueType() == OpenMPIRBuilder::Ident && GV.hasInitializer())
278 if (GV.getInitializer() == Initializer)
279 Ident = &GV;
280
281 if (!Ident) {
282 auto *GV = new GlobalVariable(
283 M, OpenMPIRBuilder::Ident,
284 /* isConstant = */ true, GlobalValue::PrivateLinkage, Initializer, "",
285 nullptr, GlobalValue::NotThreadLocal,
286 M.getDataLayout().getDefaultGlobalsAddressSpace());
287 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
288 GV->setAlignment(Align(8));
289 Ident = GV;
290 }
291 }
292
293 return Builder.CreatePointerCast(Ident, IdentPtr);
294 }
295
getOrCreateSrcLocStr(StringRef LocStr)296 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef LocStr) {
297 Constant *&SrcLocStr = SrcLocStrMap[LocStr];
298 if (!SrcLocStr) {
299 Constant *Initializer =
300 ConstantDataArray::getString(M.getContext(), LocStr);
301
302 // Look for existing encoding of the location, not needed but minimizes the
303 // difference to the existing solution while we transition.
304 for (GlobalVariable &GV : M.getGlobalList())
305 if (GV.isConstant() && GV.hasInitializer() &&
306 GV.getInitializer() == Initializer)
307 return SrcLocStr = ConstantExpr::getPointerCast(&GV, Int8Ptr);
308
309 SrcLocStr = Builder.CreateGlobalStringPtr(LocStr, /* Name */ "",
310 /* AddressSpace */ 0, &M);
311 }
312 return SrcLocStr;
313 }
314
getOrCreateSrcLocStr(StringRef FunctionName,StringRef FileName,unsigned Line,unsigned Column)315 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef FunctionName,
316 StringRef FileName,
317 unsigned Line,
318 unsigned Column) {
319 SmallString<128> Buffer;
320 Buffer.push_back(';');
321 Buffer.append(FileName);
322 Buffer.push_back(';');
323 Buffer.append(FunctionName);
324 Buffer.push_back(';');
325 Buffer.append(std::to_string(Line));
326 Buffer.push_back(';');
327 Buffer.append(std::to_string(Column));
328 Buffer.push_back(';');
329 Buffer.push_back(';');
330 return getOrCreateSrcLocStr(Buffer.str());
331 }
332
getOrCreateDefaultSrcLocStr()333 Constant *OpenMPIRBuilder::getOrCreateDefaultSrcLocStr() {
334 return getOrCreateSrcLocStr(";unknown;unknown;0;0;;");
335 }
336
getOrCreateSrcLocStr(DebugLoc DL,Function * F)337 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(DebugLoc DL, Function *F) {
338 DILocation *DIL = DL.get();
339 if (!DIL)
340 return getOrCreateDefaultSrcLocStr();
341 StringRef FileName = M.getName();
342 if (DIFile *DIF = DIL->getFile())
343 if (Optional<StringRef> Source = DIF->getSource())
344 FileName = *Source;
345 StringRef Function = DIL->getScope()->getSubprogram()->getName();
346 if (Function.empty() && F)
347 Function = F->getName();
348 return getOrCreateSrcLocStr(Function, FileName, DIL->getLine(),
349 DIL->getColumn());
350 }
351
352 Constant *
getOrCreateSrcLocStr(const LocationDescription & Loc)353 OpenMPIRBuilder::getOrCreateSrcLocStr(const LocationDescription &Loc) {
354 return getOrCreateSrcLocStr(Loc.DL, Loc.IP.getBlock()->getParent());
355 }
356
getOrCreateThreadID(Value * Ident)357 Value *OpenMPIRBuilder::getOrCreateThreadID(Value *Ident) {
358 return Builder.CreateCall(
359 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num), Ident,
360 "omp_global_thread_num");
361 }
362
363 OpenMPIRBuilder::InsertPointTy
createBarrier(const LocationDescription & Loc,Directive DK,bool ForceSimpleCall,bool CheckCancelFlag)364 OpenMPIRBuilder::createBarrier(const LocationDescription &Loc, Directive DK,
365 bool ForceSimpleCall, bool CheckCancelFlag) {
366 if (!updateToLocation(Loc))
367 return Loc.IP;
368 return emitBarrierImpl(Loc, DK, ForceSimpleCall, CheckCancelFlag);
369 }
370
371 OpenMPIRBuilder::InsertPointTy
emitBarrierImpl(const LocationDescription & Loc,Directive Kind,bool ForceSimpleCall,bool CheckCancelFlag)372 OpenMPIRBuilder::emitBarrierImpl(const LocationDescription &Loc, Directive Kind,
373 bool ForceSimpleCall, bool CheckCancelFlag) {
374 // Build call __kmpc_cancel_barrier(loc, thread_id) or
375 // __kmpc_barrier(loc, thread_id);
376
377 IdentFlag BarrierLocFlags;
378 switch (Kind) {
379 case OMPD_for:
380 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_FOR;
381 break;
382 case OMPD_sections:
383 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SECTIONS;
384 break;
385 case OMPD_single:
386 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SINGLE;
387 break;
388 case OMPD_barrier:
389 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_EXPL;
390 break;
391 default:
392 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL;
393 break;
394 }
395
396 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
397 Value *Args[] = {getOrCreateIdent(SrcLocStr, BarrierLocFlags),
398 getOrCreateThreadID(getOrCreateIdent(SrcLocStr))};
399
400 // If we are in a cancellable parallel region, barriers are cancellation
401 // points.
402 // TODO: Check why we would force simple calls or to ignore the cancel flag.
403 bool UseCancelBarrier =
404 !ForceSimpleCall && isLastFinalizationInfoCancellable(OMPD_parallel);
405
406 Value *Result =
407 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(
408 UseCancelBarrier ? OMPRTL___kmpc_cancel_barrier
409 : OMPRTL___kmpc_barrier),
410 Args);
411
412 if (UseCancelBarrier && CheckCancelFlag)
413 emitCancelationCheckImpl(Result, OMPD_parallel);
414
415 return Builder.saveIP();
416 }
417
418 OpenMPIRBuilder::InsertPointTy
createCancel(const LocationDescription & Loc,Value * IfCondition,omp::Directive CanceledDirective)419 OpenMPIRBuilder::createCancel(const LocationDescription &Loc,
420 Value *IfCondition,
421 omp::Directive CanceledDirective) {
422 if (!updateToLocation(Loc))
423 return Loc.IP;
424
425 // LLVM utilities like blocks with terminators.
426 auto *UI = Builder.CreateUnreachable();
427
428 Instruction *ThenTI = UI, *ElseTI = nullptr;
429 if (IfCondition)
430 SplitBlockAndInsertIfThenElse(IfCondition, UI, &ThenTI, &ElseTI);
431 Builder.SetInsertPoint(ThenTI);
432
433 Value *CancelKind = nullptr;
434 switch (CanceledDirective) {
435 #define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value) \
436 case DirectiveEnum: \
437 CancelKind = Builder.getInt32(Value); \
438 break;
439 #include "llvm/Frontend/OpenMP/OMPKinds.def"
440 default:
441 llvm_unreachable("Unknown cancel kind!");
442 }
443
444 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
445 Value *Ident = getOrCreateIdent(SrcLocStr);
446 Value *Args[] = {Ident, getOrCreateThreadID(Ident), CancelKind};
447 Value *Result = Builder.CreateCall(
448 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_cancel), Args);
449 auto ExitCB = [this, CanceledDirective, Loc](InsertPointTy IP) {
450 if (CanceledDirective == OMPD_parallel) {
451 IRBuilder<>::InsertPointGuard IPG(Builder);
452 Builder.restoreIP(IP);
453 createBarrier(LocationDescription(Builder.saveIP(), Loc.DL),
454 omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false,
455 /* CheckCancelFlag */ false);
456 }
457 };
458
459 // The actual cancel logic is shared with others, e.g., cancel_barriers.
460 emitCancelationCheckImpl(Result, CanceledDirective, ExitCB);
461
462 // Update the insertion point and remove the terminator we introduced.
463 Builder.SetInsertPoint(UI->getParent());
464 UI->eraseFromParent();
465
466 return Builder.saveIP();
467 }
468
emitCancelationCheckImpl(Value * CancelFlag,omp::Directive CanceledDirective,FinalizeCallbackTy ExitCB)469 void OpenMPIRBuilder::emitCancelationCheckImpl(Value *CancelFlag,
470 omp::Directive CanceledDirective,
471 FinalizeCallbackTy ExitCB) {
472 assert(isLastFinalizationInfoCancellable(CanceledDirective) &&
473 "Unexpected cancellation!");
474
475 // For a cancel barrier we create two new blocks.
476 BasicBlock *BB = Builder.GetInsertBlock();
477 BasicBlock *NonCancellationBlock;
478 if (Builder.GetInsertPoint() == BB->end()) {
479 // TODO: This branch will not be needed once we moved to the
480 // OpenMPIRBuilder codegen completely.
481 NonCancellationBlock = BasicBlock::Create(
482 BB->getContext(), BB->getName() + ".cont", BB->getParent());
483 } else {
484 NonCancellationBlock = SplitBlock(BB, &*Builder.GetInsertPoint());
485 BB->getTerminator()->eraseFromParent();
486 Builder.SetInsertPoint(BB);
487 }
488 BasicBlock *CancellationBlock = BasicBlock::Create(
489 BB->getContext(), BB->getName() + ".cncl", BB->getParent());
490
491 // Jump to them based on the return value.
492 Value *Cmp = Builder.CreateIsNull(CancelFlag);
493 Builder.CreateCondBr(Cmp, NonCancellationBlock, CancellationBlock,
494 /* TODO weight */ nullptr, nullptr);
495
496 // From the cancellation block we finalize all variables and go to the
497 // post finalization block that is known to the FiniCB callback.
498 Builder.SetInsertPoint(CancellationBlock);
499 if (ExitCB)
500 ExitCB(Builder.saveIP());
501 auto &FI = FinalizationStack.back();
502 FI.FiniCB(Builder.saveIP());
503
504 // The continuation block is where code generation continues.
505 Builder.SetInsertPoint(NonCancellationBlock, NonCancellationBlock->begin());
506 }
507
createParallel(const LocationDescription & Loc,InsertPointTy OuterAllocaIP,BodyGenCallbackTy BodyGenCB,PrivatizeCallbackTy PrivCB,FinalizeCallbackTy FiniCB,Value * IfCondition,Value * NumThreads,omp::ProcBindKind ProcBind,bool IsCancellable)508 IRBuilder<>::InsertPoint OpenMPIRBuilder::createParallel(
509 const LocationDescription &Loc, InsertPointTy OuterAllocaIP,
510 BodyGenCallbackTy BodyGenCB, PrivatizeCallbackTy PrivCB,
511 FinalizeCallbackTy FiniCB, Value *IfCondition, Value *NumThreads,
512 omp::ProcBindKind ProcBind, bool IsCancellable) {
513 if (!updateToLocation(Loc))
514 return Loc.IP;
515
516 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
517 Value *Ident = getOrCreateIdent(SrcLocStr);
518 Value *ThreadID = getOrCreateThreadID(Ident);
519
520 if (NumThreads) {
521 // Build call __kmpc_push_num_threads(&Ident, global_tid, num_threads)
522 Value *Args[] = {
523 Ident, ThreadID,
524 Builder.CreateIntCast(NumThreads, Int32, /*isSigned*/ false)};
525 Builder.CreateCall(
526 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_num_threads), Args);
527 }
528
529 if (ProcBind != OMP_PROC_BIND_default) {
530 // Build call __kmpc_push_proc_bind(&Ident, global_tid, proc_bind)
531 Value *Args[] = {
532 Ident, ThreadID,
533 ConstantInt::get(Int32, unsigned(ProcBind), /*isSigned=*/true)};
534 Builder.CreateCall(
535 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_proc_bind), Args);
536 }
537
538 BasicBlock *InsertBB = Builder.GetInsertBlock();
539 Function *OuterFn = InsertBB->getParent();
540
541 // Save the outer alloca block because the insertion iterator may get
542 // invalidated and we still need this later.
543 BasicBlock *OuterAllocaBlock = OuterAllocaIP.getBlock();
544
545 // Vector to remember instructions we used only during the modeling but which
546 // we want to delete at the end.
547 SmallVector<Instruction *, 4> ToBeDeleted;
548
549 // Change the location to the outer alloca insertion point to create and
550 // initialize the allocas we pass into the parallel region.
551 Builder.restoreIP(OuterAllocaIP);
552 AllocaInst *TIDAddr = Builder.CreateAlloca(Int32, nullptr, "tid.addr");
553 AllocaInst *ZeroAddr = Builder.CreateAlloca(Int32, nullptr, "zero.addr");
554
555 // If there is an if condition we actually use the TIDAddr and ZeroAddr in the
556 // program, otherwise we only need them for modeling purposes to get the
557 // associated arguments in the outlined function. In the former case,
558 // initialize the allocas properly, in the latter case, delete them later.
559 if (IfCondition) {
560 Builder.CreateStore(Constant::getNullValue(Int32), TIDAddr);
561 Builder.CreateStore(Constant::getNullValue(Int32), ZeroAddr);
562 } else {
563 ToBeDeleted.push_back(TIDAddr);
564 ToBeDeleted.push_back(ZeroAddr);
565 }
566
567 // Create an artificial insertion point that will also ensure the blocks we
568 // are about to split are not degenerated.
569 auto *UI = new UnreachableInst(Builder.getContext(), InsertBB);
570
571 Instruction *ThenTI = UI, *ElseTI = nullptr;
572 if (IfCondition)
573 SplitBlockAndInsertIfThenElse(IfCondition, UI, &ThenTI, &ElseTI);
574
575 BasicBlock *ThenBB = ThenTI->getParent();
576 BasicBlock *PRegEntryBB = ThenBB->splitBasicBlock(ThenTI, "omp.par.entry");
577 BasicBlock *PRegBodyBB =
578 PRegEntryBB->splitBasicBlock(ThenTI, "omp.par.region");
579 BasicBlock *PRegPreFiniBB =
580 PRegBodyBB->splitBasicBlock(ThenTI, "omp.par.pre_finalize");
581 BasicBlock *PRegExitBB =
582 PRegPreFiniBB->splitBasicBlock(ThenTI, "omp.par.exit");
583
584 auto FiniCBWrapper = [&](InsertPointTy IP) {
585 // Hide "open-ended" blocks from the given FiniCB by setting the right jump
586 // target to the region exit block.
587 if (IP.getBlock()->end() == IP.getPoint()) {
588 IRBuilder<>::InsertPointGuard IPG(Builder);
589 Builder.restoreIP(IP);
590 Instruction *I = Builder.CreateBr(PRegExitBB);
591 IP = InsertPointTy(I->getParent(), I->getIterator());
592 }
593 assert(IP.getBlock()->getTerminator()->getNumSuccessors() == 1 &&
594 IP.getBlock()->getTerminator()->getSuccessor(0) == PRegExitBB &&
595 "Unexpected insertion point for finalization call!");
596 return FiniCB(IP);
597 };
598
599 FinalizationStack.push_back({FiniCBWrapper, OMPD_parallel, IsCancellable});
600
601 // Generate the privatization allocas in the block that will become the entry
602 // of the outlined function.
603 Builder.SetInsertPoint(PRegEntryBB->getTerminator());
604 InsertPointTy InnerAllocaIP = Builder.saveIP();
605
606 AllocaInst *PrivTIDAddr =
607 Builder.CreateAlloca(Int32, nullptr, "tid.addr.local");
608 Instruction *PrivTID = Builder.CreateLoad(Int32, PrivTIDAddr, "tid");
609
610 // Add some fake uses for OpenMP provided arguments.
611 ToBeDeleted.push_back(Builder.CreateLoad(Int32, TIDAddr, "tid.addr.use"));
612 Instruction *ZeroAddrUse =
613 Builder.CreateLoad(Int32, ZeroAddr, "zero.addr.use");
614 ToBeDeleted.push_back(ZeroAddrUse);
615
616 // ThenBB
617 // |
618 // V
619 // PRegionEntryBB <- Privatization allocas are placed here.
620 // |
621 // V
622 // PRegionBodyBB <- BodeGen is invoked here.
623 // |
624 // V
625 // PRegPreFiniBB <- The block we will start finalization from.
626 // |
627 // V
628 // PRegionExitBB <- A common exit to simplify block collection.
629 //
630
631 LLVM_DEBUG(dbgs() << "Before body codegen: " << *OuterFn << "\n");
632
633 // Let the caller create the body.
634 assert(BodyGenCB && "Expected body generation callback!");
635 InsertPointTy CodeGenIP(PRegBodyBB, PRegBodyBB->begin());
636 BodyGenCB(InnerAllocaIP, CodeGenIP, *PRegPreFiniBB);
637
638 LLVM_DEBUG(dbgs() << "After body codegen: " << *OuterFn << "\n");
639
640 FunctionCallee RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_fork_call);
641 if (auto *F = dyn_cast<llvm::Function>(RTLFn.getCallee())) {
642 if (!F->hasMetadata(llvm::LLVMContext::MD_callback)) {
643 llvm::LLVMContext &Ctx = F->getContext();
644 MDBuilder MDB(Ctx);
645 // Annotate the callback behavior of the __kmpc_fork_call:
646 // - The callback callee is argument number 2 (microtask).
647 // - The first two arguments of the callback callee are unknown (-1).
648 // - All variadic arguments to the __kmpc_fork_call are passed to the
649 // callback callee.
650 F->addMetadata(
651 llvm::LLVMContext::MD_callback,
652 *llvm::MDNode::get(
653 Ctx, {MDB.createCallbackEncoding(2, {-1, -1},
654 /* VarArgsArePassed */ true)}));
655 }
656 }
657
658 OutlineInfo OI;
659 OI.PostOutlineCB = [=](Function &OutlinedFn) {
660 // Add some known attributes.
661 OutlinedFn.addParamAttr(0, Attribute::NoAlias);
662 OutlinedFn.addParamAttr(1, Attribute::NoAlias);
663 OutlinedFn.addFnAttr(Attribute::NoUnwind);
664 OutlinedFn.addFnAttr(Attribute::NoRecurse);
665
666 assert(OutlinedFn.arg_size() >= 2 &&
667 "Expected at least tid and bounded tid as arguments");
668 unsigned NumCapturedVars =
669 OutlinedFn.arg_size() - /* tid & bounded tid */ 2;
670
671 CallInst *CI = cast<CallInst>(OutlinedFn.user_back());
672 CI->getParent()->setName("omp_parallel");
673 Builder.SetInsertPoint(CI);
674
675 // Build call __kmpc_fork_call(Ident, n, microtask, var1, .., varn);
676 Value *ForkCallArgs[] = {
677 Ident, Builder.getInt32(NumCapturedVars),
678 Builder.CreateBitCast(&OutlinedFn, ParallelTaskPtr)};
679
680 SmallVector<Value *, 16> RealArgs;
681 RealArgs.append(std::begin(ForkCallArgs), std::end(ForkCallArgs));
682 RealArgs.append(CI->arg_begin() + /* tid & bound tid */ 2, CI->arg_end());
683
684 Builder.CreateCall(RTLFn, RealArgs);
685
686 LLVM_DEBUG(dbgs() << "With fork_call placed: "
687 << *Builder.GetInsertBlock()->getParent() << "\n");
688
689 InsertPointTy ExitIP(PRegExitBB, PRegExitBB->end());
690
691 // Initialize the local TID stack location with the argument value.
692 Builder.SetInsertPoint(PrivTID);
693 Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin();
694 Builder.CreateStore(Builder.CreateLoad(Int32, OutlinedAI), PrivTIDAddr);
695
696 // If no "if" clause was present we do not need the call created during
697 // outlining, otherwise we reuse it in the serialized parallel region.
698 if (!ElseTI) {
699 CI->eraseFromParent();
700 } else {
701
702 // If an "if" clause was present we are now generating the serialized
703 // version into the "else" branch.
704 Builder.SetInsertPoint(ElseTI);
705
706 // Build calls __kmpc_serialized_parallel(&Ident, GTid);
707 Value *SerializedParallelCallArgs[] = {Ident, ThreadID};
708 Builder.CreateCall(
709 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_serialized_parallel),
710 SerializedParallelCallArgs);
711
712 // OutlinedFn(>id, &zero, CapturedStruct);
713 CI->removeFromParent();
714 Builder.Insert(CI);
715
716 // __kmpc_end_serialized_parallel(&Ident, GTid);
717 Value *EndArgs[] = {Ident, ThreadID};
718 Builder.CreateCall(
719 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_serialized_parallel),
720 EndArgs);
721
722 LLVM_DEBUG(dbgs() << "With serialized parallel region: "
723 << *Builder.GetInsertBlock()->getParent() << "\n");
724 }
725
726 for (Instruction *I : ToBeDeleted)
727 I->eraseFromParent();
728 };
729
730 // Adjust the finalization stack, verify the adjustment, and call the
731 // finalize function a last time to finalize values between the pre-fini
732 // block and the exit block if we left the parallel "the normal way".
733 auto FiniInfo = FinalizationStack.pop_back_val();
734 (void)FiniInfo;
735 assert(FiniInfo.DK == OMPD_parallel &&
736 "Unexpected finalization stack state!");
737
738 Instruction *PRegPreFiniTI = PRegPreFiniBB->getTerminator();
739
740 InsertPointTy PreFiniIP(PRegPreFiniBB, PRegPreFiniTI->getIterator());
741 FiniCB(PreFiniIP);
742
743 OI.EntryBB = PRegEntryBB;
744 OI.ExitBB = PRegExitBB;
745
746 SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
747 SmallVector<BasicBlock *, 32> Blocks;
748 OI.collectBlocks(ParallelRegionBlockSet, Blocks);
749
750 // Ensure a single exit node for the outlined region by creating one.
751 // We might have multiple incoming edges to the exit now due to finalizations,
752 // e.g., cancel calls that cause the control flow to leave the region.
753 BasicBlock *PRegOutlinedExitBB = PRegExitBB;
754 PRegExitBB = SplitBlock(PRegExitBB, &*PRegExitBB->getFirstInsertionPt());
755 PRegOutlinedExitBB->setName("omp.par.outlined.exit");
756 Blocks.push_back(PRegOutlinedExitBB);
757
758 CodeExtractorAnalysisCache CEAC(*OuterFn);
759 CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr,
760 /* AggregateArgs */ false,
761 /* BlockFrequencyInfo */ nullptr,
762 /* BranchProbabilityInfo */ nullptr,
763 /* AssumptionCache */ nullptr,
764 /* AllowVarArgs */ true,
765 /* AllowAlloca */ true,
766 /* Suffix */ ".omp_par");
767
768 // Find inputs to, outputs from the code region.
769 BasicBlock *CommonExit = nullptr;
770 SetVector<Value *> Inputs, Outputs, SinkingCands, HoistingCands;
771 Extractor.findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
772 Extractor.findInputsOutputs(Inputs, Outputs, SinkingCands);
773
774 LLVM_DEBUG(dbgs() << "Before privatization: " << *OuterFn << "\n");
775
776 FunctionCallee TIDRTLFn =
777 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num);
778
779 auto PrivHelper = [&](Value &V) {
780 if (&V == TIDAddr || &V == ZeroAddr)
781 return;
782
783 SetVector<Use *> Uses;
784 for (Use &U : V.uses())
785 if (auto *UserI = dyn_cast<Instruction>(U.getUser()))
786 if (ParallelRegionBlockSet.count(UserI->getParent()))
787 Uses.insert(&U);
788
789 // __kmpc_fork_call expects extra arguments as pointers. If the input
790 // already has a pointer type, everything is fine. Otherwise, store the
791 // value onto stack and load it back inside the to-be-outlined region. This
792 // will ensure only the pointer will be passed to the function.
793 // FIXME: if there are more than 15 trailing arguments, they must be
794 // additionally packed in a struct.
795 Value *Inner = &V;
796 if (!V.getType()->isPointerTy()) {
797 IRBuilder<>::InsertPointGuard Guard(Builder);
798 LLVM_DEBUG(llvm::dbgs() << "Forwarding input as pointer: " << V << "\n");
799
800 Builder.restoreIP(OuterAllocaIP);
801 Value *Ptr =
802 Builder.CreateAlloca(V.getType(), nullptr, V.getName() + ".reloaded");
803
804 // Store to stack at end of the block that currently branches to the entry
805 // block of the to-be-outlined region.
806 Builder.SetInsertPoint(InsertBB,
807 InsertBB->getTerminator()->getIterator());
808 Builder.CreateStore(&V, Ptr);
809
810 // Load back next to allocations in the to-be-outlined region.
811 Builder.restoreIP(InnerAllocaIP);
812 Inner = Builder.CreateLoad(V.getType(), Ptr);
813 }
814
815 Value *ReplacementValue = nullptr;
816 CallInst *CI = dyn_cast<CallInst>(&V);
817 if (CI && CI->getCalledFunction() == TIDRTLFn.getCallee()) {
818 ReplacementValue = PrivTID;
819 } else {
820 Builder.restoreIP(
821 PrivCB(InnerAllocaIP, Builder.saveIP(), V, *Inner, ReplacementValue));
822 assert(ReplacementValue &&
823 "Expected copy/create callback to set replacement value!");
824 if (ReplacementValue == &V)
825 return;
826 }
827
828 for (Use *UPtr : Uses)
829 UPtr->set(ReplacementValue);
830 };
831
832 // Reset the inner alloca insertion as it will be used for loading the values
833 // wrapped into pointers before passing them into the to-be-outlined region.
834 // Configure it to insert immediately after the fake use of zero address so
835 // that they are available in the generated body and so that the
836 // OpenMP-related values (thread ID and zero address pointers) remain leading
837 // in the argument list.
838 InnerAllocaIP = IRBuilder<>::InsertPoint(
839 ZeroAddrUse->getParent(), ZeroAddrUse->getNextNode()->getIterator());
840
841 // Reset the outer alloca insertion point to the entry of the relevant block
842 // in case it was invalidated.
843 OuterAllocaIP = IRBuilder<>::InsertPoint(
844 OuterAllocaBlock, OuterAllocaBlock->getFirstInsertionPt());
845
846 for (Value *Input : Inputs) {
847 LLVM_DEBUG(dbgs() << "Captured input: " << *Input << "\n");
848 PrivHelper(*Input);
849 }
850 LLVM_DEBUG({
851 for (Value *Output : Outputs)
852 LLVM_DEBUG(dbgs() << "Captured output: " << *Output << "\n");
853 });
854 assert(Outputs.empty() &&
855 "OpenMP outlining should not produce live-out values!");
856
857 LLVM_DEBUG(dbgs() << "After privatization: " << *OuterFn << "\n");
858 LLVM_DEBUG({
859 for (auto *BB : Blocks)
860 dbgs() << " PBR: " << BB->getName() << "\n";
861 });
862
863 // Register the outlined info.
864 addOutlineInfo(std::move(OI));
865
866 InsertPointTy AfterIP(UI->getParent(), UI->getParent()->end());
867 UI->eraseFromParent();
868
869 return AfterIP;
870 }
871
emitFlush(const LocationDescription & Loc)872 void OpenMPIRBuilder::emitFlush(const LocationDescription &Loc) {
873 // Build call void __kmpc_flush(ident_t *loc)
874 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
875 Value *Args[] = {getOrCreateIdent(SrcLocStr)};
876
877 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_flush), Args);
878 }
879
createFlush(const LocationDescription & Loc)880 void OpenMPIRBuilder::createFlush(const LocationDescription &Loc) {
881 if (!updateToLocation(Loc))
882 return;
883 emitFlush(Loc);
884 }
885
emitTaskwaitImpl(const LocationDescription & Loc)886 void OpenMPIRBuilder::emitTaskwaitImpl(const LocationDescription &Loc) {
887 // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
888 // global_tid);
889 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
890 Value *Ident = getOrCreateIdent(SrcLocStr);
891 Value *Args[] = {Ident, getOrCreateThreadID(Ident)};
892
893 // Ignore return result until untied tasks are supported.
894 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskwait),
895 Args);
896 }
897
createTaskwait(const LocationDescription & Loc)898 void OpenMPIRBuilder::createTaskwait(const LocationDescription &Loc) {
899 if (!updateToLocation(Loc))
900 return;
901 emitTaskwaitImpl(Loc);
902 }
903
emitTaskyieldImpl(const LocationDescription & Loc)904 void OpenMPIRBuilder::emitTaskyieldImpl(const LocationDescription &Loc) {
905 // Build call __kmpc_omp_taskyield(loc, thread_id, 0);
906 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
907 Value *Ident = getOrCreateIdent(SrcLocStr);
908 Constant *I32Null = ConstantInt::getNullValue(Int32);
909 Value *Args[] = {Ident, getOrCreateThreadID(Ident), I32Null};
910
911 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskyield),
912 Args);
913 }
914
createTaskyield(const LocationDescription & Loc)915 void OpenMPIRBuilder::createTaskyield(const LocationDescription &Loc) {
916 if (!updateToLocation(Loc))
917 return;
918 emitTaskyieldImpl(Loc);
919 }
920
createSections(const LocationDescription & Loc,InsertPointTy AllocaIP,ArrayRef<StorableBodyGenCallbackTy> SectionCBs,PrivatizeCallbackTy PrivCB,FinalizeCallbackTy FiniCB,bool IsCancellable,bool IsNowait)921 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSections(
922 const LocationDescription &Loc, InsertPointTy AllocaIP,
923 ArrayRef<StorableBodyGenCallbackTy> SectionCBs, PrivatizeCallbackTy PrivCB,
924 FinalizeCallbackTy FiniCB, bool IsCancellable, bool IsNowait) {
925 if (!updateToLocation(Loc))
926 return Loc.IP;
927
928 auto FiniCBWrapper = [&](InsertPointTy IP) {
929 if (IP.getBlock()->end() != IP.getPoint())
930 return FiniCB(IP);
931 // This must be done otherwise any nested constructs using FinalizeOMPRegion
932 // will fail because that function requires the Finalization Basic Block to
933 // have a terminator, which is already removed by EmitOMPRegionBody.
934 // IP is currently at cancelation block.
935 // We need to backtrack to the condition block to fetch
936 // the exit block and create a branch from cancelation
937 // to exit block.
938 IRBuilder<>::InsertPointGuard IPG(Builder);
939 Builder.restoreIP(IP);
940 auto *CaseBB = IP.getBlock()->getSinglePredecessor();
941 auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor();
942 auto *ExitBB = CondBB->getTerminator()->getSuccessor(1);
943 Instruction *I = Builder.CreateBr(ExitBB);
944 IP = InsertPointTy(I->getParent(), I->getIterator());
945 return FiniCB(IP);
946 };
947
948 FinalizationStack.push_back({FiniCBWrapper, OMPD_sections, IsCancellable});
949
950 // Each section is emitted as a switch case
951 // Each finalization callback is handled from clang.EmitOMPSectionDirective()
952 // -> OMP.createSection() which generates the IR for each section
953 // Iterate through all sections and emit a switch construct:
954 // switch (IV) {
955 // case 0:
956 // <SectionStmt[0]>;
957 // break;
958 // ...
959 // case <NumSection> - 1:
960 // <SectionStmt[<NumSection> - 1]>;
961 // break;
962 // }
963 // ...
964 // section_loop.after:
965 // <FiniCB>;
966 auto LoopBodyGenCB = [&](InsertPointTy CodeGenIP, Value *IndVar) {
967 auto *CurFn = CodeGenIP.getBlock()->getParent();
968 auto *ForIncBB = CodeGenIP.getBlock()->getSingleSuccessor();
969 auto *ForExitBB = CodeGenIP.getBlock()
970 ->getSinglePredecessor()
971 ->getTerminator()
972 ->getSuccessor(1);
973 SwitchInst *SwitchStmt = Builder.CreateSwitch(IndVar, ForIncBB);
974 Builder.restoreIP(CodeGenIP);
975 unsigned CaseNumber = 0;
976 for (auto SectionCB : SectionCBs) {
977 auto *CaseBB = BasicBlock::Create(M.getContext(),
978 "omp_section_loop.body.case", CurFn);
979 SwitchStmt->addCase(Builder.getInt32(CaseNumber), CaseBB);
980 Builder.SetInsertPoint(CaseBB);
981 SectionCB(InsertPointTy(), Builder.saveIP(), *ForExitBB);
982 CaseNumber++;
983 }
984 // remove the existing terminator from body BB since there can be no
985 // terminators after switch/case
986 CodeGenIP.getBlock()->getTerminator()->eraseFromParent();
987 };
988 // Loop body ends here
989 // LowerBound, UpperBound, and STride for createCanonicalLoop
990 Type *I32Ty = Type::getInt32Ty(M.getContext());
991 Value *LB = ConstantInt::get(I32Ty, 0);
992 Value *UB = ConstantInt::get(I32Ty, SectionCBs.size());
993 Value *ST = ConstantInt::get(I32Ty, 1);
994 llvm::CanonicalLoopInfo *LoopInfo = createCanonicalLoop(
995 Loc, LoopBodyGenCB, LB, UB, ST, true, false, AllocaIP, "section_loop");
996 InsertPointTy AfterIP =
997 applyStaticWorkshareLoop(Loc.DL, LoopInfo, AllocaIP, true);
998 BasicBlock *LoopAfterBB = AfterIP.getBlock();
999 Instruction *SplitPos = LoopAfterBB->getTerminator();
1000 if (!isa_and_nonnull<BranchInst>(SplitPos))
1001 SplitPos = new UnreachableInst(Builder.getContext(), LoopAfterBB);
1002 // ExitBB after LoopAfterBB because LoopAfterBB is used for FinalizationCB,
1003 // which requires a BB with branch
1004 BasicBlock *ExitBB =
1005 LoopAfterBB->splitBasicBlock(SplitPos, "omp_sections.end");
1006 SplitPos->eraseFromParent();
1007
1008 // Apply the finalization callback in LoopAfterBB
1009 auto FiniInfo = FinalizationStack.pop_back_val();
1010 assert(FiniInfo.DK == OMPD_sections &&
1011 "Unexpected finalization stack state!");
1012 Builder.SetInsertPoint(LoopAfterBB->getTerminator());
1013 FiniInfo.FiniCB(Builder.saveIP());
1014 Builder.SetInsertPoint(ExitBB);
1015
1016 return Builder.saveIP();
1017 }
1018
1019 OpenMPIRBuilder::InsertPointTy
createSection(const LocationDescription & Loc,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB)1020 OpenMPIRBuilder::createSection(const LocationDescription &Loc,
1021 BodyGenCallbackTy BodyGenCB,
1022 FinalizeCallbackTy FiniCB) {
1023 if (!updateToLocation(Loc))
1024 return Loc.IP;
1025
1026 auto FiniCBWrapper = [&](InsertPointTy IP) {
1027 if (IP.getBlock()->end() != IP.getPoint())
1028 return FiniCB(IP);
1029 // This must be done otherwise any nested constructs using FinalizeOMPRegion
1030 // will fail because that function requires the Finalization Basic Block to
1031 // have a terminator, which is already removed by EmitOMPRegionBody.
1032 // IP is currently at cancelation block.
1033 // We need to backtrack to the condition block to fetch
1034 // the exit block and create a branch from cancelation
1035 // to exit block.
1036 IRBuilder<>::InsertPointGuard IPG(Builder);
1037 Builder.restoreIP(IP);
1038 auto *CaseBB = Loc.IP.getBlock();
1039 auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor();
1040 auto *ExitBB = CondBB->getTerminator()->getSuccessor(1);
1041 Instruction *I = Builder.CreateBr(ExitBB);
1042 IP = InsertPointTy(I->getParent(), I->getIterator());
1043 return FiniCB(IP);
1044 };
1045
1046 Directive OMPD = Directive::OMPD_sections;
1047 // Since we are using Finalization Callback here, HasFinalize
1048 // and IsCancellable have to be true
1049 return EmitOMPInlinedRegion(OMPD, nullptr, nullptr, BodyGenCB, FiniCBWrapper,
1050 /*Conditional*/ false, /*hasFinalize*/ true,
1051 /*IsCancellable*/ true);
1052 }
1053
1054 /// Create a function with a unique name and a "void (i8*, i8*)" signature in
1055 /// the given module and return it.
getFreshReductionFunc(Module & M)1056 Function *getFreshReductionFunc(Module &M) {
1057 Type *VoidTy = Type::getVoidTy(M.getContext());
1058 Type *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
1059 auto *FuncTy =
1060 FunctionType::get(VoidTy, {Int8PtrTy, Int8PtrTy}, /* IsVarArg */ false);
1061 return Function::Create(FuncTy, GlobalVariable::InternalLinkage,
1062 M.getDataLayout().getDefaultGlobalsAddressSpace(),
1063 ".omp.reduction.func", &M);
1064 }
1065
createReductions(const LocationDescription & Loc,InsertPointTy AllocaIP,ArrayRef<ReductionInfo> ReductionInfos,bool IsNoWait)1066 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createReductions(
1067 const LocationDescription &Loc, InsertPointTy AllocaIP,
1068 ArrayRef<ReductionInfo> ReductionInfos, bool IsNoWait) {
1069 for (const ReductionInfo &RI : ReductionInfos) {
1070 (void)RI;
1071 assert(RI.Variable && "expected non-null variable");
1072 assert(RI.PrivateVariable && "expected non-null private variable");
1073 assert(RI.ReductionGen && "expected non-null reduction generator callback");
1074 assert(RI.Variable->getType() == RI.PrivateVariable->getType() &&
1075 "expected variables and their private equivalents to have the same "
1076 "type");
1077 assert(RI.Variable->getType()->isPointerTy() &&
1078 "expected variables to be pointers");
1079 }
1080
1081 if (!updateToLocation(Loc))
1082 return InsertPointTy();
1083
1084 BasicBlock *InsertBlock = Loc.IP.getBlock();
1085 BasicBlock *ContinuationBlock =
1086 InsertBlock->splitBasicBlock(Loc.IP.getPoint(), "reduce.finalize");
1087 InsertBlock->getTerminator()->eraseFromParent();
1088
1089 // Create and populate array of type-erased pointers to private reduction
1090 // values.
1091 unsigned NumReductions = ReductionInfos.size();
1092 Type *RedArrayTy = ArrayType::get(Builder.getInt8PtrTy(), NumReductions);
1093 Builder.restoreIP(AllocaIP);
1094 Value *RedArray = Builder.CreateAlloca(RedArrayTy, nullptr, "red.array");
1095
1096 Builder.SetInsertPoint(InsertBlock, InsertBlock->end());
1097
1098 for (auto En : enumerate(ReductionInfos)) {
1099 unsigned Index = En.index();
1100 const ReductionInfo &RI = En.value();
1101 Value *RedArrayElemPtr = Builder.CreateConstInBoundsGEP2_64(
1102 RedArrayTy, RedArray, 0, Index, "red.array.elem." + Twine(Index));
1103 Value *Casted =
1104 Builder.CreateBitCast(RI.PrivateVariable, Builder.getInt8PtrTy(),
1105 "private.red.var." + Twine(Index) + ".casted");
1106 Builder.CreateStore(Casted, RedArrayElemPtr);
1107 }
1108
1109 // Emit a call to the runtime function that orchestrates the reduction.
1110 // Declare the reduction function in the process.
1111 Function *Func = Builder.GetInsertBlock()->getParent();
1112 Module *Module = Func->getParent();
1113 Value *RedArrayPtr =
1114 Builder.CreateBitCast(RedArray, Builder.getInt8PtrTy(), "red.array.ptr");
1115 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
1116 bool CanGenerateAtomic =
1117 llvm::all_of(ReductionInfos, [](const ReductionInfo &RI) {
1118 return RI.AtomicReductionGen;
1119 });
1120 Value *Ident = getOrCreateIdent(
1121 SrcLocStr, CanGenerateAtomic ? IdentFlag::OMP_IDENT_FLAG_ATOMIC_REDUCE
1122 : IdentFlag(0));
1123 Value *ThreadId = getOrCreateThreadID(Ident);
1124 Constant *NumVariables = Builder.getInt32(NumReductions);
1125 const DataLayout &DL = Module->getDataLayout();
1126 unsigned RedArrayByteSize = DL.getTypeStoreSize(RedArrayTy);
1127 Constant *RedArraySize = Builder.getInt64(RedArrayByteSize);
1128 Function *ReductionFunc = getFreshReductionFunc(*Module);
1129 Value *Lock = getOMPCriticalRegionLock(".reduction");
1130 Function *ReduceFunc = getOrCreateRuntimeFunctionPtr(
1131 IsNoWait ? RuntimeFunction::OMPRTL___kmpc_reduce_nowait
1132 : RuntimeFunction::OMPRTL___kmpc_reduce);
1133 CallInst *ReduceCall =
1134 Builder.CreateCall(ReduceFunc,
1135 {Ident, ThreadId, NumVariables, RedArraySize,
1136 RedArrayPtr, ReductionFunc, Lock},
1137 "reduce");
1138
1139 // Create final reduction entry blocks for the atomic and non-atomic case.
1140 // Emit IR that dispatches control flow to one of the blocks based on the
1141 // reduction supporting the atomic mode.
1142 BasicBlock *NonAtomicRedBlock =
1143 BasicBlock::Create(Module->getContext(), "reduce.switch.nonatomic", Func);
1144 BasicBlock *AtomicRedBlock =
1145 BasicBlock::Create(Module->getContext(), "reduce.switch.atomic", Func);
1146 SwitchInst *Switch =
1147 Builder.CreateSwitch(ReduceCall, ContinuationBlock, /* NumCases */ 2);
1148 Switch->addCase(Builder.getInt32(1), NonAtomicRedBlock);
1149 Switch->addCase(Builder.getInt32(2), AtomicRedBlock);
1150
1151 // Populate the non-atomic reduction using the elementwise reduction function.
1152 // This loads the elements from the global and private variables and reduces
1153 // them before storing back the result to the global variable.
1154 Builder.SetInsertPoint(NonAtomicRedBlock);
1155 for (auto En : enumerate(ReductionInfos)) {
1156 const ReductionInfo &RI = En.value();
1157 Type *ValueType = RI.getElementType();
1158 Value *RedValue = Builder.CreateLoad(ValueType, RI.Variable,
1159 "red.value." + Twine(En.index()));
1160 Value *PrivateRedValue =
1161 Builder.CreateLoad(ValueType, RI.PrivateVariable,
1162 "red.private.value." + Twine(En.index()));
1163 Value *Reduced;
1164 Builder.restoreIP(
1165 RI.ReductionGen(Builder.saveIP(), RedValue, PrivateRedValue, Reduced));
1166 if (!Builder.GetInsertBlock())
1167 return InsertPointTy();
1168 Builder.CreateStore(Reduced, RI.Variable);
1169 }
1170 Function *EndReduceFunc = getOrCreateRuntimeFunctionPtr(
1171 IsNoWait ? RuntimeFunction::OMPRTL___kmpc_end_reduce_nowait
1172 : RuntimeFunction::OMPRTL___kmpc_end_reduce);
1173 Builder.CreateCall(EndReduceFunc, {Ident, ThreadId, Lock});
1174 Builder.CreateBr(ContinuationBlock);
1175
1176 // Populate the atomic reduction using the atomic elementwise reduction
1177 // function. There are no loads/stores here because they will be happening
1178 // inside the atomic elementwise reduction.
1179 Builder.SetInsertPoint(AtomicRedBlock);
1180 if (CanGenerateAtomic) {
1181 for (const ReductionInfo &RI : ReductionInfos) {
1182 Builder.restoreIP(RI.AtomicReductionGen(Builder.saveIP(), RI.Variable,
1183 RI.PrivateVariable));
1184 if (!Builder.GetInsertBlock())
1185 return InsertPointTy();
1186 }
1187 Builder.CreateBr(ContinuationBlock);
1188 } else {
1189 Builder.CreateUnreachable();
1190 }
1191
1192 // Populate the outlined reduction function using the elementwise reduction
1193 // function. Partial values are extracted from the type-erased array of
1194 // pointers to private variables.
1195 BasicBlock *ReductionFuncBlock =
1196 BasicBlock::Create(Module->getContext(), "", ReductionFunc);
1197 Builder.SetInsertPoint(ReductionFuncBlock);
1198 Value *LHSArrayPtr = Builder.CreateBitCast(ReductionFunc->getArg(0),
1199 RedArrayTy->getPointerTo());
1200 Value *RHSArrayPtr = Builder.CreateBitCast(ReductionFunc->getArg(1),
1201 RedArrayTy->getPointerTo());
1202 for (auto En : enumerate(ReductionInfos)) {
1203 const ReductionInfo &RI = En.value();
1204 Value *LHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64(
1205 RedArrayTy, LHSArrayPtr, 0, En.index());
1206 Value *LHSI8Ptr = Builder.CreateLoad(Builder.getInt8PtrTy(), LHSI8PtrPtr);
1207 Value *LHSPtr = Builder.CreateBitCast(LHSI8Ptr, RI.Variable->getType());
1208 Value *LHS = Builder.CreateLoad(RI.getElementType(), LHSPtr);
1209 Value *RHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64(
1210 RedArrayTy, RHSArrayPtr, 0, En.index());
1211 Value *RHSI8Ptr = Builder.CreateLoad(Builder.getInt8PtrTy(), RHSI8PtrPtr);
1212 Value *RHSPtr =
1213 Builder.CreateBitCast(RHSI8Ptr, RI.PrivateVariable->getType());
1214 Value *RHS = Builder.CreateLoad(RI.getElementType(), RHSPtr);
1215 Value *Reduced;
1216 Builder.restoreIP(RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced));
1217 if (!Builder.GetInsertBlock())
1218 return InsertPointTy();
1219 Builder.CreateStore(Reduced, LHSPtr);
1220 }
1221 Builder.CreateRetVoid();
1222
1223 Builder.SetInsertPoint(ContinuationBlock);
1224 return Builder.saveIP();
1225 }
1226
1227 OpenMPIRBuilder::InsertPointTy
createMaster(const LocationDescription & Loc,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB)1228 OpenMPIRBuilder::createMaster(const LocationDescription &Loc,
1229 BodyGenCallbackTy BodyGenCB,
1230 FinalizeCallbackTy FiniCB) {
1231
1232 if (!updateToLocation(Loc))
1233 return Loc.IP;
1234
1235 Directive OMPD = Directive::OMPD_master;
1236 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
1237 Value *Ident = getOrCreateIdent(SrcLocStr);
1238 Value *ThreadId = getOrCreateThreadID(Ident);
1239 Value *Args[] = {Ident, ThreadId};
1240
1241 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_master);
1242 Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args);
1243
1244 Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_master);
1245 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args);
1246
1247 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
1248 /*Conditional*/ true, /*hasFinalize*/ true);
1249 }
1250
1251 OpenMPIRBuilder::InsertPointTy
createMasked(const LocationDescription & Loc,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB,Value * Filter)1252 OpenMPIRBuilder::createMasked(const LocationDescription &Loc,
1253 BodyGenCallbackTy BodyGenCB,
1254 FinalizeCallbackTy FiniCB, Value *Filter) {
1255 if (!updateToLocation(Loc))
1256 return Loc.IP;
1257
1258 Directive OMPD = Directive::OMPD_masked;
1259 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
1260 Value *Ident = getOrCreateIdent(SrcLocStr);
1261 Value *ThreadId = getOrCreateThreadID(Ident);
1262 Value *Args[] = {Ident, ThreadId, Filter};
1263 Value *ArgsEnd[] = {Ident, ThreadId};
1264
1265 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_masked);
1266 Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args);
1267
1268 Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_masked);
1269 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, ArgsEnd);
1270
1271 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
1272 /*Conditional*/ true, /*hasFinalize*/ true);
1273 }
1274
createLoopSkeleton(DebugLoc DL,Value * TripCount,Function * F,BasicBlock * PreInsertBefore,BasicBlock * PostInsertBefore,const Twine & Name)1275 CanonicalLoopInfo *OpenMPIRBuilder::createLoopSkeleton(
1276 DebugLoc DL, Value *TripCount, Function *F, BasicBlock *PreInsertBefore,
1277 BasicBlock *PostInsertBefore, const Twine &Name) {
1278 Module *M = F->getParent();
1279 LLVMContext &Ctx = M->getContext();
1280 Type *IndVarTy = TripCount->getType();
1281
1282 // Create the basic block structure.
1283 BasicBlock *Preheader =
1284 BasicBlock::Create(Ctx, "omp_" + Name + ".preheader", F, PreInsertBefore);
1285 BasicBlock *Header =
1286 BasicBlock::Create(Ctx, "omp_" + Name + ".header", F, PreInsertBefore);
1287 BasicBlock *Cond =
1288 BasicBlock::Create(Ctx, "omp_" + Name + ".cond", F, PreInsertBefore);
1289 BasicBlock *Body =
1290 BasicBlock::Create(Ctx, "omp_" + Name + ".body", F, PreInsertBefore);
1291 BasicBlock *Latch =
1292 BasicBlock::Create(Ctx, "omp_" + Name + ".inc", F, PostInsertBefore);
1293 BasicBlock *Exit =
1294 BasicBlock::Create(Ctx, "omp_" + Name + ".exit", F, PostInsertBefore);
1295 BasicBlock *After =
1296 BasicBlock::Create(Ctx, "omp_" + Name + ".after", F, PostInsertBefore);
1297
1298 // Use specified DebugLoc for new instructions.
1299 Builder.SetCurrentDebugLocation(DL);
1300
1301 Builder.SetInsertPoint(Preheader);
1302 Builder.CreateBr(Header);
1303
1304 Builder.SetInsertPoint(Header);
1305 PHINode *IndVarPHI = Builder.CreatePHI(IndVarTy, 2, "omp_" + Name + ".iv");
1306 IndVarPHI->addIncoming(ConstantInt::get(IndVarTy, 0), Preheader);
1307 Builder.CreateBr(Cond);
1308
1309 Builder.SetInsertPoint(Cond);
1310 Value *Cmp =
1311 Builder.CreateICmpULT(IndVarPHI, TripCount, "omp_" + Name + ".cmp");
1312 Builder.CreateCondBr(Cmp, Body, Exit);
1313
1314 Builder.SetInsertPoint(Body);
1315 Builder.CreateBr(Latch);
1316
1317 Builder.SetInsertPoint(Latch);
1318 Value *Next = Builder.CreateAdd(IndVarPHI, ConstantInt::get(IndVarTy, 1),
1319 "omp_" + Name + ".next", /*HasNUW=*/true);
1320 Builder.CreateBr(Header);
1321 IndVarPHI->addIncoming(Next, Latch);
1322
1323 Builder.SetInsertPoint(Exit);
1324 Builder.CreateBr(After);
1325
1326 // Remember and return the canonical control flow.
1327 LoopInfos.emplace_front();
1328 CanonicalLoopInfo *CL = &LoopInfos.front();
1329
1330 CL->Preheader = Preheader;
1331 CL->Header = Header;
1332 CL->Cond = Cond;
1333 CL->Body = Body;
1334 CL->Latch = Latch;
1335 CL->Exit = Exit;
1336 CL->After = After;
1337
1338 #ifndef NDEBUG
1339 CL->assertOK();
1340 #endif
1341 return CL;
1342 }
1343
1344 CanonicalLoopInfo *
createCanonicalLoop(const LocationDescription & Loc,LoopBodyGenCallbackTy BodyGenCB,Value * TripCount,const Twine & Name)1345 OpenMPIRBuilder::createCanonicalLoop(const LocationDescription &Loc,
1346 LoopBodyGenCallbackTy BodyGenCB,
1347 Value *TripCount, const Twine &Name) {
1348 BasicBlock *BB = Loc.IP.getBlock();
1349 BasicBlock *NextBB = BB->getNextNode();
1350
1351 CanonicalLoopInfo *CL = createLoopSkeleton(Loc.DL, TripCount, BB->getParent(),
1352 NextBB, NextBB, Name);
1353 BasicBlock *After = CL->getAfter();
1354
1355 // If location is not set, don't connect the loop.
1356 if (updateToLocation(Loc)) {
1357 // Split the loop at the insertion point: Branch to the preheader and move
1358 // every following instruction to after the loop (the After BB). Also, the
1359 // new successor is the loop's after block.
1360 Builder.CreateBr(CL->Preheader);
1361 After->getInstList().splice(After->begin(), BB->getInstList(),
1362 Builder.GetInsertPoint(), BB->end());
1363 After->replaceSuccessorsPhiUsesWith(BB, After);
1364 }
1365
1366 // Emit the body content. We do it after connecting the loop to the CFG to
1367 // avoid that the callback encounters degenerate BBs.
1368 BodyGenCB(CL->getBodyIP(), CL->getIndVar());
1369
1370 #ifndef NDEBUG
1371 CL->assertOK();
1372 #endif
1373 return CL;
1374 }
1375
createCanonicalLoop(const LocationDescription & Loc,LoopBodyGenCallbackTy BodyGenCB,Value * Start,Value * Stop,Value * Step,bool IsSigned,bool InclusiveStop,InsertPointTy ComputeIP,const Twine & Name)1376 CanonicalLoopInfo *OpenMPIRBuilder::createCanonicalLoop(
1377 const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB,
1378 Value *Start, Value *Stop, Value *Step, bool IsSigned, bool InclusiveStop,
1379 InsertPointTy ComputeIP, const Twine &Name) {
1380
1381 // Consider the following difficulties (assuming 8-bit signed integers):
1382 // * Adding \p Step to the loop counter which passes \p Stop may overflow:
1383 // DO I = 1, 100, 50
1384 /// * A \p Step of INT_MIN cannot not be normalized to a positive direction:
1385 // DO I = 100, 0, -128
1386
1387 // Start, Stop and Step must be of the same integer type.
1388 auto *IndVarTy = cast<IntegerType>(Start->getType());
1389 assert(IndVarTy == Stop->getType() && "Stop type mismatch");
1390 assert(IndVarTy == Step->getType() && "Step type mismatch");
1391
1392 LocationDescription ComputeLoc =
1393 ComputeIP.isSet() ? LocationDescription(ComputeIP, Loc.DL) : Loc;
1394 updateToLocation(ComputeLoc);
1395
1396 ConstantInt *Zero = ConstantInt::get(IndVarTy, 0);
1397 ConstantInt *One = ConstantInt::get(IndVarTy, 1);
1398
1399 // Like Step, but always positive.
1400 Value *Incr = Step;
1401
1402 // Distance between Start and Stop; always positive.
1403 Value *Span;
1404
1405 // Condition whether there are no iterations are executed at all, e.g. because
1406 // UB < LB.
1407 Value *ZeroCmp;
1408
1409 if (IsSigned) {
1410 // Ensure that increment is positive. If not, negate and invert LB and UB.
1411 Value *IsNeg = Builder.CreateICmpSLT(Step, Zero);
1412 Incr = Builder.CreateSelect(IsNeg, Builder.CreateNeg(Step), Step);
1413 Value *LB = Builder.CreateSelect(IsNeg, Stop, Start);
1414 Value *UB = Builder.CreateSelect(IsNeg, Start, Stop);
1415 Span = Builder.CreateSub(UB, LB, "", false, true);
1416 ZeroCmp = Builder.CreateICmp(
1417 InclusiveStop ? CmpInst::ICMP_SLT : CmpInst::ICMP_SLE, UB, LB);
1418 } else {
1419 Span = Builder.CreateSub(Stop, Start, "", true);
1420 ZeroCmp = Builder.CreateICmp(
1421 InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Stop, Start);
1422 }
1423
1424 Value *CountIfLooping;
1425 if (InclusiveStop) {
1426 CountIfLooping = Builder.CreateAdd(Builder.CreateUDiv(Span, Incr), One);
1427 } else {
1428 // Avoid incrementing past stop since it could overflow.
1429 Value *CountIfTwo = Builder.CreateAdd(
1430 Builder.CreateUDiv(Builder.CreateSub(Span, One), Incr), One);
1431 Value *OneCmp = Builder.CreateICmp(
1432 InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Span, Incr);
1433 CountIfLooping = Builder.CreateSelect(OneCmp, One, CountIfTwo);
1434 }
1435 Value *TripCount = Builder.CreateSelect(ZeroCmp, Zero, CountIfLooping,
1436 "omp_" + Name + ".tripcount");
1437
1438 auto BodyGen = [=](InsertPointTy CodeGenIP, Value *IV) {
1439 Builder.restoreIP(CodeGenIP);
1440 Value *Span = Builder.CreateMul(IV, Step);
1441 Value *IndVar = Builder.CreateAdd(Span, Start);
1442 BodyGenCB(Builder.saveIP(), IndVar);
1443 };
1444 LocationDescription LoopLoc = ComputeIP.isSet() ? Loc.IP : Builder.saveIP();
1445 return createCanonicalLoop(LoopLoc, BodyGen, TripCount, Name);
1446 }
1447
1448 // Returns an LLVM function to call for initializing loop bounds using OpenMP
1449 // static scheduling depending on `type`. Only i32 and i64 are supported by the
1450 // runtime. Always interpret integers as unsigned similarly to
1451 // CanonicalLoopInfo.
getKmpcForStaticInitForType(Type * Ty,Module & M,OpenMPIRBuilder & OMPBuilder)1452 static FunctionCallee getKmpcForStaticInitForType(Type *Ty, Module &M,
1453 OpenMPIRBuilder &OMPBuilder) {
1454 unsigned Bitwidth = Ty->getIntegerBitWidth();
1455 if (Bitwidth == 32)
1456 return OMPBuilder.getOrCreateRuntimeFunction(
1457 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_4u);
1458 if (Bitwidth == 64)
1459 return OMPBuilder.getOrCreateRuntimeFunction(
1460 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_8u);
1461 llvm_unreachable("unknown OpenMP loop iterator bitwidth");
1462 }
1463
1464 // Sets the number of loop iterations to the given value. This value must be
1465 // valid in the condition block (i.e., defined in the preheader) and is
1466 // interpreted as an unsigned integer.
setCanonicalLoopTripCount(CanonicalLoopInfo * CLI,Value * TripCount)1467 void setCanonicalLoopTripCount(CanonicalLoopInfo *CLI, Value *TripCount) {
1468 Instruction *CmpI = &CLI->getCond()->front();
1469 assert(isa<CmpInst>(CmpI) && "First inst must compare IV with TripCount");
1470 CmpI->setOperand(1, TripCount);
1471 CLI->assertOK();
1472 }
1473
1474 OpenMPIRBuilder::InsertPointTy
applyStaticWorkshareLoop(DebugLoc DL,CanonicalLoopInfo * CLI,InsertPointTy AllocaIP,bool NeedsBarrier,Value * Chunk)1475 OpenMPIRBuilder::applyStaticWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI,
1476 InsertPointTy AllocaIP,
1477 bool NeedsBarrier, Value *Chunk) {
1478 assert(CLI->isValid() && "Requires a valid canonical loop");
1479
1480 // Set up the source location value for OpenMP runtime.
1481 Builder.restoreIP(CLI->getPreheaderIP());
1482 Builder.SetCurrentDebugLocation(DL);
1483
1484 Constant *SrcLocStr = getOrCreateSrcLocStr(DL);
1485 Value *SrcLoc = getOrCreateIdent(SrcLocStr);
1486
1487 // Declare useful OpenMP runtime functions.
1488 Value *IV = CLI->getIndVar();
1489 Type *IVTy = IV->getType();
1490 FunctionCallee StaticInit = getKmpcForStaticInitForType(IVTy, M, *this);
1491 FunctionCallee StaticFini =
1492 getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini);
1493
1494 // Allocate space for computed loop bounds as expected by the "init" function.
1495 Builder.restoreIP(AllocaIP);
1496 Type *I32Type = Type::getInt32Ty(M.getContext());
1497 Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
1498 Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound");
1499 Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound");
1500 Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride");
1501
1502 // At the end of the preheader, prepare for calling the "init" function by
1503 // storing the current loop bounds into the allocated space. A canonical loop
1504 // always iterates from 0 to trip-count with step 1. Note that "init" expects
1505 // and produces an inclusive upper bound.
1506 Builder.SetInsertPoint(CLI->getPreheader()->getTerminator());
1507 Constant *Zero = ConstantInt::get(IVTy, 0);
1508 Constant *One = ConstantInt::get(IVTy, 1);
1509 Builder.CreateStore(Zero, PLowerBound);
1510 Value *UpperBound = Builder.CreateSub(CLI->getTripCount(), One);
1511 Builder.CreateStore(UpperBound, PUpperBound);
1512 Builder.CreateStore(One, PStride);
1513
1514 // FIXME: schedule(static) is NOT the same as schedule(static,1)
1515 if (!Chunk)
1516 Chunk = One;
1517
1518 Value *ThreadNum = getOrCreateThreadID(SrcLoc);
1519
1520 Constant *SchedulingType =
1521 ConstantInt::get(I32Type, static_cast<int>(OMPScheduleType::Static));
1522
1523 // Call the "init" function and update the trip count of the loop with the
1524 // value it produced.
1525 Builder.CreateCall(StaticInit,
1526 {SrcLoc, ThreadNum, SchedulingType, PLastIter, PLowerBound,
1527 PUpperBound, PStride, One, Chunk});
1528 Value *LowerBound = Builder.CreateLoad(IVTy, PLowerBound);
1529 Value *InclusiveUpperBound = Builder.CreateLoad(IVTy, PUpperBound);
1530 Value *TripCountMinusOne = Builder.CreateSub(InclusiveUpperBound, LowerBound);
1531 Value *TripCount = Builder.CreateAdd(TripCountMinusOne, One);
1532 setCanonicalLoopTripCount(CLI, TripCount);
1533
1534 // Update all uses of the induction variable except the one in the condition
1535 // block that compares it with the actual upper bound, and the increment in
1536 // the latch block.
1537 // TODO: this can eventually move to CanonicalLoopInfo or to a new
1538 // CanonicalLoopInfoUpdater interface.
1539 Builder.SetInsertPoint(CLI->getBody(), CLI->getBody()->getFirstInsertionPt());
1540 Value *UpdatedIV = Builder.CreateAdd(IV, LowerBound);
1541 IV->replaceUsesWithIf(UpdatedIV, [&](Use &U) {
1542 auto *Instr = dyn_cast<Instruction>(U.getUser());
1543 return !Instr ||
1544 (Instr->getParent() != CLI->getCond() &&
1545 Instr->getParent() != CLI->getLatch() && Instr != UpdatedIV);
1546 });
1547
1548 // In the "exit" block, call the "fini" function.
1549 Builder.SetInsertPoint(CLI->getExit(),
1550 CLI->getExit()->getTerminator()->getIterator());
1551 Builder.CreateCall(StaticFini, {SrcLoc, ThreadNum});
1552
1553 // Add the barrier if requested.
1554 if (NeedsBarrier)
1555 createBarrier(LocationDescription(Builder.saveIP(), DL),
1556 omp::Directive::OMPD_for, /* ForceSimpleCall */ false,
1557 /* CheckCancelFlag */ false);
1558
1559 InsertPointTy AfterIP = CLI->getAfterIP();
1560 CLI->invalidate();
1561
1562 return AfterIP;
1563 }
1564
1565 OpenMPIRBuilder::InsertPointTy
applyWorkshareLoop(DebugLoc DL,CanonicalLoopInfo * CLI,InsertPointTy AllocaIP,bool NeedsBarrier)1566 OpenMPIRBuilder::applyWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI,
1567 InsertPointTy AllocaIP, bool NeedsBarrier) {
1568 // Currently only supports static schedules.
1569 return applyStaticWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier);
1570 }
1571
1572 /// Returns an LLVM function to call for initializing loop bounds using OpenMP
1573 /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by
1574 /// the runtime. Always interpret integers as unsigned similarly to
1575 /// CanonicalLoopInfo.
1576 static FunctionCallee
getKmpcForDynamicInitForType(Type * Ty,Module & M,OpenMPIRBuilder & OMPBuilder)1577 getKmpcForDynamicInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
1578 unsigned Bitwidth = Ty->getIntegerBitWidth();
1579 if (Bitwidth == 32)
1580 return OMPBuilder.getOrCreateRuntimeFunction(
1581 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_4u);
1582 if (Bitwidth == 64)
1583 return OMPBuilder.getOrCreateRuntimeFunction(
1584 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_8u);
1585 llvm_unreachable("unknown OpenMP loop iterator bitwidth");
1586 }
1587
1588 /// Returns an LLVM function to call for updating the next loop using OpenMP
1589 /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by
1590 /// the runtime. Always interpret integers as unsigned similarly to
1591 /// CanonicalLoopInfo.
1592 static FunctionCallee
getKmpcForDynamicNextForType(Type * Ty,Module & M,OpenMPIRBuilder & OMPBuilder)1593 getKmpcForDynamicNextForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
1594 unsigned Bitwidth = Ty->getIntegerBitWidth();
1595 if (Bitwidth == 32)
1596 return OMPBuilder.getOrCreateRuntimeFunction(
1597 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_4u);
1598 if (Bitwidth == 64)
1599 return OMPBuilder.getOrCreateRuntimeFunction(
1600 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_8u);
1601 llvm_unreachable("unknown OpenMP loop iterator bitwidth");
1602 }
1603
applyDynamicWorkshareLoop(DebugLoc DL,CanonicalLoopInfo * CLI,InsertPointTy AllocaIP,OMPScheduleType SchedType,bool NeedsBarrier,Value * Chunk)1604 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyDynamicWorkshareLoop(
1605 DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
1606 OMPScheduleType SchedType, bool NeedsBarrier, Value *Chunk) {
1607 assert(CLI->isValid() && "Requires a valid canonical loop");
1608
1609 // Set up the source location value for OpenMP runtime.
1610 Builder.SetCurrentDebugLocation(DL);
1611
1612 Constant *SrcLocStr = getOrCreateSrcLocStr(DL);
1613 Value *SrcLoc = getOrCreateIdent(SrcLocStr);
1614
1615 // Declare useful OpenMP runtime functions.
1616 Value *IV = CLI->getIndVar();
1617 Type *IVTy = IV->getType();
1618 FunctionCallee DynamicInit = getKmpcForDynamicInitForType(IVTy, M, *this);
1619 FunctionCallee DynamicNext = getKmpcForDynamicNextForType(IVTy, M, *this);
1620
1621 // Allocate space for computed loop bounds as expected by the "init" function.
1622 Builder.restoreIP(AllocaIP);
1623 Type *I32Type = Type::getInt32Ty(M.getContext());
1624 Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
1625 Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound");
1626 Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound");
1627 Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride");
1628
1629 // At the end of the preheader, prepare for calling the "init" function by
1630 // storing the current loop bounds into the allocated space. A canonical loop
1631 // always iterates from 0 to trip-count with step 1. Note that "init" expects
1632 // and produces an inclusive upper bound.
1633 BasicBlock *PreHeader = CLI->getPreheader();
1634 Builder.SetInsertPoint(PreHeader->getTerminator());
1635 Constant *One = ConstantInt::get(IVTy, 1);
1636 Builder.CreateStore(One, PLowerBound);
1637 Value *UpperBound = CLI->getTripCount();
1638 Builder.CreateStore(UpperBound, PUpperBound);
1639 Builder.CreateStore(One, PStride);
1640
1641 BasicBlock *Header = CLI->getHeader();
1642 BasicBlock *Exit = CLI->getExit();
1643 BasicBlock *Cond = CLI->getCond();
1644 InsertPointTy AfterIP = CLI->getAfterIP();
1645
1646 // The CLI will be "broken" in the code below, as the loop is no longer
1647 // a valid canonical loop.
1648
1649 if (!Chunk)
1650 Chunk = One;
1651
1652 Value *ThreadNum = getOrCreateThreadID(SrcLoc);
1653
1654 Constant *SchedulingType =
1655 ConstantInt::get(I32Type, static_cast<int>(SchedType));
1656
1657 // Call the "init" function.
1658 Builder.CreateCall(DynamicInit,
1659 {SrcLoc, ThreadNum, SchedulingType, /* LowerBound */ One,
1660 UpperBound, /* step */ One, Chunk});
1661
1662 // An outer loop around the existing one.
1663 BasicBlock *OuterCond = BasicBlock::Create(
1664 PreHeader->getContext(), Twine(PreHeader->getName()) + ".outer.cond",
1665 PreHeader->getParent());
1666 // This needs to be 32-bit always, so can't use the IVTy Zero above.
1667 Builder.SetInsertPoint(OuterCond, OuterCond->getFirstInsertionPt());
1668 Value *Res =
1669 Builder.CreateCall(DynamicNext, {SrcLoc, ThreadNum, PLastIter,
1670 PLowerBound, PUpperBound, PStride});
1671 Constant *Zero32 = ConstantInt::get(I32Type, 0);
1672 Value *MoreWork = Builder.CreateCmp(CmpInst::ICMP_NE, Res, Zero32);
1673 Value *LowerBound =
1674 Builder.CreateSub(Builder.CreateLoad(IVTy, PLowerBound), One, "lb");
1675 Builder.CreateCondBr(MoreWork, Header, Exit);
1676
1677 // Change PHI-node in loop header to use outer cond rather than preheader,
1678 // and set IV to the LowerBound.
1679 Instruction *Phi = &Header->front();
1680 auto *PI = cast<PHINode>(Phi);
1681 PI->setIncomingBlock(0, OuterCond);
1682 PI->setIncomingValue(0, LowerBound);
1683
1684 // Then set the pre-header to jump to the OuterCond
1685 Instruction *Term = PreHeader->getTerminator();
1686 auto *Br = cast<BranchInst>(Term);
1687 Br->setSuccessor(0, OuterCond);
1688
1689 // Modify the inner condition:
1690 // * Use the UpperBound returned from the DynamicNext call.
1691 // * jump to the loop outer loop when done with one of the inner loops.
1692 Builder.SetInsertPoint(Cond, Cond->getFirstInsertionPt());
1693 UpperBound = Builder.CreateLoad(IVTy, PUpperBound, "ub");
1694 Instruction *Comp = &*Builder.GetInsertPoint();
1695 auto *CI = cast<CmpInst>(Comp);
1696 CI->setOperand(1, UpperBound);
1697 // Redirect the inner exit to branch to outer condition.
1698 Instruction *Branch = &Cond->back();
1699 auto *BI = cast<BranchInst>(Branch);
1700 assert(BI->getSuccessor(1) == Exit);
1701 BI->setSuccessor(1, OuterCond);
1702
1703 // Add the barrier if requested.
1704 if (NeedsBarrier) {
1705 Builder.SetInsertPoint(&Exit->back());
1706 createBarrier(LocationDescription(Builder.saveIP(), DL),
1707 omp::Directive::OMPD_for, /* ForceSimpleCall */ false,
1708 /* CheckCancelFlag */ false);
1709 }
1710
1711 CLI->invalidate();
1712 return AfterIP;
1713 }
1714
1715 /// Make \p Source branch to \p Target.
1716 ///
1717 /// Handles two situations:
1718 /// * \p Source already has an unconditional branch.
1719 /// * \p Source is a degenerate block (no terminator because the BB is
1720 /// the current head of the IR construction).
redirectTo(BasicBlock * Source,BasicBlock * Target,DebugLoc DL)1721 static void redirectTo(BasicBlock *Source, BasicBlock *Target, DebugLoc DL) {
1722 if (Instruction *Term = Source->getTerminator()) {
1723 auto *Br = cast<BranchInst>(Term);
1724 assert(!Br->isConditional() &&
1725 "BB's terminator must be an unconditional branch (or degenerate)");
1726 BasicBlock *Succ = Br->getSuccessor(0);
1727 Succ->removePredecessor(Source, /*KeepOneInputPHIs=*/true);
1728 Br->setSuccessor(0, Target);
1729 return;
1730 }
1731
1732 auto *NewBr = BranchInst::Create(Target, Source);
1733 NewBr->setDebugLoc(DL);
1734 }
1735
1736 /// Redirect all edges that branch to \p OldTarget to \p NewTarget. That is,
1737 /// after this \p OldTarget will be orphaned.
redirectAllPredecessorsTo(BasicBlock * OldTarget,BasicBlock * NewTarget,DebugLoc DL)1738 static void redirectAllPredecessorsTo(BasicBlock *OldTarget,
1739 BasicBlock *NewTarget, DebugLoc DL) {
1740 for (BasicBlock *Pred : make_early_inc_range(predecessors(OldTarget)))
1741 redirectTo(Pred, NewTarget, DL);
1742 }
1743
1744 /// Determine which blocks in \p BBs are reachable from outside and remove the
1745 /// ones that are not reachable from the function.
removeUnusedBlocksFromParent(ArrayRef<BasicBlock * > BBs)1746 static void removeUnusedBlocksFromParent(ArrayRef<BasicBlock *> BBs) {
1747 SmallPtrSet<BasicBlock *, 6> BBsToErase{BBs.begin(), BBs.end()};
1748 auto HasRemainingUses = [&BBsToErase](BasicBlock *BB) {
1749 for (Use &U : BB->uses()) {
1750 auto *UseInst = dyn_cast<Instruction>(U.getUser());
1751 if (!UseInst)
1752 continue;
1753 if (BBsToErase.count(UseInst->getParent()))
1754 continue;
1755 return true;
1756 }
1757 return false;
1758 };
1759
1760 while (true) {
1761 bool Changed = false;
1762 for (BasicBlock *BB : make_early_inc_range(BBsToErase)) {
1763 if (HasRemainingUses(BB)) {
1764 BBsToErase.erase(BB);
1765 Changed = true;
1766 }
1767 }
1768 if (!Changed)
1769 break;
1770 }
1771
1772 SmallVector<BasicBlock *, 7> BBVec(BBsToErase.begin(), BBsToErase.end());
1773 DeleteDeadBlocks(BBVec);
1774 }
1775
1776 CanonicalLoopInfo *
collapseLoops(DebugLoc DL,ArrayRef<CanonicalLoopInfo * > Loops,InsertPointTy ComputeIP)1777 OpenMPIRBuilder::collapseLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops,
1778 InsertPointTy ComputeIP) {
1779 assert(Loops.size() >= 1 && "At least one loop required");
1780 size_t NumLoops = Loops.size();
1781
1782 // Nothing to do if there is already just one loop.
1783 if (NumLoops == 1)
1784 return Loops.front();
1785
1786 CanonicalLoopInfo *Outermost = Loops.front();
1787 CanonicalLoopInfo *Innermost = Loops.back();
1788 BasicBlock *OrigPreheader = Outermost->getPreheader();
1789 BasicBlock *OrigAfter = Outermost->getAfter();
1790 Function *F = OrigPreheader->getParent();
1791
1792 // Setup the IRBuilder for inserting the trip count computation.
1793 Builder.SetCurrentDebugLocation(DL);
1794 if (ComputeIP.isSet())
1795 Builder.restoreIP(ComputeIP);
1796 else
1797 Builder.restoreIP(Outermost->getPreheaderIP());
1798
1799 // Derive the collapsed' loop trip count.
1800 // TODO: Find common/largest indvar type.
1801 Value *CollapsedTripCount = nullptr;
1802 for (CanonicalLoopInfo *L : Loops) {
1803 assert(L->isValid() &&
1804 "All loops to collapse must be valid canonical loops");
1805 Value *OrigTripCount = L->getTripCount();
1806 if (!CollapsedTripCount) {
1807 CollapsedTripCount = OrigTripCount;
1808 continue;
1809 }
1810
1811 // TODO: Enable UndefinedSanitizer to diagnose an overflow here.
1812 CollapsedTripCount = Builder.CreateMul(CollapsedTripCount, OrigTripCount,
1813 {}, /*HasNUW=*/true);
1814 }
1815
1816 // Create the collapsed loop control flow.
1817 CanonicalLoopInfo *Result =
1818 createLoopSkeleton(DL, CollapsedTripCount, F,
1819 OrigPreheader->getNextNode(), OrigAfter, "collapsed");
1820
1821 // Build the collapsed loop body code.
1822 // Start with deriving the input loop induction variables from the collapsed
1823 // one, using a divmod scheme. To preserve the original loops' order, the
1824 // innermost loop use the least significant bits.
1825 Builder.restoreIP(Result->getBodyIP());
1826
1827 Value *Leftover = Result->getIndVar();
1828 SmallVector<Value *> NewIndVars;
1829 NewIndVars.set_size(NumLoops);
1830 for (int i = NumLoops - 1; i >= 1; --i) {
1831 Value *OrigTripCount = Loops[i]->getTripCount();
1832
1833 Value *NewIndVar = Builder.CreateURem(Leftover, OrigTripCount);
1834 NewIndVars[i] = NewIndVar;
1835
1836 Leftover = Builder.CreateUDiv(Leftover, OrigTripCount);
1837 }
1838 // Outermost loop gets all the remaining bits.
1839 NewIndVars[0] = Leftover;
1840
1841 // Construct the loop body control flow.
1842 // We progressively construct the branch structure following in direction of
1843 // the control flow, from the leading in-between code, the loop nest body, the
1844 // trailing in-between code, and rejoining the collapsed loop's latch.
1845 // ContinueBlock and ContinuePred keep track of the source(s) of next edge. If
1846 // the ContinueBlock is set, continue with that block. If ContinuePred, use
1847 // its predecessors as sources.
1848 BasicBlock *ContinueBlock = Result->getBody();
1849 BasicBlock *ContinuePred = nullptr;
1850 auto ContinueWith = [&ContinueBlock, &ContinuePred, DL](BasicBlock *Dest,
1851 BasicBlock *NextSrc) {
1852 if (ContinueBlock)
1853 redirectTo(ContinueBlock, Dest, DL);
1854 else
1855 redirectAllPredecessorsTo(ContinuePred, Dest, DL);
1856
1857 ContinueBlock = nullptr;
1858 ContinuePred = NextSrc;
1859 };
1860
1861 // The code before the nested loop of each level.
1862 // Because we are sinking it into the nest, it will be executed more often
1863 // that the original loop. More sophisticated schemes could keep track of what
1864 // the in-between code is and instantiate it only once per thread.
1865 for (size_t i = 0; i < NumLoops - 1; ++i)
1866 ContinueWith(Loops[i]->getBody(), Loops[i + 1]->getHeader());
1867
1868 // Connect the loop nest body.
1869 ContinueWith(Innermost->getBody(), Innermost->getLatch());
1870
1871 // The code after the nested loop at each level.
1872 for (size_t i = NumLoops - 1; i > 0; --i)
1873 ContinueWith(Loops[i]->getAfter(), Loops[i - 1]->getLatch());
1874
1875 // Connect the finished loop to the collapsed loop latch.
1876 ContinueWith(Result->getLatch(), nullptr);
1877
1878 // Replace the input loops with the new collapsed loop.
1879 redirectTo(Outermost->getPreheader(), Result->getPreheader(), DL);
1880 redirectTo(Result->getAfter(), Outermost->getAfter(), DL);
1881
1882 // Replace the input loop indvars with the derived ones.
1883 for (size_t i = 0; i < NumLoops; ++i)
1884 Loops[i]->getIndVar()->replaceAllUsesWith(NewIndVars[i]);
1885
1886 // Remove unused parts of the input loops.
1887 SmallVector<BasicBlock *, 12> OldControlBBs;
1888 OldControlBBs.reserve(6 * Loops.size());
1889 for (CanonicalLoopInfo *Loop : Loops)
1890 Loop->collectControlBlocks(OldControlBBs);
1891 removeUnusedBlocksFromParent(OldControlBBs);
1892
1893 for (CanonicalLoopInfo *L : Loops)
1894 L->invalidate();
1895
1896 #ifndef NDEBUG
1897 Result->assertOK();
1898 #endif
1899 return Result;
1900 }
1901
1902 std::vector<CanonicalLoopInfo *>
tileLoops(DebugLoc DL,ArrayRef<CanonicalLoopInfo * > Loops,ArrayRef<Value * > TileSizes)1903 OpenMPIRBuilder::tileLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops,
1904 ArrayRef<Value *> TileSizes) {
1905 assert(TileSizes.size() == Loops.size() &&
1906 "Must pass as many tile sizes as there are loops");
1907 int NumLoops = Loops.size();
1908 assert(NumLoops >= 1 && "At least one loop to tile required");
1909
1910 CanonicalLoopInfo *OutermostLoop = Loops.front();
1911 CanonicalLoopInfo *InnermostLoop = Loops.back();
1912 Function *F = OutermostLoop->getBody()->getParent();
1913 BasicBlock *InnerEnter = InnermostLoop->getBody();
1914 BasicBlock *InnerLatch = InnermostLoop->getLatch();
1915
1916 // Collect original trip counts and induction variable to be accessible by
1917 // index. Also, the structure of the original loops is not preserved during
1918 // the construction of the tiled loops, so do it before we scavenge the BBs of
1919 // any original CanonicalLoopInfo.
1920 SmallVector<Value *, 4> OrigTripCounts, OrigIndVars;
1921 for (CanonicalLoopInfo *L : Loops) {
1922 assert(L->isValid() && "All input loops must be valid canonical loops");
1923 OrigTripCounts.push_back(L->getTripCount());
1924 OrigIndVars.push_back(L->getIndVar());
1925 }
1926
1927 // Collect the code between loop headers. These may contain SSA definitions
1928 // that are used in the loop nest body. To be usable with in the innermost
1929 // body, these BasicBlocks will be sunk into the loop nest body. That is,
1930 // these instructions may be executed more often than before the tiling.
1931 // TODO: It would be sufficient to only sink them into body of the
1932 // corresponding tile loop.
1933 SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> InbetweenCode;
1934 for (int i = 0; i < NumLoops - 1; ++i) {
1935 CanonicalLoopInfo *Surrounding = Loops[i];
1936 CanonicalLoopInfo *Nested = Loops[i + 1];
1937
1938 BasicBlock *EnterBB = Surrounding->getBody();
1939 BasicBlock *ExitBB = Nested->getHeader();
1940 InbetweenCode.emplace_back(EnterBB, ExitBB);
1941 }
1942
1943 // Compute the trip counts of the floor loops.
1944 Builder.SetCurrentDebugLocation(DL);
1945 Builder.restoreIP(OutermostLoop->getPreheaderIP());
1946 SmallVector<Value *, 4> FloorCount, FloorRems;
1947 for (int i = 0; i < NumLoops; ++i) {
1948 Value *TileSize = TileSizes[i];
1949 Value *OrigTripCount = OrigTripCounts[i];
1950 Type *IVType = OrigTripCount->getType();
1951
1952 Value *FloorTripCount = Builder.CreateUDiv(OrigTripCount, TileSize);
1953 Value *FloorTripRem = Builder.CreateURem(OrigTripCount, TileSize);
1954
1955 // 0 if tripcount divides the tilesize, 1 otherwise.
1956 // 1 means we need an additional iteration for a partial tile.
1957 //
1958 // Unfortunately we cannot just use the roundup-formula
1959 // (tripcount + tilesize - 1)/tilesize
1960 // because the summation might overflow. We do not want introduce undefined
1961 // behavior when the untiled loop nest did not.
1962 Value *FloorTripOverflow =
1963 Builder.CreateICmpNE(FloorTripRem, ConstantInt::get(IVType, 0));
1964
1965 FloorTripOverflow = Builder.CreateZExt(FloorTripOverflow, IVType);
1966 FloorTripCount =
1967 Builder.CreateAdd(FloorTripCount, FloorTripOverflow,
1968 "omp_floor" + Twine(i) + ".tripcount", true);
1969
1970 // Remember some values for later use.
1971 FloorCount.push_back(FloorTripCount);
1972 FloorRems.push_back(FloorTripRem);
1973 }
1974
1975 // Generate the new loop nest, from the outermost to the innermost.
1976 std::vector<CanonicalLoopInfo *> Result;
1977 Result.reserve(NumLoops * 2);
1978
1979 // The basic block of the surrounding loop that enters the nest generated
1980 // loop.
1981 BasicBlock *Enter = OutermostLoop->getPreheader();
1982
1983 // The basic block of the surrounding loop where the inner code should
1984 // continue.
1985 BasicBlock *Continue = OutermostLoop->getAfter();
1986
1987 // Where the next loop basic block should be inserted.
1988 BasicBlock *OutroInsertBefore = InnermostLoop->getExit();
1989
1990 auto EmbeddNewLoop =
1991 [this, DL, F, InnerEnter, &Enter, &Continue, &OutroInsertBefore](
1992 Value *TripCount, const Twine &Name) -> CanonicalLoopInfo * {
1993 CanonicalLoopInfo *EmbeddedLoop = createLoopSkeleton(
1994 DL, TripCount, F, InnerEnter, OutroInsertBefore, Name);
1995 redirectTo(Enter, EmbeddedLoop->getPreheader(), DL);
1996 redirectTo(EmbeddedLoop->getAfter(), Continue, DL);
1997
1998 // Setup the position where the next embedded loop connects to this loop.
1999 Enter = EmbeddedLoop->getBody();
2000 Continue = EmbeddedLoop->getLatch();
2001 OutroInsertBefore = EmbeddedLoop->getLatch();
2002 return EmbeddedLoop;
2003 };
2004
2005 auto EmbeddNewLoops = [&Result, &EmbeddNewLoop](ArrayRef<Value *> TripCounts,
2006 const Twine &NameBase) {
2007 for (auto P : enumerate(TripCounts)) {
2008 CanonicalLoopInfo *EmbeddedLoop =
2009 EmbeddNewLoop(P.value(), NameBase + Twine(P.index()));
2010 Result.push_back(EmbeddedLoop);
2011 }
2012 };
2013
2014 EmbeddNewLoops(FloorCount, "floor");
2015
2016 // Within the innermost floor loop, emit the code that computes the tile
2017 // sizes.
2018 Builder.SetInsertPoint(Enter->getTerminator());
2019 SmallVector<Value *, 4> TileCounts;
2020 for (int i = 0; i < NumLoops; ++i) {
2021 CanonicalLoopInfo *FloorLoop = Result[i];
2022 Value *TileSize = TileSizes[i];
2023
2024 Value *FloorIsEpilogue =
2025 Builder.CreateICmpEQ(FloorLoop->getIndVar(), FloorCount[i]);
2026 Value *TileTripCount =
2027 Builder.CreateSelect(FloorIsEpilogue, FloorRems[i], TileSize);
2028
2029 TileCounts.push_back(TileTripCount);
2030 }
2031
2032 // Create the tile loops.
2033 EmbeddNewLoops(TileCounts, "tile");
2034
2035 // Insert the inbetween code into the body.
2036 BasicBlock *BodyEnter = Enter;
2037 BasicBlock *BodyEntered = nullptr;
2038 for (std::pair<BasicBlock *, BasicBlock *> P : InbetweenCode) {
2039 BasicBlock *EnterBB = P.first;
2040 BasicBlock *ExitBB = P.second;
2041
2042 if (BodyEnter)
2043 redirectTo(BodyEnter, EnterBB, DL);
2044 else
2045 redirectAllPredecessorsTo(BodyEntered, EnterBB, DL);
2046
2047 BodyEnter = nullptr;
2048 BodyEntered = ExitBB;
2049 }
2050
2051 // Append the original loop nest body into the generated loop nest body.
2052 if (BodyEnter)
2053 redirectTo(BodyEnter, InnerEnter, DL);
2054 else
2055 redirectAllPredecessorsTo(BodyEntered, InnerEnter, DL);
2056 redirectAllPredecessorsTo(InnerLatch, Continue, DL);
2057
2058 // Replace the original induction variable with an induction variable computed
2059 // from the tile and floor induction variables.
2060 Builder.restoreIP(Result.back()->getBodyIP());
2061 for (int i = 0; i < NumLoops; ++i) {
2062 CanonicalLoopInfo *FloorLoop = Result[i];
2063 CanonicalLoopInfo *TileLoop = Result[NumLoops + i];
2064 Value *OrigIndVar = OrigIndVars[i];
2065 Value *Size = TileSizes[i];
2066
2067 Value *Scale =
2068 Builder.CreateMul(Size, FloorLoop->getIndVar(), {}, /*HasNUW=*/true);
2069 Value *Shift =
2070 Builder.CreateAdd(Scale, TileLoop->getIndVar(), {}, /*HasNUW=*/true);
2071 OrigIndVar->replaceAllUsesWith(Shift);
2072 }
2073
2074 // Remove unused parts of the original loops.
2075 SmallVector<BasicBlock *, 12> OldControlBBs;
2076 OldControlBBs.reserve(6 * Loops.size());
2077 for (CanonicalLoopInfo *Loop : Loops)
2078 Loop->collectControlBlocks(OldControlBBs);
2079 removeUnusedBlocksFromParent(OldControlBBs);
2080
2081 for (CanonicalLoopInfo *L : Loops)
2082 L->invalidate();
2083
2084 #ifndef NDEBUG
2085 for (CanonicalLoopInfo *GenL : Result)
2086 GenL->assertOK();
2087 #endif
2088 return Result;
2089 }
2090
2091 /// Attach loop metadata \p Properties to the loop described by \p Loop. If the
2092 /// loop already has metadata, the loop properties are appended.
addLoopMetadata(CanonicalLoopInfo * Loop,ArrayRef<Metadata * > Properties)2093 static void addLoopMetadata(CanonicalLoopInfo *Loop,
2094 ArrayRef<Metadata *> Properties) {
2095 assert(Loop->isValid() && "Expecting a valid CanonicalLoopInfo");
2096
2097 // Nothing to do if no property to attach.
2098 if (Properties.empty())
2099 return;
2100
2101 LLVMContext &Ctx = Loop->getFunction()->getContext();
2102 SmallVector<Metadata *> NewLoopProperties;
2103 NewLoopProperties.push_back(nullptr);
2104
2105 // If the loop already has metadata, prepend it to the new metadata.
2106 BasicBlock *Latch = Loop->getLatch();
2107 assert(Latch && "A valid CanonicalLoopInfo must have a unique latch");
2108 MDNode *Existing = Latch->getTerminator()->getMetadata(LLVMContext::MD_loop);
2109 if (Existing)
2110 append_range(NewLoopProperties, drop_begin(Existing->operands(), 1));
2111
2112 append_range(NewLoopProperties, Properties);
2113 MDNode *LoopID = MDNode::getDistinct(Ctx, NewLoopProperties);
2114 LoopID->replaceOperandWith(0, LoopID);
2115
2116 Latch->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID);
2117 }
2118
unrollLoopFull(DebugLoc,CanonicalLoopInfo * Loop)2119 void OpenMPIRBuilder::unrollLoopFull(DebugLoc, CanonicalLoopInfo *Loop) {
2120 LLVMContext &Ctx = Builder.getContext();
2121 addLoopMetadata(
2122 Loop, {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
2123 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.full"))});
2124 }
2125
unrollLoopHeuristic(DebugLoc,CanonicalLoopInfo * Loop)2126 void OpenMPIRBuilder::unrollLoopHeuristic(DebugLoc, CanonicalLoopInfo *Loop) {
2127 LLVMContext &Ctx = Builder.getContext();
2128 addLoopMetadata(
2129 Loop, {
2130 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
2131 });
2132 }
2133
2134 /// Create the TargetMachine object to query the backend for optimization
2135 /// preferences.
2136 ///
2137 /// Ideally, this would be passed from the front-end to the OpenMPBuilder, but
2138 /// e.g. Clang does not pass it to its CodeGen layer and creates it only when
2139 /// needed for the LLVM pass pipline. We use some default options to avoid
2140 /// having to pass too many settings from the frontend that probably do not
2141 /// matter.
2142 ///
2143 /// Currently, TargetMachine is only used sometimes by the unrollLoopPartial
2144 /// method. If we are going to use TargetMachine for more purposes, especially
2145 /// those that are sensitive to TargetOptions, RelocModel and CodeModel, it
2146 /// might become be worth requiring front-ends to pass on their TargetMachine,
2147 /// or at least cache it between methods. Note that while fontends such as Clang
2148 /// have just a single main TargetMachine per translation unit, "target-cpu" and
2149 /// "target-features" that determine the TargetMachine are per-function and can
2150 /// be overrided using __attribute__((target("OPTIONS"))).
2151 static std::unique_ptr<TargetMachine>
createTargetMachine(Function * F,CodeGenOpt::Level OptLevel)2152 createTargetMachine(Function *F, CodeGenOpt::Level OptLevel) {
2153 Module *M = F->getParent();
2154
2155 StringRef CPU = F->getFnAttribute("target-cpu").getValueAsString();
2156 StringRef Features = F->getFnAttribute("target-features").getValueAsString();
2157 const std::string &Triple = M->getTargetTriple();
2158
2159 std::string Error;
2160 const llvm::Target *TheTarget = TargetRegistry::lookupTarget(Triple, Error);
2161 if (!TheTarget)
2162 return {};
2163
2164 llvm::TargetOptions Options;
2165 return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
2166 Triple, CPU, Features, Options, /*RelocModel=*/None, /*CodeModel=*/None,
2167 OptLevel));
2168 }
2169
2170 /// Heuristically determine the best-performant unroll factor for \p CLI. This
2171 /// depends on the target processor. We are re-using the same heuristics as the
2172 /// LoopUnrollPass.
computeHeuristicUnrollFactor(CanonicalLoopInfo * CLI)2173 static int32_t computeHeuristicUnrollFactor(CanonicalLoopInfo *CLI) {
2174 Function *F = CLI->getFunction();
2175
2176 // Assume the user requests the most aggressive unrolling, even if the rest of
2177 // the code is optimized using a lower setting.
2178 CodeGenOpt::Level OptLevel = CodeGenOpt::Aggressive;
2179 std::unique_ptr<TargetMachine> TM = createTargetMachine(F, OptLevel);
2180
2181 FunctionAnalysisManager FAM;
2182 FAM.registerPass([]() { return TargetLibraryAnalysis(); });
2183 FAM.registerPass([]() { return AssumptionAnalysis(); });
2184 FAM.registerPass([]() { return DominatorTreeAnalysis(); });
2185 FAM.registerPass([]() { return LoopAnalysis(); });
2186 FAM.registerPass([]() { return ScalarEvolutionAnalysis(); });
2187 FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
2188 TargetIRAnalysis TIRA;
2189 if (TM)
2190 TIRA = TargetIRAnalysis(
2191 [&](const Function &F) { return TM->getTargetTransformInfo(F); });
2192 FAM.registerPass([&]() { return TIRA; });
2193
2194 TargetIRAnalysis::Result &&TTI = TIRA.run(*F, FAM);
2195 ScalarEvolutionAnalysis SEA;
2196 ScalarEvolution &&SE = SEA.run(*F, FAM);
2197 DominatorTreeAnalysis DTA;
2198 DominatorTree &&DT = DTA.run(*F, FAM);
2199 LoopAnalysis LIA;
2200 LoopInfo &&LI = LIA.run(*F, FAM);
2201 AssumptionAnalysis ACT;
2202 AssumptionCache &&AC = ACT.run(*F, FAM);
2203 OptimizationRemarkEmitter ORE{F};
2204
2205 Loop *L = LI.getLoopFor(CLI->getHeader());
2206 assert(L && "Expecting CanonicalLoopInfo to be recognized as a loop");
2207
2208 TargetTransformInfo::UnrollingPreferences UP =
2209 gatherUnrollingPreferences(L, SE, TTI,
2210 /*BlockFrequencyInfo=*/nullptr,
2211 /*ProfileSummaryInfo=*/nullptr, ORE, OptLevel,
2212 /*UserThreshold=*/None,
2213 /*UserCount=*/None,
2214 /*UserAllowPartial=*/true,
2215 /*UserAllowRuntime=*/true,
2216 /*UserUpperBound=*/None,
2217 /*UserFullUnrollMaxCount=*/None);
2218
2219 UP.Force = true;
2220
2221 // Account for additional optimizations taking place before the LoopUnrollPass
2222 // would unroll the loop.
2223 UP.Threshold *= UnrollThresholdFactor;
2224 UP.PartialThreshold *= UnrollThresholdFactor;
2225
2226 // Use normal unroll factors even if the rest of the code is optimized for
2227 // size.
2228 UP.OptSizeThreshold = UP.Threshold;
2229 UP.PartialOptSizeThreshold = UP.PartialThreshold;
2230
2231 LLVM_DEBUG(dbgs() << "Unroll heuristic thresholds:\n"
2232 << " Threshold=" << UP.Threshold << "\n"
2233 << " PartialThreshold=" << UP.PartialThreshold << "\n"
2234 << " OptSizeThreshold=" << UP.OptSizeThreshold << "\n"
2235 << " PartialOptSizeThreshold="
2236 << UP.PartialOptSizeThreshold << "\n");
2237
2238 // Disable peeling.
2239 TargetTransformInfo::PeelingPreferences PP =
2240 gatherPeelingPreferences(L, SE, TTI,
2241 /*UserAllowPeeling=*/false,
2242 /*UserAllowProfileBasedPeeling=*/false,
2243 /*UserUnrollingSpecficValues=*/false);
2244
2245 SmallPtrSet<const Value *, 32> EphValues;
2246 CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
2247
2248 // Assume that reads and writes to stack variables can be eliminated by
2249 // Mem2Reg, SROA or LICM. That is, don't count them towards the loop body's
2250 // size.
2251 for (BasicBlock *BB : L->blocks()) {
2252 for (Instruction &I : *BB) {
2253 Value *Ptr;
2254 if (auto *Load = dyn_cast<LoadInst>(&I)) {
2255 Ptr = Load->getPointerOperand();
2256 } else if (auto *Store = dyn_cast<StoreInst>(&I)) {
2257 Ptr = Store->getPointerOperand();
2258 } else
2259 continue;
2260
2261 Ptr = Ptr->stripPointerCasts();
2262
2263 if (auto *Alloca = dyn_cast<AllocaInst>(Ptr)) {
2264 if (Alloca->getParent() == &F->getEntryBlock())
2265 EphValues.insert(&I);
2266 }
2267 }
2268 }
2269
2270 unsigned NumInlineCandidates;
2271 bool NotDuplicatable;
2272 bool Convergent;
2273 unsigned LoopSize =
2274 ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent,
2275 TTI, EphValues, UP.BEInsns);
2276 LLVM_DEBUG(dbgs() << "Estimated loop size is " << LoopSize << "\n");
2277
2278 // Loop is not unrollable if the loop contains certain instructions.
2279 if (NotDuplicatable || Convergent) {
2280 LLVM_DEBUG(dbgs() << "Loop not considered unrollable\n");
2281 return 1;
2282 }
2283
2284 // TODO: Determine trip count of \p CLI if constant, computeUnrollCount might
2285 // be able to use it.
2286 int TripCount = 0;
2287 int MaxTripCount = 0;
2288 bool MaxOrZero = false;
2289 unsigned TripMultiple = 0;
2290
2291 bool UseUpperBound = false;
2292 computeUnrollCount(L, TTI, DT, &LI, SE, EphValues, &ORE, TripCount,
2293 MaxTripCount, MaxOrZero, TripMultiple, LoopSize, UP, PP,
2294 UseUpperBound);
2295 unsigned Factor = UP.Count;
2296 LLVM_DEBUG(dbgs() << "Suggesting unroll factor of " << Factor << "\n");
2297
2298 // This function returns 1 to signal to not unroll a loop.
2299 if (Factor == 0)
2300 return 1;
2301 return Factor;
2302 }
2303
unrollLoopPartial(DebugLoc DL,CanonicalLoopInfo * Loop,int32_t Factor,CanonicalLoopInfo ** UnrolledCLI)2304 void OpenMPIRBuilder::unrollLoopPartial(DebugLoc DL, CanonicalLoopInfo *Loop,
2305 int32_t Factor,
2306 CanonicalLoopInfo **UnrolledCLI) {
2307 assert(Factor >= 0 && "Unroll factor must not be negative");
2308
2309 Function *F = Loop->getFunction();
2310 LLVMContext &Ctx = F->getContext();
2311
2312 // If the unrolled loop is not used for another loop-associated directive, it
2313 // is sufficient to add metadata for the LoopUnrollPass.
2314 if (!UnrolledCLI) {
2315 SmallVector<Metadata *, 2> LoopMetadata;
2316 LoopMetadata.push_back(
2317 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")));
2318
2319 if (Factor >= 1) {
2320 ConstantAsMetadata *FactorConst = ConstantAsMetadata::get(
2321 ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor)));
2322 LoopMetadata.push_back(MDNode::get(
2323 Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst}));
2324 }
2325
2326 addLoopMetadata(Loop, LoopMetadata);
2327 return;
2328 }
2329
2330 // Heuristically determine the unroll factor.
2331 if (Factor == 0)
2332 Factor = computeHeuristicUnrollFactor(Loop);
2333
2334 // No change required with unroll factor 1.
2335 if (Factor == 1) {
2336 *UnrolledCLI = Loop;
2337 return;
2338 }
2339
2340 assert(Factor >= 2 &&
2341 "unrolling only makes sense with a factor of 2 or larger");
2342
2343 Type *IndVarTy = Loop->getIndVarType();
2344
2345 // Apply partial unrolling by tiling the loop by the unroll-factor, then fully
2346 // unroll the inner loop.
2347 Value *FactorVal =
2348 ConstantInt::get(IndVarTy, APInt(IndVarTy->getIntegerBitWidth(), Factor,
2349 /*isSigned=*/false));
2350 std::vector<CanonicalLoopInfo *> LoopNest =
2351 tileLoops(DL, {Loop}, {FactorVal});
2352 assert(LoopNest.size() == 2 && "Expect 2 loops after tiling");
2353 *UnrolledCLI = LoopNest[0];
2354 CanonicalLoopInfo *InnerLoop = LoopNest[1];
2355
2356 // LoopUnrollPass can only fully unroll loops with constant trip count.
2357 // Unroll by the unroll factor with a fallback epilog for the remainder
2358 // iterations if necessary.
2359 ConstantAsMetadata *FactorConst = ConstantAsMetadata::get(
2360 ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor)));
2361 addLoopMetadata(
2362 InnerLoop,
2363 {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
2364 MDNode::get(
2365 Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst})});
2366
2367 #ifndef NDEBUG
2368 (*UnrolledCLI)->assertOK();
2369 #endif
2370 }
2371
2372 OpenMPIRBuilder::InsertPointTy
createCopyPrivate(const LocationDescription & Loc,llvm::Value * BufSize,llvm::Value * CpyBuf,llvm::Value * CpyFn,llvm::Value * DidIt)2373 OpenMPIRBuilder::createCopyPrivate(const LocationDescription &Loc,
2374 llvm::Value *BufSize, llvm::Value *CpyBuf,
2375 llvm::Value *CpyFn, llvm::Value *DidIt) {
2376 if (!updateToLocation(Loc))
2377 return Loc.IP;
2378
2379 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2380 Value *Ident = getOrCreateIdent(SrcLocStr);
2381 Value *ThreadId = getOrCreateThreadID(Ident);
2382
2383 llvm::Value *DidItLD = Builder.CreateLoad(Builder.getInt32Ty(), DidIt);
2384
2385 Value *Args[] = {Ident, ThreadId, BufSize, CpyBuf, CpyFn, DidItLD};
2386
2387 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_copyprivate);
2388 Builder.CreateCall(Fn, Args);
2389
2390 return Builder.saveIP();
2391 }
2392
2393 OpenMPIRBuilder::InsertPointTy
createSingle(const LocationDescription & Loc,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB,llvm::Value * DidIt)2394 OpenMPIRBuilder::createSingle(const LocationDescription &Loc,
2395 BodyGenCallbackTy BodyGenCB,
2396 FinalizeCallbackTy FiniCB, llvm::Value *DidIt) {
2397
2398 if (!updateToLocation(Loc))
2399 return Loc.IP;
2400
2401 // If needed (i.e. not null), initialize `DidIt` with 0
2402 if (DidIt) {
2403 Builder.CreateStore(Builder.getInt32(0), DidIt);
2404 }
2405
2406 Directive OMPD = Directive::OMPD_single;
2407 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2408 Value *Ident = getOrCreateIdent(SrcLocStr);
2409 Value *ThreadId = getOrCreateThreadID(Ident);
2410 Value *Args[] = {Ident, ThreadId};
2411
2412 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_single);
2413 Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args);
2414
2415 Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_single);
2416 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args);
2417
2418 // generates the following:
2419 // if (__kmpc_single()) {
2420 // .... single region ...
2421 // __kmpc_end_single
2422 // }
2423
2424 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
2425 /*Conditional*/ true, /*hasFinalize*/ true);
2426 }
2427
createCritical(const LocationDescription & Loc,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB,StringRef CriticalName,Value * HintInst)2428 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCritical(
2429 const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
2430 FinalizeCallbackTy FiniCB, StringRef CriticalName, Value *HintInst) {
2431
2432 if (!updateToLocation(Loc))
2433 return Loc.IP;
2434
2435 Directive OMPD = Directive::OMPD_critical;
2436 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2437 Value *Ident = getOrCreateIdent(SrcLocStr);
2438 Value *ThreadId = getOrCreateThreadID(Ident);
2439 Value *LockVar = getOMPCriticalRegionLock(CriticalName);
2440 Value *Args[] = {Ident, ThreadId, LockVar};
2441
2442 SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), std::end(Args));
2443 Function *RTFn = nullptr;
2444 if (HintInst) {
2445 // Add Hint to entry Args and create call
2446 EnterArgs.push_back(HintInst);
2447 RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical_with_hint);
2448 } else {
2449 RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical);
2450 }
2451 Instruction *EntryCall = Builder.CreateCall(RTFn, EnterArgs);
2452
2453 Function *ExitRTLFn =
2454 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_critical);
2455 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args);
2456
2457 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
2458 /*Conditional*/ false, /*hasFinalize*/ true);
2459 }
2460
2461 OpenMPIRBuilder::InsertPointTy
createOrderedDepend(const LocationDescription & Loc,InsertPointTy AllocaIP,unsigned NumLoops,ArrayRef<llvm::Value * > StoreValues,const Twine & Name,bool IsDependSource)2462 OpenMPIRBuilder::createOrderedDepend(const LocationDescription &Loc,
2463 InsertPointTy AllocaIP, unsigned NumLoops,
2464 ArrayRef<llvm::Value *> StoreValues,
2465 const Twine &Name, bool IsDependSource) {
2466 if (!updateToLocation(Loc))
2467 return Loc.IP;
2468
2469 // Allocate space for vector and generate alloc instruction.
2470 auto *ArrI64Ty = ArrayType::get(Int64, NumLoops);
2471 Builder.restoreIP(AllocaIP);
2472 AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI64Ty, nullptr, Name);
2473 ArgsBase->setAlignment(Align(8));
2474 Builder.restoreIP(Loc.IP);
2475
2476 // Store the index value with offset in depend vector.
2477 for (unsigned I = 0; I < NumLoops; ++I) {
2478 Value *DependAddrGEPIter = Builder.CreateInBoundsGEP(
2479 ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(I)});
2480 Builder.CreateStore(StoreValues[I], DependAddrGEPIter);
2481 }
2482
2483 Value *DependBaseAddrGEP = Builder.CreateInBoundsGEP(
2484 ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(0)});
2485
2486 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2487 Value *Ident = getOrCreateIdent(SrcLocStr);
2488 Value *ThreadId = getOrCreateThreadID(Ident);
2489 Value *Args[] = {Ident, ThreadId, DependBaseAddrGEP};
2490
2491 Function *RTLFn = nullptr;
2492 if (IsDependSource)
2493 RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_post);
2494 else
2495 RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_wait);
2496 Builder.CreateCall(RTLFn, Args);
2497
2498 return Builder.saveIP();
2499 }
2500
createOrderedThreadsSimd(const LocationDescription & Loc,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB,bool IsThreads)2501 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createOrderedThreadsSimd(
2502 const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
2503 FinalizeCallbackTy FiniCB, bool IsThreads) {
2504 if (!updateToLocation(Loc))
2505 return Loc.IP;
2506
2507 Directive OMPD = Directive::OMPD_ordered;
2508 Instruction *EntryCall = nullptr;
2509 Instruction *ExitCall = nullptr;
2510
2511 if (IsThreads) {
2512 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2513 Value *Ident = getOrCreateIdent(SrcLocStr);
2514 Value *ThreadId = getOrCreateThreadID(Ident);
2515 Value *Args[] = {Ident, ThreadId};
2516
2517 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_ordered);
2518 EntryCall = Builder.CreateCall(EntryRTLFn, Args);
2519
2520 Function *ExitRTLFn =
2521 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_ordered);
2522 ExitCall = Builder.CreateCall(ExitRTLFn, Args);
2523 }
2524
2525 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
2526 /*Conditional*/ false, /*hasFinalize*/ true);
2527 }
2528
EmitOMPInlinedRegion(Directive OMPD,Instruction * EntryCall,Instruction * ExitCall,BodyGenCallbackTy BodyGenCB,FinalizeCallbackTy FiniCB,bool Conditional,bool HasFinalize,bool IsCancellable)2529 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::EmitOMPInlinedRegion(
2530 Directive OMPD, Instruction *EntryCall, Instruction *ExitCall,
2531 BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool Conditional,
2532 bool HasFinalize, bool IsCancellable) {
2533
2534 if (HasFinalize)
2535 FinalizationStack.push_back({FiniCB, OMPD, IsCancellable});
2536
2537 // Create inlined region's entry and body blocks, in preparation
2538 // for conditional creation
2539 BasicBlock *EntryBB = Builder.GetInsertBlock();
2540 Instruction *SplitPos = EntryBB->getTerminator();
2541 if (!isa_and_nonnull<BranchInst>(SplitPos))
2542 SplitPos = new UnreachableInst(Builder.getContext(), EntryBB);
2543 BasicBlock *ExitBB = EntryBB->splitBasicBlock(SplitPos, "omp_region.end");
2544 BasicBlock *FiniBB =
2545 EntryBB->splitBasicBlock(EntryBB->getTerminator(), "omp_region.finalize");
2546
2547 Builder.SetInsertPoint(EntryBB->getTerminator());
2548 emitCommonDirectiveEntry(OMPD, EntryCall, ExitBB, Conditional);
2549
2550 // generate body
2551 BodyGenCB(/* AllocaIP */ InsertPointTy(),
2552 /* CodeGenIP */ Builder.saveIP(), *FiniBB);
2553
2554 // If we didn't emit a branch to FiniBB during body generation, it means
2555 // FiniBB is unreachable (e.g. while(1);). stop generating all the
2556 // unreachable blocks, and remove anything we are not going to use.
2557 auto SkipEmittingRegion = FiniBB->hasNPredecessors(0);
2558 if (SkipEmittingRegion) {
2559 FiniBB->eraseFromParent();
2560 ExitCall->eraseFromParent();
2561 // Discard finalization if we have it.
2562 if (HasFinalize) {
2563 assert(!FinalizationStack.empty() &&
2564 "Unexpected finalization stack state!");
2565 FinalizationStack.pop_back();
2566 }
2567 } else {
2568 // emit exit call and do any needed finalization.
2569 auto FinIP = InsertPointTy(FiniBB, FiniBB->getFirstInsertionPt());
2570 assert(FiniBB->getTerminator()->getNumSuccessors() == 1 &&
2571 FiniBB->getTerminator()->getSuccessor(0) == ExitBB &&
2572 "Unexpected control flow graph state!!");
2573 emitCommonDirectiveExit(OMPD, FinIP, ExitCall, HasFinalize);
2574 assert(FiniBB->getUniquePredecessor()->getUniqueSuccessor() == FiniBB &&
2575 "Unexpected Control Flow State!");
2576 MergeBlockIntoPredecessor(FiniBB);
2577 }
2578
2579 // If we are skipping the region of a non conditional, remove the exit
2580 // block, and clear the builder's insertion point.
2581 assert(SplitPos->getParent() == ExitBB &&
2582 "Unexpected Insertion point location!");
2583 if (!Conditional && SkipEmittingRegion) {
2584 ExitBB->eraseFromParent();
2585 Builder.ClearInsertionPoint();
2586 } else {
2587 auto merged = MergeBlockIntoPredecessor(ExitBB);
2588 BasicBlock *ExitPredBB = SplitPos->getParent();
2589 auto InsertBB = merged ? ExitPredBB : ExitBB;
2590 if (!isa_and_nonnull<BranchInst>(SplitPos))
2591 SplitPos->eraseFromParent();
2592 Builder.SetInsertPoint(InsertBB);
2593 }
2594
2595 return Builder.saveIP();
2596 }
2597
emitCommonDirectiveEntry(Directive OMPD,Value * EntryCall,BasicBlock * ExitBB,bool Conditional)2598 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveEntry(
2599 Directive OMPD, Value *EntryCall, BasicBlock *ExitBB, bool Conditional) {
2600 // if nothing to do, Return current insertion point.
2601 if (!Conditional || !EntryCall)
2602 return Builder.saveIP();
2603
2604 BasicBlock *EntryBB = Builder.GetInsertBlock();
2605 Value *CallBool = Builder.CreateIsNotNull(EntryCall);
2606 auto *ThenBB = BasicBlock::Create(M.getContext(), "omp_region.body");
2607 auto *UI = new UnreachableInst(Builder.getContext(), ThenBB);
2608
2609 // Emit thenBB and set the Builder's insertion point there for
2610 // body generation next. Place the block after the current block.
2611 Function *CurFn = EntryBB->getParent();
2612 CurFn->getBasicBlockList().insertAfter(EntryBB->getIterator(), ThenBB);
2613
2614 // Move Entry branch to end of ThenBB, and replace with conditional
2615 // branch (If-stmt)
2616 Instruction *EntryBBTI = EntryBB->getTerminator();
2617 Builder.CreateCondBr(CallBool, ThenBB, ExitBB);
2618 EntryBBTI->removeFromParent();
2619 Builder.SetInsertPoint(UI);
2620 Builder.Insert(EntryBBTI);
2621 UI->eraseFromParent();
2622 Builder.SetInsertPoint(ThenBB->getTerminator());
2623
2624 // return an insertion point to ExitBB.
2625 return IRBuilder<>::InsertPoint(ExitBB, ExitBB->getFirstInsertionPt());
2626 }
2627
emitCommonDirectiveExit(omp::Directive OMPD,InsertPointTy FinIP,Instruction * ExitCall,bool HasFinalize)2628 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveExit(
2629 omp::Directive OMPD, InsertPointTy FinIP, Instruction *ExitCall,
2630 bool HasFinalize) {
2631
2632 Builder.restoreIP(FinIP);
2633
2634 // If there is finalization to do, emit it before the exit call
2635 if (HasFinalize) {
2636 assert(!FinalizationStack.empty() &&
2637 "Unexpected finalization stack state!");
2638
2639 FinalizationInfo Fi = FinalizationStack.pop_back_val();
2640 assert(Fi.DK == OMPD && "Unexpected Directive for Finalization call!");
2641
2642 Fi.FiniCB(FinIP);
2643
2644 BasicBlock *FiniBB = FinIP.getBlock();
2645 Instruction *FiniBBTI = FiniBB->getTerminator();
2646
2647 // set Builder IP for call creation
2648 Builder.SetInsertPoint(FiniBBTI);
2649 }
2650
2651 if (!ExitCall)
2652 return Builder.saveIP();
2653
2654 // place the Exitcall as last instruction before Finalization block terminator
2655 ExitCall->removeFromParent();
2656 Builder.Insert(ExitCall);
2657
2658 return IRBuilder<>::InsertPoint(ExitCall->getParent(),
2659 ExitCall->getIterator());
2660 }
2661
createCopyinClauseBlocks(InsertPointTy IP,Value * MasterAddr,Value * PrivateAddr,llvm::IntegerType * IntPtrTy,bool BranchtoEnd)2662 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCopyinClauseBlocks(
2663 InsertPointTy IP, Value *MasterAddr, Value *PrivateAddr,
2664 llvm::IntegerType *IntPtrTy, bool BranchtoEnd) {
2665 if (!IP.isSet())
2666 return IP;
2667
2668 IRBuilder<>::InsertPointGuard IPG(Builder);
2669
2670 // creates the following CFG structure
2671 // OMP_Entry : (MasterAddr != PrivateAddr)?
2672 // F T
2673 // | \
2674 // | copin.not.master
2675 // | /
2676 // v /
2677 // copyin.not.master.end
2678 // |
2679 // v
2680 // OMP.Entry.Next
2681
2682 BasicBlock *OMP_Entry = IP.getBlock();
2683 Function *CurFn = OMP_Entry->getParent();
2684 BasicBlock *CopyBegin =
2685 BasicBlock::Create(M.getContext(), "copyin.not.master", CurFn);
2686 BasicBlock *CopyEnd = nullptr;
2687
2688 // If entry block is terminated, split to preserve the branch to following
2689 // basic block (i.e. OMP.Entry.Next), otherwise, leave everything as is.
2690 if (isa_and_nonnull<BranchInst>(OMP_Entry->getTerminator())) {
2691 CopyEnd = OMP_Entry->splitBasicBlock(OMP_Entry->getTerminator(),
2692 "copyin.not.master.end");
2693 OMP_Entry->getTerminator()->eraseFromParent();
2694 } else {
2695 CopyEnd =
2696 BasicBlock::Create(M.getContext(), "copyin.not.master.end", CurFn);
2697 }
2698
2699 Builder.SetInsertPoint(OMP_Entry);
2700 Value *MasterPtr = Builder.CreatePtrToInt(MasterAddr, IntPtrTy);
2701 Value *PrivatePtr = Builder.CreatePtrToInt(PrivateAddr, IntPtrTy);
2702 Value *cmp = Builder.CreateICmpNE(MasterPtr, PrivatePtr);
2703 Builder.CreateCondBr(cmp, CopyBegin, CopyEnd);
2704
2705 Builder.SetInsertPoint(CopyBegin);
2706 if (BranchtoEnd)
2707 Builder.SetInsertPoint(Builder.CreateBr(CopyEnd));
2708
2709 return Builder.saveIP();
2710 }
2711
createOMPAlloc(const LocationDescription & Loc,Value * Size,Value * Allocator,std::string Name)2712 CallInst *OpenMPIRBuilder::createOMPAlloc(const LocationDescription &Loc,
2713 Value *Size, Value *Allocator,
2714 std::string Name) {
2715 IRBuilder<>::InsertPointGuard IPG(Builder);
2716 Builder.restoreIP(Loc.IP);
2717
2718 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2719 Value *Ident = getOrCreateIdent(SrcLocStr);
2720 Value *ThreadId = getOrCreateThreadID(Ident);
2721 Value *Args[] = {ThreadId, Size, Allocator};
2722
2723 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_alloc);
2724
2725 return Builder.CreateCall(Fn, Args, Name);
2726 }
2727
createOMPFree(const LocationDescription & Loc,Value * Addr,Value * Allocator,std::string Name)2728 CallInst *OpenMPIRBuilder::createOMPFree(const LocationDescription &Loc,
2729 Value *Addr, Value *Allocator,
2730 std::string Name) {
2731 IRBuilder<>::InsertPointGuard IPG(Builder);
2732 Builder.restoreIP(Loc.IP);
2733
2734 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2735 Value *Ident = getOrCreateIdent(SrcLocStr);
2736 Value *ThreadId = getOrCreateThreadID(Ident);
2737 Value *Args[] = {ThreadId, Addr, Allocator};
2738 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_free);
2739 return Builder.CreateCall(Fn, Args, Name);
2740 }
2741
createCachedThreadPrivate(const LocationDescription & Loc,llvm::Value * Pointer,llvm::ConstantInt * Size,const llvm::Twine & Name)2742 CallInst *OpenMPIRBuilder::createCachedThreadPrivate(
2743 const LocationDescription &Loc, llvm::Value *Pointer,
2744 llvm::ConstantInt *Size, const llvm::Twine &Name) {
2745 IRBuilder<>::InsertPointGuard IPG(Builder);
2746 Builder.restoreIP(Loc.IP);
2747
2748 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2749 Value *Ident = getOrCreateIdent(SrcLocStr);
2750 Value *ThreadId = getOrCreateThreadID(Ident);
2751 Constant *ThreadPrivateCache =
2752 getOrCreateOMPInternalVariable(Int8PtrPtr, Name);
2753 llvm::Value *Args[] = {Ident, ThreadId, Pointer, Size, ThreadPrivateCache};
2754
2755 Function *Fn =
2756 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_threadprivate_cached);
2757
2758 return Builder.CreateCall(Fn, Args);
2759 }
2760
2761 OpenMPIRBuilder::InsertPointTy
createTargetInit(const LocationDescription & Loc,bool IsSPMD,bool RequiresFullRuntime)2762 OpenMPIRBuilder::createTargetInit(const LocationDescription &Loc, bool IsSPMD,
2763 bool RequiresFullRuntime) {
2764 if (!updateToLocation(Loc))
2765 return Loc.IP;
2766
2767 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2768 Value *Ident = getOrCreateIdent(SrcLocStr);
2769 ConstantInt *IsSPMDVal = ConstantInt::getSigned(
2770 IntegerType::getInt8Ty(Int8->getContext()),
2771 IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC);
2772 ConstantInt *UseGenericStateMachine =
2773 ConstantInt::getBool(Int32->getContext(), !IsSPMD);
2774 ConstantInt *RequiresFullRuntimeVal =
2775 ConstantInt::getBool(Int32->getContext(), RequiresFullRuntime);
2776
2777 Function *Fn = getOrCreateRuntimeFunctionPtr(
2778 omp::RuntimeFunction::OMPRTL___kmpc_target_init);
2779
2780 CallInst *ThreadKind = Builder.CreateCall(
2781 Fn, {Ident, IsSPMDVal, UseGenericStateMachine, RequiresFullRuntimeVal});
2782
2783 Value *ExecUserCode = Builder.CreateICmpEQ(
2784 ThreadKind, ConstantInt::get(ThreadKind->getType(), -1),
2785 "exec_user_code");
2786
2787 // ThreadKind = __kmpc_target_init(...)
2788 // if (ThreadKind == -1)
2789 // user_code
2790 // else
2791 // return;
2792
2793 auto *UI = Builder.CreateUnreachable();
2794 BasicBlock *CheckBB = UI->getParent();
2795 BasicBlock *UserCodeEntryBB = CheckBB->splitBasicBlock(UI, "user_code.entry");
2796
2797 BasicBlock *WorkerExitBB = BasicBlock::Create(
2798 CheckBB->getContext(), "worker.exit", CheckBB->getParent());
2799 Builder.SetInsertPoint(WorkerExitBB);
2800 Builder.CreateRetVoid();
2801
2802 auto *CheckBBTI = CheckBB->getTerminator();
2803 Builder.SetInsertPoint(CheckBBTI);
2804 Builder.CreateCondBr(ExecUserCode, UI->getParent(), WorkerExitBB);
2805
2806 CheckBBTI->eraseFromParent();
2807 UI->eraseFromParent();
2808
2809 // Continue in the "user_code" block, see diagram above and in
2810 // openmp/libomptarget/deviceRTLs/common/include/target.h .
2811 return InsertPointTy(UserCodeEntryBB, UserCodeEntryBB->getFirstInsertionPt());
2812 }
2813
createTargetDeinit(const LocationDescription & Loc,bool IsSPMD,bool RequiresFullRuntime)2814 void OpenMPIRBuilder::createTargetDeinit(const LocationDescription &Loc,
2815 bool IsSPMD,
2816 bool RequiresFullRuntime) {
2817 if (!updateToLocation(Loc))
2818 return;
2819
2820 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc);
2821 Value *Ident = getOrCreateIdent(SrcLocStr);
2822 ConstantInt *IsSPMDVal = ConstantInt::getSigned(
2823 IntegerType::getInt8Ty(Int8->getContext()),
2824 IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC);
2825 ConstantInt *RequiresFullRuntimeVal =
2826 ConstantInt::getBool(Int32->getContext(), RequiresFullRuntime);
2827
2828 Function *Fn = getOrCreateRuntimeFunctionPtr(
2829 omp::RuntimeFunction::OMPRTL___kmpc_target_deinit);
2830
2831 Builder.CreateCall(Fn, {Ident, IsSPMDVal, RequiresFullRuntimeVal});
2832 }
2833
getNameWithSeparators(ArrayRef<StringRef> Parts,StringRef FirstSeparator,StringRef Separator)2834 std::string OpenMPIRBuilder::getNameWithSeparators(ArrayRef<StringRef> Parts,
2835 StringRef FirstSeparator,
2836 StringRef Separator) {
2837 SmallString<128> Buffer;
2838 llvm::raw_svector_ostream OS(Buffer);
2839 StringRef Sep = FirstSeparator;
2840 for (StringRef Part : Parts) {
2841 OS << Sep << Part;
2842 Sep = Separator;
2843 }
2844 return OS.str().str();
2845 }
2846
getOrCreateOMPInternalVariable(llvm::Type * Ty,const llvm::Twine & Name,unsigned AddressSpace)2847 Constant *OpenMPIRBuilder::getOrCreateOMPInternalVariable(
2848 llvm::Type *Ty, const llvm::Twine &Name, unsigned AddressSpace) {
2849 // TODO: Replace the twine arg with stringref to get rid of the conversion
2850 // logic. However This is taken from current implementation in clang as is.
2851 // Since this method is used in many places exclusively for OMP internal use
2852 // we will keep it as is for temporarily until we move all users to the
2853 // builder and then, if possible, fix it everywhere in one go.
2854 SmallString<256> Buffer;
2855 llvm::raw_svector_ostream Out(Buffer);
2856 Out << Name;
2857 StringRef RuntimeName = Out.str();
2858 auto &Elem = *InternalVars.try_emplace(RuntimeName, nullptr).first;
2859 if (Elem.second) {
2860 assert(Elem.second->getType()->getPointerElementType() == Ty &&
2861 "OMP internal variable has different type than requested");
2862 } else {
2863 // TODO: investigate the appropriate linkage type used for the global
2864 // variable for possibly changing that to internal or private, or maybe
2865 // create different versions of the function for different OMP internal
2866 // variables.
2867 Elem.second = new llvm::GlobalVariable(
2868 M, Ty, /*IsConstant*/ false, llvm::GlobalValue::CommonLinkage,
2869 llvm::Constant::getNullValue(Ty), Elem.first(),
2870 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
2871 AddressSpace);
2872 }
2873
2874 return Elem.second;
2875 }
2876
getOMPCriticalRegionLock(StringRef CriticalName)2877 Value *OpenMPIRBuilder::getOMPCriticalRegionLock(StringRef CriticalName) {
2878 std::string Prefix = Twine("gomp_critical_user_", CriticalName).str();
2879 std::string Name = getNameWithSeparators({Prefix, "var"}, ".", ".");
2880 return getOrCreateOMPInternalVariable(KmpCriticalNameTy, Name);
2881 }
2882
2883 GlobalVariable *
createOffloadMaptypes(SmallVectorImpl<uint64_t> & Mappings,std::string VarName)2884 OpenMPIRBuilder::createOffloadMaptypes(SmallVectorImpl<uint64_t> &Mappings,
2885 std::string VarName) {
2886 llvm::Constant *MaptypesArrayInit =
2887 llvm::ConstantDataArray::get(M.getContext(), Mappings);
2888 auto *MaptypesArrayGlobal = new llvm::GlobalVariable(
2889 M, MaptypesArrayInit->getType(),
2890 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MaptypesArrayInit,
2891 VarName);
2892 MaptypesArrayGlobal->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2893 return MaptypesArrayGlobal;
2894 }
2895
createMapperAllocas(const LocationDescription & Loc,InsertPointTy AllocaIP,unsigned NumOperands,struct MapperAllocas & MapperAllocas)2896 void OpenMPIRBuilder::createMapperAllocas(const LocationDescription &Loc,
2897 InsertPointTy AllocaIP,
2898 unsigned NumOperands,
2899 struct MapperAllocas &MapperAllocas) {
2900 if (!updateToLocation(Loc))
2901 return;
2902
2903 auto *ArrI8PtrTy = ArrayType::get(Int8Ptr, NumOperands);
2904 auto *ArrI64Ty = ArrayType::get(Int64, NumOperands);
2905 Builder.restoreIP(AllocaIP);
2906 AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI8PtrTy);
2907 AllocaInst *Args = Builder.CreateAlloca(ArrI8PtrTy);
2908 AllocaInst *ArgSizes = Builder.CreateAlloca(ArrI64Ty);
2909 Builder.restoreIP(Loc.IP);
2910 MapperAllocas.ArgsBase = ArgsBase;
2911 MapperAllocas.Args = Args;
2912 MapperAllocas.ArgSizes = ArgSizes;
2913 }
2914
emitMapperCall(const LocationDescription & Loc,Function * MapperFunc,Value * SrcLocInfo,Value * MaptypesArg,Value * MapnamesArg,struct MapperAllocas & MapperAllocas,int64_t DeviceID,unsigned NumOperands)2915 void OpenMPIRBuilder::emitMapperCall(const LocationDescription &Loc,
2916 Function *MapperFunc, Value *SrcLocInfo,
2917 Value *MaptypesArg, Value *MapnamesArg,
2918 struct MapperAllocas &MapperAllocas,
2919 int64_t DeviceID, unsigned NumOperands) {
2920 if (!updateToLocation(Loc))
2921 return;
2922
2923 auto *ArrI8PtrTy = ArrayType::get(Int8Ptr, NumOperands);
2924 auto *ArrI64Ty = ArrayType::get(Int64, NumOperands);
2925 Value *ArgsBaseGEP =
2926 Builder.CreateInBoundsGEP(ArrI8PtrTy, MapperAllocas.ArgsBase,
2927 {Builder.getInt32(0), Builder.getInt32(0)});
2928 Value *ArgsGEP =
2929 Builder.CreateInBoundsGEP(ArrI8PtrTy, MapperAllocas.Args,
2930 {Builder.getInt32(0), Builder.getInt32(0)});
2931 Value *ArgSizesGEP =
2932 Builder.CreateInBoundsGEP(ArrI64Ty, MapperAllocas.ArgSizes,
2933 {Builder.getInt32(0), Builder.getInt32(0)});
2934 Value *NullPtr = Constant::getNullValue(Int8Ptr->getPointerTo());
2935 Builder.CreateCall(MapperFunc,
2936 {SrcLocInfo, Builder.getInt64(DeviceID),
2937 Builder.getInt32(NumOperands), ArgsBaseGEP, ArgsGEP,
2938 ArgSizesGEP, MaptypesArg, MapnamesArg, NullPtr});
2939 }
2940
checkAndEmitFlushAfterAtomic(const LocationDescription & Loc,llvm::AtomicOrdering AO,AtomicKind AK)2941 bool OpenMPIRBuilder::checkAndEmitFlushAfterAtomic(
2942 const LocationDescription &Loc, llvm::AtomicOrdering AO, AtomicKind AK) {
2943 assert(!(AO == AtomicOrdering::NotAtomic ||
2944 AO == llvm::AtomicOrdering::Unordered) &&
2945 "Unexpected Atomic Ordering.");
2946
2947 bool Flush = false;
2948 llvm::AtomicOrdering FlushAO = AtomicOrdering::Monotonic;
2949
2950 switch (AK) {
2951 case Read:
2952 if (AO == AtomicOrdering::Acquire || AO == AtomicOrdering::AcquireRelease ||
2953 AO == AtomicOrdering::SequentiallyConsistent) {
2954 FlushAO = AtomicOrdering::Acquire;
2955 Flush = true;
2956 }
2957 break;
2958 case Write:
2959 case Update:
2960 if (AO == AtomicOrdering::Release || AO == AtomicOrdering::AcquireRelease ||
2961 AO == AtomicOrdering::SequentiallyConsistent) {
2962 FlushAO = AtomicOrdering::Release;
2963 Flush = true;
2964 }
2965 break;
2966 case Capture:
2967 switch (AO) {
2968 case AtomicOrdering::Acquire:
2969 FlushAO = AtomicOrdering::Acquire;
2970 Flush = true;
2971 break;
2972 case AtomicOrdering::Release:
2973 FlushAO = AtomicOrdering::Release;
2974 Flush = true;
2975 break;
2976 case AtomicOrdering::AcquireRelease:
2977 case AtomicOrdering::SequentiallyConsistent:
2978 FlushAO = AtomicOrdering::AcquireRelease;
2979 Flush = true;
2980 break;
2981 default:
2982 // do nothing - leave silently.
2983 break;
2984 }
2985 }
2986
2987 if (Flush) {
2988 // Currently Flush RT call still doesn't take memory_ordering, so for when
2989 // that happens, this tries to do the resolution of which atomic ordering
2990 // to use with but issue the flush call
2991 // TODO: pass `FlushAO` after memory ordering support is added
2992 (void)FlushAO;
2993 emitFlush(Loc);
2994 }
2995
2996 // for AO == AtomicOrdering::Monotonic and all other case combinations
2997 // do nothing
2998 return Flush;
2999 }
3000
3001 OpenMPIRBuilder::InsertPointTy
createAtomicRead(const LocationDescription & Loc,AtomicOpValue & X,AtomicOpValue & V,AtomicOrdering AO)3002 OpenMPIRBuilder::createAtomicRead(const LocationDescription &Loc,
3003 AtomicOpValue &X, AtomicOpValue &V,
3004 AtomicOrdering AO) {
3005 if (!updateToLocation(Loc))
3006 return Loc.IP;
3007
3008 Type *XTy = X.Var->getType();
3009 assert(XTy->isPointerTy() && "OMP Atomic expects a pointer to target memory");
3010 Type *XElemTy = XTy->getPointerElementType();
3011 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
3012 XElemTy->isPointerTy()) &&
3013 "OMP atomic read expected a scalar type");
3014
3015 Value *XRead = nullptr;
3016
3017 if (XElemTy->isIntegerTy()) {
3018 LoadInst *XLD =
3019 Builder.CreateLoad(XElemTy, X.Var, X.IsVolatile, "omp.atomic.read");
3020 XLD->setAtomic(AO);
3021 XRead = cast<Value>(XLD);
3022 } else {
3023 // We need to bitcast and perform atomic op as integer
3024 unsigned Addrspace = cast<PointerType>(XTy)->getAddressSpace();
3025 IntegerType *IntCastTy =
3026 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
3027 Value *XBCast = Builder.CreateBitCast(
3028 X.Var, IntCastTy->getPointerTo(Addrspace), "atomic.src.int.cast");
3029 LoadInst *XLoad =
3030 Builder.CreateLoad(IntCastTy, XBCast, X.IsVolatile, "omp.atomic.load");
3031 XLoad->setAtomic(AO);
3032 if (XElemTy->isFloatingPointTy()) {
3033 XRead = Builder.CreateBitCast(XLoad, XElemTy, "atomic.flt.cast");
3034 } else {
3035 XRead = Builder.CreateIntToPtr(XLoad, XElemTy, "atomic.ptr.cast");
3036 }
3037 }
3038 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Read);
3039 Builder.CreateStore(XRead, V.Var, V.IsVolatile);
3040 return Builder.saveIP();
3041 }
3042
3043 OpenMPIRBuilder::InsertPointTy
createAtomicWrite(const LocationDescription & Loc,AtomicOpValue & X,Value * Expr,AtomicOrdering AO)3044 OpenMPIRBuilder::createAtomicWrite(const LocationDescription &Loc,
3045 AtomicOpValue &X, Value *Expr,
3046 AtomicOrdering AO) {
3047 if (!updateToLocation(Loc))
3048 return Loc.IP;
3049
3050 Type *XTy = X.Var->getType();
3051 assert(XTy->isPointerTy() && "OMP Atomic expects a pointer to target memory");
3052 Type *XElemTy = XTy->getPointerElementType();
3053 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
3054 XElemTy->isPointerTy()) &&
3055 "OMP atomic write expected a scalar type");
3056
3057 if (XElemTy->isIntegerTy()) {
3058 StoreInst *XSt = Builder.CreateStore(Expr, X.Var, X.IsVolatile);
3059 XSt->setAtomic(AO);
3060 } else {
3061 // We need to bitcast and perform atomic op as integers
3062 unsigned Addrspace = cast<PointerType>(XTy)->getAddressSpace();
3063 IntegerType *IntCastTy =
3064 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
3065 Value *XBCast = Builder.CreateBitCast(
3066 X.Var, IntCastTy->getPointerTo(Addrspace), "atomic.dst.int.cast");
3067 Value *ExprCast =
3068 Builder.CreateBitCast(Expr, IntCastTy, "atomic.src.int.cast");
3069 StoreInst *XSt = Builder.CreateStore(ExprCast, XBCast, X.IsVolatile);
3070 XSt->setAtomic(AO);
3071 }
3072
3073 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Write);
3074 return Builder.saveIP();
3075 }
3076
createAtomicUpdate(const LocationDescription & Loc,Instruction * AllocIP,AtomicOpValue & X,Value * Expr,AtomicOrdering AO,AtomicRMWInst::BinOp RMWOp,AtomicUpdateCallbackTy & UpdateOp,bool IsXLHSInRHSPart)3077 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicUpdate(
3078 const LocationDescription &Loc, Instruction *AllocIP, AtomicOpValue &X,
3079 Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
3080 AtomicUpdateCallbackTy &UpdateOp, bool IsXLHSInRHSPart) {
3081 if (!updateToLocation(Loc))
3082 return Loc.IP;
3083
3084 LLVM_DEBUG({
3085 Type *XTy = X.Var->getType();
3086 assert(XTy->isPointerTy() &&
3087 "OMP Atomic expects a pointer to target memory");
3088 Type *XElemTy = XTy->getPointerElementType();
3089 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
3090 XElemTy->isPointerTy()) &&
3091 "OMP atomic update expected a scalar type");
3092 assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) &&
3093 (RMWOp != AtomicRMWInst::UMax) && (RMWOp != AtomicRMWInst::UMin) &&
3094 "OpenMP atomic does not support LT or GT operations");
3095 });
3096
3097 emitAtomicUpdate(AllocIP, X.Var, Expr, AO, RMWOp, UpdateOp, X.IsVolatile,
3098 IsXLHSInRHSPart);
3099 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Update);
3100 return Builder.saveIP();
3101 }
3102
emitRMWOpAsInstruction(Value * Src1,Value * Src2,AtomicRMWInst::BinOp RMWOp)3103 Value *OpenMPIRBuilder::emitRMWOpAsInstruction(Value *Src1, Value *Src2,
3104 AtomicRMWInst::BinOp RMWOp) {
3105 switch (RMWOp) {
3106 case AtomicRMWInst::Add:
3107 return Builder.CreateAdd(Src1, Src2);
3108 case AtomicRMWInst::Sub:
3109 return Builder.CreateSub(Src1, Src2);
3110 case AtomicRMWInst::And:
3111 return Builder.CreateAnd(Src1, Src2);
3112 case AtomicRMWInst::Nand:
3113 return Builder.CreateNeg(Builder.CreateAnd(Src1, Src2));
3114 case AtomicRMWInst::Or:
3115 return Builder.CreateOr(Src1, Src2);
3116 case AtomicRMWInst::Xor:
3117 return Builder.CreateXor(Src1, Src2);
3118 case AtomicRMWInst::Xchg:
3119 case AtomicRMWInst::FAdd:
3120 case AtomicRMWInst::FSub:
3121 case AtomicRMWInst::BAD_BINOP:
3122 case AtomicRMWInst::Max:
3123 case AtomicRMWInst::Min:
3124 case AtomicRMWInst::UMax:
3125 case AtomicRMWInst::UMin:
3126 llvm_unreachable("Unsupported atomic update operation");
3127 }
3128 llvm_unreachable("Unsupported atomic update operation");
3129 }
3130
3131 std::pair<Value *, Value *>
emitAtomicUpdate(Instruction * AllocIP,Value * X,Value * Expr,AtomicOrdering AO,AtomicRMWInst::BinOp RMWOp,AtomicUpdateCallbackTy & UpdateOp,bool VolatileX,bool IsXLHSInRHSPart)3132 OpenMPIRBuilder::emitAtomicUpdate(Instruction *AllocIP, Value *X, Value *Expr,
3133 AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
3134 AtomicUpdateCallbackTy &UpdateOp,
3135 bool VolatileX, bool IsXLHSInRHSPart) {
3136 Type *XElemTy = X->getType()->getPointerElementType();
3137
3138 bool DoCmpExch =
3139 ((RMWOp == AtomicRMWInst::BAD_BINOP) || (RMWOp == AtomicRMWInst::FAdd)) ||
3140 (RMWOp == AtomicRMWInst::FSub) ||
3141 (RMWOp == AtomicRMWInst::Sub && !IsXLHSInRHSPart);
3142
3143 std::pair<Value *, Value *> Res;
3144 if (XElemTy->isIntegerTy() && !DoCmpExch) {
3145 Res.first = Builder.CreateAtomicRMW(RMWOp, X, Expr, llvm::MaybeAlign(), AO);
3146 // not needed except in case of postfix captures. Generate anyway for
3147 // consistency with the else part. Will be removed with any DCE pass.
3148 Res.second = emitRMWOpAsInstruction(Res.first, Expr, RMWOp);
3149 } else {
3150 unsigned Addrspace = cast<PointerType>(X->getType())->getAddressSpace();
3151 IntegerType *IntCastTy =
3152 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
3153 Value *XBCast =
3154 Builder.CreateBitCast(X, IntCastTy->getPointerTo(Addrspace));
3155 LoadInst *OldVal =
3156 Builder.CreateLoad(IntCastTy, XBCast, X->getName() + ".atomic.load");
3157 OldVal->setAtomic(AO);
3158 // CurBB
3159 // | /---\
3160 // ContBB |
3161 // | \---/
3162 // ExitBB
3163 BasicBlock *CurBB = Builder.GetInsertBlock();
3164 Instruction *CurBBTI = CurBB->getTerminator();
3165 CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable();
3166 BasicBlock *ExitBB =
3167 CurBB->splitBasicBlock(CurBBTI, X->getName() + ".atomic.exit");
3168 BasicBlock *ContBB = CurBB->splitBasicBlock(CurBB->getTerminator(),
3169 X->getName() + ".atomic.cont");
3170 ContBB->getTerminator()->eraseFromParent();
3171 Builder.SetInsertPoint(ContBB);
3172 llvm::PHINode *PHI = Builder.CreatePHI(OldVal->getType(), 2);
3173 PHI->addIncoming(OldVal, CurBB);
3174 AllocaInst *NewAtomicAddr = Builder.CreateAlloca(XElemTy);
3175 NewAtomicAddr->setName(X->getName() + "x.new.val");
3176 NewAtomicAddr->moveBefore(AllocIP);
3177 IntegerType *NewAtomicCastTy =
3178 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
3179 bool IsIntTy = XElemTy->isIntegerTy();
3180 Value *NewAtomicIntAddr =
3181 (IsIntTy)
3182 ? NewAtomicAddr
3183 : Builder.CreateBitCast(NewAtomicAddr,
3184 NewAtomicCastTy->getPointerTo(Addrspace));
3185 Value *OldExprVal = PHI;
3186 if (!IsIntTy) {
3187 if (XElemTy->isFloatingPointTy()) {
3188 OldExprVal = Builder.CreateBitCast(PHI, XElemTy,
3189 X->getName() + ".atomic.fltCast");
3190 } else {
3191 OldExprVal = Builder.CreateIntToPtr(PHI, XElemTy,
3192 X->getName() + ".atomic.ptrCast");
3193 }
3194 }
3195
3196 Value *Upd = UpdateOp(OldExprVal, Builder);
3197 Builder.CreateStore(Upd, NewAtomicAddr);
3198 LoadInst *DesiredVal = Builder.CreateLoad(XElemTy, NewAtomicIntAddr);
3199 Value *XAddr =
3200 (IsIntTy)
3201 ? X
3202 : Builder.CreateBitCast(X, IntCastTy->getPointerTo(Addrspace));
3203 AtomicOrdering Failure =
3204 llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
3205 AtomicCmpXchgInst *Result = Builder.CreateAtomicCmpXchg(
3206 XAddr, OldExprVal, DesiredVal, llvm::MaybeAlign(), AO, Failure);
3207 Result->setVolatile(VolatileX);
3208 Value *PreviousVal = Builder.CreateExtractValue(Result, /*Idxs=*/0);
3209 Value *SuccessFailureVal = Builder.CreateExtractValue(Result, /*Idxs=*/1);
3210 PHI->addIncoming(PreviousVal, Builder.GetInsertBlock());
3211 Builder.CreateCondBr(SuccessFailureVal, ExitBB, ContBB);
3212
3213 Res.first = OldExprVal;
3214 Res.second = Upd;
3215
3216 // set Insertion point in exit block
3217 if (UnreachableInst *ExitTI =
3218 dyn_cast<UnreachableInst>(ExitBB->getTerminator())) {
3219 CurBBTI->eraseFromParent();
3220 Builder.SetInsertPoint(ExitBB);
3221 } else {
3222 Builder.SetInsertPoint(ExitTI);
3223 }
3224 }
3225
3226 return Res;
3227 }
3228
createAtomicCapture(const LocationDescription & Loc,Instruction * AllocIP,AtomicOpValue & X,AtomicOpValue & V,Value * Expr,AtomicOrdering AO,AtomicRMWInst::BinOp RMWOp,AtomicUpdateCallbackTy & UpdateOp,bool UpdateExpr,bool IsPostfixUpdate,bool IsXLHSInRHSPart)3229 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCapture(
3230 const LocationDescription &Loc, Instruction *AllocIP, AtomicOpValue &X,
3231 AtomicOpValue &V, Value *Expr, AtomicOrdering AO,
3232 AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp,
3233 bool UpdateExpr, bool IsPostfixUpdate, bool IsXLHSInRHSPart) {
3234 if (!updateToLocation(Loc))
3235 return Loc.IP;
3236
3237 LLVM_DEBUG({
3238 Type *XTy = X.Var->getType();
3239 assert(XTy->isPointerTy() &&
3240 "OMP Atomic expects a pointer to target memory");
3241 Type *XElemTy = XTy->getPointerElementType();
3242 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
3243 XElemTy->isPointerTy()) &&
3244 "OMP atomic capture expected a scalar type");
3245 assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) &&
3246 "OpenMP atomic does not support LT or GT operations");
3247 });
3248
3249 // If UpdateExpr is 'x' updated with some `expr` not based on 'x',
3250 // 'x' is simply atomically rewritten with 'expr'.
3251 AtomicRMWInst::BinOp AtomicOp = (UpdateExpr ? RMWOp : AtomicRMWInst::Xchg);
3252 std::pair<Value *, Value *> Result =
3253 emitAtomicUpdate(AllocIP, X.Var, Expr, AO, AtomicOp, UpdateOp,
3254 X.IsVolatile, IsXLHSInRHSPart);
3255
3256 Value *CapturedVal = (IsPostfixUpdate ? Result.first : Result.second);
3257 Builder.CreateStore(CapturedVal, V.Var, V.IsVolatile);
3258
3259 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Capture);
3260 return Builder.saveIP();
3261 }
3262
3263 GlobalVariable *
createOffloadMapnames(SmallVectorImpl<llvm::Constant * > & Names,std::string VarName)3264 OpenMPIRBuilder::createOffloadMapnames(SmallVectorImpl<llvm::Constant *> &Names,
3265 std::string VarName) {
3266 llvm::Constant *MapNamesArrayInit = llvm::ConstantArray::get(
3267 llvm::ArrayType::get(
3268 llvm::Type::getInt8Ty(M.getContext())->getPointerTo(), Names.size()),
3269 Names);
3270 auto *MapNamesArrayGlobal = new llvm::GlobalVariable(
3271 M, MapNamesArrayInit->getType(),
3272 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MapNamesArrayInit,
3273 VarName);
3274 return MapNamesArrayGlobal;
3275 }
3276
3277 // Create all simple and struct types exposed by the runtime and remember
3278 // the llvm::PointerTypes of them for easy access later.
initializeTypes(Module & M)3279 void OpenMPIRBuilder::initializeTypes(Module &M) {
3280 LLVMContext &Ctx = M.getContext();
3281 StructType *T;
3282 #define OMP_TYPE(VarName, InitValue) VarName = InitValue;
3283 #define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize) \
3284 VarName##Ty = ArrayType::get(ElemTy, ArraySize); \
3285 VarName##PtrTy = PointerType::getUnqual(VarName##Ty);
3286 #define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...) \
3287 VarName = FunctionType::get(ReturnType, {__VA_ARGS__}, IsVarArg); \
3288 VarName##Ptr = PointerType::getUnqual(VarName);
3289 #define OMP_STRUCT_TYPE(VarName, StructName, ...) \
3290 T = StructType::getTypeByName(Ctx, StructName); \
3291 if (!T) \
3292 T = StructType::create(Ctx, {__VA_ARGS__}, StructName); \
3293 VarName = T; \
3294 VarName##Ptr = PointerType::getUnqual(T);
3295 #include "llvm/Frontend/OpenMP/OMPKinds.def"
3296 }
3297
collectBlocks(SmallPtrSetImpl<BasicBlock * > & BlockSet,SmallVectorImpl<BasicBlock * > & BlockVector)3298 void OpenMPIRBuilder::OutlineInfo::collectBlocks(
3299 SmallPtrSetImpl<BasicBlock *> &BlockSet,
3300 SmallVectorImpl<BasicBlock *> &BlockVector) {
3301 SmallVector<BasicBlock *, 32> Worklist;
3302 BlockSet.insert(EntryBB);
3303 BlockSet.insert(ExitBB);
3304
3305 Worklist.push_back(EntryBB);
3306 while (!Worklist.empty()) {
3307 BasicBlock *BB = Worklist.pop_back_val();
3308 BlockVector.push_back(BB);
3309 for (BasicBlock *SuccBB : successors(BB))
3310 if (BlockSet.insert(SuccBB).second)
3311 Worklist.push_back(SuccBB);
3312 }
3313 }
3314
collectControlBlocks(SmallVectorImpl<BasicBlock * > & BBs)3315 void CanonicalLoopInfo::collectControlBlocks(
3316 SmallVectorImpl<BasicBlock *> &BBs) {
3317 // We only count those BBs as control block for which we do not need to
3318 // reverse the CFG, i.e. not the loop body which can contain arbitrary control
3319 // flow. For consistency, this also means we do not add the Body block, which
3320 // is just the entry to the body code.
3321 BBs.reserve(BBs.size() + 6);
3322 BBs.append({Preheader, Header, Cond, Latch, Exit, After});
3323 }
3324
assertOK() const3325 void CanonicalLoopInfo::assertOK() const {
3326 #ifndef NDEBUG
3327 // No constraints if this object currently does not describe a loop.
3328 if (!isValid())
3329 return;
3330
3331 // Verify standard control-flow we use for OpenMP loops.
3332 assert(Preheader);
3333 assert(isa<BranchInst>(Preheader->getTerminator()) &&
3334 "Preheader must terminate with unconditional branch");
3335 assert(Preheader->getSingleSuccessor() == Header &&
3336 "Preheader must jump to header");
3337
3338 assert(Header);
3339 assert(isa<BranchInst>(Header->getTerminator()) &&
3340 "Header must terminate with unconditional branch");
3341 assert(Header->getSingleSuccessor() == Cond &&
3342 "Header must jump to exiting block");
3343
3344 assert(Cond);
3345 assert(Cond->getSinglePredecessor() == Header &&
3346 "Exiting block only reachable from header");
3347
3348 assert(isa<BranchInst>(Cond->getTerminator()) &&
3349 "Exiting block must terminate with conditional branch");
3350 assert(size(successors(Cond)) == 2 &&
3351 "Exiting block must have two successors");
3352 assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(0) == Body &&
3353 "Exiting block's first successor jump to the body");
3354 assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(1) == Exit &&
3355 "Exiting block's second successor must exit the loop");
3356
3357 assert(Body);
3358 assert(Body->getSinglePredecessor() == Cond &&
3359 "Body only reachable from exiting block");
3360 assert(!isa<PHINode>(Body->front()));
3361
3362 assert(Latch);
3363 assert(isa<BranchInst>(Latch->getTerminator()) &&
3364 "Latch must terminate with unconditional branch");
3365 assert(Latch->getSingleSuccessor() == Header && "Latch must jump to header");
3366 // TODO: To support simple redirecting of the end of the body code that has
3367 // multiple; introduce another auxiliary basic block like preheader and after.
3368 assert(Latch->getSinglePredecessor() != nullptr);
3369 assert(!isa<PHINode>(Latch->front()));
3370
3371 assert(Exit);
3372 assert(isa<BranchInst>(Exit->getTerminator()) &&
3373 "Exit block must terminate with unconditional branch");
3374 assert(Exit->getSingleSuccessor() == After &&
3375 "Exit block must jump to after block");
3376
3377 assert(After);
3378 assert(After->getSinglePredecessor() == Exit &&
3379 "After block only reachable from exit block");
3380 assert(After->empty() || !isa<PHINode>(After->front()));
3381
3382 Instruction *IndVar = getIndVar();
3383 assert(IndVar && "Canonical induction variable not found?");
3384 assert(isa<IntegerType>(IndVar->getType()) &&
3385 "Induction variable must be an integer");
3386 assert(cast<PHINode>(IndVar)->getParent() == Header &&
3387 "Induction variable must be a PHI in the loop header");
3388 assert(cast<PHINode>(IndVar)->getIncomingBlock(0) == Preheader);
3389 assert(
3390 cast<ConstantInt>(cast<PHINode>(IndVar)->getIncomingValue(0))->isZero());
3391 assert(cast<PHINode>(IndVar)->getIncomingBlock(1) == Latch);
3392
3393 auto *NextIndVar = cast<PHINode>(IndVar)->getIncomingValue(1);
3394 assert(cast<Instruction>(NextIndVar)->getParent() == Latch);
3395 assert(cast<BinaryOperator>(NextIndVar)->getOpcode() == BinaryOperator::Add);
3396 assert(cast<BinaryOperator>(NextIndVar)->getOperand(0) == IndVar);
3397 assert(cast<ConstantInt>(cast<BinaryOperator>(NextIndVar)->getOperand(1))
3398 ->isOne());
3399
3400 Value *TripCount = getTripCount();
3401 assert(TripCount && "Loop trip count not found?");
3402 assert(IndVar->getType() == TripCount->getType() &&
3403 "Trip count and induction variable must have the same type");
3404
3405 auto *CmpI = cast<CmpInst>(&Cond->front());
3406 assert(CmpI->getPredicate() == CmpInst::ICMP_ULT &&
3407 "Exit condition must be a signed less-than comparison");
3408 assert(CmpI->getOperand(0) == IndVar &&
3409 "Exit condition must compare the induction variable");
3410 assert(CmpI->getOperand(1) == TripCount &&
3411 "Exit condition must compare with the trip count");
3412 #endif
3413 }
3414
invalidate()3415 void CanonicalLoopInfo::invalidate() {
3416 Preheader = nullptr;
3417 Header = nullptr;
3418 Cond = nullptr;
3419 Body = nullptr;
3420 Latch = nullptr;
3421 Exit = nullptr;
3422 After = nullptr;
3423 }
3424