1 //===----- CGOpenMPRuntime.h - Interface to OpenMP Runtimes -----*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This provides a class for OpenMP runtime code generation.
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIME_H
15 #define LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIME_H
18 #include "clang/AST/Type.h"
19 #include "clang/Basic/OpenMPKinds.h"
20 #include "clang/Basic/SourceLocation.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/StringMap.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/ValueHandle.h"
40 class OMPDependClause;
41 class OMPExecutableDirective;
42 class OMPLoopDirective;
44 class OMPDeclareReductionDecl;
49 class CodeGenFunction;
52 /// A basic class for pre|post-action for advanced codegen sequence for OpenMP
54 class PrePostActionTy {
56 explicit PrePostActionTy() {}
57 virtual void Enter(CodeGenFunction &CGF) {}
58 virtual void Exit(CodeGenFunction &CGF) {}
59 virtual ~PrePostActionTy() {}
62 /// Class provides a way to call simple version of codegen for OpenMP region, or
63 /// an advanced with possible pre|post-actions in codegen.
64 class RegionCodeGenTy final {
66 typedef void (*CodeGenTy)(intptr_t, CodeGenFunction &, PrePostActionTy &);
68 mutable PrePostActionTy *PrePostAction;
69 RegionCodeGenTy() = delete;
70 RegionCodeGenTy &operator=(const RegionCodeGenTy &) = delete;
71 template <typename Callable>
72 static void CallbackFn(intptr_t CodeGen, CodeGenFunction &CGF,
73 PrePostActionTy &Action) {
74 return (*reinterpret_cast<Callable *>(CodeGen))(CGF, Action);
78 template <typename Callable>
81 typename std::enable_if<
82 !std::is_same<typename std::remove_reference<Callable>::type,
83 RegionCodeGenTy>::value>::type * = nullptr)
84 : CodeGen(reinterpret_cast<intptr_t>(&CodeGen)),
85 Callback(CallbackFn<typename std::remove_reference<Callable>::type>),
86 PrePostAction(nullptr) {}
87 void setAction(PrePostActionTy &Action) const { PrePostAction = &Action; }
88 void operator()(CodeGenFunction &CGF) const;
91 struct OMPTaskDataTy final {
92 SmallVector<const Expr *, 4> PrivateVars;
93 SmallVector<const Expr *, 4> PrivateCopies;
94 SmallVector<const Expr *, 4> FirstprivateVars;
95 SmallVector<const Expr *, 4> FirstprivateCopies;
96 SmallVector<const Expr *, 4> FirstprivateInits;
97 SmallVector<const Expr *, 4> LastprivateVars;
98 SmallVector<const Expr *, 4> LastprivateCopies;
99 SmallVector<const Expr *, 4> ReductionVars;
100 SmallVector<const Expr *, 4> ReductionCopies;
101 SmallVector<const Expr *, 4> ReductionOps;
102 SmallVector<std::pair<OpenMPDependClauseKind, const Expr *>, 4> Dependences;
103 llvm::PointerIntPair<llvm::Value *, 1, bool> Final;
104 llvm::PointerIntPair<llvm::Value *, 1, bool> Schedule;
105 llvm::PointerIntPair<llvm::Value *, 1, bool> Priority;
106 llvm::Value *Reductions = nullptr;
107 unsigned NumberOfParts = 0;
109 bool Nogroup = false;
112 /// Class intended to support codegen of all kind of the reduction clauses.
113 class ReductionCodeGen {
115 /// Data required for codegen of reduction clauses.
116 struct ReductionData {
117 /// Reference to the original shared item.
118 const Expr *Ref = nullptr;
119 /// Helper expression for generation of private copy.
120 const Expr *Private = nullptr;
121 /// Helper expression for generation reduction operation.
122 const Expr *ReductionOp = nullptr;
123 ReductionData(const Expr *Ref, const Expr *Private, const Expr *ReductionOp)
124 : Ref(Ref), Private(Private), ReductionOp(ReductionOp) {}
126 /// List of reduction-based clauses.
127 SmallVector<ReductionData, 4> ClausesData;
129 /// List of addresses of original shared variables/expressions.
130 SmallVector<std::pair<LValue, LValue>, 4> SharedAddresses;
131 /// Sizes of the reduction items in chars.
132 SmallVector<std::pair<llvm::Value *, llvm::Value *>, 4> Sizes;
133 /// Base declarations for the reduction items.
134 SmallVector<const VarDecl *, 4> BaseDecls;
136 /// Emits lvalue for shared expresion.
137 LValue emitSharedLValue(CodeGenFunction &CGF, const Expr *E);
138 /// Emits upper bound for shared expression (if array section).
139 LValue emitSharedLValueUB(CodeGenFunction &CGF, const Expr *E);
140 /// Performs aggregate initialization.
141 /// \param N Number of reduction item in the common list.
142 /// \param PrivateAddr Address of the corresponding private item.
143 /// \param SharedLVal Address of the original shared variable.
144 /// \param DRD Declare reduction construct used for reduction item.
145 void emitAggregateInitialization(CodeGenFunction &CGF, unsigned N,
146 Address PrivateAddr, LValue SharedLVal,
147 const OMPDeclareReductionDecl *DRD);
150 ReductionCodeGen(ArrayRef<const Expr *> Shareds,
151 ArrayRef<const Expr *> Privates,
152 ArrayRef<const Expr *> ReductionOps);
153 /// Emits lvalue for a reduction item.
154 /// \param N Number of the reduction item.
155 void emitSharedLValue(CodeGenFunction &CGF, unsigned N);
156 /// Emits the code for the variable-modified type, if required.
157 /// \param N Number of the reduction item.
158 void emitAggregateType(CodeGenFunction &CGF, unsigned N);
159 /// Emits the code for the variable-modified type, if required.
160 /// \param N Number of the reduction item.
161 /// \param Size Size of the type in chars.
162 void emitAggregateType(CodeGenFunction &CGF, unsigned N, llvm::Value *Size);
163 /// Performs initialization of the private copy for the reduction item.
164 /// \param N Number of the reduction item.
165 /// \param PrivateAddr Address of the corresponding private item.
166 /// \param DefaultInit Default initialization sequence that should be
167 /// performed if no reduction specific initialization is found.
168 /// \param SharedLVal Address of the original shared variable.
170 emitInitialization(CodeGenFunction &CGF, unsigned N, Address PrivateAddr,
172 llvm::function_ref<bool(CodeGenFunction &)> DefaultInit);
173 /// Returns true if the private copy requires cleanups.
174 bool needCleanups(unsigned N);
175 /// Emits cleanup code for the reduction item.
176 /// \param N Number of the reduction item.
177 /// \param PrivateAddr Address of the corresponding private item.
178 void emitCleanups(CodeGenFunction &CGF, unsigned N, Address PrivateAddr);
179 /// Adjusts \p PrivatedAddr for using instead of the original variable
180 /// address in normal operations.
181 /// \param N Number of the reduction item.
182 /// \param PrivateAddr Address of the corresponding private item.
183 Address adjustPrivateAddress(CodeGenFunction &CGF, unsigned N,
184 Address PrivateAddr);
185 /// Returns LValue for the reduction item.
186 LValue getSharedLValue(unsigned N) const { return SharedAddresses[N].first; }
187 /// Returns the size of the reduction item (in chars and total number of
188 /// elements in the item), or nullptr, if the size is a constant.
189 std::pair<llvm::Value *, llvm::Value *> getSizes(unsigned N) const {
192 /// Returns the base declaration of the reduction item.
193 const VarDecl *getBaseDecl(unsigned N) const { return BaseDecls[N]; }
194 /// Returns true if the initialization of the reduction item uses initializer
195 /// from declare reduction construct.
196 bool usesReductionInitializer(unsigned N) const;
199 class CGOpenMPRuntime {
203 /// \brief Creates offloading entry for the provided entry ID \a ID,
204 /// address \a Addr, size \a Size, and flags \a Flags.
205 virtual void createOffloadEntry(llvm::Constant *ID, llvm::Constant *Addr,
206 uint64_t Size, int32_t Flags = 0);
208 /// \brief Helper to emit outlined function for 'target' directive.
209 /// \param D Directive to emit.
210 /// \param ParentName Name of the function that encloses the target region.
211 /// \param OutlinedFn Outlined function value to be defined by this call.
212 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
213 /// \param IsOffloadEntry True if the outlined function is an offload entry.
214 /// \param CodeGen Lambda codegen specific to an accelerator device.
215 /// An outlined function may not be an entry if, e.g. the if clause always
216 /// evaluates to false.
217 virtual void emitTargetOutlinedFunctionHelper(const OMPExecutableDirective &D,
218 StringRef ParentName,
219 llvm::Function *&OutlinedFn,
220 llvm::Constant *&OutlinedFnID,
222 const RegionCodeGenTy &CodeGen);
224 /// \brief Emits code for OpenMP 'if' clause using specified \a CodeGen
225 /// function. Here is the logic:
231 void emitOMPIfClause(CodeGenFunction &CGF, const Expr *Cond,
232 const RegionCodeGenTy &ThenGen,
233 const RegionCodeGenTy &ElseGen);
235 /// \brief Emits object of ident_t type with info for source location.
236 /// \param Flags Flags for OpenMP location.
238 llvm::Value *emitUpdateLocation(CodeGenFunction &CGF, SourceLocation Loc,
241 /// \brief Returns pointer to ident_t type.
242 llvm::Type *getIdentTyPointerTy();
244 /// \brief Gets thread id value for the current thread.
246 llvm::Value *getThreadID(CodeGenFunction &CGF, SourceLocation Loc);
248 /// \brief Get the function name of an outlined region.
249 // The name can be customized depending on the target.
251 virtual StringRef getOutlinedHelperName() const { return ".omp_outlined."; }
254 /// \brief Default const ident_t object used for initialization of all other
256 llvm::Constant *DefaultOpenMPPSource = nullptr;
257 /// \brief Map of flags and corresponding default locations.
258 typedef llvm::DenseMap<unsigned, llvm::Value *> OpenMPDefaultLocMapTy;
259 OpenMPDefaultLocMapTy OpenMPDefaultLocMap;
260 Address getOrCreateDefaultLocation(unsigned Flags);
262 llvm::StructType *IdentTy = nullptr;
263 /// \brief Map for SourceLocation and OpenMP runtime library debug locations.
264 typedef llvm::DenseMap<unsigned, llvm::Value *> OpenMPDebugLocMapTy;
265 OpenMPDebugLocMapTy OpenMPDebugLocMap;
266 /// \brief The type for a microtask which gets passed to __kmpc_fork_call().
267 /// Original representation is:
268 /// typedef void (kmpc_micro)(kmp_int32 global_tid, kmp_int32 bound_tid,...);
269 llvm::FunctionType *Kmpc_MicroTy = nullptr;
270 /// \brief Stores debug location and ThreadID for the function.
271 struct DebugLocThreadIdTy {
272 llvm::Value *DebugLoc;
273 llvm::Value *ThreadID;
275 /// \brief Map of local debug location, ThreadId and functions.
276 typedef llvm::DenseMap<llvm::Function *, DebugLocThreadIdTy>
277 OpenMPLocThreadIDMapTy;
278 OpenMPLocThreadIDMapTy OpenMPLocThreadIDMap;
279 /// Map of UDRs and corresponding combiner/initializer.
280 typedef llvm::DenseMap<const OMPDeclareReductionDecl *,
281 std::pair<llvm::Function *, llvm::Function *>>
284 /// Map of functions and locally defined UDRs.
285 typedef llvm::DenseMap<llvm::Function *,
286 SmallVector<const OMPDeclareReductionDecl *, 4>>
288 FunctionUDRMapTy FunctionUDRMap;
289 IdentifierInfo *In = nullptr;
290 IdentifierInfo *Out = nullptr;
291 IdentifierInfo *Priv = nullptr;
292 IdentifierInfo *Orig = nullptr;
293 /// \brief Type kmp_critical_name, originally defined as typedef kmp_int32
294 /// kmp_critical_name[8];
295 llvm::ArrayType *KmpCriticalNameTy;
296 /// \brief An ordered map of auto-generated variables to their unique names.
297 /// It stores variables with the following names: 1) ".gomp_critical_user_" +
298 /// <critical_section_name> + ".var" for "omp critical" directives; 2)
299 /// <mangled_name_for_global_var> + ".cache." for cache for threadprivate
301 llvm::StringMap<llvm::AssertingVH<llvm::Constant>, llvm::BumpPtrAllocator>
303 /// \brief Type typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *);
304 llvm::Type *KmpRoutineEntryPtrTy = nullptr;
305 QualType KmpRoutineEntryPtrQTy;
306 /// \brief Type typedef struct kmp_task {
307 /// void * shareds; /**< pointer to block of pointers to
309 /// kmp_routine_entry_t routine; /**< pointer to routine to call for
310 /// executing task */
311 /// kmp_int32 part_id; /**< part id for the task */
312 /// kmp_routine_entry_t destructors; /* pointer to function to invoke
313 /// deconstructors of firstprivate C++ objects */
315 QualType KmpTaskTQTy;
316 /// \brief Type typedef struct kmp_depend_info {
317 /// kmp_intptr_t base_addr;
323 /// } kmp_depend_info_t;
324 QualType KmpDependInfoTy;
325 /// struct kmp_dim { // loop bounds info casted to kmp_int64
326 /// kmp_int64 lo; // lower
327 /// kmp_int64 up; // upper
328 /// kmp_int64 st; // stride
331 /// \brief Type struct __tgt_offload_entry{
332 /// void *addr; // Pointer to the offload entry info.
333 /// // (function or global)
334 /// char *name; // Name of the function or global.
335 /// size_t size; // Size of the entry info (0 if it a function).
337 QualType TgtOffloadEntryQTy;
338 /// struct __tgt_device_image{
339 /// void *ImageStart; // Pointer to the target code start.
340 /// void *ImageEnd; // Pointer to the target code end.
341 /// // We also add the host entries to the device image, as it may be useful
342 /// // for the target runtime to have access to that information.
343 /// __tgt_offload_entry *EntriesBegin; // Begin of the table with all
345 /// __tgt_offload_entry *EntriesEnd; // End of the table with all the
346 /// // entries (non inclusive).
348 QualType TgtDeviceImageQTy;
349 /// struct __tgt_bin_desc{
350 /// int32_t NumDevices; // Number of devices supported.
351 /// __tgt_device_image *DeviceImages; // Arrays of device images
352 /// // (one per device).
353 /// __tgt_offload_entry *EntriesBegin; // Begin of the table with all the
355 /// __tgt_offload_entry *EntriesEnd; // End of the table with all the
356 /// // entries (non inclusive).
358 QualType TgtBinaryDescriptorQTy;
359 /// \brief Entity that registers the offloading constants that were emitted so
361 class OffloadEntriesInfoManagerTy {
364 /// \brief Number of entries registered so far.
365 unsigned OffloadingEntriesNum;
368 /// Base class of the entries info.
369 class OffloadEntryInfo {
371 /// Kind of a given entry. Currently, only target regions are
373 enum OffloadingEntryInfoKinds : unsigned {
374 // Entry is a target region.
375 OFFLOAD_ENTRY_INFO_TARGET_REGION = 0,
376 // Invalid entry info.
377 OFFLOAD_ENTRY_INFO_INVALID = ~0u
381 : Flags(0), Order(~0u), Kind(OFFLOAD_ENTRY_INFO_INVALID) {}
382 explicit OffloadEntryInfo(OffloadingEntryInfoKinds Kind, unsigned Order,
384 : Flags(Flags), Order(Order), Kind(Kind) {}
386 bool isValid() const { return Order != ~0u; }
387 unsigned getOrder() const { return Order; }
388 OffloadingEntryInfoKinds getKind() const { return Kind; }
389 int32_t getFlags() const { return Flags; }
390 void setFlags(int32_t NewFlags) { Flags = NewFlags; }
391 static bool classof(const OffloadEntryInfo *Info) { return true; }
394 /// Flags associated with the device global.
397 /// Order this entry was emitted.
400 OffloadingEntryInfoKinds Kind;
403 /// \brief Return true if a there are no entries defined.
405 /// \brief Return number of entries defined so far.
406 unsigned size() const { return OffloadingEntriesNum; }
407 OffloadEntriesInfoManagerTy(CodeGenModule &CGM)
408 : CGM(CGM), OffloadingEntriesNum(0) {}
411 /// Target region entries related.
413 /// \brief Target region entries info.
414 class OffloadEntryInfoTargetRegion : public OffloadEntryInfo {
415 // \brief Address of the entity that has to be mapped for offloading.
416 llvm::Constant *Addr;
417 // \brief Address that can be used as the ID of the entry.
421 OffloadEntryInfoTargetRegion()
422 : OffloadEntryInfo(OFFLOAD_ENTRY_INFO_TARGET_REGION, ~0u,
424 Addr(nullptr), ID(nullptr) {}
425 explicit OffloadEntryInfoTargetRegion(unsigned Order,
426 llvm::Constant *Addr,
427 llvm::Constant *ID, int32_t Flags)
428 : OffloadEntryInfo(OFFLOAD_ENTRY_INFO_TARGET_REGION, Order, Flags),
429 Addr(Addr), ID(ID) {}
431 llvm::Constant *getAddress() const { return Addr; }
432 llvm::Constant *getID() const { return ID; }
433 void setAddress(llvm::Constant *V) {
434 assert(!Addr && "Address as been set before!");
437 void setID(llvm::Constant *V) {
438 assert(!ID && "ID as been set before!");
441 static bool classof(const OffloadEntryInfo *Info) {
442 return Info->getKind() == OFFLOAD_ENTRY_INFO_TARGET_REGION;
445 /// \brief Initialize target region entry.
446 void initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
447 StringRef ParentName, unsigned LineNum,
449 /// \brief Register target region entry.
450 void registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
451 StringRef ParentName, unsigned LineNum,
452 llvm::Constant *Addr, llvm::Constant *ID,
454 /// \brief Return true if a target region entry with the provided
455 /// information exists.
456 bool hasTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
457 StringRef ParentName, unsigned LineNum) const;
458 /// brief Applies action \a Action on all registered entries.
459 typedef llvm::function_ref<void(unsigned, unsigned, StringRef, unsigned,
460 OffloadEntryInfoTargetRegion &)>
461 OffloadTargetRegionEntryInfoActTy;
462 void actOnTargetRegionEntriesInfo(
463 const OffloadTargetRegionEntryInfoActTy &Action);
466 // Storage for target region entries kind. The storage is to be indexed by
467 // file ID, device ID, parent function name and line number.
468 typedef llvm::DenseMap<unsigned, OffloadEntryInfoTargetRegion>
469 OffloadEntriesTargetRegionPerLine;
470 typedef llvm::StringMap<OffloadEntriesTargetRegionPerLine>
471 OffloadEntriesTargetRegionPerParentName;
472 typedef llvm::DenseMap<unsigned, OffloadEntriesTargetRegionPerParentName>
473 OffloadEntriesTargetRegionPerFile;
474 typedef llvm::DenseMap<unsigned, OffloadEntriesTargetRegionPerFile>
475 OffloadEntriesTargetRegionPerDevice;
476 typedef OffloadEntriesTargetRegionPerDevice OffloadEntriesTargetRegionTy;
477 OffloadEntriesTargetRegionTy OffloadEntriesTargetRegion;
479 OffloadEntriesInfoManagerTy OffloadEntriesInfoManager;
481 /// \brief Creates and registers offloading binary descriptor for the current
482 /// compilation unit. The function that does the registration is returned.
483 llvm::Function *createOffloadingBinaryDescriptorRegistration();
485 /// \brief Creates all the offload entries in the current compilation unit
486 /// along with the associated metadata.
487 void createOffloadEntriesAndInfoMetadata();
489 /// \brief Loads all the offload entries information from the host IR
491 void loadOffloadInfoMetadata();
493 /// \brief Returns __tgt_offload_entry type.
494 QualType getTgtOffloadEntryQTy();
496 /// \brief Returns __tgt_device_image type.
497 QualType getTgtDeviceImageQTy();
499 /// \brief Returns __tgt_bin_desc type.
500 QualType getTgtBinaryDescriptorQTy();
502 /// \brief Start scanning from statement \a S and and emit all target regions
503 /// found along the way.
504 /// \param S Starting statement.
505 /// \param ParentName Name of the function declaration that is being scanned.
506 void scanForTargetRegionsFunctions(const Stmt *S, StringRef ParentName);
508 /// \brief Build type kmp_routine_entry_t (if not built yet).
509 void emitKmpRoutineEntryT(QualType KmpInt32Ty);
511 /// \brief Returns pointer to kmpc_micro type.
512 llvm::Type *getKmpc_MicroPointerTy();
514 /// \brief Returns specified OpenMP runtime function.
515 /// \param Function OpenMP runtime function.
516 /// \return Specified function.
517 llvm::Constant *createRuntimeFunction(unsigned Function);
519 /// \brief Returns __kmpc_for_static_init_* runtime function for the specified
520 /// size \a IVSize and sign \a IVSigned.
521 llvm::Constant *createForStaticInitFunction(unsigned IVSize, bool IVSigned);
523 /// \brief Returns __kmpc_dispatch_init_* runtime function for the specified
524 /// size \a IVSize and sign \a IVSigned.
525 llvm::Constant *createDispatchInitFunction(unsigned IVSize, bool IVSigned);
527 /// \brief Returns __kmpc_dispatch_next_* runtime function for the specified
528 /// size \a IVSize and sign \a IVSigned.
529 llvm::Constant *createDispatchNextFunction(unsigned IVSize, bool IVSigned);
531 /// \brief Returns __kmpc_dispatch_fini_* runtime function for the specified
532 /// size \a IVSize and sign \a IVSigned.
533 llvm::Constant *createDispatchFiniFunction(unsigned IVSize, bool IVSigned);
535 /// \brief If the specified mangled name is not in the module, create and
536 /// return threadprivate cache object. This object is a pointer's worth of
537 /// storage that's reserved for use by the OpenMP runtime.
538 /// \param VD Threadprivate variable.
539 /// \return Cache variable for the specified threadprivate.
540 llvm::Constant *getOrCreateThreadPrivateCache(const VarDecl *VD);
542 /// \brief Emits address of the word in a memory where current thread id is
544 virtual Address emitThreadIDAddress(CodeGenFunction &CGF, SourceLocation Loc);
546 /// \brief Gets (if variable with the given name already exist) or creates
547 /// internal global variable with the specified Name. The created variable has
548 /// linkage CommonLinkage by default and is initialized by null value.
549 /// \param Ty Type of the global variable. If it is exist already the type
550 /// must be the same.
551 /// \param Name Name of the variable.
552 llvm::Constant *getOrCreateInternalVariable(llvm::Type *Ty,
553 const llvm::Twine &Name);
555 /// \brief Set of threadprivate variables with the generated initializer.
556 llvm::SmallPtrSet<const VarDecl *, 4> ThreadPrivateWithDefinition;
558 /// \brief Emits initialization code for the threadprivate variables.
559 /// \param VDAddr Address of the global variable \a VD.
560 /// \param Ctor Pointer to a global init function for \a VD.
561 /// \param CopyCtor Pointer to a global copy function for \a VD.
562 /// \param Dtor Pointer to a global destructor function for \a VD.
563 /// \param Loc Location of threadprivate declaration.
564 void emitThreadPrivateVarInit(CodeGenFunction &CGF, Address VDAddr,
565 llvm::Value *Ctor, llvm::Value *CopyCtor,
566 llvm::Value *Dtor, SourceLocation Loc);
568 /// \brief Returns corresponding lock object for the specified critical region
569 /// name. If the lock object does not exist it is created, otherwise the
570 /// reference to the existing copy is returned.
571 /// \param CriticalName Name of the critical region.
573 llvm::Value *getCriticalRegionLock(StringRef CriticalName);
575 struct TaskResultTy {
576 llvm::Value *NewTask = nullptr;
577 llvm::Value *TaskEntry = nullptr;
578 llvm::Value *NewTaskNewTaskTTy = nullptr;
580 RecordDecl *KmpTaskTQTyRD = nullptr;
581 llvm::Value *TaskDupFn = nullptr;
583 /// Emit task region for the task directive. The task region is emitted in
585 /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32
586 /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
587 /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the
589 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
590 /// TaskFunction(gtid, tt->part_id, tt->shareds);
593 /// 2. Copy a list of shared variables to field shareds of the resulting
594 /// structure kmp_task_t returned by the previous call (if any).
595 /// 3. Copy a pointer to destructions function to field destructions of the
596 /// resulting structure kmp_task_t.
597 /// \param D Current task directive.
598 /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32
599 /// /*part_id*/, captured_struct */*__context*/);
600 /// \param SharedsTy A type which contains references the shared variables.
601 /// \param Shareds Context with the list of shared variables from the \p
603 /// \param Data Additional data for task generation like tiednsee, final
604 /// state, list of privates etc.
605 TaskResultTy emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc,
606 const OMPExecutableDirective &D,
607 llvm::Value *TaskFunction, QualType SharedsTy,
608 Address Shareds, const OMPTaskDataTy &Data);
611 explicit CGOpenMPRuntime(CodeGenModule &CGM);
612 virtual ~CGOpenMPRuntime() {}
613 virtual void clear();
615 /// Emit code for the specified user defined reduction construct.
616 virtual void emitUserDefinedReduction(CodeGenFunction *CGF,
617 const OMPDeclareReductionDecl *D);
618 /// Get combiner/initializer for the specified user-defined reduction, if any.
619 virtual std::pair<llvm::Function *, llvm::Function *>
620 getUserDefinedReduction(const OMPDeclareReductionDecl *D);
622 /// \brief Emits outlined function for the specified OpenMP parallel directive
623 /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID,
624 /// kmp_int32 BoundID, struct context_vars*).
625 /// \param D OpenMP directive.
626 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
627 /// \param InnermostKind Kind of innermost directive (for simple directives it
628 /// is a directive itself, for combined - its innermost directive).
629 /// \param CodeGen Code generation sequence for the \a D directive.
630 virtual llvm::Value *emitParallelOutlinedFunction(
631 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
632 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen);
634 /// \brief Emits outlined function for the specified OpenMP teams directive
635 /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID,
636 /// kmp_int32 BoundID, struct context_vars*).
637 /// \param D OpenMP directive.
638 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
639 /// \param InnermostKind Kind of innermost directive (for simple directives it
640 /// is a directive itself, for combined - its innermost directive).
641 /// \param CodeGen Code generation sequence for the \a D directive.
642 virtual llvm::Value *emitTeamsOutlinedFunction(
643 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
644 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen);
646 /// \brief Emits outlined function for the OpenMP task directive \a D. This
647 /// outlined function has type void(*)(kmp_int32 ThreadID, struct task_t*
649 /// \param D OpenMP directive.
650 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
651 /// \param PartIDVar Variable for partition id in the current OpenMP untied
653 /// \param TaskTVar Variable for task_t argument.
654 /// \param InnermostKind Kind of innermost directive (for simple directives it
655 /// is a directive itself, for combined - its innermost directive).
656 /// \param CodeGen Code generation sequence for the \a D directive.
657 /// \param Tied true if task is generated for tied task, false otherwise.
658 /// \param NumberOfParts Number of parts in untied task. Ignored for tied
661 virtual llvm::Value *emitTaskOutlinedFunction(
662 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
663 const VarDecl *PartIDVar, const VarDecl *TaskTVar,
664 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
665 bool Tied, unsigned &NumberOfParts);
667 /// \brief Cleans up references to the objects in finished function.
669 void functionFinished(CodeGenFunction &CGF);
671 /// \brief Emits code for parallel or serial call of the \a OutlinedFn with
672 /// variables captured in a record which address is stored in \a
674 /// \param OutlinedFn Outlined function to be run in parallel threads. Type of
675 /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
676 /// \param CapturedVars A pointer to the record with the references to
677 /// variables used in \a OutlinedFn function.
678 /// \param IfCond Condition in the associated 'if' clause, if it was
679 /// specified, nullptr otherwise.
681 virtual void emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc,
682 llvm::Value *OutlinedFn,
683 ArrayRef<llvm::Value *> CapturedVars,
686 /// \brief Emits a critical region.
687 /// \param CriticalName Name of the critical region.
688 /// \param CriticalOpGen Generator for the statement associated with the given
690 /// \param Hint Value of the 'hint' clause (optional).
691 virtual void emitCriticalRegion(CodeGenFunction &CGF, StringRef CriticalName,
692 const RegionCodeGenTy &CriticalOpGen,
694 const Expr *Hint = nullptr);
696 /// \brief Emits a master region.
697 /// \param MasterOpGen Generator for the statement associated with the given
699 virtual void emitMasterRegion(CodeGenFunction &CGF,
700 const RegionCodeGenTy &MasterOpGen,
703 /// \brief Emits code for a taskyield directive.
704 virtual void emitTaskyieldCall(CodeGenFunction &CGF, SourceLocation Loc);
706 /// \brief Emit a taskgroup region.
707 /// \param TaskgroupOpGen Generator for the statement associated with the
708 /// given taskgroup region.
709 virtual void emitTaskgroupRegion(CodeGenFunction &CGF,
710 const RegionCodeGenTy &TaskgroupOpGen,
713 /// \brief Emits a single region.
714 /// \param SingleOpGen Generator for the statement associated with the given
716 virtual void emitSingleRegion(CodeGenFunction &CGF,
717 const RegionCodeGenTy &SingleOpGen,
719 ArrayRef<const Expr *> CopyprivateVars,
720 ArrayRef<const Expr *> DestExprs,
721 ArrayRef<const Expr *> SrcExprs,
722 ArrayRef<const Expr *> AssignmentOps);
724 /// \brief Emit an ordered region.
725 /// \param OrderedOpGen Generator for the statement associated with the given
727 virtual void emitOrderedRegion(CodeGenFunction &CGF,
728 const RegionCodeGenTy &OrderedOpGen,
729 SourceLocation Loc, bool IsThreads);
731 /// \brief Emit an implicit/explicit barrier for OpenMP threads.
732 /// \param Kind Directive for which this implicit barrier call must be
733 /// generated. Must be OMPD_barrier for explicit barrier generation.
734 /// \param EmitChecks true if need to emit checks for cancellation barriers.
735 /// \param ForceSimpleCall true simple barrier call must be emitted, false if
736 /// runtime class decides which one to emit (simple or with cancellation
739 virtual void emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc,
740 OpenMPDirectiveKind Kind,
741 bool EmitChecks = true,
742 bool ForceSimpleCall = false);
744 /// \brief Check if the specified \a ScheduleKind is static non-chunked.
745 /// This kind of worksharing directive is emitted without outer loop.
746 /// \param ScheduleKind Schedule kind specified in the 'schedule' clause.
747 /// \param Chunked True if chunk is specified in the clause.
749 virtual bool isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind,
752 /// \brief Check if the specified \a ScheduleKind is static non-chunked.
753 /// This kind of distribute directive is emitted without outer loop.
754 /// \param ScheduleKind Schedule kind specified in the 'dist_schedule' clause.
755 /// \param Chunked True if chunk is specified in the clause.
757 virtual bool isStaticNonchunked(OpenMPDistScheduleClauseKind ScheduleKind,
760 /// \brief Check if the specified \a ScheduleKind is dynamic.
761 /// This kind of worksharing directive is emitted without outer loop.
762 /// \param ScheduleKind Schedule Kind specified in the 'schedule' clause.
764 virtual bool isDynamic(OpenMPScheduleClauseKind ScheduleKind) const;
766 /// struct with the values to be passed to the dispatch runtime function
767 struct DispatchRTInput {
769 llvm::Value *LB = nullptr;
771 llvm::Value *UB = nullptr;
772 /// Chunk size specified using 'schedule' clause (nullptr if chunk
773 /// was not specified)
774 llvm::Value *Chunk = nullptr;
775 DispatchRTInput() = default;
776 DispatchRTInput(llvm::Value *LB, llvm::Value *UB, llvm::Value *Chunk)
777 : LB(LB), UB(UB), Chunk(Chunk) {}
780 /// Call the appropriate runtime routine to initialize it before start
783 /// This is used for non static scheduled types and when the ordered
784 /// clause is present on the loop construct.
785 /// Depending on the loop schedule, it is necessary to call some runtime
786 /// routine before start of the OpenMP loop to get the loop upper / lower
787 /// bounds \a LB and \a UB and stride \a ST.
789 /// \param CGF Reference to current CodeGenFunction.
790 /// \param Loc Clang source location.
791 /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause.
792 /// \param IVSize Size of the iteration variable in bits.
793 /// \param IVSigned Sign of the iteration variable.
794 /// \param Ordered true if loop is ordered, false otherwise.
795 /// \param DispatchValues struct containing llvm values for lower bound, upper
796 /// bound, and chunk expression.
797 /// For the default (nullptr) value, the chunk 1 will be used.
799 virtual void emitForDispatchInit(CodeGenFunction &CGF, SourceLocation Loc,
800 const OpenMPScheduleTy &ScheduleKind,
801 unsigned IVSize, bool IVSigned, bool Ordered,
802 const DispatchRTInput &DispatchValues);
804 /// \brief Call the appropriate runtime routine to initialize it before start
807 /// This is used only in case of static schedule, when the user did not
808 /// specify a ordered clause on the loop construct.
809 /// Depending on the loop schedule, it is necessary to call some runtime
810 /// routine before start of the OpenMP loop to get the loop upper / lower
811 /// bounds \a LB and \a UB and stride \a ST.
813 /// \param CGF Reference to current CodeGenFunction.
814 /// \param Loc Clang source location.
815 /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause.
816 /// \param IVSize Size of the iteration variable in bits.
817 /// \param IVSigned Sign of the iteration variable.
818 /// \param Ordered true if loop is ordered, false otherwise.
819 /// \param IL Address of the output variable in which the flag of the
820 /// last iteration is returned.
821 /// \param LB Address of the output variable in which the lower iteration
822 /// number is returned.
823 /// \param UB Address of the output variable in which the upper iteration
824 /// number is returned.
825 /// \param ST Address of the output variable in which the stride value is
826 /// returned necessary to generated the static_chunked scheduled loop.
827 /// \param Chunk Value of the chunk for the static_chunked scheduled loop.
828 /// For the default (nullptr) value, the chunk 1 will be used.
830 virtual void emitForStaticInit(CodeGenFunction &CGF, SourceLocation Loc,
831 const OpenMPScheduleTy &ScheduleKind,
832 unsigned IVSize, bool IVSigned, bool Ordered,
833 Address IL, Address LB, Address UB, Address ST,
834 llvm::Value *Chunk = nullptr);
837 /// \param CGF Reference to current CodeGenFunction.
838 /// \param Loc Clang source location.
839 /// \param SchedKind Schedule kind, specified by the 'dist_schedule' clause.
840 /// \param IVSize Size of the iteration variable in bits.
841 /// \param IVSigned Sign of the iteration variable.
842 /// \param Ordered true if loop is ordered, false otherwise.
843 /// \param IL Address of the output variable in which the flag of the
844 /// last iteration is returned.
845 /// \param LB Address of the output variable in which the lower iteration
846 /// number is returned.
847 /// \param UB Address of the output variable in which the upper iteration
848 /// number is returned.
849 /// \param ST Address of the output variable in which the stride value is
850 /// returned necessary to generated the static_chunked scheduled loop.
851 /// \param Chunk Value of the chunk for the static_chunked scheduled loop.
852 /// For the default (nullptr) value, the chunk 1 will be used.
854 virtual void emitDistributeStaticInit(CodeGenFunction &CGF, SourceLocation Loc,
855 OpenMPDistScheduleClauseKind SchedKind,
856 unsigned IVSize, bool IVSigned,
857 bool Ordered, Address IL, Address LB,
858 Address UB, Address ST,
859 llvm::Value *Chunk = nullptr);
861 /// \brief Call the appropriate runtime routine to notify that we finished
862 /// iteration of the ordered loop with the dynamic scheduling.
864 /// \param CGF Reference to current CodeGenFunction.
865 /// \param Loc Clang source location.
866 /// \param IVSize Size of the iteration variable in bits.
867 /// \param IVSigned Sign of the iteration variable.
869 virtual void emitForOrderedIterationEnd(CodeGenFunction &CGF,
870 SourceLocation Loc, unsigned IVSize,
873 /// \brief Call the appropriate runtime routine to notify that we finished
874 /// all the work with current loop.
876 /// \param CGF Reference to current CodeGenFunction.
877 /// \param Loc Clang source location.
879 virtual void emitForStaticFinish(CodeGenFunction &CGF, SourceLocation Loc);
881 /// Call __kmpc_dispatch_next(
882 /// ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter,
883 /// kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper,
884 /// kmp_int[32|64] *p_stride);
885 /// \param IVSize Size of the iteration variable in bits.
886 /// \param IVSigned Sign of the iteration variable.
887 /// \param IL Address of the output variable in which the flag of the
888 /// last iteration is returned.
889 /// \param LB Address of the output variable in which the lower iteration
890 /// number is returned.
891 /// \param UB Address of the output variable in which the upper iteration
892 /// number is returned.
893 /// \param ST Address of the output variable in which the stride value is
895 virtual llvm::Value *emitForNext(CodeGenFunction &CGF, SourceLocation Loc,
896 unsigned IVSize, bool IVSigned,
897 Address IL, Address LB,
898 Address UB, Address ST);
900 /// \brief Emits call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32
901 /// global_tid, kmp_int32 num_threads) to generate code for 'num_threads'
903 /// \param NumThreads An integer value of threads.
904 virtual void emitNumThreadsClause(CodeGenFunction &CGF,
905 llvm::Value *NumThreads,
908 /// \brief Emit call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32
909 /// global_tid, int proc_bind) to generate code for 'proc_bind' clause.
910 virtual void emitProcBindClause(CodeGenFunction &CGF,
911 OpenMPProcBindClauseKind ProcBind,
914 /// \brief Returns address of the threadprivate variable for the current
916 /// \param VD Threadprivate variable.
917 /// \param VDAddr Address of the global variable \a VD.
918 /// \param Loc Location of the reference to threadprivate var.
919 /// \return Address of the threadprivate variable for the current thread.
920 virtual Address getAddrOfThreadPrivate(CodeGenFunction &CGF,
925 /// \brief Emit a code for initialization of threadprivate variable. It emits
926 /// a call to runtime library which adds initial value to the newly created
927 /// threadprivate variable (if it is not constant) and registers destructor
928 /// for the variable (if any).
929 /// \param VD Threadprivate variable.
930 /// \param VDAddr Address of the global variable \a VD.
931 /// \param Loc Location of threadprivate declaration.
932 /// \param PerformInit true if initialization expression is not constant.
933 virtual llvm::Function *
934 emitThreadPrivateVarDefinition(const VarDecl *VD, Address VDAddr,
935 SourceLocation Loc, bool PerformInit,
936 CodeGenFunction *CGF = nullptr);
938 /// Creates artificial threadprivate variable with name \p Name and type \p
940 /// \param VarType Type of the artificial threadprivate variable.
941 /// \param Name Name of the artificial threadprivate variable.
942 virtual Address getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF,
946 /// \brief Emit flush of the variables specified in 'omp flush' directive.
947 /// \param Vars List of variables to flush.
948 virtual void emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *> Vars,
951 /// \brief Emit task region for the task directive. The task region is
952 /// emitted in several steps:
953 /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32
954 /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
955 /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the
957 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
958 /// TaskFunction(gtid, tt->part_id, tt->shareds);
961 /// 2. Copy a list of shared variables to field shareds of the resulting
962 /// structure kmp_task_t returned by the previous call (if any).
963 /// 3. Copy a pointer to destructions function to field destructions of the
964 /// resulting structure kmp_task_t.
965 /// 4. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid,
966 /// kmp_task_t *new_task), where new_task is a resulting structure from
968 /// \param D Current task directive.
969 /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32
970 /// /*part_id*/, captured_struct */*__context*/);
971 /// \param SharedsTy A type which contains references the shared variables.
972 /// \param Shareds Context with the list of shared variables from the \p
974 /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr
976 /// \param Data Additional data for task generation like tiednsee, final
977 /// state, list of privates etc.
978 virtual void emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc,
979 const OMPExecutableDirective &D,
980 llvm::Value *TaskFunction, QualType SharedsTy,
981 Address Shareds, const Expr *IfCond,
982 const OMPTaskDataTy &Data);
984 /// Emit task region for the taskloop directive. The taskloop region is
985 /// emitted in several steps:
986 /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32
987 /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
988 /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the
990 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
991 /// TaskFunction(gtid, tt->part_id, tt->shareds);
994 /// 2. Copy a list of shared variables to field shareds of the resulting
995 /// structure kmp_task_t returned by the previous call (if any).
996 /// 3. Copy a pointer to destructions function to field destructions of the
997 /// resulting structure kmp_task_t.
998 /// 4. Emit a call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t
999 /// *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int
1000 /// nogroup, int sched, kmp_uint64 grainsize, void *task_dup ), where new_task
1001 /// is a resulting structure from
1003 /// \param D Current task directive.
1004 /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32
1005 /// /*part_id*/, captured_struct */*__context*/);
1006 /// \param SharedsTy A type which contains references the shared variables.
1007 /// \param Shareds Context with the list of shared variables from the \p
1009 /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr
1011 /// \param Data Additional data for task generation like tiednsee, final
1012 /// state, list of privates etc.
1013 virtual void emitTaskLoopCall(
1014 CodeGenFunction &CGF, SourceLocation Loc, const OMPLoopDirective &D,
1015 llvm::Value *TaskFunction, QualType SharedsTy, Address Shareds,
1016 const Expr *IfCond, const OMPTaskDataTy &Data);
1018 /// \brief Emit code for the directive that does not require outlining.
1020 /// \param InnermostKind Kind of innermost directive (for simple directives it
1021 /// is a directive itself, for combined - its innermost directive).
1022 /// \param CodeGen Code generation sequence for the \a D directive.
1023 /// \param HasCancel true if region has inner cancel directive, false
1025 virtual void emitInlinedDirective(CodeGenFunction &CGF,
1026 OpenMPDirectiveKind InnermostKind,
1027 const RegionCodeGenTy &CodeGen,
1028 bool HasCancel = false);
1030 /// Emits reduction function.
1031 /// \param ArgsType Array type containing pointers to reduction variables.
1032 /// \param Privates List of private copies for original reduction arguments.
1033 /// \param LHSExprs List of LHS in \a ReductionOps reduction operations.
1034 /// \param RHSExprs List of RHS in \a ReductionOps reduction operations.
1035 /// \param ReductionOps List of reduction operations in form 'LHS binop RHS'
1036 /// or 'operator binop(LHS, RHS)'.
1037 llvm::Value *emitReductionFunction(CodeGenModule &CGM, llvm::Type *ArgsType,
1038 ArrayRef<const Expr *> Privates,
1039 ArrayRef<const Expr *> LHSExprs,
1040 ArrayRef<const Expr *> RHSExprs,
1041 ArrayRef<const Expr *> ReductionOps);
1043 /// Emits single reduction combiner
1044 void emitSingleReductionCombiner(CodeGenFunction &CGF,
1045 const Expr *ReductionOp,
1046 const Expr *PrivateRef,
1047 const DeclRefExpr *LHS,
1048 const DeclRefExpr *RHS);
1050 struct ReductionOptionsTy {
1052 bool SimpleReduction;
1053 OpenMPDirectiveKind ReductionKind;
1055 /// \brief Emit a code for reduction clause. Next code should be emitted for
1059 /// static kmp_critical_name lock = { 0 };
1061 /// void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
1063 /// *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]);
1068 /// void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]};
1069 /// switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
1070 /// RedList, reduce_func, &<lock>)) {
1073 /// <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
1075 /// __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
1079 /// Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
1086 /// \param Privates List of private copies for original reduction arguments.
1087 /// \param LHSExprs List of LHS in \a ReductionOps reduction operations.
1088 /// \param RHSExprs List of RHS in \a ReductionOps reduction operations.
1089 /// \param ReductionOps List of reduction operations in form 'LHS binop RHS'
1090 /// or 'operator binop(LHS, RHS)'.
1091 /// \param Options List of options for reduction codegen:
1092 /// WithNowait true if parent directive has also nowait clause, false
1094 /// SimpleReduction Emit reduction operation only. Used for omp simd
1095 /// directive on the host.
1096 /// ReductionKind The kind of reduction to perform.
1097 virtual void emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
1098 ArrayRef<const Expr *> Privates,
1099 ArrayRef<const Expr *> LHSExprs,
1100 ArrayRef<const Expr *> RHSExprs,
1101 ArrayRef<const Expr *> ReductionOps,
1102 ReductionOptionsTy Options);
1104 /// Emit a code for initialization of task reduction clause. Next code
1105 /// should be emitted for reduction:
1108 /// _task_red_item_t red_data[n];
1110 /// red_data[i].shar = &origs[i];
1111 /// red_data[i].size = sizeof(origs[i]);
1112 /// red_data[i].f_init = (void*)RedInit<i>;
1113 /// red_data[i].f_fini = (void*)RedDest<i>;
1114 /// red_data[i].f_comb = (void*)RedOp<i>;
1115 /// red_data[i].flags = <Flag_i>;
1117 /// void* tg1 = __kmpc_task_reduction_init(gtid, n, red_data);
1120 /// \param LHSExprs List of LHS in \a Data.ReductionOps reduction operations.
1121 /// \param RHSExprs List of RHS in \a Data.ReductionOps reduction operations.
1122 /// \param Data Additional data for task generation like tiedness, final
1123 /// state, list of privates, reductions etc.
1124 virtual llvm::Value *emitTaskReductionInit(CodeGenFunction &CGF,
1126 ArrayRef<const Expr *> LHSExprs,
1127 ArrayRef<const Expr *> RHSExprs,
1128 const OMPTaskDataTy &Data);
1130 /// Required to resolve existing problems in the runtime. Emits threadprivate
1131 /// variables to store the size of the VLAs/array sections for
1132 /// initializer/combiner/finalizer functions + emits threadprivate variable to
1133 /// store the pointer to the original reduction item for the custom
1134 /// initializer defined by declare reduction construct.
1135 /// \param RCG Allows to reuse an existing data for the reductions.
1136 /// \param N Reduction item for which fixups must be emitted.
1137 virtual void emitTaskReductionFixups(CodeGenFunction &CGF, SourceLocation Loc,
1138 ReductionCodeGen &RCG, unsigned N);
1140 /// Get the address of `void *` type of the privatue copy of the reduction
1141 /// item specified by the \p SharedLVal.
1142 /// \param ReductionsPtr Pointer to the reduction data returned by the
1143 /// emitTaskReductionInit function.
1144 /// \param SharedLVal Address of the original reduction item.
1145 virtual Address getTaskReductionItem(CodeGenFunction &CGF, SourceLocation Loc,
1146 llvm::Value *ReductionsPtr,
1149 /// \brief Emit code for 'taskwait' directive.
1150 virtual void emitTaskwaitCall(CodeGenFunction &CGF, SourceLocation Loc);
1152 /// \brief Emit code for 'cancellation point' construct.
1153 /// \param CancelRegion Region kind for which the cancellation point must be
1156 virtual void emitCancellationPointCall(CodeGenFunction &CGF,
1158 OpenMPDirectiveKind CancelRegion);
1160 /// \brief Emit code for 'cancel' construct.
1161 /// \param IfCond Condition in the associated 'if' clause, if it was
1162 /// specified, nullptr otherwise.
1163 /// \param CancelRegion Region kind for which the cancel must be emitted.
1165 virtual void emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc,
1167 OpenMPDirectiveKind CancelRegion);
1169 /// \brief Emit outilined function for 'target' directive.
1170 /// \param D Directive to emit.
1171 /// \param ParentName Name of the function that encloses the target region.
1172 /// \param OutlinedFn Outlined function value to be defined by this call.
1173 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
1174 /// \param IsOffloadEntry True if the outlined function is an offload entry.
1175 /// \param CodeGen Code generation sequence for the \a D directive.
1176 /// An outlined function may not be an entry if, e.g. the if clause always
1177 /// evaluates to false.
1178 virtual void emitTargetOutlinedFunction(const OMPExecutableDirective &D,
1179 StringRef ParentName,
1180 llvm::Function *&OutlinedFn,
1181 llvm::Constant *&OutlinedFnID,
1182 bool IsOffloadEntry,
1183 const RegionCodeGenTy &CodeGen);
1185 /// \brief Emit the target offloading code associated with \a D. The emitted
1186 /// code attempts offloading the execution to the device, an the event of
1187 /// a failure it executes the host version outlined in \a OutlinedFn.
1188 /// \param D Directive to emit.
1189 /// \param OutlinedFn Host version of the code to be offloaded.
1190 /// \param OutlinedFnID ID of host version of the code to be offloaded.
1191 /// \param IfCond Expression evaluated in if clause associated with the target
1192 /// directive, or null if no if clause is used.
1193 /// \param Device Expression evaluated in device clause associated with the
1194 /// target directive, or null if no device clause is used.
1195 /// \param CapturedVars Values captured in the current region.
1196 virtual void emitTargetCall(CodeGenFunction &CGF,
1197 const OMPExecutableDirective &D,
1198 llvm::Value *OutlinedFn,
1199 llvm::Value *OutlinedFnID, const Expr *IfCond,
1201 ArrayRef<llvm::Value *> CapturedVars);
1203 /// \brief Emit the target regions enclosed in \a GD function definition or
1204 /// the function itself in case it is a valid device function. Returns true if
1205 /// \a GD was dealt with successfully.
1206 /// \param GD Function to scan.
1207 virtual bool emitTargetFunctions(GlobalDecl GD);
1209 /// \brief Emit the global variable if it is a valid device global variable.
1210 /// Returns true if \a GD was dealt with successfully.
1211 /// \param GD Variable declaration to emit.
1212 virtual bool emitTargetGlobalVariable(GlobalDecl GD);
1214 /// \brief Emit the global \a GD if it is meaningful for the target. Returns
1215 /// if it was emitted successfully.
1216 /// \param GD Global to scan.
1217 virtual bool emitTargetGlobal(GlobalDecl GD);
1219 /// \brief Creates the offloading descriptor in the event any target region
1220 /// was emitted in the current module and return the function that registers
1222 virtual llvm::Function *emitRegistrationFunction();
1224 /// \brief Emits code for teams call of the \a OutlinedFn with
1225 /// variables captured in a record which address is stored in \a
1227 /// \param OutlinedFn Outlined function to be run by team masters. Type of
1228 /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
1229 /// \param CapturedVars A pointer to the record with the references to
1230 /// variables used in \a OutlinedFn function.
1232 virtual void emitTeamsCall(CodeGenFunction &CGF,
1233 const OMPExecutableDirective &D,
1234 SourceLocation Loc, llvm::Value *OutlinedFn,
1235 ArrayRef<llvm::Value *> CapturedVars);
1237 /// \brief Emits call to void __kmpc_push_num_teams(ident_t *loc, kmp_int32
1238 /// global_tid, kmp_int32 num_teams, kmp_int32 thread_limit) to generate code
1239 /// for num_teams clause.
1240 /// \param NumTeams An integer expression of teams.
1241 /// \param ThreadLimit An integer expression of threads.
1242 virtual void emitNumTeamsClause(CodeGenFunction &CGF, const Expr *NumTeams,
1243 const Expr *ThreadLimit, SourceLocation Loc);
1245 /// Struct that keeps all the relevant information that should be kept
1246 /// throughout a 'target data' region.
1247 class TargetDataInfo {
1248 /// Set to true if device pointer information have to be obtained.
1249 bool RequiresDevicePointerInfo = false;
1252 /// The array of base pointer passed to the runtime library.
1253 llvm::Value *BasePointersArray = nullptr;
1254 /// The array of section pointers passed to the runtime library.
1255 llvm::Value *PointersArray = nullptr;
1256 /// The array of sizes passed to the runtime library.
1257 llvm::Value *SizesArray = nullptr;
1258 /// The array of map types passed to the runtime library.
1259 llvm::Value *MapTypesArray = nullptr;
1260 /// The total number of pointers passed to the runtime library.
1261 unsigned NumberOfPtrs = 0u;
1262 /// Map between the a declaration of a capture and the corresponding base
1263 /// pointer address where the runtime returns the device pointers.
1264 llvm::DenseMap<const ValueDecl *, Address> CaptureDeviceAddrMap;
1266 explicit TargetDataInfo() {}
1267 explicit TargetDataInfo(bool RequiresDevicePointerInfo)
1268 : RequiresDevicePointerInfo(RequiresDevicePointerInfo) {}
1269 /// Clear information about the data arrays.
1270 void clearArrayInfo() {
1271 BasePointersArray = nullptr;
1272 PointersArray = nullptr;
1273 SizesArray = nullptr;
1274 MapTypesArray = nullptr;
1277 /// Return true if the current target data information has valid arrays.
1279 return BasePointersArray && PointersArray && SizesArray &&
1280 MapTypesArray && NumberOfPtrs;
1282 bool requiresDevicePointerInfo() { return RequiresDevicePointerInfo; }
1285 /// \brief Emit the target data mapping code associated with \a D.
1286 /// \param D Directive to emit.
1287 /// \param IfCond Expression evaluated in if clause associated with the
1288 /// target directive, or null if no device clause is used.
1289 /// \param Device Expression evaluated in device clause associated with the
1290 /// target directive, or null if no device clause is used.
1291 /// \param Info A record used to store information that needs to be preserved
1292 /// until the region is closed.
1293 virtual void emitTargetDataCalls(CodeGenFunction &CGF,
1294 const OMPExecutableDirective &D,
1295 const Expr *IfCond, const Expr *Device,
1296 const RegionCodeGenTy &CodeGen,
1297 TargetDataInfo &Info);
1299 /// \brief Emit the data mapping/movement code associated with the directive
1300 /// \a D that should be of the form 'target [{enter|exit} data | update]'.
1301 /// \param D Directive to emit.
1302 /// \param IfCond Expression evaluated in if clause associated with the target
1303 /// directive, or null if no if clause is used.
1304 /// \param Device Expression evaluated in device clause associated with the
1305 /// target directive, or null if no device clause is used.
1306 virtual void emitTargetDataStandAloneCall(CodeGenFunction &CGF,
1307 const OMPExecutableDirective &D,
1309 const Expr *Device);
1311 /// Marks function \a Fn with properly mangled versions of vector functions.
1312 /// \param FD Function marked as 'declare simd'.
1313 /// \param Fn LLVM function that must be marked with 'declare simd'
1315 virtual void emitDeclareSimdFunction(const FunctionDecl *FD,
1316 llvm::Function *Fn);
1318 /// Emit initialization for doacross loop nesting support.
1319 /// \param D Loop-based construct used in doacross nesting construct.
1320 virtual void emitDoacrossInit(CodeGenFunction &CGF,
1321 const OMPLoopDirective &D);
1323 /// Emit code for doacross ordered directive with 'depend' clause.
1324 /// \param C 'depend' clause with 'sink|source' dependency kind.
1325 virtual void emitDoacrossOrdered(CodeGenFunction &CGF,
1326 const OMPDependClause *C);
1329 } // namespace CodeGen
1330 } // namespace clang