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<std::pair<OpenMPDependClauseKind, const Expr *>, 4> Dependences;
100 llvm::PointerIntPair<llvm::Value *, 1, bool> Final;
101 llvm::PointerIntPair<llvm::Value *, 1, bool> Schedule;
102 llvm::PointerIntPair<llvm::Value *, 1, bool> Priority;
103 unsigned NumberOfParts = 0;
105 bool Nogroup = false;
108 class CGOpenMPRuntime {
112 /// \brief Creates offloading entry for the provided entry ID \a ID,
113 /// address \a Addr and size \a Size.
114 virtual void createOffloadEntry(llvm::Constant *ID, llvm::Constant *Addr,
117 /// \brief Helper to emit outlined function for 'target' directive.
118 /// \param D Directive to emit.
119 /// \param ParentName Name of the function that encloses the target region.
120 /// \param OutlinedFn Outlined function value to be defined by this call.
121 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
122 /// \param IsOffloadEntry True if the outlined function is an offload entry.
123 /// \param CodeGen Lambda codegen specific to an accelerator device.
124 /// An oulined function may not be an entry if, e.g. the if clause always
125 /// evaluates to false.
126 virtual void emitTargetOutlinedFunctionHelper(const OMPExecutableDirective &D,
127 StringRef ParentName,
128 llvm::Function *&OutlinedFn,
129 llvm::Constant *&OutlinedFnID,
131 const RegionCodeGenTy &CodeGen);
134 /// \brief Default const ident_t object used for initialization of all other
136 llvm::Constant *DefaultOpenMPPSource = nullptr;
137 /// \brief Map of flags and corresponding default locations.
138 typedef llvm::DenseMap<unsigned, llvm::Value *> OpenMPDefaultLocMapTy;
139 OpenMPDefaultLocMapTy OpenMPDefaultLocMap;
140 Address getOrCreateDefaultLocation(unsigned Flags);
142 llvm::StructType *IdentTy = nullptr;
143 /// \brief Map for SourceLocation and OpenMP runtime library debug locations.
144 typedef llvm::DenseMap<unsigned, llvm::Value *> OpenMPDebugLocMapTy;
145 OpenMPDebugLocMapTy OpenMPDebugLocMap;
146 /// \brief The type for a microtask which gets passed to __kmpc_fork_call().
147 /// Original representation is:
148 /// typedef void (kmpc_micro)(kmp_int32 global_tid, kmp_int32 bound_tid,...);
149 llvm::FunctionType *Kmpc_MicroTy = nullptr;
150 /// \brief Stores debug location and ThreadID for the function.
151 struct DebugLocThreadIdTy {
152 llvm::Value *DebugLoc;
153 llvm::Value *ThreadID;
155 /// \brief Map of local debug location, ThreadId and functions.
156 typedef llvm::DenseMap<llvm::Function *, DebugLocThreadIdTy>
157 OpenMPLocThreadIDMapTy;
158 OpenMPLocThreadIDMapTy OpenMPLocThreadIDMap;
159 /// Map of UDRs and corresponding combiner/initializer.
160 typedef llvm::DenseMap<const OMPDeclareReductionDecl *,
161 std::pair<llvm::Function *, llvm::Function *>>
164 /// Map of functions and locally defined UDRs.
165 typedef llvm::DenseMap<llvm::Function *,
166 SmallVector<const OMPDeclareReductionDecl *, 4>>
168 FunctionUDRMapTy FunctionUDRMap;
169 IdentifierInfo *In = nullptr;
170 IdentifierInfo *Out = nullptr;
171 IdentifierInfo *Priv = nullptr;
172 IdentifierInfo *Orig = nullptr;
173 /// \brief Type kmp_critical_name, originally defined as typedef kmp_int32
174 /// kmp_critical_name[8];
175 llvm::ArrayType *KmpCriticalNameTy;
176 /// \brief An ordered map of auto-generated variables to their unique names.
177 /// It stores variables with the following names: 1) ".gomp_critical_user_" +
178 /// <critical_section_name> + ".var" for "omp critical" directives; 2)
179 /// <mangled_name_for_global_var> + ".cache." for cache for threadprivate
181 llvm::StringMap<llvm::AssertingVH<llvm::Constant>, llvm::BumpPtrAllocator>
183 /// \brief Type typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *);
184 llvm::Type *KmpRoutineEntryPtrTy = nullptr;
185 QualType KmpRoutineEntryPtrQTy;
186 /// \brief Type typedef struct kmp_task {
187 /// void * shareds; /**< pointer to block of pointers to
189 /// kmp_routine_entry_t routine; /**< pointer to routine to call for
190 /// executing task */
191 /// kmp_int32 part_id; /**< part id for the task */
192 /// kmp_routine_entry_t destructors; /* pointer to function to invoke
193 /// deconstructors of firstprivate C++ objects */
195 QualType KmpTaskTQTy;
196 /// \brief Type typedef struct kmp_depend_info {
197 /// kmp_intptr_t base_addr;
203 /// } kmp_depend_info_t;
204 QualType KmpDependInfoTy;
205 /// struct kmp_dim { // loop bounds info casted to kmp_int64
206 /// kmp_int64 lo; // lower
207 /// kmp_int64 up; // upper
208 /// kmp_int64 st; // stride
211 /// \brief Type struct __tgt_offload_entry{
212 /// void *addr; // Pointer to the offload entry info.
213 /// // (function or global)
214 /// char *name; // Name of the function or global.
215 /// size_t size; // Size of the entry info (0 if it a function).
217 QualType TgtOffloadEntryQTy;
218 /// struct __tgt_device_image{
219 /// void *ImageStart; // Pointer to the target code start.
220 /// void *ImageEnd; // Pointer to the target code end.
221 /// // We also add the host entries to the device image, as it may be useful
222 /// // for the target runtime to have access to that information.
223 /// __tgt_offload_entry *EntriesBegin; // Begin of the table with all
225 /// __tgt_offload_entry *EntriesEnd; // End of the table with all the
226 /// // entries (non inclusive).
228 QualType TgtDeviceImageQTy;
229 /// struct __tgt_bin_desc{
230 /// int32_t NumDevices; // Number of devices supported.
231 /// __tgt_device_image *DeviceImages; // Arrays of device images
232 /// // (one per device).
233 /// __tgt_offload_entry *EntriesBegin; // Begin of the table with all the
235 /// __tgt_offload_entry *EntriesEnd; // End of the table with all the
236 /// // entries (non inclusive).
238 QualType TgtBinaryDescriptorQTy;
239 /// \brief Entity that registers the offloading constants that were emitted so
241 class OffloadEntriesInfoManagerTy {
244 /// \brief Number of entries registered so far.
245 unsigned OffloadingEntriesNum;
248 /// \brief Base class of the entries info.
249 class OffloadEntryInfo {
251 /// \brief Kind of a given entry. Currently, only target regions are
253 enum OffloadingEntryInfoKinds : unsigned {
254 // Entry is a target region.
255 OFFLOAD_ENTRY_INFO_TARGET_REGION = 0,
256 // Invalid entry info.
257 OFFLOAD_ENTRY_INFO_INVALID = ~0u
260 OffloadEntryInfo() : Order(~0u), Kind(OFFLOAD_ENTRY_INFO_INVALID) {}
261 explicit OffloadEntryInfo(OffloadingEntryInfoKinds Kind, unsigned Order)
262 : Order(Order), Kind(Kind) {}
264 bool isValid() const { return Order != ~0u; }
265 unsigned getOrder() const { return Order; }
266 OffloadingEntryInfoKinds getKind() const { return Kind; }
267 static bool classof(const OffloadEntryInfo *Info) { return true; }
270 // \brief Order this entry was emitted.
273 OffloadingEntryInfoKinds Kind;
276 /// \brief Return true if a there are no entries defined.
278 /// \brief Return number of entries defined so far.
279 unsigned size() const { return OffloadingEntriesNum; }
280 OffloadEntriesInfoManagerTy(CodeGenModule &CGM)
281 : CGM(CGM), OffloadingEntriesNum(0) {}
284 /// Target region entries related.
286 /// \brief Target region entries info.
287 class OffloadEntryInfoTargetRegion : public OffloadEntryInfo {
288 // \brief Address of the entity that has to be mapped for offloading.
289 llvm::Constant *Addr;
290 // \brief Address that can be used as the ID of the entry.
294 OffloadEntryInfoTargetRegion()
295 : OffloadEntryInfo(OFFLOAD_ENTRY_INFO_TARGET_REGION, ~0u),
296 Addr(nullptr), ID(nullptr) {}
297 explicit OffloadEntryInfoTargetRegion(unsigned Order,
298 llvm::Constant *Addr,
300 : OffloadEntryInfo(OFFLOAD_ENTRY_INFO_TARGET_REGION, Order),
301 Addr(Addr), ID(ID) {}
303 llvm::Constant *getAddress() const { return Addr; }
304 llvm::Constant *getID() const { return ID; }
305 void setAddress(llvm::Constant *V) {
306 assert(!Addr && "Address as been set before!");
309 void setID(llvm::Constant *V) {
310 assert(!ID && "ID as been set before!");
313 static bool classof(const OffloadEntryInfo *Info) {
314 return Info->getKind() == OFFLOAD_ENTRY_INFO_TARGET_REGION;
317 /// \brief Initialize target region entry.
318 void initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
319 StringRef ParentName, unsigned LineNum,
321 /// \brief Register target region entry.
322 void registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
323 StringRef ParentName, unsigned LineNum,
324 llvm::Constant *Addr,
326 /// \brief Return true if a target region entry with the provided
327 /// information exists.
328 bool hasTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
329 StringRef ParentName, unsigned LineNum) const;
330 /// brief Applies action \a Action on all registered entries.
331 typedef llvm::function_ref<void(unsigned, unsigned, StringRef, unsigned,
332 OffloadEntryInfoTargetRegion &)>
333 OffloadTargetRegionEntryInfoActTy;
334 void actOnTargetRegionEntriesInfo(
335 const OffloadTargetRegionEntryInfoActTy &Action);
338 // Storage for target region entries kind. The storage is to be indexed by
339 // file ID, device ID, parent function name and line number.
340 typedef llvm::DenseMap<unsigned, OffloadEntryInfoTargetRegion>
341 OffloadEntriesTargetRegionPerLine;
342 typedef llvm::StringMap<OffloadEntriesTargetRegionPerLine>
343 OffloadEntriesTargetRegionPerParentName;
344 typedef llvm::DenseMap<unsigned, OffloadEntriesTargetRegionPerParentName>
345 OffloadEntriesTargetRegionPerFile;
346 typedef llvm::DenseMap<unsigned, OffloadEntriesTargetRegionPerFile>
347 OffloadEntriesTargetRegionPerDevice;
348 typedef OffloadEntriesTargetRegionPerDevice OffloadEntriesTargetRegionTy;
349 OffloadEntriesTargetRegionTy OffloadEntriesTargetRegion;
351 OffloadEntriesInfoManagerTy OffloadEntriesInfoManager;
353 /// \brief Creates and registers offloading binary descriptor for the current
354 /// compilation unit. The function that does the registration is returned.
355 llvm::Function *createOffloadingBinaryDescriptorRegistration();
357 /// \brief Creates all the offload entries in the current compilation unit
358 /// along with the associated metadata.
359 void createOffloadEntriesAndInfoMetadata();
361 /// \brief Loads all the offload entries information from the host IR
363 void loadOffloadInfoMetadata();
365 /// \brief Returns __tgt_offload_entry type.
366 QualType getTgtOffloadEntryQTy();
368 /// \brief Returns __tgt_device_image type.
369 QualType getTgtDeviceImageQTy();
371 /// \brief Returns __tgt_bin_desc type.
372 QualType getTgtBinaryDescriptorQTy();
374 /// \brief Start scanning from statement \a S and and emit all target regions
375 /// found along the way.
376 /// \param S Starting statement.
377 /// \param ParentName Name of the function declaration that is being scanned.
378 void scanForTargetRegionsFunctions(const Stmt *S, StringRef ParentName);
380 /// \brief Build type kmp_routine_entry_t (if not built yet).
381 void emitKmpRoutineEntryT(QualType KmpInt32Ty);
383 /// \brief Emits object of ident_t type with info for source location.
384 /// \param Flags Flags for OpenMP location.
386 llvm::Value *emitUpdateLocation(CodeGenFunction &CGF, SourceLocation Loc,
389 /// \brief Returns pointer to ident_t type.
390 llvm::Type *getIdentTyPointerTy();
392 /// \brief Returns pointer to kmpc_micro type.
393 llvm::Type *getKmpc_MicroPointerTy();
395 /// \brief Returns specified OpenMP runtime function.
396 /// \param Function OpenMP runtime function.
397 /// \return Specified function.
398 llvm::Constant *createRuntimeFunction(unsigned Function);
400 /// \brief Returns __kmpc_for_static_init_* runtime function for the specified
401 /// size \a IVSize and sign \a IVSigned.
402 llvm::Constant *createForStaticInitFunction(unsigned IVSize, bool IVSigned);
404 /// \brief Returns __kmpc_dispatch_init_* runtime function for the specified
405 /// size \a IVSize and sign \a IVSigned.
406 llvm::Constant *createDispatchInitFunction(unsigned IVSize, bool IVSigned);
408 /// \brief Returns __kmpc_dispatch_next_* runtime function for the specified
409 /// size \a IVSize and sign \a IVSigned.
410 llvm::Constant *createDispatchNextFunction(unsigned IVSize, bool IVSigned);
412 /// \brief Returns __kmpc_dispatch_fini_* runtime function for the specified
413 /// size \a IVSize and sign \a IVSigned.
414 llvm::Constant *createDispatchFiniFunction(unsigned IVSize, bool IVSigned);
416 /// \brief If the specified mangled name is not in the module, create and
417 /// return threadprivate cache object. This object is a pointer's worth of
418 /// storage that's reserved for use by the OpenMP runtime.
419 /// \param VD Threadprivate variable.
420 /// \return Cache variable for the specified threadprivate.
421 llvm::Constant *getOrCreateThreadPrivateCache(const VarDecl *VD);
423 /// \brief Emits address of the word in a memory where current thread id is
425 virtual Address emitThreadIDAddress(CodeGenFunction &CGF, SourceLocation Loc);
427 /// \brief Gets thread id value for the current thread.
429 llvm::Value *getThreadID(CodeGenFunction &CGF, SourceLocation Loc);
431 /// \brief Gets (if variable with the given name already exist) or creates
432 /// internal global variable with the specified Name. The created variable has
433 /// linkage CommonLinkage by default and is initialized by null value.
434 /// \param Ty Type of the global variable. If it is exist already the type
435 /// must be the same.
436 /// \param Name Name of the variable.
437 llvm::Constant *getOrCreateInternalVariable(llvm::Type *Ty,
438 const llvm::Twine &Name);
440 /// \brief Set of threadprivate variables with the generated initializer.
441 llvm::SmallPtrSet<const VarDecl *, 4> ThreadPrivateWithDefinition;
443 /// \brief Emits initialization code for the threadprivate variables.
444 /// \param VDAddr Address of the global variable \a VD.
445 /// \param Ctor Pointer to a global init function for \a VD.
446 /// \param CopyCtor Pointer to a global copy function for \a VD.
447 /// \param Dtor Pointer to a global destructor function for \a VD.
448 /// \param Loc Location of threadprivate declaration.
449 void emitThreadPrivateVarInit(CodeGenFunction &CGF, Address VDAddr,
450 llvm::Value *Ctor, llvm::Value *CopyCtor,
451 llvm::Value *Dtor, SourceLocation Loc);
453 /// \brief Returns corresponding lock object for the specified critical region
454 /// name. If the lock object does not exist it is created, otherwise the
455 /// reference to the existing copy is returned.
456 /// \param CriticalName Name of the critical region.
458 llvm::Value *getCriticalRegionLock(StringRef CriticalName);
460 struct TaskResultTy {
461 llvm::Value *NewTask = nullptr;
462 llvm::Value *TaskEntry = nullptr;
463 llvm::Value *NewTaskNewTaskTTy = nullptr;
465 RecordDecl *KmpTaskTQTyRD = nullptr;
466 llvm::Value *TaskDupFn = nullptr;
468 /// Emit task region for the task directive. The task region is emitted in
470 /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32
471 /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
472 /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the
474 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
475 /// TaskFunction(gtid, tt->part_id, tt->shareds);
478 /// 2. Copy a list of shared variables to field shareds of the resulting
479 /// structure kmp_task_t returned by the previous call (if any).
480 /// 3. Copy a pointer to destructions function to field destructions of the
481 /// resulting structure kmp_task_t.
482 /// \param D Current task directive.
483 /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32
484 /// /*part_id*/, captured_struct */*__context*/);
485 /// \param SharedsTy A type which contains references the shared variables.
486 /// \param Shareds Context with the list of shared variables from the \p
488 /// \param Data Additional data for task generation like tiednsee, final
489 /// state, list of privates etc.
490 TaskResultTy emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc,
491 const OMPExecutableDirective &D,
492 llvm::Value *TaskFunction, QualType SharedsTy,
493 Address Shareds, const OMPTaskDataTy &Data);
496 explicit CGOpenMPRuntime(CodeGenModule &CGM);
497 virtual ~CGOpenMPRuntime() {}
498 virtual void clear();
500 /// Emit code for the specified user defined reduction construct.
501 virtual void emitUserDefinedReduction(CodeGenFunction *CGF,
502 const OMPDeclareReductionDecl *D);
503 /// Get combiner/initializer for the specified user-defined reduction, if any.
504 virtual std::pair<llvm::Function *, llvm::Function *>
505 getUserDefinedReduction(const OMPDeclareReductionDecl *D);
506 /// \brief Emits outlined function for the specified OpenMP parallel directive
507 /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID,
508 /// kmp_int32 BoundID, struct context_vars*).
509 /// \param D OpenMP directive.
510 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
511 /// \param InnermostKind Kind of innermost directive (for simple directives it
512 /// is a directive itself, for combined - its innermost directive).
513 /// \param CodeGen Code generation sequence for the \a D directive.
514 virtual llvm::Value *emitParallelOrTeamsOutlinedFunction(
515 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
516 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen);
518 /// \brief Emits outlined function for the OpenMP task directive \a D. This
519 /// outlined function has type void(*)(kmp_int32 ThreadID, struct task_t*
521 /// \param D OpenMP directive.
522 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
523 /// \param PartIDVar Variable for partition id in the current OpenMP untied
525 /// \param TaskTVar Variable for task_t argument.
526 /// \param InnermostKind Kind of innermost directive (for simple directives it
527 /// is a directive itself, for combined - its innermost directive).
528 /// \param CodeGen Code generation sequence for the \a D directive.
529 /// \param Tied true if task is generated for tied task, false otherwise.
530 /// \param NumberOfParts Number of parts in untied task. Ignored for tied
533 virtual llvm::Value *emitTaskOutlinedFunction(
534 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
535 const VarDecl *PartIDVar, const VarDecl *TaskTVar,
536 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
537 bool Tied, unsigned &NumberOfParts);
539 /// \brief Cleans up references to the objects in finished function.
541 void functionFinished(CodeGenFunction &CGF);
543 /// \brief Emits code for parallel or serial call of the \a OutlinedFn with
544 /// variables captured in a record which address is stored in \a
546 /// \param OutlinedFn Outlined function to be run in parallel threads. Type of
547 /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
548 /// \param CapturedVars A pointer to the record with the references to
549 /// variables used in \a OutlinedFn function.
550 /// \param IfCond Condition in the associated 'if' clause, if it was
551 /// specified, nullptr otherwise.
553 virtual void emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc,
554 llvm::Value *OutlinedFn,
555 ArrayRef<llvm::Value *> CapturedVars,
558 /// \brief Emits a critical region.
559 /// \param CriticalName Name of the critical region.
560 /// \param CriticalOpGen Generator for the statement associated with the given
562 /// \param Hint Value of the 'hint' clause (optional).
563 virtual void emitCriticalRegion(CodeGenFunction &CGF, StringRef CriticalName,
564 const RegionCodeGenTy &CriticalOpGen,
566 const Expr *Hint = nullptr);
568 /// \brief Emits a master region.
569 /// \param MasterOpGen Generator for the statement associated with the given
571 virtual void emitMasterRegion(CodeGenFunction &CGF,
572 const RegionCodeGenTy &MasterOpGen,
575 /// \brief Emits code for a taskyield directive.
576 virtual void emitTaskyieldCall(CodeGenFunction &CGF, SourceLocation Loc);
578 /// \brief Emit a taskgroup region.
579 /// \param TaskgroupOpGen Generator for the statement associated with the
580 /// given taskgroup region.
581 virtual void emitTaskgroupRegion(CodeGenFunction &CGF,
582 const RegionCodeGenTy &TaskgroupOpGen,
585 /// \brief Emits a single region.
586 /// \param SingleOpGen Generator for the statement associated with the given
588 virtual void emitSingleRegion(CodeGenFunction &CGF,
589 const RegionCodeGenTy &SingleOpGen,
591 ArrayRef<const Expr *> CopyprivateVars,
592 ArrayRef<const Expr *> DestExprs,
593 ArrayRef<const Expr *> SrcExprs,
594 ArrayRef<const Expr *> AssignmentOps);
596 /// \brief Emit an ordered region.
597 /// \param OrderedOpGen Generator for the statement associated with the given
599 virtual void emitOrderedRegion(CodeGenFunction &CGF,
600 const RegionCodeGenTy &OrderedOpGen,
601 SourceLocation Loc, bool IsThreads);
603 /// \brief Emit an implicit/explicit barrier for OpenMP threads.
604 /// \param Kind Directive for which this implicit barrier call must be
605 /// generated. Must be OMPD_barrier for explicit barrier generation.
606 /// \param EmitChecks true if need to emit checks for cancellation barriers.
607 /// \param ForceSimpleCall true simple barrier call must be emitted, false if
608 /// runtime class decides which one to emit (simple or with cancellation
611 virtual void emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc,
612 OpenMPDirectiveKind Kind,
613 bool EmitChecks = true,
614 bool ForceSimpleCall = false);
616 /// \brief Check if the specified \a ScheduleKind is static non-chunked.
617 /// This kind of worksharing directive is emitted without outer loop.
618 /// \param ScheduleKind Schedule kind specified in the 'schedule' clause.
619 /// \param Chunked True if chunk is specified in the clause.
621 virtual bool isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind,
624 /// \brief Check if the specified \a ScheduleKind is static non-chunked.
625 /// This kind of distribute directive is emitted without outer loop.
626 /// \param ScheduleKind Schedule kind specified in the 'dist_schedule' clause.
627 /// \param Chunked True if chunk is specified in the clause.
629 virtual bool isStaticNonchunked(OpenMPDistScheduleClauseKind ScheduleKind,
632 /// \brief Check if the specified \a ScheduleKind is dynamic.
633 /// This kind of worksharing directive is emitted without outer loop.
634 /// \param ScheduleKind Schedule Kind specified in the 'schedule' clause.
636 virtual bool isDynamic(OpenMPScheduleClauseKind ScheduleKind) const;
638 virtual void emitForDispatchInit(CodeGenFunction &CGF, SourceLocation Loc,
639 const OpenMPScheduleTy &ScheduleKind,
640 unsigned IVSize, bool IVSigned, bool Ordered,
642 llvm::Value *Chunk = nullptr);
644 /// \brief Call the appropriate runtime routine to initialize it before start
647 /// Depending on the loop schedule, it is nesessary to call some runtime
648 /// routine before start of the OpenMP loop to get the loop upper / lower
649 /// bounds \a LB and \a UB and stride \a ST.
651 /// \param CGF Reference to current CodeGenFunction.
652 /// \param Loc Clang source location.
653 /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause.
654 /// \param IVSize Size of the iteration variable in bits.
655 /// \param IVSigned Sign of the interation variable.
656 /// \param Ordered true if loop is ordered, false otherwise.
657 /// \param IL Address of the output variable in which the flag of the
658 /// last iteration is returned.
659 /// \param LB Address of the output variable in which the lower iteration
660 /// number is returned.
661 /// \param UB Address of the output variable in which the upper iteration
662 /// number is returned.
663 /// \param ST Address of the output variable in which the stride value is
664 /// returned nesessary to generated the static_chunked scheduled loop.
665 /// \param Chunk Value of the chunk for the static_chunked scheduled loop.
666 /// For the default (nullptr) value, the chunk 1 will be used.
668 virtual void emitForStaticInit(CodeGenFunction &CGF, SourceLocation Loc,
669 const OpenMPScheduleTy &ScheduleKind,
670 unsigned IVSize, bool IVSigned, bool Ordered,
671 Address IL, Address LB, Address UB, Address ST,
672 llvm::Value *Chunk = nullptr);
675 /// \param CGF Reference to current CodeGenFunction.
676 /// \param Loc Clang source location.
677 /// \param SchedKind Schedule kind, specified by the 'dist_schedule' clause.
678 /// \param IVSize Size of the iteration variable in bits.
679 /// \param IVSigned Sign of the interation variable.
680 /// \param Ordered true if loop is ordered, false otherwise.
681 /// \param IL Address of the output variable in which the flag of the
682 /// last iteration is returned.
683 /// \param LB Address of the output variable in which the lower iteration
684 /// number is returned.
685 /// \param UB Address of the output variable in which the upper iteration
686 /// number is returned.
687 /// \param ST Address of the output variable in which the stride value is
688 /// returned nesessary to generated the static_chunked scheduled loop.
689 /// \param Chunk Value of the chunk for the static_chunked scheduled loop.
690 /// For the default (nullptr) value, the chunk 1 will be used.
692 virtual void emitDistributeStaticInit(CodeGenFunction &CGF, SourceLocation Loc,
693 OpenMPDistScheduleClauseKind SchedKind,
694 unsigned IVSize, bool IVSigned,
695 bool Ordered, Address IL, Address LB,
696 Address UB, Address ST,
697 llvm::Value *Chunk = nullptr);
699 /// \brief Call the appropriate runtime routine to notify that we finished
700 /// iteration of the ordered loop with the dynamic scheduling.
702 /// \param CGF Reference to current CodeGenFunction.
703 /// \param Loc Clang source location.
704 /// \param IVSize Size of the iteration variable in bits.
705 /// \param IVSigned Sign of the interation variable.
707 virtual void emitForOrderedIterationEnd(CodeGenFunction &CGF,
708 SourceLocation Loc, unsigned IVSize,
711 /// \brief Call the appropriate runtime routine to notify that we finished
712 /// all the work with current loop.
714 /// \param CGF Reference to current CodeGenFunction.
715 /// \param Loc Clang source location.
717 virtual void emitForStaticFinish(CodeGenFunction &CGF, SourceLocation Loc);
719 /// Call __kmpc_dispatch_next(
720 /// ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter,
721 /// kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper,
722 /// kmp_int[32|64] *p_stride);
723 /// \param IVSize Size of the iteration variable in bits.
724 /// \param IVSigned Sign of the interation variable.
725 /// \param IL Address of the output variable in which the flag of the
726 /// last iteration is returned.
727 /// \param LB Address of the output variable in which the lower iteration
728 /// number is returned.
729 /// \param UB Address of the output variable in which the upper iteration
730 /// number is returned.
731 /// \param ST Address of the output variable in which the stride value is
733 virtual llvm::Value *emitForNext(CodeGenFunction &CGF, SourceLocation Loc,
734 unsigned IVSize, bool IVSigned,
735 Address IL, Address LB,
736 Address UB, Address ST);
738 /// \brief Emits call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32
739 /// global_tid, kmp_int32 num_threads) to generate code for 'num_threads'
741 /// \param NumThreads An integer value of threads.
742 virtual void emitNumThreadsClause(CodeGenFunction &CGF,
743 llvm::Value *NumThreads,
746 /// \brief Emit call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32
747 /// global_tid, int proc_bind) to generate code for 'proc_bind' clause.
748 virtual void emitProcBindClause(CodeGenFunction &CGF,
749 OpenMPProcBindClauseKind ProcBind,
752 /// \brief Returns address of the threadprivate variable for the current
754 /// \param VD Threadprivate variable.
755 /// \param VDAddr Address of the global variable \a VD.
756 /// \param Loc Location of the reference to threadprivate var.
757 /// \return Address of the threadprivate variable for the current thread.
758 virtual Address getAddrOfThreadPrivate(CodeGenFunction &CGF,
763 /// \brief Emit a code for initialization of threadprivate variable. It emits
764 /// a call to runtime library which adds initial value to the newly created
765 /// threadprivate variable (if it is not constant) and registers destructor
766 /// for the variable (if any).
767 /// \param VD Threadprivate variable.
768 /// \param VDAddr Address of the global variable \a VD.
769 /// \param Loc Location of threadprivate declaration.
770 /// \param PerformInit true if initialization expression is not constant.
771 virtual llvm::Function *
772 emitThreadPrivateVarDefinition(const VarDecl *VD, Address VDAddr,
773 SourceLocation Loc, bool PerformInit,
774 CodeGenFunction *CGF = nullptr);
776 /// \brief Emit flush of the variables specified in 'omp flush' directive.
777 /// \param Vars List of variables to flush.
778 virtual void emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *> Vars,
781 /// \brief Emit task region for the task directive. The task region is
782 /// emitted in several steps:
783 /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32
784 /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
785 /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the
787 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
788 /// TaskFunction(gtid, tt->part_id, tt->shareds);
791 /// 2. Copy a list of shared variables to field shareds of the resulting
792 /// structure kmp_task_t returned by the previous call (if any).
793 /// 3. Copy a pointer to destructions function to field destructions of the
794 /// resulting structure kmp_task_t.
795 /// 4. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid,
796 /// kmp_task_t *new_task), where new_task is a resulting structure from
798 /// \param D Current task directive.
799 /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32
800 /// /*part_id*/, captured_struct */*__context*/);
801 /// \param SharedsTy A type which contains references the shared variables.
802 /// \param Shareds Context with the list of shared variables from the \p
804 /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr
806 /// \param Data Additional data for task generation like tiednsee, final
807 /// state, list of privates etc.
808 virtual void emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc,
809 const OMPExecutableDirective &D,
810 llvm::Value *TaskFunction, QualType SharedsTy,
811 Address Shareds, const Expr *IfCond,
812 const OMPTaskDataTy &Data);
814 /// Emit task region for the taskloop directive. The taskloop region is
815 /// emitted in several steps:
816 /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32
817 /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
818 /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the
820 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
821 /// TaskFunction(gtid, tt->part_id, tt->shareds);
824 /// 2. Copy a list of shared variables to field shareds of the resulting
825 /// structure kmp_task_t returned by the previous call (if any).
826 /// 3. Copy a pointer to destructions function to field destructions of the
827 /// resulting structure kmp_task_t.
828 /// 4. Emit a call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t
829 /// *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int
830 /// nogroup, int sched, kmp_uint64 grainsize, void *task_dup ), where new_task
831 /// is a resulting structure from
833 /// \param D Current task directive.
834 /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32
835 /// /*part_id*/, captured_struct */*__context*/);
836 /// \param SharedsTy A type which contains references the shared variables.
837 /// \param Shareds Context with the list of shared variables from the \p
839 /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr
841 /// \param Data Additional data for task generation like tiednsee, final
842 /// state, list of privates etc.
843 virtual void emitTaskLoopCall(
844 CodeGenFunction &CGF, SourceLocation Loc, const OMPLoopDirective &D,
845 llvm::Value *TaskFunction, QualType SharedsTy, Address Shareds,
846 const Expr *IfCond, const OMPTaskDataTy &Data);
848 /// \brief Emit code for the directive that does not require outlining.
850 /// \param InnermostKind Kind of innermost directive (for simple directives it
851 /// is a directive itself, for combined - its innermost directive).
852 /// \param CodeGen Code generation sequence for the \a D directive.
853 /// \param HasCancel true if region has inner cancel directive, false
855 virtual void emitInlinedDirective(CodeGenFunction &CGF,
856 OpenMPDirectiveKind InnermostKind,
857 const RegionCodeGenTy &CodeGen,
858 bool HasCancel = false);
859 /// \brief Emit a code for reduction clause. Next code should be emitted for
863 /// static kmp_critical_name lock = { 0 };
865 /// void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
867 /// *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]);
872 /// void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]};
873 /// switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
874 /// RedList, reduce_func, &<lock>)) {
877 /// <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
879 /// __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
883 /// Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
890 /// \param Privates List of private copies for original reduction arguments.
891 /// \param LHSExprs List of LHS in \a ReductionOps reduction operations.
892 /// \param RHSExprs List of RHS in \a ReductionOps reduction operations.
893 /// \param ReductionOps List of reduction operations in form 'LHS binop RHS'
894 /// or 'operator binop(LHS, RHS)'.
895 /// \param WithNowait true if parent directive has also nowait clause, false
897 virtual void emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
898 ArrayRef<const Expr *> Privates,
899 ArrayRef<const Expr *> LHSExprs,
900 ArrayRef<const Expr *> RHSExprs,
901 ArrayRef<const Expr *> ReductionOps,
902 bool WithNowait, bool SimpleReduction);
904 /// \brief Emit code for 'taskwait' directive.
905 virtual void emitTaskwaitCall(CodeGenFunction &CGF, SourceLocation Loc);
907 /// \brief Emit code for 'cancellation point' construct.
908 /// \param CancelRegion Region kind for which the cancellation point must be
911 virtual void emitCancellationPointCall(CodeGenFunction &CGF,
913 OpenMPDirectiveKind CancelRegion);
915 /// \brief Emit code for 'cancel' construct.
916 /// \param IfCond Condition in the associated 'if' clause, if it was
917 /// specified, nullptr otherwise.
918 /// \param CancelRegion Region kind for which the cancel must be emitted.
920 virtual void emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc,
922 OpenMPDirectiveKind CancelRegion);
924 /// \brief Emit outilined function for 'target' directive.
925 /// \param D Directive to emit.
926 /// \param ParentName Name of the function that encloses the target region.
927 /// \param OutlinedFn Outlined function value to be defined by this call.
928 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
929 /// \param IsOffloadEntry True if the outlined function is an offload entry.
930 /// \param CodeGen Code generation sequence for the \a D directive.
931 /// An oulined function may not be an entry if, e.g. the if clause always
932 /// evaluates to false.
933 virtual void emitTargetOutlinedFunction(const OMPExecutableDirective &D,
934 StringRef ParentName,
935 llvm::Function *&OutlinedFn,
936 llvm::Constant *&OutlinedFnID,
938 const RegionCodeGenTy &CodeGen);
940 /// \brief Emit the target offloading code associated with \a D. The emitted
941 /// code attempts offloading the execution to the device, an the event of
942 /// a failure it executes the host version outlined in \a OutlinedFn.
943 /// \param D Directive to emit.
944 /// \param OutlinedFn Host version of the code to be offloaded.
945 /// \param OutlinedFnID ID of host version of the code to be offloaded.
946 /// \param IfCond Expression evaluated in if clause associated with the target
947 /// directive, or null if no if clause is used.
948 /// \param Device Expression evaluated in device clause associated with the
949 /// target directive, or null if no device clause is used.
950 /// \param CapturedVars Values captured in the current region.
951 virtual void emitTargetCall(CodeGenFunction &CGF,
952 const OMPExecutableDirective &D,
953 llvm::Value *OutlinedFn,
954 llvm::Value *OutlinedFnID, const Expr *IfCond,
956 ArrayRef<llvm::Value *> CapturedVars);
958 /// \brief Emit the target regions enclosed in \a GD function definition or
959 /// the function itself in case it is a valid device function. Returns true if
960 /// \a GD was dealt with successfully.
961 /// \param GD Function to scan.
962 virtual bool emitTargetFunctions(GlobalDecl GD);
964 /// \brief Emit the global variable if it is a valid device global variable.
965 /// Returns true if \a GD was dealt with successfully.
966 /// \param GD Variable declaration to emit.
967 virtual bool emitTargetGlobalVariable(GlobalDecl GD);
969 /// \brief Emit the global \a GD if it is meaningful for the target. Returns
970 /// if it was emitted succesfully.
971 /// \param GD Global to scan.
972 virtual bool emitTargetGlobal(GlobalDecl GD);
974 /// \brief Creates the offloading descriptor in the event any target region
975 /// was emitted in the current module and return the function that registers
977 virtual llvm::Function *emitRegistrationFunction();
979 /// \brief Emits code for teams call of the \a OutlinedFn with
980 /// variables captured in a record which address is stored in \a
982 /// \param OutlinedFn Outlined function to be run by team masters. Type of
983 /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
984 /// \param CapturedVars A pointer to the record with the references to
985 /// variables used in \a OutlinedFn function.
987 virtual void emitTeamsCall(CodeGenFunction &CGF,
988 const OMPExecutableDirective &D,
989 SourceLocation Loc, llvm::Value *OutlinedFn,
990 ArrayRef<llvm::Value *> CapturedVars);
992 /// \brief Emits call to void __kmpc_push_num_teams(ident_t *loc, kmp_int32
993 /// global_tid, kmp_int32 num_teams, kmp_int32 thread_limit) to generate code
994 /// for num_teams clause.
995 /// \param NumTeams An integer expression of teams.
996 /// \param ThreadLimit An integer expression of threads.
997 virtual void emitNumTeamsClause(CodeGenFunction &CGF, const Expr *NumTeams,
998 const Expr *ThreadLimit, SourceLocation Loc);
1000 /// \brief Emit the target data mapping code associated with \a D.
1001 /// \param D Directive to emit.
1002 /// \param IfCond Expression evaluated in if clause associated with the target
1003 /// directive, or null if no if clause is used.
1004 /// \param Device Expression evaluated in device clause associated with the
1005 /// target directive, or null if no device clause is used.
1006 /// \param CodeGen Function that emits the enclosed region.
1007 virtual void emitTargetDataCalls(CodeGenFunction &CGF,
1008 const OMPExecutableDirective &D,
1009 const Expr *IfCond, const Expr *Device,
1010 const RegionCodeGenTy &CodeGen);
1012 /// \brief Emit the data mapping/movement code associated with the directive
1013 /// \a D that should be of the form 'target [{enter|exit} data | update]'.
1014 /// \param D Directive to emit.
1015 /// \param IfCond Expression evaluated in if clause associated with the target
1016 /// directive, or null if no if clause is used.
1017 /// \param Device Expression evaluated in device clause associated with the
1018 /// target directive, or null if no device clause is used.
1019 virtual void emitTargetDataStandAloneCall(CodeGenFunction &CGF,
1020 const OMPExecutableDirective &D,
1022 const Expr *Device);
1024 /// Marks function \a Fn with properly mangled versions of vector functions.
1025 /// \param FD Function marked as 'declare simd'.
1026 /// \param Fn LLVM function that must be marked with 'declare simd'
1028 virtual void emitDeclareSimdFunction(const FunctionDecl *FD,
1029 llvm::Function *Fn);
1031 /// Emit initialization for doacross loop nesting support.
1032 /// \param D Loop-based construct used in doacross nesting construct.
1033 virtual void emitDoacrossInit(CodeGenFunction &CGF,
1034 const OMPLoopDirective &D);
1036 /// Emit code for doacross ordered directive with 'depend' clause.
1037 /// \param C 'depend' clause with 'sink|source' dependency kind.
1038 virtual void emitDoacrossOrdered(CodeGenFunction &CGF,
1039 const OMPDependClause *C);
1042 } // namespace CodeGen
1043 } // namespace clang