1 //===----- CGOpenMPRuntime.cpp - Interface to OpenMP Runtimes -------------===//
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 //===----------------------------------------------------------------------===//
15 #include "CGCleanup.h"
16 #include "CGOpenMPRuntime.h"
17 #include "CodeGenFunction.h"
18 #include "clang/CodeGen/ConstantInitBuilder.h"
19 #include "clang/AST/Decl.h"
20 #include "clang/AST/StmtOpenMP.h"
21 #include "llvm/ADT/ArrayRef.h"
22 #include "llvm/Bitcode/BitcodeReader.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/GlobalValue.h"
26 #include "llvm/IR/Value.h"
27 #include "llvm/Support/Format.h"
28 #include "llvm/Support/raw_ostream.h"
31 using namespace clang;
32 using namespace CodeGen;
35 /// \brief Base class for handling code generation inside OpenMP regions.
36 class CGOpenMPRegionInfo : public CodeGenFunction::CGCapturedStmtInfo {
38 /// \brief Kinds of OpenMP regions used in codegen.
39 enum CGOpenMPRegionKind {
40 /// \brief Region with outlined function for standalone 'parallel'
42 ParallelOutlinedRegion,
43 /// \brief Region with outlined function for standalone 'task' directive.
45 /// \brief Region for constructs that do not require function outlining,
46 /// like 'for', 'sections', 'atomic' etc. directives.
48 /// \brief Region with outlined function for standalone 'target' directive.
52 CGOpenMPRegionInfo(const CapturedStmt &CS,
53 const CGOpenMPRegionKind RegionKind,
54 const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind,
56 : CGCapturedStmtInfo(CS, CR_OpenMP), RegionKind(RegionKind),
57 CodeGen(CodeGen), Kind(Kind), HasCancel(HasCancel) {}
59 CGOpenMPRegionInfo(const CGOpenMPRegionKind RegionKind,
60 const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind,
62 : CGCapturedStmtInfo(CR_OpenMP), RegionKind(RegionKind), CodeGen(CodeGen),
63 Kind(Kind), HasCancel(HasCancel) {}
65 /// \brief Get a variable or parameter for storing global thread id
66 /// inside OpenMP construct.
67 virtual const VarDecl *getThreadIDVariable() const = 0;
69 /// \brief Emit the captured statement body.
70 void EmitBody(CodeGenFunction &CGF, const Stmt *S) override;
72 /// \brief Get an LValue for the current ThreadID variable.
73 /// \return LValue for thread id variable. This LValue always has type int32*.
74 virtual LValue getThreadIDVariableLValue(CodeGenFunction &CGF);
76 virtual void emitUntiedSwitch(CodeGenFunction & /*CGF*/) {}
78 CGOpenMPRegionKind getRegionKind() const { return RegionKind; }
80 OpenMPDirectiveKind getDirectiveKind() const { return Kind; }
82 bool hasCancel() const { return HasCancel; }
84 static bool classof(const CGCapturedStmtInfo *Info) {
85 return Info->getKind() == CR_OpenMP;
88 ~CGOpenMPRegionInfo() override = default;
91 CGOpenMPRegionKind RegionKind;
92 RegionCodeGenTy CodeGen;
93 OpenMPDirectiveKind Kind;
97 /// \brief API for captured statement code generation in OpenMP constructs.
98 class CGOpenMPOutlinedRegionInfo final : public CGOpenMPRegionInfo {
100 CGOpenMPOutlinedRegionInfo(const CapturedStmt &CS, const VarDecl *ThreadIDVar,
101 const RegionCodeGenTy &CodeGen,
102 OpenMPDirectiveKind Kind, bool HasCancel,
103 StringRef HelperName)
104 : CGOpenMPRegionInfo(CS, ParallelOutlinedRegion, CodeGen, Kind,
106 ThreadIDVar(ThreadIDVar), HelperName(HelperName) {
107 assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.");
110 /// \brief Get a variable or parameter for storing global thread id
111 /// inside OpenMP construct.
112 const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; }
114 /// \brief Get the name of the capture helper.
115 StringRef getHelperName() const override { return HelperName; }
117 static bool classof(const CGCapturedStmtInfo *Info) {
118 return CGOpenMPRegionInfo::classof(Info) &&
119 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() ==
120 ParallelOutlinedRegion;
124 /// \brief A variable or parameter storing global thread id for OpenMP
126 const VarDecl *ThreadIDVar;
127 StringRef HelperName;
130 /// \brief API for captured statement code generation in OpenMP constructs.
131 class CGOpenMPTaskOutlinedRegionInfo final : public CGOpenMPRegionInfo {
133 class UntiedTaskActionTy final : public PrePostActionTy {
135 const VarDecl *PartIDVar;
136 const RegionCodeGenTy UntiedCodeGen;
137 llvm::SwitchInst *UntiedSwitch = nullptr;
140 UntiedTaskActionTy(bool Tied, const VarDecl *PartIDVar,
141 const RegionCodeGenTy &UntiedCodeGen)
142 : Untied(!Tied), PartIDVar(PartIDVar), UntiedCodeGen(UntiedCodeGen) {}
143 void Enter(CodeGenFunction &CGF) override {
145 // Emit task switching point.
146 auto PartIdLVal = CGF.EmitLoadOfPointerLValue(
147 CGF.GetAddrOfLocalVar(PartIDVar),
148 PartIDVar->getType()->castAs<PointerType>());
149 auto *Res = CGF.EmitLoadOfScalar(PartIdLVal, SourceLocation());
150 auto *DoneBB = CGF.createBasicBlock(".untied.done.");
151 UntiedSwitch = CGF.Builder.CreateSwitch(Res, DoneBB);
152 CGF.EmitBlock(DoneBB);
153 CGF.EmitBranchThroughCleanup(CGF.ReturnBlock);
154 CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp."));
155 UntiedSwitch->addCase(CGF.Builder.getInt32(0),
156 CGF.Builder.GetInsertBlock());
157 emitUntiedSwitch(CGF);
160 void emitUntiedSwitch(CodeGenFunction &CGF) const {
162 auto PartIdLVal = CGF.EmitLoadOfPointerLValue(
163 CGF.GetAddrOfLocalVar(PartIDVar),
164 PartIDVar->getType()->castAs<PointerType>());
165 CGF.EmitStoreOfScalar(CGF.Builder.getInt32(UntiedSwitch->getNumCases()),
168 CodeGenFunction::JumpDest CurPoint =
169 CGF.getJumpDestInCurrentScope(".untied.next.");
170 CGF.EmitBranchThroughCleanup(CGF.ReturnBlock);
171 CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp."));
172 UntiedSwitch->addCase(CGF.Builder.getInt32(UntiedSwitch->getNumCases()),
173 CGF.Builder.GetInsertBlock());
174 CGF.EmitBranchThroughCleanup(CurPoint);
175 CGF.EmitBlock(CurPoint.getBlock());
178 unsigned getNumberOfParts() const { return UntiedSwitch->getNumCases(); }
180 CGOpenMPTaskOutlinedRegionInfo(const CapturedStmt &CS,
181 const VarDecl *ThreadIDVar,
182 const RegionCodeGenTy &CodeGen,
183 OpenMPDirectiveKind Kind, bool HasCancel,
184 const UntiedTaskActionTy &Action)
185 : CGOpenMPRegionInfo(CS, TaskOutlinedRegion, CodeGen, Kind, HasCancel),
186 ThreadIDVar(ThreadIDVar), Action(Action) {
187 assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.");
190 /// \brief Get a variable or parameter for storing global thread id
191 /// inside OpenMP construct.
192 const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; }
194 /// \brief Get an LValue for the current ThreadID variable.
195 LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override;
197 /// \brief Get the name of the capture helper.
198 StringRef getHelperName() const override { return ".omp_outlined."; }
200 void emitUntiedSwitch(CodeGenFunction &CGF) override {
201 Action.emitUntiedSwitch(CGF);
204 static bool classof(const CGCapturedStmtInfo *Info) {
205 return CGOpenMPRegionInfo::classof(Info) &&
206 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() ==
211 /// \brief A variable or parameter storing global thread id for OpenMP
213 const VarDecl *ThreadIDVar;
214 /// Action for emitting code for untied tasks.
215 const UntiedTaskActionTy &Action;
218 /// \brief API for inlined captured statement code generation in OpenMP
220 class CGOpenMPInlinedRegionInfo : public CGOpenMPRegionInfo {
222 CGOpenMPInlinedRegionInfo(CodeGenFunction::CGCapturedStmtInfo *OldCSI,
223 const RegionCodeGenTy &CodeGen,
224 OpenMPDirectiveKind Kind, bool HasCancel)
225 : CGOpenMPRegionInfo(InlinedRegion, CodeGen, Kind, HasCancel),
227 OuterRegionInfo(dyn_cast_or_null<CGOpenMPRegionInfo>(OldCSI)) {}
229 // \brief Retrieve the value of the context parameter.
230 llvm::Value *getContextValue() const override {
232 return OuterRegionInfo->getContextValue();
233 llvm_unreachable("No context value for inlined OpenMP region");
236 void setContextValue(llvm::Value *V) override {
237 if (OuterRegionInfo) {
238 OuterRegionInfo->setContextValue(V);
241 llvm_unreachable("No context value for inlined OpenMP region");
244 /// \brief Lookup the captured field decl for a variable.
245 const FieldDecl *lookup(const VarDecl *VD) const override {
247 return OuterRegionInfo->lookup(VD);
248 // If there is no outer outlined region,no need to lookup in a list of
249 // captured variables, we can use the original one.
253 FieldDecl *getThisFieldDecl() const override {
255 return OuterRegionInfo->getThisFieldDecl();
259 /// \brief Get a variable or parameter for storing global thread id
260 /// inside OpenMP construct.
261 const VarDecl *getThreadIDVariable() const override {
263 return OuterRegionInfo->getThreadIDVariable();
267 /// \brief Get the name of the capture helper.
268 StringRef getHelperName() const override {
269 if (auto *OuterRegionInfo = getOldCSI())
270 return OuterRegionInfo->getHelperName();
271 llvm_unreachable("No helper name for inlined OpenMP construct");
274 void emitUntiedSwitch(CodeGenFunction &CGF) override {
276 OuterRegionInfo->emitUntiedSwitch(CGF);
279 CodeGenFunction::CGCapturedStmtInfo *getOldCSI() const { return OldCSI; }
281 static bool classof(const CGCapturedStmtInfo *Info) {
282 return CGOpenMPRegionInfo::classof(Info) &&
283 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == InlinedRegion;
286 ~CGOpenMPInlinedRegionInfo() override = default;
289 /// \brief CodeGen info about outer OpenMP region.
290 CodeGenFunction::CGCapturedStmtInfo *OldCSI;
291 CGOpenMPRegionInfo *OuterRegionInfo;
294 /// \brief API for captured statement code generation in OpenMP target
295 /// constructs. For this captures, implicit parameters are used instead of the
296 /// captured fields. The name of the target region has to be unique in a given
297 /// application so it is provided by the client, because only the client has
298 /// the information to generate that.
299 class CGOpenMPTargetRegionInfo final : public CGOpenMPRegionInfo {
301 CGOpenMPTargetRegionInfo(const CapturedStmt &CS,
302 const RegionCodeGenTy &CodeGen, StringRef HelperName)
303 : CGOpenMPRegionInfo(CS, TargetRegion, CodeGen, OMPD_target,
304 /*HasCancel=*/false),
305 HelperName(HelperName) {}
307 /// \brief This is unused for target regions because each starts executing
308 /// with a single thread.
309 const VarDecl *getThreadIDVariable() const override { return nullptr; }
311 /// \brief Get the name of the capture helper.
312 StringRef getHelperName() const override { return HelperName; }
314 static bool classof(const CGCapturedStmtInfo *Info) {
315 return CGOpenMPRegionInfo::classof(Info) &&
316 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == TargetRegion;
320 StringRef HelperName;
323 static void EmptyCodeGen(CodeGenFunction &, PrePostActionTy &) {
324 llvm_unreachable("No codegen for expressions");
326 /// \brief API for generation of expressions captured in a innermost OpenMP
328 class CGOpenMPInnerExprInfo final : public CGOpenMPInlinedRegionInfo {
330 CGOpenMPInnerExprInfo(CodeGenFunction &CGF, const CapturedStmt &CS)
331 : CGOpenMPInlinedRegionInfo(CGF.CapturedStmtInfo, EmptyCodeGen,
333 /*HasCancel=*/false),
335 // Make sure the globals captured in the provided statement are local by
336 // using the privatization logic. We assume the same variable is not
337 // captured more than once.
338 for (auto &C : CS.captures()) {
339 if (!C.capturesVariable() && !C.capturesVariableByCopy())
342 const VarDecl *VD = C.getCapturedVar();
343 if (VD->isLocalVarDeclOrParm())
346 DeclRefExpr DRE(const_cast<VarDecl *>(VD),
347 /*RefersToEnclosingVariableOrCapture=*/false,
348 VD->getType().getNonReferenceType(), VK_LValue,
350 PrivScope.addPrivate(VD, [&CGF, &DRE]() -> Address {
351 return CGF.EmitLValue(&DRE).getAddress();
354 (void)PrivScope.Privatize();
357 /// \brief Lookup the captured field decl for a variable.
358 const FieldDecl *lookup(const VarDecl *VD) const override {
359 if (auto *FD = CGOpenMPInlinedRegionInfo::lookup(VD))
364 /// \brief Emit the captured statement body.
365 void EmitBody(CodeGenFunction &CGF, const Stmt *S) override {
366 llvm_unreachable("No body for expressions");
369 /// \brief Get a variable or parameter for storing global thread id
370 /// inside OpenMP construct.
371 const VarDecl *getThreadIDVariable() const override {
372 llvm_unreachable("No thread id for expressions");
375 /// \brief Get the name of the capture helper.
376 StringRef getHelperName() const override {
377 llvm_unreachable("No helper name for expressions");
380 static bool classof(const CGCapturedStmtInfo *Info) { return false; }
383 /// Private scope to capture global variables.
384 CodeGenFunction::OMPPrivateScope PrivScope;
387 /// \brief RAII for emitting code of OpenMP constructs.
388 class InlinedOpenMPRegionRAII {
389 CodeGenFunction &CGF;
390 llvm::DenseMap<const VarDecl *, FieldDecl *> LambdaCaptureFields;
391 FieldDecl *LambdaThisCaptureField = nullptr;
394 /// \brief Constructs region for combined constructs.
395 /// \param CodeGen Code generation sequence for combined directives. Includes
396 /// a list of functions used for code generation of implicitly inlined
398 InlinedOpenMPRegionRAII(CodeGenFunction &CGF, const RegionCodeGenTy &CodeGen,
399 OpenMPDirectiveKind Kind, bool HasCancel)
401 // Start emission for the construct.
402 CGF.CapturedStmtInfo = new CGOpenMPInlinedRegionInfo(
403 CGF.CapturedStmtInfo, CodeGen, Kind, HasCancel);
404 std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields);
405 LambdaThisCaptureField = CGF.LambdaThisCaptureField;
406 CGF.LambdaThisCaptureField = nullptr;
409 ~InlinedOpenMPRegionRAII() {
410 // Restore original CapturedStmtInfo only if we're done with code emission.
412 cast<CGOpenMPInlinedRegionInfo>(CGF.CapturedStmtInfo)->getOldCSI();
413 delete CGF.CapturedStmtInfo;
414 CGF.CapturedStmtInfo = OldCSI;
415 std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields);
416 CGF.LambdaThisCaptureField = LambdaThisCaptureField;
420 /// \brief Values for bit flags used in the ident_t to describe the fields.
421 /// All enumeric elements are named and described in accordance with the code
422 /// from http://llvm.org/svn/llvm-project/openmp/trunk/runtime/src/kmp.h
423 enum OpenMPLocationFlags {
424 /// \brief Use trampoline for internal microtask.
425 OMP_IDENT_IMD = 0x01,
426 /// \brief Use c-style ident structure.
427 OMP_IDENT_KMPC = 0x02,
428 /// \brief Atomic reduction option for kmpc_reduce.
429 OMP_ATOMIC_REDUCE = 0x10,
430 /// \brief Explicit 'barrier' directive.
431 OMP_IDENT_BARRIER_EXPL = 0x20,
432 /// \brief Implicit barrier in code.
433 OMP_IDENT_BARRIER_IMPL = 0x40,
434 /// \brief Implicit barrier in 'for' directive.
435 OMP_IDENT_BARRIER_IMPL_FOR = 0x40,
436 /// \brief Implicit barrier in 'sections' directive.
437 OMP_IDENT_BARRIER_IMPL_SECTIONS = 0xC0,
438 /// \brief Implicit barrier in 'single' directive.
439 OMP_IDENT_BARRIER_IMPL_SINGLE = 0x140
442 /// \brief Describes ident structure that describes a source location.
443 /// All descriptions are taken from
444 /// http://llvm.org/svn/llvm-project/openmp/trunk/runtime/src/kmp.h
445 /// Original structure:
446 /// typedef struct ident {
447 /// kmp_int32 reserved_1; /**< might be used in Fortran;
449 /// kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags;
450 /// KMP_IDENT_KMPC identifies this union
452 /// kmp_int32 reserved_2; /**< not really used in Fortran any more;
455 /// /* but currently used for storing
456 /// region-specific ITT */
457 /// /* contextual information. */
458 ///#endif /* USE_ITT_BUILD */
459 /// kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for
461 /// char const *psource; /**< String describing the source location.
462 /// The string is composed of semi-colon separated
463 // fields which describe the source file,
464 /// the function and a pair of line numbers that
465 /// delimit the construct.
468 enum IdentFieldIndex {
469 /// \brief might be used in Fortran
470 IdentField_Reserved_1,
471 /// \brief OMP_IDENT_xxx flags; OMP_IDENT_KMPC identifies this union member.
473 /// \brief Not really used in Fortran any more
474 IdentField_Reserved_2,
475 /// \brief Source[4] in Fortran, do not use for C++
476 IdentField_Reserved_3,
477 /// \brief String describing the source location. The string is composed of
478 /// semi-colon separated fields which describe the source file, the function
479 /// and a pair of line numbers that delimit the construct.
483 /// \brief Schedule types for 'omp for' loops (these enumerators are taken from
484 /// the enum sched_type in kmp.h).
485 enum OpenMPSchedType {
486 /// \brief Lower bound for default (unordered) versions.
488 OMP_sch_static_chunked = 33,
490 OMP_sch_dynamic_chunked = 35,
491 OMP_sch_guided_chunked = 36,
492 OMP_sch_runtime = 37,
494 /// static with chunk adjustment (e.g., simd)
495 OMP_sch_static_balanced_chunked = 45,
496 /// \brief Lower bound for 'ordered' versions.
498 OMP_ord_static_chunked = 65,
500 OMP_ord_dynamic_chunked = 67,
501 OMP_ord_guided_chunked = 68,
502 OMP_ord_runtime = 69,
504 OMP_sch_default = OMP_sch_static,
505 /// \brief dist_schedule types
506 OMP_dist_sch_static_chunked = 91,
507 OMP_dist_sch_static = 92,
508 /// Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers.
509 /// Set if the monotonic schedule modifier was present.
510 OMP_sch_modifier_monotonic = (1 << 29),
511 /// Set if the nonmonotonic schedule modifier was present.
512 OMP_sch_modifier_nonmonotonic = (1 << 30),
515 enum OpenMPRTLFunction {
516 /// \brief Call to void __kmpc_fork_call(ident_t *loc, kmp_int32 argc,
517 /// kmpc_micro microtask, ...);
518 OMPRTL__kmpc_fork_call,
519 /// \brief Call to void *__kmpc_threadprivate_cached(ident_t *loc,
520 /// kmp_int32 global_tid, void *data, size_t size, void ***cache);
521 OMPRTL__kmpc_threadprivate_cached,
522 /// \brief Call to void __kmpc_threadprivate_register( ident_t *,
523 /// void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor);
524 OMPRTL__kmpc_threadprivate_register,
525 // Call to __kmpc_int32 kmpc_global_thread_num(ident_t *loc);
526 OMPRTL__kmpc_global_thread_num,
527 // Call to void __kmpc_critical(ident_t *loc, kmp_int32 global_tid,
528 // kmp_critical_name *crit);
529 OMPRTL__kmpc_critical,
530 // Call to void __kmpc_critical_with_hint(ident_t *loc, kmp_int32
531 // global_tid, kmp_critical_name *crit, uintptr_t hint);
532 OMPRTL__kmpc_critical_with_hint,
533 // Call to void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid,
534 // kmp_critical_name *crit);
535 OMPRTL__kmpc_end_critical,
536 // Call to kmp_int32 __kmpc_cancel_barrier(ident_t *loc, kmp_int32
538 OMPRTL__kmpc_cancel_barrier,
539 // Call to void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
540 OMPRTL__kmpc_barrier,
541 // Call to void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
542 OMPRTL__kmpc_for_static_fini,
543 // Call to void __kmpc_serialized_parallel(ident_t *loc, kmp_int32
545 OMPRTL__kmpc_serialized_parallel,
546 // Call to void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32
548 OMPRTL__kmpc_end_serialized_parallel,
549 // Call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
550 // kmp_int32 num_threads);
551 OMPRTL__kmpc_push_num_threads,
552 // Call to void __kmpc_flush(ident_t *loc);
554 // Call to kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
556 // Call to void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
557 OMPRTL__kmpc_end_master,
558 // Call to kmp_int32 __kmpc_omp_taskyield(ident_t *, kmp_int32 global_tid,
560 OMPRTL__kmpc_omp_taskyield,
561 // Call to kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
563 // Call to void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
564 OMPRTL__kmpc_end_single,
565 // Call to kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid,
566 // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
567 // kmp_routine_entry_t *task_entry);
568 OMPRTL__kmpc_omp_task_alloc,
569 // Call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *
571 OMPRTL__kmpc_omp_task,
572 // Call to void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
573 // size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *),
575 OMPRTL__kmpc_copyprivate,
576 // Call to kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid,
577 // kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void
578 // (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck);
580 // Call to kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32
581 // global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data,
582 // void (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name
584 OMPRTL__kmpc_reduce_nowait,
585 // Call to void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
586 // kmp_critical_name *lck);
587 OMPRTL__kmpc_end_reduce,
588 // Call to void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
589 // kmp_critical_name *lck);
590 OMPRTL__kmpc_end_reduce_nowait,
591 // Call to void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid,
592 // kmp_task_t * new_task);
593 OMPRTL__kmpc_omp_task_begin_if0,
594 // Call to void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid,
595 // kmp_task_t * new_task);
596 OMPRTL__kmpc_omp_task_complete_if0,
597 // Call to void __kmpc_ordered(ident_t *loc, kmp_int32 global_tid);
598 OMPRTL__kmpc_ordered,
599 // Call to void __kmpc_end_ordered(ident_t *loc, kmp_int32 global_tid);
600 OMPRTL__kmpc_end_ordered,
601 // Call to kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
603 OMPRTL__kmpc_omp_taskwait,
604 // Call to void __kmpc_taskgroup(ident_t *loc, kmp_int32 global_tid);
605 OMPRTL__kmpc_taskgroup,
606 // Call to void __kmpc_end_taskgroup(ident_t *loc, kmp_int32 global_tid);
607 OMPRTL__kmpc_end_taskgroup,
608 // Call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
610 OMPRTL__kmpc_push_proc_bind,
611 // Call to kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32
612 // gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t
613 // *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list);
614 OMPRTL__kmpc_omp_task_with_deps,
615 // Call to void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32
616 // gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32
617 // ndeps_noalias, kmp_depend_info_t *noalias_dep_list);
618 OMPRTL__kmpc_omp_wait_deps,
619 // Call to kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32
620 // global_tid, kmp_int32 cncl_kind);
621 OMPRTL__kmpc_cancellationpoint,
622 // Call to kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid,
623 // kmp_int32 cncl_kind);
625 // Call to void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
626 // kmp_int32 num_teams, kmp_int32 thread_limit);
627 OMPRTL__kmpc_push_num_teams,
628 // Call to void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro
630 OMPRTL__kmpc_fork_teams,
631 // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int
632 // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int
633 // sched, kmp_uint64 grainsize, void *task_dup);
634 OMPRTL__kmpc_taskloop,
635 // Call to void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, kmp_int32
636 // num_dims, struct kmp_dim *dims);
637 OMPRTL__kmpc_doacross_init,
638 // Call to void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
639 OMPRTL__kmpc_doacross_fini,
640 // Call to void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, kmp_int64
642 OMPRTL__kmpc_doacross_post,
643 // Call to void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, kmp_int64
645 OMPRTL__kmpc_doacross_wait,
648 // Offloading related calls
650 // Call to int32_t __tgt_target(int32_t device_id, void *host_ptr, int32_t
651 // arg_num, void** args_base, void **args, size_t *arg_sizes, int32_t
654 // Call to int32_t __tgt_target_teams(int32_t device_id, void *host_ptr,
655 // int32_t arg_num, void** args_base, void **args, size_t *arg_sizes,
656 // int32_t *arg_types, int32_t num_teams, int32_t thread_limit);
657 OMPRTL__tgt_target_teams,
658 // Call to void __tgt_register_lib(__tgt_bin_desc *desc);
659 OMPRTL__tgt_register_lib,
660 // Call to void __tgt_unregister_lib(__tgt_bin_desc *desc);
661 OMPRTL__tgt_unregister_lib,
662 // Call to void __tgt_target_data_begin(int32_t device_id, int32_t arg_num,
663 // void** args_base, void **args, size_t *arg_sizes, int32_t *arg_types);
664 OMPRTL__tgt_target_data_begin,
665 // Call to void __tgt_target_data_end(int32_t device_id, int32_t arg_num,
666 // void** args_base, void **args, size_t *arg_sizes, int32_t *arg_types);
667 OMPRTL__tgt_target_data_end,
668 // Call to void __tgt_target_data_update(int32_t device_id, int32_t arg_num,
669 // void** args_base, void **args, size_t *arg_sizes, int32_t *arg_types);
670 OMPRTL__tgt_target_data_update,
673 /// A basic class for pre|post-action for advanced codegen sequence for OpenMP
675 class CleanupTy final : public EHScopeStack::Cleanup {
676 PrePostActionTy *Action;
679 explicit CleanupTy(PrePostActionTy *Action) : Action(Action) {}
680 void Emit(CodeGenFunction &CGF, Flags /*flags*/) override {
681 if (!CGF.HaveInsertPoint())
687 } // anonymous namespace
689 void RegionCodeGenTy::operator()(CodeGenFunction &CGF) const {
690 CodeGenFunction::RunCleanupsScope Scope(CGF);
692 CGF.EHStack.pushCleanup<CleanupTy>(NormalAndEHCleanup, PrePostAction);
693 Callback(CodeGen, CGF, *PrePostAction);
695 PrePostActionTy Action;
696 Callback(CodeGen, CGF, Action);
700 LValue CGOpenMPRegionInfo::getThreadIDVariableLValue(CodeGenFunction &CGF) {
701 return CGF.EmitLoadOfPointerLValue(
702 CGF.GetAddrOfLocalVar(getThreadIDVariable()),
703 getThreadIDVariable()->getType()->castAs<PointerType>());
706 void CGOpenMPRegionInfo::EmitBody(CodeGenFunction &CGF, const Stmt * /*S*/) {
707 if (!CGF.HaveInsertPoint())
709 // 1.2.2 OpenMP Language Terminology
710 // Structured block - An executable statement with a single entry at the
711 // top and a single exit at the bottom.
712 // The point of exit cannot be a branch out of the structured block.
713 // longjmp() and throw() must not violate the entry/exit criteria.
714 CGF.EHStack.pushTerminate();
716 CGF.EHStack.popTerminate();
719 LValue CGOpenMPTaskOutlinedRegionInfo::getThreadIDVariableLValue(
720 CodeGenFunction &CGF) {
721 return CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(getThreadIDVariable()),
722 getThreadIDVariable()->getType(),
723 AlignmentSource::Decl);
726 CGOpenMPRuntime::CGOpenMPRuntime(CodeGenModule &CGM)
727 : CGM(CGM), OffloadEntriesInfoManager(CGM) {
728 IdentTy = llvm::StructType::create(
729 "ident_t", CGM.Int32Ty /* reserved_1 */, CGM.Int32Ty /* flags */,
730 CGM.Int32Ty /* reserved_2 */, CGM.Int32Ty /* reserved_3 */,
731 CGM.Int8PtrTy /* psource */, nullptr);
732 KmpCriticalNameTy = llvm::ArrayType::get(CGM.Int32Ty, /*NumElements*/ 8);
734 loadOffloadInfoMetadata();
737 void CGOpenMPRuntime::clear() {
738 InternalVars.clear();
741 static llvm::Function *
742 emitCombinerOrInitializer(CodeGenModule &CGM, QualType Ty,
743 const Expr *CombinerInitializer, const VarDecl *In,
744 const VarDecl *Out, bool IsCombiner) {
745 // void .omp_combiner.(Ty *in, Ty *out);
746 auto &C = CGM.getContext();
747 QualType PtrTy = C.getPointerType(Ty).withRestrict();
748 FunctionArgList Args;
749 ImplicitParamDecl OmpOutParm(C, /*DC=*/nullptr, Out->getLocation(),
750 /*Id=*/nullptr, PtrTy);
751 ImplicitParamDecl OmpInParm(C, /*DC=*/nullptr, In->getLocation(),
752 /*Id=*/nullptr, PtrTy);
753 Args.push_back(&OmpOutParm);
754 Args.push_back(&OmpInParm);
756 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
757 auto *FnTy = CGM.getTypes().GetFunctionType(FnInfo);
758 auto *Fn = llvm::Function::Create(
759 FnTy, llvm::GlobalValue::InternalLinkage,
760 IsCombiner ? ".omp_combiner." : ".omp_initializer.", &CGM.getModule());
761 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, Fn, FnInfo);
762 Fn->removeFnAttr(llvm::Attribute::NoInline);
763 Fn->addFnAttr(llvm::Attribute::AlwaysInline);
764 CodeGenFunction CGF(CGM);
765 // Map "T omp_in;" variable to "*omp_in_parm" value in all expressions.
766 // Map "T omp_out;" variable to "*omp_out_parm" value in all expressions.
767 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args);
768 CodeGenFunction::OMPPrivateScope Scope(CGF);
769 Address AddrIn = CGF.GetAddrOfLocalVar(&OmpInParm);
770 Scope.addPrivate(In, [&CGF, AddrIn, PtrTy]() -> Address {
771 return CGF.EmitLoadOfPointerLValue(AddrIn, PtrTy->castAs<PointerType>())
774 Address AddrOut = CGF.GetAddrOfLocalVar(&OmpOutParm);
775 Scope.addPrivate(Out, [&CGF, AddrOut, PtrTy]() -> Address {
776 return CGF.EmitLoadOfPointerLValue(AddrOut, PtrTy->castAs<PointerType>())
779 (void)Scope.Privatize();
780 CGF.EmitIgnoredExpr(CombinerInitializer);
781 Scope.ForceCleanup();
782 CGF.FinishFunction();
786 void CGOpenMPRuntime::emitUserDefinedReduction(
787 CodeGenFunction *CGF, const OMPDeclareReductionDecl *D) {
788 if (UDRMap.count(D) > 0)
790 auto &C = CGM.getContext();
792 In = &C.Idents.get("omp_in");
793 Out = &C.Idents.get("omp_out");
795 llvm::Function *Combiner = emitCombinerOrInitializer(
796 CGM, D->getType(), D->getCombiner(), cast<VarDecl>(D->lookup(In).front()),
797 cast<VarDecl>(D->lookup(Out).front()),
798 /*IsCombiner=*/true);
799 llvm::Function *Initializer = nullptr;
800 if (auto *Init = D->getInitializer()) {
801 if (!Priv || !Orig) {
802 Priv = &C.Idents.get("omp_priv");
803 Orig = &C.Idents.get("omp_orig");
805 Initializer = emitCombinerOrInitializer(
806 CGM, D->getType(), Init, cast<VarDecl>(D->lookup(Orig).front()),
807 cast<VarDecl>(D->lookup(Priv).front()),
808 /*IsCombiner=*/false);
810 UDRMap.insert(std::make_pair(D, std::make_pair(Combiner, Initializer)));
812 auto &Decls = FunctionUDRMap.FindAndConstruct(CGF->CurFn);
813 Decls.second.push_back(D);
817 std::pair<llvm::Function *, llvm::Function *>
818 CGOpenMPRuntime::getUserDefinedReduction(const OMPDeclareReductionDecl *D) {
819 auto I = UDRMap.find(D);
820 if (I != UDRMap.end())
822 emitUserDefinedReduction(/*CGF=*/nullptr, D);
823 return UDRMap.lookup(D);
826 // Layout information for ident_t.
827 static CharUnits getIdentAlign(CodeGenModule &CGM) {
828 return CGM.getPointerAlign();
830 static CharUnits getIdentSize(CodeGenModule &CGM) {
831 assert((4 * CGM.getPointerSize()).isMultipleOf(CGM.getPointerAlign()));
832 return CharUnits::fromQuantity(16) + CGM.getPointerSize();
834 static CharUnits getOffsetOfIdentField(IdentFieldIndex Field) {
835 // All the fields except the last are i32, so this works beautifully.
836 return unsigned(Field) * CharUnits::fromQuantity(4);
838 static Address createIdentFieldGEP(CodeGenFunction &CGF, Address Addr,
839 IdentFieldIndex Field,
840 const llvm::Twine &Name = "") {
841 auto Offset = getOffsetOfIdentField(Field);
842 return CGF.Builder.CreateStructGEP(Addr, Field, Offset, Name);
845 static llvm::Value *emitParallelOrTeamsOutlinedFunction(
846 CodeGenModule &CGM, const OMPExecutableDirective &D, const CapturedStmt *CS,
847 const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
848 const StringRef OutlinedHelperName, const RegionCodeGenTy &CodeGen) {
849 assert(ThreadIDVar->getType()->isPointerType() &&
850 "thread id variable must be of type kmp_int32 *");
851 CodeGenFunction CGF(CGM, true);
852 bool HasCancel = false;
853 if (auto *OPD = dyn_cast<OMPParallelDirective>(&D))
854 HasCancel = OPD->hasCancel();
855 else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&D))
856 HasCancel = OPSD->hasCancel();
857 else if (auto *OPFD = dyn_cast<OMPParallelForDirective>(&D))
858 HasCancel = OPFD->hasCancel();
859 CGOpenMPOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, InnermostKind,
860 HasCancel, OutlinedHelperName);
861 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
862 return CGF.GenerateOpenMPCapturedStmtFunction(*CS);
865 llvm::Value *CGOpenMPRuntime::emitParallelOutlinedFunction(
866 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
867 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
868 const CapturedStmt *CS = D.getCapturedStmt(OMPD_parallel);
869 return emitParallelOrTeamsOutlinedFunction(
870 CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen);
873 llvm::Value *CGOpenMPRuntime::emitTeamsOutlinedFunction(
874 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
875 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
876 const CapturedStmt *CS = D.getCapturedStmt(OMPD_teams);
877 return emitParallelOrTeamsOutlinedFunction(
878 CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen);
881 llvm::Value *CGOpenMPRuntime::emitTaskOutlinedFunction(
882 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
883 const VarDecl *PartIDVar, const VarDecl *TaskTVar,
884 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
885 bool Tied, unsigned &NumberOfParts) {
886 auto &&UntiedCodeGen = [this, &D, TaskTVar](CodeGenFunction &CGF,
888 auto *ThreadID = getThreadID(CGF, D.getLocStart());
889 auto *UpLoc = emitUpdateLocation(CGF, D.getLocStart());
890 llvm::Value *TaskArgs[] = {
892 CGF.EmitLoadOfPointerLValue(CGF.GetAddrOfLocalVar(TaskTVar),
893 TaskTVar->getType()->castAs<PointerType>())
895 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_task), TaskArgs);
897 CGOpenMPTaskOutlinedRegionInfo::UntiedTaskActionTy Action(Tied, PartIDVar,
899 CodeGen.setAction(Action);
900 assert(!ThreadIDVar->getType()->isPointerType() &&
901 "thread id variable must be of type kmp_int32 for tasks");
902 auto *CS = cast<CapturedStmt>(D.getAssociatedStmt());
903 auto *TD = dyn_cast<OMPTaskDirective>(&D);
904 CodeGenFunction CGF(CGM, true);
905 CGOpenMPTaskOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen,
907 TD ? TD->hasCancel() : false, Action);
908 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
909 auto *Res = CGF.GenerateCapturedStmtFunction(*CS);
911 NumberOfParts = Action.getNumberOfParts();
915 Address CGOpenMPRuntime::getOrCreateDefaultLocation(unsigned Flags) {
916 CharUnits Align = getIdentAlign(CGM);
917 llvm::Value *Entry = OpenMPDefaultLocMap.lookup(Flags);
919 if (!DefaultOpenMPPSource) {
920 // Initialize default location for psource field of ident_t structure of
921 // all ident_t objects. Format is ";file;function;line;column;;".
923 // http://llvm.org/svn/llvm-project/openmp/trunk/runtime/src/kmp_str.c
924 DefaultOpenMPPSource =
925 CGM.GetAddrOfConstantCString(";unknown;unknown;0;0;;").getPointer();
926 DefaultOpenMPPSource =
927 llvm::ConstantExpr::getBitCast(DefaultOpenMPPSource, CGM.Int8PtrTy);
930 ConstantInitBuilder builder(CGM);
931 auto fields = builder.beginStruct(IdentTy);
932 fields.addInt(CGM.Int32Ty, 0);
933 fields.addInt(CGM.Int32Ty, Flags);
934 fields.addInt(CGM.Int32Ty, 0);
935 fields.addInt(CGM.Int32Ty, 0);
936 fields.add(DefaultOpenMPPSource);
937 auto DefaultOpenMPLocation =
938 fields.finishAndCreateGlobal("", Align, /*isConstant*/ true,
939 llvm::GlobalValue::PrivateLinkage);
940 DefaultOpenMPLocation->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
942 OpenMPDefaultLocMap[Flags] = Entry = DefaultOpenMPLocation;
944 return Address(Entry, Align);
947 llvm::Value *CGOpenMPRuntime::emitUpdateLocation(CodeGenFunction &CGF,
950 Flags |= OMP_IDENT_KMPC;
951 // If no debug info is generated - return global default location.
952 if (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo ||
954 return getOrCreateDefaultLocation(Flags).getPointer();
956 assert(CGF.CurFn && "No function in current CodeGenFunction.");
958 Address LocValue = Address::invalid();
959 auto I = OpenMPLocThreadIDMap.find(CGF.CurFn);
960 if (I != OpenMPLocThreadIDMap.end())
961 LocValue = Address(I->second.DebugLoc, getIdentAlign(CGF.CGM));
963 // OpenMPLocThreadIDMap may have null DebugLoc and non-null ThreadID, if
964 // GetOpenMPThreadID was called before this routine.
965 if (!LocValue.isValid()) {
966 // Generate "ident_t .kmpc_loc.addr;"
967 Address AI = CGF.CreateTempAlloca(IdentTy, getIdentAlign(CGF.CGM),
969 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
970 Elem.second.DebugLoc = AI.getPointer();
973 CGBuilderTy::InsertPointGuard IPG(CGF.Builder);
974 CGF.Builder.SetInsertPoint(CGF.AllocaInsertPt);
975 CGF.Builder.CreateMemCpy(LocValue, getOrCreateDefaultLocation(Flags),
976 CGM.getSize(getIdentSize(CGF.CGM)));
979 // char **psource = &.kmpc_loc_<flags>.addr.psource;
980 Address PSource = createIdentFieldGEP(CGF, LocValue, IdentField_PSource);
982 auto OMPDebugLoc = OpenMPDebugLocMap.lookup(Loc.getRawEncoding());
983 if (OMPDebugLoc == nullptr) {
984 SmallString<128> Buffer2;
985 llvm::raw_svector_ostream OS2(Buffer2);
986 // Build debug location
987 PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc);
988 OS2 << ";" << PLoc.getFilename() << ";";
989 if (const FunctionDecl *FD =
990 dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) {
991 OS2 << FD->getQualifiedNameAsString();
993 OS2 << ";" << PLoc.getLine() << ";" << PLoc.getColumn() << ";;";
994 OMPDebugLoc = CGF.Builder.CreateGlobalStringPtr(OS2.str());
995 OpenMPDebugLocMap[Loc.getRawEncoding()] = OMPDebugLoc;
997 // *psource = ";<File>;<Function>;<Line>;<Column>;;";
998 CGF.Builder.CreateStore(OMPDebugLoc, PSource);
1000 // Our callers always pass this to a runtime function, so for
1001 // convenience, go ahead and return a naked pointer.
1002 return LocValue.getPointer();
1005 llvm::Value *CGOpenMPRuntime::getThreadID(CodeGenFunction &CGF,
1006 SourceLocation Loc) {
1007 assert(CGF.CurFn && "No function in current CodeGenFunction.");
1009 llvm::Value *ThreadID = nullptr;
1010 // Check whether we've already cached a load of the thread id in this
1012 auto I = OpenMPLocThreadIDMap.find(CGF.CurFn);
1013 if (I != OpenMPLocThreadIDMap.end()) {
1014 ThreadID = I->second.ThreadID;
1015 if (ThreadID != nullptr)
1018 if (auto *OMPRegionInfo =
1019 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
1020 if (OMPRegionInfo->getThreadIDVariable()) {
1021 // Check if this an outlined function with thread id passed as argument.
1022 auto LVal = OMPRegionInfo->getThreadIDVariableLValue(CGF);
1023 ThreadID = CGF.EmitLoadOfLValue(LVal, Loc).getScalarVal();
1024 // If value loaded in entry block, cache it and use it everywhere in
1026 if (CGF.Builder.GetInsertBlock() == CGF.AllocaInsertPt->getParent()) {
1027 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
1028 Elem.second.ThreadID = ThreadID;
1034 // This is not an outlined function region - need to call __kmpc_int32
1035 // kmpc_global_thread_num(ident_t *loc).
1036 // Generate thread id value and cache this value for use across the
1038 CGBuilderTy::InsertPointGuard IPG(CGF.Builder);
1039 CGF.Builder.SetInsertPoint(CGF.AllocaInsertPt);
1041 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_global_thread_num),
1042 emitUpdateLocation(CGF, Loc));
1043 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
1044 Elem.second.ThreadID = ThreadID;
1048 void CGOpenMPRuntime::functionFinished(CodeGenFunction &CGF) {
1049 assert(CGF.CurFn && "No function in current CodeGenFunction.");
1050 if (OpenMPLocThreadIDMap.count(CGF.CurFn))
1051 OpenMPLocThreadIDMap.erase(CGF.CurFn);
1052 if (FunctionUDRMap.count(CGF.CurFn) > 0) {
1053 for(auto *D : FunctionUDRMap[CGF.CurFn]) {
1056 FunctionUDRMap.erase(CGF.CurFn);
1060 llvm::Type *CGOpenMPRuntime::getIdentTyPointerTy() {
1063 return llvm::PointerType::getUnqual(IdentTy);
1066 llvm::Type *CGOpenMPRuntime::getKmpc_MicroPointerTy() {
1067 if (!Kmpc_MicroTy) {
1068 // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
1069 llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty),
1070 llvm::PointerType::getUnqual(CGM.Int32Ty)};
1071 Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true);
1073 return llvm::PointerType::getUnqual(Kmpc_MicroTy);
1077 CGOpenMPRuntime::createRuntimeFunction(unsigned Function) {
1078 llvm::Constant *RTLFn = nullptr;
1079 switch (static_cast<OpenMPRTLFunction>(Function)) {
1080 case OMPRTL__kmpc_fork_call: {
1081 // Build void __kmpc_fork_call(ident_t *loc, kmp_int32 argc, kmpc_micro
1083 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1084 getKmpc_MicroPointerTy()};
1085 llvm::FunctionType *FnTy =
1086 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true);
1087 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_fork_call");
1090 case OMPRTL__kmpc_global_thread_num: {
1091 // Build kmp_int32 __kmpc_global_thread_num(ident_t *loc);
1092 llvm::Type *TypeParams[] = {getIdentTyPointerTy()};
1093 llvm::FunctionType *FnTy =
1094 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1095 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_global_thread_num");
1098 case OMPRTL__kmpc_threadprivate_cached: {
1099 // Build void *__kmpc_threadprivate_cached(ident_t *loc,
1100 // kmp_int32 global_tid, void *data, size_t size, void ***cache);
1101 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1102 CGM.VoidPtrTy, CGM.SizeTy,
1103 CGM.VoidPtrTy->getPointerTo()->getPointerTo()};
1104 llvm::FunctionType *FnTy =
1105 llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg*/ false);
1106 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_threadprivate_cached");
1109 case OMPRTL__kmpc_critical: {
1110 // Build void __kmpc_critical(ident_t *loc, kmp_int32 global_tid,
1111 // kmp_critical_name *crit);
1112 llvm::Type *TypeParams[] = {
1113 getIdentTyPointerTy(), CGM.Int32Ty,
1114 llvm::PointerType::getUnqual(KmpCriticalNameTy)};
1115 llvm::FunctionType *FnTy =
1116 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1117 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_critical");
1120 case OMPRTL__kmpc_critical_with_hint: {
1121 // Build void __kmpc_critical_with_hint(ident_t *loc, kmp_int32 global_tid,
1122 // kmp_critical_name *crit, uintptr_t hint);
1123 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1124 llvm::PointerType::getUnqual(KmpCriticalNameTy),
1126 llvm::FunctionType *FnTy =
1127 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1128 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_critical_with_hint");
1131 case OMPRTL__kmpc_threadprivate_register: {
1132 // Build void __kmpc_threadprivate_register(ident_t *, void *data,
1133 // kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor);
1134 // typedef void *(*kmpc_ctor)(void *);
1136 llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy,
1137 /*isVarArg*/ false)->getPointerTo();
1138 // typedef void *(*kmpc_cctor)(void *, void *);
1139 llvm::Type *KmpcCopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy};
1140 auto KmpcCopyCtorTy =
1141 llvm::FunctionType::get(CGM.VoidPtrTy, KmpcCopyCtorTyArgs,
1142 /*isVarArg*/ false)->getPointerTo();
1143 // typedef void (*kmpc_dtor)(void *);
1145 llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, /*isVarArg*/ false)
1147 llvm::Type *FnTyArgs[] = {getIdentTyPointerTy(), CGM.VoidPtrTy, KmpcCtorTy,
1148 KmpcCopyCtorTy, KmpcDtorTy};
1149 auto FnTy = llvm::FunctionType::get(CGM.VoidTy, FnTyArgs,
1150 /*isVarArg*/ false);
1151 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_threadprivate_register");
1154 case OMPRTL__kmpc_end_critical: {
1155 // Build void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid,
1156 // kmp_critical_name *crit);
1157 llvm::Type *TypeParams[] = {
1158 getIdentTyPointerTy(), CGM.Int32Ty,
1159 llvm::PointerType::getUnqual(KmpCriticalNameTy)};
1160 llvm::FunctionType *FnTy =
1161 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1162 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_critical");
1165 case OMPRTL__kmpc_cancel_barrier: {
1166 // Build kmp_int32 __kmpc_cancel_barrier(ident_t *loc, kmp_int32
1168 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1169 llvm::FunctionType *FnTy =
1170 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1171 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name*/ "__kmpc_cancel_barrier");
1174 case OMPRTL__kmpc_barrier: {
1175 // Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
1176 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1177 llvm::FunctionType *FnTy =
1178 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1179 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier");
1182 case OMPRTL__kmpc_for_static_fini: {
1183 // Build void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
1184 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1185 llvm::FunctionType *FnTy =
1186 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1187 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_for_static_fini");
1190 case OMPRTL__kmpc_push_num_threads: {
1191 // Build void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
1192 // kmp_int32 num_threads)
1193 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1195 llvm::FunctionType *FnTy =
1196 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1197 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_num_threads");
1200 case OMPRTL__kmpc_serialized_parallel: {
1201 // Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32
1203 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1204 llvm::FunctionType *FnTy =
1205 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1206 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel");
1209 case OMPRTL__kmpc_end_serialized_parallel: {
1210 // Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32
1212 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1213 llvm::FunctionType *FnTy =
1214 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1215 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel");
1218 case OMPRTL__kmpc_flush: {
1219 // Build void __kmpc_flush(ident_t *loc);
1220 llvm::Type *TypeParams[] = {getIdentTyPointerTy()};
1221 llvm::FunctionType *FnTy =
1222 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1223 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_flush");
1226 case OMPRTL__kmpc_master: {
1227 // Build kmp_int32 __kmpc_master(ident_t *loc, kmp_int32 global_tid);
1228 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1229 llvm::FunctionType *FnTy =
1230 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1231 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_master");
1234 case OMPRTL__kmpc_end_master: {
1235 // Build void __kmpc_end_master(ident_t *loc, kmp_int32 global_tid);
1236 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1237 llvm::FunctionType *FnTy =
1238 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1239 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_master");
1242 case OMPRTL__kmpc_omp_taskyield: {
1243 // Build kmp_int32 __kmpc_omp_taskyield(ident_t *, kmp_int32 global_tid,
1245 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy};
1246 llvm::FunctionType *FnTy =
1247 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1248 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_taskyield");
1251 case OMPRTL__kmpc_single: {
1252 // Build kmp_int32 __kmpc_single(ident_t *loc, kmp_int32 global_tid);
1253 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1254 llvm::FunctionType *FnTy =
1255 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1256 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_single");
1259 case OMPRTL__kmpc_end_single: {
1260 // Build void __kmpc_end_single(ident_t *loc, kmp_int32 global_tid);
1261 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1262 llvm::FunctionType *FnTy =
1263 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1264 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_single");
1267 case OMPRTL__kmpc_omp_task_alloc: {
1268 // Build kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid,
1269 // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
1270 // kmp_routine_entry_t *task_entry);
1271 assert(KmpRoutineEntryPtrTy != nullptr &&
1272 "Type kmp_routine_entry_t must be created.");
1273 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty,
1274 CGM.SizeTy, CGM.SizeTy, KmpRoutineEntryPtrTy};
1275 // Return void * and then cast to particular kmp_task_t type.
1276 llvm::FunctionType *FnTy =
1277 llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false);
1278 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task_alloc");
1281 case OMPRTL__kmpc_omp_task: {
1282 // Build kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t
1284 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1286 llvm::FunctionType *FnTy =
1287 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1288 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task");
1291 case OMPRTL__kmpc_copyprivate: {
1292 // Build void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
1293 // size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *),
1294 // kmp_int32 didit);
1295 llvm::Type *CpyTypeParams[] = {CGM.VoidPtrTy, CGM.VoidPtrTy};
1297 llvm::FunctionType::get(CGM.VoidTy, CpyTypeParams, /*isVarArg=*/false);
1298 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.SizeTy,
1299 CGM.VoidPtrTy, CpyFnTy->getPointerTo(),
1301 llvm::FunctionType *FnTy =
1302 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1303 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_copyprivate");
1306 case OMPRTL__kmpc_reduce: {
1307 // Build kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid,
1308 // kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void
1309 // (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck);
1310 llvm::Type *ReduceTypeParams[] = {CGM.VoidPtrTy, CGM.VoidPtrTy};
1311 auto *ReduceFnTy = llvm::FunctionType::get(CGM.VoidTy, ReduceTypeParams,
1312 /*isVarArg=*/false);
1313 llvm::Type *TypeParams[] = {
1314 getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty, CGM.SizeTy,
1315 CGM.VoidPtrTy, ReduceFnTy->getPointerTo(),
1316 llvm::PointerType::getUnqual(KmpCriticalNameTy)};
1317 llvm::FunctionType *FnTy =
1318 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1319 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_reduce");
1322 case OMPRTL__kmpc_reduce_nowait: {
1323 // Build kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32
1324 // global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data,
1325 // void (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name
1327 llvm::Type *ReduceTypeParams[] = {CGM.VoidPtrTy, CGM.VoidPtrTy};
1328 auto *ReduceFnTy = llvm::FunctionType::get(CGM.VoidTy, ReduceTypeParams,
1329 /*isVarArg=*/false);
1330 llvm::Type *TypeParams[] = {
1331 getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty, CGM.SizeTy,
1332 CGM.VoidPtrTy, ReduceFnTy->getPointerTo(),
1333 llvm::PointerType::getUnqual(KmpCriticalNameTy)};
1334 llvm::FunctionType *FnTy =
1335 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1336 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_reduce_nowait");
1339 case OMPRTL__kmpc_end_reduce: {
1340 // Build void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
1341 // kmp_critical_name *lck);
1342 llvm::Type *TypeParams[] = {
1343 getIdentTyPointerTy(), CGM.Int32Ty,
1344 llvm::PointerType::getUnqual(KmpCriticalNameTy)};
1345 llvm::FunctionType *FnTy =
1346 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1347 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_reduce");
1350 case OMPRTL__kmpc_end_reduce_nowait: {
1351 // Build __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
1352 // kmp_critical_name *lck);
1353 llvm::Type *TypeParams[] = {
1354 getIdentTyPointerTy(), CGM.Int32Ty,
1355 llvm::PointerType::getUnqual(KmpCriticalNameTy)};
1356 llvm::FunctionType *FnTy =
1357 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1359 CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_reduce_nowait");
1362 case OMPRTL__kmpc_omp_task_begin_if0: {
1363 // Build void __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t
1365 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1367 llvm::FunctionType *FnTy =
1368 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1370 CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task_begin_if0");
1373 case OMPRTL__kmpc_omp_task_complete_if0: {
1374 // Build void __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t
1376 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1378 llvm::FunctionType *FnTy =
1379 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1380 RTLFn = CGM.CreateRuntimeFunction(FnTy,
1381 /*Name=*/"__kmpc_omp_task_complete_if0");
1384 case OMPRTL__kmpc_ordered: {
1385 // Build void __kmpc_ordered(ident_t *loc, kmp_int32 global_tid);
1386 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1387 llvm::FunctionType *FnTy =
1388 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1389 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_ordered");
1392 case OMPRTL__kmpc_end_ordered: {
1393 // Build void __kmpc_end_ordered(ident_t *loc, kmp_int32 global_tid);
1394 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1395 llvm::FunctionType *FnTy =
1396 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1397 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_ordered");
1400 case OMPRTL__kmpc_omp_taskwait: {
1401 // Build kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 global_tid);
1402 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1403 llvm::FunctionType *FnTy =
1404 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1405 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_omp_taskwait");
1408 case OMPRTL__kmpc_taskgroup: {
1409 // Build void __kmpc_taskgroup(ident_t *loc, kmp_int32 global_tid);
1410 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1411 llvm::FunctionType *FnTy =
1412 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1413 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_taskgroup");
1416 case OMPRTL__kmpc_end_taskgroup: {
1417 // Build void __kmpc_end_taskgroup(ident_t *loc, kmp_int32 global_tid);
1418 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1419 llvm::FunctionType *FnTy =
1420 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1421 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_taskgroup");
1424 case OMPRTL__kmpc_push_proc_bind: {
1425 // Build void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
1427 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy};
1428 llvm::FunctionType *FnTy =
1429 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1430 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_proc_bind");
1433 case OMPRTL__kmpc_omp_task_with_deps: {
1434 // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid,
1435 // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,
1436 // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list);
1437 llvm::Type *TypeParams[] = {
1438 getIdentTyPointerTy(), CGM.Int32Ty, CGM.VoidPtrTy, CGM.Int32Ty,
1439 CGM.VoidPtrTy, CGM.Int32Ty, CGM.VoidPtrTy};
1440 llvm::FunctionType *FnTy =
1441 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1443 CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task_with_deps");
1446 case OMPRTL__kmpc_omp_wait_deps: {
1447 // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid,
1448 // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
1449 // kmp_depend_info_t *noalias_dep_list);
1450 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1451 CGM.Int32Ty, CGM.VoidPtrTy,
1452 CGM.Int32Ty, CGM.VoidPtrTy};
1453 llvm::FunctionType *FnTy =
1454 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1455 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_wait_deps");
1458 case OMPRTL__kmpc_cancellationpoint: {
1459 // Build kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32
1460 // global_tid, kmp_int32 cncl_kind)
1461 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy};
1462 llvm::FunctionType *FnTy =
1463 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1464 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_cancellationpoint");
1467 case OMPRTL__kmpc_cancel: {
1468 // Build kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid,
1469 // kmp_int32 cncl_kind)
1470 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy};
1471 llvm::FunctionType *FnTy =
1472 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1473 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_cancel");
1476 case OMPRTL__kmpc_push_num_teams: {
1477 // Build void kmpc_push_num_teams (ident_t loc, kmp_int32 global_tid,
1478 // kmp_int32 num_teams, kmp_int32 num_threads)
1479 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty,
1481 llvm::FunctionType *FnTy =
1482 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1483 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_num_teams");
1486 case OMPRTL__kmpc_fork_teams: {
1487 // Build void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro
1489 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1490 getKmpc_MicroPointerTy()};
1491 llvm::FunctionType *FnTy =
1492 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true);
1493 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_fork_teams");
1496 case OMPRTL__kmpc_taskloop: {
1497 // Build void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int
1498 // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int
1499 // sched, kmp_uint64 grainsize, void *task_dup);
1500 llvm::Type *TypeParams[] = {getIdentTyPointerTy(),
1504 CGM.Int64Ty->getPointerTo(),
1505 CGM.Int64Ty->getPointerTo(),
1511 llvm::FunctionType *FnTy =
1512 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1513 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_taskloop");
1516 case OMPRTL__kmpc_doacross_init: {
1517 // Build void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, kmp_int32
1518 // num_dims, struct kmp_dim *dims);
1519 llvm::Type *TypeParams[] = {getIdentTyPointerTy(),
1523 llvm::FunctionType *FnTy =
1524 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1525 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_init");
1528 case OMPRTL__kmpc_doacross_fini: {
1529 // Build void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
1530 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1531 llvm::FunctionType *FnTy =
1532 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1533 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_fini");
1536 case OMPRTL__kmpc_doacross_post: {
1537 // Build void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, kmp_int64
1539 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1540 CGM.Int64Ty->getPointerTo()};
1541 llvm::FunctionType *FnTy =
1542 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1543 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_post");
1546 case OMPRTL__kmpc_doacross_wait: {
1547 // Build void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, kmp_int64
1549 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
1550 CGM.Int64Ty->getPointerTo()};
1551 llvm::FunctionType *FnTy =
1552 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1553 RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_wait");
1556 case OMPRTL__tgt_target: {
1557 // Build int32_t __tgt_target(int32_t device_id, void *host_ptr, int32_t
1558 // arg_num, void** args_base, void **args, size_t *arg_sizes, int32_t
1560 llvm::Type *TypeParams[] = {CGM.Int32Ty,
1565 CGM.SizeTy->getPointerTo(),
1566 CGM.Int32Ty->getPointerTo()};
1567 llvm::FunctionType *FnTy =
1568 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1569 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target");
1572 case OMPRTL__tgt_target_teams: {
1573 // Build int32_t __tgt_target_teams(int32_t device_id, void *host_ptr,
1574 // int32_t arg_num, void** args_base, void **args, size_t *arg_sizes,
1575 // int32_t *arg_types, int32_t num_teams, int32_t thread_limit);
1576 llvm::Type *TypeParams[] = {CGM.Int32Ty,
1581 CGM.SizeTy->getPointerTo(),
1582 CGM.Int32Ty->getPointerTo(),
1585 llvm::FunctionType *FnTy =
1586 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1587 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_teams");
1590 case OMPRTL__tgt_register_lib: {
1591 // Build void __tgt_register_lib(__tgt_bin_desc *desc);
1593 CGM.getContext().getPointerType(getTgtBinaryDescriptorQTy());
1594 llvm::Type *TypeParams[] = {CGM.getTypes().ConvertTypeForMem(ParamTy)};
1595 llvm::FunctionType *FnTy =
1596 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1597 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_register_lib");
1600 case OMPRTL__tgt_unregister_lib: {
1601 // Build void __tgt_unregister_lib(__tgt_bin_desc *desc);
1603 CGM.getContext().getPointerType(getTgtBinaryDescriptorQTy());
1604 llvm::Type *TypeParams[] = {CGM.getTypes().ConvertTypeForMem(ParamTy)};
1605 llvm::FunctionType *FnTy =
1606 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1607 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_unregister_lib");
1610 case OMPRTL__tgt_target_data_begin: {
1611 // Build void __tgt_target_data_begin(int32_t device_id, int32_t arg_num,
1612 // void** args_base, void **args, size_t *arg_sizes, int32_t *arg_types);
1613 llvm::Type *TypeParams[] = {CGM.Int32Ty,
1617 CGM.SizeTy->getPointerTo(),
1618 CGM.Int32Ty->getPointerTo()};
1619 llvm::FunctionType *FnTy =
1620 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1621 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_begin");
1624 case OMPRTL__tgt_target_data_end: {
1625 // Build void __tgt_target_data_end(int32_t device_id, int32_t arg_num,
1626 // void** args_base, void **args, size_t *arg_sizes, int32_t *arg_types);
1627 llvm::Type *TypeParams[] = {CGM.Int32Ty,
1631 CGM.SizeTy->getPointerTo(),
1632 CGM.Int32Ty->getPointerTo()};
1633 llvm::FunctionType *FnTy =
1634 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1635 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_end");
1638 case OMPRTL__tgt_target_data_update: {
1639 // Build void __tgt_target_data_update(int32_t device_id, int32_t arg_num,
1640 // void** args_base, void **args, size_t *arg_sizes, int32_t *arg_types);
1641 llvm::Type *TypeParams[] = {CGM.Int32Ty,
1645 CGM.SizeTy->getPointerTo(),
1646 CGM.Int32Ty->getPointerTo()};
1647 llvm::FunctionType *FnTy =
1648 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1649 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_update");
1653 assert(RTLFn && "Unable to find OpenMP runtime function");
1657 llvm::Constant *CGOpenMPRuntime::createForStaticInitFunction(unsigned IVSize,
1659 assert((IVSize == 32 || IVSize == 64) &&
1660 "IV size is not compatible with the omp runtime");
1661 auto Name = IVSize == 32 ? (IVSigned ? "__kmpc_for_static_init_4"
1662 : "__kmpc_for_static_init_4u")
1663 : (IVSigned ? "__kmpc_for_static_init_8"
1664 : "__kmpc_for_static_init_8u");
1665 auto ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty;
1666 auto PtrTy = llvm::PointerType::getUnqual(ITy);
1667 llvm::Type *TypeParams[] = {
1668 getIdentTyPointerTy(), // loc
1670 CGM.Int32Ty, // schedtype
1671 llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter
1678 llvm::FunctionType *FnTy =
1679 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1680 return CGM.CreateRuntimeFunction(FnTy, Name);
1683 llvm::Constant *CGOpenMPRuntime::createDispatchInitFunction(unsigned IVSize,
1685 assert((IVSize == 32 || IVSize == 64) &&
1686 "IV size is not compatible with the omp runtime");
1689 ? (IVSigned ? "__kmpc_dispatch_init_4" : "__kmpc_dispatch_init_4u")
1690 : (IVSigned ? "__kmpc_dispatch_init_8" : "__kmpc_dispatch_init_8u");
1691 auto ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty;
1692 llvm::Type *TypeParams[] = { getIdentTyPointerTy(), // loc
1694 CGM.Int32Ty, // schedtype
1700 llvm::FunctionType *FnTy =
1701 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1702 return CGM.CreateRuntimeFunction(FnTy, Name);
1705 llvm::Constant *CGOpenMPRuntime::createDispatchFiniFunction(unsigned IVSize,
1707 assert((IVSize == 32 || IVSize == 64) &&
1708 "IV size is not compatible with the omp runtime");
1711 ? (IVSigned ? "__kmpc_dispatch_fini_4" : "__kmpc_dispatch_fini_4u")
1712 : (IVSigned ? "__kmpc_dispatch_fini_8" : "__kmpc_dispatch_fini_8u");
1713 llvm::Type *TypeParams[] = {
1714 getIdentTyPointerTy(), // loc
1717 llvm::FunctionType *FnTy =
1718 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1719 return CGM.CreateRuntimeFunction(FnTy, Name);
1722 llvm::Constant *CGOpenMPRuntime::createDispatchNextFunction(unsigned IVSize,
1724 assert((IVSize == 32 || IVSize == 64) &&
1725 "IV size is not compatible with the omp runtime");
1728 ? (IVSigned ? "__kmpc_dispatch_next_4" : "__kmpc_dispatch_next_4u")
1729 : (IVSigned ? "__kmpc_dispatch_next_8" : "__kmpc_dispatch_next_8u");
1730 auto ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty;
1731 auto PtrTy = llvm::PointerType::getUnqual(ITy);
1732 llvm::Type *TypeParams[] = {
1733 getIdentTyPointerTy(), // loc
1735 llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter
1740 llvm::FunctionType *FnTy =
1741 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1742 return CGM.CreateRuntimeFunction(FnTy, Name);
1746 CGOpenMPRuntime::getOrCreateThreadPrivateCache(const VarDecl *VD) {
1747 assert(!CGM.getLangOpts().OpenMPUseTLS ||
1748 !CGM.getContext().getTargetInfo().isTLSSupported());
1749 // Lookup the entry, lazily creating it if necessary.
1750 return getOrCreateInternalVariable(CGM.Int8PtrPtrTy,
1751 Twine(CGM.getMangledName(VD)) + ".cache.");
1754 Address CGOpenMPRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF,
1757 SourceLocation Loc) {
1758 if (CGM.getLangOpts().OpenMPUseTLS &&
1759 CGM.getContext().getTargetInfo().isTLSSupported())
1762 auto VarTy = VDAddr.getElementType();
1763 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
1764 CGF.Builder.CreatePointerCast(VDAddr.getPointer(),
1766 CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)),
1767 getOrCreateThreadPrivateCache(VD)};
1768 return Address(CGF.EmitRuntimeCall(
1769 createRuntimeFunction(OMPRTL__kmpc_threadprivate_cached), Args),
1770 VDAddr.getAlignment());
1773 void CGOpenMPRuntime::emitThreadPrivateVarInit(
1774 CodeGenFunction &CGF, Address VDAddr, llvm::Value *Ctor,
1775 llvm::Value *CopyCtor, llvm::Value *Dtor, SourceLocation Loc) {
1776 // Call kmp_int32 __kmpc_global_thread_num(&loc) to init OpenMP runtime
1778 auto OMPLoc = emitUpdateLocation(CGF, Loc);
1779 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_global_thread_num),
1781 // Call __kmpc_threadprivate_register(&loc, &var, ctor, cctor/*NULL*/, dtor)
1782 // to register constructor/destructor for variable.
1783 llvm::Value *Args[] = {OMPLoc,
1784 CGF.Builder.CreatePointerCast(VDAddr.getPointer(),
1786 Ctor, CopyCtor, Dtor};
1787 CGF.EmitRuntimeCall(
1788 createRuntimeFunction(OMPRTL__kmpc_threadprivate_register), Args);
1791 llvm::Function *CGOpenMPRuntime::emitThreadPrivateVarDefinition(
1792 const VarDecl *VD, Address VDAddr, SourceLocation Loc,
1793 bool PerformInit, CodeGenFunction *CGF) {
1794 if (CGM.getLangOpts().OpenMPUseTLS &&
1795 CGM.getContext().getTargetInfo().isTLSSupported())
1798 VD = VD->getDefinition(CGM.getContext());
1799 if (VD && ThreadPrivateWithDefinition.count(VD) == 0) {
1800 ThreadPrivateWithDefinition.insert(VD);
1801 QualType ASTTy = VD->getType();
1803 llvm::Value *Ctor = nullptr, *CopyCtor = nullptr, *Dtor = nullptr;
1804 auto Init = VD->getAnyInitializer();
1805 if (CGM.getLangOpts().CPlusPlus && PerformInit) {
1806 // Generate function that re-emits the declaration's initializer into the
1807 // threadprivate copy of the variable VD
1808 CodeGenFunction CtorCGF(CGM);
1809 FunctionArgList Args;
1810 ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, SourceLocation(),
1811 /*Id=*/nullptr, CGM.getContext().VoidPtrTy);
1812 Args.push_back(&Dst);
1814 auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(
1815 CGM.getContext().VoidPtrTy, Args);
1816 auto FTy = CGM.getTypes().GetFunctionType(FI);
1817 auto Fn = CGM.CreateGlobalInitOrDestructFunction(
1818 FTy, ".__kmpc_global_ctor_.", FI, Loc);
1819 CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidPtrTy, Fn, FI,
1820 Args, SourceLocation());
1821 auto ArgVal = CtorCGF.EmitLoadOfScalar(
1822 CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false,
1823 CGM.getContext().VoidPtrTy, Dst.getLocation());
1824 Address Arg = Address(ArgVal, VDAddr.getAlignment());
1825 Arg = CtorCGF.Builder.CreateElementBitCast(Arg,
1826 CtorCGF.ConvertTypeForMem(ASTTy));
1827 CtorCGF.EmitAnyExprToMem(Init, Arg, Init->getType().getQualifiers(),
1828 /*IsInitializer=*/true);
1829 ArgVal = CtorCGF.EmitLoadOfScalar(
1830 CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false,
1831 CGM.getContext().VoidPtrTy, Dst.getLocation());
1832 CtorCGF.Builder.CreateStore(ArgVal, CtorCGF.ReturnValue);
1833 CtorCGF.FinishFunction();
1836 if (VD->getType().isDestructedType() != QualType::DK_none) {
1837 // Generate function that emits destructor call for the threadprivate copy
1838 // of the variable VD
1839 CodeGenFunction DtorCGF(CGM);
1840 FunctionArgList Args;
1841 ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, SourceLocation(),
1842 /*Id=*/nullptr, CGM.getContext().VoidPtrTy);
1843 Args.push_back(&Dst);
1845 auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(
1846 CGM.getContext().VoidTy, Args);
1847 auto FTy = CGM.getTypes().GetFunctionType(FI);
1848 auto Fn = CGM.CreateGlobalInitOrDestructFunction(
1849 FTy, ".__kmpc_global_dtor_.", FI, Loc);
1850 auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF);
1851 DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, Args,
1853 // Create a scope with an artificial location for the body of this function.
1854 auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF);
1855 auto ArgVal = DtorCGF.EmitLoadOfScalar(
1856 DtorCGF.GetAddrOfLocalVar(&Dst),
1857 /*Volatile=*/false, CGM.getContext().VoidPtrTy, Dst.getLocation());
1858 DtorCGF.emitDestroy(Address(ArgVal, VDAddr.getAlignment()), ASTTy,
1859 DtorCGF.getDestroyer(ASTTy.isDestructedType()),
1860 DtorCGF.needsEHCleanup(ASTTy.isDestructedType()));
1861 DtorCGF.FinishFunction();
1864 // Do not emit init function if it is not required.
1868 llvm::Type *CopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy};
1870 llvm::FunctionType::get(CGM.VoidPtrTy, CopyCtorTyArgs,
1871 /*isVarArg=*/false)->getPointerTo();
1872 // Copying constructor for the threadprivate variable.
1873 // Must be NULL - reserved by runtime, but currently it requires that this
1874 // parameter is always NULL. Otherwise it fires assertion.
1875 CopyCtor = llvm::Constant::getNullValue(CopyCtorTy);
1876 if (Ctor == nullptr) {
1877 auto CtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy,
1878 /*isVarArg=*/false)->getPointerTo();
1879 Ctor = llvm::Constant::getNullValue(CtorTy);
1881 if (Dtor == nullptr) {
1882 auto DtorTy = llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy,
1883 /*isVarArg=*/false)->getPointerTo();
1884 Dtor = llvm::Constant::getNullValue(DtorTy);
1887 auto InitFunctionTy =
1888 llvm::FunctionType::get(CGM.VoidTy, /*isVarArg*/ false);
1889 auto InitFunction = CGM.CreateGlobalInitOrDestructFunction(
1890 InitFunctionTy, ".__omp_threadprivate_init_.",
1891 CGM.getTypes().arrangeNullaryFunction());
1892 CodeGenFunction InitCGF(CGM);
1893 FunctionArgList ArgList;
1894 InitCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, InitFunction,
1895 CGM.getTypes().arrangeNullaryFunction(), ArgList,
1897 emitThreadPrivateVarInit(InitCGF, VDAddr, Ctor, CopyCtor, Dtor, Loc);
1898 InitCGF.FinishFunction();
1899 return InitFunction;
1901 emitThreadPrivateVarInit(*CGF, VDAddr, Ctor, CopyCtor, Dtor, Loc);
1906 /// \brief Emits code for OpenMP 'if' clause using specified \a CodeGen
1907 /// function. Here is the logic:
1913 void CGOpenMPRuntime::emitOMPIfClause(CodeGenFunction &CGF, const Expr *Cond,
1914 const RegionCodeGenTy &ThenGen,
1915 const RegionCodeGenTy &ElseGen) {
1916 CodeGenFunction::LexicalScope ConditionScope(CGF, Cond->getSourceRange());
1918 // If the condition constant folds and can be elided, try to avoid emitting
1919 // the condition and the dead arm of the if/else.
1921 if (CGF.ConstantFoldsToSimpleInteger(Cond, CondConstant)) {
1929 // Otherwise, the condition did not fold, or we couldn't elide it. Just
1930 // emit the conditional branch.
1931 auto ThenBlock = CGF.createBasicBlock("omp_if.then");
1932 auto ElseBlock = CGF.createBasicBlock("omp_if.else");
1933 auto ContBlock = CGF.createBasicBlock("omp_if.end");
1934 CGF.EmitBranchOnBoolExpr(Cond, ThenBlock, ElseBlock, /*TrueCount=*/0);
1936 // Emit the 'then' code.
1937 CGF.EmitBlock(ThenBlock);
1939 CGF.EmitBranch(ContBlock);
1940 // Emit the 'else' code if present.
1941 // There is no need to emit line number for unconditional branch.
1942 (void)ApplyDebugLocation::CreateEmpty(CGF);
1943 CGF.EmitBlock(ElseBlock);
1945 // There is no need to emit line number for unconditional branch.
1946 (void)ApplyDebugLocation::CreateEmpty(CGF);
1947 CGF.EmitBranch(ContBlock);
1948 // Emit the continuation block for code after the if.
1949 CGF.EmitBlock(ContBlock, /*IsFinished=*/true);
1952 void CGOpenMPRuntime::emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc,
1953 llvm::Value *OutlinedFn,
1954 ArrayRef<llvm::Value *> CapturedVars,
1955 const Expr *IfCond) {
1956 if (!CGF.HaveInsertPoint())
1958 auto *RTLoc = emitUpdateLocation(CGF, Loc);
1959 auto &&ThenGen = [OutlinedFn, CapturedVars, RTLoc](CodeGenFunction &CGF,
1960 PrePostActionTy &) {
1961 // Build call __kmpc_fork_call(loc, n, microtask, var1, .., varn);
1962 auto &RT = CGF.CGM.getOpenMPRuntime();
1963 llvm::Value *Args[] = {
1965 CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars
1966 CGF.Builder.CreateBitCast(OutlinedFn, RT.getKmpc_MicroPointerTy())};
1967 llvm::SmallVector<llvm::Value *, 16> RealArgs;
1968 RealArgs.append(std::begin(Args), std::end(Args));
1969 RealArgs.append(CapturedVars.begin(), CapturedVars.end());
1971 auto RTLFn = RT.createRuntimeFunction(OMPRTL__kmpc_fork_call);
1972 CGF.EmitRuntimeCall(RTLFn, RealArgs);
1974 auto &&ElseGen = [OutlinedFn, CapturedVars, RTLoc, Loc](CodeGenFunction &CGF,
1975 PrePostActionTy &) {
1976 auto &RT = CGF.CGM.getOpenMPRuntime();
1977 auto ThreadID = RT.getThreadID(CGF, Loc);
1979 // __kmpc_serialized_parallel(&Loc, GTid);
1980 llvm::Value *Args[] = {RTLoc, ThreadID};
1981 CGF.EmitRuntimeCall(
1982 RT.createRuntimeFunction(OMPRTL__kmpc_serialized_parallel), Args);
1984 // OutlinedFn(>id, &zero, CapturedStruct);
1985 auto ThreadIDAddr = RT.emitThreadIDAddress(CGF, Loc);
1987 CGF.CreateTempAlloca(CGF.Int32Ty, CharUnits::fromQuantity(4),
1988 /*Name*/ ".zero.addr");
1989 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
1990 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
1991 OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
1992 OutlinedFnArgs.push_back(ZeroAddr.getPointer());
1993 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
1994 CGF.EmitCallOrInvoke(OutlinedFn, OutlinedFnArgs);
1996 // __kmpc_end_serialized_parallel(&Loc, GTid);
1997 llvm::Value *EndArgs[] = {RT.emitUpdateLocation(CGF, Loc), ThreadID};
1998 CGF.EmitRuntimeCall(
1999 RT.createRuntimeFunction(OMPRTL__kmpc_end_serialized_parallel),
2003 emitOMPIfClause(CGF, IfCond, ThenGen, ElseGen);
2005 RegionCodeGenTy ThenRCG(ThenGen);
2010 // If we're inside an (outlined) parallel region, use the region info's
2011 // thread-ID variable (it is passed in a first argument of the outlined function
2012 // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in
2013 // regular serial code region, get thread ID by calling kmp_int32
2014 // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and
2015 // return the address of that temp.
2016 Address CGOpenMPRuntime::emitThreadIDAddress(CodeGenFunction &CGF,
2017 SourceLocation Loc) {
2018 if (auto *OMPRegionInfo =
2019 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
2020 if (OMPRegionInfo->getThreadIDVariable())
2021 return OMPRegionInfo->getThreadIDVariableLValue(CGF).getAddress();
2023 auto ThreadID = getThreadID(CGF, Loc);
2025 CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true);
2026 auto ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp.");
2027 CGF.EmitStoreOfScalar(ThreadID,
2028 CGF.MakeAddrLValue(ThreadIDTemp, Int32Ty));
2030 return ThreadIDTemp;
2034 CGOpenMPRuntime::getOrCreateInternalVariable(llvm::Type *Ty,
2035 const llvm::Twine &Name) {
2036 SmallString<256> Buffer;
2037 llvm::raw_svector_ostream Out(Buffer);
2039 auto RuntimeName = Out.str();
2040 auto &Elem = *InternalVars.insert(std::make_pair(RuntimeName, nullptr)).first;
2042 assert(Elem.second->getType()->getPointerElementType() == Ty &&
2043 "OMP internal variable has different type than requested");
2044 return &*Elem.second;
2047 return Elem.second = new llvm::GlobalVariable(
2048 CGM.getModule(), Ty, /*IsConstant*/ false,
2049 llvm::GlobalValue::CommonLinkage, llvm::Constant::getNullValue(Ty),
2053 llvm::Value *CGOpenMPRuntime::getCriticalRegionLock(StringRef CriticalName) {
2054 llvm::Twine Name(".gomp_critical_user_", CriticalName);
2055 return getOrCreateInternalVariable(KmpCriticalNameTy, Name.concat(".var"));
2059 /// Common pre(post)-action for different OpenMP constructs.
2060 class CommonActionTy final : public PrePostActionTy {
2061 llvm::Value *EnterCallee;
2062 ArrayRef<llvm::Value *> EnterArgs;
2063 llvm::Value *ExitCallee;
2064 ArrayRef<llvm::Value *> ExitArgs;
2066 llvm::BasicBlock *ContBlock = nullptr;
2069 CommonActionTy(llvm::Value *EnterCallee, ArrayRef<llvm::Value *> EnterArgs,
2070 llvm::Value *ExitCallee, ArrayRef<llvm::Value *> ExitArgs,
2071 bool Conditional = false)
2072 : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
2073 ExitArgs(ExitArgs), Conditional(Conditional) {}
2074 void Enter(CodeGenFunction &CGF) override {
2075 llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
2077 llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
2078 auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
2079 ContBlock = CGF.createBasicBlock("omp_if.end");
2080 // Generate the branch (If-stmt)
2081 CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
2082 CGF.EmitBlock(ThenBlock);
2085 void Done(CodeGenFunction &CGF) {
2086 // Emit the rest of blocks/branches
2087 CGF.EmitBranch(ContBlock);
2088 CGF.EmitBlock(ContBlock, true);
2090 void Exit(CodeGenFunction &CGF) override {
2091 CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
2094 } // anonymous namespace
2096 void CGOpenMPRuntime::emitCriticalRegion(CodeGenFunction &CGF,
2097 StringRef CriticalName,
2098 const RegionCodeGenTy &CriticalOpGen,
2099 SourceLocation Loc, const Expr *Hint) {
2100 // __kmpc_critical[_with_hint](ident_t *, gtid, Lock[, hint]);
2102 // __kmpc_end_critical(ident_t *, gtid, Lock);
2103 // Prepare arguments and build a call to __kmpc_critical
2104 if (!CGF.HaveInsertPoint())
2106 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2107 getCriticalRegionLock(CriticalName)};
2108 llvm::SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args),
2111 EnterArgs.push_back(CGF.Builder.CreateIntCast(
2112 CGF.EmitScalarExpr(Hint), CGM.IntPtrTy, /*isSigned=*/false));
2114 CommonActionTy Action(
2115 createRuntimeFunction(Hint ? OMPRTL__kmpc_critical_with_hint
2116 : OMPRTL__kmpc_critical),
2117 EnterArgs, createRuntimeFunction(OMPRTL__kmpc_end_critical), Args);
2118 CriticalOpGen.setAction(Action);
2119 emitInlinedDirective(CGF, OMPD_critical, CriticalOpGen);
2122 void CGOpenMPRuntime::emitMasterRegion(CodeGenFunction &CGF,
2123 const RegionCodeGenTy &MasterOpGen,
2124 SourceLocation Loc) {
2125 if (!CGF.HaveInsertPoint())
2127 // if(__kmpc_master(ident_t *, gtid)) {
2129 // __kmpc_end_master(ident_t *, gtid);
2131 // Prepare arguments and build a call to __kmpc_master
2132 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2133 CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_master), Args,
2134 createRuntimeFunction(OMPRTL__kmpc_end_master), Args,
2135 /*Conditional=*/true);
2136 MasterOpGen.setAction(Action);
2137 emitInlinedDirective(CGF, OMPD_master, MasterOpGen);
2141 void CGOpenMPRuntime::emitTaskyieldCall(CodeGenFunction &CGF,
2142 SourceLocation Loc) {
2143 if (!CGF.HaveInsertPoint())
2145 // Build call __kmpc_omp_taskyield(loc, thread_id, 0);
2146 llvm::Value *Args[] = {
2147 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2148 llvm::ConstantInt::get(CGM.IntTy, /*V=*/0, /*isSigned=*/true)};
2149 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_taskyield), Args);
2150 if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
2151 Region->emitUntiedSwitch(CGF);
2154 void CGOpenMPRuntime::emitTaskgroupRegion(CodeGenFunction &CGF,
2155 const RegionCodeGenTy &TaskgroupOpGen,
2156 SourceLocation Loc) {
2157 if (!CGF.HaveInsertPoint())
2159 // __kmpc_taskgroup(ident_t *, gtid);
2160 // TaskgroupOpGen();
2161 // __kmpc_end_taskgroup(ident_t *, gtid);
2162 // Prepare arguments and build a call to __kmpc_taskgroup
2163 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2164 CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_taskgroup), Args,
2165 createRuntimeFunction(OMPRTL__kmpc_end_taskgroup),
2167 TaskgroupOpGen.setAction(Action);
2168 emitInlinedDirective(CGF, OMPD_taskgroup, TaskgroupOpGen);
2171 /// Given an array of pointers to variables, project the address of a
2173 static Address emitAddrOfVarFromArray(CodeGenFunction &CGF, Address Array,
2174 unsigned Index, const VarDecl *Var) {
2175 // Pull out the pointer to the variable.
2177 CGF.Builder.CreateConstArrayGEP(Array, Index, CGF.getPointerSize());
2178 llvm::Value *Ptr = CGF.Builder.CreateLoad(PtrAddr);
2180 Address Addr = Address(Ptr, CGF.getContext().getDeclAlign(Var));
2181 Addr = CGF.Builder.CreateElementBitCast(
2182 Addr, CGF.ConvertTypeForMem(Var->getType()));
2186 static llvm::Value *emitCopyprivateCopyFunction(
2187 CodeGenModule &CGM, llvm::Type *ArgsType,
2188 ArrayRef<const Expr *> CopyprivateVars, ArrayRef<const Expr *> DestExprs,
2189 ArrayRef<const Expr *> SrcExprs, ArrayRef<const Expr *> AssignmentOps) {
2190 auto &C = CGM.getContext();
2191 // void copy_func(void *LHSArg, void *RHSArg);
2192 FunctionArgList Args;
2193 ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, SourceLocation(), /*Id=*/nullptr,
2195 ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, SourceLocation(), /*Id=*/nullptr,
2197 Args.push_back(&LHSArg);
2198 Args.push_back(&RHSArg);
2199 auto &CGFI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2200 auto *Fn = llvm::Function::Create(
2201 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2202 ".omp.copyprivate.copy_func", &CGM.getModule());
2203 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, Fn, CGFI);
2204 CodeGenFunction CGF(CGM);
2205 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args);
2206 // Dest = (void*[n])(LHSArg);
2207 // Src = (void*[n])(RHSArg);
2208 Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2209 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)),
2210 ArgsType), CGF.getPointerAlign());
2211 Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2212 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)),
2213 ArgsType), CGF.getPointerAlign());
2214 // *(Type0*)Dst[0] = *(Type0*)Src[0];
2215 // *(Type1*)Dst[1] = *(Type1*)Src[1];
2217 // *(Typen*)Dst[n] = *(Typen*)Src[n];
2218 for (unsigned I = 0, E = AssignmentOps.size(); I < E; ++I) {
2219 auto DestVar = cast<VarDecl>(cast<DeclRefExpr>(DestExprs[I])->getDecl());
2220 Address DestAddr = emitAddrOfVarFromArray(CGF, LHS, I, DestVar);
2222 auto SrcVar = cast<VarDecl>(cast<DeclRefExpr>(SrcExprs[I])->getDecl());
2223 Address SrcAddr = emitAddrOfVarFromArray(CGF, RHS, I, SrcVar);
2225 auto *VD = cast<DeclRefExpr>(CopyprivateVars[I])->getDecl();
2226 QualType Type = VD->getType();
2227 CGF.EmitOMPCopy(Type, DestAddr, SrcAddr, DestVar, SrcVar, AssignmentOps[I]);
2229 CGF.FinishFunction();
2233 void CGOpenMPRuntime::emitSingleRegion(CodeGenFunction &CGF,
2234 const RegionCodeGenTy &SingleOpGen,
2236 ArrayRef<const Expr *> CopyprivateVars,
2237 ArrayRef<const Expr *> SrcExprs,
2238 ArrayRef<const Expr *> DstExprs,
2239 ArrayRef<const Expr *> AssignmentOps) {
2240 if (!CGF.HaveInsertPoint())
2242 assert(CopyprivateVars.size() == SrcExprs.size() &&
2243 CopyprivateVars.size() == DstExprs.size() &&
2244 CopyprivateVars.size() == AssignmentOps.size());
2245 auto &C = CGM.getContext();
2246 // int32 did_it = 0;
2247 // if(__kmpc_single(ident_t *, gtid)) {
2249 // __kmpc_end_single(ident_t *, gtid);
2252 // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>,
2253 // <copy_func>, did_it);
2255 Address DidIt = Address::invalid();
2256 if (!CopyprivateVars.empty()) {
2257 // int32 did_it = 0;
2258 auto KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
2259 DidIt = CGF.CreateMemTemp(KmpInt32Ty, ".omp.copyprivate.did_it");
2260 CGF.Builder.CreateStore(CGF.Builder.getInt32(0), DidIt);
2262 // Prepare arguments and build a call to __kmpc_single
2263 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2264 CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_single), Args,
2265 createRuntimeFunction(OMPRTL__kmpc_end_single), Args,
2266 /*Conditional=*/true);
2267 SingleOpGen.setAction(Action);
2268 emitInlinedDirective(CGF, OMPD_single, SingleOpGen);
2269 if (DidIt.isValid()) {
2271 CGF.Builder.CreateStore(CGF.Builder.getInt32(1), DidIt);
2274 // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>,
2275 // <copy_func>, did_it);
2276 if (DidIt.isValid()) {
2277 llvm::APInt ArraySize(/*unsigned int numBits=*/32, CopyprivateVars.size());
2278 auto CopyprivateArrayTy =
2279 C.getConstantArrayType(C.VoidPtrTy, ArraySize, ArrayType::Normal,
2280 /*IndexTypeQuals=*/0);
2281 // Create a list of all private variables for copyprivate.
2282 Address CopyprivateList =
2283 CGF.CreateMemTemp(CopyprivateArrayTy, ".omp.copyprivate.cpr_list");
2284 for (unsigned I = 0, E = CopyprivateVars.size(); I < E; ++I) {
2285 Address Elem = CGF.Builder.CreateConstArrayGEP(
2286 CopyprivateList, I, CGF.getPointerSize());
2287 CGF.Builder.CreateStore(
2288 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2289 CGF.EmitLValue(CopyprivateVars[I]).getPointer(), CGF.VoidPtrTy),
2292 // Build function that copies private values from single region to all other
2293 // threads in the corresponding parallel region.
2294 auto *CpyFn = emitCopyprivateCopyFunction(
2295 CGM, CGF.ConvertTypeForMem(CopyprivateArrayTy)->getPointerTo(),
2296 CopyprivateVars, SrcExprs, DstExprs, AssignmentOps);
2297 auto *BufSize = CGF.getTypeSize(CopyprivateArrayTy);
2299 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(CopyprivateList,
2301 auto *DidItVal = CGF.Builder.CreateLoad(DidIt);
2302 llvm::Value *Args[] = {
2303 emitUpdateLocation(CGF, Loc), // ident_t *<loc>
2304 getThreadID(CGF, Loc), // i32 <gtid>
2305 BufSize, // size_t <buf_size>
2306 CL.getPointer(), // void *<copyprivate list>
2307 CpyFn, // void (*) (void *, void *) <copy_func>
2308 DidItVal // i32 did_it
2310 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_copyprivate), Args);
2314 void CGOpenMPRuntime::emitOrderedRegion(CodeGenFunction &CGF,
2315 const RegionCodeGenTy &OrderedOpGen,
2316 SourceLocation Loc, bool IsThreads) {
2317 if (!CGF.HaveInsertPoint())
2319 // __kmpc_ordered(ident_t *, gtid);
2321 // __kmpc_end_ordered(ident_t *, gtid);
2322 // Prepare arguments and build a call to __kmpc_ordered
2324 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2325 CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_ordered), Args,
2326 createRuntimeFunction(OMPRTL__kmpc_end_ordered),
2328 OrderedOpGen.setAction(Action);
2329 emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen);
2332 emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen);
2335 void CGOpenMPRuntime::emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc,
2336 OpenMPDirectiveKind Kind, bool EmitChecks,
2337 bool ForceSimpleCall) {
2338 if (!CGF.HaveInsertPoint())
2340 // Build call __kmpc_cancel_barrier(loc, thread_id);
2341 // Build call __kmpc_barrier(loc, thread_id);
2343 if (Kind == OMPD_for)
2344 Flags = OMP_IDENT_BARRIER_IMPL_FOR;
2345 else if (Kind == OMPD_sections)
2346 Flags = OMP_IDENT_BARRIER_IMPL_SECTIONS;
2347 else if (Kind == OMPD_single)
2348 Flags = OMP_IDENT_BARRIER_IMPL_SINGLE;
2349 else if (Kind == OMPD_barrier)
2350 Flags = OMP_IDENT_BARRIER_EXPL;
2352 Flags = OMP_IDENT_BARRIER_IMPL;
2353 // Build call __kmpc_cancel_barrier(loc, thread_id) or __kmpc_barrier(loc,
2355 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
2356 getThreadID(CGF, Loc)};
2357 if (auto *OMPRegionInfo =
2358 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
2359 if (!ForceSimpleCall && OMPRegionInfo->hasCancel()) {
2360 auto *Result = CGF.EmitRuntimeCall(
2361 createRuntimeFunction(OMPRTL__kmpc_cancel_barrier), Args);
2363 // if (__kmpc_cancel_barrier()) {
2364 // exit from construct;
2366 auto *ExitBB = CGF.createBasicBlock(".cancel.exit");
2367 auto *ContBB = CGF.createBasicBlock(".cancel.continue");
2368 auto *Cmp = CGF.Builder.CreateIsNotNull(Result);
2369 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
2370 CGF.EmitBlock(ExitBB);
2371 // exit from construct;
2372 auto CancelDestination =
2373 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
2374 CGF.EmitBranchThroughCleanup(CancelDestination);
2375 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
2380 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_barrier), Args);
2383 /// \brief Map the OpenMP loop schedule to the runtime enumeration.
2384 static OpenMPSchedType getRuntimeSchedule(OpenMPScheduleClauseKind ScheduleKind,
2385 bool Chunked, bool Ordered) {
2386 switch (ScheduleKind) {
2387 case OMPC_SCHEDULE_static:
2388 return Chunked ? (Ordered ? OMP_ord_static_chunked : OMP_sch_static_chunked)
2389 : (Ordered ? OMP_ord_static : OMP_sch_static);
2390 case OMPC_SCHEDULE_dynamic:
2391 return Ordered ? OMP_ord_dynamic_chunked : OMP_sch_dynamic_chunked;
2392 case OMPC_SCHEDULE_guided:
2393 return Ordered ? OMP_ord_guided_chunked : OMP_sch_guided_chunked;
2394 case OMPC_SCHEDULE_runtime:
2395 return Ordered ? OMP_ord_runtime : OMP_sch_runtime;
2396 case OMPC_SCHEDULE_auto:
2397 return Ordered ? OMP_ord_auto : OMP_sch_auto;
2398 case OMPC_SCHEDULE_unknown:
2399 assert(!Chunked && "chunk was specified but schedule kind not known");
2400 return Ordered ? OMP_ord_static : OMP_sch_static;
2402 llvm_unreachable("Unexpected runtime schedule");
2405 /// \brief Map the OpenMP distribute schedule to the runtime enumeration.
2406 static OpenMPSchedType
2407 getRuntimeSchedule(OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) {
2408 // only static is allowed for dist_schedule
2409 return Chunked ? OMP_dist_sch_static_chunked : OMP_dist_sch_static;
2412 bool CGOpenMPRuntime::isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind,
2413 bool Chunked) const {
2414 auto Schedule = getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false);
2415 return Schedule == OMP_sch_static;
2418 bool CGOpenMPRuntime::isStaticNonchunked(
2419 OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const {
2420 auto Schedule = getRuntimeSchedule(ScheduleKind, Chunked);
2421 return Schedule == OMP_dist_sch_static;
2425 bool CGOpenMPRuntime::isDynamic(OpenMPScheduleClauseKind ScheduleKind) const {
2427 getRuntimeSchedule(ScheduleKind, /*Chunked=*/false, /*Ordered=*/false);
2428 assert(Schedule != OMP_sch_static_chunked && "cannot be chunked here");
2429 return Schedule != OMP_sch_static;
2432 static int addMonoNonMonoModifier(OpenMPSchedType Schedule,
2433 OpenMPScheduleClauseModifier M1,
2434 OpenMPScheduleClauseModifier M2) {
2437 case OMPC_SCHEDULE_MODIFIER_monotonic:
2438 Modifier = OMP_sch_modifier_monotonic;
2440 case OMPC_SCHEDULE_MODIFIER_nonmonotonic:
2441 Modifier = OMP_sch_modifier_nonmonotonic;
2443 case OMPC_SCHEDULE_MODIFIER_simd:
2444 if (Schedule == OMP_sch_static_chunked)
2445 Schedule = OMP_sch_static_balanced_chunked;
2447 case OMPC_SCHEDULE_MODIFIER_last:
2448 case OMPC_SCHEDULE_MODIFIER_unknown:
2452 case OMPC_SCHEDULE_MODIFIER_monotonic:
2453 Modifier = OMP_sch_modifier_monotonic;
2455 case OMPC_SCHEDULE_MODIFIER_nonmonotonic:
2456 Modifier = OMP_sch_modifier_nonmonotonic;
2458 case OMPC_SCHEDULE_MODIFIER_simd:
2459 if (Schedule == OMP_sch_static_chunked)
2460 Schedule = OMP_sch_static_balanced_chunked;
2462 case OMPC_SCHEDULE_MODIFIER_last:
2463 case OMPC_SCHEDULE_MODIFIER_unknown:
2466 return Schedule | Modifier;
2469 void CGOpenMPRuntime::emitForDispatchInit(CodeGenFunction &CGF,
2471 const OpenMPScheduleTy &ScheduleKind,
2472 unsigned IVSize, bool IVSigned,
2473 bool Ordered, llvm::Value *UB,
2474 llvm::Value *Chunk) {
2475 if (!CGF.HaveInsertPoint())
2477 OpenMPSchedType Schedule =
2478 getRuntimeSchedule(ScheduleKind.Schedule, Chunk != nullptr, Ordered);
2480 (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked &&
2481 Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked &&
2482 Schedule != OMP_sch_static_balanced_chunked));
2483 // Call __kmpc_dispatch_init(
2484 // ident_t *loc, kmp_int32 tid, kmp_int32 schedule,
2485 // kmp_int[32|64] lower, kmp_int[32|64] upper,
2486 // kmp_int[32|64] stride, kmp_int[32|64] chunk);
2488 // If the Chunk was not specified in the clause - use default value 1.
2489 if (Chunk == nullptr)
2490 Chunk = CGF.Builder.getIntN(IVSize, 1);
2491 llvm::Value *Args[] = {
2492 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2493 CGF.Builder.getInt32(addMonoNonMonoModifier(
2494 Schedule, ScheduleKind.M1, ScheduleKind.M2)), // Schedule type
2495 CGF.Builder.getIntN(IVSize, 0), // Lower
2497 CGF.Builder.getIntN(IVSize, 1), // Stride
2500 CGF.EmitRuntimeCall(createDispatchInitFunction(IVSize, IVSigned), Args);
2503 static void emitForStaticInitCall(
2504 CodeGenFunction &CGF, llvm::Value *UpdateLocation, llvm::Value *ThreadId,
2505 llvm::Constant *ForStaticInitFunction, OpenMPSchedType Schedule,
2506 OpenMPScheduleClauseModifier M1, OpenMPScheduleClauseModifier M2,
2507 unsigned IVSize, bool Ordered, Address IL, Address LB, Address UB,
2508 Address ST, llvm::Value *Chunk) {
2509 if (!CGF.HaveInsertPoint())
2513 assert(Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked ||
2514 Schedule == OMP_sch_static_balanced_chunked ||
2515 Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked ||
2516 Schedule == OMP_dist_sch_static ||
2517 Schedule == OMP_dist_sch_static_chunked);
2519 // Call __kmpc_for_static_init(
2520 // ident_t *loc, kmp_int32 tid, kmp_int32 schedtype,
2521 // kmp_int32 *p_lastiter, kmp_int[32|64] *p_lower,
2522 // kmp_int[32|64] *p_upper, kmp_int[32|64] *p_stride,
2523 // kmp_int[32|64] incr, kmp_int[32|64] chunk);
2524 if (Chunk == nullptr) {
2525 assert((Schedule == OMP_sch_static || Schedule == OMP_ord_static ||
2526 Schedule == OMP_dist_sch_static) &&
2527 "expected static non-chunked schedule");
2528 // If the Chunk was not specified in the clause - use default value 1.
2529 Chunk = CGF.Builder.getIntN(IVSize, 1);
2531 assert((Schedule == OMP_sch_static_chunked ||
2532 Schedule == OMP_sch_static_balanced_chunked ||
2533 Schedule == OMP_ord_static_chunked ||
2534 Schedule == OMP_dist_sch_static_chunked) &&
2535 "expected static chunked schedule");
2537 llvm::Value *Args[] = {
2538 UpdateLocation, ThreadId, CGF.Builder.getInt32(addMonoNonMonoModifier(
2539 Schedule, M1, M2)), // Schedule type
2540 IL.getPointer(), // &isLastIter
2541 LB.getPointer(), // &LB
2542 UB.getPointer(), // &UB
2543 ST.getPointer(), // &Stride
2544 CGF.Builder.getIntN(IVSize, 1), // Incr
2547 CGF.EmitRuntimeCall(ForStaticInitFunction, Args);
2550 void CGOpenMPRuntime::emitForStaticInit(CodeGenFunction &CGF,
2552 const OpenMPScheduleTy &ScheduleKind,
2553 unsigned IVSize, bool IVSigned,
2554 bool Ordered, Address IL, Address LB,
2555 Address UB, Address ST,
2556 llvm::Value *Chunk) {
2557 OpenMPSchedType ScheduleNum =
2558 getRuntimeSchedule(ScheduleKind.Schedule, Chunk != nullptr, Ordered);
2559 auto *UpdatedLocation = emitUpdateLocation(CGF, Loc);
2560 auto *ThreadId = getThreadID(CGF, Loc);
2561 auto *StaticInitFunction = createForStaticInitFunction(IVSize, IVSigned);
2562 emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction,
2563 ScheduleNum, ScheduleKind.M1, ScheduleKind.M2, IVSize,
2564 Ordered, IL, LB, UB, ST, Chunk);
2567 void CGOpenMPRuntime::emitDistributeStaticInit(
2568 CodeGenFunction &CGF, SourceLocation Loc,
2569 OpenMPDistScheduleClauseKind SchedKind, unsigned IVSize, bool IVSigned,
2570 bool Ordered, Address IL, Address LB, Address UB, Address ST,
2571 llvm::Value *Chunk) {
2572 OpenMPSchedType ScheduleNum = getRuntimeSchedule(SchedKind, Chunk != nullptr);
2573 auto *UpdatedLocation = emitUpdateLocation(CGF, Loc);
2574 auto *ThreadId = getThreadID(CGF, Loc);
2575 auto *StaticInitFunction = createForStaticInitFunction(IVSize, IVSigned);
2576 emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction,
2577 ScheduleNum, OMPC_SCHEDULE_MODIFIER_unknown,
2578 OMPC_SCHEDULE_MODIFIER_unknown, IVSize, Ordered, IL, LB,
2582 void CGOpenMPRuntime::emitForStaticFinish(CodeGenFunction &CGF,
2583 SourceLocation Loc) {
2584 if (!CGF.HaveInsertPoint())
2586 // Call __kmpc_for_static_fini(ident_t *loc, kmp_int32 tid);
2587 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2588 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_for_static_fini),
2592 void CGOpenMPRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF,
2596 if (!CGF.HaveInsertPoint())
2598 // Call __kmpc_for_dynamic_fini_(4|8)[u](ident_t *loc, kmp_int32 tid);
2599 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2600 CGF.EmitRuntimeCall(createDispatchFiniFunction(IVSize, IVSigned), Args);
2603 llvm::Value *CGOpenMPRuntime::emitForNext(CodeGenFunction &CGF,
2604 SourceLocation Loc, unsigned IVSize,
2605 bool IVSigned, Address IL,
2606 Address LB, Address UB,
2608 // Call __kmpc_dispatch_next(
2609 // ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter,
2610 // kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper,
2611 // kmp_int[32|64] *p_stride);
2612 llvm::Value *Args[] = {
2613 emitUpdateLocation(CGF, Loc),
2614 getThreadID(CGF, Loc),
2615 IL.getPointer(), // &isLastIter
2616 LB.getPointer(), // &Lower
2617 UB.getPointer(), // &Upper
2618 ST.getPointer() // &Stride
2621 CGF.EmitRuntimeCall(createDispatchNextFunction(IVSize, IVSigned), Args);
2622 return CGF.EmitScalarConversion(
2623 Call, CGF.getContext().getIntTypeForBitwidth(32, /* Signed */ true),
2624 CGF.getContext().BoolTy, Loc);
2627 void CGOpenMPRuntime::emitNumThreadsClause(CodeGenFunction &CGF,
2628 llvm::Value *NumThreads,
2629 SourceLocation Loc) {
2630 if (!CGF.HaveInsertPoint())
2632 // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads)
2633 llvm::Value *Args[] = {
2634 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2635 CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)};
2636 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_num_threads),
2640 void CGOpenMPRuntime::emitProcBindClause(CodeGenFunction &CGF,
2641 OpenMPProcBindClauseKind ProcBind,
2642 SourceLocation Loc) {
2643 if (!CGF.HaveInsertPoint())
2645 // Constants for proc bind value accepted by the runtime.
2656 case OMPC_PROC_BIND_master:
2657 RuntimeProcBind = ProcBindMaster;
2659 case OMPC_PROC_BIND_close:
2660 RuntimeProcBind = ProcBindClose;
2662 case OMPC_PROC_BIND_spread:
2663 RuntimeProcBind = ProcBindSpread;
2665 case OMPC_PROC_BIND_unknown:
2666 llvm_unreachable("Unsupported proc_bind value.");
2668 // Build call __kmpc_push_proc_bind(&loc, global_tid, proc_bind)
2669 llvm::Value *Args[] = {
2670 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2671 llvm::ConstantInt::get(CGM.IntTy, RuntimeProcBind, /*isSigned=*/true)};
2672 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_proc_bind), Args);
2675 void CGOpenMPRuntime::emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>,
2676 SourceLocation Loc) {
2677 if (!CGF.HaveInsertPoint())
2679 // Build call void __kmpc_flush(ident_t *loc)
2680 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_flush),
2681 emitUpdateLocation(CGF, Loc));
2685 /// \brief Indexes of fields for type kmp_task_t.
2686 enum KmpTaskTFields {
2687 /// \brief List of shared variables.
2689 /// \brief Task routine.
2691 /// \brief Partition id for the untied tasks.
2693 /// Function with call of destructors for private variables.
2697 /// (Taskloops only) Lower bound.
2699 /// (Taskloops only) Upper bound.
2701 /// (Taskloops only) Stride.
2703 /// (Taskloops only) Is last iteration flag.
2706 } // anonymous namespace
2708 bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::empty() const {
2709 // FIXME: Add other entries type when they become supported.
2710 return OffloadEntriesTargetRegion.empty();
2713 /// \brief Initialize target region entry.
2714 void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::
2715 initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
2716 StringRef ParentName, unsigned LineNum,
2718 assert(CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is "
2719 "only required for the device "
2720 "code generation.");
2721 OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] =
2722 OffloadEntryInfoTargetRegion(Order, /*Addr=*/nullptr, /*ID=*/nullptr,
2724 ++OffloadingEntriesNum;
2727 void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::
2728 registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
2729 StringRef ParentName, unsigned LineNum,
2730 llvm::Constant *Addr, llvm::Constant *ID,
2732 // If we are emitting code for a target, the entry is already initialized,
2733 // only has to be registered.
2734 if (CGM.getLangOpts().OpenMPIsDevice) {
2735 assert(hasTargetRegionEntryInfo(DeviceID, FileID, ParentName, LineNum) &&
2736 "Entry must exist.");
2738 OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum];
2739 assert(Entry.isValid() && "Entry not initialized!");
2740 Entry.setAddress(Addr);
2742 Entry.setFlags(Flags);
2745 OffloadEntryInfoTargetRegion Entry(OffloadingEntriesNum++, Addr, ID, Flags);
2746 OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] = Entry;
2750 bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::hasTargetRegionEntryInfo(
2751 unsigned DeviceID, unsigned FileID, StringRef ParentName,
2752 unsigned LineNum) const {
2753 auto PerDevice = OffloadEntriesTargetRegion.find(DeviceID);
2754 if (PerDevice == OffloadEntriesTargetRegion.end())
2756 auto PerFile = PerDevice->second.find(FileID);
2757 if (PerFile == PerDevice->second.end())
2759 auto PerParentName = PerFile->second.find(ParentName);
2760 if (PerParentName == PerFile->second.end())
2762 auto PerLine = PerParentName->second.find(LineNum);
2763 if (PerLine == PerParentName->second.end())
2765 // Fail if this entry is already registered.
2766 if (PerLine->second.getAddress() || PerLine->second.getID())
2771 void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::actOnTargetRegionEntriesInfo(
2772 const OffloadTargetRegionEntryInfoActTy &Action) {
2773 // Scan all target region entries and perform the provided action.
2774 for (auto &D : OffloadEntriesTargetRegion)
2775 for (auto &F : D.second)
2776 for (auto &P : F.second)
2777 for (auto &L : P.second)
2778 Action(D.first, F.first, P.first(), L.first, L.second);
2781 /// \brief Create a Ctor/Dtor-like function whose body is emitted through
2782 /// \a Codegen. This is used to emit the two functions that register and
2783 /// unregister the descriptor of the current compilation unit.
2784 static llvm::Function *
2785 createOffloadingBinaryDescriptorFunction(CodeGenModule &CGM, StringRef Name,
2786 const RegionCodeGenTy &Codegen) {
2787 auto &C = CGM.getContext();
2788 FunctionArgList Args;
2789 ImplicitParamDecl DummyPtr(C, /*DC=*/nullptr, SourceLocation(),
2790 /*Id=*/nullptr, C.VoidPtrTy);
2791 Args.push_back(&DummyPtr);
2793 CodeGenFunction CGF(CGM);
2794 auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2795 auto FTy = CGM.getTypes().GetFunctionType(FI);
2797 CGM.CreateGlobalInitOrDestructFunction(FTy, Name, FI, SourceLocation());
2798 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FI, Args, SourceLocation());
2800 CGF.FinishFunction();
2805 CGOpenMPRuntime::createOffloadingBinaryDescriptorRegistration() {
2807 // If we don't have entries or if we are emitting code for the device, we
2808 // don't need to do anything.
2809 if (CGM.getLangOpts().OpenMPIsDevice || OffloadEntriesInfoManager.empty())
2812 auto &M = CGM.getModule();
2813 auto &C = CGM.getContext();
2815 // Get list of devices we care about
2816 auto &Devices = CGM.getLangOpts().OMPTargetTriples;
2818 // We should be creating an offloading descriptor only if there are devices
2820 assert(!Devices.empty() && "No OpenMP offloading devices??");
2822 // Create the external variables that will point to the begin and end of the
2823 // host entries section. These will be defined by the linker.
2824 auto *OffloadEntryTy =
2825 CGM.getTypes().ConvertTypeForMem(getTgtOffloadEntryQTy());
2826 llvm::GlobalVariable *HostEntriesBegin = new llvm::GlobalVariable(
2827 M, OffloadEntryTy, /*isConstant=*/true,
2828 llvm::GlobalValue::ExternalLinkage, /*Initializer=*/nullptr,
2829 ".omp_offloading.entries_begin");
2830 llvm::GlobalVariable *HostEntriesEnd = new llvm::GlobalVariable(
2831 M, OffloadEntryTy, /*isConstant=*/true,
2832 llvm::GlobalValue::ExternalLinkage, /*Initializer=*/nullptr,
2833 ".omp_offloading.entries_end");
2835 // Create all device images
2836 auto *DeviceImageTy = cast<llvm::StructType>(
2837 CGM.getTypes().ConvertTypeForMem(getTgtDeviceImageQTy()));
2838 ConstantInitBuilder DeviceImagesBuilder(CGM);
2839 auto DeviceImagesEntries = DeviceImagesBuilder.beginArray(DeviceImageTy);
2841 for (unsigned i = 0; i < Devices.size(); ++i) {
2842 StringRef T = Devices[i].getTriple();
2843 auto *ImgBegin = new llvm::GlobalVariable(
2844 M, CGM.Int8Ty, /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage,
2845 /*Initializer=*/nullptr,
2846 Twine(".omp_offloading.img_start.") + Twine(T));
2847 auto *ImgEnd = new llvm::GlobalVariable(
2848 M, CGM.Int8Ty, /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage,
2849 /*Initializer=*/nullptr, Twine(".omp_offloading.img_end.") + Twine(T));
2851 auto Dev = DeviceImagesEntries.beginStruct(DeviceImageTy);
2854 Dev.add(HostEntriesBegin);
2855 Dev.add(HostEntriesEnd);
2856 Dev.finishAndAddTo(DeviceImagesEntries);
2859 // Create device images global array.
2860 llvm::GlobalVariable *DeviceImages =
2861 DeviceImagesEntries.finishAndCreateGlobal(".omp_offloading.device_images",
2862 CGM.getPointerAlign(),
2863 /*isConstant=*/true);
2864 DeviceImages->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2866 // This is a Zero array to be used in the creation of the constant expressions
2867 llvm::Constant *Index[] = {llvm::Constant::getNullValue(CGM.Int32Ty),
2868 llvm::Constant::getNullValue(CGM.Int32Ty)};
2870 // Create the target region descriptor.
2871 auto *BinaryDescriptorTy = cast<llvm::StructType>(
2872 CGM.getTypes().ConvertTypeForMem(getTgtBinaryDescriptorQTy()));
2873 ConstantInitBuilder DescBuilder(CGM);
2874 auto DescInit = DescBuilder.beginStruct(BinaryDescriptorTy);
2875 DescInit.addInt(CGM.Int32Ty, Devices.size());
2876 DescInit.add(llvm::ConstantExpr::getGetElementPtr(DeviceImages->getValueType(),
2879 DescInit.add(HostEntriesBegin);
2880 DescInit.add(HostEntriesEnd);
2882 auto *Desc = DescInit.finishAndCreateGlobal(".omp_offloading.descriptor",
2883 CGM.getPointerAlign(),
2884 /*isConstant=*/true);
2886 // Emit code to register or unregister the descriptor at execution
2887 // startup or closing, respectively.
2889 // Create a variable to drive the registration and unregistration of the
2890 // descriptor, so we can reuse the logic that emits Ctors and Dtors.
2891 auto *IdentInfo = &C.Idents.get(".omp_offloading.reg_unreg_var");
2892 ImplicitParamDecl RegUnregVar(C, C.getTranslationUnitDecl(), SourceLocation(),
2893 IdentInfo, C.CharTy);
2895 auto *UnRegFn = createOffloadingBinaryDescriptorFunction(
2896 CGM, ".omp_offloading.descriptor_unreg",
2897 [&](CodeGenFunction &CGF, PrePostActionTy &) {
2898 CGF.EmitCallOrInvoke(createRuntimeFunction(OMPRTL__tgt_unregister_lib),
2901 auto *RegFn = createOffloadingBinaryDescriptorFunction(
2902 CGM, ".omp_offloading.descriptor_reg",
2903 [&](CodeGenFunction &CGF, PrePostActionTy &) {
2904 CGF.EmitCallOrInvoke(createRuntimeFunction(OMPRTL__tgt_register_lib),
2906 CGM.getCXXABI().registerGlobalDtor(CGF, RegUnregVar, UnRegFn, Desc);
2911 void CGOpenMPRuntime::createOffloadEntry(llvm::Constant *ID,
2912 llvm::Constant *Addr, uint64_t Size,
2914 StringRef Name = Addr->getName();
2915 auto *TgtOffloadEntryType = cast<llvm::StructType>(
2916 CGM.getTypes().ConvertTypeForMem(getTgtOffloadEntryQTy()));
2917 llvm::LLVMContext &C = CGM.getModule().getContext();
2918 llvm::Module &M = CGM.getModule();
2920 // Make sure the address has the right type.
2921 llvm::Constant *AddrPtr = llvm::ConstantExpr::getBitCast(ID, CGM.VoidPtrTy);
2923 // Create constant string with the name.
2924 llvm::Constant *StrPtrInit = llvm::ConstantDataArray::getString(C, Name);
2926 llvm::GlobalVariable *Str =
2927 new llvm::GlobalVariable(M, StrPtrInit->getType(), /*isConstant=*/true,
2928 llvm::GlobalValue::InternalLinkage, StrPtrInit,
2929 ".omp_offloading.entry_name");
2930 Str->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2931 llvm::Constant *StrPtr = llvm::ConstantExpr::getBitCast(Str, CGM.Int8PtrTy);
2933 // We can't have any padding between symbols, so we need to have 1-byte
2935 auto Align = CharUnits::fromQuantity(1);
2937 // Create the entry struct.
2938 ConstantInitBuilder EntryBuilder(CGM);
2939 auto EntryInit = EntryBuilder.beginStruct(TgtOffloadEntryType);
2940 EntryInit.add(AddrPtr);
2941 EntryInit.add(StrPtr);
2942 EntryInit.addInt(CGM.SizeTy, Size);
2943 EntryInit.addInt(CGM.Int32Ty, Flags);
2944 EntryInit.addInt(CGM.Int32Ty, 0);
2945 llvm::GlobalVariable *Entry =
2946 EntryInit.finishAndCreateGlobal(".omp_offloading.entry",
2949 llvm::GlobalValue::ExternalLinkage);
2951 // The entry has to be created in the section the linker expects it to be.
2952 Entry->setSection(".omp_offloading.entries");
2955 void CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata() {
2956 // Emit the offloading entries and metadata so that the device codegen side
2957 // can easily figure out what to emit. The produced metadata looks like
2960 // !omp_offload.info = !{!1, ...}
2962 // Right now we only generate metadata for function that contain target
2965 // If we do not have entries, we dont need to do anything.
2966 if (OffloadEntriesInfoManager.empty())
2969 llvm::Module &M = CGM.getModule();
2970 llvm::LLVMContext &C = M.getContext();
2971 SmallVector<OffloadEntriesInfoManagerTy::OffloadEntryInfo *, 16>
2972 OrderedEntries(OffloadEntriesInfoManager.size());
2974 // Create the offloading info metadata node.
2975 llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("omp_offload.info");
2977 // Auxiliary methods to create metadata values and strings.
2978 auto getMDInt = [&](unsigned v) {
2979 return llvm::ConstantAsMetadata::get(
2980 llvm::ConstantInt::get(llvm::Type::getInt32Ty(C), v));
2983 auto getMDString = [&](StringRef v) { return llvm::MDString::get(C, v); };
2985 // Create function that emits metadata for each target region entry;
2986 auto &&TargetRegionMetadataEmitter = [&](
2987 unsigned DeviceID, unsigned FileID, StringRef ParentName, unsigned Line,
2988 OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion &E) {
2989 llvm::SmallVector<llvm::Metadata *, 32> Ops;
2990 // Generate metadata for target regions. Each entry of this metadata
2992 // - Entry 0 -> Kind of this type of metadata (0).
2993 // - Entry 1 -> Device ID of the file where the entry was identified.
2994 // - Entry 2 -> File ID of the file where the entry was identified.
2995 // - Entry 3 -> Mangled name of the function where the entry was identified.
2996 // - Entry 4 -> Line in the file where the entry was identified.
2997 // - Entry 5 -> Order the entry was created.
2998 // The first element of the metadata node is the kind.
2999 Ops.push_back(getMDInt(E.getKind()));
3000 Ops.push_back(getMDInt(DeviceID));
3001 Ops.push_back(getMDInt(FileID));
3002 Ops.push_back(getMDString(ParentName));
3003 Ops.push_back(getMDInt(Line));
3004 Ops.push_back(getMDInt(E.getOrder()));
3006 // Save this entry in the right position of the ordered entries array.
3007 OrderedEntries[E.getOrder()] = &E;
3009 // Add metadata to the named metadata node.
3010 MD->addOperand(llvm::MDNode::get(C, Ops));
3013 OffloadEntriesInfoManager.actOnTargetRegionEntriesInfo(
3014 TargetRegionMetadataEmitter);
3016 for (auto *E : OrderedEntries) {
3017 assert(E && "All ordered entries must exist!");
3019 dyn_cast<OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion>(
3021 assert(CE->getID() && CE->getAddress() &&
3022 "Entry ID and Addr are invalid!");
3023 createOffloadEntry(CE->getID(), CE->getAddress(), /*Size=*/0);
3025 llvm_unreachable("Unsupported entry kind.");
3029 /// \brief Loads all the offload entries information from the host IR
3031 void CGOpenMPRuntime::loadOffloadInfoMetadata() {
3032 // If we are in target mode, load the metadata from the host IR. This code has
3033 // to match the metadaata creation in createOffloadEntriesAndInfoMetadata().
3035 if (!CGM.getLangOpts().OpenMPIsDevice)
3038 if (CGM.getLangOpts().OMPHostIRFile.empty())
3041 auto Buf = llvm::MemoryBuffer::getFile(CGM.getLangOpts().OMPHostIRFile);
3045 llvm::LLVMContext C;
3046 auto ME = expectedToErrorOrAndEmitErrors(
3047 C, llvm::parseBitcodeFile(Buf.get()->getMemBufferRef(), C));
3052 llvm::NamedMDNode *MD = ME.get()->getNamedMetadata("omp_offload.info");
3056 for (auto I : MD->operands()) {
3057 llvm::MDNode *MN = cast<llvm::MDNode>(I);
3059 auto getMDInt = [&](unsigned Idx) {
3060 llvm::ConstantAsMetadata *V =
3061 cast<llvm::ConstantAsMetadata>(MN->getOperand(Idx));
3062 return cast<llvm::ConstantInt>(V->getValue())->getZExtValue();
3065 auto getMDString = [&](unsigned Idx) {
3066 llvm::MDString *V = cast<llvm::MDString>(MN->getOperand(Idx));
3067 return V->getString();
3070 switch (getMDInt(0)) {
3072 llvm_unreachable("Unexpected metadata!");
3074 case OffloadEntriesInfoManagerTy::OffloadEntryInfo::
3075 OFFLOAD_ENTRY_INFO_TARGET_REGION:
3076 OffloadEntriesInfoManager.initializeTargetRegionEntryInfo(
3077 /*DeviceID=*/getMDInt(1), /*FileID=*/getMDInt(2),
3078 /*ParentName=*/getMDString(3), /*Line=*/getMDInt(4),
3079 /*Order=*/getMDInt(5));
3085 void CGOpenMPRuntime::emitKmpRoutineEntryT(QualType KmpInt32Ty) {
3086 if (!KmpRoutineEntryPtrTy) {
3087 // Build typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); type.
3088 auto &C = CGM.getContext();
3089 QualType KmpRoutineEntryTyArgs[] = {KmpInt32Ty, C.VoidPtrTy};
3090 FunctionProtoType::ExtProtoInfo EPI;
3091 KmpRoutineEntryPtrQTy = C.getPointerType(
3092 C.getFunctionType(KmpInt32Ty, KmpRoutineEntryTyArgs, EPI));
3093 KmpRoutineEntryPtrTy = CGM.getTypes().ConvertType(KmpRoutineEntryPtrQTy);
3097 static FieldDecl *addFieldToRecordDecl(ASTContext &C, DeclContext *DC,
3099 auto *Field = FieldDecl::Create(
3100 C, DC, SourceLocation(), SourceLocation(), /*Id=*/nullptr, FieldTy,
3101 C.getTrivialTypeSourceInfo(FieldTy, SourceLocation()),
3102 /*BW=*/nullptr, /*Mutable=*/false, /*InitStyle=*/ICIS_NoInit);
3103 Field->setAccess(AS_public);
3108 QualType CGOpenMPRuntime::getTgtOffloadEntryQTy() {
3110 // Make sure the type of the entry is already created. This is the type we
3112 // struct __tgt_offload_entry{
3113 // void *addr; // Pointer to the offload entry info.
3114 // // (function or global)
3115 // char *name; // Name of the function or global.
3116 // size_t size; // Size of the entry info (0 if it a function).
3117 // int32_t flags; // Flags associated with the entry, e.g. 'link'.
3118 // int32_t reserved; // Reserved, to use by the runtime library.
3120 if (TgtOffloadEntryQTy.isNull()) {
3121 ASTContext &C = CGM.getContext();
3122 auto *RD = C.buildImplicitRecord("__tgt_offload_entry");
3123 RD->startDefinition();
3124 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3125 addFieldToRecordDecl(C, RD, C.getPointerType(C.CharTy));
3126 addFieldToRecordDecl(C, RD, C.getSizeType());
3127 addFieldToRecordDecl(
3128 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true));
3129 addFieldToRecordDecl(
3130 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true));
3131 RD->completeDefinition();
3132 TgtOffloadEntryQTy = C.getRecordType(RD);
3134 return TgtOffloadEntryQTy;
3137 QualType CGOpenMPRuntime::getTgtDeviceImageQTy() {
3138 // These are the types we need to build:
3139 // struct __tgt_device_image{
3140 // void *ImageStart; // Pointer to the target code start.
3141 // void *ImageEnd; // Pointer to the target code end.
3142 // // We also add the host entries to the device image, as it may be useful
3143 // // for the target runtime to have access to that information.
3144 // __tgt_offload_entry *EntriesBegin; // Begin of the table with all
3146 // __tgt_offload_entry *EntriesEnd; // End of the table with all the
3147 // // entries (non inclusive).
3149 if (TgtDeviceImageQTy.isNull()) {
3150 ASTContext &C = CGM.getContext();
3151 auto *RD = C.buildImplicitRecord("__tgt_device_image");
3152 RD->startDefinition();
3153 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3154 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3155 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3156 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3157 RD->completeDefinition();
3158 TgtDeviceImageQTy = C.getRecordType(RD);
3160 return TgtDeviceImageQTy;
3163 QualType CGOpenMPRuntime::getTgtBinaryDescriptorQTy() {
3164 // struct __tgt_bin_desc{
3165 // int32_t NumDevices; // Number of devices supported.
3166 // __tgt_device_image *DeviceImages; // Arrays of device images
3167 // // (one per device).
3168 // __tgt_offload_entry *EntriesBegin; // Begin of the table with all the
3170 // __tgt_offload_entry *EntriesEnd; // End of the table with all the
3171 // // entries (non inclusive).
3173 if (TgtBinaryDescriptorQTy.isNull()) {
3174 ASTContext &C = CGM.getContext();
3175 auto *RD = C.buildImplicitRecord("__tgt_bin_desc");
3176 RD->startDefinition();
3177 addFieldToRecordDecl(
3178 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true));
3179 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtDeviceImageQTy()));
3180 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3181 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3182 RD->completeDefinition();
3183 TgtBinaryDescriptorQTy = C.getRecordType(RD);
3185 return TgtBinaryDescriptorQTy;
3189 struct PrivateHelpersTy {
3190 PrivateHelpersTy(const VarDecl *Original, const VarDecl *PrivateCopy,
3191 const VarDecl *PrivateElemInit)
3192 : Original(Original), PrivateCopy(PrivateCopy),
3193 PrivateElemInit(PrivateElemInit) {}
3194 const VarDecl *Original;
3195 const VarDecl *PrivateCopy;
3196 const VarDecl *PrivateElemInit;
3198 typedef std::pair<CharUnits /*Align*/, PrivateHelpersTy> PrivateDataTy;
3199 } // anonymous namespace
3202 createPrivatesRecordDecl(CodeGenModule &CGM, ArrayRef<PrivateDataTy> Privates) {
3203 if (!Privates.empty()) {
3204 auto &C = CGM.getContext();
3205 // Build struct .kmp_privates_t. {
3206 // /* private vars */
3208 auto *RD = C.buildImplicitRecord(".kmp_privates.t");
3209 RD->startDefinition();
3210 for (auto &&Pair : Privates) {
3211 auto *VD = Pair.second.Original;
3212 auto Type = VD->getType();
3213 Type = Type.getNonReferenceType();
3214 auto *FD = addFieldToRecordDecl(C, RD, Type);
3215 if (VD->hasAttrs()) {
3216 for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
3217 E(VD->getAttrs().end());
3222 RD->completeDefinition();
3229 createKmpTaskTRecordDecl(CodeGenModule &CGM, OpenMPDirectiveKind Kind,
3230 QualType KmpInt32Ty,
3231 QualType KmpRoutineEntryPointerQTy) {
3232 auto &C = CGM.getContext();
3233 // Build struct kmp_task_t {
3235 // kmp_routine_entry_t routine;
3236 // kmp_int32 part_id;
3237 // kmp_cmplrdata_t data1;
3238 // kmp_cmplrdata_t data2;
3239 // For taskloops additional fields:
3245 auto *UD = C.buildImplicitRecord("kmp_cmplrdata_t", TTK_Union);
3246 UD->startDefinition();
3247 addFieldToRecordDecl(C, UD, KmpInt32Ty);
3248 addFieldToRecordDecl(C, UD, KmpRoutineEntryPointerQTy);
3249 UD->completeDefinition();
3250 QualType KmpCmplrdataTy = C.getRecordType(UD);
3251 auto *RD = C.buildImplicitRecord("kmp_task_t");
3252 RD->startDefinition();
3253 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3254 addFieldToRecordDecl(C, RD, KmpRoutineEntryPointerQTy);
3255 addFieldToRecordDecl(C, RD, KmpInt32Ty);
3256 addFieldToRecordDecl(C, RD, KmpCmplrdataTy);
3257 addFieldToRecordDecl(C, RD, KmpCmplrdataTy);
3258 if (isOpenMPTaskLoopDirective(Kind)) {
3259 QualType KmpUInt64Ty =
3260 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0);
3261 QualType KmpInt64Ty =
3262 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
3263 addFieldToRecordDecl(C, RD, KmpUInt64Ty);
3264 addFieldToRecordDecl(C, RD, KmpUInt64Ty);
3265 addFieldToRecordDecl(C, RD, KmpInt64Ty);
3266 addFieldToRecordDecl(C, RD, KmpInt32Ty);
3268 RD->completeDefinition();
3273 createKmpTaskTWithPrivatesRecordDecl(CodeGenModule &CGM, QualType KmpTaskTQTy,
3274 ArrayRef<PrivateDataTy> Privates) {
3275 auto &C = CGM.getContext();
3276 // Build struct kmp_task_t_with_privates {
3277 // kmp_task_t task_data;
3278 // .kmp_privates_t. privates;
3280 auto *RD = C.buildImplicitRecord("kmp_task_t_with_privates");
3281 RD->startDefinition();
3282 addFieldToRecordDecl(C, RD, KmpTaskTQTy);
3283 if (auto *PrivateRD = createPrivatesRecordDecl(CGM, Privates)) {
3284 addFieldToRecordDecl(C, RD, C.getRecordType(PrivateRD));
3286 RD->completeDefinition();
3290 /// \brief Emit a proxy function which accepts kmp_task_t as the second
3293 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
3294 /// TaskFunction(gtid, tt->part_id, &tt->privates, task_privates_map, tt,
3296 /// tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter,
3301 static llvm::Value *
3302 emitProxyTaskFunction(CodeGenModule &CGM, SourceLocation Loc,
3303 OpenMPDirectiveKind Kind, QualType KmpInt32Ty,
3304 QualType KmpTaskTWithPrivatesPtrQTy,
3305 QualType KmpTaskTWithPrivatesQTy, QualType KmpTaskTQTy,
3306 QualType SharedsPtrTy, llvm::Value *TaskFunction,
3307 llvm::Value *TaskPrivatesMap) {
3308 auto &C = CGM.getContext();
3309 FunctionArgList Args;
3310 ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty);
3311 ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc,
3313 KmpTaskTWithPrivatesPtrQTy.withRestrict());
3314 Args.push_back(&GtidArg);
3315 Args.push_back(&TaskTypeArg);
3316 auto &TaskEntryFnInfo =
3317 CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args);
3318 auto *TaskEntryTy = CGM.getTypes().GetFunctionType(TaskEntryFnInfo);
3320 llvm::Function::Create(TaskEntryTy, llvm::GlobalValue::InternalLinkage,
3321 ".omp_task_entry.", &CGM.getModule());
3322 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, TaskEntry, TaskEntryFnInfo);
3323 CodeGenFunction CGF(CGM);
3324 CGF.disableDebugInfo();
3325 CGF.StartFunction(GlobalDecl(), KmpInt32Ty, TaskEntry, TaskEntryFnInfo, Args);
3327 // TaskFunction(gtid, tt->task_data.part_id, &tt->privates, task_privates_map,
3330 // tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter,
3331 // tt->task_data.shareds);
3332 auto *GtidParam = CGF.EmitLoadOfScalar(
3333 CGF.GetAddrOfLocalVar(&GtidArg), /*Volatile=*/false, KmpInt32Ty, Loc);
3334 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3335 CGF.GetAddrOfLocalVar(&TaskTypeArg),
3336 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3337 auto *KmpTaskTWithPrivatesQTyRD =
3338 cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl());
3340 CGF.EmitLValueForField(TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3341 auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl());
3342 auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId);
3343 auto PartIdLVal = CGF.EmitLValueForField(Base, *PartIdFI);
3344 auto *PartidParam = PartIdLVal.getPointer();
3346 auto SharedsFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds);
3347 auto SharedsLVal = CGF.EmitLValueForField(Base, *SharedsFI);
3348 auto *SharedsParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3349 CGF.EmitLoadOfLValue(SharedsLVal, Loc).getScalarVal(),
3350 CGF.ConvertTypeForMem(SharedsPtrTy));
3352 auto PrivatesFI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1);
3353 llvm::Value *PrivatesParam;
3354 if (PrivatesFI != KmpTaskTWithPrivatesQTyRD->field_end()) {
3355 auto PrivatesLVal = CGF.EmitLValueForField(TDBase, *PrivatesFI);
3356 PrivatesParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3357 PrivatesLVal.getPointer(), CGF.VoidPtrTy);
3359 PrivatesParam = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
3361 llvm::Value *CommonArgs[] = {GtidParam, PartidParam, PrivatesParam,
3364 .CreatePointerBitCastOrAddrSpaceCast(
3365 TDBase.getAddress(), CGF.VoidPtrTy)
3367 SmallVector<llvm::Value *, 16> CallArgs(std::begin(CommonArgs),
3368 std::end(CommonArgs));
3369 if (isOpenMPTaskLoopDirective(Kind)) {
3370 auto LBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound);
3371 auto LBLVal = CGF.EmitLValueForField(Base, *LBFI);
3372 auto *LBParam = CGF.EmitLoadOfLValue(LBLVal, Loc).getScalarVal();
3373 auto UBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound);
3374 auto UBLVal = CGF.EmitLValueForField(Base, *UBFI);
3375 auto *UBParam = CGF.EmitLoadOfLValue(UBLVal, Loc).getScalarVal();
3376 auto StFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTStride);
3377 auto StLVal = CGF.EmitLValueForField(Base, *StFI);
3378 auto *StParam = CGF.EmitLoadOfLValue(StLVal, Loc).getScalarVal();
3379 auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter);
3380 auto LILVal = CGF.EmitLValueForField(Base, *LIFI);
3381 auto *LIParam = CGF.EmitLoadOfLValue(LILVal, Loc).getScalarVal();
3382 CallArgs.push_back(LBParam);
3383 CallArgs.push_back(UBParam);
3384 CallArgs.push_back(StParam);
3385 CallArgs.push_back(LIParam);
3387 CallArgs.push_back(SharedsParam);
3389 CGF.EmitCallOrInvoke(TaskFunction, CallArgs);
3390 CGF.EmitStoreThroughLValue(
3391 RValue::get(CGF.Builder.getInt32(/*C=*/0)),
3392 CGF.MakeAddrLValue(CGF.ReturnValue, KmpInt32Ty));
3393 CGF.FinishFunction();
3397 static llvm::Value *emitDestructorsFunction(CodeGenModule &CGM,
3399 QualType KmpInt32Ty,
3400 QualType KmpTaskTWithPrivatesPtrQTy,
3401 QualType KmpTaskTWithPrivatesQTy) {
3402 auto &C = CGM.getContext();
3403 FunctionArgList Args;
3404 ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty);
3405 ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc,
3407 KmpTaskTWithPrivatesPtrQTy.withRestrict());
3408 Args.push_back(&GtidArg);
3409 Args.push_back(&TaskTypeArg);
3410 FunctionType::ExtInfo Info;
3411 auto &DestructorFnInfo =
3412 CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args);
3413 auto *DestructorFnTy = CGM.getTypes().GetFunctionType(DestructorFnInfo);
3414 auto *DestructorFn =
3415 llvm::Function::Create(DestructorFnTy, llvm::GlobalValue::InternalLinkage,
3416 ".omp_task_destructor.", &CGM.getModule());
3417 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, DestructorFn,
3419 CodeGenFunction CGF(CGM);
3420 CGF.disableDebugInfo();
3421 CGF.StartFunction(GlobalDecl(), KmpInt32Ty, DestructorFn, DestructorFnInfo,
3424 LValue Base = CGF.EmitLoadOfPointerLValue(
3425 CGF.GetAddrOfLocalVar(&TaskTypeArg),
3426 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3427 auto *KmpTaskTWithPrivatesQTyRD =
3428 cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl());
3429 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3430 Base = CGF.EmitLValueForField(Base, *FI);
3432 cast<RecordDecl>(FI->getType()->getAsTagDecl())->fields()) {
3433 if (auto DtorKind = Field->getType().isDestructedType()) {
3434 auto FieldLValue = CGF.EmitLValueForField(Base, Field);
3435 CGF.pushDestroy(DtorKind, FieldLValue.getAddress(), Field->getType());
3438 CGF.FinishFunction();
3439 return DestructorFn;
3442 /// \brief Emit a privates mapping function for correct handling of private and
3443 /// firstprivate variables.
3445 /// void .omp_task_privates_map.(const .privates. *noalias privs, <ty1>
3446 /// **noalias priv1,..., <tyn> **noalias privn) {
3447 /// *priv1 = &.privates.priv1;
3449 /// *privn = &.privates.privn;
3452 static llvm::Value *
3453 emitTaskPrivateMappingFunction(CodeGenModule &CGM, SourceLocation Loc,
3454 ArrayRef<const Expr *> PrivateVars,
3455 ArrayRef<const Expr *> FirstprivateVars,
3456 ArrayRef<const Expr *> LastprivateVars,
3457 QualType PrivatesQTy,
3458 ArrayRef<PrivateDataTy> Privates) {
3459 auto &C = CGM.getContext();
3460 FunctionArgList Args;
3461 ImplicitParamDecl TaskPrivatesArg(
3462 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3463 C.getPointerType(PrivatesQTy).withConst().withRestrict());
3464 Args.push_back(&TaskPrivatesArg);
3465 llvm::DenseMap<const VarDecl *, unsigned> PrivateVarsPos;
3466 unsigned Counter = 1;
3467 for (auto *E: PrivateVars) {
3468 Args.push_back(ImplicitParamDecl::Create(
3469 C, /*DC=*/nullptr, Loc,
3470 /*Id=*/nullptr, C.getPointerType(C.getPointerType(E->getType()))
3473 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3474 PrivateVarsPos[VD] = Counter;
3477 for (auto *E : FirstprivateVars) {
3478 Args.push_back(ImplicitParamDecl::Create(
3479 C, /*DC=*/nullptr, Loc,
3480 /*Id=*/nullptr, C.getPointerType(C.getPointerType(E->getType()))
3483 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3484 PrivateVarsPos[VD] = Counter;
3487 for (auto *E: LastprivateVars) {
3488 Args.push_back(ImplicitParamDecl::Create(
3489 C, /*DC=*/nullptr, Loc,
3490 /*Id=*/nullptr, C.getPointerType(C.getPointerType(E->getType()))
3493 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3494 PrivateVarsPos[VD] = Counter;
3497 auto &TaskPrivatesMapFnInfo =
3498 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3499 auto *TaskPrivatesMapTy =
3500 CGM.getTypes().GetFunctionType(TaskPrivatesMapFnInfo);
3501 auto *TaskPrivatesMap = llvm::Function::Create(
3502 TaskPrivatesMapTy, llvm::GlobalValue::InternalLinkage,
3503 ".omp_task_privates_map.", &CGM.getModule());
3504 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, TaskPrivatesMap,
3505 TaskPrivatesMapFnInfo);
3506 TaskPrivatesMap->removeFnAttr(llvm::Attribute::NoInline);
3507 TaskPrivatesMap->addFnAttr(llvm::Attribute::AlwaysInline);
3508 CodeGenFunction CGF(CGM);
3509 CGF.disableDebugInfo();
3510 CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskPrivatesMap,
3511 TaskPrivatesMapFnInfo, Args);
3513 // *privi = &.privates.privi;
3514 LValue Base = CGF.EmitLoadOfPointerLValue(
3515 CGF.GetAddrOfLocalVar(&TaskPrivatesArg),
3516 TaskPrivatesArg.getType()->castAs<PointerType>());
3517 auto *PrivatesQTyRD = cast<RecordDecl>(PrivatesQTy->getAsTagDecl());
3519 for (auto *Field : PrivatesQTyRD->fields()) {
3520 auto FieldLVal = CGF.EmitLValueForField(Base, Field);
3521 auto *VD = Args[PrivateVarsPos[Privates[Counter].second.Original]];
3522 auto RefLVal = CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
3523 auto RefLoadLVal = CGF.EmitLoadOfPointerLValue(
3524 RefLVal.getAddress(), RefLVal.getType()->castAs<PointerType>());
3525 CGF.EmitStoreOfScalar(FieldLVal.getPointer(), RefLoadLVal);
3528 CGF.FinishFunction();
3529 return TaskPrivatesMap;
3532 static int array_pod_sort_comparator(const PrivateDataTy *P1,
3533 const PrivateDataTy *P2) {
3534 return P1->first < P2->first ? 1 : (P2->first < P1->first ? -1 : 0);
3537 /// Emit initialization for private variables in task-based directives.
3538 static void emitPrivatesInit(CodeGenFunction &CGF,
3539 const OMPExecutableDirective &D,
3540 Address KmpTaskSharedsPtr, LValue TDBase,
3541 const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3542 QualType SharedsTy, QualType SharedsPtrTy,
3543 const OMPTaskDataTy &Data,
3544 ArrayRef<PrivateDataTy> Privates, bool ForDup) {
3545 auto &C = CGF.getContext();
3546 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3547 LValue PrivatesBase = CGF.EmitLValueForField(TDBase, *FI);
3549 if (!Data.FirstprivateVars.empty()) {
3550 SrcBase = CGF.MakeAddrLValue(
3551 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3552 KmpTaskSharedsPtr, CGF.ConvertTypeForMem(SharedsPtrTy)),
3555 CodeGenFunction::CGCapturedStmtInfo CapturesInfo(
3556 cast<CapturedStmt>(*D.getAssociatedStmt()));
3557 FI = cast<RecordDecl>(FI->getType()->getAsTagDecl())->field_begin();
3558 for (auto &&Pair : Privates) {
3559 auto *VD = Pair.second.PrivateCopy;
3560 auto *Init = VD->getAnyInitializer();
3561 if (Init && (!ForDup || (isa<CXXConstructExpr>(Init) &&
3562 !CGF.isTrivialInitializer(Init)))) {
3563 LValue PrivateLValue = CGF.EmitLValueForField(PrivatesBase, *FI);
3564 if (auto *Elem = Pair.second.PrivateElemInit) {
3565 auto *OriginalVD = Pair.second.Original;
3566 auto *SharedField = CapturesInfo.lookup(OriginalVD);
3567 auto SharedRefLValue = CGF.EmitLValueForField(SrcBase, SharedField);
3568 SharedRefLValue = CGF.MakeAddrLValue(
3569 Address(SharedRefLValue.getPointer(), C.getDeclAlign(OriginalVD)),
3570 SharedRefLValue.getType(), AlignmentSource::Decl);
3571 QualType Type = OriginalVD->getType();
3572 if (Type->isArrayType()) {
3573 // Initialize firstprivate array.
3574 if (!isa<CXXConstructExpr>(Init) || CGF.isTrivialInitializer(Init)) {
3575 // Perform simple memcpy.
3576 CGF.EmitAggregateAssign(PrivateLValue.getAddress(),
3577 SharedRefLValue.getAddress(), Type);
3579 // Initialize firstprivate array using element-by-element
3581 CGF.EmitOMPAggregateAssign(
3582 PrivateLValue.getAddress(), SharedRefLValue.getAddress(), Type,
3583 [&CGF, Elem, Init, &CapturesInfo](Address DestElement,
3584 Address SrcElement) {
3585 // Clean up any temporaries needed by the initialization.
3586 CodeGenFunction::OMPPrivateScope InitScope(CGF);
3587 InitScope.addPrivate(
3588 Elem, [SrcElement]() -> Address { return SrcElement; });
3589 (void)InitScope.Privatize();
3590 // Emit initialization for single element.
3591 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(
3592 CGF, &CapturesInfo);
3593 CGF.EmitAnyExprToMem(Init, DestElement,
3594 Init->getType().getQualifiers(),
3595 /*IsInitializer=*/false);
3599 CodeGenFunction::OMPPrivateScope InitScope(CGF);
3600 InitScope.addPrivate(Elem, [SharedRefLValue]() -> Address {
3601 return SharedRefLValue.getAddress();
3603 (void)InitScope.Privatize();
3604 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CapturesInfo);
3605 CGF.EmitExprAsInit(Init, VD, PrivateLValue,
3606 /*capturedByInit=*/false);
3609 CGF.EmitExprAsInit(Init, VD, PrivateLValue, /*capturedByInit=*/false);
3615 /// Check if duplication function is required for taskloops.
3616 static bool checkInitIsRequired(CodeGenFunction &CGF,
3617 ArrayRef<PrivateDataTy> Privates) {
3618 bool InitRequired = false;
3619 for (auto &&Pair : Privates) {
3620 auto *VD = Pair.second.PrivateCopy;
3621 auto *Init = VD->getAnyInitializer();
3622 InitRequired = InitRequired || (Init && isa<CXXConstructExpr>(Init) &&
3623 !CGF.isTrivialInitializer(Init));
3625 return InitRequired;
3629 /// Emit task_dup function (for initialization of
3630 /// private/firstprivate/lastprivate vars and last_iter flag)
3632 /// void __task_dup_entry(kmp_task_t *task_dst, const kmp_task_t *task_src, int
3634 /// // setup lastprivate flag
3635 /// task_dst->last = lastpriv;
3636 /// // could be constructor calls here...
3639 static llvm::Value *
3640 emitTaskDupFunction(CodeGenModule &CGM, SourceLocation Loc,
3641 const OMPExecutableDirective &D,
3642 QualType KmpTaskTWithPrivatesPtrQTy,
3643 const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3644 const RecordDecl *KmpTaskTQTyRD, QualType SharedsTy,
3645 QualType SharedsPtrTy, const OMPTaskDataTy &Data,
3646 ArrayRef<PrivateDataTy> Privates, bool WithLastIter) {
3647 auto &C = CGM.getContext();
3648 FunctionArgList Args;
3649 ImplicitParamDecl DstArg(C, /*DC=*/nullptr, Loc,
3650 /*Id=*/nullptr, KmpTaskTWithPrivatesPtrQTy);
3651 ImplicitParamDecl SrcArg(C, /*DC=*/nullptr, Loc,
3652 /*Id=*/nullptr, KmpTaskTWithPrivatesPtrQTy);
3653 ImplicitParamDecl LastprivArg(C, /*DC=*/nullptr, Loc,
3654 /*Id=*/nullptr, C.IntTy);
3655 Args.push_back(&DstArg);
3656 Args.push_back(&SrcArg);
3657 Args.push_back(&LastprivArg);
3658 auto &TaskDupFnInfo =
3659 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3660 auto *TaskDupTy = CGM.getTypes().GetFunctionType(TaskDupFnInfo);
3662 llvm::Function::Create(TaskDupTy, llvm::GlobalValue::InternalLinkage,
3663 ".omp_task_dup.", &CGM.getModule());
3664 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, TaskDup, TaskDupFnInfo);
3665 CodeGenFunction CGF(CGM);
3666 CGF.disableDebugInfo();
3667 CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskDup, TaskDupFnInfo, Args);
3669 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3670 CGF.GetAddrOfLocalVar(&DstArg),
3671 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3672 // task_dst->liter = lastpriv;
3674 auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter);
3675 LValue Base = CGF.EmitLValueForField(
3676 TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3677 LValue LILVal = CGF.EmitLValueForField(Base, *LIFI);
3678 llvm::Value *Lastpriv = CGF.EmitLoadOfScalar(
3679 CGF.GetAddrOfLocalVar(&LastprivArg), /*Volatile=*/false, C.IntTy, Loc);
3680 CGF.EmitStoreOfScalar(Lastpriv, LILVal);
3683 // Emit initial values for private copies (if any).
3684 assert(!Privates.empty());
3685 Address KmpTaskSharedsPtr = Address::invalid();
3686 if (!Data.FirstprivateVars.empty()) {
3687 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3688 CGF.GetAddrOfLocalVar(&SrcArg),
3689 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3690 LValue Base = CGF.EmitLValueForField(
3691 TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3692 KmpTaskSharedsPtr = Address(
3693 CGF.EmitLoadOfScalar(CGF.EmitLValueForField(
3694 Base, *std::next(KmpTaskTQTyRD->field_begin(),
3697 CGF.getNaturalTypeAlignment(SharedsTy));
3699 emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, TDBase, KmpTaskTWithPrivatesQTyRD,
3700 SharedsTy, SharedsPtrTy, Data, Privates, /*ForDup=*/true);
3701 CGF.FinishFunction();
3705 /// Checks if destructor function is required to be generated.
3706 /// \return true if cleanups are required, false otherwise.
3708 checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD) {
3709 bool NeedsCleanup = false;
3710 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3711 auto *PrivateRD = cast<RecordDecl>(FI->getType()->getAsTagDecl());
3712 for (auto *FD : PrivateRD->fields()) {
3713 NeedsCleanup = NeedsCleanup || FD->getType().isDestructedType();
3717 return NeedsCleanup;
3720 CGOpenMPRuntime::TaskResultTy
3721 CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc,
3722 const OMPExecutableDirective &D,
3723 llvm::Value *TaskFunction, QualType SharedsTy,
3724 Address Shareds, const OMPTaskDataTy &Data) {
3725 auto &C = CGM.getContext();
3726 llvm::SmallVector<PrivateDataTy, 4> Privates;
3727 // Aggregate privates and sort them by the alignment.
3728 auto I = Data.PrivateCopies.begin();
3729 for (auto *E : Data.PrivateVars) {
3730 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3731 Privates.push_back(std::make_pair(
3733 PrivateHelpersTy(VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3734 /*PrivateElemInit=*/nullptr)));
3737 I = Data.FirstprivateCopies.begin();
3738 auto IElemInitRef = Data.FirstprivateInits.begin();
3739 for (auto *E : Data.FirstprivateVars) {
3740 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3741 Privates.push_back(std::make_pair(
3744 VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3745 cast<VarDecl>(cast<DeclRefExpr>(*IElemInitRef)->getDecl()))));
3749 I = Data.LastprivateCopies.begin();
3750 for (auto *E : Data.LastprivateVars) {
3751 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3752 Privates.push_back(std::make_pair(
3754 PrivateHelpersTy(VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3755 /*PrivateElemInit=*/nullptr)));
3758 llvm::array_pod_sort(Privates.begin(), Privates.end(),
3759 array_pod_sort_comparator);
3760 auto KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
3761 // Build type kmp_routine_entry_t (if not built yet).
3762 emitKmpRoutineEntryT(KmpInt32Ty);
3763 // Build type kmp_task_t (if not built yet).
3764 if (KmpTaskTQTy.isNull()) {
3765 KmpTaskTQTy = C.getRecordType(createKmpTaskTRecordDecl(
3766 CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy));
3768 auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl());
3769 // Build particular struct kmp_task_t for the given task.
3770 auto *KmpTaskTWithPrivatesQTyRD =
3771 createKmpTaskTWithPrivatesRecordDecl(CGM, KmpTaskTQTy, Privates);
3772 auto KmpTaskTWithPrivatesQTy = C.getRecordType(KmpTaskTWithPrivatesQTyRD);
3773 QualType KmpTaskTWithPrivatesPtrQTy =
3774 C.getPointerType(KmpTaskTWithPrivatesQTy);
3775 auto *KmpTaskTWithPrivatesTy = CGF.ConvertType(KmpTaskTWithPrivatesQTy);
3776 auto *KmpTaskTWithPrivatesPtrTy = KmpTaskTWithPrivatesTy->getPointerTo();
3777 auto *KmpTaskTWithPrivatesTySize = CGF.getTypeSize(KmpTaskTWithPrivatesQTy);
3778 QualType SharedsPtrTy = C.getPointerType(SharedsTy);
3780 // Emit initial values for private copies (if any).
3781 llvm::Value *TaskPrivatesMap = nullptr;
3782 auto *TaskPrivatesMapTy =
3783 std::next(cast<llvm::Function>(TaskFunction)->arg_begin(), 3)->getType();
3784 if (!Privates.empty()) {
3785 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3786 TaskPrivatesMap = emitTaskPrivateMappingFunction(
3787 CGM, Loc, Data.PrivateVars, Data.FirstprivateVars, Data.LastprivateVars,
3788 FI->getType(), Privates);
3789 TaskPrivatesMap = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3790 TaskPrivatesMap, TaskPrivatesMapTy);
3792 TaskPrivatesMap = llvm::ConstantPointerNull::get(
3793 cast<llvm::PointerType>(TaskPrivatesMapTy));
3795 // Build a proxy function kmp_int32 .omp_task_entry.(kmp_int32 gtid,
3797 auto *TaskEntry = emitProxyTaskFunction(
3798 CGM, Loc, D.getDirectiveKind(), KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy,
3799 KmpTaskTWithPrivatesQTy, KmpTaskTQTy, SharedsPtrTy, TaskFunction,
3802 // Build call kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid,
3803 // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
3804 // kmp_routine_entry_t *task_entry);
3805 // Task flags. Format is taken from
3806 // http://llvm.org/svn/llvm-project/openmp/trunk/runtime/src/kmp.h,
3807 // description of kmp_tasking_flags struct.
3811 DestructorsFlag = 0x8,
3814 unsigned Flags = Data.Tied ? TiedFlag : 0;
3815 bool NeedsCleanup = false;
3816 if (!Privates.empty()) {
3817 NeedsCleanup = checkDestructorsRequired(KmpTaskTWithPrivatesQTyRD);
3819 Flags = Flags | DestructorsFlag;
3821 if (Data.Priority.getInt())
3822 Flags = Flags | PriorityFlag;
3824 Data.Final.getPointer()
3825 ? CGF.Builder.CreateSelect(Data.Final.getPointer(),
3826 CGF.Builder.getInt32(FinalFlag),
3827 CGF.Builder.getInt32(/*C=*/0))
3828 : CGF.Builder.getInt32(Data.Final.getInt() ? FinalFlag : 0);
3829 TaskFlags = CGF.Builder.CreateOr(TaskFlags, CGF.Builder.getInt32(Flags));
3830 auto *SharedsSize = CGM.getSize(C.getTypeSizeInChars(SharedsTy));
3831 llvm::Value *AllocArgs[] = {emitUpdateLocation(CGF, Loc),
3832 getThreadID(CGF, Loc), TaskFlags,
3833 KmpTaskTWithPrivatesTySize, SharedsSize,
3834 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3835 TaskEntry, KmpRoutineEntryPtrTy)};
3836 auto *NewTask = CGF.EmitRuntimeCall(
3837 createRuntimeFunction(OMPRTL__kmpc_omp_task_alloc), AllocArgs);
3838 auto *NewTaskNewTaskTTy = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3839 NewTask, KmpTaskTWithPrivatesPtrTy);
3840 LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy,
3841 KmpTaskTWithPrivatesQTy);
3843 CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin());
3844 // Fill the data in the resulting kmp_task_t record.
3845 // Copy shareds if there are any.
3846 Address KmpTaskSharedsPtr = Address::invalid();
3847 if (!SharedsTy->getAsStructureType()->getDecl()->field_empty()) {
3849 Address(CGF.EmitLoadOfScalar(
3850 CGF.EmitLValueForField(
3851 TDBase, *std::next(KmpTaskTQTyRD->field_begin(),
3854 CGF.getNaturalTypeAlignment(SharedsTy));
3855 CGF.EmitAggregateCopy(KmpTaskSharedsPtr, Shareds, SharedsTy);
3857 // Emit initial values for private copies (if any).
3858 TaskResultTy Result;
3859 if (!Privates.empty()) {
3860 emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, Base, KmpTaskTWithPrivatesQTyRD,
3861 SharedsTy, SharedsPtrTy, Data, Privates,
3863 if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) &&
3864 (!Data.LastprivateVars.empty() || checkInitIsRequired(CGF, Privates))) {
3865 Result.TaskDupFn = emitTaskDupFunction(
3866 CGM, Loc, D, KmpTaskTWithPrivatesPtrQTy, KmpTaskTWithPrivatesQTyRD,
3867 KmpTaskTQTyRD, SharedsTy, SharedsPtrTy, Data, Privates,
3868 /*WithLastIter=*/!Data.LastprivateVars.empty());
3871 // Fields of union "kmp_cmplrdata_t" for destructors and priority.
3872 enum { Priority = 0, Destructors = 1 };
3873 // Provide pointer to function with destructors for privates.
3874 auto FI = std::next(KmpTaskTQTyRD->field_begin(), Data1);
3875 auto *KmpCmplrdataUD = (*FI)->getType()->getAsUnionType()->getDecl();
3877 llvm::Value *DestructorFn = emitDestructorsFunction(
3878 CGM, Loc, KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy,
3879 KmpTaskTWithPrivatesQTy);
3880 LValue Data1LV = CGF.EmitLValueForField(TDBase, *FI);
3881 LValue DestructorsLV = CGF.EmitLValueForField(
3882 Data1LV, *std::next(KmpCmplrdataUD->field_begin(), Destructors));
3883 CGF.EmitStoreOfScalar(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3884 DestructorFn, KmpRoutineEntryPtrTy),
3888 if (Data.Priority.getInt()) {
3889 LValue Data2LV = CGF.EmitLValueForField(
3890 TDBase, *std::next(KmpTaskTQTyRD->field_begin(), Data2));
3891 LValue PriorityLV = CGF.EmitLValueForField(
3892 Data2LV, *std::next(KmpCmplrdataUD->field_begin(), Priority));
3893 CGF.EmitStoreOfScalar(Data.Priority.getPointer(), PriorityLV);
3895 Result.NewTask = NewTask;
3896 Result.TaskEntry = TaskEntry;
3897 Result.NewTaskNewTaskTTy = NewTaskNewTaskTTy;
3898 Result.TDBase = TDBase;
3899 Result.KmpTaskTQTyRD = KmpTaskTQTyRD;
3903 void CGOpenMPRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc,
3904 const OMPExecutableDirective &D,
3905 llvm::Value *TaskFunction,
3906 QualType SharedsTy, Address Shareds,
3908 const OMPTaskDataTy &Data) {
3909 if (!CGF.HaveInsertPoint())
3912 TaskResultTy Result =
3913 emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data);
3914 llvm::Value *NewTask = Result.NewTask;
3915 llvm::Value *TaskEntry = Result.TaskEntry;
3916 llvm::Value *NewTaskNewTaskTTy = Result.NewTaskNewTaskTTy;
3917 LValue TDBase = Result.TDBase;
3918 RecordDecl *KmpTaskTQTyRD = Result.KmpTaskTQTyRD;
3919 auto &C = CGM.getContext();
3920 // Process list of dependences.
3921 Address DependenciesArray = Address::invalid();
3922 unsigned NumDependencies = Data.Dependences.size();
3923 if (NumDependencies) {
3924 // Dependence kind for RTL.
3925 enum RTLDependenceKindTy { DepIn = 0x01, DepInOut = 0x3 };
3926 enum RTLDependInfoFieldsTy { BaseAddr, Len, Flags };
3927 RecordDecl *KmpDependInfoRD;
3929 C.getIntTypeForBitwidth(C.getTypeSize(C.BoolTy), /*Signed=*/false);
3930 llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy);
3931 if (KmpDependInfoTy.isNull()) {
3932 KmpDependInfoRD = C.buildImplicitRecord("kmp_depend_info");
3933 KmpDependInfoRD->startDefinition();
3934 addFieldToRecordDecl(C, KmpDependInfoRD, C.getIntPtrType());
3935 addFieldToRecordDecl(C, KmpDependInfoRD, C.getSizeType());
3936 addFieldToRecordDecl(C, KmpDependInfoRD, FlagsTy);
3937 KmpDependInfoRD->completeDefinition();
3938 KmpDependInfoTy = C.getRecordType(KmpDependInfoRD);
3940 KmpDependInfoRD = cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
3941 CharUnits DependencySize = C.getTypeSizeInChars(KmpDependInfoTy);
3942 // Define type kmp_depend_info[<Dependences.size()>];
3943 QualType KmpDependInfoArrayTy = C.getConstantArrayType(
3944 KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies),
3945 ArrayType::Normal, /*IndexTypeQuals=*/0);
3946 // kmp_depend_info[<Dependences.size()>] deps;
3948 CGF.CreateMemTemp(KmpDependInfoArrayTy, ".dep.arr.addr");
3949 for (unsigned i = 0; i < NumDependencies; ++i) {
3950 const Expr *E = Data.Dependences[i].second;
3951 auto Addr = CGF.EmitLValue(E);
3953 QualType Ty = E->getType();
3954 if (auto *ASE = dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) {
3956 CGF.EmitOMPArraySectionExpr(ASE, /*LowerBound=*/false);
3957 llvm::Value *UpAddr =
3958 CGF.Builder.CreateConstGEP1_32(UpAddrLVal.getPointer(), /*Idx0=*/1);
3959 llvm::Value *LowIntPtr =
3960 CGF.Builder.CreatePtrToInt(Addr.getPointer(), CGM.SizeTy);
3961 llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGM.SizeTy);
3962 Size = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr);
3964 Size = CGF.getTypeSize(Ty);
3965 auto Base = CGF.MakeAddrLValue(
3966 CGF.Builder.CreateConstArrayGEP(DependenciesArray, i, DependencySize),
3968 // deps[i].base_addr = &<Dependences[i].second>;
3969 auto BaseAddrLVal = CGF.EmitLValueForField(
3970 Base, *std::next(KmpDependInfoRD->field_begin(), BaseAddr));
3971 CGF.EmitStoreOfScalar(
3972 CGF.Builder.CreatePtrToInt(Addr.getPointer(), CGF.IntPtrTy),
3974 // deps[i].len = sizeof(<Dependences[i].second>);
3975 auto LenLVal = CGF.EmitLValueForField(
3976 Base, *std::next(KmpDependInfoRD->field_begin(), Len));
3977 CGF.EmitStoreOfScalar(Size, LenLVal);
3978 // deps[i].flags = <Dependences[i].first>;
3979 RTLDependenceKindTy DepKind;
3980 switch (Data.Dependences[i].first) {
3981 case OMPC_DEPEND_in:
3984 // Out and InOut dependencies must use the same code.
3985 case OMPC_DEPEND_out:
3986 case OMPC_DEPEND_inout:
3989 case OMPC_DEPEND_source:
3990 case OMPC_DEPEND_sink:
3991 case OMPC_DEPEND_unknown:
3992 llvm_unreachable("Unknown task dependence type");
3994 auto FlagsLVal = CGF.EmitLValueForField(
3995 Base, *std::next(KmpDependInfoRD->field_begin(), Flags));
3996 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(LLVMFlagsTy, DepKind),
3999 DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4000 CGF.Builder.CreateStructGEP(DependenciesArray, 0, CharUnits::Zero()),
4004 // NOTE: routine and part_id fields are intialized by __kmpc_omp_task_alloc()
4006 // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid,
4007 // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,
4008 // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) if dependence
4009 // list is not empty
4010 auto *ThreadID = getThreadID(CGF, Loc);
4011 auto *UpLoc = emitUpdateLocation(CGF, Loc);
4012 llvm::Value *TaskArgs[] = { UpLoc, ThreadID, NewTask };
4013 llvm::Value *DepTaskArgs[7];
4014 if (NumDependencies) {
4015 DepTaskArgs[0] = UpLoc;
4016 DepTaskArgs[1] = ThreadID;
4017 DepTaskArgs[2] = NewTask;
4018 DepTaskArgs[3] = CGF.Builder.getInt32(NumDependencies);
4019 DepTaskArgs[4] = DependenciesArray.getPointer();
4020 DepTaskArgs[5] = CGF.Builder.getInt32(0);
4021 DepTaskArgs[6] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4023 auto &&ThenCodeGen = [this, &Data, TDBase, KmpTaskTQTyRD, NumDependencies,
4025 &DepTaskArgs](CodeGenFunction &CGF, PrePostActionTy &) {
4027 auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId);
4028 auto PartIdLVal = CGF.EmitLValueForField(TDBase, *PartIdFI);
4029 CGF.EmitStoreOfScalar(CGF.Builder.getInt32(0), PartIdLVal);
4031 if (NumDependencies) {
4032 CGF.EmitRuntimeCall(
4033 createRuntimeFunction(OMPRTL__kmpc_omp_task_with_deps), DepTaskArgs);
4035 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_task),
4038 // Check if parent region is untied and build return for untied task;
4040 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
4041 Region->emitUntiedSwitch(CGF);
4044 llvm::Value *DepWaitTaskArgs[6];
4045 if (NumDependencies) {
4046 DepWaitTaskArgs[0] = UpLoc;
4047 DepWaitTaskArgs[1] = ThreadID;
4048 DepWaitTaskArgs[2] = CGF.Builder.getInt32(NumDependencies);
4049 DepWaitTaskArgs[3] = DependenciesArray.getPointer();
4050 DepWaitTaskArgs[4] = CGF.Builder.getInt32(0);
4051 DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4053 auto &&ElseCodeGen = [&TaskArgs, ThreadID, NewTaskNewTaskTTy, TaskEntry,
4054 NumDependencies, &DepWaitTaskArgs](CodeGenFunction &CGF,
4055 PrePostActionTy &) {
4056 auto &RT = CGF.CGM.getOpenMPRuntime();
4057 CodeGenFunction::RunCleanupsScope LocalScope(CGF);
4058 // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid,
4059 // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32
4060 // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info
4062 if (NumDependencies)
4063 CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__kmpc_omp_wait_deps),
4065 // Call proxy_task_entry(gtid, new_task);
4066 auto &&CodeGen = [TaskEntry, ThreadID, NewTaskNewTaskTTy](
4067 CodeGenFunction &CGF, PrePostActionTy &Action) {
4069 llvm::Value *OutlinedFnArgs[] = {ThreadID, NewTaskNewTaskTTy};
4070 CGF.EmitCallOrInvoke(TaskEntry, OutlinedFnArgs);
4073 // Build void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid,
4074 // kmp_task_t *new_task);
4075 // Build void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid,
4076 // kmp_task_t *new_task);
4077 RegionCodeGenTy RCG(CodeGen);
4078 CommonActionTy Action(
4079 RT.createRuntimeFunction(OMPRTL__kmpc_omp_task_begin_if0), TaskArgs,
4080 RT.createRuntimeFunction(OMPRTL__kmpc_omp_task_complete_if0), TaskArgs);
4081 RCG.setAction(Action);
4086 emitOMPIfClause(CGF, IfCond, ThenCodeGen, ElseCodeGen);
4088 RegionCodeGenTy ThenRCG(ThenCodeGen);
4093 void CGOpenMPRuntime::emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc,
4094 const OMPLoopDirective &D,
4095 llvm::Value *TaskFunction,
4096 QualType SharedsTy, Address Shareds,
4098 const OMPTaskDataTy &Data) {
4099 if (!CGF.HaveInsertPoint())
4101 TaskResultTy Result =
4102 emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data);
4103 // NOTE: routine and part_id fields are intialized by __kmpc_omp_task_alloc()
4105 // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int
4106 // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int
4107 // sched, kmp_uint64 grainsize, void *task_dup);
4108 llvm::Value *ThreadID = getThreadID(CGF, Loc);
4109 llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc);
4112 IfVal = CGF.Builder.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.IntTy,
4115 IfVal = llvm::ConstantInt::getSigned(CGF.IntTy, /*V=*/1);
4117 LValue LBLVal = CGF.EmitLValueForField(
4119 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound));
4121 cast<VarDecl>(cast<DeclRefExpr>(D.getLowerBoundVariable())->getDecl());
4122 CGF.EmitAnyExprToMem(LBVar->getInit(), LBLVal.getAddress(), LBLVal.getQuals(),
4123 /*IsInitializer=*/true);
4124 LValue UBLVal = CGF.EmitLValueForField(
4126 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound));
4128 cast<VarDecl>(cast<DeclRefExpr>(D.getUpperBoundVariable())->getDecl());
4129 CGF.EmitAnyExprToMem(UBVar->getInit(), UBLVal.getAddress(), UBLVal.getQuals(),
4130 /*IsInitializer=*/true);
4131 LValue StLVal = CGF.EmitLValueForField(
4133 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTStride));
4135 cast<VarDecl>(cast<DeclRefExpr>(D.getStrideVariable())->getDecl());
4136 CGF.EmitAnyExprToMem(StVar->getInit(), StLVal.getAddress(), StLVal.getQuals(),
4137 /*IsInitializer=*/true);
4138 enum { NoSchedule = 0, Grainsize = 1, NumTasks = 2 };
4139 llvm::Value *TaskArgs[] = {
4140 UpLoc, ThreadID, Result.NewTask, IfVal, LBLVal.getPointer(),
4141 UBLVal.getPointer(), CGF.EmitLoadOfScalar(StLVal, SourceLocation()),
4142 llvm::ConstantInt::getSigned(CGF.IntTy, Data.Nogroup ? 1 : 0),
4143 llvm::ConstantInt::getSigned(
4144 CGF.IntTy, Data.Schedule.getPointer()
4145 ? Data.Schedule.getInt() ? NumTasks : Grainsize
4147 Data.Schedule.getPointer()
4148 ? CGF.Builder.CreateIntCast(Data.Schedule.getPointer(), CGF.Int64Ty,
4150 : llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/0),
4152 ? CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Result.TaskDupFn,
4154 : llvm::ConstantPointerNull::get(CGF.VoidPtrTy)};
4155 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_taskloop), TaskArgs);
4158 /// \brief Emit reduction operation for each element of array (required for
4159 /// array sections) LHS op = RHS.
4160 /// \param Type Type of array.
4161 /// \param LHSVar Variable on the left side of the reduction operation
4162 /// (references element of array in original variable).
4163 /// \param RHSVar Variable on the right side of the reduction operation
4164 /// (references element of array in original variable).
4165 /// \param RedOpGen Generator of reduction operation with use of LHSVar and
4167 static void EmitOMPAggregateReduction(
4168 CodeGenFunction &CGF, QualType Type, const VarDecl *LHSVar,
4169 const VarDecl *RHSVar,
4170 const llvm::function_ref<void(CodeGenFunction &CGF, const Expr *,
4171 const Expr *, const Expr *)> &RedOpGen,
4172 const Expr *XExpr = nullptr, const Expr *EExpr = nullptr,
4173 const Expr *UpExpr = nullptr) {
4174 // Perform element-by-element initialization.
4176 Address LHSAddr = CGF.GetAddrOfLocalVar(LHSVar);
4177 Address RHSAddr = CGF.GetAddrOfLocalVar(RHSVar);
4179 // Drill down to the base element type on both arrays.
4180 auto ArrayTy = Type->getAsArrayTypeUnsafe();
4181 auto NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, LHSAddr);
4183 auto RHSBegin = RHSAddr.getPointer();
4184 auto LHSBegin = LHSAddr.getPointer();
4185 // Cast from pointer to array type to pointer to single element.
4186 auto LHSEnd = CGF.Builder.CreateGEP(LHSBegin, NumElements);
4187 // The basic structure here is a while-do loop.
4188 auto BodyBB = CGF.createBasicBlock("omp.arraycpy.body");
4189 auto DoneBB = CGF.createBasicBlock("omp.arraycpy.done");
4191 CGF.Builder.CreateICmpEQ(LHSBegin, LHSEnd, "omp.arraycpy.isempty");
4192 CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
4194 // Enter the loop body, making that address the current address.
4195 auto EntryBB = CGF.Builder.GetInsertBlock();
4196 CGF.EmitBlock(BodyBB);
4198 CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy);
4200 llvm::PHINode *RHSElementPHI = CGF.Builder.CreatePHI(
4201 RHSBegin->getType(), 2, "omp.arraycpy.srcElementPast");
4202 RHSElementPHI->addIncoming(RHSBegin, EntryBB);
4203 Address RHSElementCurrent =
4204 Address(RHSElementPHI,
4205 RHSAddr.getAlignment().alignmentOfArrayElement(ElementSize));
4207 llvm::PHINode *LHSElementPHI = CGF.Builder.CreatePHI(
4208 LHSBegin->getType(), 2, "omp.arraycpy.destElementPast");
4209 LHSElementPHI->addIncoming(LHSBegin, EntryBB);
4210 Address LHSElementCurrent =
4211 Address(LHSElementPHI,
4212 LHSAddr.getAlignment().alignmentOfArrayElement(ElementSize));
4215 CodeGenFunction::OMPPrivateScope Scope(CGF);
4216 Scope.addPrivate(LHSVar, [=]() -> Address { return LHSElementCurrent; });
4217 Scope.addPrivate(RHSVar, [=]() -> Address { return RHSElementCurrent; });
4219 RedOpGen(CGF, XExpr, EExpr, UpExpr);
4220 Scope.ForceCleanup();
4222 // Shift the address forward by one element.
4223 auto LHSElementNext = CGF.Builder.CreateConstGEP1_32(
4224 LHSElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element");
4225 auto RHSElementNext = CGF.Builder.CreateConstGEP1_32(
4226 RHSElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element");
4227 // Check whether we've reached the end.
4229 CGF.Builder.CreateICmpEQ(LHSElementNext, LHSEnd, "omp.arraycpy.done");
4230 CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB);
4231 LHSElementPHI->addIncoming(LHSElementNext, CGF.Builder.GetInsertBlock());
4232 RHSElementPHI->addIncoming(RHSElementNext, CGF.Builder.GetInsertBlock());
4235 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
4238 /// Emit reduction combiner. If the combiner is a simple expression emit it as
4239 /// is, otherwise consider it as combiner of UDR decl and emit it as a call of
4240 /// UDR combiner function.
4241 static void emitReductionCombiner(CodeGenFunction &CGF,
4242 const Expr *ReductionOp) {
4243 if (auto *CE = dyn_cast<CallExpr>(ReductionOp))
4244 if (auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee()))
4246 dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts()))
4247 if (auto *DRD = dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) {
4248 std::pair<llvm::Function *, llvm::Function *> Reduction =
4249 CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD);
4250 RValue Func = RValue::get(Reduction.first);
4251 CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func);
4252 CGF.EmitIgnoredExpr(ReductionOp);
4255 CGF.EmitIgnoredExpr(ReductionOp);
4258 llvm::Value *CGOpenMPRuntime::emitReductionFunction(
4259 CodeGenModule &CGM, llvm::Type *ArgsType, ArrayRef<const Expr *> Privates,
4260 ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
4261 ArrayRef<const Expr *> ReductionOps) {
4262 auto &C = CGM.getContext();
4264 // void reduction_func(void *LHSArg, void *RHSArg);
4265 FunctionArgList Args;
4266 ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, SourceLocation(), /*Id=*/nullptr,
4268 ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, SourceLocation(), /*Id=*/nullptr,
4270 Args.push_back(&LHSArg);
4271 Args.push_back(&RHSArg);
4272 auto &CGFI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
4273 auto *Fn = llvm::Function::Create(
4274 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
4275 ".omp.reduction.reduction_func", &CGM.getModule());
4276 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, Fn, CGFI);
4277 CodeGenFunction CGF(CGM);
4278 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args);
4280 // Dst = (void*[n])(LHSArg);
4281 // Src = (void*[n])(RHSArg);
4282 Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4283 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)),
4284 ArgsType), CGF.getPointerAlign());
4285 Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4286 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)),
4287 ArgsType), CGF.getPointerAlign());
4290 // *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]);
4292 CodeGenFunction::OMPPrivateScope Scope(CGF);
4293 auto IPriv = Privates.begin();
4295 for (unsigned I = 0, E = ReductionOps.size(); I < E; ++I, ++IPriv, ++Idx) {
4296 auto RHSVar = cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl());
4297 Scope.addPrivate(RHSVar, [&]() -> Address {
4298 return emitAddrOfVarFromArray(CGF, RHS, Idx, RHSVar);
4300 auto LHSVar = cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl());
4301 Scope.addPrivate(LHSVar, [&]() -> Address {
4302 return emitAddrOfVarFromArray(CGF, LHS, Idx, LHSVar);
4304 QualType PrivTy = (*IPriv)->getType();
4305 if (PrivTy->isVariablyModifiedType()) {
4306 // Get array size and emit VLA type.
4309 CGF.Builder.CreateConstArrayGEP(LHS, Idx, CGF.getPointerSize());
4310 llvm::Value *Ptr = CGF.Builder.CreateLoad(Elem);
4311 auto *VLA = CGF.getContext().getAsVariableArrayType(PrivTy);
4312 auto *OVE = cast<OpaqueValueExpr>(VLA->getSizeExpr());
4313 CodeGenFunction::OpaqueValueMapping OpaqueMap(
4314 CGF, OVE, RValue::get(CGF.Builder.CreatePtrToInt(Ptr, CGF.SizeTy)));
4315 CGF.EmitVariablyModifiedType(PrivTy);
4319 IPriv = Privates.begin();
4320 auto ILHS = LHSExprs.begin();
4321 auto IRHS = RHSExprs.begin();
4322 for (auto *E : ReductionOps) {
4323 if ((*IPriv)->getType()->isArrayType()) {
4324 // Emit reduction for array section.
4325 auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
4326 auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
4327 EmitOMPAggregateReduction(
4328 CGF, (*IPriv)->getType(), LHSVar, RHSVar,
4329 [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) {
4330 emitReductionCombiner(CGF, E);
4333 // Emit reduction for array subscript or single variable.
4334 emitReductionCombiner(CGF, E);
4339 Scope.ForceCleanup();
4340 CGF.FinishFunction();
4344 void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF,
4345 const Expr *ReductionOp,
4346 const Expr *PrivateRef,
4347 const DeclRefExpr *LHS,
4348 const DeclRefExpr *RHS) {
4349 if (PrivateRef->getType()->isArrayType()) {
4350 // Emit reduction for array section.
4351 auto *LHSVar = cast<VarDecl>(LHS->getDecl());
4352 auto *RHSVar = cast<VarDecl>(RHS->getDecl());
4353 EmitOMPAggregateReduction(
4354 CGF, PrivateRef->getType(), LHSVar, RHSVar,
4355 [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) {
4356 emitReductionCombiner(CGF, ReductionOp);
4359 // Emit reduction for array subscript or single variable.
4360 emitReductionCombiner(CGF, ReductionOp);
4363 void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
4364 ArrayRef<const Expr *> Privates,
4365 ArrayRef<const Expr *> LHSExprs,
4366 ArrayRef<const Expr *> RHSExprs,
4367 ArrayRef<const Expr *> ReductionOps,
4368 ReductionOptionsTy Options) {
4369 if (!CGF.HaveInsertPoint())
4372 bool WithNowait = Options.WithNowait;
4373 bool SimpleReduction = Options.SimpleReduction;
4375 // Next code should be emitted for reduction:
4377 // static kmp_critical_name lock = { 0 };
4379 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
4380 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
4382 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
4383 // *(Type<n>-1*)rhs[<n>-1]);
4387 // void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]};
4388 // switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
4389 // RedList, reduce_func, &<lock>)) {
4392 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
4394 // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
4398 // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
4400 // [__kmpc_end_reduce(<loc>, <gtid>, &<lock>);]
4405 // if SimpleReduction is true, only the next code is generated:
4407 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
4410 auto &C = CGM.getContext();
4412 if (SimpleReduction) {
4413 CodeGenFunction::RunCleanupsScope Scope(CGF);
4414 auto IPriv = Privates.begin();
4415 auto ILHS = LHSExprs.begin();
4416 auto IRHS = RHSExprs.begin();
4417 for (auto *E : ReductionOps) {
4418 emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
4419 cast<DeclRefExpr>(*IRHS));
4427 // 1. Build a list of reduction variables.
4428 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
4429 auto Size = RHSExprs.size();
4430 for (auto *E : Privates) {
4431 if (E->getType()->isVariablyModifiedType())
4432 // Reserve place for array size.
4435 llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
4436 QualType ReductionArrayTy =
4437 C.getConstantArrayType(C.VoidPtrTy, ArraySize, ArrayType::Normal,
4438 /*IndexTypeQuals=*/0);
4439 Address ReductionList =
4440 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
4441 auto IPriv = Privates.begin();
4443 for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
4445 CGF.Builder.CreateConstArrayGEP(ReductionList, Idx, CGF.getPointerSize());
4446 CGF.Builder.CreateStore(
4447 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4448 CGF.EmitLValue(RHSExprs[I]).getPointer(), CGF.VoidPtrTy),
4450 if ((*IPriv)->getType()->isVariablyModifiedType()) {
4451 // Store array size.
4453 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx,
4454 CGF.getPointerSize());
4455 llvm::Value *Size = CGF.Builder.CreateIntCast(
4457 CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
4459 CGF.SizeTy, /*isSigned=*/false);
4460 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
4465 // 2. Emit reduce_func().
4466 auto *ReductionFn = emitReductionFunction(
4467 CGM, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates,
4468 LHSExprs, RHSExprs, ReductionOps);
4470 // 3. Create static kmp_critical_name lock = { 0 };
4471 auto *Lock = getCriticalRegionLock(".reduction");
4473 // 4. Build res = __kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
4474 // RedList, reduce_func, &<lock>);
4475 auto *IdentTLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE);
4476 auto *ThreadId = getThreadID(CGF, Loc);
4477 auto *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
4478 auto *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4479 ReductionList.getPointer(), CGF.VoidPtrTy);
4480 llvm::Value *Args[] = {
4481 IdentTLoc, // ident_t *<loc>
4482 ThreadId, // i32 <gtid>
4483 CGF.Builder.getInt32(RHSExprs.size()), // i32 <n>
4484 ReductionArrayTySize, // size_type sizeof(RedList)
4485 RL, // void *RedList
4486 ReductionFn, // void (*) (void *, void *) <reduce_func>
4487 Lock // kmp_critical_name *&<lock>
4489 auto Res = CGF.EmitRuntimeCall(
4490 createRuntimeFunction(WithNowait ? OMPRTL__kmpc_reduce_nowait
4491 : OMPRTL__kmpc_reduce),
4494 // 5. Build switch(res)
4495 auto *DefaultBB = CGF.createBasicBlock(".omp.reduction.default");
4496 auto *SwInst = CGF.Builder.CreateSwitch(Res, DefaultBB, /*NumCases=*/2);
4500 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
4502 // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
4504 auto *Case1BB = CGF.createBasicBlock(".omp.reduction.case1");
4505 SwInst->addCase(CGF.Builder.getInt32(1), Case1BB);
4506 CGF.EmitBlock(Case1BB);
4508 // Add emission of __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
4509 llvm::Value *EndArgs[] = {
4510 IdentTLoc, // ident_t *<loc>
4511 ThreadId, // i32 <gtid>
4512 Lock // kmp_critical_name *&<lock>
4514 auto &&CodeGen = [&Privates, &LHSExprs, &RHSExprs, &ReductionOps](
4515 CodeGenFunction &CGF, PrePostActionTy &Action) {
4516 auto &RT = CGF.CGM.getOpenMPRuntime();
4517 auto IPriv = Privates.begin();
4518 auto ILHS = LHSExprs.begin();
4519 auto IRHS = RHSExprs.begin();
4520 for (auto *E : ReductionOps) {
4521 RT.emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
4522 cast<DeclRefExpr>(*IRHS));
4528 RegionCodeGenTy RCG(CodeGen);
4529 CommonActionTy Action(
4530 nullptr, llvm::None,
4531 createRuntimeFunction(WithNowait ? OMPRTL__kmpc_end_reduce_nowait
4532 : OMPRTL__kmpc_end_reduce),
4534 RCG.setAction(Action);
4537 CGF.EmitBranch(DefaultBB);
4541 // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
4544 auto *Case2BB = CGF.createBasicBlock(".omp.reduction.case2");
4545 SwInst->addCase(CGF.Builder.getInt32(2), Case2BB);
4546 CGF.EmitBlock(Case2BB);
4548 auto &&AtomicCodeGen = [Loc, &Privates, &LHSExprs, &RHSExprs, &ReductionOps](
4549 CodeGenFunction &CGF, PrePostActionTy &Action) {
4550 auto ILHS = LHSExprs.begin();
4551 auto IRHS = RHSExprs.begin();
4552 auto IPriv = Privates.begin();
4553 for (auto *E : ReductionOps) {
4554 const Expr *XExpr = nullptr;
4555 const Expr *EExpr = nullptr;
4556 const Expr *UpExpr = nullptr;
4557 BinaryOperatorKind BO = BO_Comma;
4558 if (auto *BO = dyn_cast<BinaryOperator>(E)) {
4559 if (BO->getOpcode() == BO_Assign) {
4560 XExpr = BO->getLHS();
4561 UpExpr = BO->getRHS();
4564 // Try to emit update expression as a simple atomic.
4565 auto *RHSExpr = UpExpr;
4567 // Analyze RHS part of the whole expression.
4568 if (auto *ACO = dyn_cast<AbstractConditionalOperator>(
4569 RHSExpr->IgnoreParenImpCasts())) {
4570 // If this is a conditional operator, analyze its condition for
4571 // min/max reduction operator.
4572 RHSExpr = ACO->getCond();
4575 dyn_cast<BinaryOperator>(RHSExpr->IgnoreParenImpCasts())) {
4576 EExpr = BORHS->getRHS();
4577 BO = BORHS->getOpcode();
4581 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
4582 auto &&AtomicRedGen = [BO, VD,
4583 Loc](CodeGenFunction &CGF, const Expr *XExpr,
4584 const Expr *EExpr, const Expr *UpExpr) {
4585 LValue X = CGF.EmitLValue(XExpr);
4588 E = CGF.EmitAnyExpr(EExpr);
4589 CGF.EmitOMPAtomicSimpleUpdateExpr(
4590 X, E, BO, /*IsXLHSInRHSPart=*/true,
4591 llvm::AtomicOrdering::Monotonic, Loc,
4592 [&CGF, UpExpr, VD, Loc](RValue XRValue) {
4593 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4594 PrivateScope.addPrivate(
4595 VD, [&CGF, VD, XRValue, Loc]() -> Address {
4596 Address LHSTemp = CGF.CreateMemTemp(VD->getType());
4597 CGF.emitOMPSimpleStore(
4598 CGF.MakeAddrLValue(LHSTemp, VD->getType()), XRValue,
4599 VD->getType().getNonReferenceType(), Loc);
4602 (void)PrivateScope.Privatize();
4603 return CGF.EmitAnyExpr(UpExpr);
4606 if ((*IPriv)->getType()->isArrayType()) {
4607 // Emit atomic reduction for array section.
4608 auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
4609 EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), VD, RHSVar,
4610 AtomicRedGen, XExpr, EExpr, UpExpr);
4612 // Emit atomic reduction for array subscript or single variable.
4613 AtomicRedGen(CGF, XExpr, EExpr, UpExpr);
4615 // Emit as a critical region.
4616 auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *,
4617 const Expr *, const Expr *) {
4618 auto &RT = CGF.CGM.getOpenMPRuntime();
4619 RT.emitCriticalRegion(
4620 CGF, ".atomic_reduction",
4621 [=](CodeGenFunction &CGF, PrePostActionTy &Action) {
4623 emitReductionCombiner(CGF, E);
4627 if ((*IPriv)->getType()->isArrayType()) {
4628 auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
4629 auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
4630 EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), LHSVar, RHSVar,
4633 CritRedGen(CGF, nullptr, nullptr, nullptr);
4640 RegionCodeGenTy AtomicRCG(AtomicCodeGen);
4642 // Add emission of __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
4643 llvm::Value *EndArgs[] = {
4644 IdentTLoc, // ident_t *<loc>
4645 ThreadId, // i32 <gtid>
4646 Lock // kmp_critical_name *&<lock>
4648 CommonActionTy Action(nullptr, llvm::None,
4649 createRuntimeFunction(OMPRTL__kmpc_end_reduce),
4651 AtomicRCG.setAction(Action);
4656 CGF.EmitBranch(DefaultBB);
4657 CGF.EmitBlock(DefaultBB, /*IsFinished=*/true);
4660 void CGOpenMPRuntime::emitTaskwaitCall(CodeGenFunction &CGF,
4661 SourceLocation Loc) {
4662 if (!CGF.HaveInsertPoint())
4664 // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
4666 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
4667 // Ignore return result until untied tasks are supported.
4668 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_taskwait), Args);
4669 if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
4670 Region->emitUntiedSwitch(CGF);
4673 void CGOpenMPRuntime::emitInlinedDirective(CodeGenFunction &CGF,
4674 OpenMPDirectiveKind InnerKind,
4675 const RegionCodeGenTy &CodeGen,
4677 if (!CGF.HaveInsertPoint())
4679 InlinedOpenMPRegionRAII Region(CGF, CodeGen, InnerKind, HasCancel);
4680 CGF.CapturedStmtInfo->EmitBody(CGF, /*S=*/nullptr);
4691 } // anonymous namespace
4693 static RTCancelKind getCancellationKind(OpenMPDirectiveKind CancelRegion) {
4694 RTCancelKind CancelKind = CancelNoreq;
4695 if (CancelRegion == OMPD_parallel)
4696 CancelKind = CancelParallel;
4697 else if (CancelRegion == OMPD_for)
4698 CancelKind = CancelLoop;
4699 else if (CancelRegion == OMPD_sections)
4700 CancelKind = CancelSections;
4702 assert(CancelRegion == OMPD_taskgroup);
4703 CancelKind = CancelTaskgroup;
4708 void CGOpenMPRuntime::emitCancellationPointCall(
4709 CodeGenFunction &CGF, SourceLocation Loc,
4710 OpenMPDirectiveKind CancelRegion) {
4711 if (!CGF.HaveInsertPoint())
4713 // Build call kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32
4714 // global_tid, kmp_int32 cncl_kind);
4715 if (auto *OMPRegionInfo =
4716 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
4717 // For 'cancellation point taskgroup', the task region info may not have a
4718 // cancel. This may instead happen in another adjacent task.
4719 if (CancelRegion == OMPD_taskgroup || OMPRegionInfo->hasCancel()) {
4720 llvm::Value *Args[] = {
4721 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
4722 CGF.Builder.getInt32(getCancellationKind(CancelRegion))};
4723 // Ignore return result until untied tasks are supported.
4724 auto *Result = CGF.EmitRuntimeCall(
4725 createRuntimeFunction(OMPRTL__kmpc_cancellationpoint), Args);
4726 // if (__kmpc_cancellationpoint()) {
4727 // exit from construct;
4729 auto *ExitBB = CGF.createBasicBlock(".cancel.exit");
4730 auto *ContBB = CGF.createBasicBlock(".cancel.continue");
4731 auto *Cmp = CGF.Builder.CreateIsNotNull(Result);
4732 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
4733 CGF.EmitBlock(ExitBB);
4734 // exit from construct;
4736 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
4737 CGF.EmitBranchThroughCleanup(CancelDest);
4738 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
4743 void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc,
4745 OpenMPDirectiveKind CancelRegion) {
4746 if (!CGF.HaveInsertPoint())
4748 // Build call kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid,
4749 // kmp_int32 cncl_kind);
4750 if (auto *OMPRegionInfo =
4751 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
4752 auto &&ThenGen = [Loc, CancelRegion, OMPRegionInfo](CodeGenFunction &CGF,
4753 PrePostActionTy &) {
4754 auto &RT = CGF.CGM.getOpenMPRuntime();
4755 llvm::Value *Args[] = {
4756 RT.emitUpdateLocation(CGF, Loc), RT.getThreadID(CGF, Loc),
4757 CGF.Builder.getInt32(getCancellationKind(CancelRegion))};
4758 // Ignore return result until untied tasks are supported.
4759 auto *Result = CGF.EmitRuntimeCall(
4760 RT.createRuntimeFunction(OMPRTL__kmpc_cancel), Args);
4761 // if (__kmpc_cancel()) {
4762 // exit from construct;
4764 auto *ExitBB = CGF.createBasicBlock(".cancel.exit");
4765 auto *ContBB = CGF.createBasicBlock(".cancel.continue");
4766 auto *Cmp = CGF.Builder.CreateIsNotNull(Result);
4767 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
4768 CGF.EmitBlock(ExitBB);
4769 // exit from construct;
4771 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
4772 CGF.EmitBranchThroughCleanup(CancelDest);
4773 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
4776 emitOMPIfClause(CGF, IfCond, ThenGen,
4777 [](CodeGenFunction &, PrePostActionTy &) {});
4779 RegionCodeGenTy ThenRCG(ThenGen);
4785 /// \brief Obtain information that uniquely identifies a target entry. This
4786 /// consists of the file and device IDs as well as line number associated with
4787 /// the relevant entry source location.
4788 static void getTargetEntryUniqueInfo(ASTContext &C, SourceLocation Loc,
4789 unsigned &DeviceID, unsigned &FileID,
4790 unsigned &LineNum) {
4792 auto &SM = C.getSourceManager();
4794 // The loc should be always valid and have a file ID (the user cannot use
4795 // #pragma directives in macros)
4797 assert(Loc.isValid() && "Source location is expected to be always valid.");
4798 assert(Loc.isFileID() && "Source location is expected to refer to a file.");
4800 PresumedLoc PLoc = SM.getPresumedLoc(Loc);
4801 assert(PLoc.isValid() && "Source location is expected to be always valid.");
4803 llvm::sys::fs::UniqueID ID;
4804 if (llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID))
4805 llvm_unreachable("Source file with target region no longer exists!");
4807 DeviceID = ID.getDevice();
4808 FileID = ID.getFile();
4809 LineNum = PLoc.getLine();
4812 void CGOpenMPRuntime::emitTargetOutlinedFunction(
4813 const OMPExecutableDirective &D, StringRef ParentName,
4814 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
4815 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
4816 assert(!ParentName.empty() && "Invalid target region parent name!");
4818 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
4819 IsOffloadEntry, CodeGen);
4822 void CGOpenMPRuntime::emitTargetOutlinedFunctionHelper(
4823 const OMPExecutableDirective &D, StringRef ParentName,
4824 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
4825 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
4826 // Create a unique name for the entry function using the source location
4827 // information of the current target region. The name will be something like:
4829 // __omp_offloading_DD_FFFF_PP_lBB
4831 // where DD_FFFF is an ID unique to the file (device and file IDs), PP is the
4832 // mangled name of the function that encloses the target region and BB is the
4833 // line number of the target region.
4838 getTargetEntryUniqueInfo(CGM.getContext(), D.getLocStart(), DeviceID, FileID,
4840 SmallString<64> EntryFnName;
4842 llvm::raw_svector_ostream OS(EntryFnName);
4843 OS << "__omp_offloading" << llvm::format("_%x", DeviceID)
4844 << llvm::format("_%x_", FileID) << ParentName << "_l" << Line;
4847 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
4849 CodeGenFunction CGF(CGM, true);
4850 CGOpenMPTargetRegionInfo CGInfo(CS, CodeGen, EntryFnName);
4851 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
4853 OutlinedFn = CGF.GenerateOpenMPCapturedStmtFunction(CS);
4855 // If this target outline function is not an offload entry, we don't need to
4857 if (!IsOffloadEntry)
4860 // The target region ID is used by the runtime library to identify the current
4861 // target region, so it only has to be unique and not necessarily point to
4862 // anything. It could be the pointer to the outlined function that implements
4863 // the target region, but we aren't using that so that the compiler doesn't
4864 // need to keep that, and could therefore inline the host function if proven
4865 // worthwhile during optimization. In the other hand, if emitting code for the
4866 // device, the ID has to be the function address so that it can retrieved from
4867 // the offloading entry and launched by the runtime library. We also mark the
4868 // outlined function to have external linkage in case we are emitting code for
4869 // the device, because these functions will be entry points to the device.
4871 if (CGM.getLangOpts().OpenMPIsDevice) {
4872 OutlinedFnID = llvm::ConstantExpr::getBitCast(OutlinedFn, CGM.Int8PtrTy);
4873 OutlinedFn->setLinkage(llvm::GlobalValue::ExternalLinkage);
4875 OutlinedFnID = new llvm::GlobalVariable(
4876 CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
4877 llvm::GlobalValue::PrivateLinkage,
4878 llvm::Constant::getNullValue(CGM.Int8Ty), ".omp_offload.region_id");
4880 // Register the information for the entry associated with this target region.
4881 OffloadEntriesInfoManager.registerTargetRegionEntryInfo(
4882 DeviceID, FileID, ParentName, Line, OutlinedFn, OutlinedFnID,
4886 /// discard all CompoundStmts intervening between two constructs
4887 static const Stmt *ignoreCompoundStmts(const Stmt *Body) {
4888 while (auto *CS = dyn_cast_or_null<CompoundStmt>(Body))
4889 Body = CS->body_front();
4894 /// Emit the number of teams for a target directive. Inspect the num_teams
4895 /// clause associated with a teams construct combined or closely nested
4896 /// with the target directive.
4898 /// Emit a team of size one for directives such as 'target parallel' that
4899 /// have no associated teams construct.
4901 /// Otherwise, return nullptr.
4902 static llvm::Value *
4903 emitNumTeamsForTargetDirective(CGOpenMPRuntime &OMPRuntime,
4904 CodeGenFunction &CGF,
4905 const OMPExecutableDirective &D) {
4907 assert(!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the "
4908 "teams directive expected to be "
4909 "emitted only for the host!");
4911 auto &Bld = CGF.Builder;
4913 // If the target directive is combined with a teams directive:
4914 // Return the value in the num_teams clause, if any.
4915 // Otherwise, return 0 to denote the runtime default.
4916 if (isOpenMPTeamsDirective(D.getDirectiveKind())) {
4917 if (const auto *NumTeamsClause = D.getSingleClause<OMPNumTeamsClause>()) {
4918 CodeGenFunction::RunCleanupsScope NumTeamsScope(CGF);
4919 auto NumTeams = CGF.EmitScalarExpr(NumTeamsClause->getNumTeams(),
4920 /*IgnoreResultAssign*/ true);
4921 return Bld.CreateIntCast(NumTeams, CGF.Int32Ty,
4925 // The default value is 0.
4926 return Bld.getInt32(0);
4929 // If the target directive is combined with a parallel directive but not a
4930 // teams directive, start one team.
4931 if (isOpenMPParallelDirective(D.getDirectiveKind()))
4932 return Bld.getInt32(1);
4934 // If the current target region has a teams region enclosed, we need to get
4935 // the number of teams to pass to the runtime function call. This is done
4936 // by generating the expression in a inlined region. This is required because
4937 // the expression is captured in the enclosing target environment when the
4938 // teams directive is not combined with target.
4940 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
4942 // FIXME: Accommodate other combined directives with teams when they become
4944 if (auto *TeamsDir = dyn_cast_or_null<OMPTeamsDirective>(
4945 ignoreCompoundStmts(CS.getCapturedStmt()))) {
4946 if (auto *NTE = TeamsDir->getSingleClause<OMPNumTeamsClause>()) {
4947 CGOpenMPInnerExprInfo CGInfo(CGF, CS);
4948 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
4949 llvm::Value *NumTeams = CGF.EmitScalarExpr(NTE->getNumTeams());
4950 return Bld.CreateIntCast(NumTeams, CGF.Int32Ty,
4954 // If we have an enclosed teams directive but no num_teams clause we use
4955 // the default value 0.
4956 return Bld.getInt32(0);
4959 // No teams associated with the directive.
4963 /// Emit the number of threads for a target directive. Inspect the
4964 /// thread_limit clause associated with a teams construct combined or closely
4965 /// nested with the target directive.
4967 /// Emit the num_threads clause for directives such as 'target parallel' that
4968 /// have no associated teams construct.
4970 /// Otherwise, return nullptr.
4971 static llvm::Value *
4972 emitNumThreadsForTargetDirective(CGOpenMPRuntime &OMPRuntime,
4973 CodeGenFunction &CGF,
4974 const OMPExecutableDirective &D) {
4976 assert(!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the "
4977 "teams directive expected to be "
4978 "emitted only for the host!");
4980 auto &Bld = CGF.Builder;
4983 // If the target directive is combined with a teams directive:
4984 // Return the value in the thread_limit clause, if any.
4986 // If the target directive is combined with a parallel directive:
4987 // Return the value in the num_threads clause, if any.
4989 // If both clauses are set, select the minimum of the two.
4991 // If neither teams or parallel combined directives set the number of threads
4992 // in a team, return 0 to denote the runtime default.
4994 // If this is not a teams directive return nullptr.
4996 if (isOpenMPTeamsDirective(D.getDirectiveKind()) ||
4997 isOpenMPParallelDirective(D.getDirectiveKind())) {
4998 llvm::Value *DefaultThreadLimitVal = Bld.getInt32(0);
4999 llvm::Value *NumThreadsVal = nullptr;
5000 llvm::Value *ThreadLimitVal = nullptr;
5002 if (const auto *ThreadLimitClause =
5003 D.getSingleClause<OMPThreadLimitClause>()) {
5004 CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF);
5005 auto ThreadLimit = CGF.EmitScalarExpr(ThreadLimitClause->getThreadLimit(),
5006 /*IgnoreResultAssign*/ true);
5007 ThreadLimitVal = Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty,
5011 if (const auto *NumThreadsClause =
5012 D.getSingleClause<OMPNumThreadsClause>()) {
5013 CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
5014 llvm::Value *NumThreads =
5015 CGF.EmitScalarExpr(NumThreadsClause->getNumThreads(),
5016 /*IgnoreResultAssign*/ true);
5018 Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*IsSigned=*/true);
5021 // Select the lesser of thread_limit and num_threads.
5023 ThreadLimitVal = ThreadLimitVal
5024 ? Bld.CreateSelect(Bld.CreateICmpSLT(NumThreadsVal,
5026 NumThreadsVal, ThreadLimitVal)
5029 // Set default value passed to the runtime if either teams or a target
5030 // parallel type directive is found but no clause is specified.
5031 if (!ThreadLimitVal)
5032 ThreadLimitVal = DefaultThreadLimitVal;
5034 return ThreadLimitVal;
5037 // If the current target region has a teams region enclosed, we need to get
5038 // the thread limit to pass to the runtime function call. This is done
5039 // by generating the expression in a inlined region. This is required because
5040 // the expression is captured in the enclosing target environment when the
5041 // teams directive is not combined with target.
5043 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
5045 // FIXME: Accommodate other combined directives with teams when they become
5047 if (auto *TeamsDir = dyn_cast_or_null<OMPTeamsDirective>(
5048 ignoreCompoundStmts(CS.getCapturedStmt()))) {
5049 if (auto *TLE = TeamsDir->getSingleClause<OMPThreadLimitClause>()) {
5050 CGOpenMPInnerExprInfo CGInfo(CGF, CS);
5051 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
5052 llvm::Value *ThreadLimit = CGF.EmitScalarExpr(TLE->getThreadLimit());
5053 return CGF.Builder.CreateIntCast(ThreadLimit, CGF.Int32Ty,
5057 // If we have an enclosed teams directive but no thread_limit clause we use
5058 // the default value 0.
5059 return CGF.Builder.getInt32(0);
5062 // No teams associated with the directive.
5067 // \brief Utility to handle information from clauses associated with a given
5068 // construct that use mappable expressions (e.g. 'map' clause, 'to' clause).
5069 // It provides a convenient interface to obtain the information and generate
5070 // code for that information.
5071 class MappableExprsHandler {
5073 /// \brief Values for bit flags used to specify the mapping type for
5075 enum OpenMPOffloadMappingFlags {
5076 /// \brief Allocate memory on the device and move data from host to device.
5078 /// \brief Allocate memory on the device and move data from device to host.
5079 OMP_MAP_FROM = 0x02,
5080 /// \brief Always perform the requested mapping action on the element, even
5081 /// if it was already mapped before.
5082 OMP_MAP_ALWAYS = 0x04,
5083 /// \brief Delete the element from the device environment, ignoring the
5084 /// current reference count associated with the element.
5085 OMP_MAP_DELETE = 0x08,
5086 /// \brief The element being mapped is a pointer, therefore the pointee
5087 /// should be mapped as well.
5088 OMP_MAP_IS_PTR = 0x10,
5089 /// \brief This flags signals that an argument is the first one relating to
5090 /// a map/private clause expression. For some cases a single
5091 /// map/privatization results in multiple arguments passed to the runtime
5093 OMP_MAP_FIRST_REF = 0x20,
5094 /// \brief Signal that the runtime library has to return the device pointer
5095 /// in the current position for the data being mapped.
5096 OMP_MAP_RETURN_PTR = 0x40,
5097 /// \brief This flag signals that the reference being passed is a pointer to
5099 OMP_MAP_PRIVATE_PTR = 0x80,
5100 /// \brief Pass the element to the device by value.
5101 OMP_MAP_PRIVATE_VAL = 0x100,
5104 /// Class that associates information with a base pointer to be passed to the
5105 /// runtime library.
5106 class BasePointerInfo {
5107 /// The base pointer.
5108 llvm::Value *Ptr = nullptr;
5109 /// The base declaration that refers to this device pointer, or null if
5111 const ValueDecl *DevPtrDecl = nullptr;
5114 BasePointerInfo(llvm::Value *Ptr, const ValueDecl *DevPtrDecl = nullptr)
5115 : Ptr(Ptr), DevPtrDecl(DevPtrDecl) {}
5116 llvm::Value *operator*() const { return Ptr; }
5117 const ValueDecl *getDevicePtrDecl() const { return DevPtrDecl; }
5118 void setDevicePtrDecl(const ValueDecl *D) { DevPtrDecl = D; }
5121 typedef SmallVector<BasePointerInfo, 16> MapBaseValuesArrayTy;
5122 typedef SmallVector<llvm::Value *, 16> MapValuesArrayTy;
5123 typedef SmallVector<unsigned, 16> MapFlagsArrayTy;
5126 /// \brief Directive from where the map clauses were extracted.
5127 const OMPExecutableDirective &CurDir;
5129 /// \brief Function the directive is being generated for.
5130 CodeGenFunction &CGF;
5132 /// \brief Set of all first private variables in the current directive.
5133 llvm::SmallPtrSet<const VarDecl *, 8> FirstPrivateDecls;
5135 /// Map between device pointer declarations and their expression components.
5136 /// The key value for declarations in 'this' is null.
5139 SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>>
5142 llvm::Value *getExprTypeSize(const Expr *E) const {
5143 auto ExprTy = E->getType().getCanonicalType();
5145 // Reference types are ignored for mapping purposes.
5146 if (auto *RefTy = ExprTy->getAs<ReferenceType>())
5147 ExprTy = RefTy->getPointeeType().getCanonicalType();
5149 // Given that an array section is considered a built-in type, we need to
5150 // do the calculation based on the length of the section instead of relying
5151 // on CGF.getTypeSize(E->getType()).
5152 if (const auto *OAE = dyn_cast<OMPArraySectionExpr>(E)) {
5153 QualType BaseTy = OMPArraySectionExpr::getBaseOriginalType(
5154 OAE->getBase()->IgnoreParenImpCasts())
5155 .getCanonicalType();
5157 // If there is no length associated with the expression, that means we
5158 // are using the whole length of the base.
5159 if (!OAE->getLength() && OAE->getColonLoc().isValid())
5160 return CGF.getTypeSize(BaseTy);
5162 llvm::Value *ElemSize;
5163 if (auto *PTy = BaseTy->getAs<PointerType>())
5164 ElemSize = CGF.getTypeSize(PTy->getPointeeType().getCanonicalType());
5166 auto *ATy = cast<ArrayType>(BaseTy.getTypePtr());
5167 assert(ATy && "Expecting array type if not a pointer type.");
5168 ElemSize = CGF.getTypeSize(ATy->getElementType().getCanonicalType());
5171 // If we don't have a length at this point, that is because we have an
5172 // array section with a single element.
5173 if (!OAE->getLength())
5176 auto *LengthVal = CGF.EmitScalarExpr(OAE->getLength());
5178 CGF.Builder.CreateIntCast(LengthVal, CGF.SizeTy, /*isSigned=*/false);
5179 return CGF.Builder.CreateNUWMul(LengthVal, ElemSize);
5181 return CGF.getTypeSize(ExprTy);
5184 /// \brief Return the corresponding bits for a given map clause modifier. Add
5185 /// a flag marking the map as a pointer if requested. Add a flag marking the
5186 /// map as the first one of a series of maps that relate to the same map
5188 unsigned getMapTypeBits(OpenMPMapClauseKind MapType,
5189 OpenMPMapClauseKind MapTypeModifier, bool AddPtrFlag,
5190 bool AddIsFirstFlag) const {
5193 case OMPC_MAP_alloc:
5194 case OMPC_MAP_release:
5195 // alloc and release is the default behavior in the runtime library, i.e.
5196 // if we don't pass any bits alloc/release that is what the runtime is
5197 // going to do. Therefore, we don't need to signal anything for these two
5204 Bits = OMP_MAP_FROM;
5206 case OMPC_MAP_tofrom:
5207 Bits = OMP_MAP_TO | OMP_MAP_FROM;
5209 case OMPC_MAP_delete:
5210 Bits = OMP_MAP_DELETE;
5213 llvm_unreachable("Unexpected map type!");
5217 Bits |= OMP_MAP_IS_PTR;
5219 Bits |= OMP_MAP_FIRST_REF;
5220 if (MapTypeModifier == OMPC_MAP_always)
5221 Bits |= OMP_MAP_ALWAYS;
5225 /// \brief Return true if the provided expression is a final array section. A
5226 /// final array section, is one whose length can't be proved to be one.
5227 bool isFinalArraySectionExpression(const Expr *E) const {
5228 auto *OASE = dyn_cast<OMPArraySectionExpr>(E);
5230 // It is not an array section and therefore not a unity-size one.
5234 // An array section with no colon always refer to a single element.
5235 if (OASE->getColonLoc().isInvalid())
5238 auto *Length = OASE->getLength();
5240 // If we don't have a length we have to check if the array has size 1
5241 // for this dimension. Also, we should always expect a length if the
5242 // base type is pointer.
5244 auto BaseQTy = OMPArraySectionExpr::getBaseOriginalType(
5245 OASE->getBase()->IgnoreParenImpCasts())
5246 .getCanonicalType();
5247 if (auto *ATy = dyn_cast<ConstantArrayType>(BaseQTy.getTypePtr()))
5248 return ATy->getSize().getSExtValue() != 1;
5249 // If we don't have a constant dimension length, we have to consider
5250 // the current section as having any size, so it is not necessarily
5251 // unitary. If it happen to be unity size, that's user fault.
5255 // Check if the length evaluates to 1.
5256 llvm::APSInt ConstLength;
5257 if (!Length->EvaluateAsInt(ConstLength, CGF.getContext()))
5258 return true; // Can have more that size 1.
5260 return ConstLength.getSExtValue() != 1;
5263 /// \brief Generate the base pointers, section pointers, sizes and map type
5264 /// bits for the provided map type, map modifier, and expression components.
5265 /// \a IsFirstComponent should be set to true if the provided set of
5266 /// components is the first associated with a capture.
5267 void generateInfoForComponentList(
5268 OpenMPMapClauseKind MapType, OpenMPMapClauseKind MapTypeModifier,
5269 OMPClauseMappableExprCommon::MappableExprComponentListRef Components,
5270 MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers,
5271 MapValuesArrayTy &Sizes, MapFlagsArrayTy &Types,
5272 bool IsFirstComponentList) const {
5274 // The following summarizes what has to be generated for each map and the
5275 // types bellow. The generated information is expressed in this order:
5276 // base pointer, section pointer, size, flags
5277 // (to add to the ones that come from the map type and modifier).
5298 // &d, &d, sizeof(double), noflags
5301 // &i, &i, 100*sizeof(int), noflags
5304 // &i(=&i[0]), &i[1], 23*sizeof(int), noflags
5307 // &p, &p, sizeof(float*), noflags
5310 // p, &p[1], 24*sizeof(float), noflags
5313 // &s, &s, sizeof(S2), noflags
5316 // &s, &(s.i), sizeof(int), noflags
5319 // &s, &(s.i.f), 50*sizeof(int), noflags
5322 // &s, &(s.p), sizeof(double*), noflags
5324 // map(s.p[:22], s.a s.b)
5325 // &s, &(s.p), sizeof(double*), noflags
5326 // &(s.p), &(s.p[0]), 22*sizeof(double), ptr_flag + extra_flag
5329 // &s, &(s.ps), sizeof(S2*), noflags
5332 // &s, &(s.ps), sizeof(S2*), noflags
5333 // &(s.ps), &(s.ps->s.i), sizeof(int), ptr_flag + extra_flag
5336 // &s, &(s.ps), sizeof(S2*), noflags
5337 // &(s.ps), &(s.ps->ps), sizeof(S2*), ptr_flag + extra_flag
5339 // map(s.ps->ps->ps)
5340 // &s, &(s.ps), sizeof(S2*), noflags
5341 // &(s.ps), &(s.ps->ps), sizeof(S2*), ptr_flag + extra_flag
5342 // &(s.ps->ps), &(s.ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5344 // map(s.ps->ps->s.f[:22])
5345 // &s, &(s.ps), sizeof(S2*), noflags
5346 // &(s.ps), &(s.ps->ps), sizeof(S2*), ptr_flag + extra_flag
5347 // &(s.ps->ps), &(s.ps->ps->s.f[0]), 22*sizeof(float), ptr_flag + extra_flag
5350 // &ps, &ps, sizeof(S2*), noflags
5353 // ps, &(ps->i), sizeof(int), noflags
5356 // ps, &(ps->s.f[0]), 50*sizeof(float), noflags
5359 // ps, &(ps->p), sizeof(double*), noflags
5362 // ps, &(ps->p), sizeof(double*), noflags
5363 // &(ps->p), &(ps->p[0]), 22*sizeof(double), ptr_flag + extra_flag
5366 // ps, &(ps->ps), sizeof(S2*), noflags
5369 // ps, &(ps->ps), sizeof(S2*), noflags
5370 // &(ps->ps), &(ps->ps->s.i), sizeof(int), ptr_flag + extra_flag
5373 // ps, &(ps->ps), sizeof(S2*), noflags
5374 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5376 // map(ps->ps->ps->ps)
5377 // ps, &(ps->ps), sizeof(S2*), noflags
5378 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5379 // &(ps->ps->ps), &(ps->ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5381 // map(ps->ps->ps->s.f[:22])
5382 // ps, &(ps->ps), sizeof(S2*), noflags
5383 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5384 // &(ps->ps->ps), &(ps->ps->ps->s.f[0]), 22*sizeof(float), ptr_flag +
5387 // Track if the map information being generated is the first for a capture.
5388 bool IsCaptureFirstInfo = IsFirstComponentList;
5390 // Scan the components from the base to the complete expression.
5391 auto CI = Components.rbegin();
5392 auto CE = Components.rend();
5395 // Track if the map information being generated is the first for a list of
5397 bool IsExpressionFirstInfo = true;
5398 llvm::Value *BP = nullptr;
5400 if (auto *ME = dyn_cast<MemberExpr>(I->getAssociatedExpression())) {
5401 // The base is the 'this' pointer. The content of the pointer is going
5402 // to be the base of the field being mapped.
5403 BP = CGF.EmitScalarExpr(ME->getBase());
5405 // The base is the reference to the variable.
5407 BP = CGF.EmitLValue(cast<DeclRefExpr>(I->getAssociatedExpression()))
5410 // If the variable is a pointer and is being dereferenced (i.e. is not
5411 // the last component), the base has to be the pointer itself, not its
5412 // reference. References are ignored for mapping purposes.
5414 I->getAssociatedDeclaration()->getType().getNonReferenceType();
5415 if (Ty->isAnyPointerType() && std::next(I) != CE) {
5416 auto PtrAddr = CGF.MakeNaturalAlignAddrLValue(BP, Ty);
5417 BP = CGF.EmitLoadOfPointerLValue(PtrAddr.getAddress(),
5418 Ty->castAs<PointerType>())
5421 // We do not need to generate individual map information for the
5422 // pointer, it can be associated with the combined storage.
5427 for (; I != CE; ++I) {
5428 auto Next = std::next(I);
5430 // We need to generate the addresses and sizes if this is the last
5431 // component, if the component is a pointer or if it is an array section
5432 // whose length can't be proved to be one. If this is a pointer, it
5433 // becomes the base address for the following components.
5435 // A final array section, is one whose length can't be proved to be one.
5436 bool IsFinalArraySection =
5437 isFinalArraySectionExpression(I->getAssociatedExpression());
5439 // Get information on whether the element is a pointer. Have to do a
5440 // special treatment for array sections given that they are built-in
5443 dyn_cast<OMPArraySectionExpr>(I->getAssociatedExpression());
5446 OMPArraySectionExpr::getBaseOriginalType(OASE)
5448 ->isAnyPointerType()) ||
5449 I->getAssociatedExpression()->getType()->isAnyPointerType();
5451 if (Next == CE || IsPointer || IsFinalArraySection) {
5453 // If this is not the last component, we expect the pointer to be
5454 // associated with an array expression or member expression.
5455 assert((Next == CE ||
5456 isa<MemberExpr>(Next->getAssociatedExpression()) ||
5457 isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) ||
5458 isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) &&
5459 "Unexpected expression");
5461 auto *LB = CGF.EmitLValue(I->getAssociatedExpression()).getPointer();
5462 auto *Size = getExprTypeSize(I->getAssociatedExpression());
5464 // If we have a member expression and the current component is a
5465 // reference, we have to map the reference too. Whenever we have a
5466 // reference, the section that reference refers to is going to be a
5467 // load instruction from the storage assigned to the reference.
5468 if (isa<MemberExpr>(I->getAssociatedExpression()) &&
5469 I->getAssociatedDeclaration()->getType()->isReferenceType()) {
5470 auto *LI = cast<llvm::LoadInst>(LB);
5471 auto *RefAddr = LI->getPointerOperand();
5473 BasePointers.push_back(BP);
5474 Pointers.push_back(RefAddr);
5475 Sizes.push_back(CGF.getTypeSize(CGF.getContext().VoidPtrTy));
5476 Types.push_back(getMapTypeBits(
5477 /*MapType*/ OMPC_MAP_alloc, /*MapTypeModifier=*/OMPC_MAP_unknown,
5478 !IsExpressionFirstInfo, IsCaptureFirstInfo));
5479 IsExpressionFirstInfo = false;
5480 IsCaptureFirstInfo = false;
5481 // The reference will be the next base address.
5485 BasePointers.push_back(BP);
5486 Pointers.push_back(LB);
5487 Sizes.push_back(Size);
5489 // We need to add a pointer flag for each map that comes from the
5490 // same expression except for the first one. We also need to signal
5491 // this map is the first one that relates with the current capture
5492 // (there is a set of entries for each capture).
5493 Types.push_back(getMapTypeBits(MapType, MapTypeModifier,
5494 !IsExpressionFirstInfo,
5495 IsCaptureFirstInfo));
5497 // If we have a final array section, we are done with this expression.
5498 if (IsFinalArraySection)
5501 // The pointer becomes the base for the next element.
5505 IsExpressionFirstInfo = false;
5506 IsCaptureFirstInfo = false;
5512 /// \brief Return the adjusted map modifiers if the declaration a capture
5513 /// refers to appears in a first-private clause. This is expected to be used
5514 /// only with directives that start with 'target'.
5515 unsigned adjustMapModifiersForPrivateClauses(const CapturedStmt::Capture &Cap,
5516 unsigned CurrentModifiers) {
5517 assert(Cap.capturesVariable() && "Expected capture by reference only!");
5519 // A first private variable captured by reference will use only the
5520 // 'private ptr' and 'map to' flag. Return the right flags if the captured
5521 // declaration is known as first-private in this handler.
5522 if (FirstPrivateDecls.count(Cap.getCapturedVar()))
5523 return MappableExprsHandler::OMP_MAP_PRIVATE_PTR |
5524 MappableExprsHandler::OMP_MAP_TO;
5526 // We didn't modify anything.
5527 return CurrentModifiers;
5531 MappableExprsHandler(const OMPExecutableDirective &Dir, CodeGenFunction &CGF)
5532 : CurDir(Dir), CGF(CGF) {
5533 // Extract firstprivate clause information.
5534 for (const auto *C : Dir.getClausesOfKind<OMPFirstprivateClause>())
5535 for (const auto *D : C->varlists())
5536 FirstPrivateDecls.insert(
5537 cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl());
5538 // Extract device pointer clause information.
5539 for (const auto *C : Dir.getClausesOfKind<OMPIsDevicePtrClause>())
5540 for (auto L : C->component_lists())
5541 DevPointersMap[L.first].push_back(L.second);
5544 /// \brief Generate all the base pointers, section pointers, sizes and map
5545 /// types for the extracted mappable expressions. Also, for each item that
5546 /// relates with a device pointer, a pair of the relevant declaration and
5547 /// index where it occurs is appended to the device pointers info array.
5548 void generateAllInfo(MapBaseValuesArrayTy &BasePointers,
5549 MapValuesArrayTy &Pointers, MapValuesArrayTy &Sizes,
5550 MapFlagsArrayTy &Types) const {
5551 BasePointers.clear();
5557 /// Kind that defines how a device pointer has to be returned.
5558 enum ReturnPointerKind {
5559 // Don't have to return any pointer.
5561 // Pointer is the base of the declaration.
5563 // Pointer is a member of the base declaration - 'this'
5565 // Pointer is a reference and a member of the base declaration - 'this'
5566 RPK_MemberReference,
5568 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
5569 OpenMPMapClauseKind MapType;
5570 OpenMPMapClauseKind MapTypeModifier;
5571 ReturnPointerKind ReturnDevicePointer;
5574 : MapType(OMPC_MAP_unknown), MapTypeModifier(OMPC_MAP_unknown),
5575 ReturnDevicePointer(RPK_None) {}
5577 OMPClauseMappableExprCommon::MappableExprComponentListRef Components,
5578 OpenMPMapClauseKind MapType, OpenMPMapClauseKind MapTypeModifier,
5579 ReturnPointerKind ReturnDevicePointer)
5580 : Components(Components), MapType(MapType),
5581 MapTypeModifier(MapTypeModifier),
5582 ReturnDevicePointer(ReturnDevicePointer) {}
5585 // We have to process the component lists that relate with the same
5586 // declaration in a single chunk so that we can generate the map flags
5587 // correctly. Therefore, we organize all lists in a map.
5588 llvm::DenseMap<const ValueDecl *, SmallVector<MapInfo, 8>> Info;
5590 // Helper function to fill the information map for the different supported
5592 auto &&InfoGen = [&Info](
5594 OMPClauseMappableExprCommon::MappableExprComponentListRef L,
5595 OpenMPMapClauseKind MapType, OpenMPMapClauseKind MapModifier,
5596 MapInfo::ReturnPointerKind ReturnDevicePointer) {
5597 const ValueDecl *VD =
5598 D ? cast<ValueDecl>(D->getCanonicalDecl()) : nullptr;
5599 Info[VD].push_back({L, MapType, MapModifier, ReturnDevicePointer});
5602 // FIXME: MSVC 2013 seems to require this-> to find member CurDir.
5603 for (auto *C : this->CurDir.getClausesOfKind<OMPMapClause>())
5604 for (auto L : C->component_lists())
5605 InfoGen(L.first, L.second, C->getMapType(), C->getMapTypeModifier(),
5607 for (auto *C : this->CurDir.getClausesOfKind<OMPToClause>())
5608 for (auto L : C->component_lists())
5609 InfoGen(L.first, L.second, OMPC_MAP_to, OMPC_MAP_unknown,
5611 for (auto *C : this->CurDir.getClausesOfKind<OMPFromClause>())
5612 for (auto L : C->component_lists())
5613 InfoGen(L.first, L.second, OMPC_MAP_from, OMPC_MAP_unknown,
5616 // Look at the use_device_ptr clause information and mark the existing map
5617 // entries as such. If there is no map information for an entry in the
5618 // use_device_ptr list, we create one with map type 'alloc' and zero size
5619 // section. It is the user fault if that was not mapped before.
5620 // FIXME: MSVC 2013 seems to require this-> to find member CurDir.
5621 for (auto *C : this->CurDir.getClausesOfKind<OMPUseDevicePtrClause>())
5622 for (auto L : C->component_lists()) {
5623 assert(!L.second.empty() && "Not expecting empty list of components!");
5624 const ValueDecl *VD = L.second.back().getAssociatedDeclaration();
5625 VD = cast<ValueDecl>(VD->getCanonicalDecl());
5626 auto *IE = L.second.back().getAssociatedExpression();
5627 // If the first component is a member expression, we have to look into
5628 // 'this', which maps to null in the map of map information. Otherwise
5629 // look directly for the information.
5630 auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD);
5632 // We potentially have map information for this declaration already.
5633 // Look for the first set of components that refer to it.
5634 if (It != Info.end()) {
5635 auto CI = std::find_if(
5636 It->second.begin(), It->second.end(), [VD](const MapInfo &MI) {
5637 return MI.Components.back().getAssociatedDeclaration() == VD;
5639 // If we found a map entry, signal that the pointer has to be returned
5640 // and move on to the next declaration.
5641 if (CI != It->second.end()) {
5642 CI->ReturnDevicePointer = isa<MemberExpr>(IE)
5643 ? (VD->getType()->isReferenceType()
5644 ? MapInfo::RPK_MemberReference
5645 : MapInfo::RPK_Member)
5646 : MapInfo::RPK_Base;
5651 // We didn't find any match in our map information - generate a zero
5652 // size array section.
5653 // FIXME: MSVC 2013 seems to require this-> to find member CGF.
5656 .EmitLoadOfLValue(this->CGF.EmitLValue(IE), SourceLocation())
5658 BasePointers.push_back({Ptr, VD});
5659 Pointers.push_back(Ptr);
5660 Sizes.push_back(llvm::Constant::getNullValue(this->CGF.SizeTy));
5661 Types.push_back(OMP_MAP_RETURN_PTR | OMP_MAP_FIRST_REF);
5664 for (auto &M : Info) {
5665 // We need to know when we generate information for the first component
5666 // associated with a capture, because the mapping flags depend on it.
5667 bool IsFirstComponentList = true;
5668 for (MapInfo &L : M.second) {
5669 assert(!L.Components.empty() &&
5670 "Not expecting declaration with no component lists.");
5672 // Remember the current base pointer index.
5673 unsigned CurrentBasePointersIdx = BasePointers.size();
5674 // FIXME: MSVC 2013 seems to require this-> to find the member method.
5675 this->generateInfoForComponentList(L.MapType, L.MapTypeModifier,
5676 L.Components, BasePointers, Pointers,
5677 Sizes, Types, IsFirstComponentList);
5679 // If this entry relates with a device pointer, set the relevant
5680 // declaration and add the 'return pointer' flag.
5681 if (IsFirstComponentList &&
5682 L.ReturnDevicePointer != MapInfo::RPK_None) {
5683 // If the pointer is not the base of the map, we need to skip the
5684 // base. If it is a reference in a member field, we also need to skip
5685 // the map of the reference.
5686 if (L.ReturnDevicePointer != MapInfo::RPK_Base) {
5687 ++CurrentBasePointersIdx;
5688 if (L.ReturnDevicePointer == MapInfo::RPK_MemberReference)
5689 ++CurrentBasePointersIdx;
5691 assert(BasePointers.size() > CurrentBasePointersIdx &&
5692 "Unexpected number of mapped base pointers.");
5694 auto *RelevantVD = L.Components.back().getAssociatedDeclaration();
5695 assert(RelevantVD &&
5696 "No relevant declaration related with device pointer??");
5698 BasePointers[CurrentBasePointersIdx].setDevicePtrDecl(RelevantVD);
5699 Types[CurrentBasePointersIdx] |= OMP_MAP_RETURN_PTR;
5701 IsFirstComponentList = false;
5706 /// \brief Generate the base pointers, section pointers, sizes and map types
5707 /// associated to a given capture.
5708 void generateInfoForCapture(const CapturedStmt::Capture *Cap,
5710 MapBaseValuesArrayTy &BasePointers,
5711 MapValuesArrayTy &Pointers,
5712 MapValuesArrayTy &Sizes,
5713 MapFlagsArrayTy &Types) const {
5714 assert(!Cap->capturesVariableArrayType() &&
5715 "Not expecting to generate map info for a variable array type!");
5717 BasePointers.clear();
5722 // We need to know when we generating information for the first component
5723 // associated with a capture, because the mapping flags depend on it.
5724 bool IsFirstComponentList = true;
5726 const ValueDecl *VD =
5729 : cast<ValueDecl>(Cap->getCapturedVar()->getCanonicalDecl());
5731 // If this declaration appears in a is_device_ptr clause we just have to
5732 // pass the pointer by value. If it is a reference to a declaration, we just
5733 // pass its value, otherwise, if it is a member expression, we need to map
5736 auto It = DevPointersMap.find(VD);
5737 if (It != DevPointersMap.end()) {
5738 for (auto L : It->second) {
5739 generateInfoForComponentList(
5740 /*MapType=*/OMPC_MAP_to, /*MapTypeModifier=*/OMPC_MAP_unknown, L,
5741 BasePointers, Pointers, Sizes, Types, IsFirstComponentList);
5742 IsFirstComponentList = false;
5746 } else if (DevPointersMap.count(VD)) {
5747 BasePointers.push_back({Arg, VD});
5748 Pointers.push_back(Arg);
5749 Sizes.push_back(CGF.getTypeSize(CGF.getContext().VoidPtrTy));
5750 Types.push_back(OMP_MAP_PRIVATE_VAL | OMP_MAP_FIRST_REF);
5754 // FIXME: MSVC 2013 seems to require this-> to find member CurDir.
5755 for (auto *C : this->CurDir.getClausesOfKind<OMPMapClause>())
5756 for (auto L : C->decl_component_lists(VD)) {
5757 assert(L.first == VD &&
5758 "We got information for the wrong declaration??");
5759 assert(!L.second.empty() &&
5760 "Not expecting declaration with no component lists.");
5761 generateInfoForComponentList(C->getMapType(), C->getMapTypeModifier(),
5762 L.second, BasePointers, Pointers, Sizes,
5763 Types, IsFirstComponentList);
5764 IsFirstComponentList = false;
5770 /// \brief Generate the default map information for a given capture \a CI,
5771 /// record field declaration \a RI and captured value \a CV.
5772 void generateDefaultMapInfo(const CapturedStmt::Capture &CI,
5773 const FieldDecl &RI, llvm::Value *CV,
5774 MapBaseValuesArrayTy &CurBasePointers,
5775 MapValuesArrayTy &CurPointers,
5776 MapValuesArrayTy &CurSizes,
5777 MapFlagsArrayTy &CurMapTypes) {
5779 // Do the default mapping.
5780 if (CI.capturesThis()) {
5781 CurBasePointers.push_back(CV);
5782 CurPointers.push_back(CV);
5783 const PointerType *PtrTy = cast<PointerType>(RI.getType().getTypePtr());
5784 CurSizes.push_back(CGF.getTypeSize(PtrTy->getPointeeType()));
5785 // Default map type.
5786 CurMapTypes.push_back(OMP_MAP_TO | OMP_MAP_FROM);
5787 } else if (CI.capturesVariableByCopy()) {
5788 CurBasePointers.push_back(CV);
5789 CurPointers.push_back(CV);
5790 if (!RI.getType()->isAnyPointerType()) {
5791 // We have to signal to the runtime captures passed by value that are
5793 CurMapTypes.push_back(OMP_MAP_PRIVATE_VAL);
5794 CurSizes.push_back(CGF.getTypeSize(RI.getType()));
5796 // Pointers are implicitly mapped with a zero size and no flags
5797 // (other than first map that is added for all implicit maps).
5798 CurMapTypes.push_back(0u);
5799 CurSizes.push_back(llvm::Constant::getNullValue(CGF.SizeTy));
5802 assert(CI.capturesVariable() && "Expected captured reference.");
5803 CurBasePointers.push_back(CV);
5804 CurPointers.push_back(CV);
5806 const ReferenceType *PtrTy =
5807 cast<ReferenceType>(RI.getType().getTypePtr());
5808 QualType ElementType = PtrTy->getPointeeType();
5809 CurSizes.push_back(CGF.getTypeSize(ElementType));
5810 // The default map type for a scalar/complex type is 'to' because by
5811 // default the value doesn't have to be retrieved. For an aggregate
5812 // type, the default is 'tofrom'.
5813 CurMapTypes.push_back(ElementType->isAggregateType()
5814 ? (OMP_MAP_TO | OMP_MAP_FROM)
5817 // If we have a capture by reference we may need to add the private
5818 // pointer flag if the base declaration shows in some first-private
5820 CurMapTypes.back() =
5821 adjustMapModifiersForPrivateClauses(CI, CurMapTypes.back());
5823 // Every default map produces a single argument, so, it is always the
5825 CurMapTypes.back() |= OMP_MAP_FIRST_REF;
5829 enum OpenMPOffloadingReservedDeviceIDs {
5830 /// \brief Device ID if the device was not defined, runtime should get it
5831 /// from environment variables in the spec.
5832 OMP_DEVICEID_UNDEF = -1,
5834 } // anonymous namespace
5836 /// \brief Emit the arrays used to pass the captures and map information to the
5837 /// offloading runtime library. If there is no map or capture information,
5838 /// return nullptr by reference.
5840 emitOffloadingArrays(CodeGenFunction &CGF,
5841 MappableExprsHandler::MapBaseValuesArrayTy &BasePointers,
5842 MappableExprsHandler::MapValuesArrayTy &Pointers,
5843 MappableExprsHandler::MapValuesArrayTy &Sizes,
5844 MappableExprsHandler::MapFlagsArrayTy &MapTypes,
5845 CGOpenMPRuntime::TargetDataInfo &Info) {
5846 auto &CGM = CGF.CGM;
5847 auto &Ctx = CGF.getContext();
5849 // Reset the array information.
5850 Info.clearArrayInfo();
5851 Info.NumberOfPtrs = BasePointers.size();
5853 if (Info.NumberOfPtrs) {
5854 // Detect if we have any capture size requiring runtime evaluation of the
5855 // size so that a constant array could be eventually used.
5856 bool hasRuntimeEvaluationCaptureSize = false;
5857 for (auto *S : Sizes)
5858 if (!isa<llvm::Constant>(S)) {
5859 hasRuntimeEvaluationCaptureSize = true;
5863 llvm::APInt PointerNumAP(32, Info.NumberOfPtrs, /*isSigned=*/true);
5864 QualType PointerArrayType =
5865 Ctx.getConstantArrayType(Ctx.VoidPtrTy, PointerNumAP, ArrayType::Normal,
5866 /*IndexTypeQuals=*/0);
5868 Info.BasePointersArray =
5869 CGF.CreateMemTemp(PointerArrayType, ".offload_baseptrs").getPointer();
5870 Info.PointersArray =
5871 CGF.CreateMemTemp(PointerArrayType, ".offload_ptrs").getPointer();
5873 // If we don't have any VLA types or other types that require runtime
5874 // evaluation, we can use a constant array for the map sizes, otherwise we
5875 // need to fill up the arrays as we do for the pointers.
5876 if (hasRuntimeEvaluationCaptureSize) {
5877 QualType SizeArrayType = Ctx.getConstantArrayType(
5878 Ctx.getSizeType(), PointerNumAP, ArrayType::Normal,
5879 /*IndexTypeQuals=*/0);
5881 CGF.CreateMemTemp(SizeArrayType, ".offload_sizes").getPointer();
5883 // We expect all the sizes to be constant, so we collect them to create
5884 // a constant array.
5885 SmallVector<llvm::Constant *, 16> ConstSizes;
5886 for (auto S : Sizes)
5887 ConstSizes.push_back(cast<llvm::Constant>(S));
5889 auto *SizesArrayInit = llvm::ConstantArray::get(
5890 llvm::ArrayType::get(CGM.SizeTy, ConstSizes.size()), ConstSizes);
5891 auto *SizesArrayGbl = new llvm::GlobalVariable(
5892 CGM.getModule(), SizesArrayInit->getType(),
5893 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage,
5894 SizesArrayInit, ".offload_sizes");
5895 SizesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
5896 Info.SizesArray = SizesArrayGbl;
5899 // The map types are always constant so we don't need to generate code to
5900 // fill arrays. Instead, we create an array constant.
5901 llvm::Constant *MapTypesArrayInit =
5902 llvm::ConstantDataArray::get(CGF.Builder.getContext(), MapTypes);
5903 auto *MapTypesArrayGbl = new llvm::GlobalVariable(
5904 CGM.getModule(), MapTypesArrayInit->getType(),
5905 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage,
5906 MapTypesArrayInit, ".offload_maptypes");
5907 MapTypesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
5908 Info.MapTypesArray = MapTypesArrayGbl;
5910 for (unsigned i = 0; i < Info.NumberOfPtrs; ++i) {
5911 llvm::Value *BPVal = *BasePointers[i];
5912 if (BPVal->getType()->isPointerTy())
5913 BPVal = CGF.Builder.CreateBitCast(BPVal, CGM.VoidPtrTy);
5915 assert(BPVal->getType()->isIntegerTy() &&
5916 "If not a pointer, the value type must be an integer.");
5917 BPVal = CGF.Builder.CreateIntToPtr(BPVal, CGM.VoidPtrTy);
5919 llvm::Value *BP = CGF.Builder.CreateConstInBoundsGEP2_32(
5920 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5921 Info.BasePointersArray, 0, i);
5922 Address BPAddr(BP, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy));
5923 CGF.Builder.CreateStore(BPVal, BPAddr);
5925 if (Info.requiresDevicePointerInfo())
5926 if (auto *DevVD = BasePointers[i].getDevicePtrDecl())
5927 Info.CaptureDeviceAddrMap.insert(std::make_pair(DevVD, BPAddr));
5929 llvm::Value *PVal = Pointers[i];
5930 if (PVal->getType()->isPointerTy())
5931 PVal = CGF.Builder.CreateBitCast(PVal, CGM.VoidPtrTy);
5933 assert(PVal->getType()->isIntegerTy() &&
5934 "If not a pointer, the value type must be an integer.");
5935 PVal = CGF.Builder.CreateIntToPtr(PVal, CGM.VoidPtrTy);
5937 llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32(
5938 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5939 Info.PointersArray, 0, i);
5940 Address PAddr(P, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy));
5941 CGF.Builder.CreateStore(PVal, PAddr);
5943 if (hasRuntimeEvaluationCaptureSize) {
5944 llvm::Value *S = CGF.Builder.CreateConstInBoundsGEP2_32(
5945 llvm::ArrayType::get(CGM.SizeTy, Info.NumberOfPtrs),
5949 Address SAddr(S, Ctx.getTypeAlignInChars(Ctx.getSizeType()));
5950 CGF.Builder.CreateStore(
5951 CGF.Builder.CreateIntCast(Sizes[i], CGM.SizeTy, /*isSigned=*/true),
5957 /// \brief Emit the arguments to be passed to the runtime library based on the
5958 /// arrays of pointers, sizes and map types.
5959 static void emitOffloadingArraysArgument(
5960 CodeGenFunction &CGF, llvm::Value *&BasePointersArrayArg,
5961 llvm::Value *&PointersArrayArg, llvm::Value *&SizesArrayArg,
5962 llvm::Value *&MapTypesArrayArg, CGOpenMPRuntime::TargetDataInfo &Info) {
5963 auto &CGM = CGF.CGM;
5964 if (Info.NumberOfPtrs) {
5965 BasePointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5966 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5967 Info.BasePointersArray,
5968 /*Idx0=*/0, /*Idx1=*/0);
5969 PointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5970 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5974 SizesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5975 llvm::ArrayType::get(CGM.SizeTy, Info.NumberOfPtrs), Info.SizesArray,
5976 /*Idx0=*/0, /*Idx1=*/0);
5977 MapTypesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5978 llvm::ArrayType::get(CGM.Int32Ty, Info.NumberOfPtrs),
5983 BasePointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy);
5984 PointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy);
5985 SizesArrayArg = llvm::ConstantPointerNull::get(CGM.SizeTy->getPointerTo());
5987 llvm::ConstantPointerNull::get(CGM.Int32Ty->getPointerTo());
5991 void CGOpenMPRuntime::emitTargetCall(CodeGenFunction &CGF,
5992 const OMPExecutableDirective &D,
5993 llvm::Value *OutlinedFn,
5994 llvm::Value *OutlinedFnID,
5995 const Expr *IfCond, const Expr *Device,
5996 ArrayRef<llvm::Value *> CapturedVars) {
5997 if (!CGF.HaveInsertPoint())
6000 assert(OutlinedFn && "Invalid outlined function!");
6002 auto &Ctx = CGF.getContext();
6004 // Fill up the arrays with all the captured variables.
6005 MappableExprsHandler::MapValuesArrayTy KernelArgs;
6006 MappableExprsHandler::MapBaseValuesArrayTy BasePointers;
6007 MappableExprsHandler::MapValuesArrayTy Pointers;
6008 MappableExprsHandler::MapValuesArrayTy Sizes;
6009 MappableExprsHandler::MapFlagsArrayTy MapTypes;
6011 MappableExprsHandler::MapBaseValuesArrayTy CurBasePointers;
6012 MappableExprsHandler::MapValuesArrayTy CurPointers;
6013 MappableExprsHandler::MapValuesArrayTy CurSizes;
6014 MappableExprsHandler::MapFlagsArrayTy CurMapTypes;
6016 // Get mappable expression information.
6017 MappableExprsHandler MEHandler(D, CGF);
6019 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
6020 auto RI = CS.getCapturedRecordDecl()->field_begin();
6021 auto CV = CapturedVars.begin();
6022 for (CapturedStmt::const_capture_iterator CI = CS.capture_begin(),
6023 CE = CS.capture_end();
6024 CI != CE; ++CI, ++RI, ++CV) {
6028 CurBasePointers.clear();
6029 CurPointers.clear();
6031 CurMapTypes.clear();
6033 // VLA sizes are passed to the outlined region by copy and do not have map
6034 // information associated.
6035 if (CI->capturesVariableArrayType()) {
6036 CurBasePointers.push_back(*CV);
6037 CurPointers.push_back(*CV);
6038 CurSizes.push_back(CGF.getTypeSize(RI->getType()));
6039 // Copy to the device as an argument. No need to retrieve it.
6040 CurMapTypes.push_back(MappableExprsHandler::OMP_MAP_PRIVATE_VAL |
6041 MappableExprsHandler::OMP_MAP_FIRST_REF);
6043 // If we have any information in the map clause, we use it, otherwise we
6044 // just do a default mapping.
6045 MEHandler.generateInfoForCapture(CI, *CV, CurBasePointers, CurPointers,
6046 CurSizes, CurMapTypes);
6047 if (CurBasePointers.empty())
6048 MEHandler.generateDefaultMapInfo(*CI, **RI, *CV, CurBasePointers,
6049 CurPointers, CurSizes, CurMapTypes);
6051 // We expect to have at least an element of information for this capture.
6052 assert(!CurBasePointers.empty() && "Non-existing map pointer for capture!");
6053 assert(CurBasePointers.size() == CurPointers.size() &&
6054 CurBasePointers.size() == CurSizes.size() &&
6055 CurBasePointers.size() == CurMapTypes.size() &&
6056 "Inconsistent map information sizes!");
6058 // The kernel args are always the first elements of the base pointers
6059 // associated with a capture.
6060 KernelArgs.push_back(*CurBasePointers.front());
6061 // We need to append the results of this capture to what we already have.
6062 BasePointers.append(CurBasePointers.begin(), CurBasePointers.end());
6063 Pointers.append(CurPointers.begin(), CurPointers.end());
6064 Sizes.append(CurSizes.begin(), CurSizes.end());
6065 MapTypes.append(CurMapTypes.begin(), CurMapTypes.end());
6068 // Keep track on whether the host function has to be executed.
6069 auto OffloadErrorQType =
6070 Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true);
6071 auto OffloadError = CGF.MakeAddrLValue(
6072 CGF.CreateMemTemp(OffloadErrorQType, ".run_host_version"),
6074 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty),
6077 // Fill up the pointer arrays and transfer execution to the device.
6078 auto &&ThenGen = [&BasePointers, &Pointers, &Sizes, &MapTypes, Device,
6079 OutlinedFnID, OffloadError,
6080 &D](CodeGenFunction &CGF, PrePostActionTy &) {
6081 auto &RT = CGF.CGM.getOpenMPRuntime();
6082 // Emit the offloading arrays.
6083 TargetDataInfo Info;
6084 emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info);
6085 emitOffloadingArraysArgument(CGF, Info.BasePointersArray,
6086 Info.PointersArray, Info.SizesArray,
6087 Info.MapTypesArray, Info);
6089 // On top of the arrays that were filled up, the target offloading call
6090 // takes as arguments the device id as well as the host pointer. The host
6091 // pointer is used by the runtime library to identify the current target
6092 // region, so it only has to be unique and not necessarily point to
6093 // anything. It could be the pointer to the outlined function that
6094 // implements the target region, but we aren't using that so that the
6095 // compiler doesn't need to keep that, and could therefore inline the host
6096 // function if proven worthwhile during optimization.
6098 // From this point on, we need to have an ID of the target region defined.
6099 assert(OutlinedFnID && "Invalid outlined function ID!");
6101 // Emit device ID if any.
6102 llvm::Value *DeviceID;
6104 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6105 CGF.Int32Ty, /*isSigned=*/true);
6107 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6109 // Emit the number of elements in the offloading arrays.
6110 llvm::Value *PointerNum = CGF.Builder.getInt32(BasePointers.size());
6112 // Return value of the runtime offloading call.
6113 llvm::Value *Return;
6115 auto *NumTeams = emitNumTeamsForTargetDirective(RT, CGF, D);
6116 auto *NumThreads = emitNumThreadsForTargetDirective(RT, CGF, D);
6118 // The target region is an outlined function launched by the runtime
6119 // via calls __tgt_target() or __tgt_target_teams().
6121 // __tgt_target() launches a target region with one team and one thread,
6122 // executing a serial region. This master thread may in turn launch
6123 // more threads within its team upon encountering a parallel region,
6124 // however, no additional teams can be launched on the device.
6126 // __tgt_target_teams() launches a target region with one or more teams,
6127 // each with one or more threads. This call is required for target
6128 // constructs such as:
6130 // 'target' / 'teams'
6131 // 'target teams distribute parallel for'
6132 // 'target parallel'
6135 // Note that on the host and CPU targets, the runtime implementation of
6136 // these calls simply call the outlined function without forking threads.
6137 // The outlined functions themselves have runtime calls to
6138 // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by
6139 // the compiler in emitTeamsCall() and emitParallelCall().
6141 // In contrast, on the NVPTX target, the implementation of
6142 // __tgt_target_teams() launches a GPU kernel with the requested number
6143 // of teams and threads so no additional calls to the runtime are required.
6145 // If we have NumTeams defined this means that we have an enclosed teams
6146 // region. Therefore we also expect to have NumThreads defined. These two
6147 // values should be defined in the presence of a teams directive,
6148 // regardless of having any clauses associated. If the user is using teams
6149 // but no clauses, these two values will be the default that should be
6150 // passed to the runtime library - a 32-bit integer with the value zero.
6151 assert(NumThreads && "Thread limit expression should be available along "
6152 "with number of teams.");
6153 llvm::Value *OffloadingArgs[] = {
6154 DeviceID, OutlinedFnID,
6155 PointerNum, Info.BasePointersArray,
6156 Info.PointersArray, Info.SizesArray,
6157 Info.MapTypesArray, NumTeams,
6159 Return = CGF.EmitRuntimeCall(
6160 RT.createRuntimeFunction(OMPRTL__tgt_target_teams), OffloadingArgs);
6162 llvm::Value *OffloadingArgs[] = {
6163 DeviceID, OutlinedFnID,
6164 PointerNum, Info.BasePointersArray,
6165 Info.PointersArray, Info.SizesArray,
6166 Info.MapTypesArray};
6167 Return = CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__tgt_target),
6171 CGF.EmitStoreOfScalar(Return, OffloadError);
6174 // Notify that the host version must be executed.
6175 auto &&ElseGen = [OffloadError](CodeGenFunction &CGF, PrePostActionTy &) {
6176 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/-1u),
6180 // If we have a target function ID it means that we need to support
6181 // offloading, otherwise, just execute on the host. We need to execute on host
6182 // regardless of the conditional in the if clause if, e.g., the user do not
6183 // specify target triples.
6186 emitOMPIfClause(CGF, IfCond, ThenGen, ElseGen);
6188 RegionCodeGenTy ThenRCG(ThenGen);
6192 RegionCodeGenTy ElseRCG(ElseGen);
6196 // Check the error code and execute the host version if required.
6197 auto OffloadFailedBlock = CGF.createBasicBlock("omp_offload.failed");
6198 auto OffloadContBlock = CGF.createBasicBlock("omp_offload.cont");
6199 auto OffloadErrorVal = CGF.EmitLoadOfScalar(OffloadError, SourceLocation());
6200 auto Failed = CGF.Builder.CreateIsNotNull(OffloadErrorVal);
6201 CGF.Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock);
6203 CGF.EmitBlock(OffloadFailedBlock);
6204 CGF.Builder.CreateCall(OutlinedFn, KernelArgs);
6205 CGF.EmitBranch(OffloadContBlock);
6207 CGF.EmitBlock(OffloadContBlock, /*IsFinished=*/true);
6210 void CGOpenMPRuntime::scanForTargetRegionsFunctions(const Stmt *S,
6211 StringRef ParentName) {
6215 // Codegen OMP target directives that offload compute to the device.
6216 bool requiresDeviceCodegen =
6217 isa<OMPExecutableDirective>(S) &&
6218 isOpenMPTargetExecutionDirective(
6219 cast<OMPExecutableDirective>(S)->getDirectiveKind());
6221 if (requiresDeviceCodegen) {
6222 auto &E = *cast<OMPExecutableDirective>(S);
6226 getTargetEntryUniqueInfo(CGM.getContext(), E.getLocStart(), DeviceID,
6229 // Is this a target region that should not be emitted as an entry point? If
6230 // so just signal we are done with this target region.
6231 if (!OffloadEntriesInfoManager.hasTargetRegionEntryInfo(DeviceID, FileID,
6235 switch (S->getStmtClass()) {
6236 case Stmt::OMPTargetDirectiveClass:
6237 CodeGenFunction::EmitOMPTargetDeviceFunction(
6238 CGM, ParentName, cast<OMPTargetDirective>(*S));
6240 case Stmt::OMPTargetParallelDirectiveClass:
6241 CodeGenFunction::EmitOMPTargetParallelDeviceFunction(
6242 CGM, ParentName, cast<OMPTargetParallelDirective>(*S));
6244 case Stmt::OMPTargetTeamsDirectiveClass:
6245 CodeGenFunction::EmitOMPTargetTeamsDeviceFunction(
6246 CGM, ParentName, cast<OMPTargetTeamsDirective>(*S));
6249 llvm_unreachable("Unknown target directive for OpenMP device codegen.");
6254 if (const OMPExecutableDirective *E = dyn_cast<OMPExecutableDirective>(S)) {
6255 if (!E->hasAssociatedStmt())
6258 scanForTargetRegionsFunctions(
6259 cast<CapturedStmt>(E->getAssociatedStmt())->getCapturedStmt(),
6264 // If this is a lambda function, look into its body.
6265 if (auto *L = dyn_cast<LambdaExpr>(S))
6268 // Keep looking for target regions recursively.
6269 for (auto *II : S->children())
6270 scanForTargetRegionsFunctions(II, ParentName);
6273 bool CGOpenMPRuntime::emitTargetFunctions(GlobalDecl GD) {
6274 auto &FD = *cast<FunctionDecl>(GD.getDecl());
6276 // If emitting code for the host, we do not process FD here. Instead we do
6277 // the normal code generation.
6278 if (!CGM.getLangOpts().OpenMPIsDevice)
6281 // Try to detect target regions in the function.
6282 scanForTargetRegionsFunctions(FD.getBody(), CGM.getMangledName(GD));
6284 // We should not emit any function other that the ones created during the
6285 // scanning. Therefore, we signal that this function is completely dealt
6290 bool CGOpenMPRuntime::emitTargetGlobalVariable(GlobalDecl GD) {
6291 if (!CGM.getLangOpts().OpenMPIsDevice)
6294 // Check if there are Ctors/Dtors in this declaration and look for target
6295 // regions in it. We use the complete variant to produce the kernel name
6297 QualType RDTy = cast<VarDecl>(GD.getDecl())->getType();
6298 if (auto *RD = RDTy->getBaseElementTypeUnsafe()->getAsCXXRecordDecl()) {
6299 for (auto *Ctor : RD->ctors()) {
6300 StringRef ParentName =
6301 CGM.getMangledName(GlobalDecl(Ctor, Ctor_Complete));
6302 scanForTargetRegionsFunctions(Ctor->getBody(), ParentName);
6304 auto *Dtor = RD->getDestructor();
6306 StringRef ParentName =
6307 CGM.getMangledName(GlobalDecl(Dtor, Dtor_Complete));
6308 scanForTargetRegionsFunctions(Dtor->getBody(), ParentName);
6312 // If we are in target mode we do not emit any global (declare target is not
6313 // implemented yet). Therefore we signal that GD was processed in this case.
6317 bool CGOpenMPRuntime::emitTargetGlobal(GlobalDecl GD) {
6318 auto *VD = GD.getDecl();
6319 if (isa<FunctionDecl>(VD))
6320 return emitTargetFunctions(GD);
6322 return emitTargetGlobalVariable(GD);
6325 llvm::Function *CGOpenMPRuntime::emitRegistrationFunction() {
6326 // If we have offloading in the current module, we need to emit the entries
6327 // now and register the offloading descriptor.
6328 createOffloadEntriesAndInfoMetadata();
6330 // Create and register the offloading binary descriptors. This is the main
6331 // entity that captures all the information about offloading in the current
6332 // compilation unit.
6333 return createOffloadingBinaryDescriptorRegistration();
6336 void CGOpenMPRuntime::emitTeamsCall(CodeGenFunction &CGF,
6337 const OMPExecutableDirective &D,
6339 llvm::Value *OutlinedFn,
6340 ArrayRef<llvm::Value *> CapturedVars) {
6341 if (!CGF.HaveInsertPoint())
6344 auto *RTLoc = emitUpdateLocation(CGF, Loc);
6345 CodeGenFunction::RunCleanupsScope Scope(CGF);
6347 // Build call __kmpc_fork_teams(loc, n, microtask, var1, .., varn);
6348 llvm::Value *Args[] = {
6350 CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars
6351 CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy())};
6352 llvm::SmallVector<llvm::Value *, 16> RealArgs;
6353 RealArgs.append(std::begin(Args), std::end(Args));
6354 RealArgs.append(CapturedVars.begin(), CapturedVars.end());
6356 auto RTLFn = createRuntimeFunction(OMPRTL__kmpc_fork_teams);
6357 CGF.EmitRuntimeCall(RTLFn, RealArgs);
6360 void CGOpenMPRuntime::emitNumTeamsClause(CodeGenFunction &CGF,
6361 const Expr *NumTeams,
6362 const Expr *ThreadLimit,
6363 SourceLocation Loc) {
6364 if (!CGF.HaveInsertPoint())
6367 auto *RTLoc = emitUpdateLocation(CGF, Loc);
6369 llvm::Value *NumTeamsVal =
6371 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(NumTeams),
6372 CGF.CGM.Int32Ty, /* isSigned = */ true)
6373 : CGF.Builder.getInt32(0);
6375 llvm::Value *ThreadLimitVal =
6377 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(ThreadLimit),
6378 CGF.CGM.Int32Ty, /* isSigned = */ true)
6379 : CGF.Builder.getInt32(0);
6381 // Build call __kmpc_push_num_teamss(&loc, global_tid, num_teams, thread_limit)
6382 llvm::Value *PushNumTeamsArgs[] = {RTLoc, getThreadID(CGF, Loc), NumTeamsVal,
6384 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_num_teams),
6388 void CGOpenMPRuntime::emitTargetDataCalls(
6389 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
6390 const Expr *Device, const RegionCodeGenTy &CodeGen, TargetDataInfo &Info) {
6391 if (!CGF.HaveInsertPoint())
6394 // Action used to replace the default codegen action and turn privatization
6396 PrePostActionTy NoPrivAction;
6398 // Generate the code for the opening of the data environment. Capture all the
6399 // arguments of the runtime call by reference because they are used in the
6400 // closing of the region.
6401 auto &&BeginThenGen = [&D, Device, &Info, &CodeGen](CodeGenFunction &CGF,
6402 PrePostActionTy &) {
6403 // Fill up the arrays with all the mapped variables.
6404 MappableExprsHandler::MapBaseValuesArrayTy BasePointers;
6405 MappableExprsHandler::MapValuesArrayTy Pointers;
6406 MappableExprsHandler::MapValuesArrayTy Sizes;
6407 MappableExprsHandler::MapFlagsArrayTy MapTypes;
6409 // Get map clause information.
6410 MappableExprsHandler MCHandler(D, CGF);
6411 MCHandler.generateAllInfo(BasePointers, Pointers, Sizes, MapTypes);
6413 // Fill up the arrays and create the arguments.
6414 emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info);
6416 llvm::Value *BasePointersArrayArg = nullptr;
6417 llvm::Value *PointersArrayArg = nullptr;
6418 llvm::Value *SizesArrayArg = nullptr;
6419 llvm::Value *MapTypesArrayArg = nullptr;
6420 emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg,
6421 SizesArrayArg, MapTypesArrayArg, Info);
6423 // Emit device ID if any.
6424 llvm::Value *DeviceID = nullptr;
6426 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6427 CGF.Int32Ty, /*isSigned=*/true);
6429 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6431 // Emit the number of elements in the offloading arrays.
6432 auto *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs);
6434 llvm::Value *OffloadingArgs[] = {
6435 DeviceID, PointerNum, BasePointersArrayArg,
6436 PointersArrayArg, SizesArrayArg, MapTypesArrayArg};
6437 auto &RT = CGF.CGM.getOpenMPRuntime();
6438 CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__tgt_target_data_begin),
6441 // If device pointer privatization is required, emit the body of the region
6442 // here. It will have to be duplicated: with and without privatization.
6443 if (!Info.CaptureDeviceAddrMap.empty())
6447 // Generate code for the closing of the data region.
6448 auto &&EndThenGen = [Device, &Info](CodeGenFunction &CGF, PrePostActionTy &) {
6449 assert(Info.isValid() && "Invalid data environment closing arguments.");
6451 llvm::Value *BasePointersArrayArg = nullptr;
6452 llvm::Value *PointersArrayArg = nullptr;
6453 llvm::Value *SizesArrayArg = nullptr;
6454 llvm::Value *MapTypesArrayArg = nullptr;
6455 emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg,
6456 SizesArrayArg, MapTypesArrayArg, Info);
6458 // Emit device ID if any.
6459 llvm::Value *DeviceID = nullptr;
6461 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6462 CGF.Int32Ty, /*isSigned=*/true);
6464 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6466 // Emit the number of elements in the offloading arrays.
6467 auto *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs);
6469 llvm::Value *OffloadingArgs[] = {
6470 DeviceID, PointerNum, BasePointersArrayArg,
6471 PointersArrayArg, SizesArrayArg, MapTypesArrayArg};
6472 auto &RT = CGF.CGM.getOpenMPRuntime();
6473 CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__tgt_target_data_end),
6477 // If we need device pointer privatization, we need to emit the body of the
6478 // region with no privatization in the 'else' branch of the conditional.
6479 // Otherwise, we don't have to do anything.
6480 auto &&BeginElseGen = [&Info, &CodeGen, &NoPrivAction](CodeGenFunction &CGF,
6481 PrePostActionTy &) {
6482 if (!Info.CaptureDeviceAddrMap.empty()) {
6483 CodeGen.setAction(NoPrivAction);
6488 // We don't have to do anything to close the region if the if clause evaluates
6490 auto &&EndElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {};
6493 emitOMPIfClause(CGF, IfCond, BeginThenGen, BeginElseGen);
6495 RegionCodeGenTy RCG(BeginThenGen);
6499 // If we don't require privatization of device pointers, we emit the body in
6500 // between the runtime calls. This avoids duplicating the body code.
6501 if (Info.CaptureDeviceAddrMap.empty()) {
6502 CodeGen.setAction(NoPrivAction);
6507 emitOMPIfClause(CGF, IfCond, EndThenGen, EndElseGen);
6509 RegionCodeGenTy RCG(EndThenGen);
6514 void CGOpenMPRuntime::emitTargetDataStandAloneCall(
6515 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
6516 const Expr *Device) {
6517 if (!CGF.HaveInsertPoint())
6520 assert((isa<OMPTargetEnterDataDirective>(D) ||
6521 isa<OMPTargetExitDataDirective>(D) ||
6522 isa<OMPTargetUpdateDirective>(D)) &&
6523 "Expecting either target enter, exit data, or update directives.");
6525 // Generate the code for the opening of the data environment.
6526 auto &&ThenGen = [&D, Device](CodeGenFunction &CGF, PrePostActionTy &) {
6527 // Fill up the arrays with all the mapped variables.
6528 MappableExprsHandler::MapBaseValuesArrayTy BasePointers;
6529 MappableExprsHandler::MapValuesArrayTy Pointers;
6530 MappableExprsHandler::MapValuesArrayTy Sizes;
6531 MappableExprsHandler::MapFlagsArrayTy MapTypes;
6533 // Get map clause information.
6534 MappableExprsHandler MEHandler(D, CGF);
6535 MEHandler.generateAllInfo(BasePointers, Pointers, Sizes, MapTypes);
6537 // Fill up the arrays and create the arguments.
6538 TargetDataInfo Info;
6539 emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info);
6540 emitOffloadingArraysArgument(CGF, Info.BasePointersArray,
6541 Info.PointersArray, Info.SizesArray,
6542 Info.MapTypesArray, Info);
6544 // Emit device ID if any.
6545 llvm::Value *DeviceID = nullptr;
6547 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6548 CGF.Int32Ty, /*isSigned=*/true);
6550 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6552 // Emit the number of elements in the offloading arrays.
6553 auto *PointerNum = CGF.Builder.getInt32(BasePointers.size());
6555 llvm::Value *OffloadingArgs[] = {
6556 DeviceID, PointerNum, Info.BasePointersArray,
6557 Info.PointersArray, Info.SizesArray, Info.MapTypesArray};
6559 auto &RT = CGF.CGM.getOpenMPRuntime();
6560 // Select the right runtime function call for each expected standalone
6562 OpenMPRTLFunction RTLFn;
6563 switch (D.getDirectiveKind()) {
6565 llvm_unreachable("Unexpected standalone target data directive.");
6567 case OMPD_target_enter_data:
6568 RTLFn = OMPRTL__tgt_target_data_begin;
6570 case OMPD_target_exit_data:
6571 RTLFn = OMPRTL__tgt_target_data_end;
6573 case OMPD_target_update:
6574 RTLFn = OMPRTL__tgt_target_data_update;
6577 CGF.EmitRuntimeCall(RT.createRuntimeFunction(RTLFn), OffloadingArgs);
6580 // In the event we get an if clause, we don't have to take any action on the
6582 auto &&ElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {};
6585 emitOMPIfClause(CGF, IfCond, ThenGen, ElseGen);
6587 RegionCodeGenTy ThenGenRCG(ThenGen);
6593 /// Kind of parameter in a function with 'declare simd' directive.
6594 enum ParamKindTy { LinearWithVarStride, Linear, Uniform, Vector };
6595 /// Attribute set of the parameter.
6596 struct ParamAttrTy {
6597 ParamKindTy Kind = Vector;
6598 llvm::APSInt StrideOrArg;
6599 llvm::APSInt Alignment;
6603 static unsigned evaluateCDTSize(const FunctionDecl *FD,
6604 ArrayRef<ParamAttrTy> ParamAttrs) {
6605 // Every vector variant of a SIMD-enabled function has a vector length (VLEN).
6606 // If OpenMP clause "simdlen" is used, the VLEN is the value of the argument
6607 // of that clause. The VLEN value must be power of 2.
6608 // In other case the notion of the function`s "characteristic data type" (CDT)
6609 // is used to compute the vector length.
6610 // CDT is defined in the following order:
6611 // a) For non-void function, the CDT is the return type.
6612 // b) If the function has any non-uniform, non-linear parameters, then the
6613 // CDT is the type of the first such parameter.
6614 // c) If the CDT determined by a) or b) above is struct, union, or class
6615 // type which is pass-by-value (except for the type that maps to the
6616 // built-in complex data type), the characteristic data type is int.
6617 // d) If none of the above three cases is applicable, the CDT is int.
6618 // The VLEN is then determined based on the CDT and the size of vector
6619 // register of that ISA for which current vector version is generated. The
6620 // VLEN is computed using the formula below:
6621 // VLEN = sizeof(vector_register) / sizeof(CDT),
6622 // where vector register size specified in section 3.2.1 Registers and the
6623 // Stack Frame of original AMD64 ABI document.
6624 QualType RetType = FD->getReturnType();
6625 if (RetType.isNull())
6627 ASTContext &C = FD->getASTContext();
6629 if (!RetType.isNull() && !RetType->isVoidType())
6632 unsigned Offset = 0;
6633 if (auto *MD = dyn_cast<CXXMethodDecl>(FD)) {
6634 if (ParamAttrs[Offset].Kind == Vector)
6635 CDT = C.getPointerType(C.getRecordType(MD->getParent()));
6639 for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) {
6640 if (ParamAttrs[I + Offset].Kind == Vector) {
6641 CDT = FD->getParamDecl(I)->getType();
6649 CDT = CDT->getCanonicalTypeUnqualified();
6650 if (CDT->isRecordType() || CDT->isUnionType())
6652 return C.getTypeSize(CDT);
6656 emitX86DeclareSimdFunction(const FunctionDecl *FD, llvm::Function *Fn,
6657 const llvm::APSInt &VLENVal,
6658 ArrayRef<ParamAttrTy> ParamAttrs,
6659 OMPDeclareSimdDeclAttr::BranchStateTy State) {
6662 unsigned VecRegSize;
6664 ISADataTy ISAData[] = {
6678 llvm::SmallVector<char, 2> Masked;
6680 case OMPDeclareSimdDeclAttr::BS_Undefined:
6681 Masked.push_back('N');
6682 Masked.push_back('M');
6684 case OMPDeclareSimdDeclAttr::BS_Notinbranch:
6685 Masked.push_back('N');
6687 case OMPDeclareSimdDeclAttr::BS_Inbranch:
6688 Masked.push_back('M');
6691 for (auto Mask : Masked) {
6692 for (auto &Data : ISAData) {
6693 SmallString<256> Buffer;
6694 llvm::raw_svector_ostream Out(Buffer);
6695 Out << "_ZGV" << Data.ISA << Mask;
6697 Out << llvm::APSInt::getUnsigned(Data.VecRegSize /
6698 evaluateCDTSize(FD, ParamAttrs));
6701 for (auto &ParamAttr : ParamAttrs) {
6702 switch (ParamAttr.Kind){
6703 case LinearWithVarStride:
6704 Out << 's' << ParamAttr.StrideOrArg;
6708 if (!!ParamAttr.StrideOrArg)
6709 Out << ParamAttr.StrideOrArg;
6718 if (!!ParamAttr.Alignment)
6719 Out << 'a' << ParamAttr.Alignment;
6721 Out << '_' << Fn->getName();
6722 Fn->addFnAttr(Out.str());
6727 void CGOpenMPRuntime::emitDeclareSimdFunction(const FunctionDecl *FD,
6728 llvm::Function *Fn) {
6729 ASTContext &C = CGM.getContext();
6730 FD = FD->getCanonicalDecl();
6731 // Map params to their positions in function decl.
6732 llvm::DenseMap<const Decl *, unsigned> ParamPositions;
6733 if (isa<CXXMethodDecl>(FD))
6734 ParamPositions.insert({FD, 0});
6735 unsigned ParamPos = ParamPositions.size();
6736 for (auto *P : FD->parameters()) {
6737 ParamPositions.insert({P->getCanonicalDecl(), ParamPos});
6740 for (auto *Attr : FD->specific_attrs<OMPDeclareSimdDeclAttr>()) {
6741 llvm::SmallVector<ParamAttrTy, 8> ParamAttrs(ParamPositions.size());
6742 // Mark uniform parameters.
6743 for (auto *E : Attr->uniforms()) {
6744 E = E->IgnoreParenImpCasts();
6746 if (isa<CXXThisExpr>(E))
6747 Pos = ParamPositions[FD];
6749 auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
6750 ->getCanonicalDecl();
6751 Pos = ParamPositions[PVD];
6753 ParamAttrs[Pos].Kind = Uniform;
6755 // Get alignment info.
6756 auto NI = Attr->alignments_begin();
6757 for (auto *E : Attr->aligneds()) {
6758 E = E->IgnoreParenImpCasts();
6761 if (isa<CXXThisExpr>(E)) {
6762 Pos = ParamPositions[FD];
6763 ParmTy = E->getType();
6765 auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
6766 ->getCanonicalDecl();
6767 Pos = ParamPositions[PVD];
6768 ParmTy = PVD->getType();
6770 ParamAttrs[Pos].Alignment =
6771 (*NI) ? (*NI)->EvaluateKnownConstInt(C)
6772 : llvm::APSInt::getUnsigned(
6773 C.toCharUnitsFromBits(C.getOpenMPDefaultSimdAlign(ParmTy))
6777 // Mark linear parameters.
6778 auto SI = Attr->steps_begin();
6779 auto MI = Attr->modifiers_begin();
6780 for (auto *E : Attr->linears()) {
6781 E = E->IgnoreParenImpCasts();
6783 if (isa<CXXThisExpr>(E))
6784 Pos = ParamPositions[FD];
6786 auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
6787 ->getCanonicalDecl();
6788 Pos = ParamPositions[PVD];
6790 auto &ParamAttr = ParamAttrs[Pos];
6791 ParamAttr.Kind = Linear;
6793 if (!(*SI)->EvaluateAsInt(ParamAttr.StrideOrArg, C,
6794 Expr::SE_AllowSideEffects)) {
6795 if (auto *DRE = cast<DeclRefExpr>((*SI)->IgnoreParenImpCasts())) {
6796 if (auto *StridePVD = cast<ParmVarDecl>(DRE->getDecl())) {
6797 ParamAttr.Kind = LinearWithVarStride;
6798 ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(
6799 ParamPositions[StridePVD->getCanonicalDecl()]);
6807 llvm::APSInt VLENVal;
6808 if (const Expr *VLEN = Attr->getSimdlen())
6809 VLENVal = VLEN->EvaluateKnownConstInt(C);
6810 OMPDeclareSimdDeclAttr::BranchStateTy State = Attr->getBranchState();
6811 if (CGM.getTriple().getArch() == llvm::Triple::x86 ||
6812 CGM.getTriple().getArch() == llvm::Triple::x86_64)
6813 emitX86DeclareSimdFunction(FD, Fn, VLENVal, ParamAttrs, State);
6818 /// Cleanup action for doacross support.
6819 class DoacrossCleanupTy final : public EHScopeStack::Cleanup {
6821 static const int DoacrossFinArgs = 2;
6825 llvm::Value *Args[DoacrossFinArgs];
6828 DoacrossCleanupTy(llvm::Value *RTLFn, ArrayRef<llvm::Value *> CallArgs)
6830 assert(CallArgs.size() == DoacrossFinArgs);
6831 std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args));
6833 void Emit(CodeGenFunction &CGF, Flags /*flags*/) override {
6834 if (!CGF.HaveInsertPoint())
6836 CGF.EmitRuntimeCall(RTLFn, Args);
6841 void CGOpenMPRuntime::emitDoacrossInit(CodeGenFunction &CGF,
6842 const OMPLoopDirective &D) {
6843 if (!CGF.HaveInsertPoint())
6846 ASTContext &C = CGM.getContext();
6847 QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true);
6849 if (KmpDimTy.isNull()) {
6850 // Build struct kmp_dim { // loop bounds info casted to kmp_int64
6851 // kmp_int64 lo; // lower
6852 // kmp_int64 up; // upper
6853 // kmp_int64 st; // stride
6855 RD = C.buildImplicitRecord("kmp_dim");
6856 RD->startDefinition();
6857 addFieldToRecordDecl(C, RD, Int64Ty);
6858 addFieldToRecordDecl(C, RD, Int64Ty);
6859 addFieldToRecordDecl(C, RD, Int64Ty);
6860 RD->completeDefinition();
6861 KmpDimTy = C.getRecordType(RD);
6863 RD = cast<RecordDecl>(KmpDimTy->getAsTagDecl());
6865 Address DimsAddr = CGF.CreateMemTemp(KmpDimTy, "dims");
6866 CGF.EmitNullInitialization(DimsAddr, KmpDimTy);
6867 enum { LowerFD = 0, UpperFD, StrideFD };
6868 // Fill dims with data.
6869 LValue DimsLVal = CGF.MakeAddrLValue(DimsAddr, KmpDimTy);
6870 // dims.upper = num_iterations;
6872 CGF.EmitLValueForField(DimsLVal, *std::next(RD->field_begin(), UpperFD));
6873 llvm::Value *NumIterVal = CGF.EmitScalarConversion(
6874 CGF.EmitScalarExpr(D.getNumIterations()), D.getNumIterations()->getType(),
6875 Int64Ty, D.getNumIterations()->getExprLoc());
6876 CGF.EmitStoreOfScalar(NumIterVal, UpperLVal);
6879 CGF.EmitLValueForField(DimsLVal, *std::next(RD->field_begin(), StrideFD));
6880 CGF.EmitStoreOfScalar(llvm::ConstantInt::getSigned(CGM.Int64Ty, /*V=*/1),
6883 // Build call void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
6884 // kmp_int32 num_dims, struct kmp_dim * dims);
6885 llvm::Value *Args[] = {emitUpdateLocation(CGF, D.getLocStart()),
6886 getThreadID(CGF, D.getLocStart()),
6887 llvm::ConstantInt::getSigned(CGM.Int32Ty, 1),
6888 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
6889 DimsAddr.getPointer(), CGM.VoidPtrTy)};
6891 llvm::Value *RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_init);
6892 CGF.EmitRuntimeCall(RTLFn, Args);
6893 llvm::Value *FiniArgs[DoacrossCleanupTy::DoacrossFinArgs] = {
6894 emitUpdateLocation(CGF, D.getLocEnd()), getThreadID(CGF, D.getLocEnd())};
6895 llvm::Value *FiniRTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_fini);
6896 CGF.EHStack.pushCleanup<DoacrossCleanupTy>(NormalAndEHCleanup, FiniRTLFn,
6897 llvm::makeArrayRef(FiniArgs));
6900 void CGOpenMPRuntime::emitDoacrossOrdered(CodeGenFunction &CGF,
6901 const OMPDependClause *C) {
6903 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
6904 const Expr *CounterVal = C->getCounterValue();
6906 llvm::Value *CntVal = CGF.EmitScalarConversion(CGF.EmitScalarExpr(CounterVal),
6907 CounterVal->getType(), Int64Ty,
6908 CounterVal->getExprLoc());
6909 Address CntAddr = CGF.CreateMemTemp(Int64Ty, ".cnt.addr");
6910 CGF.EmitStoreOfScalar(CntVal, CntAddr, /*Volatile=*/false, Int64Ty);
6911 llvm::Value *Args[] = {emitUpdateLocation(CGF, C->getLocStart()),
6912 getThreadID(CGF, C->getLocStart()),
6913 CntAddr.getPointer()};
6915 if (C->getDependencyKind() == OMPC_DEPEND_source)
6916 RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_post);
6918 assert(C->getDependencyKind() == OMPC_DEPEND_sink);
6919 RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_wait);
6921 CGF.EmitRuntimeCall(RTLFn, Args);