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(
2470 CodeGenFunction &CGF, SourceLocation Loc,
2471 const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned,
2472 bool Ordered, const DispatchRTInput &DispatchValues) {
2473 if (!CGF.HaveInsertPoint())
2475 OpenMPSchedType Schedule = getRuntimeSchedule(
2476 ScheduleKind.Schedule, DispatchValues.Chunk != nullptr, Ordered);
2478 (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked &&
2479 Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked &&
2480 Schedule != OMP_sch_static_balanced_chunked));
2481 // Call __kmpc_dispatch_init(
2482 // ident_t *loc, kmp_int32 tid, kmp_int32 schedule,
2483 // kmp_int[32|64] lower, kmp_int[32|64] upper,
2484 // kmp_int[32|64] stride, kmp_int[32|64] chunk);
2486 // If the Chunk was not specified in the clause - use default value 1.
2487 llvm::Value *Chunk = DispatchValues.Chunk ? DispatchValues.Chunk
2488 : CGF.Builder.getIntN(IVSize, 1);
2489 llvm::Value *Args[] = {
2490 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2491 CGF.Builder.getInt32(addMonoNonMonoModifier(
2492 Schedule, ScheduleKind.M1, ScheduleKind.M2)), // Schedule type
2493 DispatchValues.LB, // Lower
2494 DispatchValues.UB, // Upper
2495 CGF.Builder.getIntN(IVSize, 1), // Stride
2498 CGF.EmitRuntimeCall(createDispatchInitFunction(IVSize, IVSigned), Args);
2501 static void emitForStaticInitCall(
2502 CodeGenFunction &CGF, llvm::Value *UpdateLocation, llvm::Value *ThreadId,
2503 llvm::Constant *ForStaticInitFunction, OpenMPSchedType Schedule,
2504 OpenMPScheduleClauseModifier M1, OpenMPScheduleClauseModifier M2,
2505 unsigned IVSize, bool Ordered, Address IL, Address LB, Address UB,
2506 Address ST, llvm::Value *Chunk) {
2507 if (!CGF.HaveInsertPoint())
2511 assert(Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked ||
2512 Schedule == OMP_sch_static_balanced_chunked ||
2513 Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked ||
2514 Schedule == OMP_dist_sch_static ||
2515 Schedule == OMP_dist_sch_static_chunked);
2517 // Call __kmpc_for_static_init(
2518 // ident_t *loc, kmp_int32 tid, kmp_int32 schedtype,
2519 // kmp_int32 *p_lastiter, kmp_int[32|64] *p_lower,
2520 // kmp_int[32|64] *p_upper, kmp_int[32|64] *p_stride,
2521 // kmp_int[32|64] incr, kmp_int[32|64] chunk);
2522 if (Chunk == nullptr) {
2523 assert((Schedule == OMP_sch_static || Schedule == OMP_ord_static ||
2524 Schedule == OMP_dist_sch_static) &&
2525 "expected static non-chunked schedule");
2526 // If the Chunk was not specified in the clause - use default value 1.
2527 Chunk = CGF.Builder.getIntN(IVSize, 1);
2529 assert((Schedule == OMP_sch_static_chunked ||
2530 Schedule == OMP_sch_static_balanced_chunked ||
2531 Schedule == OMP_ord_static_chunked ||
2532 Schedule == OMP_dist_sch_static_chunked) &&
2533 "expected static chunked schedule");
2535 llvm::Value *Args[] = {
2536 UpdateLocation, ThreadId, CGF.Builder.getInt32(addMonoNonMonoModifier(
2537 Schedule, M1, M2)), // Schedule type
2538 IL.getPointer(), // &isLastIter
2539 LB.getPointer(), // &LB
2540 UB.getPointer(), // &UB
2541 ST.getPointer(), // &Stride
2542 CGF.Builder.getIntN(IVSize, 1), // Incr
2545 CGF.EmitRuntimeCall(ForStaticInitFunction, Args);
2548 void CGOpenMPRuntime::emitForStaticInit(CodeGenFunction &CGF,
2550 const OpenMPScheduleTy &ScheduleKind,
2551 unsigned IVSize, bool IVSigned,
2552 bool Ordered, Address IL, Address LB,
2553 Address UB, Address ST,
2554 llvm::Value *Chunk) {
2555 OpenMPSchedType ScheduleNum =
2556 getRuntimeSchedule(ScheduleKind.Schedule, Chunk != nullptr, Ordered);
2557 auto *UpdatedLocation = emitUpdateLocation(CGF, Loc);
2558 auto *ThreadId = getThreadID(CGF, Loc);
2559 auto *StaticInitFunction = createForStaticInitFunction(IVSize, IVSigned);
2560 emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction,
2561 ScheduleNum, ScheduleKind.M1, ScheduleKind.M2, IVSize,
2562 Ordered, IL, LB, UB, ST, Chunk);
2565 void CGOpenMPRuntime::emitDistributeStaticInit(
2566 CodeGenFunction &CGF, SourceLocation Loc,
2567 OpenMPDistScheduleClauseKind SchedKind, unsigned IVSize, bool IVSigned,
2568 bool Ordered, Address IL, Address LB, Address UB, Address ST,
2569 llvm::Value *Chunk) {
2570 OpenMPSchedType ScheduleNum = getRuntimeSchedule(SchedKind, Chunk != nullptr);
2571 auto *UpdatedLocation = emitUpdateLocation(CGF, Loc);
2572 auto *ThreadId = getThreadID(CGF, Loc);
2573 auto *StaticInitFunction = createForStaticInitFunction(IVSize, IVSigned);
2574 emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction,
2575 ScheduleNum, OMPC_SCHEDULE_MODIFIER_unknown,
2576 OMPC_SCHEDULE_MODIFIER_unknown, IVSize, Ordered, IL, LB,
2580 void CGOpenMPRuntime::emitForStaticFinish(CodeGenFunction &CGF,
2581 SourceLocation Loc) {
2582 if (!CGF.HaveInsertPoint())
2584 // Call __kmpc_for_static_fini(ident_t *loc, kmp_int32 tid);
2585 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2586 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_for_static_fini),
2590 void CGOpenMPRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF,
2594 if (!CGF.HaveInsertPoint())
2596 // Call __kmpc_for_dynamic_fini_(4|8)[u](ident_t *loc, kmp_int32 tid);
2597 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2598 CGF.EmitRuntimeCall(createDispatchFiniFunction(IVSize, IVSigned), Args);
2601 llvm::Value *CGOpenMPRuntime::emitForNext(CodeGenFunction &CGF,
2602 SourceLocation Loc, unsigned IVSize,
2603 bool IVSigned, Address IL,
2604 Address LB, Address UB,
2606 // Call __kmpc_dispatch_next(
2607 // ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter,
2608 // kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper,
2609 // kmp_int[32|64] *p_stride);
2610 llvm::Value *Args[] = {
2611 emitUpdateLocation(CGF, Loc),
2612 getThreadID(CGF, Loc),
2613 IL.getPointer(), // &isLastIter
2614 LB.getPointer(), // &Lower
2615 UB.getPointer(), // &Upper
2616 ST.getPointer() // &Stride
2619 CGF.EmitRuntimeCall(createDispatchNextFunction(IVSize, IVSigned), Args);
2620 return CGF.EmitScalarConversion(
2621 Call, CGF.getContext().getIntTypeForBitwidth(32, /* Signed */ true),
2622 CGF.getContext().BoolTy, Loc);
2625 void CGOpenMPRuntime::emitNumThreadsClause(CodeGenFunction &CGF,
2626 llvm::Value *NumThreads,
2627 SourceLocation Loc) {
2628 if (!CGF.HaveInsertPoint())
2630 // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads)
2631 llvm::Value *Args[] = {
2632 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2633 CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)};
2634 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_num_threads),
2638 void CGOpenMPRuntime::emitProcBindClause(CodeGenFunction &CGF,
2639 OpenMPProcBindClauseKind ProcBind,
2640 SourceLocation Loc) {
2641 if (!CGF.HaveInsertPoint())
2643 // Constants for proc bind value accepted by the runtime.
2654 case OMPC_PROC_BIND_master:
2655 RuntimeProcBind = ProcBindMaster;
2657 case OMPC_PROC_BIND_close:
2658 RuntimeProcBind = ProcBindClose;
2660 case OMPC_PROC_BIND_spread:
2661 RuntimeProcBind = ProcBindSpread;
2663 case OMPC_PROC_BIND_unknown:
2664 llvm_unreachable("Unsupported proc_bind value.");
2666 // Build call __kmpc_push_proc_bind(&loc, global_tid, proc_bind)
2667 llvm::Value *Args[] = {
2668 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2669 llvm::ConstantInt::get(CGM.IntTy, RuntimeProcBind, /*isSigned=*/true)};
2670 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_proc_bind), Args);
2673 void CGOpenMPRuntime::emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>,
2674 SourceLocation Loc) {
2675 if (!CGF.HaveInsertPoint())
2677 // Build call void __kmpc_flush(ident_t *loc)
2678 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_flush),
2679 emitUpdateLocation(CGF, Loc));
2683 /// \brief Indexes of fields for type kmp_task_t.
2684 enum KmpTaskTFields {
2685 /// \brief List of shared variables.
2687 /// \brief Task routine.
2689 /// \brief Partition id for the untied tasks.
2691 /// Function with call of destructors for private variables.
2695 /// (Taskloops only) Lower bound.
2697 /// (Taskloops only) Upper bound.
2699 /// (Taskloops only) Stride.
2701 /// (Taskloops only) Is last iteration flag.
2704 } // anonymous namespace
2706 bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::empty() const {
2707 // FIXME: Add other entries type when they become supported.
2708 return OffloadEntriesTargetRegion.empty();
2711 /// \brief Initialize target region entry.
2712 void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::
2713 initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
2714 StringRef ParentName, unsigned LineNum,
2716 assert(CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is "
2717 "only required for the device "
2718 "code generation.");
2719 OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] =
2720 OffloadEntryInfoTargetRegion(Order, /*Addr=*/nullptr, /*ID=*/nullptr,
2722 ++OffloadingEntriesNum;
2725 void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::
2726 registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID,
2727 StringRef ParentName, unsigned LineNum,
2728 llvm::Constant *Addr, llvm::Constant *ID,
2730 // If we are emitting code for a target, the entry is already initialized,
2731 // only has to be registered.
2732 if (CGM.getLangOpts().OpenMPIsDevice) {
2733 assert(hasTargetRegionEntryInfo(DeviceID, FileID, ParentName, LineNum) &&
2734 "Entry must exist.");
2736 OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum];
2737 assert(Entry.isValid() && "Entry not initialized!");
2738 Entry.setAddress(Addr);
2740 Entry.setFlags(Flags);
2743 OffloadEntryInfoTargetRegion Entry(OffloadingEntriesNum++, Addr, ID, Flags);
2744 OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] = Entry;
2748 bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::hasTargetRegionEntryInfo(
2749 unsigned DeviceID, unsigned FileID, StringRef ParentName,
2750 unsigned LineNum) const {
2751 auto PerDevice = OffloadEntriesTargetRegion.find(DeviceID);
2752 if (PerDevice == OffloadEntriesTargetRegion.end())
2754 auto PerFile = PerDevice->second.find(FileID);
2755 if (PerFile == PerDevice->second.end())
2757 auto PerParentName = PerFile->second.find(ParentName);
2758 if (PerParentName == PerFile->second.end())
2760 auto PerLine = PerParentName->second.find(LineNum);
2761 if (PerLine == PerParentName->second.end())
2763 // Fail if this entry is already registered.
2764 if (PerLine->second.getAddress() || PerLine->second.getID())
2769 void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::actOnTargetRegionEntriesInfo(
2770 const OffloadTargetRegionEntryInfoActTy &Action) {
2771 // Scan all target region entries and perform the provided action.
2772 for (auto &D : OffloadEntriesTargetRegion)
2773 for (auto &F : D.second)
2774 for (auto &P : F.second)
2775 for (auto &L : P.second)
2776 Action(D.first, F.first, P.first(), L.first, L.second);
2779 /// \brief Create a Ctor/Dtor-like function whose body is emitted through
2780 /// \a Codegen. This is used to emit the two functions that register and
2781 /// unregister the descriptor of the current compilation unit.
2782 static llvm::Function *
2783 createOffloadingBinaryDescriptorFunction(CodeGenModule &CGM, StringRef Name,
2784 const RegionCodeGenTy &Codegen) {
2785 auto &C = CGM.getContext();
2786 FunctionArgList Args;
2787 ImplicitParamDecl DummyPtr(C, /*DC=*/nullptr, SourceLocation(),
2788 /*Id=*/nullptr, C.VoidPtrTy);
2789 Args.push_back(&DummyPtr);
2791 CodeGenFunction CGF(CGM);
2792 auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2793 auto FTy = CGM.getTypes().GetFunctionType(FI);
2795 CGM.CreateGlobalInitOrDestructFunction(FTy, Name, FI, SourceLocation());
2796 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FI, Args, SourceLocation());
2798 CGF.FinishFunction();
2803 CGOpenMPRuntime::createOffloadingBinaryDescriptorRegistration() {
2805 // If we don't have entries or if we are emitting code for the device, we
2806 // don't need to do anything.
2807 if (CGM.getLangOpts().OpenMPIsDevice || OffloadEntriesInfoManager.empty())
2810 auto &M = CGM.getModule();
2811 auto &C = CGM.getContext();
2813 // Get list of devices we care about
2814 auto &Devices = CGM.getLangOpts().OMPTargetTriples;
2816 // We should be creating an offloading descriptor only if there are devices
2818 assert(!Devices.empty() && "No OpenMP offloading devices??");
2820 // Create the external variables that will point to the begin and end of the
2821 // host entries section. These will be defined by the linker.
2822 auto *OffloadEntryTy =
2823 CGM.getTypes().ConvertTypeForMem(getTgtOffloadEntryQTy());
2824 llvm::GlobalVariable *HostEntriesBegin = new llvm::GlobalVariable(
2825 M, OffloadEntryTy, /*isConstant=*/true,
2826 llvm::GlobalValue::ExternalLinkage, /*Initializer=*/nullptr,
2827 ".omp_offloading.entries_begin");
2828 llvm::GlobalVariable *HostEntriesEnd = new llvm::GlobalVariable(
2829 M, OffloadEntryTy, /*isConstant=*/true,
2830 llvm::GlobalValue::ExternalLinkage, /*Initializer=*/nullptr,
2831 ".omp_offloading.entries_end");
2833 // Create all device images
2834 auto *DeviceImageTy = cast<llvm::StructType>(
2835 CGM.getTypes().ConvertTypeForMem(getTgtDeviceImageQTy()));
2836 ConstantInitBuilder DeviceImagesBuilder(CGM);
2837 auto DeviceImagesEntries = DeviceImagesBuilder.beginArray(DeviceImageTy);
2839 for (unsigned i = 0; i < Devices.size(); ++i) {
2840 StringRef T = Devices[i].getTriple();
2841 auto *ImgBegin = new llvm::GlobalVariable(
2842 M, CGM.Int8Ty, /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage,
2843 /*Initializer=*/nullptr,
2844 Twine(".omp_offloading.img_start.") + Twine(T));
2845 auto *ImgEnd = new llvm::GlobalVariable(
2846 M, CGM.Int8Ty, /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage,
2847 /*Initializer=*/nullptr, Twine(".omp_offloading.img_end.") + Twine(T));
2849 auto Dev = DeviceImagesEntries.beginStruct(DeviceImageTy);
2852 Dev.add(HostEntriesBegin);
2853 Dev.add(HostEntriesEnd);
2854 Dev.finishAndAddTo(DeviceImagesEntries);
2857 // Create device images global array.
2858 llvm::GlobalVariable *DeviceImages =
2859 DeviceImagesEntries.finishAndCreateGlobal(".omp_offloading.device_images",
2860 CGM.getPointerAlign(),
2861 /*isConstant=*/true);
2862 DeviceImages->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2864 // This is a Zero array to be used in the creation of the constant expressions
2865 llvm::Constant *Index[] = {llvm::Constant::getNullValue(CGM.Int32Ty),
2866 llvm::Constant::getNullValue(CGM.Int32Ty)};
2868 // Create the target region descriptor.
2869 auto *BinaryDescriptorTy = cast<llvm::StructType>(
2870 CGM.getTypes().ConvertTypeForMem(getTgtBinaryDescriptorQTy()));
2871 ConstantInitBuilder DescBuilder(CGM);
2872 auto DescInit = DescBuilder.beginStruct(BinaryDescriptorTy);
2873 DescInit.addInt(CGM.Int32Ty, Devices.size());
2874 DescInit.add(llvm::ConstantExpr::getGetElementPtr(DeviceImages->getValueType(),
2877 DescInit.add(HostEntriesBegin);
2878 DescInit.add(HostEntriesEnd);
2880 auto *Desc = DescInit.finishAndCreateGlobal(".omp_offloading.descriptor",
2881 CGM.getPointerAlign(),
2882 /*isConstant=*/true);
2884 // Emit code to register or unregister the descriptor at execution
2885 // startup or closing, respectively.
2887 // Create a variable to drive the registration and unregistration of the
2888 // descriptor, so we can reuse the logic that emits Ctors and Dtors.
2889 auto *IdentInfo = &C.Idents.get(".omp_offloading.reg_unreg_var");
2890 ImplicitParamDecl RegUnregVar(C, C.getTranslationUnitDecl(), SourceLocation(),
2891 IdentInfo, C.CharTy);
2893 auto *UnRegFn = createOffloadingBinaryDescriptorFunction(
2894 CGM, ".omp_offloading.descriptor_unreg",
2895 [&](CodeGenFunction &CGF, PrePostActionTy &) {
2896 CGF.EmitCallOrInvoke(createRuntimeFunction(OMPRTL__tgt_unregister_lib),
2899 auto *RegFn = createOffloadingBinaryDescriptorFunction(
2900 CGM, ".omp_offloading.descriptor_reg",
2901 [&](CodeGenFunction &CGF, PrePostActionTy &) {
2902 CGF.EmitCallOrInvoke(createRuntimeFunction(OMPRTL__tgt_register_lib),
2904 CGM.getCXXABI().registerGlobalDtor(CGF, RegUnregVar, UnRegFn, Desc);
2909 void CGOpenMPRuntime::createOffloadEntry(llvm::Constant *ID,
2910 llvm::Constant *Addr, uint64_t Size,
2912 StringRef Name = Addr->getName();
2913 auto *TgtOffloadEntryType = cast<llvm::StructType>(
2914 CGM.getTypes().ConvertTypeForMem(getTgtOffloadEntryQTy()));
2915 llvm::LLVMContext &C = CGM.getModule().getContext();
2916 llvm::Module &M = CGM.getModule();
2918 // Make sure the address has the right type.
2919 llvm::Constant *AddrPtr = llvm::ConstantExpr::getBitCast(ID, CGM.VoidPtrTy);
2921 // Create constant string with the name.
2922 llvm::Constant *StrPtrInit = llvm::ConstantDataArray::getString(C, Name);
2924 llvm::GlobalVariable *Str =
2925 new llvm::GlobalVariable(M, StrPtrInit->getType(), /*isConstant=*/true,
2926 llvm::GlobalValue::InternalLinkage, StrPtrInit,
2927 ".omp_offloading.entry_name");
2928 Str->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2929 llvm::Constant *StrPtr = llvm::ConstantExpr::getBitCast(Str, CGM.Int8PtrTy);
2931 // We can't have any padding between symbols, so we need to have 1-byte
2933 auto Align = CharUnits::fromQuantity(1);
2935 // Create the entry struct.
2936 ConstantInitBuilder EntryBuilder(CGM);
2937 auto EntryInit = EntryBuilder.beginStruct(TgtOffloadEntryType);
2938 EntryInit.add(AddrPtr);
2939 EntryInit.add(StrPtr);
2940 EntryInit.addInt(CGM.SizeTy, Size);
2941 EntryInit.addInt(CGM.Int32Ty, Flags);
2942 EntryInit.addInt(CGM.Int32Ty, 0);
2943 llvm::GlobalVariable *Entry =
2944 EntryInit.finishAndCreateGlobal(".omp_offloading.entry",
2947 llvm::GlobalValue::ExternalLinkage);
2949 // The entry has to be created in the section the linker expects it to be.
2950 Entry->setSection(".omp_offloading.entries");
2953 void CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata() {
2954 // Emit the offloading entries and metadata so that the device codegen side
2955 // can easily figure out what to emit. The produced metadata looks like
2958 // !omp_offload.info = !{!1, ...}
2960 // Right now we only generate metadata for function that contain target
2963 // If we do not have entries, we dont need to do anything.
2964 if (OffloadEntriesInfoManager.empty())
2967 llvm::Module &M = CGM.getModule();
2968 llvm::LLVMContext &C = M.getContext();
2969 SmallVector<OffloadEntriesInfoManagerTy::OffloadEntryInfo *, 16>
2970 OrderedEntries(OffloadEntriesInfoManager.size());
2972 // Create the offloading info metadata node.
2973 llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("omp_offload.info");
2975 // Auxiliary methods to create metadata values and strings.
2976 auto getMDInt = [&](unsigned v) {
2977 return llvm::ConstantAsMetadata::get(
2978 llvm::ConstantInt::get(llvm::Type::getInt32Ty(C), v));
2981 auto getMDString = [&](StringRef v) { return llvm::MDString::get(C, v); };
2983 // Create function that emits metadata for each target region entry;
2984 auto &&TargetRegionMetadataEmitter = [&](
2985 unsigned DeviceID, unsigned FileID, StringRef ParentName, unsigned Line,
2986 OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion &E) {
2987 llvm::SmallVector<llvm::Metadata *, 32> Ops;
2988 // Generate metadata for target regions. Each entry of this metadata
2990 // - Entry 0 -> Kind of this type of metadata (0).
2991 // - Entry 1 -> Device ID of the file where the entry was identified.
2992 // - Entry 2 -> File ID of the file where the entry was identified.
2993 // - Entry 3 -> Mangled name of the function where the entry was identified.
2994 // - Entry 4 -> Line in the file where the entry was identified.
2995 // - Entry 5 -> Order the entry was created.
2996 // The first element of the metadata node is the kind.
2997 Ops.push_back(getMDInt(E.getKind()));
2998 Ops.push_back(getMDInt(DeviceID));
2999 Ops.push_back(getMDInt(FileID));
3000 Ops.push_back(getMDString(ParentName));
3001 Ops.push_back(getMDInt(Line));
3002 Ops.push_back(getMDInt(E.getOrder()));
3004 // Save this entry in the right position of the ordered entries array.
3005 OrderedEntries[E.getOrder()] = &E;
3007 // Add metadata to the named metadata node.
3008 MD->addOperand(llvm::MDNode::get(C, Ops));
3011 OffloadEntriesInfoManager.actOnTargetRegionEntriesInfo(
3012 TargetRegionMetadataEmitter);
3014 for (auto *E : OrderedEntries) {
3015 assert(E && "All ordered entries must exist!");
3017 dyn_cast<OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion>(
3019 assert(CE->getID() && CE->getAddress() &&
3020 "Entry ID and Addr are invalid!");
3021 createOffloadEntry(CE->getID(), CE->getAddress(), /*Size=*/0);
3023 llvm_unreachable("Unsupported entry kind.");
3027 /// \brief Loads all the offload entries information from the host IR
3029 void CGOpenMPRuntime::loadOffloadInfoMetadata() {
3030 // If we are in target mode, load the metadata from the host IR. This code has
3031 // to match the metadaata creation in createOffloadEntriesAndInfoMetadata().
3033 if (!CGM.getLangOpts().OpenMPIsDevice)
3036 if (CGM.getLangOpts().OMPHostIRFile.empty())
3039 auto Buf = llvm::MemoryBuffer::getFile(CGM.getLangOpts().OMPHostIRFile);
3043 llvm::LLVMContext C;
3044 auto ME = expectedToErrorOrAndEmitErrors(
3045 C, llvm::parseBitcodeFile(Buf.get()->getMemBufferRef(), C));
3050 llvm::NamedMDNode *MD = ME.get()->getNamedMetadata("omp_offload.info");
3054 for (auto I : MD->operands()) {
3055 llvm::MDNode *MN = cast<llvm::MDNode>(I);
3057 auto getMDInt = [&](unsigned Idx) {
3058 llvm::ConstantAsMetadata *V =
3059 cast<llvm::ConstantAsMetadata>(MN->getOperand(Idx));
3060 return cast<llvm::ConstantInt>(V->getValue())->getZExtValue();
3063 auto getMDString = [&](unsigned Idx) {
3064 llvm::MDString *V = cast<llvm::MDString>(MN->getOperand(Idx));
3065 return V->getString();
3068 switch (getMDInt(0)) {
3070 llvm_unreachable("Unexpected metadata!");
3072 case OffloadEntriesInfoManagerTy::OffloadEntryInfo::
3073 OFFLOAD_ENTRY_INFO_TARGET_REGION:
3074 OffloadEntriesInfoManager.initializeTargetRegionEntryInfo(
3075 /*DeviceID=*/getMDInt(1), /*FileID=*/getMDInt(2),
3076 /*ParentName=*/getMDString(3), /*Line=*/getMDInt(4),
3077 /*Order=*/getMDInt(5));
3083 void CGOpenMPRuntime::emitKmpRoutineEntryT(QualType KmpInt32Ty) {
3084 if (!KmpRoutineEntryPtrTy) {
3085 // Build typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); type.
3086 auto &C = CGM.getContext();
3087 QualType KmpRoutineEntryTyArgs[] = {KmpInt32Ty, C.VoidPtrTy};
3088 FunctionProtoType::ExtProtoInfo EPI;
3089 KmpRoutineEntryPtrQTy = C.getPointerType(
3090 C.getFunctionType(KmpInt32Ty, KmpRoutineEntryTyArgs, EPI));
3091 KmpRoutineEntryPtrTy = CGM.getTypes().ConvertType(KmpRoutineEntryPtrQTy);
3095 static FieldDecl *addFieldToRecordDecl(ASTContext &C, DeclContext *DC,
3097 auto *Field = FieldDecl::Create(
3098 C, DC, SourceLocation(), SourceLocation(), /*Id=*/nullptr, FieldTy,
3099 C.getTrivialTypeSourceInfo(FieldTy, SourceLocation()),
3100 /*BW=*/nullptr, /*Mutable=*/false, /*InitStyle=*/ICIS_NoInit);
3101 Field->setAccess(AS_public);
3106 QualType CGOpenMPRuntime::getTgtOffloadEntryQTy() {
3108 // Make sure the type of the entry is already created. This is the type we
3110 // struct __tgt_offload_entry{
3111 // void *addr; // Pointer to the offload entry info.
3112 // // (function or global)
3113 // char *name; // Name of the function or global.
3114 // size_t size; // Size of the entry info (0 if it a function).
3115 // int32_t flags; // Flags associated with the entry, e.g. 'link'.
3116 // int32_t reserved; // Reserved, to use by the runtime library.
3118 if (TgtOffloadEntryQTy.isNull()) {
3119 ASTContext &C = CGM.getContext();
3120 auto *RD = C.buildImplicitRecord("__tgt_offload_entry");
3121 RD->startDefinition();
3122 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3123 addFieldToRecordDecl(C, RD, C.getPointerType(C.CharTy));
3124 addFieldToRecordDecl(C, RD, C.getSizeType());
3125 addFieldToRecordDecl(
3126 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true));
3127 addFieldToRecordDecl(
3128 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true));
3129 RD->completeDefinition();
3130 TgtOffloadEntryQTy = C.getRecordType(RD);
3132 return TgtOffloadEntryQTy;
3135 QualType CGOpenMPRuntime::getTgtDeviceImageQTy() {
3136 // These are the types we need to build:
3137 // struct __tgt_device_image{
3138 // void *ImageStart; // Pointer to the target code start.
3139 // void *ImageEnd; // Pointer to the target code end.
3140 // // We also add the host entries to the device image, as it may be useful
3141 // // for the target runtime to have access to that information.
3142 // __tgt_offload_entry *EntriesBegin; // Begin of the table with all
3144 // __tgt_offload_entry *EntriesEnd; // End of the table with all the
3145 // // entries (non inclusive).
3147 if (TgtDeviceImageQTy.isNull()) {
3148 ASTContext &C = CGM.getContext();
3149 auto *RD = C.buildImplicitRecord("__tgt_device_image");
3150 RD->startDefinition();
3151 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3152 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3153 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3154 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3155 RD->completeDefinition();
3156 TgtDeviceImageQTy = C.getRecordType(RD);
3158 return TgtDeviceImageQTy;
3161 QualType CGOpenMPRuntime::getTgtBinaryDescriptorQTy() {
3162 // struct __tgt_bin_desc{
3163 // int32_t NumDevices; // Number of devices supported.
3164 // __tgt_device_image *DeviceImages; // Arrays of device images
3165 // // (one per device).
3166 // __tgt_offload_entry *EntriesBegin; // Begin of the table with all the
3168 // __tgt_offload_entry *EntriesEnd; // End of the table with all the
3169 // // entries (non inclusive).
3171 if (TgtBinaryDescriptorQTy.isNull()) {
3172 ASTContext &C = CGM.getContext();
3173 auto *RD = C.buildImplicitRecord("__tgt_bin_desc");
3174 RD->startDefinition();
3175 addFieldToRecordDecl(
3176 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true));
3177 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtDeviceImageQTy()));
3178 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3179 addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy()));
3180 RD->completeDefinition();
3181 TgtBinaryDescriptorQTy = C.getRecordType(RD);
3183 return TgtBinaryDescriptorQTy;
3187 struct PrivateHelpersTy {
3188 PrivateHelpersTy(const VarDecl *Original, const VarDecl *PrivateCopy,
3189 const VarDecl *PrivateElemInit)
3190 : Original(Original), PrivateCopy(PrivateCopy),
3191 PrivateElemInit(PrivateElemInit) {}
3192 const VarDecl *Original;
3193 const VarDecl *PrivateCopy;
3194 const VarDecl *PrivateElemInit;
3196 typedef std::pair<CharUnits /*Align*/, PrivateHelpersTy> PrivateDataTy;
3197 } // anonymous namespace
3200 createPrivatesRecordDecl(CodeGenModule &CGM, ArrayRef<PrivateDataTy> Privates) {
3201 if (!Privates.empty()) {
3202 auto &C = CGM.getContext();
3203 // Build struct .kmp_privates_t. {
3204 // /* private vars */
3206 auto *RD = C.buildImplicitRecord(".kmp_privates.t");
3207 RD->startDefinition();
3208 for (auto &&Pair : Privates) {
3209 auto *VD = Pair.second.Original;
3210 auto Type = VD->getType();
3211 Type = Type.getNonReferenceType();
3212 auto *FD = addFieldToRecordDecl(C, RD, Type);
3213 if (VD->hasAttrs()) {
3214 for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
3215 E(VD->getAttrs().end());
3220 RD->completeDefinition();
3227 createKmpTaskTRecordDecl(CodeGenModule &CGM, OpenMPDirectiveKind Kind,
3228 QualType KmpInt32Ty,
3229 QualType KmpRoutineEntryPointerQTy) {
3230 auto &C = CGM.getContext();
3231 // Build struct kmp_task_t {
3233 // kmp_routine_entry_t routine;
3234 // kmp_int32 part_id;
3235 // kmp_cmplrdata_t data1;
3236 // kmp_cmplrdata_t data2;
3237 // For taskloops additional fields:
3243 auto *UD = C.buildImplicitRecord("kmp_cmplrdata_t", TTK_Union);
3244 UD->startDefinition();
3245 addFieldToRecordDecl(C, UD, KmpInt32Ty);
3246 addFieldToRecordDecl(C, UD, KmpRoutineEntryPointerQTy);
3247 UD->completeDefinition();
3248 QualType KmpCmplrdataTy = C.getRecordType(UD);
3249 auto *RD = C.buildImplicitRecord("kmp_task_t");
3250 RD->startDefinition();
3251 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3252 addFieldToRecordDecl(C, RD, KmpRoutineEntryPointerQTy);
3253 addFieldToRecordDecl(C, RD, KmpInt32Ty);
3254 addFieldToRecordDecl(C, RD, KmpCmplrdataTy);
3255 addFieldToRecordDecl(C, RD, KmpCmplrdataTy);
3256 if (isOpenMPTaskLoopDirective(Kind)) {
3257 QualType KmpUInt64Ty =
3258 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0);
3259 QualType KmpInt64Ty =
3260 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
3261 addFieldToRecordDecl(C, RD, KmpUInt64Ty);
3262 addFieldToRecordDecl(C, RD, KmpUInt64Ty);
3263 addFieldToRecordDecl(C, RD, KmpInt64Ty);
3264 addFieldToRecordDecl(C, RD, KmpInt32Ty);
3266 RD->completeDefinition();
3271 createKmpTaskTWithPrivatesRecordDecl(CodeGenModule &CGM, QualType KmpTaskTQTy,
3272 ArrayRef<PrivateDataTy> Privates) {
3273 auto &C = CGM.getContext();
3274 // Build struct kmp_task_t_with_privates {
3275 // kmp_task_t task_data;
3276 // .kmp_privates_t. privates;
3278 auto *RD = C.buildImplicitRecord("kmp_task_t_with_privates");
3279 RD->startDefinition();
3280 addFieldToRecordDecl(C, RD, KmpTaskTQTy);
3281 if (auto *PrivateRD = createPrivatesRecordDecl(CGM, Privates)) {
3282 addFieldToRecordDecl(C, RD, C.getRecordType(PrivateRD));
3284 RD->completeDefinition();
3288 /// \brief Emit a proxy function which accepts kmp_task_t as the second
3291 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
3292 /// TaskFunction(gtid, tt->part_id, &tt->privates, task_privates_map, tt,
3294 /// tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter,
3299 static llvm::Value *
3300 emitProxyTaskFunction(CodeGenModule &CGM, SourceLocation Loc,
3301 OpenMPDirectiveKind Kind, QualType KmpInt32Ty,
3302 QualType KmpTaskTWithPrivatesPtrQTy,
3303 QualType KmpTaskTWithPrivatesQTy, QualType KmpTaskTQTy,
3304 QualType SharedsPtrTy, llvm::Value *TaskFunction,
3305 llvm::Value *TaskPrivatesMap) {
3306 auto &C = CGM.getContext();
3307 FunctionArgList Args;
3308 ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty);
3309 ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc,
3311 KmpTaskTWithPrivatesPtrQTy.withRestrict());
3312 Args.push_back(&GtidArg);
3313 Args.push_back(&TaskTypeArg);
3314 auto &TaskEntryFnInfo =
3315 CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args);
3316 auto *TaskEntryTy = CGM.getTypes().GetFunctionType(TaskEntryFnInfo);
3318 llvm::Function::Create(TaskEntryTy, llvm::GlobalValue::InternalLinkage,
3319 ".omp_task_entry.", &CGM.getModule());
3320 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, TaskEntry, TaskEntryFnInfo);
3321 CodeGenFunction CGF(CGM);
3322 CGF.disableDebugInfo();
3323 CGF.StartFunction(GlobalDecl(), KmpInt32Ty, TaskEntry, TaskEntryFnInfo, Args);
3325 // TaskFunction(gtid, tt->task_data.part_id, &tt->privates, task_privates_map,
3328 // tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter,
3329 // tt->task_data.shareds);
3330 auto *GtidParam = CGF.EmitLoadOfScalar(
3331 CGF.GetAddrOfLocalVar(&GtidArg), /*Volatile=*/false, KmpInt32Ty, Loc);
3332 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3333 CGF.GetAddrOfLocalVar(&TaskTypeArg),
3334 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3335 auto *KmpTaskTWithPrivatesQTyRD =
3336 cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl());
3338 CGF.EmitLValueForField(TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3339 auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl());
3340 auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId);
3341 auto PartIdLVal = CGF.EmitLValueForField(Base, *PartIdFI);
3342 auto *PartidParam = PartIdLVal.getPointer();
3344 auto SharedsFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds);
3345 auto SharedsLVal = CGF.EmitLValueForField(Base, *SharedsFI);
3346 auto *SharedsParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3347 CGF.EmitLoadOfLValue(SharedsLVal, Loc).getScalarVal(),
3348 CGF.ConvertTypeForMem(SharedsPtrTy));
3350 auto PrivatesFI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1);
3351 llvm::Value *PrivatesParam;
3352 if (PrivatesFI != KmpTaskTWithPrivatesQTyRD->field_end()) {
3353 auto PrivatesLVal = CGF.EmitLValueForField(TDBase, *PrivatesFI);
3354 PrivatesParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3355 PrivatesLVal.getPointer(), CGF.VoidPtrTy);
3357 PrivatesParam = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
3359 llvm::Value *CommonArgs[] = {GtidParam, PartidParam, PrivatesParam,
3362 .CreatePointerBitCastOrAddrSpaceCast(
3363 TDBase.getAddress(), CGF.VoidPtrTy)
3365 SmallVector<llvm::Value *, 16> CallArgs(std::begin(CommonArgs),
3366 std::end(CommonArgs));
3367 if (isOpenMPTaskLoopDirective(Kind)) {
3368 auto LBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound);
3369 auto LBLVal = CGF.EmitLValueForField(Base, *LBFI);
3370 auto *LBParam = CGF.EmitLoadOfLValue(LBLVal, Loc).getScalarVal();
3371 auto UBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound);
3372 auto UBLVal = CGF.EmitLValueForField(Base, *UBFI);
3373 auto *UBParam = CGF.EmitLoadOfLValue(UBLVal, Loc).getScalarVal();
3374 auto StFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTStride);
3375 auto StLVal = CGF.EmitLValueForField(Base, *StFI);
3376 auto *StParam = CGF.EmitLoadOfLValue(StLVal, Loc).getScalarVal();
3377 auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter);
3378 auto LILVal = CGF.EmitLValueForField(Base, *LIFI);
3379 auto *LIParam = CGF.EmitLoadOfLValue(LILVal, Loc).getScalarVal();
3380 CallArgs.push_back(LBParam);
3381 CallArgs.push_back(UBParam);
3382 CallArgs.push_back(StParam);
3383 CallArgs.push_back(LIParam);
3385 CallArgs.push_back(SharedsParam);
3387 CGF.EmitCallOrInvoke(TaskFunction, CallArgs);
3388 CGF.EmitStoreThroughLValue(
3389 RValue::get(CGF.Builder.getInt32(/*C=*/0)),
3390 CGF.MakeAddrLValue(CGF.ReturnValue, KmpInt32Ty));
3391 CGF.FinishFunction();
3395 static llvm::Value *emitDestructorsFunction(CodeGenModule &CGM,
3397 QualType KmpInt32Ty,
3398 QualType KmpTaskTWithPrivatesPtrQTy,
3399 QualType KmpTaskTWithPrivatesQTy) {
3400 auto &C = CGM.getContext();
3401 FunctionArgList Args;
3402 ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty);
3403 ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc,
3405 KmpTaskTWithPrivatesPtrQTy.withRestrict());
3406 Args.push_back(&GtidArg);
3407 Args.push_back(&TaskTypeArg);
3408 FunctionType::ExtInfo Info;
3409 auto &DestructorFnInfo =
3410 CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args);
3411 auto *DestructorFnTy = CGM.getTypes().GetFunctionType(DestructorFnInfo);
3412 auto *DestructorFn =
3413 llvm::Function::Create(DestructorFnTy, llvm::GlobalValue::InternalLinkage,
3414 ".omp_task_destructor.", &CGM.getModule());
3415 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, DestructorFn,
3417 CodeGenFunction CGF(CGM);
3418 CGF.disableDebugInfo();
3419 CGF.StartFunction(GlobalDecl(), KmpInt32Ty, DestructorFn, DestructorFnInfo,
3422 LValue Base = CGF.EmitLoadOfPointerLValue(
3423 CGF.GetAddrOfLocalVar(&TaskTypeArg),
3424 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3425 auto *KmpTaskTWithPrivatesQTyRD =
3426 cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl());
3427 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3428 Base = CGF.EmitLValueForField(Base, *FI);
3430 cast<RecordDecl>(FI->getType()->getAsTagDecl())->fields()) {
3431 if (auto DtorKind = Field->getType().isDestructedType()) {
3432 auto FieldLValue = CGF.EmitLValueForField(Base, Field);
3433 CGF.pushDestroy(DtorKind, FieldLValue.getAddress(), Field->getType());
3436 CGF.FinishFunction();
3437 return DestructorFn;
3440 /// \brief Emit a privates mapping function for correct handling of private and
3441 /// firstprivate variables.
3443 /// void .omp_task_privates_map.(const .privates. *noalias privs, <ty1>
3444 /// **noalias priv1,..., <tyn> **noalias privn) {
3445 /// *priv1 = &.privates.priv1;
3447 /// *privn = &.privates.privn;
3450 static llvm::Value *
3451 emitTaskPrivateMappingFunction(CodeGenModule &CGM, SourceLocation Loc,
3452 ArrayRef<const Expr *> PrivateVars,
3453 ArrayRef<const Expr *> FirstprivateVars,
3454 ArrayRef<const Expr *> LastprivateVars,
3455 QualType PrivatesQTy,
3456 ArrayRef<PrivateDataTy> Privates) {
3457 auto &C = CGM.getContext();
3458 FunctionArgList Args;
3459 ImplicitParamDecl TaskPrivatesArg(
3460 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3461 C.getPointerType(PrivatesQTy).withConst().withRestrict());
3462 Args.push_back(&TaskPrivatesArg);
3463 llvm::DenseMap<const VarDecl *, unsigned> PrivateVarsPos;
3464 unsigned Counter = 1;
3465 for (auto *E: PrivateVars) {
3466 Args.push_back(ImplicitParamDecl::Create(
3467 C, /*DC=*/nullptr, Loc,
3468 /*Id=*/nullptr, C.getPointerType(C.getPointerType(E->getType()))
3471 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3472 PrivateVarsPos[VD] = Counter;
3475 for (auto *E : FirstprivateVars) {
3476 Args.push_back(ImplicitParamDecl::Create(
3477 C, /*DC=*/nullptr, Loc,
3478 /*Id=*/nullptr, C.getPointerType(C.getPointerType(E->getType()))
3481 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3482 PrivateVarsPos[VD] = Counter;
3485 for (auto *E: LastprivateVars) {
3486 Args.push_back(ImplicitParamDecl::Create(
3487 C, /*DC=*/nullptr, Loc,
3488 /*Id=*/nullptr, C.getPointerType(C.getPointerType(E->getType()))
3491 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3492 PrivateVarsPos[VD] = Counter;
3495 auto &TaskPrivatesMapFnInfo =
3496 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3497 auto *TaskPrivatesMapTy =
3498 CGM.getTypes().GetFunctionType(TaskPrivatesMapFnInfo);
3499 auto *TaskPrivatesMap = llvm::Function::Create(
3500 TaskPrivatesMapTy, llvm::GlobalValue::InternalLinkage,
3501 ".omp_task_privates_map.", &CGM.getModule());
3502 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, TaskPrivatesMap,
3503 TaskPrivatesMapFnInfo);
3504 TaskPrivatesMap->removeFnAttr(llvm::Attribute::NoInline);
3505 TaskPrivatesMap->addFnAttr(llvm::Attribute::AlwaysInline);
3506 CodeGenFunction CGF(CGM);
3507 CGF.disableDebugInfo();
3508 CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskPrivatesMap,
3509 TaskPrivatesMapFnInfo, Args);
3511 // *privi = &.privates.privi;
3512 LValue Base = CGF.EmitLoadOfPointerLValue(
3513 CGF.GetAddrOfLocalVar(&TaskPrivatesArg),
3514 TaskPrivatesArg.getType()->castAs<PointerType>());
3515 auto *PrivatesQTyRD = cast<RecordDecl>(PrivatesQTy->getAsTagDecl());
3517 for (auto *Field : PrivatesQTyRD->fields()) {
3518 auto FieldLVal = CGF.EmitLValueForField(Base, Field);
3519 auto *VD = Args[PrivateVarsPos[Privates[Counter].second.Original]];
3520 auto RefLVal = CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
3521 auto RefLoadLVal = CGF.EmitLoadOfPointerLValue(
3522 RefLVal.getAddress(), RefLVal.getType()->castAs<PointerType>());
3523 CGF.EmitStoreOfScalar(FieldLVal.getPointer(), RefLoadLVal);
3526 CGF.FinishFunction();
3527 return TaskPrivatesMap;
3530 static int array_pod_sort_comparator(const PrivateDataTy *P1,
3531 const PrivateDataTy *P2) {
3532 return P1->first < P2->first ? 1 : (P2->first < P1->first ? -1 : 0);
3535 /// Emit initialization for private variables in task-based directives.
3536 static void emitPrivatesInit(CodeGenFunction &CGF,
3537 const OMPExecutableDirective &D,
3538 Address KmpTaskSharedsPtr, LValue TDBase,
3539 const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3540 QualType SharedsTy, QualType SharedsPtrTy,
3541 const OMPTaskDataTy &Data,
3542 ArrayRef<PrivateDataTy> Privates, bool ForDup) {
3543 auto &C = CGF.getContext();
3544 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3545 LValue PrivatesBase = CGF.EmitLValueForField(TDBase, *FI);
3547 if (!Data.FirstprivateVars.empty()) {
3548 SrcBase = CGF.MakeAddrLValue(
3549 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3550 KmpTaskSharedsPtr, CGF.ConvertTypeForMem(SharedsPtrTy)),
3553 CodeGenFunction::CGCapturedStmtInfo CapturesInfo(
3554 cast<CapturedStmt>(*D.getAssociatedStmt()));
3555 FI = cast<RecordDecl>(FI->getType()->getAsTagDecl())->field_begin();
3556 for (auto &&Pair : Privates) {
3557 auto *VD = Pair.second.PrivateCopy;
3558 auto *Init = VD->getAnyInitializer();
3559 if (Init && (!ForDup || (isa<CXXConstructExpr>(Init) &&
3560 !CGF.isTrivialInitializer(Init)))) {
3561 LValue PrivateLValue = CGF.EmitLValueForField(PrivatesBase, *FI);
3562 if (auto *Elem = Pair.second.PrivateElemInit) {
3563 auto *OriginalVD = Pair.second.Original;
3564 auto *SharedField = CapturesInfo.lookup(OriginalVD);
3565 auto SharedRefLValue = CGF.EmitLValueForField(SrcBase, SharedField);
3566 SharedRefLValue = CGF.MakeAddrLValue(
3567 Address(SharedRefLValue.getPointer(), C.getDeclAlign(OriginalVD)),
3568 SharedRefLValue.getType(), AlignmentSource::Decl);
3569 QualType Type = OriginalVD->getType();
3570 if (Type->isArrayType()) {
3571 // Initialize firstprivate array.
3572 if (!isa<CXXConstructExpr>(Init) || CGF.isTrivialInitializer(Init)) {
3573 // Perform simple memcpy.
3574 CGF.EmitAggregateAssign(PrivateLValue.getAddress(),
3575 SharedRefLValue.getAddress(), Type);
3577 // Initialize firstprivate array using element-by-element
3579 CGF.EmitOMPAggregateAssign(
3580 PrivateLValue.getAddress(), SharedRefLValue.getAddress(), Type,
3581 [&CGF, Elem, Init, &CapturesInfo](Address DestElement,
3582 Address SrcElement) {
3583 // Clean up any temporaries needed by the initialization.
3584 CodeGenFunction::OMPPrivateScope InitScope(CGF);
3585 InitScope.addPrivate(
3586 Elem, [SrcElement]() -> Address { return SrcElement; });
3587 (void)InitScope.Privatize();
3588 // Emit initialization for single element.
3589 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(
3590 CGF, &CapturesInfo);
3591 CGF.EmitAnyExprToMem(Init, DestElement,
3592 Init->getType().getQualifiers(),
3593 /*IsInitializer=*/false);
3597 CodeGenFunction::OMPPrivateScope InitScope(CGF);
3598 InitScope.addPrivate(Elem, [SharedRefLValue]() -> Address {
3599 return SharedRefLValue.getAddress();
3601 (void)InitScope.Privatize();
3602 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CapturesInfo);
3603 CGF.EmitExprAsInit(Init, VD, PrivateLValue,
3604 /*capturedByInit=*/false);
3607 CGF.EmitExprAsInit(Init, VD, PrivateLValue, /*capturedByInit=*/false);
3613 /// Check if duplication function is required for taskloops.
3614 static bool checkInitIsRequired(CodeGenFunction &CGF,
3615 ArrayRef<PrivateDataTy> Privates) {
3616 bool InitRequired = false;
3617 for (auto &&Pair : Privates) {
3618 auto *VD = Pair.second.PrivateCopy;
3619 auto *Init = VD->getAnyInitializer();
3620 InitRequired = InitRequired || (Init && isa<CXXConstructExpr>(Init) &&
3621 !CGF.isTrivialInitializer(Init));
3623 return InitRequired;
3627 /// Emit task_dup function (for initialization of
3628 /// private/firstprivate/lastprivate vars and last_iter flag)
3630 /// void __task_dup_entry(kmp_task_t *task_dst, const kmp_task_t *task_src, int
3632 /// // setup lastprivate flag
3633 /// task_dst->last = lastpriv;
3634 /// // could be constructor calls here...
3637 static llvm::Value *
3638 emitTaskDupFunction(CodeGenModule &CGM, SourceLocation Loc,
3639 const OMPExecutableDirective &D,
3640 QualType KmpTaskTWithPrivatesPtrQTy,
3641 const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3642 const RecordDecl *KmpTaskTQTyRD, QualType SharedsTy,
3643 QualType SharedsPtrTy, const OMPTaskDataTy &Data,
3644 ArrayRef<PrivateDataTy> Privates, bool WithLastIter) {
3645 auto &C = CGM.getContext();
3646 FunctionArgList Args;
3647 ImplicitParamDecl DstArg(C, /*DC=*/nullptr, Loc,
3648 /*Id=*/nullptr, KmpTaskTWithPrivatesPtrQTy);
3649 ImplicitParamDecl SrcArg(C, /*DC=*/nullptr, Loc,
3650 /*Id=*/nullptr, KmpTaskTWithPrivatesPtrQTy);
3651 ImplicitParamDecl LastprivArg(C, /*DC=*/nullptr, Loc,
3652 /*Id=*/nullptr, C.IntTy);
3653 Args.push_back(&DstArg);
3654 Args.push_back(&SrcArg);
3655 Args.push_back(&LastprivArg);
3656 auto &TaskDupFnInfo =
3657 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3658 auto *TaskDupTy = CGM.getTypes().GetFunctionType(TaskDupFnInfo);
3660 llvm::Function::Create(TaskDupTy, llvm::GlobalValue::InternalLinkage,
3661 ".omp_task_dup.", &CGM.getModule());
3662 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, TaskDup, TaskDupFnInfo);
3663 CodeGenFunction CGF(CGM);
3664 CGF.disableDebugInfo();
3665 CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskDup, TaskDupFnInfo, Args);
3667 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3668 CGF.GetAddrOfLocalVar(&DstArg),
3669 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3670 // task_dst->liter = lastpriv;
3672 auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter);
3673 LValue Base = CGF.EmitLValueForField(
3674 TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3675 LValue LILVal = CGF.EmitLValueForField(Base, *LIFI);
3676 llvm::Value *Lastpriv = CGF.EmitLoadOfScalar(
3677 CGF.GetAddrOfLocalVar(&LastprivArg), /*Volatile=*/false, C.IntTy, Loc);
3678 CGF.EmitStoreOfScalar(Lastpriv, LILVal);
3681 // Emit initial values for private copies (if any).
3682 assert(!Privates.empty());
3683 Address KmpTaskSharedsPtr = Address::invalid();
3684 if (!Data.FirstprivateVars.empty()) {
3685 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3686 CGF.GetAddrOfLocalVar(&SrcArg),
3687 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3688 LValue Base = CGF.EmitLValueForField(
3689 TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3690 KmpTaskSharedsPtr = Address(
3691 CGF.EmitLoadOfScalar(CGF.EmitLValueForField(
3692 Base, *std::next(KmpTaskTQTyRD->field_begin(),
3695 CGF.getNaturalTypeAlignment(SharedsTy));
3697 emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, TDBase, KmpTaskTWithPrivatesQTyRD,
3698 SharedsTy, SharedsPtrTy, Data, Privates, /*ForDup=*/true);
3699 CGF.FinishFunction();
3703 /// Checks if destructor function is required to be generated.
3704 /// \return true if cleanups are required, false otherwise.
3706 checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD) {
3707 bool NeedsCleanup = false;
3708 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3709 auto *PrivateRD = cast<RecordDecl>(FI->getType()->getAsTagDecl());
3710 for (auto *FD : PrivateRD->fields()) {
3711 NeedsCleanup = NeedsCleanup || FD->getType().isDestructedType();
3715 return NeedsCleanup;
3718 CGOpenMPRuntime::TaskResultTy
3719 CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc,
3720 const OMPExecutableDirective &D,
3721 llvm::Value *TaskFunction, QualType SharedsTy,
3722 Address Shareds, const OMPTaskDataTy &Data) {
3723 auto &C = CGM.getContext();
3724 llvm::SmallVector<PrivateDataTy, 4> Privates;
3725 // Aggregate privates and sort them by the alignment.
3726 auto I = Data.PrivateCopies.begin();
3727 for (auto *E : Data.PrivateVars) {
3728 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3729 Privates.push_back(std::make_pair(
3731 PrivateHelpersTy(VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3732 /*PrivateElemInit=*/nullptr)));
3735 I = Data.FirstprivateCopies.begin();
3736 auto IElemInitRef = Data.FirstprivateInits.begin();
3737 for (auto *E : Data.FirstprivateVars) {
3738 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3739 Privates.push_back(std::make_pair(
3742 VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3743 cast<VarDecl>(cast<DeclRefExpr>(*IElemInitRef)->getDecl()))));
3747 I = Data.LastprivateCopies.begin();
3748 for (auto *E : Data.LastprivateVars) {
3749 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3750 Privates.push_back(std::make_pair(
3752 PrivateHelpersTy(VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3753 /*PrivateElemInit=*/nullptr)));
3756 llvm::array_pod_sort(Privates.begin(), Privates.end(),
3757 array_pod_sort_comparator);
3758 auto KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
3759 // Build type kmp_routine_entry_t (if not built yet).
3760 emitKmpRoutineEntryT(KmpInt32Ty);
3761 // Build type kmp_task_t (if not built yet).
3762 if (KmpTaskTQTy.isNull()) {
3763 KmpTaskTQTy = C.getRecordType(createKmpTaskTRecordDecl(
3764 CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy));
3766 auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl());
3767 // Build particular struct kmp_task_t for the given task.
3768 auto *KmpTaskTWithPrivatesQTyRD =
3769 createKmpTaskTWithPrivatesRecordDecl(CGM, KmpTaskTQTy, Privates);
3770 auto KmpTaskTWithPrivatesQTy = C.getRecordType(KmpTaskTWithPrivatesQTyRD);
3771 QualType KmpTaskTWithPrivatesPtrQTy =
3772 C.getPointerType(KmpTaskTWithPrivatesQTy);
3773 auto *KmpTaskTWithPrivatesTy = CGF.ConvertType(KmpTaskTWithPrivatesQTy);
3774 auto *KmpTaskTWithPrivatesPtrTy = KmpTaskTWithPrivatesTy->getPointerTo();
3775 auto *KmpTaskTWithPrivatesTySize = CGF.getTypeSize(KmpTaskTWithPrivatesQTy);
3776 QualType SharedsPtrTy = C.getPointerType(SharedsTy);
3778 // Emit initial values for private copies (if any).
3779 llvm::Value *TaskPrivatesMap = nullptr;
3780 auto *TaskPrivatesMapTy =
3781 std::next(cast<llvm::Function>(TaskFunction)->arg_begin(), 3)->getType();
3782 if (!Privates.empty()) {
3783 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3784 TaskPrivatesMap = emitTaskPrivateMappingFunction(
3785 CGM, Loc, Data.PrivateVars, Data.FirstprivateVars, Data.LastprivateVars,
3786 FI->getType(), Privates);
3787 TaskPrivatesMap = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3788 TaskPrivatesMap, TaskPrivatesMapTy);
3790 TaskPrivatesMap = llvm::ConstantPointerNull::get(
3791 cast<llvm::PointerType>(TaskPrivatesMapTy));
3793 // Build a proxy function kmp_int32 .omp_task_entry.(kmp_int32 gtid,
3795 auto *TaskEntry = emitProxyTaskFunction(
3796 CGM, Loc, D.getDirectiveKind(), KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy,
3797 KmpTaskTWithPrivatesQTy, KmpTaskTQTy, SharedsPtrTy, TaskFunction,
3800 // Build call kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid,
3801 // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
3802 // kmp_routine_entry_t *task_entry);
3803 // Task flags. Format is taken from
3804 // http://llvm.org/svn/llvm-project/openmp/trunk/runtime/src/kmp.h,
3805 // description of kmp_tasking_flags struct.
3809 DestructorsFlag = 0x8,
3812 unsigned Flags = Data.Tied ? TiedFlag : 0;
3813 bool NeedsCleanup = false;
3814 if (!Privates.empty()) {
3815 NeedsCleanup = checkDestructorsRequired(KmpTaskTWithPrivatesQTyRD);
3817 Flags = Flags | DestructorsFlag;
3819 if (Data.Priority.getInt())
3820 Flags = Flags | PriorityFlag;
3822 Data.Final.getPointer()
3823 ? CGF.Builder.CreateSelect(Data.Final.getPointer(),
3824 CGF.Builder.getInt32(FinalFlag),
3825 CGF.Builder.getInt32(/*C=*/0))
3826 : CGF.Builder.getInt32(Data.Final.getInt() ? FinalFlag : 0);
3827 TaskFlags = CGF.Builder.CreateOr(TaskFlags, CGF.Builder.getInt32(Flags));
3828 auto *SharedsSize = CGM.getSize(C.getTypeSizeInChars(SharedsTy));
3829 llvm::Value *AllocArgs[] = {emitUpdateLocation(CGF, Loc),
3830 getThreadID(CGF, Loc), TaskFlags,
3831 KmpTaskTWithPrivatesTySize, SharedsSize,
3832 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3833 TaskEntry, KmpRoutineEntryPtrTy)};
3834 auto *NewTask = CGF.EmitRuntimeCall(
3835 createRuntimeFunction(OMPRTL__kmpc_omp_task_alloc), AllocArgs);
3836 auto *NewTaskNewTaskTTy = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3837 NewTask, KmpTaskTWithPrivatesPtrTy);
3838 LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy,
3839 KmpTaskTWithPrivatesQTy);
3841 CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin());
3842 // Fill the data in the resulting kmp_task_t record.
3843 // Copy shareds if there are any.
3844 Address KmpTaskSharedsPtr = Address::invalid();
3845 if (!SharedsTy->getAsStructureType()->getDecl()->field_empty()) {
3847 Address(CGF.EmitLoadOfScalar(
3848 CGF.EmitLValueForField(
3849 TDBase, *std::next(KmpTaskTQTyRD->field_begin(),
3852 CGF.getNaturalTypeAlignment(SharedsTy));
3853 CGF.EmitAggregateCopy(KmpTaskSharedsPtr, Shareds, SharedsTy);
3855 // Emit initial values for private copies (if any).
3856 TaskResultTy Result;
3857 if (!Privates.empty()) {
3858 emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, Base, KmpTaskTWithPrivatesQTyRD,
3859 SharedsTy, SharedsPtrTy, Data, Privates,
3861 if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) &&
3862 (!Data.LastprivateVars.empty() || checkInitIsRequired(CGF, Privates))) {
3863 Result.TaskDupFn = emitTaskDupFunction(
3864 CGM, Loc, D, KmpTaskTWithPrivatesPtrQTy, KmpTaskTWithPrivatesQTyRD,
3865 KmpTaskTQTyRD, SharedsTy, SharedsPtrTy, Data, Privates,
3866 /*WithLastIter=*/!Data.LastprivateVars.empty());
3869 // Fields of union "kmp_cmplrdata_t" for destructors and priority.
3870 enum { Priority = 0, Destructors = 1 };
3871 // Provide pointer to function with destructors for privates.
3872 auto FI = std::next(KmpTaskTQTyRD->field_begin(), Data1);
3873 auto *KmpCmplrdataUD = (*FI)->getType()->getAsUnionType()->getDecl();
3875 llvm::Value *DestructorFn = emitDestructorsFunction(
3876 CGM, Loc, KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy,
3877 KmpTaskTWithPrivatesQTy);
3878 LValue Data1LV = CGF.EmitLValueForField(TDBase, *FI);
3879 LValue DestructorsLV = CGF.EmitLValueForField(
3880 Data1LV, *std::next(KmpCmplrdataUD->field_begin(), Destructors));
3881 CGF.EmitStoreOfScalar(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3882 DestructorFn, KmpRoutineEntryPtrTy),
3886 if (Data.Priority.getInt()) {
3887 LValue Data2LV = CGF.EmitLValueForField(
3888 TDBase, *std::next(KmpTaskTQTyRD->field_begin(), Data2));
3889 LValue PriorityLV = CGF.EmitLValueForField(
3890 Data2LV, *std::next(KmpCmplrdataUD->field_begin(), Priority));
3891 CGF.EmitStoreOfScalar(Data.Priority.getPointer(), PriorityLV);
3893 Result.NewTask = NewTask;
3894 Result.TaskEntry = TaskEntry;
3895 Result.NewTaskNewTaskTTy = NewTaskNewTaskTTy;
3896 Result.TDBase = TDBase;
3897 Result.KmpTaskTQTyRD = KmpTaskTQTyRD;
3901 void CGOpenMPRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc,
3902 const OMPExecutableDirective &D,
3903 llvm::Value *TaskFunction,
3904 QualType SharedsTy, Address Shareds,
3906 const OMPTaskDataTy &Data) {
3907 if (!CGF.HaveInsertPoint())
3910 TaskResultTy Result =
3911 emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data);
3912 llvm::Value *NewTask = Result.NewTask;
3913 llvm::Value *TaskEntry = Result.TaskEntry;
3914 llvm::Value *NewTaskNewTaskTTy = Result.NewTaskNewTaskTTy;
3915 LValue TDBase = Result.TDBase;
3916 RecordDecl *KmpTaskTQTyRD = Result.KmpTaskTQTyRD;
3917 auto &C = CGM.getContext();
3918 // Process list of dependences.
3919 Address DependenciesArray = Address::invalid();
3920 unsigned NumDependencies = Data.Dependences.size();
3921 if (NumDependencies) {
3922 // Dependence kind for RTL.
3923 enum RTLDependenceKindTy { DepIn = 0x01, DepInOut = 0x3 };
3924 enum RTLDependInfoFieldsTy { BaseAddr, Len, Flags };
3925 RecordDecl *KmpDependInfoRD;
3927 C.getIntTypeForBitwidth(C.getTypeSize(C.BoolTy), /*Signed=*/false);
3928 llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy);
3929 if (KmpDependInfoTy.isNull()) {
3930 KmpDependInfoRD = C.buildImplicitRecord("kmp_depend_info");
3931 KmpDependInfoRD->startDefinition();
3932 addFieldToRecordDecl(C, KmpDependInfoRD, C.getIntPtrType());
3933 addFieldToRecordDecl(C, KmpDependInfoRD, C.getSizeType());
3934 addFieldToRecordDecl(C, KmpDependInfoRD, FlagsTy);
3935 KmpDependInfoRD->completeDefinition();
3936 KmpDependInfoTy = C.getRecordType(KmpDependInfoRD);
3938 KmpDependInfoRD = cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
3939 CharUnits DependencySize = C.getTypeSizeInChars(KmpDependInfoTy);
3940 // Define type kmp_depend_info[<Dependences.size()>];
3941 QualType KmpDependInfoArrayTy = C.getConstantArrayType(
3942 KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies),
3943 ArrayType::Normal, /*IndexTypeQuals=*/0);
3944 // kmp_depend_info[<Dependences.size()>] deps;
3946 CGF.CreateMemTemp(KmpDependInfoArrayTy, ".dep.arr.addr");
3947 for (unsigned i = 0; i < NumDependencies; ++i) {
3948 const Expr *E = Data.Dependences[i].second;
3949 auto Addr = CGF.EmitLValue(E);
3951 QualType Ty = E->getType();
3952 if (auto *ASE = dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) {
3954 CGF.EmitOMPArraySectionExpr(ASE, /*LowerBound=*/false);
3955 llvm::Value *UpAddr =
3956 CGF.Builder.CreateConstGEP1_32(UpAddrLVal.getPointer(), /*Idx0=*/1);
3957 llvm::Value *LowIntPtr =
3958 CGF.Builder.CreatePtrToInt(Addr.getPointer(), CGM.SizeTy);
3959 llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGM.SizeTy);
3960 Size = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr);
3962 Size = CGF.getTypeSize(Ty);
3963 auto Base = CGF.MakeAddrLValue(
3964 CGF.Builder.CreateConstArrayGEP(DependenciesArray, i, DependencySize),
3966 // deps[i].base_addr = &<Dependences[i].second>;
3967 auto BaseAddrLVal = CGF.EmitLValueForField(
3968 Base, *std::next(KmpDependInfoRD->field_begin(), BaseAddr));
3969 CGF.EmitStoreOfScalar(
3970 CGF.Builder.CreatePtrToInt(Addr.getPointer(), CGF.IntPtrTy),
3972 // deps[i].len = sizeof(<Dependences[i].second>);
3973 auto LenLVal = CGF.EmitLValueForField(
3974 Base, *std::next(KmpDependInfoRD->field_begin(), Len));
3975 CGF.EmitStoreOfScalar(Size, LenLVal);
3976 // deps[i].flags = <Dependences[i].first>;
3977 RTLDependenceKindTy DepKind;
3978 switch (Data.Dependences[i].first) {
3979 case OMPC_DEPEND_in:
3982 // Out and InOut dependencies must use the same code.
3983 case OMPC_DEPEND_out:
3984 case OMPC_DEPEND_inout:
3987 case OMPC_DEPEND_source:
3988 case OMPC_DEPEND_sink:
3989 case OMPC_DEPEND_unknown:
3990 llvm_unreachable("Unknown task dependence type");
3992 auto FlagsLVal = CGF.EmitLValueForField(
3993 Base, *std::next(KmpDependInfoRD->field_begin(), Flags));
3994 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(LLVMFlagsTy, DepKind),
3997 DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3998 CGF.Builder.CreateStructGEP(DependenciesArray, 0, CharUnits::Zero()),
4002 // NOTE: routine and part_id fields are intialized by __kmpc_omp_task_alloc()
4004 // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid,
4005 // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,
4006 // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) if dependence
4007 // list is not empty
4008 auto *ThreadID = getThreadID(CGF, Loc);
4009 auto *UpLoc = emitUpdateLocation(CGF, Loc);
4010 llvm::Value *TaskArgs[] = { UpLoc, ThreadID, NewTask };
4011 llvm::Value *DepTaskArgs[7];
4012 if (NumDependencies) {
4013 DepTaskArgs[0] = UpLoc;
4014 DepTaskArgs[1] = ThreadID;
4015 DepTaskArgs[2] = NewTask;
4016 DepTaskArgs[3] = CGF.Builder.getInt32(NumDependencies);
4017 DepTaskArgs[4] = DependenciesArray.getPointer();
4018 DepTaskArgs[5] = CGF.Builder.getInt32(0);
4019 DepTaskArgs[6] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4021 auto &&ThenCodeGen = [this, &Data, TDBase, KmpTaskTQTyRD, NumDependencies,
4023 &DepTaskArgs](CodeGenFunction &CGF, PrePostActionTy &) {
4025 auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId);
4026 auto PartIdLVal = CGF.EmitLValueForField(TDBase, *PartIdFI);
4027 CGF.EmitStoreOfScalar(CGF.Builder.getInt32(0), PartIdLVal);
4029 if (NumDependencies) {
4030 CGF.EmitRuntimeCall(
4031 createRuntimeFunction(OMPRTL__kmpc_omp_task_with_deps), DepTaskArgs);
4033 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_task),
4036 // Check if parent region is untied and build return for untied task;
4038 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
4039 Region->emitUntiedSwitch(CGF);
4042 llvm::Value *DepWaitTaskArgs[6];
4043 if (NumDependencies) {
4044 DepWaitTaskArgs[0] = UpLoc;
4045 DepWaitTaskArgs[1] = ThreadID;
4046 DepWaitTaskArgs[2] = CGF.Builder.getInt32(NumDependencies);
4047 DepWaitTaskArgs[3] = DependenciesArray.getPointer();
4048 DepWaitTaskArgs[4] = CGF.Builder.getInt32(0);
4049 DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4051 auto &&ElseCodeGen = [&TaskArgs, ThreadID, NewTaskNewTaskTTy, TaskEntry,
4052 NumDependencies, &DepWaitTaskArgs](CodeGenFunction &CGF,
4053 PrePostActionTy &) {
4054 auto &RT = CGF.CGM.getOpenMPRuntime();
4055 CodeGenFunction::RunCleanupsScope LocalScope(CGF);
4056 // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid,
4057 // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32
4058 // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info
4060 if (NumDependencies)
4061 CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__kmpc_omp_wait_deps),
4063 // Call proxy_task_entry(gtid, new_task);
4064 auto &&CodeGen = [TaskEntry, ThreadID, NewTaskNewTaskTTy](
4065 CodeGenFunction &CGF, PrePostActionTy &Action) {
4067 llvm::Value *OutlinedFnArgs[] = {ThreadID, NewTaskNewTaskTTy};
4068 CGF.EmitCallOrInvoke(TaskEntry, OutlinedFnArgs);
4071 // Build void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid,
4072 // kmp_task_t *new_task);
4073 // Build void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid,
4074 // kmp_task_t *new_task);
4075 RegionCodeGenTy RCG(CodeGen);
4076 CommonActionTy Action(
4077 RT.createRuntimeFunction(OMPRTL__kmpc_omp_task_begin_if0), TaskArgs,
4078 RT.createRuntimeFunction(OMPRTL__kmpc_omp_task_complete_if0), TaskArgs);
4079 RCG.setAction(Action);
4084 emitOMPIfClause(CGF, IfCond, ThenCodeGen, ElseCodeGen);
4086 RegionCodeGenTy ThenRCG(ThenCodeGen);
4091 void CGOpenMPRuntime::emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc,
4092 const OMPLoopDirective &D,
4093 llvm::Value *TaskFunction,
4094 QualType SharedsTy, Address Shareds,
4096 const OMPTaskDataTy &Data) {
4097 if (!CGF.HaveInsertPoint())
4099 TaskResultTy Result =
4100 emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data);
4101 // NOTE: routine and part_id fields are intialized by __kmpc_omp_task_alloc()
4103 // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int
4104 // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int
4105 // sched, kmp_uint64 grainsize, void *task_dup);
4106 llvm::Value *ThreadID = getThreadID(CGF, Loc);
4107 llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc);
4110 IfVal = CGF.Builder.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.IntTy,
4113 IfVal = llvm::ConstantInt::getSigned(CGF.IntTy, /*V=*/1);
4115 LValue LBLVal = CGF.EmitLValueForField(
4117 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound));
4119 cast<VarDecl>(cast<DeclRefExpr>(D.getLowerBoundVariable())->getDecl());
4120 CGF.EmitAnyExprToMem(LBVar->getInit(), LBLVal.getAddress(), LBLVal.getQuals(),
4121 /*IsInitializer=*/true);
4122 LValue UBLVal = CGF.EmitLValueForField(
4124 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound));
4126 cast<VarDecl>(cast<DeclRefExpr>(D.getUpperBoundVariable())->getDecl());
4127 CGF.EmitAnyExprToMem(UBVar->getInit(), UBLVal.getAddress(), UBLVal.getQuals(),
4128 /*IsInitializer=*/true);
4129 LValue StLVal = CGF.EmitLValueForField(
4131 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTStride));
4133 cast<VarDecl>(cast<DeclRefExpr>(D.getStrideVariable())->getDecl());
4134 CGF.EmitAnyExprToMem(StVar->getInit(), StLVal.getAddress(), StLVal.getQuals(),
4135 /*IsInitializer=*/true);
4136 enum { NoSchedule = 0, Grainsize = 1, NumTasks = 2 };
4137 llvm::Value *TaskArgs[] = {
4138 UpLoc, ThreadID, Result.NewTask, IfVal, LBLVal.getPointer(),
4139 UBLVal.getPointer(), CGF.EmitLoadOfScalar(StLVal, SourceLocation()),
4140 llvm::ConstantInt::getSigned(CGF.IntTy, Data.Nogroup ? 1 : 0),
4141 llvm::ConstantInt::getSigned(
4142 CGF.IntTy, Data.Schedule.getPointer()
4143 ? Data.Schedule.getInt() ? NumTasks : Grainsize
4145 Data.Schedule.getPointer()
4146 ? CGF.Builder.CreateIntCast(Data.Schedule.getPointer(), CGF.Int64Ty,
4148 : llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/0),
4150 ? CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Result.TaskDupFn,
4152 : llvm::ConstantPointerNull::get(CGF.VoidPtrTy)};
4153 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_taskloop), TaskArgs);
4156 /// \brief Emit reduction operation for each element of array (required for
4157 /// array sections) LHS op = RHS.
4158 /// \param Type Type of array.
4159 /// \param LHSVar Variable on the left side of the reduction operation
4160 /// (references element of array in original variable).
4161 /// \param RHSVar Variable on the right side of the reduction operation
4162 /// (references element of array in original variable).
4163 /// \param RedOpGen Generator of reduction operation with use of LHSVar and
4165 static void EmitOMPAggregateReduction(
4166 CodeGenFunction &CGF, QualType Type, const VarDecl *LHSVar,
4167 const VarDecl *RHSVar,
4168 const llvm::function_ref<void(CodeGenFunction &CGF, const Expr *,
4169 const Expr *, const Expr *)> &RedOpGen,
4170 const Expr *XExpr = nullptr, const Expr *EExpr = nullptr,
4171 const Expr *UpExpr = nullptr) {
4172 // Perform element-by-element initialization.
4174 Address LHSAddr = CGF.GetAddrOfLocalVar(LHSVar);
4175 Address RHSAddr = CGF.GetAddrOfLocalVar(RHSVar);
4177 // Drill down to the base element type on both arrays.
4178 auto ArrayTy = Type->getAsArrayTypeUnsafe();
4179 auto NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, LHSAddr);
4181 auto RHSBegin = RHSAddr.getPointer();
4182 auto LHSBegin = LHSAddr.getPointer();
4183 // Cast from pointer to array type to pointer to single element.
4184 auto LHSEnd = CGF.Builder.CreateGEP(LHSBegin, NumElements);
4185 // The basic structure here is a while-do loop.
4186 auto BodyBB = CGF.createBasicBlock("omp.arraycpy.body");
4187 auto DoneBB = CGF.createBasicBlock("omp.arraycpy.done");
4189 CGF.Builder.CreateICmpEQ(LHSBegin, LHSEnd, "omp.arraycpy.isempty");
4190 CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
4192 // Enter the loop body, making that address the current address.
4193 auto EntryBB = CGF.Builder.GetInsertBlock();
4194 CGF.EmitBlock(BodyBB);
4196 CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy);
4198 llvm::PHINode *RHSElementPHI = CGF.Builder.CreatePHI(
4199 RHSBegin->getType(), 2, "omp.arraycpy.srcElementPast");
4200 RHSElementPHI->addIncoming(RHSBegin, EntryBB);
4201 Address RHSElementCurrent =
4202 Address(RHSElementPHI,
4203 RHSAddr.getAlignment().alignmentOfArrayElement(ElementSize));
4205 llvm::PHINode *LHSElementPHI = CGF.Builder.CreatePHI(
4206 LHSBegin->getType(), 2, "omp.arraycpy.destElementPast");
4207 LHSElementPHI->addIncoming(LHSBegin, EntryBB);
4208 Address LHSElementCurrent =
4209 Address(LHSElementPHI,
4210 LHSAddr.getAlignment().alignmentOfArrayElement(ElementSize));
4213 CodeGenFunction::OMPPrivateScope Scope(CGF);
4214 Scope.addPrivate(LHSVar, [=]() -> Address { return LHSElementCurrent; });
4215 Scope.addPrivate(RHSVar, [=]() -> Address { return RHSElementCurrent; });
4217 RedOpGen(CGF, XExpr, EExpr, UpExpr);
4218 Scope.ForceCleanup();
4220 // Shift the address forward by one element.
4221 auto LHSElementNext = CGF.Builder.CreateConstGEP1_32(
4222 LHSElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element");
4223 auto RHSElementNext = CGF.Builder.CreateConstGEP1_32(
4224 RHSElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element");
4225 // Check whether we've reached the end.
4227 CGF.Builder.CreateICmpEQ(LHSElementNext, LHSEnd, "omp.arraycpy.done");
4228 CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB);
4229 LHSElementPHI->addIncoming(LHSElementNext, CGF.Builder.GetInsertBlock());
4230 RHSElementPHI->addIncoming(RHSElementNext, CGF.Builder.GetInsertBlock());
4233 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
4236 /// Emit reduction combiner. If the combiner is a simple expression emit it as
4237 /// is, otherwise consider it as combiner of UDR decl and emit it as a call of
4238 /// UDR combiner function.
4239 static void emitReductionCombiner(CodeGenFunction &CGF,
4240 const Expr *ReductionOp) {
4241 if (auto *CE = dyn_cast<CallExpr>(ReductionOp))
4242 if (auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee()))
4244 dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts()))
4245 if (auto *DRD = dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) {
4246 std::pair<llvm::Function *, llvm::Function *> Reduction =
4247 CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD);
4248 RValue Func = RValue::get(Reduction.first);
4249 CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func);
4250 CGF.EmitIgnoredExpr(ReductionOp);
4253 CGF.EmitIgnoredExpr(ReductionOp);
4256 llvm::Value *CGOpenMPRuntime::emitReductionFunction(
4257 CodeGenModule &CGM, llvm::Type *ArgsType, ArrayRef<const Expr *> Privates,
4258 ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
4259 ArrayRef<const Expr *> ReductionOps) {
4260 auto &C = CGM.getContext();
4262 // void reduction_func(void *LHSArg, void *RHSArg);
4263 FunctionArgList Args;
4264 ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, SourceLocation(), /*Id=*/nullptr,
4266 ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, SourceLocation(), /*Id=*/nullptr,
4268 Args.push_back(&LHSArg);
4269 Args.push_back(&RHSArg);
4270 auto &CGFI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
4271 auto *Fn = llvm::Function::Create(
4272 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
4273 ".omp.reduction.reduction_func", &CGM.getModule());
4274 CGM.SetInternalFunctionAttributes(/*D=*/nullptr, Fn, CGFI);
4275 CodeGenFunction CGF(CGM);
4276 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args);
4278 // Dst = (void*[n])(LHSArg);
4279 // Src = (void*[n])(RHSArg);
4280 Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4281 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)),
4282 ArgsType), CGF.getPointerAlign());
4283 Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4284 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)),
4285 ArgsType), CGF.getPointerAlign());
4288 // *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]);
4290 CodeGenFunction::OMPPrivateScope Scope(CGF);
4291 auto IPriv = Privates.begin();
4293 for (unsigned I = 0, E = ReductionOps.size(); I < E; ++I, ++IPriv, ++Idx) {
4294 auto RHSVar = cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl());
4295 Scope.addPrivate(RHSVar, [&]() -> Address {
4296 return emitAddrOfVarFromArray(CGF, RHS, Idx, RHSVar);
4298 auto LHSVar = cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl());
4299 Scope.addPrivate(LHSVar, [&]() -> Address {
4300 return emitAddrOfVarFromArray(CGF, LHS, Idx, LHSVar);
4302 QualType PrivTy = (*IPriv)->getType();
4303 if (PrivTy->isVariablyModifiedType()) {
4304 // Get array size and emit VLA type.
4307 CGF.Builder.CreateConstArrayGEP(LHS, Idx, CGF.getPointerSize());
4308 llvm::Value *Ptr = CGF.Builder.CreateLoad(Elem);
4309 auto *VLA = CGF.getContext().getAsVariableArrayType(PrivTy);
4310 auto *OVE = cast<OpaqueValueExpr>(VLA->getSizeExpr());
4311 CodeGenFunction::OpaqueValueMapping OpaqueMap(
4312 CGF, OVE, RValue::get(CGF.Builder.CreatePtrToInt(Ptr, CGF.SizeTy)));
4313 CGF.EmitVariablyModifiedType(PrivTy);
4317 IPriv = Privates.begin();
4318 auto ILHS = LHSExprs.begin();
4319 auto IRHS = RHSExprs.begin();
4320 for (auto *E : ReductionOps) {
4321 if ((*IPriv)->getType()->isArrayType()) {
4322 // Emit reduction for array section.
4323 auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
4324 auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
4325 EmitOMPAggregateReduction(
4326 CGF, (*IPriv)->getType(), LHSVar, RHSVar,
4327 [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) {
4328 emitReductionCombiner(CGF, E);
4331 // Emit reduction for array subscript or single variable.
4332 emitReductionCombiner(CGF, E);
4337 Scope.ForceCleanup();
4338 CGF.FinishFunction();
4342 void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF,
4343 const Expr *ReductionOp,
4344 const Expr *PrivateRef,
4345 const DeclRefExpr *LHS,
4346 const DeclRefExpr *RHS) {
4347 if (PrivateRef->getType()->isArrayType()) {
4348 // Emit reduction for array section.
4349 auto *LHSVar = cast<VarDecl>(LHS->getDecl());
4350 auto *RHSVar = cast<VarDecl>(RHS->getDecl());
4351 EmitOMPAggregateReduction(
4352 CGF, PrivateRef->getType(), LHSVar, RHSVar,
4353 [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) {
4354 emitReductionCombiner(CGF, ReductionOp);
4357 // Emit reduction for array subscript or single variable.
4358 emitReductionCombiner(CGF, ReductionOp);
4361 void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
4362 ArrayRef<const Expr *> Privates,
4363 ArrayRef<const Expr *> LHSExprs,
4364 ArrayRef<const Expr *> RHSExprs,
4365 ArrayRef<const Expr *> ReductionOps,
4366 ReductionOptionsTy Options) {
4367 if (!CGF.HaveInsertPoint())
4370 bool WithNowait = Options.WithNowait;
4371 bool SimpleReduction = Options.SimpleReduction;
4373 // Next code should be emitted for reduction:
4375 // static kmp_critical_name lock = { 0 };
4377 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
4378 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
4380 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
4381 // *(Type<n>-1*)rhs[<n>-1]);
4385 // void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]};
4386 // switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
4387 // RedList, reduce_func, &<lock>)) {
4390 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
4392 // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
4396 // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
4398 // [__kmpc_end_reduce(<loc>, <gtid>, &<lock>);]
4403 // if SimpleReduction is true, only the next code is generated:
4405 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
4408 auto &C = CGM.getContext();
4410 if (SimpleReduction) {
4411 CodeGenFunction::RunCleanupsScope Scope(CGF);
4412 auto IPriv = Privates.begin();
4413 auto ILHS = LHSExprs.begin();
4414 auto IRHS = RHSExprs.begin();
4415 for (auto *E : ReductionOps) {
4416 emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
4417 cast<DeclRefExpr>(*IRHS));
4425 // 1. Build a list of reduction variables.
4426 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
4427 auto Size = RHSExprs.size();
4428 for (auto *E : Privates) {
4429 if (E->getType()->isVariablyModifiedType())
4430 // Reserve place for array size.
4433 llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
4434 QualType ReductionArrayTy =
4435 C.getConstantArrayType(C.VoidPtrTy, ArraySize, ArrayType::Normal,
4436 /*IndexTypeQuals=*/0);
4437 Address ReductionList =
4438 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
4439 auto IPriv = Privates.begin();
4441 for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
4443 CGF.Builder.CreateConstArrayGEP(ReductionList, Idx, CGF.getPointerSize());
4444 CGF.Builder.CreateStore(
4445 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4446 CGF.EmitLValue(RHSExprs[I]).getPointer(), CGF.VoidPtrTy),
4448 if ((*IPriv)->getType()->isVariablyModifiedType()) {
4449 // Store array size.
4451 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx,
4452 CGF.getPointerSize());
4453 llvm::Value *Size = CGF.Builder.CreateIntCast(
4455 CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
4457 CGF.SizeTy, /*isSigned=*/false);
4458 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
4463 // 2. Emit reduce_func().
4464 auto *ReductionFn = emitReductionFunction(
4465 CGM, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates,
4466 LHSExprs, RHSExprs, ReductionOps);
4468 // 3. Create static kmp_critical_name lock = { 0 };
4469 auto *Lock = getCriticalRegionLock(".reduction");
4471 // 4. Build res = __kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
4472 // RedList, reduce_func, &<lock>);
4473 auto *IdentTLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE);
4474 auto *ThreadId = getThreadID(CGF, Loc);
4475 auto *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
4476 auto *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4477 ReductionList.getPointer(), CGF.VoidPtrTy);
4478 llvm::Value *Args[] = {
4479 IdentTLoc, // ident_t *<loc>
4480 ThreadId, // i32 <gtid>
4481 CGF.Builder.getInt32(RHSExprs.size()), // i32 <n>
4482 ReductionArrayTySize, // size_type sizeof(RedList)
4483 RL, // void *RedList
4484 ReductionFn, // void (*) (void *, void *) <reduce_func>
4485 Lock // kmp_critical_name *&<lock>
4487 auto Res = CGF.EmitRuntimeCall(
4488 createRuntimeFunction(WithNowait ? OMPRTL__kmpc_reduce_nowait
4489 : OMPRTL__kmpc_reduce),
4492 // 5. Build switch(res)
4493 auto *DefaultBB = CGF.createBasicBlock(".omp.reduction.default");
4494 auto *SwInst = CGF.Builder.CreateSwitch(Res, DefaultBB, /*NumCases=*/2);
4498 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
4500 // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
4502 auto *Case1BB = CGF.createBasicBlock(".omp.reduction.case1");
4503 SwInst->addCase(CGF.Builder.getInt32(1), Case1BB);
4504 CGF.EmitBlock(Case1BB);
4506 // Add emission of __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
4507 llvm::Value *EndArgs[] = {
4508 IdentTLoc, // ident_t *<loc>
4509 ThreadId, // i32 <gtid>
4510 Lock // kmp_critical_name *&<lock>
4512 auto &&CodeGen = [&Privates, &LHSExprs, &RHSExprs, &ReductionOps](
4513 CodeGenFunction &CGF, PrePostActionTy &Action) {
4514 auto &RT = CGF.CGM.getOpenMPRuntime();
4515 auto IPriv = Privates.begin();
4516 auto ILHS = LHSExprs.begin();
4517 auto IRHS = RHSExprs.begin();
4518 for (auto *E : ReductionOps) {
4519 RT.emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
4520 cast<DeclRefExpr>(*IRHS));
4526 RegionCodeGenTy RCG(CodeGen);
4527 CommonActionTy Action(
4528 nullptr, llvm::None,
4529 createRuntimeFunction(WithNowait ? OMPRTL__kmpc_end_reduce_nowait
4530 : OMPRTL__kmpc_end_reduce),
4532 RCG.setAction(Action);
4535 CGF.EmitBranch(DefaultBB);
4539 // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
4542 auto *Case2BB = CGF.createBasicBlock(".omp.reduction.case2");
4543 SwInst->addCase(CGF.Builder.getInt32(2), Case2BB);
4544 CGF.EmitBlock(Case2BB);
4546 auto &&AtomicCodeGen = [Loc, &Privates, &LHSExprs, &RHSExprs, &ReductionOps](
4547 CodeGenFunction &CGF, PrePostActionTy &Action) {
4548 auto ILHS = LHSExprs.begin();
4549 auto IRHS = RHSExprs.begin();
4550 auto IPriv = Privates.begin();
4551 for (auto *E : ReductionOps) {
4552 const Expr *XExpr = nullptr;
4553 const Expr *EExpr = nullptr;
4554 const Expr *UpExpr = nullptr;
4555 BinaryOperatorKind BO = BO_Comma;
4556 if (auto *BO = dyn_cast<BinaryOperator>(E)) {
4557 if (BO->getOpcode() == BO_Assign) {
4558 XExpr = BO->getLHS();
4559 UpExpr = BO->getRHS();
4562 // Try to emit update expression as a simple atomic.
4563 auto *RHSExpr = UpExpr;
4565 // Analyze RHS part of the whole expression.
4566 if (auto *ACO = dyn_cast<AbstractConditionalOperator>(
4567 RHSExpr->IgnoreParenImpCasts())) {
4568 // If this is a conditional operator, analyze its condition for
4569 // min/max reduction operator.
4570 RHSExpr = ACO->getCond();
4573 dyn_cast<BinaryOperator>(RHSExpr->IgnoreParenImpCasts())) {
4574 EExpr = BORHS->getRHS();
4575 BO = BORHS->getOpcode();
4579 auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
4580 auto &&AtomicRedGen = [BO, VD,
4581 Loc](CodeGenFunction &CGF, const Expr *XExpr,
4582 const Expr *EExpr, const Expr *UpExpr) {
4583 LValue X = CGF.EmitLValue(XExpr);
4586 E = CGF.EmitAnyExpr(EExpr);
4587 CGF.EmitOMPAtomicSimpleUpdateExpr(
4588 X, E, BO, /*IsXLHSInRHSPart=*/true,
4589 llvm::AtomicOrdering::Monotonic, Loc,
4590 [&CGF, UpExpr, VD, Loc](RValue XRValue) {
4591 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4592 PrivateScope.addPrivate(
4593 VD, [&CGF, VD, XRValue, Loc]() -> Address {
4594 Address LHSTemp = CGF.CreateMemTemp(VD->getType());
4595 CGF.emitOMPSimpleStore(
4596 CGF.MakeAddrLValue(LHSTemp, VD->getType()), XRValue,
4597 VD->getType().getNonReferenceType(), Loc);
4600 (void)PrivateScope.Privatize();
4601 return CGF.EmitAnyExpr(UpExpr);
4604 if ((*IPriv)->getType()->isArrayType()) {
4605 // Emit atomic reduction for array section.
4606 auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
4607 EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), VD, RHSVar,
4608 AtomicRedGen, XExpr, EExpr, UpExpr);
4610 // Emit atomic reduction for array subscript or single variable.
4611 AtomicRedGen(CGF, XExpr, EExpr, UpExpr);
4613 // Emit as a critical region.
4614 auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *,
4615 const Expr *, const Expr *) {
4616 auto &RT = CGF.CGM.getOpenMPRuntime();
4617 RT.emitCriticalRegion(
4618 CGF, ".atomic_reduction",
4619 [=](CodeGenFunction &CGF, PrePostActionTy &Action) {
4621 emitReductionCombiner(CGF, E);
4625 if ((*IPriv)->getType()->isArrayType()) {
4626 auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
4627 auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
4628 EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), LHSVar, RHSVar,
4631 CritRedGen(CGF, nullptr, nullptr, nullptr);
4638 RegionCodeGenTy AtomicRCG(AtomicCodeGen);
4640 // Add emission of __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
4641 llvm::Value *EndArgs[] = {
4642 IdentTLoc, // ident_t *<loc>
4643 ThreadId, // i32 <gtid>
4644 Lock // kmp_critical_name *&<lock>
4646 CommonActionTy Action(nullptr, llvm::None,
4647 createRuntimeFunction(OMPRTL__kmpc_end_reduce),
4649 AtomicRCG.setAction(Action);
4654 CGF.EmitBranch(DefaultBB);
4655 CGF.EmitBlock(DefaultBB, /*IsFinished=*/true);
4658 void CGOpenMPRuntime::emitTaskwaitCall(CodeGenFunction &CGF,
4659 SourceLocation Loc) {
4660 if (!CGF.HaveInsertPoint())
4662 // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
4664 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
4665 // Ignore return result until untied tasks are supported.
4666 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_taskwait), Args);
4667 if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
4668 Region->emitUntiedSwitch(CGF);
4671 void CGOpenMPRuntime::emitInlinedDirective(CodeGenFunction &CGF,
4672 OpenMPDirectiveKind InnerKind,
4673 const RegionCodeGenTy &CodeGen,
4675 if (!CGF.HaveInsertPoint())
4677 InlinedOpenMPRegionRAII Region(CGF, CodeGen, InnerKind, HasCancel);
4678 CGF.CapturedStmtInfo->EmitBody(CGF, /*S=*/nullptr);
4689 } // anonymous namespace
4691 static RTCancelKind getCancellationKind(OpenMPDirectiveKind CancelRegion) {
4692 RTCancelKind CancelKind = CancelNoreq;
4693 if (CancelRegion == OMPD_parallel)
4694 CancelKind = CancelParallel;
4695 else if (CancelRegion == OMPD_for)
4696 CancelKind = CancelLoop;
4697 else if (CancelRegion == OMPD_sections)
4698 CancelKind = CancelSections;
4700 assert(CancelRegion == OMPD_taskgroup);
4701 CancelKind = CancelTaskgroup;
4706 void CGOpenMPRuntime::emitCancellationPointCall(
4707 CodeGenFunction &CGF, SourceLocation Loc,
4708 OpenMPDirectiveKind CancelRegion) {
4709 if (!CGF.HaveInsertPoint())
4711 // Build call kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32
4712 // global_tid, kmp_int32 cncl_kind);
4713 if (auto *OMPRegionInfo =
4714 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
4715 // For 'cancellation point taskgroup', the task region info may not have a
4716 // cancel. This may instead happen in another adjacent task.
4717 if (CancelRegion == OMPD_taskgroup || OMPRegionInfo->hasCancel()) {
4718 llvm::Value *Args[] = {
4719 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
4720 CGF.Builder.getInt32(getCancellationKind(CancelRegion))};
4721 // Ignore return result until untied tasks are supported.
4722 auto *Result = CGF.EmitRuntimeCall(
4723 createRuntimeFunction(OMPRTL__kmpc_cancellationpoint), Args);
4724 // if (__kmpc_cancellationpoint()) {
4725 // exit from construct;
4727 auto *ExitBB = CGF.createBasicBlock(".cancel.exit");
4728 auto *ContBB = CGF.createBasicBlock(".cancel.continue");
4729 auto *Cmp = CGF.Builder.CreateIsNotNull(Result);
4730 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
4731 CGF.EmitBlock(ExitBB);
4732 // exit from construct;
4734 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
4735 CGF.EmitBranchThroughCleanup(CancelDest);
4736 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
4741 void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc,
4743 OpenMPDirectiveKind CancelRegion) {
4744 if (!CGF.HaveInsertPoint())
4746 // Build call kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid,
4747 // kmp_int32 cncl_kind);
4748 if (auto *OMPRegionInfo =
4749 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
4750 auto &&ThenGen = [Loc, CancelRegion, OMPRegionInfo](CodeGenFunction &CGF,
4751 PrePostActionTy &) {
4752 auto &RT = CGF.CGM.getOpenMPRuntime();
4753 llvm::Value *Args[] = {
4754 RT.emitUpdateLocation(CGF, Loc), RT.getThreadID(CGF, Loc),
4755 CGF.Builder.getInt32(getCancellationKind(CancelRegion))};
4756 // Ignore return result until untied tasks are supported.
4757 auto *Result = CGF.EmitRuntimeCall(
4758 RT.createRuntimeFunction(OMPRTL__kmpc_cancel), Args);
4759 // if (__kmpc_cancel()) {
4760 // exit from construct;
4762 auto *ExitBB = CGF.createBasicBlock(".cancel.exit");
4763 auto *ContBB = CGF.createBasicBlock(".cancel.continue");
4764 auto *Cmp = CGF.Builder.CreateIsNotNull(Result);
4765 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
4766 CGF.EmitBlock(ExitBB);
4767 // exit from construct;
4769 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
4770 CGF.EmitBranchThroughCleanup(CancelDest);
4771 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
4774 emitOMPIfClause(CGF, IfCond, ThenGen,
4775 [](CodeGenFunction &, PrePostActionTy &) {});
4777 RegionCodeGenTy ThenRCG(ThenGen);
4783 /// \brief Obtain information that uniquely identifies a target entry. This
4784 /// consists of the file and device IDs as well as line number associated with
4785 /// the relevant entry source location.
4786 static void getTargetEntryUniqueInfo(ASTContext &C, SourceLocation Loc,
4787 unsigned &DeviceID, unsigned &FileID,
4788 unsigned &LineNum) {
4790 auto &SM = C.getSourceManager();
4792 // The loc should be always valid and have a file ID (the user cannot use
4793 // #pragma directives in macros)
4795 assert(Loc.isValid() && "Source location is expected to be always valid.");
4796 assert(Loc.isFileID() && "Source location is expected to refer to a file.");
4798 PresumedLoc PLoc = SM.getPresumedLoc(Loc);
4799 assert(PLoc.isValid() && "Source location is expected to be always valid.");
4801 llvm::sys::fs::UniqueID ID;
4802 if (llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID))
4803 llvm_unreachable("Source file with target region no longer exists!");
4805 DeviceID = ID.getDevice();
4806 FileID = ID.getFile();
4807 LineNum = PLoc.getLine();
4810 void CGOpenMPRuntime::emitTargetOutlinedFunction(
4811 const OMPExecutableDirective &D, StringRef ParentName,
4812 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
4813 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
4814 assert(!ParentName.empty() && "Invalid target region parent name!");
4816 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
4817 IsOffloadEntry, CodeGen);
4820 void CGOpenMPRuntime::emitTargetOutlinedFunctionHelper(
4821 const OMPExecutableDirective &D, StringRef ParentName,
4822 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
4823 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
4824 // Create a unique name for the entry function using the source location
4825 // information of the current target region. The name will be something like:
4827 // __omp_offloading_DD_FFFF_PP_lBB
4829 // where DD_FFFF is an ID unique to the file (device and file IDs), PP is the
4830 // mangled name of the function that encloses the target region and BB is the
4831 // line number of the target region.
4836 getTargetEntryUniqueInfo(CGM.getContext(), D.getLocStart(), DeviceID, FileID,
4838 SmallString<64> EntryFnName;
4840 llvm::raw_svector_ostream OS(EntryFnName);
4841 OS << "__omp_offloading" << llvm::format("_%x", DeviceID)
4842 << llvm::format("_%x_", FileID) << ParentName << "_l" << Line;
4845 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
4847 CodeGenFunction CGF(CGM, true);
4848 CGOpenMPTargetRegionInfo CGInfo(CS, CodeGen, EntryFnName);
4849 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
4851 OutlinedFn = CGF.GenerateOpenMPCapturedStmtFunction(CS);
4853 // If this target outline function is not an offload entry, we don't need to
4855 if (!IsOffloadEntry)
4858 // The target region ID is used by the runtime library to identify the current
4859 // target region, so it only has to be unique and not necessarily point to
4860 // anything. It could be the pointer to the outlined function that implements
4861 // the target region, but we aren't using that so that the compiler doesn't
4862 // need to keep that, and could therefore inline the host function if proven
4863 // worthwhile during optimization. In the other hand, if emitting code for the
4864 // device, the ID has to be the function address so that it can retrieved from
4865 // the offloading entry and launched by the runtime library. We also mark the
4866 // outlined function to have external linkage in case we are emitting code for
4867 // the device, because these functions will be entry points to the device.
4869 if (CGM.getLangOpts().OpenMPIsDevice) {
4870 OutlinedFnID = llvm::ConstantExpr::getBitCast(OutlinedFn, CGM.Int8PtrTy);
4871 OutlinedFn->setLinkage(llvm::GlobalValue::ExternalLinkage);
4873 OutlinedFnID = new llvm::GlobalVariable(
4874 CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
4875 llvm::GlobalValue::PrivateLinkage,
4876 llvm::Constant::getNullValue(CGM.Int8Ty), ".omp_offload.region_id");
4878 // Register the information for the entry associated with this target region.
4879 OffloadEntriesInfoManager.registerTargetRegionEntryInfo(
4880 DeviceID, FileID, ParentName, Line, OutlinedFn, OutlinedFnID,
4884 /// discard all CompoundStmts intervening between two constructs
4885 static const Stmt *ignoreCompoundStmts(const Stmt *Body) {
4886 while (auto *CS = dyn_cast_or_null<CompoundStmt>(Body))
4887 Body = CS->body_front();
4892 /// Emit the number of teams for a target directive. Inspect the num_teams
4893 /// clause associated with a teams construct combined or closely nested
4894 /// with the target directive.
4896 /// Emit a team of size one for directives such as 'target parallel' that
4897 /// have no associated teams construct.
4899 /// Otherwise, return nullptr.
4900 static llvm::Value *
4901 emitNumTeamsForTargetDirective(CGOpenMPRuntime &OMPRuntime,
4902 CodeGenFunction &CGF,
4903 const OMPExecutableDirective &D) {
4905 assert(!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the "
4906 "teams directive expected to be "
4907 "emitted only for the host!");
4909 auto &Bld = CGF.Builder;
4911 // If the target directive is combined with a teams directive:
4912 // Return the value in the num_teams clause, if any.
4913 // Otherwise, return 0 to denote the runtime default.
4914 if (isOpenMPTeamsDirective(D.getDirectiveKind())) {
4915 if (const auto *NumTeamsClause = D.getSingleClause<OMPNumTeamsClause>()) {
4916 CodeGenFunction::RunCleanupsScope NumTeamsScope(CGF);
4917 auto NumTeams = CGF.EmitScalarExpr(NumTeamsClause->getNumTeams(),
4918 /*IgnoreResultAssign*/ true);
4919 return Bld.CreateIntCast(NumTeams, CGF.Int32Ty,
4923 // The default value is 0.
4924 return Bld.getInt32(0);
4927 // If the target directive is combined with a parallel directive but not a
4928 // teams directive, start one team.
4929 if (isOpenMPParallelDirective(D.getDirectiveKind()))
4930 return Bld.getInt32(1);
4932 // If the current target region has a teams region enclosed, we need to get
4933 // the number of teams to pass to the runtime function call. This is done
4934 // by generating the expression in a inlined region. This is required because
4935 // the expression is captured in the enclosing target environment when the
4936 // teams directive is not combined with target.
4938 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
4940 // FIXME: Accommodate other combined directives with teams when they become
4942 if (auto *TeamsDir = dyn_cast_or_null<OMPTeamsDirective>(
4943 ignoreCompoundStmts(CS.getCapturedStmt()))) {
4944 if (auto *NTE = TeamsDir->getSingleClause<OMPNumTeamsClause>()) {
4945 CGOpenMPInnerExprInfo CGInfo(CGF, CS);
4946 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
4947 llvm::Value *NumTeams = CGF.EmitScalarExpr(NTE->getNumTeams());
4948 return Bld.CreateIntCast(NumTeams, CGF.Int32Ty,
4952 // If we have an enclosed teams directive but no num_teams clause we use
4953 // the default value 0.
4954 return Bld.getInt32(0);
4957 // No teams associated with the directive.
4961 /// Emit the number of threads for a target directive. Inspect the
4962 /// thread_limit clause associated with a teams construct combined or closely
4963 /// nested with the target directive.
4965 /// Emit the num_threads clause for directives such as 'target parallel' that
4966 /// have no associated teams construct.
4968 /// Otherwise, return nullptr.
4969 static llvm::Value *
4970 emitNumThreadsForTargetDirective(CGOpenMPRuntime &OMPRuntime,
4971 CodeGenFunction &CGF,
4972 const OMPExecutableDirective &D) {
4974 assert(!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the "
4975 "teams directive expected to be "
4976 "emitted only for the host!");
4978 auto &Bld = CGF.Builder;
4981 // If the target directive is combined with a teams directive:
4982 // Return the value in the thread_limit clause, if any.
4984 // If the target directive is combined with a parallel directive:
4985 // Return the value in the num_threads clause, if any.
4987 // If both clauses are set, select the minimum of the two.
4989 // If neither teams or parallel combined directives set the number of threads
4990 // in a team, return 0 to denote the runtime default.
4992 // If this is not a teams directive return nullptr.
4994 if (isOpenMPTeamsDirective(D.getDirectiveKind()) ||
4995 isOpenMPParallelDirective(D.getDirectiveKind())) {
4996 llvm::Value *DefaultThreadLimitVal = Bld.getInt32(0);
4997 llvm::Value *NumThreadsVal = nullptr;
4998 llvm::Value *ThreadLimitVal = nullptr;
5000 if (const auto *ThreadLimitClause =
5001 D.getSingleClause<OMPThreadLimitClause>()) {
5002 CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF);
5003 auto ThreadLimit = CGF.EmitScalarExpr(ThreadLimitClause->getThreadLimit(),
5004 /*IgnoreResultAssign*/ true);
5005 ThreadLimitVal = Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty,
5009 if (const auto *NumThreadsClause =
5010 D.getSingleClause<OMPNumThreadsClause>()) {
5011 CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
5012 llvm::Value *NumThreads =
5013 CGF.EmitScalarExpr(NumThreadsClause->getNumThreads(),
5014 /*IgnoreResultAssign*/ true);
5016 Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*IsSigned=*/true);
5019 // Select the lesser of thread_limit and num_threads.
5021 ThreadLimitVal = ThreadLimitVal
5022 ? Bld.CreateSelect(Bld.CreateICmpSLT(NumThreadsVal,
5024 NumThreadsVal, ThreadLimitVal)
5027 // Set default value passed to the runtime if either teams or a target
5028 // parallel type directive is found but no clause is specified.
5029 if (!ThreadLimitVal)
5030 ThreadLimitVal = DefaultThreadLimitVal;
5032 return ThreadLimitVal;
5035 // If the current target region has a teams region enclosed, we need to get
5036 // the thread limit to pass to the runtime function call. This is done
5037 // by generating the expression in a inlined region. This is required because
5038 // the expression is captured in the enclosing target environment when the
5039 // teams directive is not combined with target.
5041 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
5043 // FIXME: Accommodate other combined directives with teams when they become
5045 if (auto *TeamsDir = dyn_cast_or_null<OMPTeamsDirective>(
5046 ignoreCompoundStmts(CS.getCapturedStmt()))) {
5047 if (auto *TLE = TeamsDir->getSingleClause<OMPThreadLimitClause>()) {
5048 CGOpenMPInnerExprInfo CGInfo(CGF, CS);
5049 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
5050 llvm::Value *ThreadLimit = CGF.EmitScalarExpr(TLE->getThreadLimit());
5051 return CGF.Builder.CreateIntCast(ThreadLimit, CGF.Int32Ty,
5055 // If we have an enclosed teams directive but no thread_limit clause we use
5056 // the default value 0.
5057 return CGF.Builder.getInt32(0);
5060 // No teams associated with the directive.
5065 // \brief Utility to handle information from clauses associated with a given
5066 // construct that use mappable expressions (e.g. 'map' clause, 'to' clause).
5067 // It provides a convenient interface to obtain the information and generate
5068 // code for that information.
5069 class MappableExprsHandler {
5071 /// \brief Values for bit flags used to specify the mapping type for
5073 enum OpenMPOffloadMappingFlags {
5074 /// \brief Allocate memory on the device and move data from host to device.
5076 /// \brief Allocate memory on the device and move data from device to host.
5077 OMP_MAP_FROM = 0x02,
5078 /// \brief Always perform the requested mapping action on the element, even
5079 /// if it was already mapped before.
5080 OMP_MAP_ALWAYS = 0x04,
5081 /// \brief Delete the element from the device environment, ignoring the
5082 /// current reference count associated with the element.
5083 OMP_MAP_DELETE = 0x08,
5084 /// \brief The element being mapped is a pointer, therefore the pointee
5085 /// should be mapped as well.
5086 OMP_MAP_IS_PTR = 0x10,
5087 /// \brief This flags signals that an argument is the first one relating to
5088 /// a map/private clause expression. For some cases a single
5089 /// map/privatization results in multiple arguments passed to the runtime
5091 OMP_MAP_FIRST_REF = 0x20,
5092 /// \brief Signal that the runtime library has to return the device pointer
5093 /// in the current position for the data being mapped.
5094 OMP_MAP_RETURN_PTR = 0x40,
5095 /// \brief This flag signals that the reference being passed is a pointer to
5097 OMP_MAP_PRIVATE_PTR = 0x80,
5098 /// \brief Pass the element to the device by value.
5099 OMP_MAP_PRIVATE_VAL = 0x100,
5102 /// Class that associates information with a base pointer to be passed to the
5103 /// runtime library.
5104 class BasePointerInfo {
5105 /// The base pointer.
5106 llvm::Value *Ptr = nullptr;
5107 /// The base declaration that refers to this device pointer, or null if
5109 const ValueDecl *DevPtrDecl = nullptr;
5112 BasePointerInfo(llvm::Value *Ptr, const ValueDecl *DevPtrDecl = nullptr)
5113 : Ptr(Ptr), DevPtrDecl(DevPtrDecl) {}
5114 llvm::Value *operator*() const { return Ptr; }
5115 const ValueDecl *getDevicePtrDecl() const { return DevPtrDecl; }
5116 void setDevicePtrDecl(const ValueDecl *D) { DevPtrDecl = D; }
5119 typedef SmallVector<BasePointerInfo, 16> MapBaseValuesArrayTy;
5120 typedef SmallVector<llvm::Value *, 16> MapValuesArrayTy;
5121 typedef SmallVector<unsigned, 16> MapFlagsArrayTy;
5124 /// \brief Directive from where the map clauses were extracted.
5125 const OMPExecutableDirective &CurDir;
5127 /// \brief Function the directive is being generated for.
5128 CodeGenFunction &CGF;
5130 /// \brief Set of all first private variables in the current directive.
5131 llvm::SmallPtrSet<const VarDecl *, 8> FirstPrivateDecls;
5133 /// Map between device pointer declarations and their expression components.
5134 /// The key value for declarations in 'this' is null.
5137 SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>>
5140 llvm::Value *getExprTypeSize(const Expr *E) const {
5141 auto ExprTy = E->getType().getCanonicalType();
5143 // Reference types are ignored for mapping purposes.
5144 if (auto *RefTy = ExprTy->getAs<ReferenceType>())
5145 ExprTy = RefTy->getPointeeType().getCanonicalType();
5147 // Given that an array section is considered a built-in type, we need to
5148 // do the calculation based on the length of the section instead of relying
5149 // on CGF.getTypeSize(E->getType()).
5150 if (const auto *OAE = dyn_cast<OMPArraySectionExpr>(E)) {
5151 QualType BaseTy = OMPArraySectionExpr::getBaseOriginalType(
5152 OAE->getBase()->IgnoreParenImpCasts())
5153 .getCanonicalType();
5155 // If there is no length associated with the expression, that means we
5156 // are using the whole length of the base.
5157 if (!OAE->getLength() && OAE->getColonLoc().isValid())
5158 return CGF.getTypeSize(BaseTy);
5160 llvm::Value *ElemSize;
5161 if (auto *PTy = BaseTy->getAs<PointerType>())
5162 ElemSize = CGF.getTypeSize(PTy->getPointeeType().getCanonicalType());
5164 auto *ATy = cast<ArrayType>(BaseTy.getTypePtr());
5165 assert(ATy && "Expecting array type if not a pointer type.");
5166 ElemSize = CGF.getTypeSize(ATy->getElementType().getCanonicalType());
5169 // If we don't have a length at this point, that is because we have an
5170 // array section with a single element.
5171 if (!OAE->getLength())
5174 auto *LengthVal = CGF.EmitScalarExpr(OAE->getLength());
5176 CGF.Builder.CreateIntCast(LengthVal, CGF.SizeTy, /*isSigned=*/false);
5177 return CGF.Builder.CreateNUWMul(LengthVal, ElemSize);
5179 return CGF.getTypeSize(ExprTy);
5182 /// \brief Return the corresponding bits for a given map clause modifier. Add
5183 /// a flag marking the map as a pointer if requested. Add a flag marking the
5184 /// map as the first one of a series of maps that relate to the same map
5186 unsigned getMapTypeBits(OpenMPMapClauseKind MapType,
5187 OpenMPMapClauseKind MapTypeModifier, bool AddPtrFlag,
5188 bool AddIsFirstFlag) const {
5191 case OMPC_MAP_alloc:
5192 case OMPC_MAP_release:
5193 // alloc and release is the default behavior in the runtime library, i.e.
5194 // if we don't pass any bits alloc/release that is what the runtime is
5195 // going to do. Therefore, we don't need to signal anything for these two
5202 Bits = OMP_MAP_FROM;
5204 case OMPC_MAP_tofrom:
5205 Bits = OMP_MAP_TO | OMP_MAP_FROM;
5207 case OMPC_MAP_delete:
5208 Bits = OMP_MAP_DELETE;
5211 llvm_unreachable("Unexpected map type!");
5215 Bits |= OMP_MAP_IS_PTR;
5217 Bits |= OMP_MAP_FIRST_REF;
5218 if (MapTypeModifier == OMPC_MAP_always)
5219 Bits |= OMP_MAP_ALWAYS;
5223 /// \brief Return true if the provided expression is a final array section. A
5224 /// final array section, is one whose length can't be proved to be one.
5225 bool isFinalArraySectionExpression(const Expr *E) const {
5226 auto *OASE = dyn_cast<OMPArraySectionExpr>(E);
5228 // It is not an array section and therefore not a unity-size one.
5232 // An array section with no colon always refer to a single element.
5233 if (OASE->getColonLoc().isInvalid())
5236 auto *Length = OASE->getLength();
5238 // If we don't have a length we have to check if the array has size 1
5239 // for this dimension. Also, we should always expect a length if the
5240 // base type is pointer.
5242 auto BaseQTy = OMPArraySectionExpr::getBaseOriginalType(
5243 OASE->getBase()->IgnoreParenImpCasts())
5244 .getCanonicalType();
5245 if (auto *ATy = dyn_cast<ConstantArrayType>(BaseQTy.getTypePtr()))
5246 return ATy->getSize().getSExtValue() != 1;
5247 // If we don't have a constant dimension length, we have to consider
5248 // the current section as having any size, so it is not necessarily
5249 // unitary. If it happen to be unity size, that's user fault.
5253 // Check if the length evaluates to 1.
5254 llvm::APSInt ConstLength;
5255 if (!Length->EvaluateAsInt(ConstLength, CGF.getContext()))
5256 return true; // Can have more that size 1.
5258 return ConstLength.getSExtValue() != 1;
5261 /// \brief Generate the base pointers, section pointers, sizes and map type
5262 /// bits for the provided map type, map modifier, and expression components.
5263 /// \a IsFirstComponent should be set to true if the provided set of
5264 /// components is the first associated with a capture.
5265 void generateInfoForComponentList(
5266 OpenMPMapClauseKind MapType, OpenMPMapClauseKind MapTypeModifier,
5267 OMPClauseMappableExprCommon::MappableExprComponentListRef Components,
5268 MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers,
5269 MapValuesArrayTy &Sizes, MapFlagsArrayTy &Types,
5270 bool IsFirstComponentList) const {
5272 // The following summarizes what has to be generated for each map and the
5273 // types bellow. The generated information is expressed in this order:
5274 // base pointer, section pointer, size, flags
5275 // (to add to the ones that come from the map type and modifier).
5296 // &d, &d, sizeof(double), noflags
5299 // &i, &i, 100*sizeof(int), noflags
5302 // &i(=&i[0]), &i[1], 23*sizeof(int), noflags
5305 // &p, &p, sizeof(float*), noflags
5308 // p, &p[1], 24*sizeof(float), noflags
5311 // &s, &s, sizeof(S2), noflags
5314 // &s, &(s.i), sizeof(int), noflags
5317 // &s, &(s.i.f), 50*sizeof(int), noflags
5320 // &s, &(s.p), sizeof(double*), noflags
5322 // map(s.p[:22], s.a s.b)
5323 // &s, &(s.p), sizeof(double*), noflags
5324 // &(s.p), &(s.p[0]), 22*sizeof(double), ptr_flag + extra_flag
5327 // &s, &(s.ps), sizeof(S2*), noflags
5330 // &s, &(s.ps), sizeof(S2*), noflags
5331 // &(s.ps), &(s.ps->s.i), sizeof(int), ptr_flag + extra_flag
5334 // &s, &(s.ps), sizeof(S2*), noflags
5335 // &(s.ps), &(s.ps->ps), sizeof(S2*), ptr_flag + extra_flag
5337 // map(s.ps->ps->ps)
5338 // &s, &(s.ps), sizeof(S2*), noflags
5339 // &(s.ps), &(s.ps->ps), sizeof(S2*), ptr_flag + extra_flag
5340 // &(s.ps->ps), &(s.ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5342 // map(s.ps->ps->s.f[:22])
5343 // &s, &(s.ps), sizeof(S2*), noflags
5344 // &(s.ps), &(s.ps->ps), sizeof(S2*), ptr_flag + extra_flag
5345 // &(s.ps->ps), &(s.ps->ps->s.f[0]), 22*sizeof(float), ptr_flag + extra_flag
5348 // &ps, &ps, sizeof(S2*), noflags
5351 // ps, &(ps->i), sizeof(int), noflags
5354 // ps, &(ps->s.f[0]), 50*sizeof(float), noflags
5357 // ps, &(ps->p), sizeof(double*), noflags
5360 // ps, &(ps->p), sizeof(double*), noflags
5361 // &(ps->p), &(ps->p[0]), 22*sizeof(double), ptr_flag + extra_flag
5364 // ps, &(ps->ps), sizeof(S2*), noflags
5367 // ps, &(ps->ps), sizeof(S2*), noflags
5368 // &(ps->ps), &(ps->ps->s.i), sizeof(int), ptr_flag + extra_flag
5371 // ps, &(ps->ps), sizeof(S2*), noflags
5372 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5374 // map(ps->ps->ps->ps)
5375 // ps, &(ps->ps), sizeof(S2*), noflags
5376 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5377 // &(ps->ps->ps), &(ps->ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5379 // map(ps->ps->ps->s.f[:22])
5380 // ps, &(ps->ps), sizeof(S2*), noflags
5381 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), ptr_flag + extra_flag
5382 // &(ps->ps->ps), &(ps->ps->ps->s.f[0]), 22*sizeof(float), ptr_flag +
5385 // Track if the map information being generated is the first for a capture.
5386 bool IsCaptureFirstInfo = IsFirstComponentList;
5388 // Scan the components from the base to the complete expression.
5389 auto CI = Components.rbegin();
5390 auto CE = Components.rend();
5393 // Track if the map information being generated is the first for a list of
5395 bool IsExpressionFirstInfo = true;
5396 llvm::Value *BP = nullptr;
5398 if (auto *ME = dyn_cast<MemberExpr>(I->getAssociatedExpression())) {
5399 // The base is the 'this' pointer. The content of the pointer is going
5400 // to be the base of the field being mapped.
5401 BP = CGF.EmitScalarExpr(ME->getBase());
5403 // The base is the reference to the variable.
5405 BP = CGF.EmitLValue(cast<DeclRefExpr>(I->getAssociatedExpression()))
5408 // If the variable is a pointer and is being dereferenced (i.e. is not
5409 // the last component), the base has to be the pointer itself, not its
5410 // reference. References are ignored for mapping purposes.
5412 I->getAssociatedDeclaration()->getType().getNonReferenceType();
5413 if (Ty->isAnyPointerType() && std::next(I) != CE) {
5414 auto PtrAddr = CGF.MakeNaturalAlignAddrLValue(BP, Ty);
5415 BP = CGF.EmitLoadOfPointerLValue(PtrAddr.getAddress(),
5416 Ty->castAs<PointerType>())
5419 // We do not need to generate individual map information for the
5420 // pointer, it can be associated with the combined storage.
5425 for (; I != CE; ++I) {
5426 auto Next = std::next(I);
5428 // We need to generate the addresses and sizes if this is the last
5429 // component, if the component is a pointer or if it is an array section
5430 // whose length can't be proved to be one. If this is a pointer, it
5431 // becomes the base address for the following components.
5433 // A final array section, is one whose length can't be proved to be one.
5434 bool IsFinalArraySection =
5435 isFinalArraySectionExpression(I->getAssociatedExpression());
5437 // Get information on whether the element is a pointer. Have to do a
5438 // special treatment for array sections given that they are built-in
5441 dyn_cast<OMPArraySectionExpr>(I->getAssociatedExpression());
5444 OMPArraySectionExpr::getBaseOriginalType(OASE)
5446 ->isAnyPointerType()) ||
5447 I->getAssociatedExpression()->getType()->isAnyPointerType();
5449 if (Next == CE || IsPointer || IsFinalArraySection) {
5451 // If this is not the last component, we expect the pointer to be
5452 // associated with an array expression or member expression.
5453 assert((Next == CE ||
5454 isa<MemberExpr>(Next->getAssociatedExpression()) ||
5455 isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) ||
5456 isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) &&
5457 "Unexpected expression");
5459 auto *LB = CGF.EmitLValue(I->getAssociatedExpression()).getPointer();
5460 auto *Size = getExprTypeSize(I->getAssociatedExpression());
5462 // If we have a member expression and the current component is a
5463 // reference, we have to map the reference too. Whenever we have a
5464 // reference, the section that reference refers to is going to be a
5465 // load instruction from the storage assigned to the reference.
5466 if (isa<MemberExpr>(I->getAssociatedExpression()) &&
5467 I->getAssociatedDeclaration()->getType()->isReferenceType()) {
5468 auto *LI = cast<llvm::LoadInst>(LB);
5469 auto *RefAddr = LI->getPointerOperand();
5471 BasePointers.push_back(BP);
5472 Pointers.push_back(RefAddr);
5473 Sizes.push_back(CGF.getTypeSize(CGF.getContext().VoidPtrTy));
5474 Types.push_back(getMapTypeBits(
5475 /*MapType*/ OMPC_MAP_alloc, /*MapTypeModifier=*/OMPC_MAP_unknown,
5476 !IsExpressionFirstInfo, IsCaptureFirstInfo));
5477 IsExpressionFirstInfo = false;
5478 IsCaptureFirstInfo = false;
5479 // The reference will be the next base address.
5483 BasePointers.push_back(BP);
5484 Pointers.push_back(LB);
5485 Sizes.push_back(Size);
5487 // We need to add a pointer flag for each map that comes from the
5488 // same expression except for the first one. We also need to signal
5489 // this map is the first one that relates with the current capture
5490 // (there is a set of entries for each capture).
5491 Types.push_back(getMapTypeBits(MapType, MapTypeModifier,
5492 !IsExpressionFirstInfo,
5493 IsCaptureFirstInfo));
5495 // If we have a final array section, we are done with this expression.
5496 if (IsFinalArraySection)
5499 // The pointer becomes the base for the next element.
5503 IsExpressionFirstInfo = false;
5504 IsCaptureFirstInfo = false;
5510 /// \brief Return the adjusted map modifiers if the declaration a capture
5511 /// refers to appears in a first-private clause. This is expected to be used
5512 /// only with directives that start with 'target'.
5513 unsigned adjustMapModifiersForPrivateClauses(const CapturedStmt::Capture &Cap,
5514 unsigned CurrentModifiers) {
5515 assert(Cap.capturesVariable() && "Expected capture by reference only!");
5517 // A first private variable captured by reference will use only the
5518 // 'private ptr' and 'map to' flag. Return the right flags if the captured
5519 // declaration is known as first-private in this handler.
5520 if (FirstPrivateDecls.count(Cap.getCapturedVar()))
5521 return MappableExprsHandler::OMP_MAP_PRIVATE_PTR |
5522 MappableExprsHandler::OMP_MAP_TO;
5524 // We didn't modify anything.
5525 return CurrentModifiers;
5529 MappableExprsHandler(const OMPExecutableDirective &Dir, CodeGenFunction &CGF)
5530 : CurDir(Dir), CGF(CGF) {
5531 // Extract firstprivate clause information.
5532 for (const auto *C : Dir.getClausesOfKind<OMPFirstprivateClause>())
5533 for (const auto *D : C->varlists())
5534 FirstPrivateDecls.insert(
5535 cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl());
5536 // Extract device pointer clause information.
5537 for (const auto *C : Dir.getClausesOfKind<OMPIsDevicePtrClause>())
5538 for (auto L : C->component_lists())
5539 DevPointersMap[L.first].push_back(L.second);
5542 /// \brief Generate all the base pointers, section pointers, sizes and map
5543 /// types for the extracted mappable expressions. Also, for each item that
5544 /// relates with a device pointer, a pair of the relevant declaration and
5545 /// index where it occurs is appended to the device pointers info array.
5546 void generateAllInfo(MapBaseValuesArrayTy &BasePointers,
5547 MapValuesArrayTy &Pointers, MapValuesArrayTy &Sizes,
5548 MapFlagsArrayTy &Types) const {
5549 BasePointers.clear();
5555 /// Kind that defines how a device pointer has to be returned.
5556 enum ReturnPointerKind {
5557 // Don't have to return any pointer.
5559 // Pointer is the base of the declaration.
5561 // Pointer is a member of the base declaration - 'this'
5563 // Pointer is a reference and a member of the base declaration - 'this'
5564 RPK_MemberReference,
5566 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
5567 OpenMPMapClauseKind MapType;
5568 OpenMPMapClauseKind MapTypeModifier;
5569 ReturnPointerKind ReturnDevicePointer;
5572 : MapType(OMPC_MAP_unknown), MapTypeModifier(OMPC_MAP_unknown),
5573 ReturnDevicePointer(RPK_None) {}
5575 OMPClauseMappableExprCommon::MappableExprComponentListRef Components,
5576 OpenMPMapClauseKind MapType, OpenMPMapClauseKind MapTypeModifier,
5577 ReturnPointerKind ReturnDevicePointer)
5578 : Components(Components), MapType(MapType),
5579 MapTypeModifier(MapTypeModifier),
5580 ReturnDevicePointer(ReturnDevicePointer) {}
5583 // We have to process the component lists that relate with the same
5584 // declaration in a single chunk so that we can generate the map flags
5585 // correctly. Therefore, we organize all lists in a map.
5586 llvm::DenseMap<const ValueDecl *, SmallVector<MapInfo, 8>> Info;
5588 // Helper function to fill the information map for the different supported
5590 auto &&InfoGen = [&Info](
5592 OMPClauseMappableExprCommon::MappableExprComponentListRef L,
5593 OpenMPMapClauseKind MapType, OpenMPMapClauseKind MapModifier,
5594 MapInfo::ReturnPointerKind ReturnDevicePointer) {
5595 const ValueDecl *VD =
5596 D ? cast<ValueDecl>(D->getCanonicalDecl()) : nullptr;
5597 Info[VD].push_back({L, MapType, MapModifier, ReturnDevicePointer});
5600 // FIXME: MSVC 2013 seems to require this-> to find member CurDir.
5601 for (auto *C : this->CurDir.getClausesOfKind<OMPMapClause>())
5602 for (auto L : C->component_lists())
5603 InfoGen(L.first, L.second, C->getMapType(), C->getMapTypeModifier(),
5605 for (auto *C : this->CurDir.getClausesOfKind<OMPToClause>())
5606 for (auto L : C->component_lists())
5607 InfoGen(L.first, L.second, OMPC_MAP_to, OMPC_MAP_unknown,
5609 for (auto *C : this->CurDir.getClausesOfKind<OMPFromClause>())
5610 for (auto L : C->component_lists())
5611 InfoGen(L.first, L.second, OMPC_MAP_from, OMPC_MAP_unknown,
5614 // Look at the use_device_ptr clause information and mark the existing map
5615 // entries as such. If there is no map information for an entry in the
5616 // use_device_ptr list, we create one with map type 'alloc' and zero size
5617 // section. It is the user fault if that was not mapped before.
5618 // FIXME: MSVC 2013 seems to require this-> to find member CurDir.
5619 for (auto *C : this->CurDir.getClausesOfKind<OMPUseDevicePtrClause>())
5620 for (auto L : C->component_lists()) {
5621 assert(!L.second.empty() && "Not expecting empty list of components!");
5622 const ValueDecl *VD = L.second.back().getAssociatedDeclaration();
5623 VD = cast<ValueDecl>(VD->getCanonicalDecl());
5624 auto *IE = L.second.back().getAssociatedExpression();
5625 // If the first component is a member expression, we have to look into
5626 // 'this', which maps to null in the map of map information. Otherwise
5627 // look directly for the information.
5628 auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD);
5630 // We potentially have map information for this declaration already.
5631 // Look for the first set of components that refer to it.
5632 if (It != Info.end()) {
5633 auto CI = std::find_if(
5634 It->second.begin(), It->second.end(), [VD](const MapInfo &MI) {
5635 return MI.Components.back().getAssociatedDeclaration() == VD;
5637 // If we found a map entry, signal that the pointer has to be returned
5638 // and move on to the next declaration.
5639 if (CI != It->second.end()) {
5640 CI->ReturnDevicePointer = isa<MemberExpr>(IE)
5641 ? (VD->getType()->isReferenceType()
5642 ? MapInfo::RPK_MemberReference
5643 : MapInfo::RPK_Member)
5644 : MapInfo::RPK_Base;
5649 // We didn't find any match in our map information - generate a zero
5650 // size array section.
5651 // FIXME: MSVC 2013 seems to require this-> to find member CGF.
5654 .EmitLoadOfLValue(this->CGF.EmitLValue(IE), SourceLocation())
5656 BasePointers.push_back({Ptr, VD});
5657 Pointers.push_back(Ptr);
5658 Sizes.push_back(llvm::Constant::getNullValue(this->CGF.SizeTy));
5659 Types.push_back(OMP_MAP_RETURN_PTR | OMP_MAP_FIRST_REF);
5662 for (auto &M : Info) {
5663 // We need to know when we generate information for the first component
5664 // associated with a capture, because the mapping flags depend on it.
5665 bool IsFirstComponentList = true;
5666 for (MapInfo &L : M.second) {
5667 assert(!L.Components.empty() &&
5668 "Not expecting declaration with no component lists.");
5670 // Remember the current base pointer index.
5671 unsigned CurrentBasePointersIdx = BasePointers.size();
5672 // FIXME: MSVC 2013 seems to require this-> to find the member method.
5673 this->generateInfoForComponentList(L.MapType, L.MapTypeModifier,
5674 L.Components, BasePointers, Pointers,
5675 Sizes, Types, IsFirstComponentList);
5677 // If this entry relates with a device pointer, set the relevant
5678 // declaration and add the 'return pointer' flag.
5679 if (IsFirstComponentList &&
5680 L.ReturnDevicePointer != MapInfo::RPK_None) {
5681 // If the pointer is not the base of the map, we need to skip the
5682 // base. If it is a reference in a member field, we also need to skip
5683 // the map of the reference.
5684 if (L.ReturnDevicePointer != MapInfo::RPK_Base) {
5685 ++CurrentBasePointersIdx;
5686 if (L.ReturnDevicePointer == MapInfo::RPK_MemberReference)
5687 ++CurrentBasePointersIdx;
5689 assert(BasePointers.size() > CurrentBasePointersIdx &&
5690 "Unexpected number of mapped base pointers.");
5692 auto *RelevantVD = L.Components.back().getAssociatedDeclaration();
5693 assert(RelevantVD &&
5694 "No relevant declaration related with device pointer??");
5696 BasePointers[CurrentBasePointersIdx].setDevicePtrDecl(RelevantVD);
5697 Types[CurrentBasePointersIdx] |= OMP_MAP_RETURN_PTR;
5699 IsFirstComponentList = false;
5704 /// \brief Generate the base pointers, section pointers, sizes and map types
5705 /// associated to a given capture.
5706 void generateInfoForCapture(const CapturedStmt::Capture *Cap,
5708 MapBaseValuesArrayTy &BasePointers,
5709 MapValuesArrayTy &Pointers,
5710 MapValuesArrayTy &Sizes,
5711 MapFlagsArrayTy &Types) const {
5712 assert(!Cap->capturesVariableArrayType() &&
5713 "Not expecting to generate map info for a variable array type!");
5715 BasePointers.clear();
5720 // We need to know when we generating information for the first component
5721 // associated with a capture, because the mapping flags depend on it.
5722 bool IsFirstComponentList = true;
5724 const ValueDecl *VD =
5727 : cast<ValueDecl>(Cap->getCapturedVar()->getCanonicalDecl());
5729 // If this declaration appears in a is_device_ptr clause we just have to
5730 // pass the pointer by value. If it is a reference to a declaration, we just
5731 // pass its value, otherwise, if it is a member expression, we need to map
5734 auto It = DevPointersMap.find(VD);
5735 if (It != DevPointersMap.end()) {
5736 for (auto L : It->second) {
5737 generateInfoForComponentList(
5738 /*MapType=*/OMPC_MAP_to, /*MapTypeModifier=*/OMPC_MAP_unknown, L,
5739 BasePointers, Pointers, Sizes, Types, IsFirstComponentList);
5740 IsFirstComponentList = false;
5744 } else if (DevPointersMap.count(VD)) {
5745 BasePointers.push_back({Arg, VD});
5746 Pointers.push_back(Arg);
5747 Sizes.push_back(CGF.getTypeSize(CGF.getContext().VoidPtrTy));
5748 Types.push_back(OMP_MAP_PRIVATE_VAL | OMP_MAP_FIRST_REF);
5752 // FIXME: MSVC 2013 seems to require this-> to find member CurDir.
5753 for (auto *C : this->CurDir.getClausesOfKind<OMPMapClause>())
5754 for (auto L : C->decl_component_lists(VD)) {
5755 assert(L.first == VD &&
5756 "We got information for the wrong declaration??");
5757 assert(!L.second.empty() &&
5758 "Not expecting declaration with no component lists.");
5759 generateInfoForComponentList(C->getMapType(), C->getMapTypeModifier(),
5760 L.second, BasePointers, Pointers, Sizes,
5761 Types, IsFirstComponentList);
5762 IsFirstComponentList = false;
5768 /// \brief Generate the default map information for a given capture \a CI,
5769 /// record field declaration \a RI and captured value \a CV.
5770 void generateDefaultMapInfo(const CapturedStmt::Capture &CI,
5771 const FieldDecl &RI, llvm::Value *CV,
5772 MapBaseValuesArrayTy &CurBasePointers,
5773 MapValuesArrayTy &CurPointers,
5774 MapValuesArrayTy &CurSizes,
5775 MapFlagsArrayTy &CurMapTypes) {
5777 // Do the default mapping.
5778 if (CI.capturesThis()) {
5779 CurBasePointers.push_back(CV);
5780 CurPointers.push_back(CV);
5781 const PointerType *PtrTy = cast<PointerType>(RI.getType().getTypePtr());
5782 CurSizes.push_back(CGF.getTypeSize(PtrTy->getPointeeType()));
5783 // Default map type.
5784 CurMapTypes.push_back(OMP_MAP_TO | OMP_MAP_FROM);
5785 } else if (CI.capturesVariableByCopy()) {
5786 CurBasePointers.push_back(CV);
5787 CurPointers.push_back(CV);
5788 if (!RI.getType()->isAnyPointerType()) {
5789 // We have to signal to the runtime captures passed by value that are
5791 CurMapTypes.push_back(OMP_MAP_PRIVATE_VAL);
5792 CurSizes.push_back(CGF.getTypeSize(RI.getType()));
5794 // Pointers are implicitly mapped with a zero size and no flags
5795 // (other than first map that is added for all implicit maps).
5796 CurMapTypes.push_back(0u);
5797 CurSizes.push_back(llvm::Constant::getNullValue(CGF.SizeTy));
5800 assert(CI.capturesVariable() && "Expected captured reference.");
5801 CurBasePointers.push_back(CV);
5802 CurPointers.push_back(CV);
5804 const ReferenceType *PtrTy =
5805 cast<ReferenceType>(RI.getType().getTypePtr());
5806 QualType ElementType = PtrTy->getPointeeType();
5807 CurSizes.push_back(CGF.getTypeSize(ElementType));
5808 // The default map type for a scalar/complex type is 'to' because by
5809 // default the value doesn't have to be retrieved. For an aggregate
5810 // type, the default is 'tofrom'.
5811 CurMapTypes.push_back(ElementType->isAggregateType()
5812 ? (OMP_MAP_TO | OMP_MAP_FROM)
5815 // If we have a capture by reference we may need to add the private
5816 // pointer flag if the base declaration shows in some first-private
5818 CurMapTypes.back() =
5819 adjustMapModifiersForPrivateClauses(CI, CurMapTypes.back());
5821 // Every default map produces a single argument, so, it is always the
5823 CurMapTypes.back() |= OMP_MAP_FIRST_REF;
5827 enum OpenMPOffloadingReservedDeviceIDs {
5828 /// \brief Device ID if the device was not defined, runtime should get it
5829 /// from environment variables in the spec.
5830 OMP_DEVICEID_UNDEF = -1,
5832 } // anonymous namespace
5834 /// \brief Emit the arrays used to pass the captures and map information to the
5835 /// offloading runtime library. If there is no map or capture information,
5836 /// return nullptr by reference.
5838 emitOffloadingArrays(CodeGenFunction &CGF,
5839 MappableExprsHandler::MapBaseValuesArrayTy &BasePointers,
5840 MappableExprsHandler::MapValuesArrayTy &Pointers,
5841 MappableExprsHandler::MapValuesArrayTy &Sizes,
5842 MappableExprsHandler::MapFlagsArrayTy &MapTypes,
5843 CGOpenMPRuntime::TargetDataInfo &Info) {
5844 auto &CGM = CGF.CGM;
5845 auto &Ctx = CGF.getContext();
5847 // Reset the array information.
5848 Info.clearArrayInfo();
5849 Info.NumberOfPtrs = BasePointers.size();
5851 if (Info.NumberOfPtrs) {
5852 // Detect if we have any capture size requiring runtime evaluation of the
5853 // size so that a constant array could be eventually used.
5854 bool hasRuntimeEvaluationCaptureSize = false;
5855 for (auto *S : Sizes)
5856 if (!isa<llvm::Constant>(S)) {
5857 hasRuntimeEvaluationCaptureSize = true;
5861 llvm::APInt PointerNumAP(32, Info.NumberOfPtrs, /*isSigned=*/true);
5862 QualType PointerArrayType =
5863 Ctx.getConstantArrayType(Ctx.VoidPtrTy, PointerNumAP, ArrayType::Normal,
5864 /*IndexTypeQuals=*/0);
5866 Info.BasePointersArray =
5867 CGF.CreateMemTemp(PointerArrayType, ".offload_baseptrs").getPointer();
5868 Info.PointersArray =
5869 CGF.CreateMemTemp(PointerArrayType, ".offload_ptrs").getPointer();
5871 // If we don't have any VLA types or other types that require runtime
5872 // evaluation, we can use a constant array for the map sizes, otherwise we
5873 // need to fill up the arrays as we do for the pointers.
5874 if (hasRuntimeEvaluationCaptureSize) {
5875 QualType SizeArrayType = Ctx.getConstantArrayType(
5876 Ctx.getSizeType(), PointerNumAP, ArrayType::Normal,
5877 /*IndexTypeQuals=*/0);
5879 CGF.CreateMemTemp(SizeArrayType, ".offload_sizes").getPointer();
5881 // We expect all the sizes to be constant, so we collect them to create
5882 // a constant array.
5883 SmallVector<llvm::Constant *, 16> ConstSizes;
5884 for (auto S : Sizes)
5885 ConstSizes.push_back(cast<llvm::Constant>(S));
5887 auto *SizesArrayInit = llvm::ConstantArray::get(
5888 llvm::ArrayType::get(CGM.SizeTy, ConstSizes.size()), ConstSizes);
5889 auto *SizesArrayGbl = new llvm::GlobalVariable(
5890 CGM.getModule(), SizesArrayInit->getType(),
5891 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage,
5892 SizesArrayInit, ".offload_sizes");
5893 SizesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
5894 Info.SizesArray = SizesArrayGbl;
5897 // The map types are always constant so we don't need to generate code to
5898 // fill arrays. Instead, we create an array constant.
5899 llvm::Constant *MapTypesArrayInit =
5900 llvm::ConstantDataArray::get(CGF.Builder.getContext(), MapTypes);
5901 auto *MapTypesArrayGbl = new llvm::GlobalVariable(
5902 CGM.getModule(), MapTypesArrayInit->getType(),
5903 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage,
5904 MapTypesArrayInit, ".offload_maptypes");
5905 MapTypesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
5906 Info.MapTypesArray = MapTypesArrayGbl;
5908 for (unsigned i = 0; i < Info.NumberOfPtrs; ++i) {
5909 llvm::Value *BPVal = *BasePointers[i];
5910 if (BPVal->getType()->isPointerTy())
5911 BPVal = CGF.Builder.CreateBitCast(BPVal, CGM.VoidPtrTy);
5913 assert(BPVal->getType()->isIntegerTy() &&
5914 "If not a pointer, the value type must be an integer.");
5915 BPVal = CGF.Builder.CreateIntToPtr(BPVal, CGM.VoidPtrTy);
5917 llvm::Value *BP = CGF.Builder.CreateConstInBoundsGEP2_32(
5918 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5919 Info.BasePointersArray, 0, i);
5920 Address BPAddr(BP, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy));
5921 CGF.Builder.CreateStore(BPVal, BPAddr);
5923 if (Info.requiresDevicePointerInfo())
5924 if (auto *DevVD = BasePointers[i].getDevicePtrDecl())
5925 Info.CaptureDeviceAddrMap.insert(std::make_pair(DevVD, BPAddr));
5927 llvm::Value *PVal = Pointers[i];
5928 if (PVal->getType()->isPointerTy())
5929 PVal = CGF.Builder.CreateBitCast(PVal, CGM.VoidPtrTy);
5931 assert(PVal->getType()->isIntegerTy() &&
5932 "If not a pointer, the value type must be an integer.");
5933 PVal = CGF.Builder.CreateIntToPtr(PVal, CGM.VoidPtrTy);
5935 llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32(
5936 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5937 Info.PointersArray, 0, i);
5938 Address PAddr(P, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy));
5939 CGF.Builder.CreateStore(PVal, PAddr);
5941 if (hasRuntimeEvaluationCaptureSize) {
5942 llvm::Value *S = CGF.Builder.CreateConstInBoundsGEP2_32(
5943 llvm::ArrayType::get(CGM.SizeTy, Info.NumberOfPtrs),
5947 Address SAddr(S, Ctx.getTypeAlignInChars(Ctx.getSizeType()));
5948 CGF.Builder.CreateStore(
5949 CGF.Builder.CreateIntCast(Sizes[i], CGM.SizeTy, /*isSigned=*/true),
5955 /// \brief Emit the arguments to be passed to the runtime library based on the
5956 /// arrays of pointers, sizes and map types.
5957 static void emitOffloadingArraysArgument(
5958 CodeGenFunction &CGF, llvm::Value *&BasePointersArrayArg,
5959 llvm::Value *&PointersArrayArg, llvm::Value *&SizesArrayArg,
5960 llvm::Value *&MapTypesArrayArg, CGOpenMPRuntime::TargetDataInfo &Info) {
5961 auto &CGM = CGF.CGM;
5962 if (Info.NumberOfPtrs) {
5963 BasePointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5964 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5965 Info.BasePointersArray,
5966 /*Idx0=*/0, /*Idx1=*/0);
5967 PointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5968 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
5972 SizesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5973 llvm::ArrayType::get(CGM.SizeTy, Info.NumberOfPtrs), Info.SizesArray,
5974 /*Idx0=*/0, /*Idx1=*/0);
5975 MapTypesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32(
5976 llvm::ArrayType::get(CGM.Int32Ty, Info.NumberOfPtrs),
5981 BasePointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy);
5982 PointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy);
5983 SizesArrayArg = llvm::ConstantPointerNull::get(CGM.SizeTy->getPointerTo());
5985 llvm::ConstantPointerNull::get(CGM.Int32Ty->getPointerTo());
5989 void CGOpenMPRuntime::emitTargetCall(CodeGenFunction &CGF,
5990 const OMPExecutableDirective &D,
5991 llvm::Value *OutlinedFn,
5992 llvm::Value *OutlinedFnID,
5993 const Expr *IfCond, const Expr *Device,
5994 ArrayRef<llvm::Value *> CapturedVars) {
5995 if (!CGF.HaveInsertPoint())
5998 assert(OutlinedFn && "Invalid outlined function!");
6000 auto &Ctx = CGF.getContext();
6002 // Fill up the arrays with all the captured variables.
6003 MappableExprsHandler::MapValuesArrayTy KernelArgs;
6004 MappableExprsHandler::MapBaseValuesArrayTy BasePointers;
6005 MappableExprsHandler::MapValuesArrayTy Pointers;
6006 MappableExprsHandler::MapValuesArrayTy Sizes;
6007 MappableExprsHandler::MapFlagsArrayTy MapTypes;
6009 MappableExprsHandler::MapBaseValuesArrayTy CurBasePointers;
6010 MappableExprsHandler::MapValuesArrayTy CurPointers;
6011 MappableExprsHandler::MapValuesArrayTy CurSizes;
6012 MappableExprsHandler::MapFlagsArrayTy CurMapTypes;
6014 // Get mappable expression information.
6015 MappableExprsHandler MEHandler(D, CGF);
6017 const CapturedStmt &CS = *cast<CapturedStmt>(D.getAssociatedStmt());
6018 auto RI = CS.getCapturedRecordDecl()->field_begin();
6019 auto CV = CapturedVars.begin();
6020 for (CapturedStmt::const_capture_iterator CI = CS.capture_begin(),
6021 CE = CS.capture_end();
6022 CI != CE; ++CI, ++RI, ++CV) {
6026 CurBasePointers.clear();
6027 CurPointers.clear();
6029 CurMapTypes.clear();
6031 // VLA sizes are passed to the outlined region by copy and do not have map
6032 // information associated.
6033 if (CI->capturesVariableArrayType()) {
6034 CurBasePointers.push_back(*CV);
6035 CurPointers.push_back(*CV);
6036 CurSizes.push_back(CGF.getTypeSize(RI->getType()));
6037 // Copy to the device as an argument. No need to retrieve it.
6038 CurMapTypes.push_back(MappableExprsHandler::OMP_MAP_PRIVATE_VAL |
6039 MappableExprsHandler::OMP_MAP_FIRST_REF);
6041 // If we have any information in the map clause, we use it, otherwise we
6042 // just do a default mapping.
6043 MEHandler.generateInfoForCapture(CI, *CV, CurBasePointers, CurPointers,
6044 CurSizes, CurMapTypes);
6045 if (CurBasePointers.empty())
6046 MEHandler.generateDefaultMapInfo(*CI, **RI, *CV, CurBasePointers,
6047 CurPointers, CurSizes, CurMapTypes);
6049 // We expect to have at least an element of information for this capture.
6050 assert(!CurBasePointers.empty() && "Non-existing map pointer for capture!");
6051 assert(CurBasePointers.size() == CurPointers.size() &&
6052 CurBasePointers.size() == CurSizes.size() &&
6053 CurBasePointers.size() == CurMapTypes.size() &&
6054 "Inconsistent map information sizes!");
6056 // The kernel args are always the first elements of the base pointers
6057 // associated with a capture.
6058 KernelArgs.push_back(*CurBasePointers.front());
6059 // We need to append the results of this capture to what we already have.
6060 BasePointers.append(CurBasePointers.begin(), CurBasePointers.end());
6061 Pointers.append(CurPointers.begin(), CurPointers.end());
6062 Sizes.append(CurSizes.begin(), CurSizes.end());
6063 MapTypes.append(CurMapTypes.begin(), CurMapTypes.end());
6066 // Keep track on whether the host function has to be executed.
6067 auto OffloadErrorQType =
6068 Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true);
6069 auto OffloadError = CGF.MakeAddrLValue(
6070 CGF.CreateMemTemp(OffloadErrorQType, ".run_host_version"),
6072 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty),
6075 // Fill up the pointer arrays and transfer execution to the device.
6076 auto &&ThenGen = [&BasePointers, &Pointers, &Sizes, &MapTypes, Device,
6077 OutlinedFnID, OffloadError,
6078 &D](CodeGenFunction &CGF, PrePostActionTy &) {
6079 auto &RT = CGF.CGM.getOpenMPRuntime();
6080 // Emit the offloading arrays.
6081 TargetDataInfo Info;
6082 emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info);
6083 emitOffloadingArraysArgument(CGF, Info.BasePointersArray,
6084 Info.PointersArray, Info.SizesArray,
6085 Info.MapTypesArray, Info);
6087 // On top of the arrays that were filled up, the target offloading call
6088 // takes as arguments the device id as well as the host pointer. The host
6089 // pointer is used by the runtime library to identify the current target
6090 // region, so it only has to be unique and not necessarily point to
6091 // anything. It could be the pointer to the outlined function that
6092 // implements the target region, but we aren't using that so that the
6093 // compiler doesn't need to keep that, and could therefore inline the host
6094 // function if proven worthwhile during optimization.
6096 // From this point on, we need to have an ID of the target region defined.
6097 assert(OutlinedFnID && "Invalid outlined function ID!");
6099 // Emit device ID if any.
6100 llvm::Value *DeviceID;
6102 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6103 CGF.Int32Ty, /*isSigned=*/true);
6105 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6107 // Emit the number of elements in the offloading arrays.
6108 llvm::Value *PointerNum = CGF.Builder.getInt32(BasePointers.size());
6110 // Return value of the runtime offloading call.
6111 llvm::Value *Return;
6113 auto *NumTeams = emitNumTeamsForTargetDirective(RT, CGF, D);
6114 auto *NumThreads = emitNumThreadsForTargetDirective(RT, CGF, D);
6116 // The target region is an outlined function launched by the runtime
6117 // via calls __tgt_target() or __tgt_target_teams().
6119 // __tgt_target() launches a target region with one team and one thread,
6120 // executing a serial region. This master thread may in turn launch
6121 // more threads within its team upon encountering a parallel region,
6122 // however, no additional teams can be launched on the device.
6124 // __tgt_target_teams() launches a target region with one or more teams,
6125 // each with one or more threads. This call is required for target
6126 // constructs such as:
6128 // 'target' / 'teams'
6129 // 'target teams distribute parallel for'
6130 // 'target parallel'
6133 // Note that on the host and CPU targets, the runtime implementation of
6134 // these calls simply call the outlined function without forking threads.
6135 // The outlined functions themselves have runtime calls to
6136 // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by
6137 // the compiler in emitTeamsCall() and emitParallelCall().
6139 // In contrast, on the NVPTX target, the implementation of
6140 // __tgt_target_teams() launches a GPU kernel with the requested number
6141 // of teams and threads so no additional calls to the runtime are required.
6143 // If we have NumTeams defined this means that we have an enclosed teams
6144 // region. Therefore we also expect to have NumThreads defined. These two
6145 // values should be defined in the presence of a teams directive,
6146 // regardless of having any clauses associated. If the user is using teams
6147 // but no clauses, these two values will be the default that should be
6148 // passed to the runtime library - a 32-bit integer with the value zero.
6149 assert(NumThreads && "Thread limit expression should be available along "
6150 "with number of teams.");
6151 llvm::Value *OffloadingArgs[] = {
6152 DeviceID, OutlinedFnID,
6153 PointerNum, Info.BasePointersArray,
6154 Info.PointersArray, Info.SizesArray,
6155 Info.MapTypesArray, NumTeams,
6157 Return = CGF.EmitRuntimeCall(
6158 RT.createRuntimeFunction(OMPRTL__tgt_target_teams), OffloadingArgs);
6160 llvm::Value *OffloadingArgs[] = {
6161 DeviceID, OutlinedFnID,
6162 PointerNum, Info.BasePointersArray,
6163 Info.PointersArray, Info.SizesArray,
6164 Info.MapTypesArray};
6165 Return = CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__tgt_target),
6169 CGF.EmitStoreOfScalar(Return, OffloadError);
6172 // Notify that the host version must be executed.
6173 auto &&ElseGen = [OffloadError](CodeGenFunction &CGF, PrePostActionTy &) {
6174 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/-1u),
6178 // If we have a target function ID it means that we need to support
6179 // offloading, otherwise, just execute on the host. We need to execute on host
6180 // regardless of the conditional in the if clause if, e.g., the user do not
6181 // specify target triples.
6184 emitOMPIfClause(CGF, IfCond, ThenGen, ElseGen);
6186 RegionCodeGenTy ThenRCG(ThenGen);
6190 RegionCodeGenTy ElseRCG(ElseGen);
6194 // Check the error code and execute the host version if required.
6195 auto OffloadFailedBlock = CGF.createBasicBlock("omp_offload.failed");
6196 auto OffloadContBlock = CGF.createBasicBlock("omp_offload.cont");
6197 auto OffloadErrorVal = CGF.EmitLoadOfScalar(OffloadError, SourceLocation());
6198 auto Failed = CGF.Builder.CreateIsNotNull(OffloadErrorVal);
6199 CGF.Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock);
6201 CGF.EmitBlock(OffloadFailedBlock);
6202 CGF.Builder.CreateCall(OutlinedFn, KernelArgs);
6203 CGF.EmitBranch(OffloadContBlock);
6205 CGF.EmitBlock(OffloadContBlock, /*IsFinished=*/true);
6208 void CGOpenMPRuntime::scanForTargetRegionsFunctions(const Stmt *S,
6209 StringRef ParentName) {
6213 // Codegen OMP target directives that offload compute to the device.
6214 bool requiresDeviceCodegen =
6215 isa<OMPExecutableDirective>(S) &&
6216 isOpenMPTargetExecutionDirective(
6217 cast<OMPExecutableDirective>(S)->getDirectiveKind());
6219 if (requiresDeviceCodegen) {
6220 auto &E = *cast<OMPExecutableDirective>(S);
6224 getTargetEntryUniqueInfo(CGM.getContext(), E.getLocStart(), DeviceID,
6227 // Is this a target region that should not be emitted as an entry point? If
6228 // so just signal we are done with this target region.
6229 if (!OffloadEntriesInfoManager.hasTargetRegionEntryInfo(DeviceID, FileID,
6233 switch (S->getStmtClass()) {
6234 case Stmt::OMPTargetDirectiveClass:
6235 CodeGenFunction::EmitOMPTargetDeviceFunction(
6236 CGM, ParentName, cast<OMPTargetDirective>(*S));
6238 case Stmt::OMPTargetParallelDirectiveClass:
6239 CodeGenFunction::EmitOMPTargetParallelDeviceFunction(
6240 CGM, ParentName, cast<OMPTargetParallelDirective>(*S));
6242 case Stmt::OMPTargetTeamsDirectiveClass:
6243 CodeGenFunction::EmitOMPTargetTeamsDeviceFunction(
6244 CGM, ParentName, cast<OMPTargetTeamsDirective>(*S));
6247 llvm_unreachable("Unknown target directive for OpenMP device codegen.");
6252 if (const OMPExecutableDirective *E = dyn_cast<OMPExecutableDirective>(S)) {
6253 if (!E->hasAssociatedStmt())
6256 scanForTargetRegionsFunctions(
6257 cast<CapturedStmt>(E->getAssociatedStmt())->getCapturedStmt(),
6262 // If this is a lambda function, look into its body.
6263 if (auto *L = dyn_cast<LambdaExpr>(S))
6266 // Keep looking for target regions recursively.
6267 for (auto *II : S->children())
6268 scanForTargetRegionsFunctions(II, ParentName);
6271 bool CGOpenMPRuntime::emitTargetFunctions(GlobalDecl GD) {
6272 auto &FD = *cast<FunctionDecl>(GD.getDecl());
6274 // If emitting code for the host, we do not process FD here. Instead we do
6275 // the normal code generation.
6276 if (!CGM.getLangOpts().OpenMPIsDevice)
6279 // Try to detect target regions in the function.
6280 scanForTargetRegionsFunctions(FD.getBody(), CGM.getMangledName(GD));
6282 // We should not emit any function other that the ones created during the
6283 // scanning. Therefore, we signal that this function is completely dealt
6288 bool CGOpenMPRuntime::emitTargetGlobalVariable(GlobalDecl GD) {
6289 if (!CGM.getLangOpts().OpenMPIsDevice)
6292 // Check if there are Ctors/Dtors in this declaration and look for target
6293 // regions in it. We use the complete variant to produce the kernel name
6295 QualType RDTy = cast<VarDecl>(GD.getDecl())->getType();
6296 if (auto *RD = RDTy->getBaseElementTypeUnsafe()->getAsCXXRecordDecl()) {
6297 for (auto *Ctor : RD->ctors()) {
6298 StringRef ParentName =
6299 CGM.getMangledName(GlobalDecl(Ctor, Ctor_Complete));
6300 scanForTargetRegionsFunctions(Ctor->getBody(), ParentName);
6302 auto *Dtor = RD->getDestructor();
6304 StringRef ParentName =
6305 CGM.getMangledName(GlobalDecl(Dtor, Dtor_Complete));
6306 scanForTargetRegionsFunctions(Dtor->getBody(), ParentName);
6310 // If we are in target mode we do not emit any global (declare target is not
6311 // implemented yet). Therefore we signal that GD was processed in this case.
6315 bool CGOpenMPRuntime::emitTargetGlobal(GlobalDecl GD) {
6316 auto *VD = GD.getDecl();
6317 if (isa<FunctionDecl>(VD))
6318 return emitTargetFunctions(GD);
6320 return emitTargetGlobalVariable(GD);
6323 llvm::Function *CGOpenMPRuntime::emitRegistrationFunction() {
6324 // If we have offloading in the current module, we need to emit the entries
6325 // now and register the offloading descriptor.
6326 createOffloadEntriesAndInfoMetadata();
6328 // Create and register the offloading binary descriptors. This is the main
6329 // entity that captures all the information about offloading in the current
6330 // compilation unit.
6331 return createOffloadingBinaryDescriptorRegistration();
6334 void CGOpenMPRuntime::emitTeamsCall(CodeGenFunction &CGF,
6335 const OMPExecutableDirective &D,
6337 llvm::Value *OutlinedFn,
6338 ArrayRef<llvm::Value *> CapturedVars) {
6339 if (!CGF.HaveInsertPoint())
6342 auto *RTLoc = emitUpdateLocation(CGF, Loc);
6343 CodeGenFunction::RunCleanupsScope Scope(CGF);
6345 // Build call __kmpc_fork_teams(loc, n, microtask, var1, .., varn);
6346 llvm::Value *Args[] = {
6348 CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars
6349 CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy())};
6350 llvm::SmallVector<llvm::Value *, 16> RealArgs;
6351 RealArgs.append(std::begin(Args), std::end(Args));
6352 RealArgs.append(CapturedVars.begin(), CapturedVars.end());
6354 auto RTLFn = createRuntimeFunction(OMPRTL__kmpc_fork_teams);
6355 CGF.EmitRuntimeCall(RTLFn, RealArgs);
6358 void CGOpenMPRuntime::emitNumTeamsClause(CodeGenFunction &CGF,
6359 const Expr *NumTeams,
6360 const Expr *ThreadLimit,
6361 SourceLocation Loc) {
6362 if (!CGF.HaveInsertPoint())
6365 auto *RTLoc = emitUpdateLocation(CGF, Loc);
6367 llvm::Value *NumTeamsVal =
6369 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(NumTeams),
6370 CGF.CGM.Int32Ty, /* isSigned = */ true)
6371 : CGF.Builder.getInt32(0);
6373 llvm::Value *ThreadLimitVal =
6375 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(ThreadLimit),
6376 CGF.CGM.Int32Ty, /* isSigned = */ true)
6377 : CGF.Builder.getInt32(0);
6379 // Build call __kmpc_push_num_teamss(&loc, global_tid, num_teams, thread_limit)
6380 llvm::Value *PushNumTeamsArgs[] = {RTLoc, getThreadID(CGF, Loc), NumTeamsVal,
6382 CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_num_teams),
6386 void CGOpenMPRuntime::emitTargetDataCalls(
6387 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
6388 const Expr *Device, const RegionCodeGenTy &CodeGen, TargetDataInfo &Info) {
6389 if (!CGF.HaveInsertPoint())
6392 // Action used to replace the default codegen action and turn privatization
6394 PrePostActionTy NoPrivAction;
6396 // Generate the code for the opening of the data environment. Capture all the
6397 // arguments of the runtime call by reference because they are used in the
6398 // closing of the region.
6399 auto &&BeginThenGen = [&D, Device, &Info, &CodeGen](CodeGenFunction &CGF,
6400 PrePostActionTy &) {
6401 // Fill up the arrays with all the mapped variables.
6402 MappableExprsHandler::MapBaseValuesArrayTy BasePointers;
6403 MappableExprsHandler::MapValuesArrayTy Pointers;
6404 MappableExprsHandler::MapValuesArrayTy Sizes;
6405 MappableExprsHandler::MapFlagsArrayTy MapTypes;
6407 // Get map clause information.
6408 MappableExprsHandler MCHandler(D, CGF);
6409 MCHandler.generateAllInfo(BasePointers, Pointers, Sizes, MapTypes);
6411 // Fill up the arrays and create the arguments.
6412 emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info);
6414 llvm::Value *BasePointersArrayArg = nullptr;
6415 llvm::Value *PointersArrayArg = nullptr;
6416 llvm::Value *SizesArrayArg = nullptr;
6417 llvm::Value *MapTypesArrayArg = nullptr;
6418 emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg,
6419 SizesArrayArg, MapTypesArrayArg, Info);
6421 // Emit device ID if any.
6422 llvm::Value *DeviceID = nullptr;
6424 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6425 CGF.Int32Ty, /*isSigned=*/true);
6427 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6429 // Emit the number of elements in the offloading arrays.
6430 auto *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs);
6432 llvm::Value *OffloadingArgs[] = {
6433 DeviceID, PointerNum, BasePointersArrayArg,
6434 PointersArrayArg, SizesArrayArg, MapTypesArrayArg};
6435 auto &RT = CGF.CGM.getOpenMPRuntime();
6436 CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__tgt_target_data_begin),
6439 // If device pointer privatization is required, emit the body of the region
6440 // here. It will have to be duplicated: with and without privatization.
6441 if (!Info.CaptureDeviceAddrMap.empty())
6445 // Generate code for the closing of the data region.
6446 auto &&EndThenGen = [Device, &Info](CodeGenFunction &CGF, PrePostActionTy &) {
6447 assert(Info.isValid() && "Invalid data environment closing arguments.");
6449 llvm::Value *BasePointersArrayArg = nullptr;
6450 llvm::Value *PointersArrayArg = nullptr;
6451 llvm::Value *SizesArrayArg = nullptr;
6452 llvm::Value *MapTypesArrayArg = nullptr;
6453 emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg,
6454 SizesArrayArg, MapTypesArrayArg, Info);
6456 // Emit device ID if any.
6457 llvm::Value *DeviceID = nullptr;
6459 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6460 CGF.Int32Ty, /*isSigned=*/true);
6462 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6464 // Emit the number of elements in the offloading arrays.
6465 auto *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs);
6467 llvm::Value *OffloadingArgs[] = {
6468 DeviceID, PointerNum, BasePointersArrayArg,
6469 PointersArrayArg, SizesArrayArg, MapTypesArrayArg};
6470 auto &RT = CGF.CGM.getOpenMPRuntime();
6471 CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__tgt_target_data_end),
6475 // If we need device pointer privatization, we need to emit the body of the
6476 // region with no privatization in the 'else' branch of the conditional.
6477 // Otherwise, we don't have to do anything.
6478 auto &&BeginElseGen = [&Info, &CodeGen, &NoPrivAction](CodeGenFunction &CGF,
6479 PrePostActionTy &) {
6480 if (!Info.CaptureDeviceAddrMap.empty()) {
6481 CodeGen.setAction(NoPrivAction);
6486 // We don't have to do anything to close the region if the if clause evaluates
6488 auto &&EndElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {};
6491 emitOMPIfClause(CGF, IfCond, BeginThenGen, BeginElseGen);
6493 RegionCodeGenTy RCG(BeginThenGen);
6497 // If we don't require privatization of device pointers, we emit the body in
6498 // between the runtime calls. This avoids duplicating the body code.
6499 if (Info.CaptureDeviceAddrMap.empty()) {
6500 CodeGen.setAction(NoPrivAction);
6505 emitOMPIfClause(CGF, IfCond, EndThenGen, EndElseGen);
6507 RegionCodeGenTy RCG(EndThenGen);
6512 void CGOpenMPRuntime::emitTargetDataStandAloneCall(
6513 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
6514 const Expr *Device) {
6515 if (!CGF.HaveInsertPoint())
6518 assert((isa<OMPTargetEnterDataDirective>(D) ||
6519 isa<OMPTargetExitDataDirective>(D) ||
6520 isa<OMPTargetUpdateDirective>(D)) &&
6521 "Expecting either target enter, exit data, or update directives.");
6523 // Generate the code for the opening of the data environment.
6524 auto &&ThenGen = [&D, Device](CodeGenFunction &CGF, PrePostActionTy &) {
6525 // Fill up the arrays with all the mapped variables.
6526 MappableExprsHandler::MapBaseValuesArrayTy BasePointers;
6527 MappableExprsHandler::MapValuesArrayTy Pointers;
6528 MappableExprsHandler::MapValuesArrayTy Sizes;
6529 MappableExprsHandler::MapFlagsArrayTy MapTypes;
6531 // Get map clause information.
6532 MappableExprsHandler MEHandler(D, CGF);
6533 MEHandler.generateAllInfo(BasePointers, Pointers, Sizes, MapTypes);
6535 // Fill up the arrays and create the arguments.
6536 TargetDataInfo Info;
6537 emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info);
6538 emitOffloadingArraysArgument(CGF, Info.BasePointersArray,
6539 Info.PointersArray, Info.SizesArray,
6540 Info.MapTypesArray, Info);
6542 // Emit device ID if any.
6543 llvm::Value *DeviceID = nullptr;
6545 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
6546 CGF.Int32Ty, /*isSigned=*/true);
6548 DeviceID = CGF.Builder.getInt32(OMP_DEVICEID_UNDEF);
6550 // Emit the number of elements in the offloading arrays.
6551 auto *PointerNum = CGF.Builder.getInt32(BasePointers.size());
6553 llvm::Value *OffloadingArgs[] = {
6554 DeviceID, PointerNum, Info.BasePointersArray,
6555 Info.PointersArray, Info.SizesArray, Info.MapTypesArray};
6557 auto &RT = CGF.CGM.getOpenMPRuntime();
6558 // Select the right runtime function call for each expected standalone
6560 OpenMPRTLFunction RTLFn;
6561 switch (D.getDirectiveKind()) {
6563 llvm_unreachable("Unexpected standalone target data directive.");
6565 case OMPD_target_enter_data:
6566 RTLFn = OMPRTL__tgt_target_data_begin;
6568 case OMPD_target_exit_data:
6569 RTLFn = OMPRTL__tgt_target_data_end;
6571 case OMPD_target_update:
6572 RTLFn = OMPRTL__tgt_target_data_update;
6575 CGF.EmitRuntimeCall(RT.createRuntimeFunction(RTLFn), OffloadingArgs);
6578 // In the event we get an if clause, we don't have to take any action on the
6580 auto &&ElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {};
6583 emitOMPIfClause(CGF, IfCond, ThenGen, ElseGen);
6585 RegionCodeGenTy ThenGenRCG(ThenGen);
6591 /// Kind of parameter in a function with 'declare simd' directive.
6592 enum ParamKindTy { LinearWithVarStride, Linear, Uniform, Vector };
6593 /// Attribute set of the parameter.
6594 struct ParamAttrTy {
6595 ParamKindTy Kind = Vector;
6596 llvm::APSInt StrideOrArg;
6597 llvm::APSInt Alignment;
6601 static unsigned evaluateCDTSize(const FunctionDecl *FD,
6602 ArrayRef<ParamAttrTy> ParamAttrs) {
6603 // Every vector variant of a SIMD-enabled function has a vector length (VLEN).
6604 // If OpenMP clause "simdlen" is used, the VLEN is the value of the argument
6605 // of that clause. The VLEN value must be power of 2.
6606 // In other case the notion of the function`s "characteristic data type" (CDT)
6607 // is used to compute the vector length.
6608 // CDT is defined in the following order:
6609 // a) For non-void function, the CDT is the return type.
6610 // b) If the function has any non-uniform, non-linear parameters, then the
6611 // CDT is the type of the first such parameter.
6612 // c) If the CDT determined by a) or b) above is struct, union, or class
6613 // type which is pass-by-value (except for the type that maps to the
6614 // built-in complex data type), the characteristic data type is int.
6615 // d) If none of the above three cases is applicable, the CDT is int.
6616 // The VLEN is then determined based on the CDT and the size of vector
6617 // register of that ISA for which current vector version is generated. The
6618 // VLEN is computed using the formula below:
6619 // VLEN = sizeof(vector_register) / sizeof(CDT),
6620 // where vector register size specified in section 3.2.1 Registers and the
6621 // Stack Frame of original AMD64 ABI document.
6622 QualType RetType = FD->getReturnType();
6623 if (RetType.isNull())
6625 ASTContext &C = FD->getASTContext();
6627 if (!RetType.isNull() && !RetType->isVoidType())
6630 unsigned Offset = 0;
6631 if (auto *MD = dyn_cast<CXXMethodDecl>(FD)) {
6632 if (ParamAttrs[Offset].Kind == Vector)
6633 CDT = C.getPointerType(C.getRecordType(MD->getParent()));
6637 for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) {
6638 if (ParamAttrs[I + Offset].Kind == Vector) {
6639 CDT = FD->getParamDecl(I)->getType();
6647 CDT = CDT->getCanonicalTypeUnqualified();
6648 if (CDT->isRecordType() || CDT->isUnionType())
6650 return C.getTypeSize(CDT);
6654 emitX86DeclareSimdFunction(const FunctionDecl *FD, llvm::Function *Fn,
6655 const llvm::APSInt &VLENVal,
6656 ArrayRef<ParamAttrTy> ParamAttrs,
6657 OMPDeclareSimdDeclAttr::BranchStateTy State) {
6660 unsigned VecRegSize;
6662 ISADataTy ISAData[] = {
6676 llvm::SmallVector<char, 2> Masked;
6678 case OMPDeclareSimdDeclAttr::BS_Undefined:
6679 Masked.push_back('N');
6680 Masked.push_back('M');
6682 case OMPDeclareSimdDeclAttr::BS_Notinbranch:
6683 Masked.push_back('N');
6685 case OMPDeclareSimdDeclAttr::BS_Inbranch:
6686 Masked.push_back('M');
6689 for (auto Mask : Masked) {
6690 for (auto &Data : ISAData) {
6691 SmallString<256> Buffer;
6692 llvm::raw_svector_ostream Out(Buffer);
6693 Out << "_ZGV" << Data.ISA << Mask;
6695 Out << llvm::APSInt::getUnsigned(Data.VecRegSize /
6696 evaluateCDTSize(FD, ParamAttrs));
6699 for (auto &ParamAttr : ParamAttrs) {
6700 switch (ParamAttr.Kind){
6701 case LinearWithVarStride:
6702 Out << 's' << ParamAttr.StrideOrArg;
6706 if (!!ParamAttr.StrideOrArg)
6707 Out << ParamAttr.StrideOrArg;
6716 if (!!ParamAttr.Alignment)
6717 Out << 'a' << ParamAttr.Alignment;
6719 Out << '_' << Fn->getName();
6720 Fn->addFnAttr(Out.str());
6725 void CGOpenMPRuntime::emitDeclareSimdFunction(const FunctionDecl *FD,
6726 llvm::Function *Fn) {
6727 ASTContext &C = CGM.getContext();
6728 FD = FD->getCanonicalDecl();
6729 // Map params to their positions in function decl.
6730 llvm::DenseMap<const Decl *, unsigned> ParamPositions;
6731 if (isa<CXXMethodDecl>(FD))
6732 ParamPositions.insert({FD, 0});
6733 unsigned ParamPos = ParamPositions.size();
6734 for (auto *P : FD->parameters()) {
6735 ParamPositions.insert({P->getCanonicalDecl(), ParamPos});
6738 for (auto *Attr : FD->specific_attrs<OMPDeclareSimdDeclAttr>()) {
6739 llvm::SmallVector<ParamAttrTy, 8> ParamAttrs(ParamPositions.size());
6740 // Mark uniform parameters.
6741 for (auto *E : Attr->uniforms()) {
6742 E = E->IgnoreParenImpCasts();
6744 if (isa<CXXThisExpr>(E))
6745 Pos = ParamPositions[FD];
6747 auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
6748 ->getCanonicalDecl();
6749 Pos = ParamPositions[PVD];
6751 ParamAttrs[Pos].Kind = Uniform;
6753 // Get alignment info.
6754 auto NI = Attr->alignments_begin();
6755 for (auto *E : Attr->aligneds()) {
6756 E = E->IgnoreParenImpCasts();
6759 if (isa<CXXThisExpr>(E)) {
6760 Pos = ParamPositions[FD];
6761 ParmTy = E->getType();
6763 auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
6764 ->getCanonicalDecl();
6765 Pos = ParamPositions[PVD];
6766 ParmTy = PVD->getType();
6768 ParamAttrs[Pos].Alignment =
6769 (*NI) ? (*NI)->EvaluateKnownConstInt(C)
6770 : llvm::APSInt::getUnsigned(
6771 C.toCharUnitsFromBits(C.getOpenMPDefaultSimdAlign(ParmTy))
6775 // Mark linear parameters.
6776 auto SI = Attr->steps_begin();
6777 auto MI = Attr->modifiers_begin();
6778 for (auto *E : Attr->linears()) {
6779 E = E->IgnoreParenImpCasts();
6781 if (isa<CXXThisExpr>(E))
6782 Pos = ParamPositions[FD];
6784 auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
6785 ->getCanonicalDecl();
6786 Pos = ParamPositions[PVD];
6788 auto &ParamAttr = ParamAttrs[Pos];
6789 ParamAttr.Kind = Linear;
6791 if (!(*SI)->EvaluateAsInt(ParamAttr.StrideOrArg, C,
6792 Expr::SE_AllowSideEffects)) {
6793 if (auto *DRE = cast<DeclRefExpr>((*SI)->IgnoreParenImpCasts())) {
6794 if (auto *StridePVD = cast<ParmVarDecl>(DRE->getDecl())) {
6795 ParamAttr.Kind = LinearWithVarStride;
6796 ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(
6797 ParamPositions[StridePVD->getCanonicalDecl()]);
6805 llvm::APSInt VLENVal;
6806 if (const Expr *VLEN = Attr->getSimdlen())
6807 VLENVal = VLEN->EvaluateKnownConstInt(C);
6808 OMPDeclareSimdDeclAttr::BranchStateTy State = Attr->getBranchState();
6809 if (CGM.getTriple().getArch() == llvm::Triple::x86 ||
6810 CGM.getTriple().getArch() == llvm::Triple::x86_64)
6811 emitX86DeclareSimdFunction(FD, Fn, VLENVal, ParamAttrs, State);
6816 /// Cleanup action for doacross support.
6817 class DoacrossCleanupTy final : public EHScopeStack::Cleanup {
6819 static const int DoacrossFinArgs = 2;
6823 llvm::Value *Args[DoacrossFinArgs];
6826 DoacrossCleanupTy(llvm::Value *RTLFn, ArrayRef<llvm::Value *> CallArgs)
6828 assert(CallArgs.size() == DoacrossFinArgs);
6829 std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args));
6831 void Emit(CodeGenFunction &CGF, Flags /*flags*/) override {
6832 if (!CGF.HaveInsertPoint())
6834 CGF.EmitRuntimeCall(RTLFn, Args);
6839 void CGOpenMPRuntime::emitDoacrossInit(CodeGenFunction &CGF,
6840 const OMPLoopDirective &D) {
6841 if (!CGF.HaveInsertPoint())
6844 ASTContext &C = CGM.getContext();
6845 QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true);
6847 if (KmpDimTy.isNull()) {
6848 // Build struct kmp_dim { // loop bounds info casted to kmp_int64
6849 // kmp_int64 lo; // lower
6850 // kmp_int64 up; // upper
6851 // kmp_int64 st; // stride
6853 RD = C.buildImplicitRecord("kmp_dim");
6854 RD->startDefinition();
6855 addFieldToRecordDecl(C, RD, Int64Ty);
6856 addFieldToRecordDecl(C, RD, Int64Ty);
6857 addFieldToRecordDecl(C, RD, Int64Ty);
6858 RD->completeDefinition();
6859 KmpDimTy = C.getRecordType(RD);
6861 RD = cast<RecordDecl>(KmpDimTy->getAsTagDecl());
6863 Address DimsAddr = CGF.CreateMemTemp(KmpDimTy, "dims");
6864 CGF.EmitNullInitialization(DimsAddr, KmpDimTy);
6865 enum { LowerFD = 0, UpperFD, StrideFD };
6866 // Fill dims with data.
6867 LValue DimsLVal = CGF.MakeAddrLValue(DimsAddr, KmpDimTy);
6868 // dims.upper = num_iterations;
6870 CGF.EmitLValueForField(DimsLVal, *std::next(RD->field_begin(), UpperFD));
6871 llvm::Value *NumIterVal = CGF.EmitScalarConversion(
6872 CGF.EmitScalarExpr(D.getNumIterations()), D.getNumIterations()->getType(),
6873 Int64Ty, D.getNumIterations()->getExprLoc());
6874 CGF.EmitStoreOfScalar(NumIterVal, UpperLVal);
6877 CGF.EmitLValueForField(DimsLVal, *std::next(RD->field_begin(), StrideFD));
6878 CGF.EmitStoreOfScalar(llvm::ConstantInt::getSigned(CGM.Int64Ty, /*V=*/1),
6881 // Build call void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
6882 // kmp_int32 num_dims, struct kmp_dim * dims);
6883 llvm::Value *Args[] = {emitUpdateLocation(CGF, D.getLocStart()),
6884 getThreadID(CGF, D.getLocStart()),
6885 llvm::ConstantInt::getSigned(CGM.Int32Ty, 1),
6886 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
6887 DimsAddr.getPointer(), CGM.VoidPtrTy)};
6889 llvm::Value *RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_init);
6890 CGF.EmitRuntimeCall(RTLFn, Args);
6891 llvm::Value *FiniArgs[DoacrossCleanupTy::DoacrossFinArgs] = {
6892 emitUpdateLocation(CGF, D.getLocEnd()), getThreadID(CGF, D.getLocEnd())};
6893 llvm::Value *FiniRTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_fini);
6894 CGF.EHStack.pushCleanup<DoacrossCleanupTy>(NormalAndEHCleanup, FiniRTLFn,
6895 llvm::makeArrayRef(FiniArgs));
6898 void CGOpenMPRuntime::emitDoacrossOrdered(CodeGenFunction &CGF,
6899 const OMPDependClause *C) {
6901 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
6902 const Expr *CounterVal = C->getCounterValue();
6904 llvm::Value *CntVal = CGF.EmitScalarConversion(CGF.EmitScalarExpr(CounterVal),
6905 CounterVal->getType(), Int64Ty,
6906 CounterVal->getExprLoc());
6907 Address CntAddr = CGF.CreateMemTemp(Int64Ty, ".cnt.addr");
6908 CGF.EmitStoreOfScalar(CntVal, CntAddr, /*Volatile=*/false, Int64Ty);
6909 llvm::Value *Args[] = {emitUpdateLocation(CGF, C->getLocStart()),
6910 getThreadID(CGF, C->getLocStart()),
6911 CntAddr.getPointer()};
6913 if (C->getDependencyKind() == OMPC_DEPEND_source)
6914 RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_post);
6916 assert(C->getDependencyKind() == OMPC_DEPEND_sink);
6917 RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_wait);
6919 CGF.EmitRuntimeCall(RTLFn, Args);