1 //===----- CGOpenMPRuntime.cpp - Interface to OpenMP Runtimes -------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This provides a class for OpenMP runtime code generation.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CGOpenMPRuntime.h"
15 #include "CodeGenFunction.h"
16 #include "clang/AST/StmtOpenMP.h"
17 #include "clang/AST/Decl.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/IR/CallSite.h"
20 #include "llvm/IR/DerivedTypes.h"
21 #include "llvm/IR/GlobalValue.h"
22 #include "llvm/IR/Value.h"
23 #include "llvm/Support/raw_ostream.h"
24 #include <cassert>
25 
26 using namespace clang;
27 using namespace CodeGen;
28 
29 namespace {
30 /// \brief API for captured statement code generation in OpenMP constructs.
31 class CGOpenMPRegionInfo : public CodeGenFunction::CGCapturedStmtInfo {
32 public:
33   CGOpenMPRegionInfo(const OMPExecutableDirective &D, const CapturedStmt &CS,
34                      const VarDecl *ThreadIDVar)
35       : CGCapturedStmtInfo(CS, CR_OpenMP), ThreadIDVar(ThreadIDVar),
36         Directive(D) {
37     assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.");
38   }
39 
40   /// \brief Gets a variable or parameter for storing global thread id
41   /// inside OpenMP construct.
42   const VarDecl *getThreadIDVariable() const { return ThreadIDVar; }
43 
44   /// \brief Gets an LValue for the current ThreadID variable.
45   LValue getThreadIDVariableLValue(CodeGenFunction &CGF);
46 
47   static bool classof(const CGCapturedStmtInfo *Info) {
48     return Info->getKind() == CR_OpenMP;
49   }
50 
51   /// \brief Emit the captured statement body.
52   void EmitBody(CodeGenFunction &CGF, Stmt *S) override;
53 
54   /// \brief Get the name of the capture helper.
55   StringRef getHelperName() const override { return ".omp_outlined."; }
56 
57 private:
58   /// \brief A variable or parameter storing global thread id for OpenMP
59   /// constructs.
60   const VarDecl *ThreadIDVar;
61   /// \brief OpenMP executable directive associated with the region.
62   const OMPExecutableDirective &Directive;
63 };
64 } // namespace
65 
66 LValue CGOpenMPRegionInfo::getThreadIDVariableLValue(CodeGenFunction &CGF) {
67   return CGF.MakeNaturalAlignAddrLValue(
68       CGF.GetAddrOfLocalVar(ThreadIDVar),
69       CGF.getContext().getPointerType(ThreadIDVar->getType()));
70 }
71 
72 void CGOpenMPRegionInfo::EmitBody(CodeGenFunction &CGF, Stmt *S) {
73   CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
74   CGF.EmitOMPPrivateClause(Directive, PrivateScope);
75   CGF.EmitOMPFirstprivateClause(Directive, PrivateScope);
76   if (PrivateScope.Privatize()) {
77     // Emit implicit barrier to synchronize threads and avoid data races.
78     auto Flags = static_cast<CGOpenMPRuntime::OpenMPLocationFlags>(
79         CGOpenMPRuntime::OMP_IDENT_KMPC |
80         CGOpenMPRuntime::OMP_IDENT_BARRIER_IMPL);
81     CGF.CGM.getOpenMPRuntime().EmitOMPBarrierCall(CGF, Directive.getLocStart(),
82                                                   Flags);
83   }
84   CGCapturedStmtInfo::EmitBody(CGF, S);
85 }
86 
87 CGOpenMPRuntime::CGOpenMPRuntime(CodeGenModule &CGM)
88     : CGM(CGM), DefaultOpenMPPSource(nullptr) {
89   IdentTy = llvm::StructType::create(
90       "ident_t", CGM.Int32Ty /* reserved_1 */, CGM.Int32Ty /* flags */,
91       CGM.Int32Ty /* reserved_2 */, CGM.Int32Ty /* reserved_3 */,
92       CGM.Int8PtrTy /* psource */, nullptr);
93   // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
94   llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty),
95                                llvm::PointerType::getUnqual(CGM.Int32Ty)};
96   Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true);
97   KmpCriticalNameTy = llvm::ArrayType::get(CGM.Int32Ty, /*NumElements*/ 8);
98 }
99 
100 llvm::Value *
101 CGOpenMPRuntime::EmitOpenMPOutlinedFunction(const OMPExecutableDirective &D,
102                                             const VarDecl *ThreadIDVar) {
103   const CapturedStmt *CS = cast<CapturedStmt>(D.getAssociatedStmt());
104   CodeGenFunction CGF(CGM, true);
105   CGOpenMPRegionInfo CGInfo(D, *CS, ThreadIDVar);
106   CGF.CapturedStmtInfo = &CGInfo;
107   return CGF.GenerateCapturedStmtFunction(*CS);
108 }
109 
110 llvm::Value *
111 CGOpenMPRuntime::GetOrCreateDefaultOpenMPLocation(OpenMPLocationFlags Flags) {
112   llvm::Value *Entry = OpenMPDefaultLocMap.lookup(Flags);
113   if (!Entry) {
114     if (!DefaultOpenMPPSource) {
115       // Initialize default location for psource field of ident_t structure of
116       // all ident_t objects. Format is ";file;function;line;column;;".
117       // Taken from
118       // http://llvm.org/svn/llvm-project/openmp/trunk/runtime/src/kmp_str.c
119       DefaultOpenMPPSource =
120           CGM.GetAddrOfConstantCString(";unknown;unknown;0;0;;");
121       DefaultOpenMPPSource =
122           llvm::ConstantExpr::getBitCast(DefaultOpenMPPSource, CGM.Int8PtrTy);
123     }
124     auto DefaultOpenMPLocation = new llvm::GlobalVariable(
125         CGM.getModule(), IdentTy, /*isConstant*/ true,
126         llvm::GlobalValue::PrivateLinkage, /*Initializer*/ nullptr);
127     DefaultOpenMPLocation->setUnnamedAddr(true);
128 
129     llvm::Constant *Zero = llvm::ConstantInt::get(CGM.Int32Ty, 0, true);
130     llvm::Constant *Values[] = {Zero,
131                                 llvm::ConstantInt::get(CGM.Int32Ty, Flags),
132                                 Zero, Zero, DefaultOpenMPPSource};
133     llvm::Constant *Init = llvm::ConstantStruct::get(IdentTy, Values);
134     DefaultOpenMPLocation->setInitializer(Init);
135     OpenMPDefaultLocMap[Flags] = DefaultOpenMPLocation;
136     return DefaultOpenMPLocation;
137   }
138   return Entry;
139 }
140 
141 llvm::Value *CGOpenMPRuntime::EmitOpenMPUpdateLocation(
142     CodeGenFunction &CGF, SourceLocation Loc, OpenMPLocationFlags Flags) {
143   // If no debug info is generated - return global default location.
144   if (CGM.getCodeGenOpts().getDebugInfo() == CodeGenOptions::NoDebugInfo ||
145       Loc.isInvalid())
146     return GetOrCreateDefaultOpenMPLocation(Flags);
147 
148   assert(CGF.CurFn && "No function in current CodeGenFunction.");
149 
150   llvm::Value *LocValue = nullptr;
151   OpenMPLocThreadIDMapTy::iterator I = OpenMPLocThreadIDMap.find(CGF.CurFn);
152   if (I != OpenMPLocThreadIDMap.end()) {
153     LocValue = I->second.DebugLoc;
154   } else {
155     // Generate "ident_t .kmpc_loc.addr;"
156     llvm::AllocaInst *AI = CGF.CreateTempAlloca(IdentTy, ".kmpc_loc.addr");
157     AI->setAlignment(CGM.getDataLayout().getPrefTypeAlignment(IdentTy));
158     auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
159     Elem.second.DebugLoc = AI;
160     LocValue = AI;
161 
162     CGBuilderTy::InsertPointGuard IPG(CGF.Builder);
163     CGF.Builder.SetInsertPoint(CGF.AllocaInsertPt);
164     CGF.Builder.CreateMemCpy(LocValue, GetOrCreateDefaultOpenMPLocation(Flags),
165                              llvm::ConstantExpr::getSizeOf(IdentTy),
166                              CGM.PointerAlignInBytes);
167   }
168 
169   // char **psource = &.kmpc_loc_<flags>.addr.psource;
170   llvm::Value *PSource =
171       CGF.Builder.CreateConstInBoundsGEP2_32(LocValue, 0, IdentField_PSource);
172 
173   auto OMPDebugLoc = OpenMPDebugLocMap.lookup(Loc.getRawEncoding());
174   if (OMPDebugLoc == nullptr) {
175     SmallString<128> Buffer2;
176     llvm::raw_svector_ostream OS2(Buffer2);
177     // Build debug location
178     PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc);
179     OS2 << ";" << PLoc.getFilename() << ";";
180     if (const FunctionDecl *FD =
181             dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) {
182       OS2 << FD->getQualifiedNameAsString();
183     }
184     OS2 << ";" << PLoc.getLine() << ";" << PLoc.getColumn() << ";;";
185     OMPDebugLoc = CGF.Builder.CreateGlobalStringPtr(OS2.str());
186     OpenMPDebugLocMap[Loc.getRawEncoding()] = OMPDebugLoc;
187   }
188   // *psource = ";<File>;<Function>;<Line>;<Column>;;";
189   CGF.Builder.CreateStore(OMPDebugLoc, PSource);
190 
191   return LocValue;
192 }
193 
194 llvm::Value *CGOpenMPRuntime::GetOpenMPThreadID(CodeGenFunction &CGF,
195                                                 SourceLocation Loc) {
196   assert(CGF.CurFn && "No function in current CodeGenFunction.");
197 
198   llvm::Value *ThreadID = nullptr;
199   // Check whether we've already cached a load of the thread id in this
200   // function.
201   OpenMPLocThreadIDMapTy::iterator I = OpenMPLocThreadIDMap.find(CGF.CurFn);
202   if (I != OpenMPLocThreadIDMap.end()) {
203     ThreadID = I->second.ThreadID;
204     if (ThreadID != nullptr)
205       return ThreadID;
206   }
207   if (auto OMPRegionInfo =
208                  dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
209     // Check if this an outlined function with thread id passed as argument.
210     auto ThreadIDVar = OMPRegionInfo->getThreadIDVariable();
211     auto LVal = OMPRegionInfo->getThreadIDVariableLValue(CGF);
212     auto RVal = CGF.EmitLoadOfLValue(LVal, Loc);
213     LVal = CGF.MakeNaturalAlignAddrLValue(RVal.getScalarVal(),
214                                           ThreadIDVar->getType());
215     ThreadID = CGF.EmitLoadOfLValue(LVal, Loc).getScalarVal();
216     // If value loaded in entry block, cache it and use it everywhere in
217     // function.
218     if (CGF.Builder.GetInsertBlock() == CGF.AllocaInsertPt->getParent()) {
219       auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
220       Elem.second.ThreadID = ThreadID;
221     }
222   } else {
223     // This is not an outlined function region - need to call __kmpc_int32
224     // kmpc_global_thread_num(ident_t *loc).
225     // Generate thread id value and cache this value for use across the
226     // function.
227     CGBuilderTy::InsertPointGuard IPG(CGF.Builder);
228     CGF.Builder.SetInsertPoint(CGF.AllocaInsertPt);
229     llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc)};
230     ThreadID = CGF.EmitRuntimeCall(
231         CreateRuntimeFunction(OMPRTL__kmpc_global_thread_num), Args);
232     auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
233     Elem.second.ThreadID = ThreadID;
234   }
235   return ThreadID;
236 }
237 
238 void CGOpenMPRuntime::FunctionFinished(CodeGenFunction &CGF) {
239   assert(CGF.CurFn && "No function in current CodeGenFunction.");
240   if (OpenMPLocThreadIDMap.count(CGF.CurFn))
241     OpenMPLocThreadIDMap.erase(CGF.CurFn);
242 }
243 
244 llvm::Type *CGOpenMPRuntime::getIdentTyPointerTy() {
245   return llvm::PointerType::getUnqual(IdentTy);
246 }
247 
248 llvm::Type *CGOpenMPRuntime::getKmpc_MicroPointerTy() {
249   return llvm::PointerType::getUnqual(Kmpc_MicroTy);
250 }
251 
252 llvm::Constant *
253 CGOpenMPRuntime::CreateRuntimeFunction(OpenMPRTLFunction Function) {
254   llvm::Constant *RTLFn = nullptr;
255   switch (Function) {
256   case OMPRTL__kmpc_fork_call: {
257     // Build void __kmpc_fork_call(ident_t *loc, kmp_int32 argc, kmpc_micro
258     // microtask, ...);
259     llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
260                                 getKmpc_MicroPointerTy()};
261     llvm::FunctionType *FnTy =
262         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true);
263     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_fork_call");
264     break;
265   }
266   case OMPRTL__kmpc_global_thread_num: {
267     // Build kmp_int32 __kmpc_global_thread_num(ident_t *loc);
268     llvm::Type *TypeParams[] = {getIdentTyPointerTy()};
269     llvm::FunctionType *FnTy =
270         llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
271     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_global_thread_num");
272     break;
273   }
274   case OMPRTL__kmpc_critical: {
275     // Build void __kmpc_critical(ident_t *loc, kmp_int32 global_tid,
276     // kmp_critical_name *crit);
277     llvm::Type *TypeParams[] = {
278         getIdentTyPointerTy(), CGM.Int32Ty,
279         llvm::PointerType::getUnqual(KmpCriticalNameTy)};
280     llvm::FunctionType *FnTy =
281         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
282     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_critical");
283     break;
284   }
285   case OMPRTL__kmpc_end_critical: {
286     // Build void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid,
287     // kmp_critical_name *crit);
288     llvm::Type *TypeParams[] = {
289         getIdentTyPointerTy(), CGM.Int32Ty,
290         llvm::PointerType::getUnqual(KmpCriticalNameTy)};
291     llvm::FunctionType *FnTy =
292         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
293     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_critical");
294     break;
295   }
296   case OMPRTL__kmpc_barrier: {
297     // Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
298     llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
299     llvm::FunctionType *FnTy =
300         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
301     RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier");
302     break;
303   }
304   case OMPRTL__kmpc_push_num_threads: {
305     // Build void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
306     // kmp_int32 num_threads)
307     llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
308                                 CGM.Int32Ty};
309     llvm::FunctionType *FnTy =
310         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
311     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_num_threads");
312     break;
313   }
314   case OMPRTL__kmpc_serialized_parallel: {
315     // Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32
316     // global_tid);
317     llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
318     llvm::FunctionType *FnTy =
319         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
320     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel");
321     break;
322   }
323   case OMPRTL__kmpc_end_serialized_parallel: {
324     // Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32
325     // global_tid);
326     llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
327     llvm::FunctionType *FnTy =
328         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
329     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel");
330     break;
331   }
332   }
333   return RTLFn;
334 }
335 
336 void CGOpenMPRuntime::EmitOMPParallelCall(CodeGenFunction &CGF,
337                                           SourceLocation Loc,
338                                           llvm::Value *OutlinedFn,
339                                           llvm::Value *CapturedStruct) {
340   // Build call __kmpc_fork_call(loc, 1, microtask, captured_struct/*context*/)
341   llvm::Value *Args[] = {
342       EmitOpenMPUpdateLocation(CGF, Loc),
343       CGF.Builder.getInt32(1), // Number of arguments after 'microtask' argument
344       // (there is only one additional argument - 'context')
345       CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy()),
346       CGF.EmitCastToVoidPtr(CapturedStruct)};
347   auto RTLFn = CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_fork_call);
348   CGF.EmitRuntimeCall(RTLFn, Args);
349 }
350 
351 void CGOpenMPRuntime::EmitOMPSerialCall(CodeGenFunction &CGF,
352                                         SourceLocation Loc,
353                                         llvm::Value *OutlinedFn,
354                                         llvm::Value *CapturedStruct) {
355   auto ThreadID = GetOpenMPThreadID(CGF, Loc);
356   // Build calls:
357   // __kmpc_serialized_parallel(&Loc, GTid);
358   llvm::Value *SerArgs[] = {EmitOpenMPUpdateLocation(CGF, Loc), ThreadID};
359   auto RTLFn =
360       CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_serialized_parallel);
361   CGF.EmitRuntimeCall(RTLFn, SerArgs);
362 
363   // OutlinedFn(&GTid, &zero, CapturedStruct);
364   auto ThreadIDAddr = EmitThreadIDAddress(CGF, Loc);
365   auto Int32Ty =
366       CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true);
367   auto ZeroAddr = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".zero.addr");
368   CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
369   llvm::Value *OutlinedFnArgs[] = {ThreadIDAddr, ZeroAddr, CapturedStruct};
370   CGF.EmitCallOrInvoke(OutlinedFn, OutlinedFnArgs);
371 
372   // __kmpc_end_serialized_parallel(&Loc, GTid);
373   llvm::Value *EndSerArgs[] = {EmitOpenMPUpdateLocation(CGF, Loc), ThreadID};
374   RTLFn = CreateRuntimeFunction(
375       CGOpenMPRuntime::OMPRTL__kmpc_end_serialized_parallel);
376   CGF.EmitRuntimeCall(RTLFn, EndSerArgs);
377 }
378 
379 // If we're inside an (outlined) parallel region, use the region info's
380 // thread-ID variable (it is passed in a first argument of the outlined function
381 // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in
382 // regular serial code region, get thread ID by calling kmp_int32
383 // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and
384 // return the address of that temp.
385 llvm::Value *CGOpenMPRuntime::EmitThreadIDAddress(CodeGenFunction &CGF,
386                                                   SourceLocation Loc) {
387   if (auto OMPRegionInfo =
388           dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
389     return CGF.EmitLoadOfLValue(OMPRegionInfo->getThreadIDVariableLValue(CGF),
390                                 SourceLocation()).getScalarVal();
391   auto ThreadID = GetOpenMPThreadID(CGF, Loc);
392   auto Int32Ty =
393       CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true);
394   auto ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp.");
395   CGF.EmitStoreOfScalar(ThreadID,
396                         CGF.MakeNaturalAlignAddrLValue(ThreadIDTemp, Int32Ty));
397 
398   return ThreadIDTemp;
399 }
400 
401 llvm::Value *CGOpenMPRuntime::GetCriticalRegionLock(StringRef CriticalName) {
402   SmallString<256> Buffer;
403   llvm::raw_svector_ostream Out(Buffer);
404   Out << ".gomp_critical_user_" << CriticalName << ".var";
405   auto RuntimeCriticalName = Out.str();
406   auto &Elem = CriticalRegionVarNames.GetOrCreateValue(RuntimeCriticalName);
407   if (Elem.getValue() != nullptr)
408     return Elem.getValue();
409 
410   auto Lock = new llvm::GlobalVariable(
411       CGM.getModule(), KmpCriticalNameTy, /*IsConstant*/ false,
412       llvm::GlobalValue::CommonLinkage,
413       llvm::Constant::getNullValue(KmpCriticalNameTy), Elem.getKey());
414   Elem.setValue(Lock);
415   return Lock;
416 }
417 
418 void CGOpenMPRuntime::EmitOMPCriticalRegionStart(CodeGenFunction &CGF,
419                                                  llvm::Value *RegionLock,
420                                                  SourceLocation Loc) {
421   // Prepare other arguments and build a call to __kmpc_critical
422   llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc),
423                          GetOpenMPThreadID(CGF, Loc), RegionLock};
424   auto RTLFn = CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_critical);
425   CGF.EmitRuntimeCall(RTLFn, Args);
426 }
427 
428 void CGOpenMPRuntime::EmitOMPCriticalRegionEnd(CodeGenFunction &CGF,
429                                                llvm::Value *RegionLock,
430                                                SourceLocation Loc) {
431   // Prepare other arguments and build a call to __kmpc_end_critical
432   llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc),
433                          GetOpenMPThreadID(CGF, Loc), RegionLock};
434   auto RTLFn =
435       CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_end_critical);
436   CGF.EmitRuntimeCall(RTLFn, Args);
437 }
438 
439 void CGOpenMPRuntime::EmitOMPBarrierCall(CodeGenFunction &CGF,
440                                          SourceLocation Loc,
441                                          OpenMPLocationFlags Flags) {
442   // Build call __kmpc_barrier(loc, thread_id)
443   llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc, Flags),
444                          GetOpenMPThreadID(CGF, Loc)};
445   auto RTLFn = CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_barrier);
446   CGF.EmitRuntimeCall(RTLFn, Args);
447 }
448 
449 void CGOpenMPRuntime::EmitOMPNumThreadsClause(CodeGenFunction &CGF,
450                                               llvm::Value *NumThreads,
451                                               SourceLocation Loc) {
452   // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads)
453   llvm::Value *Args[] = {
454       EmitOpenMPUpdateLocation(CGF, Loc), GetOpenMPThreadID(CGF, Loc),
455       CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)};
456   llvm::Constant *RTLFn = CGF.CGM.getOpenMPRuntime().CreateRuntimeFunction(
457       CGOpenMPRuntime::OMPRTL__kmpc_push_num_threads);
458   CGF.EmitRuntimeCall(RTLFn, Args);
459 }
460 
461