Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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" |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 16 | #include "clang/AST/StmtOpenMP.h" |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 17 | #include "clang/AST/Decl.h" |
| 18 | #include "llvm/ADT/ArrayRef.h" |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 19 | #include "llvm/IR/CallSite.h" |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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" |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 24 | #include <cassert> |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 25 | |
| 26 | using namespace clang; |
| 27 | using namespace CodeGen; |
| 28 | |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 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 | |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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 */, |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 92 | CGM.Int8PtrTy /* psource */, nullptr); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 93 | // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...) |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 94 | llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty), |
| 95 | llvm::PointerType::getUnqual(CGM.Int32Ty)}; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 96 | Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true); |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 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); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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 | } |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 124 | auto DefaultOpenMPLocation = new llvm::GlobalVariable( |
| 125 | CGM.getModule(), IdentTy, /*isConstant*/ true, |
| 126 | llvm::GlobalValue::PrivateLinkage, /*Initializer*/ nullptr); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 127 | DefaultOpenMPLocation->setUnnamedAddr(true); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 128 | |
| 129 | llvm::Constant *Zero = llvm::ConstantInt::get(CGM.Int32Ty, 0, true); |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 130 | llvm::Constant *Values[] = {Zero, |
| 131 | llvm::ConstantInt::get(CGM.Int32Ty, Flags), |
| 132 | Zero, Zero, DefaultOpenMPPSource}; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 133 | llvm::Constant *Init = llvm::ConstantStruct::get(IdentTy, Values); |
| 134 | DefaultOpenMPLocation->setInitializer(Init); |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 135 | OpenMPDefaultLocMap[Flags] = DefaultOpenMPLocation; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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; |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 151 | OpenMPLocThreadIDMapTy::iterator I = OpenMPLocThreadIDMap.find(CGF.CurFn); |
| 152 | if (I != OpenMPLocThreadIDMap.end()) { |
| 153 | LocValue = I->second.DebugLoc; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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)); |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 158 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); |
| 159 | Elem.second.DebugLoc = AI; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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 | |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 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; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 187 | } |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 188 | // *psource = ";<File>;<Function>;<Line>;<Column>;;"; |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 189 | CGF.Builder.CreateStore(OMPDebugLoc, PSource); |
| 190 | |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 191 | return LocValue; |
| 192 | } |
| 193 | |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 194 | llvm::Value *CGOpenMPRuntime::GetOpenMPThreadID(CodeGenFunction &CGF, |
| 195 | SourceLocation Loc) { |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 196 | assert(CGF.CurFn && "No function in current CodeGenFunction."); |
| 197 | |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 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 | } |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 222 | } else { |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 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. |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 227 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); |
| 228 | CGF.Builder.SetInsertPoint(CGF.AllocaInsertPt); |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 229 | llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc)}; |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 230 | ThreadID = CGF.EmitRuntimeCall( |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 231 | CreateRuntimeFunction(OMPRTL__kmpc_global_thread_num), Args); |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 232 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); |
| 233 | Elem.second.ThreadID = ThreadID; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 234 | } |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 235 | return ThreadID; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 236 | } |
| 237 | |
| 238 | void CGOpenMPRuntime::FunctionFinished(CodeGenFunction &CGF) { |
| 239 | assert(CGF.CurFn && "No function in current CodeGenFunction."); |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 240 | if (OpenMPLocThreadIDMap.count(CGF.CurFn)) |
| 241 | OpenMPLocThreadIDMap.erase(CGF.CurFn); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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, ...); |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 259 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, |
| 260 | getKmpc_MicroPointerTy()}; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 261 | llvm::FunctionType *FnTy = |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 262 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 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); |
Stephen Hines | c568f1e | 2014-07-21 00:47:37 -0700 | [diff] [blame] | 268 | llvm::Type *TypeParams[] = {getIdentTyPointerTy()}; |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 269 | llvm::FunctionType *FnTy = |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 270 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 271 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_global_thread_num"); |
| 272 | break; |
| 273 | } |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 274 | case OMPRTL__kmpc_threadprivate_cached: { |
| 275 | // Build void *__kmpc_threadprivate_cached(ident_t *loc, |
| 276 | // kmp_int32 global_tid, void *data, size_t size, void ***cache); |
| 277 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, |
| 278 | CGM.VoidPtrTy, CGM.SizeTy, |
| 279 | CGM.VoidPtrTy->getPointerTo()->getPointerTo()}; |
| 280 | llvm::FunctionType *FnTy = |
| 281 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg*/ false); |
| 282 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_threadprivate_cached"); |
| 283 | break; |
| 284 | } |
| 285 | case OMPRTL__kmpc_critical: { |
| 286 | // Build void __kmpc_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_critical"); |
| 294 | break; |
| 295 | } |
| 296 | case OMPRTL__kmpc_threadprivate_register: { |
| 297 | // Build void __kmpc_threadprivate_register(ident_t *, void *data, |
| 298 | // kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor); |
| 299 | // typedef void *(*kmpc_ctor)(void *); |
| 300 | auto KmpcCtorTy = |
| 301 | llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy, |
| 302 | /*isVarArg*/ false)->getPointerTo(); |
| 303 | // typedef void *(*kmpc_cctor)(void *, void *); |
| 304 | llvm::Type *KmpcCopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; |
| 305 | auto KmpcCopyCtorTy = |
| 306 | llvm::FunctionType::get(CGM.VoidPtrTy, KmpcCopyCtorTyArgs, |
| 307 | /*isVarArg*/ false)->getPointerTo(); |
| 308 | // typedef void (*kmpc_dtor)(void *); |
| 309 | auto KmpcDtorTy = |
| 310 | llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, /*isVarArg*/ false) |
| 311 | ->getPointerTo(); |
| 312 | llvm::Type *FnTyArgs[] = {getIdentTyPointerTy(), CGM.VoidPtrTy, KmpcCtorTy, |
| 313 | KmpcCopyCtorTy, KmpcDtorTy}; |
| 314 | auto FnTy = llvm::FunctionType::get(CGM.VoidTy, FnTyArgs, |
| 315 | /*isVarArg*/ false); |
| 316 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_threadprivate_register"); |
| 317 | break; |
| 318 | } |
| 319 | case OMPRTL__kmpc_end_critical: { |
| 320 | // Build void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid, |
| 321 | // kmp_critical_name *crit); |
| 322 | llvm::Type *TypeParams[] = { |
| 323 | getIdentTyPointerTy(), CGM.Int32Ty, |
| 324 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; |
| 325 | llvm::FunctionType *FnTy = |
| 326 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); |
| 327 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_critical"); |
| 328 | break; |
| 329 | } |
| 330 | case OMPRTL__kmpc_barrier: { |
| 331 | // Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid); |
| 332 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; |
| 333 | llvm::FunctionType *FnTy = |
| 334 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); |
| 335 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier"); |
| 336 | break; |
| 337 | } |
| 338 | case OMPRTL__kmpc_push_num_threads: { |
| 339 | // Build void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, |
| 340 | // kmp_int32 num_threads) |
| 341 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, |
| 342 | CGM.Int32Ty}; |
| 343 | llvm::FunctionType *FnTy = |
| 344 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); |
| 345 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_num_threads"); |
| 346 | break; |
| 347 | } |
| 348 | case OMPRTL__kmpc_serialized_parallel: { |
| 349 | // Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 |
| 350 | // global_tid); |
| 351 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; |
| 352 | llvm::FunctionType *FnTy = |
| 353 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); |
| 354 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel"); |
| 355 | break; |
| 356 | } |
| 357 | case OMPRTL__kmpc_end_serialized_parallel: { |
| 358 | // Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 |
| 359 | // global_tid); |
| 360 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; |
| 361 | llvm::FunctionType *FnTy = |
| 362 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); |
| 363 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel"); |
| 364 | break; |
| 365 | } |
| 366 | case OMPRTL__kmpc_flush: { |
| 367 | // Build void __kmpc_flush(ident_t *loc, ...); |
| 368 | llvm::Type *TypeParams[] = {getIdentTyPointerTy()}; |
| 369 | llvm::FunctionType *FnTy = |
| 370 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true); |
| 371 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_flush"); |
| 372 | break; |
| 373 | } |
Stephen Hines | 6bcf27b | 2014-05-29 04:14:42 -0700 | [diff] [blame] | 374 | } |
| 375 | return RTLFn; |
| 376 | } |
Stephen Hines | 176edba | 2014-12-01 14:53:08 -0800 | [diff] [blame] | 377 | |
| 378 | llvm::Constant * |
| 379 | CGOpenMPRuntime::getOrCreateThreadPrivateCache(const VarDecl *VD) { |
| 380 | // Lookup the entry, lazily creating it if necessary. |
| 381 | return GetOrCreateInternalVariable(CGM.Int8PtrPtrTy, |
| 382 | Twine(CGM.getMangledName(VD)) + ".cache."); |
| 383 | } |
| 384 | |
| 385 | llvm::Value *CGOpenMPRuntime::getOMPAddrOfThreadPrivate(CodeGenFunction &CGF, |
| 386 | const VarDecl *VD, |
| 387 | llvm::Value *VDAddr, |
| 388 | SourceLocation Loc) { |
| 389 | auto VarTy = VDAddr->getType()->getPointerElementType(); |
| 390 | llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc), |
| 391 | GetOpenMPThreadID(CGF, Loc), |
| 392 | CGF.Builder.CreatePointerCast(VDAddr, CGM.Int8PtrTy), |
| 393 | CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)), |
| 394 | getOrCreateThreadPrivateCache(VD)}; |
| 395 | return CGF.EmitRuntimeCall( |
| 396 | CreateRuntimeFunction(OMPRTL__kmpc_threadprivate_cached), Args); |
| 397 | } |
| 398 | |
| 399 | void CGOpenMPRuntime::EmitOMPThreadPrivateVarInit( |
| 400 | CodeGenFunction &CGF, llvm::Value *VDAddr, llvm::Value *Ctor, |
| 401 | llvm::Value *CopyCtor, llvm::Value *Dtor, SourceLocation Loc) { |
| 402 | // Call kmp_int32 __kmpc_global_thread_num(&loc) to init OpenMP runtime |
| 403 | // library. |
| 404 | auto OMPLoc = EmitOpenMPUpdateLocation(CGF, Loc); |
| 405 | CGF.EmitRuntimeCall(CreateRuntimeFunction(OMPRTL__kmpc_global_thread_num), |
| 406 | OMPLoc); |
| 407 | // Call __kmpc_threadprivate_register(&loc, &var, ctor, cctor/*NULL*/, dtor) |
| 408 | // to register constructor/destructor for variable. |
| 409 | llvm::Value *Args[] = {OMPLoc, |
| 410 | CGF.Builder.CreatePointerCast(VDAddr, CGM.VoidPtrTy), |
| 411 | Ctor, CopyCtor, Dtor}; |
| 412 | CGF.EmitRuntimeCall(CreateRuntimeFunction( |
| 413 | CGOpenMPRuntime::OMPRTL__kmpc_threadprivate_register), |
| 414 | Args); |
| 415 | } |
| 416 | |
| 417 | llvm::Function *CGOpenMPRuntime::EmitOMPThreadPrivateVarDefinition( |
| 418 | const VarDecl *VD, llvm::Value *VDAddr, SourceLocation Loc, |
| 419 | bool PerformInit, CodeGenFunction *CGF) { |
| 420 | VD = VD->getDefinition(CGM.getContext()); |
| 421 | if (VD && ThreadPrivateWithDefinition.count(VD) == 0) { |
| 422 | ThreadPrivateWithDefinition.insert(VD); |
| 423 | QualType ASTTy = VD->getType(); |
| 424 | |
| 425 | llvm::Value *Ctor = nullptr, *CopyCtor = nullptr, *Dtor = nullptr; |
| 426 | auto Init = VD->getAnyInitializer(); |
| 427 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { |
| 428 | // Generate function that re-emits the declaration's initializer into the |
| 429 | // threadprivate copy of the variable VD |
| 430 | CodeGenFunction CtorCGF(CGM); |
| 431 | FunctionArgList Args; |
| 432 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, SourceLocation(), |
| 433 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy); |
| 434 | Args.push_back(&Dst); |
| 435 | |
| 436 | auto &FI = CGM.getTypes().arrangeFreeFunctionDeclaration( |
| 437 | CGM.getContext().VoidPtrTy, Args, FunctionType::ExtInfo(), |
| 438 | /*isVariadic=*/false); |
| 439 | auto FTy = CGM.getTypes().GetFunctionType(FI); |
| 440 | auto Fn = CGM.CreateGlobalInitOrDestructFunction( |
| 441 | FTy, ".__kmpc_global_ctor_.", Loc); |
| 442 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidPtrTy, Fn, FI, |
| 443 | Args, SourceLocation()); |
| 444 | auto ArgVal = CtorCGF.EmitLoadOfScalar( |
| 445 | CtorCGF.GetAddrOfLocalVar(&Dst), |
| 446 | /*Volatile=*/false, CGM.PointerAlignInBytes, |
| 447 | CGM.getContext().VoidPtrTy, Dst.getLocation()); |
| 448 | auto Arg = CtorCGF.Builder.CreatePointerCast( |
| 449 | ArgVal, |
| 450 | CtorCGF.ConvertTypeForMem(CGM.getContext().getPointerType(ASTTy))); |
| 451 | CtorCGF.EmitAnyExprToMem(Init, Arg, Init->getType().getQualifiers(), |
| 452 | /*IsInitializer=*/true); |
| 453 | ArgVal = CtorCGF.EmitLoadOfScalar( |
| 454 | CtorCGF.GetAddrOfLocalVar(&Dst), |
| 455 | /*Volatile=*/false, CGM.PointerAlignInBytes, |
| 456 | CGM.getContext().VoidPtrTy, Dst.getLocation()); |
| 457 | CtorCGF.Builder.CreateStore(ArgVal, CtorCGF.ReturnValue); |
| 458 | CtorCGF.FinishFunction(); |
| 459 | Ctor = Fn; |
| 460 | } |
| 461 | if (VD->getType().isDestructedType() != QualType::DK_none) { |
| 462 | // Generate function that emits destructor call for the threadprivate copy |
| 463 | // of the variable VD |
| 464 | CodeGenFunction DtorCGF(CGM); |
| 465 | FunctionArgList Args; |
| 466 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, SourceLocation(), |
| 467 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy); |
| 468 | Args.push_back(&Dst); |
| 469 | |
| 470 | auto &FI = CGM.getTypes().arrangeFreeFunctionDeclaration( |
| 471 | CGM.getContext().VoidTy, Args, FunctionType::ExtInfo(), |
| 472 | /*isVariadic=*/false); |
| 473 | auto FTy = CGM.getTypes().GetFunctionType(FI); |
| 474 | auto Fn = CGM.CreateGlobalInitOrDestructFunction( |
| 475 | FTy, ".__kmpc_global_dtor_.", Loc); |
| 476 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, Args, |
| 477 | SourceLocation()); |
| 478 | auto ArgVal = DtorCGF.EmitLoadOfScalar( |
| 479 | DtorCGF.GetAddrOfLocalVar(&Dst), |
| 480 | /*Volatile=*/false, CGM.PointerAlignInBytes, |
| 481 | CGM.getContext().VoidPtrTy, Dst.getLocation()); |
| 482 | DtorCGF.emitDestroy(ArgVal, ASTTy, |
| 483 | DtorCGF.getDestroyer(ASTTy.isDestructedType()), |
| 484 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); |
| 485 | DtorCGF.FinishFunction(); |
| 486 | Dtor = Fn; |
| 487 | } |
| 488 | // Do not emit init function if it is not required. |
| 489 | if (!Ctor && !Dtor) |
| 490 | return nullptr; |
| 491 | |
| 492 | llvm::Type *CopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; |
| 493 | auto CopyCtorTy = |
| 494 | llvm::FunctionType::get(CGM.VoidPtrTy, CopyCtorTyArgs, |
| 495 | /*isVarArg=*/false)->getPointerTo(); |
| 496 | // Copying constructor for the threadprivate variable. |
| 497 | // Must be NULL - reserved by runtime, but currently it requires that this |
| 498 | // parameter is always NULL. Otherwise it fires assertion. |
| 499 | CopyCtor = llvm::Constant::getNullValue(CopyCtorTy); |
| 500 | if (Ctor == nullptr) { |
| 501 | auto CtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy, |
| 502 | /*isVarArg=*/false)->getPointerTo(); |
| 503 | Ctor = llvm::Constant::getNullValue(CtorTy); |
| 504 | } |
| 505 | if (Dtor == nullptr) { |
| 506 | auto DtorTy = llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, |
| 507 | /*isVarArg=*/false)->getPointerTo(); |
| 508 | Dtor = llvm::Constant::getNullValue(DtorTy); |
| 509 | } |
| 510 | if (!CGF) { |
| 511 | auto InitFunctionTy = |
| 512 | llvm::FunctionType::get(CGM.VoidTy, /*isVarArg*/ false); |
| 513 | auto InitFunction = CGM.CreateGlobalInitOrDestructFunction( |
| 514 | InitFunctionTy, ".__omp_threadprivate_init_."); |
| 515 | CodeGenFunction InitCGF(CGM); |
| 516 | FunctionArgList ArgList; |
| 517 | InitCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, InitFunction, |
| 518 | CGM.getTypes().arrangeNullaryFunction(), ArgList, |
| 519 | Loc); |
| 520 | EmitOMPThreadPrivateVarInit(InitCGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); |
| 521 | InitCGF.FinishFunction(); |
| 522 | return InitFunction; |
| 523 | } |
| 524 | EmitOMPThreadPrivateVarInit(*CGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); |
| 525 | } |
| 526 | return nullptr; |
| 527 | } |
| 528 | |
| 529 | void CGOpenMPRuntime::EmitOMPParallelCall(CodeGenFunction &CGF, |
| 530 | SourceLocation Loc, |
| 531 | llvm::Value *OutlinedFn, |
| 532 | llvm::Value *CapturedStruct) { |
| 533 | // Build call __kmpc_fork_call(loc, 1, microtask, captured_struct/*context*/) |
| 534 | llvm::Value *Args[] = { |
| 535 | EmitOpenMPUpdateLocation(CGF, Loc), |
| 536 | CGF.Builder.getInt32(1), // Number of arguments after 'microtask' argument |
| 537 | // (there is only one additional argument - 'context') |
| 538 | CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy()), |
| 539 | CGF.EmitCastToVoidPtr(CapturedStruct)}; |
| 540 | auto RTLFn = CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_fork_call); |
| 541 | CGF.EmitRuntimeCall(RTLFn, Args); |
| 542 | } |
| 543 | |
| 544 | void CGOpenMPRuntime::EmitOMPSerialCall(CodeGenFunction &CGF, |
| 545 | SourceLocation Loc, |
| 546 | llvm::Value *OutlinedFn, |
| 547 | llvm::Value *CapturedStruct) { |
| 548 | auto ThreadID = GetOpenMPThreadID(CGF, Loc); |
| 549 | // Build calls: |
| 550 | // __kmpc_serialized_parallel(&Loc, GTid); |
| 551 | llvm::Value *SerArgs[] = {EmitOpenMPUpdateLocation(CGF, Loc), ThreadID}; |
| 552 | auto RTLFn = |
| 553 | CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_serialized_parallel); |
| 554 | CGF.EmitRuntimeCall(RTLFn, SerArgs); |
| 555 | |
| 556 | // OutlinedFn(>id, &zero, CapturedStruct); |
| 557 | auto ThreadIDAddr = EmitThreadIDAddress(CGF, Loc); |
| 558 | auto Int32Ty = |
| 559 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true); |
| 560 | auto ZeroAddr = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".zero.addr"); |
| 561 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); |
| 562 | llvm::Value *OutlinedFnArgs[] = {ThreadIDAddr, ZeroAddr, CapturedStruct}; |
| 563 | CGF.EmitCallOrInvoke(OutlinedFn, OutlinedFnArgs); |
| 564 | |
| 565 | // __kmpc_end_serialized_parallel(&Loc, GTid); |
| 566 | llvm::Value *EndSerArgs[] = {EmitOpenMPUpdateLocation(CGF, Loc), ThreadID}; |
| 567 | RTLFn = CreateRuntimeFunction( |
| 568 | CGOpenMPRuntime::OMPRTL__kmpc_end_serialized_parallel); |
| 569 | CGF.EmitRuntimeCall(RTLFn, EndSerArgs); |
| 570 | } |
| 571 | |
| 572 | // If we're inside an (outlined) parallel region, use the region info's |
| 573 | // thread-ID variable (it is passed in a first argument of the outlined function |
| 574 | // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in |
| 575 | // regular serial code region, get thread ID by calling kmp_int32 |
| 576 | // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and |
| 577 | // return the address of that temp. |
| 578 | llvm::Value *CGOpenMPRuntime::EmitThreadIDAddress(CodeGenFunction &CGF, |
| 579 | SourceLocation Loc) { |
| 580 | if (auto OMPRegionInfo = |
| 581 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) |
| 582 | return CGF.EmitLoadOfLValue(OMPRegionInfo->getThreadIDVariableLValue(CGF), |
| 583 | SourceLocation()).getScalarVal(); |
| 584 | auto ThreadID = GetOpenMPThreadID(CGF, Loc); |
| 585 | auto Int32Ty = |
| 586 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true); |
| 587 | auto ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp."); |
| 588 | CGF.EmitStoreOfScalar(ThreadID, |
| 589 | CGF.MakeNaturalAlignAddrLValue(ThreadIDTemp, Int32Ty)); |
| 590 | |
| 591 | return ThreadIDTemp; |
| 592 | } |
| 593 | |
| 594 | llvm::Constant * |
| 595 | CGOpenMPRuntime::GetOrCreateInternalVariable(llvm::Type *Ty, |
| 596 | const llvm::Twine &Name) { |
| 597 | SmallString<256> Buffer; |
| 598 | llvm::raw_svector_ostream Out(Buffer); |
| 599 | Out << Name; |
| 600 | auto RuntimeName = Out.str(); |
| 601 | auto &Elem = *InternalVars.insert(std::make_pair(RuntimeName, nullptr)).first; |
| 602 | if (Elem.second) { |
| 603 | assert(Elem.second->getType()->getPointerElementType() == Ty && |
| 604 | "OMP internal variable has different type than requested"); |
| 605 | return &*Elem.second; |
| 606 | } |
| 607 | |
| 608 | return Elem.second = new llvm::GlobalVariable( |
| 609 | CGM.getModule(), Ty, /*IsConstant*/ false, |
| 610 | llvm::GlobalValue::CommonLinkage, llvm::Constant::getNullValue(Ty), |
| 611 | Elem.first()); |
| 612 | } |
| 613 | |
| 614 | llvm::Value *CGOpenMPRuntime::GetCriticalRegionLock(StringRef CriticalName) { |
| 615 | llvm::Twine Name(".gomp_critical_user_", CriticalName); |
| 616 | return GetOrCreateInternalVariable(KmpCriticalNameTy, Name.concat(".var")); |
| 617 | } |
| 618 | |
| 619 | void CGOpenMPRuntime::EmitOMPCriticalRegionStart(CodeGenFunction &CGF, |
| 620 | llvm::Value *RegionLock, |
| 621 | SourceLocation Loc) { |
| 622 | // Prepare other arguments and build a call to __kmpc_critical |
| 623 | llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc), |
| 624 | GetOpenMPThreadID(CGF, Loc), RegionLock}; |
| 625 | auto RTLFn = CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_critical); |
| 626 | CGF.EmitRuntimeCall(RTLFn, Args); |
| 627 | } |
| 628 | |
| 629 | void CGOpenMPRuntime::EmitOMPCriticalRegionEnd(CodeGenFunction &CGF, |
| 630 | llvm::Value *RegionLock, |
| 631 | SourceLocation Loc) { |
| 632 | // Prepare other arguments and build a call to __kmpc_end_critical |
| 633 | llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc), |
| 634 | GetOpenMPThreadID(CGF, Loc), RegionLock}; |
| 635 | auto RTLFn = |
| 636 | CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_end_critical); |
| 637 | CGF.EmitRuntimeCall(RTLFn, Args); |
| 638 | } |
| 639 | |
| 640 | void CGOpenMPRuntime::EmitOMPBarrierCall(CodeGenFunction &CGF, |
| 641 | SourceLocation Loc, |
| 642 | OpenMPLocationFlags Flags) { |
| 643 | // Build call __kmpc_barrier(loc, thread_id) |
| 644 | llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc, Flags), |
| 645 | GetOpenMPThreadID(CGF, Loc)}; |
| 646 | auto RTLFn = CreateRuntimeFunction(CGOpenMPRuntime::OMPRTL__kmpc_barrier); |
| 647 | CGF.EmitRuntimeCall(RTLFn, Args); |
| 648 | } |
| 649 | |
| 650 | void CGOpenMPRuntime::EmitOMPNumThreadsClause(CodeGenFunction &CGF, |
| 651 | llvm::Value *NumThreads, |
| 652 | SourceLocation Loc) { |
| 653 | // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads) |
| 654 | llvm::Value *Args[] = { |
| 655 | EmitOpenMPUpdateLocation(CGF, Loc), GetOpenMPThreadID(CGF, Loc), |
| 656 | CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)}; |
| 657 | llvm::Constant *RTLFn = CGF.CGM.getOpenMPRuntime().CreateRuntimeFunction( |
| 658 | CGOpenMPRuntime::OMPRTL__kmpc_push_num_threads); |
| 659 | CGF.EmitRuntimeCall(RTLFn, Args); |
| 660 | } |
| 661 | |
| 662 | void CGOpenMPRuntime::EmitOMPFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>, |
| 663 | SourceLocation Loc) { |
| 664 | // Build call void __kmpc_flush(ident_t *loc, ...) |
| 665 | // FIXME: List of variables is ignored by libiomp5 runtime, no need to |
| 666 | // generate it, just request full memory fence. |
| 667 | llvm::Value *Args[] = {EmitOpenMPUpdateLocation(CGF, Loc), |
| 668 | llvm::ConstantInt::get(CGM.Int32Ty, 0)}; |
| 669 | auto *RTLFn = CGF.CGM.getOpenMPRuntime().CreateRuntimeFunction( |
| 670 | CGOpenMPRuntime::OMPRTL__kmpc_flush); |
| 671 | CGF.EmitRuntimeCall(RTLFn, Args); |
| 672 | } |