File: | build/source/clang/lib/CodeGen/CGOpenMPRuntime.cpp |
Warning: | line 7932, column 9 2nd function call argument is an uninitialized value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===----- CGOpenMPRuntime.cpp - Interface to OpenMP Runtimes -------------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This provides a class for OpenMP runtime code generation. | |||
10 | // | |||
11 | //===----------------------------------------------------------------------===// | |||
12 | ||||
13 | #include "CGOpenMPRuntime.h" | |||
14 | #include "CGCXXABI.h" | |||
15 | #include "CGCleanup.h" | |||
16 | #include "CGRecordLayout.h" | |||
17 | #include "CodeGenFunction.h" | |||
18 | #include "TargetInfo.h" | |||
19 | #include "clang/AST/APValue.h" | |||
20 | #include "clang/AST/Attr.h" | |||
21 | #include "clang/AST/Decl.h" | |||
22 | #include "clang/AST/OpenMPClause.h" | |||
23 | #include "clang/AST/StmtOpenMP.h" | |||
24 | #include "clang/AST/StmtVisitor.h" | |||
25 | #include "clang/Basic/BitmaskEnum.h" | |||
26 | #include "clang/Basic/FileManager.h" | |||
27 | #include "clang/Basic/OpenMPKinds.h" | |||
28 | #include "clang/Basic/SourceManager.h" | |||
29 | #include "clang/CodeGen/ConstantInitBuilder.h" | |||
30 | #include "llvm/ADT/ArrayRef.h" | |||
31 | #include "llvm/ADT/SetOperations.h" | |||
32 | #include "llvm/ADT/SmallBitVector.h" | |||
33 | #include "llvm/ADT/StringExtras.h" | |||
34 | #include "llvm/Bitcode/BitcodeReader.h" | |||
35 | #include "llvm/IR/Constants.h" | |||
36 | #include "llvm/IR/DerivedTypes.h" | |||
37 | #include "llvm/IR/GlobalValue.h" | |||
38 | #include "llvm/IR/InstrTypes.h" | |||
39 | #include "llvm/IR/Value.h" | |||
40 | #include "llvm/Support/AtomicOrdering.h" | |||
41 | #include "llvm/Support/Format.h" | |||
42 | #include "llvm/Support/raw_ostream.h" | |||
43 | #include <cassert> | |||
44 | #include <numeric> | |||
45 | ||||
46 | using namespace clang; | |||
47 | using namespace CodeGen; | |||
48 | using namespace llvm::omp; | |||
49 | ||||
50 | namespace { | |||
51 | /// Base class for handling code generation inside OpenMP regions. | |||
52 | class CGOpenMPRegionInfo : public CodeGenFunction::CGCapturedStmtInfo { | |||
53 | public: | |||
54 | /// Kinds of OpenMP regions used in codegen. | |||
55 | enum CGOpenMPRegionKind { | |||
56 | /// Region with outlined function for standalone 'parallel' | |||
57 | /// directive. | |||
58 | ParallelOutlinedRegion, | |||
59 | /// Region with outlined function for standalone 'task' directive. | |||
60 | TaskOutlinedRegion, | |||
61 | /// Region for constructs that do not require function outlining, | |||
62 | /// like 'for', 'sections', 'atomic' etc. directives. | |||
63 | InlinedRegion, | |||
64 | /// Region with outlined function for standalone 'target' directive. | |||
65 | TargetRegion, | |||
66 | }; | |||
67 | ||||
68 | CGOpenMPRegionInfo(const CapturedStmt &CS, | |||
69 | const CGOpenMPRegionKind RegionKind, | |||
70 | const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind, | |||
71 | bool HasCancel) | |||
72 | : CGCapturedStmtInfo(CS, CR_OpenMP), RegionKind(RegionKind), | |||
73 | CodeGen(CodeGen), Kind(Kind), HasCancel(HasCancel) {} | |||
74 | ||||
75 | CGOpenMPRegionInfo(const CGOpenMPRegionKind RegionKind, | |||
76 | const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind, | |||
77 | bool HasCancel) | |||
78 | : CGCapturedStmtInfo(CR_OpenMP), RegionKind(RegionKind), CodeGen(CodeGen), | |||
79 | Kind(Kind), HasCancel(HasCancel) {} | |||
80 | ||||
81 | /// Get a variable or parameter for storing global thread id | |||
82 | /// inside OpenMP construct. | |||
83 | virtual const VarDecl *getThreadIDVariable() const = 0; | |||
84 | ||||
85 | /// Emit the captured statement body. | |||
86 | void EmitBody(CodeGenFunction &CGF, const Stmt *S) override; | |||
87 | ||||
88 | /// Get an LValue for the current ThreadID variable. | |||
89 | /// \return LValue for thread id variable. This LValue always has type int32*. | |||
90 | virtual LValue getThreadIDVariableLValue(CodeGenFunction &CGF); | |||
91 | ||||
92 | virtual void emitUntiedSwitch(CodeGenFunction & /*CGF*/) {} | |||
93 | ||||
94 | CGOpenMPRegionKind getRegionKind() const { return RegionKind; } | |||
95 | ||||
96 | OpenMPDirectiveKind getDirectiveKind() const { return Kind; } | |||
97 | ||||
98 | bool hasCancel() const { return HasCancel; } | |||
99 | ||||
100 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
101 | return Info->getKind() == CR_OpenMP; | |||
102 | } | |||
103 | ||||
104 | ~CGOpenMPRegionInfo() override = default; | |||
105 | ||||
106 | protected: | |||
107 | CGOpenMPRegionKind RegionKind; | |||
108 | RegionCodeGenTy CodeGen; | |||
109 | OpenMPDirectiveKind Kind; | |||
110 | bool HasCancel; | |||
111 | }; | |||
112 | ||||
113 | /// API for captured statement code generation in OpenMP constructs. | |||
114 | class CGOpenMPOutlinedRegionInfo final : public CGOpenMPRegionInfo { | |||
115 | public: | |||
116 | CGOpenMPOutlinedRegionInfo(const CapturedStmt &CS, const VarDecl *ThreadIDVar, | |||
117 | const RegionCodeGenTy &CodeGen, | |||
118 | OpenMPDirectiveKind Kind, bool HasCancel, | |||
119 | StringRef HelperName) | |||
120 | : CGOpenMPRegionInfo(CS, ParallelOutlinedRegion, CodeGen, Kind, | |||
121 | HasCancel), | |||
122 | ThreadIDVar(ThreadIDVar), HelperName(HelperName) { | |||
123 | assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.")(static_cast <bool> (ThreadIDVar != nullptr && "No ThreadID in OpenMP region." ) ? void (0) : __assert_fail ("ThreadIDVar != nullptr && \"No ThreadID in OpenMP region.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 123, __extension__ __PRETTY_FUNCTION__)); | |||
124 | } | |||
125 | ||||
126 | /// Get a variable or parameter for storing global thread id | |||
127 | /// inside OpenMP construct. | |||
128 | const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; } | |||
129 | ||||
130 | /// Get the name of the capture helper. | |||
131 | StringRef getHelperName() const override { return HelperName; } | |||
132 | ||||
133 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
134 | return CGOpenMPRegionInfo::classof(Info) && | |||
135 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == | |||
136 | ParallelOutlinedRegion; | |||
137 | } | |||
138 | ||||
139 | private: | |||
140 | /// A variable or parameter storing global thread id for OpenMP | |||
141 | /// constructs. | |||
142 | const VarDecl *ThreadIDVar; | |||
143 | StringRef HelperName; | |||
144 | }; | |||
145 | ||||
146 | /// API for captured statement code generation in OpenMP constructs. | |||
147 | class CGOpenMPTaskOutlinedRegionInfo final : public CGOpenMPRegionInfo { | |||
148 | public: | |||
149 | class UntiedTaskActionTy final : public PrePostActionTy { | |||
150 | bool Untied; | |||
151 | const VarDecl *PartIDVar; | |||
152 | const RegionCodeGenTy UntiedCodeGen; | |||
153 | llvm::SwitchInst *UntiedSwitch = nullptr; | |||
154 | ||||
155 | public: | |||
156 | UntiedTaskActionTy(bool Tied, const VarDecl *PartIDVar, | |||
157 | const RegionCodeGenTy &UntiedCodeGen) | |||
158 | : Untied(!Tied), PartIDVar(PartIDVar), UntiedCodeGen(UntiedCodeGen) {} | |||
159 | void Enter(CodeGenFunction &CGF) override { | |||
160 | if (Untied) { | |||
161 | // Emit task switching point. | |||
162 | LValue PartIdLVal = CGF.EmitLoadOfPointerLValue( | |||
163 | CGF.GetAddrOfLocalVar(PartIDVar), | |||
164 | PartIDVar->getType()->castAs<PointerType>()); | |||
165 | llvm::Value *Res = | |||
166 | CGF.EmitLoadOfScalar(PartIdLVal, PartIDVar->getLocation()); | |||
167 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock(".untied.done."); | |||
168 | UntiedSwitch = CGF.Builder.CreateSwitch(Res, DoneBB); | |||
169 | CGF.EmitBlock(DoneBB); | |||
170 | CGF.EmitBranchThroughCleanup(CGF.ReturnBlock); | |||
171 | CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp.")); | |||
172 | UntiedSwitch->addCase(CGF.Builder.getInt32(0), | |||
173 | CGF.Builder.GetInsertBlock()); | |||
174 | emitUntiedSwitch(CGF); | |||
175 | } | |||
176 | } | |||
177 | void emitUntiedSwitch(CodeGenFunction &CGF) const { | |||
178 | if (Untied) { | |||
179 | LValue PartIdLVal = CGF.EmitLoadOfPointerLValue( | |||
180 | CGF.GetAddrOfLocalVar(PartIDVar), | |||
181 | PartIDVar->getType()->castAs<PointerType>()); | |||
182 | CGF.EmitStoreOfScalar(CGF.Builder.getInt32(UntiedSwitch->getNumCases()), | |||
183 | PartIdLVal); | |||
184 | UntiedCodeGen(CGF); | |||
185 | CodeGenFunction::JumpDest CurPoint = | |||
186 | CGF.getJumpDestInCurrentScope(".untied.next."); | |||
187 | CGF.EmitBranch(CGF.ReturnBlock.getBlock()); | |||
188 | CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp.")); | |||
189 | UntiedSwitch->addCase(CGF.Builder.getInt32(UntiedSwitch->getNumCases()), | |||
190 | CGF.Builder.GetInsertBlock()); | |||
191 | CGF.EmitBranchThroughCleanup(CurPoint); | |||
192 | CGF.EmitBlock(CurPoint.getBlock()); | |||
193 | } | |||
194 | } | |||
195 | unsigned getNumberOfParts() const { return UntiedSwitch->getNumCases(); } | |||
196 | }; | |||
197 | CGOpenMPTaskOutlinedRegionInfo(const CapturedStmt &CS, | |||
198 | const VarDecl *ThreadIDVar, | |||
199 | const RegionCodeGenTy &CodeGen, | |||
200 | OpenMPDirectiveKind Kind, bool HasCancel, | |||
201 | const UntiedTaskActionTy &Action) | |||
202 | : CGOpenMPRegionInfo(CS, TaskOutlinedRegion, CodeGen, Kind, HasCancel), | |||
203 | ThreadIDVar(ThreadIDVar), Action(Action) { | |||
204 | assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.")(static_cast <bool> (ThreadIDVar != nullptr && "No ThreadID in OpenMP region." ) ? void (0) : __assert_fail ("ThreadIDVar != nullptr && \"No ThreadID in OpenMP region.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 204, __extension__ __PRETTY_FUNCTION__)); | |||
205 | } | |||
206 | ||||
207 | /// Get a variable or parameter for storing global thread id | |||
208 | /// inside OpenMP construct. | |||
209 | const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; } | |||
210 | ||||
211 | /// Get an LValue for the current ThreadID variable. | |||
212 | LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override; | |||
213 | ||||
214 | /// Get the name of the capture helper. | |||
215 | StringRef getHelperName() const override { return ".omp_outlined."; } | |||
216 | ||||
217 | void emitUntiedSwitch(CodeGenFunction &CGF) override { | |||
218 | Action.emitUntiedSwitch(CGF); | |||
219 | } | |||
220 | ||||
221 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
222 | return CGOpenMPRegionInfo::classof(Info) && | |||
223 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == | |||
224 | TaskOutlinedRegion; | |||
225 | } | |||
226 | ||||
227 | private: | |||
228 | /// A variable or parameter storing global thread id for OpenMP | |||
229 | /// constructs. | |||
230 | const VarDecl *ThreadIDVar; | |||
231 | /// Action for emitting code for untied tasks. | |||
232 | const UntiedTaskActionTy &Action; | |||
233 | }; | |||
234 | ||||
235 | /// API for inlined captured statement code generation in OpenMP | |||
236 | /// constructs. | |||
237 | class CGOpenMPInlinedRegionInfo : public CGOpenMPRegionInfo { | |||
238 | public: | |||
239 | CGOpenMPInlinedRegionInfo(CodeGenFunction::CGCapturedStmtInfo *OldCSI, | |||
240 | const RegionCodeGenTy &CodeGen, | |||
241 | OpenMPDirectiveKind Kind, bool HasCancel) | |||
242 | : CGOpenMPRegionInfo(InlinedRegion, CodeGen, Kind, HasCancel), | |||
243 | OldCSI(OldCSI), | |||
244 | OuterRegionInfo(dyn_cast_or_null<CGOpenMPRegionInfo>(OldCSI)) {} | |||
245 | ||||
246 | // Retrieve the value of the context parameter. | |||
247 | llvm::Value *getContextValue() const override { | |||
248 | if (OuterRegionInfo) | |||
249 | return OuterRegionInfo->getContextValue(); | |||
250 | llvm_unreachable("No context value for inlined OpenMP region")::llvm::llvm_unreachable_internal("No context value for inlined OpenMP region" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 250); | |||
251 | } | |||
252 | ||||
253 | void setContextValue(llvm::Value *V) override { | |||
254 | if (OuterRegionInfo) { | |||
255 | OuterRegionInfo->setContextValue(V); | |||
256 | return; | |||
257 | } | |||
258 | llvm_unreachable("No context value for inlined OpenMP region")::llvm::llvm_unreachable_internal("No context value for inlined OpenMP region" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 258); | |||
259 | } | |||
260 | ||||
261 | /// Lookup the captured field decl for a variable. | |||
262 | const FieldDecl *lookup(const VarDecl *VD) const override { | |||
263 | if (OuterRegionInfo) | |||
264 | return OuterRegionInfo->lookup(VD); | |||
265 | // If there is no outer outlined region,no need to lookup in a list of | |||
266 | // captured variables, we can use the original one. | |||
267 | return nullptr; | |||
268 | } | |||
269 | ||||
270 | FieldDecl *getThisFieldDecl() const override { | |||
271 | if (OuterRegionInfo) | |||
272 | return OuterRegionInfo->getThisFieldDecl(); | |||
273 | return nullptr; | |||
274 | } | |||
275 | ||||
276 | /// Get a variable or parameter for storing global thread id | |||
277 | /// inside OpenMP construct. | |||
278 | const VarDecl *getThreadIDVariable() const override { | |||
279 | if (OuterRegionInfo) | |||
280 | return OuterRegionInfo->getThreadIDVariable(); | |||
281 | return nullptr; | |||
282 | } | |||
283 | ||||
284 | /// Get an LValue for the current ThreadID variable. | |||
285 | LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override { | |||
286 | if (OuterRegionInfo) | |||
287 | return OuterRegionInfo->getThreadIDVariableLValue(CGF); | |||
288 | llvm_unreachable("No LValue for inlined OpenMP construct")::llvm::llvm_unreachable_internal("No LValue for inlined OpenMP construct" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 288); | |||
289 | } | |||
290 | ||||
291 | /// Get the name of the capture helper. | |||
292 | StringRef getHelperName() const override { | |||
293 | if (auto *OuterRegionInfo = getOldCSI()) | |||
294 | return OuterRegionInfo->getHelperName(); | |||
295 | llvm_unreachable("No helper name for inlined OpenMP construct")::llvm::llvm_unreachable_internal("No helper name for inlined OpenMP construct" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 295); | |||
296 | } | |||
297 | ||||
298 | void emitUntiedSwitch(CodeGenFunction &CGF) override { | |||
299 | if (OuterRegionInfo) | |||
300 | OuterRegionInfo->emitUntiedSwitch(CGF); | |||
301 | } | |||
302 | ||||
303 | CodeGenFunction::CGCapturedStmtInfo *getOldCSI() const { return OldCSI; } | |||
304 | ||||
305 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
306 | return CGOpenMPRegionInfo::classof(Info) && | |||
307 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == InlinedRegion; | |||
308 | } | |||
309 | ||||
310 | ~CGOpenMPInlinedRegionInfo() override = default; | |||
311 | ||||
312 | private: | |||
313 | /// CodeGen info about outer OpenMP region. | |||
314 | CodeGenFunction::CGCapturedStmtInfo *OldCSI; | |||
315 | CGOpenMPRegionInfo *OuterRegionInfo; | |||
316 | }; | |||
317 | ||||
318 | /// API for captured statement code generation in OpenMP target | |||
319 | /// constructs. For this captures, implicit parameters are used instead of the | |||
320 | /// captured fields. The name of the target region has to be unique in a given | |||
321 | /// application so it is provided by the client, because only the client has | |||
322 | /// the information to generate that. | |||
323 | class CGOpenMPTargetRegionInfo final : public CGOpenMPRegionInfo { | |||
324 | public: | |||
325 | CGOpenMPTargetRegionInfo(const CapturedStmt &CS, | |||
326 | const RegionCodeGenTy &CodeGen, StringRef HelperName) | |||
327 | : CGOpenMPRegionInfo(CS, TargetRegion, CodeGen, OMPD_target, | |||
328 | /*HasCancel=*/false), | |||
329 | HelperName(HelperName) {} | |||
330 | ||||
331 | /// This is unused for target regions because each starts executing | |||
332 | /// with a single thread. | |||
333 | const VarDecl *getThreadIDVariable() const override { return nullptr; } | |||
334 | ||||
335 | /// Get the name of the capture helper. | |||
336 | StringRef getHelperName() const override { return HelperName; } | |||
337 | ||||
338 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
339 | return CGOpenMPRegionInfo::classof(Info) && | |||
340 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == TargetRegion; | |||
341 | } | |||
342 | ||||
343 | private: | |||
344 | StringRef HelperName; | |||
345 | }; | |||
346 | ||||
347 | static void EmptyCodeGen(CodeGenFunction &, PrePostActionTy &) { | |||
348 | llvm_unreachable("No codegen for expressions")::llvm::llvm_unreachable_internal("No codegen for expressions" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 348); | |||
349 | } | |||
350 | /// API for generation of expressions captured in a innermost OpenMP | |||
351 | /// region. | |||
352 | class CGOpenMPInnerExprInfo final : public CGOpenMPInlinedRegionInfo { | |||
353 | public: | |||
354 | CGOpenMPInnerExprInfo(CodeGenFunction &CGF, const CapturedStmt &CS) | |||
355 | : CGOpenMPInlinedRegionInfo(CGF.CapturedStmtInfo, EmptyCodeGen, | |||
356 | OMPD_unknown, | |||
357 | /*HasCancel=*/false), | |||
358 | PrivScope(CGF) { | |||
359 | // Make sure the globals captured in the provided statement are local by | |||
360 | // using the privatization logic. We assume the same variable is not | |||
361 | // captured more than once. | |||
362 | for (const auto &C : CS.captures()) { | |||
363 | if (!C.capturesVariable() && !C.capturesVariableByCopy()) | |||
364 | continue; | |||
365 | ||||
366 | const VarDecl *VD = C.getCapturedVar(); | |||
367 | if (VD->isLocalVarDeclOrParm()) | |||
368 | continue; | |||
369 | ||||
370 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD), | |||
371 | /*RefersToEnclosingVariableOrCapture=*/false, | |||
372 | VD->getType().getNonReferenceType(), VK_LValue, | |||
373 | C.getLocation()); | |||
374 | PrivScope.addPrivate(VD, CGF.EmitLValue(&DRE).getAddress(CGF)); | |||
375 | } | |||
376 | (void)PrivScope.Privatize(); | |||
377 | } | |||
378 | ||||
379 | /// Lookup the captured field decl for a variable. | |||
380 | const FieldDecl *lookup(const VarDecl *VD) const override { | |||
381 | if (const FieldDecl *FD = CGOpenMPInlinedRegionInfo::lookup(VD)) | |||
382 | return FD; | |||
383 | return nullptr; | |||
384 | } | |||
385 | ||||
386 | /// Emit the captured statement body. | |||
387 | void EmitBody(CodeGenFunction &CGF, const Stmt *S) override { | |||
388 | llvm_unreachable("No body for expressions")::llvm::llvm_unreachable_internal("No body for expressions", "clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 388); | |||
389 | } | |||
390 | ||||
391 | /// Get a variable or parameter for storing global thread id | |||
392 | /// inside OpenMP construct. | |||
393 | const VarDecl *getThreadIDVariable() const override { | |||
394 | llvm_unreachable("No thread id for expressions")::llvm::llvm_unreachable_internal("No thread id for expressions" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 394); | |||
395 | } | |||
396 | ||||
397 | /// Get the name of the capture helper. | |||
398 | StringRef getHelperName() const override { | |||
399 | llvm_unreachable("No helper name for expressions")::llvm::llvm_unreachable_internal("No helper name for expressions" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 399); | |||
400 | } | |||
401 | ||||
402 | static bool classof(const CGCapturedStmtInfo *Info) { return false; } | |||
403 | ||||
404 | private: | |||
405 | /// Private scope to capture global variables. | |||
406 | CodeGenFunction::OMPPrivateScope PrivScope; | |||
407 | }; | |||
408 | ||||
409 | /// RAII for emitting code of OpenMP constructs. | |||
410 | class InlinedOpenMPRegionRAII { | |||
411 | CodeGenFunction &CGF; | |||
412 | llvm::DenseMap<const ValueDecl *, FieldDecl *> LambdaCaptureFields; | |||
413 | FieldDecl *LambdaThisCaptureField = nullptr; | |||
414 | const CodeGen::CGBlockInfo *BlockInfo = nullptr; | |||
415 | bool NoInheritance = false; | |||
416 | ||||
417 | public: | |||
418 | /// Constructs region for combined constructs. | |||
419 | /// \param CodeGen Code generation sequence for combined directives. Includes | |||
420 | /// a list of functions used for code generation of implicitly inlined | |||
421 | /// regions. | |||
422 | InlinedOpenMPRegionRAII(CodeGenFunction &CGF, const RegionCodeGenTy &CodeGen, | |||
423 | OpenMPDirectiveKind Kind, bool HasCancel, | |||
424 | bool NoInheritance = true) | |||
425 | : CGF(CGF), NoInheritance(NoInheritance) { | |||
426 | // Start emission for the construct. | |||
427 | CGF.CapturedStmtInfo = new CGOpenMPInlinedRegionInfo( | |||
428 | CGF.CapturedStmtInfo, CodeGen, Kind, HasCancel); | |||
429 | if (NoInheritance) { | |||
430 | std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields); | |||
431 | LambdaThisCaptureField = CGF.LambdaThisCaptureField; | |||
432 | CGF.LambdaThisCaptureField = nullptr; | |||
433 | BlockInfo = CGF.BlockInfo; | |||
434 | CGF.BlockInfo = nullptr; | |||
435 | } | |||
436 | } | |||
437 | ||||
438 | ~InlinedOpenMPRegionRAII() { | |||
439 | // Restore original CapturedStmtInfo only if we're done with code emission. | |||
440 | auto *OldCSI = | |||
441 | cast<CGOpenMPInlinedRegionInfo>(CGF.CapturedStmtInfo)->getOldCSI(); | |||
442 | delete CGF.CapturedStmtInfo; | |||
443 | CGF.CapturedStmtInfo = OldCSI; | |||
444 | if (NoInheritance) { | |||
445 | std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields); | |||
446 | CGF.LambdaThisCaptureField = LambdaThisCaptureField; | |||
447 | CGF.BlockInfo = BlockInfo; | |||
448 | } | |||
449 | } | |||
450 | }; | |||
451 | ||||
452 | /// Values for bit flags used in the ident_t to describe the fields. | |||
453 | /// All enumeric elements are named and described in accordance with the code | |||
454 | /// from https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp.h | |||
455 | enum OpenMPLocationFlags : unsigned { | |||
456 | /// Use trampoline for internal microtask. | |||
457 | OMP_IDENT_IMD = 0x01, | |||
458 | /// Use c-style ident structure. | |||
459 | OMP_IDENT_KMPC = 0x02, | |||
460 | /// Atomic reduction option for kmpc_reduce. | |||
461 | OMP_ATOMIC_REDUCE = 0x10, | |||
462 | /// Explicit 'barrier' directive. | |||
463 | OMP_IDENT_BARRIER_EXPL = 0x20, | |||
464 | /// Implicit barrier in code. | |||
465 | OMP_IDENT_BARRIER_IMPL = 0x40, | |||
466 | /// Implicit barrier in 'for' directive. | |||
467 | OMP_IDENT_BARRIER_IMPL_FOR = 0x40, | |||
468 | /// Implicit barrier in 'sections' directive. | |||
469 | OMP_IDENT_BARRIER_IMPL_SECTIONS = 0xC0, | |||
470 | /// Implicit barrier in 'single' directive. | |||
471 | OMP_IDENT_BARRIER_IMPL_SINGLE = 0x140, | |||
472 | /// Call of __kmp_for_static_init for static loop. | |||
473 | OMP_IDENT_WORK_LOOP = 0x200, | |||
474 | /// Call of __kmp_for_static_init for sections. | |||
475 | OMP_IDENT_WORK_SECTIONS = 0x400, | |||
476 | /// Call of __kmp_for_static_init for distribute. | |||
477 | OMP_IDENT_WORK_DISTRIBUTE = 0x800, | |||
478 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_IDENT_WORK_DISTRIBUTE)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_IDENT_WORK_DISTRIBUTE | |||
479 | }; | |||
480 | ||||
481 | namespace { | |||
482 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail ::operator|; using ::llvm::BitmaskEnumDetail::operator&; using ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail ::operator|=; using ::llvm::BitmaskEnumDetail::operator&= ; using ::llvm::BitmaskEnumDetail::operator^=; | |||
483 | /// Values for bit flags for marking which requires clauses have been used. | |||
484 | enum OpenMPOffloadingRequiresDirFlags : int64_t { | |||
485 | /// flag undefined. | |||
486 | OMP_REQ_UNDEFINED = 0x000, | |||
487 | /// no requires clause present. | |||
488 | OMP_REQ_NONE = 0x001, | |||
489 | /// reverse_offload clause. | |||
490 | OMP_REQ_REVERSE_OFFLOAD = 0x002, | |||
491 | /// unified_address clause. | |||
492 | OMP_REQ_UNIFIED_ADDRESS = 0x004, | |||
493 | /// unified_shared_memory clause. | |||
494 | OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008, | |||
495 | /// dynamic_allocators clause. | |||
496 | OMP_REQ_DYNAMIC_ALLOCATORS = 0x010, | |||
497 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_REQ_DYNAMIC_ALLOCATORS)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_REQ_DYNAMIC_ALLOCATORS | |||
498 | }; | |||
499 | ||||
500 | enum OpenMPOffloadingReservedDeviceIDs { | |||
501 | /// Device ID if the device was not defined, runtime should get it | |||
502 | /// from environment variables in the spec. | |||
503 | OMP_DEVICEID_UNDEF = -1, | |||
504 | }; | |||
505 | } // anonymous namespace | |||
506 | ||||
507 | /// Describes ident structure that describes a source location. | |||
508 | /// All descriptions are taken from | |||
509 | /// https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp.h | |||
510 | /// Original structure: | |||
511 | /// typedef struct ident { | |||
512 | /// kmp_int32 reserved_1; /**< might be used in Fortran; | |||
513 | /// see above */ | |||
514 | /// kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; | |||
515 | /// KMP_IDENT_KMPC identifies this union | |||
516 | /// member */ | |||
517 | /// kmp_int32 reserved_2; /**< not really used in Fortran any more; | |||
518 | /// see above */ | |||
519 | ///#if USE_ITT_BUILD | |||
520 | /// /* but currently used for storing | |||
521 | /// region-specific ITT */ | |||
522 | /// /* contextual information. */ | |||
523 | ///#endif /* USE_ITT_BUILD */ | |||
524 | /// kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for | |||
525 | /// C++ */ | |||
526 | /// char const *psource; /**< String describing the source location. | |||
527 | /// The string is composed of semi-colon separated | |||
528 | // fields which describe the source file, | |||
529 | /// the function and a pair of line numbers that | |||
530 | /// delimit the construct. | |||
531 | /// */ | |||
532 | /// } ident_t; | |||
533 | enum IdentFieldIndex { | |||
534 | /// might be used in Fortran | |||
535 | IdentField_Reserved_1, | |||
536 | /// OMP_IDENT_xxx flags; OMP_IDENT_KMPC identifies this union member. | |||
537 | IdentField_Flags, | |||
538 | /// Not really used in Fortran any more | |||
539 | IdentField_Reserved_2, | |||
540 | /// Source[4] in Fortran, do not use for C++ | |||
541 | IdentField_Reserved_3, | |||
542 | /// String describing the source location. The string is composed of | |||
543 | /// semi-colon separated fields which describe the source file, the function | |||
544 | /// and a pair of line numbers that delimit the construct. | |||
545 | IdentField_PSource | |||
546 | }; | |||
547 | ||||
548 | /// Schedule types for 'omp for' loops (these enumerators are taken from | |||
549 | /// the enum sched_type in kmp.h). | |||
550 | enum OpenMPSchedType { | |||
551 | /// Lower bound for default (unordered) versions. | |||
552 | OMP_sch_lower = 32, | |||
553 | OMP_sch_static_chunked = 33, | |||
554 | OMP_sch_static = 34, | |||
555 | OMP_sch_dynamic_chunked = 35, | |||
556 | OMP_sch_guided_chunked = 36, | |||
557 | OMP_sch_runtime = 37, | |||
558 | OMP_sch_auto = 38, | |||
559 | /// static with chunk adjustment (e.g., simd) | |||
560 | OMP_sch_static_balanced_chunked = 45, | |||
561 | /// Lower bound for 'ordered' versions. | |||
562 | OMP_ord_lower = 64, | |||
563 | OMP_ord_static_chunked = 65, | |||
564 | OMP_ord_static = 66, | |||
565 | OMP_ord_dynamic_chunked = 67, | |||
566 | OMP_ord_guided_chunked = 68, | |||
567 | OMP_ord_runtime = 69, | |||
568 | OMP_ord_auto = 70, | |||
569 | OMP_sch_default = OMP_sch_static, | |||
570 | /// dist_schedule types | |||
571 | OMP_dist_sch_static_chunked = 91, | |||
572 | OMP_dist_sch_static = 92, | |||
573 | /// Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. | |||
574 | /// Set if the monotonic schedule modifier was present. | |||
575 | OMP_sch_modifier_monotonic = (1 << 29), | |||
576 | /// Set if the nonmonotonic schedule modifier was present. | |||
577 | OMP_sch_modifier_nonmonotonic = (1 << 30), | |||
578 | }; | |||
579 | ||||
580 | /// A basic class for pre|post-action for advanced codegen sequence for OpenMP | |||
581 | /// region. | |||
582 | class CleanupTy final : public EHScopeStack::Cleanup { | |||
583 | PrePostActionTy *Action; | |||
584 | ||||
585 | public: | |||
586 | explicit CleanupTy(PrePostActionTy *Action) : Action(Action) {} | |||
587 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | |||
588 | if (!CGF.HaveInsertPoint()) | |||
589 | return; | |||
590 | Action->Exit(CGF); | |||
591 | } | |||
592 | }; | |||
593 | ||||
594 | } // anonymous namespace | |||
595 | ||||
596 | void RegionCodeGenTy::operator()(CodeGenFunction &CGF) const { | |||
597 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
598 | if (PrePostAction) { | |||
599 | CGF.EHStack.pushCleanup<CleanupTy>(NormalAndEHCleanup, PrePostAction); | |||
600 | Callback(CodeGen, CGF, *PrePostAction); | |||
601 | } else { | |||
602 | PrePostActionTy Action; | |||
603 | Callback(CodeGen, CGF, Action); | |||
604 | } | |||
605 | } | |||
606 | ||||
607 | /// Check if the combiner is a call to UDR combiner and if it is so return the | |||
608 | /// UDR decl used for reduction. | |||
609 | static const OMPDeclareReductionDecl * | |||
610 | getReductionInit(const Expr *ReductionOp) { | |||
611 | if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) | |||
612 | if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee())) | |||
613 | if (const auto *DRE = | |||
614 | dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts())) | |||
615 | if (const auto *DRD = dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) | |||
616 | return DRD; | |||
617 | return nullptr; | |||
618 | } | |||
619 | ||||
620 | static void emitInitWithReductionInitializer(CodeGenFunction &CGF, | |||
621 | const OMPDeclareReductionDecl *DRD, | |||
622 | const Expr *InitOp, | |||
623 | Address Private, Address Original, | |||
624 | QualType Ty) { | |||
625 | if (DRD->getInitializer()) { | |||
626 | std::pair<llvm::Function *, llvm::Function *> Reduction = | |||
627 | CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD); | |||
628 | const auto *CE = cast<CallExpr>(InitOp); | |||
629 | const auto *OVE = cast<OpaqueValueExpr>(CE->getCallee()); | |||
630 | const Expr *LHS = CE->getArg(/*Arg=*/0)->IgnoreParenImpCasts(); | |||
631 | const Expr *RHS = CE->getArg(/*Arg=*/1)->IgnoreParenImpCasts(); | |||
632 | const auto *LHSDRE = | |||
633 | cast<DeclRefExpr>(cast<UnaryOperator>(LHS)->getSubExpr()); | |||
634 | const auto *RHSDRE = | |||
635 | cast<DeclRefExpr>(cast<UnaryOperator>(RHS)->getSubExpr()); | |||
636 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | |||
637 | PrivateScope.addPrivate(cast<VarDecl>(LHSDRE->getDecl()), Private); | |||
638 | PrivateScope.addPrivate(cast<VarDecl>(RHSDRE->getDecl()), Original); | |||
639 | (void)PrivateScope.Privatize(); | |||
640 | RValue Func = RValue::get(Reduction.second); | |||
641 | CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func); | |||
642 | CGF.EmitIgnoredExpr(InitOp); | |||
643 | } else { | |||
644 | llvm::Constant *Init = CGF.CGM.EmitNullConstant(Ty); | |||
645 | std::string Name = CGF.CGM.getOpenMPRuntime().getName({"init"}); | |||
646 | auto *GV = new llvm::GlobalVariable( | |||
647 | CGF.CGM.getModule(), Init->getType(), /*isConstant=*/true, | |||
648 | llvm::GlobalValue::PrivateLinkage, Init, Name); | |||
649 | LValue LV = CGF.MakeNaturalAlignAddrLValue(GV, Ty); | |||
650 | RValue InitRVal; | |||
651 | switch (CGF.getEvaluationKind(Ty)) { | |||
652 | case TEK_Scalar: | |||
653 | InitRVal = CGF.EmitLoadOfLValue(LV, DRD->getLocation()); | |||
654 | break; | |||
655 | case TEK_Complex: | |||
656 | InitRVal = | |||
657 | RValue::getComplex(CGF.EmitLoadOfComplex(LV, DRD->getLocation())); | |||
658 | break; | |||
659 | case TEK_Aggregate: { | |||
660 | OpaqueValueExpr OVE(DRD->getLocation(), Ty, VK_LValue); | |||
661 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, LV); | |||
662 | CGF.EmitAnyExprToMem(&OVE, Private, Ty.getQualifiers(), | |||
663 | /*IsInitializer=*/false); | |||
664 | return; | |||
665 | } | |||
666 | } | |||
667 | OpaqueValueExpr OVE(DRD->getLocation(), Ty, VK_PRValue); | |||
668 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, InitRVal); | |||
669 | CGF.EmitAnyExprToMem(&OVE, Private, Ty.getQualifiers(), | |||
670 | /*IsInitializer=*/false); | |||
671 | } | |||
672 | } | |||
673 | ||||
674 | /// Emit initialization of arrays of complex types. | |||
675 | /// \param DestAddr Address of the array. | |||
676 | /// \param Type Type of array. | |||
677 | /// \param Init Initial expression of array. | |||
678 | /// \param SrcAddr Address of the original array. | |||
679 | static void EmitOMPAggregateInit(CodeGenFunction &CGF, Address DestAddr, | |||
680 | QualType Type, bool EmitDeclareReductionInit, | |||
681 | const Expr *Init, | |||
682 | const OMPDeclareReductionDecl *DRD, | |||
683 | Address SrcAddr = Address::invalid()) { | |||
684 | // Perform element-by-element initialization. | |||
685 | QualType ElementTy; | |||
686 | ||||
687 | // Drill down to the base element type on both arrays. | |||
688 | const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe(); | |||
689 | llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, DestAddr); | |||
690 | if (DRD) | |||
691 | SrcAddr = | |||
692 | CGF.Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType()); | |||
693 | ||||
694 | llvm::Value *SrcBegin = nullptr; | |||
695 | if (DRD) | |||
696 | SrcBegin = SrcAddr.getPointer(); | |||
697 | llvm::Value *DestBegin = DestAddr.getPointer(); | |||
698 | // Cast from pointer to array type to pointer to single element. | |||
699 | llvm::Value *DestEnd = | |||
700 | CGF.Builder.CreateGEP(DestAddr.getElementType(), DestBegin, NumElements); | |||
701 | // The basic structure here is a while-do loop. | |||
702 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arrayinit.body"); | |||
703 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arrayinit.done"); | |||
704 | llvm::Value *IsEmpty = | |||
705 | CGF.Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arrayinit.isempty"); | |||
706 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | |||
707 | ||||
708 | // Enter the loop body, making that address the current address. | |||
709 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | |||
710 | CGF.EmitBlock(BodyBB); | |||
711 | ||||
712 | CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy); | |||
713 | ||||
714 | llvm::PHINode *SrcElementPHI = nullptr; | |||
715 | Address SrcElementCurrent = Address::invalid(); | |||
716 | if (DRD) { | |||
717 | SrcElementPHI = CGF.Builder.CreatePHI(SrcBegin->getType(), 2, | |||
718 | "omp.arraycpy.srcElementPast"); | |||
719 | SrcElementPHI->addIncoming(SrcBegin, EntryBB); | |||
720 | SrcElementCurrent = | |||
721 | Address(SrcElementPHI, SrcAddr.getElementType(), | |||
722 | SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
723 | } | |||
724 | llvm::PHINode *DestElementPHI = CGF.Builder.CreatePHI( | |||
725 | DestBegin->getType(), 2, "omp.arraycpy.destElementPast"); | |||
726 | DestElementPHI->addIncoming(DestBegin, EntryBB); | |||
727 | Address DestElementCurrent = | |||
728 | Address(DestElementPHI, DestAddr.getElementType(), | |||
729 | DestAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
730 | ||||
731 | // Emit copy. | |||
732 | { | |||
733 | CodeGenFunction::RunCleanupsScope InitScope(CGF); | |||
734 | if (EmitDeclareReductionInit) { | |||
735 | emitInitWithReductionInitializer(CGF, DRD, Init, DestElementCurrent, | |||
736 | SrcElementCurrent, ElementTy); | |||
737 | } else | |||
738 | CGF.EmitAnyExprToMem(Init, DestElementCurrent, ElementTy.getQualifiers(), | |||
739 | /*IsInitializer=*/false); | |||
740 | } | |||
741 | ||||
742 | if (DRD) { | |||
743 | // Shift the address forward by one element. | |||
744 | llvm::Value *SrcElementNext = CGF.Builder.CreateConstGEP1_32( | |||
745 | SrcAddr.getElementType(), SrcElementPHI, /*Idx0=*/1, | |||
746 | "omp.arraycpy.dest.element"); | |||
747 | SrcElementPHI->addIncoming(SrcElementNext, CGF.Builder.GetInsertBlock()); | |||
748 | } | |||
749 | ||||
750 | // Shift the address forward by one element. | |||
751 | llvm::Value *DestElementNext = CGF.Builder.CreateConstGEP1_32( | |||
752 | DestAddr.getElementType(), DestElementPHI, /*Idx0=*/1, | |||
753 | "omp.arraycpy.dest.element"); | |||
754 | // Check whether we've reached the end. | |||
755 | llvm::Value *Done = | |||
756 | CGF.Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done"); | |||
757 | CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB); | |||
758 | DestElementPHI->addIncoming(DestElementNext, CGF.Builder.GetInsertBlock()); | |||
759 | ||||
760 | // Done. | |||
761 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | |||
762 | } | |||
763 | ||||
764 | LValue ReductionCodeGen::emitSharedLValue(CodeGenFunction &CGF, const Expr *E) { | |||
765 | return CGF.EmitOMPSharedLValue(E); | |||
766 | } | |||
767 | ||||
768 | LValue ReductionCodeGen::emitSharedLValueUB(CodeGenFunction &CGF, | |||
769 | const Expr *E) { | |||
770 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(E)) | |||
771 | return CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false); | |||
772 | return LValue(); | |||
773 | } | |||
774 | ||||
775 | void ReductionCodeGen::emitAggregateInitialization( | |||
776 | CodeGenFunction &CGF, unsigned N, Address PrivateAddr, Address SharedAddr, | |||
777 | const OMPDeclareReductionDecl *DRD) { | |||
778 | // Emit VarDecl with copy init for arrays. | |||
779 | // Get the address of the original variable captured in current | |||
780 | // captured region. | |||
781 | const auto *PrivateVD = | |||
782 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
783 | bool EmitDeclareReductionInit = | |||
784 | DRD && (DRD->getInitializer() || !PrivateVD->hasInit()); | |||
785 | EmitOMPAggregateInit(CGF, PrivateAddr, PrivateVD->getType(), | |||
786 | EmitDeclareReductionInit, | |||
787 | EmitDeclareReductionInit ? ClausesData[N].ReductionOp | |||
788 | : PrivateVD->getInit(), | |||
789 | DRD, SharedAddr); | |||
790 | } | |||
791 | ||||
792 | ReductionCodeGen::ReductionCodeGen(ArrayRef<const Expr *> Shareds, | |||
793 | ArrayRef<const Expr *> Origs, | |||
794 | ArrayRef<const Expr *> Privates, | |||
795 | ArrayRef<const Expr *> ReductionOps) { | |||
796 | ClausesData.reserve(Shareds.size()); | |||
797 | SharedAddresses.reserve(Shareds.size()); | |||
798 | Sizes.reserve(Shareds.size()); | |||
799 | BaseDecls.reserve(Shareds.size()); | |||
800 | const auto *IOrig = Origs.begin(); | |||
801 | const auto *IPriv = Privates.begin(); | |||
802 | const auto *IRed = ReductionOps.begin(); | |||
803 | for (const Expr *Ref : Shareds) { | |||
804 | ClausesData.emplace_back(Ref, *IOrig, *IPriv, *IRed); | |||
805 | std::advance(IOrig, 1); | |||
806 | std::advance(IPriv, 1); | |||
807 | std::advance(IRed, 1); | |||
808 | } | |||
809 | } | |||
810 | ||||
811 | void ReductionCodeGen::emitSharedOrigLValue(CodeGenFunction &CGF, unsigned N) { | |||
812 | assert(SharedAddresses.size() == N && OrigAddresses.size() == N &&(static_cast <bool> (SharedAddresses.size() == N && OrigAddresses.size() == N && "Number of generated lvalues must be exactly N." ) ? void (0) : __assert_fail ("SharedAddresses.size() == N && OrigAddresses.size() == N && \"Number of generated lvalues must be exactly N.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 813, __extension__ __PRETTY_FUNCTION__)) | |||
813 | "Number of generated lvalues must be exactly N.")(static_cast <bool> (SharedAddresses.size() == N && OrigAddresses.size() == N && "Number of generated lvalues must be exactly N." ) ? void (0) : __assert_fail ("SharedAddresses.size() == N && OrigAddresses.size() == N && \"Number of generated lvalues must be exactly N.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 813, __extension__ __PRETTY_FUNCTION__)); | |||
814 | LValue First = emitSharedLValue(CGF, ClausesData[N].Shared); | |||
815 | LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Shared); | |||
816 | SharedAddresses.emplace_back(First, Second); | |||
817 | if (ClausesData[N].Shared == ClausesData[N].Ref) { | |||
818 | OrigAddresses.emplace_back(First, Second); | |||
819 | } else { | |||
820 | LValue First = emitSharedLValue(CGF, ClausesData[N].Ref); | |||
821 | LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Ref); | |||
822 | OrigAddresses.emplace_back(First, Second); | |||
823 | } | |||
824 | } | |||
825 | ||||
826 | void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N) { | |||
827 | QualType PrivateType = getPrivateType(N); | |||
828 | bool AsArraySection = isa<OMPArraySectionExpr>(ClausesData[N].Ref); | |||
829 | if (!PrivateType->isVariablyModifiedType()) { | |||
830 | Sizes.emplace_back( | |||
831 | CGF.getTypeSize(OrigAddresses[N].first.getType().getNonReferenceType()), | |||
832 | nullptr); | |||
833 | return; | |||
834 | } | |||
835 | llvm::Value *Size; | |||
836 | llvm::Value *SizeInChars; | |||
837 | auto *ElemType = OrigAddresses[N].first.getAddress(CGF).getElementType(); | |||
838 | auto *ElemSizeOf = llvm::ConstantExpr::getSizeOf(ElemType); | |||
839 | if (AsArraySection) { | |||
840 | Size = CGF.Builder.CreatePtrDiff(ElemType, | |||
841 | OrigAddresses[N].second.getPointer(CGF), | |||
842 | OrigAddresses[N].first.getPointer(CGF)); | |||
843 | Size = CGF.Builder.CreateNUWAdd( | |||
844 | Size, llvm::ConstantInt::get(Size->getType(), /*V=*/1)); | |||
845 | SizeInChars = CGF.Builder.CreateNUWMul(Size, ElemSizeOf); | |||
846 | } else { | |||
847 | SizeInChars = | |||
848 | CGF.getTypeSize(OrigAddresses[N].first.getType().getNonReferenceType()); | |||
849 | Size = CGF.Builder.CreateExactUDiv(SizeInChars, ElemSizeOf); | |||
850 | } | |||
851 | Sizes.emplace_back(SizeInChars, Size); | |||
852 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | |||
853 | CGF, | |||
854 | cast<OpaqueValueExpr>( | |||
855 | CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()), | |||
856 | RValue::get(Size)); | |||
857 | CGF.EmitVariablyModifiedType(PrivateType); | |||
858 | } | |||
859 | ||||
860 | void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N, | |||
861 | llvm::Value *Size) { | |||
862 | QualType PrivateType = getPrivateType(N); | |||
863 | if (!PrivateType->isVariablyModifiedType()) { | |||
864 | assert(!Size && !Sizes[N].second &&(static_cast <bool> (!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? void (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 866, __extension__ __PRETTY_FUNCTION__)) | |||
865 | "Size should be nullptr for non-variably modified reduction "(static_cast <bool> (!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? void (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 866, __extension__ __PRETTY_FUNCTION__)) | |||
866 | "items.")(static_cast <bool> (!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? void (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 866, __extension__ __PRETTY_FUNCTION__)); | |||
867 | return; | |||
868 | } | |||
869 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | |||
870 | CGF, | |||
871 | cast<OpaqueValueExpr>( | |||
872 | CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()), | |||
873 | RValue::get(Size)); | |||
874 | CGF.EmitVariablyModifiedType(PrivateType); | |||
875 | } | |||
876 | ||||
877 | void ReductionCodeGen::emitInitialization( | |||
878 | CodeGenFunction &CGF, unsigned N, Address PrivateAddr, Address SharedAddr, | |||
879 | llvm::function_ref<bool(CodeGenFunction &)> DefaultInit) { | |||
880 | assert(SharedAddresses.size() > N && "No variable was generated")(static_cast <bool> (SharedAddresses.size() > N && "No variable was generated") ? void (0) : __assert_fail ("SharedAddresses.size() > N && \"No variable was generated\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 880, __extension__ __PRETTY_FUNCTION__)); | |||
881 | const auto *PrivateVD = | |||
882 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
883 | const OMPDeclareReductionDecl *DRD = | |||
884 | getReductionInit(ClausesData[N].ReductionOp); | |||
885 | if (CGF.getContext().getAsArrayType(PrivateVD->getType())) { | |||
886 | if (DRD && DRD->getInitializer()) | |||
887 | (void)DefaultInit(CGF); | |||
888 | emitAggregateInitialization(CGF, N, PrivateAddr, SharedAddr, DRD); | |||
889 | } else if (DRD && (DRD->getInitializer() || !PrivateVD->hasInit())) { | |||
890 | (void)DefaultInit(CGF); | |||
891 | QualType SharedType = SharedAddresses[N].first.getType(); | |||
892 | emitInitWithReductionInitializer(CGF, DRD, ClausesData[N].ReductionOp, | |||
893 | PrivateAddr, SharedAddr, SharedType); | |||
894 | } else if (!DefaultInit(CGF) && PrivateVD->hasInit() && | |||
895 | !CGF.isTrivialInitializer(PrivateVD->getInit())) { | |||
896 | CGF.EmitAnyExprToMem(PrivateVD->getInit(), PrivateAddr, | |||
897 | PrivateVD->getType().getQualifiers(), | |||
898 | /*IsInitializer=*/false); | |||
899 | } | |||
900 | } | |||
901 | ||||
902 | bool ReductionCodeGen::needCleanups(unsigned N) { | |||
903 | QualType PrivateType = getPrivateType(N); | |||
904 | QualType::DestructionKind DTorKind = PrivateType.isDestructedType(); | |||
905 | return DTorKind != QualType::DK_none; | |||
906 | } | |||
907 | ||||
908 | void ReductionCodeGen::emitCleanups(CodeGenFunction &CGF, unsigned N, | |||
909 | Address PrivateAddr) { | |||
910 | QualType PrivateType = getPrivateType(N); | |||
911 | QualType::DestructionKind DTorKind = PrivateType.isDestructedType(); | |||
912 | if (needCleanups(N)) { | |||
913 | PrivateAddr = CGF.Builder.CreateElementBitCast( | |||
914 | PrivateAddr, CGF.ConvertTypeForMem(PrivateType)); | |||
915 | CGF.pushDestroy(DTorKind, PrivateAddr, PrivateType); | |||
916 | } | |||
917 | } | |||
918 | ||||
919 | static LValue loadToBegin(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy, | |||
920 | LValue BaseLV) { | |||
921 | BaseTy = BaseTy.getNonReferenceType(); | |||
922 | while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) && | |||
923 | !CGF.getContext().hasSameType(BaseTy, ElTy)) { | |||
924 | if (const auto *PtrTy = BaseTy->getAs<PointerType>()) { | |||
925 | BaseLV = CGF.EmitLoadOfPointerLValue(BaseLV.getAddress(CGF), PtrTy); | |||
926 | } else { | |||
927 | LValue RefLVal = CGF.MakeAddrLValue(BaseLV.getAddress(CGF), BaseTy); | |||
928 | BaseLV = CGF.EmitLoadOfReferenceLValue(RefLVal); | |||
929 | } | |||
930 | BaseTy = BaseTy->getPointeeType(); | |||
931 | } | |||
932 | return CGF.MakeAddrLValue( | |||
933 | CGF.Builder.CreateElementBitCast(BaseLV.getAddress(CGF), | |||
934 | CGF.ConvertTypeForMem(ElTy)), | |||
935 | BaseLV.getType(), BaseLV.getBaseInfo(), | |||
936 | CGF.CGM.getTBAAInfoForSubobject(BaseLV, BaseLV.getType())); | |||
937 | } | |||
938 | ||||
939 | static Address castToBase(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy, | |||
940 | Address OriginalBaseAddress, llvm::Value *Addr) { | |||
941 | Address Tmp = Address::invalid(); | |||
942 | Address TopTmp = Address::invalid(); | |||
943 | Address MostTopTmp = Address::invalid(); | |||
944 | BaseTy = BaseTy.getNonReferenceType(); | |||
945 | while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) && | |||
946 | !CGF.getContext().hasSameType(BaseTy, ElTy)) { | |||
947 | Tmp = CGF.CreateMemTemp(BaseTy); | |||
948 | if (TopTmp.isValid()) | |||
949 | CGF.Builder.CreateStore(Tmp.getPointer(), TopTmp); | |||
950 | else | |||
951 | MostTopTmp = Tmp; | |||
952 | TopTmp = Tmp; | |||
953 | BaseTy = BaseTy->getPointeeType(); | |||
954 | } | |||
955 | ||||
956 | if (Tmp.isValid()) { | |||
957 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
958 | Addr, Tmp.getElementType()); | |||
959 | CGF.Builder.CreateStore(Addr, Tmp); | |||
960 | return MostTopTmp; | |||
961 | } | |||
962 | ||||
963 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
964 | Addr, OriginalBaseAddress.getType()); | |||
965 | return OriginalBaseAddress.withPointer(Addr); | |||
966 | } | |||
967 | ||||
968 | static const VarDecl *getBaseDecl(const Expr *Ref, const DeclRefExpr *&DE) { | |||
969 | const VarDecl *OrigVD = nullptr; | |||
970 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(Ref)) { | |||
971 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); | |||
972 | while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) | |||
973 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); | |||
974 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | |||
975 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | |||
976 | DE = cast<DeclRefExpr>(Base); | |||
977 | OrigVD = cast<VarDecl>(DE->getDecl()); | |||
978 | } else if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Ref)) { | |||
979 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); | |||
980 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | |||
981 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | |||
982 | DE = cast<DeclRefExpr>(Base); | |||
983 | OrigVD = cast<VarDecl>(DE->getDecl()); | |||
984 | } | |||
985 | return OrigVD; | |||
986 | } | |||
987 | ||||
988 | Address ReductionCodeGen::adjustPrivateAddress(CodeGenFunction &CGF, unsigned N, | |||
989 | Address PrivateAddr) { | |||
990 | const DeclRefExpr *DE; | |||
991 | if (const VarDecl *OrigVD = ::getBaseDecl(ClausesData[N].Ref, DE)) { | |||
992 | BaseDecls.emplace_back(OrigVD); | |||
993 | LValue OriginalBaseLValue = CGF.EmitLValue(DE); | |||
994 | LValue BaseLValue = | |||
995 | loadToBegin(CGF, OrigVD->getType(), SharedAddresses[N].first.getType(), | |||
996 | OriginalBaseLValue); | |||
997 | Address SharedAddr = SharedAddresses[N].first.getAddress(CGF); | |||
998 | llvm::Value *Adjustment = CGF.Builder.CreatePtrDiff( | |||
999 | SharedAddr.getElementType(), BaseLValue.getPointer(CGF), | |||
1000 | SharedAddr.getPointer()); | |||
1001 | llvm::Value *PrivatePointer = | |||
1002 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
1003 | PrivateAddr.getPointer(), SharedAddr.getType()); | |||
1004 | llvm::Value *Ptr = CGF.Builder.CreateGEP( | |||
1005 | SharedAddr.getElementType(), PrivatePointer, Adjustment); | |||
1006 | return castToBase(CGF, OrigVD->getType(), | |||
1007 | SharedAddresses[N].first.getType(), | |||
1008 | OriginalBaseLValue.getAddress(CGF), Ptr); | |||
1009 | } | |||
1010 | BaseDecls.emplace_back( | |||
1011 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Ref)->getDecl())); | |||
1012 | return PrivateAddr; | |||
1013 | } | |||
1014 | ||||
1015 | bool ReductionCodeGen::usesReductionInitializer(unsigned N) const { | |||
1016 | const OMPDeclareReductionDecl *DRD = | |||
1017 | getReductionInit(ClausesData[N].ReductionOp); | |||
1018 | return DRD && DRD->getInitializer(); | |||
1019 | } | |||
1020 | ||||
1021 | LValue CGOpenMPRegionInfo::getThreadIDVariableLValue(CodeGenFunction &CGF) { | |||
1022 | return CGF.EmitLoadOfPointerLValue( | |||
1023 | CGF.GetAddrOfLocalVar(getThreadIDVariable()), | |||
1024 | getThreadIDVariable()->getType()->castAs<PointerType>()); | |||
1025 | } | |||
1026 | ||||
1027 | void CGOpenMPRegionInfo::EmitBody(CodeGenFunction &CGF, const Stmt *S) { | |||
1028 | if (!CGF.HaveInsertPoint()) | |||
1029 | return; | |||
1030 | // 1.2.2 OpenMP Language Terminology | |||
1031 | // Structured block - An executable statement with a single entry at the | |||
1032 | // top and a single exit at the bottom. | |||
1033 | // The point of exit cannot be a branch out of the structured block. | |||
1034 | // longjmp() and throw() must not violate the entry/exit criteria. | |||
1035 | CGF.EHStack.pushTerminate(); | |||
1036 | if (S) | |||
1037 | CGF.incrementProfileCounter(S); | |||
1038 | CodeGen(CGF); | |||
1039 | CGF.EHStack.popTerminate(); | |||
1040 | } | |||
1041 | ||||
1042 | LValue CGOpenMPTaskOutlinedRegionInfo::getThreadIDVariableLValue( | |||
1043 | CodeGenFunction &CGF) { | |||
1044 | return CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(getThreadIDVariable()), | |||
1045 | getThreadIDVariable()->getType(), | |||
1046 | AlignmentSource::Decl); | |||
1047 | } | |||
1048 | ||||
1049 | static FieldDecl *addFieldToRecordDecl(ASTContext &C, DeclContext *DC, | |||
1050 | QualType FieldTy) { | |||
1051 | auto *Field = FieldDecl::Create( | |||
1052 | C, DC, SourceLocation(), SourceLocation(), /*Id=*/nullptr, FieldTy, | |||
1053 | C.getTrivialTypeSourceInfo(FieldTy, SourceLocation()), | |||
1054 | /*BW=*/nullptr, /*Mutable=*/false, /*InitStyle=*/ICIS_NoInit); | |||
1055 | Field->setAccess(AS_public); | |||
1056 | DC->addDecl(Field); | |||
1057 | return Field; | |||
1058 | } | |||
1059 | ||||
1060 | CGOpenMPRuntime::CGOpenMPRuntime(CodeGenModule &CGM, StringRef FirstSeparator, | |||
1061 | StringRef Separator) | |||
1062 | : CGM(CGM), FirstSeparator(FirstSeparator), Separator(Separator), | |||
1063 | OMPBuilder(CGM.getModule()), OffloadEntriesInfoManager() { | |||
1064 | KmpCriticalNameTy = llvm::ArrayType::get(CGM.Int32Ty, /*NumElements*/ 8); | |||
1065 | ||||
1066 | // Initialize Types used in OpenMPIRBuilder from OMPKinds.def | |||
1067 | OMPBuilder.initialize(); | |||
1068 | loadOffloadInfoMetadata(); | |||
1069 | } | |||
1070 | ||||
1071 | void CGOpenMPRuntime::clear() { | |||
1072 | InternalVars.clear(); | |||
1073 | // Clean non-target variable declarations possibly used only in debug info. | |||
1074 | for (const auto &Data : EmittedNonTargetVariables) { | |||
1075 | if (!Data.getValue().pointsToAliveValue()) | |||
1076 | continue; | |||
1077 | auto *GV = dyn_cast<llvm::GlobalVariable>(Data.getValue()); | |||
1078 | if (!GV) | |||
1079 | continue; | |||
1080 | if (!GV->isDeclaration() || GV->getNumUses() > 0) | |||
1081 | continue; | |||
1082 | GV->eraseFromParent(); | |||
1083 | } | |||
1084 | } | |||
1085 | ||||
1086 | std::string CGOpenMPRuntime::getName(ArrayRef<StringRef> Parts) const { | |||
1087 | SmallString<128> Buffer; | |||
1088 | llvm::raw_svector_ostream OS(Buffer); | |||
1089 | StringRef Sep = FirstSeparator; | |||
1090 | for (StringRef Part : Parts) { | |||
1091 | OS << Sep << Part; | |||
1092 | Sep = Separator; | |||
1093 | } | |||
1094 | return std::string(OS.str()); | |||
1095 | } | |||
1096 | ||||
1097 | static llvm::Function * | |||
1098 | emitCombinerOrInitializer(CodeGenModule &CGM, QualType Ty, | |||
1099 | const Expr *CombinerInitializer, const VarDecl *In, | |||
1100 | const VarDecl *Out, bool IsCombiner) { | |||
1101 | // void .omp_combiner.(Ty *in, Ty *out); | |||
1102 | ASTContext &C = CGM.getContext(); | |||
1103 | QualType PtrTy = C.getPointerType(Ty).withRestrict(); | |||
1104 | FunctionArgList Args; | |||
1105 | ImplicitParamDecl OmpOutParm(C, /*DC=*/nullptr, Out->getLocation(), | |||
1106 | /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other); | |||
1107 | ImplicitParamDecl OmpInParm(C, /*DC=*/nullptr, In->getLocation(), | |||
1108 | /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other); | |||
1109 | Args.push_back(&OmpOutParm); | |||
1110 | Args.push_back(&OmpInParm); | |||
1111 | const CGFunctionInfo &FnInfo = | |||
1112 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
1113 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
1114 | std::string Name = CGM.getOpenMPRuntime().getName( | |||
1115 | {IsCombiner ? "omp_combiner" : "omp_initializer", ""}); | |||
1116 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
1117 | Name, &CGM.getModule()); | |||
1118 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
1119 | if (CGM.getLangOpts().Optimize) { | |||
1120 | Fn->removeFnAttr(llvm::Attribute::NoInline); | |||
1121 | Fn->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
1122 | Fn->addFnAttr(llvm::Attribute::AlwaysInline); | |||
1123 | } | |||
1124 | CodeGenFunction CGF(CGM); | |||
1125 | // Map "T omp_in;" variable to "*omp_in_parm" value in all expressions. | |||
1126 | // Map "T omp_out;" variable to "*omp_out_parm" value in all expressions. | |||
1127 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, In->getLocation(), | |||
1128 | Out->getLocation()); | |||
1129 | CodeGenFunction::OMPPrivateScope Scope(CGF); | |||
1130 | Address AddrIn = CGF.GetAddrOfLocalVar(&OmpInParm); | |||
1131 | Scope.addPrivate( | |||
1132 | In, CGF.EmitLoadOfPointerLValue(AddrIn, PtrTy->castAs<PointerType>()) | |||
1133 | .getAddress(CGF)); | |||
1134 | Address AddrOut = CGF.GetAddrOfLocalVar(&OmpOutParm); | |||
1135 | Scope.addPrivate( | |||
1136 | Out, CGF.EmitLoadOfPointerLValue(AddrOut, PtrTy->castAs<PointerType>()) | |||
1137 | .getAddress(CGF)); | |||
1138 | (void)Scope.Privatize(); | |||
1139 | if (!IsCombiner && Out->hasInit() && | |||
1140 | !CGF.isTrivialInitializer(Out->getInit())) { | |||
1141 | CGF.EmitAnyExprToMem(Out->getInit(), CGF.GetAddrOfLocalVar(Out), | |||
1142 | Out->getType().getQualifiers(), | |||
1143 | /*IsInitializer=*/true); | |||
1144 | } | |||
1145 | if (CombinerInitializer) | |||
1146 | CGF.EmitIgnoredExpr(CombinerInitializer); | |||
1147 | Scope.ForceCleanup(); | |||
1148 | CGF.FinishFunction(); | |||
1149 | return Fn; | |||
1150 | } | |||
1151 | ||||
1152 | void CGOpenMPRuntime::emitUserDefinedReduction( | |||
1153 | CodeGenFunction *CGF, const OMPDeclareReductionDecl *D) { | |||
1154 | if (UDRMap.count(D) > 0) | |||
1155 | return; | |||
1156 | llvm::Function *Combiner = emitCombinerOrInitializer( | |||
1157 | CGM, D->getType(), D->getCombiner(), | |||
1158 | cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerIn())->getDecl()), | |||
1159 | cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerOut())->getDecl()), | |||
1160 | /*IsCombiner=*/true); | |||
1161 | llvm::Function *Initializer = nullptr; | |||
1162 | if (const Expr *Init = D->getInitializer()) { | |||
1163 | Initializer = emitCombinerOrInitializer( | |||
1164 | CGM, D->getType(), | |||
1165 | D->getInitializerKind() == OMPDeclareReductionDecl::CallInit ? Init | |||
1166 | : nullptr, | |||
1167 | cast<VarDecl>(cast<DeclRefExpr>(D->getInitOrig())->getDecl()), | |||
1168 | cast<VarDecl>(cast<DeclRefExpr>(D->getInitPriv())->getDecl()), | |||
1169 | /*IsCombiner=*/false); | |||
1170 | } | |||
1171 | UDRMap.try_emplace(D, Combiner, Initializer); | |||
1172 | if (CGF) { | |||
1173 | auto &Decls = FunctionUDRMap.FindAndConstruct(CGF->CurFn); | |||
1174 | Decls.second.push_back(D); | |||
1175 | } | |||
1176 | } | |||
1177 | ||||
1178 | std::pair<llvm::Function *, llvm::Function *> | |||
1179 | CGOpenMPRuntime::getUserDefinedReduction(const OMPDeclareReductionDecl *D) { | |||
1180 | auto I = UDRMap.find(D); | |||
1181 | if (I != UDRMap.end()) | |||
1182 | return I->second; | |||
1183 | emitUserDefinedReduction(/*CGF=*/nullptr, D); | |||
1184 | return UDRMap.lookup(D); | |||
1185 | } | |||
1186 | ||||
1187 | namespace { | |||
1188 | // Temporary RAII solution to perform a push/pop stack event on the OpenMP IR | |||
1189 | // Builder if one is present. | |||
1190 | struct PushAndPopStackRAII { | |||
1191 | PushAndPopStackRAII(llvm::OpenMPIRBuilder *OMPBuilder, CodeGenFunction &CGF, | |||
1192 | bool HasCancel, llvm::omp::Directive Kind) | |||
1193 | : OMPBuilder(OMPBuilder) { | |||
1194 | if (!OMPBuilder) | |||
1195 | return; | |||
1196 | ||||
1197 | // The following callback is the crucial part of clangs cleanup process. | |||
1198 | // | |||
1199 | // NOTE: | |||
1200 | // Once the OpenMPIRBuilder is used to create parallel regions (and | |||
1201 | // similar), the cancellation destination (Dest below) is determined via | |||
1202 | // IP. That means if we have variables to finalize we split the block at IP, | |||
1203 | // use the new block (=BB) as destination to build a JumpDest (via | |||
1204 | // getJumpDestInCurrentScope(BB)) which then is fed to | |||
1205 | // EmitBranchThroughCleanup. Furthermore, there will not be the need | |||
1206 | // to push & pop an FinalizationInfo object. | |||
1207 | // The FiniCB will still be needed but at the point where the | |||
1208 | // OpenMPIRBuilder is asked to construct a parallel (or similar) construct. | |||
1209 | auto FiniCB = [&CGF](llvm::OpenMPIRBuilder::InsertPointTy IP) { | |||
1210 | assert(IP.getBlock()->end() == IP.getPoint() &&(static_cast <bool> (IP.getBlock()->end() == IP.getPoint () && "Clang CG should cause non-terminated block!") ? void (0) : __assert_fail ("IP.getBlock()->end() == IP.getPoint() && \"Clang CG should cause non-terminated block!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1211, __extension__ __PRETTY_FUNCTION__)) | |||
1211 | "Clang CG should cause non-terminated block!")(static_cast <bool> (IP.getBlock()->end() == IP.getPoint () && "Clang CG should cause non-terminated block!") ? void (0) : __assert_fail ("IP.getBlock()->end() == IP.getPoint() && \"Clang CG should cause non-terminated block!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1211, __extension__ __PRETTY_FUNCTION__)); | |||
1212 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); | |||
1213 | CGF.Builder.restoreIP(IP); | |||
1214 | CodeGenFunction::JumpDest Dest = | |||
1215 | CGF.getOMPCancelDestination(OMPD_parallel); | |||
1216 | CGF.EmitBranchThroughCleanup(Dest); | |||
1217 | }; | |||
1218 | ||||
1219 | // TODO: Remove this once we emit parallel regions through the | |||
1220 | // OpenMPIRBuilder as it can do this setup internally. | |||
1221 | llvm::OpenMPIRBuilder::FinalizationInfo FI({FiniCB, Kind, HasCancel}); | |||
1222 | OMPBuilder->pushFinalizationCB(std::move(FI)); | |||
1223 | } | |||
1224 | ~PushAndPopStackRAII() { | |||
1225 | if (OMPBuilder) | |||
1226 | OMPBuilder->popFinalizationCB(); | |||
1227 | } | |||
1228 | llvm::OpenMPIRBuilder *OMPBuilder; | |||
1229 | }; | |||
1230 | } // namespace | |||
1231 | ||||
1232 | static llvm::Function *emitParallelOrTeamsOutlinedFunction( | |||
1233 | CodeGenModule &CGM, const OMPExecutableDirective &D, const CapturedStmt *CS, | |||
1234 | const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, | |||
1235 | const StringRef OutlinedHelperName, const RegionCodeGenTy &CodeGen) { | |||
1236 | assert(ThreadIDVar->getType()->isPointerType() &&(static_cast <bool> (ThreadIDVar->getType()->isPointerType () && "thread id variable must be of type kmp_int32 *" ) ? void (0) : __assert_fail ("ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 *\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1237, __extension__ __PRETTY_FUNCTION__)) | |||
1237 | "thread id variable must be of type kmp_int32 *")(static_cast <bool> (ThreadIDVar->getType()->isPointerType () && "thread id variable must be of type kmp_int32 *" ) ? void (0) : __assert_fail ("ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 *\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1237, __extension__ __PRETTY_FUNCTION__)); | |||
1238 | CodeGenFunction CGF(CGM, true); | |||
1239 | bool HasCancel = false; | |||
1240 | if (const auto *OPD = dyn_cast<OMPParallelDirective>(&D)) | |||
1241 | HasCancel = OPD->hasCancel(); | |||
1242 | else if (const auto *OPD = dyn_cast<OMPTargetParallelDirective>(&D)) | |||
1243 | HasCancel = OPD->hasCancel(); | |||
1244 | else if (const auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&D)) | |||
1245 | HasCancel = OPSD->hasCancel(); | |||
1246 | else if (const auto *OPFD = dyn_cast<OMPParallelForDirective>(&D)) | |||
1247 | HasCancel = OPFD->hasCancel(); | |||
1248 | else if (const auto *OPFD = dyn_cast<OMPTargetParallelForDirective>(&D)) | |||
1249 | HasCancel = OPFD->hasCancel(); | |||
1250 | else if (const auto *OPFD = dyn_cast<OMPDistributeParallelForDirective>(&D)) | |||
1251 | HasCancel = OPFD->hasCancel(); | |||
1252 | else if (const auto *OPFD = | |||
1253 | dyn_cast<OMPTeamsDistributeParallelForDirective>(&D)) | |||
1254 | HasCancel = OPFD->hasCancel(); | |||
1255 | else if (const auto *OPFD = | |||
1256 | dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&D)) | |||
1257 | HasCancel = OPFD->hasCancel(); | |||
1258 | ||||
1259 | // TODO: Temporarily inform the OpenMPIRBuilder, if any, about the new | |||
1260 | // parallel region to make cancellation barriers work properly. | |||
1261 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); | |||
1262 | PushAndPopStackRAII PSR(&OMPBuilder, CGF, HasCancel, InnermostKind); | |||
1263 | CGOpenMPOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, InnermostKind, | |||
1264 | HasCancel, OutlinedHelperName); | |||
1265 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
1266 | return CGF.GenerateOpenMPCapturedStmtFunction(*CS, D.getBeginLoc()); | |||
1267 | } | |||
1268 | ||||
1269 | llvm::Function *CGOpenMPRuntime::emitParallelOutlinedFunction( | |||
1270 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1271 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
1272 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_parallel); | |||
1273 | return emitParallelOrTeamsOutlinedFunction( | |||
1274 | CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen); | |||
1275 | } | |||
1276 | ||||
1277 | llvm::Function *CGOpenMPRuntime::emitTeamsOutlinedFunction( | |||
1278 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1279 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
1280 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_teams); | |||
1281 | return emitParallelOrTeamsOutlinedFunction( | |||
1282 | CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen); | |||
1283 | } | |||
1284 | ||||
1285 | llvm::Function *CGOpenMPRuntime::emitTaskOutlinedFunction( | |||
1286 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1287 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, | |||
1288 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, | |||
1289 | bool Tied, unsigned &NumberOfParts) { | |||
1290 | auto &&UntiedCodeGen = [this, &D, TaskTVar](CodeGenFunction &CGF, | |||
1291 | PrePostActionTy &) { | |||
1292 | llvm::Value *ThreadID = getThreadID(CGF, D.getBeginLoc()); | |||
1293 | llvm::Value *UpLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | |||
1294 | llvm::Value *TaskArgs[] = { | |||
1295 | UpLoc, ThreadID, | |||
1296 | CGF.EmitLoadOfPointerLValue(CGF.GetAddrOfLocalVar(TaskTVar), | |||
1297 | TaskTVar->getType()->castAs<PointerType>()) | |||
1298 | .getPointer(CGF)}; | |||
1299 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
1300 | CGM.getModule(), OMPRTL___kmpc_omp_task), | |||
1301 | TaskArgs); | |||
1302 | }; | |||
1303 | CGOpenMPTaskOutlinedRegionInfo::UntiedTaskActionTy Action(Tied, PartIDVar, | |||
1304 | UntiedCodeGen); | |||
1305 | CodeGen.setAction(Action); | |||
1306 | assert(!ThreadIDVar->getType()->isPointerType() &&(static_cast <bool> (!ThreadIDVar->getType()->isPointerType () && "thread id variable must be of type kmp_int32 for tasks" ) ? void (0) : __assert_fail ("!ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 for tasks\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1307, __extension__ __PRETTY_FUNCTION__)) | |||
1307 | "thread id variable must be of type kmp_int32 for tasks")(static_cast <bool> (!ThreadIDVar->getType()->isPointerType () && "thread id variable must be of type kmp_int32 for tasks" ) ? void (0) : __assert_fail ("!ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 for tasks\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1307, __extension__ __PRETTY_FUNCTION__)); | |||
1308 | const OpenMPDirectiveKind Region = | |||
1309 | isOpenMPTaskLoopDirective(D.getDirectiveKind()) ? OMPD_taskloop | |||
1310 | : OMPD_task; | |||
1311 | const CapturedStmt *CS = D.getCapturedStmt(Region); | |||
1312 | bool HasCancel = false; | |||
1313 | if (const auto *TD = dyn_cast<OMPTaskDirective>(&D)) | |||
1314 | HasCancel = TD->hasCancel(); | |||
1315 | else if (const auto *TD = dyn_cast<OMPTaskLoopDirective>(&D)) | |||
1316 | HasCancel = TD->hasCancel(); | |||
1317 | else if (const auto *TD = dyn_cast<OMPMasterTaskLoopDirective>(&D)) | |||
1318 | HasCancel = TD->hasCancel(); | |||
1319 | else if (const auto *TD = dyn_cast<OMPParallelMasterTaskLoopDirective>(&D)) | |||
1320 | HasCancel = TD->hasCancel(); | |||
1321 | ||||
1322 | CodeGenFunction CGF(CGM, true); | |||
1323 | CGOpenMPTaskOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, | |||
1324 | InnermostKind, HasCancel, Action); | |||
1325 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
1326 | llvm::Function *Res = CGF.GenerateCapturedStmtFunction(*CS); | |||
1327 | if (!Tied) | |||
1328 | NumberOfParts = Action.getNumberOfParts(); | |||
1329 | return Res; | |||
1330 | } | |||
1331 | ||||
1332 | void CGOpenMPRuntime::setLocThreadIdInsertPt(CodeGenFunction &CGF, | |||
1333 | bool AtCurrentPoint) { | |||
1334 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1335 | assert(!Elem.second.ServiceInsertPt && "Insert point is set already.")(static_cast <bool> (!Elem.second.ServiceInsertPt && "Insert point is set already.") ? void (0) : __assert_fail ( "!Elem.second.ServiceInsertPt && \"Insert point is set already.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1335, __extension__ __PRETTY_FUNCTION__)); | |||
1336 | ||||
1337 | llvm::Value *Undef = llvm::UndefValue::get(CGF.Int32Ty); | |||
1338 | if (AtCurrentPoint) { | |||
1339 | Elem.second.ServiceInsertPt = new llvm::BitCastInst( | |||
1340 | Undef, CGF.Int32Ty, "svcpt", CGF.Builder.GetInsertBlock()); | |||
1341 | } else { | |||
1342 | Elem.second.ServiceInsertPt = | |||
1343 | new llvm::BitCastInst(Undef, CGF.Int32Ty, "svcpt"); | |||
1344 | Elem.second.ServiceInsertPt->insertAfter(CGF.AllocaInsertPt); | |||
1345 | } | |||
1346 | } | |||
1347 | ||||
1348 | void CGOpenMPRuntime::clearLocThreadIdInsertPt(CodeGenFunction &CGF) { | |||
1349 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1350 | if (Elem.second.ServiceInsertPt) { | |||
1351 | llvm::Instruction *Ptr = Elem.second.ServiceInsertPt; | |||
1352 | Elem.second.ServiceInsertPt = nullptr; | |||
1353 | Ptr->eraseFromParent(); | |||
1354 | } | |||
1355 | } | |||
1356 | ||||
1357 | static StringRef getIdentStringFromSourceLocation(CodeGenFunction &CGF, | |||
1358 | SourceLocation Loc, | |||
1359 | SmallString<128> &Buffer) { | |||
1360 | llvm::raw_svector_ostream OS(Buffer); | |||
1361 | // Build debug location | |||
1362 | PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc); | |||
1363 | OS << ";" << PLoc.getFilename() << ";"; | |||
1364 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) | |||
1365 | OS << FD->getQualifiedNameAsString(); | |||
1366 | OS << ";" << PLoc.getLine() << ";" << PLoc.getColumn() << ";;"; | |||
1367 | return OS.str(); | |||
1368 | } | |||
1369 | ||||
1370 | llvm::Value *CGOpenMPRuntime::emitUpdateLocation(CodeGenFunction &CGF, | |||
1371 | SourceLocation Loc, | |||
1372 | unsigned Flags) { | |||
1373 | uint32_t SrcLocStrSize; | |||
1374 | llvm::Constant *SrcLocStr; | |||
1375 | if (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo || | |||
1376 | Loc.isInvalid()) { | |||
1377 | SrcLocStr = OMPBuilder.getOrCreateDefaultSrcLocStr(SrcLocStrSize); | |||
1378 | } else { | |||
1379 | std::string FunctionName; | |||
1380 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) | |||
1381 | FunctionName = FD->getQualifiedNameAsString(); | |||
1382 | PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc); | |||
1383 | const char *FileName = PLoc.getFilename(); | |||
1384 | unsigned Line = PLoc.getLine(); | |||
1385 | unsigned Column = PLoc.getColumn(); | |||
1386 | SrcLocStr = OMPBuilder.getOrCreateSrcLocStr(FunctionName, FileName, Line, | |||
1387 | Column, SrcLocStrSize); | |||
1388 | } | |||
1389 | unsigned Reserved2Flags = getDefaultLocationReserved2Flags(); | |||
1390 | return OMPBuilder.getOrCreateIdent( | |||
1391 | SrcLocStr, SrcLocStrSize, llvm::omp::IdentFlag(Flags), Reserved2Flags); | |||
1392 | } | |||
1393 | ||||
1394 | llvm::Value *CGOpenMPRuntime::getThreadID(CodeGenFunction &CGF, | |||
1395 | SourceLocation Loc) { | |||
1396 | assert(CGF.CurFn && "No function in current CodeGenFunction.")(static_cast <bool> (CGF.CurFn && "No function in current CodeGenFunction." ) ? void (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1396, __extension__ __PRETTY_FUNCTION__)); | |||
1397 | // If the OpenMPIRBuilder is used we need to use it for all thread id calls as | |||
1398 | // the clang invariants used below might be broken. | |||
1399 | if (CGM.getLangOpts().OpenMPIRBuilder) { | |||
1400 | SmallString<128> Buffer; | |||
1401 | OMPBuilder.updateToLocation(CGF.Builder.saveIP()); | |||
1402 | uint32_t SrcLocStrSize; | |||
1403 | auto *SrcLocStr = OMPBuilder.getOrCreateSrcLocStr( | |||
1404 | getIdentStringFromSourceLocation(CGF, Loc, Buffer), SrcLocStrSize); | |||
1405 | return OMPBuilder.getOrCreateThreadID( | |||
1406 | OMPBuilder.getOrCreateIdent(SrcLocStr, SrcLocStrSize)); | |||
1407 | } | |||
1408 | ||||
1409 | llvm::Value *ThreadID = nullptr; | |||
1410 | // Check whether we've already cached a load of the thread id in this | |||
1411 | // function. | |||
1412 | auto I = OpenMPLocThreadIDMap.find(CGF.CurFn); | |||
1413 | if (I != OpenMPLocThreadIDMap.end()) { | |||
1414 | ThreadID = I->second.ThreadID; | |||
1415 | if (ThreadID != nullptr) | |||
1416 | return ThreadID; | |||
1417 | } | |||
1418 | // If exceptions are enabled, do not use parameter to avoid possible crash. | |||
1419 | if (auto *OMPRegionInfo = | |||
1420 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
1421 | if (OMPRegionInfo->getThreadIDVariable()) { | |||
1422 | // Check if this an outlined function with thread id passed as argument. | |||
1423 | LValue LVal = OMPRegionInfo->getThreadIDVariableLValue(CGF); | |||
1424 | llvm::BasicBlock *TopBlock = CGF.AllocaInsertPt->getParent(); | |||
1425 | if (!CGF.EHStack.requiresLandingPad() || !CGF.getLangOpts().Exceptions || | |||
1426 | !CGF.getLangOpts().CXXExceptions || | |||
1427 | CGF.Builder.GetInsertBlock() == TopBlock || | |||
1428 | !isa<llvm::Instruction>(LVal.getPointer(CGF)) || | |||
1429 | cast<llvm::Instruction>(LVal.getPointer(CGF))->getParent() == | |||
1430 | TopBlock || | |||
1431 | cast<llvm::Instruction>(LVal.getPointer(CGF))->getParent() == | |||
1432 | CGF.Builder.GetInsertBlock()) { | |||
1433 | ThreadID = CGF.EmitLoadOfScalar(LVal, Loc); | |||
1434 | // If value loaded in entry block, cache it and use it everywhere in | |||
1435 | // function. | |||
1436 | if (CGF.Builder.GetInsertBlock() == TopBlock) { | |||
1437 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1438 | Elem.second.ThreadID = ThreadID; | |||
1439 | } | |||
1440 | return ThreadID; | |||
1441 | } | |||
1442 | } | |||
1443 | } | |||
1444 | ||||
1445 | // This is not an outlined function region - need to call __kmpc_int32 | |||
1446 | // kmpc_global_thread_num(ident_t *loc). | |||
1447 | // Generate thread id value and cache this value for use across the | |||
1448 | // function. | |||
1449 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1450 | if (!Elem.second.ServiceInsertPt) | |||
1451 | setLocThreadIdInsertPt(CGF); | |||
1452 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); | |||
1453 | CGF.Builder.SetInsertPoint(Elem.second.ServiceInsertPt); | |||
1454 | llvm::CallInst *Call = CGF.Builder.CreateCall( | |||
1455 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
1456 | OMPRTL___kmpc_global_thread_num), | |||
1457 | emitUpdateLocation(CGF, Loc)); | |||
1458 | Call->setCallingConv(CGF.getRuntimeCC()); | |||
1459 | Elem.second.ThreadID = Call; | |||
1460 | return Call; | |||
1461 | } | |||
1462 | ||||
1463 | void CGOpenMPRuntime::functionFinished(CodeGenFunction &CGF) { | |||
1464 | assert(CGF.CurFn && "No function in current CodeGenFunction.")(static_cast <bool> (CGF.CurFn && "No function in current CodeGenFunction." ) ? void (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1464, __extension__ __PRETTY_FUNCTION__)); | |||
1465 | if (OpenMPLocThreadIDMap.count(CGF.CurFn)) { | |||
1466 | clearLocThreadIdInsertPt(CGF); | |||
1467 | OpenMPLocThreadIDMap.erase(CGF.CurFn); | |||
1468 | } | |||
1469 | if (FunctionUDRMap.count(CGF.CurFn) > 0) { | |||
1470 | for(const auto *D : FunctionUDRMap[CGF.CurFn]) | |||
1471 | UDRMap.erase(D); | |||
1472 | FunctionUDRMap.erase(CGF.CurFn); | |||
1473 | } | |||
1474 | auto I = FunctionUDMMap.find(CGF.CurFn); | |||
1475 | if (I != FunctionUDMMap.end()) { | |||
1476 | for(const auto *D : I->second) | |||
1477 | UDMMap.erase(D); | |||
1478 | FunctionUDMMap.erase(I); | |||
1479 | } | |||
1480 | LastprivateConditionalToTypes.erase(CGF.CurFn); | |||
1481 | FunctionToUntiedTaskStackMap.erase(CGF.CurFn); | |||
1482 | } | |||
1483 | ||||
1484 | llvm::Type *CGOpenMPRuntime::getIdentTyPointerTy() { | |||
1485 | return OMPBuilder.IdentPtr; | |||
1486 | } | |||
1487 | ||||
1488 | llvm::Type *CGOpenMPRuntime::getKmpc_MicroPointerTy() { | |||
1489 | if (!Kmpc_MicroTy) { | |||
1490 | // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...) | |||
1491 | llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty), | |||
1492 | llvm::PointerType::getUnqual(CGM.Int32Ty)}; | |||
1493 | Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true); | |||
1494 | } | |||
1495 | return llvm::PointerType::getUnqual(Kmpc_MicroTy); | |||
1496 | } | |||
1497 | ||||
1498 | llvm::FunctionCallee | |||
1499 | CGOpenMPRuntime::createForStaticInitFunction(unsigned IVSize, bool IVSigned, | |||
1500 | bool IsGPUDistribute) { | |||
1501 | assert((IVSize == 32 || IVSize == 64) &&(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1502, __extension__ __PRETTY_FUNCTION__)) | |||
1502 | "IV size is not compatible with the omp runtime")(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1502, __extension__ __PRETTY_FUNCTION__)); | |||
1503 | StringRef Name; | |||
1504 | if (IsGPUDistribute) | |||
1505 | Name = IVSize == 32 ? (IVSigned ? "__kmpc_distribute_static_init_4" | |||
1506 | : "__kmpc_distribute_static_init_4u") | |||
1507 | : (IVSigned ? "__kmpc_distribute_static_init_8" | |||
1508 | : "__kmpc_distribute_static_init_8u"); | |||
1509 | else | |||
1510 | Name = IVSize == 32 ? (IVSigned ? "__kmpc_for_static_init_4" | |||
1511 | : "__kmpc_for_static_init_4u") | |||
1512 | : (IVSigned ? "__kmpc_for_static_init_8" | |||
1513 | : "__kmpc_for_static_init_8u"); | |||
1514 | ||||
1515 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | |||
1516 | auto *PtrTy = llvm::PointerType::getUnqual(ITy); | |||
1517 | llvm::Type *TypeParams[] = { | |||
1518 | getIdentTyPointerTy(), // loc | |||
1519 | CGM.Int32Ty, // tid | |||
1520 | CGM.Int32Ty, // schedtype | |||
1521 | llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter | |||
1522 | PtrTy, // p_lower | |||
1523 | PtrTy, // p_upper | |||
1524 | PtrTy, // p_stride | |||
1525 | ITy, // incr | |||
1526 | ITy // chunk | |||
1527 | }; | |||
1528 | auto *FnTy = | |||
1529 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1530 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
1531 | } | |||
1532 | ||||
1533 | llvm::FunctionCallee | |||
1534 | CGOpenMPRuntime::createDispatchInitFunction(unsigned IVSize, bool IVSigned) { | |||
1535 | assert((IVSize == 32 || IVSize == 64) &&(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1536, __extension__ __PRETTY_FUNCTION__)) | |||
1536 | "IV size is not compatible with the omp runtime")(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1536, __extension__ __PRETTY_FUNCTION__)); | |||
1537 | StringRef Name = | |||
1538 | IVSize == 32 | |||
1539 | ? (IVSigned ? "__kmpc_dispatch_init_4" : "__kmpc_dispatch_init_4u") | |||
1540 | : (IVSigned ? "__kmpc_dispatch_init_8" : "__kmpc_dispatch_init_8u"); | |||
1541 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | |||
1542 | llvm::Type *TypeParams[] = { getIdentTyPointerTy(), // loc | |||
1543 | CGM.Int32Ty, // tid | |||
1544 | CGM.Int32Ty, // schedtype | |||
1545 | ITy, // lower | |||
1546 | ITy, // upper | |||
1547 | ITy, // stride | |||
1548 | ITy // chunk | |||
1549 | }; | |||
1550 | auto *FnTy = | |||
1551 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1552 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
1553 | } | |||
1554 | ||||
1555 | llvm::FunctionCallee | |||
1556 | CGOpenMPRuntime::createDispatchFiniFunction(unsigned IVSize, bool IVSigned) { | |||
1557 | assert((IVSize == 32 || IVSize == 64) &&(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1558, __extension__ __PRETTY_FUNCTION__)) | |||
1558 | "IV size is not compatible with the omp runtime")(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1558, __extension__ __PRETTY_FUNCTION__)); | |||
1559 | StringRef Name = | |||
1560 | IVSize == 32 | |||
1561 | ? (IVSigned ? "__kmpc_dispatch_fini_4" : "__kmpc_dispatch_fini_4u") | |||
1562 | : (IVSigned ? "__kmpc_dispatch_fini_8" : "__kmpc_dispatch_fini_8u"); | |||
1563 | llvm::Type *TypeParams[] = { | |||
1564 | getIdentTyPointerTy(), // loc | |||
1565 | CGM.Int32Ty, // tid | |||
1566 | }; | |||
1567 | auto *FnTy = | |||
1568 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1569 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
1570 | } | |||
1571 | ||||
1572 | llvm::FunctionCallee | |||
1573 | CGOpenMPRuntime::createDispatchNextFunction(unsigned IVSize, bool IVSigned) { | |||
1574 | assert((IVSize == 32 || IVSize == 64) &&(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1575, __extension__ __PRETTY_FUNCTION__)) | |||
1575 | "IV size is not compatible with the omp runtime")(static_cast <bool> ((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime") ? void (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1575, __extension__ __PRETTY_FUNCTION__)); | |||
1576 | StringRef Name = | |||
1577 | IVSize == 32 | |||
1578 | ? (IVSigned ? "__kmpc_dispatch_next_4" : "__kmpc_dispatch_next_4u") | |||
1579 | : (IVSigned ? "__kmpc_dispatch_next_8" : "__kmpc_dispatch_next_8u"); | |||
1580 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | |||
1581 | auto *PtrTy = llvm::PointerType::getUnqual(ITy); | |||
1582 | llvm::Type *TypeParams[] = { | |||
1583 | getIdentTyPointerTy(), // loc | |||
1584 | CGM.Int32Ty, // tid | |||
1585 | llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter | |||
1586 | PtrTy, // p_lower | |||
1587 | PtrTy, // p_upper | |||
1588 | PtrTy // p_stride | |||
1589 | }; | |||
1590 | auto *FnTy = | |||
1591 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
1592 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
1593 | } | |||
1594 | ||||
1595 | /// Obtain information that uniquely identifies a target entry. This | |||
1596 | /// consists of the file and device IDs as well as line number associated with | |||
1597 | /// the relevant entry source location. | |||
1598 | static llvm::TargetRegionEntryInfo | |||
1599 | getTargetEntryUniqueInfo(ASTContext &C, SourceLocation Loc, | |||
1600 | StringRef ParentName = "") { | |||
1601 | SourceManager &SM = C.getSourceManager(); | |||
1602 | ||||
1603 | // The loc should be always valid and have a file ID (the user cannot use | |||
1604 | // #pragma directives in macros) | |||
1605 | ||||
1606 | assert(Loc.isValid() && "Source location is expected to be always valid.")(static_cast <bool> (Loc.isValid() && "Source location is expected to be always valid." ) ? void (0) : __assert_fail ("Loc.isValid() && \"Source location is expected to be always valid.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1606, __extension__ __PRETTY_FUNCTION__)); | |||
1607 | ||||
1608 | PresumedLoc PLoc = SM.getPresumedLoc(Loc); | |||
1609 | assert(PLoc.isValid() && "Source location is expected to be always valid.")(static_cast <bool> (PLoc.isValid() && "Source location is expected to be always valid." ) ? void (0) : __assert_fail ("PLoc.isValid() && \"Source location is expected to be always valid.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1609, __extension__ __PRETTY_FUNCTION__)); | |||
1610 | ||||
1611 | llvm::sys::fs::UniqueID ID; | |||
1612 | if (auto EC = llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID)) { | |||
1613 | PLoc = SM.getPresumedLoc(Loc, /*UseLineDirectives=*/false); | |||
1614 | assert(PLoc.isValid() && "Source location is expected to be always valid.")(static_cast <bool> (PLoc.isValid() && "Source location is expected to be always valid." ) ? void (0) : __assert_fail ("PLoc.isValid() && \"Source location is expected to be always valid.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1614, __extension__ __PRETTY_FUNCTION__)); | |||
1615 | if (auto EC = llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID)) | |||
1616 | SM.getDiagnostics().Report(diag::err_cannot_open_file) | |||
1617 | << PLoc.getFilename() << EC.message(); | |||
1618 | } | |||
1619 | ||||
1620 | return llvm::TargetRegionEntryInfo(ParentName, ID.getDevice(), ID.getFile(), | |||
1621 | PLoc.getLine()); | |||
1622 | } | |||
1623 | ||||
1624 | Address CGOpenMPRuntime::getAddrOfDeclareTargetVar(const VarDecl *VD) { | |||
1625 | if (CGM.getLangOpts().OpenMPSimd) | |||
1626 | return Address::invalid(); | |||
1627 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
1628 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
1629 | if (Res && (*Res == OMPDeclareTargetDeclAttr::MT_Link || | |||
1630 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | |||
1631 | HasRequiresUnifiedSharedMemory))) { | |||
1632 | SmallString<64> PtrName; | |||
1633 | { | |||
1634 | llvm::raw_svector_ostream OS(PtrName); | |||
1635 | OS << CGM.getMangledName(GlobalDecl(VD)); | |||
1636 | if (!VD->isExternallyVisible()) { | |||
1637 | auto EntryInfo = getTargetEntryUniqueInfo( | |||
1638 | CGM.getContext(), VD->getCanonicalDecl()->getBeginLoc()); | |||
1639 | OS << llvm::format("_%x", EntryInfo.FileID); | |||
1640 | } | |||
1641 | OS << "_decl_tgt_ref_ptr"; | |||
1642 | } | |||
1643 | llvm::Value *Ptr = CGM.getModule().getNamedValue(PtrName); | |||
1644 | QualType PtrTy = CGM.getContext().getPointerType(VD->getType()); | |||
1645 | llvm::Type *LlvmPtrTy = CGM.getTypes().ConvertTypeForMem(PtrTy); | |||
1646 | if (!Ptr) { | |||
1647 | Ptr = getOrCreateInternalVariable(LlvmPtrTy, PtrName); | |||
1648 | ||||
1649 | auto *GV = cast<llvm::GlobalVariable>(Ptr); | |||
1650 | GV->setLinkage(llvm::GlobalValue::WeakAnyLinkage); | |||
1651 | ||||
1652 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
1653 | GV->setInitializer(CGM.GetAddrOfGlobal(VD)); | |||
1654 | registerTargetGlobalVariable(VD, cast<llvm::Constant>(Ptr)); | |||
1655 | } | |||
1656 | return Address(Ptr, LlvmPtrTy, CGM.getContext().getDeclAlign(VD)); | |||
1657 | } | |||
1658 | return Address::invalid(); | |||
1659 | } | |||
1660 | ||||
1661 | llvm::Constant * | |||
1662 | CGOpenMPRuntime::getOrCreateThreadPrivateCache(const VarDecl *VD) { | |||
1663 | assert(!CGM.getLangOpts().OpenMPUseTLS ||(static_cast <bool> (!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()) ? void ( 0) : __assert_fail ("!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1664, __extension__ __PRETTY_FUNCTION__)) | |||
1664 | !CGM.getContext().getTargetInfo().isTLSSupported())(static_cast <bool> (!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()) ? void ( 0) : __assert_fail ("!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1664, __extension__ __PRETTY_FUNCTION__)); | |||
1665 | // Lookup the entry, lazily creating it if necessary. | |||
1666 | std::string Suffix = getName({"cache", ""}); | |||
1667 | return getOrCreateInternalVariable( | |||
1668 | CGM.Int8PtrPtrTy, Twine(CGM.getMangledName(VD)).concat(Suffix)); | |||
1669 | } | |||
1670 | ||||
1671 | Address CGOpenMPRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF, | |||
1672 | const VarDecl *VD, | |||
1673 | Address VDAddr, | |||
1674 | SourceLocation Loc) { | |||
1675 | if (CGM.getLangOpts().OpenMPUseTLS && | |||
1676 | CGM.getContext().getTargetInfo().isTLSSupported()) | |||
1677 | return VDAddr; | |||
1678 | ||||
1679 | llvm::Type *VarTy = VDAddr.getElementType(); | |||
1680 | llvm::Value *Args[] = { | |||
1681 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
1682 | CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.Int8PtrTy), | |||
1683 | CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)), | |||
1684 | getOrCreateThreadPrivateCache(VD)}; | |||
1685 | return Address( | |||
1686 | CGF.EmitRuntimeCall( | |||
1687 | OMPBuilder.getOrCreateRuntimeFunction( | |||
1688 | CGM.getModule(), OMPRTL___kmpc_threadprivate_cached), | |||
1689 | Args), | |||
1690 | CGF.Int8Ty, VDAddr.getAlignment()); | |||
1691 | } | |||
1692 | ||||
1693 | void CGOpenMPRuntime::emitThreadPrivateVarInit( | |||
1694 | CodeGenFunction &CGF, Address VDAddr, llvm::Value *Ctor, | |||
1695 | llvm::Value *CopyCtor, llvm::Value *Dtor, SourceLocation Loc) { | |||
1696 | // Call kmp_int32 __kmpc_global_thread_num(&loc) to init OpenMP runtime | |||
1697 | // library. | |||
1698 | llvm::Value *OMPLoc = emitUpdateLocation(CGF, Loc); | |||
1699 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
1700 | CGM.getModule(), OMPRTL___kmpc_global_thread_num), | |||
1701 | OMPLoc); | |||
1702 | // Call __kmpc_threadprivate_register(&loc, &var, ctor, cctor/*NULL*/, dtor) | |||
1703 | // to register constructor/destructor for variable. | |||
1704 | llvm::Value *Args[] = { | |||
1705 | OMPLoc, CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.VoidPtrTy), | |||
1706 | Ctor, CopyCtor, Dtor}; | |||
1707 | CGF.EmitRuntimeCall( | |||
1708 | OMPBuilder.getOrCreateRuntimeFunction( | |||
1709 | CGM.getModule(), OMPRTL___kmpc_threadprivate_register), | |||
1710 | Args); | |||
1711 | } | |||
1712 | ||||
1713 | llvm::Function *CGOpenMPRuntime::emitThreadPrivateVarDefinition( | |||
1714 | const VarDecl *VD, Address VDAddr, SourceLocation Loc, | |||
1715 | bool PerformInit, CodeGenFunction *CGF) { | |||
1716 | if (CGM.getLangOpts().OpenMPUseTLS && | |||
1717 | CGM.getContext().getTargetInfo().isTLSSupported()) | |||
1718 | return nullptr; | |||
1719 | ||||
1720 | VD = VD->getDefinition(CGM.getContext()); | |||
1721 | if (VD && ThreadPrivateWithDefinition.insert(CGM.getMangledName(VD)).second) { | |||
1722 | QualType ASTTy = VD->getType(); | |||
1723 | ||||
1724 | llvm::Value *Ctor = nullptr, *CopyCtor = nullptr, *Dtor = nullptr; | |||
1725 | const Expr *Init = VD->getAnyInitializer(); | |||
1726 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { | |||
1727 | // Generate function that re-emits the declaration's initializer into the | |||
1728 | // threadprivate copy of the variable VD | |||
1729 | CodeGenFunction CtorCGF(CGM); | |||
1730 | FunctionArgList Args; | |||
1731 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc, | |||
1732 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy, | |||
1733 | ImplicitParamDecl::Other); | |||
1734 | Args.push_back(&Dst); | |||
1735 | ||||
1736 | const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration( | |||
1737 | CGM.getContext().VoidPtrTy, Args); | |||
1738 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
1739 | std::string Name = getName({"__kmpc_global_ctor_", ""}); | |||
1740 | llvm::Function *Fn = | |||
1741 | CGM.CreateGlobalInitOrCleanUpFunction(FTy, Name, FI, Loc); | |||
1742 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidPtrTy, Fn, FI, | |||
1743 | Args, Loc, Loc); | |||
1744 | llvm::Value *ArgVal = CtorCGF.EmitLoadOfScalar( | |||
1745 | CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false, | |||
1746 | CGM.getContext().VoidPtrTy, Dst.getLocation()); | |||
1747 | Address Arg(ArgVal, CtorCGF.Int8Ty, VDAddr.getAlignment()); | |||
1748 | Arg = CtorCGF.Builder.CreateElementBitCast( | |||
1749 | Arg, CtorCGF.ConvertTypeForMem(ASTTy)); | |||
1750 | CtorCGF.EmitAnyExprToMem(Init, Arg, Init->getType().getQualifiers(), | |||
1751 | /*IsInitializer=*/true); | |||
1752 | ArgVal = CtorCGF.EmitLoadOfScalar( | |||
1753 | CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false, | |||
1754 | CGM.getContext().VoidPtrTy, Dst.getLocation()); | |||
1755 | CtorCGF.Builder.CreateStore(ArgVal, CtorCGF.ReturnValue); | |||
1756 | CtorCGF.FinishFunction(); | |||
1757 | Ctor = Fn; | |||
1758 | } | |||
1759 | if (VD->getType().isDestructedType() != QualType::DK_none) { | |||
1760 | // Generate function that emits destructor call for the threadprivate copy | |||
1761 | // of the variable VD | |||
1762 | CodeGenFunction DtorCGF(CGM); | |||
1763 | FunctionArgList Args; | |||
1764 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc, | |||
1765 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy, | |||
1766 | ImplicitParamDecl::Other); | |||
1767 | Args.push_back(&Dst); | |||
1768 | ||||
1769 | const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration( | |||
1770 | CGM.getContext().VoidTy, Args); | |||
1771 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
1772 | std::string Name = getName({"__kmpc_global_dtor_", ""}); | |||
1773 | llvm::Function *Fn = | |||
1774 | CGM.CreateGlobalInitOrCleanUpFunction(FTy, Name, FI, Loc); | |||
1775 | auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF); | |||
1776 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, Args, | |||
1777 | Loc, Loc); | |||
1778 | // Create a scope with an artificial location for the body of this function. | |||
1779 | auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF); | |||
1780 | llvm::Value *ArgVal = DtorCGF.EmitLoadOfScalar( | |||
1781 | DtorCGF.GetAddrOfLocalVar(&Dst), | |||
1782 | /*Volatile=*/false, CGM.getContext().VoidPtrTy, Dst.getLocation()); | |||
1783 | DtorCGF.emitDestroy( | |||
1784 | Address(ArgVal, DtorCGF.Int8Ty, VDAddr.getAlignment()), ASTTy, | |||
1785 | DtorCGF.getDestroyer(ASTTy.isDestructedType()), | |||
1786 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); | |||
1787 | DtorCGF.FinishFunction(); | |||
1788 | Dtor = Fn; | |||
1789 | } | |||
1790 | // Do not emit init function if it is not required. | |||
1791 | if (!Ctor && !Dtor) | |||
1792 | return nullptr; | |||
1793 | ||||
1794 | llvm::Type *CopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
1795 | auto *CopyCtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CopyCtorTyArgs, | |||
1796 | /*isVarArg=*/false) | |||
1797 | ->getPointerTo(); | |||
1798 | // Copying constructor for the threadprivate variable. | |||
1799 | // Must be NULL - reserved by runtime, but currently it requires that this | |||
1800 | // parameter is always NULL. Otherwise it fires assertion. | |||
1801 | CopyCtor = llvm::Constant::getNullValue(CopyCtorTy); | |||
1802 | if (Ctor == nullptr) { | |||
1803 | auto *CtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy, | |||
1804 | /*isVarArg=*/false) | |||
1805 | ->getPointerTo(); | |||
1806 | Ctor = llvm::Constant::getNullValue(CtorTy); | |||
1807 | } | |||
1808 | if (Dtor == nullptr) { | |||
1809 | auto *DtorTy = llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, | |||
1810 | /*isVarArg=*/false) | |||
1811 | ->getPointerTo(); | |||
1812 | Dtor = llvm::Constant::getNullValue(DtorTy); | |||
1813 | } | |||
1814 | if (!CGF) { | |||
1815 | auto *InitFunctionTy = | |||
1816 | llvm::FunctionType::get(CGM.VoidTy, /*isVarArg*/ false); | |||
1817 | std::string Name = getName({"__omp_threadprivate_init_", ""}); | |||
1818 | llvm::Function *InitFunction = CGM.CreateGlobalInitOrCleanUpFunction( | |||
1819 | InitFunctionTy, Name, CGM.getTypes().arrangeNullaryFunction()); | |||
1820 | CodeGenFunction InitCGF(CGM); | |||
1821 | FunctionArgList ArgList; | |||
1822 | InitCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, InitFunction, | |||
1823 | CGM.getTypes().arrangeNullaryFunction(), ArgList, | |||
1824 | Loc, Loc); | |||
1825 | emitThreadPrivateVarInit(InitCGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); | |||
1826 | InitCGF.FinishFunction(); | |||
1827 | return InitFunction; | |||
1828 | } | |||
1829 | emitThreadPrivateVarInit(*CGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); | |||
1830 | } | |||
1831 | return nullptr; | |||
1832 | } | |||
1833 | ||||
1834 | bool CGOpenMPRuntime::emitDeclareTargetVarDefinition(const VarDecl *VD, | |||
1835 | llvm::GlobalVariable *Addr, | |||
1836 | bool PerformInit) { | |||
1837 | if (CGM.getLangOpts().OMPTargetTriples.empty() && | |||
1838 | !CGM.getLangOpts().OpenMPIsDevice) | |||
1839 | return false; | |||
1840 | Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
1841 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
1842 | if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link || | |||
1843 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | |||
1844 | HasRequiresUnifiedSharedMemory)) | |||
1845 | return CGM.getLangOpts().OpenMPIsDevice; | |||
1846 | VD = VD->getDefinition(CGM.getContext()); | |||
1847 | assert(VD && "Unknown VarDecl")(static_cast <bool> (VD && "Unknown VarDecl") ? void (0) : __assert_fail ("VD && \"Unknown VarDecl\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 1847, __extension__ __PRETTY_FUNCTION__)); | |||
1848 | ||||
1849 | if (!DeclareTargetWithDefinition.insert(CGM.getMangledName(VD)).second) | |||
1850 | return CGM.getLangOpts().OpenMPIsDevice; | |||
1851 | ||||
1852 | QualType ASTTy = VD->getType(); | |||
1853 | SourceLocation Loc = VD->getCanonicalDecl()->getBeginLoc(); | |||
1854 | ||||
1855 | // Produce the unique prefix to identify the new target regions. We use | |||
1856 | // the source location of the variable declaration which we know to not | |||
1857 | // conflict with any target region. | |||
1858 | auto EntryInfo = | |||
1859 | getTargetEntryUniqueInfo(CGM.getContext(), Loc, VD->getName()); | |||
1860 | SmallString<128> Buffer, Out; | |||
1861 | OffloadEntriesInfoManager.getTargetRegionEntryFnName(Buffer, EntryInfo); | |||
1862 | ||||
1863 | const Expr *Init = VD->getAnyInitializer(); | |||
1864 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { | |||
1865 | llvm::Constant *Ctor; | |||
1866 | llvm::Constant *ID; | |||
1867 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
1868 | // Generate function that re-emits the declaration's initializer into | |||
1869 | // the threadprivate copy of the variable VD | |||
1870 | CodeGenFunction CtorCGF(CGM); | |||
1871 | ||||
1872 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
1873 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
1874 | llvm::Function *Fn = CGM.CreateGlobalInitOrCleanUpFunction( | |||
1875 | FTy, Twine(Buffer, "_ctor"), FI, Loc, false, | |||
1876 | llvm::GlobalValue::WeakODRLinkage); | |||
1877 | Fn->setVisibility(llvm::GlobalValue::ProtectedVisibility); | |||
1878 | if (CGM.getTriple().isAMDGCN()) | |||
1879 | Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL); | |||
1880 | auto NL = ApplyDebugLocation::CreateEmpty(CtorCGF); | |||
1881 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, | |||
1882 | FunctionArgList(), Loc, Loc); | |||
1883 | auto AL = ApplyDebugLocation::CreateArtificial(CtorCGF); | |||
1884 | llvm::Constant *AddrInAS0 = Addr; | |||
1885 | if (Addr->getAddressSpace() != 0) | |||
1886 | AddrInAS0 = llvm::ConstantExpr::getAddrSpaceCast( | |||
1887 | Addr, llvm::PointerType::getWithSamePointeeType( | |||
1888 | cast<llvm::PointerType>(Addr->getType()), 0)); | |||
1889 | CtorCGF.EmitAnyExprToMem(Init, | |||
1890 | Address(AddrInAS0, Addr->getValueType(), | |||
1891 | CGM.getContext().getDeclAlign(VD)), | |||
1892 | Init->getType().getQualifiers(), | |||
1893 | /*IsInitializer=*/true); | |||
1894 | CtorCGF.FinishFunction(); | |||
1895 | Ctor = Fn; | |||
1896 | ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy); | |||
1897 | } else { | |||
1898 | Ctor = new llvm::GlobalVariable( | |||
1899 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
1900 | llvm::GlobalValue::PrivateLinkage, | |||
1901 | llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_ctor")); | |||
1902 | ID = Ctor; | |||
1903 | } | |||
1904 | ||||
1905 | // Register the information for the entry associated with the constructor. | |||
1906 | Out.clear(); | |||
1907 | auto CtorEntryInfo = EntryInfo; | |||
1908 | CtorEntryInfo.ParentName = Twine(Buffer, "_ctor").toStringRef(Out); | |||
1909 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | |||
1910 | CtorEntryInfo, Ctor, ID, | |||
1911 | llvm::OffloadEntriesInfoManager::OMPTargetRegionEntryCtor, | |||
1912 | CGM.getLangOpts().OpenMPIsDevice); | |||
1913 | } | |||
1914 | if (VD->getType().isDestructedType() != QualType::DK_none) { | |||
1915 | llvm::Constant *Dtor; | |||
1916 | llvm::Constant *ID; | |||
1917 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
1918 | // Generate function that emits destructor call for the threadprivate | |||
1919 | // copy of the variable VD | |||
1920 | CodeGenFunction DtorCGF(CGM); | |||
1921 | ||||
1922 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
1923 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
1924 | llvm::Function *Fn = CGM.CreateGlobalInitOrCleanUpFunction( | |||
1925 | FTy, Twine(Buffer, "_dtor"), FI, Loc, false, | |||
1926 | llvm::GlobalValue::WeakODRLinkage); | |||
1927 | Fn->setVisibility(llvm::GlobalValue::ProtectedVisibility); | |||
1928 | if (CGM.getTriple().isAMDGCN()) | |||
1929 | Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL); | |||
1930 | auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF); | |||
1931 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, | |||
1932 | FunctionArgList(), Loc, Loc); | |||
1933 | // Create a scope with an artificial location for the body of this | |||
1934 | // function. | |||
1935 | auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF); | |||
1936 | llvm::Constant *AddrInAS0 = Addr; | |||
1937 | if (Addr->getAddressSpace() != 0) | |||
1938 | AddrInAS0 = llvm::ConstantExpr::getAddrSpaceCast( | |||
1939 | Addr, llvm::PointerType::getWithSamePointeeType( | |||
1940 | cast<llvm::PointerType>(Addr->getType()), 0)); | |||
1941 | DtorCGF.emitDestroy(Address(AddrInAS0, Addr->getValueType(), | |||
1942 | CGM.getContext().getDeclAlign(VD)), | |||
1943 | ASTTy, DtorCGF.getDestroyer(ASTTy.isDestructedType()), | |||
1944 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); | |||
1945 | DtorCGF.FinishFunction(); | |||
1946 | Dtor = Fn; | |||
1947 | ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy); | |||
1948 | } else { | |||
1949 | Dtor = new llvm::GlobalVariable( | |||
1950 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
1951 | llvm::GlobalValue::PrivateLinkage, | |||
1952 | llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_dtor")); | |||
1953 | ID = Dtor; | |||
1954 | } | |||
1955 | // Register the information for the entry associated with the destructor. | |||
1956 | Out.clear(); | |||
1957 | auto DtorEntryInfo = EntryInfo; | |||
1958 | DtorEntryInfo.ParentName = Twine(Buffer, "_dtor").toStringRef(Out); | |||
1959 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | |||
1960 | DtorEntryInfo, Dtor, ID, | |||
1961 | llvm::OffloadEntriesInfoManager::OMPTargetRegionEntryDtor, | |||
1962 | CGM.getLangOpts().OpenMPIsDevice); | |||
1963 | } | |||
1964 | return CGM.getLangOpts().OpenMPIsDevice; | |||
1965 | } | |||
1966 | ||||
1967 | Address CGOpenMPRuntime::getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF, | |||
1968 | QualType VarType, | |||
1969 | StringRef Name) { | |||
1970 | std::string Suffix = getName({"artificial", ""}); | |||
1971 | llvm::Type *VarLVType = CGF.ConvertTypeForMem(VarType); | |||
1972 | llvm::GlobalVariable *GAddr = | |||
1973 | getOrCreateInternalVariable(VarLVType, Twine(Name).concat(Suffix)); | |||
1974 | if (CGM.getLangOpts().OpenMP && CGM.getLangOpts().OpenMPUseTLS && | |||
1975 | CGM.getTarget().isTLSSupported()) { | |||
1976 | GAddr->setThreadLocal(/*Val=*/true); | |||
1977 | return Address(GAddr, GAddr->getValueType(), | |||
1978 | CGM.getContext().getTypeAlignInChars(VarType)); | |||
1979 | } | |||
1980 | std::string CacheSuffix = getName({"cache", ""}); | |||
1981 | llvm::Value *Args[] = { | |||
1982 | emitUpdateLocation(CGF, SourceLocation()), | |||
1983 | getThreadID(CGF, SourceLocation()), | |||
1984 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(GAddr, CGM.VoidPtrTy), | |||
1985 | CGF.Builder.CreateIntCast(CGF.getTypeSize(VarType), CGM.SizeTy, | |||
1986 | /*isSigned=*/false), | |||
1987 | getOrCreateInternalVariable( | |||
1988 | CGM.VoidPtrPtrTy, Twine(Name).concat(Suffix).concat(CacheSuffix))}; | |||
1989 | return Address( | |||
1990 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
1991 | CGF.EmitRuntimeCall( | |||
1992 | OMPBuilder.getOrCreateRuntimeFunction( | |||
1993 | CGM.getModule(), OMPRTL___kmpc_threadprivate_cached), | |||
1994 | Args), | |||
1995 | VarLVType->getPointerTo(/*AddrSpace=*/0)), | |||
1996 | VarLVType, CGM.getContext().getTypeAlignInChars(VarType)); | |||
1997 | } | |||
1998 | ||||
1999 | void CGOpenMPRuntime::emitIfClause(CodeGenFunction &CGF, const Expr *Cond, | |||
2000 | const RegionCodeGenTy &ThenGen, | |||
2001 | const RegionCodeGenTy &ElseGen) { | |||
2002 | CodeGenFunction::LexicalScope ConditionScope(CGF, Cond->getSourceRange()); | |||
2003 | ||||
2004 | // If the condition constant folds and can be elided, try to avoid emitting | |||
2005 | // the condition and the dead arm of the if/else. | |||
2006 | bool CondConstant; | |||
2007 | if (CGF.ConstantFoldsToSimpleInteger(Cond, CondConstant)) { | |||
2008 | if (CondConstant) | |||
2009 | ThenGen(CGF); | |||
2010 | else | |||
2011 | ElseGen(CGF); | |||
2012 | return; | |||
2013 | } | |||
2014 | ||||
2015 | // Otherwise, the condition did not fold, or we couldn't elide it. Just | |||
2016 | // emit the conditional branch. | |||
2017 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("omp_if.then"); | |||
2018 | llvm::BasicBlock *ElseBlock = CGF.createBasicBlock("omp_if.else"); | |||
2019 | llvm::BasicBlock *ContBlock = CGF.createBasicBlock("omp_if.end"); | |||
2020 | CGF.EmitBranchOnBoolExpr(Cond, ThenBlock, ElseBlock, /*TrueCount=*/0); | |||
2021 | ||||
2022 | // Emit the 'then' code. | |||
2023 | CGF.EmitBlock(ThenBlock); | |||
2024 | ThenGen(CGF); | |||
2025 | CGF.EmitBranch(ContBlock); | |||
2026 | // Emit the 'else' code if present. | |||
2027 | // There is no need to emit line number for unconditional branch. | |||
2028 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2029 | CGF.EmitBlock(ElseBlock); | |||
2030 | ElseGen(CGF); | |||
2031 | // There is no need to emit line number for unconditional branch. | |||
2032 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2033 | CGF.EmitBranch(ContBlock); | |||
2034 | // Emit the continuation block for code after the if. | |||
2035 | CGF.EmitBlock(ContBlock, /*IsFinished=*/true); | |||
2036 | } | |||
2037 | ||||
2038 | void CGOpenMPRuntime::emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
2039 | llvm::Function *OutlinedFn, | |||
2040 | ArrayRef<llvm::Value *> CapturedVars, | |||
2041 | const Expr *IfCond, | |||
2042 | llvm::Value *NumThreads) { | |||
2043 | if (!CGF.HaveInsertPoint()) | |||
2044 | return; | |||
2045 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
2046 | auto &M = CGM.getModule(); | |||
2047 | auto &&ThenGen = [&M, OutlinedFn, CapturedVars, RTLoc, | |||
2048 | this](CodeGenFunction &CGF, PrePostActionTy &) { | |||
2049 | // Build call __kmpc_fork_call(loc, n, microtask, var1, .., varn); | |||
2050 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
2051 | llvm::Value *Args[] = { | |||
2052 | RTLoc, | |||
2053 | CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars | |||
2054 | CGF.Builder.CreateBitCast(OutlinedFn, RT.getKmpc_MicroPointerTy())}; | |||
2055 | llvm::SmallVector<llvm::Value *, 16> RealArgs; | |||
2056 | RealArgs.append(std::begin(Args), std::end(Args)); | |||
2057 | RealArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2058 | ||||
2059 | llvm::FunctionCallee RTLFn = | |||
2060 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_fork_call); | |||
2061 | CGF.EmitRuntimeCall(RTLFn, RealArgs); | |||
2062 | }; | |||
2063 | auto &&ElseGen = [&M, OutlinedFn, CapturedVars, RTLoc, Loc, | |||
2064 | this](CodeGenFunction &CGF, PrePostActionTy &) { | |||
2065 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
2066 | llvm::Value *ThreadID = RT.getThreadID(CGF, Loc); | |||
2067 | // Build calls: | |||
2068 | // __kmpc_serialized_parallel(&Loc, GTid); | |||
2069 | llvm::Value *Args[] = {RTLoc, ThreadID}; | |||
2070 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2071 | M, OMPRTL___kmpc_serialized_parallel), | |||
2072 | Args); | |||
2073 | ||||
2074 | // OutlinedFn(>id, &zero_bound, CapturedStruct); | |||
2075 | Address ThreadIDAddr = RT.emitThreadIDAddress(CGF, Loc); | |||
2076 | Address ZeroAddrBound = | |||
2077 | CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2078 | /*Name=*/".bound.zero.addr"); | |||
2079 | CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddrBound); | |||
2080 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | |||
2081 | // ThreadId for serialized parallels is 0. | |||
2082 | OutlinedFnArgs.push_back(ThreadIDAddr.getPointer()); | |||
2083 | OutlinedFnArgs.push_back(ZeroAddrBound.getPointer()); | |||
2084 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2085 | ||||
2086 | // Ensure we do not inline the function. This is trivially true for the ones | |||
2087 | // passed to __kmpc_fork_call but the ones called in serialized regions | |||
2088 | // could be inlined. This is not a perfect but it is closer to the invariant | |||
2089 | // we want, namely, every data environment starts with a new function. | |||
2090 | // TODO: We should pass the if condition to the runtime function and do the | |||
2091 | // handling there. Much cleaner code. | |||
2092 | OutlinedFn->removeFnAttr(llvm::Attribute::AlwaysInline); | |||
2093 | OutlinedFn->addFnAttr(llvm::Attribute::NoInline); | |||
2094 | RT.emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); | |||
2095 | ||||
2096 | // __kmpc_end_serialized_parallel(&Loc, GTid); | |||
2097 | llvm::Value *EndArgs[] = {RT.emitUpdateLocation(CGF, Loc), ThreadID}; | |||
2098 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2099 | M, OMPRTL___kmpc_end_serialized_parallel), | |||
2100 | EndArgs); | |||
2101 | }; | |||
2102 | if (IfCond) { | |||
2103 | emitIfClause(CGF, IfCond, ThenGen, ElseGen); | |||
2104 | } else { | |||
2105 | RegionCodeGenTy ThenRCG(ThenGen); | |||
2106 | ThenRCG(CGF); | |||
2107 | } | |||
2108 | } | |||
2109 | ||||
2110 | // If we're inside an (outlined) parallel region, use the region info's | |||
2111 | // thread-ID variable (it is passed in a first argument of the outlined function | |||
2112 | // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in | |||
2113 | // regular serial code region, get thread ID by calling kmp_int32 | |||
2114 | // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and | |||
2115 | // return the address of that temp. | |||
2116 | Address CGOpenMPRuntime::emitThreadIDAddress(CodeGenFunction &CGF, | |||
2117 | SourceLocation Loc) { | |||
2118 | if (auto *OMPRegionInfo = | |||
2119 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
2120 | if (OMPRegionInfo->getThreadIDVariable()) | |||
2121 | return OMPRegionInfo->getThreadIDVariableLValue(CGF).getAddress(CGF); | |||
2122 | ||||
2123 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
2124 | QualType Int32Ty = | |||
2125 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true); | |||
2126 | Address ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp."); | |||
2127 | CGF.EmitStoreOfScalar(ThreadID, | |||
2128 | CGF.MakeAddrLValue(ThreadIDTemp, Int32Ty)); | |||
2129 | ||||
2130 | return ThreadIDTemp; | |||
2131 | } | |||
2132 | ||||
2133 | llvm::GlobalVariable *CGOpenMPRuntime::getOrCreateInternalVariable( | |||
2134 | llvm::Type *Ty, const llvm::Twine &Name, unsigned AddressSpace) { | |||
2135 | SmallString<256> Buffer; | |||
2136 | llvm::raw_svector_ostream Out(Buffer); | |||
2137 | Out << Name; | |||
2138 | StringRef RuntimeName = Out.str(); | |||
2139 | auto &Elem = *InternalVars.try_emplace(RuntimeName, nullptr).first; | |||
2140 | if (Elem.second) { | |||
2141 | assert(Elem.second->getType()->isOpaqueOrPointeeTypeMatches(Ty) &&(static_cast <bool> (Elem.second->getType()->isOpaqueOrPointeeTypeMatches (Ty) && "OMP internal variable has different type than requested" ) ? void (0) : __assert_fail ("Elem.second->getType()->isOpaqueOrPointeeTypeMatches(Ty) && \"OMP internal variable has different type than requested\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2142, __extension__ __PRETTY_FUNCTION__)) | |||
2142 | "OMP internal variable has different type than requested")(static_cast <bool> (Elem.second->getType()->isOpaqueOrPointeeTypeMatches (Ty) && "OMP internal variable has different type than requested" ) ? void (0) : __assert_fail ("Elem.second->getType()->isOpaqueOrPointeeTypeMatches(Ty) && \"OMP internal variable has different type than requested\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2142, __extension__ __PRETTY_FUNCTION__)); | |||
2143 | return &*Elem.second; | |||
2144 | } | |||
2145 | ||||
2146 | return Elem.second = new llvm::GlobalVariable( | |||
2147 | CGM.getModule(), Ty, /*IsConstant*/ false, | |||
2148 | llvm::GlobalValue::CommonLinkage, llvm::Constant::getNullValue(Ty), | |||
2149 | Elem.first(), /*InsertBefore=*/nullptr, | |||
2150 | llvm::GlobalValue::NotThreadLocal, AddressSpace); | |||
2151 | } | |||
2152 | ||||
2153 | llvm::Value *CGOpenMPRuntime::getCriticalRegionLock(StringRef CriticalName) { | |||
2154 | std::string Prefix = Twine("gomp_critical_user_", CriticalName).str(); | |||
2155 | std::string Name = getName({Prefix, "var"}); | |||
2156 | return getOrCreateInternalVariable(KmpCriticalNameTy, Name); | |||
2157 | } | |||
2158 | ||||
2159 | namespace { | |||
2160 | /// Common pre(post)-action for different OpenMP constructs. | |||
2161 | class CommonActionTy final : public PrePostActionTy { | |||
2162 | llvm::FunctionCallee EnterCallee; | |||
2163 | ArrayRef<llvm::Value *> EnterArgs; | |||
2164 | llvm::FunctionCallee ExitCallee; | |||
2165 | ArrayRef<llvm::Value *> ExitArgs; | |||
2166 | bool Conditional; | |||
2167 | llvm::BasicBlock *ContBlock = nullptr; | |||
2168 | ||||
2169 | public: | |||
2170 | CommonActionTy(llvm::FunctionCallee EnterCallee, | |||
2171 | ArrayRef<llvm::Value *> EnterArgs, | |||
2172 | llvm::FunctionCallee ExitCallee, | |||
2173 | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) | |||
2174 | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), | |||
2175 | ExitArgs(ExitArgs), Conditional(Conditional) {} | |||
2176 | void Enter(CodeGenFunction &CGF) override { | |||
2177 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); | |||
2178 | if (Conditional) { | |||
2179 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); | |||
2180 | auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); | |||
2181 | ContBlock = CGF.createBasicBlock("omp_if.end"); | |||
2182 | // Generate the branch (If-stmt) | |||
2183 | CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); | |||
2184 | CGF.EmitBlock(ThenBlock); | |||
2185 | } | |||
2186 | } | |||
2187 | void Done(CodeGenFunction &CGF) { | |||
2188 | // Emit the rest of blocks/branches | |||
2189 | CGF.EmitBranch(ContBlock); | |||
2190 | CGF.EmitBlock(ContBlock, true); | |||
2191 | } | |||
2192 | void Exit(CodeGenFunction &CGF) override { | |||
2193 | CGF.EmitRuntimeCall(ExitCallee, ExitArgs); | |||
2194 | } | |||
2195 | }; | |||
2196 | } // anonymous namespace | |||
2197 | ||||
2198 | void CGOpenMPRuntime::emitCriticalRegion(CodeGenFunction &CGF, | |||
2199 | StringRef CriticalName, | |||
2200 | const RegionCodeGenTy &CriticalOpGen, | |||
2201 | SourceLocation Loc, const Expr *Hint) { | |||
2202 | // __kmpc_critical[_with_hint](ident_t *, gtid, Lock[, hint]); | |||
2203 | // CriticalOpGen(); | |||
2204 | // __kmpc_end_critical(ident_t *, gtid, Lock); | |||
2205 | // Prepare arguments and build a call to __kmpc_critical | |||
2206 | if (!CGF.HaveInsertPoint()) | |||
2207 | return; | |||
2208 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
2209 | getCriticalRegionLock(CriticalName)}; | |||
2210 | llvm::SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), | |||
2211 | std::end(Args)); | |||
2212 | if (Hint) { | |||
2213 | EnterArgs.push_back(CGF.Builder.CreateIntCast( | |||
2214 | CGF.EmitScalarExpr(Hint), CGM.Int32Ty, /*isSigned=*/false)); | |||
2215 | } | |||
2216 | CommonActionTy Action( | |||
2217 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2218 | CGM.getModule(), | |||
2219 | Hint ? OMPRTL___kmpc_critical_with_hint : OMPRTL___kmpc_critical), | |||
2220 | EnterArgs, | |||
2221 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
2222 | OMPRTL___kmpc_end_critical), | |||
2223 | Args); | |||
2224 | CriticalOpGen.setAction(Action); | |||
2225 | emitInlinedDirective(CGF, OMPD_critical, CriticalOpGen); | |||
2226 | } | |||
2227 | ||||
2228 | void CGOpenMPRuntime::emitMasterRegion(CodeGenFunction &CGF, | |||
2229 | const RegionCodeGenTy &MasterOpGen, | |||
2230 | SourceLocation Loc) { | |||
2231 | if (!CGF.HaveInsertPoint()) | |||
2232 | return; | |||
2233 | // if(__kmpc_master(ident_t *, gtid)) { | |||
2234 | // MasterOpGen(); | |||
2235 | // __kmpc_end_master(ident_t *, gtid); | |||
2236 | // } | |||
2237 | // Prepare arguments and build a call to __kmpc_master | |||
2238 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
2239 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | |||
2240 | CGM.getModule(), OMPRTL___kmpc_master), | |||
2241 | Args, | |||
2242 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2243 | CGM.getModule(), OMPRTL___kmpc_end_master), | |||
2244 | Args, | |||
2245 | /*Conditional=*/true); | |||
2246 | MasterOpGen.setAction(Action); | |||
2247 | emitInlinedDirective(CGF, OMPD_master, MasterOpGen); | |||
2248 | Action.Done(CGF); | |||
2249 | } | |||
2250 | ||||
2251 | void CGOpenMPRuntime::emitMaskedRegion(CodeGenFunction &CGF, | |||
2252 | const RegionCodeGenTy &MaskedOpGen, | |||
2253 | SourceLocation Loc, const Expr *Filter) { | |||
2254 | if (!CGF.HaveInsertPoint()) | |||
2255 | return; | |||
2256 | // if(__kmpc_masked(ident_t *, gtid, filter)) { | |||
2257 | // MaskedOpGen(); | |||
2258 | // __kmpc_end_masked(iden_t *, gtid); | |||
2259 | // } | |||
2260 | // Prepare arguments and build a call to __kmpc_masked | |||
2261 | llvm::Value *FilterVal = Filter | |||
2262 | ? CGF.EmitScalarExpr(Filter, CGF.Int32Ty) | |||
2263 | : llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/0); | |||
2264 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
2265 | FilterVal}; | |||
2266 | llvm::Value *ArgsEnd[] = {emitUpdateLocation(CGF, Loc), | |||
2267 | getThreadID(CGF, Loc)}; | |||
2268 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | |||
2269 | CGM.getModule(), OMPRTL___kmpc_masked), | |||
2270 | Args, | |||
2271 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2272 | CGM.getModule(), OMPRTL___kmpc_end_masked), | |||
2273 | ArgsEnd, | |||
2274 | /*Conditional=*/true); | |||
2275 | MaskedOpGen.setAction(Action); | |||
2276 | emitInlinedDirective(CGF, OMPD_masked, MaskedOpGen); | |||
2277 | Action.Done(CGF); | |||
2278 | } | |||
2279 | ||||
2280 | void CGOpenMPRuntime::emitTaskyieldCall(CodeGenFunction &CGF, | |||
2281 | SourceLocation Loc) { | |||
2282 | if (!CGF.HaveInsertPoint()) | |||
2283 | return; | |||
2284 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | |||
2285 | OMPBuilder.createTaskyield(CGF.Builder); | |||
2286 | } else { | |||
2287 | // Build call __kmpc_omp_taskyield(loc, thread_id, 0); | |||
2288 | llvm::Value *Args[] = { | |||
2289 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
2290 | llvm::ConstantInt::get(CGM.IntTy, /*V=*/0, /*isSigned=*/true)}; | |||
2291 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2292 | CGM.getModule(), OMPRTL___kmpc_omp_taskyield), | |||
2293 | Args); | |||
2294 | } | |||
2295 | ||||
2296 | if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
2297 | Region->emitUntiedSwitch(CGF); | |||
2298 | } | |||
2299 | ||||
2300 | void CGOpenMPRuntime::emitTaskgroupRegion(CodeGenFunction &CGF, | |||
2301 | const RegionCodeGenTy &TaskgroupOpGen, | |||
2302 | SourceLocation Loc) { | |||
2303 | if (!CGF.HaveInsertPoint()) | |||
2304 | return; | |||
2305 | // __kmpc_taskgroup(ident_t *, gtid); | |||
2306 | // TaskgroupOpGen(); | |||
2307 | // __kmpc_end_taskgroup(ident_t *, gtid); | |||
2308 | // Prepare arguments and build a call to __kmpc_taskgroup | |||
2309 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
2310 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | |||
2311 | CGM.getModule(), OMPRTL___kmpc_taskgroup), | |||
2312 | Args, | |||
2313 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2314 | CGM.getModule(), OMPRTL___kmpc_end_taskgroup), | |||
2315 | Args); | |||
2316 | TaskgroupOpGen.setAction(Action); | |||
2317 | emitInlinedDirective(CGF, OMPD_taskgroup, TaskgroupOpGen); | |||
2318 | } | |||
2319 | ||||
2320 | /// Given an array of pointers to variables, project the address of a | |||
2321 | /// given variable. | |||
2322 | static Address emitAddrOfVarFromArray(CodeGenFunction &CGF, Address Array, | |||
2323 | unsigned Index, const VarDecl *Var) { | |||
2324 | // Pull out the pointer to the variable. | |||
2325 | Address PtrAddr = CGF.Builder.CreateConstArrayGEP(Array, Index); | |||
2326 | llvm::Value *Ptr = CGF.Builder.CreateLoad(PtrAddr); | |||
2327 | ||||
2328 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(Var->getType()); | |||
2329 | return Address( | |||
2330 | CGF.Builder.CreateBitCast( | |||
2331 | Ptr, ElemTy->getPointerTo(Ptr->getType()->getPointerAddressSpace())), | |||
2332 | ElemTy, CGF.getContext().getDeclAlign(Var)); | |||
2333 | } | |||
2334 | ||||
2335 | static llvm::Value *emitCopyprivateCopyFunction( | |||
2336 | CodeGenModule &CGM, llvm::Type *ArgsElemType, | |||
2337 | ArrayRef<const Expr *> CopyprivateVars, ArrayRef<const Expr *> DestExprs, | |||
2338 | ArrayRef<const Expr *> SrcExprs, ArrayRef<const Expr *> AssignmentOps, | |||
2339 | SourceLocation Loc) { | |||
2340 | ASTContext &C = CGM.getContext(); | |||
2341 | // void copy_func(void *LHSArg, void *RHSArg); | |||
2342 | FunctionArgList Args; | |||
2343 | ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
2344 | ImplicitParamDecl::Other); | |||
2345 | ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
2346 | ImplicitParamDecl::Other); | |||
2347 | Args.push_back(&LHSArg); | |||
2348 | Args.push_back(&RHSArg); | |||
2349 | const auto &CGFI = | |||
2350 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
2351 | std::string Name = | |||
2352 | CGM.getOpenMPRuntime().getName({"omp", "copyprivate", "copy_func"}); | |||
2353 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | |||
2354 | llvm::GlobalValue::InternalLinkage, Name, | |||
2355 | &CGM.getModule()); | |||
2356 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
2357 | Fn->setDoesNotRecurse(); | |||
2358 | CodeGenFunction CGF(CGM); | |||
2359 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
2360 | // Dest = (void*[n])(LHSArg); | |||
2361 | // Src = (void*[n])(RHSArg); | |||
2362 | Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
2363 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)), | |||
2364 | ArgsElemType->getPointerTo()), | |||
2365 | ArgsElemType, CGF.getPointerAlign()); | |||
2366 | Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
2367 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)), | |||
2368 | ArgsElemType->getPointerTo()), | |||
2369 | ArgsElemType, CGF.getPointerAlign()); | |||
2370 | // *(Type0*)Dst[0] = *(Type0*)Src[0]; | |||
2371 | // *(Type1*)Dst[1] = *(Type1*)Src[1]; | |||
2372 | // ... | |||
2373 | // *(Typen*)Dst[n] = *(Typen*)Src[n]; | |||
2374 | for (unsigned I = 0, E = AssignmentOps.size(); I < E; ++I) { | |||
2375 | const auto *DestVar = | |||
2376 | cast<VarDecl>(cast<DeclRefExpr>(DestExprs[I])->getDecl()); | |||
2377 | Address DestAddr = emitAddrOfVarFromArray(CGF, LHS, I, DestVar); | |||
2378 | ||||
2379 | const auto *SrcVar = | |||
2380 | cast<VarDecl>(cast<DeclRefExpr>(SrcExprs[I])->getDecl()); | |||
2381 | Address SrcAddr = emitAddrOfVarFromArray(CGF, RHS, I, SrcVar); | |||
2382 | ||||
2383 | const auto *VD = cast<DeclRefExpr>(CopyprivateVars[I])->getDecl(); | |||
2384 | QualType Type = VD->getType(); | |||
2385 | CGF.EmitOMPCopy(Type, DestAddr, SrcAddr, DestVar, SrcVar, AssignmentOps[I]); | |||
2386 | } | |||
2387 | CGF.FinishFunction(); | |||
2388 | return Fn; | |||
2389 | } | |||
2390 | ||||
2391 | void CGOpenMPRuntime::emitSingleRegion(CodeGenFunction &CGF, | |||
2392 | const RegionCodeGenTy &SingleOpGen, | |||
2393 | SourceLocation Loc, | |||
2394 | ArrayRef<const Expr *> CopyprivateVars, | |||
2395 | ArrayRef<const Expr *> SrcExprs, | |||
2396 | ArrayRef<const Expr *> DstExprs, | |||
2397 | ArrayRef<const Expr *> AssignmentOps) { | |||
2398 | if (!CGF.HaveInsertPoint()) | |||
2399 | return; | |||
2400 | assert(CopyprivateVars.size() == SrcExprs.size() &&(static_cast <bool> (CopyprivateVars.size() == SrcExprs .size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? void (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2402, __extension__ __PRETTY_FUNCTION__)) | |||
2401 | CopyprivateVars.size() == DstExprs.size() &&(static_cast <bool> (CopyprivateVars.size() == SrcExprs .size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? void (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2402, __extension__ __PRETTY_FUNCTION__)) | |||
2402 | CopyprivateVars.size() == AssignmentOps.size())(static_cast <bool> (CopyprivateVars.size() == SrcExprs .size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? void (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2402, __extension__ __PRETTY_FUNCTION__)); | |||
2403 | ASTContext &C = CGM.getContext(); | |||
2404 | // int32 did_it = 0; | |||
2405 | // if(__kmpc_single(ident_t *, gtid)) { | |||
2406 | // SingleOpGen(); | |||
2407 | // __kmpc_end_single(ident_t *, gtid); | |||
2408 | // did_it = 1; | |||
2409 | // } | |||
2410 | // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>, | |||
2411 | // <copy_func>, did_it); | |||
2412 | ||||
2413 | Address DidIt = Address::invalid(); | |||
2414 | if (!CopyprivateVars.empty()) { | |||
2415 | // int32 did_it = 0; | |||
2416 | QualType KmpInt32Ty = | |||
2417 | C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | |||
2418 | DidIt = CGF.CreateMemTemp(KmpInt32Ty, ".omp.copyprivate.did_it"); | |||
2419 | CGF.Builder.CreateStore(CGF.Builder.getInt32(0), DidIt); | |||
2420 | } | |||
2421 | // Prepare arguments and build a call to __kmpc_single | |||
2422 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
2423 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | |||
2424 | CGM.getModule(), OMPRTL___kmpc_single), | |||
2425 | Args, | |||
2426 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2427 | CGM.getModule(), OMPRTL___kmpc_end_single), | |||
2428 | Args, | |||
2429 | /*Conditional=*/true); | |||
2430 | SingleOpGen.setAction(Action); | |||
2431 | emitInlinedDirective(CGF, OMPD_single, SingleOpGen); | |||
2432 | if (DidIt.isValid()) { | |||
2433 | // did_it = 1; | |||
2434 | CGF.Builder.CreateStore(CGF.Builder.getInt32(1), DidIt); | |||
2435 | } | |||
2436 | Action.Done(CGF); | |||
2437 | // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>, | |||
2438 | // <copy_func>, did_it); | |||
2439 | if (DidIt.isValid()) { | |||
2440 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, CopyprivateVars.size()); | |||
2441 | QualType CopyprivateArrayTy = C.getConstantArrayType( | |||
2442 | C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, | |||
2443 | /*IndexTypeQuals=*/0); | |||
2444 | // Create a list of all private variables for copyprivate. | |||
2445 | Address CopyprivateList = | |||
2446 | CGF.CreateMemTemp(CopyprivateArrayTy, ".omp.copyprivate.cpr_list"); | |||
2447 | for (unsigned I = 0, E = CopyprivateVars.size(); I < E; ++I) { | |||
2448 | Address Elem = CGF.Builder.CreateConstArrayGEP(CopyprivateList, I); | |||
2449 | CGF.Builder.CreateStore( | |||
2450 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
2451 | CGF.EmitLValue(CopyprivateVars[I]).getPointer(CGF), | |||
2452 | CGF.VoidPtrTy), | |||
2453 | Elem); | |||
2454 | } | |||
2455 | // Build function that copies private values from single region to all other | |||
2456 | // threads in the corresponding parallel region. | |||
2457 | llvm::Value *CpyFn = emitCopyprivateCopyFunction( | |||
2458 | CGM, CGF.ConvertTypeForMem(CopyprivateArrayTy), CopyprivateVars, | |||
2459 | SrcExprs, DstExprs, AssignmentOps, Loc); | |||
2460 | llvm::Value *BufSize = CGF.getTypeSize(CopyprivateArrayTy); | |||
2461 | Address CL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
2462 | CopyprivateList, CGF.VoidPtrTy, CGF.Int8Ty); | |||
2463 | llvm::Value *DidItVal = CGF.Builder.CreateLoad(DidIt); | |||
2464 | llvm::Value *Args[] = { | |||
2465 | emitUpdateLocation(CGF, Loc), // ident_t *<loc> | |||
2466 | getThreadID(CGF, Loc), // i32 <gtid> | |||
2467 | BufSize, // size_t <buf_size> | |||
2468 | CL.getPointer(), // void *<copyprivate list> | |||
2469 | CpyFn, // void (*) (void *, void *) <copy_func> | |||
2470 | DidItVal // i32 did_it | |||
2471 | }; | |||
2472 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2473 | CGM.getModule(), OMPRTL___kmpc_copyprivate), | |||
2474 | Args); | |||
2475 | } | |||
2476 | } | |||
2477 | ||||
2478 | void CGOpenMPRuntime::emitOrderedRegion(CodeGenFunction &CGF, | |||
2479 | const RegionCodeGenTy &OrderedOpGen, | |||
2480 | SourceLocation Loc, bool IsThreads) { | |||
2481 | if (!CGF.HaveInsertPoint()) | |||
2482 | return; | |||
2483 | // __kmpc_ordered(ident_t *, gtid); | |||
2484 | // OrderedOpGen(); | |||
2485 | // __kmpc_end_ordered(ident_t *, gtid); | |||
2486 | // Prepare arguments and build a call to __kmpc_ordered | |||
2487 | if (IsThreads) { | |||
2488 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
2489 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | |||
2490 | CGM.getModule(), OMPRTL___kmpc_ordered), | |||
2491 | Args, | |||
2492 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2493 | CGM.getModule(), OMPRTL___kmpc_end_ordered), | |||
2494 | Args); | |||
2495 | OrderedOpGen.setAction(Action); | |||
2496 | emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen); | |||
2497 | return; | |||
2498 | } | |||
2499 | emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen); | |||
2500 | } | |||
2501 | ||||
2502 | unsigned CGOpenMPRuntime::getDefaultFlagsForBarriers(OpenMPDirectiveKind Kind) { | |||
2503 | unsigned Flags; | |||
2504 | if (Kind == OMPD_for) | |||
2505 | Flags = OMP_IDENT_BARRIER_IMPL_FOR; | |||
2506 | else if (Kind == OMPD_sections) | |||
2507 | Flags = OMP_IDENT_BARRIER_IMPL_SECTIONS; | |||
2508 | else if (Kind == OMPD_single) | |||
2509 | Flags = OMP_IDENT_BARRIER_IMPL_SINGLE; | |||
2510 | else if (Kind == OMPD_barrier) | |||
2511 | Flags = OMP_IDENT_BARRIER_EXPL; | |||
2512 | else | |||
2513 | Flags = OMP_IDENT_BARRIER_IMPL; | |||
2514 | return Flags; | |||
2515 | } | |||
2516 | ||||
2517 | void CGOpenMPRuntime::getDefaultScheduleAndChunk( | |||
2518 | CodeGenFunction &CGF, const OMPLoopDirective &S, | |||
2519 | OpenMPScheduleClauseKind &ScheduleKind, const Expr *&ChunkExpr) const { | |||
2520 | // Check if the loop directive is actually a doacross loop directive. In this | |||
2521 | // case choose static, 1 schedule. | |||
2522 | if (llvm::any_of( | |||
2523 | S.getClausesOfKind<OMPOrderedClause>(), | |||
2524 | [](const OMPOrderedClause *C) { return C->getNumForLoops(); })) { | |||
2525 | ScheduleKind = OMPC_SCHEDULE_static; | |||
2526 | // Chunk size is 1 in this case. | |||
2527 | llvm::APInt ChunkSize(32, 1); | |||
2528 | ChunkExpr = IntegerLiteral::Create( | |||
2529 | CGF.getContext(), ChunkSize, | |||
2530 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), | |||
2531 | SourceLocation()); | |||
2532 | } | |||
2533 | } | |||
2534 | ||||
2535 | void CGOpenMPRuntime::emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
2536 | OpenMPDirectiveKind Kind, bool EmitChecks, | |||
2537 | bool ForceSimpleCall) { | |||
2538 | // Check if we should use the OMPBuilder | |||
2539 | auto *OMPRegionInfo = | |||
2540 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo); | |||
2541 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | |||
2542 | CGF.Builder.restoreIP(OMPBuilder.createBarrier( | |||
2543 | CGF.Builder, Kind, ForceSimpleCall, EmitChecks)); | |||
2544 | return; | |||
2545 | } | |||
2546 | ||||
2547 | if (!CGF.HaveInsertPoint()) | |||
2548 | return; | |||
2549 | // Build call __kmpc_cancel_barrier(loc, thread_id); | |||
2550 | // Build call __kmpc_barrier(loc, thread_id); | |||
2551 | unsigned Flags = getDefaultFlagsForBarriers(Kind); | |||
2552 | // Build call __kmpc_cancel_barrier(loc, thread_id) or __kmpc_barrier(loc, | |||
2553 | // thread_id); | |||
2554 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), | |||
2555 | getThreadID(CGF, Loc)}; | |||
2556 | if (OMPRegionInfo) { | |||
2557 | if (!ForceSimpleCall && OMPRegionInfo->hasCancel()) { | |||
2558 | llvm::Value *Result = CGF.EmitRuntimeCall( | |||
2559 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
2560 | OMPRTL___kmpc_cancel_barrier), | |||
2561 | Args); | |||
2562 | if (EmitChecks) { | |||
2563 | // if (__kmpc_cancel_barrier()) { | |||
2564 | // exit from construct; | |||
2565 | // } | |||
2566 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | |||
2567 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | |||
2568 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | |||
2569 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | |||
2570 | CGF.EmitBlock(ExitBB); | |||
2571 | // exit from construct; | |||
2572 | CodeGenFunction::JumpDest CancelDestination = | |||
2573 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | |||
2574 | CGF.EmitBranchThroughCleanup(CancelDestination); | |||
2575 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | |||
2576 | } | |||
2577 | return; | |||
2578 | } | |||
2579 | } | |||
2580 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2581 | CGM.getModule(), OMPRTL___kmpc_barrier), | |||
2582 | Args); | |||
2583 | } | |||
2584 | ||||
2585 | /// Map the OpenMP loop schedule to the runtime enumeration. | |||
2586 | static OpenMPSchedType getRuntimeSchedule(OpenMPScheduleClauseKind ScheduleKind, | |||
2587 | bool Chunked, bool Ordered) { | |||
2588 | switch (ScheduleKind) { | |||
2589 | case OMPC_SCHEDULE_static: | |||
2590 | return Chunked ? (Ordered ? OMP_ord_static_chunked : OMP_sch_static_chunked) | |||
2591 | : (Ordered ? OMP_ord_static : OMP_sch_static); | |||
2592 | case OMPC_SCHEDULE_dynamic: | |||
2593 | return Ordered ? OMP_ord_dynamic_chunked : OMP_sch_dynamic_chunked; | |||
2594 | case OMPC_SCHEDULE_guided: | |||
2595 | return Ordered ? OMP_ord_guided_chunked : OMP_sch_guided_chunked; | |||
2596 | case OMPC_SCHEDULE_runtime: | |||
2597 | return Ordered ? OMP_ord_runtime : OMP_sch_runtime; | |||
2598 | case OMPC_SCHEDULE_auto: | |||
2599 | return Ordered ? OMP_ord_auto : OMP_sch_auto; | |||
2600 | case OMPC_SCHEDULE_unknown: | |||
2601 | assert(!Chunked && "chunk was specified but schedule kind not known")(static_cast <bool> (!Chunked && "chunk was specified but schedule kind not known" ) ? void (0) : __assert_fail ("!Chunked && \"chunk was specified but schedule kind not known\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2601, __extension__ __PRETTY_FUNCTION__)); | |||
2602 | return Ordered ? OMP_ord_static : OMP_sch_static; | |||
2603 | } | |||
2604 | llvm_unreachable("Unexpected runtime schedule")::llvm::llvm_unreachable_internal("Unexpected runtime schedule" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2604); | |||
2605 | } | |||
2606 | ||||
2607 | /// Map the OpenMP distribute schedule to the runtime enumeration. | |||
2608 | static OpenMPSchedType | |||
2609 | getRuntimeSchedule(OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) { | |||
2610 | // only static is allowed for dist_schedule | |||
2611 | return Chunked ? OMP_dist_sch_static_chunked : OMP_dist_sch_static; | |||
2612 | } | |||
2613 | ||||
2614 | bool CGOpenMPRuntime::isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind, | |||
2615 | bool Chunked) const { | |||
2616 | OpenMPSchedType Schedule = | |||
2617 | getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false); | |||
2618 | return Schedule == OMP_sch_static; | |||
2619 | } | |||
2620 | ||||
2621 | bool CGOpenMPRuntime::isStaticNonchunked( | |||
2622 | OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const { | |||
2623 | OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked); | |||
2624 | return Schedule == OMP_dist_sch_static; | |||
2625 | } | |||
2626 | ||||
2627 | bool CGOpenMPRuntime::isStaticChunked(OpenMPScheduleClauseKind ScheduleKind, | |||
2628 | bool Chunked) const { | |||
2629 | OpenMPSchedType Schedule = | |||
2630 | getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false); | |||
2631 | return Schedule == OMP_sch_static_chunked; | |||
2632 | } | |||
2633 | ||||
2634 | bool CGOpenMPRuntime::isStaticChunked( | |||
2635 | OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const { | |||
2636 | OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked); | |||
2637 | return Schedule == OMP_dist_sch_static_chunked; | |||
2638 | } | |||
2639 | ||||
2640 | bool CGOpenMPRuntime::isDynamic(OpenMPScheduleClauseKind ScheduleKind) const { | |||
2641 | OpenMPSchedType Schedule = | |||
2642 | getRuntimeSchedule(ScheduleKind, /*Chunked=*/false, /*Ordered=*/false); | |||
2643 | assert(Schedule != OMP_sch_static_chunked && "cannot be chunked here")(static_cast <bool> (Schedule != OMP_sch_static_chunked && "cannot be chunked here") ? void (0) : __assert_fail ("Schedule != OMP_sch_static_chunked && \"cannot be chunked here\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2643, __extension__ __PRETTY_FUNCTION__)); | |||
2644 | return Schedule != OMP_sch_static; | |||
2645 | } | |||
2646 | ||||
2647 | static int addMonoNonMonoModifier(CodeGenModule &CGM, OpenMPSchedType Schedule, | |||
2648 | OpenMPScheduleClauseModifier M1, | |||
2649 | OpenMPScheduleClauseModifier M2) { | |||
2650 | int Modifier = 0; | |||
2651 | switch (M1) { | |||
2652 | case OMPC_SCHEDULE_MODIFIER_monotonic: | |||
2653 | Modifier = OMP_sch_modifier_monotonic; | |||
2654 | break; | |||
2655 | case OMPC_SCHEDULE_MODIFIER_nonmonotonic: | |||
2656 | Modifier = OMP_sch_modifier_nonmonotonic; | |||
2657 | break; | |||
2658 | case OMPC_SCHEDULE_MODIFIER_simd: | |||
2659 | if (Schedule == OMP_sch_static_chunked) | |||
2660 | Schedule = OMP_sch_static_balanced_chunked; | |||
2661 | break; | |||
2662 | case OMPC_SCHEDULE_MODIFIER_last: | |||
2663 | case OMPC_SCHEDULE_MODIFIER_unknown: | |||
2664 | break; | |||
2665 | } | |||
2666 | switch (M2) { | |||
2667 | case OMPC_SCHEDULE_MODIFIER_monotonic: | |||
2668 | Modifier = OMP_sch_modifier_monotonic; | |||
2669 | break; | |||
2670 | case OMPC_SCHEDULE_MODIFIER_nonmonotonic: | |||
2671 | Modifier = OMP_sch_modifier_nonmonotonic; | |||
2672 | break; | |||
2673 | case OMPC_SCHEDULE_MODIFIER_simd: | |||
2674 | if (Schedule == OMP_sch_static_chunked) | |||
2675 | Schedule = OMP_sch_static_balanced_chunked; | |||
2676 | break; | |||
2677 | case OMPC_SCHEDULE_MODIFIER_last: | |||
2678 | case OMPC_SCHEDULE_MODIFIER_unknown: | |||
2679 | break; | |||
2680 | } | |||
2681 | // OpenMP 5.0, 2.9.2 Worksharing-Loop Construct, Desription. | |||
2682 | // If the static schedule kind is specified or if the ordered clause is | |||
2683 | // specified, and if the nonmonotonic modifier is not specified, the effect is | |||
2684 | // as if the monotonic modifier is specified. Otherwise, unless the monotonic | |||
2685 | // modifier is specified, the effect is as if the nonmonotonic modifier is | |||
2686 | // specified. | |||
2687 | if (CGM.getLangOpts().OpenMP >= 50 && Modifier == 0) { | |||
2688 | if (!(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static || | |||
2689 | Schedule == OMP_sch_static_balanced_chunked || | |||
2690 | Schedule == OMP_ord_static_chunked || Schedule == OMP_ord_static || | |||
2691 | Schedule == OMP_dist_sch_static_chunked || | |||
2692 | Schedule == OMP_dist_sch_static)) | |||
2693 | Modifier = OMP_sch_modifier_nonmonotonic; | |||
2694 | } | |||
2695 | return Schedule | Modifier; | |||
2696 | } | |||
2697 | ||||
2698 | void CGOpenMPRuntime::emitForDispatchInit( | |||
2699 | CodeGenFunction &CGF, SourceLocation Loc, | |||
2700 | const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned, | |||
2701 | bool Ordered, const DispatchRTInput &DispatchValues) { | |||
2702 | if (!CGF.HaveInsertPoint()) | |||
2703 | return; | |||
2704 | OpenMPSchedType Schedule = getRuntimeSchedule( | |||
2705 | ScheduleKind.Schedule, DispatchValues.Chunk != nullptr, Ordered); | |||
2706 | assert(Ordered ||(static_cast <bool> (Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? void (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2709, __extension__ __PRETTY_FUNCTION__)) | |||
2707 | (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked &&(static_cast <bool> (Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? void (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2709, __extension__ __PRETTY_FUNCTION__)) | |||
2708 | Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked &&(static_cast <bool> (Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? void (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2709, __extension__ __PRETTY_FUNCTION__)) | |||
2709 | Schedule != OMP_sch_static_balanced_chunked))(static_cast <bool> (Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? void (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2709, __extension__ __PRETTY_FUNCTION__)); | |||
2710 | // Call __kmpc_dispatch_init( | |||
2711 | // ident_t *loc, kmp_int32 tid, kmp_int32 schedule, | |||
2712 | // kmp_int[32|64] lower, kmp_int[32|64] upper, | |||
2713 | // kmp_int[32|64] stride, kmp_int[32|64] chunk); | |||
2714 | ||||
2715 | // If the Chunk was not specified in the clause - use default value 1. | |||
2716 | llvm::Value *Chunk = DispatchValues.Chunk ? DispatchValues.Chunk | |||
2717 | : CGF.Builder.getIntN(IVSize, 1); | |||
2718 | llvm::Value *Args[] = { | |||
2719 | emitUpdateLocation(CGF, Loc), | |||
2720 | getThreadID(CGF, Loc), | |||
2721 | CGF.Builder.getInt32(addMonoNonMonoModifier( | |||
2722 | CGM, Schedule, ScheduleKind.M1, ScheduleKind.M2)), // Schedule type | |||
2723 | DispatchValues.LB, // Lower | |||
2724 | DispatchValues.UB, // Upper | |||
2725 | CGF.Builder.getIntN(IVSize, 1), // Stride | |||
2726 | Chunk // Chunk | |||
2727 | }; | |||
2728 | CGF.EmitRuntimeCall(createDispatchInitFunction(IVSize, IVSigned), Args); | |||
2729 | } | |||
2730 | ||||
2731 | static void emitForStaticInitCall( | |||
2732 | CodeGenFunction &CGF, llvm::Value *UpdateLocation, llvm::Value *ThreadId, | |||
2733 | llvm::FunctionCallee ForStaticInitFunction, OpenMPSchedType Schedule, | |||
2734 | OpenMPScheduleClauseModifier M1, OpenMPScheduleClauseModifier M2, | |||
2735 | const CGOpenMPRuntime::StaticRTInput &Values) { | |||
2736 | if (!CGF.HaveInsertPoint()) | |||
2737 | return; | |||
2738 | ||||
2739 | assert(!Values.Ordered)(static_cast <bool> (!Values.Ordered) ? void (0) : __assert_fail ("!Values.Ordered", "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2739, __extension__ __PRETTY_FUNCTION__)); | |||
2740 | assert(Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked ||(static_cast <bool> (Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? void (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2744, __extension__ __PRETTY_FUNCTION__)) | |||
2741 | Schedule == OMP_sch_static_balanced_chunked ||(static_cast <bool> (Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? void (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2744, __extension__ __PRETTY_FUNCTION__)) | |||
2742 | Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked ||(static_cast <bool> (Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? void (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2744, __extension__ __PRETTY_FUNCTION__)) | |||
2743 | Schedule == OMP_dist_sch_static ||(static_cast <bool> (Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? void (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2744, __extension__ __PRETTY_FUNCTION__)) | |||
2744 | Schedule == OMP_dist_sch_static_chunked)(static_cast <bool> (Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? void (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2744, __extension__ __PRETTY_FUNCTION__)); | |||
2745 | ||||
2746 | // Call __kmpc_for_static_init( | |||
2747 | // ident_t *loc, kmp_int32 tid, kmp_int32 schedtype, | |||
2748 | // kmp_int32 *p_lastiter, kmp_int[32|64] *p_lower, | |||
2749 | // kmp_int[32|64] *p_upper, kmp_int[32|64] *p_stride, | |||
2750 | // kmp_int[32|64] incr, kmp_int[32|64] chunk); | |||
2751 | llvm::Value *Chunk = Values.Chunk; | |||
2752 | if (Chunk == nullptr) { | |||
2753 | assert((Schedule == OMP_sch_static || Schedule == OMP_ord_static ||(static_cast <bool> ((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2755, __extension__ __PRETTY_FUNCTION__)) | |||
2754 | Schedule == OMP_dist_sch_static) &&(static_cast <bool> ((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2755, __extension__ __PRETTY_FUNCTION__)) | |||
2755 | "expected static non-chunked schedule")(static_cast <bool> ((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2755, __extension__ __PRETTY_FUNCTION__)); | |||
2756 | // If the Chunk was not specified in the clause - use default value 1. | |||
2757 | Chunk = CGF.Builder.getIntN(Values.IVSize, 1); | |||
2758 | } else { | |||
2759 | assert((Schedule == OMP_sch_static_chunked ||(static_cast <bool> ((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2763, __extension__ __PRETTY_FUNCTION__)) | |||
2760 | Schedule == OMP_sch_static_balanced_chunked ||(static_cast <bool> ((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2763, __extension__ __PRETTY_FUNCTION__)) | |||
2761 | Schedule == OMP_ord_static_chunked ||(static_cast <bool> ((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2763, __extension__ __PRETTY_FUNCTION__)) | |||
2762 | Schedule == OMP_dist_sch_static_chunked) &&(static_cast <bool> ((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2763, __extension__ __PRETTY_FUNCTION__)) | |||
2763 | "expected static chunked schedule")(static_cast <bool> ((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? void (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2763, __extension__ __PRETTY_FUNCTION__)); | |||
2764 | } | |||
2765 | llvm::Value *Args[] = { | |||
2766 | UpdateLocation, | |||
2767 | ThreadId, | |||
2768 | CGF.Builder.getInt32(addMonoNonMonoModifier(CGF.CGM, Schedule, M1, | |||
2769 | M2)), // Schedule type | |||
2770 | Values.IL.getPointer(), // &isLastIter | |||
2771 | Values.LB.getPointer(), // &LB | |||
2772 | Values.UB.getPointer(), // &UB | |||
2773 | Values.ST.getPointer(), // &Stride | |||
2774 | CGF.Builder.getIntN(Values.IVSize, 1), // Incr | |||
2775 | Chunk // Chunk | |||
2776 | }; | |||
2777 | CGF.EmitRuntimeCall(ForStaticInitFunction, Args); | |||
2778 | } | |||
2779 | ||||
2780 | void CGOpenMPRuntime::emitForStaticInit(CodeGenFunction &CGF, | |||
2781 | SourceLocation Loc, | |||
2782 | OpenMPDirectiveKind DKind, | |||
2783 | const OpenMPScheduleTy &ScheduleKind, | |||
2784 | const StaticRTInput &Values) { | |||
2785 | OpenMPSchedType ScheduleNum = getRuntimeSchedule( | |||
2786 | ScheduleKind.Schedule, Values.Chunk != nullptr, Values.Ordered); | |||
2787 | assert(isOpenMPWorksharingDirective(DKind) &&(static_cast <bool> (isOpenMPWorksharingDirective(DKind ) && "Expected loop-based or sections-based directive." ) ? void (0) : __assert_fail ("isOpenMPWorksharingDirective(DKind) && \"Expected loop-based or sections-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2788, __extension__ __PRETTY_FUNCTION__)) | |||
2788 | "Expected loop-based or sections-based directive.")(static_cast <bool> (isOpenMPWorksharingDirective(DKind ) && "Expected loop-based or sections-based directive." ) ? void (0) : __assert_fail ("isOpenMPWorksharingDirective(DKind) && \"Expected loop-based or sections-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2788, __extension__ __PRETTY_FUNCTION__)); | |||
2789 | llvm::Value *UpdatedLocation = emitUpdateLocation(CGF, Loc, | |||
2790 | isOpenMPLoopDirective(DKind) | |||
2791 | ? OMP_IDENT_WORK_LOOP | |||
2792 | : OMP_IDENT_WORK_SECTIONS); | |||
2793 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
2794 | llvm::FunctionCallee StaticInitFunction = | |||
2795 | createForStaticInitFunction(Values.IVSize, Values.IVSigned, false); | |||
2796 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | |||
2797 | emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction, | |||
2798 | ScheduleNum, ScheduleKind.M1, ScheduleKind.M2, Values); | |||
2799 | } | |||
2800 | ||||
2801 | void CGOpenMPRuntime::emitDistributeStaticInit( | |||
2802 | CodeGenFunction &CGF, SourceLocation Loc, | |||
2803 | OpenMPDistScheduleClauseKind SchedKind, | |||
2804 | const CGOpenMPRuntime::StaticRTInput &Values) { | |||
2805 | OpenMPSchedType ScheduleNum = | |||
2806 | getRuntimeSchedule(SchedKind, Values.Chunk != nullptr); | |||
2807 | llvm::Value *UpdatedLocation = | |||
2808 | emitUpdateLocation(CGF, Loc, OMP_IDENT_WORK_DISTRIBUTE); | |||
2809 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
2810 | llvm::FunctionCallee StaticInitFunction; | |||
2811 | bool isGPUDistribute = | |||
2812 | CGM.getLangOpts().OpenMPIsDevice && | |||
2813 | (CGM.getTriple().isAMDGCN() || CGM.getTriple().isNVPTX()); | |||
2814 | StaticInitFunction = createForStaticInitFunction( | |||
2815 | Values.IVSize, Values.IVSigned, isGPUDistribute); | |||
2816 | ||||
2817 | emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction, | |||
2818 | ScheduleNum, OMPC_SCHEDULE_MODIFIER_unknown, | |||
2819 | OMPC_SCHEDULE_MODIFIER_unknown, Values); | |||
2820 | } | |||
2821 | ||||
2822 | void CGOpenMPRuntime::emitForStaticFinish(CodeGenFunction &CGF, | |||
2823 | SourceLocation Loc, | |||
2824 | OpenMPDirectiveKind DKind) { | |||
2825 | if (!CGF.HaveInsertPoint()) | |||
2826 | return; | |||
2827 | // Call __kmpc_for_static_fini(ident_t *loc, kmp_int32 tid); | |||
2828 | llvm::Value *Args[] = { | |||
2829 | emitUpdateLocation(CGF, Loc, | |||
2830 | isOpenMPDistributeDirective(DKind) | |||
2831 | ? OMP_IDENT_WORK_DISTRIBUTE | |||
2832 | : isOpenMPLoopDirective(DKind) | |||
2833 | ? OMP_IDENT_WORK_LOOP | |||
2834 | : OMP_IDENT_WORK_SECTIONS), | |||
2835 | getThreadID(CGF, Loc)}; | |||
2836 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | |||
2837 | if (isOpenMPDistributeDirective(DKind) && CGM.getLangOpts().OpenMPIsDevice && | |||
2838 | (CGM.getTriple().isAMDGCN() || CGM.getTriple().isNVPTX())) | |||
2839 | CGF.EmitRuntimeCall( | |||
2840 | OMPBuilder.getOrCreateRuntimeFunction( | |||
2841 | CGM.getModule(), OMPRTL___kmpc_distribute_static_fini), | |||
2842 | Args); | |||
2843 | else | |||
2844 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2845 | CGM.getModule(), OMPRTL___kmpc_for_static_fini), | |||
2846 | Args); | |||
2847 | } | |||
2848 | ||||
2849 | void CGOpenMPRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF, | |||
2850 | SourceLocation Loc, | |||
2851 | unsigned IVSize, | |||
2852 | bool IVSigned) { | |||
2853 | if (!CGF.HaveInsertPoint()) | |||
2854 | return; | |||
2855 | // Call __kmpc_for_dynamic_fini_(4|8)[u](ident_t *loc, kmp_int32 tid); | |||
2856 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
2857 | CGF.EmitRuntimeCall(createDispatchFiniFunction(IVSize, IVSigned), Args); | |||
2858 | } | |||
2859 | ||||
2860 | llvm::Value *CGOpenMPRuntime::emitForNext(CodeGenFunction &CGF, | |||
2861 | SourceLocation Loc, unsigned IVSize, | |||
2862 | bool IVSigned, Address IL, | |||
2863 | Address LB, Address UB, | |||
2864 | Address ST) { | |||
2865 | // Call __kmpc_dispatch_next( | |||
2866 | // ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter, | |||
2867 | // kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper, | |||
2868 | // kmp_int[32|64] *p_stride); | |||
2869 | llvm::Value *Args[] = { | |||
2870 | emitUpdateLocation(CGF, Loc), | |||
2871 | getThreadID(CGF, Loc), | |||
2872 | IL.getPointer(), // &isLastIter | |||
2873 | LB.getPointer(), // &Lower | |||
2874 | UB.getPointer(), // &Upper | |||
2875 | ST.getPointer() // &Stride | |||
2876 | }; | |||
2877 | llvm::Value *Call = | |||
2878 | CGF.EmitRuntimeCall(createDispatchNextFunction(IVSize, IVSigned), Args); | |||
2879 | return CGF.EmitScalarConversion( | |||
2880 | Call, CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/1), | |||
2881 | CGF.getContext().BoolTy, Loc); | |||
2882 | } | |||
2883 | ||||
2884 | void CGOpenMPRuntime::emitNumThreadsClause(CodeGenFunction &CGF, | |||
2885 | llvm::Value *NumThreads, | |||
2886 | SourceLocation Loc) { | |||
2887 | if (!CGF.HaveInsertPoint()) | |||
2888 | return; | |||
2889 | // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads) | |||
2890 | llvm::Value *Args[] = { | |||
2891 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
2892 | CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)}; | |||
2893 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2894 | CGM.getModule(), OMPRTL___kmpc_push_num_threads), | |||
2895 | Args); | |||
2896 | } | |||
2897 | ||||
2898 | void CGOpenMPRuntime::emitProcBindClause(CodeGenFunction &CGF, | |||
2899 | ProcBindKind ProcBind, | |||
2900 | SourceLocation Loc) { | |||
2901 | if (!CGF.HaveInsertPoint()) | |||
2902 | return; | |||
2903 | assert(ProcBind != OMP_PROC_BIND_unknown && "Unsupported proc_bind value.")(static_cast <bool> (ProcBind != OMP_PROC_BIND_unknown && "Unsupported proc_bind value.") ? void (0) : __assert_fail ( "ProcBind != OMP_PROC_BIND_unknown && \"Unsupported proc_bind value.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 2903, __extension__ __PRETTY_FUNCTION__)); | |||
2904 | // Build call __kmpc_push_proc_bind(&loc, global_tid, proc_bind) | |||
2905 | llvm::Value *Args[] = { | |||
2906 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
2907 | llvm::ConstantInt::get(CGM.IntTy, unsigned(ProcBind), /*isSigned=*/true)}; | |||
2908 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2909 | CGM.getModule(), OMPRTL___kmpc_push_proc_bind), | |||
2910 | Args); | |||
2911 | } | |||
2912 | ||||
2913 | void CGOpenMPRuntime::emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>, | |||
2914 | SourceLocation Loc, llvm::AtomicOrdering AO) { | |||
2915 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | |||
2916 | OMPBuilder.createFlush(CGF.Builder); | |||
2917 | } else { | |||
2918 | if (!CGF.HaveInsertPoint()) | |||
2919 | return; | |||
2920 | // Build call void __kmpc_flush(ident_t *loc) | |||
2921 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
2922 | CGM.getModule(), OMPRTL___kmpc_flush), | |||
2923 | emitUpdateLocation(CGF, Loc)); | |||
2924 | } | |||
2925 | } | |||
2926 | ||||
2927 | namespace { | |||
2928 | /// Indexes of fields for type kmp_task_t. | |||
2929 | enum KmpTaskTFields { | |||
2930 | /// List of shared variables. | |||
2931 | KmpTaskTShareds, | |||
2932 | /// Task routine. | |||
2933 | KmpTaskTRoutine, | |||
2934 | /// Partition id for the untied tasks. | |||
2935 | KmpTaskTPartId, | |||
2936 | /// Function with call of destructors for private variables. | |||
2937 | Data1, | |||
2938 | /// Task priority. | |||
2939 | Data2, | |||
2940 | /// (Taskloops only) Lower bound. | |||
2941 | KmpTaskTLowerBound, | |||
2942 | /// (Taskloops only) Upper bound. | |||
2943 | KmpTaskTUpperBound, | |||
2944 | /// (Taskloops only) Stride. | |||
2945 | KmpTaskTStride, | |||
2946 | /// (Taskloops only) Is last iteration flag. | |||
2947 | KmpTaskTLastIter, | |||
2948 | /// (Taskloops only) Reduction data. | |||
2949 | KmpTaskTReductions, | |||
2950 | }; | |||
2951 | } // anonymous namespace | |||
2952 | ||||
2953 | void CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata() { | |||
2954 | // If we are in simd mode or there are no entries, we don't need to do | |||
2955 | // anything. | |||
2956 | if (CGM.getLangOpts().OpenMPSimd || OffloadEntriesInfoManager.empty()) | |||
2957 | return; | |||
2958 | ||||
2959 | llvm::OpenMPIRBuilder::EmitMetadataErrorReportFunctionTy &&ErrorReportFn = | |||
2960 | [this](llvm::OpenMPIRBuilder::EmitMetadataErrorKind Kind, | |||
2961 | const llvm::TargetRegionEntryInfo &EntryInfo) -> void { | |||
2962 | SourceLocation Loc; | |||
2963 | if (Kind != llvm::OpenMPIRBuilder::EMIT_MD_GLOBAL_VAR_LINK_ERROR) { | |||
2964 | for (auto I = CGM.getContext().getSourceManager().fileinfo_begin(), | |||
2965 | E = CGM.getContext().getSourceManager().fileinfo_end(); | |||
2966 | I != E; ++I) { | |||
2967 | if (I->getFirst()->getUniqueID().getDevice() == EntryInfo.DeviceID && | |||
2968 | I->getFirst()->getUniqueID().getFile() == EntryInfo.FileID) { | |||
2969 | Loc = CGM.getContext().getSourceManager().translateFileLineCol( | |||
2970 | I->getFirst(), EntryInfo.Line, 1); | |||
2971 | break; | |||
2972 | } | |||
2973 | } | |||
2974 | } | |||
2975 | switch (Kind) { | |||
2976 | case llvm::OpenMPIRBuilder::EMIT_MD_TARGET_REGION_ERROR: { | |||
2977 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
2978 | DiagnosticsEngine::Error, "Offloading entry for target region in " | |||
2979 | "%0 is incorrect: either the " | |||
2980 | "address or the ID is invalid."); | |||
2981 | CGM.getDiags().Report(Loc, DiagID) << EntryInfo.ParentName; | |||
2982 | } break; | |||
2983 | case llvm::OpenMPIRBuilder::EMIT_MD_DECLARE_TARGET_ERROR: { | |||
2984 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
2985 | DiagnosticsEngine::Error, "Offloading entry for declare target " | |||
2986 | "variable %0 is incorrect: the " | |||
2987 | "address is invalid."); | |||
2988 | CGM.getDiags().Report(Loc, DiagID) << EntryInfo.ParentName; | |||
2989 | } break; | |||
2990 | case llvm::OpenMPIRBuilder::EMIT_MD_GLOBAL_VAR_LINK_ERROR: { | |||
2991 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
2992 | DiagnosticsEngine::Error, | |||
2993 | "Offloading entry for declare target variable is incorrect: the " | |||
2994 | "address is invalid."); | |||
2995 | CGM.getDiags().Report(DiagID); | |||
2996 | } break; | |||
2997 | } | |||
2998 | }; | |||
2999 | ||||
3000 | OMPBuilder.createOffloadEntriesAndInfoMetadata( | |||
3001 | OffloadEntriesInfoManager, isTargetCodegen(), | |||
3002 | CGM.getLangOpts().OpenMPIsDevice, | |||
3003 | CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory(), ErrorReportFn); | |||
3004 | } | |||
3005 | ||||
3006 | /// Loads all the offload entries information from the host IR | |||
3007 | /// metadata. | |||
3008 | void CGOpenMPRuntime::loadOffloadInfoMetadata() { | |||
3009 | // If we are in target mode, load the metadata from the host IR. This code has | |||
3010 | // to match the metadaata creation in createOffloadEntriesAndInfoMetadata(). | |||
3011 | ||||
3012 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
3013 | return; | |||
3014 | ||||
3015 | if (CGM.getLangOpts().OMPHostIRFile.empty()) | |||
3016 | return; | |||
3017 | ||||
3018 | auto Buf = llvm::MemoryBuffer::getFile(CGM.getLangOpts().OMPHostIRFile); | |||
3019 | if (auto EC = Buf.getError()) { | |||
3020 | CGM.getDiags().Report(diag::err_cannot_open_file) | |||
3021 | << CGM.getLangOpts().OMPHostIRFile << EC.message(); | |||
3022 | return; | |||
3023 | } | |||
3024 | ||||
3025 | llvm::LLVMContext C; | |||
3026 | auto ME = expectedToErrorOrAndEmitErrors( | |||
3027 | C, llvm::parseBitcodeFile(Buf.get()->getMemBufferRef(), C)); | |||
3028 | ||||
3029 | if (auto EC = ME.getError()) { | |||
3030 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
3031 | DiagnosticsEngine::Error, "Unable to parse host IR file '%0':'%1'"); | |||
3032 | CGM.getDiags().Report(DiagID) | |||
3033 | << CGM.getLangOpts().OMPHostIRFile << EC.message(); | |||
3034 | return; | |||
3035 | } | |||
3036 | ||||
3037 | OMPBuilder.loadOffloadInfoMetadata(*ME.get(), OffloadEntriesInfoManager); | |||
3038 | } | |||
3039 | ||||
3040 | void CGOpenMPRuntime::emitKmpRoutineEntryT(QualType KmpInt32Ty) { | |||
3041 | if (!KmpRoutineEntryPtrTy) { | |||
3042 | // Build typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); type. | |||
3043 | ASTContext &C = CGM.getContext(); | |||
3044 | QualType KmpRoutineEntryTyArgs[] = {KmpInt32Ty, C.VoidPtrTy}; | |||
3045 | FunctionProtoType::ExtProtoInfo EPI; | |||
3046 | KmpRoutineEntryPtrQTy = C.getPointerType( | |||
3047 | C.getFunctionType(KmpInt32Ty, KmpRoutineEntryTyArgs, EPI)); | |||
3048 | KmpRoutineEntryPtrTy = CGM.getTypes().ConvertType(KmpRoutineEntryPtrQTy); | |||
3049 | } | |||
3050 | } | |||
3051 | ||||
3052 | namespace { | |||
3053 | struct PrivateHelpersTy { | |||
3054 | PrivateHelpersTy(const Expr *OriginalRef, const VarDecl *Original, | |||
3055 | const VarDecl *PrivateCopy, const VarDecl *PrivateElemInit) | |||
3056 | : OriginalRef(OriginalRef), Original(Original), PrivateCopy(PrivateCopy), | |||
3057 | PrivateElemInit(PrivateElemInit) {} | |||
3058 | PrivateHelpersTy(const VarDecl *Original) : Original(Original) {} | |||
3059 | const Expr *OriginalRef = nullptr; | |||
3060 | const VarDecl *Original = nullptr; | |||
3061 | const VarDecl *PrivateCopy = nullptr; | |||
3062 | const VarDecl *PrivateElemInit = nullptr; | |||
3063 | bool isLocalPrivate() const { | |||
3064 | return !OriginalRef && !PrivateCopy && !PrivateElemInit; | |||
3065 | } | |||
3066 | }; | |||
3067 | typedef std::pair<CharUnits /*Align*/, PrivateHelpersTy> PrivateDataTy; | |||
3068 | } // anonymous namespace | |||
3069 | ||||
3070 | static bool isAllocatableDecl(const VarDecl *VD) { | |||
3071 | const VarDecl *CVD = VD->getCanonicalDecl(); | |||
3072 | if (!CVD->hasAttr<OMPAllocateDeclAttr>()) | |||
3073 | return false; | |||
3074 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); | |||
3075 | // Use the default allocation. | |||
3076 | return !(AA->getAllocatorType() == OMPAllocateDeclAttr::OMPDefaultMemAlloc && | |||
3077 | !AA->getAllocator()); | |||
3078 | } | |||
3079 | ||||
3080 | static RecordDecl * | |||
3081 | createPrivatesRecordDecl(CodeGenModule &CGM, ArrayRef<PrivateDataTy> Privates) { | |||
3082 | if (!Privates.empty()) { | |||
3083 | ASTContext &C = CGM.getContext(); | |||
3084 | // Build struct .kmp_privates_t. { | |||
3085 | // /* private vars */ | |||
3086 | // }; | |||
3087 | RecordDecl *RD = C.buildImplicitRecord(".kmp_privates.t"); | |||
3088 | RD->startDefinition(); | |||
3089 | for (const auto &Pair : Privates) { | |||
3090 | const VarDecl *VD = Pair.second.Original; | |||
3091 | QualType Type = VD->getType().getNonReferenceType(); | |||
3092 | // If the private variable is a local variable with lvalue ref type, | |||
3093 | // allocate the pointer instead of the pointee type. | |||
3094 | if (Pair.second.isLocalPrivate()) { | |||
3095 | if (VD->getType()->isLValueReferenceType()) | |||
3096 | Type = C.getPointerType(Type); | |||
3097 | if (isAllocatableDecl(VD)) | |||
3098 | Type = C.getPointerType(Type); | |||
3099 | } | |||
3100 | FieldDecl *FD = addFieldToRecordDecl(C, RD, Type); | |||
3101 | if (VD->hasAttrs()) { | |||
3102 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), | |||
3103 | E(VD->getAttrs().end()); | |||
3104 | I != E; ++I) | |||
3105 | FD->addAttr(*I); | |||
3106 | } | |||
3107 | } | |||
3108 | RD->completeDefinition(); | |||
3109 | return RD; | |||
3110 | } | |||
3111 | return nullptr; | |||
3112 | } | |||
3113 | ||||
3114 | static RecordDecl * | |||
3115 | createKmpTaskTRecordDecl(CodeGenModule &CGM, OpenMPDirectiveKind Kind, | |||
3116 | QualType KmpInt32Ty, | |||
3117 | QualType KmpRoutineEntryPointerQTy) { | |||
3118 | ASTContext &C = CGM.getContext(); | |||
3119 | // Build struct kmp_task_t { | |||
3120 | // void * shareds; | |||
3121 | // kmp_routine_entry_t routine; | |||
3122 | // kmp_int32 part_id; | |||
3123 | // kmp_cmplrdata_t data1; | |||
3124 | // kmp_cmplrdata_t data2; | |||
3125 | // For taskloops additional fields: | |||
3126 | // kmp_uint64 lb; | |||
3127 | // kmp_uint64 ub; | |||
3128 | // kmp_int64 st; | |||
3129 | // kmp_int32 liter; | |||
3130 | // void * reductions; | |||
3131 | // }; | |||
3132 | RecordDecl *UD = C.buildImplicitRecord("kmp_cmplrdata_t", TTK_Union); | |||
3133 | UD->startDefinition(); | |||
3134 | addFieldToRecordDecl(C, UD, KmpInt32Ty); | |||
3135 | addFieldToRecordDecl(C, UD, KmpRoutineEntryPointerQTy); | |||
3136 | UD->completeDefinition(); | |||
3137 | QualType KmpCmplrdataTy = C.getRecordType(UD); | |||
3138 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_t"); | |||
3139 | RD->startDefinition(); | |||
3140 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
3141 | addFieldToRecordDecl(C, RD, KmpRoutineEntryPointerQTy); | |||
3142 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
3143 | addFieldToRecordDecl(C, RD, KmpCmplrdataTy); | |||
3144 | addFieldToRecordDecl(C, RD, KmpCmplrdataTy); | |||
3145 | if (isOpenMPTaskLoopDirective(Kind)) { | |||
3146 | QualType KmpUInt64Ty = | |||
3147 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0); | |||
3148 | QualType KmpInt64Ty = | |||
3149 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | |||
3150 | addFieldToRecordDecl(C, RD, KmpUInt64Ty); | |||
3151 | addFieldToRecordDecl(C, RD, KmpUInt64Ty); | |||
3152 | addFieldToRecordDecl(C, RD, KmpInt64Ty); | |||
3153 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
3154 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
3155 | } | |||
3156 | RD->completeDefinition(); | |||
3157 | return RD; | |||
3158 | } | |||
3159 | ||||
3160 | static RecordDecl * | |||
3161 | createKmpTaskTWithPrivatesRecordDecl(CodeGenModule &CGM, QualType KmpTaskTQTy, | |||
3162 | ArrayRef<PrivateDataTy> Privates) { | |||
3163 | ASTContext &C = CGM.getContext(); | |||
3164 | // Build struct kmp_task_t_with_privates { | |||
3165 | // kmp_task_t task_data; | |||
3166 | // .kmp_privates_t. privates; | |||
3167 | // }; | |||
3168 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_t_with_privates"); | |||
3169 | RD->startDefinition(); | |||
3170 | addFieldToRecordDecl(C, RD, KmpTaskTQTy); | |||
3171 | if (const RecordDecl *PrivateRD = createPrivatesRecordDecl(CGM, Privates)) | |||
3172 | addFieldToRecordDecl(C, RD, C.getRecordType(PrivateRD)); | |||
3173 | RD->completeDefinition(); | |||
3174 | return RD; | |||
3175 | } | |||
3176 | ||||
3177 | /// Emit a proxy function which accepts kmp_task_t as the second | |||
3178 | /// argument. | |||
3179 | /// \code | |||
3180 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { | |||
3181 | /// TaskFunction(gtid, tt->part_id, &tt->privates, task_privates_map, tt, | |||
3182 | /// For taskloops: | |||
3183 | /// tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter, | |||
3184 | /// tt->reductions, tt->shareds); | |||
3185 | /// return 0; | |||
3186 | /// } | |||
3187 | /// \endcode | |||
3188 | static llvm::Function * | |||
3189 | emitProxyTaskFunction(CodeGenModule &CGM, SourceLocation Loc, | |||
3190 | OpenMPDirectiveKind Kind, QualType KmpInt32Ty, | |||
3191 | QualType KmpTaskTWithPrivatesPtrQTy, | |||
3192 | QualType KmpTaskTWithPrivatesQTy, QualType KmpTaskTQTy, | |||
3193 | QualType SharedsPtrTy, llvm::Function *TaskFunction, | |||
3194 | llvm::Value *TaskPrivatesMap) { | |||
3195 | ASTContext &C = CGM.getContext(); | |||
3196 | FunctionArgList Args; | |||
3197 | ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty, | |||
3198 | ImplicitParamDecl::Other); | |||
3199 | ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3200 | KmpTaskTWithPrivatesPtrQTy.withRestrict(), | |||
3201 | ImplicitParamDecl::Other); | |||
3202 | Args.push_back(&GtidArg); | |||
3203 | Args.push_back(&TaskTypeArg); | |||
3204 | const auto &TaskEntryFnInfo = | |||
3205 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args); | |||
3206 | llvm::FunctionType *TaskEntryTy = | |||
3207 | CGM.getTypes().GetFunctionType(TaskEntryFnInfo); | |||
3208 | std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_entry", ""}); | |||
3209 | auto *TaskEntry = llvm::Function::Create( | |||
3210 | TaskEntryTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule()); | |||
3211 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskEntry, TaskEntryFnInfo); | |||
3212 | TaskEntry->setDoesNotRecurse(); | |||
3213 | CodeGenFunction CGF(CGM); | |||
3214 | CGF.StartFunction(GlobalDecl(), KmpInt32Ty, TaskEntry, TaskEntryFnInfo, Args, | |||
3215 | Loc, Loc); | |||
3216 | ||||
3217 | // TaskFunction(gtid, tt->task_data.part_id, &tt->privates, task_privates_map, | |||
3218 | // tt, | |||
3219 | // For taskloops: | |||
3220 | // tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter, | |||
3221 | // tt->task_data.shareds); | |||
3222 | llvm::Value *GtidParam = CGF.EmitLoadOfScalar( | |||
3223 | CGF.GetAddrOfLocalVar(&GtidArg), /*Volatile=*/false, KmpInt32Ty, Loc); | |||
3224 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | |||
3225 | CGF.GetAddrOfLocalVar(&TaskTypeArg), | |||
3226 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
3227 | const auto *KmpTaskTWithPrivatesQTyRD = | |||
3228 | cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl()); | |||
3229 | LValue Base = | |||
3230 | CGF.EmitLValueForField(TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
3231 | const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl()); | |||
3232 | auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId); | |||
3233 | LValue PartIdLVal = CGF.EmitLValueForField(Base, *PartIdFI); | |||
3234 | llvm::Value *PartidParam = PartIdLVal.getPointer(CGF); | |||
3235 | ||||
3236 | auto SharedsFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds); | |||
3237 | LValue SharedsLVal = CGF.EmitLValueForField(Base, *SharedsFI); | |||
3238 | llvm::Value *SharedsParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3239 | CGF.EmitLoadOfScalar(SharedsLVal, Loc), | |||
3240 | CGF.ConvertTypeForMem(SharedsPtrTy)); | |||
3241 | ||||
3242 | auto PrivatesFI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1); | |||
3243 | llvm::Value *PrivatesParam; | |||
3244 | if (PrivatesFI != KmpTaskTWithPrivatesQTyRD->field_end()) { | |||
3245 | LValue PrivatesLVal = CGF.EmitLValueForField(TDBase, *PrivatesFI); | |||
3246 | PrivatesParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3247 | PrivatesLVal.getPointer(CGF), CGF.VoidPtrTy); | |||
3248 | } else { | |||
3249 | PrivatesParam = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
3250 | } | |||
3251 | ||||
3252 | llvm::Value *CommonArgs[] = { | |||
3253 | GtidParam, PartidParam, PrivatesParam, TaskPrivatesMap, | |||
3254 | CGF.Builder | |||
3255 | .CreatePointerBitCastOrAddrSpaceCast(TDBase.getAddress(CGF), | |||
3256 | CGF.VoidPtrTy, CGF.Int8Ty) | |||
3257 | .getPointer()}; | |||
3258 | SmallVector<llvm::Value *, 16> CallArgs(std::begin(CommonArgs), | |||
3259 | std::end(CommonArgs)); | |||
3260 | if (isOpenMPTaskLoopDirective(Kind)) { | |||
3261 | auto LBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound); | |||
3262 | LValue LBLVal = CGF.EmitLValueForField(Base, *LBFI); | |||
3263 | llvm::Value *LBParam = CGF.EmitLoadOfScalar(LBLVal, Loc); | |||
3264 | auto UBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound); | |||
3265 | LValue UBLVal = CGF.EmitLValueForField(Base, *UBFI); | |||
3266 | llvm::Value *UBParam = CGF.EmitLoadOfScalar(UBLVal, Loc); | |||
3267 | auto StFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTStride); | |||
3268 | LValue StLVal = CGF.EmitLValueForField(Base, *StFI); | |||
3269 | llvm::Value *StParam = CGF.EmitLoadOfScalar(StLVal, Loc); | |||
3270 | auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter); | |||
3271 | LValue LILVal = CGF.EmitLValueForField(Base, *LIFI); | |||
3272 | llvm::Value *LIParam = CGF.EmitLoadOfScalar(LILVal, Loc); | |||
3273 | auto RFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTReductions); | |||
3274 | LValue RLVal = CGF.EmitLValueForField(Base, *RFI); | |||
3275 | llvm::Value *RParam = CGF.EmitLoadOfScalar(RLVal, Loc); | |||
3276 | CallArgs.push_back(LBParam); | |||
3277 | CallArgs.push_back(UBParam); | |||
3278 | CallArgs.push_back(StParam); | |||
3279 | CallArgs.push_back(LIParam); | |||
3280 | CallArgs.push_back(RParam); | |||
3281 | } | |||
3282 | CallArgs.push_back(SharedsParam); | |||
3283 | ||||
3284 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskFunction, | |||
3285 | CallArgs); | |||
3286 | CGF.EmitStoreThroughLValue(RValue::get(CGF.Builder.getInt32(/*C=*/0)), | |||
3287 | CGF.MakeAddrLValue(CGF.ReturnValue, KmpInt32Ty)); | |||
3288 | CGF.FinishFunction(); | |||
3289 | return TaskEntry; | |||
3290 | } | |||
3291 | ||||
3292 | static llvm::Value *emitDestructorsFunction(CodeGenModule &CGM, | |||
3293 | SourceLocation Loc, | |||
3294 | QualType KmpInt32Ty, | |||
3295 | QualType KmpTaskTWithPrivatesPtrQTy, | |||
3296 | QualType KmpTaskTWithPrivatesQTy) { | |||
3297 | ASTContext &C = CGM.getContext(); | |||
3298 | FunctionArgList Args; | |||
3299 | ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty, | |||
3300 | ImplicitParamDecl::Other); | |||
3301 | ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3302 | KmpTaskTWithPrivatesPtrQTy.withRestrict(), | |||
3303 | ImplicitParamDecl::Other); | |||
3304 | Args.push_back(&GtidArg); | |||
3305 | Args.push_back(&TaskTypeArg); | |||
3306 | const auto &DestructorFnInfo = | |||
3307 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args); | |||
3308 | llvm::FunctionType *DestructorFnTy = | |||
3309 | CGM.getTypes().GetFunctionType(DestructorFnInfo); | |||
3310 | std::string Name = | |||
3311 | CGM.getOpenMPRuntime().getName({"omp_task_destructor", ""}); | |||
3312 | auto *DestructorFn = | |||
3313 | llvm::Function::Create(DestructorFnTy, llvm::GlobalValue::InternalLinkage, | |||
3314 | Name, &CGM.getModule()); | |||
3315 | CGM.SetInternalFunctionAttributes(GlobalDecl(), DestructorFn, | |||
3316 | DestructorFnInfo); | |||
3317 | DestructorFn->setDoesNotRecurse(); | |||
3318 | CodeGenFunction CGF(CGM); | |||
3319 | CGF.StartFunction(GlobalDecl(), KmpInt32Ty, DestructorFn, DestructorFnInfo, | |||
3320 | Args, Loc, Loc); | |||
3321 | ||||
3322 | LValue Base = CGF.EmitLoadOfPointerLValue( | |||
3323 | CGF.GetAddrOfLocalVar(&TaskTypeArg), | |||
3324 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
3325 | const auto *KmpTaskTWithPrivatesQTyRD = | |||
3326 | cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl()); | |||
3327 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
3328 | Base = CGF.EmitLValueForField(Base, *FI); | |||
3329 | for (const auto *Field : | |||
3330 | cast<RecordDecl>(FI->getType()->getAsTagDecl())->fields()) { | |||
3331 | if (QualType::DestructionKind DtorKind = | |||
3332 | Field->getType().isDestructedType()) { | |||
3333 | LValue FieldLValue = CGF.EmitLValueForField(Base, Field); | |||
3334 | CGF.pushDestroy(DtorKind, FieldLValue.getAddress(CGF), Field->getType()); | |||
3335 | } | |||
3336 | } | |||
3337 | CGF.FinishFunction(); | |||
3338 | return DestructorFn; | |||
3339 | } | |||
3340 | ||||
3341 | /// Emit a privates mapping function for correct handling of private and | |||
3342 | /// firstprivate variables. | |||
3343 | /// \code | |||
3344 | /// void .omp_task_privates_map.(const .privates. *noalias privs, <ty1> | |||
3345 | /// **noalias priv1,..., <tyn> **noalias privn) { | |||
3346 | /// *priv1 = &.privates.priv1; | |||
3347 | /// ...; | |||
3348 | /// *privn = &.privates.privn; | |||
3349 | /// } | |||
3350 | /// \endcode | |||
3351 | static llvm::Value * | |||
3352 | emitTaskPrivateMappingFunction(CodeGenModule &CGM, SourceLocation Loc, | |||
3353 | const OMPTaskDataTy &Data, QualType PrivatesQTy, | |||
3354 | ArrayRef<PrivateDataTy> Privates) { | |||
3355 | ASTContext &C = CGM.getContext(); | |||
3356 | FunctionArgList Args; | |||
3357 | ImplicitParamDecl TaskPrivatesArg( | |||
3358 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3359 | C.getPointerType(PrivatesQTy).withConst().withRestrict(), | |||
3360 | ImplicitParamDecl::Other); | |||
3361 | Args.push_back(&TaskPrivatesArg); | |||
3362 | llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, unsigned> PrivateVarsPos; | |||
3363 | unsigned Counter = 1; | |||
3364 | for (const Expr *E : Data.PrivateVars) { | |||
3365 | Args.push_back(ImplicitParamDecl::Create( | |||
3366 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3367 | C.getPointerType(C.getPointerType(E->getType())) | |||
3368 | .withConst() | |||
3369 | .withRestrict(), | |||
3370 | ImplicitParamDecl::Other)); | |||
3371 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
3372 | PrivateVarsPos[VD] = Counter; | |||
3373 | ++Counter; | |||
3374 | } | |||
3375 | for (const Expr *E : Data.FirstprivateVars) { | |||
3376 | Args.push_back(ImplicitParamDecl::Create( | |||
3377 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3378 | C.getPointerType(C.getPointerType(E->getType())) | |||
3379 | .withConst() | |||
3380 | .withRestrict(), | |||
3381 | ImplicitParamDecl::Other)); | |||
3382 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
3383 | PrivateVarsPos[VD] = Counter; | |||
3384 | ++Counter; | |||
3385 | } | |||
3386 | for (const Expr *E : Data.LastprivateVars) { | |||
3387 | Args.push_back(ImplicitParamDecl::Create( | |||
3388 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3389 | C.getPointerType(C.getPointerType(E->getType())) | |||
3390 | .withConst() | |||
3391 | .withRestrict(), | |||
3392 | ImplicitParamDecl::Other)); | |||
3393 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
3394 | PrivateVarsPos[VD] = Counter; | |||
3395 | ++Counter; | |||
3396 | } | |||
3397 | for (const VarDecl *VD : Data.PrivateLocals) { | |||
3398 | QualType Ty = VD->getType().getNonReferenceType(); | |||
3399 | if (VD->getType()->isLValueReferenceType()) | |||
3400 | Ty = C.getPointerType(Ty); | |||
3401 | if (isAllocatableDecl(VD)) | |||
3402 | Ty = C.getPointerType(Ty); | |||
3403 | Args.push_back(ImplicitParamDecl::Create( | |||
3404 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3405 | C.getPointerType(C.getPointerType(Ty)).withConst().withRestrict(), | |||
3406 | ImplicitParamDecl::Other)); | |||
3407 | PrivateVarsPos[VD] = Counter; | |||
3408 | ++Counter; | |||
3409 | } | |||
3410 | const auto &TaskPrivatesMapFnInfo = | |||
3411 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3412 | llvm::FunctionType *TaskPrivatesMapTy = | |||
3413 | CGM.getTypes().GetFunctionType(TaskPrivatesMapFnInfo); | |||
3414 | std::string Name = | |||
3415 | CGM.getOpenMPRuntime().getName({"omp_task_privates_map", ""}); | |||
3416 | auto *TaskPrivatesMap = llvm::Function::Create( | |||
3417 | TaskPrivatesMapTy, llvm::GlobalValue::InternalLinkage, Name, | |||
3418 | &CGM.getModule()); | |||
3419 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskPrivatesMap, | |||
3420 | TaskPrivatesMapFnInfo); | |||
3421 | if (CGM.getLangOpts().Optimize) { | |||
3422 | TaskPrivatesMap->removeFnAttr(llvm::Attribute::NoInline); | |||
3423 | TaskPrivatesMap->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
3424 | TaskPrivatesMap->addFnAttr(llvm::Attribute::AlwaysInline); | |||
3425 | } | |||
3426 | CodeGenFunction CGF(CGM); | |||
3427 | CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskPrivatesMap, | |||
3428 | TaskPrivatesMapFnInfo, Args, Loc, Loc); | |||
3429 | ||||
3430 | // *privi = &.privates.privi; | |||
3431 | LValue Base = CGF.EmitLoadOfPointerLValue( | |||
3432 | CGF.GetAddrOfLocalVar(&TaskPrivatesArg), | |||
3433 | TaskPrivatesArg.getType()->castAs<PointerType>()); | |||
3434 | const auto *PrivatesQTyRD = cast<RecordDecl>(PrivatesQTy->getAsTagDecl()); | |||
3435 | Counter = 0; | |||
3436 | for (const FieldDecl *Field : PrivatesQTyRD->fields()) { | |||
3437 | LValue FieldLVal = CGF.EmitLValueForField(Base, Field); | |||
3438 | const VarDecl *VD = Args[PrivateVarsPos[Privates[Counter].second.Original]]; | |||
3439 | LValue RefLVal = | |||
3440 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); | |||
3441 | LValue RefLoadLVal = CGF.EmitLoadOfPointerLValue( | |||
3442 | RefLVal.getAddress(CGF), RefLVal.getType()->castAs<PointerType>()); | |||
3443 | CGF.EmitStoreOfScalar(FieldLVal.getPointer(CGF), RefLoadLVal); | |||
3444 | ++Counter; | |||
3445 | } | |||
3446 | CGF.FinishFunction(); | |||
3447 | return TaskPrivatesMap; | |||
3448 | } | |||
3449 | ||||
3450 | /// Emit initialization for private variables in task-based directives. | |||
3451 | static void emitPrivatesInit(CodeGenFunction &CGF, | |||
3452 | const OMPExecutableDirective &D, | |||
3453 | Address KmpTaskSharedsPtr, LValue TDBase, | |||
3454 | const RecordDecl *KmpTaskTWithPrivatesQTyRD, | |||
3455 | QualType SharedsTy, QualType SharedsPtrTy, | |||
3456 | const OMPTaskDataTy &Data, | |||
3457 | ArrayRef<PrivateDataTy> Privates, bool ForDup) { | |||
3458 | ASTContext &C = CGF.getContext(); | |||
3459 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
3460 | LValue PrivatesBase = CGF.EmitLValueForField(TDBase, *FI); | |||
3461 | OpenMPDirectiveKind Kind = isOpenMPTaskLoopDirective(D.getDirectiveKind()) | |||
3462 | ? OMPD_taskloop | |||
3463 | : OMPD_task; | |||
3464 | const CapturedStmt &CS = *D.getCapturedStmt(Kind); | |||
3465 | CodeGenFunction::CGCapturedStmtInfo CapturesInfo(CS); | |||
3466 | LValue SrcBase; | |||
3467 | bool IsTargetTask = | |||
3468 | isOpenMPTargetDataManagementDirective(D.getDirectiveKind()) || | |||
3469 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()); | |||
3470 | // For target-based directives skip 4 firstprivate arrays BasePointersArray, | |||
3471 | // PointersArray, SizesArray, and MappersArray. The original variables for | |||
3472 | // these arrays are not captured and we get their addresses explicitly. | |||
3473 | if ((!IsTargetTask && !Data.FirstprivateVars.empty() && ForDup) || | |||
3474 | (IsTargetTask && KmpTaskSharedsPtr.isValid())) { | |||
3475 | SrcBase = CGF.MakeAddrLValue( | |||
3476 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3477 | KmpTaskSharedsPtr, CGF.ConvertTypeForMem(SharedsPtrTy), | |||
3478 | CGF.ConvertTypeForMem(SharedsTy)), | |||
3479 | SharedsTy); | |||
3480 | } | |||
3481 | FI = cast<RecordDecl>(FI->getType()->getAsTagDecl())->field_begin(); | |||
3482 | for (const PrivateDataTy &Pair : Privates) { | |||
3483 | // Do not initialize private locals. | |||
3484 | if (Pair.second.isLocalPrivate()) { | |||
3485 | ++FI; | |||
3486 | continue; | |||
3487 | } | |||
3488 | const VarDecl *VD = Pair.second.PrivateCopy; | |||
3489 | const Expr *Init = VD->getAnyInitializer(); | |||
3490 | if (Init && (!ForDup || (isa<CXXConstructExpr>(Init) && | |||
3491 | !CGF.isTrivialInitializer(Init)))) { | |||
3492 | LValue PrivateLValue = CGF.EmitLValueForField(PrivatesBase, *FI); | |||
3493 | if (const VarDecl *Elem = Pair.second.PrivateElemInit) { | |||
3494 | const VarDecl *OriginalVD = Pair.second.Original; | |||
3495 | // Check if the variable is the target-based BasePointersArray, | |||
3496 | // PointersArray, SizesArray, or MappersArray. | |||
3497 | LValue SharedRefLValue; | |||
3498 | QualType Type = PrivateLValue.getType(); | |||
3499 | const FieldDecl *SharedField = CapturesInfo.lookup(OriginalVD); | |||
3500 | if (IsTargetTask && !SharedField) { | |||
3501 | assert(isa<ImplicitParamDecl>(OriginalVD) &&(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3502 | isa<CapturedDecl>(OriginalVD->getDeclContext()) &&(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3503 | cast<CapturedDecl>(OriginalVD->getDeclContext())(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3504 | ->getNumParams() == 0 &&(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3505 | isa<TranslationUnitDecl>((static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3506 | cast<CapturedDecl>(OriginalVD->getDeclContext())(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3507 | ->getDeclContext()) &&(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)) | |||
3508 | "Expected artificial target data variable.")(static_cast <bool> (isa<ImplicitParamDecl>(OriginalVD ) && isa<CapturedDecl>(OriginalVD->getDeclContext ()) && cast<CapturedDecl>(OriginalVD->getDeclContext ()) ->getNumParams() == 0 && isa<TranslationUnitDecl >( cast<CapturedDecl>(OriginalVD->getDeclContext( )) ->getDeclContext()) && "Expected artificial target data variable." ) ? void (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3508, __extension__ __PRETTY_FUNCTION__)); | |||
3509 | SharedRefLValue = | |||
3510 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(OriginalVD), Type); | |||
3511 | } else if (ForDup) { | |||
3512 | SharedRefLValue = CGF.EmitLValueForField(SrcBase, SharedField); | |||
3513 | SharedRefLValue = CGF.MakeAddrLValue( | |||
3514 | SharedRefLValue.getAddress(CGF).withAlignment( | |||
3515 | C.getDeclAlign(OriginalVD)), | |||
3516 | SharedRefLValue.getType(), LValueBaseInfo(AlignmentSource::Decl), | |||
3517 | SharedRefLValue.getTBAAInfo()); | |||
3518 | } else if (CGF.LambdaCaptureFields.count( | |||
3519 | Pair.second.Original->getCanonicalDecl()) > 0 || | |||
3520 | isa_and_nonnull<BlockDecl>(CGF.CurCodeDecl)) { | |||
3521 | SharedRefLValue = CGF.EmitLValue(Pair.second.OriginalRef); | |||
3522 | } else { | |||
3523 | // Processing for implicitly captured variables. | |||
3524 | InlinedOpenMPRegionRAII Region( | |||
3525 | CGF, [](CodeGenFunction &, PrePostActionTy &) {}, OMPD_unknown, | |||
3526 | /*HasCancel=*/false, /*NoInheritance=*/true); | |||
3527 | SharedRefLValue = CGF.EmitLValue(Pair.second.OriginalRef); | |||
3528 | } | |||
3529 | if (Type->isArrayType()) { | |||
3530 | // Initialize firstprivate array. | |||
3531 | if (!isa<CXXConstructExpr>(Init) || CGF.isTrivialInitializer(Init)) { | |||
3532 | // Perform simple memcpy. | |||
3533 | CGF.EmitAggregateAssign(PrivateLValue, SharedRefLValue, Type); | |||
3534 | } else { | |||
3535 | // Initialize firstprivate array using element-by-element | |||
3536 | // initialization. | |||
3537 | CGF.EmitOMPAggregateAssign( | |||
3538 | PrivateLValue.getAddress(CGF), SharedRefLValue.getAddress(CGF), | |||
3539 | Type, | |||
3540 | [&CGF, Elem, Init, &CapturesInfo](Address DestElement, | |||
3541 | Address SrcElement) { | |||
3542 | // Clean up any temporaries needed by the initialization. | |||
3543 | CodeGenFunction::OMPPrivateScope InitScope(CGF); | |||
3544 | InitScope.addPrivate(Elem, SrcElement); | |||
3545 | (void)InitScope.Privatize(); | |||
3546 | // Emit initialization for single element. | |||
3547 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII( | |||
3548 | CGF, &CapturesInfo); | |||
3549 | CGF.EmitAnyExprToMem(Init, DestElement, | |||
3550 | Init->getType().getQualifiers(), | |||
3551 | /*IsInitializer=*/false); | |||
3552 | }); | |||
3553 | } | |||
3554 | } else { | |||
3555 | CodeGenFunction::OMPPrivateScope InitScope(CGF); | |||
3556 | InitScope.addPrivate(Elem, SharedRefLValue.getAddress(CGF)); | |||
3557 | (void)InitScope.Privatize(); | |||
3558 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CapturesInfo); | |||
3559 | CGF.EmitExprAsInit(Init, VD, PrivateLValue, | |||
3560 | /*capturedByInit=*/false); | |||
3561 | } | |||
3562 | } else { | |||
3563 | CGF.EmitExprAsInit(Init, VD, PrivateLValue, /*capturedByInit=*/false); | |||
3564 | } | |||
3565 | } | |||
3566 | ++FI; | |||
3567 | } | |||
3568 | } | |||
3569 | ||||
3570 | /// Check if duplication function is required for taskloops. | |||
3571 | static bool checkInitIsRequired(CodeGenFunction &CGF, | |||
3572 | ArrayRef<PrivateDataTy> Privates) { | |||
3573 | bool InitRequired = false; | |||
3574 | for (const PrivateDataTy &Pair : Privates) { | |||
3575 | if (Pair.second.isLocalPrivate()) | |||
3576 | continue; | |||
3577 | const VarDecl *VD = Pair.second.PrivateCopy; | |||
3578 | const Expr *Init = VD->getAnyInitializer(); | |||
3579 | InitRequired = InitRequired || (isa_and_nonnull<CXXConstructExpr>(Init) && | |||
3580 | !CGF.isTrivialInitializer(Init)); | |||
3581 | if (InitRequired) | |||
3582 | break; | |||
3583 | } | |||
3584 | return InitRequired; | |||
3585 | } | |||
3586 | ||||
3587 | ||||
3588 | /// Emit task_dup function (for initialization of | |||
3589 | /// private/firstprivate/lastprivate vars and last_iter flag) | |||
3590 | /// \code | |||
3591 | /// void __task_dup_entry(kmp_task_t *task_dst, const kmp_task_t *task_src, int | |||
3592 | /// lastpriv) { | |||
3593 | /// // setup lastprivate flag | |||
3594 | /// task_dst->last = lastpriv; | |||
3595 | /// // could be constructor calls here... | |||
3596 | /// } | |||
3597 | /// \endcode | |||
3598 | static llvm::Value * | |||
3599 | emitTaskDupFunction(CodeGenModule &CGM, SourceLocation Loc, | |||
3600 | const OMPExecutableDirective &D, | |||
3601 | QualType KmpTaskTWithPrivatesPtrQTy, | |||
3602 | const RecordDecl *KmpTaskTWithPrivatesQTyRD, | |||
3603 | const RecordDecl *KmpTaskTQTyRD, QualType SharedsTy, | |||
3604 | QualType SharedsPtrTy, const OMPTaskDataTy &Data, | |||
3605 | ArrayRef<PrivateDataTy> Privates, bool WithLastIter) { | |||
3606 | ASTContext &C = CGM.getContext(); | |||
3607 | FunctionArgList Args; | |||
3608 | ImplicitParamDecl DstArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3609 | KmpTaskTWithPrivatesPtrQTy, | |||
3610 | ImplicitParamDecl::Other); | |||
3611 | ImplicitParamDecl SrcArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3612 | KmpTaskTWithPrivatesPtrQTy, | |||
3613 | ImplicitParamDecl::Other); | |||
3614 | ImplicitParamDecl LastprivArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | |||
3615 | ImplicitParamDecl::Other); | |||
3616 | Args.push_back(&DstArg); | |||
3617 | Args.push_back(&SrcArg); | |||
3618 | Args.push_back(&LastprivArg); | |||
3619 | const auto &TaskDupFnInfo = | |||
3620 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3621 | llvm::FunctionType *TaskDupTy = CGM.getTypes().GetFunctionType(TaskDupFnInfo); | |||
3622 | std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_dup", ""}); | |||
3623 | auto *TaskDup = llvm::Function::Create( | |||
3624 | TaskDupTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule()); | |||
3625 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskDup, TaskDupFnInfo); | |||
3626 | TaskDup->setDoesNotRecurse(); | |||
3627 | CodeGenFunction CGF(CGM); | |||
3628 | CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskDup, TaskDupFnInfo, Args, Loc, | |||
3629 | Loc); | |||
3630 | ||||
3631 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | |||
3632 | CGF.GetAddrOfLocalVar(&DstArg), | |||
3633 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
3634 | // task_dst->liter = lastpriv; | |||
3635 | if (WithLastIter) { | |||
3636 | auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter); | |||
3637 | LValue Base = CGF.EmitLValueForField( | |||
3638 | TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
3639 | LValue LILVal = CGF.EmitLValueForField(Base, *LIFI); | |||
3640 | llvm::Value *Lastpriv = CGF.EmitLoadOfScalar( | |||
3641 | CGF.GetAddrOfLocalVar(&LastprivArg), /*Volatile=*/false, C.IntTy, Loc); | |||
3642 | CGF.EmitStoreOfScalar(Lastpriv, LILVal); | |||
3643 | } | |||
3644 | ||||
3645 | // Emit initial values for private copies (if any). | |||
3646 | assert(!Privates.empty())(static_cast <bool> (!Privates.empty()) ? void (0) : __assert_fail ("!Privates.empty()", "clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3646, __extension__ __PRETTY_FUNCTION__)); | |||
3647 | Address KmpTaskSharedsPtr = Address::invalid(); | |||
3648 | if (!Data.FirstprivateVars.empty()) { | |||
3649 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | |||
3650 | CGF.GetAddrOfLocalVar(&SrcArg), | |||
3651 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
3652 | LValue Base = CGF.EmitLValueForField( | |||
3653 | TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
3654 | KmpTaskSharedsPtr = Address( | |||
3655 | CGF.EmitLoadOfScalar(CGF.EmitLValueForField( | |||
3656 | Base, *std::next(KmpTaskTQTyRD->field_begin(), | |||
3657 | KmpTaskTShareds)), | |||
3658 | Loc), | |||
3659 | CGF.Int8Ty, CGM.getNaturalTypeAlignment(SharedsTy)); | |||
3660 | } | |||
3661 | emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, TDBase, KmpTaskTWithPrivatesQTyRD, | |||
3662 | SharedsTy, SharedsPtrTy, Data, Privates, /*ForDup=*/true); | |||
3663 | CGF.FinishFunction(); | |||
3664 | return TaskDup; | |||
3665 | } | |||
3666 | ||||
3667 | /// Checks if destructor function is required to be generated. | |||
3668 | /// \return true if cleanups are required, false otherwise. | |||
3669 | static bool | |||
3670 | checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD, | |||
3671 | ArrayRef<PrivateDataTy> Privates) { | |||
3672 | for (const PrivateDataTy &P : Privates) { | |||
3673 | if (P.second.isLocalPrivate()) | |||
3674 | continue; | |||
3675 | QualType Ty = P.second.Original->getType().getNonReferenceType(); | |||
3676 | if (Ty.isDestructedType()) | |||
3677 | return true; | |||
3678 | } | |||
3679 | return false; | |||
3680 | } | |||
3681 | ||||
3682 | namespace { | |||
3683 | /// Loop generator for OpenMP iterator expression. | |||
3684 | class OMPIteratorGeneratorScope final | |||
3685 | : public CodeGenFunction::OMPPrivateScope { | |||
3686 | CodeGenFunction &CGF; | |||
3687 | const OMPIteratorExpr *E = nullptr; | |||
3688 | SmallVector<CodeGenFunction::JumpDest, 4> ContDests; | |||
3689 | SmallVector<CodeGenFunction::JumpDest, 4> ExitDests; | |||
3690 | OMPIteratorGeneratorScope() = delete; | |||
3691 | OMPIteratorGeneratorScope(OMPIteratorGeneratorScope &) = delete; | |||
3692 | ||||
3693 | public: | |||
3694 | OMPIteratorGeneratorScope(CodeGenFunction &CGF, const OMPIteratorExpr *E) | |||
3695 | : CodeGenFunction::OMPPrivateScope(CGF), CGF(CGF), E(E) { | |||
3696 | if (!E) | |||
3697 | return; | |||
3698 | SmallVector<llvm::Value *, 4> Uppers; | |||
3699 | for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) { | |||
3700 | Uppers.push_back(CGF.EmitScalarExpr(E->getHelper(I).Upper)); | |||
3701 | const auto *VD = cast<VarDecl>(E->getIteratorDecl(I)); | |||
3702 | addPrivate(VD, CGF.CreateMemTemp(VD->getType(), VD->getName())); | |||
3703 | const OMPIteratorHelperData &HelperData = E->getHelper(I); | |||
3704 | addPrivate( | |||
3705 | HelperData.CounterVD, | |||
3706 | CGF.CreateMemTemp(HelperData.CounterVD->getType(), "counter.addr")); | |||
3707 | } | |||
3708 | Privatize(); | |||
3709 | ||||
3710 | for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) { | |||
3711 | const OMPIteratorHelperData &HelperData = E->getHelper(I); | |||
3712 | LValue CLVal = | |||
3713 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(HelperData.CounterVD), | |||
3714 | HelperData.CounterVD->getType()); | |||
3715 | // Counter = 0; | |||
3716 | CGF.EmitStoreOfScalar( | |||
3717 | llvm::ConstantInt::get(CLVal.getAddress(CGF).getElementType(), 0), | |||
3718 | CLVal); | |||
3719 | CodeGenFunction::JumpDest &ContDest = | |||
3720 | ContDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.cont")); | |||
3721 | CodeGenFunction::JumpDest &ExitDest = | |||
3722 | ExitDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.exit")); | |||
3723 | // N = <number-of_iterations>; | |||
3724 | llvm::Value *N = Uppers[I]; | |||
3725 | // cont: | |||
3726 | // if (Counter < N) goto body; else goto exit; | |||
3727 | CGF.EmitBlock(ContDest.getBlock()); | |||
3728 | auto *CVal = | |||
3729 | CGF.EmitLoadOfScalar(CLVal, HelperData.CounterVD->getLocation()); | |||
3730 | llvm::Value *Cmp = | |||
3731 | HelperData.CounterVD->getType()->isSignedIntegerOrEnumerationType() | |||
3732 | ? CGF.Builder.CreateICmpSLT(CVal, N) | |||
3733 | : CGF.Builder.CreateICmpULT(CVal, N); | |||
3734 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("iter.body"); | |||
3735 | CGF.Builder.CreateCondBr(Cmp, BodyBB, ExitDest.getBlock()); | |||
3736 | // body: | |||
3737 | CGF.EmitBlock(BodyBB); | |||
3738 | // Iteri = Begini + Counter * Stepi; | |||
3739 | CGF.EmitIgnoredExpr(HelperData.Update); | |||
3740 | } | |||
3741 | } | |||
3742 | ~OMPIteratorGeneratorScope() { | |||
3743 | if (!E) | |||
3744 | return; | |||
3745 | for (unsigned I = E->numOfIterators(); I > 0; --I) { | |||
3746 | // Counter = Counter + 1; | |||
3747 | const OMPIteratorHelperData &HelperData = E->getHelper(I - 1); | |||
3748 | CGF.EmitIgnoredExpr(HelperData.CounterUpdate); | |||
3749 | // goto cont; | |||
3750 | CGF.EmitBranchThroughCleanup(ContDests[I - 1]); | |||
3751 | // exit: | |||
3752 | CGF.EmitBlock(ExitDests[I - 1].getBlock(), /*IsFinished=*/I == 1); | |||
3753 | } | |||
3754 | } | |||
3755 | }; | |||
3756 | } // namespace | |||
3757 | ||||
3758 | static std::pair<llvm::Value *, llvm::Value *> | |||
3759 | getPointerAndSize(CodeGenFunction &CGF, const Expr *E) { | |||
3760 | const auto *OASE = dyn_cast<OMPArrayShapingExpr>(E); | |||
3761 | llvm::Value *Addr; | |||
3762 | if (OASE) { | |||
3763 | const Expr *Base = OASE->getBase(); | |||
3764 | Addr = CGF.EmitScalarExpr(Base); | |||
3765 | } else { | |||
3766 | Addr = CGF.EmitLValue(E).getPointer(CGF); | |||
3767 | } | |||
3768 | llvm::Value *SizeVal; | |||
3769 | QualType Ty = E->getType(); | |||
3770 | if (OASE) { | |||
3771 | SizeVal = CGF.getTypeSize(OASE->getBase()->getType()->getPointeeType()); | |||
3772 | for (const Expr *SE : OASE->getDimensions()) { | |||
3773 | llvm::Value *Sz = CGF.EmitScalarExpr(SE); | |||
3774 | Sz = CGF.EmitScalarConversion( | |||
3775 | Sz, SE->getType(), CGF.getContext().getSizeType(), SE->getExprLoc()); | |||
3776 | SizeVal = CGF.Builder.CreateNUWMul(SizeVal, Sz); | |||
3777 | } | |||
3778 | } else if (const auto *ASE = | |||
3779 | dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) { | |||
3780 | LValue UpAddrLVal = | |||
3781 | CGF.EmitOMPArraySectionExpr(ASE, /*IsLowerBound=*/false); | |||
3782 | Address UpAddrAddress = UpAddrLVal.getAddress(CGF); | |||
3783 | llvm::Value *UpAddr = CGF.Builder.CreateConstGEP1_32( | |||
3784 | UpAddrAddress.getElementType(), UpAddrAddress.getPointer(), /*Idx0=*/1); | |||
3785 | llvm::Value *LowIntPtr = CGF.Builder.CreatePtrToInt(Addr, CGF.SizeTy); | |||
3786 | llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGF.SizeTy); | |||
3787 | SizeVal = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr); | |||
3788 | } else { | |||
3789 | SizeVal = CGF.getTypeSize(Ty); | |||
3790 | } | |||
3791 | return std::make_pair(Addr, SizeVal); | |||
3792 | } | |||
3793 | ||||
3794 | /// Builds kmp_depend_info, if it is not built yet, and builds flags type. | |||
3795 | static void getKmpAffinityType(ASTContext &C, QualType &KmpTaskAffinityInfoTy) { | |||
3796 | QualType FlagsTy = C.getIntTypeForBitwidth(32, /*Signed=*/false); | |||
3797 | if (KmpTaskAffinityInfoTy.isNull()) { | |||
3798 | RecordDecl *KmpAffinityInfoRD = | |||
3799 | C.buildImplicitRecord("kmp_task_affinity_info_t"); | |||
3800 | KmpAffinityInfoRD->startDefinition(); | |||
3801 | addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getIntPtrType()); | |||
3802 | addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getSizeType()); | |||
3803 | addFieldToRecordDecl(C, KmpAffinityInfoRD, FlagsTy); | |||
3804 | KmpAffinityInfoRD->completeDefinition(); | |||
3805 | KmpTaskAffinityInfoTy = C.getRecordType(KmpAffinityInfoRD); | |||
3806 | } | |||
3807 | } | |||
3808 | ||||
3809 | CGOpenMPRuntime::TaskResultTy | |||
3810 | CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc, | |||
3811 | const OMPExecutableDirective &D, | |||
3812 | llvm::Function *TaskFunction, QualType SharedsTy, | |||
3813 | Address Shareds, const OMPTaskDataTy &Data) { | |||
3814 | ASTContext &C = CGM.getContext(); | |||
3815 | llvm::SmallVector<PrivateDataTy, 4> Privates; | |||
3816 | // Aggregate privates and sort them by the alignment. | |||
3817 | const auto *I = Data.PrivateCopies.begin(); | |||
3818 | for (const Expr *E : Data.PrivateVars) { | |||
3819 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
3820 | Privates.emplace_back( | |||
3821 | C.getDeclAlign(VD), | |||
3822 | PrivateHelpersTy(E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | |||
3823 | /*PrivateElemInit=*/nullptr)); | |||
3824 | ++I; | |||
3825 | } | |||
3826 | I = Data.FirstprivateCopies.begin(); | |||
3827 | const auto *IElemInitRef = Data.FirstprivateInits.begin(); | |||
3828 | for (const Expr *E : Data.FirstprivateVars) { | |||
3829 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
3830 | Privates.emplace_back( | |||
3831 | C.getDeclAlign(VD), | |||
3832 | PrivateHelpersTy( | |||
3833 | E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | |||
3834 | cast<VarDecl>(cast<DeclRefExpr>(*IElemInitRef)->getDecl()))); | |||
3835 | ++I; | |||
3836 | ++IElemInitRef; | |||
3837 | } | |||
3838 | I = Data.LastprivateCopies.begin(); | |||
3839 | for (const Expr *E : Data.LastprivateVars) { | |||
3840 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
3841 | Privates.emplace_back( | |||
3842 | C.getDeclAlign(VD), | |||
3843 | PrivateHelpersTy(E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | |||
3844 | /*PrivateElemInit=*/nullptr)); | |||
3845 | ++I; | |||
3846 | } | |||
3847 | for (const VarDecl *VD : Data.PrivateLocals) { | |||
3848 | if (isAllocatableDecl(VD)) | |||
3849 | Privates.emplace_back(CGM.getPointerAlign(), PrivateHelpersTy(VD)); | |||
3850 | else | |||
3851 | Privates.emplace_back(C.getDeclAlign(VD), PrivateHelpersTy(VD)); | |||
3852 | } | |||
3853 | llvm::stable_sort(Privates, | |||
3854 | [](const PrivateDataTy &L, const PrivateDataTy &R) { | |||
3855 | return L.first > R.first; | |||
3856 | }); | |||
3857 | QualType KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | |||
3858 | // Build type kmp_routine_entry_t (if not built yet). | |||
3859 | emitKmpRoutineEntryT(KmpInt32Ty); | |||
3860 | // Build type kmp_task_t (if not built yet). | |||
3861 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind())) { | |||
3862 | if (SavedKmpTaskloopTQTy.isNull()) { | |||
3863 | SavedKmpTaskloopTQTy = C.getRecordType(createKmpTaskTRecordDecl( | |||
3864 | CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy)); | |||
3865 | } | |||
3866 | KmpTaskTQTy = SavedKmpTaskloopTQTy; | |||
3867 | } else { | |||
3868 | assert((D.getDirectiveKind() == OMPD_task ||(static_cast <bool> ((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && "Expected taskloop, task or target directive") ? void (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3871, __extension__ __PRETTY_FUNCTION__)) | |||
3869 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()) ||(static_cast <bool> ((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && "Expected taskloop, task or target directive") ? void (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3871, __extension__ __PRETTY_FUNCTION__)) | |||
3870 | isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) &&(static_cast <bool> ((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && "Expected taskloop, task or target directive") ? void (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3871, __extension__ __PRETTY_FUNCTION__)) | |||
3871 | "Expected taskloop, task or target directive")(static_cast <bool> ((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && "Expected taskloop, task or target directive") ? void (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 3871, __extension__ __PRETTY_FUNCTION__)); | |||
3872 | if (SavedKmpTaskTQTy.isNull()) { | |||
3873 | SavedKmpTaskTQTy = C.getRecordType(createKmpTaskTRecordDecl( | |||
3874 | CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy)); | |||
3875 | } | |||
3876 | KmpTaskTQTy = SavedKmpTaskTQTy; | |||
3877 | } | |||
3878 | const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl()); | |||
3879 | // Build particular struct kmp_task_t for the given task. | |||
3880 | const RecordDecl *KmpTaskTWithPrivatesQTyRD = | |||
3881 | createKmpTaskTWithPrivatesRecordDecl(CGM, KmpTaskTQTy, Privates); | |||
3882 | QualType KmpTaskTWithPrivatesQTy = C.getRecordType(KmpTaskTWithPrivatesQTyRD); | |||
3883 | QualType KmpTaskTWithPrivatesPtrQTy = | |||
3884 | C.getPointerType(KmpTaskTWithPrivatesQTy); | |||
3885 | llvm::Type *KmpTaskTWithPrivatesTy = CGF.ConvertType(KmpTaskTWithPrivatesQTy); | |||
3886 | llvm::Type *KmpTaskTWithPrivatesPtrTy = | |||
3887 | KmpTaskTWithPrivatesTy->getPointerTo(); | |||
3888 | llvm::Value *KmpTaskTWithPrivatesTySize = | |||
3889 | CGF.getTypeSize(KmpTaskTWithPrivatesQTy); | |||
3890 | QualType SharedsPtrTy = C.getPointerType(SharedsTy); | |||
3891 | ||||
3892 | // Emit initial values for private copies (if any). | |||
3893 | llvm::Value *TaskPrivatesMap = nullptr; | |||
3894 | llvm::Type *TaskPrivatesMapTy = | |||
3895 | std::next(TaskFunction->arg_begin(), 3)->getType(); | |||
3896 | if (!Privates.empty()) { | |||
3897 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
3898 | TaskPrivatesMap = | |||
3899 | emitTaskPrivateMappingFunction(CGM, Loc, Data, FI->getType(), Privates); | |||
3900 | TaskPrivatesMap = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3901 | TaskPrivatesMap, TaskPrivatesMapTy); | |||
3902 | } else { | |||
3903 | TaskPrivatesMap = llvm::ConstantPointerNull::get( | |||
3904 | cast<llvm::PointerType>(TaskPrivatesMapTy)); | |||
3905 | } | |||
3906 | // Build a proxy function kmp_int32 .omp_task_entry.(kmp_int32 gtid, | |||
3907 | // kmp_task_t *tt); | |||
3908 | llvm::Function *TaskEntry = emitProxyTaskFunction( | |||
3909 | CGM, Loc, D.getDirectiveKind(), KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy, | |||
3910 | KmpTaskTWithPrivatesQTy, KmpTaskTQTy, SharedsPtrTy, TaskFunction, | |||
3911 | TaskPrivatesMap); | |||
3912 | ||||
3913 | // Build call kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, | |||
3914 | // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
3915 | // kmp_routine_entry_t *task_entry); | |||
3916 | // Task flags. Format is taken from | |||
3917 | // https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp.h, | |||
3918 | // description of kmp_tasking_flags struct. | |||
3919 | enum { | |||
3920 | TiedFlag = 0x1, | |||
3921 | FinalFlag = 0x2, | |||
3922 | DestructorsFlag = 0x8, | |||
3923 | PriorityFlag = 0x20, | |||
3924 | DetachableFlag = 0x40, | |||
3925 | }; | |||
3926 | unsigned Flags = Data.Tied ? TiedFlag : 0; | |||
3927 | bool NeedsCleanup = false; | |||
3928 | if (!Privates.empty()) { | |||
3929 | NeedsCleanup = | |||
3930 | checkDestructorsRequired(KmpTaskTWithPrivatesQTyRD, Privates); | |||
3931 | if (NeedsCleanup) | |||
3932 | Flags = Flags | DestructorsFlag; | |||
3933 | } | |||
3934 | if (Data.Priority.getInt()) | |||
3935 | Flags = Flags | PriorityFlag; | |||
3936 | if (D.hasClausesOfKind<OMPDetachClause>()) | |||
3937 | Flags = Flags | DetachableFlag; | |||
3938 | llvm::Value *TaskFlags = | |||
3939 | Data.Final.getPointer() | |||
3940 | ? CGF.Builder.CreateSelect(Data.Final.getPointer(), | |||
3941 | CGF.Builder.getInt32(FinalFlag), | |||
3942 | CGF.Builder.getInt32(/*C=*/0)) | |||
3943 | : CGF.Builder.getInt32(Data.Final.getInt() ? FinalFlag : 0); | |||
3944 | TaskFlags = CGF.Builder.CreateOr(TaskFlags, CGF.Builder.getInt32(Flags)); | |||
3945 | llvm::Value *SharedsSize = CGM.getSize(C.getTypeSizeInChars(SharedsTy)); | |||
3946 | SmallVector<llvm::Value *, 8> AllocArgs = {emitUpdateLocation(CGF, Loc), | |||
3947 | getThreadID(CGF, Loc), TaskFlags, KmpTaskTWithPrivatesTySize, | |||
3948 | SharedsSize, CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3949 | TaskEntry, KmpRoutineEntryPtrTy)}; | |||
3950 | llvm::Value *NewTask; | |||
3951 | if (D.hasClausesOfKind<OMPNowaitClause>()) { | |||
3952 | // Check if we have any device clause associated with the directive. | |||
3953 | const Expr *Device = nullptr; | |||
3954 | if (auto *C = D.getSingleClause<OMPDeviceClause>()) | |||
3955 | Device = C->getDevice(); | |||
3956 | // Emit device ID if any otherwise use default value. | |||
3957 | llvm::Value *DeviceID; | |||
3958 | if (Device) | |||
3959 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
3960 | CGF.Int64Ty, /*isSigned=*/true); | |||
3961 | else | |||
3962 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
3963 | AllocArgs.push_back(DeviceID); | |||
3964 | NewTask = CGF.EmitRuntimeCall( | |||
3965 | OMPBuilder.getOrCreateRuntimeFunction( | |||
3966 | CGM.getModule(), OMPRTL___kmpc_omp_target_task_alloc), | |||
3967 | AllocArgs); | |||
3968 | } else { | |||
3969 | NewTask = | |||
3970 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
3971 | CGM.getModule(), OMPRTL___kmpc_omp_task_alloc), | |||
3972 | AllocArgs); | |||
3973 | } | |||
3974 | // Emit detach clause initialization. | |||
3975 | // evt = (typeof(evt))__kmpc_task_allow_completion_event(loc, tid, | |||
3976 | // task_descriptor); | |||
3977 | if (const auto *DC = D.getSingleClause<OMPDetachClause>()) { | |||
3978 | const Expr *Evt = DC->getEventHandler()->IgnoreParenImpCasts(); | |||
3979 | LValue EvtLVal = CGF.EmitLValue(Evt); | |||
3980 | ||||
3981 | // Build kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, | |||
3982 | // int gtid, kmp_task_t *task); | |||
3983 | llvm::Value *Loc = emitUpdateLocation(CGF, DC->getBeginLoc()); | |||
3984 | llvm::Value *Tid = getThreadID(CGF, DC->getBeginLoc()); | |||
3985 | Tid = CGF.Builder.CreateIntCast(Tid, CGF.IntTy, /*isSigned=*/false); | |||
3986 | llvm::Value *EvtVal = CGF.EmitRuntimeCall( | |||
3987 | OMPBuilder.getOrCreateRuntimeFunction( | |||
3988 | CGM.getModule(), OMPRTL___kmpc_task_allow_completion_event), | |||
3989 | {Loc, Tid, NewTask}); | |||
3990 | EvtVal = CGF.EmitScalarConversion(EvtVal, C.VoidPtrTy, Evt->getType(), | |||
3991 | Evt->getExprLoc()); | |||
3992 | CGF.EmitStoreOfScalar(EvtVal, EvtLVal); | |||
3993 | } | |||
3994 | // Process affinity clauses. | |||
3995 | if (D.hasClausesOfKind<OMPAffinityClause>()) { | |||
3996 | // Process list of affinity data. | |||
3997 | ASTContext &C = CGM.getContext(); | |||
3998 | Address AffinitiesArray = Address::invalid(); | |||
3999 | // Calculate number of elements to form the array of affinity data. | |||
4000 | llvm::Value *NumOfElements = nullptr; | |||
4001 | unsigned NumAffinities = 0; | |||
4002 | for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) { | |||
4003 | if (const Expr *Modifier = C->getModifier()) { | |||
4004 | const auto *IE = cast<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts()); | |||
4005 | for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) { | |||
4006 | llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper); | |||
4007 | Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false); | |||
4008 | NumOfElements = | |||
4009 | NumOfElements ? CGF.Builder.CreateNUWMul(NumOfElements, Sz) : Sz; | |||
4010 | } | |||
4011 | } else { | |||
4012 | NumAffinities += C->varlist_size(); | |||
4013 | } | |||
4014 | } | |||
4015 | getKmpAffinityType(CGM.getContext(), KmpTaskAffinityInfoTy); | |||
4016 | // Fields ids in kmp_task_affinity_info record. | |||
4017 | enum RTLAffinityInfoFieldsTy { BaseAddr, Len, Flags }; | |||
4018 | ||||
4019 | QualType KmpTaskAffinityInfoArrayTy; | |||
4020 | if (NumOfElements) { | |||
4021 | NumOfElements = CGF.Builder.CreateNUWAdd( | |||
4022 | llvm::ConstantInt::get(CGF.SizeTy, NumAffinities), NumOfElements); | |||
4023 | auto *OVE = new (C) OpaqueValueExpr( | |||
4024 | Loc, | |||
4025 | C.getIntTypeForBitwidth(C.getTypeSize(C.getSizeType()), /*Signed=*/0), | |||
4026 | VK_PRValue); | |||
4027 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, OVE, | |||
4028 | RValue::get(NumOfElements)); | |||
4029 | KmpTaskAffinityInfoArrayTy = | |||
4030 | C.getVariableArrayType(KmpTaskAffinityInfoTy, OVE, ArrayType::Normal, | |||
4031 | /*IndexTypeQuals=*/0, SourceRange(Loc, Loc)); | |||
4032 | // Properly emit variable-sized array. | |||
4033 | auto *PD = ImplicitParamDecl::Create(C, KmpTaskAffinityInfoArrayTy, | |||
4034 | ImplicitParamDecl::Other); | |||
4035 | CGF.EmitVarDecl(*PD); | |||
4036 | AffinitiesArray = CGF.GetAddrOfLocalVar(PD); | |||
4037 | NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty, | |||
4038 | /*isSigned=*/false); | |||
4039 | } else { | |||
4040 | KmpTaskAffinityInfoArrayTy = C.getConstantArrayType( | |||
4041 | KmpTaskAffinityInfoTy, | |||
4042 | llvm::APInt(C.getTypeSize(C.getSizeType()), NumAffinities), nullptr, | |||
4043 | ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
4044 | AffinitiesArray = | |||
4045 | CGF.CreateMemTemp(KmpTaskAffinityInfoArrayTy, ".affs.arr.addr"); | |||
4046 | AffinitiesArray = CGF.Builder.CreateConstArrayGEP(AffinitiesArray, 0); | |||
4047 | NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumAffinities, | |||
4048 | /*isSigned=*/false); | |||
4049 | } | |||
4050 | ||||
4051 | const auto *KmpAffinityInfoRD = KmpTaskAffinityInfoTy->getAsRecordDecl(); | |||
4052 | // Fill array by elements without iterators. | |||
4053 | unsigned Pos = 0; | |||
4054 | bool HasIterator = false; | |||
4055 | for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) { | |||
4056 | if (C->getModifier()) { | |||
4057 | HasIterator = true; | |||
4058 | continue; | |||
4059 | } | |||
4060 | for (const Expr *E : C->varlists()) { | |||
4061 | llvm::Value *Addr; | |||
4062 | llvm::Value *Size; | |||
4063 | std::tie(Addr, Size) = getPointerAndSize(CGF, E); | |||
4064 | LValue Base = | |||
4065 | CGF.MakeAddrLValue(CGF.Builder.CreateConstGEP(AffinitiesArray, Pos), | |||
4066 | KmpTaskAffinityInfoTy); | |||
4067 | // affs[i].base_addr = &<Affinities[i].second>; | |||
4068 | LValue BaseAddrLVal = CGF.EmitLValueForField( | |||
4069 | Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr)); | |||
4070 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy), | |||
4071 | BaseAddrLVal); | |||
4072 | // affs[i].len = sizeof(<Affinities[i].second>); | |||
4073 | LValue LenLVal = CGF.EmitLValueForField( | |||
4074 | Base, *std::next(KmpAffinityInfoRD->field_begin(), Len)); | |||
4075 | CGF.EmitStoreOfScalar(Size, LenLVal); | |||
4076 | ++Pos; | |||
4077 | } | |||
4078 | } | |||
4079 | LValue PosLVal; | |||
4080 | if (HasIterator) { | |||
4081 | PosLVal = CGF.MakeAddrLValue( | |||
4082 | CGF.CreateMemTemp(C.getSizeType(), "affs.counter.addr"), | |||
4083 | C.getSizeType()); | |||
4084 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal); | |||
4085 | } | |||
4086 | // Process elements with iterators. | |||
4087 | for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) { | |||
4088 | const Expr *Modifier = C->getModifier(); | |||
4089 | if (!Modifier) | |||
4090 | continue; | |||
4091 | OMPIteratorGeneratorScope IteratorScope( | |||
4092 | CGF, cast_or_null<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts())); | |||
4093 | for (const Expr *E : C->varlists()) { | |||
4094 | llvm::Value *Addr; | |||
4095 | llvm::Value *Size; | |||
4096 | std::tie(Addr, Size) = getPointerAndSize(CGF, E); | |||
4097 | llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | |||
4098 | LValue Base = CGF.MakeAddrLValue( | |||
4099 | CGF.Builder.CreateGEP(AffinitiesArray, Idx), KmpTaskAffinityInfoTy); | |||
4100 | // affs[i].base_addr = &<Affinities[i].second>; | |||
4101 | LValue BaseAddrLVal = CGF.EmitLValueForField( | |||
4102 | Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr)); | |||
4103 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy), | |||
4104 | BaseAddrLVal); | |||
4105 | // affs[i].len = sizeof(<Affinities[i].second>); | |||
4106 | LValue LenLVal = CGF.EmitLValueForField( | |||
4107 | Base, *std::next(KmpAffinityInfoRD->field_begin(), Len)); | |||
4108 | CGF.EmitStoreOfScalar(Size, LenLVal); | |||
4109 | Idx = CGF.Builder.CreateNUWAdd( | |||
4110 | Idx, llvm::ConstantInt::get(Idx->getType(), 1)); | |||
4111 | CGF.EmitStoreOfScalar(Idx, PosLVal); | |||
4112 | } | |||
4113 | } | |||
4114 | // Call to kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, | |||
4115 | // kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 | |||
4116 | // naffins, kmp_task_affinity_info_t *affin_list); | |||
4117 | llvm::Value *LocRef = emitUpdateLocation(CGF, Loc); | |||
4118 | llvm::Value *GTid = getThreadID(CGF, Loc); | |||
4119 | llvm::Value *AffinListPtr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4120 | AffinitiesArray.getPointer(), CGM.VoidPtrTy); | |||
4121 | // FIXME: Emit the function and ignore its result for now unless the | |||
4122 | // runtime function is properly implemented. | |||
4123 | (void)CGF.EmitRuntimeCall( | |||
4124 | OMPBuilder.getOrCreateRuntimeFunction( | |||
4125 | CGM.getModule(), OMPRTL___kmpc_omp_reg_task_with_affinity), | |||
4126 | {LocRef, GTid, NewTask, NumOfElements, AffinListPtr}); | |||
4127 | } | |||
4128 | llvm::Value *NewTaskNewTaskTTy = | |||
4129 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4130 | NewTask, KmpTaskTWithPrivatesPtrTy); | |||
4131 | LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy, | |||
4132 | KmpTaskTWithPrivatesQTy); | |||
4133 | LValue TDBase = | |||
4134 | CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
4135 | // Fill the data in the resulting kmp_task_t record. | |||
4136 | // Copy shareds if there are any. | |||
4137 | Address KmpTaskSharedsPtr = Address::invalid(); | |||
4138 | if (!SharedsTy->getAsStructureType()->getDecl()->field_empty()) { | |||
4139 | KmpTaskSharedsPtr = Address( | |||
4140 | CGF.EmitLoadOfScalar( | |||
4141 | CGF.EmitLValueForField( | |||
4142 | TDBase, | |||
4143 | *std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds)), | |||
4144 | Loc), | |||
4145 | CGF.Int8Ty, CGM.getNaturalTypeAlignment(SharedsTy)); | |||
4146 | LValue Dest = CGF.MakeAddrLValue(KmpTaskSharedsPtr, SharedsTy); | |||
4147 | LValue Src = CGF.MakeAddrLValue(Shareds, SharedsTy); | |||
4148 | CGF.EmitAggregateCopy(Dest, Src, SharedsTy, AggValueSlot::DoesNotOverlap); | |||
4149 | } | |||
4150 | // Emit initial values for private copies (if any). | |||
4151 | TaskResultTy Result; | |||
4152 | if (!Privates.empty()) { | |||
4153 | emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, Base, KmpTaskTWithPrivatesQTyRD, | |||
4154 | SharedsTy, SharedsPtrTy, Data, Privates, | |||
4155 | /*ForDup=*/false); | |||
4156 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) && | |||
4157 | (!Data.LastprivateVars.empty() || checkInitIsRequired(CGF, Privates))) { | |||
4158 | Result.TaskDupFn = emitTaskDupFunction( | |||
4159 | CGM, Loc, D, KmpTaskTWithPrivatesPtrQTy, KmpTaskTWithPrivatesQTyRD, | |||
4160 | KmpTaskTQTyRD, SharedsTy, SharedsPtrTy, Data, Privates, | |||
4161 | /*WithLastIter=*/!Data.LastprivateVars.empty()); | |||
4162 | } | |||
4163 | } | |||
4164 | // Fields of union "kmp_cmplrdata_t" for destructors and priority. | |||
4165 | enum { Priority = 0, Destructors = 1 }; | |||
4166 | // Provide pointer to function with destructors for privates. | |||
4167 | auto FI = std::next(KmpTaskTQTyRD->field_begin(), Data1); | |||
4168 | const RecordDecl *KmpCmplrdataUD = | |||
4169 | (*FI)->getType()->getAsUnionType()->getDecl(); | |||
4170 | if (NeedsCleanup) { | |||
4171 | llvm::Value *DestructorFn = emitDestructorsFunction( | |||
4172 | CGM, Loc, KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy, | |||
4173 | KmpTaskTWithPrivatesQTy); | |||
4174 | LValue Data1LV = CGF.EmitLValueForField(TDBase, *FI); | |||
4175 | LValue DestructorsLV = CGF.EmitLValueForField( | |||
4176 | Data1LV, *std::next(KmpCmplrdataUD->field_begin(), Destructors)); | |||
4177 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4178 | DestructorFn, KmpRoutineEntryPtrTy), | |||
4179 | DestructorsLV); | |||
4180 | } | |||
4181 | // Set priority. | |||
4182 | if (Data.Priority.getInt()) { | |||
4183 | LValue Data2LV = CGF.EmitLValueForField( | |||
4184 | TDBase, *std::next(KmpTaskTQTyRD->field_begin(), Data2)); | |||
4185 | LValue PriorityLV = CGF.EmitLValueForField( | |||
4186 | Data2LV, *std::next(KmpCmplrdataUD->field_begin(), Priority)); | |||
4187 | CGF.EmitStoreOfScalar(Data.Priority.getPointer(), PriorityLV); | |||
4188 | } | |||
4189 | Result.NewTask = NewTask; | |||
4190 | Result.TaskEntry = TaskEntry; | |||
4191 | Result.NewTaskNewTaskTTy = NewTaskNewTaskTTy; | |||
4192 | Result.TDBase = TDBase; | |||
4193 | Result.KmpTaskTQTyRD = KmpTaskTQTyRD; | |||
4194 | return Result; | |||
4195 | } | |||
4196 | ||||
4197 | /// Translates internal dependency kind into the runtime kind. | |||
4198 | static RTLDependenceKindTy translateDependencyKind(OpenMPDependClauseKind K) { | |||
4199 | RTLDependenceKindTy DepKind; | |||
4200 | switch (K) { | |||
4201 | case OMPC_DEPEND_in: | |||
4202 | DepKind = RTLDependenceKindTy::DepIn; | |||
4203 | break; | |||
4204 | // Out and InOut dependencies must use the same code. | |||
4205 | case OMPC_DEPEND_out: | |||
4206 | case OMPC_DEPEND_inout: | |||
4207 | DepKind = RTLDependenceKindTy::DepInOut; | |||
4208 | break; | |||
4209 | case OMPC_DEPEND_mutexinoutset: | |||
4210 | DepKind = RTLDependenceKindTy::DepMutexInOutSet; | |||
4211 | break; | |||
4212 | case OMPC_DEPEND_inoutset: | |||
4213 | DepKind = RTLDependenceKindTy::DepInOutSet; | |||
4214 | break; | |||
4215 | case OMPC_DEPEND_outallmemory: | |||
4216 | DepKind = RTLDependenceKindTy::DepOmpAllMem; | |||
4217 | break; | |||
4218 | case OMPC_DEPEND_source: | |||
4219 | case OMPC_DEPEND_sink: | |||
4220 | case OMPC_DEPEND_depobj: | |||
4221 | case OMPC_DEPEND_inoutallmemory: | |||
4222 | case OMPC_DEPEND_unknown: | |||
4223 | llvm_unreachable("Unknown task dependence type")::llvm::llvm_unreachable_internal("Unknown task dependence type" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 4223); | |||
4224 | } | |||
4225 | return DepKind; | |||
4226 | } | |||
4227 | ||||
4228 | /// Builds kmp_depend_info, if it is not built yet, and builds flags type. | |||
4229 | static void getDependTypes(ASTContext &C, QualType &KmpDependInfoTy, | |||
4230 | QualType &FlagsTy) { | |||
4231 | FlagsTy = C.getIntTypeForBitwidth(C.getTypeSize(C.BoolTy), /*Signed=*/false); | |||
4232 | if (KmpDependInfoTy.isNull()) { | |||
4233 | RecordDecl *KmpDependInfoRD = C.buildImplicitRecord("kmp_depend_info"); | |||
4234 | KmpDependInfoRD->startDefinition(); | |||
4235 | addFieldToRecordDecl(C, KmpDependInfoRD, C.getIntPtrType()); | |||
4236 | addFieldToRecordDecl(C, KmpDependInfoRD, C.getSizeType()); | |||
4237 | addFieldToRecordDecl(C, KmpDependInfoRD, FlagsTy); | |||
4238 | KmpDependInfoRD->completeDefinition(); | |||
4239 | KmpDependInfoTy = C.getRecordType(KmpDependInfoRD); | |||
4240 | } | |||
4241 | } | |||
4242 | ||||
4243 | std::pair<llvm::Value *, LValue> | |||
4244 | CGOpenMPRuntime::getDepobjElements(CodeGenFunction &CGF, LValue DepobjLVal, | |||
4245 | SourceLocation Loc) { | |||
4246 | ASTContext &C = CGM.getContext(); | |||
4247 | QualType FlagsTy; | |||
4248 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | |||
4249 | RecordDecl *KmpDependInfoRD = | |||
4250 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | |||
4251 | QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy); | |||
4252 | LValue Base = CGF.EmitLoadOfPointerLValue( | |||
4253 | CGF.Builder.CreateElementBitCast( | |||
4254 | DepobjLVal.getAddress(CGF), | |||
4255 | CGF.ConvertTypeForMem(KmpDependInfoPtrTy)), | |||
4256 | KmpDependInfoPtrTy->castAs<PointerType>()); | |||
4257 | Address DepObjAddr = CGF.Builder.CreateGEP( | |||
4258 | Base.getAddress(CGF), | |||
4259 | llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true)); | |||
4260 | LValue NumDepsBase = CGF.MakeAddrLValue( | |||
4261 | DepObjAddr, KmpDependInfoTy, Base.getBaseInfo(), Base.getTBAAInfo()); | |||
4262 | // NumDeps = deps[i].base_addr; | |||
4263 | LValue BaseAddrLVal = CGF.EmitLValueForField( | |||
4264 | NumDepsBase, | |||
4265 | *std::next(KmpDependInfoRD->field_begin(), | |||
4266 | static_cast<unsigned int>(RTLDependInfoFields::BaseAddr))); | |||
4267 | llvm::Value *NumDeps = CGF.EmitLoadOfScalar(BaseAddrLVal, Loc); | |||
4268 | return std::make_pair(NumDeps, Base); | |||
4269 | } | |||
4270 | ||||
4271 | static void emitDependData(CodeGenFunction &CGF, QualType &KmpDependInfoTy, | |||
4272 | llvm::PointerUnion<unsigned *, LValue *> Pos, | |||
4273 | const OMPTaskDataTy::DependData &Data, | |||
4274 | Address DependenciesArray) { | |||
4275 | CodeGenModule &CGM = CGF.CGM; | |||
4276 | ASTContext &C = CGM.getContext(); | |||
4277 | QualType FlagsTy; | |||
4278 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | |||
4279 | RecordDecl *KmpDependInfoRD = | |||
4280 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | |||
4281 | llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy); | |||
4282 | ||||
4283 | OMPIteratorGeneratorScope IteratorScope( | |||
4284 | CGF, cast_or_null<OMPIteratorExpr>( | |||
4285 | Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts() | |||
4286 | : nullptr)); | |||
4287 | for (const Expr *E : Data.DepExprs) { | |||
4288 | llvm::Value *Addr; | |||
4289 | llvm::Value *Size; | |||
4290 | ||||
4291 | // The expression will be a nullptr in the 'omp_all_memory' case. | |||
4292 | if (E) { | |||
4293 | std::tie(Addr, Size) = getPointerAndSize(CGF, E); | |||
4294 | Addr = CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy); | |||
4295 | } else { | |||
4296 | Addr = llvm::ConstantInt::get(CGF.IntPtrTy, 0); | |||
4297 | Size = llvm::ConstantInt::get(CGF.SizeTy, 0); | |||
4298 | } | |||
4299 | LValue Base; | |||
4300 | if (unsigned *P = Pos.dyn_cast<unsigned *>()) { | |||
4301 | Base = CGF.MakeAddrLValue( | |||
4302 | CGF.Builder.CreateConstGEP(DependenciesArray, *P), KmpDependInfoTy); | |||
4303 | } else { | |||
4304 | assert(E && "Expected a non-null expression")(static_cast <bool> (E && "Expected a non-null expression" ) ? void (0) : __assert_fail ("E && \"Expected a non-null expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 4304, __extension__ __PRETTY_FUNCTION__)); | |||
4305 | LValue &PosLVal = *Pos.get<LValue *>(); | |||
4306 | llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | |||
4307 | Base = CGF.MakeAddrLValue( | |||
4308 | CGF.Builder.CreateGEP(DependenciesArray, Idx), KmpDependInfoTy); | |||
4309 | } | |||
4310 | // deps[i].base_addr = &<Dependencies[i].second>; | |||
4311 | LValue BaseAddrLVal = CGF.EmitLValueForField( | |||
4312 | Base, | |||
4313 | *std::next(KmpDependInfoRD->field_begin(), | |||
4314 | static_cast<unsigned int>(RTLDependInfoFields::BaseAddr))); | |||
4315 | CGF.EmitStoreOfScalar(Addr, BaseAddrLVal); | |||
4316 | // deps[i].len = sizeof(<Dependencies[i].second>); | |||
4317 | LValue LenLVal = CGF.EmitLValueForField( | |||
4318 | Base, *std::next(KmpDependInfoRD->field_begin(), | |||
4319 | static_cast<unsigned int>(RTLDependInfoFields::Len))); | |||
4320 | CGF.EmitStoreOfScalar(Size, LenLVal); | |||
4321 | // deps[i].flags = <Dependencies[i].first>; | |||
4322 | RTLDependenceKindTy DepKind = translateDependencyKind(Data.DepKind); | |||
4323 | LValue FlagsLVal = CGF.EmitLValueForField( | |||
4324 | Base, | |||
4325 | *std::next(KmpDependInfoRD->field_begin(), | |||
4326 | static_cast<unsigned int>(RTLDependInfoFields::Flags))); | |||
4327 | CGF.EmitStoreOfScalar( | |||
4328 | llvm::ConstantInt::get(LLVMFlagsTy, static_cast<unsigned int>(DepKind)), | |||
4329 | FlagsLVal); | |||
4330 | if (unsigned *P = Pos.dyn_cast<unsigned *>()) { | |||
4331 | ++(*P); | |||
4332 | } else { | |||
4333 | LValue &PosLVal = *Pos.get<LValue *>(); | |||
4334 | llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | |||
4335 | Idx = CGF.Builder.CreateNUWAdd(Idx, | |||
4336 | llvm::ConstantInt::get(Idx->getType(), 1)); | |||
4337 | CGF.EmitStoreOfScalar(Idx, PosLVal); | |||
4338 | } | |||
4339 | } | |||
4340 | } | |||
4341 | ||||
4342 | SmallVector<llvm::Value *, 4> CGOpenMPRuntime::emitDepobjElementsSizes( | |||
4343 | CodeGenFunction &CGF, QualType &KmpDependInfoTy, | |||
4344 | const OMPTaskDataTy::DependData &Data) { | |||
4345 | assert(Data.DepKind == OMPC_DEPEND_depobj &&(static_cast <bool> (Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependency kind.") ? void (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependency kind.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 4346, __extension__ __PRETTY_FUNCTION__)) | |||
4346 | "Expected depobj dependency kind.")(static_cast <bool> (Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependency kind.") ? void (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependency kind.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 4346, __extension__ __PRETTY_FUNCTION__)); | |||
4347 | SmallVector<llvm::Value *, 4> Sizes; | |||
4348 | SmallVector<LValue, 4> SizeLVals; | |||
4349 | ASTContext &C = CGF.getContext(); | |||
4350 | { | |||
4351 | OMPIteratorGeneratorScope IteratorScope( | |||
4352 | CGF, cast_or_null<OMPIteratorExpr>( | |||
4353 | Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts() | |||
4354 | : nullptr)); | |||
4355 | for (const Expr *E : Data.DepExprs) { | |||
4356 | llvm::Value *NumDeps; | |||
4357 | LValue Base; | |||
4358 | LValue DepobjLVal = CGF.EmitLValue(E->IgnoreParenImpCasts()); | |||
4359 | std::tie(NumDeps, Base) = | |||
4360 | getDepobjElements(CGF, DepobjLVal, E->getExprLoc()); | |||
4361 | LValue NumLVal = CGF.MakeAddrLValue( | |||
4362 | CGF.CreateMemTemp(C.getUIntPtrType(), "depobj.size.addr"), | |||
4363 | C.getUIntPtrType()); | |||
4364 | CGF.Builder.CreateStore(llvm::ConstantInt::get(CGF.IntPtrTy, 0), | |||
4365 | NumLVal.getAddress(CGF)); | |||
4366 | llvm::Value *PrevVal = CGF.EmitLoadOfScalar(NumLVal, E->getExprLoc()); | |||
4367 | llvm::Value *Add = CGF.Builder.CreateNUWAdd(PrevVal, NumDeps); | |||
4368 | CGF.EmitStoreOfScalar(Add, NumLVal); | |||
4369 | SizeLVals.push_back(NumLVal); | |||
4370 | } | |||
4371 | } | |||
4372 | for (unsigned I = 0, E = SizeLVals.size(); I < E; ++I) { | |||
4373 | llvm::Value *Size = | |||
4374 | CGF.EmitLoadOfScalar(SizeLVals[I], Data.DepExprs[I]->getExprLoc()); | |||
4375 | Sizes.push_back(Size); | |||
4376 | } | |||
4377 | return Sizes; | |||
4378 | } | |||
4379 | ||||
4380 | void CGOpenMPRuntime::emitDepobjElements(CodeGenFunction &CGF, | |||
4381 | QualType &KmpDependInfoTy, | |||
4382 | LValue PosLVal, | |||
4383 | const OMPTaskDataTy::DependData &Data, | |||
4384 | Address DependenciesArray) { | |||
4385 | assert(Data.DepKind == OMPC_DEPEND_depobj &&(static_cast <bool> (Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependency kind.") ? void (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependency kind.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 4386, __extension__ __PRETTY_FUNCTION__)) | |||
4386 | "Expected depobj dependency kind.")(static_cast <bool> (Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependency kind.") ? void (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependency kind.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 4386, __extension__ __PRETTY_FUNCTION__)); | |||
4387 | llvm::Value *ElSize = CGF.getTypeSize(KmpDependInfoTy); | |||
4388 | { | |||
4389 | OMPIteratorGeneratorScope IteratorScope( | |||
4390 | CGF, cast_or_null<OMPIteratorExpr>( | |||
4391 | Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts() | |||
4392 | : nullptr)); | |||
4393 | for (unsigned I = 0, End = Data.DepExprs.size(); I < End; ++I) { | |||
4394 | const Expr *E = Data.DepExprs[I]; | |||
4395 | llvm::Value *NumDeps; | |||
4396 | LValue Base; | |||
4397 | LValue DepobjLVal = CGF.EmitLValue(E->IgnoreParenImpCasts()); | |||
4398 | std::tie(NumDeps, Base) = | |||
4399 | getDepobjElements(CGF, DepobjLVal, E->getExprLoc()); | |||
4400 | ||||
4401 | // memcopy dependency data. | |||
4402 | llvm::Value *Size = CGF.Builder.CreateNUWMul( | |||
4403 | ElSize, | |||
4404 | CGF.Builder.CreateIntCast(NumDeps, CGF.SizeTy, /*isSigned=*/false)); | |||
4405 | llvm::Value *Pos = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | |||
4406 | Address DepAddr = CGF.Builder.CreateGEP(DependenciesArray, Pos); | |||
4407 | CGF.Builder.CreateMemCpy(DepAddr, Base.getAddress(CGF), Size); | |||
4408 | ||||
4409 | // Increase pos. | |||
4410 | // pos += size; | |||
4411 | llvm::Value *Add = CGF.Builder.CreateNUWAdd(Pos, NumDeps); | |||
4412 | CGF.EmitStoreOfScalar(Add, PosLVal); | |||
4413 | } | |||
4414 | } | |||
4415 | } | |||
4416 | ||||
4417 | std::pair<llvm::Value *, Address> CGOpenMPRuntime::emitDependClause( | |||
4418 | CodeGenFunction &CGF, ArrayRef<OMPTaskDataTy::DependData> Dependencies, | |||
4419 | SourceLocation Loc) { | |||
4420 | if (llvm::all_of(Dependencies, [](const OMPTaskDataTy::DependData &D) { | |||
4421 | return D.DepExprs.empty(); | |||
4422 | })) | |||
4423 | return std::make_pair(nullptr, Address::invalid()); | |||
4424 | // Process list of dependencies. | |||
4425 | ASTContext &C = CGM.getContext(); | |||
4426 | Address DependenciesArray = Address::invalid(); | |||
4427 | llvm::Value *NumOfElements = nullptr; | |||
4428 | unsigned NumDependencies = std::accumulate( | |||
4429 | Dependencies.begin(), Dependencies.end(), 0, | |||
4430 | [](unsigned V, const OMPTaskDataTy::DependData &D) { | |||
4431 | return D.DepKind == OMPC_DEPEND_depobj | |||
4432 | ? V | |||
4433 | : (V + (D.IteratorExpr ? 0 : D.DepExprs.size())); | |||
4434 | }); | |||
4435 | QualType FlagsTy; | |||
4436 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | |||
4437 | bool HasDepobjDeps = false; | |||
4438 | bool HasRegularWithIterators = false; | |||
4439 | llvm::Value *NumOfDepobjElements = llvm::ConstantInt::get(CGF.IntPtrTy, 0); | |||
4440 | llvm::Value *NumOfRegularWithIterators = | |||
4441 | llvm::ConstantInt::get(CGF.IntPtrTy, 0); | |||
4442 | // Calculate number of depobj dependencies and regular deps with the | |||
4443 | // iterators. | |||
4444 | for (const OMPTaskDataTy::DependData &D : Dependencies) { | |||
4445 | if (D.DepKind == OMPC_DEPEND_depobj) { | |||
4446 | SmallVector<llvm::Value *, 4> Sizes = | |||
4447 | emitDepobjElementsSizes(CGF, KmpDependInfoTy, D); | |||
4448 | for (llvm::Value *Size : Sizes) { | |||
4449 | NumOfDepobjElements = | |||
4450 | CGF.Builder.CreateNUWAdd(NumOfDepobjElements, Size); | |||
4451 | } | |||
4452 | HasDepobjDeps = true; | |||
4453 | continue; | |||
4454 | } | |||
4455 | // Include number of iterations, if any. | |||
4456 | ||||
4457 | if (const auto *IE = cast_or_null<OMPIteratorExpr>(D.IteratorExpr)) { | |||
4458 | for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) { | |||
4459 | llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper); | |||
4460 | Sz = CGF.Builder.CreateIntCast(Sz, CGF.IntPtrTy, /*isSigned=*/false); | |||
4461 | llvm::Value *NumClauseDeps = CGF.Builder.CreateNUWMul( | |||
4462 | Sz, llvm::ConstantInt::get(CGF.IntPtrTy, D.DepExprs.size())); | |||
4463 | NumOfRegularWithIterators = | |||
4464 | CGF.Builder.CreateNUWAdd(NumOfRegularWithIterators, NumClauseDeps); | |||
4465 | } | |||
4466 | HasRegularWithIterators = true; | |||
4467 | continue; | |||
4468 | } | |||
4469 | } | |||
4470 | ||||
4471 | QualType KmpDependInfoArrayTy; | |||
4472 | if (HasDepobjDeps || HasRegularWithIterators) { | |||
4473 | NumOfElements = llvm::ConstantInt::get(CGM.IntPtrTy, NumDependencies, | |||
4474 | /*isSigned=*/false); | |||
4475 | if (HasDepobjDeps) { | |||
4476 | NumOfElements = | |||
4477 | CGF.Builder.CreateNUWAdd(NumOfDepobjElements, NumOfElements); | |||
4478 | } | |||
4479 | if (HasRegularWithIterators) { | |||
4480 | NumOfElements = | |||
4481 | CGF.Builder.CreateNUWAdd(NumOfRegularWithIterators, NumOfElements); | |||
4482 | } | |||
4483 | auto *OVE = new (C) OpaqueValueExpr( | |||
4484 | Loc, C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0), | |||
4485 | VK_PRValue); | |||
4486 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, OVE, | |||
4487 | RValue::get(NumOfElements)); | |||
4488 | KmpDependInfoArrayTy = | |||
4489 | C.getVariableArrayType(KmpDependInfoTy, OVE, ArrayType::Normal, | |||
4490 | /*IndexTypeQuals=*/0, SourceRange(Loc, Loc)); | |||
4491 | // CGF.EmitVariablyModifiedType(KmpDependInfoArrayTy); | |||
4492 | // Properly emit variable-sized array. | |||
4493 | auto *PD = ImplicitParamDecl::Create(C, KmpDependInfoArrayTy, | |||
4494 | ImplicitParamDecl::Other); | |||
4495 | CGF.EmitVarDecl(*PD); | |||
4496 | DependenciesArray = CGF.GetAddrOfLocalVar(PD); | |||
4497 | NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty, | |||
4498 | /*isSigned=*/false); | |||
4499 | } else { | |||
4500 | KmpDependInfoArrayTy = C.getConstantArrayType( | |||
4501 | KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies), nullptr, | |||
4502 | ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
4503 | DependenciesArray = | |||
4504 | CGF.CreateMemTemp(KmpDependInfoArrayTy, ".dep.arr.addr"); | |||
4505 | DependenciesArray = CGF.Builder.CreateConstArrayGEP(DependenciesArray, 0); | |||
4506 | NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumDependencies, | |||
4507 | /*isSigned=*/false); | |||
4508 | } | |||
4509 | unsigned Pos = 0; | |||
4510 | for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) { | |||
4511 | if (Dependencies[I].DepKind == OMPC_DEPEND_depobj || | |||
4512 | Dependencies[I].IteratorExpr) | |||
4513 | continue; | |||
4514 | emitDependData(CGF, KmpDependInfoTy, &Pos, Dependencies[I], | |||
4515 | DependenciesArray); | |||
4516 | } | |||
4517 | // Copy regular dependencies with iterators. | |||
4518 | LValue PosLVal = CGF.MakeAddrLValue( | |||
4519 | CGF.CreateMemTemp(C.getSizeType(), "dep.counter.addr"), C.getSizeType()); | |||
4520 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal); | |||
4521 | for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) { | |||
4522 | if (Dependencies[I].DepKind == OMPC_DEPEND_depobj || | |||
4523 | !Dependencies[I].IteratorExpr) | |||
4524 | continue; | |||
4525 | emitDependData(CGF, KmpDependInfoTy, &PosLVal, Dependencies[I], | |||
4526 | DependenciesArray); | |||
4527 | } | |||
4528 | // Copy final depobj arrays without iterators. | |||
4529 | if (HasDepobjDeps) { | |||
4530 | for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) { | |||
4531 | if (Dependencies[I].DepKind != OMPC_DEPEND_depobj) | |||
4532 | continue; | |||
4533 | emitDepobjElements(CGF, KmpDependInfoTy, PosLVal, Dependencies[I], | |||
4534 | DependenciesArray); | |||
4535 | } | |||
4536 | } | |||
4537 | DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4538 | DependenciesArray, CGF.VoidPtrTy, CGF.Int8Ty); | |||
4539 | return std::make_pair(NumOfElements, DependenciesArray); | |||
4540 | } | |||
4541 | ||||
4542 | Address CGOpenMPRuntime::emitDepobjDependClause( | |||
4543 | CodeGenFunction &CGF, const OMPTaskDataTy::DependData &Dependencies, | |||
4544 | SourceLocation Loc) { | |||
4545 | if (Dependencies.DepExprs.empty()) | |||
4546 | return Address::invalid(); | |||
4547 | // Process list of dependencies. | |||
4548 | ASTContext &C = CGM.getContext(); | |||
4549 | Address DependenciesArray = Address::invalid(); | |||
4550 | unsigned NumDependencies = Dependencies.DepExprs.size(); | |||
4551 | QualType FlagsTy; | |||
4552 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | |||
4553 | RecordDecl *KmpDependInfoRD = | |||
4554 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | |||
4555 | ||||
4556 | llvm::Value *Size; | |||
4557 | // Define type kmp_depend_info[<Dependencies.size()>]; | |||
4558 | // For depobj reserve one extra element to store the number of elements. | |||
4559 | // It is required to handle depobj(x) update(in) construct. | |||
4560 | // kmp_depend_info[<Dependencies.size()>] deps; | |||
4561 | llvm::Value *NumDepsVal; | |||
4562 | CharUnits Align = C.getTypeAlignInChars(KmpDependInfoTy); | |||
4563 | if (const auto *IE = | |||
4564 | cast_or_null<OMPIteratorExpr>(Dependencies.IteratorExpr)) { | |||
4565 | NumDepsVal = llvm::ConstantInt::get(CGF.SizeTy, 1); | |||
4566 | for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) { | |||
4567 | llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper); | |||
4568 | Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false); | |||
4569 | NumDepsVal = CGF.Builder.CreateNUWMul(NumDepsVal, Sz); | |||
4570 | } | |||
4571 | Size = CGF.Builder.CreateNUWAdd(llvm::ConstantInt::get(CGF.SizeTy, 1), | |||
4572 | NumDepsVal); | |||
4573 | CharUnits SizeInBytes = | |||
4574 | C.getTypeSizeInChars(KmpDependInfoTy).alignTo(Align); | |||
4575 | llvm::Value *RecSize = CGM.getSize(SizeInBytes); | |||
4576 | Size = CGF.Builder.CreateNUWMul(Size, RecSize); | |||
4577 | NumDepsVal = | |||
4578 | CGF.Builder.CreateIntCast(NumDepsVal, CGF.IntPtrTy, /*isSigned=*/false); | |||
4579 | } else { | |||
4580 | QualType KmpDependInfoArrayTy = C.getConstantArrayType( | |||
4581 | KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies + 1), | |||
4582 | nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
4583 | CharUnits Sz = C.getTypeSizeInChars(KmpDependInfoArrayTy); | |||
4584 | Size = CGM.getSize(Sz.alignTo(Align)); | |||
4585 | NumDepsVal = llvm::ConstantInt::get(CGF.IntPtrTy, NumDependencies); | |||
4586 | } | |||
4587 | // Need to allocate on the dynamic memory. | |||
4588 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
4589 | // Use default allocator. | |||
4590 | llvm::Value *Allocator = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
4591 | llvm::Value *Args[] = {ThreadID, Size, Allocator}; | |||
4592 | ||||
4593 | llvm::Value *Addr = | |||
4594 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
4595 | CGM.getModule(), OMPRTL___kmpc_alloc), | |||
4596 | Args, ".dep.arr.addr"); | |||
4597 | llvm::Type *KmpDependInfoLlvmTy = CGF.ConvertTypeForMem(KmpDependInfoTy); | |||
4598 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4599 | Addr, KmpDependInfoLlvmTy->getPointerTo()); | |||
4600 | DependenciesArray = Address(Addr, KmpDependInfoLlvmTy, Align); | |||
4601 | // Write number of elements in the first element of array for depobj. | |||
4602 | LValue Base = CGF.MakeAddrLValue(DependenciesArray, KmpDependInfoTy); | |||
4603 | // deps[i].base_addr = NumDependencies; | |||
4604 | LValue BaseAddrLVal = CGF.EmitLValueForField( | |||
4605 | Base, | |||
4606 | *std::next(KmpDependInfoRD->field_begin(), | |||
4607 | static_cast<unsigned int>(RTLDependInfoFields::BaseAddr))); | |||
4608 | CGF.EmitStoreOfScalar(NumDepsVal, BaseAddrLVal); | |||
4609 | llvm::PointerUnion<unsigned *, LValue *> Pos; | |||
4610 | unsigned Idx = 1; | |||
4611 | LValue PosLVal; | |||
4612 | if (Dependencies.IteratorExpr) { | |||
4613 | PosLVal = CGF.MakeAddrLValue( | |||
4614 | CGF.CreateMemTemp(C.getSizeType(), "iterator.counter.addr"), | |||
4615 | C.getSizeType()); | |||
4616 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Idx), PosLVal, | |||
4617 | /*IsInit=*/true); | |||
4618 | Pos = &PosLVal; | |||
4619 | } else { | |||
4620 | Pos = &Idx; | |||
4621 | } | |||
4622 | emitDependData(CGF, KmpDependInfoTy, Pos, Dependencies, DependenciesArray); | |||
4623 | DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4624 | CGF.Builder.CreateConstGEP(DependenciesArray, 1), CGF.VoidPtrTy, | |||
4625 | CGF.Int8Ty); | |||
4626 | return DependenciesArray; | |||
4627 | } | |||
4628 | ||||
4629 | void CGOpenMPRuntime::emitDestroyClause(CodeGenFunction &CGF, LValue DepobjLVal, | |||
4630 | SourceLocation Loc) { | |||
4631 | ASTContext &C = CGM.getContext(); | |||
4632 | QualType FlagsTy; | |||
4633 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | |||
4634 | LValue Base = CGF.EmitLoadOfPointerLValue( | |||
4635 | DepobjLVal.getAddress(CGF), C.VoidPtrTy.castAs<PointerType>()); | |||
4636 | QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy); | |||
4637 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4638 | Base.getAddress(CGF), CGF.ConvertTypeForMem(KmpDependInfoPtrTy), | |||
4639 | CGF.ConvertTypeForMem(KmpDependInfoTy)); | |||
4640 | llvm::Value *DepObjAddr = CGF.Builder.CreateGEP( | |||
4641 | Addr.getElementType(), Addr.getPointer(), | |||
4642 | llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true)); | |||
4643 | DepObjAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(DepObjAddr, | |||
4644 | CGF.VoidPtrTy); | |||
4645 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
4646 | // Use default allocator. | |||
4647 | llvm::Value *Allocator = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
4648 | llvm::Value *Args[] = {ThreadID, DepObjAddr, Allocator}; | |||
4649 | ||||
4650 | // _kmpc_free(gtid, addr, nullptr); | |||
4651 | (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
4652 | CGM.getModule(), OMPRTL___kmpc_free), | |||
4653 | Args); | |||
4654 | } | |||
4655 | ||||
4656 | void CGOpenMPRuntime::emitUpdateClause(CodeGenFunction &CGF, LValue DepobjLVal, | |||
4657 | OpenMPDependClauseKind NewDepKind, | |||
4658 | SourceLocation Loc) { | |||
4659 | ASTContext &C = CGM.getContext(); | |||
4660 | QualType FlagsTy; | |||
4661 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | |||
4662 | RecordDecl *KmpDependInfoRD = | |||
4663 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | |||
4664 | llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy); | |||
4665 | llvm::Value *NumDeps; | |||
4666 | LValue Base; | |||
4667 | std::tie(NumDeps, Base) = getDepobjElements(CGF, DepobjLVal, Loc); | |||
4668 | ||||
4669 | Address Begin = Base.getAddress(CGF); | |||
4670 | // Cast from pointer to array type to pointer to single element. | |||
4671 | llvm::Value *End = CGF.Builder.CreateGEP( | |||
4672 | Begin.getElementType(), Begin.getPointer(), NumDeps); | |||
4673 | // The basic structure here is a while-do loop. | |||
4674 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.body"); | |||
4675 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.done"); | |||
4676 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | |||
4677 | CGF.EmitBlock(BodyBB); | |||
4678 | llvm::PHINode *ElementPHI = | |||
4679 | CGF.Builder.CreatePHI(Begin.getType(), 2, "omp.elementPast"); | |||
4680 | ElementPHI->addIncoming(Begin.getPointer(), EntryBB); | |||
4681 | Begin = Begin.withPointer(ElementPHI); | |||
4682 | Base = CGF.MakeAddrLValue(Begin, KmpDependInfoTy, Base.getBaseInfo(), | |||
4683 | Base.getTBAAInfo()); | |||
4684 | // deps[i].flags = NewDepKind; | |||
4685 | RTLDependenceKindTy DepKind = translateDependencyKind(NewDepKind); | |||
4686 | LValue FlagsLVal = CGF.EmitLValueForField( | |||
4687 | Base, *std::next(KmpDependInfoRD->field_begin(), | |||
4688 | static_cast<unsigned int>(RTLDependInfoFields::Flags))); | |||
4689 | CGF.EmitStoreOfScalar( | |||
4690 | llvm::ConstantInt::get(LLVMFlagsTy, static_cast<unsigned int>(DepKind)), | |||
4691 | FlagsLVal); | |||
4692 | ||||
4693 | // Shift the address forward by one element. | |||
4694 | Address ElementNext = | |||
4695 | CGF.Builder.CreateConstGEP(Begin, /*Index=*/1, "omp.elementNext"); | |||
4696 | ElementPHI->addIncoming(ElementNext.getPointer(), | |||
4697 | CGF.Builder.GetInsertBlock()); | |||
4698 | llvm::Value *IsEmpty = | |||
4699 | CGF.Builder.CreateICmpEQ(ElementNext.getPointer(), End, "omp.isempty"); | |||
4700 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | |||
4701 | // Done. | |||
4702 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | |||
4703 | } | |||
4704 | ||||
4705 | void CGOpenMPRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
4706 | const OMPExecutableDirective &D, | |||
4707 | llvm::Function *TaskFunction, | |||
4708 | QualType SharedsTy, Address Shareds, | |||
4709 | const Expr *IfCond, | |||
4710 | const OMPTaskDataTy &Data) { | |||
4711 | if (!CGF.HaveInsertPoint()) | |||
4712 | return; | |||
4713 | ||||
4714 | TaskResultTy Result = | |||
4715 | emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data); | |||
4716 | llvm::Value *NewTask = Result.NewTask; | |||
4717 | llvm::Function *TaskEntry = Result.TaskEntry; | |||
4718 | llvm::Value *NewTaskNewTaskTTy = Result.NewTaskNewTaskTTy; | |||
4719 | LValue TDBase = Result.TDBase; | |||
4720 | const RecordDecl *KmpTaskTQTyRD = Result.KmpTaskTQTyRD; | |||
4721 | // Process list of dependences. | |||
4722 | Address DependenciesArray = Address::invalid(); | |||
4723 | llvm::Value *NumOfElements; | |||
4724 | std::tie(NumOfElements, DependenciesArray) = | |||
4725 | emitDependClause(CGF, Data.Dependences, Loc); | |||
4726 | ||||
4727 | // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc() | |||
4728 | // libcall. | |||
4729 | // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid, | |||
4730 | // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, | |||
4731 | // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) if dependence | |||
4732 | // list is not empty | |||
4733 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
4734 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | |||
4735 | llvm::Value *TaskArgs[] = { UpLoc, ThreadID, NewTask }; | |||
4736 | llvm::Value *DepTaskArgs[7]; | |||
4737 | if (!Data.Dependences.empty()) { | |||
4738 | DepTaskArgs[0] = UpLoc; | |||
4739 | DepTaskArgs[1] = ThreadID; | |||
4740 | DepTaskArgs[2] = NewTask; | |||
4741 | DepTaskArgs[3] = NumOfElements; | |||
4742 | DepTaskArgs[4] = DependenciesArray.getPointer(); | |||
4743 | DepTaskArgs[5] = CGF.Builder.getInt32(0); | |||
4744 | DepTaskArgs[6] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
4745 | } | |||
4746 | auto &&ThenCodeGen = [this, &Data, TDBase, KmpTaskTQTyRD, &TaskArgs, | |||
4747 | &DepTaskArgs](CodeGenFunction &CGF, PrePostActionTy &) { | |||
4748 | if (!Data.Tied) { | |||
4749 | auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId); | |||
4750 | LValue PartIdLVal = CGF.EmitLValueForField(TDBase, *PartIdFI); | |||
4751 | CGF.EmitStoreOfScalar(CGF.Builder.getInt32(0), PartIdLVal); | |||
4752 | } | |||
4753 | if (!Data.Dependences.empty()) { | |||
4754 | CGF.EmitRuntimeCall( | |||
4755 | OMPBuilder.getOrCreateRuntimeFunction( | |||
4756 | CGM.getModule(), OMPRTL___kmpc_omp_task_with_deps), | |||
4757 | DepTaskArgs); | |||
4758 | } else { | |||
4759 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
4760 | CGM.getModule(), OMPRTL___kmpc_omp_task), | |||
4761 | TaskArgs); | |||
4762 | } | |||
4763 | // Check if parent region is untied and build return for untied task; | |||
4764 | if (auto *Region = | |||
4765 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
4766 | Region->emitUntiedSwitch(CGF); | |||
4767 | }; | |||
4768 | ||||
4769 | llvm::Value *DepWaitTaskArgs[6]; | |||
4770 | if (!Data.Dependences.empty()) { | |||
4771 | DepWaitTaskArgs[0] = UpLoc; | |||
4772 | DepWaitTaskArgs[1] = ThreadID; | |||
4773 | DepWaitTaskArgs[2] = NumOfElements; | |||
4774 | DepWaitTaskArgs[3] = DependenciesArray.getPointer(); | |||
4775 | DepWaitTaskArgs[4] = CGF.Builder.getInt32(0); | |||
4776 | DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
4777 | } | |||
4778 | auto &M = CGM.getModule(); | |||
4779 | auto &&ElseCodeGen = [this, &M, &TaskArgs, ThreadID, NewTaskNewTaskTTy, | |||
4780 | TaskEntry, &Data, &DepWaitTaskArgs, | |||
4781 | Loc](CodeGenFunction &CGF, PrePostActionTy &) { | |||
4782 | CodeGenFunction::RunCleanupsScope LocalScope(CGF); | |||
4783 | // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid, | |||
4784 | // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 | |||
4785 | // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info | |||
4786 | // is specified. | |||
4787 | if (!Data.Dependences.empty()) | |||
4788 | CGF.EmitRuntimeCall( | |||
4789 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_omp_wait_deps), | |||
4790 | DepWaitTaskArgs); | |||
4791 | // Call proxy_task_entry(gtid, new_task); | |||
4792 | auto &&CodeGen = [TaskEntry, ThreadID, NewTaskNewTaskTTy, | |||
4793 | Loc](CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
4794 | Action.Enter(CGF); | |||
4795 | llvm::Value *OutlinedFnArgs[] = {ThreadID, NewTaskNewTaskTTy}; | |||
4796 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskEntry, | |||
4797 | OutlinedFnArgs); | |||
4798 | }; | |||
4799 | ||||
4800 | // Build void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid, | |||
4801 | // kmp_task_t *new_task); | |||
4802 | // Build void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid, | |||
4803 | // kmp_task_t *new_task); | |||
4804 | RegionCodeGenTy RCG(CodeGen); | |||
4805 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | |||
4806 | M, OMPRTL___kmpc_omp_task_begin_if0), | |||
4807 | TaskArgs, | |||
4808 | OMPBuilder.getOrCreateRuntimeFunction( | |||
4809 | M, OMPRTL___kmpc_omp_task_complete_if0), | |||
4810 | TaskArgs); | |||
4811 | RCG.setAction(Action); | |||
4812 | RCG(CGF); | |||
4813 | }; | |||
4814 | ||||
4815 | if (IfCond) { | |||
4816 | emitIfClause(CGF, IfCond, ThenCodeGen, ElseCodeGen); | |||
4817 | } else { | |||
4818 | RegionCodeGenTy ThenRCG(ThenCodeGen); | |||
4819 | ThenRCG(CGF); | |||
4820 | } | |||
4821 | } | |||
4822 | ||||
4823 | void CGOpenMPRuntime::emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
4824 | const OMPLoopDirective &D, | |||
4825 | llvm::Function *TaskFunction, | |||
4826 | QualType SharedsTy, Address Shareds, | |||
4827 | const Expr *IfCond, | |||
4828 | const OMPTaskDataTy &Data) { | |||
4829 | if (!CGF.HaveInsertPoint()) | |||
4830 | return; | |||
4831 | TaskResultTy Result = | |||
4832 | emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data); | |||
4833 | // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc() | |||
4834 | // libcall. | |||
4835 | // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int | |||
4836 | // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int | |||
4837 | // sched, kmp_uint64 grainsize, void *task_dup); | |||
4838 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
4839 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | |||
4840 | llvm::Value *IfVal; | |||
4841 | if (IfCond) { | |||
4842 | IfVal = CGF.Builder.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.IntTy, | |||
4843 | /*isSigned=*/true); | |||
4844 | } else { | |||
4845 | IfVal = llvm::ConstantInt::getSigned(CGF.IntTy, /*V=*/1); | |||
4846 | } | |||
4847 | ||||
4848 | LValue LBLVal = CGF.EmitLValueForField( | |||
4849 | Result.TDBase, | |||
4850 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound)); | |||
4851 | const auto *LBVar = | |||
4852 | cast<VarDecl>(cast<DeclRefExpr>(D.getLowerBoundVariable())->getDecl()); | |||
4853 | CGF.EmitAnyExprToMem(LBVar->getInit(), LBLVal.getAddress(CGF), | |||
4854 | LBLVal.getQuals(), | |||
4855 | /*IsInitializer=*/true); | |||
4856 | LValue UBLVal = CGF.EmitLValueForField( | |||
4857 | Result.TDBase, | |||
4858 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound)); | |||
4859 | const auto *UBVar = | |||
4860 | cast<VarDecl>(cast<DeclRefExpr>(D.getUpperBoundVariable())->getDecl()); | |||
4861 | CGF.EmitAnyExprToMem(UBVar->getInit(), UBLVal.getAddress(CGF), | |||
4862 | UBLVal.getQuals(), | |||
4863 | /*IsInitializer=*/true); | |||
4864 | LValue StLVal = CGF.EmitLValueForField( | |||
4865 | Result.TDBase, | |||
4866 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTStride)); | |||
4867 | const auto *StVar = | |||
4868 | cast<VarDecl>(cast<DeclRefExpr>(D.getStrideVariable())->getDecl()); | |||
4869 | CGF.EmitAnyExprToMem(StVar->getInit(), StLVal.getAddress(CGF), | |||
4870 | StLVal.getQuals(), | |||
4871 | /*IsInitializer=*/true); | |||
4872 | // Store reductions address. | |||
4873 | LValue RedLVal = CGF.EmitLValueForField( | |||
4874 | Result.TDBase, | |||
4875 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTReductions)); | |||
4876 | if (Data.Reductions) { | |||
4877 | CGF.EmitStoreOfScalar(Data.Reductions, RedLVal); | |||
4878 | } else { | |||
4879 | CGF.EmitNullInitialization(RedLVal.getAddress(CGF), | |||
4880 | CGF.getContext().VoidPtrTy); | |||
4881 | } | |||
4882 | enum { NoSchedule = 0, Grainsize = 1, NumTasks = 2 }; | |||
4883 | llvm::Value *TaskArgs[] = { | |||
4884 | UpLoc, | |||
4885 | ThreadID, | |||
4886 | Result.NewTask, | |||
4887 | IfVal, | |||
4888 | LBLVal.getPointer(CGF), | |||
4889 | UBLVal.getPointer(CGF), | |||
4890 | CGF.EmitLoadOfScalar(StLVal, Loc), | |||
4891 | llvm::ConstantInt::getSigned( | |||
4892 | CGF.IntTy, 1), // Always 1 because taskgroup emitted by the compiler | |||
4893 | llvm::ConstantInt::getSigned( | |||
4894 | CGF.IntTy, Data.Schedule.getPointer() | |||
4895 | ? Data.Schedule.getInt() ? NumTasks : Grainsize | |||
4896 | : NoSchedule), | |||
4897 | Data.Schedule.getPointer() | |||
4898 | ? CGF.Builder.CreateIntCast(Data.Schedule.getPointer(), CGF.Int64Ty, | |||
4899 | /*isSigned=*/false) | |||
4900 | : llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/0), | |||
4901 | Result.TaskDupFn ? CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4902 | Result.TaskDupFn, CGF.VoidPtrTy) | |||
4903 | : llvm::ConstantPointerNull::get(CGF.VoidPtrTy)}; | |||
4904 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
4905 | CGM.getModule(), OMPRTL___kmpc_taskloop), | |||
4906 | TaskArgs); | |||
4907 | } | |||
4908 | ||||
4909 | /// Emit reduction operation for each element of array (required for | |||
4910 | /// array sections) LHS op = RHS. | |||
4911 | /// \param Type Type of array. | |||
4912 | /// \param LHSVar Variable on the left side of the reduction operation | |||
4913 | /// (references element of array in original variable). | |||
4914 | /// \param RHSVar Variable on the right side of the reduction operation | |||
4915 | /// (references element of array in original variable). | |||
4916 | /// \param RedOpGen Generator of reduction operation with use of LHSVar and | |||
4917 | /// RHSVar. | |||
4918 | static void EmitOMPAggregateReduction( | |||
4919 | CodeGenFunction &CGF, QualType Type, const VarDecl *LHSVar, | |||
4920 | const VarDecl *RHSVar, | |||
4921 | const llvm::function_ref<void(CodeGenFunction &CGF, const Expr *, | |||
4922 | const Expr *, const Expr *)> &RedOpGen, | |||
4923 | const Expr *XExpr = nullptr, const Expr *EExpr = nullptr, | |||
4924 | const Expr *UpExpr = nullptr) { | |||
4925 | // Perform element-by-element initialization. | |||
4926 | QualType ElementTy; | |||
4927 | Address LHSAddr = CGF.GetAddrOfLocalVar(LHSVar); | |||
4928 | Address RHSAddr = CGF.GetAddrOfLocalVar(RHSVar); | |||
4929 | ||||
4930 | // Drill down to the base element type on both arrays. | |||
4931 | const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe(); | |||
4932 | llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, LHSAddr); | |||
4933 | ||||
4934 | llvm::Value *RHSBegin = RHSAddr.getPointer(); | |||
4935 | llvm::Value *LHSBegin = LHSAddr.getPointer(); | |||
4936 | // Cast from pointer to array type to pointer to single element. | |||
4937 | llvm::Value *LHSEnd = | |||
4938 | CGF.Builder.CreateGEP(LHSAddr.getElementType(), LHSBegin, NumElements); | |||
4939 | // The basic structure here is a while-do loop. | |||
4940 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arraycpy.body"); | |||
4941 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arraycpy.done"); | |||
4942 | llvm::Value *IsEmpty = | |||
4943 | CGF.Builder.CreateICmpEQ(LHSBegin, LHSEnd, "omp.arraycpy.isempty"); | |||
4944 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | |||
4945 | ||||
4946 | // Enter the loop body, making that address the current address. | |||
4947 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | |||
4948 | CGF.EmitBlock(BodyBB); | |||
4949 | ||||
4950 | CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy); | |||
4951 | ||||
4952 | llvm::PHINode *RHSElementPHI = CGF.Builder.CreatePHI( | |||
4953 | RHSBegin->getType(), 2, "omp.arraycpy.srcElementPast"); | |||
4954 | RHSElementPHI->addIncoming(RHSBegin, EntryBB); | |||
4955 | Address RHSElementCurrent( | |||
4956 | RHSElementPHI, RHSAddr.getElementType(), | |||
4957 | RHSAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
4958 | ||||
4959 | llvm::PHINode *LHSElementPHI = CGF.Builder.CreatePHI( | |||
4960 | LHSBegin->getType(), 2, "omp.arraycpy.destElementPast"); | |||
4961 | LHSElementPHI->addIncoming(LHSBegin, EntryBB); | |||
4962 | Address LHSElementCurrent( | |||
4963 | LHSElementPHI, LHSAddr.getElementType(), | |||
4964 | LHSAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
4965 | ||||
4966 | // Emit copy. | |||
4967 | CodeGenFunction::OMPPrivateScope Scope(CGF); | |||
4968 | Scope.addPrivate(LHSVar, LHSElementCurrent); | |||
4969 | Scope.addPrivate(RHSVar, RHSElementCurrent); | |||
4970 | Scope.Privatize(); | |||
4971 | RedOpGen(CGF, XExpr, EExpr, UpExpr); | |||
4972 | Scope.ForceCleanup(); | |||
4973 | ||||
4974 | // Shift the address forward by one element. | |||
4975 | llvm::Value *LHSElementNext = CGF.Builder.CreateConstGEP1_32( | |||
4976 | LHSAddr.getElementType(), LHSElementPHI, /*Idx0=*/1, | |||
4977 | "omp.arraycpy.dest.element"); | |||
4978 | llvm::Value *RHSElementNext = CGF.Builder.CreateConstGEP1_32( | |||
4979 | RHSAddr.getElementType(), RHSElementPHI, /*Idx0=*/1, | |||
4980 | "omp.arraycpy.src.element"); | |||
4981 | // Check whether we've reached the end. | |||
4982 | llvm::Value *Done = | |||
4983 | CGF.Builder.CreateICmpEQ(LHSElementNext, LHSEnd, "omp.arraycpy.done"); | |||
4984 | CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB); | |||
4985 | LHSElementPHI->addIncoming(LHSElementNext, CGF.Builder.GetInsertBlock()); | |||
4986 | RHSElementPHI->addIncoming(RHSElementNext, CGF.Builder.GetInsertBlock()); | |||
4987 | ||||
4988 | // Done. | |||
4989 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | |||
4990 | } | |||
4991 | ||||
4992 | /// Emit reduction combiner. If the combiner is a simple expression emit it as | |||
4993 | /// is, otherwise consider it as combiner of UDR decl and emit it as a call of | |||
4994 | /// UDR combiner function. | |||
4995 | static void emitReductionCombiner(CodeGenFunction &CGF, | |||
4996 | const Expr *ReductionOp) { | |||
4997 | if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) | |||
4998 | if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee())) | |||
4999 | if (const auto *DRE = | |||
5000 | dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts())) | |||
5001 | if (const auto *DRD = | |||
5002 | dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) { | |||
5003 | std::pair<llvm::Function *, llvm::Function *> Reduction = | |||
5004 | CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD); | |||
5005 | RValue Func = RValue::get(Reduction.first); | |||
5006 | CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func); | |||
5007 | CGF.EmitIgnoredExpr(ReductionOp); | |||
5008 | return; | |||
5009 | } | |||
5010 | CGF.EmitIgnoredExpr(ReductionOp); | |||
5011 | } | |||
5012 | ||||
5013 | llvm::Function *CGOpenMPRuntime::emitReductionFunction( | |||
5014 | SourceLocation Loc, llvm::Type *ArgsElemType, | |||
5015 | ArrayRef<const Expr *> Privates, ArrayRef<const Expr *> LHSExprs, | |||
5016 | ArrayRef<const Expr *> RHSExprs, ArrayRef<const Expr *> ReductionOps) { | |||
5017 | ASTContext &C = CGM.getContext(); | |||
5018 | ||||
5019 | // void reduction_func(void *LHSArg, void *RHSArg); | |||
5020 | FunctionArgList Args; | |||
5021 | ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5022 | ImplicitParamDecl::Other); | |||
5023 | ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5024 | ImplicitParamDecl::Other); | |||
5025 | Args.push_back(&LHSArg); | |||
5026 | Args.push_back(&RHSArg); | |||
5027 | const auto &CGFI = | |||
5028 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
5029 | std::string Name = getName({"omp", "reduction", "reduction_func"}); | |||
5030 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | |||
5031 | llvm::GlobalValue::InternalLinkage, Name, | |||
5032 | &CGM.getModule()); | |||
5033 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
5034 | Fn->setDoesNotRecurse(); | |||
5035 | CodeGenFunction CGF(CGM); | |||
5036 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
5037 | ||||
5038 | // Dst = (void*[n])(LHSArg); | |||
5039 | // Src = (void*[n])(RHSArg); | |||
5040 | Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5041 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)), | |||
5042 | ArgsElemType->getPointerTo()), | |||
5043 | ArgsElemType, CGF.getPointerAlign()); | |||
5044 | Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5045 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)), | |||
5046 | ArgsElemType->getPointerTo()), | |||
5047 | ArgsElemType, CGF.getPointerAlign()); | |||
5048 | ||||
5049 | // ... | |||
5050 | // *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]); | |||
5051 | // ... | |||
5052 | CodeGenFunction::OMPPrivateScope Scope(CGF); | |||
5053 | const auto *IPriv = Privates.begin(); | |||
5054 | unsigned Idx = 0; | |||
5055 | for (unsigned I = 0, E = ReductionOps.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
5056 | const auto *RHSVar = | |||
5057 | cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl()); | |||
5058 | Scope.addPrivate(RHSVar, emitAddrOfVarFromArray(CGF, RHS, Idx, RHSVar)); | |||
5059 | const auto *LHSVar = | |||
5060 | cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl()); | |||
5061 | Scope.addPrivate(LHSVar, emitAddrOfVarFromArray(CGF, LHS, Idx, LHSVar)); | |||
5062 | QualType PrivTy = (*IPriv)->getType(); | |||
5063 | if (PrivTy->isVariablyModifiedType()) { | |||
5064 | // Get array size and emit VLA type. | |||
5065 | ++Idx; | |||
5066 | Address Elem = CGF.Builder.CreateConstArrayGEP(LHS, Idx); | |||
5067 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Elem); | |||
5068 | const VariableArrayType *VLA = | |||
5069 | CGF.getContext().getAsVariableArrayType(PrivTy); | |||
5070 | const auto *OVE = cast<OpaqueValueExpr>(VLA->getSizeExpr()); | |||
5071 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | |||
5072 | CGF, OVE, RValue::get(CGF.Builder.CreatePtrToInt(Ptr, CGF.SizeTy))); | |||
5073 | CGF.EmitVariablyModifiedType(PrivTy); | |||
5074 | } | |||
5075 | } | |||
5076 | Scope.Privatize(); | |||
5077 | IPriv = Privates.begin(); | |||
5078 | const auto *ILHS = LHSExprs.begin(); | |||
5079 | const auto *IRHS = RHSExprs.begin(); | |||
5080 | for (const Expr *E : ReductionOps) { | |||
5081 | if ((*IPriv)->getType()->isArrayType()) { | |||
5082 | // Emit reduction for array section. | |||
5083 | const auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | |||
5084 | const auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | |||
5085 | EmitOMPAggregateReduction( | |||
5086 | CGF, (*IPriv)->getType(), LHSVar, RHSVar, | |||
5087 | [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) { | |||
5088 | emitReductionCombiner(CGF, E); | |||
5089 | }); | |||
5090 | } else { | |||
5091 | // Emit reduction for array subscript or single variable. | |||
5092 | emitReductionCombiner(CGF, E); | |||
5093 | } | |||
5094 | ++IPriv; | |||
5095 | ++ILHS; | |||
5096 | ++IRHS; | |||
5097 | } | |||
5098 | Scope.ForceCleanup(); | |||
5099 | CGF.FinishFunction(); | |||
5100 | return Fn; | |||
5101 | } | |||
5102 | ||||
5103 | void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF, | |||
5104 | const Expr *ReductionOp, | |||
5105 | const Expr *PrivateRef, | |||
5106 | const DeclRefExpr *LHS, | |||
5107 | const DeclRefExpr *RHS) { | |||
5108 | if (PrivateRef->getType()->isArrayType()) { | |||
5109 | // Emit reduction for array section. | |||
5110 | const auto *LHSVar = cast<VarDecl>(LHS->getDecl()); | |||
5111 | const auto *RHSVar = cast<VarDecl>(RHS->getDecl()); | |||
5112 | EmitOMPAggregateReduction( | |||
5113 | CGF, PrivateRef->getType(), LHSVar, RHSVar, | |||
5114 | [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) { | |||
5115 | emitReductionCombiner(CGF, ReductionOp); | |||
5116 | }); | |||
5117 | } else { | |||
5118 | // Emit reduction for array subscript or single variable. | |||
5119 | emitReductionCombiner(CGF, ReductionOp); | |||
5120 | } | |||
5121 | } | |||
5122 | ||||
5123 | void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc, | |||
5124 | ArrayRef<const Expr *> Privates, | |||
5125 | ArrayRef<const Expr *> LHSExprs, | |||
5126 | ArrayRef<const Expr *> RHSExprs, | |||
5127 | ArrayRef<const Expr *> ReductionOps, | |||
5128 | ReductionOptionsTy Options) { | |||
5129 | if (!CGF.HaveInsertPoint()) | |||
5130 | return; | |||
5131 | ||||
5132 | bool WithNowait = Options.WithNowait; | |||
5133 | bool SimpleReduction = Options.SimpleReduction; | |||
5134 | ||||
5135 | // Next code should be emitted for reduction: | |||
5136 | // | |||
5137 | // static kmp_critical_name lock = { 0 }; | |||
5138 | // | |||
5139 | // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { | |||
5140 | // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); | |||
5141 | // ... | |||
5142 | // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], | |||
5143 | // *(Type<n>-1*)rhs[<n>-1]); | |||
5144 | // } | |||
5145 | // | |||
5146 | // ... | |||
5147 | // void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; | |||
5148 | // switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), | |||
5149 | // RedList, reduce_func, &<lock>)) { | |||
5150 | // case 1: | |||
5151 | // ... | |||
5152 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | |||
5153 | // ... | |||
5154 | // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | |||
5155 | // break; | |||
5156 | // case 2: | |||
5157 | // ... | |||
5158 | // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); | |||
5159 | // ... | |||
5160 | // [__kmpc_end_reduce(<loc>, <gtid>, &<lock>);] | |||
5161 | // break; | |||
5162 | // default:; | |||
5163 | // } | |||
5164 | // | |||
5165 | // if SimpleReduction is true, only the next code is generated: | |||
5166 | // ... | |||
5167 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | |||
5168 | // ... | |||
5169 | ||||
5170 | ASTContext &C = CGM.getContext(); | |||
5171 | ||||
5172 | if (SimpleReduction) { | |||
5173 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
5174 | const auto *IPriv = Privates.begin(); | |||
5175 | const auto *ILHS = LHSExprs.begin(); | |||
5176 | const auto *IRHS = RHSExprs.begin(); | |||
5177 | for (const Expr *E : ReductionOps) { | |||
5178 | emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | |||
5179 | cast<DeclRefExpr>(*IRHS)); | |||
5180 | ++IPriv; | |||
5181 | ++ILHS; | |||
5182 | ++IRHS; | |||
5183 | } | |||
5184 | return; | |||
5185 | } | |||
5186 | ||||
5187 | // 1. Build a list of reduction variables. | |||
5188 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; | |||
5189 | auto Size = RHSExprs.size(); | |||
5190 | for (const Expr *E : Privates) { | |||
5191 | if (E->getType()->isVariablyModifiedType()) | |||
5192 | // Reserve place for array size. | |||
5193 | ++Size; | |||
5194 | } | |||
5195 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); | |||
5196 | QualType ReductionArrayTy = | |||
5197 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, | |||
5198 | /*IndexTypeQuals=*/0); | |||
5199 | Address ReductionList = | |||
5200 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); | |||
5201 | const auto *IPriv = Privates.begin(); | |||
5202 | unsigned Idx = 0; | |||
5203 | for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
5204 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
5205 | CGF.Builder.CreateStore( | |||
5206 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5207 | CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), | |||
5208 | Elem); | |||
5209 | if ((*IPriv)->getType()->isVariablyModifiedType()) { | |||
5210 | // Store array size. | |||
5211 | ++Idx; | |||
5212 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
5213 | llvm::Value *Size = CGF.Builder.CreateIntCast( | |||
5214 | CGF.getVLASize( | |||
5215 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) | |||
5216 | .NumElts, | |||
5217 | CGF.SizeTy, /*isSigned=*/false); | |||
5218 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), | |||
5219 | Elem); | |||
5220 | } | |||
5221 | } | |||
5222 | ||||
5223 | // 2. Emit reduce_func(). | |||
5224 | llvm::Function *ReductionFn = | |||
5225 | emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy), | |||
5226 | Privates, LHSExprs, RHSExprs, ReductionOps); | |||
5227 | ||||
5228 | // 3. Create static kmp_critical_name lock = { 0 }; | |||
5229 | std::string Name = getName({"reduction"}); | |||
5230 | llvm::Value *Lock = getCriticalRegionLock(Name); | |||
5231 | ||||
5232 | // 4. Build res = __kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), | |||
5233 | // RedList, reduce_func, &<lock>); | |||
5234 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE); | |||
5235 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
5236 | llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); | |||
5237 | llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5238 | ReductionList.getPointer(), CGF.VoidPtrTy); | |||
5239 | llvm::Value *Args[] = { | |||
5240 | IdentTLoc, // ident_t *<loc> | |||
5241 | ThreadId, // i32 <gtid> | |||
5242 | CGF.Builder.getInt32(RHSExprs.size()), // i32 <n> | |||
5243 | ReductionArrayTySize, // size_type sizeof(RedList) | |||
5244 | RL, // void *RedList | |||
5245 | ReductionFn, // void (*) (void *, void *) <reduce_func> | |||
5246 | Lock // kmp_critical_name *&<lock> | |||
5247 | }; | |||
5248 | llvm::Value *Res = CGF.EmitRuntimeCall( | |||
5249 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5250 | CGM.getModule(), | |||
5251 | WithNowait ? OMPRTL___kmpc_reduce_nowait : OMPRTL___kmpc_reduce), | |||
5252 | Args); | |||
5253 | ||||
5254 | // 5. Build switch(res) | |||
5255 | llvm::BasicBlock *DefaultBB = CGF.createBasicBlock(".omp.reduction.default"); | |||
5256 | llvm::SwitchInst *SwInst = | |||
5257 | CGF.Builder.CreateSwitch(Res, DefaultBB, /*NumCases=*/2); | |||
5258 | ||||
5259 | // 6. Build case 1: | |||
5260 | // ... | |||
5261 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | |||
5262 | // ... | |||
5263 | // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | |||
5264 | // break; | |||
5265 | llvm::BasicBlock *Case1BB = CGF.createBasicBlock(".omp.reduction.case1"); | |||
5266 | SwInst->addCase(CGF.Builder.getInt32(1), Case1BB); | |||
5267 | CGF.EmitBlock(Case1BB); | |||
5268 | ||||
5269 | // Add emission of __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | |||
5270 | llvm::Value *EndArgs[] = { | |||
5271 | IdentTLoc, // ident_t *<loc> | |||
5272 | ThreadId, // i32 <gtid> | |||
5273 | Lock // kmp_critical_name *&<lock> | |||
5274 | }; | |||
5275 | auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps]( | |||
5276 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5277 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
5278 | const auto *IPriv = Privates.begin(); | |||
5279 | const auto *ILHS = LHSExprs.begin(); | |||
5280 | const auto *IRHS = RHSExprs.begin(); | |||
5281 | for (const Expr *E : ReductionOps) { | |||
5282 | RT.emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | |||
5283 | cast<DeclRefExpr>(*IRHS)); | |||
5284 | ++IPriv; | |||
5285 | ++ILHS; | |||
5286 | ++IRHS; | |||
5287 | } | |||
5288 | }; | |||
5289 | RegionCodeGenTy RCG(CodeGen); | |||
5290 | CommonActionTy Action( | |||
5291 | nullptr, llvm::None, | |||
5292 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5293 | CGM.getModule(), WithNowait ? OMPRTL___kmpc_end_reduce_nowait | |||
5294 | : OMPRTL___kmpc_end_reduce), | |||
5295 | EndArgs); | |||
5296 | RCG.setAction(Action); | |||
5297 | RCG(CGF); | |||
5298 | ||||
5299 | CGF.EmitBranch(DefaultBB); | |||
5300 | ||||
5301 | // 7. Build case 2: | |||
5302 | // ... | |||
5303 | // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); | |||
5304 | // ... | |||
5305 | // break; | |||
5306 | llvm::BasicBlock *Case2BB = CGF.createBasicBlock(".omp.reduction.case2"); | |||
5307 | SwInst->addCase(CGF.Builder.getInt32(2), Case2BB); | |||
5308 | CGF.EmitBlock(Case2BB); | |||
5309 | ||||
5310 | auto &&AtomicCodeGen = [Loc, Privates, LHSExprs, RHSExprs, ReductionOps]( | |||
5311 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5312 | const auto *ILHS = LHSExprs.begin(); | |||
5313 | const auto *IRHS = RHSExprs.begin(); | |||
5314 | const auto *IPriv = Privates.begin(); | |||
5315 | for (const Expr *E : ReductionOps) { | |||
5316 | const Expr *XExpr = nullptr; | |||
5317 | const Expr *EExpr = nullptr; | |||
5318 | const Expr *UpExpr = nullptr; | |||
5319 | BinaryOperatorKind BO = BO_Comma; | |||
5320 | if (const auto *BO = dyn_cast<BinaryOperator>(E)) { | |||
5321 | if (BO->getOpcode() == BO_Assign) { | |||
5322 | XExpr = BO->getLHS(); | |||
5323 | UpExpr = BO->getRHS(); | |||
5324 | } | |||
5325 | } | |||
5326 | // Try to emit update expression as a simple atomic. | |||
5327 | const Expr *RHSExpr = UpExpr; | |||
5328 | if (RHSExpr) { | |||
5329 | // Analyze RHS part of the whole expression. | |||
5330 | if (const auto *ACO = dyn_cast<AbstractConditionalOperator>( | |||
5331 | RHSExpr->IgnoreParenImpCasts())) { | |||
5332 | // If this is a conditional operator, analyze its condition for | |||
5333 | // min/max reduction operator. | |||
5334 | RHSExpr = ACO->getCond(); | |||
5335 | } | |||
5336 | if (const auto *BORHS = | |||
5337 | dyn_cast<BinaryOperator>(RHSExpr->IgnoreParenImpCasts())) { | |||
5338 | EExpr = BORHS->getRHS(); | |||
5339 | BO = BORHS->getOpcode(); | |||
5340 | } | |||
5341 | } | |||
5342 | if (XExpr) { | |||
5343 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | |||
5344 | auto &&AtomicRedGen = [BO, VD, | |||
5345 | Loc](CodeGenFunction &CGF, const Expr *XExpr, | |||
5346 | const Expr *EExpr, const Expr *UpExpr) { | |||
5347 | LValue X = CGF.EmitLValue(XExpr); | |||
5348 | RValue E; | |||
5349 | if (EExpr) | |||
5350 | E = CGF.EmitAnyExpr(EExpr); | |||
5351 | CGF.EmitOMPAtomicSimpleUpdateExpr( | |||
5352 | X, E, BO, /*IsXLHSInRHSPart=*/true, | |||
5353 | llvm::AtomicOrdering::Monotonic, Loc, | |||
5354 | [&CGF, UpExpr, VD, Loc](RValue XRValue) { | |||
5355 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | |||
5356 | Address LHSTemp = CGF.CreateMemTemp(VD->getType()); | |||
5357 | CGF.emitOMPSimpleStore( | |||
5358 | CGF.MakeAddrLValue(LHSTemp, VD->getType()), XRValue, | |||
5359 | VD->getType().getNonReferenceType(), Loc); | |||
5360 | PrivateScope.addPrivate(VD, LHSTemp); | |||
5361 | (void)PrivateScope.Privatize(); | |||
5362 | return CGF.EmitAnyExpr(UpExpr); | |||
5363 | }); | |||
5364 | }; | |||
5365 | if ((*IPriv)->getType()->isArrayType()) { | |||
5366 | // Emit atomic reduction for array section. | |||
5367 | const auto *RHSVar = | |||
5368 | cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | |||
5369 | EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), VD, RHSVar, | |||
5370 | AtomicRedGen, XExpr, EExpr, UpExpr); | |||
5371 | } else { | |||
5372 | // Emit atomic reduction for array subscript or single variable. | |||
5373 | AtomicRedGen(CGF, XExpr, EExpr, UpExpr); | |||
5374 | } | |||
5375 | } else { | |||
5376 | // Emit as a critical region. | |||
5377 | auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *, | |||
5378 | const Expr *, const Expr *) { | |||
5379 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
5380 | std::string Name = RT.getName({"atomic_reduction"}); | |||
5381 | RT.emitCriticalRegion( | |||
5382 | CGF, Name, | |||
5383 | [=](CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5384 | Action.Enter(CGF); | |||
5385 | emitReductionCombiner(CGF, E); | |||
5386 | }, | |||
5387 | Loc); | |||
5388 | }; | |||
5389 | if ((*IPriv)->getType()->isArrayType()) { | |||
5390 | const auto *LHSVar = | |||
5391 | cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | |||
5392 | const auto *RHSVar = | |||
5393 | cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | |||
5394 | EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), LHSVar, RHSVar, | |||
5395 | CritRedGen); | |||
5396 | } else { | |||
5397 | CritRedGen(CGF, nullptr, nullptr, nullptr); | |||
5398 | } | |||
5399 | } | |||
5400 | ++ILHS; | |||
5401 | ++IRHS; | |||
5402 | ++IPriv; | |||
5403 | } | |||
5404 | }; | |||
5405 | RegionCodeGenTy AtomicRCG(AtomicCodeGen); | |||
5406 | if (!WithNowait) { | |||
5407 | // Add emission of __kmpc_end_reduce(<loc>, <gtid>, &<lock>); | |||
5408 | llvm::Value *EndArgs[] = { | |||
5409 | IdentTLoc, // ident_t *<loc> | |||
5410 | ThreadId, // i32 <gtid> | |||
5411 | Lock // kmp_critical_name *&<lock> | |||
5412 | }; | |||
5413 | CommonActionTy Action(nullptr, llvm::None, | |||
5414 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5415 | CGM.getModule(), OMPRTL___kmpc_end_reduce), | |||
5416 | EndArgs); | |||
5417 | AtomicRCG.setAction(Action); | |||
5418 | AtomicRCG(CGF); | |||
5419 | } else { | |||
5420 | AtomicRCG(CGF); | |||
5421 | } | |||
5422 | ||||
5423 | CGF.EmitBranch(DefaultBB); | |||
5424 | CGF.EmitBlock(DefaultBB, /*IsFinished=*/true); | |||
5425 | } | |||
5426 | ||||
5427 | /// Generates unique name for artificial threadprivate variables. | |||
5428 | /// Format is: <Prefix> "." <Decl_mangled_name> "_" "<Decl_start_loc_raw_enc>" | |||
5429 | static std::string generateUniqueName(CodeGenModule &CGM, StringRef Prefix, | |||
5430 | const Expr *Ref) { | |||
5431 | SmallString<256> Buffer; | |||
5432 | llvm::raw_svector_ostream Out(Buffer); | |||
5433 | const clang::DeclRefExpr *DE; | |||
5434 | const VarDecl *D = ::getBaseDecl(Ref, DE); | |||
5435 | if (!D) | |||
5436 | D = cast<VarDecl>(cast<DeclRefExpr>(Ref)->getDecl()); | |||
5437 | D = D->getCanonicalDecl(); | |||
5438 | std::string Name = CGM.getOpenMPRuntime().getName( | |||
5439 | {D->isLocalVarDeclOrParm() ? D->getName() : CGM.getMangledName(D)}); | |||
5440 | Out << Prefix << Name << "_" | |||
5441 | << D->getCanonicalDecl()->getBeginLoc().getRawEncoding(); | |||
5442 | return std::string(Out.str()); | |||
5443 | } | |||
5444 | ||||
5445 | /// Emits reduction initializer function: | |||
5446 | /// \code | |||
5447 | /// void @.red_init(void* %arg, void* %orig) { | |||
5448 | /// %0 = bitcast void* %arg to <type>* | |||
5449 | /// store <type> <init>, <type>* %0 | |||
5450 | /// ret void | |||
5451 | /// } | |||
5452 | /// \endcode | |||
5453 | static llvm::Value *emitReduceInitFunction(CodeGenModule &CGM, | |||
5454 | SourceLocation Loc, | |||
5455 | ReductionCodeGen &RCG, unsigned N) { | |||
5456 | ASTContext &C = CGM.getContext(); | |||
5457 | QualType VoidPtrTy = C.VoidPtrTy; | |||
5458 | VoidPtrTy.addRestrict(); | |||
5459 | FunctionArgList Args; | |||
5460 | ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, VoidPtrTy, | |||
5461 | ImplicitParamDecl::Other); | |||
5462 | ImplicitParamDecl ParamOrig(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, VoidPtrTy, | |||
5463 | ImplicitParamDecl::Other); | |||
5464 | Args.emplace_back(&Param); | |||
5465 | Args.emplace_back(&ParamOrig); | |||
5466 | const auto &FnInfo = | |||
5467 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
5468 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
5469 | std::string Name = CGM.getOpenMPRuntime().getName({"red_init", ""}); | |||
5470 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
5471 | Name, &CGM.getModule()); | |||
5472 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
5473 | Fn->setDoesNotRecurse(); | |||
5474 | CodeGenFunction CGF(CGM); | |||
5475 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
5476 | QualType PrivateType = RCG.getPrivateType(N); | |||
5477 | Address PrivateAddr = CGF.EmitLoadOfPointer( | |||
5478 | CGF.Builder.CreateElementBitCast( | |||
5479 | CGF.GetAddrOfLocalVar(&Param), | |||
5480 | CGF.ConvertTypeForMem(PrivateType)->getPointerTo()), | |||
5481 | C.getPointerType(PrivateType)->castAs<PointerType>()); | |||
5482 | llvm::Value *Size = nullptr; | |||
5483 | // If the size of the reduction item is non-constant, load it from global | |||
5484 | // threadprivate variable. | |||
5485 | if (RCG.getSizes(N).second) { | |||
5486 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
5487 | CGF, CGM.getContext().getSizeType(), | |||
5488 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
5489 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | |||
5490 | CGM.getContext().getSizeType(), Loc); | |||
5491 | } | |||
5492 | RCG.emitAggregateType(CGF, N, Size); | |||
5493 | Address OrigAddr = Address::invalid(); | |||
5494 | // If initializer uses initializer from declare reduction construct, emit a | |||
5495 | // pointer to the address of the original reduction item (reuired by reduction | |||
5496 | // initializer) | |||
5497 | if (RCG.usesReductionInitializer(N)) { | |||
5498 | Address SharedAddr = CGF.GetAddrOfLocalVar(&ParamOrig); | |||
5499 | OrigAddr = CGF.EmitLoadOfPointer( | |||
5500 | SharedAddr, | |||
5501 | CGM.getContext().VoidPtrTy.castAs<PointerType>()->getTypePtr()); | |||
5502 | } | |||
5503 | // Emit the initializer: | |||
5504 | // %0 = bitcast void* %arg to <type>* | |||
5505 | // store <type> <init>, <type>* %0 | |||
5506 | RCG.emitInitialization(CGF, N, PrivateAddr, OrigAddr, | |||
5507 | [](CodeGenFunction &) { return false; }); | |||
5508 | CGF.FinishFunction(); | |||
5509 | return Fn; | |||
5510 | } | |||
5511 | ||||
5512 | /// Emits reduction combiner function: | |||
5513 | /// \code | |||
5514 | /// void @.red_comb(void* %arg0, void* %arg1) { | |||
5515 | /// %lhs = bitcast void* %arg0 to <type>* | |||
5516 | /// %rhs = bitcast void* %arg1 to <type>* | |||
5517 | /// %2 = <ReductionOp>(<type>* %lhs, <type>* %rhs) | |||
5518 | /// store <type> %2, <type>* %lhs | |||
5519 | /// ret void | |||
5520 | /// } | |||
5521 | /// \endcode | |||
5522 | static llvm::Value *emitReduceCombFunction(CodeGenModule &CGM, | |||
5523 | SourceLocation Loc, | |||
5524 | ReductionCodeGen &RCG, unsigned N, | |||
5525 | const Expr *ReductionOp, | |||
5526 | const Expr *LHS, const Expr *RHS, | |||
5527 | const Expr *PrivateRef) { | |||
5528 | ASTContext &C = CGM.getContext(); | |||
5529 | const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(LHS)->getDecl()); | |||
5530 | const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(RHS)->getDecl()); | |||
5531 | FunctionArgList Args; | |||
5532 | ImplicitParamDecl ParamInOut(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
5533 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
5534 | ImplicitParamDecl ParamIn(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5535 | ImplicitParamDecl::Other); | |||
5536 | Args.emplace_back(&ParamInOut); | |||
5537 | Args.emplace_back(&ParamIn); | |||
5538 | const auto &FnInfo = | |||
5539 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
5540 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
5541 | std::string Name = CGM.getOpenMPRuntime().getName({"red_comb", ""}); | |||
5542 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
5543 | Name, &CGM.getModule()); | |||
5544 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
5545 | Fn->setDoesNotRecurse(); | |||
5546 | CodeGenFunction CGF(CGM); | |||
5547 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
5548 | llvm::Value *Size = nullptr; | |||
5549 | // If the size of the reduction item is non-constant, load it from global | |||
5550 | // threadprivate variable. | |||
5551 | if (RCG.getSizes(N).second) { | |||
5552 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
5553 | CGF, CGM.getContext().getSizeType(), | |||
5554 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
5555 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | |||
5556 | CGM.getContext().getSizeType(), Loc); | |||
5557 | } | |||
5558 | RCG.emitAggregateType(CGF, N, Size); | |||
5559 | // Remap lhs and rhs variables to the addresses of the function arguments. | |||
5560 | // %lhs = bitcast void* %arg0 to <type>* | |||
5561 | // %rhs = bitcast void* %arg1 to <type>* | |||
5562 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | |||
5563 | PrivateScope.addPrivate( | |||
5564 | LHSVD, | |||
5565 | // Pull out the pointer to the variable. | |||
5566 | CGF.EmitLoadOfPointer( | |||
5567 | CGF.Builder.CreateElementBitCast( | |||
5568 | CGF.GetAddrOfLocalVar(&ParamInOut), | |||
5569 | CGF.ConvertTypeForMem(LHSVD->getType())->getPointerTo()), | |||
5570 | C.getPointerType(LHSVD->getType())->castAs<PointerType>())); | |||
5571 | PrivateScope.addPrivate( | |||
5572 | RHSVD, | |||
5573 | // Pull out the pointer to the variable. | |||
5574 | CGF.EmitLoadOfPointer( | |||
5575 | CGF.Builder.CreateElementBitCast( | |||
5576 | CGF.GetAddrOfLocalVar(&ParamIn), | |||
5577 | CGF.ConvertTypeForMem(RHSVD->getType())->getPointerTo()), | |||
5578 | C.getPointerType(RHSVD->getType())->castAs<PointerType>())); | |||
5579 | PrivateScope.Privatize(); | |||
5580 | // Emit the combiner body: | |||
5581 | // %2 = <ReductionOp>(<type> *%lhs, <type> *%rhs) | |||
5582 | // store <type> %2, <type>* %lhs | |||
5583 | CGM.getOpenMPRuntime().emitSingleReductionCombiner( | |||
5584 | CGF, ReductionOp, PrivateRef, cast<DeclRefExpr>(LHS), | |||
5585 | cast<DeclRefExpr>(RHS)); | |||
5586 | CGF.FinishFunction(); | |||
5587 | return Fn; | |||
5588 | } | |||
5589 | ||||
5590 | /// Emits reduction finalizer function: | |||
5591 | /// \code | |||
5592 | /// void @.red_fini(void* %arg) { | |||
5593 | /// %0 = bitcast void* %arg to <type>* | |||
5594 | /// <destroy>(<type>* %0) | |||
5595 | /// ret void | |||
5596 | /// } | |||
5597 | /// \endcode | |||
5598 | static llvm::Value *emitReduceFiniFunction(CodeGenModule &CGM, | |||
5599 | SourceLocation Loc, | |||
5600 | ReductionCodeGen &RCG, unsigned N) { | |||
5601 | if (!RCG.needCleanups(N)) | |||
5602 | return nullptr; | |||
5603 | ASTContext &C = CGM.getContext(); | |||
5604 | FunctionArgList Args; | |||
5605 | ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5606 | ImplicitParamDecl::Other); | |||
5607 | Args.emplace_back(&Param); | |||
5608 | const auto &FnInfo = | |||
5609 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
5610 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
5611 | std::string Name = CGM.getOpenMPRuntime().getName({"red_fini", ""}); | |||
5612 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
5613 | Name, &CGM.getModule()); | |||
5614 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
5615 | Fn->setDoesNotRecurse(); | |||
5616 | CodeGenFunction CGF(CGM); | |||
5617 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
5618 | Address PrivateAddr = CGF.EmitLoadOfPointer( | |||
5619 | CGF.GetAddrOfLocalVar(&Param), C.VoidPtrTy.castAs<PointerType>()); | |||
5620 | llvm::Value *Size = nullptr; | |||
5621 | // If the size of the reduction item is non-constant, load it from global | |||
5622 | // threadprivate variable. | |||
5623 | if (RCG.getSizes(N).second) { | |||
5624 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
5625 | CGF, CGM.getContext().getSizeType(), | |||
5626 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
5627 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | |||
5628 | CGM.getContext().getSizeType(), Loc); | |||
5629 | } | |||
5630 | RCG.emitAggregateType(CGF, N, Size); | |||
5631 | // Emit the finalizer body: | |||
5632 | // <destroy>(<type>* %0) | |||
5633 | RCG.emitCleanups(CGF, N, PrivateAddr); | |||
5634 | CGF.FinishFunction(Loc); | |||
5635 | return Fn; | |||
5636 | } | |||
5637 | ||||
5638 | llvm::Value *CGOpenMPRuntime::emitTaskReductionInit( | |||
5639 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs, | |||
5640 | ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) { | |||
5641 | if (!CGF.HaveInsertPoint() || Data.ReductionVars.empty()) | |||
5642 | return nullptr; | |||
5643 | ||||
5644 | // Build typedef struct: | |||
5645 | // kmp_taskred_input { | |||
5646 | // void *reduce_shar; // shared reduction item | |||
5647 | // void *reduce_orig; // original reduction item used for initialization | |||
5648 | // size_t reduce_size; // size of data item | |||
5649 | // void *reduce_init; // data initialization routine | |||
5650 | // void *reduce_fini; // data finalization routine | |||
5651 | // void *reduce_comb; // data combiner routine | |||
5652 | // kmp_task_red_flags_t flags; // flags for additional info from compiler | |||
5653 | // } kmp_taskred_input_t; | |||
5654 | ASTContext &C = CGM.getContext(); | |||
5655 | RecordDecl *RD = C.buildImplicitRecord("kmp_taskred_input_t"); | |||
5656 | RD->startDefinition(); | |||
5657 | const FieldDecl *SharedFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
5658 | const FieldDecl *OrigFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
5659 | const FieldDecl *SizeFD = addFieldToRecordDecl(C, RD, C.getSizeType()); | |||
5660 | const FieldDecl *InitFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
5661 | const FieldDecl *FiniFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
5662 | const FieldDecl *CombFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
5663 | const FieldDecl *FlagsFD = addFieldToRecordDecl( | |||
5664 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false)); | |||
5665 | RD->completeDefinition(); | |||
5666 | QualType RDType = C.getRecordType(RD); | |||
5667 | unsigned Size = Data.ReductionVars.size(); | |||
5668 | llvm::APInt ArraySize(/*numBits=*/64, Size); | |||
5669 | QualType ArrayRDType = C.getConstantArrayType( | |||
5670 | RDType, ArraySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
5671 | // kmp_task_red_input_t .rd_input.[Size]; | |||
5672 | Address TaskRedInput = CGF.CreateMemTemp(ArrayRDType, ".rd_input."); | |||
5673 | ReductionCodeGen RCG(Data.ReductionVars, Data.ReductionOrigs, | |||
5674 | Data.ReductionCopies, Data.ReductionOps); | |||
5675 | for (unsigned Cnt = 0; Cnt < Size; ++Cnt) { | |||
5676 | // kmp_task_red_input_t &ElemLVal = .rd_input.[Cnt]; | |||
5677 | llvm::Value *Idxs[] = {llvm::ConstantInt::get(CGM.SizeTy, /*V=*/0), | |||
5678 | llvm::ConstantInt::get(CGM.SizeTy, Cnt)}; | |||
5679 | llvm::Value *GEP = CGF.EmitCheckedInBoundsGEP( | |||
5680 | TaskRedInput.getElementType(), TaskRedInput.getPointer(), Idxs, | |||
5681 | /*SignedIndices=*/false, /*IsSubtraction=*/false, Loc, | |||
5682 | ".rd_input.gep."); | |||
5683 | LValue ElemLVal = CGF.MakeNaturalAlignAddrLValue(GEP, RDType); | |||
5684 | // ElemLVal.reduce_shar = &Shareds[Cnt]; | |||
5685 | LValue SharedLVal = CGF.EmitLValueForField(ElemLVal, SharedFD); | |||
5686 | RCG.emitSharedOrigLValue(CGF, Cnt); | |||
5687 | llvm::Value *CastedShared = | |||
5688 | CGF.EmitCastToVoidPtr(RCG.getSharedLValue(Cnt).getPointer(CGF)); | |||
5689 | CGF.EmitStoreOfScalar(CastedShared, SharedLVal); | |||
5690 | // ElemLVal.reduce_orig = &Origs[Cnt]; | |||
5691 | LValue OrigLVal = CGF.EmitLValueForField(ElemLVal, OrigFD); | |||
5692 | llvm::Value *CastedOrig = | |||
5693 | CGF.EmitCastToVoidPtr(RCG.getOrigLValue(Cnt).getPointer(CGF)); | |||
5694 | CGF.EmitStoreOfScalar(CastedOrig, OrigLVal); | |||
5695 | RCG.emitAggregateType(CGF, Cnt); | |||
5696 | llvm::Value *SizeValInChars; | |||
5697 | llvm::Value *SizeVal; | |||
5698 | std::tie(SizeValInChars, SizeVal) = RCG.getSizes(Cnt); | |||
5699 | // We use delayed creation/initialization for VLAs and array sections. It is | |||
5700 | // required because runtime does not provide the way to pass the sizes of | |||
5701 | // VLAs/array sections to initializer/combiner/finalizer functions. Instead | |||
5702 | // threadprivate global variables are used to store these values and use | |||
5703 | // them in the functions. | |||
5704 | bool DelayedCreation = !!SizeVal; | |||
5705 | SizeValInChars = CGF.Builder.CreateIntCast(SizeValInChars, CGM.SizeTy, | |||
5706 | /*isSigned=*/false); | |||
5707 | LValue SizeLVal = CGF.EmitLValueForField(ElemLVal, SizeFD); | |||
5708 | CGF.EmitStoreOfScalar(SizeValInChars, SizeLVal); | |||
5709 | // ElemLVal.reduce_init = init; | |||
5710 | LValue InitLVal = CGF.EmitLValueForField(ElemLVal, InitFD); | |||
5711 | llvm::Value *InitAddr = | |||
5712 | CGF.EmitCastToVoidPtr(emitReduceInitFunction(CGM, Loc, RCG, Cnt)); | |||
5713 | CGF.EmitStoreOfScalar(InitAddr, InitLVal); | |||
5714 | // ElemLVal.reduce_fini = fini; | |||
5715 | LValue FiniLVal = CGF.EmitLValueForField(ElemLVal, FiniFD); | |||
5716 | llvm::Value *Fini = emitReduceFiniFunction(CGM, Loc, RCG, Cnt); | |||
5717 | llvm::Value *FiniAddr = Fini | |||
5718 | ? CGF.EmitCastToVoidPtr(Fini) | |||
5719 | : llvm::ConstantPointerNull::get(CGM.VoidPtrTy); | |||
5720 | CGF.EmitStoreOfScalar(FiniAddr, FiniLVal); | |||
5721 | // ElemLVal.reduce_comb = comb; | |||
5722 | LValue CombLVal = CGF.EmitLValueForField(ElemLVal, CombFD); | |||
5723 | llvm::Value *CombAddr = CGF.EmitCastToVoidPtr(emitReduceCombFunction( | |||
5724 | CGM, Loc, RCG, Cnt, Data.ReductionOps[Cnt], LHSExprs[Cnt], | |||
5725 | RHSExprs[Cnt], Data.ReductionCopies[Cnt])); | |||
5726 | CGF.EmitStoreOfScalar(CombAddr, CombLVal); | |||
5727 | // ElemLVal.flags = 0; | |||
5728 | LValue FlagsLVal = CGF.EmitLValueForField(ElemLVal, FlagsFD); | |||
5729 | if (DelayedCreation) { | |||
5730 | CGF.EmitStoreOfScalar( | |||
5731 | llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1, /*isSigned=*/true), | |||
5732 | FlagsLVal); | |||
5733 | } else | |||
5734 | CGF.EmitNullInitialization(FlagsLVal.getAddress(CGF), | |||
5735 | FlagsLVal.getType()); | |||
5736 | } | |||
5737 | if (Data.IsReductionWithTaskMod) { | |||
5738 | // Build call void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int | |||
5739 | // is_ws, int num, void *data); | |||
5740 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc); | |||
5741 | llvm::Value *GTid = CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), | |||
5742 | CGM.IntTy, /*isSigned=*/true); | |||
5743 | llvm::Value *Args[] = { | |||
5744 | IdentTLoc, GTid, | |||
5745 | llvm::ConstantInt::get(CGM.IntTy, Data.IsWorksharingReduction ? 1 : 0, | |||
5746 | /*isSigned=*/true), | |||
5747 | llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true), | |||
5748 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5749 | TaskRedInput.getPointer(), CGM.VoidPtrTy)}; | |||
5750 | return CGF.EmitRuntimeCall( | |||
5751 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5752 | CGM.getModule(), OMPRTL___kmpc_taskred_modifier_init), | |||
5753 | Args); | |||
5754 | } | |||
5755 | // Build call void *__kmpc_taskred_init(int gtid, int num_data, void *data); | |||
5756 | llvm::Value *Args[] = { | |||
5757 | CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), CGM.IntTy, | |||
5758 | /*isSigned=*/true), | |||
5759 | llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true), | |||
5760 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TaskRedInput.getPointer(), | |||
5761 | CGM.VoidPtrTy)}; | |||
5762 | return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
5763 | CGM.getModule(), OMPRTL___kmpc_taskred_init), | |||
5764 | Args); | |||
5765 | } | |||
5766 | ||||
5767 | void CGOpenMPRuntime::emitTaskReductionFini(CodeGenFunction &CGF, | |||
5768 | SourceLocation Loc, | |||
5769 | bool IsWorksharingReduction) { | |||
5770 | // Build call void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int | |||
5771 | // is_ws, int num, void *data); | |||
5772 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc); | |||
5773 | llvm::Value *GTid = CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), | |||
5774 | CGM.IntTy, /*isSigned=*/true); | |||
5775 | llvm::Value *Args[] = {IdentTLoc, GTid, | |||
5776 | llvm::ConstantInt::get(CGM.IntTy, | |||
5777 | IsWorksharingReduction ? 1 : 0, | |||
5778 | /*isSigned=*/true)}; | |||
5779 | (void)CGF.EmitRuntimeCall( | |||
5780 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5781 | CGM.getModule(), OMPRTL___kmpc_task_reduction_modifier_fini), | |||
5782 | Args); | |||
5783 | } | |||
5784 | ||||
5785 | void CGOpenMPRuntime::emitTaskReductionFixups(CodeGenFunction &CGF, | |||
5786 | SourceLocation Loc, | |||
5787 | ReductionCodeGen &RCG, | |||
5788 | unsigned N) { | |||
5789 | auto Sizes = RCG.getSizes(N); | |||
5790 | // Emit threadprivate global variable if the type is non-constant | |||
5791 | // (Sizes.second = nullptr). | |||
5792 | if (Sizes.second) { | |||
5793 | llvm::Value *SizeVal = CGF.Builder.CreateIntCast(Sizes.second, CGM.SizeTy, | |||
5794 | /*isSigned=*/false); | |||
5795 | Address SizeAddr = getAddrOfArtificialThreadPrivate( | |||
5796 | CGF, CGM.getContext().getSizeType(), | |||
5797 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
5798 | CGF.Builder.CreateStore(SizeVal, SizeAddr, /*IsVolatile=*/false); | |||
5799 | } | |||
5800 | } | |||
5801 | ||||
5802 | Address CGOpenMPRuntime::getTaskReductionItem(CodeGenFunction &CGF, | |||
5803 | SourceLocation Loc, | |||
5804 | llvm::Value *ReductionsPtr, | |||
5805 | LValue SharedLVal) { | |||
5806 | // Build call void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void | |||
5807 | // *d); | |||
5808 | llvm::Value *Args[] = {CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), | |||
5809 | CGM.IntTy, | |||
5810 | /*isSigned=*/true), | |||
5811 | ReductionsPtr, | |||
5812 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5813 | SharedLVal.getPointer(CGF), CGM.VoidPtrTy)}; | |||
5814 | return Address( | |||
5815 | CGF.EmitRuntimeCall( | |||
5816 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5817 | CGM.getModule(), OMPRTL___kmpc_task_reduction_get_th_data), | |||
5818 | Args), | |||
5819 | CGF.Int8Ty, SharedLVal.getAlignment()); | |||
5820 | } | |||
5821 | ||||
5822 | void CGOpenMPRuntime::emitTaskwaitCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
5823 | const OMPTaskDataTy &Data) { | |||
5824 | if (!CGF.HaveInsertPoint()) | |||
5825 | return; | |||
5826 | ||||
5827 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder && Data.Dependences.empty()) { | |||
5828 | // TODO: Need to support taskwait with dependences in the OpenMPIRBuilder. | |||
5829 | OMPBuilder.createTaskwait(CGF.Builder); | |||
5830 | } else { | |||
5831 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
5832 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | |||
5833 | auto &M = CGM.getModule(); | |||
5834 | Address DependenciesArray = Address::invalid(); | |||
5835 | llvm::Value *NumOfElements; | |||
5836 | std::tie(NumOfElements, DependenciesArray) = | |||
5837 | emitDependClause(CGF, Data.Dependences, Loc); | |||
5838 | llvm::Value *DepWaitTaskArgs[6]; | |||
5839 | if (!Data.Dependences.empty()) { | |||
5840 | DepWaitTaskArgs[0] = UpLoc; | |||
5841 | DepWaitTaskArgs[1] = ThreadID; | |||
5842 | DepWaitTaskArgs[2] = NumOfElements; | |||
5843 | DepWaitTaskArgs[3] = DependenciesArray.getPointer(); | |||
5844 | DepWaitTaskArgs[4] = CGF.Builder.getInt32(0); | |||
5845 | DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
5846 | ||||
5847 | CodeGenFunction::RunCleanupsScope LocalScope(CGF); | |||
5848 | ||||
5849 | // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid, | |||
5850 | // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 | |||
5851 | // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info | |||
5852 | // is specified. | |||
5853 | CGF.EmitRuntimeCall( | |||
5854 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_omp_wait_deps), | |||
5855 | DepWaitTaskArgs); | |||
5856 | ||||
5857 | } else { | |||
5858 | ||||
5859 | // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 | |||
5860 | // global_tid); | |||
5861 | llvm::Value *Args[] = {UpLoc, ThreadID}; | |||
5862 | // Ignore return result until untied tasks are supported. | |||
5863 | CGF.EmitRuntimeCall( | |||
5864 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_omp_taskwait), | |||
5865 | Args); | |||
5866 | } | |||
5867 | } | |||
5868 | ||||
5869 | if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
5870 | Region->emitUntiedSwitch(CGF); | |||
5871 | } | |||
5872 | ||||
5873 | void CGOpenMPRuntime::emitInlinedDirective(CodeGenFunction &CGF, | |||
5874 | OpenMPDirectiveKind InnerKind, | |||
5875 | const RegionCodeGenTy &CodeGen, | |||
5876 | bool HasCancel) { | |||
5877 | if (!CGF.HaveInsertPoint()) | |||
5878 | return; | |||
5879 | InlinedOpenMPRegionRAII Region(CGF, CodeGen, InnerKind, HasCancel, | |||
5880 | InnerKind != OMPD_critical && | |||
5881 | InnerKind != OMPD_master && | |||
5882 | InnerKind != OMPD_masked); | |||
5883 | CGF.CapturedStmtInfo->EmitBody(CGF, /*S=*/nullptr); | |||
5884 | } | |||
5885 | ||||
5886 | namespace { | |||
5887 | enum RTCancelKind { | |||
5888 | CancelNoreq = 0, | |||
5889 | CancelParallel = 1, | |||
5890 | CancelLoop = 2, | |||
5891 | CancelSections = 3, | |||
5892 | CancelTaskgroup = 4 | |||
5893 | }; | |||
5894 | } // anonymous namespace | |||
5895 | ||||
5896 | static RTCancelKind getCancellationKind(OpenMPDirectiveKind CancelRegion) { | |||
5897 | RTCancelKind CancelKind = CancelNoreq; | |||
5898 | if (CancelRegion == OMPD_parallel) | |||
5899 | CancelKind = CancelParallel; | |||
5900 | else if (CancelRegion == OMPD_for) | |||
5901 | CancelKind = CancelLoop; | |||
5902 | else if (CancelRegion == OMPD_sections) | |||
5903 | CancelKind = CancelSections; | |||
5904 | else { | |||
5905 | assert(CancelRegion == OMPD_taskgroup)(static_cast <bool> (CancelRegion == OMPD_taskgroup) ? void (0) : __assert_fail ("CancelRegion == OMPD_taskgroup", "clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 5905, __extension__ __PRETTY_FUNCTION__)); | |||
5906 | CancelKind = CancelTaskgroup; | |||
5907 | } | |||
5908 | return CancelKind; | |||
5909 | } | |||
5910 | ||||
5911 | void CGOpenMPRuntime::emitCancellationPointCall( | |||
5912 | CodeGenFunction &CGF, SourceLocation Loc, | |||
5913 | OpenMPDirectiveKind CancelRegion) { | |||
5914 | if (!CGF.HaveInsertPoint()) | |||
5915 | return; | |||
5916 | // Build call kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32 | |||
5917 | // global_tid, kmp_int32 cncl_kind); | |||
5918 | if (auto *OMPRegionInfo = | |||
5919 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
5920 | // For 'cancellation point taskgroup', the task region info may not have a | |||
5921 | // cancel. This may instead happen in another adjacent task. | |||
5922 | if (CancelRegion == OMPD_taskgroup || OMPRegionInfo->hasCancel()) { | |||
5923 | llvm::Value *Args[] = { | |||
5924 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
5925 | CGF.Builder.getInt32(getCancellationKind(CancelRegion))}; | |||
5926 | // Ignore return result until untied tasks are supported. | |||
5927 | llvm::Value *Result = CGF.EmitRuntimeCall( | |||
5928 | OMPBuilder.getOrCreateRuntimeFunction( | |||
5929 | CGM.getModule(), OMPRTL___kmpc_cancellationpoint), | |||
5930 | Args); | |||
5931 | // if (__kmpc_cancellationpoint()) { | |||
5932 | // call i32 @__kmpc_cancel_barrier( // for parallel cancellation only | |||
5933 | // exit from construct; | |||
5934 | // } | |||
5935 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | |||
5936 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | |||
5937 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | |||
5938 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | |||
5939 | CGF.EmitBlock(ExitBB); | |||
5940 | if (CancelRegion == OMPD_parallel) | |||
5941 | emitBarrierCall(CGF, Loc, OMPD_unknown, /*EmitChecks=*/false); | |||
5942 | // exit from construct; | |||
5943 | CodeGenFunction::JumpDest CancelDest = | |||
5944 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | |||
5945 | CGF.EmitBranchThroughCleanup(CancelDest); | |||
5946 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | |||
5947 | } | |||
5948 | } | |||
5949 | } | |||
5950 | ||||
5951 | void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
5952 | const Expr *IfCond, | |||
5953 | OpenMPDirectiveKind CancelRegion) { | |||
5954 | if (!CGF.HaveInsertPoint()) | |||
5955 | return; | |||
5956 | // Build call kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid, | |||
5957 | // kmp_int32 cncl_kind); | |||
5958 | auto &M = CGM.getModule(); | |||
5959 | if (auto *OMPRegionInfo = | |||
5960 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
5961 | auto &&ThenGen = [this, &M, Loc, CancelRegion, | |||
5962 | OMPRegionInfo](CodeGenFunction &CGF, PrePostActionTy &) { | |||
5963 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
5964 | llvm::Value *Args[] = { | |||
5965 | RT.emitUpdateLocation(CGF, Loc), RT.getThreadID(CGF, Loc), | |||
5966 | CGF.Builder.getInt32(getCancellationKind(CancelRegion))}; | |||
5967 | // Ignore return result until untied tasks are supported. | |||
5968 | llvm::Value *Result = CGF.EmitRuntimeCall( | |||
5969 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_cancel), Args); | |||
5970 | // if (__kmpc_cancel()) { | |||
5971 | // call i32 @__kmpc_cancel_barrier( // for parallel cancellation only | |||
5972 | // exit from construct; | |||
5973 | // } | |||
5974 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | |||
5975 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | |||
5976 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | |||
5977 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | |||
5978 | CGF.EmitBlock(ExitBB); | |||
5979 | if (CancelRegion == OMPD_parallel) | |||
5980 | RT.emitBarrierCall(CGF, Loc, OMPD_unknown, /*EmitChecks=*/false); | |||
5981 | // exit from construct; | |||
5982 | CodeGenFunction::JumpDest CancelDest = | |||
5983 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | |||
5984 | CGF.EmitBranchThroughCleanup(CancelDest); | |||
5985 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | |||
5986 | }; | |||
5987 | if (IfCond) { | |||
5988 | emitIfClause(CGF, IfCond, ThenGen, | |||
5989 | [](CodeGenFunction &, PrePostActionTy &) {}); | |||
5990 | } else { | |||
5991 | RegionCodeGenTy ThenRCG(ThenGen); | |||
5992 | ThenRCG(CGF); | |||
5993 | } | |||
5994 | } | |||
5995 | } | |||
5996 | ||||
5997 | namespace { | |||
5998 | /// Cleanup action for uses_allocators support. | |||
5999 | class OMPUsesAllocatorsActionTy final : public PrePostActionTy { | |||
6000 | ArrayRef<std::pair<const Expr *, const Expr *>> Allocators; | |||
6001 | ||||
6002 | public: | |||
6003 | OMPUsesAllocatorsActionTy( | |||
6004 | ArrayRef<std::pair<const Expr *, const Expr *>> Allocators) | |||
6005 | : Allocators(Allocators) {} | |||
6006 | void Enter(CodeGenFunction &CGF) override { | |||
6007 | if (!CGF.HaveInsertPoint()) | |||
6008 | return; | |||
6009 | for (const auto &AllocatorData : Allocators) { | |||
6010 | CGF.CGM.getOpenMPRuntime().emitUsesAllocatorsInit( | |||
6011 | CGF, AllocatorData.first, AllocatorData.second); | |||
6012 | } | |||
6013 | } | |||
6014 | void Exit(CodeGenFunction &CGF) override { | |||
6015 | if (!CGF.HaveInsertPoint()) | |||
6016 | return; | |||
6017 | for (const auto &AllocatorData : Allocators) { | |||
6018 | CGF.CGM.getOpenMPRuntime().emitUsesAllocatorsFini(CGF, | |||
6019 | AllocatorData.first); | |||
6020 | } | |||
6021 | } | |||
6022 | }; | |||
6023 | } // namespace | |||
6024 | ||||
6025 | void CGOpenMPRuntime::emitTargetOutlinedFunction( | |||
6026 | const OMPExecutableDirective &D, StringRef ParentName, | |||
6027 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
6028 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
6029 | assert(!ParentName.empty() && "Invalid target entry parent name!")(static_cast <bool> (!ParentName.empty() && "Invalid target entry parent name!" ) ? void (0) : __assert_fail ("!ParentName.empty() && \"Invalid target entry parent name!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6029, __extension__ __PRETTY_FUNCTION__)); | |||
6030 | HasEmittedTargetRegion = true; | |||
6031 | SmallVector<std::pair<const Expr *, const Expr *>, 4> Allocators; | |||
6032 | for (const auto *C : D.getClausesOfKind<OMPUsesAllocatorsClause>()) { | |||
6033 | for (unsigned I = 0, E = C->getNumberOfAllocators(); I < E; ++I) { | |||
6034 | const OMPUsesAllocatorsClause::Data D = C->getAllocatorData(I); | |||
6035 | if (!D.AllocatorTraits) | |||
6036 | continue; | |||
6037 | Allocators.emplace_back(D.Allocator, D.AllocatorTraits); | |||
6038 | } | |||
6039 | } | |||
6040 | OMPUsesAllocatorsActionTy UsesAllocatorAction(Allocators); | |||
6041 | CodeGen.setAction(UsesAllocatorAction); | |||
6042 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, | |||
6043 | IsOffloadEntry, CodeGen); | |||
6044 | } | |||
6045 | ||||
6046 | void CGOpenMPRuntime::emitUsesAllocatorsInit(CodeGenFunction &CGF, | |||
6047 | const Expr *Allocator, | |||
6048 | const Expr *AllocatorTraits) { | |||
6049 | llvm::Value *ThreadId = getThreadID(CGF, Allocator->getExprLoc()); | |||
6050 | ThreadId = CGF.Builder.CreateIntCast(ThreadId, CGF.IntTy, /*isSigned=*/true); | |||
6051 | // Use default memspace handle. | |||
6052 | llvm::Value *MemSpaceHandle = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
6053 | llvm::Value *NumTraits = llvm::ConstantInt::get( | |||
6054 | CGF.IntTy, cast<ConstantArrayType>( | |||
6055 | AllocatorTraits->getType()->getAsArrayTypeUnsafe()) | |||
6056 | ->getSize() | |||
6057 | .getLimitedValue()); | |||
6058 | LValue AllocatorTraitsLVal = CGF.EmitLValue(AllocatorTraits); | |||
6059 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
6060 | AllocatorTraitsLVal.getAddress(CGF), CGF.VoidPtrPtrTy, CGF.VoidPtrTy); | |||
6061 | AllocatorTraitsLVal = CGF.MakeAddrLValue(Addr, CGF.getContext().VoidPtrTy, | |||
6062 | AllocatorTraitsLVal.getBaseInfo(), | |||
6063 | AllocatorTraitsLVal.getTBAAInfo()); | |||
6064 | llvm::Value *Traits = | |||
6065 | CGF.EmitLoadOfScalar(AllocatorTraitsLVal, AllocatorTraits->getExprLoc()); | |||
6066 | ||||
6067 | llvm::Value *AllocatorVal = | |||
6068 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
6069 | CGM.getModule(), OMPRTL___kmpc_init_allocator), | |||
6070 | {ThreadId, MemSpaceHandle, NumTraits, Traits}); | |||
6071 | // Store to allocator. | |||
6072 | CGF.EmitVarDecl(*cast<VarDecl>( | |||
6073 | cast<DeclRefExpr>(Allocator->IgnoreParenImpCasts())->getDecl())); | |||
6074 | LValue AllocatorLVal = CGF.EmitLValue(Allocator->IgnoreParenImpCasts()); | |||
6075 | AllocatorVal = | |||
6076 | CGF.EmitScalarConversion(AllocatorVal, CGF.getContext().VoidPtrTy, | |||
6077 | Allocator->getType(), Allocator->getExprLoc()); | |||
6078 | CGF.EmitStoreOfScalar(AllocatorVal, AllocatorLVal); | |||
6079 | } | |||
6080 | ||||
6081 | void CGOpenMPRuntime::emitUsesAllocatorsFini(CodeGenFunction &CGF, | |||
6082 | const Expr *Allocator) { | |||
6083 | llvm::Value *ThreadId = getThreadID(CGF, Allocator->getExprLoc()); | |||
6084 | ThreadId = CGF.Builder.CreateIntCast(ThreadId, CGF.IntTy, /*isSigned=*/true); | |||
6085 | LValue AllocatorLVal = CGF.EmitLValue(Allocator->IgnoreParenImpCasts()); | |||
6086 | llvm::Value *AllocatorVal = | |||
6087 | CGF.EmitLoadOfScalar(AllocatorLVal, Allocator->getExprLoc()); | |||
6088 | AllocatorVal = CGF.EmitScalarConversion(AllocatorVal, Allocator->getType(), | |||
6089 | CGF.getContext().VoidPtrTy, | |||
6090 | Allocator->getExprLoc()); | |||
6091 | (void)CGF.EmitRuntimeCall( | |||
6092 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
6093 | OMPRTL___kmpc_destroy_allocator), | |||
6094 | {ThreadId, AllocatorVal}); | |||
6095 | } | |||
6096 | ||||
6097 | void CGOpenMPRuntime::emitTargetOutlinedFunctionHelper( | |||
6098 | const OMPExecutableDirective &D, StringRef ParentName, | |||
6099 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
6100 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
6101 | // Create a unique name for the entry function using the source location | |||
6102 | // information of the current target region. The name will be something like: | |||
6103 | // | |||
6104 | // __omp_offloading_DD_FFFF_PP_lBB[_CC] | |||
6105 | // | |||
6106 | // where DD_FFFF is an ID unique to the file (device and file IDs), PP is the | |||
6107 | // mangled name of the function that encloses the target region and BB is the | |||
6108 | // line number of the target region. CC is a count added when more than one | |||
6109 | // region is located at the same location. | |||
6110 | ||||
6111 | const bool BuildOutlinedFn = CGM.getLangOpts().OpenMPIsDevice || | |||
6112 | !CGM.getLangOpts().OpenMPOffloadMandatory; | |||
6113 | auto EntryInfo = | |||
6114 | getTargetEntryUniqueInfo(CGM.getContext(), D.getBeginLoc(), ParentName); | |||
6115 | ||||
6116 | SmallString<64> EntryFnName; | |||
6117 | OffloadEntriesInfoManager.getTargetRegionEntryFnName(EntryFnName, EntryInfo); | |||
6118 | ||||
6119 | const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target); | |||
6120 | ||||
6121 | CodeGenFunction CGF(CGM, true); | |||
6122 | CGOpenMPTargetRegionInfo CGInfo(CS, CodeGen, EntryFnName); | |||
6123 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6124 | ||||
6125 | if (BuildOutlinedFn) | |||
6126 | OutlinedFn = CGF.GenerateOpenMPCapturedStmtFunction(CS, D.getBeginLoc()); | |||
6127 | ||||
6128 | // If this target outline function is not an offload entry, we don't need to | |||
6129 | // register it. | |||
6130 | if (!IsOffloadEntry) | |||
6131 | return; | |||
6132 | ||||
6133 | // The target region ID is used by the runtime library to identify the current | |||
6134 | // target region, so it only has to be unique and not necessarily point to | |||
6135 | // anything. It could be the pointer to the outlined function that implements | |||
6136 | // the target region, but we aren't using that so that the compiler doesn't | |||
6137 | // need to keep that, and could therefore inline the host function if proven | |||
6138 | // worthwhile during optimization. In the other hand, if emitting code for the | |||
6139 | // device, the ID has to be the function address so that it can retrieved from | |||
6140 | // the offloading entry and launched by the runtime library. We also mark the | |||
6141 | // outlined function to have external linkage in case we are emitting code for | |||
6142 | // the device, because these functions will be entry points to the device. | |||
6143 | ||||
6144 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
6145 | OutlinedFnID = llvm::ConstantExpr::getBitCast(OutlinedFn, CGM.Int8PtrTy); | |||
6146 | OutlinedFn->setLinkage(llvm::GlobalValue::WeakODRLinkage); | |||
6147 | OutlinedFn->setDSOLocal(false); | |||
6148 | OutlinedFn->setVisibility(llvm::GlobalValue::ProtectedVisibility); | |||
6149 | if (CGM.getTriple().isAMDGCN()) | |||
6150 | OutlinedFn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL); | |||
6151 | } else { | |||
6152 | std::string Name = getName({EntryFnName, "region_id"}); | |||
6153 | OutlinedFnID = new llvm::GlobalVariable( | |||
6154 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
6155 | llvm::GlobalValue::WeakAnyLinkage, | |||
6156 | llvm::Constant::getNullValue(CGM.Int8Ty), Name); | |||
6157 | } | |||
6158 | ||||
6159 | // If we do not allow host fallback we still need a named address to use. | |||
6160 | llvm::Constant *TargetRegionEntryAddr = OutlinedFn; | |||
6161 | if (!BuildOutlinedFn) { | |||
6162 | assert(!CGM.getModule().getGlobalVariable(EntryFnName, true) &&(static_cast <bool> (!CGM.getModule().getGlobalVariable (EntryFnName, true) && "Named kernel already exists?" ) ? void (0) : __assert_fail ("!CGM.getModule().getGlobalVariable(EntryFnName, true) && \"Named kernel already exists?\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6163, __extension__ __PRETTY_FUNCTION__)) | |||
6163 | "Named kernel already exists?")(static_cast <bool> (!CGM.getModule().getGlobalVariable (EntryFnName, true) && "Named kernel already exists?" ) ? void (0) : __assert_fail ("!CGM.getModule().getGlobalVariable(EntryFnName, true) && \"Named kernel already exists?\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6163, __extension__ __PRETTY_FUNCTION__)); | |||
6164 | TargetRegionEntryAddr = new llvm::GlobalVariable( | |||
6165 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
6166 | llvm::GlobalValue::InternalLinkage, | |||
6167 | llvm::Constant::getNullValue(CGM.Int8Ty), EntryFnName); | |||
6168 | } | |||
6169 | ||||
6170 | // Register the information for the entry associated with this target region. | |||
6171 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | |||
6172 | EntryInfo, TargetRegionEntryAddr, OutlinedFnID, | |||
6173 | llvm::OffloadEntriesInfoManager::OMPTargetRegionEntryTargetRegion, | |||
6174 | CGM.getLangOpts().OpenMPIsDevice); | |||
6175 | ||||
6176 | // Add NumTeams and ThreadLimit attributes to the outlined GPU function | |||
6177 | int32_t DefaultValTeams = -1; | |||
6178 | getNumTeamsExprForTargetDirective(CGF, D, DefaultValTeams); | |||
6179 | if (DefaultValTeams > 0 && OutlinedFn) { | |||
6180 | OutlinedFn->addFnAttr("omp_target_num_teams", | |||
6181 | std::to_string(DefaultValTeams)); | |||
6182 | } | |||
6183 | int32_t DefaultValThreads = -1; | |||
6184 | getNumThreadsExprForTargetDirective(CGF, D, DefaultValThreads); | |||
6185 | if (DefaultValThreads > 0 && OutlinedFn) { | |||
6186 | OutlinedFn->addFnAttr("omp_target_thread_limit", | |||
6187 | std::to_string(DefaultValThreads)); | |||
6188 | } | |||
6189 | ||||
6190 | if (BuildOutlinedFn) | |||
6191 | CGM.getTargetCodeGenInfo().setTargetAttributes(nullptr, OutlinedFn, CGM); | |||
6192 | } | |||
6193 | ||||
6194 | /// Checks if the expression is constant or does not have non-trivial function | |||
6195 | /// calls. | |||
6196 | static bool isTrivial(ASTContext &Ctx, const Expr * E) { | |||
6197 | // We can skip constant expressions. | |||
6198 | // We can skip expressions with trivial calls or simple expressions. | |||
6199 | return (E->isEvaluatable(Ctx, Expr::SE_AllowUndefinedBehavior) || | |||
6200 | !E->hasNonTrivialCall(Ctx)) && | |||
6201 | !E->HasSideEffects(Ctx, /*IncludePossibleEffects=*/true); | |||
6202 | } | |||
6203 | ||||
6204 | const Stmt *CGOpenMPRuntime::getSingleCompoundChild(ASTContext &Ctx, | |||
6205 | const Stmt *Body) { | |||
6206 | const Stmt *Child = Body->IgnoreContainers(); | |||
6207 | while (const auto *C = dyn_cast_or_null<CompoundStmt>(Child)) { | |||
6208 | Child = nullptr; | |||
6209 | for (const Stmt *S : C->body()) { | |||
6210 | if (const auto *E = dyn_cast<Expr>(S)) { | |||
6211 | if (isTrivial(Ctx, E)) | |||
6212 | continue; | |||
6213 | } | |||
6214 | // Some of the statements can be ignored. | |||
6215 | if (isa<AsmStmt>(S) || isa<NullStmt>(S) || isa<OMPFlushDirective>(S) || | |||
6216 | isa<OMPBarrierDirective>(S) || isa<OMPTaskyieldDirective>(S)) | |||
6217 | continue; | |||
6218 | // Analyze declarations. | |||
6219 | if (const auto *DS = dyn_cast<DeclStmt>(S)) { | |||
6220 | if (llvm::all_of(DS->decls(), [](const Decl *D) { | |||
6221 | if (isa<EmptyDecl>(D) || isa<DeclContext>(D) || | |||
6222 | isa<TypeDecl>(D) || isa<PragmaCommentDecl>(D) || | |||
6223 | isa<PragmaDetectMismatchDecl>(D) || isa<UsingDecl>(D) || | |||
6224 | isa<UsingDirectiveDecl>(D) || | |||
6225 | isa<OMPDeclareReductionDecl>(D) || | |||
6226 | isa<OMPThreadPrivateDecl>(D) || isa<OMPAllocateDecl>(D)) | |||
6227 | return true; | |||
6228 | const auto *VD = dyn_cast<VarDecl>(D); | |||
6229 | if (!VD) | |||
6230 | return false; | |||
6231 | return VD->hasGlobalStorage() || !VD->isUsed(); | |||
6232 | })) | |||
6233 | continue; | |||
6234 | } | |||
6235 | // Found multiple children - cannot get the one child only. | |||
6236 | if (Child) | |||
6237 | return nullptr; | |||
6238 | Child = S; | |||
6239 | } | |||
6240 | if (Child) | |||
6241 | Child = Child->IgnoreContainers(); | |||
6242 | } | |||
6243 | return Child; | |||
6244 | } | |||
6245 | ||||
6246 | const Expr *CGOpenMPRuntime::getNumTeamsExprForTargetDirective( | |||
6247 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | |||
6248 | int32_t &DefaultVal) { | |||
6249 | ||||
6250 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
6251 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&(static_cast <bool> (isOpenMPTargetExecutionDirective(DirectiveKind ) && "Expected target-based executable directive.") ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6252, __extension__ __PRETTY_FUNCTION__)) | |||
6252 | "Expected target-based executable directive.")(static_cast <bool> (isOpenMPTargetExecutionDirective(DirectiveKind ) && "Expected target-based executable directive.") ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6252, __extension__ __PRETTY_FUNCTION__)); | |||
6253 | switch (DirectiveKind) { | |||
6254 | case OMPD_target: { | |||
6255 | const auto *CS = D.getInnermostCapturedStmt(); | |||
6256 | const auto *Body = | |||
6257 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | |||
6258 | const Stmt *ChildStmt = | |||
6259 | CGOpenMPRuntime::getSingleCompoundChild(CGF.getContext(), Body); | |||
6260 | if (const auto *NestedDir = | |||
6261 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
6262 | if (isOpenMPTeamsDirective(NestedDir->getDirectiveKind())) { | |||
6263 | if (NestedDir->hasClausesOfKind<OMPNumTeamsClause>()) { | |||
6264 | const Expr *NumTeams = | |||
6265 | NestedDir->getSingleClause<OMPNumTeamsClause>()->getNumTeams(); | |||
6266 | if (NumTeams->isIntegerConstantExpr(CGF.getContext())) | |||
6267 | if (auto Constant = | |||
6268 | NumTeams->getIntegerConstantExpr(CGF.getContext())) | |||
6269 | DefaultVal = Constant->getExtValue(); | |||
6270 | return NumTeams; | |||
6271 | } | |||
6272 | DefaultVal = 0; | |||
6273 | return nullptr; | |||
6274 | } | |||
6275 | if (isOpenMPParallelDirective(NestedDir->getDirectiveKind()) || | |||
6276 | isOpenMPSimdDirective(NestedDir->getDirectiveKind())) { | |||
6277 | DefaultVal = 1; | |||
6278 | return nullptr; | |||
6279 | } | |||
6280 | DefaultVal = 1; | |||
6281 | return nullptr; | |||
6282 | } | |||
6283 | // A value of -1 is used to check if we need to emit no teams region | |||
6284 | DefaultVal = -1; | |||
6285 | return nullptr; | |||
6286 | } | |||
6287 | case OMPD_target_teams: | |||
6288 | case OMPD_target_teams_distribute: | |||
6289 | case OMPD_target_teams_distribute_simd: | |||
6290 | case OMPD_target_teams_distribute_parallel_for: | |||
6291 | case OMPD_target_teams_distribute_parallel_for_simd: { | |||
6292 | if (D.hasClausesOfKind<OMPNumTeamsClause>()) { | |||
6293 | const Expr *NumTeams = | |||
6294 | D.getSingleClause<OMPNumTeamsClause>()->getNumTeams(); | |||
6295 | if (NumTeams->isIntegerConstantExpr(CGF.getContext())) | |||
6296 | if (auto Constant = NumTeams->getIntegerConstantExpr(CGF.getContext())) | |||
6297 | DefaultVal = Constant->getExtValue(); | |||
6298 | return NumTeams; | |||
6299 | } | |||
6300 | DefaultVal = 0; | |||
6301 | return nullptr; | |||
6302 | } | |||
6303 | case OMPD_target_parallel: | |||
6304 | case OMPD_target_parallel_for: | |||
6305 | case OMPD_target_parallel_for_simd: | |||
6306 | case OMPD_target_simd: | |||
6307 | DefaultVal = 1; | |||
6308 | return nullptr; | |||
6309 | case OMPD_parallel: | |||
6310 | case OMPD_for: | |||
6311 | case OMPD_parallel_for: | |||
6312 | case OMPD_parallel_master: | |||
6313 | case OMPD_parallel_sections: | |||
6314 | case OMPD_for_simd: | |||
6315 | case OMPD_parallel_for_simd: | |||
6316 | case OMPD_cancel: | |||
6317 | case OMPD_cancellation_point: | |||
6318 | case OMPD_ordered: | |||
6319 | case OMPD_threadprivate: | |||
6320 | case OMPD_allocate: | |||
6321 | case OMPD_task: | |||
6322 | case OMPD_simd: | |||
6323 | case OMPD_tile: | |||
6324 | case OMPD_unroll: | |||
6325 | case OMPD_sections: | |||
6326 | case OMPD_section: | |||
6327 | case OMPD_single: | |||
6328 | case OMPD_master: | |||
6329 | case OMPD_critical: | |||
6330 | case OMPD_taskyield: | |||
6331 | case OMPD_barrier: | |||
6332 | case OMPD_taskwait: | |||
6333 | case OMPD_taskgroup: | |||
6334 | case OMPD_atomic: | |||
6335 | case OMPD_flush: | |||
6336 | case OMPD_depobj: | |||
6337 | case OMPD_scan: | |||
6338 | case OMPD_teams: | |||
6339 | case OMPD_target_data: | |||
6340 | case OMPD_target_exit_data: | |||
6341 | case OMPD_target_enter_data: | |||
6342 | case OMPD_distribute: | |||
6343 | case OMPD_distribute_simd: | |||
6344 | case OMPD_distribute_parallel_for: | |||
6345 | case OMPD_distribute_parallel_for_simd: | |||
6346 | case OMPD_teams_distribute: | |||
6347 | case OMPD_teams_distribute_simd: | |||
6348 | case OMPD_teams_distribute_parallel_for: | |||
6349 | case OMPD_teams_distribute_parallel_for_simd: | |||
6350 | case OMPD_target_update: | |||
6351 | case OMPD_declare_simd: | |||
6352 | case OMPD_declare_variant: | |||
6353 | case OMPD_begin_declare_variant: | |||
6354 | case OMPD_end_declare_variant: | |||
6355 | case OMPD_declare_target: | |||
6356 | case OMPD_end_declare_target: | |||
6357 | case OMPD_declare_reduction: | |||
6358 | case OMPD_declare_mapper: | |||
6359 | case OMPD_taskloop: | |||
6360 | case OMPD_taskloop_simd: | |||
6361 | case OMPD_master_taskloop: | |||
6362 | case OMPD_master_taskloop_simd: | |||
6363 | case OMPD_parallel_master_taskloop: | |||
6364 | case OMPD_parallel_master_taskloop_simd: | |||
6365 | case OMPD_requires: | |||
6366 | case OMPD_metadirective: | |||
6367 | case OMPD_unknown: | |||
6368 | break; | |||
6369 | default: | |||
6370 | break; | |||
6371 | } | |||
6372 | llvm_unreachable("Unexpected directive kind.")::llvm::llvm_unreachable_internal("Unexpected directive kind." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6372); | |||
6373 | } | |||
6374 | ||||
6375 | llvm::Value *CGOpenMPRuntime::emitNumTeamsForTargetDirective( | |||
6376 | CodeGenFunction &CGF, const OMPExecutableDirective &D) { | |||
6377 | assert(!CGF.getLangOpts().OpenMPIsDevice &&(static_cast <bool> (!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? void (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6379, __extension__ __PRETTY_FUNCTION__)) | |||
6378 | "Clauses associated with the teams directive expected to be emitted "(static_cast <bool> (!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? void (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6379, __extension__ __PRETTY_FUNCTION__)) | |||
6379 | "only for the host!")(static_cast <bool> (!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? void (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6379, __extension__ __PRETTY_FUNCTION__)); | |||
6380 | CGBuilderTy &Bld = CGF.Builder; | |||
6381 | int32_t DefaultNT = -1; | |||
6382 | const Expr *NumTeams = getNumTeamsExprForTargetDirective(CGF, D, DefaultNT); | |||
6383 | if (NumTeams != nullptr) { | |||
6384 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
6385 | ||||
6386 | switch (DirectiveKind) { | |||
6387 | case OMPD_target: { | |||
6388 | const auto *CS = D.getInnermostCapturedStmt(); | |||
6389 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6390 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6391 | llvm::Value *NumTeamsVal = CGF.EmitScalarExpr(NumTeams, | |||
6392 | /*IgnoreResultAssign*/ true); | |||
6393 | return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty, | |||
6394 | /*isSigned=*/true); | |||
6395 | } | |||
6396 | case OMPD_target_teams: | |||
6397 | case OMPD_target_teams_distribute: | |||
6398 | case OMPD_target_teams_distribute_simd: | |||
6399 | case OMPD_target_teams_distribute_parallel_for: | |||
6400 | case OMPD_target_teams_distribute_parallel_for_simd: { | |||
6401 | CodeGenFunction::RunCleanupsScope NumTeamsScope(CGF); | |||
6402 | llvm::Value *NumTeamsVal = CGF.EmitScalarExpr(NumTeams, | |||
6403 | /*IgnoreResultAssign*/ true); | |||
6404 | return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty, | |||
6405 | /*isSigned=*/true); | |||
6406 | } | |||
6407 | default: | |||
6408 | break; | |||
6409 | } | |||
6410 | } | |||
6411 | ||||
6412 | return llvm::ConstantInt::get(CGF.Int32Ty, DefaultNT); | |||
6413 | } | |||
6414 | ||||
6415 | static llvm::Value *getNumThreads(CodeGenFunction &CGF, const CapturedStmt *CS, | |||
6416 | llvm::Value *DefaultThreadLimitVal) { | |||
6417 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6418 | CGF.getContext(), CS->getCapturedStmt()); | |||
6419 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | |||
6420 | if (isOpenMPParallelDirective(Dir->getDirectiveKind())) { | |||
6421 | llvm::Value *NumThreads = nullptr; | |||
6422 | llvm::Value *CondVal = nullptr; | |||
6423 | // Handle if clause. If if clause present, the number of threads is | |||
6424 | // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1. | |||
6425 | if (Dir->hasClausesOfKind<OMPIfClause>()) { | |||
6426 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6427 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6428 | const OMPIfClause *IfClause = nullptr; | |||
6429 | for (const auto *C : Dir->getClausesOfKind<OMPIfClause>()) { | |||
6430 | if (C->getNameModifier() == OMPD_unknown || | |||
6431 | C->getNameModifier() == OMPD_parallel) { | |||
6432 | IfClause = C; | |||
6433 | break; | |||
6434 | } | |||
6435 | } | |||
6436 | if (IfClause) { | |||
6437 | const Expr *Cond = IfClause->getCondition(); | |||
6438 | bool Result; | |||
6439 | if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) { | |||
6440 | if (!Result) | |||
6441 | return CGF.Builder.getInt32(1); | |||
6442 | } else { | |||
6443 | CodeGenFunction::LexicalScope Scope(CGF, Cond->getSourceRange()); | |||
6444 | if (const auto *PreInit = | |||
6445 | cast_or_null<DeclStmt>(IfClause->getPreInitStmt())) { | |||
6446 | for (const auto *I : PreInit->decls()) { | |||
6447 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | |||
6448 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | |||
6449 | } else { | |||
6450 | CodeGenFunction::AutoVarEmission Emission = | |||
6451 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | |||
6452 | CGF.EmitAutoVarCleanups(Emission); | |||
6453 | } | |||
6454 | } | |||
6455 | } | |||
6456 | CondVal = CGF.EvaluateExprAsBool(Cond); | |||
6457 | } | |||
6458 | } | |||
6459 | } | |||
6460 | // Check the value of num_threads clause iff if clause was not specified | |||
6461 | // or is not evaluated to false. | |||
6462 | if (Dir->hasClausesOfKind<OMPNumThreadsClause>()) { | |||
6463 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6464 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6465 | const auto *NumThreadsClause = | |||
6466 | Dir->getSingleClause<OMPNumThreadsClause>(); | |||
6467 | CodeGenFunction::LexicalScope Scope( | |||
6468 | CGF, NumThreadsClause->getNumThreads()->getSourceRange()); | |||
6469 | if (const auto *PreInit = | |||
6470 | cast_or_null<DeclStmt>(NumThreadsClause->getPreInitStmt())) { | |||
6471 | for (const auto *I : PreInit->decls()) { | |||
6472 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | |||
6473 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | |||
6474 | } else { | |||
6475 | CodeGenFunction::AutoVarEmission Emission = | |||
6476 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | |||
6477 | CGF.EmitAutoVarCleanups(Emission); | |||
6478 | } | |||
6479 | } | |||
6480 | } | |||
6481 | NumThreads = CGF.EmitScalarExpr(NumThreadsClause->getNumThreads()); | |||
6482 | NumThreads = CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, | |||
6483 | /*isSigned=*/false); | |||
6484 | if (DefaultThreadLimitVal) | |||
6485 | NumThreads = CGF.Builder.CreateSelect( | |||
6486 | CGF.Builder.CreateICmpULT(DefaultThreadLimitVal, NumThreads), | |||
6487 | DefaultThreadLimitVal, NumThreads); | |||
6488 | } else { | |||
6489 | NumThreads = DefaultThreadLimitVal ? DefaultThreadLimitVal | |||
6490 | : CGF.Builder.getInt32(0); | |||
6491 | } | |||
6492 | // Process condition of the if clause. | |||
6493 | if (CondVal) { | |||
6494 | NumThreads = CGF.Builder.CreateSelect(CondVal, NumThreads, | |||
6495 | CGF.Builder.getInt32(1)); | |||
6496 | } | |||
6497 | return NumThreads; | |||
6498 | } | |||
6499 | if (isOpenMPSimdDirective(Dir->getDirectiveKind())) | |||
6500 | return CGF.Builder.getInt32(1); | |||
6501 | return DefaultThreadLimitVal; | |||
6502 | } | |||
6503 | return DefaultThreadLimitVal ? DefaultThreadLimitVal | |||
6504 | : CGF.Builder.getInt32(0); | |||
6505 | } | |||
6506 | ||||
6507 | const Expr *CGOpenMPRuntime::getNumThreadsExprForTargetDirective( | |||
6508 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | |||
6509 | int32_t &DefaultVal) { | |||
6510 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
6511 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&(static_cast <bool> (isOpenMPTargetExecutionDirective(DirectiveKind ) && "Expected target-based executable directive.") ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6512, __extension__ __PRETTY_FUNCTION__)) | |||
6512 | "Expected target-based executable directive.")(static_cast <bool> (isOpenMPTargetExecutionDirective(DirectiveKind ) && "Expected target-based executable directive.") ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6512, __extension__ __PRETTY_FUNCTION__)); | |||
6513 | ||||
6514 | switch (DirectiveKind) { | |||
6515 | case OMPD_target: | |||
6516 | // Teams have no clause thread_limit | |||
6517 | return nullptr; | |||
6518 | case OMPD_target_teams: | |||
6519 | case OMPD_target_teams_distribute: | |||
6520 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6521 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6522 | const Expr *ThreadLimit = ThreadLimitClause->getThreadLimit(); | |||
6523 | if (ThreadLimit->isIntegerConstantExpr(CGF.getContext())) | |||
6524 | if (auto Constant = | |||
6525 | ThreadLimit->getIntegerConstantExpr(CGF.getContext())) | |||
6526 | DefaultVal = Constant->getExtValue(); | |||
6527 | return ThreadLimit; | |||
6528 | } | |||
6529 | return nullptr; | |||
6530 | case OMPD_target_parallel: | |||
6531 | case OMPD_target_parallel_for: | |||
6532 | case OMPD_target_parallel_for_simd: | |||
6533 | case OMPD_target_teams_distribute_parallel_for: | |||
6534 | case OMPD_target_teams_distribute_parallel_for_simd: { | |||
6535 | Expr *ThreadLimit = nullptr; | |||
6536 | Expr *NumThreads = nullptr; | |||
6537 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6538 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6539 | ThreadLimit = ThreadLimitClause->getThreadLimit(); | |||
6540 | if (ThreadLimit->isIntegerConstantExpr(CGF.getContext())) | |||
6541 | if (auto Constant = | |||
6542 | ThreadLimit->getIntegerConstantExpr(CGF.getContext())) | |||
6543 | DefaultVal = Constant->getExtValue(); | |||
6544 | } | |||
6545 | if (D.hasClausesOfKind<OMPNumThreadsClause>()) { | |||
6546 | const auto *NumThreadsClause = D.getSingleClause<OMPNumThreadsClause>(); | |||
6547 | NumThreads = NumThreadsClause->getNumThreads(); | |||
6548 | if (NumThreads->isIntegerConstantExpr(CGF.getContext())) { | |||
6549 | if (auto Constant = | |||
6550 | NumThreads->getIntegerConstantExpr(CGF.getContext())) { | |||
6551 | if (Constant->getExtValue() < DefaultVal) { | |||
6552 | DefaultVal = Constant->getExtValue(); | |||
6553 | ThreadLimit = NumThreads; | |||
6554 | } | |||
6555 | } | |||
6556 | } | |||
6557 | } | |||
6558 | return ThreadLimit; | |||
6559 | } | |||
6560 | case OMPD_target_teams_distribute_simd: | |||
6561 | case OMPD_target_simd: | |||
6562 | DefaultVal = 1; | |||
6563 | return nullptr; | |||
6564 | case OMPD_parallel: | |||
6565 | case OMPD_for: | |||
6566 | case OMPD_parallel_for: | |||
6567 | case OMPD_parallel_master: | |||
6568 | case OMPD_parallel_sections: | |||
6569 | case OMPD_for_simd: | |||
6570 | case OMPD_parallel_for_simd: | |||
6571 | case OMPD_cancel: | |||
6572 | case OMPD_cancellation_point: | |||
6573 | case OMPD_ordered: | |||
6574 | case OMPD_threadprivate: | |||
6575 | case OMPD_allocate: | |||
6576 | case OMPD_task: | |||
6577 | case OMPD_simd: | |||
6578 | case OMPD_tile: | |||
6579 | case OMPD_unroll: | |||
6580 | case OMPD_sections: | |||
6581 | case OMPD_section: | |||
6582 | case OMPD_single: | |||
6583 | case OMPD_master: | |||
6584 | case OMPD_critical: | |||
6585 | case OMPD_taskyield: | |||
6586 | case OMPD_barrier: | |||
6587 | case OMPD_taskwait: | |||
6588 | case OMPD_taskgroup: | |||
6589 | case OMPD_atomic: | |||
6590 | case OMPD_flush: | |||
6591 | case OMPD_depobj: | |||
6592 | case OMPD_scan: | |||
6593 | case OMPD_teams: | |||
6594 | case OMPD_target_data: | |||
6595 | case OMPD_target_exit_data: | |||
6596 | case OMPD_target_enter_data: | |||
6597 | case OMPD_distribute: | |||
6598 | case OMPD_distribute_simd: | |||
6599 | case OMPD_distribute_parallel_for: | |||
6600 | case OMPD_distribute_parallel_for_simd: | |||
6601 | case OMPD_teams_distribute: | |||
6602 | case OMPD_teams_distribute_simd: | |||
6603 | case OMPD_teams_distribute_parallel_for: | |||
6604 | case OMPD_teams_distribute_parallel_for_simd: | |||
6605 | case OMPD_target_update: | |||
6606 | case OMPD_declare_simd: | |||
6607 | case OMPD_declare_variant: | |||
6608 | case OMPD_begin_declare_variant: | |||
6609 | case OMPD_end_declare_variant: | |||
6610 | case OMPD_declare_target: | |||
6611 | case OMPD_end_declare_target: | |||
6612 | case OMPD_declare_reduction: | |||
6613 | case OMPD_declare_mapper: | |||
6614 | case OMPD_taskloop: | |||
6615 | case OMPD_taskloop_simd: | |||
6616 | case OMPD_master_taskloop: | |||
6617 | case OMPD_master_taskloop_simd: | |||
6618 | case OMPD_parallel_master_taskloop: | |||
6619 | case OMPD_parallel_master_taskloop_simd: | |||
6620 | case OMPD_requires: | |||
6621 | case OMPD_unknown: | |||
6622 | break; | |||
6623 | default: | |||
6624 | break; | |||
6625 | } | |||
6626 | llvm_unreachable("Unsupported directive kind.")::llvm::llvm_unreachable_internal("Unsupported directive kind." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6626); | |||
6627 | } | |||
6628 | ||||
6629 | llvm::Value *CGOpenMPRuntime::emitNumThreadsForTargetDirective( | |||
6630 | CodeGenFunction &CGF, const OMPExecutableDirective &D) { | |||
6631 | assert(!CGF.getLangOpts().OpenMPIsDevice &&(static_cast <bool> (!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? void (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6633, __extension__ __PRETTY_FUNCTION__)) | |||
6632 | "Clauses associated with the teams directive expected to be emitted "(static_cast <bool> (!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? void (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6633, __extension__ __PRETTY_FUNCTION__)) | |||
6633 | "only for the host!")(static_cast <bool> (!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? void (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6633, __extension__ __PRETTY_FUNCTION__)); | |||
6634 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
6635 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&(static_cast <bool> (isOpenMPTargetExecutionDirective(DirectiveKind ) && "Expected target-based executable directive.") ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6636, __extension__ __PRETTY_FUNCTION__)) | |||
6636 | "Expected target-based executable directive.")(static_cast <bool> (isOpenMPTargetExecutionDirective(DirectiveKind ) && "Expected target-based executable directive.") ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6636, __extension__ __PRETTY_FUNCTION__)); | |||
6637 | CGBuilderTy &Bld = CGF.Builder; | |||
6638 | llvm::Value *ThreadLimitVal = nullptr; | |||
6639 | llvm::Value *NumThreadsVal = nullptr; | |||
6640 | switch (DirectiveKind) { | |||
6641 | case OMPD_target: { | |||
6642 | const CapturedStmt *CS = D.getInnermostCapturedStmt(); | |||
6643 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6644 | return NumThreads; | |||
6645 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6646 | CGF.getContext(), CS->getCapturedStmt()); | |||
6647 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | |||
6648 | if (Dir->hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6649 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6650 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6651 | const auto *ThreadLimitClause = | |||
6652 | Dir->getSingleClause<OMPThreadLimitClause>(); | |||
6653 | CodeGenFunction::LexicalScope Scope( | |||
6654 | CGF, ThreadLimitClause->getThreadLimit()->getSourceRange()); | |||
6655 | if (const auto *PreInit = | |||
6656 | cast_or_null<DeclStmt>(ThreadLimitClause->getPreInitStmt())) { | |||
6657 | for (const auto *I : PreInit->decls()) { | |||
6658 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | |||
6659 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | |||
6660 | } else { | |||
6661 | CodeGenFunction::AutoVarEmission Emission = | |||
6662 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | |||
6663 | CGF.EmitAutoVarCleanups(Emission); | |||
6664 | } | |||
6665 | } | |||
6666 | } | |||
6667 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6668 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6669 | ThreadLimitVal = | |||
6670 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | |||
6671 | } | |||
6672 | if (isOpenMPTeamsDirective(Dir->getDirectiveKind()) && | |||
6673 | !isOpenMPDistributeDirective(Dir->getDirectiveKind())) { | |||
6674 | CS = Dir->getInnermostCapturedStmt(); | |||
6675 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6676 | CGF.getContext(), CS->getCapturedStmt()); | |||
6677 | Dir = dyn_cast_or_null<OMPExecutableDirective>(Child); | |||
6678 | } | |||
6679 | if (Dir && isOpenMPDistributeDirective(Dir->getDirectiveKind()) && | |||
6680 | !isOpenMPSimdDirective(Dir->getDirectiveKind())) { | |||
6681 | CS = Dir->getInnermostCapturedStmt(); | |||
6682 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6683 | return NumThreads; | |||
6684 | } | |||
6685 | if (Dir && isOpenMPSimdDirective(Dir->getDirectiveKind())) | |||
6686 | return Bld.getInt32(1); | |||
6687 | } | |||
6688 | return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0); | |||
6689 | } | |||
6690 | case OMPD_target_teams: { | |||
6691 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6692 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | |||
6693 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6694 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6695 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6696 | ThreadLimitVal = | |||
6697 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | |||
6698 | } | |||
6699 | const CapturedStmt *CS = D.getInnermostCapturedStmt(); | |||
6700 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6701 | return NumThreads; | |||
6702 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6703 | CGF.getContext(), CS->getCapturedStmt()); | |||
6704 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | |||
6705 | if (Dir->getDirectiveKind() == OMPD_distribute) { | |||
6706 | CS = Dir->getInnermostCapturedStmt(); | |||
6707 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6708 | return NumThreads; | |||
6709 | } | |||
6710 | } | |||
6711 | return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0); | |||
6712 | } | |||
6713 | case OMPD_target_teams_distribute: | |||
6714 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6715 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | |||
6716 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6717 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6718 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6719 | ThreadLimitVal = | |||
6720 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | |||
6721 | } | |||
6722 | return getNumThreads(CGF, D.getInnermostCapturedStmt(), ThreadLimitVal); | |||
6723 | case OMPD_target_parallel: | |||
6724 | case OMPD_target_parallel_for: | |||
6725 | case OMPD_target_parallel_for_simd: | |||
6726 | case OMPD_target_teams_distribute_parallel_for: | |||
6727 | case OMPD_target_teams_distribute_parallel_for_simd: { | |||
6728 | llvm::Value *CondVal = nullptr; | |||
6729 | // Handle if clause. If if clause present, the number of threads is | |||
6730 | // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1. | |||
6731 | if (D.hasClausesOfKind<OMPIfClause>()) { | |||
6732 | const OMPIfClause *IfClause = nullptr; | |||
6733 | for (const auto *C : D.getClausesOfKind<OMPIfClause>()) { | |||
6734 | if (C->getNameModifier() == OMPD_unknown || | |||
6735 | C->getNameModifier() == OMPD_parallel) { | |||
6736 | IfClause = C; | |||
6737 | break; | |||
6738 | } | |||
6739 | } | |||
6740 | if (IfClause) { | |||
6741 | const Expr *Cond = IfClause->getCondition(); | |||
6742 | bool Result; | |||
6743 | if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) { | |||
6744 | if (!Result) | |||
6745 | return Bld.getInt32(1); | |||
6746 | } else { | |||
6747 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
6748 | CondVal = CGF.EvaluateExprAsBool(Cond); | |||
6749 | } | |||
6750 | } | |||
6751 | } | |||
6752 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6753 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | |||
6754 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6755 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6756 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6757 | ThreadLimitVal = | |||
6758 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | |||
6759 | } | |||
6760 | if (D.hasClausesOfKind<OMPNumThreadsClause>()) { | |||
6761 | CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF); | |||
6762 | const auto *NumThreadsClause = D.getSingleClause<OMPNumThreadsClause>(); | |||
6763 | llvm::Value *NumThreads = CGF.EmitScalarExpr( | |||
6764 | NumThreadsClause->getNumThreads(), /*IgnoreResultAssign=*/true); | |||
6765 | NumThreadsVal = | |||
6766 | Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned=*/false); | |||
6767 | ThreadLimitVal = ThreadLimitVal | |||
6768 | ? Bld.CreateSelect(Bld.CreateICmpULT(NumThreadsVal, | |||
6769 | ThreadLimitVal), | |||
6770 | NumThreadsVal, ThreadLimitVal) | |||
6771 | : NumThreadsVal; | |||
6772 | } | |||
6773 | if (!ThreadLimitVal) | |||
6774 | ThreadLimitVal = Bld.getInt32(0); | |||
6775 | if (CondVal) | |||
6776 | return Bld.CreateSelect(CondVal, ThreadLimitVal, Bld.getInt32(1)); | |||
6777 | return ThreadLimitVal; | |||
6778 | } | |||
6779 | case OMPD_target_teams_distribute_simd: | |||
6780 | case OMPD_target_simd: | |||
6781 | return Bld.getInt32(1); | |||
6782 | case OMPD_parallel: | |||
6783 | case OMPD_for: | |||
6784 | case OMPD_parallel_for: | |||
6785 | case OMPD_parallel_master: | |||
6786 | case OMPD_parallel_sections: | |||
6787 | case OMPD_for_simd: | |||
6788 | case OMPD_parallel_for_simd: | |||
6789 | case OMPD_cancel: | |||
6790 | case OMPD_cancellation_point: | |||
6791 | case OMPD_ordered: | |||
6792 | case OMPD_threadprivate: | |||
6793 | case OMPD_allocate: | |||
6794 | case OMPD_task: | |||
6795 | case OMPD_simd: | |||
6796 | case OMPD_tile: | |||
6797 | case OMPD_unroll: | |||
6798 | case OMPD_sections: | |||
6799 | case OMPD_section: | |||
6800 | case OMPD_single: | |||
6801 | case OMPD_master: | |||
6802 | case OMPD_critical: | |||
6803 | case OMPD_taskyield: | |||
6804 | case OMPD_barrier: | |||
6805 | case OMPD_taskwait: | |||
6806 | case OMPD_taskgroup: | |||
6807 | case OMPD_atomic: | |||
6808 | case OMPD_flush: | |||
6809 | case OMPD_depobj: | |||
6810 | case OMPD_scan: | |||
6811 | case OMPD_teams: | |||
6812 | case OMPD_target_data: | |||
6813 | case OMPD_target_exit_data: | |||
6814 | case OMPD_target_enter_data: | |||
6815 | case OMPD_distribute: | |||
6816 | case OMPD_distribute_simd: | |||
6817 | case OMPD_distribute_parallel_for: | |||
6818 | case OMPD_distribute_parallel_for_simd: | |||
6819 | case OMPD_teams_distribute: | |||
6820 | case OMPD_teams_distribute_simd: | |||
6821 | case OMPD_teams_distribute_parallel_for: | |||
6822 | case OMPD_teams_distribute_parallel_for_simd: | |||
6823 | case OMPD_target_update: | |||
6824 | case OMPD_declare_simd: | |||
6825 | case OMPD_declare_variant: | |||
6826 | case OMPD_begin_declare_variant: | |||
6827 | case OMPD_end_declare_variant: | |||
6828 | case OMPD_declare_target: | |||
6829 | case OMPD_end_declare_target: | |||
6830 | case OMPD_declare_reduction: | |||
6831 | case OMPD_declare_mapper: | |||
6832 | case OMPD_taskloop: | |||
6833 | case OMPD_taskloop_simd: | |||
6834 | case OMPD_master_taskloop: | |||
6835 | case OMPD_master_taskloop_simd: | |||
6836 | case OMPD_parallel_master_taskloop: | |||
6837 | case OMPD_parallel_master_taskloop_simd: | |||
6838 | case OMPD_requires: | |||
6839 | case OMPD_metadirective: | |||
6840 | case OMPD_unknown: | |||
6841 | break; | |||
6842 | default: | |||
6843 | break; | |||
6844 | } | |||
6845 | llvm_unreachable("Unsupported directive kind.")::llvm::llvm_unreachable_internal("Unsupported directive kind." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 6845); | |||
6846 | } | |||
6847 | ||||
6848 | namespace { | |||
6849 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail ::operator|; using ::llvm::BitmaskEnumDetail::operator&; using ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail ::operator|=; using ::llvm::BitmaskEnumDetail::operator&= ; using ::llvm::BitmaskEnumDetail::operator^=; | |||
6850 | ||||
6851 | // Utility to handle information from clauses associated with a given | |||
6852 | // construct that use mappable expressions (e.g. 'map' clause, 'to' clause). | |||
6853 | // It provides a convenient interface to obtain the information and generate | |||
6854 | // code for that information. | |||
6855 | class MappableExprsHandler { | |||
6856 | public: | |||
6857 | /// Values for bit flags used to specify the mapping type for | |||
6858 | /// offloading. | |||
6859 | enum OpenMPOffloadMappingFlags : uint64_t { | |||
6860 | /// No flags | |||
6861 | OMP_MAP_NONE = 0x0, | |||
6862 | /// Allocate memory on the device and move data from host to device. | |||
6863 | OMP_MAP_TO = 0x01, | |||
6864 | /// Allocate memory on the device and move data from device to host. | |||
6865 | OMP_MAP_FROM = 0x02, | |||
6866 | /// Always perform the requested mapping action on the element, even | |||
6867 | /// if it was already mapped before. | |||
6868 | OMP_MAP_ALWAYS = 0x04, | |||
6869 | /// Delete the element from the device environment, ignoring the | |||
6870 | /// current reference count associated with the element. | |||
6871 | OMP_MAP_DELETE = 0x08, | |||
6872 | /// The element being mapped is a pointer-pointee pair; both the | |||
6873 | /// pointer and the pointee should be mapped. | |||
6874 | OMP_MAP_PTR_AND_OBJ = 0x10, | |||
6875 | /// This flags signals that the base address of an entry should be | |||
6876 | /// passed to the target kernel as an argument. | |||
6877 | OMP_MAP_TARGET_PARAM = 0x20, | |||
6878 | /// Signal that the runtime library has to return the device pointer | |||
6879 | /// in the current position for the data being mapped. Used when we have the | |||
6880 | /// use_device_ptr or use_device_addr clause. | |||
6881 | OMP_MAP_RETURN_PARAM = 0x40, | |||
6882 | /// This flag signals that the reference being passed is a pointer to | |||
6883 | /// private data. | |||
6884 | OMP_MAP_PRIVATE = 0x80, | |||
6885 | /// Pass the element to the device by value. | |||
6886 | OMP_MAP_LITERAL = 0x100, | |||
6887 | /// Implicit map | |||
6888 | OMP_MAP_IMPLICIT = 0x200, | |||
6889 | /// Close is a hint to the runtime to allocate memory close to | |||
6890 | /// the target device. | |||
6891 | OMP_MAP_CLOSE = 0x400, | |||
6892 | /// 0x800 is reserved for compatibility with XLC. | |||
6893 | /// Produce a runtime error if the data is not already allocated. | |||
6894 | OMP_MAP_PRESENT = 0x1000, | |||
6895 | // Increment and decrement a separate reference counter so that the data | |||
6896 | // cannot be unmapped within the associated region. Thus, this flag is | |||
6897 | // intended to be used on 'target' and 'target data' directives because they | |||
6898 | // are inherently structured. It is not intended to be used on 'target | |||
6899 | // enter data' and 'target exit data' directives because they are inherently | |||
6900 | // dynamic. | |||
6901 | // This is an OpenMP extension for the sake of OpenACC support. | |||
6902 | OMP_MAP_OMPX_HOLD = 0x2000, | |||
6903 | /// Signal that the runtime library should use args as an array of | |||
6904 | /// descriptor_dim pointers and use args_size as dims. Used when we have | |||
6905 | /// non-contiguous list items in target update directive | |||
6906 | OMP_MAP_NON_CONTIG = 0x100000000000, | |||
6907 | /// The 16 MSBs of the flags indicate whether the entry is member of some | |||
6908 | /// struct/class. | |||
6909 | OMP_MAP_MEMBER_OF = 0xffff000000000000, | |||
6910 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ OMP_MAP_MEMBER_OF)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_MAP_MEMBER_OF, | |||
6911 | }; | |||
6912 | ||||
6913 | /// Get the offset of the OMP_MAP_MEMBER_OF field. | |||
6914 | static unsigned getFlagMemberOffset() { | |||
6915 | unsigned Offset = 0; | |||
6916 | for (uint64_t Remain = OMP_MAP_MEMBER_OF; !(Remain & 1); | |||
6917 | Remain = Remain >> 1) | |||
6918 | Offset++; | |||
6919 | return Offset; | |||
6920 | } | |||
6921 | ||||
6922 | /// Class that holds debugging information for a data mapping to be passed to | |||
6923 | /// the runtime library. | |||
6924 | class MappingExprInfo { | |||
6925 | /// The variable declaration used for the data mapping. | |||
6926 | const ValueDecl *MapDecl = nullptr; | |||
6927 | /// The original expression used in the map clause, or null if there is | |||
6928 | /// none. | |||
6929 | const Expr *MapExpr = nullptr; | |||
6930 | ||||
6931 | public: | |||
6932 | MappingExprInfo(const ValueDecl *MapDecl, const Expr *MapExpr = nullptr) | |||
6933 | : MapDecl(MapDecl), MapExpr(MapExpr) {} | |||
6934 | ||||
6935 | const ValueDecl *getMapDecl() const { return MapDecl; } | |||
6936 | const Expr *getMapExpr() const { return MapExpr; } | |||
6937 | }; | |||
6938 | ||||
6939 | /// Class that associates information with a base pointer to be passed to the | |||
6940 | /// runtime library. | |||
6941 | class BasePointerInfo { | |||
6942 | /// The base pointer. | |||
6943 | llvm::Value *Ptr = nullptr; | |||
6944 | /// The base declaration that refers to this device pointer, or null if | |||
6945 | /// there is none. | |||
6946 | const ValueDecl *DevPtrDecl = nullptr; | |||
6947 | ||||
6948 | public: | |||
6949 | BasePointerInfo(llvm::Value *Ptr, const ValueDecl *DevPtrDecl = nullptr) | |||
6950 | : Ptr(Ptr), DevPtrDecl(DevPtrDecl) {} | |||
6951 | llvm::Value *operator*() const { return Ptr; } | |||
6952 | const ValueDecl *getDevicePtrDecl() const { return DevPtrDecl; } | |||
6953 | void setDevicePtrDecl(const ValueDecl *D) { DevPtrDecl = D; } | |||
6954 | }; | |||
6955 | ||||
6956 | using MapExprsArrayTy = SmallVector<MappingExprInfo, 4>; | |||
6957 | using MapBaseValuesArrayTy = SmallVector<BasePointerInfo, 4>; | |||
6958 | using MapValuesArrayTy = SmallVector<llvm::Value *, 4>; | |||
6959 | using MapFlagsArrayTy = SmallVector<OpenMPOffloadMappingFlags, 4>; | |||
6960 | using MapMappersArrayTy = SmallVector<const ValueDecl *, 4>; | |||
6961 | using MapDimArrayTy = SmallVector<uint64_t, 4>; | |||
6962 | using MapNonContiguousArrayTy = SmallVector<MapValuesArrayTy, 4>; | |||
6963 | ||||
6964 | /// This structure contains combined information generated for mappable | |||
6965 | /// clauses, including base pointers, pointers, sizes, map types, user-defined | |||
6966 | /// mappers, and non-contiguous information. | |||
6967 | struct MapCombinedInfoTy { | |||
6968 | struct StructNonContiguousInfo { | |||
6969 | bool IsNonContiguous = false; | |||
6970 | MapDimArrayTy Dims; | |||
6971 | MapNonContiguousArrayTy Offsets; | |||
6972 | MapNonContiguousArrayTy Counts; | |||
6973 | MapNonContiguousArrayTy Strides; | |||
6974 | }; | |||
6975 | MapExprsArrayTy Exprs; | |||
6976 | MapBaseValuesArrayTy BasePointers; | |||
6977 | MapValuesArrayTy Pointers; | |||
6978 | MapValuesArrayTy Sizes; | |||
6979 | MapFlagsArrayTy Types; | |||
6980 | MapMappersArrayTy Mappers; | |||
6981 | StructNonContiguousInfo NonContigInfo; | |||
6982 | ||||
6983 | /// Append arrays in \a CurInfo. | |||
6984 | void append(MapCombinedInfoTy &CurInfo) { | |||
6985 | Exprs.append(CurInfo.Exprs.begin(), CurInfo.Exprs.end()); | |||
6986 | BasePointers.append(CurInfo.BasePointers.begin(), | |||
6987 | CurInfo.BasePointers.end()); | |||
6988 | Pointers.append(CurInfo.Pointers.begin(), CurInfo.Pointers.end()); | |||
6989 | Sizes.append(CurInfo.Sizes.begin(), CurInfo.Sizes.end()); | |||
6990 | Types.append(CurInfo.Types.begin(), CurInfo.Types.end()); | |||
6991 | Mappers.append(CurInfo.Mappers.begin(), CurInfo.Mappers.end()); | |||
6992 | NonContigInfo.Dims.append(CurInfo.NonContigInfo.Dims.begin(), | |||
6993 | CurInfo.NonContigInfo.Dims.end()); | |||
6994 | NonContigInfo.Offsets.append(CurInfo.NonContigInfo.Offsets.begin(), | |||
6995 | CurInfo.NonContigInfo.Offsets.end()); | |||
6996 | NonContigInfo.Counts.append(CurInfo.NonContigInfo.Counts.begin(), | |||
6997 | CurInfo.NonContigInfo.Counts.end()); | |||
6998 | NonContigInfo.Strides.append(CurInfo.NonContigInfo.Strides.begin(), | |||
6999 | CurInfo.NonContigInfo.Strides.end()); | |||
7000 | } | |||
7001 | }; | |||
7002 | ||||
7003 | /// Map between a struct and the its lowest & highest elements which have been | |||
7004 | /// mapped. | |||
7005 | /// [ValueDecl *] --> {LE(FieldIndex, Pointer), | |||
7006 | /// HE(FieldIndex, Pointer)} | |||
7007 | struct StructRangeInfoTy { | |||
7008 | MapCombinedInfoTy PreliminaryMapData; | |||
7009 | std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> LowestElem = { | |||
7010 | 0, Address::invalid()}; | |||
7011 | std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> HighestElem = { | |||
7012 | 0, Address::invalid()}; | |||
7013 | Address Base = Address::invalid(); | |||
7014 | Address LB = Address::invalid(); | |||
7015 | bool IsArraySection = false; | |||
7016 | bool HasCompleteRecord = false; | |||
7017 | }; | |||
7018 | ||||
7019 | private: | |||
7020 | /// Kind that defines how a device pointer has to be returned. | |||
7021 | struct MapInfo { | |||
7022 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
7023 | OpenMPMapClauseKind MapType = OMPC_MAP_unknown; | |||
7024 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
7025 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers; | |||
7026 | bool ReturnDevicePointer = false; | |||
7027 | bool IsImplicit = false; | |||
7028 | const ValueDecl *Mapper = nullptr; | |||
7029 | const Expr *VarRef = nullptr; | |||
7030 | bool ForDeviceAddr = false; | |||
7031 | ||||
7032 | MapInfo() = default; | |||
7033 | MapInfo( | |||
7034 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components, | |||
7035 | OpenMPMapClauseKind MapType, | |||
7036 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7037 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, | |||
7038 | bool ReturnDevicePointer, bool IsImplicit, | |||
7039 | const ValueDecl *Mapper = nullptr, const Expr *VarRef = nullptr, | |||
7040 | bool ForDeviceAddr = false) | |||
7041 | : Components(Components), MapType(MapType), MapModifiers(MapModifiers), | |||
7042 | MotionModifiers(MotionModifiers), | |||
7043 | ReturnDevicePointer(ReturnDevicePointer), IsImplicit(IsImplicit), | |||
7044 | Mapper(Mapper), VarRef(VarRef), ForDeviceAddr(ForDeviceAddr) {} | |||
7045 | }; | |||
7046 | ||||
7047 | /// If use_device_ptr or use_device_addr is used on a decl which is a struct | |||
7048 | /// member and there is no map information about it, then emission of that | |||
7049 | /// entry is deferred until the whole struct has been processed. | |||
7050 | struct DeferredDevicePtrEntryTy { | |||
7051 | const Expr *IE = nullptr; | |||
7052 | const ValueDecl *VD = nullptr; | |||
7053 | bool ForDeviceAddr = false; | |||
7054 | ||||
7055 | DeferredDevicePtrEntryTy(const Expr *IE, const ValueDecl *VD, | |||
7056 | bool ForDeviceAddr) | |||
7057 | : IE(IE), VD(VD), ForDeviceAddr(ForDeviceAddr) {} | |||
7058 | }; | |||
7059 | ||||
7060 | /// The target directive from where the mappable clauses were extracted. It | |||
7061 | /// is either a executable directive or a user-defined mapper directive. | |||
7062 | llvm::PointerUnion<const OMPExecutableDirective *, | |||
7063 | const OMPDeclareMapperDecl *> | |||
7064 | CurDir; | |||
7065 | ||||
7066 | /// Function the directive is being generated for. | |||
7067 | CodeGenFunction &CGF; | |||
7068 | ||||
7069 | /// Set of all first private variables in the current directive. | |||
7070 | /// bool data is set to true if the variable is implicitly marked as | |||
7071 | /// firstprivate, false otherwise. | |||
7072 | llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, bool> FirstPrivateDecls; | |||
7073 | ||||
7074 | /// Map between device pointer declarations and their expression components. | |||
7075 | /// The key value for declarations in 'this' is null. | |||
7076 | llvm::DenseMap< | |||
7077 | const ValueDecl *, | |||
7078 | SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>> | |||
7079 | DevPointersMap; | |||
7080 | ||||
7081 | /// Map between device addr declarations and their expression components. | |||
7082 | /// The key value for declarations in 'this' is null. | |||
7083 | llvm::DenseMap< | |||
7084 | const ValueDecl *, | |||
7085 | SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>> | |||
7086 | HasDevAddrsMap; | |||
7087 | ||||
7088 | /// Map between lambda declarations and their map type. | |||
7089 | llvm::DenseMap<const ValueDecl *, const OMPMapClause *> LambdasMap; | |||
7090 | ||||
7091 | llvm::Value *getExprTypeSize(const Expr *E) const { | |||
7092 | QualType ExprTy = E->getType().getCanonicalType(); | |||
7093 | ||||
7094 | // Calculate the size for array shaping expression. | |||
7095 | if (const auto *OAE = dyn_cast<OMPArrayShapingExpr>(E)) { | |||
7096 | llvm::Value *Size = | |||
7097 | CGF.getTypeSize(OAE->getBase()->getType()->getPointeeType()); | |||
7098 | for (const Expr *SE : OAE->getDimensions()) { | |||
7099 | llvm::Value *Sz = CGF.EmitScalarExpr(SE); | |||
7100 | Sz = CGF.EmitScalarConversion(Sz, SE->getType(), | |||
7101 | CGF.getContext().getSizeType(), | |||
7102 | SE->getExprLoc()); | |||
7103 | Size = CGF.Builder.CreateNUWMul(Size, Sz); | |||
7104 | } | |||
7105 | return Size; | |||
7106 | } | |||
7107 | ||||
7108 | // Reference types are ignored for mapping purposes. | |||
7109 | if (const auto *RefTy = ExprTy->getAs<ReferenceType>()) | |||
7110 | ExprTy = RefTy->getPointeeType().getCanonicalType(); | |||
7111 | ||||
7112 | // Given that an array section is considered a built-in type, we need to | |||
7113 | // do the calculation based on the length of the section instead of relying | |||
7114 | // on CGF.getTypeSize(E->getType()). | |||
7115 | if (const auto *OAE = dyn_cast<OMPArraySectionExpr>(E)) { | |||
7116 | QualType BaseTy = OMPArraySectionExpr::getBaseOriginalType( | |||
7117 | OAE->getBase()->IgnoreParenImpCasts()) | |||
7118 | .getCanonicalType(); | |||
7119 | ||||
7120 | // If there is no length associated with the expression and lower bound is | |||
7121 | // not specified too, that means we are using the whole length of the | |||
7122 | // base. | |||
7123 | if (!OAE->getLength() && OAE->getColonLocFirst().isValid() && | |||
7124 | !OAE->getLowerBound()) | |||
7125 | return CGF.getTypeSize(BaseTy); | |||
7126 | ||||
7127 | llvm::Value *ElemSize; | |||
7128 | if (const auto *PTy = BaseTy->getAs<PointerType>()) { | |||
7129 | ElemSize = CGF.getTypeSize(PTy->getPointeeType().getCanonicalType()); | |||
7130 | } else { | |||
7131 | const auto *ATy = cast<ArrayType>(BaseTy.getTypePtr()); | |||
7132 | assert(ATy && "Expecting array type if not a pointer type.")(static_cast <bool> (ATy && "Expecting array type if not a pointer type." ) ? void (0) : __assert_fail ("ATy && \"Expecting array type if not a pointer type.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7132, __extension__ __PRETTY_FUNCTION__)); | |||
7133 | ElemSize = CGF.getTypeSize(ATy->getElementType().getCanonicalType()); | |||
7134 | } | |||
7135 | ||||
7136 | // If we don't have a length at this point, that is because we have an | |||
7137 | // array section with a single element. | |||
7138 | if (!OAE->getLength() && OAE->getColonLocFirst().isInvalid()) | |||
7139 | return ElemSize; | |||
7140 | ||||
7141 | if (const Expr *LenExpr = OAE->getLength()) { | |||
7142 | llvm::Value *LengthVal = CGF.EmitScalarExpr(LenExpr); | |||
7143 | LengthVal = CGF.EmitScalarConversion(LengthVal, LenExpr->getType(), | |||
7144 | CGF.getContext().getSizeType(), | |||
7145 | LenExpr->getExprLoc()); | |||
7146 | return CGF.Builder.CreateNUWMul(LengthVal, ElemSize); | |||
7147 | } | |||
7148 | assert(!OAE->getLength() && OAE->getColonLocFirst().isValid() &&(static_cast <bool> (!OAE->getLength() && OAE ->getColonLocFirst().isValid() && OAE->getLowerBound () && "expected array_section[lb:].") ? void (0) : __assert_fail ("!OAE->getLength() && OAE->getColonLocFirst().isValid() && OAE->getLowerBound() && \"expected array_section[lb:].\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7149, __extension__ __PRETTY_FUNCTION__)) | |||
7149 | OAE->getLowerBound() && "expected array_section[lb:].")(static_cast <bool> (!OAE->getLength() && OAE ->getColonLocFirst().isValid() && OAE->getLowerBound () && "expected array_section[lb:].") ? void (0) : __assert_fail ("!OAE->getLength() && OAE->getColonLocFirst().isValid() && OAE->getLowerBound() && \"expected array_section[lb:].\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7149, __extension__ __PRETTY_FUNCTION__)); | |||
7150 | // Size = sizetype - lb * elemtype; | |||
7151 | llvm::Value *LengthVal = CGF.getTypeSize(BaseTy); | |||
7152 | llvm::Value *LBVal = CGF.EmitScalarExpr(OAE->getLowerBound()); | |||
7153 | LBVal = CGF.EmitScalarConversion(LBVal, OAE->getLowerBound()->getType(), | |||
7154 | CGF.getContext().getSizeType(), | |||
7155 | OAE->getLowerBound()->getExprLoc()); | |||
7156 | LBVal = CGF.Builder.CreateNUWMul(LBVal, ElemSize); | |||
7157 | llvm::Value *Cmp = CGF.Builder.CreateICmpUGT(LengthVal, LBVal); | |||
7158 | llvm::Value *TrueVal = CGF.Builder.CreateNUWSub(LengthVal, LBVal); | |||
7159 | LengthVal = CGF.Builder.CreateSelect( | |||
7160 | Cmp, TrueVal, llvm::ConstantInt::get(CGF.SizeTy, 0)); | |||
7161 | return LengthVal; | |||
7162 | } | |||
7163 | return CGF.getTypeSize(ExprTy); | |||
7164 | } | |||
7165 | ||||
7166 | /// Return the corresponding bits for a given map clause modifier. Add | |||
7167 | /// a flag marking the map as a pointer if requested. Add a flag marking the | |||
7168 | /// map as the first one of a series of maps that relate to the same map | |||
7169 | /// expression. | |||
7170 | OpenMPOffloadMappingFlags getMapTypeBits( | |||
7171 | OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7172 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, bool IsImplicit, | |||
7173 | bool AddPtrFlag, bool AddIsTargetParamFlag, bool IsNonContiguous) const { | |||
7174 | OpenMPOffloadMappingFlags Bits = | |||
7175 | IsImplicit ? OMP_MAP_IMPLICIT : OMP_MAP_NONE; | |||
7176 | switch (MapType) { | |||
7177 | case OMPC_MAP_alloc: | |||
7178 | case OMPC_MAP_release: | |||
7179 | // alloc and release is the default behavior in the runtime library, i.e. | |||
7180 | // if we don't pass any bits alloc/release that is what the runtime is | |||
7181 | // going to do. Therefore, we don't need to signal anything for these two | |||
7182 | // type modifiers. | |||
7183 | break; | |||
7184 | case OMPC_MAP_to: | |||
7185 | Bits |= OMP_MAP_TO; | |||
7186 | break; | |||
7187 | case OMPC_MAP_from: | |||
7188 | Bits |= OMP_MAP_FROM; | |||
7189 | break; | |||
7190 | case OMPC_MAP_tofrom: | |||
7191 | Bits |= OMP_MAP_TO | OMP_MAP_FROM; | |||
7192 | break; | |||
7193 | case OMPC_MAP_delete: | |||
7194 | Bits |= OMP_MAP_DELETE; | |||
7195 | break; | |||
7196 | case OMPC_MAP_unknown: | |||
7197 | llvm_unreachable("Unexpected map type!")::llvm::llvm_unreachable_internal("Unexpected map type!", "clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7197); | |||
7198 | } | |||
7199 | if (AddPtrFlag) | |||
7200 | Bits |= OMP_MAP_PTR_AND_OBJ; | |||
7201 | if (AddIsTargetParamFlag) | |||
7202 | Bits |= OMP_MAP_TARGET_PARAM; | |||
7203 | if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_always)) | |||
7204 | Bits |= OMP_MAP_ALWAYS; | |||
7205 | if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_close)) | |||
7206 | Bits |= OMP_MAP_CLOSE; | |||
7207 | if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_present) || | |||
7208 | llvm::is_contained(MotionModifiers, OMPC_MOTION_MODIFIER_present)) | |||
7209 | Bits |= OMP_MAP_PRESENT; | |||
7210 | if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_ompx_hold)) | |||
7211 | Bits |= OMP_MAP_OMPX_HOLD; | |||
7212 | if (IsNonContiguous) | |||
7213 | Bits |= OMP_MAP_NON_CONTIG; | |||
7214 | return Bits; | |||
7215 | } | |||
7216 | ||||
7217 | /// Return true if the provided expression is a final array section. A | |||
7218 | /// final array section, is one whose length can't be proved to be one. | |||
7219 | bool isFinalArraySectionExpression(const Expr *E) const { | |||
7220 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(E); | |||
7221 | ||||
7222 | // It is not an array section and therefore not a unity-size one. | |||
7223 | if (!OASE) | |||
7224 | return false; | |||
7225 | ||||
7226 | // An array section with no colon always refer to a single element. | |||
7227 | if (OASE->getColonLocFirst().isInvalid()) | |||
7228 | return false; | |||
7229 | ||||
7230 | const Expr *Length = OASE->getLength(); | |||
7231 | ||||
7232 | // If we don't have a length we have to check if the array has size 1 | |||
7233 | // for this dimension. Also, we should always expect a length if the | |||
7234 | // base type is pointer. | |||
7235 | if (!Length) { | |||
7236 | QualType BaseQTy = OMPArraySectionExpr::getBaseOriginalType( | |||
7237 | OASE->getBase()->IgnoreParenImpCasts()) | |||
7238 | .getCanonicalType(); | |||
7239 | if (const auto *ATy = dyn_cast<ConstantArrayType>(BaseQTy.getTypePtr())) | |||
7240 | return ATy->getSize().getSExtValue() != 1; | |||
7241 | // If we don't have a constant dimension length, we have to consider | |||
7242 | // the current section as having any size, so it is not necessarily | |||
7243 | // unitary. If it happen to be unity size, that's user fault. | |||
7244 | return true; | |||
7245 | } | |||
7246 | ||||
7247 | // Check if the length evaluates to 1. | |||
7248 | Expr::EvalResult Result; | |||
7249 | if (!Length->EvaluateAsInt(Result, CGF.getContext())) | |||
7250 | return true; // Can have more that size 1. | |||
7251 | ||||
7252 | llvm::APSInt ConstLength = Result.Val.getInt(); | |||
7253 | return ConstLength.getSExtValue() != 1; | |||
7254 | } | |||
7255 | ||||
7256 | /// Generate the base pointers, section pointers, sizes, map type bits, and | |||
7257 | /// user-defined mappers (all included in \a CombinedInfo) for the provided | |||
7258 | /// map type, map or motion modifiers, and expression components. | |||
7259 | /// \a IsFirstComponent should be set to true if the provided set of | |||
7260 | /// components is the first associated with a capture. | |||
7261 | void generateInfoForComponentList( | |||
7262 | OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7263 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, | |||
7264 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components, | |||
7265 | MapCombinedInfoTy &CombinedInfo, StructRangeInfoTy &PartialStruct, | |||
7266 | bool IsFirstComponentList, bool IsImplicit, | |||
7267 | const ValueDecl *Mapper = nullptr, bool ForDeviceAddr = false, | |||
7268 | const ValueDecl *BaseDecl = nullptr, const Expr *MapExpr = nullptr, | |||
7269 | ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> | |||
7270 | OverlappedElements = llvm::None) const { | |||
7271 | // The following summarizes what has to be generated for each map and the | |||
7272 | // types below. The generated information is expressed in this order: | |||
7273 | // base pointer, section pointer, size, flags | |||
7274 | // (to add to the ones that come from the map type and modifier). | |||
7275 | // | |||
7276 | // double d; | |||
7277 | // int i[100]; | |||
7278 | // float *p; | |||
7279 | // | |||
7280 | // struct S1 { | |||
7281 | // int i; | |||
7282 | // float f[50]; | |||
7283 | // } | |||
7284 | // struct S2 { | |||
7285 | // int i; | |||
7286 | // float f[50]; | |||
7287 | // S1 s; | |||
7288 | // double *p; | |||
7289 | // struct S2 *ps; | |||
7290 | // int &ref; | |||
7291 | // } | |||
7292 | // S2 s; | |||
7293 | // S2 *ps; | |||
7294 | // | |||
7295 | // map(d) | |||
7296 | // &d, &d, sizeof(double), TARGET_PARAM | TO | FROM | |||
7297 | // | |||
7298 | // map(i) | |||
7299 | // &i, &i, 100*sizeof(int), TARGET_PARAM | TO | FROM | |||
7300 | // | |||
7301 | // map(i[1:23]) | |||
7302 | // &i(=&i[0]), &i[1], 23*sizeof(int), TARGET_PARAM | TO | FROM | |||
7303 | // | |||
7304 | // map(p) | |||
7305 | // &p, &p, sizeof(float*), TARGET_PARAM | TO | FROM | |||
7306 | // | |||
7307 | // map(p[1:24]) | |||
7308 | // &p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM | PTR_AND_OBJ | |||
7309 | // in unified shared memory mode or for local pointers | |||
7310 | // p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM | |||
7311 | // | |||
7312 | // map(s) | |||
7313 | // &s, &s, sizeof(S2), TARGET_PARAM | TO | FROM | |||
7314 | // | |||
7315 | // map(s.i) | |||
7316 | // &s, &(s.i), sizeof(int), TARGET_PARAM | TO | FROM | |||
7317 | // | |||
7318 | // map(s.s.f) | |||
7319 | // &s, &(s.s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM | |||
7320 | // | |||
7321 | // map(s.p) | |||
7322 | // &s, &(s.p), sizeof(double*), TARGET_PARAM | TO | FROM | |||
7323 | // | |||
7324 | // map(to: s.p[:22]) | |||
7325 | // &s, &(s.p), sizeof(double*), TARGET_PARAM (*) | |||
7326 | // &s, &(s.p), sizeof(double*), MEMBER_OF(1) (**) | |||
7327 | // &(s.p), &(s.p[0]), 22*sizeof(double), | |||
7328 | // MEMBER_OF(1) | PTR_AND_OBJ | TO (***) | |||
7329 | // (*) alloc space for struct members, only this is a target parameter | |||
7330 | // (**) map the pointer (nothing to be mapped in this example) (the compiler | |||
7331 | // optimizes this entry out, same in the examples below) | |||
7332 | // (***) map the pointee (map: to) | |||
7333 | // | |||
7334 | // map(to: s.ref) | |||
7335 | // &s, &(s.ref), sizeof(int*), TARGET_PARAM (*) | |||
7336 | // &s, &(s.ref), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | TO (***) | |||
7337 | // (*) alloc space for struct members, only this is a target parameter | |||
7338 | // (**) map the pointer (nothing to be mapped in this example) (the compiler | |||
7339 | // optimizes this entry out, same in the examples below) | |||
7340 | // (***) map the pointee (map: to) | |||
7341 | // | |||
7342 | // map(s.ps) | |||
7343 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | TO | FROM | |||
7344 | // | |||
7345 | // map(from: s.ps->s.i) | |||
7346 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7347 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7348 | // &(s.ps), &(s.ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7349 | // | |||
7350 | // map(to: s.ps->ps) | |||
7351 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7352 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7353 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | TO | |||
7354 | // | |||
7355 | // map(s.ps->ps->ps) | |||
7356 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7357 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7358 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7359 | // &(s.ps->ps), &(s.ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM | |||
7360 | // | |||
7361 | // map(to: s.ps->ps->s.f[:22]) | |||
7362 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7363 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7364 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7365 | // &(s.ps->ps), &(s.ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO | |||
7366 | // | |||
7367 | // map(ps) | |||
7368 | // &ps, &ps, sizeof(S2*), TARGET_PARAM | TO | FROM | |||
7369 | // | |||
7370 | // map(ps->i) | |||
7371 | // ps, &(ps->i), sizeof(int), TARGET_PARAM | TO | FROM | |||
7372 | // | |||
7373 | // map(ps->s.f) | |||
7374 | // ps, &(ps->s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM | |||
7375 | // | |||
7376 | // map(from: ps->p) | |||
7377 | // ps, &(ps->p), sizeof(double*), TARGET_PARAM | FROM | |||
7378 | // | |||
7379 | // map(to: ps->p[:22]) | |||
7380 | // ps, &(ps->p), sizeof(double*), TARGET_PARAM | |||
7381 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(1) | |||
7382 | // &(ps->p), &(ps->p[0]), 22*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | TO | |||
7383 | // | |||
7384 | // map(ps->ps) | |||
7385 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | TO | FROM | |||
7386 | // | |||
7387 | // map(from: ps->ps->s.i) | |||
7388 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7389 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7390 | // &(ps->ps), &(ps->ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7391 | // | |||
7392 | // map(from: ps->ps->ps) | |||
7393 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7394 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7395 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7396 | // | |||
7397 | // map(ps->ps->ps->ps) | |||
7398 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7399 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7400 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7401 | // &(ps->ps->ps), &(ps->ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM | |||
7402 | // | |||
7403 | // map(to: ps->ps->ps->s.f[:22]) | |||
7404 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7405 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7406 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7407 | // &(ps->ps->ps), &(ps->ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO | |||
7408 | // | |||
7409 | // map(to: s.f[:22]) map(from: s.p[:33]) | |||
7410 | // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1) + | |||
7411 | // sizeof(double*) (**), TARGET_PARAM | |||
7412 | // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | TO | |||
7413 | // &s, &(s.p), sizeof(double*), MEMBER_OF(1) | |||
7414 | // &(s.p), &(s.p[0]), 33*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7415 | // (*) allocate contiguous space needed to fit all mapped members even if | |||
7416 | // we allocate space for members not mapped (in this example, | |||
7417 | // s.f[22..49] and s.s are not mapped, yet we must allocate space for | |||
7418 | // them as well because they fall between &s.f[0] and &s.p) | |||
7419 | // | |||
7420 | // map(from: s.f[:22]) map(to: ps->p[:33]) | |||
7421 | // &s, &(s.f[0]), 22*sizeof(float), TARGET_PARAM | FROM | |||
7422 | // ps, &(ps->p), sizeof(S2*), TARGET_PARAM | |||
7423 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(2) (*) | |||
7424 | // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(2) | PTR_AND_OBJ | TO | |||
7425 | // (*) the struct this entry pertains to is the 2nd element in the list of | |||
7426 | // arguments, hence MEMBER_OF(2) | |||
7427 | // | |||
7428 | // map(from: s.f[:22], s.s) map(to: ps->p[:33]) | |||
7429 | // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1), TARGET_PARAM | |||
7430 | // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | FROM | |||
7431 | // &s, &(s.s), sizeof(struct S1), MEMBER_OF(1) | FROM | |||
7432 | // ps, &(ps->p), sizeof(S2*), TARGET_PARAM | |||
7433 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(4) (*) | |||
7434 | // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(4) | PTR_AND_OBJ | TO | |||
7435 | // (*) the struct this entry pertains to is the 4th element in the list | |||
7436 | // of arguments, hence MEMBER_OF(4) | |||
7437 | ||||
7438 | // Track if the map information being generated is the first for a capture. | |||
7439 | bool IsCaptureFirstInfo = IsFirstComponentList; | |||
7440 | // When the variable is on a declare target link or in a to clause with | |||
7441 | // unified memory, a reference is needed to hold the host/device address | |||
7442 | // of the variable. | |||
7443 | bool RequiresReference = false; | |||
7444 | ||||
7445 | // Scan the components from the base to the complete expression. | |||
7446 | auto CI = Components.rbegin(); | |||
7447 | auto CE = Components.rend(); | |||
7448 | auto I = CI; | |||
7449 | ||||
7450 | // Track if the map information being generated is the first for a list of | |||
7451 | // components. | |||
7452 | bool IsExpressionFirstInfo = true; | |||
7453 | bool FirstPointerInComplexData = false; | |||
7454 | Address BP = Address::invalid(); | |||
7455 | const Expr *AssocExpr = I->getAssociatedExpression(); | |||
7456 | const auto *AE = dyn_cast<ArraySubscriptExpr>(AssocExpr); | |||
| ||||
7457 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr); | |||
7458 | const auto *OAShE = dyn_cast<OMPArrayShapingExpr>(AssocExpr); | |||
7459 | ||||
7460 | if (isa<MemberExpr>(AssocExpr)) { | |||
7461 | // The base is the 'this' pointer. The content of the pointer is going | |||
7462 | // to be the base of the field being mapped. | |||
7463 | BP = CGF.LoadCXXThisAddress(); | |||
7464 | } else if ((AE && isa<CXXThisExpr>(AE->getBase()->IgnoreParenImpCasts())) || | |||
7465 | (OASE && | |||
7466 | isa<CXXThisExpr>(OASE->getBase()->IgnoreParenImpCasts()))) { | |||
7467 | BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(CGF); | |||
7468 | } else if (OAShE && | |||
7469 | isa<CXXThisExpr>(OAShE->getBase()->IgnoreParenCasts())) { | |||
7470 | BP = Address( | |||
7471 | CGF.EmitScalarExpr(OAShE->getBase()), | |||
7472 | CGF.ConvertTypeForMem(OAShE->getBase()->getType()->getPointeeType()), | |||
7473 | CGF.getContext().getTypeAlignInChars(OAShE->getBase()->getType())); | |||
7474 | } else { | |||
7475 | // The base is the reference to the variable. | |||
7476 | // BP = &Var. | |||
7477 | BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(CGF); | |||
7478 | if (const auto *VD = | |||
7479 | dyn_cast_or_null<VarDecl>(I->getAssociatedDeclaration())) { | |||
7480 | if (llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
7481 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) { | |||
7482 | if ((*Res == OMPDeclareTargetDeclAttr::MT_Link) || | |||
7483 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | |||
7484 | CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory())) { | |||
7485 | RequiresReference = true; | |||
7486 | BP = CGF.CGM.getOpenMPRuntime().getAddrOfDeclareTargetVar(VD); | |||
7487 | } | |||
7488 | } | |||
7489 | } | |||
7490 | ||||
7491 | // If the variable is a pointer and is being dereferenced (i.e. is not | |||
7492 | // the last component), the base has to be the pointer itself, not its | |||
7493 | // reference. References are ignored for mapping purposes. | |||
7494 | QualType Ty = | |||
7495 | I->getAssociatedDeclaration()->getType().getNonReferenceType(); | |||
7496 | if (Ty->isAnyPointerType() && std::next(I) != CE) { | |||
7497 | // No need to generate individual map information for the pointer, it | |||
7498 | // can be associated with the combined storage if shared memory mode is | |||
7499 | // active or the base declaration is not global variable. | |||
7500 | const auto *VD = dyn_cast<VarDecl>(I->getAssociatedDeclaration()); | |||
7501 | if (CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory() || | |||
7502 | !VD || VD->hasLocalStorage()) | |||
7503 | BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>()); | |||
7504 | else | |||
7505 | FirstPointerInComplexData = true; | |||
7506 | ++I; | |||
7507 | } | |||
7508 | } | |||
7509 | ||||
7510 | // Track whether a component of the list should be marked as MEMBER_OF some | |||
7511 | // combined entry (for partial structs). Only the first PTR_AND_OBJ entry | |||
7512 | // in a component list should be marked as MEMBER_OF, all subsequent entries | |||
7513 | // do not belong to the base struct. E.g. | |||
7514 | // struct S2 s; | |||
7515 | // s.ps->ps->ps->f[:] | |||
7516 | // (1) (2) (3) (4) | |||
7517 | // ps(1) is a member pointer, ps(2) is a pointee of ps(1), so it is a | |||
7518 | // PTR_AND_OBJ entry; the PTR is ps(1), so MEMBER_OF the base struct. ps(3) | |||
7519 | // is the pointee of ps(2) which is not member of struct s, so it should not | |||
7520 | // be marked as such (it is still PTR_AND_OBJ). | |||
7521 | // The variable is initialized to false so that PTR_AND_OBJ entries which | |||
7522 | // are not struct members are not considered (e.g. array of pointers to | |||
7523 | // data). | |||
7524 | bool ShouldBeMemberOf = false; | |||
7525 | ||||
7526 | // Variable keeping track of whether or not we have encountered a component | |||
7527 | // in the component list which is a member expression. Useful when we have a | |||
7528 | // pointer or a final array section, in which case it is the previous | |||
7529 | // component in the list which tells us whether we have a member expression. | |||
7530 | // E.g. X.f[:] | |||
7531 | // While processing the final array section "[:]" it is "f" which tells us | |||
7532 | // whether we are dealing with a member of a declared struct. | |||
7533 | const MemberExpr *EncounteredME = nullptr; | |||
7534 | ||||
7535 | // Track for the total number of dimension. Start from one for the dummy | |||
7536 | // dimension. | |||
7537 | uint64_t DimSize = 1; | |||
7538 | ||||
7539 | bool IsNonContiguous = CombinedInfo.NonContigInfo.IsNonContiguous; | |||
7540 | bool IsPrevMemberReference = false; | |||
7541 | ||||
7542 | for (; I != CE; ++I) { | |||
7543 | // If the current component is member of a struct (parent struct) mark it. | |||
7544 | if (!EncounteredME) { | |||
7545 | EncounteredME = dyn_cast<MemberExpr>(I->getAssociatedExpression()); | |||
7546 | // If we encounter a PTR_AND_OBJ entry from now on it should be marked | |||
7547 | // as MEMBER_OF the parent struct. | |||
7548 | if (EncounteredME) { | |||
7549 | ShouldBeMemberOf = true; | |||
7550 | // Do not emit as complex pointer if this is actually not array-like | |||
7551 | // expression. | |||
7552 | if (FirstPointerInComplexData) { | |||
7553 | QualType Ty = std::prev(I) | |||
7554 | ->getAssociatedDeclaration() | |||
7555 | ->getType() | |||
7556 | .getNonReferenceType(); | |||
7557 | BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>()); | |||
7558 | FirstPointerInComplexData = false; | |||
7559 | } | |||
7560 | } | |||
7561 | } | |||
7562 | ||||
7563 | auto Next = std::next(I); | |||
7564 | ||||
7565 | // We need to generate the addresses and sizes if this is the last | |||
7566 | // component, if the component is a pointer or if it is an array section | |||
7567 | // whose length can't be proved to be one. If this is a pointer, it | |||
7568 | // becomes the base address for the following components. | |||
7569 | ||||
7570 | // A final array section, is one whose length can't be proved to be one. | |||
7571 | // If the map item is non-contiguous then we don't treat any array section | |||
7572 | // as final array section. | |||
7573 | bool IsFinalArraySection = | |||
7574 | !IsNonContiguous && | |||
7575 | isFinalArraySectionExpression(I->getAssociatedExpression()); | |||
7576 | ||||
7577 | // If we have a declaration for the mapping use that, otherwise use | |||
7578 | // the base declaration of the map clause. | |||
7579 | const ValueDecl *MapDecl = (I->getAssociatedDeclaration()) | |||
7580 | ? I->getAssociatedDeclaration() | |||
7581 | : BaseDecl; | |||
7582 | MapExpr = (I->getAssociatedExpression()) ? I->getAssociatedExpression() | |||
7583 | : MapExpr; | |||
7584 | ||||
7585 | // Get information on whether the element is a pointer. Have to do a | |||
7586 | // special treatment for array sections given that they are built-in | |||
7587 | // types. | |||
7588 | const auto *OASE = | |||
7589 | dyn_cast<OMPArraySectionExpr>(I->getAssociatedExpression()); | |||
7590 | const auto *OAShE = | |||
7591 | dyn_cast<OMPArrayShapingExpr>(I->getAssociatedExpression()); | |||
7592 | const auto *UO = dyn_cast<UnaryOperator>(I->getAssociatedExpression()); | |||
7593 | const auto *BO = dyn_cast<BinaryOperator>(I->getAssociatedExpression()); | |||
7594 | bool IsPointer = | |||
7595 | OAShE || | |||
7596 | (OASE && OMPArraySectionExpr::getBaseOriginalType(OASE) | |||
7597 | .getCanonicalType() | |||
7598 | ->isAnyPointerType()) || | |||
7599 | I->getAssociatedExpression()->getType()->isAnyPointerType(); | |||
7600 | bool IsMemberReference = isa<MemberExpr>(I->getAssociatedExpression()) && | |||
7601 | MapDecl && | |||
7602 | MapDecl->getType()->isLValueReferenceType(); | |||
7603 | bool IsNonDerefPointer = IsPointer && !UO && !BO && !IsNonContiguous; | |||
7604 | ||||
7605 | if (OASE) | |||
7606 | ++DimSize; | |||
7607 | ||||
7608 | if (Next == CE || IsMemberReference || IsNonDerefPointer || | |||
7609 | IsFinalArraySection) { | |||
7610 | // If this is not the last component, we expect the pointer to be | |||
7611 | // associated with an array expression or member expression. | |||
7612 | assert((Next == CE ||(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7613 | isa<MemberExpr>(Next->getAssociatedExpression()) ||(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7614 | isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) ||(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7615 | isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) ||(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7616 | isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) ||(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7617 | isa<UnaryOperator>(Next->getAssociatedExpression()) ||(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7618 | isa<BinaryOperator>(Next->getAssociatedExpression())) &&(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7619 | "Unexpected expression")(static_cast <bool> ((Next == CE || isa<MemberExpr> (Next->getAssociatedExpression()) || isa<ArraySubscriptExpr >(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr >(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr >(Next->getAssociatedExpression()) || isa<UnaryOperator >(Next->getAssociatedExpression()) || isa<BinaryOperator >(Next->getAssociatedExpression())) && "Unexpected expression" ) ? void (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) || isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) || isa<UnaryOperator>(Next->getAssociatedExpression()) || isa<BinaryOperator>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7619, __extension__ __PRETTY_FUNCTION__)); | |||
7620 | ||||
7621 | Address LB = Address::invalid(); | |||
7622 | Address LowestElem = Address::invalid(); | |||
7623 | auto &&EmitMemberExprBase = [](CodeGenFunction &CGF, | |||
7624 | const MemberExpr *E) { | |||
7625 | const Expr *BaseExpr = E->getBase(); | |||
7626 | // If this is s.x, emit s as an lvalue. If it is s->x, emit s as a | |||
7627 | // scalar. | |||
7628 | LValue BaseLV; | |||
7629 | if (E->isArrow()) { | |||
7630 | LValueBaseInfo BaseInfo; | |||
7631 | TBAAAccessInfo TBAAInfo; | |||
7632 | Address Addr = | |||
7633 | CGF.EmitPointerWithAlignment(BaseExpr, &BaseInfo, &TBAAInfo); | |||
7634 | QualType PtrTy = BaseExpr->getType()->getPointeeType(); | |||
7635 | BaseLV = CGF.MakeAddrLValue(Addr, PtrTy, BaseInfo, TBAAInfo); | |||
7636 | } else { | |||
7637 | BaseLV = CGF.EmitOMPSharedLValue(BaseExpr); | |||
7638 | } | |||
7639 | return BaseLV; | |||
7640 | }; | |||
7641 | if (OAShE) { | |||
7642 | LowestElem = LB = | |||
7643 | Address(CGF.EmitScalarExpr(OAShE->getBase()), | |||
7644 | CGF.ConvertTypeForMem( | |||
7645 | OAShE->getBase()->getType()->getPointeeType()), | |||
7646 | CGF.getContext().getTypeAlignInChars( | |||
7647 | OAShE->getBase()->getType())); | |||
7648 | } else if (IsMemberReference) { | |||
7649 | const auto *ME = cast<MemberExpr>(I->getAssociatedExpression()); | |||
7650 | LValue BaseLVal = EmitMemberExprBase(CGF, ME); | |||
7651 | LowestElem = CGF.EmitLValueForFieldInitialization( | |||
7652 | BaseLVal, cast<FieldDecl>(MapDecl)) | |||
7653 | .getAddress(CGF); | |||
7654 | LB = CGF.EmitLoadOfReferenceLValue(LowestElem, MapDecl->getType()) | |||
7655 | .getAddress(CGF); | |||
7656 | } else { | |||
7657 | LowestElem = LB = | |||
7658 | CGF.EmitOMPSharedLValue(I->getAssociatedExpression()) | |||
7659 | .getAddress(CGF); | |||
7660 | } | |||
7661 | ||||
7662 | // If this component is a pointer inside the base struct then we don't | |||
7663 | // need to create any entry for it - it will be combined with the object | |||
7664 | // it is pointing to into a single PTR_AND_OBJ entry. | |||
7665 | bool IsMemberPointerOrAddr = | |||
7666 | EncounteredME && | |||
7667 | (((IsPointer || ForDeviceAddr) && | |||
7668 | I->getAssociatedExpression() == EncounteredME) || | |||
7669 | (IsPrevMemberReference && !IsPointer) || | |||
7670 | (IsMemberReference && Next != CE && | |||
7671 | !Next->getAssociatedExpression()->getType()->isPointerType())); | |||
7672 | if (!OverlappedElements.empty() && Next == CE) { | |||
7673 | // Handle base element with the info for overlapped elements. | |||
7674 | assert(!PartialStruct.Base.isValid() && "The base element is set.")(static_cast <bool> (!PartialStruct.Base.isValid() && "The base element is set.") ? void (0) : __assert_fail ("!PartialStruct.Base.isValid() && \"The base element is set.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7674, __extension__ __PRETTY_FUNCTION__)); | |||
7675 | assert(!IsPointer &&(static_cast <bool> (!IsPointer && "Unexpected base element with the pointer type." ) ? void (0) : __assert_fail ("!IsPointer && \"Unexpected base element with the pointer type.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7676, __extension__ __PRETTY_FUNCTION__)) | |||
7676 | "Unexpected base element with the pointer type.")(static_cast <bool> (!IsPointer && "Unexpected base element with the pointer type." ) ? void (0) : __assert_fail ("!IsPointer && \"Unexpected base element with the pointer type.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7676, __extension__ __PRETTY_FUNCTION__)); | |||
7677 | // Mark the whole struct as the struct that requires allocation on the | |||
7678 | // device. | |||
7679 | PartialStruct.LowestElem = {0, LowestElem}; | |||
7680 | CharUnits TypeSize = CGF.getContext().getTypeSizeInChars( | |||
7681 | I->getAssociatedExpression()->getType()); | |||
7682 | Address HB = CGF.Builder.CreateConstGEP( | |||
7683 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
7684 | LowestElem, CGF.VoidPtrTy, CGF.Int8Ty), | |||
7685 | TypeSize.getQuantity() - 1); | |||
7686 | PartialStruct.HighestElem = { | |||
7687 | std::numeric_limits<decltype( | |||
7688 | PartialStruct.HighestElem.first)>::max(), | |||
7689 | HB}; | |||
7690 | PartialStruct.Base = BP; | |||
7691 | PartialStruct.LB = LB; | |||
7692 | assert((static_cast <bool> (PartialStruct.PreliminaryMapData.BasePointers .empty() && "Overlapped elements must be used only once for the variable." ) ? void (0) : __assert_fail ("PartialStruct.PreliminaryMapData.BasePointers.empty() && \"Overlapped elements must be used only once for the variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7694, __extension__ __PRETTY_FUNCTION__)) | |||
7693 | PartialStruct.PreliminaryMapData.BasePointers.empty() &&(static_cast <bool> (PartialStruct.PreliminaryMapData.BasePointers .empty() && "Overlapped elements must be used only once for the variable." ) ? void (0) : __assert_fail ("PartialStruct.PreliminaryMapData.BasePointers.empty() && \"Overlapped elements must be used only once for the variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7694, __extension__ __PRETTY_FUNCTION__)) | |||
7694 | "Overlapped elements must be used only once for the variable.")(static_cast <bool> (PartialStruct.PreliminaryMapData.BasePointers .empty() && "Overlapped elements must be used only once for the variable." ) ? void (0) : __assert_fail ("PartialStruct.PreliminaryMapData.BasePointers.empty() && \"Overlapped elements must be used only once for the variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7694, __extension__ __PRETTY_FUNCTION__)); | |||
7695 | std::swap(PartialStruct.PreliminaryMapData, CombinedInfo); | |||
7696 | // Emit data for non-overlapped data. | |||
7697 | OpenMPOffloadMappingFlags Flags = | |||
7698 | OMP_MAP_MEMBER_OF | | |||
7699 | getMapTypeBits(MapType, MapModifiers, MotionModifiers, IsImplicit, | |||
7700 | /*AddPtrFlag=*/false, | |||
7701 | /*AddIsTargetParamFlag=*/false, IsNonContiguous); | |||
7702 | llvm::Value *Size = nullptr; | |||
7703 | // Do bitcopy of all non-overlapped structure elements. | |||
7704 | for (OMPClauseMappableExprCommon::MappableExprComponentListRef | |||
7705 | Component : OverlappedElements) { | |||
7706 | Address ComponentLB = Address::invalid(); | |||
7707 | for (const OMPClauseMappableExprCommon::MappableComponent &MC : | |||
7708 | Component) { | |||
7709 | if (const ValueDecl *VD = MC.getAssociatedDeclaration()) { | |||
7710 | const auto *FD = dyn_cast<FieldDecl>(VD); | |||
7711 | if (FD && FD->getType()->isLValueReferenceType()) { | |||
7712 | const auto *ME = | |||
7713 | cast<MemberExpr>(MC.getAssociatedExpression()); | |||
7714 | LValue BaseLVal = EmitMemberExprBase(CGF, ME); | |||
7715 | ComponentLB = | |||
7716 | CGF.EmitLValueForFieldInitialization(BaseLVal, FD) | |||
7717 | .getAddress(CGF); | |||
7718 | } else { | |||
7719 | ComponentLB = | |||
7720 | CGF.EmitOMPSharedLValue(MC.getAssociatedExpression()) | |||
7721 | .getAddress(CGF); | |||
7722 | } | |||
7723 | Size = CGF.Builder.CreatePtrDiff( | |||
7724 | CGF.Int8Ty, CGF.EmitCastToVoidPtr(ComponentLB.getPointer()), | |||
7725 | CGF.EmitCastToVoidPtr(LB.getPointer())); | |||
7726 | break; | |||
7727 | } | |||
7728 | } | |||
7729 | assert(Size && "Failed to determine structure size")(static_cast <bool> (Size && "Failed to determine structure size" ) ? void (0) : __assert_fail ("Size && \"Failed to determine structure size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7729, __extension__ __PRETTY_FUNCTION__)); | |||
7730 | CombinedInfo.Exprs.emplace_back(MapDecl, MapExpr); | |||
7731 | CombinedInfo.BasePointers.push_back(BP.getPointer()); | |||
7732 | CombinedInfo.Pointers.push_back(LB.getPointer()); | |||
7733 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | |||
7734 | Size, CGF.Int64Ty, /*isSigned=*/true)); | |||
7735 | CombinedInfo.Types.push_back(Flags); | |||
7736 | CombinedInfo.Mappers.push_back(nullptr); | |||
7737 | CombinedInfo.NonContigInfo.Dims.push_back(IsNonContiguous ? DimSize | |||
7738 | : 1); | |||
7739 | LB = CGF.Builder.CreateConstGEP(ComponentLB, 1); | |||
7740 | } | |||
7741 | CombinedInfo.Exprs.emplace_back(MapDecl, MapExpr); | |||
7742 | CombinedInfo.BasePointers.push_back(BP.getPointer()); | |||
7743 | CombinedInfo.Pointers.push_back(LB.getPointer()); | |||
7744 | Size = CGF.Builder.CreatePtrDiff( | |||
7745 | CGF.Int8Ty, CGF.Builder.CreateConstGEP(HB, 1).getPointer(), | |||
7746 | CGF.EmitCastToVoidPtr(LB.getPointer())); | |||
7747 | CombinedInfo.Sizes.push_back( | |||
7748 | CGF.Builder.CreateIntCast(Size, CGF.Int64Ty, /*isSigned=*/true)); | |||
7749 | CombinedInfo.Types.push_back(Flags); | |||
7750 | CombinedInfo.Mappers.push_back(nullptr); | |||
7751 | CombinedInfo.NonContigInfo.Dims.push_back(IsNonContiguous ? DimSize | |||
7752 | : 1); | |||
7753 | break; | |||
7754 | } | |||
7755 | llvm::Value *Size = getExprTypeSize(I->getAssociatedExpression()); | |||
7756 | if (!IsMemberPointerOrAddr || | |||
7757 | (Next == CE && MapType != OMPC_MAP_unknown)) { | |||
7758 | CombinedInfo.Exprs.emplace_back(MapDecl, MapExpr); | |||
7759 | CombinedInfo.BasePointers.push_back(BP.getPointer()); | |||
7760 | CombinedInfo.Pointers.push_back(LB.getPointer()); | |||
7761 | CombinedInfo.Sizes.push_back( | |||
7762 | CGF.Builder.CreateIntCast(Size, CGF.Int64Ty, /*isSigned=*/true)); | |||
7763 | CombinedInfo.NonContigInfo.Dims.push_back(IsNonContiguous ? DimSize | |||
7764 | : 1); | |||
7765 | ||||
7766 | // If Mapper is valid, the last component inherits the mapper. | |||
7767 | bool HasMapper = Mapper && Next == CE; | |||
7768 | CombinedInfo.Mappers.push_back(HasMapper ? Mapper : nullptr); | |||
7769 | ||||
7770 | // We need to add a pointer flag for each map that comes from the | |||
7771 | // same expression except for the first one. We also need to signal | |||
7772 | // this map is the first one that relates with the current capture | |||
7773 | // (there is a set of entries for each capture). | |||
7774 | OpenMPOffloadMappingFlags Flags = getMapTypeBits( | |||
7775 | MapType, MapModifiers, MotionModifiers, IsImplicit, | |||
7776 | !IsExpressionFirstInfo || RequiresReference || | |||
7777 | FirstPointerInComplexData || IsMemberReference, | |||
7778 | IsCaptureFirstInfo && !RequiresReference, IsNonContiguous); | |||
7779 | ||||
7780 | if (!IsExpressionFirstInfo || IsMemberReference) { | |||
7781 | // If we have a PTR_AND_OBJ pair where the OBJ is a pointer as well, | |||
7782 | // then we reset the TO/FROM/ALWAYS/DELETE/CLOSE flags. | |||
7783 | if (IsPointer || (IsMemberReference && Next != CE)) | |||
7784 | Flags &= ~(OMP_MAP_TO | OMP_MAP_FROM | OMP_MAP_ALWAYS | | |||
7785 | OMP_MAP_DELETE | OMP_MAP_CLOSE); | |||
7786 | ||||
7787 | if (ShouldBeMemberOf) { | |||
7788 | // Set placeholder value MEMBER_OF=FFFF to indicate that the flag | |||
7789 | // should be later updated with the correct value of MEMBER_OF. | |||
7790 | Flags |= OMP_MAP_MEMBER_OF; | |||
7791 | // From now on, all subsequent PTR_AND_OBJ entries should not be | |||
7792 | // marked as MEMBER_OF. | |||
7793 | ShouldBeMemberOf = false; | |||
7794 | } | |||
7795 | } | |||
7796 | ||||
7797 | CombinedInfo.Types.push_back(Flags); | |||
7798 | } | |||
7799 | ||||
7800 | // If we have encountered a member expression so far, keep track of the | |||
7801 | // mapped member. If the parent is "*this", then the value declaration | |||
7802 | // is nullptr. | |||
7803 | if (EncounteredME) { | |||
7804 | const auto *FD = cast<FieldDecl>(EncounteredME->getMemberDecl()); | |||
7805 | unsigned FieldIndex = FD->getFieldIndex(); | |||
7806 | ||||
7807 | // Update info about the lowest and highest elements for this struct | |||
7808 | if (!PartialStruct.Base.isValid()) { | |||
7809 | PartialStruct.LowestElem = {FieldIndex, LowestElem}; | |||
7810 | if (IsFinalArraySection) { | |||
7811 | Address HB = | |||
7812 | CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false) | |||
7813 | .getAddress(CGF); | |||
7814 | PartialStruct.HighestElem = {FieldIndex, HB}; | |||
7815 | } else { | |||
7816 | PartialStruct.HighestElem = {FieldIndex, LowestElem}; | |||
7817 | } | |||
7818 | PartialStruct.Base = BP; | |||
7819 | PartialStruct.LB = BP; | |||
7820 | } else if (FieldIndex < PartialStruct.LowestElem.first) { | |||
7821 | PartialStruct.LowestElem = {FieldIndex, LowestElem}; | |||
7822 | } else if (FieldIndex > PartialStruct.HighestElem.first) { | |||
7823 | PartialStruct.HighestElem = {FieldIndex, LowestElem}; | |||
7824 | } | |||
7825 | } | |||
7826 | ||||
7827 | // Need to emit combined struct for array sections. | |||
7828 | if (IsFinalArraySection || IsNonContiguous) | |||
7829 | PartialStruct.IsArraySection = true; | |||
7830 | ||||
7831 | // If we have a final array section, we are done with this expression. | |||
7832 | if (IsFinalArraySection) | |||
7833 | break; | |||
7834 | ||||
7835 | // The pointer becomes the base for the next element. | |||
7836 | if (Next != CE) | |||
7837 | BP = IsMemberReference ? LowestElem : LB; | |||
7838 | ||||
7839 | IsExpressionFirstInfo = false; | |||
7840 | IsCaptureFirstInfo = false; | |||
7841 | FirstPointerInComplexData = false; | |||
7842 | IsPrevMemberReference = IsMemberReference; | |||
7843 | } else if (FirstPointerInComplexData) { | |||
7844 | QualType Ty = Components.rbegin() | |||
7845 | ->getAssociatedDeclaration() | |||
7846 | ->getType() | |||
7847 | .getNonReferenceType(); | |||
7848 | BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>()); | |||
7849 | FirstPointerInComplexData = false; | |||
7850 | } | |||
7851 | } | |||
7852 | // If ran into the whole component - allocate the space for the whole | |||
7853 | // record. | |||
7854 | if (!EncounteredME
| |||
7855 | PartialStruct.HasCompleteRecord = true; | |||
7856 | ||||
7857 | if (!IsNonContiguous) | |||
7858 | return; | |||
7859 | ||||
7860 | const ASTContext &Context = CGF.getContext(); | |||
7861 | ||||
7862 | // For supporting stride in array section, we need to initialize the first | |||
7863 | // dimension size as 1, first offset as 0, and first count as 1 | |||
7864 | MapValuesArrayTy CurOffsets = {llvm::ConstantInt::get(CGF.CGM.Int64Ty, 0)}; | |||
7865 | MapValuesArrayTy CurCounts = {llvm::ConstantInt::get(CGF.CGM.Int64Ty, 1)}; | |||
7866 | MapValuesArrayTy CurStrides; | |||
7867 | MapValuesArrayTy DimSizes{llvm::ConstantInt::get(CGF.CGM.Int64Ty, 1)}; | |||
7868 | uint64_t ElementTypeSize; | |||
7869 | ||||
7870 | // Collect Size information for each dimension and get the element size as | |||
7871 | // the first Stride. For example, for `int arr[10][10]`, the DimSizes | |||
7872 | // should be [10, 10] and the first stride is 4 btyes. | |||
7873 | for (const OMPClauseMappableExprCommon::MappableComponent &Component : | |||
7874 | Components) { | |||
7875 | const Expr *AssocExpr = Component.getAssociatedExpression(); | |||
7876 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr); | |||
7877 | ||||
7878 | if (!OASE) | |||
7879 | continue; | |||
7880 | ||||
7881 | QualType Ty = OMPArraySectionExpr::getBaseOriginalType(OASE->getBase()); | |||
7882 | auto *CAT = Context.getAsConstantArrayType(Ty); | |||
7883 | auto *VAT = Context.getAsVariableArrayType(Ty); | |||
7884 | ||||
7885 | // We need all the dimension size except for the last dimension. | |||
7886 | assert((VAT || CAT || &Component == &*Components.begin()) &&(static_cast <bool> ((VAT || CAT || &Component == & *Components.begin()) && "Should be either ConstantArray or VariableArray if not the " "first Component") ? void (0) : __assert_fail ("(VAT || CAT || &Component == &*Components.begin()) && \"Should be either ConstantArray or VariableArray if not the \" \"first Component\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7888, __extension__ __PRETTY_FUNCTION__)) | |||
7887 | "Should be either ConstantArray or VariableArray if not the "(static_cast <bool> ((VAT || CAT || &Component == & *Components.begin()) && "Should be either ConstantArray or VariableArray if not the " "first Component") ? void (0) : __assert_fail ("(VAT || CAT || &Component == &*Components.begin()) && \"Should be either ConstantArray or VariableArray if not the \" \"first Component\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7888, __extension__ __PRETTY_FUNCTION__)) | |||
7888 | "first Component")(static_cast <bool> ((VAT || CAT || &Component == & *Components.begin()) && "Should be either ConstantArray or VariableArray if not the " "first Component") ? void (0) : __assert_fail ("(VAT || CAT || &Component == &*Components.begin()) && \"Should be either ConstantArray or VariableArray if not the \" \"first Component\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7888, __extension__ __PRETTY_FUNCTION__)); | |||
7889 | ||||
7890 | // Get element size if CurStrides is empty. | |||
7891 | if (CurStrides.empty()) { | |||
7892 | const Type *ElementType = nullptr; | |||
7893 | if (CAT) | |||
7894 | ElementType = CAT->getElementType().getTypePtr(); | |||
7895 | else if (VAT) | |||
7896 | ElementType = VAT->getElementType().getTypePtr(); | |||
7897 | else | |||
7898 | assert(&Component == &*Components.begin() &&(static_cast <bool> (&Component == &*Components .begin() && "Only expect pointer (non CAT or VAT) when this is the " "first Component") ? void (0) : __assert_fail ("&Component == &*Components.begin() && \"Only expect pointer (non CAT or VAT) when this is the \" \"first Component\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7900, __extension__ __PRETTY_FUNCTION__)) | |||
7899 | "Only expect pointer (non CAT or VAT) when this is the "(static_cast <bool> (&Component == &*Components .begin() && "Only expect pointer (non CAT or VAT) when this is the " "first Component") ? void (0) : __assert_fail ("&Component == &*Components.begin() && \"Only expect pointer (non CAT or VAT) when this is the \" \"first Component\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7900, __extension__ __PRETTY_FUNCTION__)) | |||
7900 | "first Component")(static_cast <bool> (&Component == &*Components .begin() && "Only expect pointer (non CAT or VAT) when this is the " "first Component") ? void (0) : __assert_fail ("&Component == &*Components.begin() && \"Only expect pointer (non CAT or VAT) when this is the \" \"first Component\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 7900, __extension__ __PRETTY_FUNCTION__)); | |||
7901 | // If ElementType is null, then it means the base is a pointer | |||
7902 | // (neither CAT nor VAT) and we'll attempt to get ElementType again | |||
7903 | // for next iteration. | |||
7904 | if (ElementType) { | |||
7905 | // For the case that having pointer as base, we need to remove one | |||
7906 | // level of indirection. | |||
7907 | if (&Component != &*Components.begin()) | |||
7908 | ElementType = ElementType->getPointeeOrArrayElementType(); | |||
7909 | ElementTypeSize = | |||
7910 | Context.getTypeSizeInChars(ElementType).getQuantity(); | |||
7911 | CurStrides.push_back( | |||
7912 | llvm::ConstantInt::get(CGF.Int64Ty, ElementTypeSize)); | |||
7913 | } | |||
7914 | } | |||
7915 | // Get dimension value except for the last dimension since we don't need | |||
7916 | // it. | |||
7917 | if (DimSizes.size() < Components.size() - 1) { | |||
7918 | if (CAT) | |||
7919 | DimSizes.push_back(llvm::ConstantInt::get( | |||
7920 | CGF.Int64Ty, CAT->getSize().getZExtValue())); | |||
7921 | else if (VAT) | |||
7922 | DimSizes.push_back(CGF.Builder.CreateIntCast( | |||
7923 | CGF.EmitScalarExpr(VAT->getSizeExpr()), CGF.Int64Ty, | |||
7924 | /*IsSigned=*/false)); | |||
7925 | } | |||
7926 | } | |||
7927 | ||||
7928 | // Skip the dummy dimension since we have already have its information. | |||
7929 | auto *DI = DimSizes.begin() + 1; | |||
7930 | // Product of dimension. | |||
7931 | llvm::Value *DimProd = | |||
7932 | llvm::ConstantInt::get(CGF.CGM.Int64Ty, ElementTypeSize); | |||
| ||||
7933 | ||||
7934 | // Collect info for non-contiguous. Notice that offset, count, and stride | |||
7935 | // are only meaningful for array-section, so we insert a null for anything | |||
7936 | // other than array-section. | |||
7937 | // Also, the size of offset, count, and stride are not the same as | |||
7938 | // pointers, base_pointers, sizes, or dims. Instead, the size of offset, | |||
7939 | // count, and stride are the same as the number of non-contiguous | |||
7940 | // declaration in target update to/from clause. | |||
7941 | for (const OMPClauseMappableExprCommon::MappableComponent &Component : | |||
7942 | Components) { | |||
7943 | const Expr *AssocExpr = Component.getAssociatedExpression(); | |||
7944 | ||||
7945 | if (const auto *AE = dyn_cast<ArraySubscriptExpr>(AssocExpr)) { | |||
7946 | llvm::Value *Offset = CGF.Builder.CreateIntCast( | |||
7947 | CGF.EmitScalarExpr(AE->getIdx()), CGF.Int64Ty, | |||
7948 | /*isSigned=*/false); | |||
7949 | CurOffsets.push_back(Offset); | |||
7950 | CurCounts.push_back(llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/1)); | |||
7951 | CurStrides.push_back(CurStrides.back()); | |||
7952 | continue; | |||
7953 | } | |||
7954 | ||||
7955 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr); | |||
7956 | ||||
7957 | if (!OASE) | |||
7958 | continue; | |||
7959 | ||||
7960 | // Offset | |||
7961 | const Expr *OffsetExpr = OASE->getLowerBound(); | |||
7962 | llvm::Value *Offset = nullptr; | |||
7963 | if (!OffsetExpr) { | |||
7964 | // If offset is absent, then we just set it to zero. | |||
7965 | Offset = llvm::ConstantInt::get(CGF.Int64Ty, 0); | |||
7966 | } else { | |||
7967 | Offset = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(OffsetExpr), | |||
7968 | CGF.Int64Ty, | |||
7969 | /*isSigned=*/false); | |||
7970 | } | |||
7971 | CurOffsets.push_back(Offset); | |||
7972 | ||||
7973 | // Count | |||
7974 | const Expr *CountExpr = OASE->getLength(); | |||
7975 | llvm::Value *Count = nullptr; | |||
7976 | if (!CountExpr) { | |||
7977 | // In Clang, once a high dimension is an array section, we construct all | |||
7978 | // the lower dimension as array section, however, for case like | |||
7979 | // arr[0:2][2], Clang construct the inner dimension as an array section | |||
7980 | // but it actually is not in an array section form according to spec. | |||
7981 | if (!OASE->getColonLocFirst().isValid() && | |||
7982 | !OASE->getColonLocSecond().isValid()) { | |||
7983 | Count = llvm::ConstantInt::get(CGF.Int64Ty, 1); | |||
7984 | } else { | |||
7985 | // OpenMP 5.0, 2.1.5 Array Sections, Description. | |||
7986 | // When the length is absent it defaults to ⌈(size − | |||
7987 | // lower-bound)/stride⌉, where size is the size of the array | |||
7988 | // dimension. | |||
7989 | const Expr *StrideExpr = OASE->getStride(); | |||
7990 | llvm::Value *Stride = | |||
7991 | StrideExpr | |||
7992 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(StrideExpr), | |||
7993 | CGF.Int64Ty, /*isSigned=*/false) | |||
7994 | : nullptr; | |||
7995 | if (Stride) | |||
7996 | Count = CGF.Builder.CreateUDiv( | |||
7997 | CGF.Builder.CreateNUWSub(*DI, Offset), Stride); | |||
7998 | else | |||
7999 | Count = CGF.Builder.CreateNUWSub(*DI, Offset); | |||
8000 | } | |||
8001 | } else { | |||
8002 | Count = CGF.EmitScalarExpr(CountExpr); | |||
8003 | } | |||
8004 | Count = CGF.Builder.CreateIntCast(Count, CGF.Int64Ty, /*isSigned=*/false); | |||
8005 | CurCounts.push_back(Count); | |||
8006 | ||||
8007 | // Stride_n' = Stride_n * (D_0 * D_1 ... * D_n-1) * Unit size | |||
8008 | // Take `int arr[5][5][5]` and `arr[0:2:2][1:2:1][0:2:2]` as an example: | |||
8009 | // Offset Count Stride | |||
8010 | // D0 0 1 4 (int) <- dummy dimension | |||
8011 | // D1 0 2 8 (2 * (1) * 4) | |||
8012 | // D2 1 2 20 (1 * (1 * 5) * 4) | |||
8013 | // D3 0 2 200 (2 * (1 * 5 * 4) * 4) | |||
8014 | const Expr *StrideExpr = OASE->getStride(); | |||
8015 | llvm::Value *Stride = | |||
8016 | StrideExpr | |||
8017 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(StrideExpr), | |||
8018 | CGF.Int64Ty, /*isSigned=*/false) | |||
8019 | : nullptr; | |||
8020 | DimProd = CGF.Builder.CreateNUWMul(DimProd, *(DI - 1)); | |||
8021 | if (Stride) | |||
8022 | CurStrides.push_back(CGF.Builder.CreateNUWMul(DimProd, Stride)); | |||
8023 | else | |||
8024 | CurStrides.push_back(DimProd); | |||
8025 | if (DI != DimSizes.end()) | |||
8026 | ++DI; | |||
8027 | } | |||
8028 | ||||
8029 | CombinedInfo.NonContigInfo.Offsets.push_back(CurOffsets); | |||
8030 | CombinedInfo.NonContigInfo.Counts.push_back(CurCounts); | |||
8031 | CombinedInfo.NonContigInfo.Strides.push_back(CurStrides); | |||
8032 | } | |||
8033 | ||||
8034 | /// Return the adjusted map modifiers if the declaration a capture refers to | |||
8035 | /// appears in a first-private clause. This is expected to be used only with | |||
8036 | /// directives that start with 'target'. | |||
8037 | MappableExprsHandler::OpenMPOffloadMappingFlags | |||
8038 | getMapModifiersForPrivateClauses(const CapturedStmt::Capture &Cap) const { | |||
8039 | assert(Cap.capturesVariable() && "Expected capture by reference only!")(static_cast <bool> (Cap.capturesVariable() && "Expected capture by reference only!" ) ? void (0) : __assert_fail ("Cap.capturesVariable() && \"Expected capture by reference only!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8039, __extension__ __PRETTY_FUNCTION__)); | |||
8040 | ||||
8041 | // A first private variable captured by reference will use only the | |||
8042 | // 'private ptr' and 'map to' flag. Return the right flags if the captured | |||
8043 | // declaration is known as first-private in this handler. | |||
8044 | if (FirstPrivateDecls.count(Cap.getCapturedVar())) { | |||
8045 | if (Cap.getCapturedVar()->getType()->isAnyPointerType()) | |||
8046 | return MappableExprsHandler::OMP_MAP_TO | | |||
8047 | MappableExprsHandler::OMP_MAP_PTR_AND_OBJ; | |||
8048 | return MappableExprsHandler::OMP_MAP_PRIVATE | | |||
8049 | MappableExprsHandler::OMP_MAP_TO; | |||
8050 | } | |||
8051 | auto I = LambdasMap.find(Cap.getCapturedVar()->getCanonicalDecl()); | |||
8052 | if (I != LambdasMap.end()) | |||
8053 | // for map(to: lambda): using user specified map type. | |||
8054 | return getMapTypeBits( | |||
8055 | I->getSecond()->getMapType(), I->getSecond()->getMapTypeModifiers(), | |||
8056 | /*MotionModifiers=*/llvm::None, I->getSecond()->isImplicit(), | |||
8057 | /*AddPtrFlag=*/false, | |||
8058 | /*AddIsTargetParamFlag=*/false, | |||
8059 | /*isNonContiguous=*/false); | |||
8060 | return MappableExprsHandler::OMP_MAP_TO | | |||
8061 | MappableExprsHandler::OMP_MAP_FROM; | |||
8062 | } | |||
8063 | ||||
8064 | static OpenMPOffloadMappingFlags getMemberOfFlag(unsigned Position) { | |||
8065 | // Rotate by getFlagMemberOffset() bits. | |||
8066 | return static_cast<OpenMPOffloadMappingFlags>(((uint64_t)Position + 1) | |||
8067 | << getFlagMemberOffset()); | |||
8068 | } | |||
8069 | ||||
8070 | static void setCorrectMemberOfFlag(OpenMPOffloadMappingFlags &Flags, | |||
8071 | OpenMPOffloadMappingFlags MemberOfFlag) { | |||
8072 | // If the entry is PTR_AND_OBJ but has not been marked with the special | |||
8073 | // placeholder value 0xFFFF in the MEMBER_OF field, then it should not be | |||
8074 | // marked as MEMBER_OF. | |||
8075 | if ((Flags & OMP_MAP_PTR_AND_OBJ) && | |||
8076 | ((Flags & OMP_MAP_MEMBER_OF) != OMP_MAP_MEMBER_OF)) | |||
8077 | return; | |||
8078 | ||||
8079 | // Reset the placeholder value to prepare the flag for the assignment of the | |||
8080 | // proper MEMBER_OF value. | |||
8081 | Flags &= ~OMP_MAP_MEMBER_OF; | |||
8082 | Flags |= MemberOfFlag; | |||
8083 | } | |||
8084 | ||||
8085 | void getPlainLayout(const CXXRecordDecl *RD, | |||
8086 | llvm::SmallVectorImpl<const FieldDecl *> &Layout, | |||
8087 | bool AsBase) const { | |||
8088 | const CGRecordLayout &RL = CGF.getTypes().getCGRecordLayout(RD); | |||
8089 | ||||
8090 | llvm::StructType *St = | |||
8091 | AsBase ? RL.getBaseSubobjectLLVMType() : RL.getLLVMType(); | |||
8092 | ||||
8093 | unsigned NumElements = St->getNumElements(); | |||
8094 | llvm::SmallVector< | |||
8095 | llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *>, 4> | |||
8096 | RecordLayout(NumElements); | |||
8097 | ||||
8098 | // Fill bases. | |||
8099 | for (const auto &I : RD->bases()) { | |||
8100 | if (I.isVirtual()) | |||
8101 | continue; | |||
8102 | const auto *Base = I.getType()->getAsCXXRecordDecl(); | |||
8103 | // Ignore empty bases. | |||
8104 | if (Base->isEmpty() || CGF.getContext() | |||
8105 | .getASTRecordLayout(Base) | |||
8106 | .getNonVirtualSize() | |||
8107 | .isZero()) | |||
8108 | continue; | |||
8109 | ||||
8110 | unsigned FieldIndex = RL.getNonVirtualBaseLLVMFieldNo(Base); | |||
8111 | RecordLayout[FieldIndex] = Base; | |||
8112 | } | |||
8113 | // Fill in virtual bases. | |||
8114 | for (const auto &I : RD->vbases()) { | |||
8115 | const auto *Base = I.getType()->getAsCXXRecordDecl(); | |||
8116 | // Ignore empty bases. | |||
8117 | if (Base->isEmpty()) | |||
8118 | continue; | |||
8119 | unsigned FieldIndex = RL.getVirtualBaseIndex(Base); | |||
8120 | if (RecordLayout[FieldIndex]) | |||
8121 | continue; | |||
8122 | RecordLayout[FieldIndex] = Base; | |||
8123 | } | |||
8124 | // Fill in all the fields. | |||
8125 | assert(!RD->isUnion() && "Unexpected union.")(static_cast <bool> (!RD->isUnion() && "Unexpected union." ) ? void (0) : __assert_fail ("!RD->isUnion() && \"Unexpected union.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8125, __extension__ __PRETTY_FUNCTION__)); | |||
8126 | for (const auto *Field : RD->fields()) { | |||
8127 | // Fill in non-bitfields. (Bitfields always use a zero pattern, which we | |||
8128 | // will fill in later.) | |||
8129 | if (!Field->isBitField() && !Field->isZeroSize(CGF.getContext())) { | |||
8130 | unsigned FieldIndex = RL.getLLVMFieldNo(Field); | |||
8131 | RecordLayout[FieldIndex] = Field; | |||
8132 | } | |||
8133 | } | |||
8134 | for (const llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *> | |||
8135 | &Data : RecordLayout) { | |||
8136 | if (Data.isNull()) | |||
8137 | continue; | |||
8138 | if (const auto *Base = Data.dyn_cast<const CXXRecordDecl *>()) | |||
8139 | getPlainLayout(Base, Layout, /*AsBase=*/true); | |||
8140 | else | |||
8141 | Layout.push_back(Data.get<const FieldDecl *>()); | |||
8142 | } | |||
8143 | } | |||
8144 | ||||
8145 | /// Generate all the base pointers, section pointers, sizes, map types, and | |||
8146 | /// mappers for the extracted mappable expressions (all included in \a | |||
8147 | /// CombinedInfo). Also, for each item that relates with a device pointer, a | |||
8148 | /// pair of the relevant declaration and index where it occurs is appended to | |||
8149 | /// the device pointers info array. | |||
8150 | void generateAllInfoForClauses( | |||
8151 | ArrayRef<const OMPClause *> Clauses, MapCombinedInfoTy &CombinedInfo, | |||
8152 | const llvm::DenseSet<CanonicalDeclPtr<const Decl>> &SkipVarSet = | |||
8153 | llvm::DenseSet<CanonicalDeclPtr<const Decl>>()) const { | |||
8154 | // We have to process the component lists that relate with the same | |||
8155 | // declaration in a single chunk so that we can generate the map flags | |||
8156 | // correctly. Therefore, we organize all lists in a map. | |||
8157 | enum MapKind { Present, Allocs, Other, Total }; | |||
8158 | llvm::MapVector<CanonicalDeclPtr<const Decl>, | |||
8159 | SmallVector<SmallVector<MapInfo, 8>, 4>> | |||
8160 | Info; | |||
8161 | ||||
8162 | // Helper function to fill the information map for the different supported | |||
8163 | // clauses. | |||
8164 | auto &&InfoGen = | |||
8165 | [&Info, &SkipVarSet]( | |||
8166 | const ValueDecl *D, MapKind Kind, | |||
8167 | OMPClauseMappableExprCommon::MappableExprComponentListRef L, | |||
8168 | OpenMPMapClauseKind MapType, | |||
8169 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
8170 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, | |||
8171 | bool ReturnDevicePointer, bool IsImplicit, const ValueDecl *Mapper, | |||
8172 | const Expr *VarRef = nullptr, bool ForDeviceAddr = false) { | |||
8173 | if (SkipVarSet.contains(D)) | |||
8174 | return; | |||
8175 | auto It = Info.find(D); | |||
8176 | if (It == Info.end()) | |||
8177 | It = Info | |||
8178 | .insert(std::make_pair( | |||
8179 | D, SmallVector<SmallVector<MapInfo, 8>, 4>(Total))) | |||
8180 | .first; | |||
8181 | It->second[Kind].emplace_back( | |||
8182 | L, MapType, MapModifiers, MotionModifiers, ReturnDevicePointer, | |||
8183 | IsImplicit, Mapper, VarRef, ForDeviceAddr); | |||
8184 | }; | |||
8185 | ||||
8186 | for (const auto *Cl : Clauses) { | |||
8187 | const auto *C = dyn_cast<OMPMapClause>(Cl); | |||
8188 | if (!C) | |||
8189 | continue; | |||
8190 | MapKind Kind = Other; | |||
8191 | if (llvm::is_contained(C->getMapTypeModifiers(), | |||
8192 | OMPC_MAP_MODIFIER_present)) | |||
8193 | Kind = Present; | |||
8194 | else if (C->getMapType() == OMPC_MAP_alloc) | |||
8195 | Kind = Allocs; | |||
8196 | const auto *EI = C->getVarRefs().begin(); | |||
8197 | for (const auto L : C->component_lists()) { | |||
8198 | const Expr *E = (C->getMapLoc().isValid()) ? *EI : nullptr; | |||
8199 | InfoGen(std::get<0>(L), Kind, std::get<1>(L), C->getMapType(), | |||
8200 | C->getMapTypeModifiers(), llvm::None, | |||
8201 | /*ReturnDevicePointer=*/false, C->isImplicit(), std::get<2>(L), | |||
8202 | E); | |||
8203 | ++EI; | |||
8204 | } | |||
8205 | } | |||
8206 | for (const auto *Cl : Clauses) { | |||
8207 | const auto *C = dyn_cast<OMPToClause>(Cl); | |||
8208 | if (!C) | |||
8209 | continue; | |||
8210 | MapKind Kind = Other; | |||
8211 | if (llvm::is_contained(C->getMotionModifiers(), | |||
8212 | OMPC_MOTION_MODIFIER_present)) | |||
8213 | Kind = Present; | |||
8214 | const auto *EI = C->getVarRefs().begin(); | |||
8215 | for (const auto L : C->component_lists()) { | |||
8216 | InfoGen(std::get<0>(L), Kind, std::get<1>(L), OMPC_MAP_to, llvm::None, | |||
8217 | C->getMotionModifiers(), /*ReturnDevicePointer=*/false, | |||
8218 | C->isImplicit(), std::get<2>(L), *EI); | |||
8219 | ++EI; | |||
8220 | } | |||
8221 | } | |||
8222 | for (const auto *Cl : Clauses) { | |||
8223 | const auto *C = dyn_cast<OMPFromClause>(Cl); | |||
8224 | if (!C) | |||
8225 | continue; | |||
8226 | MapKind Kind = Other; | |||
8227 | if (llvm::is_contained(C->getMotionModifiers(), | |||
8228 | OMPC_MOTION_MODIFIER_present)) | |||
8229 | Kind = Present; | |||
8230 | const auto *EI = C->getVarRefs().begin(); | |||
8231 | for (const auto L : C->component_lists()) { | |||
8232 | InfoGen(std::get<0>(L), Kind, std::get<1>(L), OMPC_MAP_from, llvm::None, | |||
8233 | C->getMotionModifiers(), /*ReturnDevicePointer=*/false, | |||
8234 | C->isImplicit(), std::get<2>(L), *EI); | |||
8235 | ++EI; | |||
8236 | } | |||
8237 | } | |||
8238 | ||||
8239 | // Look at the use_device_ptr and use_device_addr clauses information and | |||
8240 | // mark the existing map entries as such. If there is no map information for | |||
8241 | // an entry in the use_device_ptr and use_device_addr list, we create one | |||
8242 | // with map type 'alloc' and zero size section. It is the user fault if that | |||
8243 | // was not mapped before. If there is no map information and the pointer is | |||
8244 | // a struct member, then we defer the emission of that entry until the whole | |||
8245 | // struct has been processed. | |||
8246 | llvm::MapVector<CanonicalDeclPtr<const Decl>, | |||
8247 | SmallVector<DeferredDevicePtrEntryTy, 4>> | |||
8248 | DeferredInfo; | |||
8249 | MapCombinedInfoTy UseDeviceDataCombinedInfo; | |||
8250 | ||||
8251 | auto &&UseDeviceDataCombinedInfoGen = | |||
8252 | [&UseDeviceDataCombinedInfo](const ValueDecl *VD, llvm::Value *Ptr, | |||
8253 | CodeGenFunction &CGF) { | |||
8254 | UseDeviceDataCombinedInfo.Exprs.push_back(VD); | |||
8255 | UseDeviceDataCombinedInfo.BasePointers.emplace_back(Ptr, VD); | |||
8256 | UseDeviceDataCombinedInfo.Pointers.push_back(Ptr); | |||
8257 | UseDeviceDataCombinedInfo.Sizes.push_back( | |||
8258 | llvm::Constant::getNullValue(CGF.Int64Ty)); | |||
8259 | UseDeviceDataCombinedInfo.Types.push_back(OMP_MAP_RETURN_PARAM); | |||
8260 | UseDeviceDataCombinedInfo.Mappers.push_back(nullptr); | |||
8261 | }; | |||
8262 | ||||
8263 | auto &&MapInfoGen = | |||
8264 | [&DeferredInfo, &UseDeviceDataCombinedInfoGen, | |||
8265 | &InfoGen](CodeGenFunction &CGF, const Expr *IE, const ValueDecl *VD, | |||
8266 | OMPClauseMappableExprCommon::MappableExprComponentListRef | |||
8267 | Components, | |||
8268 | bool IsImplicit, bool IsDevAddr) { | |||
8269 | // We didn't find any match in our map information - generate a zero | |||
8270 | // size array section - if the pointer is a struct member we defer | |||
8271 | // this action until the whole struct has been processed. | |||
8272 | if (isa<MemberExpr>(IE)) { | |||
8273 | // Insert the pointer into Info to be processed by | |||
8274 | // generateInfoForComponentList. Because it is a member pointer | |||
8275 | // without a pointee, no entry will be generated for it, therefore | |||
8276 | // we need to generate one after the whole struct has been | |||
8277 | // processed. Nonetheless, generateInfoForComponentList must be | |||
8278 | // called to take the pointer into account for the calculation of | |||
8279 | // the range of the partial struct. | |||
8280 | InfoGen(nullptr, Other, Components, OMPC_MAP_unknown, llvm::None, | |||
8281 | llvm::None, /*ReturnDevicePointer=*/false, IsImplicit, | |||
8282 | nullptr, nullptr, IsDevAddr); | |||
8283 | DeferredInfo[nullptr].emplace_back(IE, VD, IsDevAddr); | |||
8284 | } else { | |||
8285 | llvm::Value *Ptr; | |||
8286 | if (IsDevAddr) { | |||
8287 | if (IE->isGLValue()) | |||
8288 | Ptr = CGF.EmitLValue(IE).getPointer(CGF); | |||
8289 | else | |||
8290 | Ptr = CGF.EmitScalarExpr(IE); | |||
8291 | } else { | |||
8292 | Ptr = CGF.EmitLoadOfScalar(CGF.EmitLValue(IE), IE->getExprLoc()); | |||
8293 | } | |||
8294 | UseDeviceDataCombinedInfoGen(VD, Ptr, CGF); | |||
8295 | } | |||
8296 | }; | |||
8297 | ||||
8298 | auto &&IsMapInfoExist = [&Info](CodeGenFunction &CGF, const ValueDecl *VD, | |||
8299 | const Expr *IE, bool IsDevAddr) -> bool { | |||
8300 | // We potentially have map information for this declaration already. | |||
8301 | // Look for the first set of components that refer to it. If found, | |||
8302 | // return true. | |||
8303 | // If the first component is a member expression, we have to look into | |||
8304 | // 'this', which maps to null in the map of map information. Otherwise | |||
8305 | // look directly for the information. | |||
8306 | auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD); | |||
8307 | if (It != Info.end()) { | |||
8308 | bool Found = false; | |||
8309 | for (auto &Data : It->second) { | |||
8310 | auto *CI = llvm::find_if(Data, [VD](const MapInfo &MI) { | |||
8311 | return MI.Components.back().getAssociatedDeclaration() == VD; | |||
8312 | }); | |||
8313 | // If we found a map entry, signal that the pointer has to be | |||
8314 | // returned and move on to the next declaration. Exclude cases where | |||
8315 | // the base pointer is mapped as array subscript, array section or | |||
8316 | // array shaping. The base address is passed as a pointer to base in | |||
8317 | // this case and cannot be used as a base for use_device_ptr list | |||
8318 | // item. | |||
8319 | if (CI != Data.end()) { | |||
8320 | if (IsDevAddr) { | |||
8321 | CI->ReturnDevicePointer = true; | |||
8322 | Found = true; | |||
8323 | break; | |||
8324 | } else { | |||
8325 | auto PrevCI = std::next(CI->Components.rbegin()); | |||
8326 | const auto *VarD = dyn_cast<VarDecl>(VD); | |||
8327 | if (CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory() || | |||
8328 | isa<MemberExpr>(IE) || | |||
8329 | !VD->getType().getNonReferenceType()->isPointerType() || | |||
8330 | PrevCI == CI->Components.rend() || | |||
8331 | isa<MemberExpr>(PrevCI->getAssociatedExpression()) || !VarD || | |||
8332 | VarD->hasLocalStorage()) { | |||
8333 | CI->ReturnDevicePointer = true; | |||
8334 | Found = true; | |||
8335 | break; | |||
8336 | } | |||
8337 | } | |||
8338 | } | |||
8339 | } | |||
8340 | return Found; | |||
8341 | } | |||
8342 | return false; | |||
8343 | }; | |||
8344 | ||||
8345 | // Look at the use_device_ptr clause information and mark the existing map | |||
8346 | // entries as such. If there is no map information for an entry in the | |||
8347 | // use_device_ptr list, we create one with map type 'alloc' and zero size | |||
8348 | // section. It is the user fault if that was not mapped before. If there is | |||
8349 | // no map information and the pointer is a struct member, then we defer the | |||
8350 | // emission of that entry until the whole struct has been processed. | |||
8351 | for (const auto *Cl : Clauses) { | |||
8352 | const auto *C = dyn_cast<OMPUseDevicePtrClause>(Cl); | |||
8353 | if (!C) | |||
8354 | continue; | |||
8355 | for (const auto L : C->component_lists()) { | |||
8356 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components = | |||
8357 | std::get<1>(L); | |||
8358 | assert(!Components.empty() &&(static_cast <bool> (!Components.empty() && "Not expecting empty list of components!" ) ? void (0) : __assert_fail ("!Components.empty() && \"Not expecting empty list of components!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8359, __extension__ __PRETTY_FUNCTION__)) | |||
8359 | "Not expecting empty list of components!")(static_cast <bool> (!Components.empty() && "Not expecting empty list of components!" ) ? void (0) : __assert_fail ("!Components.empty() && \"Not expecting empty list of components!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8359, __extension__ __PRETTY_FUNCTION__)); | |||
8360 | const ValueDecl *VD = Components.back().getAssociatedDeclaration(); | |||
8361 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); | |||
8362 | const Expr *IE = Components.back().getAssociatedExpression(); | |||
8363 | if (IsMapInfoExist(CGF, VD, IE, /*IsDevAddr=*/false)) | |||
8364 | continue; | |||
8365 | MapInfoGen(CGF, IE, VD, Components, C->isImplicit(), | |||
8366 | /*IsDevAddr=*/false); | |||
8367 | } | |||
8368 | } | |||
8369 | ||||
8370 | llvm::SmallDenseSet<CanonicalDeclPtr<const Decl>, 4> Processed; | |||
8371 | for (const auto *Cl : Clauses) { | |||
8372 | const auto *C = dyn_cast<OMPUseDeviceAddrClause>(Cl); | |||
8373 | if (!C) | |||
8374 | continue; | |||
8375 | for (const auto L : C->component_lists()) { | |||
8376 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components = | |||
8377 | std::get<1>(L); | |||
8378 | assert(!std::get<1>(L).empty() &&(static_cast <bool> (!std::get<1>(L).empty() && "Not expecting empty list of components!") ? void (0) : __assert_fail ("!std::get<1>(L).empty() && \"Not expecting empty list of components!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8379, __extension__ __PRETTY_FUNCTION__)) | |||
8379 | "Not expecting empty list of components!")(static_cast <bool> (!std::get<1>(L).empty() && "Not expecting empty list of components!") ? void (0) : __assert_fail ("!std::get<1>(L).empty() && \"Not expecting empty list of components!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8379, __extension__ __PRETTY_FUNCTION__)); | |||
8380 | const ValueDecl *VD = std::get<1>(L).back().getAssociatedDeclaration(); | |||
8381 | if (!Processed.insert(VD).second) | |||
8382 | continue; | |||
8383 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); | |||
8384 | const Expr *IE = std::get<1>(L).back().getAssociatedExpression(); | |||
8385 | if (IsMapInfoExist(CGF, VD, IE, /*IsDevAddr=*/true)) | |||
8386 | continue; | |||
8387 | MapInfoGen(CGF, IE, VD, Components, C->isImplicit(), | |||
8388 | /*IsDevAddr=*/true); | |||
8389 | } | |||
8390 | } | |||
8391 | ||||
8392 | for (const auto &Data : Info) { | |||
8393 | StructRangeInfoTy PartialStruct; | |||
8394 | // Temporary generated information. | |||
8395 | MapCombinedInfoTy CurInfo; | |||
8396 | const Decl *D = Data.first; | |||
8397 | const ValueDecl *VD = cast_or_null<ValueDecl>(D); | |||
8398 | for (const auto &M : Data.second) { | |||
8399 | for (const MapInfo &L : M) { | |||
8400 | assert(!L.Components.empty() &&(static_cast <bool> (!L.Components.empty() && "Not expecting declaration with no component lists." ) ? void (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8401, __extension__ __PRETTY_FUNCTION__)) | |||
8401 | "Not expecting declaration with no component lists.")(static_cast <bool> (!L.Components.empty() && "Not expecting declaration with no component lists." ) ? void (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8401, __extension__ __PRETTY_FUNCTION__)); | |||
8402 | ||||
8403 | // Remember the current base pointer index. | |||
8404 | unsigned CurrentBasePointersIdx = CurInfo.BasePointers.size(); | |||
8405 | CurInfo.NonContigInfo.IsNonContiguous = | |||
8406 | L.Components.back().isNonContiguous(); | |||
8407 | generateInfoForComponentList( | |||
8408 | L.MapType, L.MapModifiers, L.MotionModifiers, L.Components, | |||
8409 | CurInfo, PartialStruct, /*IsFirstComponentList=*/false, | |||
8410 | L.IsImplicit, L.Mapper, L.ForDeviceAddr, VD, L.VarRef); | |||
8411 | ||||
8412 | // If this entry relates with a device pointer, set the relevant | |||
8413 | // declaration and add the 'return pointer' flag. | |||
8414 | if (L.ReturnDevicePointer) { | |||
8415 | assert(CurInfo.BasePointers.size() > CurrentBasePointersIdx &&(static_cast <bool> (CurInfo.BasePointers.size() > CurrentBasePointersIdx && "Unexpected number of mapped base pointers.") ? void (0) : __assert_fail ("CurInfo.BasePointers.size() > CurrentBasePointersIdx && \"Unexpected number of mapped base pointers.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8416, __extension__ __PRETTY_FUNCTION__)) | |||
8416 | "Unexpected number of mapped base pointers.")(static_cast <bool> (CurInfo.BasePointers.size() > CurrentBasePointersIdx && "Unexpected number of mapped base pointers.") ? void (0) : __assert_fail ("CurInfo.BasePointers.size() > CurrentBasePointersIdx && \"Unexpected number of mapped base pointers.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8416, __extension__ __PRETTY_FUNCTION__)); | |||
8417 | ||||
8418 | const ValueDecl *RelevantVD = | |||
8419 | L.Components.back().getAssociatedDeclaration(); | |||
8420 | assert(RelevantVD &&(static_cast <bool> (RelevantVD && "No relevant declaration related with device pointer??" ) ? void (0) : __assert_fail ("RelevantVD && \"No relevant declaration related with device pointer??\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8421, __extension__ __PRETTY_FUNCTION__)) | |||
8421 | "No relevant declaration related with device pointer??")(static_cast <bool> (RelevantVD && "No relevant declaration related with device pointer??" ) ? void (0) : __assert_fail ("RelevantVD && \"No relevant declaration related with device pointer??\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8421, __extension__ __PRETTY_FUNCTION__)); | |||
8422 | ||||
8423 | CurInfo.BasePointers[CurrentBasePointersIdx].setDevicePtrDecl( | |||
8424 | RelevantVD); | |||
8425 | CurInfo.Types[CurrentBasePointersIdx] |= OMP_MAP_RETURN_PARAM; | |||
8426 | } | |||
8427 | } | |||
8428 | } | |||
8429 | ||||
8430 | // Append any pending zero-length pointers which are struct members and | |||
8431 | // used with use_device_ptr or use_device_addr. | |||
8432 | auto CI = DeferredInfo.find(Data.first); | |||
8433 | if (CI != DeferredInfo.end()) { | |||
8434 | for (const DeferredDevicePtrEntryTy &L : CI->second) { | |||
8435 | llvm::Value *BasePtr; | |||
8436 | llvm::Value *Ptr; | |||
8437 | if (L.ForDeviceAddr) { | |||
8438 | if (L.IE->isGLValue()) | |||
8439 | Ptr = this->CGF.EmitLValue(L.IE).getPointer(CGF); | |||
8440 | else | |||
8441 | Ptr = this->CGF.EmitScalarExpr(L.IE); | |||
8442 | BasePtr = Ptr; | |||
8443 | // Entry is RETURN_PARAM. Also, set the placeholder value | |||
8444 | // MEMBER_OF=FFFF so that the entry is later updated with the | |||
8445 | // correct value of MEMBER_OF. | |||
8446 | CurInfo.Types.push_back(OMP_MAP_RETURN_PARAM | OMP_MAP_MEMBER_OF); | |||
8447 | } else { | |||
8448 | BasePtr = this->CGF.EmitLValue(L.IE).getPointer(CGF); | |||
8449 | Ptr = this->CGF.EmitLoadOfScalar(this->CGF.EmitLValue(L.IE), | |||
8450 | L.IE->getExprLoc()); | |||
8451 | // Entry is PTR_AND_OBJ and RETURN_PARAM. Also, set the | |||
8452 | // placeholder value MEMBER_OF=FFFF so that the entry is later | |||
8453 | // updated with the correct value of MEMBER_OF. | |||
8454 | CurInfo.Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_RETURN_PARAM | | |||
8455 | OMP_MAP_MEMBER_OF); | |||
8456 | } | |||
8457 | CurInfo.Exprs.push_back(L.VD); | |||
8458 | CurInfo.BasePointers.emplace_back(BasePtr, L.VD); | |||
8459 | CurInfo.Pointers.push_back(Ptr); | |||
8460 | CurInfo.Sizes.push_back( | |||
8461 | llvm::Constant::getNullValue(this->CGF.Int64Ty)); | |||
8462 | CurInfo.Mappers.push_back(nullptr); | |||
8463 | } | |||
8464 | } | |||
8465 | // If there is an entry in PartialStruct it means we have a struct with | |||
8466 | // individual members mapped. Emit an extra combined entry. | |||
8467 | if (PartialStruct.Base.isValid()) { | |||
8468 | CurInfo.NonContigInfo.Dims.push_back(0); | |||
8469 | emitCombinedEntry(CombinedInfo, CurInfo.Types, PartialStruct, VD); | |||
8470 | } | |||
8471 | ||||
8472 | // We need to append the results of this capture to what we already | |||
8473 | // have. | |||
8474 | CombinedInfo.append(CurInfo); | |||
8475 | } | |||
8476 | // Append data for use_device_ptr clauses. | |||
8477 | CombinedInfo.append(UseDeviceDataCombinedInfo); | |||
8478 | } | |||
8479 | ||||
8480 | public: | |||
8481 | MappableExprsHandler(const OMPExecutableDirective &Dir, CodeGenFunction &CGF) | |||
8482 | : CurDir(&Dir), CGF(CGF) { | |||
8483 | // Extract firstprivate clause information. | |||
8484 | for (const auto *C : Dir.getClausesOfKind<OMPFirstprivateClause>()) | |||
8485 | for (const auto *D : C->varlists()) | |||
8486 | FirstPrivateDecls.try_emplace( | |||
8487 | cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl()), C->isImplicit()); | |||
8488 | // Extract implicit firstprivates from uses_allocators clauses. | |||
8489 | for (const auto *C : Dir.getClausesOfKind<OMPUsesAllocatorsClause>()) { | |||
8490 | for (unsigned I = 0, E = C->getNumberOfAllocators(); I < E; ++I) { | |||
8491 | OMPUsesAllocatorsClause::Data D = C->getAllocatorData(I); | |||
8492 | if (const auto *DRE = dyn_cast_or_null<DeclRefExpr>(D.AllocatorTraits)) | |||
8493 | FirstPrivateDecls.try_emplace(cast<VarDecl>(DRE->getDecl()), | |||
8494 | /*Implicit=*/true); | |||
8495 | else if (const auto *VD = dyn_cast<VarDecl>( | |||
8496 | cast<DeclRefExpr>(D.Allocator->IgnoreParenImpCasts()) | |||
8497 | ->getDecl())) | |||
8498 | FirstPrivateDecls.try_emplace(VD, /*Implicit=*/true); | |||
8499 | } | |||
8500 | } | |||
8501 | // Extract device pointer clause information. | |||
8502 | for (const auto *C : Dir.getClausesOfKind<OMPIsDevicePtrClause>()) | |||
8503 | for (auto L : C->component_lists()) | |||
8504 | DevPointersMap[std::get<0>(L)].push_back(std::get<1>(L)); | |||
8505 | // Extract device addr clause information. | |||
8506 | for (const auto *C : Dir.getClausesOfKind<OMPHasDeviceAddrClause>()) | |||
8507 | for (auto L : C->component_lists()) | |||
8508 | HasDevAddrsMap[std::get<0>(L)].push_back(std::get<1>(L)); | |||
8509 | // Extract map information. | |||
8510 | for (const auto *C : Dir.getClausesOfKind<OMPMapClause>()) { | |||
8511 | if (C->getMapType() != OMPC_MAP_to) | |||
8512 | continue; | |||
8513 | for (auto L : C->component_lists()) { | |||
8514 | const ValueDecl *VD = std::get<0>(L); | |||
8515 | const auto *RD = VD ? VD->getType() | |||
8516 | .getCanonicalType() | |||
8517 | .getNonReferenceType() | |||
8518 | ->getAsCXXRecordDecl() | |||
8519 | : nullptr; | |||
8520 | if (RD && RD->isLambda()) | |||
8521 | LambdasMap.try_emplace(std::get<0>(L), C); | |||
8522 | } | |||
8523 | } | |||
8524 | } | |||
8525 | ||||
8526 | /// Constructor for the declare mapper directive. | |||
8527 | MappableExprsHandler(const OMPDeclareMapperDecl &Dir, CodeGenFunction &CGF) | |||
8528 | : CurDir(&Dir), CGF(CGF) {} | |||
8529 | ||||
8530 | /// Generate code for the combined entry if we have a partially mapped struct | |||
8531 | /// and take care of the mapping flags of the arguments corresponding to | |||
8532 | /// individual struct members. | |||
8533 | void emitCombinedEntry(MapCombinedInfoTy &CombinedInfo, | |||
8534 | MapFlagsArrayTy &CurTypes, | |||
8535 | const StructRangeInfoTy &PartialStruct, | |||
8536 | const ValueDecl *VD = nullptr, | |||
8537 | bool NotTargetParams = true) const { | |||
8538 | if (CurTypes.size() == 1 && | |||
8539 | ((CurTypes.back() & OMP_MAP_MEMBER_OF) != OMP_MAP_MEMBER_OF) && | |||
8540 | !PartialStruct.IsArraySection) | |||
8541 | return; | |||
8542 | Address LBAddr = PartialStruct.LowestElem.second; | |||
8543 | Address HBAddr = PartialStruct.HighestElem.second; | |||
8544 | if (PartialStruct.HasCompleteRecord) { | |||
8545 | LBAddr = PartialStruct.LB; | |||
8546 | HBAddr = PartialStruct.LB; | |||
8547 | } | |||
8548 | CombinedInfo.Exprs.push_back(VD); | |||
8549 | // Base is the base of the struct | |||
8550 | CombinedInfo.BasePointers.push_back(PartialStruct.Base.getPointer()); | |||
8551 | // Pointer is the address of the lowest element | |||
8552 | llvm::Value *LB = LBAddr.getPointer(); | |||
8553 | CombinedInfo.Pointers.push_back(LB); | |||
8554 | // There should not be a mapper for a combined entry. | |||
8555 | CombinedInfo.Mappers.push_back(nullptr); | |||
8556 | // Size is (addr of {highest+1} element) - (addr of lowest element) | |||
8557 | llvm::Value *HB = HBAddr.getPointer(); | |||
8558 | llvm::Value *HAddr = | |||
8559 | CGF.Builder.CreateConstGEP1_32(HBAddr.getElementType(), HB, /*Idx0=*/1); | |||
8560 | llvm::Value *CLAddr = CGF.Builder.CreatePointerCast(LB, CGF.VoidPtrTy); | |||
8561 | llvm::Value *CHAddr = CGF.Builder.CreatePointerCast(HAddr, CGF.VoidPtrTy); | |||
8562 | llvm::Value *Diff = CGF.Builder.CreatePtrDiff(CGF.Int8Ty, CHAddr, CLAddr); | |||
8563 | llvm::Value *Size = CGF.Builder.CreateIntCast(Diff, CGF.Int64Ty, | |||
8564 | /*isSigned=*/false); | |||
8565 | CombinedInfo.Sizes.push_back(Size); | |||
8566 | // Map type is always TARGET_PARAM, if generate info for captures. | |||
8567 | CombinedInfo.Types.push_back(NotTargetParams ? OMP_MAP_NONE | |||
8568 | : OMP_MAP_TARGET_PARAM); | |||
8569 | // If any element has the present modifier, then make sure the runtime | |||
8570 | // doesn't attempt to allocate the struct. | |||
8571 | if (CurTypes.end() != | |||
8572 | llvm::find_if(CurTypes, [](OpenMPOffloadMappingFlags Type) { | |||
8573 | return Type & OMP_MAP_PRESENT; | |||
8574 | })) | |||
8575 | CombinedInfo.Types.back() |= OMP_MAP_PRESENT; | |||
8576 | // Remove TARGET_PARAM flag from the first element | |||
8577 | (*CurTypes.begin()) &= ~OMP_MAP_TARGET_PARAM; | |||
8578 | // If any element has the ompx_hold modifier, then make sure the runtime | |||
8579 | // uses the hold reference count for the struct as a whole so that it won't | |||
8580 | // be unmapped by an extra dynamic reference count decrement. Add it to all | |||
8581 | // elements as well so the runtime knows which reference count to check | |||
8582 | // when determining whether it's time for device-to-host transfers of | |||
8583 | // individual elements. | |||
8584 | if (CurTypes.end() != | |||
8585 | llvm::find_if(CurTypes, [](OpenMPOffloadMappingFlags Type) { | |||
8586 | return Type & OMP_MAP_OMPX_HOLD; | |||
8587 | })) { | |||
8588 | CombinedInfo.Types.back() |= OMP_MAP_OMPX_HOLD; | |||
8589 | for (auto &M : CurTypes) | |||
8590 | M |= OMP_MAP_OMPX_HOLD; | |||
8591 | } | |||
8592 | ||||
8593 | // All other current entries will be MEMBER_OF the combined entry | |||
8594 | // (except for PTR_AND_OBJ entries which do not have a placeholder value | |||
8595 | // 0xFFFF in the MEMBER_OF field). | |||
8596 | OpenMPOffloadMappingFlags MemberOfFlag = | |||
8597 | getMemberOfFlag(CombinedInfo.BasePointers.size() - 1); | |||
8598 | for (auto &M : CurTypes) | |||
8599 | setCorrectMemberOfFlag(M, MemberOfFlag); | |||
8600 | } | |||
8601 | ||||
8602 | /// Generate all the base pointers, section pointers, sizes, map types, and | |||
8603 | /// mappers for the extracted mappable expressions (all included in \a | |||
8604 | /// CombinedInfo). Also, for each item that relates with a device pointer, a | |||
8605 | /// pair of the relevant declaration and index where it occurs is appended to | |||
8606 | /// the device pointers info array. | |||
8607 | void generateAllInfo( | |||
8608 | MapCombinedInfoTy &CombinedInfo, | |||
8609 | const llvm::DenseSet<CanonicalDeclPtr<const Decl>> &SkipVarSet = | |||
8610 | llvm::DenseSet<CanonicalDeclPtr<const Decl>>()) const { | |||
8611 | assert(CurDir.is<const OMPExecutableDirective *>() &&(static_cast <bool> (CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? void ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8612, __extension__ __PRETTY_FUNCTION__)) | |||
8612 | "Expect a executable directive")(static_cast <bool> (CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? void ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8612, __extension__ __PRETTY_FUNCTION__)); | |||
8613 | const auto *CurExecDir = CurDir.get<const OMPExecutableDirective *>(); | |||
8614 | generateAllInfoForClauses(CurExecDir->clauses(), CombinedInfo, SkipVarSet); | |||
8615 | } | |||
8616 | ||||
8617 | /// Generate all the base pointers, section pointers, sizes, map types, and | |||
8618 | /// mappers for the extracted map clauses of user-defined mapper (all included | |||
8619 | /// in \a CombinedInfo). | |||
8620 | void generateAllInfoForMapper(MapCombinedInfoTy &CombinedInfo) const { | |||
8621 | assert(CurDir.is<const OMPDeclareMapperDecl *>() &&(static_cast <bool> (CurDir.is<const OMPDeclareMapperDecl *>() && "Expect a declare mapper directive") ? void (0) : __assert_fail ("CurDir.is<const OMPDeclareMapperDecl *>() && \"Expect a declare mapper directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8622, __extension__ __PRETTY_FUNCTION__)) | |||
8622 | "Expect a declare mapper directive")(static_cast <bool> (CurDir.is<const OMPDeclareMapperDecl *>() && "Expect a declare mapper directive") ? void (0) : __assert_fail ("CurDir.is<const OMPDeclareMapperDecl *>() && \"Expect a declare mapper directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8622, __extension__ __PRETTY_FUNCTION__)); | |||
8623 | const auto *CurMapperDir = CurDir.get<const OMPDeclareMapperDecl *>(); | |||
8624 | generateAllInfoForClauses(CurMapperDir->clauses(), CombinedInfo); | |||
8625 | } | |||
8626 | ||||
8627 | /// Emit capture info for lambdas for variables captured by reference. | |||
8628 | void generateInfoForLambdaCaptures( | |||
8629 | const ValueDecl *VD, llvm::Value *Arg, MapCombinedInfoTy &CombinedInfo, | |||
8630 | llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers) const { | |||
8631 | QualType VDType = VD->getType().getCanonicalType().getNonReferenceType(); | |||
8632 | const auto *RD = VDType->getAsCXXRecordDecl(); | |||
8633 | if (!RD || !RD->isLambda()) | |||
8634 | return; | |||
8635 | Address VDAddr(Arg, CGF.ConvertTypeForMem(VDType), | |||
8636 | CGF.getContext().getDeclAlign(VD)); | |||
8637 | LValue VDLVal = CGF.MakeAddrLValue(VDAddr, VDType); | |||
8638 | llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures; | |||
8639 | FieldDecl *ThisCapture = nullptr; | |||
8640 | RD->getCaptureFields(Captures, ThisCapture); | |||
8641 | if (ThisCapture) { | |||
8642 | LValue ThisLVal = | |||
8643 | CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); | |||
8644 | LValue ThisLValVal = CGF.EmitLValueForField(VDLVal, ThisCapture); | |||
8645 | LambdaPointers.try_emplace(ThisLVal.getPointer(CGF), | |||
8646 | VDLVal.getPointer(CGF)); | |||
8647 | CombinedInfo.Exprs.push_back(VD); | |||
8648 | CombinedInfo.BasePointers.push_back(ThisLVal.getPointer(CGF)); | |||
8649 | CombinedInfo.Pointers.push_back(ThisLValVal.getPointer(CGF)); | |||
8650 | CombinedInfo.Sizes.push_back( | |||
8651 | CGF.Builder.CreateIntCast(CGF.getTypeSize(CGF.getContext().VoidPtrTy), | |||
8652 | CGF.Int64Ty, /*isSigned=*/true)); | |||
8653 | CombinedInfo.Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | |||
8654 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT); | |||
8655 | CombinedInfo.Mappers.push_back(nullptr); | |||
8656 | } | |||
8657 | for (const LambdaCapture &LC : RD->captures()) { | |||
8658 | if (!LC.capturesVariable()) | |||
8659 | continue; | |||
8660 | const VarDecl *VD = cast<VarDecl>(LC.getCapturedVar()); | |||
8661 | if (LC.getCaptureKind() != LCK_ByRef && !VD->getType()->isPointerType()) | |||
8662 | continue; | |||
8663 | auto It = Captures.find(VD); | |||
8664 | assert(It != Captures.end() && "Found lambda capture without field.")(static_cast <bool> (It != Captures.end() && "Found lambda capture without field." ) ? void (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8664, __extension__ __PRETTY_FUNCTION__)); | |||
8665 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); | |||
8666 | if (LC.getCaptureKind() == LCK_ByRef) { | |||
8667 | LValue VarLValVal = CGF.EmitLValueForField(VDLVal, It->second); | |||
8668 | LambdaPointers.try_emplace(VarLVal.getPointer(CGF), | |||
8669 | VDLVal.getPointer(CGF)); | |||
8670 | CombinedInfo.Exprs.push_back(VD); | |||
8671 | CombinedInfo.BasePointers.push_back(VarLVal.getPointer(CGF)); | |||
8672 | CombinedInfo.Pointers.push_back(VarLValVal.getPointer(CGF)); | |||
8673 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | |||
8674 | CGF.getTypeSize( | |||
8675 | VD->getType().getCanonicalType().getNonReferenceType()), | |||
8676 | CGF.Int64Ty, /*isSigned=*/true)); | |||
8677 | } else { | |||
8678 | RValue VarRVal = CGF.EmitLoadOfLValue(VarLVal, RD->getLocation()); | |||
8679 | LambdaPointers.try_emplace(VarLVal.getPointer(CGF), | |||
8680 | VDLVal.getPointer(CGF)); | |||
8681 | CombinedInfo.Exprs.push_back(VD); | |||
8682 | CombinedInfo.BasePointers.push_back(VarLVal.getPointer(CGF)); | |||
8683 | CombinedInfo.Pointers.push_back(VarRVal.getScalarVal()); | |||
8684 | CombinedInfo.Sizes.push_back(llvm::ConstantInt::get(CGF.Int64Ty, 0)); | |||
8685 | } | |||
8686 | CombinedInfo.Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | |||
8687 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT); | |||
8688 | CombinedInfo.Mappers.push_back(nullptr); | |||
8689 | } | |||
8690 | } | |||
8691 | ||||
8692 | /// Set correct indices for lambdas captures. | |||
8693 | void adjustMemberOfForLambdaCaptures( | |||
8694 | const llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers, | |||
8695 | MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers, | |||
8696 | MapFlagsArrayTy &Types) const { | |||
8697 | for (unsigned I = 0, E = Types.size(); I < E; ++I) { | |||
8698 | // Set correct member_of idx for all implicit lambda captures. | |||
8699 | if (Types[I] != (OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | |||
8700 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT)) | |||
8701 | continue; | |||
8702 | llvm::Value *BasePtr = LambdaPointers.lookup(*BasePointers[I]); | |||
8703 | assert(BasePtr && "Unable to find base lambda address.")(static_cast <bool> (BasePtr && "Unable to find base lambda address." ) ? void (0) : __assert_fail ("BasePtr && \"Unable to find base lambda address.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8703, __extension__ __PRETTY_FUNCTION__)); | |||
8704 | int TgtIdx = -1; | |||
8705 | for (unsigned J = I; J > 0; --J) { | |||
8706 | unsigned Idx = J - 1; | |||
8707 | if (Pointers[Idx] != BasePtr) | |||
8708 | continue; | |||
8709 | TgtIdx = Idx; | |||
8710 | break; | |||
8711 | } | |||
8712 | assert(TgtIdx != -1 && "Unable to find parent lambda.")(static_cast <bool> (TgtIdx != -1 && "Unable to find parent lambda." ) ? void (0) : __assert_fail ("TgtIdx != -1 && \"Unable to find parent lambda.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8712, __extension__ __PRETTY_FUNCTION__)); | |||
8713 | // All other current entries will be MEMBER_OF the combined entry | |||
8714 | // (except for PTR_AND_OBJ entries which do not have a placeholder value | |||
8715 | // 0xFFFF in the MEMBER_OF field). | |||
8716 | OpenMPOffloadMappingFlags MemberOfFlag = getMemberOfFlag(TgtIdx); | |||
8717 | setCorrectMemberOfFlag(Types[I], MemberOfFlag); | |||
8718 | } | |||
8719 | } | |||
8720 | ||||
8721 | /// Generate the base pointers, section pointers, sizes, map types, and | |||
8722 | /// mappers associated to a given capture (all included in \a CombinedInfo). | |||
8723 | void generateInfoForCapture(const CapturedStmt::Capture *Cap, | |||
8724 | llvm::Value *Arg, MapCombinedInfoTy &CombinedInfo, | |||
8725 | StructRangeInfoTy &PartialStruct) const { | |||
8726 | assert(!Cap->capturesVariableArrayType() &&(static_cast <bool> (!Cap->capturesVariableArrayType () && "Not expecting to generate map info for a variable array type!" ) ? void (0) : __assert_fail ("!Cap->capturesVariableArrayType() && \"Not expecting to generate map info for a variable array type!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8727, __extension__ __PRETTY_FUNCTION__)) | |||
8727 | "Not expecting to generate map info for a variable array type!")(static_cast <bool> (!Cap->capturesVariableArrayType () && "Not expecting to generate map info for a variable array type!" ) ? void (0) : __assert_fail ("!Cap->capturesVariableArrayType() && \"Not expecting to generate map info for a variable array type!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8727, __extension__ __PRETTY_FUNCTION__)); | |||
8728 | ||||
8729 | // We need to know when we generating information for the first component | |||
8730 | const ValueDecl *VD = Cap->capturesThis() | |||
8731 | ? nullptr | |||
8732 | : Cap->getCapturedVar()->getCanonicalDecl(); | |||
8733 | ||||
8734 | // for map(to: lambda): skip here, processing it in | |||
8735 | // generateDefaultMapInfo | |||
8736 | if (LambdasMap.count(VD)) | |||
8737 | return; | |||
8738 | ||||
8739 | // If this declaration appears in a is_device_ptr clause we just have to | |||
8740 | // pass the pointer by value. If it is a reference to a declaration, we just | |||
8741 | // pass its value. | |||
8742 | if (VD && (DevPointersMap.count(VD) || HasDevAddrsMap.count(VD))) { | |||
8743 | CombinedInfo.Exprs.push_back(VD); | |||
8744 | CombinedInfo.BasePointers.emplace_back(Arg, VD); | |||
8745 | CombinedInfo.Pointers.push_back(Arg); | |||
8746 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | |||
8747 | CGF.getTypeSize(CGF.getContext().VoidPtrTy), CGF.Int64Ty, | |||
8748 | /*isSigned=*/true)); | |||
8749 | CombinedInfo.Types.push_back( | |||
8750 | (Cap->capturesVariable() ? OMP_MAP_TO : OMP_MAP_LITERAL) | | |||
8751 | OMP_MAP_TARGET_PARAM); | |||
8752 | CombinedInfo.Mappers.push_back(nullptr); | |||
8753 | return; | |||
8754 | } | |||
8755 | ||||
8756 | using MapData = | |||
8757 | std::tuple<OMPClauseMappableExprCommon::MappableExprComponentListRef, | |||
8758 | OpenMPMapClauseKind, ArrayRef<OpenMPMapModifierKind>, bool, | |||
8759 | const ValueDecl *, const Expr *>; | |||
8760 | SmallVector<MapData, 4> DeclComponentLists; | |||
8761 | // For member fields list in is_device_ptr, store it in | |||
8762 | // DeclComponentLists for generating components info. | |||
8763 | static const OpenMPMapModifierKind Unknown = OMPC_MAP_MODIFIER_unknown; | |||
8764 | auto It = DevPointersMap.find(VD); | |||
8765 | if (It != DevPointersMap.end()) | |||
8766 | for (const auto &MCL : It->second) | |||
8767 | DeclComponentLists.emplace_back(MCL, OMPC_MAP_to, Unknown, | |||
8768 | /*IsImpicit = */ true, nullptr, | |||
8769 | nullptr); | |||
8770 | auto I = HasDevAddrsMap.find(VD); | |||
8771 | if (I != HasDevAddrsMap.end()) | |||
8772 | for (const auto &MCL : I->second) | |||
8773 | DeclComponentLists.emplace_back(MCL, OMPC_MAP_tofrom, Unknown, | |||
8774 | /*IsImpicit = */ true, nullptr, | |||
8775 | nullptr); | |||
8776 | assert(CurDir.is<const OMPExecutableDirective *>() &&(static_cast <bool> (CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? void ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8777, __extension__ __PRETTY_FUNCTION__)) | |||
8777 | "Expect a executable directive")(static_cast <bool> (CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? void ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8777, __extension__ __PRETTY_FUNCTION__)); | |||
8778 | const auto *CurExecDir = CurDir.get<const OMPExecutableDirective *>(); | |||
8779 | for (const auto *C : CurExecDir->getClausesOfKind<OMPMapClause>()) { | |||
8780 | const auto *EI = C->getVarRefs().begin(); | |||
8781 | for (const auto L : C->decl_component_lists(VD)) { | |||
8782 | const ValueDecl *VDecl, *Mapper; | |||
8783 | // The Expression is not correct if the mapping is implicit | |||
8784 | const Expr *E = (C->getMapLoc().isValid()) ? *EI : nullptr; | |||
8785 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8786 | std::tie(VDecl, Components, Mapper) = L; | |||
8787 | assert(VDecl == VD && "We got information for the wrong declaration??")(static_cast <bool> (VDecl == VD && "We got information for the wrong declaration??" ) ? void (0) : __assert_fail ("VDecl == VD && \"We got information for the wrong declaration??\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8787, __extension__ __PRETTY_FUNCTION__)); | |||
8788 | assert(!Components.empty() &&(static_cast <bool> (!Components.empty() && "Not expecting declaration with no component lists." ) ? void (0) : __assert_fail ("!Components.empty() && \"Not expecting declaration with no component lists.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8789, __extension__ __PRETTY_FUNCTION__)) | |||
8789 | "Not expecting declaration with no component lists.")(static_cast <bool> (!Components.empty() && "Not expecting declaration with no component lists." ) ? void (0) : __assert_fail ("!Components.empty() && \"Not expecting declaration with no component lists.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 8789, __extension__ __PRETTY_FUNCTION__)); | |||
8790 | DeclComponentLists.emplace_back(Components, C->getMapType(), | |||
8791 | C->getMapTypeModifiers(), | |||
8792 | C->isImplicit(), Mapper, E); | |||
8793 | ++EI; | |||
8794 | } | |||
8795 | } | |||
8796 | llvm::stable_sort(DeclComponentLists, [](const MapData &LHS, | |||
8797 | const MapData &RHS) { | |||
8798 | ArrayRef<OpenMPMapModifierKind> MapModifiers = std::get<2>(LHS); | |||
8799 | OpenMPMapClauseKind MapType = std::get<1>(RHS); | |||
8800 | bool HasPresent = | |||
8801 | llvm::is_contained(MapModifiers, clang::OMPC_MAP_MODIFIER_present); | |||
8802 | bool HasAllocs = MapType == OMPC_MAP_alloc; | |||
8803 | MapModifiers = std::get<2>(RHS); | |||
8804 | MapType = std::get<1>(LHS); | |||
8805 | bool HasPresentR = | |||
8806 | llvm::is_contained(MapModifiers, clang::OMPC_MAP_MODIFIER_present); | |||
8807 | bool HasAllocsR = MapType == OMPC_MAP_alloc; | |||
8808 | return (HasPresent && !HasPresentR) || (HasAllocs && !HasAllocsR); | |||
8809 | }); | |||
8810 | ||||
8811 | // Find overlapping elements (including the offset from the base element). | |||
8812 | llvm::SmallDenseMap< | |||
8813 | const MapData *, | |||
8814 | llvm::SmallVector< | |||
8815 | OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>, | |||
8816 | 4> | |||
8817 | OverlappedData; | |||
8818 | size_t Count = 0; | |||
8819 | for (const MapData &L : DeclComponentLists) { | |||
8820 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8821 | OpenMPMapClauseKind MapType; | |||
8822 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
8823 | bool IsImplicit; | |||
8824 | const ValueDecl *Mapper; | |||
8825 | const Expr *VarRef; | |||
8826 | std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper, VarRef) = | |||
8827 | L; | |||
8828 | ++Count; | |||
8829 | for (const MapData &L1 : makeArrayRef(DeclComponentLists).slice(Count)) { | |||
8830 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components1; | |||
8831 | std::tie(Components1, MapType, MapModifiers, IsImplicit, Mapper, | |||
8832 | VarRef) = L1; | |||
8833 | auto CI = Components.rbegin(); | |||
8834 | auto CE = Components.rend(); | |||
8835 | auto SI = Components1.rbegin(); | |||
8836 | auto SE = Components1.rend(); | |||
8837 | for (; CI != CE && SI != SE; ++CI, ++SI) { | |||
8838 | if (CI->getAssociatedExpression()->getStmtClass() != | |||
8839 | SI->getAssociatedExpression()->getStmtClass()) | |||
8840 | break; | |||
8841 | // Are we dealing with different variables/fields? | |||
8842 | if (CI->getAssociatedDeclaration() != SI->getAssociatedDeclaration()) | |||
8843 | break; | |||
8844 | } | |||
8845 | // Found overlapping if, at least for one component, reached the head | |||
8846 | // of the components list. | |||
8847 | if (CI == CE || SI == SE) { | |||
8848 | // Ignore it if it is the same component. | |||
8849 | if (CI == CE && SI == SE) | |||
8850 | continue; | |||
8851 | const auto It = (SI == SE) ? CI : SI; | |||
8852 | // If one component is a pointer and another one is a kind of | |||
8853 | // dereference of this pointer (array subscript, section, dereference, | |||
8854 | // etc.), it is not an overlapping. | |||
8855 | // Same, if one component is a base and another component is a | |||
8856 | // dereferenced pointer memberexpr with the same base. | |||
8857 | if (!isa<MemberExpr>(It->getAssociatedExpression()) || | |||
8858 | (std::prev(It)->getAssociatedDeclaration() && | |||
8859 | std::prev(It) | |||
8860 | ->getAssociatedDeclaration() | |||
8861 | ->getType() | |||
8862 | ->isPointerType()) || | |||
8863 | (It->getAssociatedDeclaration() && | |||
8864 | It->getAssociatedDeclaration()->getType()->isPointerType() && | |||
8865 | std::next(It) != CE && std::next(It) != SE)) | |||
8866 | continue; | |||
8867 | const MapData &BaseData = CI == CE ? L : L1; | |||
8868 | OMPClauseMappableExprCommon::MappableExprComponentListRef SubData = | |||
8869 | SI == SE ? Components : Components1; | |||
8870 | auto &OverlappedElements = OverlappedData.FindAndConstruct(&BaseData); | |||
8871 | OverlappedElements.getSecond().push_back(SubData); | |||
8872 | } | |||
8873 | } | |||
8874 | } | |||
8875 | // Sort the overlapped elements for each item. | |||
8876 | llvm::SmallVector<const FieldDecl *, 4> Layout; | |||
8877 | if (!OverlappedData.empty()) { | |||
8878 | const Type *BaseType = VD->getType().getCanonicalType().getTypePtr(); | |||
8879 | const Type *OrigType = BaseType->getPointeeOrArrayElementType(); | |||
8880 | while (BaseType != OrigType) { | |||
8881 | BaseType = OrigType->getCanonicalTypeInternal().getTypePtr(); | |||
8882 | OrigType = BaseType->getPointeeOrArrayElementType(); | |||
8883 | } | |||
8884 | ||||
8885 | if (const auto *CRD = BaseType->getAsCXXRecordDecl()) | |||
8886 | getPlainLayout(CRD, Layout, /*AsBase=*/false); | |||
8887 | else { | |||
8888 | const auto *RD = BaseType->getAsRecordDecl(); | |||
8889 | Layout.append(RD->field_begin(), RD->field_end()); | |||
8890 | } | |||
8891 | } | |||
8892 | for (auto &Pair : OverlappedData) { | |||
8893 | llvm::stable_sort( | |||
8894 | Pair.getSecond(), | |||
8895 | [&Layout]( | |||
8896 | OMPClauseMappableExprCommon::MappableExprComponentListRef First, | |||
8897 | OMPClauseMappableExprCommon::MappableExprComponentListRef | |||
8898 | Second) { | |||
8899 | auto CI = First.rbegin(); | |||
8900 | auto CE = First.rend(); | |||
8901 | auto SI = Second.rbegin(); | |||
8902 | auto SE = Second.rend(); | |||
8903 | for (; CI != CE && SI != SE; ++CI, ++SI) { | |||
8904 | if (CI->getAssociatedExpression()->getStmtClass() != | |||
8905 | SI->getAssociatedExpression()->getStmtClass()) | |||
8906 | break; | |||
8907 | // Are we dealing with different variables/fields? | |||
8908 | if (CI->getAssociatedDeclaration() != | |||
8909 | SI->getAssociatedDeclaration()) | |||
8910 | break; | |||
8911 | } | |||
8912 | ||||
8913 | // Lists contain the same elements. | |||
8914 | if (CI == CE && SI == SE) | |||
8915 | return false; | |||
8916 | ||||
8917 | // List with less elements is less than list with more elements. | |||
8918 | if (CI == CE || SI == SE) | |||
8919 | return CI == CE; | |||
8920 | ||||
8921 | const auto *FD1 = cast<FieldDecl>(CI->getAssociatedDeclaration()); | |||
8922 | const auto *FD2 = cast<FieldDecl>(SI->getAssociatedDeclaration()); | |||
8923 | if (FD1->getParent() == FD2->getParent()) | |||
8924 | return FD1->getFieldIndex() < FD2->getFieldIndex(); | |||
8925 | const auto *It = | |||
8926 | llvm::find_if(Layout, [FD1, FD2](const FieldDecl *FD) { | |||
8927 | return FD == FD1 || FD == FD2; | |||
8928 | }); | |||
8929 | return *It == FD1; | |||
8930 | }); | |||
8931 | } | |||
8932 | ||||
8933 | // Associated with a capture, because the mapping flags depend on it. | |||
8934 | // Go through all of the elements with the overlapped elements. | |||
8935 | bool IsFirstComponentList = true; | |||
8936 | for (const auto &Pair : OverlappedData) { | |||
8937 | const MapData &L = *Pair.getFirst(); | |||
8938 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8939 | OpenMPMapClauseKind MapType; | |||
8940 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
8941 | bool IsImplicit; | |||
8942 | const ValueDecl *Mapper; | |||
8943 | const Expr *VarRef; | |||
8944 | std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper, VarRef) = | |||
8945 | L; | |||
8946 | ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> | |||
8947 | OverlappedComponents = Pair.getSecond(); | |||
8948 | generateInfoForComponentList( | |||
8949 | MapType, MapModifiers, llvm::None, Components, CombinedInfo, | |||
8950 | PartialStruct, IsFirstComponentList, IsImplicit, Mapper, | |||
8951 | /*ForDeviceAddr=*/false, VD, VarRef, OverlappedComponents); | |||
8952 | IsFirstComponentList = false; | |||
8953 | } | |||
8954 | // Go through other elements without overlapped elements. | |||
8955 | for (const MapData &L : DeclComponentLists) { | |||
8956 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8957 | OpenMPMapClauseKind MapType; | |||
8958 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
8959 | bool IsImplicit; | |||
8960 | const ValueDecl *Mapper; | |||
8961 | const Expr *VarRef; | |||
8962 | std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper, VarRef) = | |||
8963 | L; | |||
8964 | auto It = OverlappedData.find(&L); | |||
8965 | if (It == OverlappedData.end()) | |||
8966 | generateInfoForComponentList(MapType, MapModifiers, llvm::None, | |||
8967 | Components, CombinedInfo, PartialStruct, | |||
8968 | IsFirstComponentList, IsImplicit, Mapper, | |||
8969 | /*ForDeviceAddr=*/false, VD, VarRef); | |||
8970 | IsFirstComponentList = false; | |||
8971 | } | |||
8972 | } | |||
8973 | ||||
8974 | /// Generate the default map information for a given capture \a CI, | |||
8975 | /// record field declaration \a RI and captured value \a CV. | |||
8976 | void generateDefaultMapInfo(const CapturedStmt::Capture &CI, | |||
8977 | const FieldDecl &RI, llvm::Value *CV, | |||
8978 | MapCombinedInfoTy &CombinedInfo) const { | |||
8979 | bool IsImplicit = true; | |||
8980 | // Do the default mapping. | |||
8981 | if (CI.capturesThis()) { | |||
8982 | CombinedInfo.Exprs.push_back(nullptr); | |||
8983 | CombinedInfo.BasePointers.push_back(CV); | |||
8984 | CombinedInfo.Pointers.push_back(CV); | |||
8985 | const auto *PtrTy = cast<PointerType>(RI.getType().getTypePtr()); | |||
8986 | CombinedInfo.Sizes.push_back( | |||
8987 | CGF.Builder.CreateIntCast(CGF.getTypeSize(PtrTy->getPointeeType()), | |||
8988 | CGF.Int64Ty, /*isSigned=*/true)); | |||
8989 | // Default map type. | |||
8990 | CombinedInfo.Types.push_back(OMP_MAP_TO | OMP_MAP_FROM); | |||
8991 | } else if (CI.capturesVariableByCopy()) { | |||
8992 | const VarDecl *VD = CI.getCapturedVar(); | |||
8993 | CombinedInfo.Exprs.push_back(VD->getCanonicalDecl()); | |||
8994 | CombinedInfo.BasePointers.push_back(CV); | |||
8995 | CombinedInfo.Pointers.push_back(CV); | |||
8996 | if (!RI.getType()->isAnyPointerType()) { | |||
8997 | // We have to signal to the runtime captures passed by value that are | |||
8998 | // not pointers. | |||
8999 | CombinedInfo.Types.push_back(OMP_MAP_LITERAL); | |||
9000 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | |||
9001 | CGF.getTypeSize(RI.getType()), CGF.Int64Ty, /*isSigned=*/true)); | |||
9002 | } else { | |||
9003 | // Pointers are implicitly mapped with a zero size and no flags | |||
9004 | // (other than first map that is added for all implicit maps). | |||
9005 | CombinedInfo.Types.push_back(OMP_MAP_NONE); | |||
9006 | CombinedInfo.Sizes.push_back(llvm::Constant::getNullValue(CGF.Int64Ty)); | |||
9007 | } | |||
9008 | auto I = FirstPrivateDecls.find(VD); | |||
9009 | if (I != FirstPrivateDecls.end()) | |||
9010 | IsImplicit = I->getSecond(); | |||
9011 | } else { | |||
9012 | assert(CI.capturesVariable() && "Expected captured reference.")(static_cast <bool> (CI.capturesVariable() && "Expected captured reference." ) ? void (0) : __assert_fail ("CI.capturesVariable() && \"Expected captured reference.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9012, __extension__ __PRETTY_FUNCTION__)); | |||
9013 | const auto *PtrTy = cast<ReferenceType>(RI.getType().getTypePtr()); | |||
9014 | QualType ElementType = PtrTy->getPointeeType(); | |||
9015 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | |||
9016 | CGF.getTypeSize(ElementType), CGF.Int64Ty, /*isSigned=*/true)); | |||
9017 | // The default map type for a scalar/complex type is 'to' because by | |||
9018 | // default the value doesn't have to be retrieved. For an aggregate | |||
9019 | // type, the default is 'tofrom'. | |||
9020 | CombinedInfo.Types.push_back(getMapModifiersForPrivateClauses(CI)); | |||
9021 | const VarDecl *VD = CI.getCapturedVar(); | |||
9022 | auto I = FirstPrivateDecls.find(VD); | |||
9023 | CombinedInfo.Exprs.push_back(VD->getCanonicalDecl()); | |||
9024 | CombinedInfo.BasePointers.push_back(CV); | |||
9025 | if (I != FirstPrivateDecls.end() && ElementType->isAnyPointerType()) { | |||
9026 | Address PtrAddr = CGF.EmitLoadOfReference(CGF.MakeAddrLValue( | |||
9027 | CV, ElementType, CGF.getContext().getDeclAlign(VD), | |||
9028 | AlignmentSource::Decl)); | |||
9029 | CombinedInfo.Pointers.push_back(PtrAddr.getPointer()); | |||
9030 | } else { | |||
9031 | CombinedInfo.Pointers.push_back(CV); | |||
9032 | } | |||
9033 | if (I != FirstPrivateDecls.end()) | |||
9034 | IsImplicit = I->getSecond(); | |||
9035 | } | |||
9036 | // Every default map produces a single argument which is a target parameter. | |||
9037 | CombinedInfo.Types.back() |= OMP_MAP_TARGET_PARAM; | |||
9038 | ||||
9039 | // Add flag stating this is an implicit map. | |||
9040 | if (IsImplicit) | |||
9041 | CombinedInfo.Types.back() |= OMP_MAP_IMPLICIT; | |||
9042 | ||||
9043 | // No user-defined mapper for default mapping. | |||
9044 | CombinedInfo.Mappers.push_back(nullptr); | |||
9045 | } | |||
9046 | }; | |||
9047 | } // anonymous namespace | |||
9048 | ||||
9049 | static void emitNonContiguousDescriptor( | |||
9050 | CodeGenFunction &CGF, MappableExprsHandler::MapCombinedInfoTy &CombinedInfo, | |||
9051 | CGOpenMPRuntime::TargetDataInfo &Info) { | |||
9052 | CodeGenModule &CGM = CGF.CGM; | |||
9053 | MappableExprsHandler::MapCombinedInfoTy::StructNonContiguousInfo | |||
9054 | &NonContigInfo = CombinedInfo.NonContigInfo; | |||
9055 | ||||
9056 | // Build an array of struct descriptor_dim and then assign it to | |||
9057 | // offload_args. | |||
9058 | // | |||
9059 | // struct descriptor_dim { | |||
9060 | // uint64_t offset; | |||
9061 | // uint64_t count; | |||
9062 | // uint64_t stride | |||
9063 | // }; | |||
9064 | ASTContext &C = CGF.getContext(); | |||
9065 | QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0); | |||
9066 | RecordDecl *RD; | |||
9067 | RD = C.buildImplicitRecord("descriptor_dim"); | |||
9068 | RD->startDefinition(); | |||
9069 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
9070 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
9071 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
9072 | RD->completeDefinition(); | |||
9073 | QualType DimTy = C.getRecordType(RD); | |||
9074 | ||||
9075 | enum { OffsetFD = 0, CountFD, StrideFD }; | |||
9076 | // We need two index variable here since the size of "Dims" is the same as the | |||
9077 | // size of Components, however, the size of offset, count, and stride is equal | |||
9078 | // to the size of base declaration that is non-contiguous. | |||
9079 | for (unsigned I = 0, L = 0, E = NonContigInfo.Dims.size(); I < E; ++I) { | |||
9080 | // Skip emitting ir if dimension size is 1 since it cannot be | |||
9081 | // non-contiguous. | |||
9082 | if (NonContigInfo.Dims[I] == 1) | |||
9083 | continue; | |||
9084 | llvm::APInt Size(/*numBits=*/32, NonContigInfo.Dims[I]); | |||
9085 | QualType ArrayTy = | |||
9086 | C.getConstantArrayType(DimTy, Size, nullptr, ArrayType::Normal, 0); | |||
9087 | Address DimsAddr = CGF.CreateMemTemp(ArrayTy, "dims"); | |||
9088 | for (unsigned II = 0, EE = NonContigInfo.Dims[I]; II < EE; ++II) { | |||
9089 | unsigned RevIdx = EE - II - 1; | |||
9090 | LValue DimsLVal = CGF.MakeAddrLValue( | |||
9091 | CGF.Builder.CreateConstArrayGEP(DimsAddr, II), DimTy); | |||
9092 | // Offset | |||
9093 | LValue OffsetLVal = CGF.EmitLValueForField( | |||
9094 | DimsLVal, *std::next(RD->field_begin(), OffsetFD)); | |||
9095 | CGF.EmitStoreOfScalar(NonContigInfo.Offsets[L][RevIdx], OffsetLVal); | |||
9096 | // Count | |||
9097 | LValue CountLVal = CGF.EmitLValueForField( | |||
9098 | DimsLVal, *std::next(RD->field_begin(), CountFD)); | |||
9099 | CGF.EmitStoreOfScalar(NonContigInfo.Counts[L][RevIdx], CountLVal); | |||
9100 | // Stride | |||
9101 | LValue StrideLVal = CGF.EmitLValueForField( | |||
9102 | DimsLVal, *std::next(RD->field_begin(), StrideFD)); | |||
9103 | CGF.EmitStoreOfScalar(NonContigInfo.Strides[L][RevIdx], StrideLVal); | |||
9104 | } | |||
9105 | // args[I] = &dims | |||
9106 | Address DAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
9107 | DimsAddr, CGM.Int8PtrTy, CGM.Int8Ty); | |||
9108 | llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
9109 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
9110 | Info.RTArgs.PointersArray, 0, I); | |||
9111 | Address PAddr(P, CGM.VoidPtrTy, CGF.getPointerAlign()); | |||
9112 | CGF.Builder.CreateStore(DAddr.getPointer(), PAddr); | |||
9113 | ++L; | |||
9114 | } | |||
9115 | } | |||
9116 | ||||
9117 | // Try to extract the base declaration from a `this->x` expression if possible. | |||
9118 | static ValueDecl *getDeclFromThisExpr(const Expr *E) { | |||
9119 | if (!E) | |||
9120 | return nullptr; | |||
9121 | ||||
9122 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(E->IgnoreParenCasts())) | |||
9123 | if (const MemberExpr *ME = | |||
9124 | dyn_cast<MemberExpr>(OASE->getBase()->IgnoreParenImpCasts())) | |||
9125 | return ME->getMemberDecl(); | |||
9126 | return nullptr; | |||
9127 | } | |||
9128 | ||||
9129 | /// Emit a string constant containing the names of the values mapped to the | |||
9130 | /// offloading runtime library. | |||
9131 | llvm::Constant * | |||
9132 | emitMappingInformation(CodeGenFunction &CGF, llvm::OpenMPIRBuilder &OMPBuilder, | |||
9133 | MappableExprsHandler::MappingExprInfo &MapExprs) { | |||
9134 | ||||
9135 | uint32_t SrcLocStrSize; | |||
9136 | if (!MapExprs.getMapDecl() && !MapExprs.getMapExpr()) | |||
9137 | return OMPBuilder.getOrCreateDefaultSrcLocStr(SrcLocStrSize); | |||
9138 | ||||
9139 | SourceLocation Loc; | |||
9140 | if (!MapExprs.getMapDecl() && MapExprs.getMapExpr()) { | |||
9141 | if (const ValueDecl *VD = getDeclFromThisExpr(MapExprs.getMapExpr())) | |||
9142 | Loc = VD->getLocation(); | |||
9143 | else | |||
9144 | Loc = MapExprs.getMapExpr()->getExprLoc(); | |||
9145 | } else { | |||
9146 | Loc = MapExprs.getMapDecl()->getLocation(); | |||
9147 | } | |||
9148 | ||||
9149 | std::string ExprName; | |||
9150 | if (MapExprs.getMapExpr()) { | |||
9151 | PrintingPolicy P(CGF.getContext().getLangOpts()); | |||
9152 | llvm::raw_string_ostream OS(ExprName); | |||
9153 | MapExprs.getMapExpr()->printPretty(OS, nullptr, P); | |||
9154 | OS.flush(); | |||
9155 | } else { | |||
9156 | ExprName = MapExprs.getMapDecl()->getNameAsString(); | |||
9157 | } | |||
9158 | ||||
9159 | PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc); | |||
9160 | return OMPBuilder.getOrCreateSrcLocStr(PLoc.getFilename(), ExprName, | |||
9161 | PLoc.getLine(), PLoc.getColumn(), | |||
9162 | SrcLocStrSize); | |||
9163 | } | |||
9164 | ||||
9165 | /// Emit the arrays used to pass the captures and map information to the | |||
9166 | /// offloading runtime library. If there is no map or capture information, | |||
9167 | /// return nullptr by reference. | |||
9168 | static void emitOffloadingArrays( | |||
9169 | CodeGenFunction &CGF, MappableExprsHandler::MapCombinedInfoTy &CombinedInfo, | |||
9170 | CGOpenMPRuntime::TargetDataInfo &Info, llvm::OpenMPIRBuilder &OMPBuilder, | |||
9171 | bool IsNonContiguous = false) { | |||
9172 | CodeGenModule &CGM = CGF.CGM; | |||
9173 | ASTContext &Ctx = CGF.getContext(); | |||
9174 | ||||
9175 | // Reset the array information. | |||
9176 | Info.clearArrayInfo(); | |||
9177 | Info.NumberOfPtrs = CombinedInfo.BasePointers.size(); | |||
9178 | ||||
9179 | if (Info.NumberOfPtrs) { | |||
9180 | // Detect if we have any capture size requiring runtime evaluation of the | |||
9181 | // size so that a constant array could be eventually used. | |||
9182 | ||||
9183 | llvm::APInt PointerNumAP(32, Info.NumberOfPtrs, /*isSigned=*/true); | |||
9184 | QualType PointerArrayType = Ctx.getConstantArrayType( | |||
9185 | Ctx.VoidPtrTy, PointerNumAP, nullptr, ArrayType::Normal, | |||
9186 | /*IndexTypeQuals=*/0); | |||
9187 | ||||
9188 | Info.RTArgs.BasePointersArray = | |||
9189 | CGF.CreateMemTemp(PointerArrayType, ".offload_baseptrs").getPointer(); | |||
9190 | Info.RTArgs.PointersArray = | |||
9191 | CGF.CreateMemTemp(PointerArrayType, ".offload_ptrs").getPointer(); | |||
9192 | Address MappersArray = | |||
9193 | CGF.CreateMemTemp(PointerArrayType, ".offload_mappers"); | |||
9194 | Info.RTArgs.MappersArray = MappersArray.getPointer(); | |||
9195 | ||||
9196 | // If we don't have any VLA types or other types that require runtime | |||
9197 | // evaluation, we can use a constant array for the map sizes, otherwise we | |||
9198 | // need to fill up the arrays as we do for the pointers. | |||
9199 | QualType Int64Ty = | |||
9200 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | |||
9201 | SmallVector<llvm::Constant *> ConstSizes( | |||
9202 | CombinedInfo.Sizes.size(), llvm::ConstantInt::get(CGF.Int64Ty, 0)); | |||
9203 | llvm::SmallBitVector RuntimeSizes(CombinedInfo.Sizes.size()); | |||
9204 | for (unsigned I = 0, E = CombinedInfo.Sizes.size(); I < E; ++I) { | |||
9205 | if (auto *CI = dyn_cast<llvm::Constant>(CombinedInfo.Sizes[I])) { | |||
9206 | if (!isa<llvm::ConstantExpr>(CI) && !isa<llvm::GlobalValue>(CI)) { | |||
9207 | if (IsNonContiguous && (CombinedInfo.Types[I] & | |||
9208 | MappableExprsHandler::OMP_MAP_NON_CONTIG)) | |||
9209 | ConstSizes[I] = llvm::ConstantInt::get( | |||
9210 | CGF.Int64Ty, CombinedInfo.NonContigInfo.Dims[I]); | |||
9211 | else | |||
9212 | ConstSizes[I] = CI; | |||
9213 | continue; | |||
9214 | } | |||
9215 | } | |||
9216 | RuntimeSizes.set(I); | |||
9217 | } | |||
9218 | ||||
9219 | if (RuntimeSizes.all()) { | |||
9220 | QualType SizeArrayType = Ctx.getConstantArrayType( | |||
9221 | Int64Ty, PointerNumAP, nullptr, ArrayType::Normal, | |||
9222 | /*IndexTypeQuals=*/0); | |||
9223 | Info.RTArgs.SizesArray = | |||
9224 | CGF.CreateMemTemp(SizeArrayType, ".offload_sizes").getPointer(); | |||
9225 | } else { | |||
9226 | auto *SizesArrayInit = llvm::ConstantArray::get( | |||
9227 | llvm::ArrayType::get(CGM.Int64Ty, ConstSizes.size()), ConstSizes); | |||
9228 | std::string Name = CGM.getOpenMPRuntime().getName({"offload_sizes"}); | |||
9229 | auto *SizesArrayGbl = new llvm::GlobalVariable( | |||
9230 | CGM.getModule(), SizesArrayInit->getType(), /*isConstant=*/true, | |||
9231 | llvm::GlobalValue::PrivateLinkage, SizesArrayInit, Name); | |||
9232 | SizesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
9233 | if (RuntimeSizes.any()) { | |||
9234 | QualType SizeArrayType = Ctx.getConstantArrayType( | |||
9235 | Int64Ty, PointerNumAP, nullptr, ArrayType::Normal, | |||
9236 | /*IndexTypeQuals=*/0); | |||
9237 | Address Buffer = CGF.CreateMemTemp(SizeArrayType, ".offload_sizes"); | |||
9238 | llvm::Value *GblConstPtr = | |||
9239 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
9240 | SizesArrayGbl, CGM.Int64Ty->getPointerTo()); | |||
9241 | CGF.Builder.CreateMemCpy( | |||
9242 | Buffer, | |||
9243 | Address(GblConstPtr, CGM.Int64Ty, | |||
9244 | CGM.getNaturalTypeAlignment(Ctx.getIntTypeForBitwidth( | |||
9245 | /*DestWidth=*/64, /*Signed=*/false))), | |||
9246 | CGF.getTypeSize(SizeArrayType)); | |||
9247 | Info.RTArgs.SizesArray = Buffer.getPointer(); | |||
9248 | } else { | |||
9249 | Info.RTArgs.SizesArray = SizesArrayGbl; | |||
9250 | } | |||
9251 | } | |||
9252 | ||||
9253 | // The map types are always constant so we don't need to generate code to | |||
9254 | // fill arrays. Instead, we create an array constant. | |||
9255 | SmallVector<uint64_t, 4> Mapping(CombinedInfo.Types.size(), 0); | |||
9256 | llvm::copy(CombinedInfo.Types, Mapping.begin()); | |||
9257 | std::string MaptypesName = | |||
9258 | CGM.getOpenMPRuntime().getName({"offload_maptypes"}); | |||
9259 | auto *MapTypesArrayGbl = | |||
9260 | OMPBuilder.createOffloadMaptypes(Mapping, MaptypesName); | |||
9261 | Info.RTArgs.MapTypesArray = MapTypesArrayGbl; | |||
9262 | ||||
9263 | // The information types are only built if there is debug information | |||
9264 | // requested. | |||
9265 | if (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo) { | |||
9266 | Info.RTArgs.MapNamesArray = llvm::Constant::getNullValue( | |||
9267 | llvm::Type::getInt8Ty(CGF.Builder.getContext())->getPointerTo()); | |||
9268 | } else { | |||
9269 | auto fillInfoMap = [&](MappableExprsHandler::MappingExprInfo &MapExpr) { | |||
9270 | return emitMappingInformation(CGF, OMPBuilder, MapExpr); | |||
9271 | }; | |||
9272 | SmallVector<llvm::Constant *, 4> InfoMap(CombinedInfo.Exprs.size()); | |||
9273 | llvm::transform(CombinedInfo.Exprs, InfoMap.begin(), fillInfoMap); | |||
9274 | std::string MapnamesName = | |||
9275 | CGM.getOpenMPRuntime().getName({"offload_mapnames"}); | |||
9276 | auto *MapNamesArrayGbl = | |||
9277 | OMPBuilder.createOffloadMapnames(InfoMap, MapnamesName); | |||
9278 | Info.RTArgs.MapNamesArray = MapNamesArrayGbl; | |||
9279 | } | |||
9280 | ||||
9281 | // If there's a present map type modifier, it must not be applied to the end | |||
9282 | // of a region, so generate a separate map type array in that case. | |||
9283 | if (Info.separateBeginEndCalls()) { | |||
9284 | bool EndMapTypesDiffer = false; | |||
9285 | for (uint64_t &Type : Mapping) { | |||
9286 | if (Type & MappableExprsHandler::OMP_MAP_PRESENT) { | |||
9287 | Type &= ~MappableExprsHandler::OMP_MAP_PRESENT; | |||
9288 | EndMapTypesDiffer = true; | |||
9289 | } | |||
9290 | } | |||
9291 | if (EndMapTypesDiffer) { | |||
9292 | MapTypesArrayGbl = | |||
9293 | OMPBuilder.createOffloadMaptypes(Mapping, MaptypesName); | |||
9294 | Info.RTArgs.MapTypesArrayEnd = MapTypesArrayGbl; | |||
9295 | } | |||
9296 | } | |||
9297 | ||||
9298 | for (unsigned I = 0; I < Info.NumberOfPtrs; ++I) { | |||
9299 | llvm::Value *BPVal = *CombinedInfo.BasePointers[I]; | |||
9300 | llvm::Value *BP = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
9301 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
9302 | Info.RTArgs.BasePointersArray, 0, I); | |||
9303 | BP = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
9304 | BP, BPVal->getType()->getPointerTo(/*AddrSpace=*/0)); | |||
9305 | Address BPAddr(BP, BPVal->getType(), | |||
9306 | Ctx.getTypeAlignInChars(Ctx.VoidPtrTy)); | |||
9307 | CGF.Builder.CreateStore(BPVal, BPAddr); | |||
9308 | ||||
9309 | if (Info.requiresDevicePointerInfo()) | |||
9310 | if (const ValueDecl *DevVD = | |||
9311 | CombinedInfo.BasePointers[I].getDevicePtrDecl()) | |||
9312 | Info.CaptureDeviceAddrMap.try_emplace(DevVD, BPAddr); | |||
9313 | ||||
9314 | llvm::Value *PVal = CombinedInfo.Pointers[I]; | |||
9315 | llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
9316 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
9317 | Info.RTArgs.PointersArray, 0, I); | |||
9318 | P = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
9319 | P, PVal->getType()->getPointerTo(/*AddrSpace=*/0)); | |||
9320 | Address PAddr(P, PVal->getType(), Ctx.getTypeAlignInChars(Ctx.VoidPtrTy)); | |||
9321 | CGF.Builder.CreateStore(PVal, PAddr); | |||
9322 | ||||
9323 | if (RuntimeSizes.test(I)) { | |||
9324 | llvm::Value *S = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
9325 | llvm::ArrayType::get(CGM.Int64Ty, Info.NumberOfPtrs), | |||
9326 | Info.RTArgs.SizesArray, | |||
9327 | /*Idx0=*/0, | |||
9328 | /*Idx1=*/I); | |||
9329 | Address SAddr(S, CGM.Int64Ty, Ctx.getTypeAlignInChars(Int64Ty)); | |||
9330 | CGF.Builder.CreateStore(CGF.Builder.CreateIntCast(CombinedInfo.Sizes[I], | |||
9331 | CGM.Int64Ty, | |||
9332 | /*isSigned=*/true), | |||
9333 | SAddr); | |||
9334 | } | |||
9335 | ||||
9336 | // Fill up the mapper array. | |||
9337 | llvm::Value *MFunc = llvm::ConstantPointerNull::get(CGM.VoidPtrTy); | |||
9338 | if (CombinedInfo.Mappers[I]) { | |||
9339 | MFunc = CGM.getOpenMPRuntime().getOrCreateUserDefinedMapperFunc( | |||
9340 | cast<OMPDeclareMapperDecl>(CombinedInfo.Mappers[I])); | |||
9341 | MFunc = CGF.Builder.CreatePointerCast(MFunc, CGM.VoidPtrTy); | |||
9342 | Info.HasMapper = true; | |||
9343 | } | |||
9344 | Address MAddr = CGF.Builder.CreateConstArrayGEP(MappersArray, I); | |||
9345 | CGF.Builder.CreateStore(MFunc, MAddr); | |||
9346 | } | |||
9347 | } | |||
9348 | ||||
9349 | if (!IsNonContiguous || CombinedInfo.NonContigInfo.Offsets.empty() || | |||
9350 | Info.NumberOfPtrs == 0) | |||
9351 | return; | |||
9352 | ||||
9353 | emitNonContiguousDescriptor(CGF, CombinedInfo, Info); | |||
9354 | } | |||
9355 | ||||
9356 | /// Check for inner distribute directive. | |||
9357 | static const OMPExecutableDirective * | |||
9358 | getNestedDistributeDirective(ASTContext &Ctx, const OMPExecutableDirective &D) { | |||
9359 | const auto *CS = D.getInnermostCapturedStmt(); | |||
9360 | const auto *Body = | |||
9361 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | |||
9362 | const Stmt *ChildStmt = | |||
9363 | CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body); | |||
9364 | ||||
9365 | if (const auto *NestedDir = | |||
9366 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
9367 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); | |||
9368 | switch (D.getDirectiveKind()) { | |||
9369 | case OMPD_target: | |||
9370 | if (isOpenMPDistributeDirective(DKind)) | |||
9371 | return NestedDir; | |||
9372 | if (DKind == OMPD_teams) { | |||
9373 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | |||
9374 | /*IgnoreCaptured=*/true); | |||
9375 | if (!Body) | |||
9376 | return nullptr; | |||
9377 | ChildStmt = CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body); | |||
9378 | if (const auto *NND = | |||
9379 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
9380 | DKind = NND->getDirectiveKind(); | |||
9381 | if (isOpenMPDistributeDirective(DKind)) | |||
9382 | return NND; | |||
9383 | } | |||
9384 | } | |||
9385 | return nullptr; | |||
9386 | case OMPD_target_teams: | |||
9387 | if (isOpenMPDistributeDirective(DKind)) | |||
9388 | return NestedDir; | |||
9389 | return nullptr; | |||
9390 | case OMPD_target_parallel: | |||
9391 | case OMPD_target_simd: | |||
9392 | case OMPD_target_parallel_for: | |||
9393 | case OMPD_target_parallel_for_simd: | |||
9394 | return nullptr; | |||
9395 | case OMPD_target_teams_distribute: | |||
9396 | case OMPD_target_teams_distribute_simd: | |||
9397 | case OMPD_target_teams_distribute_parallel_for: | |||
9398 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
9399 | case OMPD_parallel: | |||
9400 | case OMPD_for: | |||
9401 | case OMPD_parallel_for: | |||
9402 | case OMPD_parallel_master: | |||
9403 | case OMPD_parallel_sections: | |||
9404 | case OMPD_for_simd: | |||
9405 | case OMPD_parallel_for_simd: | |||
9406 | case OMPD_cancel: | |||
9407 | case OMPD_cancellation_point: | |||
9408 | case OMPD_ordered: | |||
9409 | case OMPD_threadprivate: | |||
9410 | case OMPD_allocate: | |||
9411 | case OMPD_task: | |||
9412 | case OMPD_simd: | |||
9413 | case OMPD_tile: | |||
9414 | case OMPD_unroll: | |||
9415 | case OMPD_sections: | |||
9416 | case OMPD_section: | |||
9417 | case OMPD_single: | |||
9418 | case OMPD_master: | |||
9419 | case OMPD_critical: | |||
9420 | case OMPD_taskyield: | |||
9421 | case OMPD_barrier: | |||
9422 | case OMPD_taskwait: | |||
9423 | case OMPD_taskgroup: | |||
9424 | case OMPD_atomic: | |||
9425 | case OMPD_flush: | |||
9426 | case OMPD_depobj: | |||
9427 | case OMPD_scan: | |||
9428 | case OMPD_teams: | |||
9429 | case OMPD_target_data: | |||
9430 | case OMPD_target_exit_data: | |||
9431 | case OMPD_target_enter_data: | |||
9432 | case OMPD_distribute: | |||
9433 | case OMPD_distribute_simd: | |||
9434 | case OMPD_distribute_parallel_for: | |||
9435 | case OMPD_distribute_parallel_for_simd: | |||
9436 | case OMPD_teams_distribute: | |||
9437 | case OMPD_teams_distribute_simd: | |||
9438 | case OMPD_teams_distribute_parallel_for: | |||
9439 | case OMPD_teams_distribute_parallel_for_simd: | |||
9440 | case OMPD_target_update: | |||
9441 | case OMPD_declare_simd: | |||
9442 | case OMPD_declare_variant: | |||
9443 | case OMPD_begin_declare_variant: | |||
9444 | case OMPD_end_declare_variant: | |||
9445 | case OMPD_declare_target: | |||
9446 | case OMPD_end_declare_target: | |||
9447 | case OMPD_declare_reduction: | |||
9448 | case OMPD_declare_mapper: | |||
9449 | case OMPD_taskloop: | |||
9450 | case OMPD_taskloop_simd: | |||
9451 | case OMPD_master_taskloop: | |||
9452 | case OMPD_master_taskloop_simd: | |||
9453 | case OMPD_parallel_master_taskloop: | |||
9454 | case OMPD_parallel_master_taskloop_simd: | |||
9455 | case OMPD_requires: | |||
9456 | case OMPD_metadirective: | |||
9457 | case OMPD_unknown: | |||
9458 | default: | |||
9459 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9459); | |||
9460 | } | |||
9461 | } | |||
9462 | ||||
9463 | return nullptr; | |||
9464 | } | |||
9465 | ||||
9466 | /// Emit the user-defined mapper function. The code generation follows the | |||
9467 | /// pattern in the example below. | |||
9468 | /// \code | |||
9469 | /// void .omp_mapper.<type_name>.<mapper_id>.(void *rt_mapper_handle, | |||
9470 | /// void *base, void *begin, | |||
9471 | /// int64_t size, int64_t type, | |||
9472 | /// void *name = nullptr) { | |||
9473 | /// // Allocate space for an array section first or add a base/begin for | |||
9474 | /// // pointer dereference. | |||
9475 | /// if ((size > 1 || (base != begin && maptype.IsPtrAndObj)) && | |||
9476 | /// !maptype.IsDelete) | |||
9477 | /// __tgt_push_mapper_component(rt_mapper_handle, base, begin, | |||
9478 | /// size*sizeof(Ty), clearToFromMember(type)); | |||
9479 | /// // Map members. | |||
9480 | /// for (unsigned i = 0; i < size; i++) { | |||
9481 | /// // For each component specified by this mapper: | |||
9482 | /// for (auto c : begin[i]->all_components) { | |||
9483 | /// if (c.hasMapper()) | |||
9484 | /// (*c.Mapper())(rt_mapper_handle, c.arg_base, c.arg_begin, c.arg_size, | |||
9485 | /// c.arg_type, c.arg_name); | |||
9486 | /// else | |||
9487 | /// __tgt_push_mapper_component(rt_mapper_handle, c.arg_base, | |||
9488 | /// c.arg_begin, c.arg_size, c.arg_type, | |||
9489 | /// c.arg_name); | |||
9490 | /// } | |||
9491 | /// } | |||
9492 | /// // Delete the array section. | |||
9493 | /// if (size > 1 && maptype.IsDelete) | |||
9494 | /// __tgt_push_mapper_component(rt_mapper_handle, base, begin, | |||
9495 | /// size*sizeof(Ty), clearToFromMember(type)); | |||
9496 | /// } | |||
9497 | /// \endcode | |||
9498 | void CGOpenMPRuntime::emitUserDefinedMapper(const OMPDeclareMapperDecl *D, | |||
9499 | CodeGenFunction *CGF) { | |||
9500 | if (UDMMap.count(D) > 0) | |||
9501 | return; | |||
9502 | ASTContext &C = CGM.getContext(); | |||
9503 | QualType Ty = D->getType(); | |||
9504 | QualType PtrTy = C.getPointerType(Ty).withRestrict(); | |||
9505 | QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true); | |||
9506 | auto *MapperVarDecl = | |||
9507 | cast<VarDecl>(cast<DeclRefExpr>(D->getMapperVarRef())->getDecl()); | |||
9508 | SourceLocation Loc = D->getLocation(); | |||
9509 | CharUnits ElementSize = C.getTypeSizeInChars(Ty); | |||
9510 | llvm::Type *ElemTy = CGM.getTypes().ConvertTypeForMem(Ty); | |||
9511 | ||||
9512 | // Prepare mapper function arguments and attributes. | |||
9513 | ImplicitParamDecl HandleArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
9514 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
9515 | ImplicitParamDecl BaseArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
9516 | ImplicitParamDecl::Other); | |||
9517 | ImplicitParamDecl BeginArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
9518 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
9519 | ImplicitParamDecl SizeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, Int64Ty, | |||
9520 | ImplicitParamDecl::Other); | |||
9521 | ImplicitParamDecl TypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, Int64Ty, | |||
9522 | ImplicitParamDecl::Other); | |||
9523 | ImplicitParamDecl NameArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
9524 | ImplicitParamDecl::Other); | |||
9525 | FunctionArgList Args; | |||
9526 | Args.push_back(&HandleArg); | |||
9527 | Args.push_back(&BaseArg); | |||
9528 | Args.push_back(&BeginArg); | |||
9529 | Args.push_back(&SizeArg); | |||
9530 | Args.push_back(&TypeArg); | |||
9531 | Args.push_back(&NameArg); | |||
9532 | const CGFunctionInfo &FnInfo = | |||
9533 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
9534 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
9535 | SmallString<64> TyStr; | |||
9536 | llvm::raw_svector_ostream Out(TyStr); | |||
9537 | CGM.getCXXABI().getMangleContext().mangleTypeName(Ty, Out); | |||
9538 | std::string Name = getName({"omp_mapper", TyStr, D->getName()}); | |||
9539 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
9540 | Name, &CGM.getModule()); | |||
9541 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
9542 | Fn->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
9543 | // Start the mapper function code generation. | |||
9544 | CodeGenFunction MapperCGF(CGM); | |||
9545 | MapperCGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
9546 | // Compute the starting and end addresses of array elements. | |||
9547 | llvm::Value *Size = MapperCGF.EmitLoadOfScalar( | |||
9548 | MapperCGF.GetAddrOfLocalVar(&SizeArg), /*Volatile=*/false, | |||
9549 | C.getPointerType(Int64Ty), Loc); | |||
9550 | // Prepare common arguments for array initiation and deletion. | |||
9551 | llvm::Value *Handle = MapperCGF.EmitLoadOfScalar( | |||
9552 | MapperCGF.GetAddrOfLocalVar(&HandleArg), | |||
9553 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | |||
9554 | llvm::Value *BaseIn = MapperCGF.EmitLoadOfScalar( | |||
9555 | MapperCGF.GetAddrOfLocalVar(&BaseArg), | |||
9556 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | |||
9557 | llvm::Value *BeginIn = MapperCGF.EmitLoadOfScalar( | |||
9558 | MapperCGF.GetAddrOfLocalVar(&BeginArg), | |||
9559 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | |||
9560 | // Convert the size in bytes into the number of array elements. | |||
9561 | Size = MapperCGF.Builder.CreateExactUDiv( | |||
9562 | Size, MapperCGF.Builder.getInt64(ElementSize.getQuantity())); | |||
9563 | llvm::Value *PtrBegin = MapperCGF.Builder.CreateBitCast( | |||
9564 | BeginIn, CGM.getTypes().ConvertTypeForMem(PtrTy)); | |||
9565 | llvm::Value *PtrEnd = MapperCGF.Builder.CreateGEP(ElemTy, PtrBegin, Size); | |||
9566 | llvm::Value *MapType = MapperCGF.EmitLoadOfScalar( | |||
9567 | MapperCGF.GetAddrOfLocalVar(&TypeArg), /*Volatile=*/false, | |||
9568 | C.getPointerType(Int64Ty), Loc); | |||
9569 | llvm::Value *MapName = MapperCGF.EmitLoadOfScalar( | |||
9570 | MapperCGF.GetAddrOfLocalVar(&NameArg), | |||
9571 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | |||
9572 | ||||
9573 | // Emit array initiation if this is an array section and \p MapType indicates | |||
9574 | // that memory allocation is required. | |||
9575 | llvm::BasicBlock *HeadBB = MapperCGF.createBasicBlock("omp.arraymap.head"); | |||
9576 | emitUDMapperArrayInitOrDel(MapperCGF, Handle, BaseIn, BeginIn, Size, MapType, | |||
9577 | MapName, ElementSize, HeadBB, /*IsInit=*/true); | |||
9578 | ||||
9579 | // Emit a for loop to iterate through SizeArg of elements and map all of them. | |||
9580 | ||||
9581 | // Emit the loop header block. | |||
9582 | MapperCGF.EmitBlock(HeadBB); | |||
9583 | llvm::BasicBlock *BodyBB = MapperCGF.createBasicBlock("omp.arraymap.body"); | |||
9584 | llvm::BasicBlock *DoneBB = MapperCGF.createBasicBlock("omp.done"); | |||
9585 | // Evaluate whether the initial condition is satisfied. | |||
9586 | llvm::Value *IsEmpty = | |||
9587 | MapperCGF.Builder.CreateICmpEQ(PtrBegin, PtrEnd, "omp.arraymap.isempty"); | |||
9588 | MapperCGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | |||
9589 | llvm::BasicBlock *EntryBB = MapperCGF.Builder.GetInsertBlock(); | |||
9590 | ||||
9591 | // Emit the loop body block. | |||
9592 | MapperCGF.EmitBlock(BodyBB); | |||
9593 | llvm::BasicBlock *LastBB = BodyBB; | |||
9594 | llvm::PHINode *PtrPHI = MapperCGF.Builder.CreatePHI( | |||
9595 | PtrBegin->getType(), 2, "omp.arraymap.ptrcurrent"); | |||
9596 | PtrPHI->addIncoming(PtrBegin, EntryBB); | |||
9597 | Address PtrCurrent(PtrPHI, ElemTy, | |||
9598 | MapperCGF.GetAddrOfLocalVar(&BeginArg) | |||
9599 | .getAlignment() | |||
9600 | .alignmentOfArrayElement(ElementSize)); | |||
9601 | // Privatize the declared variable of mapper to be the current array element. | |||
9602 | CodeGenFunction::OMPPrivateScope Scope(MapperCGF); | |||
9603 | Scope.addPrivate(MapperVarDecl, PtrCurrent); | |||
9604 | (void)Scope.Privatize(); | |||
9605 | ||||
9606 | // Get map clause information. Fill up the arrays with all mapped variables. | |||
9607 | MappableExprsHandler::MapCombinedInfoTy Info; | |||
9608 | MappableExprsHandler MEHandler(*D, MapperCGF); | |||
9609 | MEHandler.generateAllInfoForMapper(Info); | |||
9610 | ||||
9611 | // Call the runtime API __tgt_mapper_num_components to get the number of | |||
9612 | // pre-existing components. | |||
9613 | llvm::Value *OffloadingArgs[] = {Handle}; | |||
9614 | llvm::Value *PreviousSize = MapperCGF.EmitRuntimeCall( | |||
9615 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
9616 | OMPRTL___tgt_mapper_num_components), | |||
9617 | OffloadingArgs); | |||
9618 | llvm::Value *ShiftedPreviousSize = MapperCGF.Builder.CreateShl( | |||
9619 | PreviousSize, | |||
9620 | MapperCGF.Builder.getInt64(MappableExprsHandler::getFlagMemberOffset())); | |||
9621 | ||||
9622 | // Fill up the runtime mapper handle for all components. | |||
9623 | for (unsigned I = 0; I < Info.BasePointers.size(); ++I) { | |||
9624 | llvm::Value *CurBaseArg = MapperCGF.Builder.CreateBitCast( | |||
9625 | *Info.BasePointers[I], CGM.getTypes().ConvertTypeForMem(C.VoidPtrTy)); | |||
9626 | llvm::Value *CurBeginArg = MapperCGF.Builder.CreateBitCast( | |||
9627 | Info.Pointers[I], CGM.getTypes().ConvertTypeForMem(C.VoidPtrTy)); | |||
9628 | llvm::Value *CurSizeArg = Info.Sizes[I]; | |||
9629 | llvm::Value *CurNameArg = | |||
9630 | (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo) | |||
9631 | ? llvm::ConstantPointerNull::get(CGM.VoidPtrTy) | |||
9632 | : emitMappingInformation(MapperCGF, OMPBuilder, Info.Exprs[I]); | |||
9633 | ||||
9634 | // Extract the MEMBER_OF field from the map type. | |||
9635 | llvm::Value *OriMapType = MapperCGF.Builder.getInt64(Info.Types[I]); | |||
9636 | llvm::Value *MemberMapType = | |||
9637 | MapperCGF.Builder.CreateNUWAdd(OriMapType, ShiftedPreviousSize); | |||
9638 | ||||
9639 | // Combine the map type inherited from user-defined mapper with that | |||
9640 | // specified in the program. According to the OMP_MAP_TO and OMP_MAP_FROM | |||
9641 | // bits of the \a MapType, which is the input argument of the mapper | |||
9642 | // function, the following code will set the OMP_MAP_TO and OMP_MAP_FROM | |||
9643 | // bits of MemberMapType. | |||
9644 | // [OpenMP 5.0], 1.2.6. map-type decay. | |||
9645 | // | alloc | to | from | tofrom | release | delete | |||
9646 | // ---------------------------------------------------------- | |||
9647 | // alloc | alloc | alloc | alloc | alloc | release | delete | |||
9648 | // to | alloc | to | alloc | to | release | delete | |||
9649 | // from | alloc | alloc | from | from | release | delete | |||
9650 | // tofrom | alloc | to | from | tofrom | release | delete | |||
9651 | llvm::Value *LeftToFrom = MapperCGF.Builder.CreateAnd( | |||
9652 | MapType, | |||
9653 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_TO | | |||
9654 | MappableExprsHandler::OMP_MAP_FROM)); | |||
9655 | llvm::BasicBlock *AllocBB = MapperCGF.createBasicBlock("omp.type.alloc"); | |||
9656 | llvm::BasicBlock *AllocElseBB = | |||
9657 | MapperCGF.createBasicBlock("omp.type.alloc.else"); | |||
9658 | llvm::BasicBlock *ToBB = MapperCGF.createBasicBlock("omp.type.to"); | |||
9659 | llvm::BasicBlock *ToElseBB = MapperCGF.createBasicBlock("omp.type.to.else"); | |||
9660 | llvm::BasicBlock *FromBB = MapperCGF.createBasicBlock("omp.type.from"); | |||
9661 | llvm::BasicBlock *EndBB = MapperCGF.createBasicBlock("omp.type.end"); | |||
9662 | llvm::Value *IsAlloc = MapperCGF.Builder.CreateIsNull(LeftToFrom); | |||
9663 | MapperCGF.Builder.CreateCondBr(IsAlloc, AllocBB, AllocElseBB); | |||
9664 | // In case of alloc, clear OMP_MAP_TO and OMP_MAP_FROM. | |||
9665 | MapperCGF.EmitBlock(AllocBB); | |||
9666 | llvm::Value *AllocMapType = MapperCGF.Builder.CreateAnd( | |||
9667 | MemberMapType, | |||
9668 | MapperCGF.Builder.getInt64(~(MappableExprsHandler::OMP_MAP_TO | | |||
9669 | MappableExprsHandler::OMP_MAP_FROM))); | |||
9670 | MapperCGF.Builder.CreateBr(EndBB); | |||
9671 | MapperCGF.EmitBlock(AllocElseBB); | |||
9672 | llvm::Value *IsTo = MapperCGF.Builder.CreateICmpEQ( | |||
9673 | LeftToFrom, | |||
9674 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_TO)); | |||
9675 | MapperCGF.Builder.CreateCondBr(IsTo, ToBB, ToElseBB); | |||
9676 | // In case of to, clear OMP_MAP_FROM. | |||
9677 | MapperCGF.EmitBlock(ToBB); | |||
9678 | llvm::Value *ToMapType = MapperCGF.Builder.CreateAnd( | |||
9679 | MemberMapType, | |||
9680 | MapperCGF.Builder.getInt64(~MappableExprsHandler::OMP_MAP_FROM)); | |||
9681 | MapperCGF.Builder.CreateBr(EndBB); | |||
9682 | MapperCGF.EmitBlock(ToElseBB); | |||
9683 | llvm::Value *IsFrom = MapperCGF.Builder.CreateICmpEQ( | |||
9684 | LeftToFrom, | |||
9685 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_FROM)); | |||
9686 | MapperCGF.Builder.CreateCondBr(IsFrom, FromBB, EndBB); | |||
9687 | // In case of from, clear OMP_MAP_TO. | |||
9688 | MapperCGF.EmitBlock(FromBB); | |||
9689 | llvm::Value *FromMapType = MapperCGF.Builder.CreateAnd( | |||
9690 | MemberMapType, | |||
9691 | MapperCGF.Builder.getInt64(~MappableExprsHandler::OMP_MAP_TO)); | |||
9692 | // In case of tofrom, do nothing. | |||
9693 | MapperCGF.EmitBlock(EndBB); | |||
9694 | LastBB = EndBB; | |||
9695 | llvm::PHINode *CurMapType = | |||
9696 | MapperCGF.Builder.CreatePHI(CGM.Int64Ty, 4, "omp.maptype"); | |||
9697 | CurMapType->addIncoming(AllocMapType, AllocBB); | |||
9698 | CurMapType->addIncoming(ToMapType, ToBB); | |||
9699 | CurMapType->addIncoming(FromMapType, FromBB); | |||
9700 | CurMapType->addIncoming(MemberMapType, ToElseBB); | |||
9701 | ||||
9702 | llvm::Value *OffloadingArgs[] = {Handle, CurBaseArg, CurBeginArg, | |||
9703 | CurSizeArg, CurMapType, CurNameArg}; | |||
9704 | if (Info.Mappers[I]) { | |||
9705 | // Call the corresponding mapper function. | |||
9706 | llvm::Function *MapperFunc = getOrCreateUserDefinedMapperFunc( | |||
9707 | cast<OMPDeclareMapperDecl>(Info.Mappers[I])); | |||
9708 | assert(MapperFunc && "Expect a valid mapper function is available.")(static_cast <bool> (MapperFunc && "Expect a valid mapper function is available." ) ? void (0) : __assert_fail ("MapperFunc && \"Expect a valid mapper function is available.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9708, __extension__ __PRETTY_FUNCTION__)); | |||
9709 | MapperCGF.EmitNounwindRuntimeCall(MapperFunc, OffloadingArgs); | |||
9710 | } else { | |||
9711 | // Call the runtime API __tgt_push_mapper_component to fill up the runtime | |||
9712 | // data structure. | |||
9713 | MapperCGF.EmitRuntimeCall( | |||
9714 | OMPBuilder.getOrCreateRuntimeFunction( | |||
9715 | CGM.getModule(), OMPRTL___tgt_push_mapper_component), | |||
9716 | OffloadingArgs); | |||
9717 | } | |||
9718 | } | |||
9719 | ||||
9720 | // Update the pointer to point to the next element that needs to be mapped, | |||
9721 | // and check whether we have mapped all elements. | |||
9722 | llvm::Value *PtrNext = MapperCGF.Builder.CreateConstGEP1_32( | |||
9723 | ElemTy, PtrPHI, /*Idx0=*/1, "omp.arraymap.next"); | |||
9724 | PtrPHI->addIncoming(PtrNext, LastBB); | |||
9725 | llvm::Value *IsDone = | |||
9726 | MapperCGF.Builder.CreateICmpEQ(PtrNext, PtrEnd, "omp.arraymap.isdone"); | |||
9727 | llvm::BasicBlock *ExitBB = MapperCGF.createBasicBlock("omp.arraymap.exit"); | |||
9728 | MapperCGF.Builder.CreateCondBr(IsDone, ExitBB, BodyBB); | |||
9729 | ||||
9730 | MapperCGF.EmitBlock(ExitBB); | |||
9731 | // Emit array deletion if this is an array section and \p MapType indicates | |||
9732 | // that deletion is required. | |||
9733 | emitUDMapperArrayInitOrDel(MapperCGF, Handle, BaseIn, BeginIn, Size, MapType, | |||
9734 | MapName, ElementSize, DoneBB, /*IsInit=*/false); | |||
9735 | ||||
9736 | // Emit the function exit block. | |||
9737 | MapperCGF.EmitBlock(DoneBB, /*IsFinished=*/true); | |||
9738 | MapperCGF.FinishFunction(); | |||
9739 | UDMMap.try_emplace(D, Fn); | |||
9740 | if (CGF) { | |||
9741 | auto &Decls = FunctionUDMMap.FindAndConstruct(CGF->CurFn); | |||
9742 | Decls.second.push_back(D); | |||
9743 | } | |||
9744 | } | |||
9745 | ||||
9746 | /// Emit the array initialization or deletion portion for user-defined mapper | |||
9747 | /// code generation. First, it evaluates whether an array section is mapped and | |||
9748 | /// whether the \a MapType instructs to delete this section. If \a IsInit is | |||
9749 | /// true, and \a MapType indicates to not delete this array, array | |||
9750 | /// initialization code is generated. If \a IsInit is false, and \a MapType | |||
9751 | /// indicates to not this array, array deletion code is generated. | |||
9752 | void CGOpenMPRuntime::emitUDMapperArrayInitOrDel( | |||
9753 | CodeGenFunction &MapperCGF, llvm::Value *Handle, llvm::Value *Base, | |||
9754 | llvm::Value *Begin, llvm::Value *Size, llvm::Value *MapType, | |||
9755 | llvm::Value *MapName, CharUnits ElementSize, llvm::BasicBlock *ExitBB, | |||
9756 | bool IsInit) { | |||
9757 | StringRef Prefix = IsInit ? ".init" : ".del"; | |||
9758 | ||||
9759 | // Evaluate if this is an array section. | |||
9760 | llvm::BasicBlock *BodyBB = | |||
9761 | MapperCGF.createBasicBlock(getName({"omp.array", Prefix})); | |||
9762 | llvm::Value *IsArray = MapperCGF.Builder.CreateICmpSGT( | |||
9763 | Size, MapperCGF.Builder.getInt64(1), "omp.arrayinit.isarray"); | |||
9764 | llvm::Value *DeleteBit = MapperCGF.Builder.CreateAnd( | |||
9765 | MapType, | |||
9766 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_DELETE)); | |||
9767 | llvm::Value *DeleteCond; | |||
9768 | llvm::Value *Cond; | |||
9769 | if (IsInit) { | |||
9770 | // base != begin? | |||
9771 | llvm::Value *BaseIsBegin = MapperCGF.Builder.CreateICmpNE(Base, Begin); | |||
9772 | // IsPtrAndObj? | |||
9773 | llvm::Value *PtrAndObjBit = MapperCGF.Builder.CreateAnd( | |||
9774 | MapType, | |||
9775 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_PTR_AND_OBJ)); | |||
9776 | PtrAndObjBit = MapperCGF.Builder.CreateIsNotNull(PtrAndObjBit); | |||
9777 | BaseIsBegin = MapperCGF.Builder.CreateAnd(BaseIsBegin, PtrAndObjBit); | |||
9778 | Cond = MapperCGF.Builder.CreateOr(IsArray, BaseIsBegin); | |||
9779 | DeleteCond = MapperCGF.Builder.CreateIsNull( | |||
9780 | DeleteBit, getName({"omp.array", Prefix, ".delete"})); | |||
9781 | } else { | |||
9782 | Cond = IsArray; | |||
9783 | DeleteCond = MapperCGF.Builder.CreateIsNotNull( | |||
9784 | DeleteBit, getName({"omp.array", Prefix, ".delete"})); | |||
9785 | } | |||
9786 | Cond = MapperCGF.Builder.CreateAnd(Cond, DeleteCond); | |||
9787 | MapperCGF.Builder.CreateCondBr(Cond, BodyBB, ExitBB); | |||
9788 | ||||
9789 | MapperCGF.EmitBlock(BodyBB); | |||
9790 | // Get the array size by multiplying element size and element number (i.e., \p | |||
9791 | // Size). | |||
9792 | llvm::Value *ArraySize = MapperCGF.Builder.CreateNUWMul( | |||
9793 | Size, MapperCGF.Builder.getInt64(ElementSize.getQuantity())); | |||
9794 | // Remove OMP_MAP_TO and OMP_MAP_FROM from the map type, so that it achieves | |||
9795 | // memory allocation/deletion purpose only. | |||
9796 | llvm::Value *MapTypeArg = MapperCGF.Builder.CreateAnd( | |||
9797 | MapType, | |||
9798 | MapperCGF.Builder.getInt64(~(MappableExprsHandler::OMP_MAP_TO | | |||
9799 | MappableExprsHandler::OMP_MAP_FROM))); | |||
9800 | MapTypeArg = MapperCGF.Builder.CreateOr( | |||
9801 | MapTypeArg, | |||
9802 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_IMPLICIT)); | |||
9803 | ||||
9804 | // Call the runtime API __tgt_push_mapper_component to fill up the runtime | |||
9805 | // data structure. | |||
9806 | llvm::Value *OffloadingArgs[] = {Handle, Base, Begin, | |||
9807 | ArraySize, MapTypeArg, MapName}; | |||
9808 | MapperCGF.EmitRuntimeCall( | |||
9809 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
9810 | OMPRTL___tgt_push_mapper_component), | |||
9811 | OffloadingArgs); | |||
9812 | } | |||
9813 | ||||
9814 | llvm::Function *CGOpenMPRuntime::getOrCreateUserDefinedMapperFunc( | |||
9815 | const OMPDeclareMapperDecl *D) { | |||
9816 | auto I = UDMMap.find(D); | |||
9817 | if (I != UDMMap.end()) | |||
9818 | return I->second; | |||
9819 | emitUserDefinedMapper(D); | |||
9820 | return UDMMap.lookup(D); | |||
9821 | } | |||
9822 | ||||
9823 | llvm::Value *CGOpenMPRuntime::emitTargetNumIterationsCall( | |||
9824 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | |||
9825 | llvm::function_ref<llvm::Value *(CodeGenFunction &CGF, | |||
9826 | const OMPLoopDirective &D)> | |||
9827 | SizeEmitter) { | |||
9828 | OpenMPDirectiveKind Kind = D.getDirectiveKind(); | |||
9829 | const OMPExecutableDirective *TD = &D; | |||
9830 | // Get nested teams distribute kind directive, if any. | |||
9831 | if (!isOpenMPDistributeDirective(Kind) || !isOpenMPTeamsDirective(Kind)) | |||
9832 | TD = getNestedDistributeDirective(CGM.getContext(), D); | |||
9833 | if (!TD) | |||
9834 | return llvm::ConstantInt::get(CGF.Int64Ty, 0); | |||
9835 | ||||
9836 | const auto *LD = cast<OMPLoopDirective>(TD); | |||
9837 | if (llvm::Value *NumIterations = SizeEmitter(CGF, *LD)) | |||
9838 | return NumIterations; | |||
9839 | return llvm::ConstantInt::get(CGF.Int64Ty, 0); | |||
9840 | } | |||
9841 | ||||
9842 | void CGOpenMPRuntime::emitTargetCall( | |||
9843 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | |||
9844 | llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, const Expr *IfCond, | |||
9845 | llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device, | |||
9846 | llvm::function_ref<llvm::Value *(CodeGenFunction &CGF, | |||
9847 | const OMPLoopDirective &D)> | |||
9848 | SizeEmitter) { | |||
9849 | if (!CGF.HaveInsertPoint()) | |||
9850 | return; | |||
9851 | ||||
9852 | const bool OffloadingMandatory = !CGM.getLangOpts().OpenMPIsDevice && | |||
9853 | CGM.getLangOpts().OpenMPOffloadMandatory; | |||
9854 | ||||
9855 | assert((OffloadingMandatory || OutlinedFn) && "Invalid outlined function!")(static_cast <bool> ((OffloadingMandatory || OutlinedFn ) && "Invalid outlined function!") ? void (0) : __assert_fail ("(OffloadingMandatory || OutlinedFn) && \"Invalid outlined function!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9855, __extension__ __PRETTY_FUNCTION__)); | |||
9856 | ||||
9857 | const bool RequiresOuterTask = D.hasClausesOfKind<OMPDependClause>() || | |||
9858 | D.hasClausesOfKind<OMPNowaitClause>() || | |||
9859 | D.hasClausesOfKind<OMPInReductionClause>(); | |||
9860 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; | |||
9861 | const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target); | |||
9862 | auto &&ArgsCodegen = [&CS, &CapturedVars](CodeGenFunction &CGF, | |||
9863 | PrePostActionTy &) { | |||
9864 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | |||
9865 | }; | |||
9866 | emitInlinedDirective(CGF, OMPD_unknown, ArgsCodegen); | |||
9867 | ||||
9868 | CodeGenFunction::OMPTargetDataInfo InputInfo; | |||
9869 | llvm::Value *MapTypesArray = nullptr; | |||
9870 | llvm::Value *MapNamesArray = nullptr; | |||
9871 | // Generate code for the host fallback function. | |||
9872 | auto &&FallbackGen = [this, OutlinedFn, &D, &CapturedVars, RequiresOuterTask, | |||
9873 | &CS, OffloadingMandatory](CodeGenFunction &CGF) { | |||
9874 | if (OffloadingMandatory) { | |||
9875 | CGF.Builder.CreateUnreachable(); | |||
9876 | } else { | |||
9877 | if (RequiresOuterTask) { | |||
9878 | CapturedVars.clear(); | |||
9879 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | |||
9880 | } | |||
9881 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars); | |||
9882 | } | |||
9883 | }; | |||
9884 | // Fill up the pointer arrays and transfer execution to the device. | |||
9885 | auto &&ThenGen = [this, Device, OutlinedFnID, &D, &InputInfo, &MapTypesArray, | |||
9886 | &MapNamesArray, SizeEmitter, | |||
9887 | FallbackGen](CodeGenFunction &CGF, PrePostActionTy &) { | |||
9888 | if (Device.getInt() == OMPC_DEVICE_ancestor) { | |||
9889 | // Reverse offloading is not supported, so just execute on the host. | |||
9890 | FallbackGen(CGF); | |||
9891 | return; | |||
9892 | } | |||
9893 | ||||
9894 | // On top of the arrays that were filled up, the target offloading call | |||
9895 | // takes as arguments the device id as well as the host pointer. The host | |||
9896 | // pointer is used by the runtime library to identify the current target | |||
9897 | // region, so it only has to be unique and not necessarily point to | |||
9898 | // anything. It could be the pointer to the outlined function that | |||
9899 | // implements the target region, but we aren't using that so that the | |||
9900 | // compiler doesn't need to keep that, and could therefore inline the host | |||
9901 | // function if proven worthwhile during optimization. | |||
9902 | ||||
9903 | // From this point on, we need to have an ID of the target region defined. | |||
9904 | assert(OutlinedFnID && "Invalid outlined function ID!")(static_cast <bool> (OutlinedFnID && "Invalid outlined function ID!" ) ? void (0) : __assert_fail ("OutlinedFnID && \"Invalid outlined function ID!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9904, __extension__ __PRETTY_FUNCTION__)); | |||
9905 | (void)OutlinedFnID; | |||
9906 | ||||
9907 | // Emit device ID if any. | |||
9908 | llvm::Value *DeviceID; | |||
9909 | if (Device.getPointer()) { | |||
9910 | assert((Device.getInt() == OMPC_DEVICE_unknown ||(static_cast <bool> ((Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && "Expected device_num modifier." ) ? void (0) : __assert_fail ("(Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && \"Expected device_num modifier.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9912, __extension__ __PRETTY_FUNCTION__)) | |||
9911 | Device.getInt() == OMPC_DEVICE_device_num) &&(static_cast <bool> ((Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && "Expected device_num modifier." ) ? void (0) : __assert_fail ("(Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && \"Expected device_num modifier.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9912, __extension__ __PRETTY_FUNCTION__)) | |||
9912 | "Expected device_num modifier.")(static_cast <bool> ((Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && "Expected device_num modifier." ) ? void (0) : __assert_fail ("(Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && \"Expected device_num modifier.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 9912, __extension__ __PRETTY_FUNCTION__)); | |||
9913 | llvm::Value *DevVal = CGF.EmitScalarExpr(Device.getPointer()); | |||
9914 | DeviceID = | |||
9915 | CGF.Builder.CreateIntCast(DevVal, CGF.Int64Ty, /*isSigned=*/true); | |||
9916 | } else { | |||
9917 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
9918 | } | |||
9919 | ||||
9920 | // Emit the number of elements in the offloading arrays. | |||
9921 | llvm::Value *PointerNum = | |||
9922 | CGF.Builder.getInt32(InputInfo.NumberOfTargetItems); | |||
9923 | ||||
9924 | // Return value of the runtime offloading call. | |||
9925 | llvm::Value *Return; | |||
9926 | ||||
9927 | llvm::Value *NumTeams = emitNumTeamsForTargetDirective(CGF, D); | |||
9928 | llvm::Value *NumThreads = emitNumThreadsForTargetDirective(CGF, D); | |||
9929 | ||||
9930 | // Source location for the ident struct | |||
9931 | llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | |||
9932 | ||||
9933 | // Get tripcount for the target loop-based directive. | |||
9934 | llvm::Value *NumIterations = | |||
9935 | emitTargetNumIterationsCall(CGF, D, SizeEmitter); | |||
9936 | ||||
9937 | // Arguments for the target kernel. | |||
9938 | SmallVector<llvm::Value *> KernelArgs{ | |||
9939 | CGF.Builder.getInt32(/* Version */ 1), | |||
9940 | PointerNum, | |||
9941 | InputInfo.BasePointersArray.getPointer(), | |||
9942 | InputInfo.PointersArray.getPointer(), | |||
9943 | InputInfo.SizesArray.getPointer(), | |||
9944 | MapTypesArray, | |||
9945 | MapNamesArray, | |||
9946 | InputInfo.MappersArray.getPointer(), | |||
9947 | NumIterations}; | |||
9948 | ||||
9949 | // Arguments passed to the 'nowait' variant. | |||
9950 | SmallVector<llvm::Value *> NoWaitKernelArgs{ | |||
9951 | CGF.Builder.getInt32(0), | |||
9952 | llvm::ConstantPointerNull::get(CGM.VoidPtrTy), | |||
9953 | CGF.Builder.getInt32(0), | |||
9954 | llvm::ConstantPointerNull::get(CGM.VoidPtrTy), | |||
9955 | }; | |||
9956 | ||||
9957 | bool HasNoWait = D.hasClausesOfKind<OMPNowaitClause>(); | |||
9958 | ||||
9959 | // The target region is an outlined function launched by the runtime | |||
9960 | // via calls to __tgt_target_kernel(). | |||
9961 | // | |||
9962 | // Note that on the host and CPU targets, the runtime implementation of | |||
9963 | // these calls simply call the outlined function without forking threads. | |||
9964 | // The outlined functions themselves have runtime calls to | |||
9965 | // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by | |||
9966 | // the compiler in emitTeamsCall() and emitParallelCall(). | |||
9967 | // | |||
9968 | // In contrast, on the NVPTX target, the implementation of | |||
9969 | // __tgt_target_teams() launches a GPU kernel with the requested number | |||
9970 | // of teams and threads so no additional calls to the runtime are required. | |||
9971 | // Check the error code and execute the host version if required. | |||
9972 | CGF.Builder.restoreIP( | |||
9973 | HasNoWait ? OMPBuilder.emitTargetKernel( | |||
9974 | CGF.Builder, Return, RTLoc, DeviceID, NumTeams, | |||
9975 | NumThreads, OutlinedFnID, KernelArgs, NoWaitKernelArgs) | |||
9976 | : OMPBuilder.emitTargetKernel(CGF.Builder, Return, RTLoc, | |||
9977 | DeviceID, NumTeams, NumThreads, | |||
9978 | OutlinedFnID, KernelArgs)); | |||
9979 | ||||
9980 | llvm::BasicBlock *OffloadFailedBlock = | |||
9981 | CGF.createBasicBlock("omp_offload.failed"); | |||
9982 | llvm::BasicBlock *OffloadContBlock = | |||
9983 | CGF.createBasicBlock("omp_offload.cont"); | |||
9984 | llvm::Value *Failed = CGF.Builder.CreateIsNotNull(Return); | |||
9985 | CGF.Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock); | |||
9986 | ||||
9987 | CGF.EmitBlock(OffloadFailedBlock); | |||
9988 | FallbackGen(CGF); | |||
9989 | ||||
9990 | CGF.EmitBranch(OffloadContBlock); | |||
9991 | ||||
9992 | CGF.EmitBlock(OffloadContBlock, /*IsFinished=*/true); | |||
9993 | }; | |||
9994 | ||||
9995 | // Notify that the host version must be executed. | |||
9996 | auto &&ElseGen = [FallbackGen](CodeGenFunction &CGF, PrePostActionTy &) { | |||
9997 | FallbackGen(CGF); | |||
9998 | }; | |||
9999 | ||||
10000 | auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray, | |||
10001 | &MapNamesArray, &CapturedVars, RequiresOuterTask, | |||
10002 | &CS](CodeGenFunction &CGF, PrePostActionTy &) { | |||
10003 | // Fill up the arrays with all the captured variables. | |||
10004 | MappableExprsHandler::MapCombinedInfoTy CombinedInfo; | |||
10005 | ||||
10006 | // Get mappable expression information. | |||
10007 | MappableExprsHandler MEHandler(D, CGF); | |||
10008 | llvm::DenseMap<llvm::Value *, llvm::Value *> LambdaPointers; | |||
10009 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> MappedVarSet; | |||
10010 | ||||
10011 | auto RI = CS.getCapturedRecordDecl()->field_begin(); | |||
10012 | auto *CV = CapturedVars.begin(); | |||
10013 | for (CapturedStmt::const_capture_iterator CI = CS.capture_begin(), | |||
10014 | CE = CS.capture_end(); | |||
10015 | CI != CE; ++CI, ++RI, ++CV) { | |||
10016 | MappableExprsHandler::MapCombinedInfoTy CurInfo; | |||
10017 | MappableExprsHandler::StructRangeInfoTy PartialStruct; | |||
10018 | ||||
10019 | // VLA sizes are passed to the outlined region by copy and do not have map | |||
10020 | // information associated. | |||
10021 | if (CI->capturesVariableArrayType()) { | |||
10022 | CurInfo.Exprs.push_back(nullptr); | |||
10023 | CurInfo.BasePointers.push_back(*CV); | |||
10024 | CurInfo.Pointers.push_back(*CV); | |||
10025 | CurInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | |||
10026 | CGF.getTypeSize(RI->getType()), CGF.Int64Ty, /*isSigned=*/true)); | |||
10027 | // Copy to the device as an argument. No need to retrieve it. | |||
10028 | CurInfo.Types.push_back(MappableExprsHandler::OMP_MAP_LITERAL | | |||
10029 | MappableExprsHandler::OMP_MAP_TARGET_PARAM | | |||
10030 | MappableExprsHandler::OMP_MAP_IMPLICIT); | |||
10031 | CurInfo.Mappers.push_back(nullptr); | |||
10032 | } else { | |||
10033 | // If we have any information in the map clause, we use it, otherwise we | |||
10034 | // just do a default mapping. | |||
10035 | MEHandler.generateInfoForCapture(CI, *CV, CurInfo, PartialStruct); | |||
10036 | if (!CI->capturesThis()) | |||
10037 | MappedVarSet.insert(CI->getCapturedVar()); | |||
10038 | else | |||
10039 | MappedVarSet.insert(nullptr); | |||
10040 | if (CurInfo.BasePointers.empty() && !PartialStruct.Base.isValid()) | |||
10041 | MEHandler.generateDefaultMapInfo(*CI, **RI, *CV, CurInfo); | |||
10042 | // Generate correct mapping for variables captured by reference in | |||
10043 | // lambdas. | |||
10044 | if (CI->capturesVariable()) | |||
10045 | MEHandler.generateInfoForLambdaCaptures(CI->getCapturedVar(), *CV, | |||
10046 | CurInfo, LambdaPointers); | |||
10047 | } | |||
10048 | // We expect to have at least an element of information for this capture. | |||
10049 | assert((!CurInfo.BasePointers.empty() || PartialStruct.Base.isValid()) &&(static_cast <bool> ((!CurInfo.BasePointers.empty() || PartialStruct .Base.isValid()) && "Non-existing map pointer for capture!" ) ? void (0) : __assert_fail ("(!CurInfo.BasePointers.empty() || PartialStruct.Base.isValid()) && \"Non-existing map pointer for capture!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10050, __extension__ __PRETTY_FUNCTION__)) | |||
10050 | "Non-existing map pointer for capture!")(static_cast <bool> ((!CurInfo.BasePointers.empty() || PartialStruct .Base.isValid()) && "Non-existing map pointer for capture!" ) ? void (0) : __assert_fail ("(!CurInfo.BasePointers.empty() || PartialStruct.Base.isValid()) && \"Non-existing map pointer for capture!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10050, __extension__ __PRETTY_FUNCTION__)); | |||
10051 | assert(CurInfo.BasePointers.size() == CurInfo.Pointers.size() &&(static_cast <bool> (CurInfo.BasePointers.size() == CurInfo .Pointers.size() && CurInfo.BasePointers.size() == CurInfo .Sizes.size() && CurInfo.BasePointers.size() == CurInfo .Types.size() && CurInfo.BasePointers.size() == CurInfo .Mappers.size() && "Inconsistent map information sizes!" ) ? void (0) : __assert_fail ("CurInfo.BasePointers.size() == CurInfo.Pointers.size() && CurInfo.BasePointers.size() == CurInfo.Sizes.size() && CurInfo.BasePointers.size() == CurInfo.Types.size() && CurInfo.BasePointers.size() == CurInfo.Mappers.size() && \"Inconsistent map information sizes!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10055, __extension__ __PRETTY_FUNCTION__)) | |||
10052 | CurInfo.BasePointers.size() == CurInfo.Sizes.size() &&(static_cast <bool> (CurInfo.BasePointers.size() == CurInfo .Pointers.size() && CurInfo.BasePointers.size() == CurInfo .Sizes.size() && CurInfo.BasePointers.size() == CurInfo .Types.size() && CurInfo.BasePointers.size() == CurInfo .Mappers.size() && "Inconsistent map information sizes!" ) ? void (0) : __assert_fail ("CurInfo.BasePointers.size() == CurInfo.Pointers.size() && CurInfo.BasePointers.size() == CurInfo.Sizes.size() && CurInfo.BasePointers.size() == CurInfo.Types.size() && CurInfo.BasePointers.size() == CurInfo.Mappers.size() && \"Inconsistent map information sizes!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10055, __extension__ __PRETTY_FUNCTION__)) | |||
10053 | CurInfo.BasePointers.size() == CurInfo.Types.size() &&(static_cast <bool> (CurInfo.BasePointers.size() == CurInfo .Pointers.size() && CurInfo.BasePointers.size() == CurInfo .Sizes.size() && CurInfo.BasePointers.size() == CurInfo .Types.size() && CurInfo.BasePointers.size() == CurInfo .Mappers.size() && "Inconsistent map information sizes!" ) ? void (0) : __assert_fail ("CurInfo.BasePointers.size() == CurInfo.Pointers.size() && CurInfo.BasePointers.size() == CurInfo.Sizes.size() && CurInfo.BasePointers.size() == CurInfo.Types.size() && CurInfo.BasePointers.size() == CurInfo.Mappers.size() && \"Inconsistent map information sizes!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10055, __extension__ __PRETTY_FUNCTION__)) | |||
10054 | CurInfo.BasePointers.size() == CurInfo.Mappers.size() &&(static_cast <bool> (CurInfo.BasePointers.size() == CurInfo .Pointers.size() && CurInfo.BasePointers.size() == CurInfo .Sizes.size() && CurInfo.BasePointers.size() == CurInfo .Types.size() && CurInfo.BasePointers.size() == CurInfo .Mappers.size() && "Inconsistent map information sizes!" ) ? void (0) : __assert_fail ("CurInfo.BasePointers.size() == CurInfo.Pointers.size() && CurInfo.BasePointers.size() == CurInfo.Sizes.size() && CurInfo.BasePointers.size() == CurInfo.Types.size() && CurInfo.BasePointers.size() == CurInfo.Mappers.size() && \"Inconsistent map information sizes!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10055, __extension__ __PRETTY_FUNCTION__)) | |||
10055 | "Inconsistent map information sizes!")(static_cast <bool> (CurInfo.BasePointers.size() == CurInfo .Pointers.size() && CurInfo.BasePointers.size() == CurInfo .Sizes.size() && CurInfo.BasePointers.size() == CurInfo .Types.size() && CurInfo.BasePointers.size() == CurInfo .Mappers.size() && "Inconsistent map information sizes!" ) ? void (0) : __assert_fail ("CurInfo.BasePointers.size() == CurInfo.Pointers.size() && CurInfo.BasePointers.size() == CurInfo.Sizes.size() && CurInfo.BasePointers.size() == CurInfo.Types.size() && CurInfo.BasePointers.size() == CurInfo.Mappers.size() && \"Inconsistent map information sizes!\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10055, __extension__ __PRETTY_FUNCTION__)); | |||
10056 | ||||
10057 | // If there is an entry in PartialStruct it means we have a struct with | |||
10058 | // individual members mapped. Emit an extra combined entry. | |||
10059 | if (PartialStruct.Base.isValid()) { | |||
10060 | CombinedInfo.append(PartialStruct.PreliminaryMapData); | |||
10061 | MEHandler.emitCombinedEntry( | |||
10062 | CombinedInfo, CurInfo.Types, PartialStruct, nullptr, | |||
10063 | !PartialStruct.PreliminaryMapData.BasePointers.empty()); | |||
10064 | } | |||
10065 | ||||
10066 | // We need to append the results of this capture to what we already have. | |||
10067 | CombinedInfo.append(CurInfo); | |||
10068 | } | |||
10069 | // Adjust MEMBER_OF flags for the lambdas captures. | |||
10070 | MEHandler.adjustMemberOfForLambdaCaptures( | |||
10071 | LambdaPointers, CombinedInfo.BasePointers, CombinedInfo.Pointers, | |||
10072 | CombinedInfo.Types); | |||
10073 | // Map any list items in a map clause that were not captures because they | |||
10074 | // weren't referenced within the construct. | |||
10075 | MEHandler.generateAllInfo(CombinedInfo, MappedVarSet); | |||
10076 | ||||
10077 | CGOpenMPRuntime::TargetDataInfo Info; | |||
10078 | // Fill up the arrays and create the arguments. | |||
10079 | emitOffloadingArrays(CGF, CombinedInfo, Info, OMPBuilder); | |||
10080 | bool EmitDebug = | |||
10081 | CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo; | |||
10082 | OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, Info.RTArgs, Info, | |||
10083 | EmitDebug, | |||
10084 | /*ForEndCall=*/false); | |||
10085 | ||||
10086 | InputInfo.NumberOfTargetItems = Info.NumberOfPtrs; | |||
10087 | InputInfo.BasePointersArray = Address(Info.RTArgs.BasePointersArray, | |||
10088 | CGF.VoidPtrTy, CGM.getPointerAlign()); | |||
10089 | InputInfo.PointersArray = Address(Info.RTArgs.PointersArray, CGF.VoidPtrTy, | |||
10090 | CGM.getPointerAlign()); | |||
10091 | InputInfo.SizesArray = | |||
10092 | Address(Info.RTArgs.SizesArray, CGF.Int64Ty, CGM.getPointerAlign()); | |||
10093 | InputInfo.MappersArray = | |||
10094 | Address(Info.RTArgs.MappersArray, CGF.VoidPtrTy, CGM.getPointerAlign()); | |||
10095 | MapTypesArray = Info.RTArgs.MapTypesArray; | |||
10096 | MapNamesArray = Info.RTArgs.MapNamesArray; | |||
10097 | if (RequiresOuterTask) | |||
10098 | CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo); | |||
10099 | else | |||
10100 | emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen); | |||
10101 | }; | |||
10102 | ||||
10103 | auto &&TargetElseGen = [this, &ElseGen, &D, RequiresOuterTask]( | |||
10104 | CodeGenFunction &CGF, PrePostActionTy &) { | |||
10105 | if (RequiresOuterTask) { | |||
10106 | CodeGenFunction::OMPTargetDataInfo InputInfo; | |||
10107 | CGF.EmitOMPTargetTaskBasedDirective(D, ElseGen, InputInfo); | |||
10108 | } else { | |||
10109 | emitInlinedDirective(CGF, D.getDirectiveKind(), ElseGen); | |||
10110 | } | |||
10111 | }; | |||
10112 | ||||
10113 | // If we have a target function ID it means that we need to support | |||
10114 | // offloading, otherwise, just execute on the host. We need to execute on host | |||
10115 | // regardless of the conditional in the if clause if, e.g., the user do not | |||
10116 | // specify target triples. | |||
10117 | if (OutlinedFnID) { | |||
10118 | if (IfCond) { | |||
10119 | emitIfClause(CGF, IfCond, TargetThenGen, TargetElseGen); | |||
10120 | } else { | |||
10121 | RegionCodeGenTy ThenRCG(TargetThenGen); | |||
10122 | ThenRCG(CGF); | |||
10123 | } | |||
10124 | } else { | |||
10125 | RegionCodeGenTy ElseRCG(TargetElseGen); | |||
10126 | ElseRCG(CGF); | |||
10127 | } | |||
10128 | } | |||
10129 | ||||
10130 | void CGOpenMPRuntime::scanForTargetRegionsFunctions(const Stmt *S, | |||
10131 | StringRef ParentName) { | |||
10132 | if (!S) | |||
10133 | return; | |||
10134 | ||||
10135 | // Codegen OMP target directives that offload compute to the device. | |||
10136 | bool RequiresDeviceCodegen = | |||
10137 | isa<OMPExecutableDirective>(S) && | |||
10138 | isOpenMPTargetExecutionDirective( | |||
10139 | cast<OMPExecutableDirective>(S)->getDirectiveKind()); | |||
10140 | ||||
10141 | if (RequiresDeviceCodegen) { | |||
10142 | const auto &E = *cast<OMPExecutableDirective>(S); | |||
10143 | auto EntryInfo = | |||
10144 | getTargetEntryUniqueInfo(CGM.getContext(), E.getBeginLoc(), ParentName); | |||
10145 | ||||
10146 | // Is this a target region that should not be emitted as an entry point? If | |||
10147 | // so just signal we are done with this target region. | |||
10148 | if (!OffloadEntriesInfoManager.hasTargetRegionEntryInfo(EntryInfo)) | |||
10149 | return; | |||
10150 | ||||
10151 | switch (E.getDirectiveKind()) { | |||
10152 | case OMPD_target: | |||
10153 | CodeGenFunction::EmitOMPTargetDeviceFunction(CGM, ParentName, | |||
10154 | cast<OMPTargetDirective>(E)); | |||
10155 | break; | |||
10156 | case OMPD_target_parallel: | |||
10157 | CodeGenFunction::EmitOMPTargetParallelDeviceFunction( | |||
10158 | CGM, ParentName, cast<OMPTargetParallelDirective>(E)); | |||
10159 | break; | |||
10160 | case OMPD_target_teams: | |||
10161 | CodeGenFunction::EmitOMPTargetTeamsDeviceFunction( | |||
10162 | CGM, ParentName, cast<OMPTargetTeamsDirective>(E)); | |||
10163 | break; | |||
10164 | case OMPD_target_teams_distribute: | |||
10165 | CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction( | |||
10166 | CGM, ParentName, cast<OMPTargetTeamsDistributeDirective>(E)); | |||
10167 | break; | |||
10168 | case OMPD_target_teams_distribute_simd: | |||
10169 | CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction( | |||
10170 | CGM, ParentName, cast<OMPTargetTeamsDistributeSimdDirective>(E)); | |||
10171 | break; | |||
10172 | case OMPD_target_parallel_for: | |||
10173 | CodeGenFunction::EmitOMPTargetParallelForDeviceFunction( | |||
10174 | CGM, ParentName, cast<OMPTargetParallelForDirective>(E)); | |||
10175 | break; | |||
10176 | case OMPD_target_parallel_for_simd: | |||
10177 | CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction( | |||
10178 | CGM, ParentName, cast<OMPTargetParallelForSimdDirective>(E)); | |||
10179 | break; | |||
10180 | case OMPD_target_simd: | |||
10181 | CodeGenFunction::EmitOMPTargetSimdDeviceFunction( | |||
10182 | CGM, ParentName, cast<OMPTargetSimdDirective>(E)); | |||
10183 | break; | |||
10184 | case OMPD_target_teams_distribute_parallel_for: | |||
10185 | CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction( | |||
10186 | CGM, ParentName, | |||
10187 | cast<OMPTargetTeamsDistributeParallelForDirective>(E)); | |||
10188 | break; | |||
10189 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
10190 | CodeGenFunction:: | |||
10191 | EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction( | |||
10192 | CGM, ParentName, | |||
10193 | cast<OMPTargetTeamsDistributeParallelForSimdDirective>(E)); | |||
10194 | break; | |||
10195 | case OMPD_parallel: | |||
10196 | case OMPD_for: | |||
10197 | case OMPD_parallel_for: | |||
10198 | case OMPD_parallel_master: | |||
10199 | case OMPD_parallel_sections: | |||
10200 | case OMPD_for_simd: | |||
10201 | case OMPD_parallel_for_simd: | |||
10202 | case OMPD_cancel: | |||
10203 | case OMPD_cancellation_point: | |||
10204 | case OMPD_ordered: | |||
10205 | case OMPD_threadprivate: | |||
10206 | case OMPD_allocate: | |||
10207 | case OMPD_task: | |||
10208 | case OMPD_simd: | |||
10209 | case OMPD_tile: | |||
10210 | case OMPD_unroll: | |||
10211 | case OMPD_sections: | |||
10212 | case OMPD_section: | |||
10213 | case OMPD_single: | |||
10214 | case OMPD_master: | |||
10215 | case OMPD_critical: | |||
10216 | case OMPD_taskyield: | |||
10217 | case OMPD_barrier: | |||
10218 | case OMPD_taskwait: | |||
10219 | case OMPD_taskgroup: | |||
10220 | case OMPD_atomic: | |||
10221 | case OMPD_flush: | |||
10222 | case OMPD_depobj: | |||
10223 | case OMPD_scan: | |||
10224 | case OMPD_teams: | |||
10225 | case OMPD_target_data: | |||
10226 | case OMPD_target_exit_data: | |||
10227 | case OMPD_target_enter_data: | |||
10228 | case OMPD_distribute: | |||
10229 | case OMPD_distribute_simd: | |||
10230 | case OMPD_distribute_parallel_for: | |||
10231 | case OMPD_distribute_parallel_for_simd: | |||
10232 | case OMPD_teams_distribute: | |||
10233 | case OMPD_teams_distribute_simd: | |||
10234 | case OMPD_teams_distribute_parallel_for: | |||
10235 | case OMPD_teams_distribute_parallel_for_simd: | |||
10236 | case OMPD_target_update: | |||
10237 | case OMPD_declare_simd: | |||
10238 | case OMPD_declare_variant: | |||
10239 | case OMPD_begin_declare_variant: | |||
10240 | case OMPD_end_declare_variant: | |||
10241 | case OMPD_declare_target: | |||
10242 | case OMPD_end_declare_target: | |||
10243 | case OMPD_declare_reduction: | |||
10244 | case OMPD_declare_mapper: | |||
10245 | case OMPD_taskloop: | |||
10246 | case OMPD_taskloop_simd: | |||
10247 | case OMPD_master_taskloop: | |||
10248 | case OMPD_master_taskloop_simd: | |||
10249 | case OMPD_parallel_master_taskloop: | |||
10250 | case OMPD_parallel_master_taskloop_simd: | |||
10251 | case OMPD_requires: | |||
10252 | case OMPD_metadirective: | |||
10253 | case OMPD_unknown: | |||
10254 | default: | |||
10255 | llvm_unreachable("Unknown target directive for OpenMP device codegen.")::llvm::llvm_unreachable_internal("Unknown target directive for OpenMP device codegen." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10255); | |||
10256 | } | |||
10257 | return; | |||
10258 | } | |||
10259 | ||||
10260 | if (const auto *E = dyn_cast<OMPExecutableDirective>(S)) { | |||
10261 | if (!E->hasAssociatedStmt() || !E->getAssociatedStmt()) | |||
10262 | return; | |||
10263 | ||||
10264 | scanForTargetRegionsFunctions(E->getRawStmt(), ParentName); | |||
10265 | return; | |||
10266 | } | |||
10267 | ||||
10268 | // If this is a lambda function, look into its body. | |||
10269 | if (const auto *L = dyn_cast<LambdaExpr>(S)) | |||
10270 | S = L->getBody(); | |||
10271 | ||||
10272 | // Keep looking for target regions recursively. | |||
10273 | for (const Stmt *II : S->children()) | |||
10274 | scanForTargetRegionsFunctions(II, ParentName); | |||
10275 | } | |||
10276 | ||||
10277 | static bool isAssumedToBeNotEmitted(const ValueDecl *VD, bool IsDevice) { | |||
10278 | Optional<OMPDeclareTargetDeclAttr::DevTypeTy> DevTy = | |||
10279 | OMPDeclareTargetDeclAttr::getDeviceType(VD); | |||
10280 | if (!DevTy) | |||
10281 | return false; | |||
10282 | // Do not emit device_type(nohost) functions for the host. | |||
10283 | if (!IsDevice && DevTy == OMPDeclareTargetDeclAttr::DT_NoHost) | |||
10284 | return true; | |||
10285 | // Do not emit device_type(host) functions for the device. | |||
10286 | if (IsDevice && DevTy == OMPDeclareTargetDeclAttr::DT_Host) | |||
10287 | return true; | |||
10288 | return false; | |||
10289 | } | |||
10290 | ||||
10291 | bool CGOpenMPRuntime::emitTargetFunctions(GlobalDecl GD) { | |||
10292 | // If emitting code for the host, we do not process FD here. Instead we do | |||
10293 | // the normal code generation. | |||
10294 | if (!CGM.getLangOpts().OpenMPIsDevice) { | |||
10295 | if (const auto *FD = dyn_cast<FunctionDecl>(GD.getDecl())) | |||
10296 | if (isAssumedToBeNotEmitted(cast<ValueDecl>(FD), | |||
10297 | CGM.getLangOpts().OpenMPIsDevice)) | |||
10298 | return true; | |||
10299 | return false; | |||
10300 | } | |||
10301 | ||||
10302 | const ValueDecl *VD = cast<ValueDecl>(GD.getDecl()); | |||
10303 | // Try to detect target regions in the function. | |||
10304 | if (const auto *FD = dyn_cast<FunctionDecl>(VD)) { | |||
10305 | StringRef Name = CGM.getMangledName(GD); | |||
10306 | scanForTargetRegionsFunctions(FD->getBody(), Name); | |||
10307 | if (isAssumedToBeNotEmitted(cast<ValueDecl>(FD), | |||
10308 | CGM.getLangOpts().OpenMPIsDevice)) | |||
10309 | return true; | |||
10310 | } | |||
10311 | ||||
10312 | // Do not to emit function if it is not marked as declare target. | |||
10313 | return !OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD) && | |||
10314 | AlreadyEmittedTargetDecls.count(VD) == 0; | |||
10315 | } | |||
10316 | ||||
10317 | bool CGOpenMPRuntime::emitTargetGlobalVariable(GlobalDecl GD) { | |||
10318 | if (isAssumedToBeNotEmitted(cast<ValueDecl>(GD.getDecl()), | |||
10319 | CGM.getLangOpts().OpenMPIsDevice)) | |||
10320 | return true; | |||
10321 | ||||
10322 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
10323 | return false; | |||
10324 | ||||
10325 | // Check if there are Ctors/Dtors in this declaration and look for target | |||
10326 | // regions in it. We use the complete variant to produce the kernel name | |||
10327 | // mangling. | |||
10328 | QualType RDTy = cast<VarDecl>(GD.getDecl())->getType(); | |||
10329 | if (const auto *RD = RDTy->getBaseElementTypeUnsafe()->getAsCXXRecordDecl()) { | |||
10330 | for (const CXXConstructorDecl *Ctor : RD->ctors()) { | |||
10331 | StringRef ParentName = | |||
10332 | CGM.getMangledName(GlobalDecl(Ctor, Ctor_Complete)); | |||
10333 | scanForTargetRegionsFunctions(Ctor->getBody(), ParentName); | |||
10334 | } | |||
10335 | if (const CXXDestructorDecl *Dtor = RD->getDestructor()) { | |||
10336 | StringRef ParentName = | |||
10337 | CGM.getMangledName(GlobalDecl(Dtor, Dtor_Complete)); | |||
10338 | scanForTargetRegionsFunctions(Dtor->getBody(), ParentName); | |||
10339 | } | |||
10340 | } | |||
10341 | ||||
10342 | // Do not to emit variable if it is not marked as declare target. | |||
10343 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
10344 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration( | |||
10345 | cast<VarDecl>(GD.getDecl())); | |||
10346 | if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link || | |||
10347 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | |||
10348 | HasRequiresUnifiedSharedMemory)) { | |||
10349 | DeferredGlobalVariables.insert(cast<VarDecl>(GD.getDecl())); | |||
10350 | return true; | |||
10351 | } | |||
10352 | return false; | |||
10353 | } | |||
10354 | ||||
10355 | void CGOpenMPRuntime::registerTargetGlobalVariable(const VarDecl *VD, | |||
10356 | llvm::Constant *Addr) { | |||
10357 | if (CGM.getLangOpts().OMPTargetTriples.empty() && | |||
10358 | !CGM.getLangOpts().OpenMPIsDevice) | |||
10359 | return; | |||
10360 | ||||
10361 | // If we have host/nohost variables, they do not need to be registered. | |||
10362 | Optional<OMPDeclareTargetDeclAttr::DevTypeTy> DevTy = | |||
10363 | OMPDeclareTargetDeclAttr::getDeviceType(VD); | |||
10364 | if (DevTy && *DevTy != OMPDeclareTargetDeclAttr::DT_Any) | |||
10365 | return; | |||
10366 | ||||
10367 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
10368 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
10369 | if (!Res) { | |||
10370 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
10371 | // Register non-target variables being emitted in device code (debug info | |||
10372 | // may cause this). | |||
10373 | StringRef VarName = CGM.getMangledName(VD); | |||
10374 | EmittedNonTargetVariables.try_emplace(VarName, Addr); | |||
10375 | } | |||
10376 | return; | |||
10377 | } | |||
10378 | // Register declare target variables. | |||
10379 | llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind Flags; | |||
10380 | StringRef VarName; | |||
10381 | int64_t VarSize; | |||
10382 | llvm::GlobalValue::LinkageTypes Linkage; | |||
10383 | ||||
10384 | if (*Res == OMPDeclareTargetDeclAttr::MT_To && | |||
10385 | !HasRequiresUnifiedSharedMemory) { | |||
10386 | Flags = llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo; | |||
10387 | VarName = CGM.getMangledName(VD); | |||
10388 | if (VD->hasDefinition(CGM.getContext()) != VarDecl::DeclarationOnly) { | |||
10389 | VarSize = | |||
10390 | CGM.getContext().getTypeSizeInChars(VD->getType()).getQuantity(); | |||
10391 | assert(VarSize != 0 && "Expected non-zero size of the variable")(static_cast <bool> (VarSize != 0 && "Expected non-zero size of the variable" ) ? void (0) : __assert_fail ("VarSize != 0 && \"Expected non-zero size of the variable\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10391, __extension__ __PRETTY_FUNCTION__)); | |||
10392 | } else { | |||
10393 | VarSize = 0; | |||
10394 | } | |||
10395 | Linkage = CGM.getLLVMLinkageVarDefinition(VD, /*IsConstant=*/false); | |||
10396 | // Temp solution to prevent optimizations of the internal variables. | |||
10397 | if (CGM.getLangOpts().OpenMPIsDevice && !VD->isExternallyVisible()) { | |||
10398 | // Do not create a "ref-variable" if the original is not also available | |||
10399 | // on the host. | |||
10400 | if (!OffloadEntriesInfoManager.hasDeviceGlobalVarEntryInfo(VarName)) | |||
10401 | return; | |||
10402 | std::string RefName = getName({VarName, "ref"}); | |||
10403 | if (!CGM.GetGlobalValue(RefName)) { | |||
10404 | llvm::Constant *AddrRef = | |||
10405 | getOrCreateInternalVariable(Addr->getType(), RefName); | |||
10406 | auto *GVAddrRef = cast<llvm::GlobalVariable>(AddrRef); | |||
10407 | GVAddrRef->setConstant(/*Val=*/true); | |||
10408 | GVAddrRef->setLinkage(llvm::GlobalValue::InternalLinkage); | |||
10409 | GVAddrRef->setInitializer(Addr); | |||
10410 | CGM.addCompilerUsedGlobal(GVAddrRef); | |||
10411 | } | |||
10412 | } | |||
10413 | } else { | |||
10414 | assert(((*Res == OMPDeclareTargetDeclAttr::MT_Link) ||(static_cast <bool> (((*Res == OMPDeclareTargetDeclAttr ::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? void (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10417, __extension__ __PRETTY_FUNCTION__)) | |||
10415 | (*Res == OMPDeclareTargetDeclAttr::MT_To &&(static_cast <bool> (((*Res == OMPDeclareTargetDeclAttr ::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? void (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10417, __extension__ __PRETTY_FUNCTION__)) | |||
10416 | HasRequiresUnifiedSharedMemory)) &&(static_cast <bool> (((*Res == OMPDeclareTargetDeclAttr ::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? void (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10417, __extension__ __PRETTY_FUNCTION__)) | |||
10417 | "Declare target attribute must link or to with unified memory.")(static_cast <bool> (((*Res == OMPDeclareTargetDeclAttr ::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? void (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10417, __extension__ __PRETTY_FUNCTION__)); | |||
10418 | if (*Res == OMPDeclareTargetDeclAttr::MT_Link) | |||
10419 | Flags = llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink; | |||
10420 | else | |||
10421 | Flags = llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo; | |||
10422 | ||||
10423 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
10424 | VarName = Addr->getName(); | |||
10425 | Addr = nullptr; | |||
10426 | } else { | |||
10427 | VarName = getAddrOfDeclareTargetVar(VD).getName(); | |||
10428 | Addr = cast<llvm::Constant>(getAddrOfDeclareTargetVar(VD).getPointer()); | |||
10429 | } | |||
10430 | VarSize = CGM.getPointerSize().getQuantity(); | |||
10431 | Linkage = llvm::GlobalValue::WeakAnyLinkage; | |||
10432 | } | |||
10433 | ||||
10434 | OffloadEntriesInfoManager.registerDeviceGlobalVarEntryInfo( | |||
10435 | VarName, Addr, VarSize, Flags, Linkage, CGM.getLangOpts().OpenMPIsDevice); | |||
10436 | } | |||
10437 | ||||
10438 | bool CGOpenMPRuntime::emitTargetGlobal(GlobalDecl GD) { | |||
10439 | if (isa<FunctionDecl>(GD.getDecl()) || | |||
10440 | isa<OMPDeclareReductionDecl>(GD.getDecl())) | |||
10441 | return emitTargetFunctions(GD); | |||
10442 | ||||
10443 | return emitTargetGlobalVariable(GD); | |||
10444 | } | |||
10445 | ||||
10446 | void CGOpenMPRuntime::emitDeferredTargetDecls() const { | |||
10447 | for (const VarDecl *VD : DeferredGlobalVariables) { | |||
10448 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
10449 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
10450 | if (!Res) | |||
10451 | continue; | |||
10452 | if (*Res == OMPDeclareTargetDeclAttr::MT_To && | |||
10453 | !HasRequiresUnifiedSharedMemory) { | |||
10454 | CGM.EmitGlobal(VD); | |||
10455 | } else { | |||
10456 | assert((*Res == OMPDeclareTargetDeclAttr::MT_Link ||(static_cast <bool> ((*Res == OMPDeclareTargetDeclAttr:: MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory." ) ? void (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10459, __extension__ __PRETTY_FUNCTION__)) | |||
10457 | (*Res == OMPDeclareTargetDeclAttr::MT_To &&(static_cast <bool> ((*Res == OMPDeclareTargetDeclAttr:: MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory." ) ? void (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10459, __extension__ __PRETTY_FUNCTION__)) | |||
10458 | HasRequiresUnifiedSharedMemory)) &&(static_cast <bool> ((*Res == OMPDeclareTargetDeclAttr:: MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory." ) ? void (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10459, __extension__ __PRETTY_FUNCTION__)) | |||
10459 | "Expected link clause or to clause with unified memory.")(static_cast <bool> ((*Res == OMPDeclareTargetDeclAttr:: MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory." ) ? void (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10459, __extension__ __PRETTY_FUNCTION__)); | |||
10460 | (void)CGM.getOpenMPRuntime().getAddrOfDeclareTargetVar(VD); | |||
10461 | } | |||
10462 | } | |||
10463 | } | |||
10464 | ||||
10465 | void CGOpenMPRuntime::adjustTargetSpecificDataForLambdas( | |||
10466 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { | |||
10467 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTargetExecutionDirective(D .getDirectiveKind()) && " Expected target-based directive." ) ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10468, __extension__ __PRETTY_FUNCTION__)) | |||
10468 | " Expected target-based directive.")(static_cast <bool> (isOpenMPTargetExecutionDirective(D .getDirectiveKind()) && " Expected target-based directive." ) ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10468, __extension__ __PRETTY_FUNCTION__)); | |||
10469 | } | |||
10470 | ||||
10471 | void CGOpenMPRuntime::processRequiresDirective(const OMPRequiresDecl *D) { | |||
10472 | for (const OMPClause *Clause : D->clauselists()) { | |||
10473 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { | |||
10474 | HasRequiresUnifiedSharedMemory = true; | |||
10475 | } else if (const auto *AC = | |||
10476 | dyn_cast<OMPAtomicDefaultMemOrderClause>(Clause)) { | |||
10477 | switch (AC->getAtomicDefaultMemOrderKind()) { | |||
10478 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_acq_rel: | |||
10479 | RequiresAtomicOrdering = llvm::AtomicOrdering::AcquireRelease; | |||
10480 | break; | |||
10481 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_seq_cst: | |||
10482 | RequiresAtomicOrdering = llvm::AtomicOrdering::SequentiallyConsistent; | |||
10483 | break; | |||
10484 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_relaxed: | |||
10485 | RequiresAtomicOrdering = llvm::AtomicOrdering::Monotonic; | |||
10486 | break; | |||
10487 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_unknown: | |||
10488 | break; | |||
10489 | } | |||
10490 | } | |||
10491 | } | |||
10492 | } | |||
10493 | ||||
10494 | llvm::AtomicOrdering CGOpenMPRuntime::getDefaultMemoryOrdering() const { | |||
10495 | return RequiresAtomicOrdering; | |||
10496 | } | |||
10497 | ||||
10498 | bool CGOpenMPRuntime::hasAllocateAttributeForGlobalVar(const VarDecl *VD, | |||
10499 | LangAS &AS) { | |||
10500 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) | |||
10501 | return false; | |||
10502 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); | |||
10503 | switch(A->getAllocatorType()) { | |||
10504 | case OMPAllocateDeclAttr::OMPNullMemAlloc: | |||
10505 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: | |||
10506 | // Not supported, fallback to the default mem space. | |||
10507 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: | |||
10508 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: | |||
10509 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: | |||
10510 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: | |||
10511 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: | |||
10512 | case OMPAllocateDeclAttr::OMPConstMemAlloc: | |||
10513 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: | |||
10514 | AS = LangAS::Default; | |||
10515 | return true; | |||
10516 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: | |||
10517 | llvm_unreachable("Expected predefined allocator for the variables with the "::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10518 ) | |||
10518 | "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10518 ); | |||
10519 | } | |||
10520 | return false; | |||
10521 | } | |||
10522 | ||||
10523 | bool CGOpenMPRuntime::hasRequiresUnifiedSharedMemory() const { | |||
10524 | return HasRequiresUnifiedSharedMemory; | |||
10525 | } | |||
10526 | ||||
10527 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII::DisableAutoDeclareTargetRAII( | |||
10528 | CodeGenModule &CGM) | |||
10529 | : CGM(CGM) { | |||
10530 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
10531 | SavedShouldMarkAsGlobal = CGM.getOpenMPRuntime().ShouldMarkAsGlobal; | |||
10532 | CGM.getOpenMPRuntime().ShouldMarkAsGlobal = false; | |||
10533 | } | |||
10534 | } | |||
10535 | ||||
10536 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII::~DisableAutoDeclareTargetRAII() { | |||
10537 | if (CGM.getLangOpts().OpenMPIsDevice) | |||
10538 | CGM.getOpenMPRuntime().ShouldMarkAsGlobal = SavedShouldMarkAsGlobal; | |||
10539 | } | |||
10540 | ||||
10541 | bool CGOpenMPRuntime::markAsGlobalTarget(GlobalDecl GD) { | |||
10542 | if (!CGM.getLangOpts().OpenMPIsDevice || !ShouldMarkAsGlobal) | |||
10543 | return true; | |||
10544 | ||||
10545 | const auto *D = cast<FunctionDecl>(GD.getDecl()); | |||
10546 | // Do not to emit function if it is marked as declare target as it was already | |||
10547 | // emitted. | |||
10548 | if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(D)) { | |||
10549 | if (D->hasBody() && AlreadyEmittedTargetDecls.count(D) == 0) { | |||
10550 | if (auto *F = dyn_cast_or_null<llvm::Function>( | |||
10551 | CGM.GetGlobalValue(CGM.getMangledName(GD)))) | |||
10552 | return !F->isDeclaration(); | |||
10553 | return false; | |||
10554 | } | |||
10555 | return true; | |||
10556 | } | |||
10557 | ||||
10558 | return !AlreadyEmittedTargetDecls.insert(D).second; | |||
10559 | } | |||
10560 | ||||
10561 | llvm::Function *CGOpenMPRuntime::emitRequiresDirectiveRegFun() { | |||
10562 | // If we don't have entries or if we are emitting code for the device, we | |||
10563 | // don't need to do anything. | |||
10564 | if (CGM.getLangOpts().OMPTargetTriples.empty() || | |||
10565 | CGM.getLangOpts().OpenMPSimd || CGM.getLangOpts().OpenMPIsDevice || | |||
10566 | (OffloadEntriesInfoManager.empty() && | |||
10567 | !HasEmittedDeclareTargetRegion && | |||
10568 | !HasEmittedTargetRegion)) | |||
10569 | return nullptr; | |||
10570 | ||||
10571 | // Create and register the function that handles the requires directives. | |||
10572 | ASTContext &C = CGM.getContext(); | |||
10573 | ||||
10574 | llvm::Function *RequiresRegFn; | |||
10575 | { | |||
10576 | CodeGenFunction CGF(CGM); | |||
10577 | const auto &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
10578 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
10579 | std::string ReqName = getName({"omp_offloading", "requires_reg"}); | |||
10580 | RequiresRegFn = CGM.CreateGlobalInitOrCleanUpFunction(FTy, ReqName, FI); | |||
10581 | CGF.StartFunction(GlobalDecl(), C.VoidTy, RequiresRegFn, FI, {}); | |||
10582 | OpenMPOffloadingRequiresDirFlags Flags = OMP_REQ_NONE; | |||
10583 | // TODO: check for other requires clauses. | |||
10584 | // The requires directive takes effect only when a target region is | |||
10585 | // present in the compilation unit. Otherwise it is ignored and not | |||
10586 | // passed to the runtime. This avoids the runtime from throwing an error | |||
10587 | // for mismatching requires clauses across compilation units that don't | |||
10588 | // contain at least 1 target region. | |||
10589 | assert((HasEmittedTargetRegion ||(static_cast <bool> ((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? void (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10592, __extension__ __PRETTY_FUNCTION__)) | |||
10590 | HasEmittedDeclareTargetRegion ||(static_cast <bool> ((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? void (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10592, __extension__ __PRETTY_FUNCTION__)) | |||
10591 | !OffloadEntriesInfoManager.empty()) &&(static_cast <bool> ((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? void (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10592, __extension__ __PRETTY_FUNCTION__)) | |||
10592 | "Target or declare target region expected.")(static_cast <bool> ((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? void (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10592, __extension__ __PRETTY_FUNCTION__)); | |||
10593 | if (HasRequiresUnifiedSharedMemory) | |||
10594 | Flags = OMP_REQ_UNIFIED_SHARED_MEMORY; | |||
10595 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
10596 | CGM.getModule(), OMPRTL___tgt_register_requires), | |||
10597 | llvm::ConstantInt::get(CGM.Int64Ty, Flags)); | |||
10598 | CGF.FinishFunction(); | |||
10599 | } | |||
10600 | return RequiresRegFn; | |||
10601 | } | |||
10602 | ||||
10603 | void CGOpenMPRuntime::emitTeamsCall(CodeGenFunction &CGF, | |||
10604 | const OMPExecutableDirective &D, | |||
10605 | SourceLocation Loc, | |||
10606 | llvm::Function *OutlinedFn, | |||
10607 | ArrayRef<llvm::Value *> CapturedVars) { | |||
10608 | if (!CGF.HaveInsertPoint()) | |||
10609 | return; | |||
10610 | ||||
10611 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
10612 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
10613 | ||||
10614 | // Build call __kmpc_fork_teams(loc, n, microtask, var1, .., varn); | |||
10615 | llvm::Value *Args[] = { | |||
10616 | RTLoc, | |||
10617 | CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars | |||
10618 | CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy())}; | |||
10619 | llvm::SmallVector<llvm::Value *, 16> RealArgs; | |||
10620 | RealArgs.append(std::begin(Args), std::end(Args)); | |||
10621 | RealArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
10622 | ||||
10623 | llvm::FunctionCallee RTLFn = OMPBuilder.getOrCreateRuntimeFunction( | |||
10624 | CGM.getModule(), OMPRTL___kmpc_fork_teams); | |||
10625 | CGF.EmitRuntimeCall(RTLFn, RealArgs); | |||
10626 | } | |||
10627 | ||||
10628 | void CGOpenMPRuntime::emitNumTeamsClause(CodeGenFunction &CGF, | |||
10629 | const Expr *NumTeams, | |||
10630 | const Expr *ThreadLimit, | |||
10631 | SourceLocation Loc) { | |||
10632 | if (!CGF.HaveInsertPoint()) | |||
10633 | return; | |||
10634 | ||||
10635 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
10636 | ||||
10637 | llvm::Value *NumTeamsVal = | |||
10638 | NumTeams | |||
10639 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(NumTeams), | |||
10640 | CGF.CGM.Int32Ty, /* isSigned = */ true) | |||
10641 | : CGF.Builder.getInt32(0); | |||
10642 | ||||
10643 | llvm::Value *ThreadLimitVal = | |||
10644 | ThreadLimit | |||
10645 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(ThreadLimit), | |||
10646 | CGF.CGM.Int32Ty, /* isSigned = */ true) | |||
10647 | : CGF.Builder.getInt32(0); | |||
10648 | ||||
10649 | // Build call __kmpc_push_num_teamss(&loc, global_tid, num_teams, thread_limit) | |||
10650 | llvm::Value *PushNumTeamsArgs[] = {RTLoc, getThreadID(CGF, Loc), NumTeamsVal, | |||
10651 | ThreadLimitVal}; | |||
10652 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | |||
10653 | CGM.getModule(), OMPRTL___kmpc_push_num_teams), | |||
10654 | PushNumTeamsArgs); | |||
10655 | } | |||
10656 | ||||
10657 | void CGOpenMPRuntime::emitTargetDataCalls( | |||
10658 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
10659 | const Expr *Device, const RegionCodeGenTy &CodeGen, | |||
10660 | CGOpenMPRuntime::TargetDataInfo &Info) { | |||
10661 | if (!CGF.HaveInsertPoint()) | |||
10662 | return; | |||
10663 | ||||
10664 | // Action used to replace the default codegen action and turn privatization | |||
10665 | // off. | |||
10666 | PrePostActionTy NoPrivAction; | |||
10667 | ||||
10668 | // Generate the code for the opening of the data environment. Capture all the | |||
10669 | // arguments of the runtime call by reference because they are used in the | |||
10670 | // closing of the region. | |||
10671 | auto &&BeginThenGen = [this, &D, Device, &Info, | |||
10672 | &CodeGen](CodeGenFunction &CGF, PrePostActionTy &) { | |||
10673 | // Fill up the arrays with all the mapped variables. | |||
10674 | MappableExprsHandler::MapCombinedInfoTy CombinedInfo; | |||
10675 | ||||
10676 | // Get map clause information. | |||
10677 | MappableExprsHandler MEHandler(D, CGF); | |||
10678 | MEHandler.generateAllInfo(CombinedInfo); | |||
10679 | ||||
10680 | // Fill up the arrays and create the arguments. | |||
10681 | emitOffloadingArrays(CGF, CombinedInfo, Info, OMPBuilder, | |||
10682 | /*IsNonContiguous=*/true); | |||
10683 | ||||
10684 | llvm::OpenMPIRBuilder::TargetDataRTArgs RTArgs; | |||
10685 | bool EmitDebug = | |||
10686 | CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo; | |||
10687 | OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, RTArgs, Info, | |||
10688 | EmitDebug); | |||
10689 | ||||
10690 | // Emit device ID if any. | |||
10691 | llvm::Value *DeviceID = nullptr; | |||
10692 | if (Device) { | |||
10693 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
10694 | CGF.Int64Ty, /*isSigned=*/true); | |||
10695 | } else { | |||
10696 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
10697 | } | |||
10698 | ||||
10699 | // Emit the number of elements in the offloading arrays. | |||
10700 | llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs); | |||
10701 | // | |||
10702 | // Source location for the ident struct | |||
10703 | llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | |||
10704 | ||||
10705 | llvm::Value *OffloadingArgs[] = {RTLoc, | |||
10706 | DeviceID, | |||
10707 | PointerNum, | |||
10708 | RTArgs.BasePointersArray, | |||
10709 | RTArgs.PointersArray, | |||
10710 | RTArgs.SizesArray, | |||
10711 | RTArgs.MapTypesArray, | |||
10712 | RTArgs.MapNamesArray, | |||
10713 | RTArgs.MappersArray}; | |||
10714 | CGF.EmitRuntimeCall( | |||
10715 | OMPBuilder.getOrCreateRuntimeFunction( | |||
10716 | CGM.getModule(), OMPRTL___tgt_target_data_begin_mapper), | |||
10717 | OffloadingArgs); | |||
10718 | ||||
10719 | // If device pointer privatization is required, emit the body of the region | |||
10720 | // here. It will have to be duplicated: with and without privatization. | |||
10721 | if (!Info.CaptureDeviceAddrMap.empty()) | |||
10722 | CodeGen(CGF); | |||
10723 | }; | |||
10724 | ||||
10725 | // Generate code for the closing of the data region. | |||
10726 | auto &&EndThenGen = [this, Device, &Info, &D](CodeGenFunction &CGF, | |||
10727 | PrePostActionTy &) { | |||
10728 | assert(Info.isValid() && "Invalid data environment closing arguments.")(static_cast <bool> (Info.isValid() && "Invalid data environment closing arguments." ) ? void (0) : __assert_fail ("Info.isValid() && \"Invalid data environment closing arguments.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10728, __extension__ __PRETTY_FUNCTION__)); | |||
10729 | ||||
10730 | llvm::OpenMPIRBuilder::TargetDataRTArgs RTArgs; | |||
10731 | bool EmitDebug = | |||
10732 | CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo; | |||
10733 | OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, RTArgs, Info, | |||
10734 | EmitDebug, | |||
10735 | /*ForEndCall=*/true); | |||
10736 | ||||
10737 | // Emit device ID if any. | |||
10738 | llvm::Value *DeviceID = nullptr; | |||
10739 | if (Device) { | |||
10740 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
10741 | CGF.Int64Ty, /*isSigned=*/true); | |||
10742 | } else { | |||
10743 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
10744 | } | |||
10745 | ||||
10746 | // Emit the number of elements in the offloading arrays. | |||
10747 | llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs); | |||
10748 | ||||
10749 | // Source location for the ident struct | |||
10750 | llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | |||
10751 | ||||
10752 | llvm::Value *OffloadingArgs[] = {RTLoc, | |||
10753 | DeviceID, | |||
10754 | PointerNum, | |||
10755 | RTArgs.BasePointersArray, | |||
10756 | RTArgs.PointersArray, | |||
10757 | RTArgs.SizesArray, | |||
10758 | RTArgs.MapTypesArray, | |||
10759 | RTArgs.MapNamesArray, | |||
10760 | RTArgs.MappersArray}; | |||
10761 | CGF.EmitRuntimeCall( | |||
10762 | OMPBuilder.getOrCreateRuntimeFunction( | |||
10763 | CGM.getModule(), OMPRTL___tgt_target_data_end_mapper), | |||
10764 | OffloadingArgs); | |||
10765 | }; | |||
10766 | ||||
10767 | // If we need device pointer privatization, we need to emit the body of the | |||
10768 | // region with no privatization in the 'else' branch of the conditional. | |||
10769 | // Otherwise, we don't have to do anything. | |||
10770 | auto &&BeginElseGen = [&Info, &CodeGen, &NoPrivAction](CodeGenFunction &CGF, | |||
10771 | PrePostActionTy &) { | |||
10772 | if (!Info.CaptureDeviceAddrMap.empty()) { | |||
10773 | CodeGen.setAction(NoPrivAction); | |||
10774 | CodeGen(CGF); | |||
10775 | } | |||
10776 | }; | |||
10777 | ||||
10778 | // We don't have to do anything to close the region if the if clause evaluates | |||
10779 | // to false. | |||
10780 | auto &&EndElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {}; | |||
10781 | ||||
10782 | if (IfCond) { | |||
10783 | emitIfClause(CGF, IfCond, BeginThenGen, BeginElseGen); | |||
10784 | } else { | |||
10785 | RegionCodeGenTy RCG(BeginThenGen); | |||
10786 | RCG(CGF); | |||
10787 | } | |||
10788 | ||||
10789 | // If we don't require privatization of device pointers, we emit the body in | |||
10790 | // between the runtime calls. This avoids duplicating the body code. | |||
10791 | if (Info.CaptureDeviceAddrMap.empty()) { | |||
10792 | CodeGen.setAction(NoPrivAction); | |||
10793 | CodeGen(CGF); | |||
10794 | } | |||
10795 | ||||
10796 | if (IfCond) { | |||
10797 | emitIfClause(CGF, IfCond, EndThenGen, EndElseGen); | |||
10798 | } else { | |||
10799 | RegionCodeGenTy RCG(EndThenGen); | |||
10800 | RCG(CGF); | |||
10801 | } | |||
10802 | } | |||
10803 | ||||
10804 | void CGOpenMPRuntime::emitTargetDataStandAloneCall( | |||
10805 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
10806 | const Expr *Device) { | |||
10807 | if (!CGF.HaveInsertPoint()) | |||
10808 | return; | |||
10809 | ||||
10810 | assert((isa<OMPTargetEnterDataDirective>(D) ||(static_cast <bool> ((isa<OMPTargetEnterDataDirective >(D) || isa<OMPTargetExitDataDirective>(D) || isa< OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? void (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10813, __extension__ __PRETTY_FUNCTION__)) | |||
10811 | isa<OMPTargetExitDataDirective>(D) ||(static_cast <bool> ((isa<OMPTargetEnterDataDirective >(D) || isa<OMPTargetExitDataDirective>(D) || isa< OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? void (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10813, __extension__ __PRETTY_FUNCTION__)) | |||
10812 | isa<OMPTargetUpdateDirective>(D)) &&(static_cast <bool> ((isa<OMPTargetEnterDataDirective >(D) || isa<OMPTargetExitDataDirective>(D) || isa< OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? void (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10813, __extension__ __PRETTY_FUNCTION__)) | |||
10813 | "Expecting either target enter, exit data, or update directives.")(static_cast <bool> ((isa<OMPTargetEnterDataDirective >(D) || isa<OMPTargetExitDataDirective>(D) || isa< OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? void (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10813, __extension__ __PRETTY_FUNCTION__)); | |||
10814 | ||||
10815 | CodeGenFunction::OMPTargetDataInfo InputInfo; | |||
10816 | llvm::Value *MapTypesArray = nullptr; | |||
10817 | llvm::Value *MapNamesArray = nullptr; | |||
10818 | // Generate the code for the opening of the data environment. | |||
10819 | auto &&ThenGen = [this, &D, Device, &InputInfo, &MapTypesArray, | |||
10820 | &MapNamesArray](CodeGenFunction &CGF, PrePostActionTy &) { | |||
10821 | // Emit device ID if any. | |||
10822 | llvm::Value *DeviceID = nullptr; | |||
10823 | if (Device) { | |||
10824 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
10825 | CGF.Int64Ty, /*isSigned=*/true); | |||
10826 | } else { | |||
10827 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
10828 | } | |||
10829 | ||||
10830 | // Emit the number of elements in the offloading arrays. | |||
10831 | llvm::Constant *PointerNum = | |||
10832 | CGF.Builder.getInt32(InputInfo.NumberOfTargetItems); | |||
10833 | ||||
10834 | // Source location for the ident struct | |||
10835 | llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | |||
10836 | ||||
10837 | llvm::Value *OffloadingArgs[] = {RTLoc, | |||
10838 | DeviceID, | |||
10839 | PointerNum, | |||
10840 | InputInfo.BasePointersArray.getPointer(), | |||
10841 | InputInfo.PointersArray.getPointer(), | |||
10842 | InputInfo.SizesArray.getPointer(), | |||
10843 | MapTypesArray, | |||
10844 | MapNamesArray, | |||
10845 | InputInfo.MappersArray.getPointer()}; | |||
10846 | ||||
10847 | // Select the right runtime function call for each standalone | |||
10848 | // directive. | |||
10849 | const bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>(); | |||
10850 | RuntimeFunction RTLFn; | |||
10851 | switch (D.getDirectiveKind()) { | |||
10852 | case OMPD_target_enter_data: | |||
10853 | RTLFn = HasNowait ? OMPRTL___tgt_target_data_begin_nowait_mapper | |||
10854 | : OMPRTL___tgt_target_data_begin_mapper; | |||
10855 | break; | |||
10856 | case OMPD_target_exit_data: | |||
10857 | RTLFn = HasNowait ? OMPRTL___tgt_target_data_end_nowait_mapper | |||
10858 | : OMPRTL___tgt_target_data_end_mapper; | |||
10859 | break; | |||
10860 | case OMPD_target_update: | |||
10861 | RTLFn = HasNowait ? OMPRTL___tgt_target_data_update_nowait_mapper | |||
10862 | : OMPRTL___tgt_target_data_update_mapper; | |||
10863 | break; | |||
10864 | case OMPD_parallel: | |||
10865 | case OMPD_for: | |||
10866 | case OMPD_parallel_for: | |||
10867 | case OMPD_parallel_master: | |||
10868 | case OMPD_parallel_sections: | |||
10869 | case OMPD_for_simd: | |||
10870 | case OMPD_parallel_for_simd: | |||
10871 | case OMPD_cancel: | |||
10872 | case OMPD_cancellation_point: | |||
10873 | case OMPD_ordered: | |||
10874 | case OMPD_threadprivate: | |||
10875 | case OMPD_allocate: | |||
10876 | case OMPD_task: | |||
10877 | case OMPD_simd: | |||
10878 | case OMPD_tile: | |||
10879 | case OMPD_unroll: | |||
10880 | case OMPD_sections: | |||
10881 | case OMPD_section: | |||
10882 | case OMPD_single: | |||
10883 | case OMPD_master: | |||
10884 | case OMPD_critical: | |||
10885 | case OMPD_taskyield: | |||
10886 | case OMPD_barrier: | |||
10887 | case OMPD_taskwait: | |||
10888 | case OMPD_taskgroup: | |||
10889 | case OMPD_atomic: | |||
10890 | case OMPD_flush: | |||
10891 | case OMPD_depobj: | |||
10892 | case OMPD_scan: | |||
10893 | case OMPD_teams: | |||
10894 | case OMPD_target_data: | |||
10895 | case OMPD_distribute: | |||
10896 | case OMPD_distribute_simd: | |||
10897 | case OMPD_distribute_parallel_for: | |||
10898 | case OMPD_distribute_parallel_for_simd: | |||
10899 | case OMPD_teams_distribute: | |||
10900 | case OMPD_teams_distribute_simd: | |||
10901 | case OMPD_teams_distribute_parallel_for: | |||
10902 | case OMPD_teams_distribute_parallel_for_simd: | |||
10903 | case OMPD_declare_simd: | |||
10904 | case OMPD_declare_variant: | |||
10905 | case OMPD_begin_declare_variant: | |||
10906 | case OMPD_end_declare_variant: | |||
10907 | case OMPD_declare_target: | |||
10908 | case OMPD_end_declare_target: | |||
10909 | case OMPD_declare_reduction: | |||
10910 | case OMPD_declare_mapper: | |||
10911 | case OMPD_taskloop: | |||
10912 | case OMPD_taskloop_simd: | |||
10913 | case OMPD_master_taskloop: | |||
10914 | case OMPD_master_taskloop_simd: | |||
10915 | case OMPD_parallel_master_taskloop: | |||
10916 | case OMPD_parallel_master_taskloop_simd: | |||
10917 | case OMPD_target: | |||
10918 | case OMPD_target_simd: | |||
10919 | case OMPD_target_teams_distribute: | |||
10920 | case OMPD_target_teams_distribute_simd: | |||
10921 | case OMPD_target_teams_distribute_parallel_for: | |||
10922 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
10923 | case OMPD_target_teams: | |||
10924 | case OMPD_target_parallel: | |||
10925 | case OMPD_target_parallel_for: | |||
10926 | case OMPD_target_parallel_for_simd: | |||
10927 | case OMPD_requires: | |||
10928 | case OMPD_metadirective: | |||
10929 | case OMPD_unknown: | |||
10930 | default: | |||
10931 | llvm_unreachable("Unexpected standalone target data directive.")::llvm::llvm_unreachable_internal("Unexpected standalone target data directive." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 10931); | |||
10932 | break; | |||
10933 | } | |||
10934 | CGF.EmitRuntimeCall( | |||
10935 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), RTLFn), | |||
10936 | OffloadingArgs); | |||
10937 | }; | |||
10938 | ||||
10939 | auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray, | |||
10940 | &MapNamesArray](CodeGenFunction &CGF, | |||
10941 | PrePostActionTy &) { | |||
10942 | // Fill up the arrays with all the mapped variables. | |||
10943 | MappableExprsHandler::MapCombinedInfoTy CombinedInfo; | |||
10944 | ||||
10945 | // Get map clause information. | |||
10946 | MappableExprsHandler MEHandler(D, CGF); | |||
10947 | MEHandler.generateAllInfo(CombinedInfo); | |||
10948 | ||||
10949 | CGOpenMPRuntime::TargetDataInfo Info; | |||
10950 | // Fill up the arrays and create the arguments. | |||
10951 | emitOffloadingArrays(CGF, CombinedInfo, Info, OMPBuilder, | |||
10952 | /*IsNonContiguous=*/true); | |||
10953 | bool RequiresOuterTask = D.hasClausesOfKind<OMPDependClause>() || | |||
10954 | D.hasClausesOfKind<OMPNowaitClause>(); | |||
10955 | bool EmitDebug = | |||
10956 | CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo; | |||
10957 | OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, Info.RTArgs, Info, | |||
10958 | EmitDebug, | |||
10959 | /*ForEndCall=*/false); | |||
10960 | InputInfo.NumberOfTargetItems = Info.NumberOfPtrs; | |||
10961 | InputInfo.BasePointersArray = Address(Info.RTArgs.BasePointersArray, | |||
10962 | CGF.VoidPtrTy, CGM.getPointerAlign()); | |||
10963 | InputInfo.PointersArray = Address(Info.RTArgs.PointersArray, CGF.VoidPtrTy, | |||
10964 | CGM.getPointerAlign()); | |||
10965 | InputInfo.SizesArray = | |||
10966 | Address(Info.RTArgs.SizesArray, CGF.Int64Ty, CGM.getPointerAlign()); | |||
10967 | InputInfo.MappersArray = | |||
10968 | Address(Info.RTArgs.MappersArray, CGF.VoidPtrTy, CGM.getPointerAlign()); | |||
10969 | MapTypesArray = Info.RTArgs.MapTypesArray; | |||
10970 | MapNamesArray = Info.RTArgs.MapNamesArray; | |||
10971 | if (RequiresOuterTask) | |||
10972 | CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo); | |||
10973 | else | |||
10974 | emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen); | |||
10975 | }; | |||
10976 | ||||
10977 | if (IfCond) { | |||
10978 | emitIfClause(CGF, IfCond, TargetThenGen, | |||
10979 | [](CodeGenFunction &CGF, PrePostActionTy &) {}); | |||
10980 | } else { | |||
10981 | RegionCodeGenTy ThenRCG(TargetThenGen); | |||
10982 | ThenRCG(CGF); | |||
10983 | } | |||
10984 | } | |||
10985 | ||||
10986 | namespace { | |||
10987 | /// Kind of parameter in a function with 'declare simd' directive. | |||
10988 | enum ParamKindTy { | |||
10989 | Linear, | |||
10990 | LinearRef, | |||
10991 | LinearUVal, | |||
10992 | LinearVal, | |||
10993 | Uniform, | |||
10994 | Vector, | |||
10995 | }; | |||
10996 | /// Attribute set of the parameter. | |||
10997 | struct ParamAttrTy { | |||
10998 | ParamKindTy Kind = Vector; | |||
10999 | llvm::APSInt StrideOrArg; | |||
11000 | llvm::APSInt Alignment; | |||
11001 | bool HasVarStride = false; | |||
11002 | }; | |||
11003 | } // namespace | |||
11004 | ||||
11005 | static unsigned evaluateCDTSize(const FunctionDecl *FD, | |||
11006 | ArrayRef<ParamAttrTy> ParamAttrs) { | |||
11007 | // Every vector variant of a SIMD-enabled function has a vector length (VLEN). | |||
11008 | // If OpenMP clause "simdlen" is used, the VLEN is the value of the argument | |||
11009 | // of that clause. The VLEN value must be power of 2. | |||
11010 | // In other case the notion of the function`s "characteristic data type" (CDT) | |||
11011 | // is used to compute the vector length. | |||
11012 | // CDT is defined in the following order: | |||
11013 | // a) For non-void function, the CDT is the return type. | |||
11014 | // b) If the function has any non-uniform, non-linear parameters, then the | |||
11015 | // CDT is the type of the first such parameter. | |||
11016 | // c) If the CDT determined by a) or b) above is struct, union, or class | |||
11017 | // type which is pass-by-value (except for the type that maps to the | |||
11018 | // built-in complex data type), the characteristic data type is int. | |||
11019 | // d) If none of the above three cases is applicable, the CDT is int. | |||
11020 | // The VLEN is then determined based on the CDT and the size of vector | |||
11021 | // register of that ISA for which current vector version is generated. The | |||
11022 | // VLEN is computed using the formula below: | |||
11023 | // VLEN = sizeof(vector_register) / sizeof(CDT), | |||
11024 | // where vector register size specified in section 3.2.1 Registers and the | |||
11025 | // Stack Frame of original AMD64 ABI document. | |||
11026 | QualType RetType = FD->getReturnType(); | |||
11027 | if (RetType.isNull()) | |||
11028 | return 0; | |||
11029 | ASTContext &C = FD->getASTContext(); | |||
11030 | QualType CDT; | |||
11031 | if (!RetType.isNull() && !RetType->isVoidType()) { | |||
11032 | CDT = RetType; | |||
11033 | } else { | |||
11034 | unsigned Offset = 0; | |||
11035 | if (const auto *MD = dyn_cast<CXXMethodDecl>(FD)) { | |||
11036 | if (ParamAttrs[Offset].Kind == Vector) | |||
11037 | CDT = C.getPointerType(C.getRecordType(MD->getParent())); | |||
11038 | ++Offset; | |||
11039 | } | |||
11040 | if (CDT.isNull()) { | |||
11041 | for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) { | |||
11042 | if (ParamAttrs[I + Offset].Kind == Vector) { | |||
11043 | CDT = FD->getParamDecl(I)->getType(); | |||
11044 | break; | |||
11045 | } | |||
11046 | } | |||
11047 | } | |||
11048 | } | |||
11049 | if (CDT.isNull()) | |||
11050 | CDT = C.IntTy; | |||
11051 | CDT = CDT->getCanonicalTypeUnqualified(); | |||
11052 | if (CDT->isRecordType() || CDT->isUnionType()) | |||
11053 | CDT = C.IntTy; | |||
11054 | return C.getTypeSize(CDT); | |||
11055 | } | |||
11056 | ||||
11057 | /// Mangle the parameter part of the vector function name according to | |||
11058 | /// their OpenMP classification. The mangling function is defined in | |||
11059 | /// section 4.5 of the AAVFABI(2021Q1). | |||
11060 | static std::string mangleVectorParameters(ArrayRef<ParamAttrTy> ParamAttrs) { | |||
11061 | SmallString<256> Buffer; | |||
11062 | llvm::raw_svector_ostream Out(Buffer); | |||
11063 | for (const auto &ParamAttr : ParamAttrs) { | |||
11064 | switch (ParamAttr.Kind) { | |||
11065 | case Linear: | |||
11066 | Out << 'l'; | |||
11067 | break; | |||
11068 | case LinearRef: | |||
11069 | Out << 'R'; | |||
11070 | break; | |||
11071 | case LinearUVal: | |||
11072 | Out << 'U'; | |||
11073 | break; | |||
11074 | case LinearVal: | |||
11075 | Out << 'L'; | |||
11076 | break; | |||
11077 | case Uniform: | |||
11078 | Out << 'u'; | |||
11079 | break; | |||
11080 | case Vector: | |||
11081 | Out << 'v'; | |||
11082 | break; | |||
11083 | } | |||
11084 | if (ParamAttr.HasVarStride) | |||
11085 | Out << "s" << ParamAttr.StrideOrArg; | |||
11086 | else if (ParamAttr.Kind == Linear || ParamAttr.Kind == LinearRef || | |||
11087 | ParamAttr.Kind == LinearUVal || ParamAttr.Kind == LinearVal) { | |||
11088 | // Don't print the step value if it is not present or if it is | |||
11089 | // equal to 1. | |||
11090 | if (ParamAttr.StrideOrArg < 0) | |||
11091 | Out << 'n' << -ParamAttr.StrideOrArg; | |||
11092 | else if (ParamAttr.StrideOrArg != 1) | |||
11093 | Out << ParamAttr.StrideOrArg; | |||
11094 | } | |||
11095 | ||||
11096 | if (!!ParamAttr.Alignment) | |||
11097 | Out << 'a' << ParamAttr.Alignment; | |||
11098 | } | |||
11099 | ||||
11100 | return std::string(Out.str()); | |||
11101 | } | |||
11102 | ||||
11103 | static void | |||
11104 | emitX86DeclareSimdFunction(const FunctionDecl *FD, llvm::Function *Fn, | |||
11105 | const llvm::APSInt &VLENVal, | |||
11106 | ArrayRef<ParamAttrTy> ParamAttrs, | |||
11107 | OMPDeclareSimdDeclAttr::BranchStateTy State) { | |||
11108 | struct ISADataTy { | |||
11109 | char ISA; | |||
11110 | unsigned VecRegSize; | |||
11111 | }; | |||
11112 | ISADataTy ISAData[] = { | |||
11113 | { | |||
11114 | 'b', 128 | |||
11115 | }, // SSE | |||
11116 | { | |||
11117 | 'c', 256 | |||
11118 | }, // AVX | |||
11119 | { | |||
11120 | 'd', 256 | |||
11121 | }, // AVX2 | |||
11122 | { | |||
11123 | 'e', 512 | |||
11124 | }, // AVX512 | |||
11125 | }; | |||
11126 | llvm::SmallVector<char, 2> Masked; | |||
11127 | switch (State) { | |||
11128 | case OMPDeclareSimdDeclAttr::BS_Undefined: | |||
11129 | Masked.push_back('N'); | |||
11130 | Masked.push_back('M'); | |||
11131 | break; | |||
11132 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | |||
11133 | Masked.push_back('N'); | |||
11134 | break; | |||
11135 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | |||
11136 | Masked.push_back('M'); | |||
11137 | break; | |||
11138 | } | |||
11139 | for (char Mask : Masked) { | |||
11140 | for (const ISADataTy &Data : ISAData) { | |||
11141 | SmallString<256> Buffer; | |||
11142 | llvm::raw_svector_ostream Out(Buffer); | |||
11143 | Out << "_ZGV" << Data.ISA << Mask; | |||
11144 | if (!VLENVal) { | |||
11145 | unsigned NumElts = evaluateCDTSize(FD, ParamAttrs); | |||
11146 | assert(NumElts && "Non-zero simdlen/cdtsize expected")(static_cast <bool> (NumElts && "Non-zero simdlen/cdtsize expected" ) ? void (0) : __assert_fail ("NumElts && \"Non-zero simdlen/cdtsize expected\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11146, __extension__ __PRETTY_FUNCTION__)); | |||
11147 | Out << llvm::APSInt::getUnsigned(Data.VecRegSize / NumElts); | |||
11148 | } else { | |||
11149 | Out << VLENVal; | |||
11150 | } | |||
11151 | Out << mangleVectorParameters(ParamAttrs); | |||
11152 | Out << '_' << Fn->getName(); | |||
11153 | Fn->addFnAttr(Out.str()); | |||
11154 | } | |||
11155 | } | |||
11156 | } | |||
11157 | ||||
11158 | // This are the Functions that are needed to mangle the name of the | |||
11159 | // vector functions generated by the compiler, according to the rules | |||
11160 | // defined in the "Vector Function ABI specifications for AArch64", | |||
11161 | // available at | |||
11162 | // https://developer.arm.com/products/software-development-tools/hpc/arm-compiler-for-hpc/vector-function-abi. | |||
11163 | ||||
11164 | /// Maps To Vector (MTV), as defined in 4.1.1 of the AAVFABI (2021Q1). | |||
11165 | static bool getAArch64MTV(QualType QT, ParamKindTy Kind) { | |||
11166 | QT = QT.getCanonicalType(); | |||
11167 | ||||
11168 | if (QT->isVoidType()) | |||
11169 | return false; | |||
11170 | ||||
11171 | if (Kind == ParamKindTy::Uniform) | |||
11172 | return false; | |||
11173 | ||||
11174 | if (Kind == ParamKindTy::LinearUVal || ParamKindTy::LinearRef) | |||
11175 | return false; | |||
11176 | ||||
11177 | if ((Kind == ParamKindTy::Linear || Kind == ParamKindTy::LinearVal) && | |||
11178 | !QT->isReferenceType()) | |||
11179 | return false; | |||
11180 | ||||
11181 | return true; | |||
11182 | } | |||
11183 | ||||
11184 | /// Pass By Value (PBV), as defined in 3.1.2 of the AAVFABI. | |||
11185 | static bool getAArch64PBV(QualType QT, ASTContext &C) { | |||
11186 | QT = QT.getCanonicalType(); | |||
11187 | unsigned Size = C.getTypeSize(QT); | |||
11188 | ||||
11189 | // Only scalars and complex within 16 bytes wide set PVB to true. | |||
11190 | if (Size != 8 && Size != 16 && Size != 32 && Size != 64 && Size != 128) | |||
11191 | return false; | |||
11192 | ||||
11193 | if (QT->isFloatingType()) | |||
11194 | return true; | |||
11195 | ||||
11196 | if (QT->isIntegerType()) | |||
11197 | return true; | |||
11198 | ||||
11199 | if (QT->isPointerType()) | |||
11200 | return true; | |||
11201 | ||||
11202 | // TODO: Add support for complex types (section 3.1.2, item 2). | |||
11203 | ||||
11204 | return false; | |||
11205 | } | |||
11206 | ||||
11207 | /// Computes the lane size (LS) of a return type or of an input parameter, | |||
11208 | /// as defined by `LS(P)` in 3.2.1 of the AAVFABI. | |||
11209 | /// TODO: Add support for references, section 3.2.1, item 1. | |||
11210 | static unsigned getAArch64LS(QualType QT, ParamKindTy Kind, ASTContext &C) { | |||
11211 | if (!getAArch64MTV(QT, Kind) && QT.getCanonicalType()->isPointerType()) { | |||
11212 | QualType PTy = QT.getCanonicalType()->getPointeeType(); | |||
11213 | if (getAArch64PBV(PTy, C)) | |||
11214 | return C.getTypeSize(PTy); | |||
11215 | } | |||
11216 | if (getAArch64PBV(QT, C)) | |||
11217 | return C.getTypeSize(QT); | |||
11218 | ||||
11219 | return C.getTypeSize(C.getUIntPtrType()); | |||
11220 | } | |||
11221 | ||||
11222 | // Get Narrowest Data Size (NDS) and Widest Data Size (WDS) from the | |||
11223 | // signature of the scalar function, as defined in 3.2.2 of the | |||
11224 | // AAVFABI. | |||
11225 | static std::tuple<unsigned, unsigned, bool> | |||
11226 | getNDSWDS(const FunctionDecl *FD, ArrayRef<ParamAttrTy> ParamAttrs) { | |||
11227 | QualType RetType = FD->getReturnType().getCanonicalType(); | |||
11228 | ||||
11229 | ASTContext &C = FD->getASTContext(); | |||
11230 | ||||
11231 | bool OutputBecomesInput = false; | |||
11232 | ||||
11233 | llvm::SmallVector<unsigned, 8> Sizes; | |||
11234 | if (!RetType->isVoidType()) { | |||
11235 | Sizes.push_back(getAArch64LS(RetType, ParamKindTy::Vector, C)); | |||
11236 | if (!getAArch64PBV(RetType, C) && getAArch64MTV(RetType, {})) | |||
11237 | OutputBecomesInput = true; | |||
11238 | } | |||
11239 | for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) { | |||
11240 | QualType QT = FD->getParamDecl(I)->getType().getCanonicalType(); | |||
11241 | Sizes.push_back(getAArch64LS(QT, ParamAttrs[I].Kind, C)); | |||
11242 | } | |||
11243 | ||||
11244 | assert(!Sizes.empty() && "Unable to determine NDS and WDS.")(static_cast <bool> (!Sizes.empty() && "Unable to determine NDS and WDS." ) ? void (0) : __assert_fail ("!Sizes.empty() && \"Unable to determine NDS and WDS.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11244, __extension__ __PRETTY_FUNCTION__)); | |||
11245 | // The LS of a function parameter / return value can only be a power | |||
11246 | // of 2, starting from 8 bits, up to 128. | |||
11247 | assert(llvm::all_of(Sizes,(static_cast <bool> (llvm::all_of(Sizes, [](unsigned Size ) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? void (0) : __assert_fail ("llvm::all_of(Sizes, [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11252, __extension__ __PRETTY_FUNCTION__)) | |||
11248 | [](unsigned Size) {(static_cast <bool> (llvm::all_of(Sizes, [](unsigned Size ) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? void (0) : __assert_fail ("llvm::all_of(Sizes, [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11252, __extension__ __PRETTY_FUNCTION__)) | |||
11249 | return Size == 8 || Size == 16 || Size == 32 ||(static_cast <bool> (llvm::all_of(Sizes, [](unsigned Size ) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? void (0) : __assert_fail ("llvm::all_of(Sizes, [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11252, __extension__ __PRETTY_FUNCTION__)) | |||
11250 | Size == 64 || Size == 128;(static_cast <bool> (llvm::all_of(Sizes, [](unsigned Size ) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? void (0) : __assert_fail ("llvm::all_of(Sizes, [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11252, __extension__ __PRETTY_FUNCTION__)) | |||
11251 | }) &&(static_cast <bool> (llvm::all_of(Sizes, [](unsigned Size ) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? void (0) : __assert_fail ("llvm::all_of(Sizes, [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11252, __extension__ __PRETTY_FUNCTION__)) | |||
11252 | "Invalid size")(static_cast <bool> (llvm::all_of(Sizes, [](unsigned Size ) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? void (0) : __assert_fail ("llvm::all_of(Sizes, [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11252, __extension__ __PRETTY_FUNCTION__)); | |||
11253 | ||||
11254 | return std::make_tuple(*std::min_element(std::begin(Sizes), std::end(Sizes)), | |||
11255 | *std::max_element(std::begin(Sizes), std::end(Sizes)), | |||
11256 | OutputBecomesInput); | |||
11257 | } | |||
11258 | ||||
11259 | // Function used to add the attribute. The parameter `VLEN` is | |||
11260 | // templated to allow the use of "x" when targeting scalable functions | |||
11261 | // for SVE. | |||
11262 | template <typename T> | |||
11263 | static void addAArch64VectorName(T VLEN, StringRef LMask, StringRef Prefix, | |||
11264 | char ISA, StringRef ParSeq, | |||
11265 | StringRef MangledName, bool OutputBecomesInput, | |||
11266 | llvm::Function *Fn) { | |||
11267 | SmallString<256> Buffer; | |||
11268 | llvm::raw_svector_ostream Out(Buffer); | |||
11269 | Out << Prefix << ISA << LMask << VLEN; | |||
11270 | if (OutputBecomesInput) | |||
11271 | Out << "v"; | |||
11272 | Out << ParSeq << "_" << MangledName; | |||
11273 | Fn->addFnAttr(Out.str()); | |||
11274 | } | |||
11275 | ||||
11276 | // Helper function to generate the Advanced SIMD names depending on | |||
11277 | // the value of the NDS when simdlen is not present. | |||
11278 | static void addAArch64AdvSIMDNDSNames(unsigned NDS, StringRef Mask, | |||
11279 | StringRef Prefix, char ISA, | |||
11280 | StringRef ParSeq, StringRef MangledName, | |||
11281 | bool OutputBecomesInput, | |||
11282 | llvm::Function *Fn) { | |||
11283 | switch (NDS) { | |||
11284 | case 8: | |||
11285 | addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11286 | OutputBecomesInput, Fn); | |||
11287 | addAArch64VectorName(16, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11288 | OutputBecomesInput, Fn); | |||
11289 | break; | |||
11290 | case 16: | |||
11291 | addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11292 | OutputBecomesInput, Fn); | |||
11293 | addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11294 | OutputBecomesInput, Fn); | |||
11295 | break; | |||
11296 | case 32: | |||
11297 | addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11298 | OutputBecomesInput, Fn); | |||
11299 | addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11300 | OutputBecomesInput, Fn); | |||
11301 | break; | |||
11302 | case 64: | |||
11303 | case 128: | |||
11304 | addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName, | |||
11305 | OutputBecomesInput, Fn); | |||
11306 | break; | |||
11307 | default: | |||
11308 | llvm_unreachable("Scalar type is too wide.")::llvm::llvm_unreachable_internal("Scalar type is too wide.", "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11308); | |||
11309 | } | |||
11310 | } | |||
11311 | ||||
11312 | /// Emit vector function attributes for AArch64, as defined in the AAVFABI. | |||
11313 | static void emitAArch64DeclareSimdFunction( | |||
11314 | CodeGenModule &CGM, const FunctionDecl *FD, unsigned UserVLEN, | |||
11315 | ArrayRef<ParamAttrTy> ParamAttrs, | |||
11316 | OMPDeclareSimdDeclAttr::BranchStateTy State, StringRef MangledName, | |||
11317 | char ISA, unsigned VecRegSize, llvm::Function *Fn, SourceLocation SLoc) { | |||
11318 | ||||
11319 | // Get basic data for building the vector signature. | |||
11320 | const auto Data = getNDSWDS(FD, ParamAttrs); | |||
11321 | const unsigned NDS = std::get<0>(Data); | |||
11322 | const unsigned WDS = std::get<1>(Data); | |||
11323 | const bool OutputBecomesInput = std::get<2>(Data); | |||
11324 | ||||
11325 | // Check the values provided via `simdlen` by the user. | |||
11326 | // 1. A `simdlen(1)` doesn't produce vector signatures, | |||
11327 | if (UserVLEN == 1) { | |||
11328 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
11329 | DiagnosticsEngine::Warning, | |||
11330 | "The clause simdlen(1) has no effect when targeting aarch64."); | |||
11331 | CGM.getDiags().Report(SLoc, DiagID); | |||
11332 | return; | |||
11333 | } | |||
11334 | ||||
11335 | // 2. Section 3.3.1, item 1: user input must be a power of 2 for | |||
11336 | // Advanced SIMD output. | |||
11337 | if (ISA == 'n' && UserVLEN && !llvm::isPowerOf2_32(UserVLEN)) { | |||
11338 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
11339 | DiagnosticsEngine::Warning, "The value specified in simdlen must be a " | |||
11340 | "power of 2 when targeting Advanced SIMD."); | |||
11341 | CGM.getDiags().Report(SLoc, DiagID); | |||
11342 | return; | |||
11343 | } | |||
11344 | ||||
11345 | // 3. Section 3.4.1. SVE fixed lengh must obey the architectural | |||
11346 | // limits. | |||
11347 | if (ISA == 's' && UserVLEN != 0) { | |||
11348 | if ((UserVLEN * WDS > 2048) || (UserVLEN * WDS % 128 != 0)) { | |||
11349 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
11350 | DiagnosticsEngine::Warning, "The clause simdlen must fit the %0-bit " | |||
11351 | "lanes in the architectural constraints " | |||
11352 | "for SVE (min is 128-bit, max is " | |||
11353 | "2048-bit, by steps of 128-bit)"); | |||
11354 | CGM.getDiags().Report(SLoc, DiagID) << WDS; | |||
11355 | return; | |||
11356 | } | |||
11357 | } | |||
11358 | ||||
11359 | // Sort out parameter sequence. | |||
11360 | const std::string ParSeq = mangleVectorParameters(ParamAttrs); | |||
11361 | StringRef Prefix = "_ZGV"; | |||
11362 | // Generate simdlen from user input (if any). | |||
11363 | if (UserVLEN) { | |||
11364 | if (ISA == 's') { | |||
11365 | // SVE generates only a masked function. | |||
11366 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | |||
11367 | OutputBecomesInput, Fn); | |||
11368 | } else { | |||
11369 | assert(ISA == 'n' && "Expected ISA either 's' or 'n'.")(static_cast <bool> (ISA == 'n' && "Expected ISA either 's' or 'n'." ) ? void (0) : __assert_fail ("ISA == 'n' && \"Expected ISA either 's' or 'n'.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11369, __extension__ __PRETTY_FUNCTION__)); | |||
11370 | // Advanced SIMD generates one or two functions, depending on | |||
11371 | // the `[not]inbranch` clause. | |||
11372 | switch (State) { | |||
11373 | case OMPDeclareSimdDeclAttr::BS_Undefined: | |||
11374 | addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName, | |||
11375 | OutputBecomesInput, Fn); | |||
11376 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | |||
11377 | OutputBecomesInput, Fn); | |||
11378 | break; | |||
11379 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | |||
11380 | addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName, | |||
11381 | OutputBecomesInput, Fn); | |||
11382 | break; | |||
11383 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | |||
11384 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | |||
11385 | OutputBecomesInput, Fn); | |||
11386 | break; | |||
11387 | } | |||
11388 | } | |||
11389 | } else { | |||
11390 | // If no user simdlen is provided, follow the AAVFABI rules for | |||
11391 | // generating the vector length. | |||
11392 | if (ISA == 's') { | |||
11393 | // SVE, section 3.4.1, item 1. | |||
11394 | addAArch64VectorName("x", "M", Prefix, ISA, ParSeq, MangledName, | |||
11395 | OutputBecomesInput, Fn); | |||
11396 | } else { | |||
11397 | assert(ISA == 'n' && "Expected ISA either 's' or 'n'.")(static_cast <bool> (ISA == 'n' && "Expected ISA either 's' or 'n'." ) ? void (0) : __assert_fail ("ISA == 'n' && \"Expected ISA either 's' or 'n'.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11397, __extension__ __PRETTY_FUNCTION__)); | |||
11398 | // Advanced SIMD, Section 3.3.1 of the AAVFABI, generates one or | |||
11399 | // two vector names depending on the use of the clause | |||
11400 | // `[not]inbranch`. | |||
11401 | switch (State) { | |||
11402 | case OMPDeclareSimdDeclAttr::BS_Undefined: | |||
11403 | addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName, | |||
11404 | OutputBecomesInput, Fn); | |||
11405 | addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName, | |||
11406 | OutputBecomesInput, Fn); | |||
11407 | break; | |||
11408 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | |||
11409 | addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName, | |||
11410 | OutputBecomesInput, Fn); | |||
11411 | break; | |||
11412 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | |||
11413 | addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName, | |||
11414 | OutputBecomesInput, Fn); | |||
11415 | break; | |||
11416 | } | |||
11417 | } | |||
11418 | } | |||
11419 | } | |||
11420 | ||||
11421 | void CGOpenMPRuntime::emitDeclareSimdFunction(const FunctionDecl *FD, | |||
11422 | llvm::Function *Fn) { | |||
11423 | ASTContext &C = CGM.getContext(); | |||
11424 | FD = FD->getMostRecentDecl(); | |||
11425 | while (FD) { | |||
11426 | // Map params to their positions in function decl. | |||
11427 | llvm::DenseMap<const Decl *, unsigned> ParamPositions; | |||
11428 | if (isa<CXXMethodDecl>(FD)) | |||
11429 | ParamPositions.try_emplace(FD, 0); | |||
11430 | unsigned ParamPos = ParamPositions.size(); | |||
11431 | for (const ParmVarDecl *P : FD->parameters()) { | |||
11432 | ParamPositions.try_emplace(P->getCanonicalDecl(), ParamPos); | |||
11433 | ++ParamPos; | |||
11434 | } | |||
11435 | for (const auto *Attr : FD->specific_attrs<OMPDeclareSimdDeclAttr>()) { | |||
11436 | llvm::SmallVector<ParamAttrTy, 8> ParamAttrs(ParamPositions.size()); | |||
11437 | // Mark uniform parameters. | |||
11438 | for (const Expr *E : Attr->uniforms()) { | |||
11439 | E = E->IgnoreParenImpCasts(); | |||
11440 | unsigned Pos; | |||
11441 | if (isa<CXXThisExpr>(E)) { | |||
11442 | Pos = ParamPositions[FD]; | |||
11443 | } else { | |||
11444 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | |||
11445 | ->getCanonicalDecl(); | |||
11446 | auto It = ParamPositions.find(PVD); | |||
11447 | assert(It != ParamPositions.end() && "Function parameter not found")(static_cast <bool> (It != ParamPositions.end() && "Function parameter not found") ? void (0) : __assert_fail ( "It != ParamPositions.end() && \"Function parameter not found\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11447, __extension__ __PRETTY_FUNCTION__)); | |||
11448 | Pos = It->second; | |||
11449 | } | |||
11450 | ParamAttrs[Pos].Kind = Uniform; | |||
11451 | } | |||
11452 | // Get alignment info. | |||
11453 | auto *NI = Attr->alignments_begin(); | |||
11454 | for (const Expr *E : Attr->aligneds()) { | |||
11455 | E = E->IgnoreParenImpCasts(); | |||
11456 | unsigned Pos; | |||
11457 | QualType ParmTy; | |||
11458 | if (isa<CXXThisExpr>(E)) { | |||
11459 | Pos = ParamPositions[FD]; | |||
11460 | ParmTy = E->getType(); | |||
11461 | } else { | |||
11462 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | |||
11463 | ->getCanonicalDecl(); | |||
11464 | auto It = ParamPositions.find(PVD); | |||
11465 | assert(It != ParamPositions.end() && "Function parameter not found")(static_cast <bool> (It != ParamPositions.end() && "Function parameter not found") ? void (0) : __assert_fail ( "It != ParamPositions.end() && \"Function parameter not found\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11465, __extension__ __PRETTY_FUNCTION__)); | |||
11466 | Pos = It->second; | |||
11467 | ParmTy = PVD->getType(); | |||
11468 | } | |||
11469 | ParamAttrs[Pos].Alignment = | |||
11470 | (*NI) | |||
11471 | ? (*NI)->EvaluateKnownConstInt(C) | |||
11472 | : llvm::APSInt::getUnsigned( | |||
11473 | C.toCharUnitsFromBits(C.getOpenMPDefaultSimdAlign(ParmTy)) | |||
11474 | .getQuantity()); | |||
11475 | ++NI; | |||
11476 | } | |||
11477 | // Mark linear parameters. | |||
11478 | auto *SI = Attr->steps_begin(); | |||
11479 | auto *MI = Attr->modifiers_begin(); | |||
11480 | for (const Expr *E : Attr->linears()) { | |||
11481 | E = E->IgnoreParenImpCasts(); | |||
11482 | unsigned Pos; | |||
11483 | bool IsReferenceType = false; | |||
11484 | // Rescaling factor needed to compute the linear parameter | |||
11485 | // value in the mangled name. | |||
11486 | unsigned PtrRescalingFactor = 1; | |||
11487 | if (isa<CXXThisExpr>(E)) { | |||
11488 | Pos = ParamPositions[FD]; | |||
11489 | auto *P = cast<PointerType>(E->getType()); | |||
11490 | PtrRescalingFactor = CGM.getContext() | |||
11491 | .getTypeSizeInChars(P->getPointeeType()) | |||
11492 | .getQuantity(); | |||
11493 | } else { | |||
11494 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | |||
11495 | ->getCanonicalDecl(); | |||
11496 | auto It = ParamPositions.find(PVD); | |||
11497 | assert(It != ParamPositions.end() && "Function parameter not found")(static_cast <bool> (It != ParamPositions.end() && "Function parameter not found") ? void (0) : __assert_fail ( "It != ParamPositions.end() && \"Function parameter not found\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11497, __extension__ __PRETTY_FUNCTION__)); | |||
11498 | Pos = It->second; | |||
11499 | if (auto *P = dyn_cast<PointerType>(PVD->getType())) | |||
11500 | PtrRescalingFactor = CGM.getContext() | |||
11501 | .getTypeSizeInChars(P->getPointeeType()) | |||
11502 | .getQuantity(); | |||
11503 | else if (PVD->getType()->isReferenceType()) { | |||
11504 | IsReferenceType = true; | |||
11505 | PtrRescalingFactor = | |||
11506 | CGM.getContext() | |||
11507 | .getTypeSizeInChars(PVD->getType().getNonReferenceType()) | |||
11508 | .getQuantity(); | |||
11509 | } | |||
11510 | } | |||
11511 | ParamAttrTy &ParamAttr = ParamAttrs[Pos]; | |||
11512 | if (*MI == OMPC_LINEAR_ref) | |||
11513 | ParamAttr.Kind = LinearRef; | |||
11514 | else if (*MI == OMPC_LINEAR_uval) | |||
11515 | ParamAttr.Kind = LinearUVal; | |||
11516 | else if (IsReferenceType) | |||
11517 | ParamAttr.Kind = LinearVal; | |||
11518 | else | |||
11519 | ParamAttr.Kind = Linear; | |||
11520 | // Assuming a stride of 1, for `linear` without modifiers. | |||
11521 | ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(1); | |||
11522 | if (*SI) { | |||
11523 | Expr::EvalResult Result; | |||
11524 | if (!(*SI)->EvaluateAsInt(Result, C, Expr::SE_AllowSideEffects)) { | |||
11525 | if (const auto *DRE = | |||
11526 | cast<DeclRefExpr>((*SI)->IgnoreParenImpCasts())) { | |||
11527 | if (const auto *StridePVD = | |||
11528 | dyn_cast<ParmVarDecl>(DRE->getDecl())) { | |||
11529 | ParamAttr.HasVarStride = true; | |||
11530 | auto It = ParamPositions.find(StridePVD->getCanonicalDecl()); | |||
11531 | assert(It != ParamPositions.end() &&(static_cast <bool> (It != ParamPositions.end() && "Function parameter not found") ? void (0) : __assert_fail ( "It != ParamPositions.end() && \"Function parameter not found\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11532, __extension__ __PRETTY_FUNCTION__)) | |||
11532 | "Function parameter not found")(static_cast <bool> (It != ParamPositions.end() && "Function parameter not found") ? void (0) : __assert_fail ( "It != ParamPositions.end() && \"Function parameter not found\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11532, __extension__ __PRETTY_FUNCTION__)); | |||
11533 | ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(It->second); | |||
11534 | } | |||
11535 | } | |||
11536 | } else { | |||
11537 | ParamAttr.StrideOrArg = Result.Val.getInt(); | |||
11538 | } | |||
11539 | } | |||
11540 | // If we are using a linear clause on a pointer, we need to | |||
11541 | // rescale the value of linear_step with the byte size of the | |||
11542 | // pointee type. | |||
11543 | if (!ParamAttr.HasVarStride && | |||
11544 | (ParamAttr.Kind == Linear || ParamAttr.Kind == LinearRef)) | |||
11545 | ParamAttr.StrideOrArg = ParamAttr.StrideOrArg * PtrRescalingFactor; | |||
11546 | ++SI; | |||
11547 | ++MI; | |||
11548 | } | |||
11549 | llvm::APSInt VLENVal; | |||
11550 | SourceLocation ExprLoc; | |||
11551 | const Expr *VLENExpr = Attr->getSimdlen(); | |||
11552 | if (VLENExpr) { | |||
11553 | VLENVal = VLENExpr->EvaluateKnownConstInt(C); | |||
11554 | ExprLoc = VLENExpr->getExprLoc(); | |||
11555 | } | |||
11556 | OMPDeclareSimdDeclAttr::BranchStateTy State = Attr->getBranchState(); | |||
11557 | if (CGM.getTriple().isX86()) { | |||
11558 | emitX86DeclareSimdFunction(FD, Fn, VLENVal, ParamAttrs, State); | |||
11559 | } else if (CGM.getTriple().getArch() == llvm::Triple::aarch64) { | |||
11560 | unsigned VLEN = VLENVal.getExtValue(); | |||
11561 | StringRef MangledName = Fn->getName(); | |||
11562 | if (CGM.getTarget().hasFeature("sve")) | |||
11563 | emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State, | |||
11564 | MangledName, 's', 128, Fn, ExprLoc); | |||
11565 | if (CGM.getTarget().hasFeature("neon")) | |||
11566 | emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State, | |||
11567 | MangledName, 'n', 128, Fn, ExprLoc); | |||
11568 | } | |||
11569 | } | |||
11570 | FD = FD->getPreviousDecl(); | |||
11571 | } | |||
11572 | } | |||
11573 | ||||
11574 | namespace { | |||
11575 | /// Cleanup action for doacross support. | |||
11576 | class DoacrossCleanupTy final : public EHScopeStack::Cleanup { | |||
11577 | public: | |||
11578 | static const int DoacrossFinArgs = 2; | |||
11579 | ||||
11580 | private: | |||
11581 | llvm::FunctionCallee RTLFn; | |||
11582 | llvm::Value *Args[DoacrossFinArgs]; | |||
11583 | ||||
11584 | public: | |||
11585 | DoacrossCleanupTy(llvm::FunctionCallee RTLFn, | |||
11586 | ArrayRef<llvm::Value *> CallArgs) | |||
11587 | : RTLFn(RTLFn) { | |||
11588 | assert(CallArgs.size() == DoacrossFinArgs)(static_cast <bool> (CallArgs.size() == DoacrossFinArgs ) ? void (0) : __assert_fail ("CallArgs.size() == DoacrossFinArgs" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11588, __extension__ __PRETTY_FUNCTION__)); | |||
11589 | std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args)); | |||
11590 | } | |||
11591 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | |||
11592 | if (!CGF.HaveInsertPoint()) | |||
11593 | return; | |||
11594 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
11595 | } | |||
11596 | }; | |||
11597 | } // namespace | |||
11598 | ||||
11599 | void CGOpenMPRuntime::emitDoacrossInit(CodeGenFunction &CGF, | |||
11600 | const OMPLoopDirective &D, | |||
11601 | ArrayRef<Expr *> NumIterations) { | |||
11602 | if (!CGF.HaveInsertPoint()) | |||
11603 | return; | |||
11604 | ||||
11605 | ASTContext &C = CGM.getContext(); | |||
11606 | QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true); | |||
11607 | RecordDecl *RD; | |||
11608 | if (KmpDimTy.isNull()) { | |||
11609 | // Build struct kmp_dim { // loop bounds info casted to kmp_int64 | |||
11610 | // kmp_int64 lo; // lower | |||
11611 | // kmp_int64 up; // upper | |||
11612 | // kmp_int64 st; // stride | |||
11613 | // }; | |||
11614 | RD = C.buildImplicitRecord("kmp_dim"); | |||
11615 | RD->startDefinition(); | |||
11616 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
11617 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
11618 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
11619 | RD->completeDefinition(); | |||
11620 | KmpDimTy = C.getRecordType(RD); | |||
11621 | } else { | |||
11622 | RD = cast<RecordDecl>(KmpDimTy->getAsTagDecl()); | |||
11623 | } | |||
11624 | llvm::APInt Size(/*numBits=*/32, NumIterations.size()); | |||
11625 | QualType ArrayTy = | |||
11626 | C.getConstantArrayType(KmpDimTy, Size, nullptr, ArrayType::Normal, 0); | |||
11627 | ||||
11628 | Address DimsAddr = CGF.CreateMemTemp(ArrayTy, "dims"); | |||
11629 | CGF.EmitNullInitialization(DimsAddr, ArrayTy); | |||
11630 | enum { LowerFD = 0, UpperFD, StrideFD }; | |||
11631 | // Fill dims with data. | |||
11632 | for (unsigned I = 0, E = NumIterations.size(); I < E; ++I) { | |||
11633 | LValue DimsLVal = CGF.MakeAddrLValue( | |||
11634 | CGF.Builder.CreateConstArrayGEP(DimsAddr, I), KmpDimTy); | |||
11635 | // dims.upper = num_iterations; | |||
11636 | LValue UpperLVal = CGF.EmitLValueForField( | |||
11637 | DimsLVal, *std::next(RD->field_begin(), UpperFD)); | |||
11638 | llvm::Value *NumIterVal = CGF.EmitScalarConversion( | |||
11639 | CGF.EmitScalarExpr(NumIterations[I]), NumIterations[I]->getType(), | |||
11640 | Int64Ty, NumIterations[I]->getExprLoc()); | |||
11641 | CGF.EmitStoreOfScalar(NumIterVal, UpperLVal); | |||
11642 | // dims.stride = 1; | |||
11643 | LValue StrideLVal = CGF.EmitLValueForField( | |||
11644 | DimsLVal, *std::next(RD->field_begin(), StrideFD)); | |||
11645 | CGF.EmitStoreOfScalar(llvm::ConstantInt::getSigned(CGM.Int64Ty, /*V=*/1), | |||
11646 | StrideLVal); | |||
11647 | } | |||
11648 | ||||
11649 | // Build call void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, | |||
11650 | // kmp_int32 num_dims, struct kmp_dim * dims); | |||
11651 | llvm::Value *Args[] = { | |||
11652 | emitUpdateLocation(CGF, D.getBeginLoc()), | |||
11653 | getThreadID(CGF, D.getBeginLoc()), | |||
11654 | llvm::ConstantInt::getSigned(CGM.Int32Ty, NumIterations.size()), | |||
11655 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
11656 | CGF.Builder.CreateConstArrayGEP(DimsAddr, 0).getPointer(), | |||
11657 | CGM.VoidPtrTy)}; | |||
11658 | ||||
11659 | llvm::FunctionCallee RTLFn = OMPBuilder.getOrCreateRuntimeFunction( | |||
11660 | CGM.getModule(), OMPRTL___kmpc_doacross_init); | |||
11661 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
11662 | llvm::Value *FiniArgs[DoacrossCleanupTy::DoacrossFinArgs] = { | |||
11663 | emitUpdateLocation(CGF, D.getEndLoc()), getThreadID(CGF, D.getEndLoc())}; | |||
11664 | llvm::FunctionCallee FiniRTLFn = OMPBuilder.getOrCreateRuntimeFunction( | |||
11665 | CGM.getModule(), OMPRTL___kmpc_doacross_fini); | |||
11666 | CGF.EHStack.pushCleanup<DoacrossCleanupTy>(NormalAndEHCleanup, FiniRTLFn, | |||
11667 | llvm::makeArrayRef(FiniArgs)); | |||
11668 | } | |||
11669 | ||||
11670 | void CGOpenMPRuntime::emitDoacrossOrdered(CodeGenFunction &CGF, | |||
11671 | const OMPDependClause *C) { | |||
11672 | QualType Int64Ty = | |||
11673 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | |||
11674 | llvm::APInt Size(/*numBits=*/32, C->getNumLoops()); | |||
11675 | QualType ArrayTy = CGM.getContext().getConstantArrayType( | |||
11676 | Int64Ty, Size, nullptr, ArrayType::Normal, 0); | |||
11677 | Address CntAddr = CGF.CreateMemTemp(ArrayTy, ".cnt.addr"); | |||
11678 | for (unsigned I = 0, E = C->getNumLoops(); I < E; ++I) { | |||
11679 | const Expr *CounterVal = C->getLoopData(I); | |||
11680 | assert(CounterVal)(static_cast <bool> (CounterVal) ? void (0) : __assert_fail ("CounterVal", "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11680 , __extension__ __PRETTY_FUNCTION__)); | |||
11681 | llvm::Value *CntVal = CGF.EmitScalarConversion( | |||
11682 | CGF.EmitScalarExpr(CounterVal), CounterVal->getType(), Int64Ty, | |||
11683 | CounterVal->getExprLoc()); | |||
11684 | CGF.EmitStoreOfScalar(CntVal, CGF.Builder.CreateConstArrayGEP(CntAddr, I), | |||
11685 | /*Volatile=*/false, Int64Ty); | |||
11686 | } | |||
11687 | llvm::Value *Args[] = { | |||
11688 | emitUpdateLocation(CGF, C->getBeginLoc()), | |||
11689 | getThreadID(CGF, C->getBeginLoc()), | |||
11690 | CGF.Builder.CreateConstArrayGEP(CntAddr, 0).getPointer()}; | |||
11691 | llvm::FunctionCallee RTLFn; | |||
11692 | if (C->getDependencyKind() == OMPC_DEPEND_source) { | |||
11693 | RTLFn = OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
11694 | OMPRTL___kmpc_doacross_post); | |||
11695 | } else { | |||
11696 | assert(C->getDependencyKind() == OMPC_DEPEND_sink)(static_cast <bool> (C->getDependencyKind() == OMPC_DEPEND_sink ) ? void (0) : __assert_fail ("C->getDependencyKind() == OMPC_DEPEND_sink" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11696, __extension__ __PRETTY_FUNCTION__)); | |||
11697 | RTLFn = OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | |||
11698 | OMPRTL___kmpc_doacross_wait); | |||
11699 | } | |||
11700 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
11701 | } | |||
11702 | ||||
11703 | void CGOpenMPRuntime::emitCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
11704 | llvm::FunctionCallee Callee, | |||
11705 | ArrayRef<llvm::Value *> Args) const { | |||
11706 | assert(Loc.isValid() && "Outlined function call location must be valid.")(static_cast <bool> (Loc.isValid() && "Outlined function call location must be valid." ) ? void (0) : __assert_fail ("Loc.isValid() && \"Outlined function call location must be valid.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11706, __extension__ __PRETTY_FUNCTION__)); | |||
11707 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | |||
11708 | ||||
11709 | if (auto *Fn = dyn_cast<llvm::Function>(Callee.getCallee())) { | |||
11710 | if (Fn->doesNotThrow()) { | |||
11711 | CGF.EmitNounwindRuntimeCall(Fn, Args); | |||
11712 | return; | |||
11713 | } | |||
11714 | } | |||
11715 | CGF.EmitRuntimeCall(Callee, Args); | |||
11716 | } | |||
11717 | ||||
11718 | void CGOpenMPRuntime::emitOutlinedFunctionCall( | |||
11719 | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, | |||
11720 | ArrayRef<llvm::Value *> Args) const { | |||
11721 | emitCall(CGF, Loc, OutlinedFn, Args); | |||
11722 | } | |||
11723 | ||||
11724 | void CGOpenMPRuntime::emitFunctionProlog(CodeGenFunction &CGF, const Decl *D) { | |||
11725 | if (const auto *FD = dyn_cast<FunctionDecl>(D)) | |||
11726 | if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(FD)) | |||
11727 | HasEmittedDeclareTargetRegion = true; | |||
11728 | } | |||
11729 | ||||
11730 | Address CGOpenMPRuntime::getParameterAddress(CodeGenFunction &CGF, | |||
11731 | const VarDecl *NativeParam, | |||
11732 | const VarDecl *TargetParam) const { | |||
11733 | return CGF.GetAddrOfLocalVar(NativeParam); | |||
11734 | } | |||
11735 | ||||
11736 | /// Return allocator value from expression, or return a null allocator (default | |||
11737 | /// when no allocator specified). | |||
11738 | static llvm::Value *getAllocatorVal(CodeGenFunction &CGF, | |||
11739 | const Expr *Allocator) { | |||
11740 | llvm::Value *AllocVal; | |||
11741 | if (Allocator) { | |||
11742 | AllocVal = CGF.EmitScalarExpr(Allocator); | |||
11743 | // According to the standard, the original allocator type is a enum | |||
11744 | // (integer). Convert to pointer type, if required. | |||
11745 | AllocVal = CGF.EmitScalarConversion(AllocVal, Allocator->getType(), | |||
11746 | CGF.getContext().VoidPtrTy, | |||
11747 | Allocator->getExprLoc()); | |||
11748 | } else { | |||
11749 | // If no allocator specified, it defaults to the null allocator. | |||
11750 | AllocVal = llvm::Constant::getNullValue( | |||
11751 | CGF.CGM.getTypes().ConvertType(CGF.getContext().VoidPtrTy)); | |||
11752 | } | |||
11753 | return AllocVal; | |||
11754 | } | |||
11755 | ||||
11756 | /// Return the alignment from an allocate directive if present. | |||
11757 | static llvm::Value *getAlignmentValue(CodeGenModule &CGM, const VarDecl *VD) { | |||
11758 | llvm::Optional<CharUnits> AllocateAlignment = CGM.getOMPAllocateAlignment(VD); | |||
11759 | ||||
11760 | if (!AllocateAlignment) | |||
11761 | return nullptr; | |||
11762 | ||||
11763 | return llvm::ConstantInt::get(CGM.SizeTy, AllocateAlignment->getQuantity()); | |||
11764 | } | |||
11765 | ||||
11766 | Address CGOpenMPRuntime::getAddressOfLocalVariable(CodeGenFunction &CGF, | |||
11767 | const VarDecl *VD) { | |||
11768 | if (!VD) | |||
11769 | return Address::invalid(); | |||
11770 | Address UntiedAddr = Address::invalid(); | |||
11771 | Address UntiedRealAddr = Address::invalid(); | |||
11772 | auto It = FunctionToUntiedTaskStackMap.find(CGF.CurFn); | |||
11773 | if (It != FunctionToUntiedTaskStackMap.end()) { | |||
11774 | const UntiedLocalVarsAddressesMap &UntiedData = | |||
11775 | UntiedLocalVarsStack[It->second]; | |||
11776 | auto I = UntiedData.find(VD); | |||
11777 | if (I != UntiedData.end()) { | |||
11778 | UntiedAddr = I->second.first; | |||
11779 | UntiedRealAddr = I->second.second; | |||
11780 | } | |||
11781 | } | |||
11782 | const VarDecl *CVD = VD->getCanonicalDecl(); | |||
11783 | if (CVD->hasAttr<OMPAllocateDeclAttr>()) { | |||
11784 | // Use the default allocation. | |||
11785 | if (!isAllocatableDecl(VD)) | |||
11786 | return UntiedAddr; | |||
11787 | llvm::Value *Size; | |||
11788 | CharUnits Align = CGM.getContext().getDeclAlign(CVD); | |||
11789 | if (CVD->getType()->isVariablyModifiedType()) { | |||
11790 | Size = CGF.getTypeSize(CVD->getType()); | |||
11791 | // Align the size: ((size + align - 1) / align) * align | |||
11792 | Size = CGF.Builder.CreateNUWAdd( | |||
11793 | Size, CGM.getSize(Align - CharUnits::fromQuantity(1))); | |||
11794 | Size = CGF.Builder.CreateUDiv(Size, CGM.getSize(Align)); | |||
11795 | Size = CGF.Builder.CreateNUWMul(Size, CGM.getSize(Align)); | |||
11796 | } else { | |||
11797 | CharUnits Sz = CGM.getContext().getTypeSizeInChars(CVD->getType()); | |||
11798 | Size = CGM.getSize(Sz.alignTo(Align)); | |||
11799 | } | |||
11800 | llvm::Value *ThreadID = getThreadID(CGF, CVD->getBeginLoc()); | |||
11801 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); | |||
11802 | const Expr *Allocator = AA->getAllocator(); | |||
11803 | llvm::Value *AllocVal = getAllocatorVal(CGF, Allocator); | |||
11804 | llvm::Value *Alignment = getAlignmentValue(CGM, CVD); | |||
11805 | SmallVector<llvm::Value *, 4> Args; | |||
11806 | Args.push_back(ThreadID); | |||
11807 | if (Alignment) | |||
11808 | Args.push_back(Alignment); | |||
11809 | Args.push_back(Size); | |||
11810 | Args.push_back(AllocVal); | |||
11811 | llvm::omp::RuntimeFunction FnID = | |||
11812 | Alignment ? OMPRTL___kmpc_aligned_alloc : OMPRTL___kmpc_alloc; | |||
11813 | llvm::Value *Addr = CGF.EmitRuntimeCall( | |||
11814 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), FnID), Args, | |||
11815 | getName({CVD->getName(), ".void.addr"})); | |||
11816 | llvm::FunctionCallee FiniRTLFn = OMPBuilder.getOrCreateRuntimeFunction( | |||
11817 | CGM.getModule(), OMPRTL___kmpc_free); | |||
11818 | QualType Ty = CGM.getContext().getPointerType(CVD->getType()); | |||
11819 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
11820 | Addr, CGF.ConvertTypeForMem(Ty), getName({CVD->getName(), ".addr"})); | |||
11821 | if (UntiedAddr.isValid()) | |||
11822 | CGF.EmitStoreOfScalar(Addr, UntiedAddr, /*Volatile=*/false, Ty); | |||
11823 | ||||
11824 | // Cleanup action for allocate support. | |||
11825 | class OMPAllocateCleanupTy final : public EHScopeStack::Cleanup { | |||
11826 | llvm::FunctionCallee RTLFn; | |||
11827 | SourceLocation::UIntTy LocEncoding; | |||
11828 | Address Addr; | |||
11829 | const Expr *AllocExpr; | |||
11830 | ||||
11831 | public: | |||
11832 | OMPAllocateCleanupTy(llvm::FunctionCallee RTLFn, | |||
11833 | SourceLocation::UIntTy LocEncoding, Address Addr, | |||
11834 | const Expr *AllocExpr) | |||
11835 | : RTLFn(RTLFn), LocEncoding(LocEncoding), Addr(Addr), | |||
11836 | AllocExpr(AllocExpr) {} | |||
11837 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | |||
11838 | if (!CGF.HaveInsertPoint()) | |||
11839 | return; | |||
11840 | llvm::Value *Args[3]; | |||
11841 | Args[0] = CGF.CGM.getOpenMPRuntime().getThreadID( | |||
11842 | CGF, SourceLocation::getFromRawEncoding(LocEncoding)); | |||
11843 | Args[1] = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
11844 | Addr.getPointer(), CGF.VoidPtrTy); | |||
11845 | llvm::Value *AllocVal = getAllocatorVal(CGF, AllocExpr); | |||
11846 | Args[2] = AllocVal; | |||
11847 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
11848 | } | |||
11849 | }; | |||
11850 | Address VDAddr = | |||
11851 | UntiedRealAddr.isValid() | |||
11852 | ? UntiedRealAddr | |||
11853 | : Address(Addr, CGF.ConvertTypeForMem(CVD->getType()), Align); | |||
11854 | CGF.EHStack.pushCleanup<OMPAllocateCleanupTy>( | |||
11855 | NormalAndEHCleanup, FiniRTLFn, CVD->getLocation().getRawEncoding(), | |||
11856 | VDAddr, Allocator); | |||
11857 | if (UntiedRealAddr.isValid()) | |||
11858 | if (auto *Region = | |||
11859 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
11860 | Region->emitUntiedSwitch(CGF); | |||
11861 | return VDAddr; | |||
11862 | } | |||
11863 | return UntiedAddr; | |||
11864 | } | |||
11865 | ||||
11866 | bool CGOpenMPRuntime::isLocalVarInUntiedTask(CodeGenFunction &CGF, | |||
11867 | const VarDecl *VD) const { | |||
11868 | auto It = FunctionToUntiedTaskStackMap.find(CGF.CurFn); | |||
11869 | if (It == FunctionToUntiedTaskStackMap.end()) | |||
11870 | return false; | |||
11871 | return UntiedLocalVarsStack[It->second].count(VD) > 0; | |||
11872 | } | |||
11873 | ||||
11874 | CGOpenMPRuntime::NontemporalDeclsRAII::NontemporalDeclsRAII( | |||
11875 | CodeGenModule &CGM, const OMPLoopDirective &S) | |||
11876 | : CGM(CGM), NeedToPush(S.hasClausesOfKind<OMPNontemporalClause>()) { | |||
11877 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")(static_cast <bool> (CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? void (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11877, __extension__ __PRETTY_FUNCTION__)); | |||
11878 | if (!NeedToPush) | |||
11879 | return; | |||
11880 | NontemporalDeclsSet &DS = | |||
11881 | CGM.getOpenMPRuntime().NontemporalDeclsStack.emplace_back(); | |||
11882 | for (const auto *C : S.getClausesOfKind<OMPNontemporalClause>()) { | |||
11883 | for (const Stmt *Ref : C->private_refs()) { | |||
11884 | const auto *SimpleRefExpr = cast<Expr>(Ref)->IgnoreParenImpCasts(); | |||
11885 | const ValueDecl *VD; | |||
11886 | if (const auto *DRE = dyn_cast<DeclRefExpr>(SimpleRefExpr)) { | |||
11887 | VD = DRE->getDecl(); | |||
11888 | } else { | |||
11889 | const auto *ME = cast<MemberExpr>(SimpleRefExpr); | |||
11890 | assert((ME->isImplicitCXXThis() ||(static_cast <bool> ((ME->isImplicitCXXThis() || isa <CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts( ))) && "Expected member of current class.") ? void (0 ) : __assert_fail ("(ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) && \"Expected member of current class.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11892, __extension__ __PRETTY_FUNCTION__)) | |||
11891 | isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) &&(static_cast <bool> ((ME->isImplicitCXXThis() || isa <CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts( ))) && "Expected member of current class.") ? void (0 ) : __assert_fail ("(ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) && \"Expected member of current class.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11892, __extension__ __PRETTY_FUNCTION__)) | |||
11892 | "Expected member of current class.")(static_cast <bool> ((ME->isImplicitCXXThis() || isa <CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts( ))) && "Expected member of current class.") ? void (0 ) : __assert_fail ("(ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) && \"Expected member of current class.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11892, __extension__ __PRETTY_FUNCTION__)); | |||
11893 | VD = ME->getMemberDecl(); | |||
11894 | } | |||
11895 | DS.insert(VD); | |||
11896 | } | |||
11897 | } | |||
11898 | } | |||
11899 | ||||
11900 | CGOpenMPRuntime::NontemporalDeclsRAII::~NontemporalDeclsRAII() { | |||
11901 | if (!NeedToPush) | |||
11902 | return; | |||
11903 | CGM.getOpenMPRuntime().NontemporalDeclsStack.pop_back(); | |||
11904 | } | |||
11905 | ||||
11906 | CGOpenMPRuntime::UntiedTaskLocalDeclsRAII::UntiedTaskLocalDeclsRAII( | |||
11907 | CodeGenFunction &CGF, | |||
11908 | const llvm::MapVector<CanonicalDeclPtr<const VarDecl>, | |||
11909 | std::pair<Address, Address>> &LocalVars) | |||
11910 | : CGM(CGF.CGM), NeedToPush(!LocalVars.empty()) { | |||
11911 | if (!NeedToPush) | |||
11912 | return; | |||
11913 | CGM.getOpenMPRuntime().FunctionToUntiedTaskStackMap.try_emplace( | |||
11914 | CGF.CurFn, CGM.getOpenMPRuntime().UntiedLocalVarsStack.size()); | |||
11915 | CGM.getOpenMPRuntime().UntiedLocalVarsStack.push_back(LocalVars); | |||
11916 | } | |||
11917 | ||||
11918 | CGOpenMPRuntime::UntiedTaskLocalDeclsRAII::~UntiedTaskLocalDeclsRAII() { | |||
11919 | if (!NeedToPush) | |||
11920 | return; | |||
11921 | CGM.getOpenMPRuntime().UntiedLocalVarsStack.pop_back(); | |||
11922 | } | |||
11923 | ||||
11924 | bool CGOpenMPRuntime::isNontemporalDecl(const ValueDecl *VD) const { | |||
11925 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")(static_cast <bool> (CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? void (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 11925, __extension__ __PRETTY_FUNCTION__)); | |||
11926 | ||||
11927 | return llvm::any_of( | |||
11928 | CGM.getOpenMPRuntime().NontemporalDeclsStack, | |||
11929 | [VD](const NontemporalDeclsSet &Set) { return Set.contains(VD); }); | |||
11930 | } | |||
11931 | ||||
11932 | void CGOpenMPRuntime::LastprivateConditionalRAII::tryToDisableInnerAnalysis( | |||
11933 | const OMPExecutableDirective &S, | |||
11934 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> &NeedToAddForLPCsAsDisabled) | |||
11935 | const { | |||
11936 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> NeedToCheckForLPCs; | |||
11937 | // Vars in target/task regions must be excluded completely. | |||
11938 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()) || | |||
11939 | isOpenMPTaskingDirective(S.getDirectiveKind())) { | |||
11940 | SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; | |||
11941 | getOpenMPCaptureRegions(CaptureRegions, S.getDirectiveKind()); | |||
11942 | const CapturedStmt *CS = S.getCapturedStmt(CaptureRegions.front()); | |||
11943 | for (const CapturedStmt::Capture &Cap : CS->captures()) { | |||
11944 | if (Cap.capturesVariable() || Cap.capturesVariableByCopy()) | |||
11945 | NeedToCheckForLPCs.insert(Cap.getCapturedVar()); | |||
11946 | } | |||
11947 | } | |||
11948 | // Exclude vars in private clauses. | |||
11949 | for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) { | |||
11950 | for (const Expr *Ref : C->varlists()) { | |||
11951 | if (!Ref->getType()->isScalarType()) | |||
11952 | continue; | |||
11953 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | |||
11954 | if (!DRE) | |||
11955 | continue; | |||
11956 | NeedToCheckForLPCs.insert(DRE->getDecl()); | |||
11957 | } | |||
11958 | } | |||
11959 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { | |||
11960 | for (const Expr *Ref : C->varlists()) { | |||
11961 | if (!Ref->getType()->isScalarType()) | |||
11962 | continue; | |||
11963 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | |||
11964 | if (!DRE) | |||
11965 | continue; | |||
11966 | NeedToCheckForLPCs.insert(DRE->getDecl()); | |||
11967 | } | |||
11968 | } | |||
11969 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { | |||
11970 | for (const Expr *Ref : C->varlists()) { | |||
11971 | if (!Ref->getType()->isScalarType()) | |||
11972 | continue; | |||
11973 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | |||
11974 | if (!DRE) | |||
11975 | continue; | |||
11976 | NeedToCheckForLPCs.insert(DRE->getDecl()); | |||
11977 | } | |||
11978 | } | |||
11979 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { | |||
11980 | for (const Expr *Ref : C->varlists()) { | |||
11981 | if (!Ref->getType()->isScalarType()) | |||
11982 | continue; | |||
11983 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | |||
11984 | if (!DRE) | |||
11985 | continue; | |||
11986 | NeedToCheckForLPCs.insert(DRE->getDecl()); | |||
11987 | } | |||
11988 | } | |||
11989 | for (const auto *C : S.getClausesOfKind<OMPLinearClause>()) { | |||
11990 | for (const Expr *Ref : C->varlists()) { | |||
11991 | if (!Ref->getType()->isScalarType()) | |||
11992 | continue; | |||
11993 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | |||
11994 | if (!DRE) | |||
11995 | continue; | |||
11996 | NeedToCheckForLPCs.insert(DRE->getDecl()); | |||
11997 | } | |||
11998 | } | |||
11999 | for (const Decl *VD : NeedToCheckForLPCs) { | |||
12000 | for (const LastprivateConditionalData &Data : | |||
12001 | llvm::reverse(CGM.getOpenMPRuntime().LastprivateConditionalStack)) { | |||
12002 | if (Data.DeclToUniqueName.count(VD) > 0) { | |||
12003 | if (!Data.Disabled) | |||
12004 | NeedToAddForLPCsAsDisabled.insert(VD); | |||
12005 | break; | |||
12006 | } | |||
12007 | } | |||
12008 | } | |||
12009 | } | |||
12010 | ||||
12011 | CGOpenMPRuntime::LastprivateConditionalRAII::LastprivateConditionalRAII( | |||
12012 | CodeGenFunction &CGF, const OMPExecutableDirective &S, LValue IVLVal) | |||
12013 | : CGM(CGF.CGM), | |||
12014 | Action((CGM.getLangOpts().OpenMP >= 50 && | |||
12015 | llvm::any_of(S.getClausesOfKind<OMPLastprivateClause>(), | |||
12016 | [](const OMPLastprivateClause *C) { | |||
12017 | return C->getKind() == | |||
12018 | OMPC_LASTPRIVATE_conditional; | |||
12019 | })) | |||
12020 | ? ActionToDo::PushAsLastprivateConditional | |||
12021 | : ActionToDo::DoNotPush) { | |||
12022 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")(static_cast <bool> (CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? void (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12022, __extension__ __PRETTY_FUNCTION__)); | |||
12023 | if (CGM.getLangOpts().OpenMP < 50 || Action == ActionToDo::DoNotPush) | |||
12024 | return; | |||
12025 | assert(Action == ActionToDo::PushAsLastprivateConditional &&(static_cast <bool> (Action == ActionToDo::PushAsLastprivateConditional && "Expected a push action.") ? void (0) : __assert_fail ("Action == ActionToDo::PushAsLastprivateConditional && \"Expected a push action.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12026, __extension__ __PRETTY_FUNCTION__)) | |||
12026 | "Expected a push action.")(static_cast <bool> (Action == ActionToDo::PushAsLastprivateConditional && "Expected a push action.") ? void (0) : __assert_fail ("Action == ActionToDo::PushAsLastprivateConditional && \"Expected a push action.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12026, __extension__ __PRETTY_FUNCTION__)); | |||
12027 | LastprivateConditionalData &Data = | |||
12028 | CGM.getOpenMPRuntime().LastprivateConditionalStack.emplace_back(); | |||
12029 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { | |||
12030 | if (C->getKind() != OMPC_LASTPRIVATE_conditional) | |||
12031 | continue; | |||
12032 | ||||
12033 | for (const Expr *Ref : C->varlists()) { | |||
12034 | Data.DeclToUniqueName.insert(std::make_pair( | |||
12035 | cast<DeclRefExpr>(Ref->IgnoreParenImpCasts())->getDecl(), | |||
12036 | SmallString<16>(generateUniqueName(CGM, "pl_cond", Ref)))); | |||
12037 | } | |||
12038 | } | |||
12039 | Data.IVLVal = IVLVal; | |||
12040 | Data.Fn = CGF.CurFn; | |||
12041 | } | |||
12042 | ||||
12043 | CGOpenMPRuntime::LastprivateConditionalRAII::LastprivateConditionalRAII( | |||
12044 | CodeGenFunction &CGF, const OMPExecutableDirective &S) | |||
12045 | : CGM(CGF.CGM), Action(ActionToDo::DoNotPush) { | |||
12046 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")(static_cast <bool> (CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? void (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12046, __extension__ __PRETTY_FUNCTION__)); | |||
12047 | if (CGM.getLangOpts().OpenMP < 50) | |||
12048 | return; | |||
12049 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> NeedToAddForLPCsAsDisabled; | |||
12050 | tryToDisableInnerAnalysis(S, NeedToAddForLPCsAsDisabled); | |||
12051 | if (!NeedToAddForLPCsAsDisabled.empty()) { | |||
12052 | Action = ActionToDo::DisableLastprivateConditional; | |||
12053 | LastprivateConditionalData &Data = | |||
12054 | CGM.getOpenMPRuntime().LastprivateConditionalStack.emplace_back(); | |||
12055 | for (const Decl *VD : NeedToAddForLPCsAsDisabled) | |||
12056 | Data.DeclToUniqueName.insert(std::make_pair(VD, SmallString<16>())); | |||
12057 | Data.Fn = CGF.CurFn; | |||
12058 | Data.Disabled = true; | |||
12059 | } | |||
12060 | } | |||
12061 | ||||
12062 | CGOpenMPRuntime::LastprivateConditionalRAII | |||
12063 | CGOpenMPRuntime::LastprivateConditionalRAII::disable( | |||
12064 | CodeGenFunction &CGF, const OMPExecutableDirective &S) { | |||
12065 | return LastprivateConditionalRAII(CGF, S); | |||
12066 | } | |||
12067 | ||||
12068 | CGOpenMPRuntime::LastprivateConditionalRAII::~LastprivateConditionalRAII() { | |||
12069 | if (CGM.getLangOpts().OpenMP < 50) | |||
12070 | return; | |||
12071 | if (Action == ActionToDo::DisableLastprivateConditional) { | |||
12072 | assert(CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled &&(static_cast <bool> (CGM.getOpenMPRuntime().LastprivateConditionalStack .back().Disabled && "Expected list of disabled private vars." ) ? void (0) : __assert_fail ("CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of disabled private vars.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12073, __extension__ __PRETTY_FUNCTION__)) | |||
12073 | "Expected list of disabled private vars.")(static_cast <bool> (CGM.getOpenMPRuntime().LastprivateConditionalStack .back().Disabled && "Expected list of disabled private vars." ) ? void (0) : __assert_fail ("CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of disabled private vars.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12073, __extension__ __PRETTY_FUNCTION__)); | |||
12074 | CGM.getOpenMPRuntime().LastprivateConditionalStack.pop_back(); | |||
12075 | } | |||
12076 | if (Action == ActionToDo::PushAsLastprivateConditional) { | |||
12077 | assert((static_cast <bool> (!CGM.getOpenMPRuntime().LastprivateConditionalStack .back().Disabled && "Expected list of lastprivate conditional vars." ) ? void (0) : __assert_fail ("!CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of lastprivate conditional vars.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12079, __extension__ __PRETTY_FUNCTION__)) | |||
12078 | !CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled &&(static_cast <bool> (!CGM.getOpenMPRuntime().LastprivateConditionalStack .back().Disabled && "Expected list of lastprivate conditional vars." ) ? void (0) : __assert_fail ("!CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of lastprivate conditional vars.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12079, __extension__ __PRETTY_FUNCTION__)) | |||
12079 | "Expected list of lastprivate conditional vars.")(static_cast <bool> (!CGM.getOpenMPRuntime().LastprivateConditionalStack .back().Disabled && "Expected list of lastprivate conditional vars." ) ? void (0) : __assert_fail ("!CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of lastprivate conditional vars.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12079, __extension__ __PRETTY_FUNCTION__)); | |||
12080 | CGM.getOpenMPRuntime().LastprivateConditionalStack.pop_back(); | |||
12081 | } | |||
12082 | } | |||
12083 | ||||
12084 | Address CGOpenMPRuntime::emitLastprivateConditionalInit(CodeGenFunction &CGF, | |||
12085 | const VarDecl *VD) { | |||
12086 | ASTContext &C = CGM.getContext(); | |||
12087 | auto I = LastprivateConditionalToTypes.find(CGF.CurFn); | |||
12088 | if (I == LastprivateConditionalToTypes.end()) | |||
12089 | I = LastprivateConditionalToTypes.try_emplace(CGF.CurFn).first; | |||
12090 | QualType NewType; | |||
12091 | const FieldDecl *VDField; | |||
12092 | const FieldDecl *FiredField; | |||
12093 | LValue BaseLVal; | |||
12094 | auto VI = I->getSecond().find(VD); | |||
12095 | if (VI == I->getSecond().end()) { | |||
12096 | RecordDecl *RD = C.buildImplicitRecord("lasprivate.conditional"); | |||
12097 | RD->startDefinition(); | |||
12098 | VDField = addFieldToRecordDecl(C, RD, VD->getType().getNonReferenceType()); | |||
12099 | FiredField = addFieldToRecordDecl(C, RD, C.CharTy); | |||
12100 | RD->completeDefinition(); | |||
12101 | NewType = C.getRecordType(RD); | |||
12102 | Address Addr = CGF.CreateMemTemp(NewType, C.getDeclAlign(VD), VD->getName()); | |||
12103 | BaseLVal = CGF.MakeAddrLValue(Addr, NewType, AlignmentSource::Decl); | |||
12104 | I->getSecond().try_emplace(VD, NewType, VDField, FiredField, BaseLVal); | |||
12105 | } else { | |||
12106 | NewType = std::get<0>(VI->getSecond()); | |||
12107 | VDField = std::get<1>(VI->getSecond()); | |||
12108 | FiredField = std::get<2>(VI->getSecond()); | |||
12109 | BaseLVal = std::get<3>(VI->getSecond()); | |||
12110 | } | |||
12111 | LValue FiredLVal = | |||
12112 | CGF.EmitLValueForField(BaseLVal, FiredField); | |||
12113 | CGF.EmitStoreOfScalar( | |||
12114 | llvm::ConstantInt::getNullValue(CGF.ConvertTypeForMem(C.CharTy)), | |||
12115 | FiredLVal); | |||
12116 | return CGF.EmitLValueForField(BaseLVal, VDField).getAddress(CGF); | |||
12117 | } | |||
12118 | ||||
12119 | namespace { | |||
12120 | /// Checks if the lastprivate conditional variable is referenced in LHS. | |||
12121 | class LastprivateConditionalRefChecker final | |||
12122 | : public ConstStmtVisitor<LastprivateConditionalRefChecker, bool> { | |||
12123 | ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM; | |||
12124 | const Expr *FoundE = nullptr; | |||
12125 | const Decl *FoundD = nullptr; | |||
12126 | StringRef UniqueDeclName; | |||
12127 | LValue IVLVal; | |||
12128 | llvm::Function *FoundFn = nullptr; | |||
12129 | SourceLocation Loc; | |||
12130 | ||||
12131 | public: | |||
12132 | bool VisitDeclRefExpr(const DeclRefExpr *E) { | |||
12133 | for (const CGOpenMPRuntime::LastprivateConditionalData &D : | |||
12134 | llvm::reverse(LPM)) { | |||
12135 | auto It = D.DeclToUniqueName.find(E->getDecl()); | |||
12136 | if (It == D.DeclToUniqueName.end()) | |||
12137 | continue; | |||
12138 | if (D.Disabled) | |||
12139 | return false; | |||
12140 | FoundE = E; | |||
12141 | FoundD = E->getDecl()->getCanonicalDecl(); | |||
12142 | UniqueDeclName = It->second; | |||
12143 | IVLVal = D.IVLVal; | |||
12144 | FoundFn = D.Fn; | |||
12145 | break; | |||
12146 | } | |||
12147 | return FoundE == E; | |||
12148 | } | |||
12149 | bool VisitMemberExpr(const MemberExpr *E) { | |||
12150 | if (!CodeGenFunction::IsWrappedCXXThis(E->getBase())) | |||
12151 | return false; | |||
12152 | for (const CGOpenMPRuntime::LastprivateConditionalData &D : | |||
12153 | llvm::reverse(LPM)) { | |||
12154 | auto It = D.DeclToUniqueName.find(E->getMemberDecl()); | |||
12155 | if (It == D.DeclToUniqueName.end()) | |||
12156 | continue; | |||
12157 | if (D.Disabled) | |||
12158 | return false; | |||
12159 | FoundE = E; | |||
12160 | FoundD = E->getMemberDecl()->getCanonicalDecl(); | |||
12161 | UniqueDeclName = It->second; | |||
12162 | IVLVal = D.IVLVal; | |||
12163 | FoundFn = D.Fn; | |||
12164 | break; | |||
12165 | } | |||
12166 | return FoundE == E; | |||
12167 | } | |||
12168 | bool VisitStmt(const Stmt *S) { | |||
12169 | for (const Stmt *Child : S->children()) { | |||
12170 | if (!Child) | |||
12171 | continue; | |||
12172 | if (const auto *E = dyn_cast<Expr>(Child)) | |||
12173 | if (!E->isGLValue()) | |||
12174 | continue; | |||
12175 | if (Visit(Child)) | |||
12176 | return true; | |||
12177 | } | |||
12178 | return false; | |||
12179 | } | |||
12180 | explicit LastprivateConditionalRefChecker( | |||
12181 | ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM) | |||
12182 | : LPM(LPM) {} | |||
12183 | std::tuple<const Expr *, const Decl *, StringRef, LValue, llvm::Function *> | |||
12184 | getFoundData() const { | |||
12185 | return std::make_tuple(FoundE, FoundD, UniqueDeclName, IVLVal, FoundFn); | |||
12186 | } | |||
12187 | }; | |||
12188 | } // namespace | |||
12189 | ||||
12190 | void CGOpenMPRuntime::emitLastprivateConditionalUpdate(CodeGenFunction &CGF, | |||
12191 | LValue IVLVal, | |||
12192 | StringRef UniqueDeclName, | |||
12193 | LValue LVal, | |||
12194 | SourceLocation Loc) { | |||
12195 | // Last updated loop counter for the lastprivate conditional var. | |||
12196 | // int<xx> last_iv = 0; | |||
12197 | llvm::Type *LLIVTy = CGF.ConvertTypeForMem(IVLVal.getType()); | |||
12198 | llvm::Constant *LastIV = | |||
12199 | getOrCreateInternalVariable(LLIVTy, getName({UniqueDeclName, "iv"})); | |||
12200 | cast<llvm::GlobalVariable>(LastIV)->setAlignment( | |||
12201 | IVLVal.getAlignment().getAsAlign()); | |||
12202 | LValue LastIVLVal = CGF.MakeNaturalAlignAddrLValue(LastIV, IVLVal.getType()); | |||
12203 | ||||
12204 | // Last value of the lastprivate conditional. | |||
12205 | // decltype(priv_a) last_a; | |||
12206 | llvm::GlobalVariable *Last = getOrCreateInternalVariable( | |||
12207 | CGF.ConvertTypeForMem(LVal.getType()), UniqueDeclName); | |||
12208 | Last->setAlignment(LVal.getAlignment().getAsAlign()); | |||
12209 | LValue LastLVal = CGF.MakeAddrLValue( | |||
12210 | Address(Last, Last->getValueType(), LVal.getAlignment()), LVal.getType()); | |||
12211 | ||||
12212 | // Global loop counter. Required to handle inner parallel-for regions. | |||
12213 | // iv | |||
12214 | llvm::Value *IVVal = CGF.EmitLoadOfScalar(IVLVal, Loc); | |||
12215 | ||||
12216 | // #pragma omp critical(a) | |||
12217 | // if (last_iv <= iv) { | |||
12218 | // last_iv = iv; | |||
12219 | // last_a = priv_a; | |||
12220 | // } | |||
12221 | auto &&CodeGen = [&LastIVLVal, &IVLVal, IVVal, &LVal, &LastLVal, | |||
12222 | Loc](CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
12223 | Action.Enter(CGF); | |||
12224 | llvm::Value *LastIVVal = CGF.EmitLoadOfScalar(LastIVLVal, Loc); | |||
12225 | // (last_iv <= iv) ? Check if the variable is updated and store new | |||
12226 | // value in global var. | |||
12227 | llvm::Value *CmpRes; | |||
12228 | if (IVLVal.getType()->isSignedIntegerType()) { | |||
12229 | CmpRes = CGF.Builder.CreateICmpSLE(LastIVVal, IVVal); | |||
12230 | } else { | |||
12231 | assert(IVLVal.getType()->isUnsignedIntegerType() &&(static_cast <bool> (IVLVal.getType()->isUnsignedIntegerType () && "Loop iteration variable must be integer.") ? void (0) : __assert_fail ("IVLVal.getType()->isUnsignedIntegerType() && \"Loop iteration variable must be integer.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12232, __extension__ __PRETTY_FUNCTION__)) | |||
12232 | "Loop iteration variable must be integer.")(static_cast <bool> (IVLVal.getType()->isUnsignedIntegerType () && "Loop iteration variable must be integer.") ? void (0) : __assert_fail ("IVLVal.getType()->isUnsignedIntegerType() && \"Loop iteration variable must be integer.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12232, __extension__ __PRETTY_FUNCTION__)); | |||
12233 | CmpRes = CGF.Builder.CreateICmpULE(LastIVVal, IVVal); | |||
12234 | } | |||
12235 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("lp_cond_then"); | |||
12236 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock("lp_cond_exit"); | |||
12237 | CGF.Builder.CreateCondBr(CmpRes, ThenBB, ExitBB); | |||
12238 | // { | |||
12239 | CGF.EmitBlock(ThenBB); | |||
12240 | ||||
12241 | // last_iv = iv; | |||
12242 | CGF.EmitStoreOfScalar(IVVal, LastIVLVal); | |||
12243 | ||||
12244 | // last_a = priv_a; | |||
12245 | switch (CGF.getEvaluationKind(LVal.getType())) { | |||
12246 | case TEK_Scalar: { | |||
12247 | llvm::Value *PrivVal = CGF.EmitLoadOfScalar(LVal, Loc); | |||
12248 | CGF.EmitStoreOfScalar(PrivVal, LastLVal); | |||
12249 | break; | |||
12250 | } | |||
12251 | case TEK_Complex: { | |||
12252 | CodeGenFunction::ComplexPairTy PrivVal = CGF.EmitLoadOfComplex(LVal, Loc); | |||
12253 | CGF.EmitStoreOfComplex(PrivVal, LastLVal, /*isInit=*/false); | |||
12254 | break; | |||
12255 | } | |||
12256 | case TEK_Aggregate: | |||
12257 | llvm_unreachable(::llvm::llvm_unreachable_internal("Aggregates are not supported in lastprivate conditional." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12258) | |||
12258 | "Aggregates are not supported in lastprivate conditional.")::llvm::llvm_unreachable_internal("Aggregates are not supported in lastprivate conditional." , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12258); | |||
12259 | } | |||
12260 | // } | |||
12261 | CGF.EmitBranch(ExitBB); | |||
12262 | // There is no need to emit line number for unconditional branch. | |||
12263 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
12264 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); | |||
12265 | }; | |||
12266 | ||||
12267 | if (CGM.getLangOpts().OpenMPSimd) { | |||
12268 | // Do not emit as a critical region as no parallel region could be emitted. | |||
12269 | RegionCodeGenTy ThenRCG(CodeGen); | |||
12270 | ThenRCG(CGF); | |||
12271 | } else { | |||
12272 | emitCriticalRegion(CGF, UniqueDeclName, CodeGen, Loc); | |||
12273 | } | |||
12274 | } | |||
12275 | ||||
12276 | void CGOpenMPRuntime::checkAndEmitLastprivateConditional(CodeGenFunction &CGF, | |||
12277 | const Expr *LHS) { | |||
12278 | if (CGF.getLangOpts().OpenMP < 50 || LastprivateConditionalStack.empty()) | |||
12279 | return; | |||
12280 | LastprivateConditionalRefChecker Checker(LastprivateConditionalStack); | |||
12281 | if (!Checker.Visit(LHS)) | |||
12282 | return; | |||
12283 | const Expr *FoundE; | |||
12284 | const Decl *FoundD; | |||
12285 | StringRef UniqueDeclName; | |||
12286 | LValue IVLVal; | |||
12287 | llvm::Function *FoundFn; | |||
12288 | std::tie(FoundE, FoundD, UniqueDeclName, IVLVal, FoundFn) = | |||
12289 | Checker.getFoundData(); | |||
12290 | if (FoundFn != CGF.CurFn) { | |||
12291 | // Special codegen for inner parallel regions. | |||
12292 | // ((struct.lastprivate.conditional*)&priv_a)->Fired = 1; | |||
12293 | auto It = LastprivateConditionalToTypes[FoundFn].find(FoundD); | |||
12294 | assert(It != LastprivateConditionalToTypes[FoundFn].end() &&(static_cast <bool> (It != LastprivateConditionalToTypes [FoundFn].end() && "Lastprivate conditional is not found in outer region." ) ? void (0) : __assert_fail ("It != LastprivateConditionalToTypes[FoundFn].end() && \"Lastprivate conditional is not found in outer region.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12295, __extension__ __PRETTY_FUNCTION__)) | |||
12295 | "Lastprivate conditional is not found in outer region.")(static_cast <bool> (It != LastprivateConditionalToTypes [FoundFn].end() && "Lastprivate conditional is not found in outer region." ) ? void (0) : __assert_fail ("It != LastprivateConditionalToTypes[FoundFn].end() && \"Lastprivate conditional is not found in outer region.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12295, __extension__ __PRETTY_FUNCTION__)); | |||
12296 | QualType StructTy = std::get<0>(It->getSecond()); | |||
12297 | const FieldDecl* FiredDecl = std::get<2>(It->getSecond()); | |||
12298 | LValue PrivLVal = CGF.EmitLValue(FoundE); | |||
12299 | Address StructAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
12300 | PrivLVal.getAddress(CGF), | |||
12301 | CGF.ConvertTypeForMem(CGF.getContext().getPointerType(StructTy)), | |||
12302 | CGF.ConvertTypeForMem(StructTy)); | |||
12303 | LValue BaseLVal = | |||
12304 | CGF.MakeAddrLValue(StructAddr, StructTy, AlignmentSource::Decl); | |||
12305 | LValue FiredLVal = CGF.EmitLValueForField(BaseLVal, FiredDecl); | |||
12306 | CGF.EmitAtomicStore(RValue::get(llvm::ConstantInt::get( | |||
12307 | CGF.ConvertTypeForMem(FiredDecl->getType()), 1)), | |||
12308 | FiredLVal, llvm::AtomicOrdering::Unordered, | |||
12309 | /*IsVolatile=*/true, /*isInit=*/false); | |||
12310 | return; | |||
12311 | } | |||
12312 | ||||
12313 | // Private address of the lastprivate conditional in the current context. | |||
12314 | // priv_a | |||
12315 | LValue LVal = CGF.EmitLValue(FoundE); | |||
12316 | emitLastprivateConditionalUpdate(CGF, IVLVal, UniqueDeclName, LVal, | |||
12317 | FoundE->getExprLoc()); | |||
12318 | } | |||
12319 | ||||
12320 | void CGOpenMPRuntime::checkAndEmitSharedLastprivateConditional( | |||
12321 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | |||
12322 | const llvm::DenseSet<CanonicalDeclPtr<const VarDecl>> &IgnoredDecls) { | |||
12323 | if (CGF.getLangOpts().OpenMP < 50 || LastprivateConditionalStack.empty()) | |||
12324 | return; | |||
12325 | auto Range = llvm::reverse(LastprivateConditionalStack); | |||
12326 | auto It = llvm::find_if( | |||
12327 | Range, [](const LastprivateConditionalData &D) { return !D.Disabled; }); | |||
12328 | if (It == Range.end() || It->Fn != CGF.CurFn) | |||
12329 | return; | |||
12330 | auto LPCI = LastprivateConditionalToTypes.find(It->Fn); | |||
12331 | assert(LPCI != LastprivateConditionalToTypes.end() &&(static_cast <bool> (LPCI != LastprivateConditionalToTypes .end() && "Lastprivates must be registered already.") ? void (0) : __assert_fail ("LPCI != LastprivateConditionalToTypes.end() && \"Lastprivates must be registered already.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12332, __extension__ __PRETTY_FUNCTION__)) | |||
12332 | "Lastprivates must be registered already.")(static_cast <bool> (LPCI != LastprivateConditionalToTypes .end() && "Lastprivates must be registered already.") ? void (0) : __assert_fail ("LPCI != LastprivateConditionalToTypes.end() && \"Lastprivates must be registered already.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12332, __extension__ __PRETTY_FUNCTION__)); | |||
12333 | SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; | |||
12334 | getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind()); | |||
12335 | const CapturedStmt *CS = D.getCapturedStmt(CaptureRegions.back()); | |||
12336 | for (const auto &Pair : It->DeclToUniqueName) { | |||
12337 | const auto *VD = cast<VarDecl>(Pair.first->getCanonicalDecl()); | |||
12338 | if (!CS->capturesVariable(VD) || IgnoredDecls.contains(VD)) | |||
12339 | continue; | |||
12340 | auto I = LPCI->getSecond().find(Pair.first); | |||
12341 | assert(I != LPCI->getSecond().end() &&(static_cast <bool> (I != LPCI->getSecond().end() && "Lastprivate must be rehistered already.") ? void (0) : __assert_fail ("I != LPCI->getSecond().end() && \"Lastprivate must be rehistered already.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12342, __extension__ __PRETTY_FUNCTION__)) | |||
12342 | "Lastprivate must be rehistered already.")(static_cast <bool> (I != LPCI->getSecond().end() && "Lastprivate must be rehistered already.") ? void (0) : __assert_fail ("I != LPCI->getSecond().end() && \"Lastprivate must be rehistered already.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12342, __extension__ __PRETTY_FUNCTION__)); | |||
12343 | // bool Cmp = priv_a.Fired != 0; | |||
12344 | LValue BaseLVal = std::get<3>(I->getSecond()); | |||
12345 | LValue FiredLVal = | |||
12346 | CGF.EmitLValueForField(BaseLVal, std::get<2>(I->getSecond())); | |||
12347 | llvm::Value *Res = CGF.EmitLoadOfScalar(FiredLVal, D.getBeginLoc()); | |||
12348 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Res); | |||
12349 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("lpc.then"); | |||
12350 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("lpc.done"); | |||
12351 | // if (Cmp) { | |||
12352 | CGF.Builder.CreateCondBr(Cmp, ThenBB, DoneBB); | |||
12353 | CGF.EmitBlock(ThenBB); | |||
12354 | Address Addr = CGF.GetAddrOfLocalVar(VD); | |||
12355 | LValue LVal; | |||
12356 | if (VD->getType()->isReferenceType()) | |||
12357 | LVal = CGF.EmitLoadOfReferenceLValue(Addr, VD->getType(), | |||
12358 | AlignmentSource::Decl); | |||
12359 | else | |||
12360 | LVal = CGF.MakeAddrLValue(Addr, VD->getType().getNonReferenceType(), | |||
12361 | AlignmentSource::Decl); | |||
12362 | emitLastprivateConditionalUpdate(CGF, It->IVLVal, Pair.second, LVal, | |||
12363 | D.getBeginLoc()); | |||
12364 | auto AL = ApplyDebugLocation::CreateArtificial(CGF); | |||
12365 | CGF.EmitBlock(DoneBB, /*IsFinal=*/true); | |||
12366 | // } | |||
12367 | } | |||
12368 | } | |||
12369 | ||||
12370 | void CGOpenMPRuntime::emitLastprivateConditionalFinalUpdate( | |||
12371 | CodeGenFunction &CGF, LValue PrivLVal, const VarDecl *VD, | |||
12372 | SourceLocation Loc) { | |||
12373 | if (CGF.getLangOpts().OpenMP < 50) | |||
12374 | return; | |||
12375 | auto It = LastprivateConditionalStack.back().DeclToUniqueName.find(VD); | |||
12376 | assert(It != LastprivateConditionalStack.back().DeclToUniqueName.end() &&(static_cast <bool> (It != LastprivateConditionalStack. back().DeclToUniqueName.end() && "Unknown lastprivate conditional variable." ) ? void (0) : __assert_fail ("It != LastprivateConditionalStack.back().DeclToUniqueName.end() && \"Unknown lastprivate conditional variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12377, __extension__ __PRETTY_FUNCTION__)) | |||
12377 | "Unknown lastprivate conditional variable.")(static_cast <bool> (It != LastprivateConditionalStack. back().DeclToUniqueName.end() && "Unknown lastprivate conditional variable." ) ? void (0) : __assert_fail ("It != LastprivateConditionalStack.back().DeclToUniqueName.end() && \"Unknown lastprivate conditional variable.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12377, __extension__ __PRETTY_FUNCTION__)); | |||
12378 | StringRef UniqueName = It->second; | |||
12379 | llvm::GlobalVariable *GV = CGM.getModule().getNamedGlobal(UniqueName); | |||
12380 | // The variable was not updated in the region - exit. | |||
12381 | if (!GV) | |||
12382 | return; | |||
12383 | LValue LPLVal = CGF.MakeAddrLValue( | |||
12384 | Address(GV, GV->getValueType(), PrivLVal.getAlignment()), | |||
12385 | PrivLVal.getType().getNonReferenceType()); | |||
12386 | llvm::Value *Res = CGF.EmitLoadOfScalar(LPLVal, Loc); | |||
12387 | CGF.EmitStoreOfScalar(Res, PrivLVal); | |||
12388 | } | |||
12389 | ||||
12390 | llvm::Function *CGOpenMPSIMDRuntime::emitParallelOutlinedFunction( | |||
12391 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
12392 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
12393 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12393); | |||
12394 | } | |||
12395 | ||||
12396 | llvm::Function *CGOpenMPSIMDRuntime::emitTeamsOutlinedFunction( | |||
12397 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
12398 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
12399 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12399); | |||
12400 | } | |||
12401 | ||||
12402 | llvm::Function *CGOpenMPSIMDRuntime::emitTaskOutlinedFunction( | |||
12403 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
12404 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, | |||
12405 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, | |||
12406 | bool Tied, unsigned &NumberOfParts) { | |||
12407 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12407); | |||
12408 | } | |||
12409 | ||||
12410 | void CGOpenMPSIMDRuntime::emitParallelCall(CodeGenFunction &CGF, | |||
12411 | SourceLocation Loc, | |||
12412 | llvm::Function *OutlinedFn, | |||
12413 | ArrayRef<llvm::Value *> CapturedVars, | |||
12414 | const Expr *IfCond, | |||
12415 | llvm::Value *NumThreads) { | |||
12416 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12416); | |||
12417 | } | |||
12418 | ||||
12419 | void CGOpenMPSIMDRuntime::emitCriticalRegion( | |||
12420 | CodeGenFunction &CGF, StringRef CriticalName, | |||
12421 | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, | |||
12422 | const Expr *Hint) { | |||
12423 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12423); | |||
12424 | } | |||
12425 | ||||
12426 | void CGOpenMPSIMDRuntime::emitMasterRegion(CodeGenFunction &CGF, | |||
12427 | const RegionCodeGenTy &MasterOpGen, | |||
12428 | SourceLocation Loc) { | |||
12429 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12429); | |||
12430 | } | |||
12431 | ||||
12432 | void CGOpenMPSIMDRuntime::emitMaskedRegion(CodeGenFunction &CGF, | |||
12433 | const RegionCodeGenTy &MasterOpGen, | |||
12434 | SourceLocation Loc, | |||
12435 | const Expr *Filter) { | |||
12436 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12436); | |||
12437 | } | |||
12438 | ||||
12439 | void CGOpenMPSIMDRuntime::emitTaskyieldCall(CodeGenFunction &CGF, | |||
12440 | SourceLocation Loc) { | |||
12441 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12441); | |||
12442 | } | |||
12443 | ||||
12444 | void CGOpenMPSIMDRuntime::emitTaskgroupRegion( | |||
12445 | CodeGenFunction &CGF, const RegionCodeGenTy &TaskgroupOpGen, | |||
12446 | SourceLocation Loc) { | |||
12447 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12447); | |||
12448 | } | |||
12449 | ||||
12450 | void CGOpenMPSIMDRuntime::emitSingleRegion( | |||
12451 | CodeGenFunction &CGF, const RegionCodeGenTy &SingleOpGen, | |||
12452 | SourceLocation Loc, ArrayRef<const Expr *> CopyprivateVars, | |||
12453 | ArrayRef<const Expr *> DestExprs, ArrayRef<const Expr *> SrcExprs, | |||
12454 | ArrayRef<const Expr *> AssignmentOps) { | |||
12455 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12455); | |||
12456 | } | |||
12457 | ||||
12458 | void CGOpenMPSIMDRuntime::emitOrderedRegion(CodeGenFunction &CGF, | |||
12459 | const RegionCodeGenTy &OrderedOpGen, | |||
12460 | SourceLocation Loc, | |||
12461 | bool IsThreads) { | |||
12462 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12462); | |||
12463 | } | |||
12464 | ||||
12465 | void CGOpenMPSIMDRuntime::emitBarrierCall(CodeGenFunction &CGF, | |||
12466 | SourceLocation Loc, | |||
12467 | OpenMPDirectiveKind Kind, | |||
12468 | bool EmitChecks, | |||
12469 | bool ForceSimpleCall) { | |||
12470 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12470); | |||
12471 | } | |||
12472 | ||||
12473 | void CGOpenMPSIMDRuntime::emitForDispatchInit( | |||
12474 | CodeGenFunction &CGF, SourceLocation Loc, | |||
12475 | const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned, | |||
12476 | bool Ordered, const DispatchRTInput &DispatchValues) { | |||
12477 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12477); | |||
12478 | } | |||
12479 | ||||
12480 | void CGOpenMPSIMDRuntime::emitForStaticInit( | |||
12481 | CodeGenFunction &CGF, SourceLocation Loc, OpenMPDirectiveKind DKind, | |||
12482 | const OpenMPScheduleTy &ScheduleKind, const StaticRTInput &Values) { | |||
12483 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12483); | |||
12484 | } | |||
12485 | ||||
12486 | void CGOpenMPSIMDRuntime::emitDistributeStaticInit( | |||
12487 | CodeGenFunction &CGF, SourceLocation Loc, | |||
12488 | OpenMPDistScheduleClauseKind SchedKind, const StaticRTInput &Values) { | |||
12489 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12489); | |||
12490 | } | |||
12491 | ||||
12492 | void CGOpenMPSIMDRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF, | |||
12493 | SourceLocation Loc, | |||
12494 | unsigned IVSize, | |||
12495 | bool IVSigned) { | |||
12496 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12496); | |||
12497 | } | |||
12498 | ||||
12499 | void CGOpenMPSIMDRuntime::emitForStaticFinish(CodeGenFunction &CGF, | |||
12500 | SourceLocation Loc, | |||
12501 | OpenMPDirectiveKind DKind) { | |||
12502 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12502); | |||
12503 | } | |||
12504 | ||||
12505 | llvm::Value *CGOpenMPSIMDRuntime::emitForNext(CodeGenFunction &CGF, | |||
12506 | SourceLocation Loc, | |||
12507 | unsigned IVSize, bool IVSigned, | |||
12508 | Address IL, Address LB, | |||
12509 | Address UB, Address ST) { | |||
12510 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12510); | |||
12511 | } | |||
12512 | ||||
12513 | void CGOpenMPSIMDRuntime::emitNumThreadsClause(CodeGenFunction &CGF, | |||
12514 | llvm::Value *NumThreads, | |||
12515 | SourceLocation Loc) { | |||
12516 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12516); | |||
12517 | } | |||
12518 | ||||
12519 | void CGOpenMPSIMDRuntime::emitProcBindClause(CodeGenFunction &CGF, | |||
12520 | ProcBindKind ProcBind, | |||
12521 | SourceLocation Loc) { | |||
12522 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12522); | |||
12523 | } | |||
12524 | ||||
12525 | Address CGOpenMPSIMDRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF, | |||
12526 | const VarDecl *VD, | |||
12527 | Address VDAddr, | |||
12528 | SourceLocation Loc) { | |||
12529 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12529); | |||
12530 | } | |||
12531 | ||||
12532 | llvm::Function *CGOpenMPSIMDRuntime::emitThreadPrivateVarDefinition( | |||
12533 | const VarDecl *VD, Address VDAddr, SourceLocation Loc, bool PerformInit, | |||
12534 | CodeGenFunction *CGF) { | |||
12535 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12535); | |||
12536 | } | |||
12537 | ||||
12538 | Address CGOpenMPSIMDRuntime::getAddrOfArtificialThreadPrivate( | |||
12539 | CodeGenFunction &CGF, QualType VarType, StringRef Name) { | |||
12540 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12540); | |||
12541 | } | |||
12542 | ||||
12543 | void CGOpenMPSIMDRuntime::emitFlush(CodeGenFunction &CGF, | |||
12544 | ArrayRef<const Expr *> Vars, | |||
12545 | SourceLocation Loc, | |||
12546 | llvm::AtomicOrdering AO) { | |||
12547 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12547); | |||
12548 | } | |||
12549 | ||||
12550 | void CGOpenMPSIMDRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
12551 | const OMPExecutableDirective &D, | |||
12552 | llvm::Function *TaskFunction, | |||
12553 | QualType SharedsTy, Address Shareds, | |||
12554 | const Expr *IfCond, | |||
12555 | const OMPTaskDataTy &Data) { | |||
12556 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12556); | |||
12557 | } | |||
12558 | ||||
12559 | void CGOpenMPSIMDRuntime::emitTaskLoopCall( | |||
12560 | CodeGenFunction &CGF, SourceLocation Loc, const OMPLoopDirective &D, | |||
12561 | llvm::Function *TaskFunction, QualType SharedsTy, Address Shareds, | |||
12562 | const Expr *IfCond, const OMPTaskDataTy &Data) { | |||
12563 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12563); | |||
12564 | } | |||
12565 | ||||
12566 | void CGOpenMPSIMDRuntime::emitReduction( | |||
12567 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, | |||
12568 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, | |||
12569 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { | |||
12570 | assert(Options.SimpleReduction && "Only simple reduction is expected.")(static_cast <bool> (Options.SimpleReduction && "Only simple reduction is expected.") ? void (0) : __assert_fail ("Options.SimpleReduction && \"Only simple reduction is expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12570, __extension__ __PRETTY_FUNCTION__)); | |||
12571 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, | |||
12572 | ReductionOps, Options); | |||
12573 | } | |||
12574 | ||||
12575 | llvm::Value *CGOpenMPSIMDRuntime::emitTaskReductionInit( | |||
12576 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs, | |||
12577 | ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) { | |||
12578 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12578); | |||
12579 | } | |||
12580 | ||||
12581 | void CGOpenMPSIMDRuntime::emitTaskReductionFini(CodeGenFunction &CGF, | |||
12582 | SourceLocation Loc, | |||
12583 | bool IsWorksharingReduction) { | |||
12584 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12584); | |||
12585 | } | |||
12586 | ||||
12587 | void CGOpenMPSIMDRuntime::emitTaskReductionFixups(CodeGenFunction &CGF, | |||
12588 | SourceLocation Loc, | |||
12589 | ReductionCodeGen &RCG, | |||
12590 | unsigned N) { | |||
12591 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12591); | |||
12592 | } | |||
12593 | ||||
12594 | Address CGOpenMPSIMDRuntime::getTaskReductionItem(CodeGenFunction &CGF, | |||
12595 | SourceLocation Loc, | |||
12596 | llvm::Value *ReductionsPtr, | |||
12597 | LValue SharedLVal) { | |||
12598 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12598); | |||
12599 | } | |||
12600 | ||||
12601 | void CGOpenMPSIMDRuntime::emitTaskwaitCall(CodeGenFunction &CGF, | |||
12602 | SourceLocation Loc, | |||
12603 | const OMPTaskDataTy &Data) { | |||
12604 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12604); | |||
12605 | } | |||
12606 | ||||
12607 | void CGOpenMPSIMDRuntime::emitCancellationPointCall( | |||
12608 | CodeGenFunction &CGF, SourceLocation Loc, | |||
12609 | OpenMPDirectiveKind CancelRegion) { | |||
12610 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12610); | |||
12611 | } | |||
12612 | ||||
12613 | void CGOpenMPSIMDRuntime::emitCancelCall(CodeGenFunction &CGF, | |||
12614 | SourceLocation Loc, const Expr *IfCond, | |||
12615 | OpenMPDirectiveKind CancelRegion) { | |||
12616 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12616); | |||
12617 | } | |||
12618 | ||||
12619 | void CGOpenMPSIMDRuntime::emitTargetOutlinedFunction( | |||
12620 | const OMPExecutableDirective &D, StringRef ParentName, | |||
12621 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
12622 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
12623 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12623); | |||
12624 | } | |||
12625 | ||||
12626 | void CGOpenMPSIMDRuntime::emitTargetCall( | |||
12627 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | |||
12628 | llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, const Expr *IfCond, | |||
12629 | llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device, | |||
12630 | llvm::function_ref<llvm::Value *(CodeGenFunction &CGF, | |||
12631 | const OMPLoopDirective &D)> | |||
12632 | SizeEmitter) { | |||
12633 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12633); | |||
12634 | } | |||
12635 | ||||
12636 | bool CGOpenMPSIMDRuntime::emitTargetFunctions(GlobalDecl GD) { | |||
12637 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12637); | |||
12638 | } | |||
12639 | ||||
12640 | bool CGOpenMPSIMDRuntime::emitTargetGlobalVariable(GlobalDecl GD) { | |||
12641 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12641); | |||
12642 | } | |||
12643 | ||||
12644 | bool CGOpenMPSIMDRuntime::emitTargetGlobal(GlobalDecl GD) { | |||
12645 | return false; | |||
12646 | } | |||
12647 | ||||
12648 | void CGOpenMPSIMDRuntime::emitTeamsCall(CodeGenFunction &CGF, | |||
12649 | const OMPExecutableDirective &D, | |||
12650 | SourceLocation Loc, | |||
12651 | llvm::Function *OutlinedFn, | |||
12652 | ArrayRef<llvm::Value *> CapturedVars) { | |||
12653 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12653); | |||
12654 | } | |||
12655 | ||||
12656 | void CGOpenMPSIMDRuntime::emitNumTeamsClause(CodeGenFunction &CGF, | |||
12657 | const Expr *NumTeams, | |||
12658 | const Expr *ThreadLimit, | |||
12659 | SourceLocation Loc) { | |||
12660 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12660); | |||
12661 | } | |||
12662 | ||||
12663 | void CGOpenMPSIMDRuntime::emitTargetDataCalls( | |||
12664 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
12665 | const Expr *Device, const RegionCodeGenTy &CodeGen, | |||
12666 | CGOpenMPRuntime::TargetDataInfo &Info) { | |||
12667 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12667); | |||
12668 | } | |||
12669 | ||||
12670 | void CGOpenMPSIMDRuntime::emitTargetDataStandAloneCall( | |||
12671 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
12672 | const Expr *Device) { | |||
12673 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12673); | |||
12674 | } | |||
12675 | ||||
12676 | void CGOpenMPSIMDRuntime::emitDoacrossInit(CodeGenFunction &CGF, | |||
12677 | const OMPLoopDirective &D, | |||
12678 | ArrayRef<Expr *> NumIterations) { | |||
12679 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12679); | |||
12680 | } | |||
12681 | ||||
12682 | void CGOpenMPSIMDRuntime::emitDoacrossOrdered(CodeGenFunction &CGF, | |||
12683 | const OMPDependClause *C) { | |||
12684 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12684); | |||
12685 | } | |||
12686 | ||||
12687 | const VarDecl * | |||
12688 | CGOpenMPSIMDRuntime::translateParameter(const FieldDecl *FD, | |||
12689 | const VarDecl *NativeParam) const { | |||
12690 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12690); | |||
12691 | } | |||
12692 | ||||
12693 | Address | |||
12694 | CGOpenMPSIMDRuntime::getParameterAddress(CodeGenFunction &CGF, | |||
12695 | const VarDecl *NativeParam, | |||
12696 | const VarDecl *TargetParam) const { | |||
12697 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "clang/lib/CodeGen/CGOpenMPRuntime.cpp", 12697); | |||
12698 | } |