File: | llvm/include/llvm/IR/IRBuilder.h |
Warning: | line 2124, column 9 Called C++ object pointer is null |
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 "clang/AST/Attr.h" | ||||||
19 | #include "clang/AST/Decl.h" | ||||||
20 | #include "clang/AST/OpenMPClause.h" | ||||||
21 | #include "clang/AST/StmtOpenMP.h" | ||||||
22 | #include "clang/AST/StmtVisitor.h" | ||||||
23 | #include "clang/Basic/BitmaskEnum.h" | ||||||
24 | #include "clang/Basic/FileManager.h" | ||||||
25 | #include "clang/Basic/OpenMPKinds.h" | ||||||
26 | #include "clang/Basic/SourceManager.h" | ||||||
27 | #include "clang/CodeGen/ConstantInitBuilder.h" | ||||||
28 | #include "llvm/ADT/ArrayRef.h" | ||||||
29 | #include "llvm/ADT/SetOperations.h" | ||||||
30 | #include "llvm/ADT/StringExtras.h" | ||||||
31 | #include "llvm/Bitcode/BitcodeReader.h" | ||||||
32 | #include "llvm/IR/Constants.h" | ||||||
33 | #include "llvm/IR/DerivedTypes.h" | ||||||
34 | #include "llvm/IR/GlobalValue.h" | ||||||
35 | #include "llvm/IR/Value.h" | ||||||
36 | #include "llvm/Support/AtomicOrdering.h" | ||||||
37 | #include "llvm/Support/Format.h" | ||||||
38 | #include "llvm/Support/raw_ostream.h" | ||||||
39 | #include <cassert> | ||||||
40 | #include <numeric> | ||||||
41 | |||||||
42 | using namespace clang; | ||||||
43 | using namespace CodeGen; | ||||||
44 | using namespace llvm::omp; | ||||||
45 | |||||||
46 | namespace { | ||||||
47 | /// Base class for handling code generation inside OpenMP regions. | ||||||
48 | class CGOpenMPRegionInfo : public CodeGenFunction::CGCapturedStmtInfo { | ||||||
49 | public: | ||||||
50 | /// Kinds of OpenMP regions used in codegen. | ||||||
51 | enum CGOpenMPRegionKind { | ||||||
52 | /// Region with outlined function for standalone 'parallel' | ||||||
53 | /// directive. | ||||||
54 | ParallelOutlinedRegion, | ||||||
55 | /// Region with outlined function for standalone 'task' directive. | ||||||
56 | TaskOutlinedRegion, | ||||||
57 | /// Region for constructs that do not require function outlining, | ||||||
58 | /// like 'for', 'sections', 'atomic' etc. directives. | ||||||
59 | InlinedRegion, | ||||||
60 | /// Region with outlined function for standalone 'target' directive. | ||||||
61 | TargetRegion, | ||||||
62 | }; | ||||||
63 | |||||||
64 | CGOpenMPRegionInfo(const CapturedStmt &CS, | ||||||
65 | const CGOpenMPRegionKind RegionKind, | ||||||
66 | const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind, | ||||||
67 | bool HasCancel) | ||||||
68 | : CGCapturedStmtInfo(CS, CR_OpenMP), RegionKind(RegionKind), | ||||||
69 | CodeGen(CodeGen), Kind(Kind), HasCancel(HasCancel) {} | ||||||
70 | |||||||
71 | CGOpenMPRegionInfo(const CGOpenMPRegionKind RegionKind, | ||||||
72 | const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind, | ||||||
73 | bool HasCancel) | ||||||
74 | : CGCapturedStmtInfo(CR_OpenMP), RegionKind(RegionKind), CodeGen(CodeGen), | ||||||
75 | Kind(Kind), HasCancel(HasCancel) {} | ||||||
76 | |||||||
77 | /// Get a variable or parameter for storing global thread id | ||||||
78 | /// inside OpenMP construct. | ||||||
79 | virtual const VarDecl *getThreadIDVariable() const = 0; | ||||||
80 | |||||||
81 | /// Emit the captured statement body. | ||||||
82 | void EmitBody(CodeGenFunction &CGF, const Stmt *S) override; | ||||||
83 | |||||||
84 | /// Get an LValue for the current ThreadID variable. | ||||||
85 | /// \return LValue for thread id variable. This LValue always has type int32*. | ||||||
86 | virtual LValue getThreadIDVariableLValue(CodeGenFunction &CGF); | ||||||
87 | |||||||
88 | virtual void emitUntiedSwitch(CodeGenFunction & /*CGF*/) {} | ||||||
89 | |||||||
90 | CGOpenMPRegionKind getRegionKind() const { return RegionKind; } | ||||||
91 | |||||||
92 | OpenMPDirectiveKind getDirectiveKind() const { return Kind; } | ||||||
93 | |||||||
94 | bool hasCancel() const { return HasCancel; } | ||||||
95 | |||||||
96 | static bool classof(const CGCapturedStmtInfo *Info) { | ||||||
97 | return Info->getKind() == CR_OpenMP; | ||||||
98 | } | ||||||
99 | |||||||
100 | ~CGOpenMPRegionInfo() override = default; | ||||||
101 | |||||||
102 | protected: | ||||||
103 | CGOpenMPRegionKind RegionKind; | ||||||
104 | RegionCodeGenTy CodeGen; | ||||||
105 | OpenMPDirectiveKind Kind; | ||||||
106 | bool HasCancel; | ||||||
107 | }; | ||||||
108 | |||||||
109 | /// API for captured statement code generation in OpenMP constructs. | ||||||
110 | class CGOpenMPOutlinedRegionInfo final : public CGOpenMPRegionInfo { | ||||||
111 | public: | ||||||
112 | CGOpenMPOutlinedRegionInfo(const CapturedStmt &CS, const VarDecl *ThreadIDVar, | ||||||
113 | const RegionCodeGenTy &CodeGen, | ||||||
114 | OpenMPDirectiveKind Kind, bool HasCancel, | ||||||
115 | StringRef HelperName) | ||||||
116 | : CGOpenMPRegionInfo(CS, ParallelOutlinedRegion, CodeGen, Kind, | ||||||
117 | HasCancel), | ||||||
118 | ThreadIDVar(ThreadIDVar), HelperName(HelperName) { | ||||||
119 | assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.")((ThreadIDVar != nullptr && "No ThreadID in OpenMP region." ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar != nullptr && \"No ThreadID in OpenMP region.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 119, __PRETTY_FUNCTION__)); | ||||||
120 | } | ||||||
121 | |||||||
122 | /// Get a variable or parameter for storing global thread id | ||||||
123 | /// inside OpenMP construct. | ||||||
124 | const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; } | ||||||
125 | |||||||
126 | /// Get the name of the capture helper. | ||||||
127 | StringRef getHelperName() const override { return HelperName; } | ||||||
128 | |||||||
129 | static bool classof(const CGCapturedStmtInfo *Info) { | ||||||
130 | return CGOpenMPRegionInfo::classof(Info) && | ||||||
131 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == | ||||||
132 | ParallelOutlinedRegion; | ||||||
133 | } | ||||||
134 | |||||||
135 | private: | ||||||
136 | /// A variable or parameter storing global thread id for OpenMP | ||||||
137 | /// constructs. | ||||||
138 | const VarDecl *ThreadIDVar; | ||||||
139 | StringRef HelperName; | ||||||
140 | }; | ||||||
141 | |||||||
142 | /// API for captured statement code generation in OpenMP constructs. | ||||||
143 | class CGOpenMPTaskOutlinedRegionInfo final : public CGOpenMPRegionInfo { | ||||||
144 | public: | ||||||
145 | class UntiedTaskActionTy final : public PrePostActionTy { | ||||||
146 | bool Untied; | ||||||
147 | const VarDecl *PartIDVar; | ||||||
148 | const RegionCodeGenTy UntiedCodeGen; | ||||||
149 | llvm::SwitchInst *UntiedSwitch = nullptr; | ||||||
150 | |||||||
151 | public: | ||||||
152 | UntiedTaskActionTy(bool Tied, const VarDecl *PartIDVar, | ||||||
153 | const RegionCodeGenTy &UntiedCodeGen) | ||||||
154 | : Untied(!Tied), PartIDVar(PartIDVar), UntiedCodeGen(UntiedCodeGen) {} | ||||||
155 | void Enter(CodeGenFunction &CGF) override { | ||||||
156 | if (Untied) { | ||||||
157 | // Emit task switching point. | ||||||
158 | LValue PartIdLVal = CGF.EmitLoadOfPointerLValue( | ||||||
159 | CGF.GetAddrOfLocalVar(PartIDVar), | ||||||
160 | PartIDVar->getType()->castAs<PointerType>()); | ||||||
161 | llvm::Value *Res = | ||||||
162 | CGF.EmitLoadOfScalar(PartIdLVal, PartIDVar->getLocation()); | ||||||
163 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock(".untied.done."); | ||||||
164 | UntiedSwitch = CGF.Builder.CreateSwitch(Res, DoneBB); | ||||||
165 | CGF.EmitBlock(DoneBB); | ||||||
166 | CGF.EmitBranchThroughCleanup(CGF.ReturnBlock); | ||||||
167 | CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp.")); | ||||||
168 | UntiedSwitch->addCase(CGF.Builder.getInt32(0), | ||||||
169 | CGF.Builder.GetInsertBlock()); | ||||||
170 | emitUntiedSwitch(CGF); | ||||||
171 | } | ||||||
172 | } | ||||||
173 | void emitUntiedSwitch(CodeGenFunction &CGF) const { | ||||||
174 | if (Untied) { | ||||||
175 | LValue PartIdLVal = CGF.EmitLoadOfPointerLValue( | ||||||
176 | CGF.GetAddrOfLocalVar(PartIDVar), | ||||||
177 | PartIDVar->getType()->castAs<PointerType>()); | ||||||
178 | CGF.EmitStoreOfScalar(CGF.Builder.getInt32(UntiedSwitch->getNumCases()), | ||||||
179 | PartIdLVal); | ||||||
180 | UntiedCodeGen(CGF); | ||||||
181 | CodeGenFunction::JumpDest CurPoint = | ||||||
182 | CGF.getJumpDestInCurrentScope(".untied.next."); | ||||||
183 | CGF.EmitBranch(CGF.ReturnBlock.getBlock()); | ||||||
184 | CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp.")); | ||||||
185 | UntiedSwitch->addCase(CGF.Builder.getInt32(UntiedSwitch->getNumCases()), | ||||||
186 | CGF.Builder.GetInsertBlock()); | ||||||
187 | CGF.EmitBranchThroughCleanup(CurPoint); | ||||||
188 | CGF.EmitBlock(CurPoint.getBlock()); | ||||||
189 | } | ||||||
190 | } | ||||||
191 | unsigned getNumberOfParts() const { return UntiedSwitch->getNumCases(); } | ||||||
192 | }; | ||||||
193 | CGOpenMPTaskOutlinedRegionInfo(const CapturedStmt &CS, | ||||||
194 | const VarDecl *ThreadIDVar, | ||||||
195 | const RegionCodeGenTy &CodeGen, | ||||||
196 | OpenMPDirectiveKind Kind, bool HasCancel, | ||||||
197 | const UntiedTaskActionTy &Action) | ||||||
198 | : CGOpenMPRegionInfo(CS, TaskOutlinedRegion, CodeGen, Kind, HasCancel), | ||||||
199 | ThreadIDVar(ThreadIDVar), Action(Action) { | ||||||
200 | assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.")((ThreadIDVar != nullptr && "No ThreadID in OpenMP region." ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar != nullptr && \"No ThreadID in OpenMP region.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 200, __PRETTY_FUNCTION__)); | ||||||
201 | } | ||||||
202 | |||||||
203 | /// Get a variable or parameter for storing global thread id | ||||||
204 | /// inside OpenMP construct. | ||||||
205 | const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; } | ||||||
206 | |||||||
207 | /// Get an LValue for the current ThreadID variable. | ||||||
208 | LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override; | ||||||
209 | |||||||
210 | /// Get the name of the capture helper. | ||||||
211 | StringRef getHelperName() const override { return ".omp_outlined."; } | ||||||
212 | |||||||
213 | void emitUntiedSwitch(CodeGenFunction &CGF) override { | ||||||
214 | Action.emitUntiedSwitch(CGF); | ||||||
215 | } | ||||||
216 | |||||||
217 | static bool classof(const CGCapturedStmtInfo *Info) { | ||||||
218 | return CGOpenMPRegionInfo::classof(Info) && | ||||||
219 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == | ||||||
220 | TaskOutlinedRegion; | ||||||
221 | } | ||||||
222 | |||||||
223 | private: | ||||||
224 | /// A variable or parameter storing global thread id for OpenMP | ||||||
225 | /// constructs. | ||||||
226 | const VarDecl *ThreadIDVar; | ||||||
227 | /// Action for emitting code for untied tasks. | ||||||
228 | const UntiedTaskActionTy &Action; | ||||||
229 | }; | ||||||
230 | |||||||
231 | /// API for inlined captured statement code generation in OpenMP | ||||||
232 | /// constructs. | ||||||
233 | class CGOpenMPInlinedRegionInfo : public CGOpenMPRegionInfo { | ||||||
234 | public: | ||||||
235 | CGOpenMPInlinedRegionInfo(CodeGenFunction::CGCapturedStmtInfo *OldCSI, | ||||||
236 | const RegionCodeGenTy &CodeGen, | ||||||
237 | OpenMPDirectiveKind Kind, bool HasCancel) | ||||||
238 | : CGOpenMPRegionInfo(InlinedRegion, CodeGen, Kind, HasCancel), | ||||||
239 | OldCSI(OldCSI), | ||||||
240 | OuterRegionInfo(dyn_cast_or_null<CGOpenMPRegionInfo>(OldCSI)) {} | ||||||
241 | |||||||
242 | // Retrieve the value of the context parameter. | ||||||
243 | llvm::Value *getContextValue() const override { | ||||||
244 | if (OuterRegionInfo) | ||||||
245 | return OuterRegionInfo->getContextValue(); | ||||||
246 | llvm_unreachable("No context value for inlined OpenMP region")::llvm::llvm_unreachable_internal("No context value for inlined OpenMP region" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 246); | ||||||
247 | } | ||||||
248 | |||||||
249 | void setContextValue(llvm::Value *V) override { | ||||||
250 | if (OuterRegionInfo) { | ||||||
251 | OuterRegionInfo->setContextValue(V); | ||||||
252 | return; | ||||||
253 | } | ||||||
254 | llvm_unreachable("No context value for inlined OpenMP region")::llvm::llvm_unreachable_internal("No context value for inlined OpenMP region" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 254); | ||||||
255 | } | ||||||
256 | |||||||
257 | /// Lookup the captured field decl for a variable. | ||||||
258 | const FieldDecl *lookup(const VarDecl *VD) const override { | ||||||
259 | if (OuterRegionInfo) | ||||||
260 | return OuterRegionInfo->lookup(VD); | ||||||
261 | // If there is no outer outlined region,no need to lookup in a list of | ||||||
262 | // captured variables, we can use the original one. | ||||||
263 | return nullptr; | ||||||
264 | } | ||||||
265 | |||||||
266 | FieldDecl *getThisFieldDecl() const override { | ||||||
267 | if (OuterRegionInfo) | ||||||
268 | return OuterRegionInfo->getThisFieldDecl(); | ||||||
269 | return nullptr; | ||||||
270 | } | ||||||
271 | |||||||
272 | /// Get a variable or parameter for storing global thread id | ||||||
273 | /// inside OpenMP construct. | ||||||
274 | const VarDecl *getThreadIDVariable() const override { | ||||||
275 | if (OuterRegionInfo) | ||||||
276 | return OuterRegionInfo->getThreadIDVariable(); | ||||||
277 | return nullptr; | ||||||
278 | } | ||||||
279 | |||||||
280 | /// Get an LValue for the current ThreadID variable. | ||||||
281 | LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override { | ||||||
282 | if (OuterRegionInfo) | ||||||
283 | return OuterRegionInfo->getThreadIDVariableLValue(CGF); | ||||||
284 | llvm_unreachable("No LValue for inlined OpenMP construct")::llvm::llvm_unreachable_internal("No LValue for inlined OpenMP construct" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 284); | ||||||
285 | } | ||||||
286 | |||||||
287 | /// Get the name of the capture helper. | ||||||
288 | StringRef getHelperName() const override { | ||||||
289 | if (auto *OuterRegionInfo = getOldCSI()) | ||||||
290 | return OuterRegionInfo->getHelperName(); | ||||||
291 | llvm_unreachable("No helper name for inlined OpenMP construct")::llvm::llvm_unreachable_internal("No helper name for inlined OpenMP construct" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 291); | ||||||
292 | } | ||||||
293 | |||||||
294 | void emitUntiedSwitch(CodeGenFunction &CGF) override { | ||||||
295 | if (OuterRegionInfo) | ||||||
296 | OuterRegionInfo->emitUntiedSwitch(CGF); | ||||||
297 | } | ||||||
298 | |||||||
299 | CodeGenFunction::CGCapturedStmtInfo *getOldCSI() const { return OldCSI; } | ||||||
300 | |||||||
301 | static bool classof(const CGCapturedStmtInfo *Info) { | ||||||
302 | return CGOpenMPRegionInfo::classof(Info) && | ||||||
303 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == InlinedRegion; | ||||||
304 | } | ||||||
305 | |||||||
306 | ~CGOpenMPInlinedRegionInfo() override = default; | ||||||
307 | |||||||
308 | private: | ||||||
309 | /// CodeGen info about outer OpenMP region. | ||||||
310 | CodeGenFunction::CGCapturedStmtInfo *OldCSI; | ||||||
311 | CGOpenMPRegionInfo *OuterRegionInfo; | ||||||
312 | }; | ||||||
313 | |||||||
314 | /// API for captured statement code generation in OpenMP target | ||||||
315 | /// constructs. For this captures, implicit parameters are used instead of the | ||||||
316 | /// captured fields. The name of the target region has to be unique in a given | ||||||
317 | /// application so it is provided by the client, because only the client has | ||||||
318 | /// the information to generate that. | ||||||
319 | class CGOpenMPTargetRegionInfo final : public CGOpenMPRegionInfo { | ||||||
320 | public: | ||||||
321 | CGOpenMPTargetRegionInfo(const CapturedStmt &CS, | ||||||
322 | const RegionCodeGenTy &CodeGen, StringRef HelperName) | ||||||
323 | : CGOpenMPRegionInfo(CS, TargetRegion, CodeGen, OMPD_target, | ||||||
324 | /*HasCancel=*/false), | ||||||
325 | HelperName(HelperName) {} | ||||||
326 | |||||||
327 | /// This is unused for target regions because each starts executing | ||||||
328 | /// with a single thread. | ||||||
329 | const VarDecl *getThreadIDVariable() const override { return nullptr; } | ||||||
330 | |||||||
331 | /// Get the name of the capture helper. | ||||||
332 | StringRef getHelperName() const override { return HelperName; } | ||||||
333 | |||||||
334 | static bool classof(const CGCapturedStmtInfo *Info) { | ||||||
335 | return CGOpenMPRegionInfo::classof(Info) && | ||||||
336 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == TargetRegion; | ||||||
337 | } | ||||||
338 | |||||||
339 | private: | ||||||
340 | StringRef HelperName; | ||||||
341 | }; | ||||||
342 | |||||||
343 | static void EmptyCodeGen(CodeGenFunction &, PrePostActionTy &) { | ||||||
344 | llvm_unreachable("No codegen for expressions")::llvm::llvm_unreachable_internal("No codegen for expressions" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 344); | ||||||
345 | } | ||||||
346 | /// API for generation of expressions captured in a innermost OpenMP | ||||||
347 | /// region. | ||||||
348 | class CGOpenMPInnerExprInfo final : public CGOpenMPInlinedRegionInfo { | ||||||
349 | public: | ||||||
350 | CGOpenMPInnerExprInfo(CodeGenFunction &CGF, const CapturedStmt &CS) | ||||||
351 | : CGOpenMPInlinedRegionInfo(CGF.CapturedStmtInfo, EmptyCodeGen, | ||||||
352 | OMPD_unknown, | ||||||
353 | /*HasCancel=*/false), | ||||||
354 | PrivScope(CGF) { | ||||||
355 | // Make sure the globals captured in the provided statement are local by | ||||||
356 | // using the privatization logic. We assume the same variable is not | ||||||
357 | // captured more than once. | ||||||
358 | for (const auto &C : CS.captures()) { | ||||||
359 | if (!C.capturesVariable() && !C.capturesVariableByCopy()) | ||||||
360 | continue; | ||||||
361 | |||||||
362 | const VarDecl *VD = C.getCapturedVar(); | ||||||
363 | if (VD->isLocalVarDeclOrParm()) | ||||||
364 | continue; | ||||||
365 | |||||||
366 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD), | ||||||
367 | /*RefersToEnclosingVariableOrCapture=*/false, | ||||||
368 | VD->getType().getNonReferenceType(), VK_LValue, | ||||||
369 | C.getLocation()); | ||||||
370 | PrivScope.addPrivate( | ||||||
371 | VD, [&CGF, &DRE]() { return CGF.EmitLValue(&DRE).getAddress(CGF); }); | ||||||
372 | } | ||||||
373 | (void)PrivScope.Privatize(); | ||||||
374 | } | ||||||
375 | |||||||
376 | /// Lookup the captured field decl for a variable. | ||||||
377 | const FieldDecl *lookup(const VarDecl *VD) const override { | ||||||
378 | if (const FieldDecl *FD = CGOpenMPInlinedRegionInfo::lookup(VD)) | ||||||
379 | return FD; | ||||||
380 | return nullptr; | ||||||
381 | } | ||||||
382 | |||||||
383 | /// Emit the captured statement body. | ||||||
384 | void EmitBody(CodeGenFunction &CGF, const Stmt *S) override { | ||||||
385 | llvm_unreachable("No body for expressions")::llvm::llvm_unreachable_internal("No body for expressions", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 385); | ||||||
386 | } | ||||||
387 | |||||||
388 | /// Get a variable or parameter for storing global thread id | ||||||
389 | /// inside OpenMP construct. | ||||||
390 | const VarDecl *getThreadIDVariable() const override { | ||||||
391 | llvm_unreachable("No thread id for expressions")::llvm::llvm_unreachable_internal("No thread id for expressions" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 391); | ||||||
392 | } | ||||||
393 | |||||||
394 | /// Get the name of the capture helper. | ||||||
395 | StringRef getHelperName() const override { | ||||||
396 | llvm_unreachable("No helper name for expressions")::llvm::llvm_unreachable_internal("No helper name for expressions" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 396); | ||||||
397 | } | ||||||
398 | |||||||
399 | static bool classof(const CGCapturedStmtInfo *Info) { return false; } | ||||||
400 | |||||||
401 | private: | ||||||
402 | /// Private scope to capture global variables. | ||||||
403 | CodeGenFunction::OMPPrivateScope PrivScope; | ||||||
404 | }; | ||||||
405 | |||||||
406 | /// RAII for emitting code of OpenMP constructs. | ||||||
407 | class InlinedOpenMPRegionRAII { | ||||||
408 | CodeGenFunction &CGF; | ||||||
409 | llvm::DenseMap<const VarDecl *, FieldDecl *> LambdaCaptureFields; | ||||||
410 | FieldDecl *LambdaThisCaptureField = nullptr; | ||||||
411 | const CodeGen::CGBlockInfo *BlockInfo = nullptr; | ||||||
412 | |||||||
413 | public: | ||||||
414 | /// Constructs region for combined constructs. | ||||||
415 | /// \param CodeGen Code generation sequence for combined directives. Includes | ||||||
416 | /// a list of functions used for code generation of implicitly inlined | ||||||
417 | /// regions. | ||||||
418 | InlinedOpenMPRegionRAII(CodeGenFunction &CGF, const RegionCodeGenTy &CodeGen, | ||||||
419 | OpenMPDirectiveKind Kind, bool HasCancel) | ||||||
420 | : CGF(CGF) { | ||||||
421 | // Start emission for the construct. | ||||||
422 | CGF.CapturedStmtInfo = new CGOpenMPInlinedRegionInfo( | ||||||
423 | CGF.CapturedStmtInfo, CodeGen, Kind, HasCancel); | ||||||
424 | std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields); | ||||||
425 | LambdaThisCaptureField = CGF.LambdaThisCaptureField; | ||||||
426 | CGF.LambdaThisCaptureField = nullptr; | ||||||
427 | BlockInfo = CGF.BlockInfo; | ||||||
428 | CGF.BlockInfo = nullptr; | ||||||
429 | } | ||||||
430 | |||||||
431 | ~InlinedOpenMPRegionRAII() { | ||||||
432 | // Restore original CapturedStmtInfo only if we're done with code emission. | ||||||
433 | auto *OldCSI = | ||||||
434 | cast<CGOpenMPInlinedRegionInfo>(CGF.CapturedStmtInfo)->getOldCSI(); | ||||||
435 | delete CGF.CapturedStmtInfo; | ||||||
436 | CGF.CapturedStmtInfo = OldCSI; | ||||||
437 | std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields); | ||||||
438 | CGF.LambdaThisCaptureField = LambdaThisCaptureField; | ||||||
439 | CGF.BlockInfo = BlockInfo; | ||||||
440 | } | ||||||
441 | }; | ||||||
442 | |||||||
443 | /// Values for bit flags used in the ident_t to describe the fields. | ||||||
444 | /// All enumeric elements are named and described in accordance with the code | ||||||
445 | /// from https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp.h | ||||||
446 | enum OpenMPLocationFlags : unsigned { | ||||||
447 | /// Use trampoline for internal microtask. | ||||||
448 | OMP_IDENT_IMD = 0x01, | ||||||
449 | /// Use c-style ident structure. | ||||||
450 | OMP_IDENT_KMPC = 0x02, | ||||||
451 | /// Atomic reduction option for kmpc_reduce. | ||||||
452 | OMP_ATOMIC_REDUCE = 0x10, | ||||||
453 | /// Explicit 'barrier' directive. | ||||||
454 | OMP_IDENT_BARRIER_EXPL = 0x20, | ||||||
455 | /// Implicit barrier in code. | ||||||
456 | OMP_IDENT_BARRIER_IMPL = 0x40, | ||||||
457 | /// Implicit barrier in 'for' directive. | ||||||
458 | OMP_IDENT_BARRIER_IMPL_FOR = 0x40, | ||||||
459 | /// Implicit barrier in 'sections' directive. | ||||||
460 | OMP_IDENT_BARRIER_IMPL_SECTIONS = 0xC0, | ||||||
461 | /// Implicit barrier in 'single' directive. | ||||||
462 | OMP_IDENT_BARRIER_IMPL_SINGLE = 0x140, | ||||||
463 | /// Call of __kmp_for_static_init for static loop. | ||||||
464 | OMP_IDENT_WORK_LOOP = 0x200, | ||||||
465 | /// Call of __kmp_for_static_init for sections. | ||||||
466 | OMP_IDENT_WORK_SECTIONS = 0x400, | ||||||
467 | /// Call of __kmp_for_static_init for distribute. | ||||||
468 | OMP_IDENT_WORK_DISTRIBUTE = 0x800, | ||||||
469 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_IDENT_WORK_DISTRIBUTE)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_IDENT_WORK_DISTRIBUTE | ||||||
470 | }; | ||||||
471 | |||||||
472 | namespace { | ||||||
473 | 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^=; | ||||||
474 | /// Values for bit flags for marking which requires clauses have been used. | ||||||
475 | enum OpenMPOffloadingRequiresDirFlags : int64_t { | ||||||
476 | /// flag undefined. | ||||||
477 | OMP_REQ_UNDEFINED = 0x000, | ||||||
478 | /// no requires clause present. | ||||||
479 | OMP_REQ_NONE = 0x001, | ||||||
480 | /// reverse_offload clause. | ||||||
481 | OMP_REQ_REVERSE_OFFLOAD = 0x002, | ||||||
482 | /// unified_address clause. | ||||||
483 | OMP_REQ_UNIFIED_ADDRESS = 0x004, | ||||||
484 | /// unified_shared_memory clause. | ||||||
485 | OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008, | ||||||
486 | /// dynamic_allocators clause. | ||||||
487 | OMP_REQ_DYNAMIC_ALLOCATORS = 0x010, | ||||||
488 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_REQ_DYNAMIC_ALLOCATORS)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_REQ_DYNAMIC_ALLOCATORS | ||||||
489 | }; | ||||||
490 | |||||||
491 | enum OpenMPOffloadingReservedDeviceIDs { | ||||||
492 | /// Device ID if the device was not defined, runtime should get it | ||||||
493 | /// from environment variables in the spec. | ||||||
494 | OMP_DEVICEID_UNDEF = -1, | ||||||
495 | }; | ||||||
496 | } // anonymous namespace | ||||||
497 | |||||||
498 | /// Describes ident structure that describes a source location. | ||||||
499 | /// All descriptions are taken from | ||||||
500 | /// https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp.h | ||||||
501 | /// Original structure: | ||||||
502 | /// typedef struct ident { | ||||||
503 | /// kmp_int32 reserved_1; /**< might be used in Fortran; | ||||||
504 | /// see above */ | ||||||
505 | /// kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; | ||||||
506 | /// KMP_IDENT_KMPC identifies this union | ||||||
507 | /// member */ | ||||||
508 | /// kmp_int32 reserved_2; /**< not really used in Fortran any more; | ||||||
509 | /// see above */ | ||||||
510 | ///#if USE_ITT_BUILD | ||||||
511 | /// /* but currently used for storing | ||||||
512 | /// region-specific ITT */ | ||||||
513 | /// /* contextual information. */ | ||||||
514 | ///#endif /* USE_ITT_BUILD */ | ||||||
515 | /// kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for | ||||||
516 | /// C++ */ | ||||||
517 | /// char const *psource; /**< String describing the source location. | ||||||
518 | /// The string is composed of semi-colon separated | ||||||
519 | // fields which describe the source file, | ||||||
520 | /// the function and a pair of line numbers that | ||||||
521 | /// delimit the construct. | ||||||
522 | /// */ | ||||||
523 | /// } ident_t; | ||||||
524 | enum IdentFieldIndex { | ||||||
525 | /// might be used in Fortran | ||||||
526 | IdentField_Reserved_1, | ||||||
527 | /// OMP_IDENT_xxx flags; OMP_IDENT_KMPC identifies this union member. | ||||||
528 | IdentField_Flags, | ||||||
529 | /// Not really used in Fortran any more | ||||||
530 | IdentField_Reserved_2, | ||||||
531 | /// Source[4] in Fortran, do not use for C++ | ||||||
532 | IdentField_Reserved_3, | ||||||
533 | /// String describing the source location. The string is composed of | ||||||
534 | /// semi-colon separated fields which describe the source file, the function | ||||||
535 | /// and a pair of line numbers that delimit the construct. | ||||||
536 | IdentField_PSource | ||||||
537 | }; | ||||||
538 | |||||||
539 | /// Schedule types for 'omp for' loops (these enumerators are taken from | ||||||
540 | /// the enum sched_type in kmp.h). | ||||||
541 | enum OpenMPSchedType { | ||||||
542 | /// Lower bound for default (unordered) versions. | ||||||
543 | OMP_sch_lower = 32, | ||||||
544 | OMP_sch_static_chunked = 33, | ||||||
545 | OMP_sch_static = 34, | ||||||
546 | OMP_sch_dynamic_chunked = 35, | ||||||
547 | OMP_sch_guided_chunked = 36, | ||||||
548 | OMP_sch_runtime = 37, | ||||||
549 | OMP_sch_auto = 38, | ||||||
550 | /// static with chunk adjustment (e.g., simd) | ||||||
551 | OMP_sch_static_balanced_chunked = 45, | ||||||
552 | /// Lower bound for 'ordered' versions. | ||||||
553 | OMP_ord_lower = 64, | ||||||
554 | OMP_ord_static_chunked = 65, | ||||||
555 | OMP_ord_static = 66, | ||||||
556 | OMP_ord_dynamic_chunked = 67, | ||||||
557 | OMP_ord_guided_chunked = 68, | ||||||
558 | OMP_ord_runtime = 69, | ||||||
559 | OMP_ord_auto = 70, | ||||||
560 | OMP_sch_default = OMP_sch_static, | ||||||
561 | /// dist_schedule types | ||||||
562 | OMP_dist_sch_static_chunked = 91, | ||||||
563 | OMP_dist_sch_static = 92, | ||||||
564 | /// Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. | ||||||
565 | /// Set if the monotonic schedule modifier was present. | ||||||
566 | OMP_sch_modifier_monotonic = (1 << 29), | ||||||
567 | /// Set if the nonmonotonic schedule modifier was present. | ||||||
568 | OMP_sch_modifier_nonmonotonic = (1 << 30), | ||||||
569 | }; | ||||||
570 | |||||||
571 | /// A basic class for pre|post-action for advanced codegen sequence for OpenMP | ||||||
572 | /// region. | ||||||
573 | class CleanupTy final : public EHScopeStack::Cleanup { | ||||||
574 | PrePostActionTy *Action; | ||||||
575 | |||||||
576 | public: | ||||||
577 | explicit CleanupTy(PrePostActionTy *Action) : Action(Action) {} | ||||||
578 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | ||||||
579 | if (!CGF.HaveInsertPoint()) | ||||||
580 | return; | ||||||
581 | Action->Exit(CGF); | ||||||
582 | } | ||||||
583 | }; | ||||||
584 | |||||||
585 | } // anonymous namespace | ||||||
586 | |||||||
587 | void RegionCodeGenTy::operator()(CodeGenFunction &CGF) const { | ||||||
588 | CodeGenFunction::RunCleanupsScope Scope(CGF); | ||||||
589 | if (PrePostAction) { | ||||||
590 | CGF.EHStack.pushCleanup<CleanupTy>(NormalAndEHCleanup, PrePostAction); | ||||||
591 | Callback(CodeGen, CGF, *PrePostAction); | ||||||
592 | } else { | ||||||
593 | PrePostActionTy Action; | ||||||
594 | Callback(CodeGen, CGF, Action); | ||||||
595 | } | ||||||
596 | } | ||||||
597 | |||||||
598 | /// Check if the combiner is a call to UDR combiner and if it is so return the | ||||||
599 | /// UDR decl used for reduction. | ||||||
600 | static const OMPDeclareReductionDecl * | ||||||
601 | getReductionInit(const Expr *ReductionOp) { | ||||||
602 | if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) | ||||||
603 | if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee())) | ||||||
604 | if (const auto *DRE = | ||||||
605 | dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts())) | ||||||
606 | if (const auto *DRD = dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) | ||||||
607 | return DRD; | ||||||
608 | return nullptr; | ||||||
609 | } | ||||||
610 | |||||||
611 | static void emitInitWithReductionInitializer(CodeGenFunction &CGF, | ||||||
612 | const OMPDeclareReductionDecl *DRD, | ||||||
613 | const Expr *InitOp, | ||||||
614 | Address Private, Address Original, | ||||||
615 | QualType Ty) { | ||||||
616 | if (DRD->getInitializer()) { | ||||||
617 | std::pair<llvm::Function *, llvm::Function *> Reduction = | ||||||
618 | CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD); | ||||||
619 | const auto *CE = cast<CallExpr>(InitOp); | ||||||
620 | const auto *OVE = cast<OpaqueValueExpr>(CE->getCallee()); | ||||||
621 | const Expr *LHS = CE->getArg(/*Arg=*/0)->IgnoreParenImpCasts(); | ||||||
622 | const Expr *RHS = CE->getArg(/*Arg=*/1)->IgnoreParenImpCasts(); | ||||||
623 | const auto *LHSDRE = | ||||||
624 | cast<DeclRefExpr>(cast<UnaryOperator>(LHS)->getSubExpr()); | ||||||
625 | const auto *RHSDRE = | ||||||
626 | cast<DeclRefExpr>(cast<UnaryOperator>(RHS)->getSubExpr()); | ||||||
627 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | ||||||
628 | PrivateScope.addPrivate(cast<VarDecl>(LHSDRE->getDecl()), | ||||||
629 | [=]() { return Private; }); | ||||||
630 | PrivateScope.addPrivate(cast<VarDecl>(RHSDRE->getDecl()), | ||||||
631 | [=]() { return Original; }); | ||||||
632 | (void)PrivateScope.Privatize(); | ||||||
633 | RValue Func = RValue::get(Reduction.second); | ||||||
634 | CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func); | ||||||
635 | CGF.EmitIgnoredExpr(InitOp); | ||||||
636 | } else { | ||||||
637 | llvm::Constant *Init = CGF.CGM.EmitNullConstant(Ty); | ||||||
638 | std::string Name = CGF.CGM.getOpenMPRuntime().getName({"init"}); | ||||||
639 | auto *GV = new llvm::GlobalVariable( | ||||||
640 | CGF.CGM.getModule(), Init->getType(), /*isConstant=*/true, | ||||||
641 | llvm::GlobalValue::PrivateLinkage, Init, Name); | ||||||
642 | LValue LV = CGF.MakeNaturalAlignAddrLValue(GV, Ty); | ||||||
643 | RValue InitRVal; | ||||||
644 | switch (CGF.getEvaluationKind(Ty)) { | ||||||
645 | case TEK_Scalar: | ||||||
646 | InitRVal = CGF.EmitLoadOfLValue(LV, DRD->getLocation()); | ||||||
647 | break; | ||||||
648 | case TEK_Complex: | ||||||
649 | InitRVal = | ||||||
650 | RValue::getComplex(CGF.EmitLoadOfComplex(LV, DRD->getLocation())); | ||||||
651 | break; | ||||||
652 | case TEK_Aggregate: | ||||||
653 | InitRVal = RValue::getAggregate(LV.getAddress(CGF)); | ||||||
654 | break; | ||||||
655 | } | ||||||
656 | OpaqueValueExpr OVE(DRD->getLocation(), Ty, VK_RValue); | ||||||
657 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, InitRVal); | ||||||
658 | CGF.EmitAnyExprToMem(&OVE, Private, Ty.getQualifiers(), | ||||||
659 | /*IsInitializer=*/false); | ||||||
660 | } | ||||||
661 | } | ||||||
662 | |||||||
663 | /// Emit initialization of arrays of complex types. | ||||||
664 | /// \param DestAddr Address of the array. | ||||||
665 | /// \param Type Type of array. | ||||||
666 | /// \param Init Initial expression of array. | ||||||
667 | /// \param SrcAddr Address of the original array. | ||||||
668 | static void EmitOMPAggregateInit(CodeGenFunction &CGF, Address DestAddr, | ||||||
669 | QualType Type, bool EmitDeclareReductionInit, | ||||||
670 | const Expr *Init, | ||||||
671 | const OMPDeclareReductionDecl *DRD, | ||||||
672 | Address SrcAddr = Address::invalid()) { | ||||||
673 | // Perform element-by-element initialization. | ||||||
674 | QualType ElementTy; | ||||||
675 | |||||||
676 | // Drill down to the base element type on both arrays. | ||||||
677 | const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe(); | ||||||
678 | llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, DestAddr); | ||||||
679 | DestAddr = | ||||||
680 | CGF.Builder.CreateElementBitCast(DestAddr, DestAddr.getElementType()); | ||||||
681 | if (DRD) | ||||||
682 | SrcAddr = | ||||||
683 | CGF.Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType()); | ||||||
684 | |||||||
685 | llvm::Value *SrcBegin = nullptr; | ||||||
686 | if (DRD) | ||||||
687 | SrcBegin = SrcAddr.getPointer(); | ||||||
688 | llvm::Value *DestBegin = DestAddr.getPointer(); | ||||||
689 | // Cast from pointer to array type to pointer to single element. | ||||||
690 | llvm::Value *DestEnd = CGF.Builder.CreateGEP(DestBegin, NumElements); | ||||||
691 | // The basic structure here is a while-do loop. | ||||||
692 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arrayinit.body"); | ||||||
693 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arrayinit.done"); | ||||||
694 | llvm::Value *IsEmpty = | ||||||
695 | CGF.Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arrayinit.isempty"); | ||||||
696 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | ||||||
697 | |||||||
698 | // Enter the loop body, making that address the current address. | ||||||
699 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | ||||||
700 | CGF.EmitBlock(BodyBB); | ||||||
701 | |||||||
702 | CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy); | ||||||
703 | |||||||
704 | llvm::PHINode *SrcElementPHI = nullptr; | ||||||
705 | Address SrcElementCurrent = Address::invalid(); | ||||||
706 | if (DRD) { | ||||||
707 | SrcElementPHI = CGF.Builder.CreatePHI(SrcBegin->getType(), 2, | ||||||
708 | "omp.arraycpy.srcElementPast"); | ||||||
709 | SrcElementPHI->addIncoming(SrcBegin, EntryBB); | ||||||
710 | SrcElementCurrent = | ||||||
711 | Address(SrcElementPHI, | ||||||
712 | SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | ||||||
713 | } | ||||||
714 | llvm::PHINode *DestElementPHI = CGF.Builder.CreatePHI( | ||||||
715 | DestBegin->getType(), 2, "omp.arraycpy.destElementPast"); | ||||||
716 | DestElementPHI->addIncoming(DestBegin, EntryBB); | ||||||
717 | Address DestElementCurrent = | ||||||
718 | Address(DestElementPHI, | ||||||
719 | DestAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | ||||||
720 | |||||||
721 | // Emit copy. | ||||||
722 | { | ||||||
723 | CodeGenFunction::RunCleanupsScope InitScope(CGF); | ||||||
724 | if (EmitDeclareReductionInit) { | ||||||
725 | emitInitWithReductionInitializer(CGF, DRD, Init, DestElementCurrent, | ||||||
726 | SrcElementCurrent, ElementTy); | ||||||
727 | } else | ||||||
728 | CGF.EmitAnyExprToMem(Init, DestElementCurrent, ElementTy.getQualifiers(), | ||||||
729 | /*IsInitializer=*/false); | ||||||
730 | } | ||||||
731 | |||||||
732 | if (DRD) { | ||||||
733 | // Shift the address forward by one element. | ||||||
734 | llvm::Value *SrcElementNext = CGF.Builder.CreateConstGEP1_32( | ||||||
735 | SrcElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element"); | ||||||
736 | SrcElementPHI->addIncoming(SrcElementNext, CGF.Builder.GetInsertBlock()); | ||||||
737 | } | ||||||
738 | |||||||
739 | // Shift the address forward by one element. | ||||||
740 | llvm::Value *DestElementNext = CGF.Builder.CreateConstGEP1_32( | ||||||
741 | DestElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element"); | ||||||
742 | // Check whether we've reached the end. | ||||||
743 | llvm::Value *Done = | ||||||
744 | CGF.Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done"); | ||||||
745 | CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB); | ||||||
746 | DestElementPHI->addIncoming(DestElementNext, CGF.Builder.GetInsertBlock()); | ||||||
747 | |||||||
748 | // Done. | ||||||
749 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | ||||||
750 | } | ||||||
751 | |||||||
752 | LValue ReductionCodeGen::emitSharedLValue(CodeGenFunction &CGF, const Expr *E) { | ||||||
753 | return CGF.EmitOMPSharedLValue(E); | ||||||
754 | } | ||||||
755 | |||||||
756 | LValue ReductionCodeGen::emitSharedLValueUB(CodeGenFunction &CGF, | ||||||
757 | const Expr *E) { | ||||||
758 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(E)) | ||||||
759 | return CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false); | ||||||
760 | return LValue(); | ||||||
761 | } | ||||||
762 | |||||||
763 | void ReductionCodeGen::emitAggregateInitialization( | ||||||
764 | CodeGenFunction &CGF, unsigned N, Address PrivateAddr, LValue SharedLVal, | ||||||
765 | const OMPDeclareReductionDecl *DRD) { | ||||||
766 | // Emit VarDecl with copy init for arrays. | ||||||
767 | // Get the address of the original variable captured in current | ||||||
768 | // captured region. | ||||||
769 | const auto *PrivateVD = | ||||||
770 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | ||||||
771 | bool EmitDeclareReductionInit = | ||||||
772 | DRD && (DRD->getInitializer() || !PrivateVD->hasInit()); | ||||||
773 | EmitOMPAggregateInit(CGF, PrivateAddr, PrivateVD->getType(), | ||||||
774 | EmitDeclareReductionInit, | ||||||
775 | EmitDeclareReductionInit ? ClausesData[N].ReductionOp | ||||||
776 | : PrivateVD->getInit(), | ||||||
777 | DRD, SharedLVal.getAddress(CGF)); | ||||||
778 | } | ||||||
779 | |||||||
780 | ReductionCodeGen::ReductionCodeGen(ArrayRef<const Expr *> Shareds, | ||||||
781 | ArrayRef<const Expr *> Origs, | ||||||
782 | ArrayRef<const Expr *> Privates, | ||||||
783 | ArrayRef<const Expr *> ReductionOps) { | ||||||
784 | ClausesData.reserve(Shareds.size()); | ||||||
785 | SharedAddresses.reserve(Shareds.size()); | ||||||
786 | Sizes.reserve(Shareds.size()); | ||||||
787 | BaseDecls.reserve(Shareds.size()); | ||||||
788 | const auto *IOrig = Origs.begin(); | ||||||
789 | const auto *IPriv = Privates.begin(); | ||||||
790 | const auto *IRed = ReductionOps.begin(); | ||||||
791 | for (const Expr *Ref : Shareds) { | ||||||
792 | ClausesData.emplace_back(Ref, *IOrig, *IPriv, *IRed); | ||||||
793 | std::advance(IOrig, 1); | ||||||
794 | std::advance(IPriv, 1); | ||||||
795 | std::advance(IRed, 1); | ||||||
796 | } | ||||||
797 | } | ||||||
798 | |||||||
799 | void ReductionCodeGen::emitSharedOrigLValue(CodeGenFunction &CGF, unsigned N) { | ||||||
800 | assert(SharedAddresses.size() == N && OrigAddresses.size() == N &&((SharedAddresses.size() == N && OrigAddresses.size() == N && "Number of generated lvalues must be exactly N." ) ? static_cast<void> (0) : __assert_fail ("SharedAddresses.size() == N && OrigAddresses.size() == N && \"Number of generated lvalues must be exactly N.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 801, __PRETTY_FUNCTION__)) | ||||||
801 | "Number of generated lvalues must be exactly N.")((SharedAddresses.size() == N && OrigAddresses.size() == N && "Number of generated lvalues must be exactly N." ) ? static_cast<void> (0) : __assert_fail ("SharedAddresses.size() == N && OrigAddresses.size() == N && \"Number of generated lvalues must be exactly N.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 801, __PRETTY_FUNCTION__)); | ||||||
802 | LValue First = emitSharedLValue(CGF, ClausesData[N].Shared); | ||||||
803 | LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Shared); | ||||||
804 | SharedAddresses.emplace_back(First, Second); | ||||||
805 | if (ClausesData[N].Shared == ClausesData[N].Ref) { | ||||||
806 | OrigAddresses.emplace_back(First, Second); | ||||||
807 | } else { | ||||||
808 | LValue First = emitSharedLValue(CGF, ClausesData[N].Ref); | ||||||
809 | LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Ref); | ||||||
810 | OrigAddresses.emplace_back(First, Second); | ||||||
811 | } | ||||||
812 | } | ||||||
813 | |||||||
814 | void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N) { | ||||||
815 | const auto *PrivateVD = | ||||||
816 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | ||||||
817 | QualType PrivateType = PrivateVD->getType(); | ||||||
818 | bool AsArraySection = isa<OMPArraySectionExpr>(ClausesData[N].Ref); | ||||||
819 | if (!PrivateType->isVariablyModifiedType()) { | ||||||
820 | Sizes.emplace_back( | ||||||
821 | CGF.getTypeSize(OrigAddresses[N].first.getType().getNonReferenceType()), | ||||||
822 | nullptr); | ||||||
823 | return; | ||||||
824 | } | ||||||
825 | llvm::Value *Size; | ||||||
826 | llvm::Value *SizeInChars; | ||||||
827 | auto *ElemType = | ||||||
828 | cast<llvm::PointerType>(OrigAddresses[N].first.getPointer(CGF)->getType()) | ||||||
829 | ->getElementType(); | ||||||
830 | auto *ElemSizeOf = llvm::ConstantExpr::getSizeOf(ElemType); | ||||||
831 | if (AsArraySection) { | ||||||
832 | Size = CGF.Builder.CreatePtrDiff(OrigAddresses[N].second.getPointer(CGF), | ||||||
833 | OrigAddresses[N].first.getPointer(CGF)); | ||||||
834 | Size = CGF.Builder.CreateNUWAdd( | ||||||
835 | Size, llvm::ConstantInt::get(Size->getType(), /*V=*/1)); | ||||||
836 | SizeInChars = CGF.Builder.CreateNUWMul(Size, ElemSizeOf); | ||||||
837 | } else { | ||||||
838 | SizeInChars = | ||||||
839 | CGF.getTypeSize(OrigAddresses[N].first.getType().getNonReferenceType()); | ||||||
840 | Size = CGF.Builder.CreateExactUDiv(SizeInChars, ElemSizeOf); | ||||||
841 | } | ||||||
842 | Sizes.emplace_back(SizeInChars, Size); | ||||||
843 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | ||||||
844 | CGF, | ||||||
845 | cast<OpaqueValueExpr>( | ||||||
846 | CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()), | ||||||
847 | RValue::get(Size)); | ||||||
848 | CGF.EmitVariablyModifiedType(PrivateType); | ||||||
849 | } | ||||||
850 | |||||||
851 | void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N, | ||||||
852 | llvm::Value *Size) { | ||||||
853 | const auto *PrivateVD = | ||||||
854 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | ||||||
855 | QualType PrivateType = PrivateVD->getType(); | ||||||
856 | if (!PrivateType->isVariablyModifiedType()) { | ||||||
857 | assert(!Size && !Sizes[N].second &&((!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? static_cast<void> (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 859, __PRETTY_FUNCTION__)) | ||||||
858 | "Size should be nullptr for non-variably modified reduction "((!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? static_cast<void> (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 859, __PRETTY_FUNCTION__)) | ||||||
859 | "items.")((!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? static_cast<void> (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 859, __PRETTY_FUNCTION__)); | ||||||
860 | return; | ||||||
861 | } | ||||||
862 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | ||||||
863 | CGF, | ||||||
864 | cast<OpaqueValueExpr>( | ||||||
865 | CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()), | ||||||
866 | RValue::get(Size)); | ||||||
867 | CGF.EmitVariablyModifiedType(PrivateType); | ||||||
868 | } | ||||||
869 | |||||||
870 | void ReductionCodeGen::emitInitialization( | ||||||
871 | CodeGenFunction &CGF, unsigned N, Address PrivateAddr, LValue SharedLVal, | ||||||
872 | llvm::function_ref<bool(CodeGenFunction &)> DefaultInit) { | ||||||
873 | assert(SharedAddresses.size() > N && "No variable was generated")((SharedAddresses.size() > N && "No variable was generated" ) ? static_cast<void> (0) : __assert_fail ("SharedAddresses.size() > N && \"No variable was generated\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 873, __PRETTY_FUNCTION__)); | ||||||
874 | const auto *PrivateVD = | ||||||
875 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | ||||||
876 | const OMPDeclareReductionDecl *DRD = | ||||||
877 | getReductionInit(ClausesData[N].ReductionOp); | ||||||
878 | QualType PrivateType = PrivateVD->getType(); | ||||||
879 | PrivateAddr = CGF.Builder.CreateElementBitCast( | ||||||
880 | PrivateAddr, CGF.ConvertTypeForMem(PrivateType)); | ||||||
881 | QualType SharedType = SharedAddresses[N].first.getType(); | ||||||
882 | SharedLVal = CGF.MakeAddrLValue( | ||||||
883 | CGF.Builder.CreateElementBitCast(SharedLVal.getAddress(CGF), | ||||||
884 | CGF.ConvertTypeForMem(SharedType)), | ||||||
885 | SharedType, SharedAddresses[N].first.getBaseInfo(), | ||||||
886 | CGF.CGM.getTBAAInfoForSubobject(SharedAddresses[N].first, SharedType)); | ||||||
887 | if (CGF.getContext().getAsArrayType(PrivateVD->getType())) { | ||||||
888 | if (DRD && DRD->getInitializer()) | ||||||
889 | (void)DefaultInit(CGF); | ||||||
890 | emitAggregateInitialization(CGF, N, PrivateAddr, SharedLVal, DRD); | ||||||
891 | } else if (DRD && (DRD->getInitializer() || !PrivateVD->hasInit())) { | ||||||
892 | (void)DefaultInit(CGF); | ||||||
893 | emitInitWithReductionInitializer(CGF, DRD, ClausesData[N].ReductionOp, | ||||||
894 | PrivateAddr, SharedLVal.getAddress(CGF), | ||||||
895 | SharedLVal.getType()); | ||||||
896 | } else if (!DefaultInit(CGF) && PrivateVD->hasInit() && | ||||||
897 | !CGF.isTrivialInitializer(PrivateVD->getInit())) { | ||||||
898 | CGF.EmitAnyExprToMem(PrivateVD->getInit(), PrivateAddr, | ||||||
899 | PrivateVD->getType().getQualifiers(), | ||||||
900 | /*IsInitializer=*/false); | ||||||
901 | } | ||||||
902 | } | ||||||
903 | |||||||
904 | bool ReductionCodeGen::needCleanups(unsigned N) { | ||||||
905 | const auto *PrivateVD = | ||||||
906 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | ||||||
907 | QualType PrivateType = PrivateVD->getType(); | ||||||
908 | QualType::DestructionKind DTorKind = PrivateType.isDestructedType(); | ||||||
909 | return DTorKind != QualType::DK_none; | ||||||
910 | } | ||||||
911 | |||||||
912 | void ReductionCodeGen::emitCleanups(CodeGenFunction &CGF, unsigned N, | ||||||
913 | Address PrivateAddr) { | ||||||
914 | const auto *PrivateVD = | ||||||
915 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | ||||||
916 | QualType PrivateType = PrivateVD->getType(); | ||||||
917 | QualType::DestructionKind DTorKind = PrivateType.isDestructedType(); | ||||||
918 | if (needCleanups(N)) { | ||||||
919 | PrivateAddr = CGF.Builder.CreateElementBitCast( | ||||||
920 | PrivateAddr, CGF.ConvertTypeForMem(PrivateType)); | ||||||
921 | CGF.pushDestroy(DTorKind, PrivateAddr, PrivateType); | ||||||
922 | } | ||||||
923 | } | ||||||
924 | |||||||
925 | static LValue loadToBegin(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy, | ||||||
926 | LValue BaseLV) { | ||||||
927 | BaseTy = BaseTy.getNonReferenceType(); | ||||||
928 | while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) && | ||||||
929 | !CGF.getContext().hasSameType(BaseTy, ElTy)) { | ||||||
930 | if (const auto *PtrTy = BaseTy->getAs<PointerType>()) { | ||||||
931 | BaseLV = CGF.EmitLoadOfPointerLValue(BaseLV.getAddress(CGF), PtrTy); | ||||||
932 | } else { | ||||||
933 | LValue RefLVal = CGF.MakeAddrLValue(BaseLV.getAddress(CGF), BaseTy); | ||||||
934 | BaseLV = CGF.EmitLoadOfReferenceLValue(RefLVal); | ||||||
935 | } | ||||||
936 | BaseTy = BaseTy->getPointeeType(); | ||||||
937 | } | ||||||
938 | return CGF.MakeAddrLValue( | ||||||
939 | CGF.Builder.CreateElementBitCast(BaseLV.getAddress(CGF), | ||||||
940 | CGF.ConvertTypeForMem(ElTy)), | ||||||
941 | BaseLV.getType(), BaseLV.getBaseInfo(), | ||||||
942 | CGF.CGM.getTBAAInfoForSubobject(BaseLV, BaseLV.getType())); | ||||||
943 | } | ||||||
944 | |||||||
945 | static Address castToBase(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy, | ||||||
946 | llvm::Type *BaseLVType, CharUnits BaseLVAlignment, | ||||||
947 | llvm::Value *Addr) { | ||||||
948 | Address Tmp = Address::invalid(); | ||||||
949 | Address TopTmp = Address::invalid(); | ||||||
950 | Address MostTopTmp = Address::invalid(); | ||||||
951 | BaseTy = BaseTy.getNonReferenceType(); | ||||||
952 | while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) && | ||||||
953 | !CGF.getContext().hasSameType(BaseTy, ElTy)) { | ||||||
954 | Tmp = CGF.CreateMemTemp(BaseTy); | ||||||
955 | if (TopTmp.isValid()) | ||||||
956 | CGF.Builder.CreateStore(Tmp.getPointer(), TopTmp); | ||||||
957 | else | ||||||
958 | MostTopTmp = Tmp; | ||||||
959 | TopTmp = Tmp; | ||||||
960 | BaseTy = BaseTy->getPointeeType(); | ||||||
961 | } | ||||||
962 | llvm::Type *Ty = BaseLVType; | ||||||
963 | if (Tmp.isValid()) | ||||||
964 | Ty = Tmp.getElementType(); | ||||||
965 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Addr, Ty); | ||||||
966 | if (Tmp.isValid()) { | ||||||
967 | CGF.Builder.CreateStore(Addr, Tmp); | ||||||
968 | return MostTopTmp; | ||||||
969 | } | ||||||
970 | return Address(Addr, BaseLVAlignment); | ||||||
971 | } | ||||||
972 | |||||||
973 | static const VarDecl *getBaseDecl(const Expr *Ref, const DeclRefExpr *&DE) { | ||||||
974 | const VarDecl *OrigVD = nullptr; | ||||||
975 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(Ref)) { | ||||||
976 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); | ||||||
977 | while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) | ||||||
978 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); | ||||||
979 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | ||||||
980 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | ||||||
981 | DE = cast<DeclRefExpr>(Base); | ||||||
982 | OrigVD = cast<VarDecl>(DE->getDecl()); | ||||||
983 | } else if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Ref)) { | ||||||
984 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); | ||||||
985 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | ||||||
986 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | ||||||
987 | DE = cast<DeclRefExpr>(Base); | ||||||
988 | OrigVD = cast<VarDecl>(DE->getDecl()); | ||||||
989 | } | ||||||
990 | return OrigVD; | ||||||
991 | } | ||||||
992 | |||||||
993 | Address ReductionCodeGen::adjustPrivateAddress(CodeGenFunction &CGF, unsigned N, | ||||||
994 | Address PrivateAddr) { | ||||||
995 | const DeclRefExpr *DE; | ||||||
996 | if (const VarDecl *OrigVD = ::getBaseDecl(ClausesData[N].Ref, DE)) { | ||||||
997 | BaseDecls.emplace_back(OrigVD); | ||||||
998 | LValue OriginalBaseLValue = CGF.EmitLValue(DE); | ||||||
999 | LValue BaseLValue = | ||||||
1000 | loadToBegin(CGF, OrigVD->getType(), SharedAddresses[N].first.getType(), | ||||||
1001 | OriginalBaseLValue); | ||||||
1002 | llvm::Value *Adjustment = CGF.Builder.CreatePtrDiff( | ||||||
1003 | BaseLValue.getPointer(CGF), SharedAddresses[N].first.getPointer(CGF)); | ||||||
1004 | llvm::Value *PrivatePointer = | ||||||
1005 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
1006 | PrivateAddr.getPointer(), | ||||||
1007 | SharedAddresses[N].first.getAddress(CGF).getType()); | ||||||
1008 | llvm::Value *Ptr = CGF.Builder.CreateGEP(PrivatePointer, Adjustment); | ||||||
1009 | return castToBase(CGF, OrigVD->getType(), | ||||||
1010 | SharedAddresses[N].first.getType(), | ||||||
1011 | OriginalBaseLValue.getAddress(CGF).getType(), | ||||||
1012 | OriginalBaseLValue.getAlignment(), Ptr); | ||||||
1013 | } | ||||||
1014 | BaseDecls.emplace_back( | ||||||
1015 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Ref)->getDecl())); | ||||||
1016 | return PrivateAddr; | ||||||
1017 | } | ||||||
1018 | |||||||
1019 | bool ReductionCodeGen::usesReductionInitializer(unsigned N) const { | ||||||
1020 | const OMPDeclareReductionDecl *DRD = | ||||||
1021 | getReductionInit(ClausesData[N].ReductionOp); | ||||||
1022 | return DRD && DRD->getInitializer(); | ||||||
1023 | } | ||||||
1024 | |||||||
1025 | LValue CGOpenMPRegionInfo::getThreadIDVariableLValue(CodeGenFunction &CGF) { | ||||||
1026 | return CGF.EmitLoadOfPointerLValue( | ||||||
1027 | CGF.GetAddrOfLocalVar(getThreadIDVariable()), | ||||||
1028 | getThreadIDVariable()->getType()->castAs<PointerType>()); | ||||||
1029 | } | ||||||
1030 | |||||||
1031 | void CGOpenMPRegionInfo::EmitBody(CodeGenFunction &CGF, const Stmt * /*S*/) { | ||||||
1032 | if (!CGF.HaveInsertPoint()) | ||||||
1033 | return; | ||||||
1034 | // 1.2.2 OpenMP Language Terminology | ||||||
1035 | // Structured block - An executable statement with a single entry at the | ||||||
1036 | // top and a single exit at the bottom. | ||||||
1037 | // The point of exit cannot be a branch out of the structured block. | ||||||
1038 | // longjmp() and throw() must not violate the entry/exit criteria. | ||||||
1039 | CGF.EHStack.pushTerminate(); | ||||||
1040 | CodeGen(CGF); | ||||||
1041 | CGF.EHStack.popTerminate(); | ||||||
1042 | } | ||||||
1043 | |||||||
1044 | LValue CGOpenMPTaskOutlinedRegionInfo::getThreadIDVariableLValue( | ||||||
1045 | CodeGenFunction &CGF) { | ||||||
1046 | return CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(getThreadIDVariable()), | ||||||
1047 | getThreadIDVariable()->getType(), | ||||||
1048 | AlignmentSource::Decl); | ||||||
1049 | } | ||||||
1050 | |||||||
1051 | static FieldDecl *addFieldToRecordDecl(ASTContext &C, DeclContext *DC, | ||||||
1052 | QualType FieldTy) { | ||||||
1053 | auto *Field = FieldDecl::Create( | ||||||
1054 | C, DC, SourceLocation(), SourceLocation(), /*Id=*/nullptr, FieldTy, | ||||||
1055 | C.getTrivialTypeSourceInfo(FieldTy, SourceLocation()), | ||||||
1056 | /*BW=*/nullptr, /*Mutable=*/false, /*InitStyle=*/ICIS_NoInit); | ||||||
1057 | Field->setAccess(AS_public); | ||||||
1058 | DC->addDecl(Field); | ||||||
1059 | return Field; | ||||||
1060 | } | ||||||
1061 | |||||||
1062 | CGOpenMPRuntime::CGOpenMPRuntime(CodeGenModule &CGM, StringRef FirstSeparator, | ||||||
1063 | StringRef Separator) | ||||||
1064 | : CGM(CGM), FirstSeparator(FirstSeparator), Separator(Separator), | ||||||
1065 | OMPBuilder(CGM.getModule()), OffloadEntriesInfoManager(CGM) { | ||||||
1066 | KmpCriticalNameTy = llvm::ArrayType::get(CGM.Int32Ty, /*NumElements*/ 8); | ||||||
1067 | |||||||
1068 | // Initialize Types used in OpenMPIRBuilder from OMPKinds.def | ||||||
1069 | OMPBuilder.initialize(); | ||||||
1070 | loadOffloadInfoMetadata(); | ||||||
1071 | } | ||||||
1072 | |||||||
1073 | void CGOpenMPRuntime::clear() { | ||||||
1074 | InternalVars.clear(); | ||||||
1075 | // Clean non-target variable declarations possibly used only in debug info. | ||||||
1076 | for (const auto &Data : EmittedNonTargetVariables) { | ||||||
1077 | if (!Data.getValue().pointsToAliveValue()) | ||||||
1078 | continue; | ||||||
1079 | auto *GV = dyn_cast<llvm::GlobalVariable>(Data.getValue()); | ||||||
1080 | if (!GV) | ||||||
1081 | continue; | ||||||
1082 | if (!GV->isDeclaration() || GV->getNumUses() > 0) | ||||||
1083 | continue; | ||||||
1084 | GV->eraseFromParent(); | ||||||
1085 | } | ||||||
1086 | } | ||||||
1087 | |||||||
1088 | std::string CGOpenMPRuntime::getName(ArrayRef<StringRef> Parts) const { | ||||||
1089 | SmallString<128> Buffer; | ||||||
1090 | llvm::raw_svector_ostream OS(Buffer); | ||||||
1091 | StringRef Sep = FirstSeparator; | ||||||
1092 | for (StringRef Part : Parts) { | ||||||
1093 | OS << Sep << Part; | ||||||
1094 | Sep = Separator; | ||||||
1095 | } | ||||||
1096 | return std::string(OS.str()); | ||||||
1097 | } | ||||||
1098 | |||||||
1099 | static llvm::Function * | ||||||
1100 | emitCombinerOrInitializer(CodeGenModule &CGM, QualType Ty, | ||||||
1101 | const Expr *CombinerInitializer, const VarDecl *In, | ||||||
1102 | const VarDecl *Out, bool IsCombiner) { | ||||||
1103 | // void .omp_combiner.(Ty *in, Ty *out); | ||||||
1104 | ASTContext &C = CGM.getContext(); | ||||||
1105 | QualType PtrTy = C.getPointerType(Ty).withRestrict(); | ||||||
1106 | FunctionArgList Args; | ||||||
1107 | ImplicitParamDecl OmpOutParm(C, /*DC=*/nullptr, Out->getLocation(), | ||||||
1108 | /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other); | ||||||
1109 | ImplicitParamDecl OmpInParm(C, /*DC=*/nullptr, In->getLocation(), | ||||||
1110 | /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other); | ||||||
1111 | Args.push_back(&OmpOutParm); | ||||||
1112 | Args.push_back(&OmpInParm); | ||||||
1113 | const CGFunctionInfo &FnInfo = | ||||||
1114 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
1115 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | ||||||
1116 | std::string Name = CGM.getOpenMPRuntime().getName( | ||||||
1117 | {IsCombiner ? "omp_combiner" : "omp_initializer", ""}); | ||||||
1118 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | ||||||
1119 | Name, &CGM.getModule()); | ||||||
1120 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | ||||||
1121 | if (CGM.getLangOpts().Optimize) { | ||||||
1122 | Fn->removeFnAttr(llvm::Attribute::NoInline); | ||||||
1123 | Fn->removeFnAttr(llvm::Attribute::OptimizeNone); | ||||||
1124 | Fn->addFnAttr(llvm::Attribute::AlwaysInline); | ||||||
1125 | } | ||||||
1126 | CodeGenFunction CGF(CGM); | ||||||
1127 | // Map "T omp_in;" variable to "*omp_in_parm" value in all expressions. | ||||||
1128 | // Map "T omp_out;" variable to "*omp_out_parm" value in all expressions. | ||||||
1129 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, In->getLocation(), | ||||||
1130 | Out->getLocation()); | ||||||
1131 | CodeGenFunction::OMPPrivateScope Scope(CGF); | ||||||
1132 | Address AddrIn = CGF.GetAddrOfLocalVar(&OmpInParm); | ||||||
1133 | Scope.addPrivate(In, [&CGF, AddrIn, PtrTy]() { | ||||||
1134 | return CGF.EmitLoadOfPointerLValue(AddrIn, PtrTy->castAs<PointerType>()) | ||||||
1135 | .getAddress(CGF); | ||||||
1136 | }); | ||||||
1137 | Address AddrOut = CGF.GetAddrOfLocalVar(&OmpOutParm); | ||||||
1138 | Scope.addPrivate(Out, [&CGF, AddrOut, PtrTy]() { | ||||||
1139 | return CGF.EmitLoadOfPointerLValue(AddrOut, PtrTy->castAs<PointerType>()) | ||||||
1140 | .getAddress(CGF); | ||||||
1141 | }); | ||||||
1142 | (void)Scope.Privatize(); | ||||||
1143 | if (!IsCombiner && Out->hasInit() && | ||||||
1144 | !CGF.isTrivialInitializer(Out->getInit())) { | ||||||
1145 | CGF.EmitAnyExprToMem(Out->getInit(), CGF.GetAddrOfLocalVar(Out), | ||||||
1146 | Out->getType().getQualifiers(), | ||||||
1147 | /*IsInitializer=*/true); | ||||||
1148 | } | ||||||
1149 | if (CombinerInitializer) | ||||||
1150 | CGF.EmitIgnoredExpr(CombinerInitializer); | ||||||
1151 | Scope.ForceCleanup(); | ||||||
1152 | CGF.FinishFunction(); | ||||||
1153 | return Fn; | ||||||
1154 | } | ||||||
1155 | |||||||
1156 | void CGOpenMPRuntime::emitUserDefinedReduction( | ||||||
1157 | CodeGenFunction *CGF, const OMPDeclareReductionDecl *D) { | ||||||
1158 | if (UDRMap.count(D) > 0) | ||||||
1159 | return; | ||||||
1160 | llvm::Function *Combiner = emitCombinerOrInitializer( | ||||||
1161 | CGM, D->getType(), D->getCombiner(), | ||||||
1162 | cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerIn())->getDecl()), | ||||||
1163 | cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerOut())->getDecl()), | ||||||
1164 | /*IsCombiner=*/true); | ||||||
1165 | llvm::Function *Initializer = nullptr; | ||||||
1166 | if (const Expr *Init = D->getInitializer()) { | ||||||
1167 | Initializer = emitCombinerOrInitializer( | ||||||
1168 | CGM, D->getType(), | ||||||
1169 | D->getInitializerKind() == OMPDeclareReductionDecl::CallInit ? Init | ||||||
1170 | : nullptr, | ||||||
1171 | cast<VarDecl>(cast<DeclRefExpr>(D->getInitOrig())->getDecl()), | ||||||
1172 | cast<VarDecl>(cast<DeclRefExpr>(D->getInitPriv())->getDecl()), | ||||||
1173 | /*IsCombiner=*/false); | ||||||
1174 | } | ||||||
1175 | UDRMap.try_emplace(D, Combiner, Initializer); | ||||||
1176 | if (CGF) { | ||||||
1177 | auto &Decls = FunctionUDRMap.FindAndConstruct(CGF->CurFn); | ||||||
1178 | Decls.second.push_back(D); | ||||||
1179 | } | ||||||
1180 | } | ||||||
1181 | |||||||
1182 | std::pair<llvm::Function *, llvm::Function *> | ||||||
1183 | CGOpenMPRuntime::getUserDefinedReduction(const OMPDeclareReductionDecl *D) { | ||||||
1184 | auto I = UDRMap.find(D); | ||||||
1185 | if (I != UDRMap.end()) | ||||||
1186 | return I->second; | ||||||
1187 | emitUserDefinedReduction(/*CGF=*/nullptr, D); | ||||||
1188 | return UDRMap.lookup(D); | ||||||
1189 | } | ||||||
1190 | |||||||
1191 | namespace { | ||||||
1192 | // Temporary RAII solution to perform a push/pop stack event on the OpenMP IR | ||||||
1193 | // Builder if one is present. | ||||||
1194 | struct PushAndPopStackRAII { | ||||||
1195 | PushAndPopStackRAII(llvm::OpenMPIRBuilder *OMPBuilder, CodeGenFunction &CGF, | ||||||
1196 | bool HasCancel) | ||||||
1197 | : OMPBuilder(OMPBuilder) { | ||||||
1198 | if (!OMPBuilder) | ||||||
1199 | return; | ||||||
1200 | |||||||
1201 | // The following callback is the crucial part of clangs cleanup process. | ||||||
1202 | // | ||||||
1203 | // NOTE: | ||||||
1204 | // Once the OpenMPIRBuilder is used to create parallel regions (and | ||||||
1205 | // similar), the cancellation destination (Dest below) is determined via | ||||||
1206 | // IP. That means if we have variables to finalize we split the block at IP, | ||||||
1207 | // use the new block (=BB) as destination to build a JumpDest (via | ||||||
1208 | // getJumpDestInCurrentScope(BB)) which then is fed to | ||||||
1209 | // EmitBranchThroughCleanup. Furthermore, there will not be the need | ||||||
1210 | // to push & pop an FinalizationInfo object. | ||||||
1211 | // The FiniCB will still be needed but at the point where the | ||||||
1212 | // OpenMPIRBuilder is asked to construct a parallel (or similar) construct. | ||||||
1213 | auto FiniCB = [&CGF](llvm::OpenMPIRBuilder::InsertPointTy IP) { | ||||||
1214 | assert(IP.getBlock()->end() == IP.getPoint() &&((IP.getBlock()->end() == IP.getPoint() && "Clang CG should cause non-terminated block!" ) ? static_cast<void> (0) : __assert_fail ("IP.getBlock()->end() == IP.getPoint() && \"Clang CG should cause non-terminated block!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1215, __PRETTY_FUNCTION__)) | ||||||
1215 | "Clang CG should cause non-terminated block!")((IP.getBlock()->end() == IP.getPoint() && "Clang CG should cause non-terminated block!" ) ? static_cast<void> (0) : __assert_fail ("IP.getBlock()->end() == IP.getPoint() && \"Clang CG should cause non-terminated block!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1215, __PRETTY_FUNCTION__)); | ||||||
1216 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); | ||||||
1217 | CGF.Builder.restoreIP(IP); | ||||||
1218 | CodeGenFunction::JumpDest Dest = | ||||||
1219 | CGF.getOMPCancelDestination(OMPD_parallel); | ||||||
1220 | CGF.EmitBranchThroughCleanup(Dest); | ||||||
1221 | }; | ||||||
1222 | |||||||
1223 | // TODO: Remove this once we emit parallel regions through the | ||||||
1224 | // OpenMPIRBuilder as it can do this setup internally. | ||||||
1225 | llvm::OpenMPIRBuilder::FinalizationInfo FI( | ||||||
1226 | {FiniCB, OMPD_parallel, HasCancel}); | ||||||
1227 | OMPBuilder->pushFinalizationCB(std::move(FI)); | ||||||
1228 | } | ||||||
1229 | ~PushAndPopStackRAII() { | ||||||
1230 | if (OMPBuilder) | ||||||
1231 | OMPBuilder->popFinalizationCB(); | ||||||
1232 | } | ||||||
1233 | llvm::OpenMPIRBuilder *OMPBuilder; | ||||||
1234 | }; | ||||||
1235 | } // namespace | ||||||
1236 | |||||||
1237 | static llvm::Function *emitParallelOrTeamsOutlinedFunction( | ||||||
1238 | CodeGenModule &CGM, const OMPExecutableDirective &D, const CapturedStmt *CS, | ||||||
1239 | const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, | ||||||
1240 | const StringRef OutlinedHelperName, const RegionCodeGenTy &CodeGen) { | ||||||
1241 | assert(ThreadIDVar->getType()->isPointerType() &&((ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 *" ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 *\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1242, __PRETTY_FUNCTION__)) | ||||||
1242 | "thread id variable must be of type kmp_int32 *")((ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 *" ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 *\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1242, __PRETTY_FUNCTION__)); | ||||||
1243 | CodeGenFunction CGF(CGM, true); | ||||||
1244 | bool HasCancel = false; | ||||||
1245 | if (const auto *OPD = dyn_cast<OMPParallelDirective>(&D)) | ||||||
1246 | HasCancel = OPD->hasCancel(); | ||||||
1247 | else if (const auto *OPD = dyn_cast<OMPTargetParallelDirective>(&D)) | ||||||
1248 | HasCancel = OPD->hasCancel(); | ||||||
1249 | else if (const auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&D)) | ||||||
1250 | HasCancel = OPSD->hasCancel(); | ||||||
1251 | else if (const auto *OPFD = dyn_cast<OMPParallelForDirective>(&D)) | ||||||
1252 | HasCancel = OPFD->hasCancel(); | ||||||
1253 | else if (const auto *OPFD = dyn_cast<OMPTargetParallelForDirective>(&D)) | ||||||
1254 | HasCancel = OPFD->hasCancel(); | ||||||
1255 | else if (const auto *OPFD = dyn_cast<OMPDistributeParallelForDirective>(&D)) | ||||||
1256 | HasCancel = OPFD->hasCancel(); | ||||||
1257 | else if (const auto *OPFD = | ||||||
1258 | dyn_cast<OMPTeamsDistributeParallelForDirective>(&D)) | ||||||
1259 | HasCancel = OPFD->hasCancel(); | ||||||
1260 | else if (const auto *OPFD = | ||||||
1261 | dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&D)) | ||||||
1262 | HasCancel = OPFD->hasCancel(); | ||||||
1263 | |||||||
1264 | // TODO: Temporarily inform the OpenMPIRBuilder, if any, about the new | ||||||
1265 | // parallel region to make cancellation barriers work properly. | ||||||
1266 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); | ||||||
1267 | PushAndPopStackRAII PSR(&OMPBuilder, CGF, HasCancel); | ||||||
1268 | CGOpenMPOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, InnermostKind, | ||||||
1269 | HasCancel, OutlinedHelperName); | ||||||
1270 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
1271 | return CGF.GenerateOpenMPCapturedStmtFunction(*CS, D.getBeginLoc()); | ||||||
1272 | } | ||||||
1273 | |||||||
1274 | llvm::Function *CGOpenMPRuntime::emitParallelOutlinedFunction( | ||||||
1275 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | ||||||
1276 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | ||||||
1277 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_parallel); | ||||||
1278 | return emitParallelOrTeamsOutlinedFunction( | ||||||
1279 | CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen); | ||||||
1280 | } | ||||||
1281 | |||||||
1282 | llvm::Function *CGOpenMPRuntime::emitTeamsOutlinedFunction( | ||||||
1283 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | ||||||
1284 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | ||||||
1285 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_teams); | ||||||
1286 | return emitParallelOrTeamsOutlinedFunction( | ||||||
1287 | CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen); | ||||||
1288 | } | ||||||
1289 | |||||||
1290 | llvm::Function *CGOpenMPRuntime::emitTaskOutlinedFunction( | ||||||
1291 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | ||||||
1292 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, | ||||||
1293 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, | ||||||
1294 | bool Tied, unsigned &NumberOfParts) { | ||||||
1295 | auto &&UntiedCodeGen = [this, &D, TaskTVar](CodeGenFunction &CGF, | ||||||
1296 | PrePostActionTy &) { | ||||||
1297 | llvm::Value *ThreadID = getThreadID(CGF, D.getBeginLoc()); | ||||||
1298 | llvm::Value *UpLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | ||||||
1299 | llvm::Value *TaskArgs[] = { | ||||||
1300 | UpLoc, ThreadID, | ||||||
1301 | CGF.EmitLoadOfPointerLValue(CGF.GetAddrOfLocalVar(TaskTVar), | ||||||
1302 | TaskTVar->getType()->castAs<PointerType>()) | ||||||
1303 | .getPointer(CGF)}; | ||||||
1304 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
1305 | CGM.getModule(), OMPRTL___kmpc_omp_task), | ||||||
1306 | TaskArgs); | ||||||
1307 | }; | ||||||
1308 | CGOpenMPTaskOutlinedRegionInfo::UntiedTaskActionTy Action(Tied, PartIDVar, | ||||||
1309 | UntiedCodeGen); | ||||||
1310 | CodeGen.setAction(Action); | ||||||
1311 | assert(!ThreadIDVar->getType()->isPointerType() &&((!ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 for tasks" ) ? static_cast<void> (0) : __assert_fail ("!ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 for tasks\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1312, __PRETTY_FUNCTION__)) | ||||||
1312 | "thread id variable must be of type kmp_int32 for tasks")((!ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 for tasks" ) ? static_cast<void> (0) : __assert_fail ("!ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 for tasks\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1312, __PRETTY_FUNCTION__)); | ||||||
1313 | const OpenMPDirectiveKind Region = | ||||||
1314 | isOpenMPTaskLoopDirective(D.getDirectiveKind()) ? OMPD_taskloop | ||||||
1315 | : OMPD_task; | ||||||
1316 | const CapturedStmt *CS = D.getCapturedStmt(Region); | ||||||
1317 | bool HasCancel = false; | ||||||
1318 | if (const auto *TD = dyn_cast<OMPTaskDirective>(&D)) | ||||||
1319 | HasCancel = TD->hasCancel(); | ||||||
1320 | else if (const auto *TD = dyn_cast<OMPTaskLoopDirective>(&D)) | ||||||
1321 | HasCancel = TD->hasCancel(); | ||||||
1322 | else if (const auto *TD = dyn_cast<OMPMasterTaskLoopDirective>(&D)) | ||||||
1323 | HasCancel = TD->hasCancel(); | ||||||
1324 | else if (const auto *TD = dyn_cast<OMPParallelMasterTaskLoopDirective>(&D)) | ||||||
1325 | HasCancel = TD->hasCancel(); | ||||||
1326 | |||||||
1327 | CodeGenFunction CGF(CGM, true); | ||||||
1328 | CGOpenMPTaskOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, | ||||||
1329 | InnermostKind, HasCancel, Action); | ||||||
1330 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
1331 | llvm::Function *Res = CGF.GenerateCapturedStmtFunction(*CS); | ||||||
1332 | if (!Tied) | ||||||
1333 | NumberOfParts = Action.getNumberOfParts(); | ||||||
1334 | return Res; | ||||||
1335 | } | ||||||
1336 | |||||||
1337 | static void buildStructValue(ConstantStructBuilder &Fields, CodeGenModule &CGM, | ||||||
1338 | const RecordDecl *RD, const CGRecordLayout &RL, | ||||||
1339 | ArrayRef<llvm::Constant *> Data) { | ||||||
1340 | llvm::StructType *StructTy = RL.getLLVMType(); | ||||||
1341 | unsigned PrevIdx = 0; | ||||||
1342 | ConstantInitBuilder CIBuilder(CGM); | ||||||
1343 | auto DI = Data.begin(); | ||||||
1344 | for (const FieldDecl *FD : RD->fields()) { | ||||||
1345 | unsigned Idx = RL.getLLVMFieldNo(FD); | ||||||
1346 | // Fill the alignment. | ||||||
1347 | for (unsigned I = PrevIdx; I < Idx; ++I) | ||||||
1348 | Fields.add(llvm::Constant::getNullValue(StructTy->getElementType(I))); | ||||||
1349 | PrevIdx = Idx + 1; | ||||||
1350 | Fields.add(*DI); | ||||||
1351 | ++DI; | ||||||
1352 | } | ||||||
1353 | } | ||||||
1354 | |||||||
1355 | template <class... As> | ||||||
1356 | static llvm::GlobalVariable * | ||||||
1357 | createGlobalStruct(CodeGenModule &CGM, QualType Ty, bool IsConstant, | ||||||
1358 | ArrayRef<llvm::Constant *> Data, const Twine &Name, | ||||||
1359 | As &&... Args) { | ||||||
1360 | const auto *RD = cast<RecordDecl>(Ty->getAsTagDecl()); | ||||||
1361 | const CGRecordLayout &RL = CGM.getTypes().getCGRecordLayout(RD); | ||||||
1362 | ConstantInitBuilder CIBuilder(CGM); | ||||||
1363 | ConstantStructBuilder Fields = CIBuilder.beginStruct(RL.getLLVMType()); | ||||||
1364 | buildStructValue(Fields, CGM, RD, RL, Data); | ||||||
1365 | return Fields.finishAndCreateGlobal( | ||||||
1366 | Name, CGM.getContext().getAlignOfGlobalVarInChars(Ty), IsConstant, | ||||||
1367 | std::forward<As>(Args)...); | ||||||
1368 | } | ||||||
1369 | |||||||
1370 | template <typename T> | ||||||
1371 | static void | ||||||
1372 | createConstantGlobalStructAndAddToParent(CodeGenModule &CGM, QualType Ty, | ||||||
1373 | ArrayRef<llvm::Constant *> Data, | ||||||
1374 | T &Parent) { | ||||||
1375 | const auto *RD = cast<RecordDecl>(Ty->getAsTagDecl()); | ||||||
1376 | const CGRecordLayout &RL = CGM.getTypes().getCGRecordLayout(RD); | ||||||
1377 | ConstantStructBuilder Fields = Parent.beginStruct(RL.getLLVMType()); | ||||||
1378 | buildStructValue(Fields, CGM, RD, RL, Data); | ||||||
1379 | Fields.finishAndAddTo(Parent); | ||||||
1380 | } | ||||||
1381 | |||||||
1382 | void CGOpenMPRuntime::setLocThreadIdInsertPt(CodeGenFunction &CGF, | ||||||
1383 | bool AtCurrentPoint) { | ||||||
1384 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | ||||||
1385 | assert(!Elem.second.ServiceInsertPt && "Insert point is set already.")((!Elem.second.ServiceInsertPt && "Insert point is set already." ) ? static_cast<void> (0) : __assert_fail ("!Elem.second.ServiceInsertPt && \"Insert point is set already.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1385, __PRETTY_FUNCTION__)); | ||||||
1386 | |||||||
1387 | llvm::Value *Undef = llvm::UndefValue::get(CGF.Int32Ty); | ||||||
1388 | if (AtCurrentPoint) { | ||||||
1389 | Elem.second.ServiceInsertPt = new llvm::BitCastInst( | ||||||
1390 | Undef, CGF.Int32Ty, "svcpt", CGF.Builder.GetInsertBlock()); | ||||||
1391 | } else { | ||||||
1392 | Elem.second.ServiceInsertPt = | ||||||
1393 | new llvm::BitCastInst(Undef, CGF.Int32Ty, "svcpt"); | ||||||
1394 | Elem.second.ServiceInsertPt->insertAfter(CGF.AllocaInsertPt); | ||||||
1395 | } | ||||||
1396 | } | ||||||
1397 | |||||||
1398 | void CGOpenMPRuntime::clearLocThreadIdInsertPt(CodeGenFunction &CGF) { | ||||||
1399 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | ||||||
1400 | if (Elem.second.ServiceInsertPt) { | ||||||
1401 | llvm::Instruction *Ptr = Elem.second.ServiceInsertPt; | ||||||
1402 | Elem.second.ServiceInsertPt = nullptr; | ||||||
1403 | Ptr->eraseFromParent(); | ||||||
1404 | } | ||||||
1405 | } | ||||||
1406 | |||||||
1407 | static StringRef getIdentStringFromSourceLocation(CodeGenFunction &CGF, | ||||||
1408 | SourceLocation Loc, | ||||||
1409 | SmallString<128> &Buffer) { | ||||||
1410 | llvm::raw_svector_ostream OS(Buffer); | ||||||
1411 | // Build debug location | ||||||
1412 | PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc); | ||||||
1413 | OS << ";" << PLoc.getFilename() << ";"; | ||||||
1414 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) | ||||||
1415 | OS << FD->getQualifiedNameAsString(); | ||||||
1416 | OS << ";" << PLoc.getLine() << ";" << PLoc.getColumn() << ";;"; | ||||||
1417 | return OS.str(); | ||||||
1418 | } | ||||||
1419 | |||||||
1420 | llvm::Value *CGOpenMPRuntime::emitUpdateLocation(CodeGenFunction &CGF, | ||||||
1421 | SourceLocation Loc, | ||||||
1422 | unsigned Flags) { | ||||||
1423 | llvm::Constant *SrcLocStr; | ||||||
1424 | if (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo || | ||||||
1425 | Loc.isInvalid()) { | ||||||
1426 | SrcLocStr = OMPBuilder.getOrCreateDefaultSrcLocStr(); | ||||||
1427 | } else { | ||||||
1428 | std::string FunctionName = ""; | ||||||
1429 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) | ||||||
1430 | FunctionName = FD->getQualifiedNameAsString(); | ||||||
1431 | PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc); | ||||||
1432 | const char *FileName = PLoc.getFilename(); | ||||||
1433 | unsigned Line = PLoc.getLine(); | ||||||
1434 | unsigned Column = PLoc.getColumn(); | ||||||
1435 | SrcLocStr = OMPBuilder.getOrCreateSrcLocStr(FunctionName.c_str(), FileName, | ||||||
1436 | Line, Column); | ||||||
1437 | } | ||||||
1438 | unsigned Reserved2Flags = getDefaultLocationReserved2Flags(); | ||||||
1439 | return OMPBuilder.getOrCreateIdent(SrcLocStr, llvm::omp::IdentFlag(Flags), | ||||||
1440 | Reserved2Flags); | ||||||
1441 | } | ||||||
1442 | |||||||
1443 | llvm::Value *CGOpenMPRuntime::getThreadID(CodeGenFunction &CGF, | ||||||
1444 | SourceLocation Loc) { | ||||||
1445 | assert(CGF.CurFn && "No function in current CodeGenFunction.")((CGF.CurFn && "No function in current CodeGenFunction." ) ? static_cast<void> (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1445, __PRETTY_FUNCTION__)); | ||||||
1446 | // If the OpenMPIRBuilder is used we need to use it for all thread id calls as | ||||||
1447 | // the clang invariants used below might be broken. | ||||||
1448 | if (CGM.getLangOpts().OpenMPIRBuilder) { | ||||||
1449 | SmallString<128> Buffer; | ||||||
1450 | OMPBuilder.updateToLocation(CGF.Builder.saveIP()); | ||||||
1451 | auto *SrcLocStr = OMPBuilder.getOrCreateSrcLocStr( | ||||||
1452 | getIdentStringFromSourceLocation(CGF, Loc, Buffer)); | ||||||
1453 | return OMPBuilder.getOrCreateThreadID( | ||||||
1454 | OMPBuilder.getOrCreateIdent(SrcLocStr)); | ||||||
1455 | } | ||||||
1456 | |||||||
1457 | llvm::Value *ThreadID = nullptr; | ||||||
1458 | // Check whether we've already cached a load of the thread id in this | ||||||
1459 | // function. | ||||||
1460 | auto I = OpenMPLocThreadIDMap.find(CGF.CurFn); | ||||||
1461 | if (I != OpenMPLocThreadIDMap.end()) { | ||||||
1462 | ThreadID = I->second.ThreadID; | ||||||
1463 | if (ThreadID != nullptr) | ||||||
1464 | return ThreadID; | ||||||
1465 | } | ||||||
1466 | // If exceptions are enabled, do not use parameter to avoid possible crash. | ||||||
1467 | if (auto *OMPRegionInfo = | ||||||
1468 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | ||||||
1469 | if (OMPRegionInfo->getThreadIDVariable()) { | ||||||
1470 | // Check if this an outlined function with thread id passed as argument. | ||||||
1471 | LValue LVal = OMPRegionInfo->getThreadIDVariableLValue(CGF); | ||||||
1472 | llvm::BasicBlock *TopBlock = CGF.AllocaInsertPt->getParent(); | ||||||
1473 | if (!CGF.EHStack.requiresLandingPad() || !CGF.getLangOpts().Exceptions || | ||||||
1474 | !CGF.getLangOpts().CXXExceptions || | ||||||
1475 | CGF.Builder.GetInsertBlock() == TopBlock || | ||||||
1476 | !isa<llvm::Instruction>(LVal.getPointer(CGF)) || | ||||||
1477 | cast<llvm::Instruction>(LVal.getPointer(CGF))->getParent() == | ||||||
1478 | TopBlock || | ||||||
1479 | cast<llvm::Instruction>(LVal.getPointer(CGF))->getParent() == | ||||||
1480 | CGF.Builder.GetInsertBlock()) { | ||||||
1481 | ThreadID = CGF.EmitLoadOfScalar(LVal, Loc); | ||||||
1482 | // If value loaded in entry block, cache it and use it everywhere in | ||||||
1483 | // function. | ||||||
1484 | if (CGF.Builder.GetInsertBlock() == TopBlock) { | ||||||
1485 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | ||||||
1486 | Elem.second.ThreadID = ThreadID; | ||||||
1487 | } | ||||||
1488 | return ThreadID; | ||||||
1489 | } | ||||||
1490 | } | ||||||
1491 | } | ||||||
1492 | |||||||
1493 | // This is not an outlined function region - need to call __kmpc_int32 | ||||||
1494 | // kmpc_global_thread_num(ident_t *loc). | ||||||
1495 | // Generate thread id value and cache this value for use across the | ||||||
1496 | // function. | ||||||
1497 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | ||||||
1498 | if (!Elem.second.ServiceInsertPt) | ||||||
1499 | setLocThreadIdInsertPt(CGF); | ||||||
1500 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); | ||||||
1501 | CGF.Builder.SetInsertPoint(Elem.second.ServiceInsertPt); | ||||||
1502 | llvm::CallInst *Call = CGF.Builder.CreateCall( | ||||||
1503 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
1504 | OMPRTL___kmpc_global_thread_num), | ||||||
1505 | emitUpdateLocation(CGF, Loc)); | ||||||
1506 | Call->setCallingConv(CGF.getRuntimeCC()); | ||||||
1507 | Elem.second.ThreadID = Call; | ||||||
1508 | return Call; | ||||||
1509 | } | ||||||
1510 | |||||||
1511 | void CGOpenMPRuntime::functionFinished(CodeGenFunction &CGF) { | ||||||
1512 | assert(CGF.CurFn && "No function in current CodeGenFunction.")((CGF.CurFn && "No function in current CodeGenFunction." ) ? static_cast<void> (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1512, __PRETTY_FUNCTION__)); | ||||||
1513 | if (OpenMPLocThreadIDMap.count(CGF.CurFn)) { | ||||||
1514 | clearLocThreadIdInsertPt(CGF); | ||||||
1515 | OpenMPLocThreadIDMap.erase(CGF.CurFn); | ||||||
1516 | } | ||||||
1517 | if (FunctionUDRMap.count(CGF.CurFn) > 0) { | ||||||
1518 | for(const auto *D : FunctionUDRMap[CGF.CurFn]) | ||||||
1519 | UDRMap.erase(D); | ||||||
1520 | FunctionUDRMap.erase(CGF.CurFn); | ||||||
1521 | } | ||||||
1522 | auto I = FunctionUDMMap.find(CGF.CurFn); | ||||||
1523 | if (I != FunctionUDMMap.end()) { | ||||||
1524 | for(const auto *D : I->second) | ||||||
1525 | UDMMap.erase(D); | ||||||
1526 | FunctionUDMMap.erase(I); | ||||||
1527 | } | ||||||
1528 | LastprivateConditionalToTypes.erase(CGF.CurFn); | ||||||
1529 | } | ||||||
1530 | |||||||
1531 | llvm::Type *CGOpenMPRuntime::getIdentTyPointerTy() { | ||||||
1532 | return OMPBuilder.IdentPtr; | ||||||
1533 | } | ||||||
1534 | |||||||
1535 | llvm::Type *CGOpenMPRuntime::getKmpc_MicroPointerTy() { | ||||||
1536 | if (!Kmpc_MicroTy) { | ||||||
1537 | // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...) | ||||||
1538 | llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty), | ||||||
1539 | llvm::PointerType::getUnqual(CGM.Int32Ty)}; | ||||||
1540 | Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true); | ||||||
1541 | } | ||||||
1542 | return llvm::PointerType::getUnqual(Kmpc_MicroTy); | ||||||
1543 | } | ||||||
1544 | |||||||
1545 | llvm::FunctionCallee | ||||||
1546 | CGOpenMPRuntime::createForStaticInitFunction(unsigned IVSize, bool IVSigned) { | ||||||
1547 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1548, __PRETTY_FUNCTION__)) | ||||||
1548 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1548, __PRETTY_FUNCTION__)); | ||||||
1549 | StringRef Name = IVSize == 32 ? (IVSigned ? "__kmpc_for_static_init_4" | ||||||
1550 | : "__kmpc_for_static_init_4u") | ||||||
1551 | : (IVSigned ? "__kmpc_for_static_init_8" | ||||||
1552 | : "__kmpc_for_static_init_8u"); | ||||||
1553 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | ||||||
1554 | auto *PtrTy = llvm::PointerType::getUnqual(ITy); | ||||||
1555 | llvm::Type *TypeParams[] = { | ||||||
1556 | getIdentTyPointerTy(), // loc | ||||||
1557 | CGM.Int32Ty, // tid | ||||||
1558 | CGM.Int32Ty, // schedtype | ||||||
1559 | llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter | ||||||
1560 | PtrTy, // p_lower | ||||||
1561 | PtrTy, // p_upper | ||||||
1562 | PtrTy, // p_stride | ||||||
1563 | ITy, // incr | ||||||
1564 | ITy // chunk | ||||||
1565 | }; | ||||||
1566 | auto *FnTy = | ||||||
1567 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | ||||||
1568 | return CGM.CreateRuntimeFunction(FnTy, Name); | ||||||
1569 | } | ||||||
1570 | |||||||
1571 | llvm::FunctionCallee | ||||||
1572 | CGOpenMPRuntime::createDispatchInitFunction(unsigned IVSize, bool IVSigned) { | ||||||
1573 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1574, __PRETTY_FUNCTION__)) | ||||||
1574 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1574, __PRETTY_FUNCTION__)); | ||||||
1575 | StringRef Name = | ||||||
1576 | IVSize == 32 | ||||||
1577 | ? (IVSigned ? "__kmpc_dispatch_init_4" : "__kmpc_dispatch_init_4u") | ||||||
1578 | : (IVSigned ? "__kmpc_dispatch_init_8" : "__kmpc_dispatch_init_8u"); | ||||||
1579 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | ||||||
1580 | llvm::Type *TypeParams[] = { getIdentTyPointerTy(), // loc | ||||||
1581 | CGM.Int32Ty, // tid | ||||||
1582 | CGM.Int32Ty, // schedtype | ||||||
1583 | ITy, // lower | ||||||
1584 | ITy, // upper | ||||||
1585 | ITy, // stride | ||||||
1586 | ITy // chunk | ||||||
1587 | }; | ||||||
1588 | auto *FnTy = | ||||||
1589 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | ||||||
1590 | return CGM.CreateRuntimeFunction(FnTy, Name); | ||||||
1591 | } | ||||||
1592 | |||||||
1593 | llvm::FunctionCallee | ||||||
1594 | CGOpenMPRuntime::createDispatchFiniFunction(unsigned IVSize, bool IVSigned) { | ||||||
1595 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1596, __PRETTY_FUNCTION__)) | ||||||
1596 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1596, __PRETTY_FUNCTION__)); | ||||||
1597 | StringRef Name = | ||||||
1598 | IVSize == 32 | ||||||
1599 | ? (IVSigned ? "__kmpc_dispatch_fini_4" : "__kmpc_dispatch_fini_4u") | ||||||
1600 | : (IVSigned ? "__kmpc_dispatch_fini_8" : "__kmpc_dispatch_fini_8u"); | ||||||
1601 | llvm::Type *TypeParams[] = { | ||||||
1602 | getIdentTyPointerTy(), // loc | ||||||
1603 | CGM.Int32Ty, // tid | ||||||
1604 | }; | ||||||
1605 | auto *FnTy = | ||||||
1606 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | ||||||
1607 | return CGM.CreateRuntimeFunction(FnTy, Name); | ||||||
1608 | } | ||||||
1609 | |||||||
1610 | llvm::FunctionCallee | ||||||
1611 | CGOpenMPRuntime::createDispatchNextFunction(unsigned IVSize, bool IVSigned) { | ||||||
1612 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1613, __PRETTY_FUNCTION__)) | ||||||
1613 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1613, __PRETTY_FUNCTION__)); | ||||||
1614 | StringRef Name = | ||||||
1615 | IVSize == 32 | ||||||
1616 | ? (IVSigned ? "__kmpc_dispatch_next_4" : "__kmpc_dispatch_next_4u") | ||||||
1617 | : (IVSigned ? "__kmpc_dispatch_next_8" : "__kmpc_dispatch_next_8u"); | ||||||
1618 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | ||||||
1619 | auto *PtrTy = llvm::PointerType::getUnqual(ITy); | ||||||
1620 | llvm::Type *TypeParams[] = { | ||||||
1621 | getIdentTyPointerTy(), // loc | ||||||
1622 | CGM.Int32Ty, // tid | ||||||
1623 | llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter | ||||||
1624 | PtrTy, // p_lower | ||||||
1625 | PtrTy, // p_upper | ||||||
1626 | PtrTy // p_stride | ||||||
1627 | }; | ||||||
1628 | auto *FnTy = | ||||||
1629 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | ||||||
1630 | return CGM.CreateRuntimeFunction(FnTy, Name); | ||||||
1631 | } | ||||||
1632 | |||||||
1633 | /// Obtain information that uniquely identifies a target entry. This | ||||||
1634 | /// consists of the file and device IDs as well as line number associated with | ||||||
1635 | /// the relevant entry source location. | ||||||
1636 | static void getTargetEntryUniqueInfo(ASTContext &C, SourceLocation Loc, | ||||||
1637 | unsigned &DeviceID, unsigned &FileID, | ||||||
1638 | unsigned &LineNum) { | ||||||
1639 | SourceManager &SM = C.getSourceManager(); | ||||||
1640 | |||||||
1641 | // The loc should be always valid and have a file ID (the user cannot use | ||||||
1642 | // #pragma directives in macros) | ||||||
1643 | |||||||
1644 | assert(Loc.isValid() && "Source location is expected to be always valid.")((Loc.isValid() && "Source location is expected to be always valid." ) ? static_cast<void> (0) : __assert_fail ("Loc.isValid() && \"Source location is expected to be always valid.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1644, __PRETTY_FUNCTION__)); | ||||||
1645 | |||||||
1646 | PresumedLoc PLoc = SM.getPresumedLoc(Loc); | ||||||
1647 | assert(PLoc.isValid() && "Source location is expected to be always valid.")((PLoc.isValid() && "Source location is expected to be always valid." ) ? static_cast<void> (0) : __assert_fail ("PLoc.isValid() && \"Source location is expected to be always valid.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1647, __PRETTY_FUNCTION__)); | ||||||
1648 | |||||||
1649 | llvm::sys::fs::UniqueID ID; | ||||||
1650 | if (auto EC = llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID)) | ||||||
1651 | SM.getDiagnostics().Report(diag::err_cannot_open_file) | ||||||
1652 | << PLoc.getFilename() << EC.message(); | ||||||
1653 | |||||||
1654 | DeviceID = ID.getDevice(); | ||||||
1655 | FileID = ID.getFile(); | ||||||
1656 | LineNum = PLoc.getLine(); | ||||||
1657 | } | ||||||
1658 | |||||||
1659 | Address CGOpenMPRuntime::getAddrOfDeclareTargetVar(const VarDecl *VD) { | ||||||
1660 | if (CGM.getLangOpts().OpenMPSimd) | ||||||
1661 | return Address::invalid(); | ||||||
1662 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | ||||||
1663 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | ||||||
1664 | if (Res && (*Res == OMPDeclareTargetDeclAttr::MT_Link || | ||||||
1665 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | ||||||
1666 | HasRequiresUnifiedSharedMemory))) { | ||||||
1667 | SmallString<64> PtrName; | ||||||
1668 | { | ||||||
1669 | llvm::raw_svector_ostream OS(PtrName); | ||||||
1670 | OS << CGM.getMangledName(GlobalDecl(VD)); | ||||||
1671 | if (!VD->isExternallyVisible()) { | ||||||
1672 | unsigned DeviceID, FileID, Line; | ||||||
1673 | getTargetEntryUniqueInfo(CGM.getContext(), | ||||||
1674 | VD->getCanonicalDecl()->getBeginLoc(), | ||||||
1675 | DeviceID, FileID, Line); | ||||||
1676 | OS << llvm::format("_%x", FileID); | ||||||
1677 | } | ||||||
1678 | OS << "_decl_tgt_ref_ptr"; | ||||||
1679 | } | ||||||
1680 | llvm::Value *Ptr = CGM.getModule().getNamedValue(PtrName); | ||||||
1681 | if (!Ptr) { | ||||||
1682 | QualType PtrTy = CGM.getContext().getPointerType(VD->getType()); | ||||||
1683 | Ptr = getOrCreateInternalVariable(CGM.getTypes().ConvertTypeForMem(PtrTy), | ||||||
1684 | PtrName); | ||||||
1685 | |||||||
1686 | auto *GV = cast<llvm::GlobalVariable>(Ptr); | ||||||
1687 | GV->setLinkage(llvm::GlobalValue::WeakAnyLinkage); | ||||||
1688 | |||||||
1689 | if (!CGM.getLangOpts().OpenMPIsDevice) | ||||||
1690 | GV->setInitializer(CGM.GetAddrOfGlobal(VD)); | ||||||
1691 | registerTargetGlobalVariable(VD, cast<llvm::Constant>(Ptr)); | ||||||
1692 | } | ||||||
1693 | return Address(Ptr, CGM.getContext().getDeclAlign(VD)); | ||||||
1694 | } | ||||||
1695 | return Address::invalid(); | ||||||
1696 | } | ||||||
1697 | |||||||
1698 | llvm::Constant * | ||||||
1699 | CGOpenMPRuntime::getOrCreateThreadPrivateCache(const VarDecl *VD) { | ||||||
1700 | assert(!CGM.getLangOpts().OpenMPUseTLS ||((!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo ().isTLSSupported()) ? static_cast<void> (0) : __assert_fail ("!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1701, __PRETTY_FUNCTION__)) | ||||||
1701 | !CGM.getContext().getTargetInfo().isTLSSupported())((!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo ().isTLSSupported()) ? static_cast<void> (0) : __assert_fail ("!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1701, __PRETTY_FUNCTION__)); | ||||||
1702 | // Lookup the entry, lazily creating it if necessary. | ||||||
1703 | std::string Suffix = getName({"cache", ""}); | ||||||
1704 | return getOrCreateInternalVariable( | ||||||
1705 | CGM.Int8PtrPtrTy, Twine(CGM.getMangledName(VD)).concat(Suffix)); | ||||||
1706 | } | ||||||
1707 | |||||||
1708 | Address CGOpenMPRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF, | ||||||
1709 | const VarDecl *VD, | ||||||
1710 | Address VDAddr, | ||||||
1711 | SourceLocation Loc) { | ||||||
1712 | if (CGM.getLangOpts().OpenMPUseTLS && | ||||||
1713 | CGM.getContext().getTargetInfo().isTLSSupported()) | ||||||
1714 | return VDAddr; | ||||||
1715 | |||||||
1716 | llvm::Type *VarTy = VDAddr.getElementType(); | ||||||
1717 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | ||||||
1718 | CGF.Builder.CreatePointerCast(VDAddr.getPointer(), | ||||||
1719 | CGM.Int8PtrTy), | ||||||
1720 | CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)), | ||||||
1721 | getOrCreateThreadPrivateCache(VD)}; | ||||||
1722 | return Address(CGF.EmitRuntimeCall( | ||||||
1723 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
1724 | CGM.getModule(), OMPRTL___kmpc_threadprivate_cached), | ||||||
1725 | Args), | ||||||
1726 | VDAddr.getAlignment()); | ||||||
1727 | } | ||||||
1728 | |||||||
1729 | void CGOpenMPRuntime::emitThreadPrivateVarInit( | ||||||
1730 | CodeGenFunction &CGF, Address VDAddr, llvm::Value *Ctor, | ||||||
1731 | llvm::Value *CopyCtor, llvm::Value *Dtor, SourceLocation Loc) { | ||||||
1732 | // Call kmp_int32 __kmpc_global_thread_num(&loc) to init OpenMP runtime | ||||||
1733 | // library. | ||||||
1734 | llvm::Value *OMPLoc = emitUpdateLocation(CGF, Loc); | ||||||
1735 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
1736 | CGM.getModule(), OMPRTL___kmpc_global_thread_num), | ||||||
1737 | OMPLoc); | ||||||
1738 | // Call __kmpc_threadprivate_register(&loc, &var, ctor, cctor/*NULL*/, dtor) | ||||||
1739 | // to register constructor/destructor for variable. | ||||||
1740 | llvm::Value *Args[] = { | ||||||
1741 | OMPLoc, CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.VoidPtrTy), | ||||||
1742 | Ctor, CopyCtor, Dtor}; | ||||||
1743 | CGF.EmitRuntimeCall( | ||||||
1744 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
1745 | CGM.getModule(), OMPRTL___kmpc_threadprivate_register), | ||||||
1746 | Args); | ||||||
1747 | } | ||||||
1748 | |||||||
1749 | llvm::Function *CGOpenMPRuntime::emitThreadPrivateVarDefinition( | ||||||
1750 | const VarDecl *VD, Address VDAddr, SourceLocation Loc, | ||||||
1751 | bool PerformInit, CodeGenFunction *CGF) { | ||||||
1752 | if (CGM.getLangOpts().OpenMPUseTLS && | ||||||
1753 | CGM.getContext().getTargetInfo().isTLSSupported()) | ||||||
1754 | return nullptr; | ||||||
1755 | |||||||
1756 | VD = VD->getDefinition(CGM.getContext()); | ||||||
1757 | if (VD && ThreadPrivateWithDefinition.insert(CGM.getMangledName(VD)).second) { | ||||||
1758 | QualType ASTTy = VD->getType(); | ||||||
1759 | |||||||
1760 | llvm::Value *Ctor = nullptr, *CopyCtor = nullptr, *Dtor = nullptr; | ||||||
1761 | const Expr *Init = VD->getAnyInitializer(); | ||||||
1762 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { | ||||||
1763 | // Generate function that re-emits the declaration's initializer into the | ||||||
1764 | // threadprivate copy of the variable VD | ||||||
1765 | CodeGenFunction CtorCGF(CGM); | ||||||
1766 | FunctionArgList Args; | ||||||
1767 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc, | ||||||
1768 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy, | ||||||
1769 | ImplicitParamDecl::Other); | ||||||
1770 | Args.push_back(&Dst); | ||||||
1771 | |||||||
1772 | const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration( | ||||||
1773 | CGM.getContext().VoidPtrTy, Args); | ||||||
1774 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | ||||||
1775 | std::string Name = getName({"__kmpc_global_ctor_", ""}); | ||||||
1776 | llvm::Function *Fn = | ||||||
1777 | CGM.CreateGlobalInitOrCleanUpFunction(FTy, Name, FI, Loc); | ||||||
1778 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidPtrTy, Fn, FI, | ||||||
1779 | Args, Loc, Loc); | ||||||
1780 | llvm::Value *ArgVal = CtorCGF.EmitLoadOfScalar( | ||||||
1781 | CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false, | ||||||
1782 | CGM.getContext().VoidPtrTy, Dst.getLocation()); | ||||||
1783 | Address Arg = Address(ArgVal, VDAddr.getAlignment()); | ||||||
1784 | Arg = CtorCGF.Builder.CreateElementBitCast( | ||||||
1785 | Arg, CtorCGF.ConvertTypeForMem(ASTTy)); | ||||||
1786 | CtorCGF.EmitAnyExprToMem(Init, Arg, Init->getType().getQualifiers(), | ||||||
1787 | /*IsInitializer=*/true); | ||||||
1788 | ArgVal = CtorCGF.EmitLoadOfScalar( | ||||||
1789 | CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false, | ||||||
1790 | CGM.getContext().VoidPtrTy, Dst.getLocation()); | ||||||
1791 | CtorCGF.Builder.CreateStore(ArgVal, CtorCGF.ReturnValue); | ||||||
1792 | CtorCGF.FinishFunction(); | ||||||
1793 | Ctor = Fn; | ||||||
1794 | } | ||||||
1795 | if (VD->getType().isDestructedType() != QualType::DK_none) { | ||||||
1796 | // Generate function that emits destructor call for the threadprivate copy | ||||||
1797 | // of the variable VD | ||||||
1798 | CodeGenFunction DtorCGF(CGM); | ||||||
1799 | FunctionArgList Args; | ||||||
1800 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc, | ||||||
1801 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy, | ||||||
1802 | ImplicitParamDecl::Other); | ||||||
1803 | Args.push_back(&Dst); | ||||||
1804 | |||||||
1805 | const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration( | ||||||
1806 | CGM.getContext().VoidTy, Args); | ||||||
1807 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | ||||||
1808 | std::string Name = getName({"__kmpc_global_dtor_", ""}); | ||||||
1809 | llvm::Function *Fn = | ||||||
1810 | CGM.CreateGlobalInitOrCleanUpFunction(FTy, Name, FI, Loc); | ||||||
1811 | auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF); | ||||||
1812 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, Args, | ||||||
1813 | Loc, Loc); | ||||||
1814 | // Create a scope with an artificial location for the body of this function. | ||||||
1815 | auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF); | ||||||
1816 | llvm::Value *ArgVal = DtorCGF.EmitLoadOfScalar( | ||||||
1817 | DtorCGF.GetAddrOfLocalVar(&Dst), | ||||||
1818 | /*Volatile=*/false, CGM.getContext().VoidPtrTy, Dst.getLocation()); | ||||||
1819 | DtorCGF.emitDestroy(Address(ArgVal, VDAddr.getAlignment()), ASTTy, | ||||||
1820 | DtorCGF.getDestroyer(ASTTy.isDestructedType()), | ||||||
1821 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); | ||||||
1822 | DtorCGF.FinishFunction(); | ||||||
1823 | Dtor = Fn; | ||||||
1824 | } | ||||||
1825 | // Do not emit init function if it is not required. | ||||||
1826 | if (!Ctor && !Dtor) | ||||||
1827 | return nullptr; | ||||||
1828 | |||||||
1829 | llvm::Type *CopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | ||||||
1830 | auto *CopyCtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CopyCtorTyArgs, | ||||||
1831 | /*isVarArg=*/false) | ||||||
1832 | ->getPointerTo(); | ||||||
1833 | // Copying constructor for the threadprivate variable. | ||||||
1834 | // Must be NULL - reserved by runtime, but currently it requires that this | ||||||
1835 | // parameter is always NULL. Otherwise it fires assertion. | ||||||
1836 | CopyCtor = llvm::Constant::getNullValue(CopyCtorTy); | ||||||
1837 | if (Ctor == nullptr) { | ||||||
1838 | auto *CtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy, | ||||||
1839 | /*isVarArg=*/false) | ||||||
1840 | ->getPointerTo(); | ||||||
1841 | Ctor = llvm::Constant::getNullValue(CtorTy); | ||||||
1842 | } | ||||||
1843 | if (Dtor == nullptr) { | ||||||
1844 | auto *DtorTy = llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, | ||||||
1845 | /*isVarArg=*/false) | ||||||
1846 | ->getPointerTo(); | ||||||
1847 | Dtor = llvm::Constant::getNullValue(DtorTy); | ||||||
1848 | } | ||||||
1849 | if (!CGF) { | ||||||
1850 | auto *InitFunctionTy = | ||||||
1851 | llvm::FunctionType::get(CGM.VoidTy, /*isVarArg*/ false); | ||||||
1852 | std::string Name = getName({"__omp_threadprivate_init_", ""}); | ||||||
1853 | llvm::Function *InitFunction = CGM.CreateGlobalInitOrCleanUpFunction( | ||||||
1854 | InitFunctionTy, Name, CGM.getTypes().arrangeNullaryFunction()); | ||||||
1855 | CodeGenFunction InitCGF(CGM); | ||||||
1856 | FunctionArgList ArgList; | ||||||
1857 | InitCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, InitFunction, | ||||||
1858 | CGM.getTypes().arrangeNullaryFunction(), ArgList, | ||||||
1859 | Loc, Loc); | ||||||
1860 | emitThreadPrivateVarInit(InitCGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); | ||||||
1861 | InitCGF.FinishFunction(); | ||||||
1862 | return InitFunction; | ||||||
1863 | } | ||||||
1864 | emitThreadPrivateVarInit(*CGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); | ||||||
1865 | } | ||||||
1866 | return nullptr; | ||||||
1867 | } | ||||||
1868 | |||||||
1869 | bool CGOpenMPRuntime::emitDeclareTargetVarDefinition(const VarDecl *VD, | ||||||
1870 | llvm::GlobalVariable *Addr, | ||||||
1871 | bool PerformInit) { | ||||||
1872 | if (CGM.getLangOpts().OMPTargetTriples.empty() && | ||||||
1873 | !CGM.getLangOpts().OpenMPIsDevice) | ||||||
1874 | return false; | ||||||
1875 | Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | ||||||
1876 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | ||||||
1877 | if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link || | ||||||
1878 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | ||||||
1879 | HasRequiresUnifiedSharedMemory)) | ||||||
1880 | return CGM.getLangOpts().OpenMPIsDevice; | ||||||
1881 | VD = VD->getDefinition(CGM.getContext()); | ||||||
1882 | assert(VD && "Unknown VarDecl")((VD && "Unknown VarDecl") ? static_cast<void> ( 0) : __assert_fail ("VD && \"Unknown VarDecl\"", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1882, __PRETTY_FUNCTION__)); | ||||||
1883 | |||||||
1884 | if (!DeclareTargetWithDefinition.insert(CGM.getMangledName(VD)).second) | ||||||
1885 | return CGM.getLangOpts().OpenMPIsDevice; | ||||||
1886 | |||||||
1887 | QualType ASTTy = VD->getType(); | ||||||
1888 | SourceLocation Loc = VD->getCanonicalDecl()->getBeginLoc(); | ||||||
1889 | |||||||
1890 | // Produce the unique prefix to identify the new target regions. We use | ||||||
1891 | // the source location of the variable declaration which we know to not | ||||||
1892 | // conflict with any target region. | ||||||
1893 | unsigned DeviceID; | ||||||
1894 | unsigned FileID; | ||||||
1895 | unsigned Line; | ||||||
1896 | getTargetEntryUniqueInfo(CGM.getContext(), Loc, DeviceID, FileID, Line); | ||||||
1897 | SmallString<128> Buffer, Out; | ||||||
1898 | { | ||||||
1899 | llvm::raw_svector_ostream OS(Buffer); | ||||||
1900 | OS << "__omp_offloading_" << llvm::format("_%x", DeviceID) | ||||||
1901 | << llvm::format("_%x_", FileID) << VD->getName() << "_l" << Line; | ||||||
1902 | } | ||||||
1903 | |||||||
1904 | const Expr *Init = VD->getAnyInitializer(); | ||||||
1905 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { | ||||||
1906 | llvm::Constant *Ctor; | ||||||
1907 | llvm::Constant *ID; | ||||||
1908 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
1909 | // Generate function that re-emits the declaration's initializer into | ||||||
1910 | // the threadprivate copy of the variable VD | ||||||
1911 | CodeGenFunction CtorCGF(CGM); | ||||||
1912 | |||||||
1913 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); | ||||||
1914 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | ||||||
1915 | llvm::Function *Fn = CGM.CreateGlobalInitOrCleanUpFunction( | ||||||
1916 | FTy, Twine(Buffer, "_ctor"), FI, Loc); | ||||||
1917 | auto NL = ApplyDebugLocation::CreateEmpty(CtorCGF); | ||||||
1918 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, | ||||||
1919 | FunctionArgList(), Loc, Loc); | ||||||
1920 | auto AL = ApplyDebugLocation::CreateArtificial(CtorCGF); | ||||||
1921 | CtorCGF.EmitAnyExprToMem(Init, | ||||||
1922 | Address(Addr, CGM.getContext().getDeclAlign(VD)), | ||||||
1923 | Init->getType().getQualifiers(), | ||||||
1924 | /*IsInitializer=*/true); | ||||||
1925 | CtorCGF.FinishFunction(); | ||||||
1926 | Ctor = Fn; | ||||||
1927 | ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy); | ||||||
1928 | CGM.addUsedGlobal(cast<llvm::GlobalValue>(Ctor)); | ||||||
1929 | } else { | ||||||
1930 | Ctor = new llvm::GlobalVariable( | ||||||
1931 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | ||||||
1932 | llvm::GlobalValue::PrivateLinkage, | ||||||
1933 | llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_ctor")); | ||||||
1934 | ID = Ctor; | ||||||
1935 | } | ||||||
1936 | |||||||
1937 | // Register the information for the entry associated with the constructor. | ||||||
1938 | Out.clear(); | ||||||
1939 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | ||||||
1940 | DeviceID, FileID, Twine(Buffer, "_ctor").toStringRef(Out), Line, Ctor, | ||||||
1941 | ID, OffloadEntriesInfoManagerTy::OMPTargetRegionEntryCtor); | ||||||
1942 | } | ||||||
1943 | if (VD->getType().isDestructedType() != QualType::DK_none) { | ||||||
1944 | llvm::Constant *Dtor; | ||||||
1945 | llvm::Constant *ID; | ||||||
1946 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
1947 | // Generate function that emits destructor call for the threadprivate | ||||||
1948 | // copy of the variable VD | ||||||
1949 | CodeGenFunction DtorCGF(CGM); | ||||||
1950 | |||||||
1951 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); | ||||||
1952 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | ||||||
1953 | llvm::Function *Fn = CGM.CreateGlobalInitOrCleanUpFunction( | ||||||
1954 | FTy, Twine(Buffer, "_dtor"), FI, Loc); | ||||||
1955 | auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF); | ||||||
1956 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, | ||||||
1957 | FunctionArgList(), Loc, Loc); | ||||||
1958 | // Create a scope with an artificial location for the body of this | ||||||
1959 | // function. | ||||||
1960 | auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF); | ||||||
1961 | DtorCGF.emitDestroy(Address(Addr, CGM.getContext().getDeclAlign(VD)), | ||||||
1962 | ASTTy, DtorCGF.getDestroyer(ASTTy.isDestructedType()), | ||||||
1963 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); | ||||||
1964 | DtorCGF.FinishFunction(); | ||||||
1965 | Dtor = Fn; | ||||||
1966 | ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy); | ||||||
1967 | CGM.addUsedGlobal(cast<llvm::GlobalValue>(Dtor)); | ||||||
1968 | } else { | ||||||
1969 | Dtor = new llvm::GlobalVariable( | ||||||
1970 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | ||||||
1971 | llvm::GlobalValue::PrivateLinkage, | ||||||
1972 | llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_dtor")); | ||||||
1973 | ID = Dtor; | ||||||
1974 | } | ||||||
1975 | // Register the information for the entry associated with the destructor. | ||||||
1976 | Out.clear(); | ||||||
1977 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | ||||||
1978 | DeviceID, FileID, Twine(Buffer, "_dtor").toStringRef(Out), Line, Dtor, | ||||||
1979 | ID, OffloadEntriesInfoManagerTy::OMPTargetRegionEntryDtor); | ||||||
1980 | } | ||||||
1981 | return CGM.getLangOpts().OpenMPIsDevice; | ||||||
1982 | } | ||||||
1983 | |||||||
1984 | Address CGOpenMPRuntime::getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF, | ||||||
1985 | QualType VarType, | ||||||
1986 | StringRef Name) { | ||||||
1987 | std::string Suffix = getName({"artificial", ""}); | ||||||
1988 | llvm::Type *VarLVType = CGF.ConvertTypeForMem(VarType); | ||||||
1989 | llvm::Value *GAddr = | ||||||
1990 | getOrCreateInternalVariable(VarLVType, Twine(Name).concat(Suffix)); | ||||||
1991 | if (CGM.getLangOpts().OpenMP && CGM.getLangOpts().OpenMPUseTLS && | ||||||
1992 | CGM.getTarget().isTLSSupported()) { | ||||||
1993 | cast<llvm::GlobalVariable>(GAddr)->setThreadLocal(/*Val=*/true); | ||||||
1994 | return Address(GAddr, CGM.getContext().getTypeAlignInChars(VarType)); | ||||||
1995 | } | ||||||
1996 | std::string CacheSuffix = getName({"cache", ""}); | ||||||
1997 | llvm::Value *Args[] = { | ||||||
1998 | emitUpdateLocation(CGF, SourceLocation()), | ||||||
1999 | getThreadID(CGF, SourceLocation()), | ||||||
2000 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(GAddr, CGM.VoidPtrTy), | ||||||
2001 | CGF.Builder.CreateIntCast(CGF.getTypeSize(VarType), CGM.SizeTy, | ||||||
2002 | /*isSigned=*/false), | ||||||
2003 | getOrCreateInternalVariable( | ||||||
2004 | CGM.VoidPtrPtrTy, Twine(Name).concat(Suffix).concat(CacheSuffix))}; | ||||||
2005 | return Address( | ||||||
2006 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
2007 | CGF.EmitRuntimeCall( | ||||||
2008 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2009 | CGM.getModule(), OMPRTL___kmpc_threadprivate_cached), | ||||||
2010 | Args), | ||||||
2011 | VarLVType->getPointerTo(/*AddrSpace=*/0)), | ||||||
2012 | CGM.getContext().getTypeAlignInChars(VarType)); | ||||||
2013 | } | ||||||
2014 | |||||||
2015 | void CGOpenMPRuntime::emitIfClause(CodeGenFunction &CGF, const Expr *Cond, | ||||||
2016 | const RegionCodeGenTy &ThenGen, | ||||||
2017 | const RegionCodeGenTy &ElseGen) { | ||||||
2018 | CodeGenFunction::LexicalScope ConditionScope(CGF, Cond->getSourceRange()); | ||||||
2019 | |||||||
2020 | // If the condition constant folds and can be elided, try to avoid emitting | ||||||
2021 | // the condition and the dead arm of the if/else. | ||||||
2022 | bool CondConstant; | ||||||
2023 | if (CGF.ConstantFoldsToSimpleInteger(Cond, CondConstant)) { | ||||||
2024 | if (CondConstant) | ||||||
2025 | ThenGen(CGF); | ||||||
2026 | else | ||||||
2027 | ElseGen(CGF); | ||||||
2028 | return; | ||||||
2029 | } | ||||||
2030 | |||||||
2031 | // Otherwise, the condition did not fold, or we couldn't elide it. Just | ||||||
2032 | // emit the conditional branch. | ||||||
2033 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("omp_if.then"); | ||||||
2034 | llvm::BasicBlock *ElseBlock = CGF.createBasicBlock("omp_if.else"); | ||||||
2035 | llvm::BasicBlock *ContBlock = CGF.createBasicBlock("omp_if.end"); | ||||||
2036 | CGF.EmitBranchOnBoolExpr(Cond, ThenBlock, ElseBlock, /*TrueCount=*/0); | ||||||
2037 | |||||||
2038 | // Emit the 'then' code. | ||||||
2039 | CGF.EmitBlock(ThenBlock); | ||||||
2040 | ThenGen(CGF); | ||||||
2041 | CGF.EmitBranch(ContBlock); | ||||||
2042 | // Emit the 'else' code if present. | ||||||
2043 | // There is no need to emit line number for unconditional branch. | ||||||
2044 | (void)ApplyDebugLocation::CreateEmpty(CGF); | ||||||
2045 | CGF.EmitBlock(ElseBlock); | ||||||
2046 | ElseGen(CGF); | ||||||
2047 | // There is no need to emit line number for unconditional branch. | ||||||
2048 | (void)ApplyDebugLocation::CreateEmpty(CGF); | ||||||
2049 | CGF.EmitBranch(ContBlock); | ||||||
2050 | // Emit the continuation block for code after the if. | ||||||
2051 | CGF.EmitBlock(ContBlock, /*IsFinished=*/true); | ||||||
2052 | } | ||||||
2053 | |||||||
2054 | void CGOpenMPRuntime::emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
2055 | llvm::Function *OutlinedFn, | ||||||
2056 | ArrayRef<llvm::Value *> CapturedVars, | ||||||
2057 | const Expr *IfCond) { | ||||||
2058 | if (!CGF.HaveInsertPoint()) | ||||||
2059 | return; | ||||||
2060 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | ||||||
2061 | auto &M = CGM.getModule(); | ||||||
2062 | auto &&ThenGen = [&M, OutlinedFn, CapturedVars, RTLoc, | ||||||
2063 | this](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
2064 | // Build call __kmpc_fork_call(loc, n, microtask, var1, .., varn); | ||||||
2065 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | ||||||
2066 | llvm::Value *Args[] = { | ||||||
2067 | RTLoc, | ||||||
2068 | CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars | ||||||
2069 | CGF.Builder.CreateBitCast(OutlinedFn, RT.getKmpc_MicroPointerTy())}; | ||||||
2070 | llvm::SmallVector<llvm::Value *, 16> RealArgs; | ||||||
2071 | RealArgs.append(std::begin(Args), std::end(Args)); | ||||||
2072 | RealArgs.append(CapturedVars.begin(), CapturedVars.end()); | ||||||
2073 | |||||||
2074 | llvm::FunctionCallee RTLFn = | ||||||
2075 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_fork_call); | ||||||
2076 | CGF.EmitRuntimeCall(RTLFn, RealArgs); | ||||||
2077 | }; | ||||||
2078 | auto &&ElseGen = [&M, OutlinedFn, CapturedVars, RTLoc, Loc, | ||||||
2079 | this](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
2080 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | ||||||
2081 | llvm::Value *ThreadID = RT.getThreadID(CGF, Loc); | ||||||
2082 | // Build calls: | ||||||
2083 | // __kmpc_serialized_parallel(&Loc, GTid); | ||||||
2084 | llvm::Value *Args[] = {RTLoc, ThreadID}; | ||||||
2085 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2086 | M, OMPRTL___kmpc_serialized_parallel), | ||||||
2087 | Args); | ||||||
2088 | |||||||
2089 | // OutlinedFn(>id, &zero_bound, CapturedStruct); | ||||||
2090 | Address ThreadIDAddr = RT.emitThreadIDAddress(CGF, Loc); | ||||||
2091 | Address ZeroAddrBound = | ||||||
2092 | CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | ||||||
2093 | /*Name=*/".bound.zero.addr"); | ||||||
2094 | CGF.InitTempAlloca(ZeroAddrBound, CGF.Builder.getInt32(/*C*/ 0)); | ||||||
2095 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | ||||||
2096 | // ThreadId for serialized parallels is 0. | ||||||
2097 | OutlinedFnArgs.push_back(ThreadIDAddr.getPointer()); | ||||||
2098 | OutlinedFnArgs.push_back(ZeroAddrBound.getPointer()); | ||||||
2099 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); | ||||||
2100 | RT.emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); | ||||||
2101 | |||||||
2102 | // __kmpc_end_serialized_parallel(&Loc, GTid); | ||||||
2103 | llvm::Value *EndArgs[] = {RT.emitUpdateLocation(CGF, Loc), ThreadID}; | ||||||
2104 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2105 | M, OMPRTL___kmpc_end_serialized_parallel), | ||||||
2106 | EndArgs); | ||||||
2107 | }; | ||||||
2108 | if (IfCond) { | ||||||
2109 | emitIfClause(CGF, IfCond, ThenGen, ElseGen); | ||||||
2110 | } else { | ||||||
2111 | RegionCodeGenTy ThenRCG(ThenGen); | ||||||
2112 | ThenRCG(CGF); | ||||||
2113 | } | ||||||
2114 | } | ||||||
2115 | |||||||
2116 | // If we're inside an (outlined) parallel region, use the region info's | ||||||
2117 | // thread-ID variable (it is passed in a first argument of the outlined function | ||||||
2118 | // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in | ||||||
2119 | // regular serial code region, get thread ID by calling kmp_int32 | ||||||
2120 | // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and | ||||||
2121 | // return the address of that temp. | ||||||
2122 | Address CGOpenMPRuntime::emitThreadIDAddress(CodeGenFunction &CGF, | ||||||
2123 | SourceLocation Loc) { | ||||||
2124 | if (auto *OMPRegionInfo = | ||||||
2125 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | ||||||
2126 | if (OMPRegionInfo->getThreadIDVariable()) | ||||||
2127 | return OMPRegionInfo->getThreadIDVariableLValue(CGF).getAddress(CGF); | ||||||
2128 | |||||||
2129 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | ||||||
2130 | QualType Int32Ty = | ||||||
2131 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true); | ||||||
2132 | Address ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp."); | ||||||
2133 | CGF.EmitStoreOfScalar(ThreadID, | ||||||
2134 | CGF.MakeAddrLValue(ThreadIDTemp, Int32Ty)); | ||||||
2135 | |||||||
2136 | return ThreadIDTemp; | ||||||
2137 | } | ||||||
2138 | |||||||
2139 | llvm::Constant *CGOpenMPRuntime::getOrCreateInternalVariable( | ||||||
2140 | llvm::Type *Ty, const llvm::Twine &Name, unsigned AddressSpace) { | ||||||
2141 | SmallString<256> Buffer; | ||||||
2142 | llvm::raw_svector_ostream Out(Buffer); | ||||||
2143 | Out << Name; | ||||||
2144 | StringRef RuntimeName = Out.str(); | ||||||
2145 | auto &Elem = *InternalVars.try_emplace(RuntimeName, nullptr).first; | ||||||
2146 | if (Elem.second) { | ||||||
2147 | assert(Elem.second->getType()->getPointerElementType() == Ty &&((Elem.second->getType()->getPointerElementType() == Ty && "OMP internal variable has different type than requested" ) ? static_cast<void> (0) : __assert_fail ("Elem.second->getType()->getPointerElementType() == Ty && \"OMP internal variable has different type than requested\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2148, __PRETTY_FUNCTION__)) | ||||||
2148 | "OMP internal variable has different type than requested")((Elem.second->getType()->getPointerElementType() == Ty && "OMP internal variable has different type than requested" ) ? static_cast<void> (0) : __assert_fail ("Elem.second->getType()->getPointerElementType() == Ty && \"OMP internal variable has different type than requested\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2148, __PRETTY_FUNCTION__)); | ||||||
2149 | return &*Elem.second; | ||||||
2150 | } | ||||||
2151 | |||||||
2152 | return Elem.second = new llvm::GlobalVariable( | ||||||
2153 | CGM.getModule(), Ty, /*IsConstant*/ false, | ||||||
2154 | llvm::GlobalValue::CommonLinkage, llvm::Constant::getNullValue(Ty), | ||||||
2155 | Elem.first(), /*InsertBefore=*/nullptr, | ||||||
2156 | llvm::GlobalValue::NotThreadLocal, AddressSpace); | ||||||
2157 | } | ||||||
2158 | |||||||
2159 | llvm::Value *CGOpenMPRuntime::getCriticalRegionLock(StringRef CriticalName) { | ||||||
2160 | std::string Prefix = Twine("gomp_critical_user_", CriticalName).str(); | ||||||
2161 | std::string Name = getName({Prefix, "var"}); | ||||||
2162 | return getOrCreateInternalVariable(KmpCriticalNameTy, Name); | ||||||
2163 | } | ||||||
2164 | |||||||
2165 | namespace { | ||||||
2166 | /// Common pre(post)-action for different OpenMP constructs. | ||||||
2167 | class CommonActionTy final : public PrePostActionTy { | ||||||
2168 | llvm::FunctionCallee EnterCallee; | ||||||
2169 | ArrayRef<llvm::Value *> EnterArgs; | ||||||
2170 | llvm::FunctionCallee ExitCallee; | ||||||
2171 | ArrayRef<llvm::Value *> ExitArgs; | ||||||
2172 | bool Conditional; | ||||||
2173 | llvm::BasicBlock *ContBlock = nullptr; | ||||||
2174 | |||||||
2175 | public: | ||||||
2176 | CommonActionTy(llvm::FunctionCallee EnterCallee, | ||||||
2177 | ArrayRef<llvm::Value *> EnterArgs, | ||||||
2178 | llvm::FunctionCallee ExitCallee, | ||||||
2179 | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) | ||||||
2180 | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), | ||||||
2181 | ExitArgs(ExitArgs), Conditional(Conditional) {} | ||||||
2182 | void Enter(CodeGenFunction &CGF) override { | ||||||
2183 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); | ||||||
2184 | if (Conditional) { | ||||||
2185 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); | ||||||
2186 | auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); | ||||||
2187 | ContBlock = CGF.createBasicBlock("omp_if.end"); | ||||||
2188 | // Generate the branch (If-stmt) | ||||||
2189 | CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); | ||||||
2190 | CGF.EmitBlock(ThenBlock); | ||||||
2191 | } | ||||||
2192 | } | ||||||
2193 | void Done(CodeGenFunction &CGF) { | ||||||
2194 | // Emit the rest of blocks/branches | ||||||
2195 | CGF.EmitBranch(ContBlock); | ||||||
2196 | CGF.EmitBlock(ContBlock, true); | ||||||
2197 | } | ||||||
2198 | void Exit(CodeGenFunction &CGF) override { | ||||||
2199 | CGF.EmitRuntimeCall(ExitCallee, ExitArgs); | ||||||
2200 | } | ||||||
2201 | }; | ||||||
2202 | } // anonymous namespace | ||||||
2203 | |||||||
2204 | void CGOpenMPRuntime::emitCriticalRegion(CodeGenFunction &CGF, | ||||||
2205 | StringRef CriticalName, | ||||||
2206 | const RegionCodeGenTy &CriticalOpGen, | ||||||
2207 | SourceLocation Loc, const Expr *Hint) { | ||||||
2208 | // __kmpc_critical[_with_hint](ident_t *, gtid, Lock[, hint]); | ||||||
2209 | // CriticalOpGen(); | ||||||
2210 | // __kmpc_end_critical(ident_t *, gtid, Lock); | ||||||
2211 | // Prepare arguments and build a call to __kmpc_critical | ||||||
2212 | if (!CGF.HaveInsertPoint()) | ||||||
2213 | return; | ||||||
2214 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | ||||||
2215 | getCriticalRegionLock(CriticalName)}; | ||||||
2216 | llvm::SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), | ||||||
2217 | std::end(Args)); | ||||||
2218 | if (Hint) { | ||||||
2219 | EnterArgs.push_back(CGF.Builder.CreateIntCast( | ||||||
2220 | CGF.EmitScalarExpr(Hint), CGM.Int32Ty, /*isSigned=*/false)); | ||||||
2221 | } | ||||||
2222 | CommonActionTy Action( | ||||||
2223 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2224 | CGM.getModule(), | ||||||
2225 | Hint ? OMPRTL___kmpc_critical_with_hint : OMPRTL___kmpc_critical), | ||||||
2226 | EnterArgs, | ||||||
2227 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
2228 | OMPRTL___kmpc_end_critical), | ||||||
2229 | Args); | ||||||
2230 | CriticalOpGen.setAction(Action); | ||||||
2231 | emitInlinedDirective(CGF, OMPD_critical, CriticalOpGen); | ||||||
2232 | } | ||||||
2233 | |||||||
2234 | void CGOpenMPRuntime::emitMasterRegion(CodeGenFunction &CGF, | ||||||
2235 | const RegionCodeGenTy &MasterOpGen, | ||||||
2236 | SourceLocation Loc) { | ||||||
2237 | if (!CGF.HaveInsertPoint()) | ||||||
2238 | return; | ||||||
2239 | // if(__kmpc_master(ident_t *, gtid)) { | ||||||
2240 | // MasterOpGen(); | ||||||
2241 | // __kmpc_end_master(ident_t *, gtid); | ||||||
2242 | // } | ||||||
2243 | // Prepare arguments and build a call to __kmpc_master | ||||||
2244 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | ||||||
2245 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2246 | CGM.getModule(), OMPRTL___kmpc_master), | ||||||
2247 | Args, | ||||||
2248 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2249 | CGM.getModule(), OMPRTL___kmpc_end_master), | ||||||
2250 | Args, | ||||||
2251 | /*Conditional=*/true); | ||||||
2252 | MasterOpGen.setAction(Action); | ||||||
2253 | emitInlinedDirective(CGF, OMPD_master, MasterOpGen); | ||||||
2254 | Action.Done(CGF); | ||||||
2255 | } | ||||||
2256 | |||||||
2257 | void CGOpenMPRuntime::emitTaskyieldCall(CodeGenFunction &CGF, | ||||||
2258 | SourceLocation Loc) { | ||||||
2259 | if (!CGF.HaveInsertPoint()) | ||||||
2260 | return; | ||||||
2261 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | ||||||
2262 | OMPBuilder.CreateTaskyield(CGF.Builder); | ||||||
2263 | } else { | ||||||
2264 | // Build call __kmpc_omp_taskyield(loc, thread_id, 0); | ||||||
2265 | llvm::Value *Args[] = { | ||||||
2266 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | ||||||
2267 | llvm::ConstantInt::get(CGM.IntTy, /*V=*/0, /*isSigned=*/true)}; | ||||||
2268 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2269 | CGM.getModule(), OMPRTL___kmpc_omp_taskyield), | ||||||
2270 | Args); | ||||||
2271 | } | ||||||
2272 | |||||||
2273 | if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | ||||||
2274 | Region->emitUntiedSwitch(CGF); | ||||||
2275 | } | ||||||
2276 | |||||||
2277 | void CGOpenMPRuntime::emitTaskgroupRegion(CodeGenFunction &CGF, | ||||||
2278 | const RegionCodeGenTy &TaskgroupOpGen, | ||||||
2279 | SourceLocation Loc) { | ||||||
2280 | if (!CGF.HaveInsertPoint()) | ||||||
2281 | return; | ||||||
2282 | // __kmpc_taskgroup(ident_t *, gtid); | ||||||
2283 | // TaskgroupOpGen(); | ||||||
2284 | // __kmpc_end_taskgroup(ident_t *, gtid); | ||||||
2285 | // Prepare arguments and build a call to __kmpc_taskgroup | ||||||
2286 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | ||||||
2287 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2288 | CGM.getModule(), OMPRTL___kmpc_taskgroup), | ||||||
2289 | Args, | ||||||
2290 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2291 | CGM.getModule(), OMPRTL___kmpc_end_taskgroup), | ||||||
2292 | Args); | ||||||
2293 | TaskgroupOpGen.setAction(Action); | ||||||
2294 | emitInlinedDirective(CGF, OMPD_taskgroup, TaskgroupOpGen); | ||||||
2295 | } | ||||||
2296 | |||||||
2297 | /// Given an array of pointers to variables, project the address of a | ||||||
2298 | /// given variable. | ||||||
2299 | static Address emitAddrOfVarFromArray(CodeGenFunction &CGF, Address Array, | ||||||
2300 | unsigned Index, const VarDecl *Var) { | ||||||
2301 | // Pull out the pointer to the variable. | ||||||
2302 | Address PtrAddr = CGF.Builder.CreateConstArrayGEP(Array, Index); | ||||||
2303 | llvm::Value *Ptr = CGF.Builder.CreateLoad(PtrAddr); | ||||||
2304 | |||||||
2305 | Address Addr = Address(Ptr, CGF.getContext().getDeclAlign(Var)); | ||||||
2306 | Addr = CGF.Builder.CreateElementBitCast( | ||||||
2307 | Addr, CGF.ConvertTypeForMem(Var->getType())); | ||||||
2308 | return Addr; | ||||||
2309 | } | ||||||
2310 | |||||||
2311 | static llvm::Value *emitCopyprivateCopyFunction( | ||||||
2312 | CodeGenModule &CGM, llvm::Type *ArgsType, | ||||||
2313 | ArrayRef<const Expr *> CopyprivateVars, ArrayRef<const Expr *> DestExprs, | ||||||
2314 | ArrayRef<const Expr *> SrcExprs, ArrayRef<const Expr *> AssignmentOps, | ||||||
2315 | SourceLocation Loc) { | ||||||
2316 | ASTContext &C = CGM.getContext(); | ||||||
2317 | // void copy_func(void *LHSArg, void *RHSArg); | ||||||
2318 | FunctionArgList Args; | ||||||
2319 | ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
2320 | ImplicitParamDecl::Other); | ||||||
2321 | ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
2322 | ImplicitParamDecl::Other); | ||||||
2323 | Args.push_back(&LHSArg); | ||||||
2324 | Args.push_back(&RHSArg); | ||||||
2325 | const auto &CGFI = | ||||||
2326 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
2327 | std::string Name = | ||||||
2328 | CGM.getOpenMPRuntime().getName({"omp", "copyprivate", "copy_func"}); | ||||||
2329 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | ||||||
2330 | llvm::GlobalValue::InternalLinkage, Name, | ||||||
2331 | &CGM.getModule()); | ||||||
2332 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | ||||||
2333 | Fn->setDoesNotRecurse(); | ||||||
2334 | CodeGenFunction CGF(CGM); | ||||||
2335 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | ||||||
2336 | // Dest = (void*[n])(LHSArg); | ||||||
2337 | // Src = (void*[n])(RHSArg); | ||||||
2338 | Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
2339 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)), | ||||||
2340 | ArgsType), CGF.getPointerAlign()); | ||||||
2341 | Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
2342 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)), | ||||||
2343 | ArgsType), CGF.getPointerAlign()); | ||||||
2344 | // *(Type0*)Dst[0] = *(Type0*)Src[0]; | ||||||
2345 | // *(Type1*)Dst[1] = *(Type1*)Src[1]; | ||||||
2346 | // ... | ||||||
2347 | // *(Typen*)Dst[n] = *(Typen*)Src[n]; | ||||||
2348 | for (unsigned I = 0, E = AssignmentOps.size(); I < E; ++I) { | ||||||
2349 | const auto *DestVar = | ||||||
2350 | cast<VarDecl>(cast<DeclRefExpr>(DestExprs[I])->getDecl()); | ||||||
2351 | Address DestAddr = emitAddrOfVarFromArray(CGF, LHS, I, DestVar); | ||||||
2352 | |||||||
2353 | const auto *SrcVar = | ||||||
2354 | cast<VarDecl>(cast<DeclRefExpr>(SrcExprs[I])->getDecl()); | ||||||
2355 | Address SrcAddr = emitAddrOfVarFromArray(CGF, RHS, I, SrcVar); | ||||||
2356 | |||||||
2357 | const auto *VD = cast<DeclRefExpr>(CopyprivateVars[I])->getDecl(); | ||||||
2358 | QualType Type = VD->getType(); | ||||||
2359 | CGF.EmitOMPCopy(Type, DestAddr, SrcAddr, DestVar, SrcVar, AssignmentOps[I]); | ||||||
2360 | } | ||||||
2361 | CGF.FinishFunction(); | ||||||
2362 | return Fn; | ||||||
2363 | } | ||||||
2364 | |||||||
2365 | void CGOpenMPRuntime::emitSingleRegion(CodeGenFunction &CGF, | ||||||
2366 | const RegionCodeGenTy &SingleOpGen, | ||||||
2367 | SourceLocation Loc, | ||||||
2368 | ArrayRef<const Expr *> CopyprivateVars, | ||||||
2369 | ArrayRef<const Expr *> SrcExprs, | ||||||
2370 | ArrayRef<const Expr *> DstExprs, | ||||||
2371 | ArrayRef<const Expr *> AssignmentOps) { | ||||||
2372 | if (!CGF.HaveInsertPoint()) | ||||||
2373 | return; | ||||||
2374 | assert(CopyprivateVars.size() == SrcExprs.size() &&((CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars .size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? static_cast<void> (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2376, __PRETTY_FUNCTION__)) | ||||||
2375 | CopyprivateVars.size() == DstExprs.size() &&((CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars .size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? static_cast<void> (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2376, __PRETTY_FUNCTION__)) | ||||||
2376 | CopyprivateVars.size() == AssignmentOps.size())((CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars .size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? static_cast<void> (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2376, __PRETTY_FUNCTION__)); | ||||||
2377 | ASTContext &C = CGM.getContext(); | ||||||
2378 | // int32 did_it = 0; | ||||||
2379 | // if(__kmpc_single(ident_t *, gtid)) { | ||||||
2380 | // SingleOpGen(); | ||||||
2381 | // __kmpc_end_single(ident_t *, gtid); | ||||||
2382 | // did_it = 1; | ||||||
2383 | // } | ||||||
2384 | // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>, | ||||||
2385 | // <copy_func>, did_it); | ||||||
2386 | |||||||
2387 | Address DidIt = Address::invalid(); | ||||||
2388 | if (!CopyprivateVars.empty()) { | ||||||
2389 | // int32 did_it = 0; | ||||||
2390 | QualType KmpInt32Ty = | ||||||
2391 | C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | ||||||
2392 | DidIt = CGF.CreateMemTemp(KmpInt32Ty, ".omp.copyprivate.did_it"); | ||||||
2393 | CGF.Builder.CreateStore(CGF.Builder.getInt32(0), DidIt); | ||||||
2394 | } | ||||||
2395 | // Prepare arguments and build a call to __kmpc_single | ||||||
2396 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | ||||||
2397 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2398 | CGM.getModule(), OMPRTL___kmpc_single), | ||||||
2399 | Args, | ||||||
2400 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2401 | CGM.getModule(), OMPRTL___kmpc_end_single), | ||||||
2402 | Args, | ||||||
2403 | /*Conditional=*/true); | ||||||
2404 | SingleOpGen.setAction(Action); | ||||||
2405 | emitInlinedDirective(CGF, OMPD_single, SingleOpGen); | ||||||
2406 | if (DidIt.isValid()) { | ||||||
2407 | // did_it = 1; | ||||||
2408 | CGF.Builder.CreateStore(CGF.Builder.getInt32(1), DidIt); | ||||||
2409 | } | ||||||
2410 | Action.Done(CGF); | ||||||
2411 | // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>, | ||||||
2412 | // <copy_func>, did_it); | ||||||
2413 | if (DidIt.isValid()) { | ||||||
2414 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, CopyprivateVars.size()); | ||||||
2415 | QualType CopyprivateArrayTy = C.getConstantArrayType( | ||||||
2416 | C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, | ||||||
2417 | /*IndexTypeQuals=*/0); | ||||||
2418 | // Create a list of all private variables for copyprivate. | ||||||
2419 | Address CopyprivateList = | ||||||
2420 | CGF.CreateMemTemp(CopyprivateArrayTy, ".omp.copyprivate.cpr_list"); | ||||||
2421 | for (unsigned I = 0, E = CopyprivateVars.size(); I < E; ++I) { | ||||||
2422 | Address Elem = CGF.Builder.CreateConstArrayGEP(CopyprivateList, I); | ||||||
2423 | CGF.Builder.CreateStore( | ||||||
2424 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
2425 | CGF.EmitLValue(CopyprivateVars[I]).getPointer(CGF), | ||||||
2426 | CGF.VoidPtrTy), | ||||||
2427 | Elem); | ||||||
2428 | } | ||||||
2429 | // Build function that copies private values from single region to all other | ||||||
2430 | // threads in the corresponding parallel region. | ||||||
2431 | llvm::Value *CpyFn = emitCopyprivateCopyFunction( | ||||||
2432 | CGM, CGF.ConvertTypeForMem(CopyprivateArrayTy)->getPointerTo(), | ||||||
2433 | CopyprivateVars, SrcExprs, DstExprs, AssignmentOps, Loc); | ||||||
2434 | llvm::Value *BufSize = CGF.getTypeSize(CopyprivateArrayTy); | ||||||
2435 | Address CL = | ||||||
2436 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(CopyprivateList, | ||||||
2437 | CGF.VoidPtrTy); | ||||||
2438 | llvm::Value *DidItVal = CGF.Builder.CreateLoad(DidIt); | ||||||
2439 | llvm::Value *Args[] = { | ||||||
2440 | emitUpdateLocation(CGF, Loc), // ident_t *<loc> | ||||||
2441 | getThreadID(CGF, Loc), // i32 <gtid> | ||||||
2442 | BufSize, // size_t <buf_size> | ||||||
2443 | CL.getPointer(), // void *<copyprivate list> | ||||||
2444 | CpyFn, // void (*) (void *, void *) <copy_func> | ||||||
2445 | DidItVal // i32 did_it | ||||||
2446 | }; | ||||||
2447 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2448 | CGM.getModule(), OMPRTL___kmpc_copyprivate), | ||||||
2449 | Args); | ||||||
2450 | } | ||||||
2451 | } | ||||||
2452 | |||||||
2453 | void CGOpenMPRuntime::emitOrderedRegion(CodeGenFunction &CGF, | ||||||
2454 | const RegionCodeGenTy &OrderedOpGen, | ||||||
2455 | SourceLocation Loc, bool IsThreads) { | ||||||
2456 | if (!CGF.HaveInsertPoint()) | ||||||
2457 | return; | ||||||
2458 | // __kmpc_ordered(ident_t *, gtid); | ||||||
2459 | // OrderedOpGen(); | ||||||
2460 | // __kmpc_end_ordered(ident_t *, gtid); | ||||||
2461 | // Prepare arguments and build a call to __kmpc_ordered | ||||||
2462 | if (IsThreads) { | ||||||
2463 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | ||||||
2464 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2465 | CGM.getModule(), OMPRTL___kmpc_ordered), | ||||||
2466 | Args, | ||||||
2467 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2468 | CGM.getModule(), OMPRTL___kmpc_end_ordered), | ||||||
2469 | Args); | ||||||
2470 | OrderedOpGen.setAction(Action); | ||||||
2471 | emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen); | ||||||
2472 | return; | ||||||
2473 | } | ||||||
2474 | emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen); | ||||||
2475 | } | ||||||
2476 | |||||||
2477 | unsigned CGOpenMPRuntime::getDefaultFlagsForBarriers(OpenMPDirectiveKind Kind) { | ||||||
2478 | unsigned Flags; | ||||||
2479 | if (Kind == OMPD_for) | ||||||
2480 | Flags = OMP_IDENT_BARRIER_IMPL_FOR; | ||||||
2481 | else if (Kind == OMPD_sections) | ||||||
2482 | Flags = OMP_IDENT_BARRIER_IMPL_SECTIONS; | ||||||
2483 | else if (Kind == OMPD_single) | ||||||
2484 | Flags = OMP_IDENT_BARRIER_IMPL_SINGLE; | ||||||
2485 | else if (Kind == OMPD_barrier) | ||||||
2486 | Flags = OMP_IDENT_BARRIER_EXPL; | ||||||
2487 | else | ||||||
2488 | Flags = OMP_IDENT_BARRIER_IMPL; | ||||||
2489 | return Flags; | ||||||
2490 | } | ||||||
2491 | |||||||
2492 | void CGOpenMPRuntime::getDefaultScheduleAndChunk( | ||||||
2493 | CodeGenFunction &CGF, const OMPLoopDirective &S, | ||||||
2494 | OpenMPScheduleClauseKind &ScheduleKind, const Expr *&ChunkExpr) const { | ||||||
2495 | // Check if the loop directive is actually a doacross loop directive. In this | ||||||
2496 | // case choose static, 1 schedule. | ||||||
2497 | if (llvm::any_of( | ||||||
2498 | S.getClausesOfKind<OMPOrderedClause>(), | ||||||
2499 | [](const OMPOrderedClause *C) { return C->getNumForLoops(); })) { | ||||||
2500 | ScheduleKind = OMPC_SCHEDULE_static; | ||||||
2501 | // Chunk size is 1 in this case. | ||||||
2502 | llvm::APInt ChunkSize(32, 1); | ||||||
2503 | ChunkExpr = IntegerLiteral::Create( | ||||||
2504 | CGF.getContext(), ChunkSize, | ||||||
2505 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), | ||||||
2506 | SourceLocation()); | ||||||
2507 | } | ||||||
2508 | } | ||||||
2509 | |||||||
2510 | void CGOpenMPRuntime::emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
2511 | OpenMPDirectiveKind Kind, bool EmitChecks, | ||||||
2512 | bool ForceSimpleCall) { | ||||||
2513 | // Check if we should use the OMPBuilder | ||||||
2514 | auto *OMPRegionInfo = | ||||||
2515 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo); | ||||||
2516 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | ||||||
2517 | CGF.Builder.restoreIP(OMPBuilder.CreateBarrier( | ||||||
2518 | CGF.Builder, Kind, ForceSimpleCall, EmitChecks)); | ||||||
2519 | return; | ||||||
2520 | } | ||||||
2521 | |||||||
2522 | if (!CGF.HaveInsertPoint()) | ||||||
2523 | return; | ||||||
2524 | // Build call __kmpc_cancel_barrier(loc, thread_id); | ||||||
2525 | // Build call __kmpc_barrier(loc, thread_id); | ||||||
2526 | unsigned Flags = getDefaultFlagsForBarriers(Kind); | ||||||
2527 | // Build call __kmpc_cancel_barrier(loc, thread_id) or __kmpc_barrier(loc, | ||||||
2528 | // thread_id); | ||||||
2529 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), | ||||||
2530 | getThreadID(CGF, Loc)}; | ||||||
2531 | if (OMPRegionInfo) { | ||||||
2532 | if (!ForceSimpleCall && OMPRegionInfo->hasCancel()) { | ||||||
2533 | llvm::Value *Result = CGF.EmitRuntimeCall( | ||||||
2534 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
2535 | OMPRTL___kmpc_cancel_barrier), | ||||||
2536 | Args); | ||||||
2537 | if (EmitChecks) { | ||||||
2538 | // if (__kmpc_cancel_barrier()) { | ||||||
2539 | // exit from construct; | ||||||
2540 | // } | ||||||
2541 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | ||||||
2542 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | ||||||
2543 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | ||||||
2544 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | ||||||
2545 | CGF.EmitBlock(ExitBB); | ||||||
2546 | // exit from construct; | ||||||
2547 | CodeGenFunction::JumpDest CancelDestination = | ||||||
2548 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | ||||||
2549 | CGF.EmitBranchThroughCleanup(CancelDestination); | ||||||
2550 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | ||||||
2551 | } | ||||||
2552 | return; | ||||||
2553 | } | ||||||
2554 | } | ||||||
2555 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2556 | CGM.getModule(), OMPRTL___kmpc_barrier), | ||||||
2557 | Args); | ||||||
2558 | } | ||||||
2559 | |||||||
2560 | /// Map the OpenMP loop schedule to the runtime enumeration. | ||||||
2561 | static OpenMPSchedType getRuntimeSchedule(OpenMPScheduleClauseKind ScheduleKind, | ||||||
2562 | bool Chunked, bool Ordered) { | ||||||
2563 | switch (ScheduleKind) { | ||||||
2564 | case OMPC_SCHEDULE_static: | ||||||
2565 | return Chunked ? (Ordered ? OMP_ord_static_chunked : OMP_sch_static_chunked) | ||||||
2566 | : (Ordered ? OMP_ord_static : OMP_sch_static); | ||||||
2567 | case OMPC_SCHEDULE_dynamic: | ||||||
2568 | return Ordered ? OMP_ord_dynamic_chunked : OMP_sch_dynamic_chunked; | ||||||
2569 | case OMPC_SCHEDULE_guided: | ||||||
2570 | return Ordered ? OMP_ord_guided_chunked : OMP_sch_guided_chunked; | ||||||
2571 | case OMPC_SCHEDULE_runtime: | ||||||
2572 | return Ordered ? OMP_ord_runtime : OMP_sch_runtime; | ||||||
2573 | case OMPC_SCHEDULE_auto: | ||||||
2574 | return Ordered ? OMP_ord_auto : OMP_sch_auto; | ||||||
2575 | case OMPC_SCHEDULE_unknown: | ||||||
2576 | assert(!Chunked && "chunk was specified but schedule kind not known")((!Chunked && "chunk was specified but schedule kind not known" ) ? static_cast<void> (0) : __assert_fail ("!Chunked && \"chunk was specified but schedule kind not known\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2576, __PRETTY_FUNCTION__)); | ||||||
2577 | return Ordered ? OMP_ord_static : OMP_sch_static; | ||||||
2578 | } | ||||||
2579 | llvm_unreachable("Unexpected runtime schedule")::llvm::llvm_unreachable_internal("Unexpected runtime schedule" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2579); | ||||||
2580 | } | ||||||
2581 | |||||||
2582 | /// Map the OpenMP distribute schedule to the runtime enumeration. | ||||||
2583 | static OpenMPSchedType | ||||||
2584 | getRuntimeSchedule(OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) { | ||||||
2585 | // only static is allowed for dist_schedule | ||||||
2586 | return Chunked ? OMP_dist_sch_static_chunked : OMP_dist_sch_static; | ||||||
2587 | } | ||||||
2588 | |||||||
2589 | bool CGOpenMPRuntime::isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind, | ||||||
2590 | bool Chunked) const { | ||||||
2591 | OpenMPSchedType Schedule = | ||||||
2592 | getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false); | ||||||
2593 | return Schedule == OMP_sch_static; | ||||||
2594 | } | ||||||
2595 | |||||||
2596 | bool CGOpenMPRuntime::isStaticNonchunked( | ||||||
2597 | OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const { | ||||||
2598 | OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked); | ||||||
2599 | return Schedule == OMP_dist_sch_static; | ||||||
2600 | } | ||||||
2601 | |||||||
2602 | bool CGOpenMPRuntime::isStaticChunked(OpenMPScheduleClauseKind ScheduleKind, | ||||||
2603 | bool Chunked) const { | ||||||
2604 | OpenMPSchedType Schedule = | ||||||
2605 | getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false); | ||||||
2606 | return Schedule == OMP_sch_static_chunked; | ||||||
2607 | } | ||||||
2608 | |||||||
2609 | bool CGOpenMPRuntime::isStaticChunked( | ||||||
2610 | OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const { | ||||||
2611 | OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked); | ||||||
2612 | return Schedule == OMP_dist_sch_static_chunked; | ||||||
2613 | } | ||||||
2614 | |||||||
2615 | bool CGOpenMPRuntime::isDynamic(OpenMPScheduleClauseKind ScheduleKind) const { | ||||||
2616 | OpenMPSchedType Schedule = | ||||||
2617 | getRuntimeSchedule(ScheduleKind, /*Chunked=*/false, /*Ordered=*/false); | ||||||
2618 | assert(Schedule != OMP_sch_static_chunked && "cannot be chunked here")((Schedule != OMP_sch_static_chunked && "cannot be chunked here" ) ? static_cast<void> (0) : __assert_fail ("Schedule != OMP_sch_static_chunked && \"cannot be chunked here\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2618, __PRETTY_FUNCTION__)); | ||||||
2619 | return Schedule != OMP_sch_static; | ||||||
2620 | } | ||||||
2621 | |||||||
2622 | static int addMonoNonMonoModifier(CodeGenModule &CGM, OpenMPSchedType Schedule, | ||||||
2623 | OpenMPScheduleClauseModifier M1, | ||||||
2624 | OpenMPScheduleClauseModifier M2) { | ||||||
2625 | int Modifier = 0; | ||||||
2626 | switch (M1) { | ||||||
2627 | case OMPC_SCHEDULE_MODIFIER_monotonic: | ||||||
2628 | Modifier = OMP_sch_modifier_monotonic; | ||||||
2629 | break; | ||||||
2630 | case OMPC_SCHEDULE_MODIFIER_nonmonotonic: | ||||||
2631 | Modifier = OMP_sch_modifier_nonmonotonic; | ||||||
2632 | break; | ||||||
2633 | case OMPC_SCHEDULE_MODIFIER_simd: | ||||||
2634 | if (Schedule == OMP_sch_static_chunked) | ||||||
2635 | Schedule = OMP_sch_static_balanced_chunked; | ||||||
2636 | break; | ||||||
2637 | case OMPC_SCHEDULE_MODIFIER_last: | ||||||
2638 | case OMPC_SCHEDULE_MODIFIER_unknown: | ||||||
2639 | break; | ||||||
2640 | } | ||||||
2641 | switch (M2) { | ||||||
2642 | case OMPC_SCHEDULE_MODIFIER_monotonic: | ||||||
2643 | Modifier = OMP_sch_modifier_monotonic; | ||||||
2644 | break; | ||||||
2645 | case OMPC_SCHEDULE_MODIFIER_nonmonotonic: | ||||||
2646 | Modifier = OMP_sch_modifier_nonmonotonic; | ||||||
2647 | break; | ||||||
2648 | case OMPC_SCHEDULE_MODIFIER_simd: | ||||||
2649 | if (Schedule == OMP_sch_static_chunked) | ||||||
2650 | Schedule = OMP_sch_static_balanced_chunked; | ||||||
2651 | break; | ||||||
2652 | case OMPC_SCHEDULE_MODIFIER_last: | ||||||
2653 | case OMPC_SCHEDULE_MODIFIER_unknown: | ||||||
2654 | break; | ||||||
2655 | } | ||||||
2656 | // OpenMP 5.0, 2.9.2 Worksharing-Loop Construct, Desription. | ||||||
2657 | // If the static schedule kind is specified or if the ordered clause is | ||||||
2658 | // specified, and if the nonmonotonic modifier is not specified, the effect is | ||||||
2659 | // as if the monotonic modifier is specified. Otherwise, unless the monotonic | ||||||
2660 | // modifier is specified, the effect is as if the nonmonotonic modifier is | ||||||
2661 | // specified. | ||||||
2662 | if (CGM.getLangOpts().OpenMP >= 50 && Modifier == 0) { | ||||||
2663 | if (!(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static || | ||||||
2664 | Schedule == OMP_sch_static_balanced_chunked || | ||||||
2665 | Schedule == OMP_ord_static_chunked || Schedule == OMP_ord_static || | ||||||
2666 | Schedule == OMP_dist_sch_static_chunked || | ||||||
2667 | Schedule == OMP_dist_sch_static)) | ||||||
2668 | Modifier = OMP_sch_modifier_nonmonotonic; | ||||||
2669 | } | ||||||
2670 | return Schedule | Modifier; | ||||||
2671 | } | ||||||
2672 | |||||||
2673 | void CGOpenMPRuntime::emitForDispatchInit( | ||||||
2674 | CodeGenFunction &CGF, SourceLocation Loc, | ||||||
2675 | const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned, | ||||||
2676 | bool Ordered, const DispatchRTInput &DispatchValues) { | ||||||
2677 | if (!CGF.HaveInsertPoint()) | ||||||
2678 | return; | ||||||
2679 | OpenMPSchedType Schedule = getRuntimeSchedule( | ||||||
2680 | ScheduleKind.Schedule, DispatchValues.Chunk != nullptr, Ordered); | ||||||
2681 | assert(Ordered ||((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)) ? static_cast<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)" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2684, __PRETTY_FUNCTION__)) | ||||||
2682 | (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked &&((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)) ? static_cast<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)" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2684, __PRETTY_FUNCTION__)) | ||||||
2683 | Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked &&((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)) ? static_cast<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)" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2684, __PRETTY_FUNCTION__)) | ||||||
2684 | Schedule != OMP_sch_static_balanced_chunked))((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)) ? static_cast<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)" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2684, __PRETTY_FUNCTION__)); | ||||||
2685 | // Call __kmpc_dispatch_init( | ||||||
2686 | // ident_t *loc, kmp_int32 tid, kmp_int32 schedule, | ||||||
2687 | // kmp_int[32|64] lower, kmp_int[32|64] upper, | ||||||
2688 | // kmp_int[32|64] stride, kmp_int[32|64] chunk); | ||||||
2689 | |||||||
2690 | // If the Chunk was not specified in the clause - use default value 1. | ||||||
2691 | llvm::Value *Chunk = DispatchValues.Chunk ? DispatchValues.Chunk | ||||||
2692 | : CGF.Builder.getIntN(IVSize, 1); | ||||||
2693 | llvm::Value *Args[] = { | ||||||
2694 | emitUpdateLocation(CGF, Loc), | ||||||
2695 | getThreadID(CGF, Loc), | ||||||
2696 | CGF.Builder.getInt32(addMonoNonMonoModifier( | ||||||
2697 | CGM, Schedule, ScheduleKind.M1, ScheduleKind.M2)), // Schedule type | ||||||
2698 | DispatchValues.LB, // Lower | ||||||
2699 | DispatchValues.UB, // Upper | ||||||
2700 | CGF.Builder.getIntN(IVSize, 1), // Stride | ||||||
2701 | Chunk // Chunk | ||||||
2702 | }; | ||||||
2703 | CGF.EmitRuntimeCall(createDispatchInitFunction(IVSize, IVSigned), Args); | ||||||
2704 | } | ||||||
2705 | |||||||
2706 | static void emitForStaticInitCall( | ||||||
2707 | CodeGenFunction &CGF, llvm::Value *UpdateLocation, llvm::Value *ThreadId, | ||||||
2708 | llvm::FunctionCallee ForStaticInitFunction, OpenMPSchedType Schedule, | ||||||
2709 | OpenMPScheduleClauseModifier M1, OpenMPScheduleClauseModifier M2, | ||||||
2710 | const CGOpenMPRuntime::StaticRTInput &Values) { | ||||||
2711 | if (!CGF.HaveInsertPoint()) | ||||||
2712 | return; | ||||||
2713 | |||||||
2714 | assert(!Values.Ordered)((!Values.Ordered) ? static_cast<void> (0) : __assert_fail ("!Values.Ordered", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2714, __PRETTY_FUNCTION__)); | ||||||
2715 | assert(Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked ||((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 ) ? static_cast<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" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2719, __PRETTY_FUNCTION__)) | ||||||
2716 | Schedule == OMP_sch_static_balanced_chunked ||((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 ) ? static_cast<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" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2719, __PRETTY_FUNCTION__)) | ||||||
2717 | Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked ||((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 ) ? static_cast<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" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2719, __PRETTY_FUNCTION__)) | ||||||
2718 | Schedule == OMP_dist_sch_static ||((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 ) ? static_cast<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" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2719, __PRETTY_FUNCTION__)) | ||||||
2719 | Schedule == OMP_dist_sch_static_chunked)((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 ) ? static_cast<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" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2719, __PRETTY_FUNCTION__)); | ||||||
2720 | |||||||
2721 | // Call __kmpc_for_static_init( | ||||||
2722 | // ident_t *loc, kmp_int32 tid, kmp_int32 schedtype, | ||||||
2723 | // kmp_int32 *p_lastiter, kmp_int[32|64] *p_lower, | ||||||
2724 | // kmp_int[32|64] *p_upper, kmp_int[32|64] *p_stride, | ||||||
2725 | // kmp_int[32|64] incr, kmp_int[32|64] chunk); | ||||||
2726 | llvm::Value *Chunk = Values.Chunk; | ||||||
2727 | if (Chunk == nullptr) { | ||||||
2728 | assert((Schedule == OMP_sch_static || Schedule == OMP_ord_static ||(((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule" ) ? static_cast<void> (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2730, __PRETTY_FUNCTION__)) | ||||||
2729 | Schedule == OMP_dist_sch_static) &&(((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule" ) ? static_cast<void> (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2730, __PRETTY_FUNCTION__)) | ||||||
2730 | "expected static non-chunked schedule")(((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule" ) ? static_cast<void> (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2730, __PRETTY_FUNCTION__)); | ||||||
2731 | // If the Chunk was not specified in the clause - use default value 1. | ||||||
2732 | Chunk = CGF.Builder.getIntN(Values.IVSize, 1); | ||||||
2733 | } else { | ||||||
2734 | assert((Schedule == OMP_sch_static_chunked ||(((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") ? static_cast <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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2738, __PRETTY_FUNCTION__)) | ||||||
2735 | Schedule == OMP_sch_static_balanced_chunked ||(((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") ? static_cast <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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2738, __PRETTY_FUNCTION__)) | ||||||
2736 | Schedule == OMP_ord_static_chunked ||(((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") ? static_cast <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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2738, __PRETTY_FUNCTION__)) | ||||||
2737 | Schedule == OMP_dist_sch_static_chunked) &&(((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") ? static_cast <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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2738, __PRETTY_FUNCTION__)) | ||||||
2738 | "expected static chunked schedule")(((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") ? static_cast <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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2738, __PRETTY_FUNCTION__)); | ||||||
2739 | } | ||||||
2740 | llvm::Value *Args[] = { | ||||||
2741 | UpdateLocation, | ||||||
2742 | ThreadId, | ||||||
2743 | CGF.Builder.getInt32(addMonoNonMonoModifier(CGF.CGM, Schedule, M1, | ||||||
2744 | M2)), // Schedule type | ||||||
2745 | Values.IL.getPointer(), // &isLastIter | ||||||
2746 | Values.LB.getPointer(), // &LB | ||||||
2747 | Values.UB.getPointer(), // &UB | ||||||
2748 | Values.ST.getPointer(), // &Stride | ||||||
2749 | CGF.Builder.getIntN(Values.IVSize, 1), // Incr | ||||||
2750 | Chunk // Chunk | ||||||
2751 | }; | ||||||
2752 | CGF.EmitRuntimeCall(ForStaticInitFunction, Args); | ||||||
2753 | } | ||||||
2754 | |||||||
2755 | void CGOpenMPRuntime::emitForStaticInit(CodeGenFunction &CGF, | ||||||
2756 | SourceLocation Loc, | ||||||
2757 | OpenMPDirectiveKind DKind, | ||||||
2758 | const OpenMPScheduleTy &ScheduleKind, | ||||||
2759 | const StaticRTInput &Values) { | ||||||
2760 | OpenMPSchedType ScheduleNum = getRuntimeSchedule( | ||||||
2761 | ScheduleKind.Schedule, Values.Chunk != nullptr, Values.Ordered); | ||||||
2762 | assert(isOpenMPWorksharingDirective(DKind) &&((isOpenMPWorksharingDirective(DKind) && "Expected loop-based or sections-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(DKind) && \"Expected loop-based or sections-based directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2763, __PRETTY_FUNCTION__)) | ||||||
2763 | "Expected loop-based or sections-based directive.")((isOpenMPWorksharingDirective(DKind) && "Expected loop-based or sections-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(DKind) && \"Expected loop-based or sections-based directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2763, __PRETTY_FUNCTION__)); | ||||||
2764 | llvm::Value *UpdatedLocation = emitUpdateLocation(CGF, Loc, | ||||||
2765 | isOpenMPLoopDirective(DKind) | ||||||
2766 | ? OMP_IDENT_WORK_LOOP | ||||||
2767 | : OMP_IDENT_WORK_SECTIONS); | ||||||
2768 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | ||||||
2769 | llvm::FunctionCallee StaticInitFunction = | ||||||
2770 | createForStaticInitFunction(Values.IVSize, Values.IVSigned); | ||||||
2771 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | ||||||
2772 | emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction, | ||||||
2773 | ScheduleNum, ScheduleKind.M1, ScheduleKind.M2, Values); | ||||||
2774 | } | ||||||
2775 | |||||||
2776 | void CGOpenMPRuntime::emitDistributeStaticInit( | ||||||
2777 | CodeGenFunction &CGF, SourceLocation Loc, | ||||||
2778 | OpenMPDistScheduleClauseKind SchedKind, | ||||||
2779 | const CGOpenMPRuntime::StaticRTInput &Values) { | ||||||
2780 | OpenMPSchedType ScheduleNum = | ||||||
2781 | getRuntimeSchedule(SchedKind, Values.Chunk != nullptr); | ||||||
2782 | llvm::Value *UpdatedLocation = | ||||||
2783 | emitUpdateLocation(CGF, Loc, OMP_IDENT_WORK_DISTRIBUTE); | ||||||
2784 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | ||||||
2785 | llvm::FunctionCallee StaticInitFunction = | ||||||
2786 | createForStaticInitFunction(Values.IVSize, Values.IVSigned); | ||||||
2787 | emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction, | ||||||
2788 | ScheduleNum, OMPC_SCHEDULE_MODIFIER_unknown, | ||||||
2789 | OMPC_SCHEDULE_MODIFIER_unknown, Values); | ||||||
2790 | } | ||||||
2791 | |||||||
2792 | void CGOpenMPRuntime::emitForStaticFinish(CodeGenFunction &CGF, | ||||||
2793 | SourceLocation Loc, | ||||||
2794 | OpenMPDirectiveKind DKind) { | ||||||
2795 | if (!CGF.HaveInsertPoint()) | ||||||
2796 | return; | ||||||
2797 | // Call __kmpc_for_static_fini(ident_t *loc, kmp_int32 tid); | ||||||
2798 | llvm::Value *Args[] = { | ||||||
2799 | emitUpdateLocation(CGF, Loc, | ||||||
2800 | isOpenMPDistributeDirective(DKind) | ||||||
2801 | ? OMP_IDENT_WORK_DISTRIBUTE | ||||||
2802 | : isOpenMPLoopDirective(DKind) | ||||||
2803 | ? OMP_IDENT_WORK_LOOP | ||||||
2804 | : OMP_IDENT_WORK_SECTIONS), | ||||||
2805 | getThreadID(CGF, Loc)}; | ||||||
2806 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | ||||||
2807 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2808 | CGM.getModule(), OMPRTL___kmpc_for_static_fini), | ||||||
2809 | Args); | ||||||
2810 | } | ||||||
2811 | |||||||
2812 | void CGOpenMPRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF, | ||||||
2813 | SourceLocation Loc, | ||||||
2814 | unsigned IVSize, | ||||||
2815 | bool IVSigned) { | ||||||
2816 | if (!CGF.HaveInsertPoint()) | ||||||
2817 | return; | ||||||
2818 | // Call __kmpc_for_dynamic_fini_(4|8)[u](ident_t *loc, kmp_int32 tid); | ||||||
2819 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | ||||||
2820 | CGF.EmitRuntimeCall(createDispatchFiniFunction(IVSize, IVSigned), Args); | ||||||
2821 | } | ||||||
2822 | |||||||
2823 | llvm::Value *CGOpenMPRuntime::emitForNext(CodeGenFunction &CGF, | ||||||
2824 | SourceLocation Loc, unsigned IVSize, | ||||||
2825 | bool IVSigned, Address IL, | ||||||
2826 | Address LB, Address UB, | ||||||
2827 | Address ST) { | ||||||
2828 | // Call __kmpc_dispatch_next( | ||||||
2829 | // ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter, | ||||||
2830 | // kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper, | ||||||
2831 | // kmp_int[32|64] *p_stride); | ||||||
2832 | llvm::Value *Args[] = { | ||||||
2833 | emitUpdateLocation(CGF, Loc), | ||||||
2834 | getThreadID(CGF, Loc), | ||||||
2835 | IL.getPointer(), // &isLastIter | ||||||
2836 | LB.getPointer(), // &Lower | ||||||
2837 | UB.getPointer(), // &Upper | ||||||
2838 | ST.getPointer() // &Stride | ||||||
2839 | }; | ||||||
2840 | llvm::Value *Call = | ||||||
2841 | CGF.EmitRuntimeCall(createDispatchNextFunction(IVSize, IVSigned), Args); | ||||||
2842 | return CGF.EmitScalarConversion( | ||||||
2843 | Call, CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/1), | ||||||
2844 | CGF.getContext().BoolTy, Loc); | ||||||
2845 | } | ||||||
2846 | |||||||
2847 | void CGOpenMPRuntime::emitNumThreadsClause(CodeGenFunction &CGF, | ||||||
2848 | llvm::Value *NumThreads, | ||||||
2849 | SourceLocation Loc) { | ||||||
2850 | if (!CGF.HaveInsertPoint()) | ||||||
2851 | return; | ||||||
2852 | // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads) | ||||||
2853 | llvm::Value *Args[] = { | ||||||
2854 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | ||||||
2855 | CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)}; | ||||||
2856 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2857 | CGM.getModule(), OMPRTL___kmpc_push_num_threads), | ||||||
2858 | Args); | ||||||
2859 | } | ||||||
2860 | |||||||
2861 | void CGOpenMPRuntime::emitProcBindClause(CodeGenFunction &CGF, | ||||||
2862 | ProcBindKind ProcBind, | ||||||
2863 | SourceLocation Loc) { | ||||||
2864 | if (!CGF.HaveInsertPoint()) | ||||||
2865 | return; | ||||||
2866 | assert(ProcBind != OMP_PROC_BIND_unknown && "Unsupported proc_bind value.")((ProcBind != OMP_PROC_BIND_unknown && "Unsupported proc_bind value." ) ? static_cast<void> (0) : __assert_fail ("ProcBind != OMP_PROC_BIND_unknown && \"Unsupported proc_bind value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2866, __PRETTY_FUNCTION__)); | ||||||
2867 | // Build call __kmpc_push_proc_bind(&loc, global_tid, proc_bind) | ||||||
2868 | llvm::Value *Args[] = { | ||||||
2869 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | ||||||
2870 | llvm::ConstantInt::get(CGM.IntTy, unsigned(ProcBind), /*isSigned=*/true)}; | ||||||
2871 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2872 | CGM.getModule(), OMPRTL___kmpc_push_proc_bind), | ||||||
2873 | Args); | ||||||
2874 | } | ||||||
2875 | |||||||
2876 | void CGOpenMPRuntime::emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>, | ||||||
2877 | SourceLocation Loc, llvm::AtomicOrdering AO) { | ||||||
2878 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | ||||||
2879 | OMPBuilder.CreateFlush(CGF.Builder); | ||||||
2880 | } else { | ||||||
2881 | if (!CGF.HaveInsertPoint()) | ||||||
2882 | return; | ||||||
2883 | // Build call void __kmpc_flush(ident_t *loc) | ||||||
2884 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
2885 | CGM.getModule(), OMPRTL___kmpc_flush), | ||||||
2886 | emitUpdateLocation(CGF, Loc)); | ||||||
2887 | } | ||||||
2888 | } | ||||||
2889 | |||||||
2890 | namespace { | ||||||
2891 | /// Indexes of fields for type kmp_task_t. | ||||||
2892 | enum KmpTaskTFields { | ||||||
2893 | /// List of shared variables. | ||||||
2894 | KmpTaskTShareds, | ||||||
2895 | /// Task routine. | ||||||
2896 | KmpTaskTRoutine, | ||||||
2897 | /// Partition id for the untied tasks. | ||||||
2898 | KmpTaskTPartId, | ||||||
2899 | /// Function with call of destructors for private variables. | ||||||
2900 | Data1, | ||||||
2901 | /// Task priority. | ||||||
2902 | Data2, | ||||||
2903 | /// (Taskloops only) Lower bound. | ||||||
2904 | KmpTaskTLowerBound, | ||||||
2905 | /// (Taskloops only) Upper bound. | ||||||
2906 | KmpTaskTUpperBound, | ||||||
2907 | /// (Taskloops only) Stride. | ||||||
2908 | KmpTaskTStride, | ||||||
2909 | /// (Taskloops only) Is last iteration flag. | ||||||
2910 | KmpTaskTLastIter, | ||||||
2911 | /// (Taskloops only) Reduction data. | ||||||
2912 | KmpTaskTReductions, | ||||||
2913 | }; | ||||||
2914 | } // anonymous namespace | ||||||
2915 | |||||||
2916 | bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::empty() const { | ||||||
2917 | return OffloadEntriesTargetRegion.empty() && | ||||||
2918 | OffloadEntriesDeviceGlobalVar.empty(); | ||||||
2919 | } | ||||||
2920 | |||||||
2921 | /// Initialize target region entry. | ||||||
2922 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | ||||||
2923 | initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, | ||||||
2924 | StringRef ParentName, unsigned LineNum, | ||||||
2925 | unsigned Order) { | ||||||
2926 | assert(CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2928, __PRETTY_FUNCTION__)) | ||||||
2927 | "only required for the device "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2928, __PRETTY_FUNCTION__)) | ||||||
2928 | "code generation.")((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2928, __PRETTY_FUNCTION__)); | ||||||
2929 | OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] = | ||||||
2930 | OffloadEntryInfoTargetRegion(Order, /*Addr=*/nullptr, /*ID=*/nullptr, | ||||||
2931 | OMPTargetRegionEntryTargetRegion); | ||||||
2932 | ++OffloadingEntriesNum; | ||||||
2933 | } | ||||||
2934 | |||||||
2935 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | ||||||
2936 | registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, | ||||||
2937 | StringRef ParentName, unsigned LineNum, | ||||||
2938 | llvm::Constant *Addr, llvm::Constant *ID, | ||||||
2939 | OMPTargetRegionEntryKind Flags) { | ||||||
2940 | // If we are emitting code for a target, the entry is already initialized, | ||||||
2941 | // only has to be registered. | ||||||
2942 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
2943 | if (!hasTargetRegionEntryInfo(DeviceID, FileID, ParentName, LineNum)) { | ||||||
2944 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
2945 | DiagnosticsEngine::Error, | ||||||
2946 | "Unable to find target region on line '%0' in the device code."); | ||||||
2947 | CGM.getDiags().Report(DiagID) << LineNum; | ||||||
2948 | return; | ||||||
2949 | } | ||||||
2950 | auto &Entry = | ||||||
2951 | OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum]; | ||||||
2952 | assert(Entry.isValid() && "Entry not initialized!")((Entry.isValid() && "Entry not initialized!") ? static_cast <void> (0) : __assert_fail ("Entry.isValid() && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2952, __PRETTY_FUNCTION__)); | ||||||
2953 | Entry.setAddress(Addr); | ||||||
2954 | Entry.setID(ID); | ||||||
2955 | Entry.setFlags(Flags); | ||||||
2956 | } else { | ||||||
2957 | OffloadEntryInfoTargetRegion Entry(OffloadingEntriesNum, Addr, ID, Flags); | ||||||
2958 | OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] = Entry; | ||||||
2959 | ++OffloadingEntriesNum; | ||||||
2960 | } | ||||||
2961 | } | ||||||
2962 | |||||||
2963 | bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::hasTargetRegionEntryInfo( | ||||||
2964 | unsigned DeviceID, unsigned FileID, StringRef ParentName, | ||||||
2965 | unsigned LineNum) const { | ||||||
2966 | auto PerDevice = OffloadEntriesTargetRegion.find(DeviceID); | ||||||
2967 | if (PerDevice == OffloadEntriesTargetRegion.end()) | ||||||
2968 | return false; | ||||||
2969 | auto PerFile = PerDevice->second.find(FileID); | ||||||
2970 | if (PerFile == PerDevice->second.end()) | ||||||
2971 | return false; | ||||||
2972 | auto PerParentName = PerFile->second.find(ParentName); | ||||||
2973 | if (PerParentName == PerFile->second.end()) | ||||||
2974 | return false; | ||||||
2975 | auto PerLine = PerParentName->second.find(LineNum); | ||||||
2976 | if (PerLine == PerParentName->second.end()) | ||||||
2977 | return false; | ||||||
2978 | // Fail if this entry is already registered. | ||||||
2979 | if (PerLine->second.getAddress() || PerLine->second.getID()) | ||||||
2980 | return false; | ||||||
2981 | return true; | ||||||
2982 | } | ||||||
2983 | |||||||
2984 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::actOnTargetRegionEntriesInfo( | ||||||
2985 | const OffloadTargetRegionEntryInfoActTy &Action) { | ||||||
2986 | // Scan all target region entries and perform the provided action. | ||||||
2987 | for (const auto &D : OffloadEntriesTargetRegion) | ||||||
2988 | for (const auto &F : D.second) | ||||||
2989 | for (const auto &P : F.second) | ||||||
2990 | for (const auto &L : P.second) | ||||||
2991 | Action(D.first, F.first, P.first(), L.first, L.second); | ||||||
2992 | } | ||||||
2993 | |||||||
2994 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | ||||||
2995 | initializeDeviceGlobalVarEntryInfo(StringRef Name, | ||||||
2996 | OMPTargetGlobalVarEntryKind Flags, | ||||||
2997 | unsigned Order) { | ||||||
2998 | assert(CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3000, __PRETTY_FUNCTION__)) | ||||||
2999 | "only required for the device "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3000, __PRETTY_FUNCTION__)) | ||||||
3000 | "code generation.")((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3000, __PRETTY_FUNCTION__)); | ||||||
3001 | OffloadEntriesDeviceGlobalVar.try_emplace(Name, Order, Flags); | ||||||
3002 | ++OffloadingEntriesNum; | ||||||
3003 | } | ||||||
3004 | |||||||
3005 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | ||||||
3006 | registerDeviceGlobalVarEntryInfo(StringRef VarName, llvm::Constant *Addr, | ||||||
3007 | CharUnits VarSize, | ||||||
3008 | OMPTargetGlobalVarEntryKind Flags, | ||||||
3009 | llvm::GlobalValue::LinkageTypes Linkage) { | ||||||
3010 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
3011 | auto &Entry = OffloadEntriesDeviceGlobalVar[VarName]; | ||||||
3012 | assert(Entry.isValid() && Entry.getFlags() == Flags &&((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3013, __PRETTY_FUNCTION__)) | ||||||
3013 | "Entry not initialized!")((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3013, __PRETTY_FUNCTION__)); | ||||||
3014 | assert((!Entry.getAddress() || Entry.getAddress() == Addr) &&(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3015, __PRETTY_FUNCTION__)) | ||||||
3015 | "Resetting with the new address.")(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3015, __PRETTY_FUNCTION__)); | ||||||
3016 | if (Entry.getAddress() && hasDeviceGlobalVarEntryInfo(VarName)) { | ||||||
3017 | if (Entry.getVarSize().isZero()) { | ||||||
3018 | Entry.setVarSize(VarSize); | ||||||
3019 | Entry.setLinkage(Linkage); | ||||||
3020 | } | ||||||
3021 | return; | ||||||
3022 | } | ||||||
3023 | Entry.setVarSize(VarSize); | ||||||
3024 | Entry.setLinkage(Linkage); | ||||||
3025 | Entry.setAddress(Addr); | ||||||
3026 | } else { | ||||||
3027 | if (hasDeviceGlobalVarEntryInfo(VarName)) { | ||||||
3028 | auto &Entry = OffloadEntriesDeviceGlobalVar[VarName]; | ||||||
3029 | assert(Entry.isValid() && Entry.getFlags() == Flags &&((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3030, __PRETTY_FUNCTION__)) | ||||||
3030 | "Entry not initialized!")((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3030, __PRETTY_FUNCTION__)); | ||||||
3031 | assert((!Entry.getAddress() || Entry.getAddress() == Addr) &&(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3032, __PRETTY_FUNCTION__)) | ||||||
3032 | "Resetting with the new address.")(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3032, __PRETTY_FUNCTION__)); | ||||||
3033 | if (Entry.getVarSize().isZero()) { | ||||||
3034 | Entry.setVarSize(VarSize); | ||||||
3035 | Entry.setLinkage(Linkage); | ||||||
3036 | } | ||||||
3037 | return; | ||||||
3038 | } | ||||||
3039 | OffloadEntriesDeviceGlobalVar.try_emplace( | ||||||
3040 | VarName, OffloadingEntriesNum, Addr, VarSize, Flags, Linkage); | ||||||
3041 | ++OffloadingEntriesNum; | ||||||
3042 | } | ||||||
3043 | } | ||||||
3044 | |||||||
3045 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | ||||||
3046 | actOnDeviceGlobalVarEntriesInfo( | ||||||
3047 | const OffloadDeviceGlobalVarEntryInfoActTy &Action) { | ||||||
3048 | // Scan all target region entries and perform the provided action. | ||||||
3049 | for (const auto &E : OffloadEntriesDeviceGlobalVar) | ||||||
3050 | Action(E.getKey(), E.getValue()); | ||||||
3051 | } | ||||||
3052 | |||||||
3053 | void CGOpenMPRuntime::createOffloadEntry( | ||||||
3054 | llvm::Constant *ID, llvm::Constant *Addr, uint64_t Size, int32_t Flags, | ||||||
3055 | llvm::GlobalValue::LinkageTypes Linkage) { | ||||||
3056 | StringRef Name = Addr->getName(); | ||||||
3057 | llvm::Module &M = CGM.getModule(); | ||||||
3058 | llvm::LLVMContext &C = M.getContext(); | ||||||
3059 | |||||||
3060 | // Create constant string with the name. | ||||||
3061 | llvm::Constant *StrPtrInit = llvm::ConstantDataArray::getString(C, Name); | ||||||
3062 | |||||||
3063 | std::string StringName = getName({"omp_offloading", "entry_name"}); | ||||||
3064 | auto *Str = new llvm::GlobalVariable( | ||||||
3065 | M, StrPtrInit->getType(), /*isConstant=*/true, | ||||||
3066 | llvm::GlobalValue::InternalLinkage, StrPtrInit, StringName); | ||||||
3067 | Str->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | ||||||
3068 | |||||||
3069 | llvm::Constant *Data[] = {llvm::ConstantExpr::getBitCast(ID, CGM.VoidPtrTy), | ||||||
3070 | llvm::ConstantExpr::getBitCast(Str, CGM.Int8PtrTy), | ||||||
3071 | llvm::ConstantInt::get(CGM.SizeTy, Size), | ||||||
3072 | llvm::ConstantInt::get(CGM.Int32Ty, Flags), | ||||||
3073 | llvm::ConstantInt::get(CGM.Int32Ty, 0)}; | ||||||
3074 | std::string EntryName = getName({"omp_offloading", "entry", ""}); | ||||||
3075 | llvm::GlobalVariable *Entry = createGlobalStruct( | ||||||
3076 | CGM, getTgtOffloadEntryQTy(), /*IsConstant=*/true, Data, | ||||||
3077 | Twine(EntryName).concat(Name), llvm::GlobalValue::WeakAnyLinkage); | ||||||
3078 | |||||||
3079 | // The entry has to be created in the section the linker expects it to be. | ||||||
3080 | Entry->setSection("omp_offloading_entries"); | ||||||
3081 | } | ||||||
3082 | |||||||
3083 | void CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata() { | ||||||
3084 | // Emit the offloading entries and metadata so that the device codegen side | ||||||
3085 | // can easily figure out what to emit. The produced metadata looks like | ||||||
3086 | // this: | ||||||
3087 | // | ||||||
3088 | // !omp_offload.info = !{!1, ...} | ||||||
3089 | // | ||||||
3090 | // Right now we only generate metadata for function that contain target | ||||||
3091 | // regions. | ||||||
3092 | |||||||
3093 | // If we are in simd mode or there are no entries, we don't need to do | ||||||
3094 | // anything. | ||||||
3095 | if (CGM.getLangOpts().OpenMPSimd || OffloadEntriesInfoManager.empty()) | ||||||
3096 | return; | ||||||
3097 | |||||||
3098 | llvm::Module &M = CGM.getModule(); | ||||||
3099 | llvm::LLVMContext &C = M.getContext(); | ||||||
3100 | SmallVector<std::tuple<const OffloadEntriesInfoManagerTy::OffloadEntryInfo *, | ||||||
3101 | SourceLocation, StringRef>, | ||||||
3102 | 16> | ||||||
3103 | OrderedEntries(OffloadEntriesInfoManager.size()); | ||||||
3104 | llvm::SmallVector<StringRef, 16> ParentFunctions( | ||||||
3105 | OffloadEntriesInfoManager.size()); | ||||||
3106 | |||||||
3107 | // Auxiliary methods to create metadata values and strings. | ||||||
3108 | auto &&GetMDInt = [this](unsigned V) { | ||||||
3109 | return llvm::ConstantAsMetadata::get( | ||||||
3110 | llvm::ConstantInt::get(CGM.Int32Ty, V)); | ||||||
3111 | }; | ||||||
3112 | |||||||
3113 | auto &&GetMDString = [&C](StringRef V) { return llvm::MDString::get(C, V); }; | ||||||
3114 | |||||||
3115 | // Create the offloading info metadata node. | ||||||
3116 | llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("omp_offload.info"); | ||||||
3117 | |||||||
3118 | // Create function that emits metadata for each target region entry; | ||||||
3119 | auto &&TargetRegionMetadataEmitter = | ||||||
3120 | [this, &C, MD, &OrderedEntries, &ParentFunctions, &GetMDInt, | ||||||
3121 | &GetMDString]( | ||||||
3122 | unsigned DeviceID, unsigned FileID, StringRef ParentName, | ||||||
3123 | unsigned Line, | ||||||
3124 | const OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion &E) { | ||||||
3125 | // Generate metadata for target regions. Each entry of this metadata | ||||||
3126 | // contains: | ||||||
3127 | // - Entry 0 -> Kind of this type of metadata (0). | ||||||
3128 | // - Entry 1 -> Device ID of the file where the entry was identified. | ||||||
3129 | // - Entry 2 -> File ID of the file where the entry was identified. | ||||||
3130 | // - Entry 3 -> Mangled name of the function where the entry was | ||||||
3131 | // identified. | ||||||
3132 | // - Entry 4 -> Line in the file where the entry was identified. | ||||||
3133 | // - Entry 5 -> Order the entry was created. | ||||||
3134 | // The first element of the metadata node is the kind. | ||||||
3135 | llvm::Metadata *Ops[] = {GetMDInt(E.getKind()), GetMDInt(DeviceID), | ||||||
3136 | GetMDInt(FileID), GetMDString(ParentName), | ||||||
3137 | GetMDInt(Line), GetMDInt(E.getOrder())}; | ||||||
3138 | |||||||
3139 | SourceLocation Loc; | ||||||
3140 | for (auto I = CGM.getContext().getSourceManager().fileinfo_begin(), | ||||||
3141 | E = CGM.getContext().getSourceManager().fileinfo_end(); | ||||||
3142 | I != E; ++I) { | ||||||
3143 | if (I->getFirst()->getUniqueID().getDevice() == DeviceID && | ||||||
3144 | I->getFirst()->getUniqueID().getFile() == FileID) { | ||||||
3145 | Loc = CGM.getContext().getSourceManager().translateFileLineCol( | ||||||
3146 | I->getFirst(), Line, 1); | ||||||
3147 | break; | ||||||
3148 | } | ||||||
3149 | } | ||||||
3150 | // Save this entry in the right position of the ordered entries array. | ||||||
3151 | OrderedEntries[E.getOrder()] = std::make_tuple(&E, Loc, ParentName); | ||||||
3152 | ParentFunctions[E.getOrder()] = ParentName; | ||||||
3153 | |||||||
3154 | // Add metadata to the named metadata node. | ||||||
3155 | MD->addOperand(llvm::MDNode::get(C, Ops)); | ||||||
3156 | }; | ||||||
3157 | |||||||
3158 | OffloadEntriesInfoManager.actOnTargetRegionEntriesInfo( | ||||||
3159 | TargetRegionMetadataEmitter); | ||||||
3160 | |||||||
3161 | // Create function that emits metadata for each device global variable entry; | ||||||
3162 | auto &&DeviceGlobalVarMetadataEmitter = | ||||||
3163 | [&C, &OrderedEntries, &GetMDInt, &GetMDString, | ||||||
3164 | MD](StringRef MangledName, | ||||||
3165 | const OffloadEntriesInfoManagerTy::OffloadEntryInfoDeviceGlobalVar | ||||||
3166 | &E) { | ||||||
3167 | // Generate metadata for global variables. Each entry of this metadata | ||||||
3168 | // contains: | ||||||
3169 | // - Entry 0 -> Kind of this type of metadata (1). | ||||||
3170 | // - Entry 1 -> Mangled name of the variable. | ||||||
3171 | // - Entry 2 -> Declare target kind. | ||||||
3172 | // - Entry 3 -> Order the entry was created. | ||||||
3173 | // The first element of the metadata node is the kind. | ||||||
3174 | llvm::Metadata *Ops[] = { | ||||||
3175 | GetMDInt(E.getKind()), GetMDString(MangledName), | ||||||
3176 | GetMDInt(E.getFlags()), GetMDInt(E.getOrder())}; | ||||||
3177 | |||||||
3178 | // Save this entry in the right position of the ordered entries array. | ||||||
3179 | OrderedEntries[E.getOrder()] = | ||||||
3180 | std::make_tuple(&E, SourceLocation(), MangledName); | ||||||
3181 | |||||||
3182 | // Add metadata to the named metadata node. | ||||||
3183 | MD->addOperand(llvm::MDNode::get(C, Ops)); | ||||||
3184 | }; | ||||||
3185 | |||||||
3186 | OffloadEntriesInfoManager.actOnDeviceGlobalVarEntriesInfo( | ||||||
3187 | DeviceGlobalVarMetadataEmitter); | ||||||
3188 | |||||||
3189 | for (const auto &E : OrderedEntries) { | ||||||
3190 | assert(std::get<0>(E) && "All ordered entries must exist!")((std::get<0>(E) && "All ordered entries must exist!" ) ? static_cast<void> (0) : __assert_fail ("std::get<0>(E) && \"All ordered entries must exist!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3190, __PRETTY_FUNCTION__)); | ||||||
3191 | if (const auto *CE = | ||||||
3192 | dyn_cast<OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion>( | ||||||
3193 | std::get<0>(E))) { | ||||||
3194 | if (!CE->getID() || !CE->getAddress()) { | ||||||
3195 | // Do not blame the entry if the parent funtion is not emitted. | ||||||
3196 | StringRef FnName = ParentFunctions[CE->getOrder()]; | ||||||
3197 | if (!CGM.GetGlobalValue(FnName)) | ||||||
3198 | continue; | ||||||
3199 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
3200 | DiagnosticsEngine::Error, | ||||||
3201 | "Offloading entry for target region in %0 is incorrect: either the " | ||||||
3202 | "address or the ID is invalid."); | ||||||
3203 | CGM.getDiags().Report(std::get<1>(E), DiagID) << FnName; | ||||||
3204 | continue; | ||||||
3205 | } | ||||||
3206 | createOffloadEntry(CE->getID(), CE->getAddress(), /*Size=*/0, | ||||||
3207 | CE->getFlags(), llvm::GlobalValue::WeakAnyLinkage); | ||||||
3208 | } else if (const auto *CE = dyn_cast<OffloadEntriesInfoManagerTy:: | ||||||
3209 | OffloadEntryInfoDeviceGlobalVar>( | ||||||
3210 | std::get<0>(E))) { | ||||||
3211 | OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind Flags = | ||||||
3212 | static_cast<OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind>( | ||||||
3213 | CE->getFlags()); | ||||||
3214 | switch (Flags) { | ||||||
3215 | case OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo: { | ||||||
3216 | if (CGM.getLangOpts().OpenMPIsDevice && | ||||||
3217 | CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory()) | ||||||
3218 | continue; | ||||||
3219 | if (!CE->getAddress()) { | ||||||
3220 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
3221 | DiagnosticsEngine::Error, "Offloading entry for declare target " | ||||||
3222 | "variable %0 is incorrect: the " | ||||||
3223 | "address is invalid."); | ||||||
3224 | CGM.getDiags().Report(std::get<1>(E), DiagID) << std::get<2>(E); | ||||||
3225 | continue; | ||||||
3226 | } | ||||||
3227 | // The vaiable has no definition - no need to add the entry. | ||||||
3228 | if (CE->getVarSize().isZero()) | ||||||
3229 | continue; | ||||||
3230 | break; | ||||||
3231 | } | ||||||
3232 | case OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryLink: | ||||||
3233 | assert(((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) ||((((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress ()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress ())) && "Declaret target link address is set.") ? static_cast <void> (0) : __assert_fail ("((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) && \"Declaret target link address is set.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3235, __PRETTY_FUNCTION__)) | ||||||
3234 | (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) &&((((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress ()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress ())) && "Declaret target link address is set.") ? static_cast <void> (0) : __assert_fail ("((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) && \"Declaret target link address is set.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3235, __PRETTY_FUNCTION__)) | ||||||
3235 | "Declaret target link address is set.")((((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress ()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress ())) && "Declaret target link address is set.") ? static_cast <void> (0) : __assert_fail ("((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) && \"Declaret target link address is set.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3235, __PRETTY_FUNCTION__)); | ||||||
3236 | if (CGM.getLangOpts().OpenMPIsDevice) | ||||||
3237 | continue; | ||||||
3238 | if (!CE->getAddress()) { | ||||||
3239 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
3240 | DiagnosticsEngine::Error, | ||||||
3241 | "Offloading entry for declare target variable is incorrect: the " | ||||||
3242 | "address is invalid."); | ||||||
3243 | CGM.getDiags().Report(DiagID); | ||||||
3244 | continue; | ||||||
3245 | } | ||||||
3246 | break; | ||||||
3247 | } | ||||||
3248 | createOffloadEntry(CE->getAddress(), CE->getAddress(), | ||||||
3249 | CE->getVarSize().getQuantity(), Flags, | ||||||
3250 | CE->getLinkage()); | ||||||
3251 | } else { | ||||||
3252 | llvm_unreachable("Unsupported entry kind.")::llvm::llvm_unreachable_internal("Unsupported entry kind.", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3252); | ||||||
3253 | } | ||||||
3254 | } | ||||||
3255 | } | ||||||
3256 | |||||||
3257 | /// Loads all the offload entries information from the host IR | ||||||
3258 | /// metadata. | ||||||
3259 | void CGOpenMPRuntime::loadOffloadInfoMetadata() { | ||||||
3260 | // If we are in target mode, load the metadata from the host IR. This code has | ||||||
3261 | // to match the metadaata creation in createOffloadEntriesAndInfoMetadata(). | ||||||
3262 | |||||||
3263 | if (!CGM.getLangOpts().OpenMPIsDevice) | ||||||
3264 | return; | ||||||
3265 | |||||||
3266 | if (CGM.getLangOpts().OMPHostIRFile.empty()) | ||||||
3267 | return; | ||||||
3268 | |||||||
3269 | auto Buf = llvm::MemoryBuffer::getFile(CGM.getLangOpts().OMPHostIRFile); | ||||||
3270 | if (auto EC = Buf.getError()) { | ||||||
3271 | CGM.getDiags().Report(diag::err_cannot_open_file) | ||||||
3272 | << CGM.getLangOpts().OMPHostIRFile << EC.message(); | ||||||
3273 | return; | ||||||
3274 | } | ||||||
3275 | |||||||
3276 | llvm::LLVMContext C; | ||||||
3277 | auto ME = expectedToErrorOrAndEmitErrors( | ||||||
3278 | C, llvm::parseBitcodeFile(Buf.get()->getMemBufferRef(), C)); | ||||||
3279 | |||||||
3280 | if (auto EC = ME.getError()) { | ||||||
3281 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
3282 | DiagnosticsEngine::Error, "Unable to parse host IR file '%0':'%1'"); | ||||||
3283 | CGM.getDiags().Report(DiagID) | ||||||
3284 | << CGM.getLangOpts().OMPHostIRFile << EC.message(); | ||||||
3285 | return; | ||||||
3286 | } | ||||||
3287 | |||||||
3288 | llvm::NamedMDNode *MD = ME.get()->getNamedMetadata("omp_offload.info"); | ||||||
3289 | if (!MD) | ||||||
3290 | return; | ||||||
3291 | |||||||
3292 | for (llvm::MDNode *MN : MD->operands()) { | ||||||
3293 | auto &&GetMDInt = [MN](unsigned Idx) { | ||||||
3294 | auto *V = cast<llvm::ConstantAsMetadata>(MN->getOperand(Idx)); | ||||||
3295 | return cast<llvm::ConstantInt>(V->getValue())->getZExtValue(); | ||||||
3296 | }; | ||||||
3297 | |||||||
3298 | auto &&GetMDString = [MN](unsigned Idx) { | ||||||
3299 | auto *V = cast<llvm::MDString>(MN->getOperand(Idx)); | ||||||
3300 | return V->getString(); | ||||||
3301 | }; | ||||||
3302 | |||||||
3303 | switch (GetMDInt(0)) { | ||||||
3304 | default: | ||||||
3305 | llvm_unreachable("Unexpected metadata!")::llvm::llvm_unreachable_internal("Unexpected metadata!", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3305); | ||||||
3306 | break; | ||||||
3307 | case OffloadEntriesInfoManagerTy::OffloadEntryInfo:: | ||||||
3308 | OffloadingEntryInfoTargetRegion: | ||||||
3309 | OffloadEntriesInfoManager.initializeTargetRegionEntryInfo( | ||||||
3310 | /*DeviceID=*/GetMDInt(1), /*FileID=*/GetMDInt(2), | ||||||
3311 | /*ParentName=*/GetMDString(3), /*Line=*/GetMDInt(4), | ||||||
3312 | /*Order=*/GetMDInt(5)); | ||||||
3313 | break; | ||||||
3314 | case OffloadEntriesInfoManagerTy::OffloadEntryInfo:: | ||||||
3315 | OffloadingEntryInfoDeviceGlobalVar: | ||||||
3316 | OffloadEntriesInfoManager.initializeDeviceGlobalVarEntryInfo( | ||||||
3317 | /*MangledName=*/GetMDString(1), | ||||||
3318 | static_cast<OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind>( | ||||||
3319 | /*Flags=*/GetMDInt(2)), | ||||||
3320 | /*Order=*/GetMDInt(3)); | ||||||
3321 | break; | ||||||
3322 | } | ||||||
3323 | } | ||||||
3324 | } | ||||||
3325 | |||||||
3326 | void CGOpenMPRuntime::emitKmpRoutineEntryT(QualType KmpInt32Ty) { | ||||||
3327 | if (!KmpRoutineEntryPtrTy) { | ||||||
3328 | // Build typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); type. | ||||||
3329 | ASTContext &C = CGM.getContext(); | ||||||
3330 | QualType KmpRoutineEntryTyArgs[] = {KmpInt32Ty, C.VoidPtrTy}; | ||||||
3331 | FunctionProtoType::ExtProtoInfo EPI; | ||||||
3332 | KmpRoutineEntryPtrQTy = C.getPointerType( | ||||||
3333 | C.getFunctionType(KmpInt32Ty, KmpRoutineEntryTyArgs, EPI)); | ||||||
3334 | KmpRoutineEntryPtrTy = CGM.getTypes().ConvertType(KmpRoutineEntryPtrQTy); | ||||||
3335 | } | ||||||
3336 | } | ||||||
3337 | |||||||
3338 | QualType CGOpenMPRuntime::getTgtOffloadEntryQTy() { | ||||||
3339 | // Make sure the type of the entry is already created. This is the type we | ||||||
3340 | // have to create: | ||||||
3341 | // struct __tgt_offload_entry{ | ||||||
3342 | // void *addr; // Pointer to the offload entry info. | ||||||
3343 | // // (function or global) | ||||||
3344 | // char *name; // Name of the function or global. | ||||||
3345 | // size_t size; // Size of the entry info (0 if it a function). | ||||||
3346 | // int32_t flags; // Flags associated with the entry, e.g. 'link'. | ||||||
3347 | // int32_t reserved; // Reserved, to use by the runtime library. | ||||||
3348 | // }; | ||||||
3349 | if (TgtOffloadEntryQTy.isNull()) { | ||||||
3350 | ASTContext &C = CGM.getContext(); | ||||||
3351 | RecordDecl *RD = C.buildImplicitRecord("__tgt_offload_entry"); | ||||||
3352 | RD->startDefinition(); | ||||||
3353 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
3354 | addFieldToRecordDecl(C, RD, C.getPointerType(C.CharTy)); | ||||||
3355 | addFieldToRecordDecl(C, RD, C.getSizeType()); | ||||||
3356 | addFieldToRecordDecl( | ||||||
3357 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true)); | ||||||
3358 | addFieldToRecordDecl( | ||||||
3359 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true)); | ||||||
3360 | RD->completeDefinition(); | ||||||
3361 | RD->addAttr(PackedAttr::CreateImplicit(C)); | ||||||
3362 | TgtOffloadEntryQTy = C.getRecordType(RD); | ||||||
3363 | } | ||||||
3364 | return TgtOffloadEntryQTy; | ||||||
3365 | } | ||||||
3366 | |||||||
3367 | namespace { | ||||||
3368 | struct PrivateHelpersTy { | ||||||
3369 | PrivateHelpersTy(const Expr *OriginalRef, const VarDecl *Original, | ||||||
3370 | const VarDecl *PrivateCopy, const VarDecl *PrivateElemInit) | ||||||
3371 | : OriginalRef(OriginalRef), Original(Original), PrivateCopy(PrivateCopy), | ||||||
3372 | PrivateElemInit(PrivateElemInit) {} | ||||||
3373 | PrivateHelpersTy(const VarDecl *Original) : Original(Original) {} | ||||||
3374 | const Expr *OriginalRef = nullptr; | ||||||
3375 | const VarDecl *Original = nullptr; | ||||||
3376 | const VarDecl *PrivateCopy = nullptr; | ||||||
3377 | const VarDecl *PrivateElemInit = nullptr; | ||||||
3378 | bool isLocalPrivate() const { | ||||||
3379 | return !OriginalRef && !PrivateCopy && !PrivateElemInit; | ||||||
3380 | } | ||||||
3381 | }; | ||||||
3382 | typedef std::pair<CharUnits /*Align*/, PrivateHelpersTy> PrivateDataTy; | ||||||
3383 | } // anonymous namespace | ||||||
3384 | |||||||
3385 | static RecordDecl * | ||||||
3386 | createPrivatesRecordDecl(CodeGenModule &CGM, ArrayRef<PrivateDataTy> Privates) { | ||||||
3387 | if (!Privates.empty()) { | ||||||
3388 | ASTContext &C = CGM.getContext(); | ||||||
3389 | // Build struct .kmp_privates_t. { | ||||||
3390 | // /* private vars */ | ||||||
3391 | // }; | ||||||
3392 | RecordDecl *RD = C.buildImplicitRecord(".kmp_privates.t"); | ||||||
3393 | RD->startDefinition(); | ||||||
3394 | for (const auto &Pair : Privates) { | ||||||
3395 | const VarDecl *VD = Pair.second.Original; | ||||||
3396 | QualType Type = VD->getType().getNonReferenceType(); | ||||||
3397 | // If the private variable is a local variable with lvalue ref type, | ||||||
3398 | // allocate the pointer instead of the pointee type. | ||||||
3399 | if (Pair.second.isLocalPrivate() && | ||||||
3400 | VD->getType()->isLValueReferenceType()) | ||||||
3401 | Type = C.getPointerType(Type); | ||||||
3402 | FieldDecl *FD = addFieldToRecordDecl(C, RD, Type); | ||||||
3403 | if (VD->hasAttrs()) { | ||||||
3404 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), | ||||||
3405 | E(VD->getAttrs().end()); | ||||||
3406 | I != E; ++I) | ||||||
3407 | FD->addAttr(*I); | ||||||
3408 | } | ||||||
3409 | } | ||||||
3410 | RD->completeDefinition(); | ||||||
3411 | return RD; | ||||||
3412 | } | ||||||
3413 | return nullptr; | ||||||
3414 | } | ||||||
3415 | |||||||
3416 | static RecordDecl * | ||||||
3417 | createKmpTaskTRecordDecl(CodeGenModule &CGM, OpenMPDirectiveKind Kind, | ||||||
3418 | QualType KmpInt32Ty, | ||||||
3419 | QualType KmpRoutineEntryPointerQTy) { | ||||||
3420 | ASTContext &C = CGM.getContext(); | ||||||
3421 | // Build struct kmp_task_t { | ||||||
3422 | // void * shareds; | ||||||
3423 | // kmp_routine_entry_t routine; | ||||||
3424 | // kmp_int32 part_id; | ||||||
3425 | // kmp_cmplrdata_t data1; | ||||||
3426 | // kmp_cmplrdata_t data2; | ||||||
3427 | // For taskloops additional fields: | ||||||
3428 | // kmp_uint64 lb; | ||||||
3429 | // kmp_uint64 ub; | ||||||
3430 | // kmp_int64 st; | ||||||
3431 | // kmp_int32 liter; | ||||||
3432 | // void * reductions; | ||||||
3433 | // }; | ||||||
3434 | RecordDecl *UD = C.buildImplicitRecord("kmp_cmplrdata_t", TTK_Union); | ||||||
3435 | UD->startDefinition(); | ||||||
3436 | addFieldToRecordDecl(C, UD, KmpInt32Ty); | ||||||
3437 | addFieldToRecordDecl(C, UD, KmpRoutineEntryPointerQTy); | ||||||
3438 | UD->completeDefinition(); | ||||||
3439 | QualType KmpCmplrdataTy = C.getRecordType(UD); | ||||||
3440 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_t"); | ||||||
3441 | RD->startDefinition(); | ||||||
3442 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
3443 | addFieldToRecordDecl(C, RD, KmpRoutineEntryPointerQTy); | ||||||
3444 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | ||||||
3445 | addFieldToRecordDecl(C, RD, KmpCmplrdataTy); | ||||||
3446 | addFieldToRecordDecl(C, RD, KmpCmplrdataTy); | ||||||
3447 | if (isOpenMPTaskLoopDirective(Kind)) { | ||||||
3448 | QualType KmpUInt64Ty = | ||||||
3449 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0); | ||||||
3450 | QualType KmpInt64Ty = | ||||||
3451 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | ||||||
3452 | addFieldToRecordDecl(C, RD, KmpUInt64Ty); | ||||||
3453 | addFieldToRecordDecl(C, RD, KmpUInt64Ty); | ||||||
3454 | addFieldToRecordDecl(C, RD, KmpInt64Ty); | ||||||
3455 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | ||||||
3456 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
3457 | } | ||||||
3458 | RD->completeDefinition(); | ||||||
3459 | return RD; | ||||||
3460 | } | ||||||
3461 | |||||||
3462 | static RecordDecl * | ||||||
3463 | createKmpTaskTWithPrivatesRecordDecl(CodeGenModule &CGM, QualType KmpTaskTQTy, | ||||||
3464 | ArrayRef<PrivateDataTy> Privates) { | ||||||
3465 | ASTContext &C = CGM.getContext(); | ||||||
3466 | // Build struct kmp_task_t_with_privates { | ||||||
3467 | // kmp_task_t task_data; | ||||||
3468 | // .kmp_privates_t. privates; | ||||||
3469 | // }; | ||||||
3470 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_t_with_privates"); | ||||||
3471 | RD->startDefinition(); | ||||||
3472 | addFieldToRecordDecl(C, RD, KmpTaskTQTy); | ||||||
3473 | if (const RecordDecl *PrivateRD = createPrivatesRecordDecl(CGM, Privates)) | ||||||
3474 | addFieldToRecordDecl(C, RD, C.getRecordType(PrivateRD)); | ||||||
3475 | RD->completeDefinition(); | ||||||
3476 | return RD; | ||||||
3477 | } | ||||||
3478 | |||||||
3479 | /// Emit a proxy function which accepts kmp_task_t as the second | ||||||
3480 | /// argument. | ||||||
3481 | /// \code | ||||||
3482 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { | ||||||
3483 | /// TaskFunction(gtid, tt->part_id, &tt->privates, task_privates_map, tt, | ||||||
3484 | /// For taskloops: | ||||||
3485 | /// tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter, | ||||||
3486 | /// tt->reductions, tt->shareds); | ||||||
3487 | /// return 0; | ||||||
3488 | /// } | ||||||
3489 | /// \endcode | ||||||
3490 | static llvm::Function * | ||||||
3491 | emitProxyTaskFunction(CodeGenModule &CGM, SourceLocation Loc, | ||||||
3492 | OpenMPDirectiveKind Kind, QualType KmpInt32Ty, | ||||||
3493 | QualType KmpTaskTWithPrivatesPtrQTy, | ||||||
3494 | QualType KmpTaskTWithPrivatesQTy, QualType KmpTaskTQTy, | ||||||
3495 | QualType SharedsPtrTy, llvm::Function *TaskFunction, | ||||||
3496 | llvm::Value *TaskPrivatesMap) { | ||||||
3497 | ASTContext &C = CGM.getContext(); | ||||||
3498 | FunctionArgList Args; | ||||||
3499 | ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty, | ||||||
3500 | ImplicitParamDecl::Other); | ||||||
3501 | ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3502 | KmpTaskTWithPrivatesPtrQTy.withRestrict(), | ||||||
3503 | ImplicitParamDecl::Other); | ||||||
3504 | Args.push_back(&GtidArg); | ||||||
3505 | Args.push_back(&TaskTypeArg); | ||||||
3506 | const auto &TaskEntryFnInfo = | ||||||
3507 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args); | ||||||
3508 | llvm::FunctionType *TaskEntryTy = | ||||||
3509 | CGM.getTypes().GetFunctionType(TaskEntryFnInfo); | ||||||
3510 | std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_entry", ""}); | ||||||
3511 | auto *TaskEntry = llvm::Function::Create( | ||||||
3512 | TaskEntryTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule()); | ||||||
3513 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskEntry, TaskEntryFnInfo); | ||||||
3514 | TaskEntry->setDoesNotRecurse(); | ||||||
3515 | CodeGenFunction CGF(CGM); | ||||||
3516 | CGF.StartFunction(GlobalDecl(), KmpInt32Ty, TaskEntry, TaskEntryFnInfo, Args, | ||||||
3517 | Loc, Loc); | ||||||
3518 | |||||||
3519 | // TaskFunction(gtid, tt->task_data.part_id, &tt->privates, task_privates_map, | ||||||
3520 | // tt, | ||||||
3521 | // For taskloops: | ||||||
3522 | // tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter, | ||||||
3523 | // tt->task_data.shareds); | ||||||
3524 | llvm::Value *GtidParam = CGF.EmitLoadOfScalar( | ||||||
3525 | CGF.GetAddrOfLocalVar(&GtidArg), /*Volatile=*/false, KmpInt32Ty, Loc); | ||||||
3526 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | ||||||
3527 | CGF.GetAddrOfLocalVar(&TaskTypeArg), | ||||||
3528 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | ||||||
3529 | const auto *KmpTaskTWithPrivatesQTyRD = | ||||||
3530 | cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl()); | ||||||
3531 | LValue Base = | ||||||
3532 | CGF.EmitLValueForField(TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
3533 | const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl()); | ||||||
3534 | auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId); | ||||||
3535 | LValue PartIdLVal = CGF.EmitLValueForField(Base, *PartIdFI); | ||||||
3536 | llvm::Value *PartidParam = PartIdLVal.getPointer(CGF); | ||||||
3537 | |||||||
3538 | auto SharedsFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds); | ||||||
3539 | LValue SharedsLVal = CGF.EmitLValueForField(Base, *SharedsFI); | ||||||
3540 | llvm::Value *SharedsParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
3541 | CGF.EmitLoadOfScalar(SharedsLVal, Loc), | ||||||
3542 | CGF.ConvertTypeForMem(SharedsPtrTy)); | ||||||
3543 | |||||||
3544 | auto PrivatesFI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1); | ||||||
3545 | llvm::Value *PrivatesParam; | ||||||
3546 | if (PrivatesFI != KmpTaskTWithPrivatesQTyRD->field_end()) { | ||||||
3547 | LValue PrivatesLVal = CGF.EmitLValueForField(TDBase, *PrivatesFI); | ||||||
3548 | PrivatesParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
3549 | PrivatesLVal.getPointer(CGF), CGF.VoidPtrTy); | ||||||
3550 | } else { | ||||||
3551 | PrivatesParam = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | ||||||
3552 | } | ||||||
3553 | |||||||
3554 | llvm::Value *CommonArgs[] = {GtidParam, PartidParam, PrivatesParam, | ||||||
3555 | TaskPrivatesMap, | ||||||
3556 | CGF.Builder | ||||||
3557 | .CreatePointerBitCastOrAddrSpaceCast( | ||||||
3558 | TDBase.getAddress(CGF), CGF.VoidPtrTy) | ||||||
3559 | .getPointer()}; | ||||||
3560 | SmallVector<llvm::Value *, 16> CallArgs(std::begin(CommonArgs), | ||||||
3561 | std::end(CommonArgs)); | ||||||
3562 | if (isOpenMPTaskLoopDirective(Kind)) { | ||||||
3563 | auto LBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound); | ||||||
3564 | LValue LBLVal = CGF.EmitLValueForField(Base, *LBFI); | ||||||
3565 | llvm::Value *LBParam = CGF.EmitLoadOfScalar(LBLVal, Loc); | ||||||
3566 | auto UBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound); | ||||||
3567 | LValue UBLVal = CGF.EmitLValueForField(Base, *UBFI); | ||||||
3568 | llvm::Value *UBParam = CGF.EmitLoadOfScalar(UBLVal, Loc); | ||||||
3569 | auto StFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTStride); | ||||||
3570 | LValue StLVal = CGF.EmitLValueForField(Base, *StFI); | ||||||
3571 | llvm::Value *StParam = CGF.EmitLoadOfScalar(StLVal, Loc); | ||||||
3572 | auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter); | ||||||
3573 | LValue LILVal = CGF.EmitLValueForField(Base, *LIFI); | ||||||
3574 | llvm::Value *LIParam = CGF.EmitLoadOfScalar(LILVal, Loc); | ||||||
3575 | auto RFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTReductions); | ||||||
3576 | LValue RLVal = CGF.EmitLValueForField(Base, *RFI); | ||||||
3577 | llvm::Value *RParam = CGF.EmitLoadOfScalar(RLVal, Loc); | ||||||
3578 | CallArgs.push_back(LBParam); | ||||||
3579 | CallArgs.push_back(UBParam); | ||||||
3580 | CallArgs.push_back(StParam); | ||||||
3581 | CallArgs.push_back(LIParam); | ||||||
3582 | CallArgs.push_back(RParam); | ||||||
3583 | } | ||||||
3584 | CallArgs.push_back(SharedsParam); | ||||||
3585 | |||||||
3586 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskFunction, | ||||||
3587 | CallArgs); | ||||||
3588 | CGF.EmitStoreThroughLValue(RValue::get(CGF.Builder.getInt32(/*C=*/0)), | ||||||
3589 | CGF.MakeAddrLValue(CGF.ReturnValue, KmpInt32Ty)); | ||||||
3590 | CGF.FinishFunction(); | ||||||
3591 | return TaskEntry; | ||||||
3592 | } | ||||||
3593 | |||||||
3594 | static llvm::Value *emitDestructorsFunction(CodeGenModule &CGM, | ||||||
3595 | SourceLocation Loc, | ||||||
3596 | QualType KmpInt32Ty, | ||||||
3597 | QualType KmpTaskTWithPrivatesPtrQTy, | ||||||
3598 | QualType KmpTaskTWithPrivatesQTy) { | ||||||
3599 | ASTContext &C = CGM.getContext(); | ||||||
3600 | FunctionArgList Args; | ||||||
3601 | ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty, | ||||||
3602 | ImplicitParamDecl::Other); | ||||||
3603 | ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3604 | KmpTaskTWithPrivatesPtrQTy.withRestrict(), | ||||||
3605 | ImplicitParamDecl::Other); | ||||||
3606 | Args.push_back(&GtidArg); | ||||||
3607 | Args.push_back(&TaskTypeArg); | ||||||
3608 | const auto &DestructorFnInfo = | ||||||
3609 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args); | ||||||
3610 | llvm::FunctionType *DestructorFnTy = | ||||||
3611 | CGM.getTypes().GetFunctionType(DestructorFnInfo); | ||||||
3612 | std::string Name = | ||||||
3613 | CGM.getOpenMPRuntime().getName({"omp_task_destructor", ""}); | ||||||
3614 | auto *DestructorFn = | ||||||
3615 | llvm::Function::Create(DestructorFnTy, llvm::GlobalValue::InternalLinkage, | ||||||
3616 | Name, &CGM.getModule()); | ||||||
3617 | CGM.SetInternalFunctionAttributes(GlobalDecl(), DestructorFn, | ||||||
3618 | DestructorFnInfo); | ||||||
3619 | DestructorFn->setDoesNotRecurse(); | ||||||
3620 | CodeGenFunction CGF(CGM); | ||||||
3621 | CGF.StartFunction(GlobalDecl(), KmpInt32Ty, DestructorFn, DestructorFnInfo, | ||||||
3622 | Args, Loc, Loc); | ||||||
3623 | |||||||
3624 | LValue Base = CGF.EmitLoadOfPointerLValue( | ||||||
3625 | CGF.GetAddrOfLocalVar(&TaskTypeArg), | ||||||
3626 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | ||||||
3627 | const auto *KmpTaskTWithPrivatesQTyRD = | ||||||
3628 | cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl()); | ||||||
3629 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
3630 | Base = CGF.EmitLValueForField(Base, *FI); | ||||||
3631 | for (const auto *Field : | ||||||
3632 | cast<RecordDecl>(FI->getType()->getAsTagDecl())->fields()) { | ||||||
3633 | if (QualType::DestructionKind DtorKind = | ||||||
3634 | Field->getType().isDestructedType()) { | ||||||
3635 | LValue FieldLValue = CGF.EmitLValueForField(Base, Field); | ||||||
3636 | CGF.pushDestroy(DtorKind, FieldLValue.getAddress(CGF), Field->getType()); | ||||||
3637 | } | ||||||
3638 | } | ||||||
3639 | CGF.FinishFunction(); | ||||||
3640 | return DestructorFn; | ||||||
3641 | } | ||||||
3642 | |||||||
3643 | /// Emit a privates mapping function for correct handling of private and | ||||||
3644 | /// firstprivate variables. | ||||||
3645 | /// \code | ||||||
3646 | /// void .omp_task_privates_map.(const .privates. *noalias privs, <ty1> | ||||||
3647 | /// **noalias priv1,..., <tyn> **noalias privn) { | ||||||
3648 | /// *priv1 = &.privates.priv1; | ||||||
3649 | /// ...; | ||||||
3650 | /// *privn = &.privates.privn; | ||||||
3651 | /// } | ||||||
3652 | /// \endcode | ||||||
3653 | static llvm::Value * | ||||||
3654 | emitTaskPrivateMappingFunction(CodeGenModule &CGM, SourceLocation Loc, | ||||||
3655 | const OMPTaskDataTy &Data, QualType PrivatesQTy, | ||||||
3656 | ArrayRef<PrivateDataTy> Privates) { | ||||||
3657 | ASTContext &C = CGM.getContext(); | ||||||
3658 | FunctionArgList Args; | ||||||
3659 | ImplicitParamDecl TaskPrivatesArg( | ||||||
3660 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3661 | C.getPointerType(PrivatesQTy).withConst().withRestrict(), | ||||||
3662 | ImplicitParamDecl::Other); | ||||||
3663 | Args.push_back(&TaskPrivatesArg); | ||||||
3664 | llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, unsigned> PrivateVarsPos; | ||||||
3665 | unsigned Counter = 1; | ||||||
3666 | for (const Expr *E : Data.PrivateVars) { | ||||||
3667 | Args.push_back(ImplicitParamDecl::Create( | ||||||
3668 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3669 | C.getPointerType(C.getPointerType(E->getType())) | ||||||
3670 | .withConst() | ||||||
3671 | .withRestrict(), | ||||||
3672 | ImplicitParamDecl::Other)); | ||||||
3673 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | ||||||
3674 | PrivateVarsPos[VD] = Counter; | ||||||
3675 | ++Counter; | ||||||
3676 | } | ||||||
3677 | for (const Expr *E : Data.FirstprivateVars) { | ||||||
3678 | Args.push_back(ImplicitParamDecl::Create( | ||||||
3679 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3680 | C.getPointerType(C.getPointerType(E->getType())) | ||||||
3681 | .withConst() | ||||||
3682 | .withRestrict(), | ||||||
3683 | ImplicitParamDecl::Other)); | ||||||
3684 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | ||||||
3685 | PrivateVarsPos[VD] = Counter; | ||||||
3686 | ++Counter; | ||||||
3687 | } | ||||||
3688 | for (const Expr *E : Data.LastprivateVars) { | ||||||
3689 | Args.push_back(ImplicitParamDecl::Create( | ||||||
3690 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3691 | C.getPointerType(C.getPointerType(E->getType())) | ||||||
3692 | .withConst() | ||||||
3693 | .withRestrict(), | ||||||
3694 | ImplicitParamDecl::Other)); | ||||||
3695 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | ||||||
3696 | PrivateVarsPos[VD] = Counter; | ||||||
3697 | ++Counter; | ||||||
3698 | } | ||||||
3699 | for (const VarDecl *VD : Data.PrivateLocals) { | ||||||
3700 | QualType Ty = VD->getType().getNonReferenceType(); | ||||||
3701 | if (VD->getType()->isLValueReferenceType()) | ||||||
3702 | Ty = C.getPointerType(Ty); | ||||||
3703 | Args.push_back(ImplicitParamDecl::Create( | ||||||
3704 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3705 | C.getPointerType(C.getPointerType(Ty)).withConst().withRestrict(), | ||||||
3706 | ImplicitParamDecl::Other)); | ||||||
3707 | PrivateVarsPos[VD] = Counter; | ||||||
3708 | ++Counter; | ||||||
3709 | } | ||||||
3710 | const auto &TaskPrivatesMapFnInfo = | ||||||
3711 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
3712 | llvm::FunctionType *TaskPrivatesMapTy = | ||||||
3713 | CGM.getTypes().GetFunctionType(TaskPrivatesMapFnInfo); | ||||||
3714 | std::string Name = | ||||||
3715 | CGM.getOpenMPRuntime().getName({"omp_task_privates_map", ""}); | ||||||
3716 | auto *TaskPrivatesMap = llvm::Function::Create( | ||||||
3717 | TaskPrivatesMapTy, llvm::GlobalValue::InternalLinkage, Name, | ||||||
3718 | &CGM.getModule()); | ||||||
3719 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskPrivatesMap, | ||||||
3720 | TaskPrivatesMapFnInfo); | ||||||
3721 | if (CGM.getLangOpts().Optimize) { | ||||||
3722 | TaskPrivatesMap->removeFnAttr(llvm::Attribute::NoInline); | ||||||
3723 | TaskPrivatesMap->removeFnAttr(llvm::Attribute::OptimizeNone); | ||||||
3724 | TaskPrivatesMap->addFnAttr(llvm::Attribute::AlwaysInline); | ||||||
3725 | } | ||||||
3726 | CodeGenFunction CGF(CGM); | ||||||
3727 | CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskPrivatesMap, | ||||||
3728 | TaskPrivatesMapFnInfo, Args, Loc, Loc); | ||||||
3729 | |||||||
3730 | // *privi = &.privates.privi; | ||||||
3731 | LValue Base = CGF.EmitLoadOfPointerLValue( | ||||||
3732 | CGF.GetAddrOfLocalVar(&TaskPrivatesArg), | ||||||
3733 | TaskPrivatesArg.getType()->castAs<PointerType>()); | ||||||
3734 | const auto *PrivatesQTyRD = cast<RecordDecl>(PrivatesQTy->getAsTagDecl()); | ||||||
3735 | Counter = 0; | ||||||
3736 | for (const FieldDecl *Field : PrivatesQTyRD->fields()) { | ||||||
3737 | LValue FieldLVal = CGF.EmitLValueForField(Base, Field); | ||||||
3738 | const VarDecl *VD = Args[PrivateVarsPos[Privates[Counter].second.Original]]; | ||||||
3739 | LValue RefLVal = | ||||||
3740 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); | ||||||
3741 | LValue RefLoadLVal = CGF.EmitLoadOfPointerLValue( | ||||||
3742 | RefLVal.getAddress(CGF), RefLVal.getType()->castAs<PointerType>()); | ||||||
3743 | CGF.EmitStoreOfScalar(FieldLVal.getPointer(CGF), RefLoadLVal); | ||||||
3744 | ++Counter; | ||||||
3745 | } | ||||||
3746 | CGF.FinishFunction(); | ||||||
3747 | return TaskPrivatesMap; | ||||||
3748 | } | ||||||
3749 | |||||||
3750 | /// Emit initialization for private variables in task-based directives. | ||||||
3751 | static void emitPrivatesInit(CodeGenFunction &CGF, | ||||||
3752 | const OMPExecutableDirective &D, | ||||||
3753 | Address KmpTaskSharedsPtr, LValue TDBase, | ||||||
3754 | const RecordDecl *KmpTaskTWithPrivatesQTyRD, | ||||||
3755 | QualType SharedsTy, QualType SharedsPtrTy, | ||||||
3756 | const OMPTaskDataTy &Data, | ||||||
3757 | ArrayRef<PrivateDataTy> Privates, bool ForDup) { | ||||||
3758 | ASTContext &C = CGF.getContext(); | ||||||
3759 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
3760 | LValue PrivatesBase = CGF.EmitLValueForField(TDBase, *FI); | ||||||
3761 | OpenMPDirectiveKind Kind = isOpenMPTaskLoopDirective(D.getDirectiveKind()) | ||||||
3762 | ? OMPD_taskloop | ||||||
3763 | : OMPD_task; | ||||||
3764 | const CapturedStmt &CS = *D.getCapturedStmt(Kind); | ||||||
3765 | CodeGenFunction::CGCapturedStmtInfo CapturesInfo(CS); | ||||||
3766 | LValue SrcBase; | ||||||
3767 | bool IsTargetTask = | ||||||
3768 | isOpenMPTargetDataManagementDirective(D.getDirectiveKind()) || | ||||||
3769 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()); | ||||||
3770 | // For target-based directives skip 3 firstprivate arrays BasePointersArray, | ||||||
3771 | // PointersArray and SizesArray. The original variables for these arrays are | ||||||
3772 | // not captured and we get their addresses explicitly. | ||||||
3773 | if ((!IsTargetTask && !Data.FirstprivateVars.empty() && ForDup) || | ||||||
3774 | (IsTargetTask && KmpTaskSharedsPtr.isValid())) { | ||||||
3775 | SrcBase = CGF.MakeAddrLValue( | ||||||
3776 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
3777 | KmpTaskSharedsPtr, CGF.ConvertTypeForMem(SharedsPtrTy)), | ||||||
3778 | SharedsTy); | ||||||
3779 | } | ||||||
3780 | FI = cast<RecordDecl>(FI->getType()->getAsTagDecl())->field_begin(); | ||||||
3781 | for (const PrivateDataTy &Pair : Privates) { | ||||||
3782 | // Do not initialize private locals. | ||||||
3783 | if (Pair.second.isLocalPrivate()) | ||||||
3784 | continue; | ||||||
3785 | const VarDecl *VD = Pair.second.PrivateCopy; | ||||||
3786 | const Expr *Init = VD->getAnyInitializer(); | ||||||
3787 | if (Init && (!ForDup || (isa<CXXConstructExpr>(Init) && | ||||||
3788 | !CGF.isTrivialInitializer(Init)))) { | ||||||
3789 | LValue PrivateLValue = CGF.EmitLValueForField(PrivatesBase, *FI); | ||||||
3790 | if (const VarDecl *Elem = Pair.second.PrivateElemInit) { | ||||||
3791 | const VarDecl *OriginalVD = Pair.second.Original; | ||||||
3792 | // Check if the variable is the target-based BasePointersArray, | ||||||
3793 | // PointersArray or SizesArray. | ||||||
3794 | LValue SharedRefLValue; | ||||||
3795 | QualType Type = PrivateLValue.getType(); | ||||||
3796 | const FieldDecl *SharedField = CapturesInfo.lookup(OriginalVD); | ||||||
3797 | if (IsTargetTask && !SharedField) { | ||||||
3798 | assert(isa<ImplicitParamDecl>(OriginalVD) &&((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3799 | isa<CapturedDecl>(OriginalVD->getDeclContext()) &&((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3800 | cast<CapturedDecl>(OriginalVD->getDeclContext())((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3801 | ->getNumParams() == 0 &&((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3802 | isa<TranslationUnitDecl>(((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3803 | cast<CapturedDecl>(OriginalVD->getDeclContext())((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3804 | ->getDeclContext()) &&((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)) | ||||||
3805 | "Expected artificial target data variable.")((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.") ? static_cast< 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.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3805, __PRETTY_FUNCTION__)); | ||||||
3806 | SharedRefLValue = | ||||||
3807 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(OriginalVD), Type); | ||||||
3808 | } else if (ForDup) { | ||||||
3809 | SharedRefLValue = CGF.EmitLValueForField(SrcBase, SharedField); | ||||||
3810 | SharedRefLValue = CGF.MakeAddrLValue( | ||||||
3811 | Address(SharedRefLValue.getPointer(CGF), | ||||||
3812 | C.getDeclAlign(OriginalVD)), | ||||||
3813 | SharedRefLValue.getType(), LValueBaseInfo(AlignmentSource::Decl), | ||||||
3814 | SharedRefLValue.getTBAAInfo()); | ||||||
3815 | } else if (CGF.LambdaCaptureFields.count( | ||||||
3816 | Pair.second.Original->getCanonicalDecl()) > 0 || | ||||||
3817 | dyn_cast_or_null<BlockDecl>(CGF.CurCodeDecl)) { | ||||||
3818 | SharedRefLValue = CGF.EmitLValue(Pair.second.OriginalRef); | ||||||
3819 | } else { | ||||||
3820 | // Processing for implicitly captured variables. | ||||||
3821 | InlinedOpenMPRegionRAII Region( | ||||||
3822 | CGF, [](CodeGenFunction &, PrePostActionTy &) {}, OMPD_unknown, | ||||||
3823 | /*HasCancel=*/false); | ||||||
3824 | SharedRefLValue = CGF.EmitLValue(Pair.second.OriginalRef); | ||||||
3825 | } | ||||||
3826 | if (Type->isArrayType()) { | ||||||
3827 | // Initialize firstprivate array. | ||||||
3828 | if (!isa<CXXConstructExpr>(Init) || CGF.isTrivialInitializer(Init)) { | ||||||
3829 | // Perform simple memcpy. | ||||||
3830 | CGF.EmitAggregateAssign(PrivateLValue, SharedRefLValue, Type); | ||||||
3831 | } else { | ||||||
3832 | // Initialize firstprivate array using element-by-element | ||||||
3833 | // initialization. | ||||||
3834 | CGF.EmitOMPAggregateAssign( | ||||||
3835 | PrivateLValue.getAddress(CGF), SharedRefLValue.getAddress(CGF), | ||||||
3836 | Type, | ||||||
3837 | [&CGF, Elem, Init, &CapturesInfo](Address DestElement, | ||||||
3838 | Address SrcElement) { | ||||||
3839 | // Clean up any temporaries needed by the initialization. | ||||||
3840 | CodeGenFunction::OMPPrivateScope InitScope(CGF); | ||||||
3841 | InitScope.addPrivate( | ||||||
3842 | Elem, [SrcElement]() -> Address { return SrcElement; }); | ||||||
3843 | (void)InitScope.Privatize(); | ||||||
3844 | // Emit initialization for single element. | ||||||
3845 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII( | ||||||
3846 | CGF, &CapturesInfo); | ||||||
3847 | CGF.EmitAnyExprToMem(Init, DestElement, | ||||||
3848 | Init->getType().getQualifiers(), | ||||||
3849 | /*IsInitializer=*/false); | ||||||
3850 | }); | ||||||
3851 | } | ||||||
3852 | } else { | ||||||
3853 | CodeGenFunction::OMPPrivateScope InitScope(CGF); | ||||||
3854 | InitScope.addPrivate(Elem, [SharedRefLValue, &CGF]() -> Address { | ||||||
3855 | return SharedRefLValue.getAddress(CGF); | ||||||
3856 | }); | ||||||
3857 | (void)InitScope.Privatize(); | ||||||
3858 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CapturesInfo); | ||||||
3859 | CGF.EmitExprAsInit(Init, VD, PrivateLValue, | ||||||
3860 | /*capturedByInit=*/false); | ||||||
3861 | } | ||||||
3862 | } else { | ||||||
3863 | CGF.EmitExprAsInit(Init, VD, PrivateLValue, /*capturedByInit=*/false); | ||||||
3864 | } | ||||||
3865 | } | ||||||
3866 | ++FI; | ||||||
3867 | } | ||||||
3868 | } | ||||||
3869 | |||||||
3870 | /// Check if duplication function is required for taskloops. | ||||||
3871 | static bool checkInitIsRequired(CodeGenFunction &CGF, | ||||||
3872 | ArrayRef<PrivateDataTy> Privates) { | ||||||
3873 | bool InitRequired = false; | ||||||
3874 | for (const PrivateDataTy &Pair : Privates) { | ||||||
3875 | if (Pair.second.isLocalPrivate()) | ||||||
3876 | continue; | ||||||
3877 | const VarDecl *VD = Pair.second.PrivateCopy; | ||||||
3878 | const Expr *Init = VD->getAnyInitializer(); | ||||||
3879 | InitRequired = InitRequired || (Init && isa<CXXConstructExpr>(Init) && | ||||||
3880 | !CGF.isTrivialInitializer(Init)); | ||||||
3881 | if (InitRequired) | ||||||
3882 | break; | ||||||
3883 | } | ||||||
3884 | return InitRequired; | ||||||
3885 | } | ||||||
3886 | |||||||
3887 | |||||||
3888 | /// Emit task_dup function (for initialization of | ||||||
3889 | /// private/firstprivate/lastprivate vars and last_iter flag) | ||||||
3890 | /// \code | ||||||
3891 | /// void __task_dup_entry(kmp_task_t *task_dst, const kmp_task_t *task_src, int | ||||||
3892 | /// lastpriv) { | ||||||
3893 | /// // setup lastprivate flag | ||||||
3894 | /// task_dst->last = lastpriv; | ||||||
3895 | /// // could be constructor calls here... | ||||||
3896 | /// } | ||||||
3897 | /// \endcode | ||||||
3898 | static llvm::Value * | ||||||
3899 | emitTaskDupFunction(CodeGenModule &CGM, SourceLocation Loc, | ||||||
3900 | const OMPExecutableDirective &D, | ||||||
3901 | QualType KmpTaskTWithPrivatesPtrQTy, | ||||||
3902 | const RecordDecl *KmpTaskTWithPrivatesQTyRD, | ||||||
3903 | const RecordDecl *KmpTaskTQTyRD, QualType SharedsTy, | ||||||
3904 | QualType SharedsPtrTy, const OMPTaskDataTy &Data, | ||||||
3905 | ArrayRef<PrivateDataTy> Privates, bool WithLastIter) { | ||||||
3906 | ASTContext &C = CGM.getContext(); | ||||||
3907 | FunctionArgList Args; | ||||||
3908 | ImplicitParamDecl DstArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3909 | KmpTaskTWithPrivatesPtrQTy, | ||||||
3910 | ImplicitParamDecl::Other); | ||||||
3911 | ImplicitParamDecl SrcArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
3912 | KmpTaskTWithPrivatesPtrQTy, | ||||||
3913 | ImplicitParamDecl::Other); | ||||||
3914 | ImplicitParamDecl LastprivArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | ||||||
3915 | ImplicitParamDecl::Other); | ||||||
3916 | Args.push_back(&DstArg); | ||||||
3917 | Args.push_back(&SrcArg); | ||||||
3918 | Args.push_back(&LastprivArg); | ||||||
3919 | const auto &TaskDupFnInfo = | ||||||
3920 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
3921 | llvm::FunctionType *TaskDupTy = CGM.getTypes().GetFunctionType(TaskDupFnInfo); | ||||||
3922 | std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_dup", ""}); | ||||||
3923 | auto *TaskDup = llvm::Function::Create( | ||||||
3924 | TaskDupTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule()); | ||||||
3925 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskDup, TaskDupFnInfo); | ||||||
3926 | TaskDup->setDoesNotRecurse(); | ||||||
3927 | CodeGenFunction CGF(CGM); | ||||||
3928 | CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskDup, TaskDupFnInfo, Args, Loc, | ||||||
3929 | Loc); | ||||||
3930 | |||||||
3931 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | ||||||
3932 | CGF.GetAddrOfLocalVar(&DstArg), | ||||||
3933 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | ||||||
3934 | // task_dst->liter = lastpriv; | ||||||
3935 | if (WithLastIter) { | ||||||
3936 | auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter); | ||||||
3937 | LValue Base = CGF.EmitLValueForField( | ||||||
3938 | TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
3939 | LValue LILVal = CGF.EmitLValueForField(Base, *LIFI); | ||||||
3940 | llvm::Value *Lastpriv = CGF.EmitLoadOfScalar( | ||||||
3941 | CGF.GetAddrOfLocalVar(&LastprivArg), /*Volatile=*/false, C.IntTy, Loc); | ||||||
3942 | CGF.EmitStoreOfScalar(Lastpriv, LILVal); | ||||||
3943 | } | ||||||
3944 | |||||||
3945 | // Emit initial values for private copies (if any). | ||||||
3946 | assert(!Privates.empty())((!Privates.empty()) ? static_cast<void> (0) : __assert_fail ("!Privates.empty()", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3946, __PRETTY_FUNCTION__)); | ||||||
3947 | Address KmpTaskSharedsPtr = Address::invalid(); | ||||||
3948 | if (!Data.FirstprivateVars.empty()) { | ||||||
3949 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | ||||||
3950 | CGF.GetAddrOfLocalVar(&SrcArg), | ||||||
3951 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | ||||||
3952 | LValue Base = CGF.EmitLValueForField( | ||||||
3953 | TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
3954 | KmpTaskSharedsPtr = Address( | ||||||
3955 | CGF.EmitLoadOfScalar(CGF.EmitLValueForField( | ||||||
3956 | Base, *std::next(KmpTaskTQTyRD->field_begin(), | ||||||
3957 | KmpTaskTShareds)), | ||||||
3958 | Loc), | ||||||
3959 | CGM.getNaturalTypeAlignment(SharedsTy)); | ||||||
3960 | } | ||||||
3961 | emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, TDBase, KmpTaskTWithPrivatesQTyRD, | ||||||
3962 | SharedsTy, SharedsPtrTy, Data, Privates, /*ForDup=*/true); | ||||||
3963 | CGF.FinishFunction(); | ||||||
3964 | return TaskDup; | ||||||
3965 | } | ||||||
3966 | |||||||
3967 | /// Checks if destructor function is required to be generated. | ||||||
3968 | /// \return true if cleanups are required, false otherwise. | ||||||
3969 | static bool | ||||||
3970 | checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD, | ||||||
3971 | ArrayRef<PrivateDataTy> Privates) { | ||||||
3972 | for (const PrivateDataTy &P : Privates) { | ||||||
3973 | if (P.second.isLocalPrivate()) | ||||||
3974 | continue; | ||||||
3975 | QualType Ty = P.second.Original->getType().getNonReferenceType(); | ||||||
3976 | if (Ty.isDestructedType()) | ||||||
3977 | return true; | ||||||
3978 | } | ||||||
3979 | return false; | ||||||
3980 | } | ||||||
3981 | |||||||
3982 | namespace { | ||||||
3983 | /// Loop generator for OpenMP iterator expression. | ||||||
3984 | class OMPIteratorGeneratorScope final | ||||||
3985 | : public CodeGenFunction::OMPPrivateScope { | ||||||
3986 | CodeGenFunction &CGF; | ||||||
3987 | const OMPIteratorExpr *E = nullptr; | ||||||
3988 | SmallVector<CodeGenFunction::JumpDest, 4> ContDests; | ||||||
3989 | SmallVector<CodeGenFunction::JumpDest, 4> ExitDests; | ||||||
3990 | OMPIteratorGeneratorScope() = delete; | ||||||
3991 | OMPIteratorGeneratorScope(OMPIteratorGeneratorScope &) = delete; | ||||||
3992 | |||||||
3993 | public: | ||||||
3994 | OMPIteratorGeneratorScope(CodeGenFunction &CGF, const OMPIteratorExpr *E) | ||||||
3995 | : CodeGenFunction::OMPPrivateScope(CGF), CGF(CGF), E(E) { | ||||||
3996 | if (!E) | ||||||
3997 | return; | ||||||
3998 | SmallVector<llvm::Value *, 4> Uppers; | ||||||
3999 | for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) { | ||||||
4000 | Uppers.push_back(CGF.EmitScalarExpr(E->getHelper(I).Upper)); | ||||||
4001 | const auto *VD = cast<VarDecl>(E->getIteratorDecl(I)); | ||||||
4002 | addPrivate(VD, [&CGF, VD]() { | ||||||
4003 | return CGF.CreateMemTemp(VD->getType(), VD->getName()); | ||||||
4004 | }); | ||||||
4005 | const OMPIteratorHelperData &HelperData = E->getHelper(I); | ||||||
4006 | addPrivate(HelperData.CounterVD, [&CGF, &HelperData]() { | ||||||
4007 | return CGF.CreateMemTemp(HelperData.CounterVD->getType(), | ||||||
4008 | "counter.addr"); | ||||||
4009 | }); | ||||||
4010 | } | ||||||
4011 | Privatize(); | ||||||
4012 | |||||||
4013 | for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) { | ||||||
4014 | const OMPIteratorHelperData &HelperData = E->getHelper(I); | ||||||
4015 | LValue CLVal = | ||||||
4016 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(HelperData.CounterVD), | ||||||
4017 | HelperData.CounterVD->getType()); | ||||||
4018 | // Counter = 0; | ||||||
4019 | CGF.EmitStoreOfScalar( | ||||||
4020 | llvm::ConstantInt::get(CLVal.getAddress(CGF).getElementType(), 0), | ||||||
4021 | CLVal); | ||||||
4022 | CodeGenFunction::JumpDest &ContDest = | ||||||
4023 | ContDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.cont")); | ||||||
4024 | CodeGenFunction::JumpDest &ExitDest = | ||||||
4025 | ExitDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.exit")); | ||||||
4026 | // N = <number-of_iterations>; | ||||||
4027 | llvm::Value *N = Uppers[I]; | ||||||
4028 | // cont: | ||||||
4029 | // if (Counter < N) goto body; else goto exit; | ||||||
4030 | CGF.EmitBlock(ContDest.getBlock()); | ||||||
4031 | auto *CVal = | ||||||
4032 | CGF.EmitLoadOfScalar(CLVal, HelperData.CounterVD->getLocation()); | ||||||
4033 | llvm::Value *Cmp = | ||||||
4034 | HelperData.CounterVD->getType()->isSignedIntegerOrEnumerationType() | ||||||
4035 | ? CGF.Builder.CreateICmpSLT(CVal, N) | ||||||
4036 | : CGF.Builder.CreateICmpULT(CVal, N); | ||||||
4037 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("iter.body"); | ||||||
4038 | CGF.Builder.CreateCondBr(Cmp, BodyBB, ExitDest.getBlock()); | ||||||
4039 | // body: | ||||||
4040 | CGF.EmitBlock(BodyBB); | ||||||
4041 | // Iteri = Begini + Counter * Stepi; | ||||||
4042 | CGF.EmitIgnoredExpr(HelperData.Update); | ||||||
4043 | } | ||||||
4044 | } | ||||||
4045 | ~OMPIteratorGeneratorScope() { | ||||||
4046 | if (!E) | ||||||
4047 | return; | ||||||
4048 | for (unsigned I = E->numOfIterators(); I > 0; --I) { | ||||||
4049 | // Counter = Counter + 1; | ||||||
4050 | const OMPIteratorHelperData &HelperData = E->getHelper(I - 1); | ||||||
4051 | CGF.EmitIgnoredExpr(HelperData.CounterUpdate); | ||||||
4052 | // goto cont; | ||||||
4053 | CGF.EmitBranchThroughCleanup(ContDests[I - 1]); | ||||||
4054 | // exit: | ||||||
4055 | CGF.EmitBlock(ExitDests[I - 1].getBlock(), /*IsFinished=*/I == 1); | ||||||
4056 | } | ||||||
4057 | } | ||||||
4058 | }; | ||||||
4059 | } // namespace | ||||||
4060 | |||||||
4061 | static std::pair<llvm::Value *, llvm::Value *> | ||||||
4062 | getPointerAndSize(CodeGenFunction &CGF, const Expr *E) { | ||||||
4063 | const auto *OASE = dyn_cast<OMPArrayShapingExpr>(E); | ||||||
4064 | llvm::Value *Addr; | ||||||
4065 | if (OASE) { | ||||||
4066 | const Expr *Base = OASE->getBase(); | ||||||
4067 | Addr = CGF.EmitScalarExpr(Base); | ||||||
4068 | } else { | ||||||
4069 | Addr = CGF.EmitLValue(E).getPointer(CGF); | ||||||
4070 | } | ||||||
4071 | llvm::Value *SizeVal; | ||||||
4072 | QualType Ty = E->getType(); | ||||||
4073 | if (OASE) { | ||||||
4074 | SizeVal = CGF.getTypeSize(OASE->getBase()->getType()->getPointeeType()); | ||||||
4075 | for (const Expr *SE : OASE->getDimensions()) { | ||||||
4076 | llvm::Value *Sz = CGF.EmitScalarExpr(SE); | ||||||
4077 | Sz = CGF.EmitScalarConversion( | ||||||
4078 | Sz, SE->getType(), CGF.getContext().getSizeType(), SE->getExprLoc()); | ||||||
4079 | SizeVal = CGF.Builder.CreateNUWMul(SizeVal, Sz); | ||||||
4080 | } | ||||||
4081 | } else if (const auto *ASE = | ||||||
4082 | dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) { | ||||||
4083 | LValue UpAddrLVal = | ||||||
4084 | CGF.EmitOMPArraySectionExpr(ASE, /*IsLowerBound=*/false); | ||||||
4085 | llvm::Value *UpAddr = | ||||||
4086 | CGF.Builder.CreateConstGEP1_32(UpAddrLVal.getPointer(CGF), /*Idx0=*/1); | ||||||
4087 | llvm::Value *LowIntPtr = CGF.Builder.CreatePtrToInt(Addr, CGF.SizeTy); | ||||||
4088 | llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGF.SizeTy); | ||||||
4089 | SizeVal = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr); | ||||||
4090 | } else { | ||||||
4091 | SizeVal = CGF.getTypeSize(Ty); | ||||||
4092 | } | ||||||
4093 | return std::make_pair(Addr, SizeVal); | ||||||
4094 | } | ||||||
4095 | |||||||
4096 | /// Builds kmp_depend_info, if it is not built yet, and builds flags type. | ||||||
4097 | static void getKmpAffinityType(ASTContext &C, QualType &KmpTaskAffinityInfoTy) { | ||||||
4098 | QualType FlagsTy = C.getIntTypeForBitwidth(32, /*Signed=*/false); | ||||||
4099 | if (KmpTaskAffinityInfoTy.isNull()) { | ||||||
4100 | RecordDecl *KmpAffinityInfoRD = | ||||||
4101 | C.buildImplicitRecord("kmp_task_affinity_info_t"); | ||||||
4102 | KmpAffinityInfoRD->startDefinition(); | ||||||
4103 | addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getIntPtrType()); | ||||||
4104 | addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getSizeType()); | ||||||
4105 | addFieldToRecordDecl(C, KmpAffinityInfoRD, FlagsTy); | ||||||
4106 | KmpAffinityInfoRD->completeDefinition(); | ||||||
4107 | KmpTaskAffinityInfoTy = C.getRecordType(KmpAffinityInfoRD); | ||||||
4108 | } | ||||||
4109 | } | ||||||
4110 | |||||||
4111 | CGOpenMPRuntime::TaskResultTy | ||||||
4112 | CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
4113 | const OMPExecutableDirective &D, | ||||||
4114 | llvm::Function *TaskFunction, QualType SharedsTy, | ||||||
4115 | Address Shareds, const OMPTaskDataTy &Data) { | ||||||
4116 | ASTContext &C = CGM.getContext(); | ||||||
4117 | llvm::SmallVector<PrivateDataTy, 4> Privates; | ||||||
4118 | // Aggregate privates and sort them by the alignment. | ||||||
4119 | const auto *I = Data.PrivateCopies.begin(); | ||||||
4120 | for (const Expr *E : Data.PrivateVars) { | ||||||
4121 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | ||||||
4122 | Privates.emplace_back( | ||||||
4123 | C.getDeclAlign(VD), | ||||||
4124 | PrivateHelpersTy(E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | ||||||
4125 | /*PrivateElemInit=*/nullptr)); | ||||||
4126 | ++I; | ||||||
4127 | } | ||||||
4128 | I = Data.FirstprivateCopies.begin(); | ||||||
4129 | const auto *IElemInitRef = Data.FirstprivateInits.begin(); | ||||||
4130 | for (const Expr *E : Data.FirstprivateVars) { | ||||||
4131 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | ||||||
4132 | Privates.emplace_back( | ||||||
4133 | C.getDeclAlign(VD), | ||||||
4134 | PrivateHelpersTy( | ||||||
4135 | E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | ||||||
4136 | cast<VarDecl>(cast<DeclRefExpr>(*IElemInitRef)->getDecl()))); | ||||||
4137 | ++I; | ||||||
4138 | ++IElemInitRef; | ||||||
4139 | } | ||||||
4140 | I = Data.LastprivateCopies.begin(); | ||||||
4141 | for (const Expr *E : Data.LastprivateVars) { | ||||||
4142 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | ||||||
4143 | Privates.emplace_back( | ||||||
4144 | C.getDeclAlign(VD), | ||||||
4145 | PrivateHelpersTy(E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | ||||||
4146 | /*PrivateElemInit=*/nullptr)); | ||||||
4147 | ++I; | ||||||
4148 | } | ||||||
4149 | for (const VarDecl *VD : Data.PrivateLocals) | ||||||
4150 | Privates.emplace_back(C.getDeclAlign(VD), PrivateHelpersTy(VD)); | ||||||
4151 | llvm::stable_sort(Privates, | ||||||
4152 | [](const PrivateDataTy &L, const PrivateDataTy &R) { | ||||||
4153 | return L.first > R.first; | ||||||
4154 | }); | ||||||
4155 | QualType KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | ||||||
4156 | // Build type kmp_routine_entry_t (if not built yet). | ||||||
4157 | emitKmpRoutineEntryT(KmpInt32Ty); | ||||||
4158 | // Build type kmp_task_t (if not built yet). | ||||||
4159 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind())) { | ||||||
4160 | if (SavedKmpTaskloopTQTy.isNull()) { | ||||||
4161 | SavedKmpTaskloopTQTy = C.getRecordType(createKmpTaskTRecordDecl( | ||||||
4162 | CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy)); | ||||||
4163 | } | ||||||
4164 | KmpTaskTQTy = SavedKmpTaskloopTQTy; | ||||||
4165 | } else { | ||||||
4166 | assert((D.getDirectiveKind() == OMPD_task ||(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4169, __PRETTY_FUNCTION__)) | ||||||
4167 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()) ||(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4169, __PRETTY_FUNCTION__)) | ||||||
4168 | isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) &&(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4169, __PRETTY_FUNCTION__)) | ||||||
4169 | "Expected taskloop, task or target directive")(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4169, __PRETTY_FUNCTION__)); | ||||||
4170 | if (SavedKmpTaskTQTy.isNull()) { | ||||||
4171 | SavedKmpTaskTQTy = C.getRecordType(createKmpTaskTRecordDecl( | ||||||
4172 | CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy)); | ||||||
4173 | } | ||||||
4174 | KmpTaskTQTy = SavedKmpTaskTQTy; | ||||||
4175 | } | ||||||
4176 | const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl()); | ||||||
4177 | // Build particular struct kmp_task_t for the given task. | ||||||
4178 | const RecordDecl *KmpTaskTWithPrivatesQTyRD = | ||||||
4179 | createKmpTaskTWithPrivatesRecordDecl(CGM, KmpTaskTQTy, Privates); | ||||||
4180 | QualType KmpTaskTWithPrivatesQTy = C.getRecordType(KmpTaskTWithPrivatesQTyRD); | ||||||
4181 | QualType KmpTaskTWithPrivatesPtrQTy = | ||||||
4182 | C.getPointerType(KmpTaskTWithPrivatesQTy); | ||||||
4183 | llvm::Type *KmpTaskTWithPrivatesTy = CGF.ConvertType(KmpTaskTWithPrivatesQTy); | ||||||
4184 | llvm::Type *KmpTaskTWithPrivatesPtrTy = | ||||||
4185 | KmpTaskTWithPrivatesTy->getPointerTo(); | ||||||
4186 | llvm::Value *KmpTaskTWithPrivatesTySize = | ||||||
4187 | CGF.getTypeSize(KmpTaskTWithPrivatesQTy); | ||||||
4188 | QualType SharedsPtrTy = C.getPointerType(SharedsTy); | ||||||
4189 | |||||||
4190 | // Emit initial values for private copies (if any). | ||||||
4191 | llvm::Value *TaskPrivatesMap = nullptr; | ||||||
4192 | llvm::Type *TaskPrivatesMapTy = | ||||||
4193 | std::next(TaskFunction->arg_begin(), 3)->getType(); | ||||||
4194 | if (!Privates.empty()) { | ||||||
4195 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
4196 | TaskPrivatesMap = | ||||||
4197 | emitTaskPrivateMappingFunction(CGM, Loc, Data, FI->getType(), Privates); | ||||||
4198 | TaskPrivatesMap = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4199 | TaskPrivatesMap, TaskPrivatesMapTy); | ||||||
4200 | } else { | ||||||
4201 | TaskPrivatesMap = llvm::ConstantPointerNull::get( | ||||||
4202 | cast<llvm::PointerType>(TaskPrivatesMapTy)); | ||||||
4203 | } | ||||||
4204 | // Build a proxy function kmp_int32 .omp_task_entry.(kmp_int32 gtid, | ||||||
4205 | // kmp_task_t *tt); | ||||||
4206 | llvm::Function *TaskEntry = emitProxyTaskFunction( | ||||||
4207 | CGM, Loc, D.getDirectiveKind(), KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy, | ||||||
4208 | KmpTaskTWithPrivatesQTy, KmpTaskTQTy, SharedsPtrTy, TaskFunction, | ||||||
4209 | TaskPrivatesMap); | ||||||
4210 | |||||||
4211 | // Build call kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, | ||||||
4212 | // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, | ||||||
4213 | // kmp_routine_entry_t *task_entry); | ||||||
4214 | // Task flags. Format is taken from | ||||||
4215 | // https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp.h, | ||||||
4216 | // description of kmp_tasking_flags struct. | ||||||
4217 | enum { | ||||||
4218 | TiedFlag = 0x1, | ||||||
4219 | FinalFlag = 0x2, | ||||||
4220 | DestructorsFlag = 0x8, | ||||||
4221 | PriorityFlag = 0x20, | ||||||
4222 | DetachableFlag = 0x40, | ||||||
4223 | }; | ||||||
4224 | unsigned Flags = Data.Tied ? TiedFlag : 0; | ||||||
4225 | bool NeedsCleanup = false; | ||||||
4226 | if (!Privates.empty()) { | ||||||
4227 | NeedsCleanup = | ||||||
4228 | checkDestructorsRequired(KmpTaskTWithPrivatesQTyRD, Privates); | ||||||
4229 | if (NeedsCleanup) | ||||||
4230 | Flags = Flags | DestructorsFlag; | ||||||
4231 | } | ||||||
4232 | if (Data.Priority.getInt()) | ||||||
4233 | Flags = Flags | PriorityFlag; | ||||||
4234 | if (D.hasClausesOfKind<OMPDetachClause>()) | ||||||
4235 | Flags = Flags | DetachableFlag; | ||||||
4236 | llvm::Value *TaskFlags = | ||||||
4237 | Data.Final.getPointer() | ||||||
4238 | ? CGF.Builder.CreateSelect(Data.Final.getPointer(), | ||||||
4239 | CGF.Builder.getInt32(FinalFlag), | ||||||
4240 | CGF.Builder.getInt32(/*C=*/0)) | ||||||
4241 | : CGF.Builder.getInt32(Data.Final.getInt() ? FinalFlag : 0); | ||||||
4242 | TaskFlags = CGF.Builder.CreateOr(TaskFlags, CGF.Builder.getInt32(Flags)); | ||||||
4243 | llvm::Value *SharedsSize = CGM.getSize(C.getTypeSizeInChars(SharedsTy)); | ||||||
4244 | SmallVector<llvm::Value *, 8> AllocArgs = {emitUpdateLocation(CGF, Loc), | ||||||
4245 | getThreadID(CGF, Loc), TaskFlags, KmpTaskTWithPrivatesTySize, | ||||||
4246 | SharedsSize, CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4247 | TaskEntry, KmpRoutineEntryPtrTy)}; | ||||||
4248 | llvm::Value *NewTask; | ||||||
4249 | if (D.hasClausesOfKind<OMPNowaitClause>()) { | ||||||
4250 | // Check if we have any device clause associated with the directive. | ||||||
4251 | const Expr *Device = nullptr; | ||||||
4252 | if (auto *C = D.getSingleClause<OMPDeviceClause>()) | ||||||
4253 | Device = C->getDevice(); | ||||||
4254 | // Emit device ID if any otherwise use default value. | ||||||
4255 | llvm::Value *DeviceID; | ||||||
4256 | if (Device) | ||||||
4257 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | ||||||
4258 | CGF.Int64Ty, /*isSigned=*/true); | ||||||
4259 | else | ||||||
4260 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | ||||||
4261 | AllocArgs.push_back(DeviceID); | ||||||
4262 | NewTask = CGF.EmitRuntimeCall( | ||||||
4263 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
4264 | CGM.getModule(), OMPRTL___kmpc_omp_target_task_alloc), | ||||||
4265 | AllocArgs); | ||||||
4266 | } else { | ||||||
4267 | NewTask = | ||||||
4268 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
4269 | CGM.getModule(), OMPRTL___kmpc_omp_task_alloc), | ||||||
4270 | AllocArgs); | ||||||
4271 | } | ||||||
4272 | // Emit detach clause initialization. | ||||||
4273 | // evt = (typeof(evt))__kmpc_task_allow_completion_event(loc, tid, | ||||||
4274 | // task_descriptor); | ||||||
4275 | if (const auto *DC = D.getSingleClause<OMPDetachClause>()) { | ||||||
4276 | const Expr *Evt = DC->getEventHandler()->IgnoreParenImpCasts(); | ||||||
4277 | LValue EvtLVal = CGF.EmitLValue(Evt); | ||||||
4278 | |||||||
4279 | // Build kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, | ||||||
4280 | // int gtid, kmp_task_t *task); | ||||||
4281 | llvm::Value *Loc = emitUpdateLocation(CGF, DC->getBeginLoc()); | ||||||
4282 | llvm::Value *Tid = getThreadID(CGF, DC->getBeginLoc()); | ||||||
4283 | Tid = CGF.Builder.CreateIntCast(Tid, CGF.IntTy, /*isSigned=*/false); | ||||||
4284 | llvm::Value *EvtVal = CGF.EmitRuntimeCall( | ||||||
4285 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
4286 | CGM.getModule(), OMPRTL___kmpc_task_allow_completion_event), | ||||||
4287 | {Loc, Tid, NewTask}); | ||||||
4288 | EvtVal = CGF.EmitScalarConversion(EvtVal, C.VoidPtrTy, Evt->getType(), | ||||||
4289 | Evt->getExprLoc()); | ||||||
4290 | CGF.EmitStoreOfScalar(EvtVal, EvtLVal); | ||||||
4291 | } | ||||||
4292 | // Process affinity clauses. | ||||||
4293 | if (D.hasClausesOfKind<OMPAffinityClause>()) { | ||||||
4294 | // Process list of affinity data. | ||||||
4295 | ASTContext &C = CGM.getContext(); | ||||||
4296 | Address AffinitiesArray = Address::invalid(); | ||||||
4297 | // Calculate number of elements to form the array of affinity data. | ||||||
4298 | llvm::Value *NumOfElements = nullptr; | ||||||
4299 | unsigned NumAffinities = 0; | ||||||
4300 | for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) { | ||||||
4301 | if (const Expr *Modifier = C->getModifier()) { | ||||||
4302 | const auto *IE = cast<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts()); | ||||||
4303 | for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) { | ||||||
4304 | llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper); | ||||||
4305 | Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false); | ||||||
4306 | NumOfElements = | ||||||
4307 | NumOfElements ? CGF.Builder.CreateNUWMul(NumOfElements, Sz) : Sz; | ||||||
4308 | } | ||||||
4309 | } else { | ||||||
4310 | NumAffinities += C->varlist_size(); | ||||||
4311 | } | ||||||
4312 | } | ||||||
4313 | getKmpAffinityType(CGM.getContext(), KmpTaskAffinityInfoTy); | ||||||
4314 | // Fields ids in kmp_task_affinity_info record. | ||||||
4315 | enum RTLAffinityInfoFieldsTy { BaseAddr, Len, Flags }; | ||||||
4316 | |||||||
4317 | QualType KmpTaskAffinityInfoArrayTy; | ||||||
4318 | if (NumOfElements) { | ||||||
4319 | NumOfElements = CGF.Builder.CreateNUWAdd( | ||||||
4320 | llvm::ConstantInt::get(CGF.SizeTy, NumAffinities), NumOfElements); | ||||||
4321 | OpaqueValueExpr OVE( | ||||||
4322 | Loc, | ||||||
4323 | C.getIntTypeForBitwidth(C.getTypeSize(C.getSizeType()), /*Signed=*/0), | ||||||
4324 | VK_RValue); | ||||||
4325 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, | ||||||
4326 | RValue::get(NumOfElements)); | ||||||
4327 | KmpTaskAffinityInfoArrayTy = | ||||||
4328 | C.getVariableArrayType(KmpTaskAffinityInfoTy, &OVE, ArrayType::Normal, | ||||||
4329 | /*IndexTypeQuals=*/0, SourceRange(Loc, Loc)); | ||||||
4330 | // Properly emit variable-sized array. | ||||||
4331 | auto *PD = ImplicitParamDecl::Create(C, KmpTaskAffinityInfoArrayTy, | ||||||
4332 | ImplicitParamDecl::Other); | ||||||
4333 | CGF.EmitVarDecl(*PD); | ||||||
4334 | AffinitiesArray = CGF.GetAddrOfLocalVar(PD); | ||||||
4335 | NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty, | ||||||
4336 | /*isSigned=*/false); | ||||||
4337 | } else { | ||||||
4338 | KmpTaskAffinityInfoArrayTy = C.getConstantArrayType( | ||||||
4339 | KmpTaskAffinityInfoTy, | ||||||
4340 | llvm::APInt(C.getTypeSize(C.getSizeType()), NumAffinities), nullptr, | ||||||
4341 | ArrayType::Normal, /*IndexTypeQuals=*/0); | ||||||
4342 | AffinitiesArray = | ||||||
4343 | CGF.CreateMemTemp(KmpTaskAffinityInfoArrayTy, ".affs.arr.addr"); | ||||||
4344 | AffinitiesArray = CGF.Builder.CreateConstArrayGEP(AffinitiesArray, 0); | ||||||
4345 | NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumAffinities, | ||||||
4346 | /*isSigned=*/false); | ||||||
4347 | } | ||||||
4348 | |||||||
4349 | const auto *KmpAffinityInfoRD = KmpTaskAffinityInfoTy->getAsRecordDecl(); | ||||||
4350 | // Fill array by elements without iterators. | ||||||
4351 | unsigned Pos = 0; | ||||||
4352 | bool HasIterator = false; | ||||||
4353 | for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) { | ||||||
4354 | if (C->getModifier()) { | ||||||
4355 | HasIterator = true; | ||||||
4356 | continue; | ||||||
4357 | } | ||||||
4358 | for (const Expr *E : C->varlists()) { | ||||||
4359 | llvm::Value *Addr; | ||||||
4360 | llvm::Value *Size; | ||||||
4361 | std::tie(Addr, Size) = getPointerAndSize(CGF, E); | ||||||
4362 | LValue Base = | ||||||
4363 | CGF.MakeAddrLValue(CGF.Builder.CreateConstGEP(AffinitiesArray, Pos), | ||||||
4364 | KmpTaskAffinityInfoTy); | ||||||
4365 | // affs[i].base_addr = &<Affinities[i].second>; | ||||||
4366 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4367 | Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr)); | ||||||
4368 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy), | ||||||
4369 | BaseAddrLVal); | ||||||
4370 | // affs[i].len = sizeof(<Affinities[i].second>); | ||||||
4371 | LValue LenLVal = CGF.EmitLValueForField( | ||||||
4372 | Base, *std::next(KmpAffinityInfoRD->field_begin(), Len)); | ||||||
4373 | CGF.EmitStoreOfScalar(Size, LenLVal); | ||||||
4374 | ++Pos; | ||||||
4375 | } | ||||||
4376 | } | ||||||
4377 | LValue PosLVal; | ||||||
4378 | if (HasIterator) { | ||||||
4379 | PosLVal = CGF.MakeAddrLValue( | ||||||
4380 | CGF.CreateMemTemp(C.getSizeType(), "affs.counter.addr"), | ||||||
4381 | C.getSizeType()); | ||||||
4382 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal); | ||||||
4383 | } | ||||||
4384 | // Process elements with iterators. | ||||||
4385 | for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) { | ||||||
4386 | const Expr *Modifier = C->getModifier(); | ||||||
4387 | if (!Modifier) | ||||||
4388 | continue; | ||||||
4389 | OMPIteratorGeneratorScope IteratorScope( | ||||||
4390 | CGF, cast_or_null<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts())); | ||||||
4391 | for (const Expr *E : C->varlists()) { | ||||||
4392 | llvm::Value *Addr; | ||||||
4393 | llvm::Value *Size; | ||||||
4394 | std::tie(Addr, Size) = getPointerAndSize(CGF, E); | ||||||
4395 | llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | ||||||
4396 | LValue Base = CGF.MakeAddrLValue( | ||||||
4397 | Address(CGF.Builder.CreateGEP(AffinitiesArray.getPointer(), Idx), | ||||||
4398 | AffinitiesArray.getAlignment()), | ||||||
4399 | KmpTaskAffinityInfoTy); | ||||||
4400 | // affs[i].base_addr = &<Affinities[i].second>; | ||||||
4401 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4402 | Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr)); | ||||||
4403 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy), | ||||||
4404 | BaseAddrLVal); | ||||||
4405 | // affs[i].len = sizeof(<Affinities[i].second>); | ||||||
4406 | LValue LenLVal = CGF.EmitLValueForField( | ||||||
4407 | Base, *std::next(KmpAffinityInfoRD->field_begin(), Len)); | ||||||
4408 | CGF.EmitStoreOfScalar(Size, LenLVal); | ||||||
4409 | Idx = CGF.Builder.CreateNUWAdd( | ||||||
4410 | Idx, llvm::ConstantInt::get(Idx->getType(), 1)); | ||||||
4411 | CGF.EmitStoreOfScalar(Idx, PosLVal); | ||||||
4412 | } | ||||||
4413 | } | ||||||
4414 | // Call to kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, | ||||||
4415 | // kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 | ||||||
4416 | // naffins, kmp_task_affinity_info_t *affin_list); | ||||||
4417 | llvm::Value *LocRef = emitUpdateLocation(CGF, Loc); | ||||||
4418 | llvm::Value *GTid = getThreadID(CGF, Loc); | ||||||
4419 | llvm::Value *AffinListPtr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4420 | AffinitiesArray.getPointer(), CGM.VoidPtrTy); | ||||||
4421 | // FIXME: Emit the function and ignore its result for now unless the | ||||||
4422 | // runtime function is properly implemented. | ||||||
4423 | (void)CGF.EmitRuntimeCall( | ||||||
4424 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
4425 | CGM.getModule(), OMPRTL___kmpc_omp_reg_task_with_affinity), | ||||||
4426 | {LocRef, GTid, NewTask, NumOfElements, AffinListPtr}); | ||||||
4427 | } | ||||||
4428 | llvm::Value *NewTaskNewTaskTTy = | ||||||
4429 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4430 | NewTask, KmpTaskTWithPrivatesPtrTy); | ||||||
4431 | LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy, | ||||||
4432 | KmpTaskTWithPrivatesQTy); | ||||||
4433 | LValue TDBase = | ||||||
4434 | CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin()); | ||||||
4435 | // Fill the data in the resulting kmp_task_t record. | ||||||
4436 | // Copy shareds if there are any. | ||||||
4437 | Address KmpTaskSharedsPtr = Address::invalid(); | ||||||
4438 | if (!SharedsTy->getAsStructureType()->getDecl()->field_empty()) { | ||||||
4439 | KmpTaskSharedsPtr = | ||||||
4440 | Address(CGF.EmitLoadOfScalar( | ||||||
4441 | CGF.EmitLValueForField( | ||||||
4442 | TDBase, *std::next(KmpTaskTQTyRD->field_begin(), | ||||||
4443 | KmpTaskTShareds)), | ||||||
4444 | Loc), | ||||||
4445 | CGM.getNaturalTypeAlignment(SharedsTy)); | ||||||
4446 | LValue Dest = CGF.MakeAddrLValue(KmpTaskSharedsPtr, SharedsTy); | ||||||
4447 | LValue Src = CGF.MakeAddrLValue(Shareds, SharedsTy); | ||||||
4448 | CGF.EmitAggregateCopy(Dest, Src, SharedsTy, AggValueSlot::DoesNotOverlap); | ||||||
4449 | } | ||||||
4450 | // Emit initial values for private copies (if any). | ||||||
4451 | TaskResultTy Result; | ||||||
4452 | if (!Privates.empty()) { | ||||||
4453 | emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, Base, KmpTaskTWithPrivatesQTyRD, | ||||||
4454 | SharedsTy, SharedsPtrTy, Data, Privates, | ||||||
4455 | /*ForDup=*/false); | ||||||
4456 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) && | ||||||
4457 | (!Data.LastprivateVars.empty() || checkInitIsRequired(CGF, Privates))) { | ||||||
4458 | Result.TaskDupFn = emitTaskDupFunction( | ||||||
4459 | CGM, Loc, D, KmpTaskTWithPrivatesPtrQTy, KmpTaskTWithPrivatesQTyRD, | ||||||
4460 | KmpTaskTQTyRD, SharedsTy, SharedsPtrTy, Data, Privates, | ||||||
4461 | /*WithLastIter=*/!Data.LastprivateVars.empty()); | ||||||
4462 | } | ||||||
4463 | } | ||||||
4464 | // Fields of union "kmp_cmplrdata_t" for destructors and priority. | ||||||
4465 | enum { Priority = 0, Destructors = 1 }; | ||||||
4466 | // Provide pointer to function with destructors for privates. | ||||||
4467 | auto FI = std::next(KmpTaskTQTyRD->field_begin(), Data1); | ||||||
4468 | const RecordDecl *KmpCmplrdataUD = | ||||||
4469 | (*FI)->getType()->getAsUnionType()->getDecl(); | ||||||
4470 | if (NeedsCleanup) { | ||||||
4471 | llvm::Value *DestructorFn = emitDestructorsFunction( | ||||||
4472 | CGM, Loc, KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy, | ||||||
4473 | KmpTaskTWithPrivatesQTy); | ||||||
4474 | LValue Data1LV = CGF.EmitLValueForField(TDBase, *FI); | ||||||
4475 | LValue DestructorsLV = CGF.EmitLValueForField( | ||||||
4476 | Data1LV, *std::next(KmpCmplrdataUD->field_begin(), Destructors)); | ||||||
4477 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4478 | DestructorFn, KmpRoutineEntryPtrTy), | ||||||
4479 | DestructorsLV); | ||||||
4480 | } | ||||||
4481 | // Set priority. | ||||||
4482 | if (Data.Priority.getInt()) { | ||||||
4483 | LValue Data2LV = CGF.EmitLValueForField( | ||||||
4484 | TDBase, *std::next(KmpTaskTQTyRD->field_begin(), Data2)); | ||||||
4485 | LValue PriorityLV = CGF.EmitLValueForField( | ||||||
4486 | Data2LV, *std::next(KmpCmplrdataUD->field_begin(), Priority)); | ||||||
4487 | CGF.EmitStoreOfScalar(Data.Priority.getPointer(), PriorityLV); | ||||||
4488 | } | ||||||
4489 | Result.NewTask = NewTask; | ||||||
4490 | Result.TaskEntry = TaskEntry; | ||||||
4491 | Result.NewTaskNewTaskTTy = NewTaskNewTaskTTy; | ||||||
4492 | Result.TDBase = TDBase; | ||||||
4493 | Result.KmpTaskTQTyRD = KmpTaskTQTyRD; | ||||||
4494 | return Result; | ||||||
4495 | } | ||||||
4496 | |||||||
4497 | namespace { | ||||||
4498 | /// Dependence kind for RTL. | ||||||
4499 | enum RTLDependenceKindTy { | ||||||
4500 | DepIn = 0x01, | ||||||
4501 | DepInOut = 0x3, | ||||||
4502 | DepMutexInOutSet = 0x4 | ||||||
4503 | }; | ||||||
4504 | /// Fields ids in kmp_depend_info record. | ||||||
4505 | enum RTLDependInfoFieldsTy { BaseAddr, Len, Flags }; | ||||||
4506 | } // namespace | ||||||
4507 | |||||||
4508 | /// Translates internal dependency kind into the runtime kind. | ||||||
4509 | static RTLDependenceKindTy translateDependencyKind(OpenMPDependClauseKind K) { | ||||||
4510 | RTLDependenceKindTy DepKind; | ||||||
4511 | switch (K) { | ||||||
4512 | case OMPC_DEPEND_in: | ||||||
4513 | DepKind = DepIn; | ||||||
4514 | break; | ||||||
4515 | // Out and InOut dependencies must use the same code. | ||||||
4516 | case OMPC_DEPEND_out: | ||||||
4517 | case OMPC_DEPEND_inout: | ||||||
4518 | DepKind = DepInOut; | ||||||
4519 | break; | ||||||
4520 | case OMPC_DEPEND_mutexinoutset: | ||||||
4521 | DepKind = DepMutexInOutSet; | ||||||
4522 | break; | ||||||
4523 | case OMPC_DEPEND_source: | ||||||
4524 | case OMPC_DEPEND_sink: | ||||||
4525 | case OMPC_DEPEND_depobj: | ||||||
4526 | case OMPC_DEPEND_unknown: | ||||||
4527 | llvm_unreachable("Unknown task dependence type")::llvm::llvm_unreachable_internal("Unknown task dependence type" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4527); | ||||||
4528 | } | ||||||
4529 | return DepKind; | ||||||
4530 | } | ||||||
4531 | |||||||
4532 | /// Builds kmp_depend_info, if it is not built yet, and builds flags type. | ||||||
4533 | static void getDependTypes(ASTContext &C, QualType &KmpDependInfoTy, | ||||||
4534 | QualType &FlagsTy) { | ||||||
4535 | FlagsTy = C.getIntTypeForBitwidth(C.getTypeSize(C.BoolTy), /*Signed=*/false); | ||||||
4536 | if (KmpDependInfoTy.isNull()) { | ||||||
4537 | RecordDecl *KmpDependInfoRD = C.buildImplicitRecord("kmp_depend_info"); | ||||||
4538 | KmpDependInfoRD->startDefinition(); | ||||||
4539 | addFieldToRecordDecl(C, KmpDependInfoRD, C.getIntPtrType()); | ||||||
4540 | addFieldToRecordDecl(C, KmpDependInfoRD, C.getSizeType()); | ||||||
4541 | addFieldToRecordDecl(C, KmpDependInfoRD, FlagsTy); | ||||||
4542 | KmpDependInfoRD->completeDefinition(); | ||||||
4543 | KmpDependInfoTy = C.getRecordType(KmpDependInfoRD); | ||||||
4544 | } | ||||||
4545 | } | ||||||
4546 | |||||||
4547 | std::pair<llvm::Value *, LValue> | ||||||
4548 | CGOpenMPRuntime::getDepobjElements(CodeGenFunction &CGF, LValue DepobjLVal, | ||||||
4549 | SourceLocation Loc) { | ||||||
4550 | ASTContext &C = CGM.getContext(); | ||||||
4551 | QualType FlagsTy; | ||||||
4552 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4553 | RecordDecl *KmpDependInfoRD = | ||||||
4554 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | ||||||
4555 | LValue Base = CGF.EmitLoadOfPointerLValue( | ||||||
4556 | DepobjLVal.getAddress(CGF), | ||||||
4557 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
4558 | QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy); | ||||||
4559 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4560 | Base.getAddress(CGF), CGF.ConvertTypeForMem(KmpDependInfoPtrTy)); | ||||||
4561 | Base = CGF.MakeAddrLValue(Addr, KmpDependInfoTy, Base.getBaseInfo(), | ||||||
4562 | Base.getTBAAInfo()); | ||||||
4563 | llvm::Value *DepObjAddr = CGF.Builder.CreateGEP( | ||||||
4564 | Addr.getPointer(), | ||||||
4565 | llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true)); | ||||||
4566 | LValue NumDepsBase = CGF.MakeAddrLValue( | ||||||
4567 | Address(DepObjAddr, Addr.getAlignment()), KmpDependInfoTy, | ||||||
4568 | Base.getBaseInfo(), Base.getTBAAInfo()); | ||||||
4569 | // NumDeps = deps[i].base_addr; | ||||||
4570 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4571 | NumDepsBase, *std::next(KmpDependInfoRD->field_begin(), BaseAddr)); | ||||||
4572 | llvm::Value *NumDeps = CGF.EmitLoadOfScalar(BaseAddrLVal, Loc); | ||||||
4573 | return std::make_pair(NumDeps, Base); | ||||||
4574 | } | ||||||
4575 | |||||||
4576 | static void emitDependData(CodeGenFunction &CGF, QualType &KmpDependInfoTy, | ||||||
4577 | llvm::PointerUnion<unsigned *, LValue *> Pos, | ||||||
4578 | const OMPTaskDataTy::DependData &Data, | ||||||
4579 | Address DependenciesArray) { | ||||||
4580 | CodeGenModule &CGM = CGF.CGM; | ||||||
4581 | ASTContext &C = CGM.getContext(); | ||||||
4582 | QualType FlagsTy; | ||||||
4583 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4584 | RecordDecl *KmpDependInfoRD = | ||||||
4585 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | ||||||
4586 | llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy); | ||||||
4587 | |||||||
4588 | OMPIteratorGeneratorScope IteratorScope( | ||||||
4589 | CGF, cast_or_null<OMPIteratorExpr>( | ||||||
4590 | Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts() | ||||||
4591 | : nullptr)); | ||||||
4592 | for (const Expr *E : Data.DepExprs) { | ||||||
4593 | llvm::Value *Addr; | ||||||
4594 | llvm::Value *Size; | ||||||
4595 | std::tie(Addr, Size) = getPointerAndSize(CGF, E); | ||||||
4596 | LValue Base; | ||||||
4597 | if (unsigned *P = Pos.dyn_cast<unsigned *>()) { | ||||||
4598 | Base = CGF.MakeAddrLValue( | ||||||
4599 | CGF.Builder.CreateConstGEP(DependenciesArray, *P), KmpDependInfoTy); | ||||||
4600 | } else { | ||||||
4601 | LValue &PosLVal = *Pos.get<LValue *>(); | ||||||
4602 | llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | ||||||
4603 | Base = CGF.MakeAddrLValue( | ||||||
4604 | Address(CGF.Builder.CreateGEP(DependenciesArray.getPointer(), Idx), | ||||||
4605 | DependenciesArray.getAlignment()), | ||||||
4606 | KmpDependInfoTy); | ||||||
4607 | } | ||||||
4608 | // deps[i].base_addr = &<Dependencies[i].second>; | ||||||
4609 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4610 | Base, *std::next(KmpDependInfoRD->field_begin(), BaseAddr)); | ||||||
4611 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy), | ||||||
4612 | BaseAddrLVal); | ||||||
4613 | // deps[i].len = sizeof(<Dependencies[i].second>); | ||||||
4614 | LValue LenLVal = CGF.EmitLValueForField( | ||||||
4615 | Base, *std::next(KmpDependInfoRD->field_begin(), Len)); | ||||||
4616 | CGF.EmitStoreOfScalar(Size, LenLVal); | ||||||
4617 | // deps[i].flags = <Dependencies[i].first>; | ||||||
4618 | RTLDependenceKindTy DepKind = translateDependencyKind(Data.DepKind); | ||||||
4619 | LValue FlagsLVal = CGF.EmitLValueForField( | ||||||
4620 | Base, *std::next(KmpDependInfoRD->field_begin(), Flags)); | ||||||
4621 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(LLVMFlagsTy, DepKind), | ||||||
4622 | FlagsLVal); | ||||||
4623 | if (unsigned *P = Pos.dyn_cast<unsigned *>()) { | ||||||
4624 | ++(*P); | ||||||
4625 | } else { | ||||||
4626 | LValue &PosLVal = *Pos.get<LValue *>(); | ||||||
4627 | llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | ||||||
4628 | Idx = CGF.Builder.CreateNUWAdd(Idx, | ||||||
4629 | llvm::ConstantInt::get(Idx->getType(), 1)); | ||||||
4630 | CGF.EmitStoreOfScalar(Idx, PosLVal); | ||||||
4631 | } | ||||||
4632 | } | ||||||
4633 | } | ||||||
4634 | |||||||
4635 | static SmallVector<llvm::Value *, 4> | ||||||
4636 | emitDepobjElementsSizes(CodeGenFunction &CGF, QualType &KmpDependInfoTy, | ||||||
4637 | const OMPTaskDataTy::DependData &Data) { | ||||||
4638 | assert(Data.DepKind == OMPC_DEPEND_depobj &&((Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependecy kind." ) ? static_cast<void> (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependecy kind.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4639, __PRETTY_FUNCTION__)) | ||||||
4639 | "Expected depobj dependecy kind.")((Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependecy kind." ) ? static_cast<void> (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependecy kind.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4639, __PRETTY_FUNCTION__)); | ||||||
4640 | SmallVector<llvm::Value *, 4> Sizes; | ||||||
4641 | SmallVector<LValue, 4> SizeLVals; | ||||||
4642 | ASTContext &C = CGF.getContext(); | ||||||
4643 | QualType FlagsTy; | ||||||
4644 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4645 | RecordDecl *KmpDependInfoRD = | ||||||
4646 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | ||||||
4647 | QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy); | ||||||
4648 | llvm::Type *KmpDependInfoPtrT = CGF.ConvertTypeForMem(KmpDependInfoPtrTy); | ||||||
4649 | { | ||||||
4650 | OMPIteratorGeneratorScope IteratorScope( | ||||||
4651 | CGF, cast_or_null<OMPIteratorExpr>( | ||||||
4652 | Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts() | ||||||
4653 | : nullptr)); | ||||||
4654 | for (const Expr *E : Data.DepExprs) { | ||||||
4655 | LValue DepobjLVal = CGF.EmitLValue(E->IgnoreParenImpCasts()); | ||||||
4656 | LValue Base = CGF.EmitLoadOfPointerLValue( | ||||||
4657 | DepobjLVal.getAddress(CGF), | ||||||
4658 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
4659 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4660 | Base.getAddress(CGF), KmpDependInfoPtrT); | ||||||
4661 | Base = CGF.MakeAddrLValue(Addr, KmpDependInfoTy, Base.getBaseInfo(), | ||||||
4662 | Base.getTBAAInfo()); | ||||||
4663 | llvm::Value *DepObjAddr = CGF.Builder.CreateGEP( | ||||||
4664 | Addr.getPointer(), | ||||||
4665 | llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true)); | ||||||
4666 | LValue NumDepsBase = CGF.MakeAddrLValue( | ||||||
4667 | Address(DepObjAddr, Addr.getAlignment()), KmpDependInfoTy, | ||||||
4668 | Base.getBaseInfo(), Base.getTBAAInfo()); | ||||||
4669 | // NumDeps = deps[i].base_addr; | ||||||
4670 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4671 | NumDepsBase, *std::next(KmpDependInfoRD->field_begin(), BaseAddr)); | ||||||
4672 | llvm::Value *NumDeps = | ||||||
4673 | CGF.EmitLoadOfScalar(BaseAddrLVal, E->getExprLoc()); | ||||||
4674 | LValue NumLVal = CGF.MakeAddrLValue( | ||||||
4675 | CGF.CreateMemTemp(C.getUIntPtrType(), "depobj.size.addr"), | ||||||
4676 | C.getUIntPtrType()); | ||||||
4677 | CGF.InitTempAlloca(NumLVal.getAddress(CGF), | ||||||
4678 | llvm::ConstantInt::get(CGF.IntPtrTy, 0)); | ||||||
4679 | llvm::Value *PrevVal = CGF.EmitLoadOfScalar(NumLVal, E->getExprLoc()); | ||||||
4680 | llvm::Value *Add = CGF.Builder.CreateNUWAdd(PrevVal, NumDeps); | ||||||
4681 | CGF.EmitStoreOfScalar(Add, NumLVal); | ||||||
4682 | SizeLVals.push_back(NumLVal); | ||||||
4683 | } | ||||||
4684 | } | ||||||
4685 | for (unsigned I = 0, E = SizeLVals.size(); I < E; ++I) { | ||||||
4686 | llvm::Value *Size = | ||||||
4687 | CGF.EmitLoadOfScalar(SizeLVals[I], Data.DepExprs[I]->getExprLoc()); | ||||||
4688 | Sizes.push_back(Size); | ||||||
4689 | } | ||||||
4690 | return Sizes; | ||||||
4691 | } | ||||||
4692 | |||||||
4693 | static void emitDepobjElements(CodeGenFunction &CGF, QualType &KmpDependInfoTy, | ||||||
4694 | LValue PosLVal, | ||||||
4695 | const OMPTaskDataTy::DependData &Data, | ||||||
4696 | Address DependenciesArray) { | ||||||
4697 | assert(Data.DepKind == OMPC_DEPEND_depobj &&((Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependecy kind." ) ? static_cast<void> (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependecy kind.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4698, __PRETTY_FUNCTION__)) | ||||||
4698 | "Expected depobj dependecy kind.")((Data.DepKind == OMPC_DEPEND_depobj && "Expected depobj dependecy kind." ) ? static_cast<void> (0) : __assert_fail ("Data.DepKind == OMPC_DEPEND_depobj && \"Expected depobj dependecy kind.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4698, __PRETTY_FUNCTION__)); | ||||||
4699 | ASTContext &C = CGF.getContext(); | ||||||
4700 | QualType FlagsTy; | ||||||
4701 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4702 | RecordDecl *KmpDependInfoRD = | ||||||
4703 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | ||||||
4704 | QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy); | ||||||
4705 | llvm::Type *KmpDependInfoPtrT = CGF.ConvertTypeForMem(KmpDependInfoPtrTy); | ||||||
4706 | llvm::Value *ElSize = CGF.getTypeSize(KmpDependInfoTy); | ||||||
4707 | { | ||||||
4708 | OMPIteratorGeneratorScope IteratorScope( | ||||||
4709 | CGF, cast_or_null<OMPIteratorExpr>( | ||||||
4710 | Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts() | ||||||
4711 | : nullptr)); | ||||||
4712 | for (unsigned I = 0, End = Data.DepExprs.size(); I < End; ++I) { | ||||||
4713 | const Expr *E = Data.DepExprs[I]; | ||||||
4714 | LValue DepobjLVal = CGF.EmitLValue(E->IgnoreParenImpCasts()); | ||||||
4715 | LValue Base = CGF.EmitLoadOfPointerLValue( | ||||||
4716 | DepobjLVal.getAddress(CGF), | ||||||
4717 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
4718 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4719 | Base.getAddress(CGF), KmpDependInfoPtrT); | ||||||
4720 | Base = CGF.MakeAddrLValue(Addr, KmpDependInfoTy, Base.getBaseInfo(), | ||||||
4721 | Base.getTBAAInfo()); | ||||||
4722 | |||||||
4723 | // Get number of elements in a single depobj. | ||||||
4724 | llvm::Value *DepObjAddr = CGF.Builder.CreateGEP( | ||||||
4725 | Addr.getPointer(), | ||||||
4726 | llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true)); | ||||||
4727 | LValue NumDepsBase = CGF.MakeAddrLValue( | ||||||
4728 | Address(DepObjAddr, Addr.getAlignment()), KmpDependInfoTy, | ||||||
4729 | Base.getBaseInfo(), Base.getTBAAInfo()); | ||||||
4730 | // NumDeps = deps[i].base_addr; | ||||||
4731 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4732 | NumDepsBase, *std::next(KmpDependInfoRD->field_begin(), BaseAddr)); | ||||||
4733 | llvm::Value *NumDeps = | ||||||
4734 | CGF.EmitLoadOfScalar(BaseAddrLVal, E->getExprLoc()); | ||||||
4735 | |||||||
4736 | // memcopy dependency data. | ||||||
4737 | llvm::Value *Size = CGF.Builder.CreateNUWMul( | ||||||
4738 | ElSize, | ||||||
4739 | CGF.Builder.CreateIntCast(NumDeps, CGF.SizeTy, /*isSigned=*/false)); | ||||||
4740 | llvm::Value *Pos = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc()); | ||||||
4741 | Address DepAddr = | ||||||
4742 | Address(CGF.Builder.CreateGEP(DependenciesArray.getPointer(), Pos), | ||||||
4743 | DependenciesArray.getAlignment()); | ||||||
4744 | CGF.Builder.CreateMemCpy(DepAddr, Base.getAddress(CGF), Size); | ||||||
4745 | |||||||
4746 | // Increase pos. | ||||||
4747 | // pos += size; | ||||||
4748 | llvm::Value *Add = CGF.Builder.CreateNUWAdd(Pos, NumDeps); | ||||||
4749 | CGF.EmitStoreOfScalar(Add, PosLVal); | ||||||
4750 | } | ||||||
4751 | } | ||||||
4752 | } | ||||||
4753 | |||||||
4754 | std::pair<llvm::Value *, Address> CGOpenMPRuntime::emitDependClause( | ||||||
4755 | CodeGenFunction &CGF, ArrayRef<OMPTaskDataTy::DependData> Dependencies, | ||||||
4756 | SourceLocation Loc) { | ||||||
4757 | if (llvm::all_of(Dependencies, [](const OMPTaskDataTy::DependData &D) { | ||||||
4758 | return D.DepExprs.empty(); | ||||||
4759 | })) | ||||||
4760 | return std::make_pair(nullptr, Address::invalid()); | ||||||
4761 | // Process list of dependencies. | ||||||
4762 | ASTContext &C = CGM.getContext(); | ||||||
4763 | Address DependenciesArray = Address::invalid(); | ||||||
4764 | llvm::Value *NumOfElements = nullptr; | ||||||
4765 | unsigned NumDependencies = std::accumulate( | ||||||
4766 | Dependencies.begin(), Dependencies.end(), 0, | ||||||
4767 | [](unsigned V, const OMPTaskDataTy::DependData &D) { | ||||||
4768 | return D.DepKind == OMPC_DEPEND_depobj | ||||||
4769 | ? V | ||||||
4770 | : (V + (D.IteratorExpr ? 0 : D.DepExprs.size())); | ||||||
4771 | }); | ||||||
4772 | QualType FlagsTy; | ||||||
4773 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4774 | bool HasDepobjDeps = false; | ||||||
4775 | bool HasRegularWithIterators = false; | ||||||
4776 | llvm::Value *NumOfDepobjElements = llvm::ConstantInt::get(CGF.IntPtrTy, 0); | ||||||
4777 | llvm::Value *NumOfRegularWithIterators = | ||||||
4778 | llvm::ConstantInt::get(CGF.IntPtrTy, 1); | ||||||
4779 | // Calculate number of depobj dependecies and regular deps with the iterators. | ||||||
4780 | for (const OMPTaskDataTy::DependData &D : Dependencies) { | ||||||
4781 | if (D.DepKind == OMPC_DEPEND_depobj) { | ||||||
4782 | SmallVector<llvm::Value *, 4> Sizes = | ||||||
4783 | emitDepobjElementsSizes(CGF, KmpDependInfoTy, D); | ||||||
4784 | for (llvm::Value *Size : Sizes) { | ||||||
4785 | NumOfDepobjElements = | ||||||
4786 | CGF.Builder.CreateNUWAdd(NumOfDepobjElements, Size); | ||||||
4787 | } | ||||||
4788 | HasDepobjDeps = true; | ||||||
4789 | continue; | ||||||
4790 | } | ||||||
4791 | // Include number of iterations, if any. | ||||||
4792 | if (const auto *IE = cast_or_null<OMPIteratorExpr>(D.IteratorExpr)) { | ||||||
4793 | for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) { | ||||||
4794 | llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper); | ||||||
4795 | Sz = CGF.Builder.CreateIntCast(Sz, CGF.IntPtrTy, /*isSigned=*/false); | ||||||
4796 | NumOfRegularWithIterators = | ||||||
4797 | CGF.Builder.CreateNUWMul(NumOfRegularWithIterators, Sz); | ||||||
4798 | } | ||||||
4799 | HasRegularWithIterators = true; | ||||||
4800 | continue; | ||||||
4801 | } | ||||||
4802 | } | ||||||
4803 | |||||||
4804 | QualType KmpDependInfoArrayTy; | ||||||
4805 | if (HasDepobjDeps || HasRegularWithIterators) { | ||||||
4806 | NumOfElements = llvm::ConstantInt::get(CGM.IntPtrTy, NumDependencies, | ||||||
4807 | /*isSigned=*/false); | ||||||
4808 | if (HasDepobjDeps) { | ||||||
4809 | NumOfElements = | ||||||
4810 | CGF.Builder.CreateNUWAdd(NumOfDepobjElements, NumOfElements); | ||||||
4811 | } | ||||||
4812 | if (HasRegularWithIterators) { | ||||||
4813 | NumOfElements = | ||||||
4814 | CGF.Builder.CreateNUWAdd(NumOfRegularWithIterators, NumOfElements); | ||||||
4815 | } | ||||||
4816 | OpaqueValueExpr OVE(Loc, | ||||||
4817 | C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0), | ||||||
4818 | VK_RValue); | ||||||
4819 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, | ||||||
4820 | RValue::get(NumOfElements)); | ||||||
4821 | KmpDependInfoArrayTy = | ||||||
4822 | C.getVariableArrayType(KmpDependInfoTy, &OVE, ArrayType::Normal, | ||||||
4823 | /*IndexTypeQuals=*/0, SourceRange(Loc, Loc)); | ||||||
4824 | // CGF.EmitVariablyModifiedType(KmpDependInfoArrayTy); | ||||||
4825 | // Properly emit variable-sized array. | ||||||
4826 | auto *PD = ImplicitParamDecl::Create(C, KmpDependInfoArrayTy, | ||||||
4827 | ImplicitParamDecl::Other); | ||||||
4828 | CGF.EmitVarDecl(*PD); | ||||||
4829 | DependenciesArray = CGF.GetAddrOfLocalVar(PD); | ||||||
4830 | NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty, | ||||||
4831 | /*isSigned=*/false); | ||||||
4832 | } else { | ||||||
4833 | KmpDependInfoArrayTy = C.getConstantArrayType( | ||||||
4834 | KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies), nullptr, | ||||||
4835 | ArrayType::Normal, /*IndexTypeQuals=*/0); | ||||||
4836 | DependenciesArray = | ||||||
4837 | CGF.CreateMemTemp(KmpDependInfoArrayTy, ".dep.arr.addr"); | ||||||
4838 | DependenciesArray = CGF.Builder.CreateConstArrayGEP(DependenciesArray, 0); | ||||||
4839 | NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumDependencies, | ||||||
4840 | /*isSigned=*/false); | ||||||
4841 | } | ||||||
4842 | unsigned Pos = 0; | ||||||
4843 | for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) { | ||||||
4844 | if (Dependencies[I].DepKind == OMPC_DEPEND_depobj || | ||||||
4845 | Dependencies[I].IteratorExpr) | ||||||
4846 | continue; | ||||||
4847 | emitDependData(CGF, KmpDependInfoTy, &Pos, Dependencies[I], | ||||||
4848 | DependenciesArray); | ||||||
4849 | } | ||||||
4850 | // Copy regular dependecies with iterators. | ||||||
4851 | LValue PosLVal = CGF.MakeAddrLValue( | ||||||
4852 | CGF.CreateMemTemp(C.getSizeType(), "dep.counter.addr"), C.getSizeType()); | ||||||
4853 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal); | ||||||
4854 | for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) { | ||||||
4855 | if (Dependencies[I].DepKind == OMPC_DEPEND_depobj || | ||||||
4856 | !Dependencies[I].IteratorExpr) | ||||||
4857 | continue; | ||||||
4858 | emitDependData(CGF, KmpDependInfoTy, &PosLVal, Dependencies[I], | ||||||
4859 | DependenciesArray); | ||||||
4860 | } | ||||||
4861 | // Copy final depobj arrays without iterators. | ||||||
4862 | if (HasDepobjDeps) { | ||||||
4863 | for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) { | ||||||
4864 | if (Dependencies[I].DepKind != OMPC_DEPEND_depobj) | ||||||
4865 | continue; | ||||||
4866 | emitDepobjElements(CGF, KmpDependInfoTy, PosLVal, Dependencies[I], | ||||||
4867 | DependenciesArray); | ||||||
4868 | } | ||||||
4869 | } | ||||||
4870 | DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4871 | DependenciesArray, CGF.VoidPtrTy); | ||||||
4872 | return std::make_pair(NumOfElements, DependenciesArray); | ||||||
4873 | } | ||||||
4874 | |||||||
4875 | Address CGOpenMPRuntime::emitDepobjDependClause( | ||||||
4876 | CodeGenFunction &CGF, const OMPTaskDataTy::DependData &Dependencies, | ||||||
4877 | SourceLocation Loc) { | ||||||
4878 | if (Dependencies.DepExprs.empty()) | ||||||
4879 | return Address::invalid(); | ||||||
4880 | // Process list of dependencies. | ||||||
4881 | ASTContext &C = CGM.getContext(); | ||||||
4882 | Address DependenciesArray = Address::invalid(); | ||||||
4883 | unsigned NumDependencies = Dependencies.DepExprs.size(); | ||||||
4884 | QualType FlagsTy; | ||||||
4885 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4886 | RecordDecl *KmpDependInfoRD = | ||||||
4887 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | ||||||
4888 | |||||||
4889 | llvm::Value *Size; | ||||||
4890 | // Define type kmp_depend_info[<Dependencies.size()>]; | ||||||
4891 | // For depobj reserve one extra element to store the number of elements. | ||||||
4892 | // It is required to handle depobj(x) update(in) construct. | ||||||
4893 | // kmp_depend_info[<Dependencies.size()>] deps; | ||||||
4894 | llvm::Value *NumDepsVal; | ||||||
4895 | CharUnits Align = C.getTypeAlignInChars(KmpDependInfoTy); | ||||||
4896 | if (const auto *IE = | ||||||
4897 | cast_or_null<OMPIteratorExpr>(Dependencies.IteratorExpr)) { | ||||||
4898 | NumDepsVal = llvm::ConstantInt::get(CGF.SizeTy, 1); | ||||||
4899 | for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) { | ||||||
4900 | llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper); | ||||||
4901 | Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false); | ||||||
4902 | NumDepsVal = CGF.Builder.CreateNUWMul(NumDepsVal, Sz); | ||||||
4903 | } | ||||||
4904 | Size = CGF.Builder.CreateNUWAdd(llvm::ConstantInt::get(CGF.SizeTy, 1), | ||||||
4905 | NumDepsVal); | ||||||
4906 | CharUnits SizeInBytes = | ||||||
4907 | C.getTypeSizeInChars(KmpDependInfoTy).alignTo(Align); | ||||||
4908 | llvm::Value *RecSize = CGM.getSize(SizeInBytes); | ||||||
4909 | Size = CGF.Builder.CreateNUWMul(Size, RecSize); | ||||||
4910 | NumDepsVal = | ||||||
4911 | CGF.Builder.CreateIntCast(NumDepsVal, CGF.IntPtrTy, /*isSigned=*/false); | ||||||
4912 | } else { | ||||||
4913 | QualType KmpDependInfoArrayTy = C.getConstantArrayType( | ||||||
4914 | KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies + 1), | ||||||
4915 | nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0); | ||||||
4916 | CharUnits Sz = C.getTypeSizeInChars(KmpDependInfoArrayTy); | ||||||
4917 | Size = CGM.getSize(Sz.alignTo(Align)); | ||||||
4918 | NumDepsVal = llvm::ConstantInt::get(CGF.IntPtrTy, NumDependencies); | ||||||
4919 | } | ||||||
4920 | // Need to allocate on the dynamic memory. | ||||||
4921 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | ||||||
4922 | // Use default allocator. | ||||||
4923 | llvm::Value *Allocator = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | ||||||
4924 | llvm::Value *Args[] = {ThreadID, Size, Allocator}; | ||||||
4925 | |||||||
4926 | llvm::Value *Addr = | ||||||
4927 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
4928 | CGM.getModule(), OMPRTL___kmpc_alloc), | ||||||
4929 | Args, ".dep.arr.addr"); | ||||||
4930 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4931 | Addr, CGF.ConvertTypeForMem(KmpDependInfoTy)->getPointerTo()); | ||||||
4932 | DependenciesArray = Address(Addr, Align); | ||||||
4933 | // Write number of elements in the first element of array for depobj. | ||||||
4934 | LValue Base = CGF.MakeAddrLValue(DependenciesArray, KmpDependInfoTy); | ||||||
4935 | // deps[i].base_addr = NumDependencies; | ||||||
4936 | LValue BaseAddrLVal = CGF.EmitLValueForField( | ||||||
4937 | Base, *std::next(KmpDependInfoRD->field_begin(), BaseAddr)); | ||||||
4938 | CGF.EmitStoreOfScalar(NumDepsVal, BaseAddrLVal); | ||||||
4939 | llvm::PointerUnion<unsigned *, LValue *> Pos; | ||||||
4940 | unsigned Idx = 1; | ||||||
4941 | LValue PosLVal; | ||||||
4942 | if (Dependencies.IteratorExpr) { | ||||||
4943 | PosLVal = CGF.MakeAddrLValue( | ||||||
4944 | CGF.CreateMemTemp(C.getSizeType(), "iterator.counter.addr"), | ||||||
4945 | C.getSizeType()); | ||||||
4946 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Idx), PosLVal, | ||||||
4947 | /*IsInit=*/true); | ||||||
4948 | Pos = &PosLVal; | ||||||
4949 | } else { | ||||||
4950 | Pos = &Idx; | ||||||
4951 | } | ||||||
4952 | emitDependData(CGF, KmpDependInfoTy, Pos, Dependencies, DependenciesArray); | ||||||
4953 | DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4954 | CGF.Builder.CreateConstGEP(DependenciesArray, 1), CGF.VoidPtrTy); | ||||||
4955 | return DependenciesArray; | ||||||
4956 | } | ||||||
4957 | |||||||
4958 | void CGOpenMPRuntime::emitDestroyClause(CodeGenFunction &CGF, LValue DepobjLVal, | ||||||
4959 | SourceLocation Loc) { | ||||||
4960 | ASTContext &C = CGM.getContext(); | ||||||
4961 | QualType FlagsTy; | ||||||
4962 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4963 | LValue Base = CGF.EmitLoadOfPointerLValue( | ||||||
4964 | DepobjLVal.getAddress(CGF), | ||||||
4965 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
4966 | QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy); | ||||||
4967 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
4968 | Base.getAddress(CGF), CGF.ConvertTypeForMem(KmpDependInfoPtrTy)); | ||||||
4969 | llvm::Value *DepObjAddr = CGF.Builder.CreateGEP( | ||||||
4970 | Addr.getPointer(), | ||||||
4971 | llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true)); | ||||||
4972 | DepObjAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(DepObjAddr, | ||||||
4973 | CGF.VoidPtrTy); | ||||||
4974 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | ||||||
4975 | // Use default allocator. | ||||||
4976 | llvm::Value *Allocator = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | ||||||
4977 | llvm::Value *Args[] = {ThreadID, DepObjAddr, Allocator}; | ||||||
4978 | |||||||
4979 | // _kmpc_free(gtid, addr, nullptr); | ||||||
4980 | (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
4981 | CGM.getModule(), OMPRTL___kmpc_free), | ||||||
4982 | Args); | ||||||
4983 | } | ||||||
4984 | |||||||
4985 | void CGOpenMPRuntime::emitUpdateClause(CodeGenFunction &CGF, LValue DepobjLVal, | ||||||
4986 | OpenMPDependClauseKind NewDepKind, | ||||||
4987 | SourceLocation Loc) { | ||||||
4988 | ASTContext &C = CGM.getContext(); | ||||||
4989 | QualType FlagsTy; | ||||||
4990 | getDependTypes(C, KmpDependInfoTy, FlagsTy); | ||||||
4991 | RecordDecl *KmpDependInfoRD = | ||||||
4992 | cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | ||||||
4993 | llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy); | ||||||
4994 | llvm::Value *NumDeps; | ||||||
4995 | LValue Base; | ||||||
4996 | std::tie(NumDeps, Base) = getDepobjElements(CGF, DepobjLVal, Loc); | ||||||
4997 | |||||||
4998 | Address Begin = Base.getAddress(CGF); | ||||||
4999 | // Cast from pointer to array type to pointer to single element. | ||||||
5000 | llvm::Value *End = CGF.Builder.CreateGEP(Begin.getPointer(), NumDeps); | ||||||
5001 | // The basic structure here is a while-do loop. | ||||||
5002 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.body"); | ||||||
5003 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.done"); | ||||||
5004 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | ||||||
5005 | CGF.EmitBlock(BodyBB); | ||||||
5006 | llvm::PHINode *ElementPHI = | ||||||
5007 | CGF.Builder.CreatePHI(Begin.getType(), 2, "omp.elementPast"); | ||||||
5008 | ElementPHI->addIncoming(Begin.getPointer(), EntryBB); | ||||||
5009 | Begin = Address(ElementPHI, Begin.getAlignment()); | ||||||
5010 | Base = CGF.MakeAddrLValue(Begin, KmpDependInfoTy, Base.getBaseInfo(), | ||||||
5011 | Base.getTBAAInfo()); | ||||||
5012 | // deps[i].flags = NewDepKind; | ||||||
5013 | RTLDependenceKindTy DepKind = translateDependencyKind(NewDepKind); | ||||||
5014 | LValue FlagsLVal = CGF.EmitLValueForField( | ||||||
5015 | Base, *std::next(KmpDependInfoRD->field_begin(), Flags)); | ||||||
5016 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(LLVMFlagsTy, DepKind), | ||||||
5017 | FlagsLVal); | ||||||
5018 | |||||||
5019 | // Shift the address forward by one element. | ||||||
5020 | Address ElementNext = | ||||||
5021 | CGF.Builder.CreateConstGEP(Begin, /*Index=*/1, "omp.elementNext"); | ||||||
5022 | ElementPHI->addIncoming(ElementNext.getPointer(), | ||||||
5023 | CGF.Builder.GetInsertBlock()); | ||||||
5024 | llvm::Value *IsEmpty = | ||||||
5025 | CGF.Builder.CreateICmpEQ(ElementNext.getPointer(), End, "omp.isempty"); | ||||||
5026 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | ||||||
5027 | // Done. | ||||||
5028 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | ||||||
5029 | } | ||||||
5030 | |||||||
5031 | void CGOpenMPRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
5032 | const OMPExecutableDirective &D, | ||||||
5033 | llvm::Function *TaskFunction, | ||||||
5034 | QualType SharedsTy, Address Shareds, | ||||||
5035 | const Expr *IfCond, | ||||||
5036 | const OMPTaskDataTy &Data) { | ||||||
5037 | if (!CGF.HaveInsertPoint()) | ||||||
5038 | return; | ||||||
5039 | |||||||
5040 | TaskResultTy Result = | ||||||
5041 | emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data); | ||||||
5042 | llvm::Value *NewTask = Result.NewTask; | ||||||
5043 | llvm::Function *TaskEntry = Result.TaskEntry; | ||||||
5044 | llvm::Value *NewTaskNewTaskTTy = Result.NewTaskNewTaskTTy; | ||||||
5045 | LValue TDBase = Result.TDBase; | ||||||
5046 | const RecordDecl *KmpTaskTQTyRD = Result.KmpTaskTQTyRD; | ||||||
5047 | // Process list of dependences. | ||||||
5048 | Address DependenciesArray = Address::invalid(); | ||||||
5049 | llvm::Value *NumOfElements; | ||||||
5050 | std::tie(NumOfElements, DependenciesArray) = | ||||||
5051 | emitDependClause(CGF, Data.Dependences, Loc); | ||||||
5052 | |||||||
5053 | // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc() | ||||||
5054 | // libcall. | ||||||
5055 | // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid, | ||||||
5056 | // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, | ||||||
5057 | // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) if dependence | ||||||
5058 | // list is not empty | ||||||
5059 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | ||||||
5060 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | ||||||
5061 | llvm::Value *TaskArgs[] = { UpLoc, ThreadID, NewTask }; | ||||||
5062 | llvm::Value *DepTaskArgs[7]; | ||||||
5063 | if (!Data.Dependences.empty()) { | ||||||
5064 | DepTaskArgs[0] = UpLoc; | ||||||
5065 | DepTaskArgs[1] = ThreadID; | ||||||
5066 | DepTaskArgs[2] = NewTask; | ||||||
5067 | DepTaskArgs[3] = NumOfElements; | ||||||
5068 | DepTaskArgs[4] = DependenciesArray.getPointer(); | ||||||
5069 | DepTaskArgs[5] = CGF.Builder.getInt32(0); | ||||||
5070 | DepTaskArgs[6] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | ||||||
5071 | } | ||||||
5072 | auto &&ThenCodeGen = [this, &Data, TDBase, KmpTaskTQTyRD, &TaskArgs, | ||||||
5073 | &DepTaskArgs](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
5074 | if (!Data.Tied) { | ||||||
5075 | auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId); | ||||||
5076 | LValue PartIdLVal = CGF.EmitLValueForField(TDBase, *PartIdFI); | ||||||
5077 | CGF.EmitStoreOfScalar(CGF.Builder.getInt32(0), PartIdLVal); | ||||||
5078 | } | ||||||
5079 | if (!Data.Dependences.empty()) { | ||||||
5080 | CGF.EmitRuntimeCall( | ||||||
5081 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5082 | CGM.getModule(), OMPRTL___kmpc_omp_task_with_deps), | ||||||
5083 | DepTaskArgs); | ||||||
5084 | } else { | ||||||
5085 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5086 | CGM.getModule(), OMPRTL___kmpc_omp_task), | ||||||
5087 | TaskArgs); | ||||||
5088 | } | ||||||
5089 | // Check if parent region is untied and build return for untied task; | ||||||
5090 | if (auto *Region = | ||||||
5091 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | ||||||
5092 | Region->emitUntiedSwitch(CGF); | ||||||
5093 | }; | ||||||
5094 | |||||||
5095 | llvm::Value *DepWaitTaskArgs[6]; | ||||||
5096 | if (!Data.Dependences.empty()) { | ||||||
5097 | DepWaitTaskArgs[0] = UpLoc; | ||||||
5098 | DepWaitTaskArgs[1] = ThreadID; | ||||||
5099 | DepWaitTaskArgs[2] = NumOfElements; | ||||||
5100 | DepWaitTaskArgs[3] = DependenciesArray.getPointer(); | ||||||
5101 | DepWaitTaskArgs[4] = CGF.Builder.getInt32(0); | ||||||
5102 | DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | ||||||
5103 | } | ||||||
5104 | auto &M = CGM.getModule(); | ||||||
5105 | auto &&ElseCodeGen = [this, &M, &TaskArgs, ThreadID, NewTaskNewTaskTTy, | ||||||
5106 | TaskEntry, &Data, &DepWaitTaskArgs, | ||||||
5107 | Loc](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
5108 | CodeGenFunction::RunCleanupsScope LocalScope(CGF); | ||||||
5109 | // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid, | ||||||
5110 | // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 | ||||||
5111 | // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info | ||||||
5112 | // is specified. | ||||||
5113 | if (!Data.Dependences.empty()) | ||||||
5114 | CGF.EmitRuntimeCall( | ||||||
5115 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_omp_wait_deps), | ||||||
5116 | DepWaitTaskArgs); | ||||||
5117 | // Call proxy_task_entry(gtid, new_task); | ||||||
5118 | auto &&CodeGen = [TaskEntry, ThreadID, NewTaskNewTaskTTy, | ||||||
5119 | Loc](CodeGenFunction &CGF, PrePostActionTy &Action) { | ||||||
5120 | Action.Enter(CGF); | ||||||
5121 | llvm::Value *OutlinedFnArgs[] = {ThreadID, NewTaskNewTaskTTy}; | ||||||
5122 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskEntry, | ||||||
5123 | OutlinedFnArgs); | ||||||
5124 | }; | ||||||
5125 | |||||||
5126 | // Build void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid, | ||||||
5127 | // kmp_task_t *new_task); | ||||||
5128 | // Build void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid, | ||||||
5129 | // kmp_task_t *new_task); | ||||||
5130 | RegionCodeGenTy RCG(CodeGen); | ||||||
5131 | CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5132 | M, OMPRTL___kmpc_omp_task_begin_if0), | ||||||
5133 | TaskArgs, | ||||||
5134 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5135 | M, OMPRTL___kmpc_omp_task_complete_if0), | ||||||
5136 | TaskArgs); | ||||||
5137 | RCG.setAction(Action); | ||||||
5138 | RCG(CGF); | ||||||
5139 | }; | ||||||
5140 | |||||||
5141 | if (IfCond) { | ||||||
5142 | emitIfClause(CGF, IfCond, ThenCodeGen, ElseCodeGen); | ||||||
5143 | } else { | ||||||
5144 | RegionCodeGenTy ThenRCG(ThenCodeGen); | ||||||
5145 | ThenRCG(CGF); | ||||||
5146 | } | ||||||
5147 | } | ||||||
5148 | |||||||
5149 | void CGOpenMPRuntime::emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
5150 | const OMPLoopDirective &D, | ||||||
5151 | llvm::Function *TaskFunction, | ||||||
5152 | QualType SharedsTy, Address Shareds, | ||||||
5153 | const Expr *IfCond, | ||||||
5154 | const OMPTaskDataTy &Data) { | ||||||
5155 | if (!CGF.HaveInsertPoint()) | ||||||
5156 | return; | ||||||
5157 | TaskResultTy Result = | ||||||
5158 | emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data); | ||||||
5159 | // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc() | ||||||
5160 | // libcall. | ||||||
5161 | // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int | ||||||
5162 | // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int | ||||||
5163 | // sched, kmp_uint64 grainsize, void *task_dup); | ||||||
5164 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | ||||||
5165 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | ||||||
5166 | llvm::Value *IfVal; | ||||||
5167 | if (IfCond) { | ||||||
5168 | IfVal = CGF.Builder.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.IntTy, | ||||||
5169 | /*isSigned=*/true); | ||||||
5170 | } else { | ||||||
5171 | IfVal = llvm::ConstantInt::getSigned(CGF.IntTy, /*V=*/1); | ||||||
5172 | } | ||||||
5173 | |||||||
5174 | LValue LBLVal = CGF.EmitLValueForField( | ||||||
5175 | Result.TDBase, | ||||||
5176 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound)); | ||||||
5177 | const auto *LBVar = | ||||||
5178 | cast<VarDecl>(cast<DeclRefExpr>(D.getLowerBoundVariable())->getDecl()); | ||||||
5179 | CGF.EmitAnyExprToMem(LBVar->getInit(), LBLVal.getAddress(CGF), | ||||||
5180 | LBLVal.getQuals(), | ||||||
5181 | /*IsInitializer=*/true); | ||||||
5182 | LValue UBLVal = CGF.EmitLValueForField( | ||||||
5183 | Result.TDBase, | ||||||
5184 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound)); | ||||||
5185 | const auto *UBVar = | ||||||
5186 | cast<VarDecl>(cast<DeclRefExpr>(D.getUpperBoundVariable())->getDecl()); | ||||||
5187 | CGF.EmitAnyExprToMem(UBVar->getInit(), UBLVal.getAddress(CGF), | ||||||
5188 | UBLVal.getQuals(), | ||||||
5189 | /*IsInitializer=*/true); | ||||||
5190 | LValue StLVal = CGF.EmitLValueForField( | ||||||
5191 | Result.TDBase, | ||||||
5192 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTStride)); | ||||||
5193 | const auto *StVar = | ||||||
5194 | cast<VarDecl>(cast<DeclRefExpr>(D.getStrideVariable())->getDecl()); | ||||||
5195 | CGF.EmitAnyExprToMem(StVar->getInit(), StLVal.getAddress(CGF), | ||||||
5196 | StLVal.getQuals(), | ||||||
5197 | /*IsInitializer=*/true); | ||||||
5198 | // Store reductions address. | ||||||
5199 | LValue RedLVal = CGF.EmitLValueForField( | ||||||
5200 | Result.TDBase, | ||||||
5201 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTReductions)); | ||||||
5202 | if (Data.Reductions) { | ||||||
5203 | CGF.EmitStoreOfScalar(Data.Reductions, RedLVal); | ||||||
5204 | } else { | ||||||
5205 | CGF.EmitNullInitialization(RedLVal.getAddress(CGF), | ||||||
5206 | CGF.getContext().VoidPtrTy); | ||||||
5207 | } | ||||||
5208 | enum { NoSchedule = 0, Grainsize = 1, NumTasks = 2 }; | ||||||
5209 | llvm::Value *TaskArgs[] = { | ||||||
5210 | UpLoc, | ||||||
5211 | ThreadID, | ||||||
5212 | Result.NewTask, | ||||||
5213 | IfVal, | ||||||
5214 | LBLVal.getPointer(CGF), | ||||||
5215 | UBLVal.getPointer(CGF), | ||||||
5216 | CGF.EmitLoadOfScalar(StLVal, Loc), | ||||||
5217 | llvm::ConstantInt::getSigned( | ||||||
5218 | CGF.IntTy, 1), // Always 1 because taskgroup emitted by the compiler | ||||||
5219 | llvm::ConstantInt::getSigned( | ||||||
5220 | CGF.IntTy, Data.Schedule.getPointer() | ||||||
5221 | ? Data.Schedule.getInt() ? NumTasks : Grainsize | ||||||
5222 | : NoSchedule), | ||||||
5223 | Data.Schedule.getPointer() | ||||||
5224 | ? CGF.Builder.CreateIntCast(Data.Schedule.getPointer(), CGF.Int64Ty, | ||||||
5225 | /*isSigned=*/false) | ||||||
5226 | : llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/0), | ||||||
5227 | Result.TaskDupFn ? CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
5228 | Result.TaskDupFn, CGF.VoidPtrTy) | ||||||
5229 | : llvm::ConstantPointerNull::get(CGF.VoidPtrTy)}; | ||||||
5230 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5231 | CGM.getModule(), OMPRTL___kmpc_taskloop), | ||||||
5232 | TaskArgs); | ||||||
5233 | } | ||||||
5234 | |||||||
5235 | /// Emit reduction operation for each element of array (required for | ||||||
5236 | /// array sections) LHS op = RHS. | ||||||
5237 | /// \param Type Type of array. | ||||||
5238 | /// \param LHSVar Variable on the left side of the reduction operation | ||||||
5239 | /// (references element of array in original variable). | ||||||
5240 | /// \param RHSVar Variable on the right side of the reduction operation | ||||||
5241 | /// (references element of array in original variable). | ||||||
5242 | /// \param RedOpGen Generator of reduction operation with use of LHSVar and | ||||||
5243 | /// RHSVar. | ||||||
5244 | static void EmitOMPAggregateReduction( | ||||||
5245 | CodeGenFunction &CGF, QualType Type, const VarDecl *LHSVar, | ||||||
5246 | const VarDecl *RHSVar, | ||||||
5247 | const llvm::function_ref<void(CodeGenFunction &CGF, const Expr *, | ||||||
5248 | const Expr *, const Expr *)> &RedOpGen, | ||||||
5249 | const Expr *XExpr = nullptr, const Expr *EExpr = nullptr, | ||||||
5250 | const Expr *UpExpr = nullptr) { | ||||||
5251 | // Perform element-by-element initialization. | ||||||
5252 | QualType ElementTy; | ||||||
5253 | Address LHSAddr = CGF.GetAddrOfLocalVar(LHSVar); | ||||||
5254 | Address RHSAddr = CGF.GetAddrOfLocalVar(RHSVar); | ||||||
5255 | |||||||
5256 | // Drill down to the base element type on both arrays. | ||||||
5257 | const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe(); | ||||||
5258 | llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, LHSAddr); | ||||||
5259 | |||||||
5260 | llvm::Value *RHSBegin = RHSAddr.getPointer(); | ||||||
5261 | llvm::Value *LHSBegin = LHSAddr.getPointer(); | ||||||
5262 | // Cast from pointer to array type to pointer to single element. | ||||||
5263 | llvm::Value *LHSEnd = CGF.Builder.CreateGEP(LHSBegin, NumElements); | ||||||
5264 | // The basic structure here is a while-do loop. | ||||||
5265 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arraycpy.body"); | ||||||
5266 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arraycpy.done"); | ||||||
5267 | llvm::Value *IsEmpty = | ||||||
5268 | CGF.Builder.CreateICmpEQ(LHSBegin, LHSEnd, "omp.arraycpy.isempty"); | ||||||
5269 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | ||||||
5270 | |||||||
5271 | // Enter the loop body, making that address the current address. | ||||||
5272 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | ||||||
5273 | CGF.EmitBlock(BodyBB); | ||||||
5274 | |||||||
5275 | CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy); | ||||||
5276 | |||||||
5277 | llvm::PHINode *RHSElementPHI = CGF.Builder.CreatePHI( | ||||||
5278 | RHSBegin->getType(), 2, "omp.arraycpy.srcElementPast"); | ||||||
5279 | RHSElementPHI->addIncoming(RHSBegin, EntryBB); | ||||||
5280 | Address RHSElementCurrent = | ||||||
5281 | Address(RHSElementPHI, | ||||||
5282 | RHSAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | ||||||
5283 | |||||||
5284 | llvm::PHINode *LHSElementPHI = CGF.Builder.CreatePHI( | ||||||
5285 | LHSBegin->getType(), 2, "omp.arraycpy.destElementPast"); | ||||||
5286 | LHSElementPHI->addIncoming(LHSBegin, EntryBB); | ||||||
5287 | Address LHSElementCurrent = | ||||||
5288 | Address(LHSElementPHI, | ||||||
5289 | LHSAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | ||||||
5290 | |||||||
5291 | // Emit copy. | ||||||
5292 | CodeGenFunction::OMPPrivateScope Scope(CGF); | ||||||
5293 | Scope.addPrivate(LHSVar, [=]() { return LHSElementCurrent; }); | ||||||
5294 | Scope.addPrivate(RHSVar, [=]() { return RHSElementCurrent; }); | ||||||
5295 | Scope.Privatize(); | ||||||
5296 | RedOpGen(CGF, XExpr, EExpr, UpExpr); | ||||||
5297 | Scope.ForceCleanup(); | ||||||
5298 | |||||||
5299 | // Shift the address forward by one element. | ||||||
5300 | llvm::Value *LHSElementNext = CGF.Builder.CreateConstGEP1_32( | ||||||
5301 | LHSElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element"); | ||||||
5302 | llvm::Value *RHSElementNext = CGF.Builder.CreateConstGEP1_32( | ||||||
5303 | RHSElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element"); | ||||||
5304 | // Check whether we've reached the end. | ||||||
5305 | llvm::Value *Done = | ||||||
5306 | CGF.Builder.CreateICmpEQ(LHSElementNext, LHSEnd, "omp.arraycpy.done"); | ||||||
5307 | CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB); | ||||||
5308 | LHSElementPHI->addIncoming(LHSElementNext, CGF.Builder.GetInsertBlock()); | ||||||
5309 | RHSElementPHI->addIncoming(RHSElementNext, CGF.Builder.GetInsertBlock()); | ||||||
5310 | |||||||
5311 | // Done. | ||||||
5312 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | ||||||
5313 | } | ||||||
5314 | |||||||
5315 | /// Emit reduction combiner. If the combiner is a simple expression emit it as | ||||||
5316 | /// is, otherwise consider it as combiner of UDR decl and emit it as a call of | ||||||
5317 | /// UDR combiner function. | ||||||
5318 | static void emitReductionCombiner(CodeGenFunction &CGF, | ||||||
5319 | const Expr *ReductionOp) { | ||||||
5320 | if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) | ||||||
5321 | if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee())) | ||||||
5322 | if (const auto *DRE = | ||||||
5323 | dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts())) | ||||||
5324 | if (const auto *DRD = | ||||||
5325 | dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) { | ||||||
5326 | std::pair<llvm::Function *, llvm::Function *> Reduction = | ||||||
5327 | CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD); | ||||||
5328 | RValue Func = RValue::get(Reduction.first); | ||||||
5329 | CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func); | ||||||
5330 | CGF.EmitIgnoredExpr(ReductionOp); | ||||||
5331 | return; | ||||||
5332 | } | ||||||
5333 | CGF.EmitIgnoredExpr(ReductionOp); | ||||||
5334 | } | ||||||
5335 | |||||||
5336 | llvm::Function *CGOpenMPRuntime::emitReductionFunction( | ||||||
5337 | SourceLocation Loc, llvm::Type *ArgsType, ArrayRef<const Expr *> Privates, | ||||||
5338 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, | ||||||
5339 | ArrayRef<const Expr *> ReductionOps) { | ||||||
5340 | ASTContext &C = CGM.getContext(); | ||||||
5341 | |||||||
5342 | // void reduction_func(void *LHSArg, void *RHSArg); | ||||||
5343 | FunctionArgList Args; | ||||||
5344 | ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
5345 | ImplicitParamDecl::Other); | ||||||
5346 | ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
5347 | ImplicitParamDecl::Other); | ||||||
5348 | Args.push_back(&LHSArg); | ||||||
5349 | Args.push_back(&RHSArg); | ||||||
5350 | const auto &CGFI = | ||||||
5351 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
5352 | std::string Name = getName({"omp", "reduction", "reduction_func"}); | ||||||
5353 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | ||||||
5354 | llvm::GlobalValue::InternalLinkage, Name, | ||||||
5355 | &CGM.getModule()); | ||||||
5356 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | ||||||
5357 | Fn->setDoesNotRecurse(); | ||||||
5358 | CodeGenFunction CGF(CGM); | ||||||
5359 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | ||||||
5360 | |||||||
5361 | // Dst = (void*[n])(LHSArg); | ||||||
5362 | // Src = (void*[n])(RHSArg); | ||||||
5363 | Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
5364 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)), | ||||||
5365 | ArgsType), CGF.getPointerAlign()); | ||||||
5366 | Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
5367 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)), | ||||||
5368 | ArgsType), CGF.getPointerAlign()); | ||||||
5369 | |||||||
5370 | // ... | ||||||
5371 | // *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]); | ||||||
5372 | // ... | ||||||
5373 | CodeGenFunction::OMPPrivateScope Scope(CGF); | ||||||
5374 | auto IPriv = Privates.begin(); | ||||||
5375 | unsigned Idx = 0; | ||||||
5376 | for (unsigned I = 0, E = ReductionOps.size(); I < E; ++I, ++IPriv, ++Idx) { | ||||||
5377 | const auto *RHSVar = | ||||||
5378 | cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl()); | ||||||
5379 | Scope.addPrivate(RHSVar, [&CGF, RHS, Idx, RHSVar]() { | ||||||
5380 | return emitAddrOfVarFromArray(CGF, RHS, Idx, RHSVar); | ||||||
5381 | }); | ||||||
5382 | const auto *LHSVar = | ||||||
5383 | cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl()); | ||||||
5384 | Scope.addPrivate(LHSVar, [&CGF, LHS, Idx, LHSVar]() { | ||||||
5385 | return emitAddrOfVarFromArray(CGF, LHS, Idx, LHSVar); | ||||||
5386 | }); | ||||||
5387 | QualType PrivTy = (*IPriv)->getType(); | ||||||
5388 | if (PrivTy->isVariablyModifiedType()) { | ||||||
5389 | // Get array size and emit VLA type. | ||||||
5390 | ++Idx; | ||||||
5391 | Address Elem = CGF.Builder.CreateConstArrayGEP(LHS, Idx); | ||||||
5392 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Elem); | ||||||
5393 | const VariableArrayType *VLA = | ||||||
5394 | CGF.getContext().getAsVariableArrayType(PrivTy); | ||||||
5395 | const auto *OVE = cast<OpaqueValueExpr>(VLA->getSizeExpr()); | ||||||
5396 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | ||||||
5397 | CGF, OVE, RValue::get(CGF.Builder.CreatePtrToInt(Ptr, CGF.SizeTy))); | ||||||
5398 | CGF.EmitVariablyModifiedType(PrivTy); | ||||||
5399 | } | ||||||
5400 | } | ||||||
5401 | Scope.Privatize(); | ||||||
5402 | IPriv = Privates.begin(); | ||||||
5403 | auto ILHS = LHSExprs.begin(); | ||||||
5404 | auto IRHS = RHSExprs.begin(); | ||||||
5405 | for (const Expr *E : ReductionOps) { | ||||||
5406 | if ((*IPriv)->getType()->isArrayType()) { | ||||||
5407 | // Emit reduction for array section. | ||||||
5408 | const auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | ||||||
5409 | const auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | ||||||
5410 | EmitOMPAggregateReduction( | ||||||
5411 | CGF, (*IPriv)->getType(), LHSVar, RHSVar, | ||||||
5412 | [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) { | ||||||
5413 | emitReductionCombiner(CGF, E); | ||||||
5414 | }); | ||||||
5415 | } else { | ||||||
5416 | // Emit reduction for array subscript or single variable. | ||||||
5417 | emitReductionCombiner(CGF, E); | ||||||
5418 | } | ||||||
5419 | ++IPriv; | ||||||
5420 | ++ILHS; | ||||||
5421 | ++IRHS; | ||||||
5422 | } | ||||||
5423 | Scope.ForceCleanup(); | ||||||
5424 | CGF.FinishFunction(); | ||||||
5425 | return Fn; | ||||||
5426 | } | ||||||
5427 | |||||||
5428 | void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF, | ||||||
5429 | const Expr *ReductionOp, | ||||||
5430 | const Expr *PrivateRef, | ||||||
5431 | const DeclRefExpr *LHS, | ||||||
5432 | const DeclRefExpr *RHS) { | ||||||
5433 | if (PrivateRef->getType()->isArrayType()) { | ||||||
5434 | // Emit reduction for array section. | ||||||
5435 | const auto *LHSVar = cast<VarDecl>(LHS->getDecl()); | ||||||
5436 | const auto *RHSVar = cast<VarDecl>(RHS->getDecl()); | ||||||
5437 | EmitOMPAggregateReduction( | ||||||
5438 | CGF, PrivateRef->getType(), LHSVar, RHSVar, | ||||||
5439 | [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) { | ||||||
5440 | emitReductionCombiner(CGF, ReductionOp); | ||||||
5441 | }); | ||||||
5442 | } else { | ||||||
5443 | // Emit reduction for array subscript or single variable. | ||||||
5444 | emitReductionCombiner(CGF, ReductionOp); | ||||||
5445 | } | ||||||
5446 | } | ||||||
5447 | |||||||
5448 | void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
5449 | ArrayRef<const Expr *> Privates, | ||||||
5450 | ArrayRef<const Expr *> LHSExprs, | ||||||
5451 | ArrayRef<const Expr *> RHSExprs, | ||||||
5452 | ArrayRef<const Expr *> ReductionOps, | ||||||
5453 | ReductionOptionsTy Options) { | ||||||
5454 | if (!CGF.HaveInsertPoint()) | ||||||
5455 | return; | ||||||
5456 | |||||||
5457 | bool WithNowait = Options.WithNowait; | ||||||
5458 | bool SimpleReduction = Options.SimpleReduction; | ||||||
5459 | |||||||
5460 | // Next code should be emitted for reduction: | ||||||
5461 | // | ||||||
5462 | // static kmp_critical_name lock = { 0 }; | ||||||
5463 | // | ||||||
5464 | // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { | ||||||
5465 | // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); | ||||||
5466 | // ... | ||||||
5467 | // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], | ||||||
5468 | // *(Type<n>-1*)rhs[<n>-1]); | ||||||
5469 | // } | ||||||
5470 | // | ||||||
5471 | // ... | ||||||
5472 | // void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; | ||||||
5473 | // switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), | ||||||
5474 | // RedList, reduce_func, &<lock>)) { | ||||||
5475 | // case 1: | ||||||
5476 | // ... | ||||||
5477 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | ||||||
5478 | // ... | ||||||
5479 | // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | ||||||
5480 | // break; | ||||||
5481 | // case 2: | ||||||
5482 | // ... | ||||||
5483 | // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); | ||||||
5484 | // ... | ||||||
5485 | // [__kmpc_end_reduce(<loc>, <gtid>, &<lock>);] | ||||||
5486 | // break; | ||||||
5487 | // default:; | ||||||
5488 | // } | ||||||
5489 | // | ||||||
5490 | // if SimpleReduction is true, only the next code is generated: | ||||||
5491 | // ... | ||||||
5492 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | ||||||
5493 | // ... | ||||||
5494 | |||||||
5495 | ASTContext &C = CGM.getContext(); | ||||||
5496 | |||||||
5497 | if (SimpleReduction) { | ||||||
5498 | CodeGenFunction::RunCleanupsScope Scope(CGF); | ||||||
5499 | auto IPriv = Privates.begin(); | ||||||
5500 | auto ILHS = LHSExprs.begin(); | ||||||
5501 | auto IRHS = RHSExprs.begin(); | ||||||
5502 | for (const Expr *E : ReductionOps) { | ||||||
5503 | emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | ||||||
5504 | cast<DeclRefExpr>(*IRHS)); | ||||||
5505 | ++IPriv; | ||||||
5506 | ++ILHS; | ||||||
5507 | ++IRHS; | ||||||
5508 | } | ||||||
5509 | return; | ||||||
5510 | } | ||||||
5511 | |||||||
5512 | // 1. Build a list of reduction variables. | ||||||
5513 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; | ||||||
5514 | auto Size = RHSExprs.size(); | ||||||
5515 | for (const Expr *E : Privates) { | ||||||
5516 | if (E->getType()->isVariablyModifiedType()) | ||||||
5517 | // Reserve place for array size. | ||||||
5518 | ++Size; | ||||||
5519 | } | ||||||
5520 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); | ||||||
5521 | QualType ReductionArrayTy = | ||||||
5522 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, | ||||||
5523 | /*IndexTypeQuals=*/0); | ||||||
5524 | Address ReductionList = | ||||||
5525 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); | ||||||
5526 | auto IPriv = Privates.begin(); | ||||||
5527 | unsigned Idx = 0; | ||||||
5528 | for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { | ||||||
5529 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | ||||||
5530 | CGF.Builder.CreateStore( | ||||||
5531 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
5532 | CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), | ||||||
5533 | Elem); | ||||||
5534 | if ((*IPriv)->getType()->isVariablyModifiedType()) { | ||||||
5535 | // Store array size. | ||||||
5536 | ++Idx; | ||||||
5537 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | ||||||
5538 | llvm::Value *Size = CGF.Builder.CreateIntCast( | ||||||
5539 | CGF.getVLASize( | ||||||
5540 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) | ||||||
5541 | .NumElts, | ||||||
5542 | CGF.SizeTy, /*isSigned=*/false); | ||||||
5543 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), | ||||||
5544 | Elem); | ||||||
5545 | } | ||||||
5546 | } | ||||||
5547 | |||||||
5548 | // 2. Emit reduce_func(). | ||||||
5549 | llvm::Function *ReductionFn = emitReductionFunction( | ||||||
5550 | Loc, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates, | ||||||
5551 | LHSExprs, RHSExprs, ReductionOps); | ||||||
5552 | |||||||
5553 | // 3. Create static kmp_critical_name lock = { 0 }; | ||||||
5554 | std::string Name = getName({"reduction"}); | ||||||
5555 | llvm::Value *Lock = getCriticalRegionLock(Name); | ||||||
5556 | |||||||
5557 | // 4. Build res = __kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), | ||||||
5558 | // RedList, reduce_func, &<lock>); | ||||||
5559 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE); | ||||||
5560 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | ||||||
5561 | llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); | ||||||
5562 | llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
5563 | ReductionList.getPointer(), CGF.VoidPtrTy); | ||||||
5564 | llvm::Value *Args[] = { | ||||||
5565 | IdentTLoc, // ident_t *<loc> | ||||||
5566 | ThreadId, // i32 <gtid> | ||||||
5567 | CGF.Builder.getInt32(RHSExprs.size()), // i32 <n> | ||||||
5568 | ReductionArrayTySize, // size_type sizeof(RedList) | ||||||
5569 | RL, // void *RedList | ||||||
5570 | ReductionFn, // void (*) (void *, void *) <reduce_func> | ||||||
5571 | Lock // kmp_critical_name *&<lock> | ||||||
5572 | }; | ||||||
5573 | llvm::Value *Res = CGF.EmitRuntimeCall( | ||||||
5574 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5575 | CGM.getModule(), | ||||||
5576 | WithNowait ? OMPRTL___kmpc_reduce_nowait : OMPRTL___kmpc_reduce), | ||||||
5577 | Args); | ||||||
5578 | |||||||
5579 | // 5. Build switch(res) | ||||||
5580 | llvm::BasicBlock *DefaultBB = CGF.createBasicBlock(".omp.reduction.default"); | ||||||
5581 | llvm::SwitchInst *SwInst = | ||||||
5582 | CGF.Builder.CreateSwitch(Res, DefaultBB, /*NumCases=*/2); | ||||||
5583 | |||||||
5584 | // 6. Build case 1: | ||||||
5585 | // ... | ||||||
5586 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | ||||||
5587 | // ... | ||||||
5588 | // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | ||||||
5589 | // break; | ||||||
5590 | llvm::BasicBlock *Case1BB = CGF.createBasicBlock(".omp.reduction.case1"); | ||||||
5591 | SwInst->addCase(CGF.Builder.getInt32(1), Case1BB); | ||||||
5592 | CGF.EmitBlock(Case1BB); | ||||||
5593 | |||||||
5594 | // Add emission of __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | ||||||
5595 | llvm::Value *EndArgs[] = { | ||||||
5596 | IdentTLoc, // ident_t *<loc> | ||||||
5597 | ThreadId, // i32 <gtid> | ||||||
5598 | Lock // kmp_critical_name *&<lock> | ||||||
5599 | }; | ||||||
5600 | auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps]( | ||||||
5601 | CodeGenFunction &CGF, PrePostActionTy &Action) { | ||||||
5602 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | ||||||
5603 | auto IPriv = Privates.begin(); | ||||||
5604 | auto ILHS = LHSExprs.begin(); | ||||||
5605 | auto IRHS = RHSExprs.begin(); | ||||||
5606 | for (const Expr *E : ReductionOps) { | ||||||
5607 | RT.emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | ||||||
5608 | cast<DeclRefExpr>(*IRHS)); | ||||||
5609 | ++IPriv; | ||||||
5610 | ++ILHS; | ||||||
5611 | ++IRHS; | ||||||
5612 | } | ||||||
5613 | }; | ||||||
5614 | RegionCodeGenTy RCG(CodeGen); | ||||||
5615 | CommonActionTy Action( | ||||||
5616 | nullptr, llvm::None, | ||||||
5617 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5618 | CGM.getModule(), WithNowait ? OMPRTL___kmpc_end_reduce_nowait | ||||||
5619 | : OMPRTL___kmpc_end_reduce), | ||||||
5620 | EndArgs); | ||||||
5621 | RCG.setAction(Action); | ||||||
5622 | RCG(CGF); | ||||||
5623 | |||||||
5624 | CGF.EmitBranch(DefaultBB); | ||||||
5625 | |||||||
5626 | // 7. Build case 2: | ||||||
5627 | // ... | ||||||
5628 | // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); | ||||||
5629 | // ... | ||||||
5630 | // break; | ||||||
5631 | llvm::BasicBlock *Case2BB = CGF.createBasicBlock(".omp.reduction.case2"); | ||||||
5632 | SwInst->addCase(CGF.Builder.getInt32(2), Case2BB); | ||||||
5633 | CGF.EmitBlock(Case2BB); | ||||||
5634 | |||||||
5635 | auto &&AtomicCodeGen = [Loc, Privates, LHSExprs, RHSExprs, ReductionOps]( | ||||||
5636 | CodeGenFunction &CGF, PrePostActionTy &Action) { | ||||||
5637 | auto ILHS = LHSExprs.begin(); | ||||||
5638 | auto IRHS = RHSExprs.begin(); | ||||||
5639 | auto IPriv = Privates.begin(); | ||||||
5640 | for (const Expr *E : ReductionOps) { | ||||||
5641 | const Expr *XExpr = nullptr; | ||||||
5642 | const Expr *EExpr = nullptr; | ||||||
5643 | const Expr *UpExpr = nullptr; | ||||||
5644 | BinaryOperatorKind BO = BO_Comma; | ||||||
5645 | if (const auto *BO = dyn_cast<BinaryOperator>(E)) { | ||||||
5646 | if (BO->getOpcode() == BO_Assign) { | ||||||
5647 | XExpr = BO->getLHS(); | ||||||
5648 | UpExpr = BO->getRHS(); | ||||||
5649 | } | ||||||
5650 | } | ||||||
5651 | // Try to emit update expression as a simple atomic. | ||||||
5652 | const Expr *RHSExpr = UpExpr; | ||||||
5653 | if (RHSExpr) { | ||||||
5654 | // Analyze RHS part of the whole expression. | ||||||
5655 | if (const auto *ACO = dyn_cast<AbstractConditionalOperator>( | ||||||
5656 | RHSExpr->IgnoreParenImpCasts())) { | ||||||
5657 | // If this is a conditional operator, analyze its condition for | ||||||
5658 | // min/max reduction operator. | ||||||
5659 | RHSExpr = ACO->getCond(); | ||||||
5660 | } | ||||||
5661 | if (const auto *BORHS = | ||||||
5662 | dyn_cast<BinaryOperator>(RHSExpr->IgnoreParenImpCasts())) { | ||||||
5663 | EExpr = BORHS->getRHS(); | ||||||
5664 | BO = BORHS->getOpcode(); | ||||||
5665 | } | ||||||
5666 | } | ||||||
5667 | if (XExpr) { | ||||||
5668 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | ||||||
5669 | auto &&AtomicRedGen = [BO, VD, | ||||||
5670 | Loc](CodeGenFunction &CGF, const Expr *XExpr, | ||||||
5671 | const Expr *EExpr, const Expr *UpExpr) { | ||||||
5672 | LValue X = CGF.EmitLValue(XExpr); | ||||||
5673 | RValue E; | ||||||
5674 | if (EExpr) | ||||||
5675 | E = CGF.EmitAnyExpr(EExpr); | ||||||
5676 | CGF.EmitOMPAtomicSimpleUpdateExpr( | ||||||
5677 | X, E, BO, /*IsXLHSInRHSPart=*/true, | ||||||
5678 | llvm::AtomicOrdering::Monotonic, Loc, | ||||||
5679 | [&CGF, UpExpr, VD, Loc](RValue XRValue) { | ||||||
5680 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | ||||||
5681 | PrivateScope.addPrivate( | ||||||
5682 | VD, [&CGF, VD, XRValue, Loc]() { | ||||||
5683 | Address LHSTemp = CGF.CreateMemTemp(VD->getType()); | ||||||
5684 | CGF.emitOMPSimpleStore( | ||||||
5685 | CGF.MakeAddrLValue(LHSTemp, VD->getType()), XRValue, | ||||||
5686 | VD->getType().getNonReferenceType(), Loc); | ||||||
5687 | return LHSTemp; | ||||||
5688 | }); | ||||||
5689 | (void)PrivateScope.Privatize(); | ||||||
5690 | return CGF.EmitAnyExpr(UpExpr); | ||||||
5691 | }); | ||||||
5692 | }; | ||||||
5693 | if ((*IPriv)->getType()->isArrayType()) { | ||||||
5694 | // Emit atomic reduction for array section. | ||||||
5695 | const auto *RHSVar = | ||||||
5696 | cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | ||||||
5697 | EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), VD, RHSVar, | ||||||
5698 | AtomicRedGen, XExpr, EExpr, UpExpr); | ||||||
5699 | } else { | ||||||
5700 | // Emit atomic reduction for array subscript or single variable. | ||||||
5701 | AtomicRedGen(CGF, XExpr, EExpr, UpExpr); | ||||||
5702 | } | ||||||
5703 | } else { | ||||||
5704 | // Emit as a critical region. | ||||||
5705 | auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *, | ||||||
5706 | const Expr *, const Expr *) { | ||||||
5707 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | ||||||
5708 | std::string Name = RT.getName({"atomic_reduction"}); | ||||||
5709 | RT.emitCriticalRegion( | ||||||
5710 | CGF, Name, | ||||||
5711 | [=](CodeGenFunction &CGF, PrePostActionTy &Action) { | ||||||
5712 | Action.Enter(CGF); | ||||||
5713 | emitReductionCombiner(CGF, E); | ||||||
5714 | }, | ||||||
5715 | Loc); | ||||||
5716 | }; | ||||||
5717 | if ((*IPriv)->getType()->isArrayType()) { | ||||||
5718 | const auto *LHSVar = | ||||||
5719 | cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | ||||||
5720 | const auto *RHSVar = | ||||||
5721 | cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | ||||||
5722 | EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), LHSVar, RHSVar, | ||||||
5723 | CritRedGen); | ||||||
5724 | } else { | ||||||
5725 | CritRedGen(CGF, nullptr, nullptr, nullptr); | ||||||
5726 | } | ||||||
5727 | } | ||||||
5728 | ++ILHS; | ||||||
5729 | ++IRHS; | ||||||
5730 | ++IPriv; | ||||||
5731 | } | ||||||
5732 | }; | ||||||
5733 | RegionCodeGenTy AtomicRCG(AtomicCodeGen); | ||||||
5734 | if (!WithNowait) { | ||||||
5735 | // Add emission of __kmpc_end_reduce(<loc>, <gtid>, &<lock>); | ||||||
5736 | llvm::Value *EndArgs[] = { | ||||||
5737 | IdentTLoc, // ident_t *<loc> | ||||||
5738 | ThreadId, // i32 <gtid> | ||||||
5739 | Lock // kmp_critical_name *&<lock> | ||||||
5740 | }; | ||||||
5741 | CommonActionTy Action(nullptr, llvm::None, | ||||||
5742 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
5743 | CGM.getModule(), OMPRTL___kmpc_end_reduce), | ||||||
5744 | EndArgs); | ||||||
5745 | AtomicRCG.setAction(Action); | ||||||
5746 | AtomicRCG(CGF); | ||||||
5747 | } else { | ||||||
5748 | AtomicRCG(CGF); | ||||||
5749 | } | ||||||
5750 | |||||||
5751 | CGF.EmitBranch(DefaultBB); | ||||||
5752 | CGF.EmitBlock(DefaultBB, /*IsFinished=*/true); | ||||||
5753 | } | ||||||
5754 | |||||||
5755 | /// Generates unique name for artificial threadprivate variables. | ||||||
5756 | /// Format is: <Prefix> "." <Decl_mangled_name> "_" "<Decl_start_loc_raw_enc>" | ||||||
5757 | static std::string generateUniqueName(CodeGenModule &CGM, StringRef Prefix, | ||||||
5758 | const Expr *Ref) { | ||||||
5759 | SmallString<256> Buffer; | ||||||
5760 | llvm::raw_svector_ostream Out(Buffer); | ||||||
5761 | const clang::DeclRefExpr *DE; | ||||||
5762 | const VarDecl *D = ::getBaseDecl(Ref, DE); | ||||||
5763 | if (!D) | ||||||
5764 | D = cast<VarDecl>(cast<DeclRefExpr>(Ref)->getDecl()); | ||||||
5765 | D = D->getCanonicalDecl(); | ||||||
5766 | std::string Name = CGM.getOpenMPRuntime().getName( | ||||||
5767 | {D->isLocalVarDeclOrParm() ? D->getName() : CGM.getMangledName(D)}); | ||||||
5768 | Out << Prefix << Name << "_" | ||||||
5769 | << D->getCanonicalDecl()->getBeginLoc().getRawEncoding(); | ||||||
5770 | return std::string(Out.str()); | ||||||
5771 | } | ||||||
5772 | |||||||
5773 | /// Emits reduction initializer function: | ||||||
5774 | /// \code | ||||||
5775 | /// void @.red_init(void* %arg, void* %orig) { | ||||||
5776 | /// %0 = bitcast void* %arg to <type>* | ||||||
5777 | /// store <type> <init>, <type>* %0 | ||||||
5778 | /// ret void | ||||||
5779 | /// } | ||||||
5780 | /// \endcode | ||||||
5781 | static llvm::Value *emitReduceInitFunction(CodeGenModule &CGM, | ||||||
5782 | SourceLocation Loc, | ||||||
5783 | ReductionCodeGen &RCG, unsigned N) { | ||||||
5784 | ASTContext &C = CGM.getContext(); | ||||||
5785 | QualType VoidPtrTy = C.VoidPtrTy; | ||||||
5786 | VoidPtrTy.addRestrict(); | ||||||
5787 | FunctionArgList Args; | ||||||
5788 | ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, VoidPtrTy, | ||||||
5789 | ImplicitParamDecl::Other); | ||||||
5790 | ImplicitParamDecl ParamOrig(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, VoidPtrTy, | ||||||
5791 | ImplicitParamDecl::Other); | ||||||
5792 | Args.emplace_back(&Param); | ||||||
5793 | Args.emplace_back(&ParamOrig); | ||||||
5794 | const auto &FnInfo = | ||||||
5795 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
5796 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | ||||||
5797 | std::string Name = CGM.getOpenMPRuntime().getName({"red_init", ""}); | ||||||
5798 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | ||||||
5799 | Name, &CGM.getModule()); | ||||||
5800 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | ||||||
5801 | Fn->setDoesNotRecurse(); | ||||||
5802 | CodeGenFunction CGF(CGM); | ||||||
5803 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | ||||||
5804 | Address PrivateAddr = CGF.EmitLoadOfPointer( | ||||||
5805 | CGF.GetAddrOfLocalVar(&Param), | ||||||
5806 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
5807 | llvm::Value *Size = nullptr; | ||||||
5808 | // If the size of the reduction item is non-constant, load it from global | ||||||
5809 | // threadprivate variable. | ||||||
5810 | if (RCG.getSizes(N).second) { | ||||||
5811 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | ||||||
5812 | CGF, CGM.getContext().getSizeType(), | ||||||
5813 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | ||||||
5814 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | ||||||
5815 | CGM.getContext().getSizeType(), Loc); | ||||||
5816 | } | ||||||
5817 | RCG.emitAggregateType(CGF, N, Size); | ||||||
5818 | LValue OrigLVal; | ||||||
5819 | // If initializer uses initializer from declare reduction construct, emit a | ||||||
5820 | // pointer to the address of the original reduction item (reuired by reduction | ||||||
5821 | // initializer) | ||||||
5822 | if (RCG.usesReductionInitializer(N)) { | ||||||
5823 | Address SharedAddr = CGF.GetAddrOfLocalVar(&ParamOrig); | ||||||
5824 | SharedAddr = CGF.EmitLoadOfPointer( | ||||||
5825 | SharedAddr, | ||||||
5826 | CGM.getContext().VoidPtrTy.castAs<PointerType>()->getTypePtr()); | ||||||
5827 | OrigLVal = CGF.MakeAddrLValue(SharedAddr, CGM.getContext().VoidPtrTy); | ||||||
5828 | } else { | ||||||
5829 | OrigLVal = CGF.MakeNaturalAlignAddrLValue( | ||||||
5830 | llvm::ConstantPointerNull::get(CGM.VoidPtrTy), | ||||||
5831 | CGM.getContext().VoidPtrTy); | ||||||
5832 | } | ||||||
5833 | // Emit the initializer: | ||||||
5834 | // %0 = bitcast void* %arg to <type>* | ||||||
5835 | // store <type> <init>, <type>* %0 | ||||||
5836 | RCG.emitInitialization(CGF, N, PrivateAddr, OrigLVal, | ||||||
5837 | [](CodeGenFunction &) { return false; }); | ||||||
5838 | CGF.FinishFunction(); | ||||||
5839 | return Fn; | ||||||
5840 | } | ||||||
5841 | |||||||
5842 | /// Emits reduction combiner function: | ||||||
5843 | /// \code | ||||||
5844 | /// void @.red_comb(void* %arg0, void* %arg1) { | ||||||
5845 | /// %lhs = bitcast void* %arg0 to <type>* | ||||||
5846 | /// %rhs = bitcast void* %arg1 to <type>* | ||||||
5847 | /// %2 = <ReductionOp>(<type>* %lhs, <type>* %rhs) | ||||||
5848 | /// store <type> %2, <type>* %lhs | ||||||
5849 | /// ret void | ||||||
5850 | /// } | ||||||
5851 | /// \endcode | ||||||
5852 | static llvm::Value *emitReduceCombFunction(CodeGenModule &CGM, | ||||||
5853 | SourceLocation Loc, | ||||||
5854 | ReductionCodeGen &RCG, unsigned N, | ||||||
5855 | const Expr *ReductionOp, | ||||||
5856 | const Expr *LHS, const Expr *RHS, | ||||||
5857 | const Expr *PrivateRef) { | ||||||
5858 | ASTContext &C = CGM.getContext(); | ||||||
5859 | const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(LHS)->getDecl()); | ||||||
5860 | const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(RHS)->getDecl()); | ||||||
5861 | FunctionArgList Args; | ||||||
5862 | ImplicitParamDecl ParamInOut(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
5863 | C.VoidPtrTy, ImplicitParamDecl::Other); | ||||||
5864 | ImplicitParamDecl ParamIn(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
5865 | ImplicitParamDecl::Other); | ||||||
5866 | Args.emplace_back(&ParamInOut); | ||||||
5867 | Args.emplace_back(&ParamIn); | ||||||
5868 | const auto &FnInfo = | ||||||
5869 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
5870 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | ||||||
5871 | std::string Name = CGM.getOpenMPRuntime().getName({"red_comb", ""}); | ||||||
5872 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | ||||||
5873 | Name, &CGM.getModule()); | ||||||
5874 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | ||||||
5875 | Fn->setDoesNotRecurse(); | ||||||
5876 | CodeGenFunction CGF(CGM); | ||||||
5877 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | ||||||
5878 | llvm::Value *Size = nullptr; | ||||||
5879 | // If the size of the reduction item is non-constant, load it from global | ||||||
5880 | // threadprivate variable. | ||||||
5881 | if (RCG.getSizes(N).second) { | ||||||
5882 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | ||||||
5883 | CGF, CGM.getContext().getSizeType(), | ||||||
5884 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | ||||||
5885 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | ||||||
5886 | CGM.getContext().getSizeType(), Loc); | ||||||
5887 | } | ||||||
5888 | RCG.emitAggregateType(CGF, N, Size); | ||||||
5889 | // Remap lhs and rhs variables to the addresses of the function arguments. | ||||||
5890 | // %lhs = bitcast void* %arg0 to <type>* | ||||||
5891 | // %rhs = bitcast void* %arg1 to <type>* | ||||||
5892 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | ||||||
5893 | PrivateScope.addPrivate(LHSVD, [&C, &CGF, &ParamInOut, LHSVD]() { | ||||||
5894 | // Pull out the pointer to the variable. | ||||||
5895 | Address PtrAddr = CGF.EmitLoadOfPointer( | ||||||
5896 | CGF.GetAddrOfLocalVar(&ParamInOut), | ||||||
5897 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
5898 | return CGF.Builder.CreateElementBitCast( | ||||||
5899 | PtrAddr, CGF.ConvertTypeForMem(LHSVD->getType())); | ||||||
5900 | }); | ||||||
5901 | PrivateScope.addPrivate(RHSVD, [&C, &CGF, &ParamIn, RHSVD]() { | ||||||
5902 | // Pull out the pointer to the variable. | ||||||
5903 | Address PtrAddr = CGF.EmitLoadOfPointer( | ||||||
5904 | CGF.GetAddrOfLocalVar(&ParamIn), | ||||||
5905 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
5906 | return CGF.Builder.CreateElementBitCast( | ||||||
5907 | PtrAddr, CGF.ConvertTypeForMem(RHSVD->getType())); | ||||||
5908 | }); | ||||||
5909 | PrivateScope.Privatize(); | ||||||
5910 | // Emit the combiner body: | ||||||
5911 | // %2 = <ReductionOp>(<type> *%lhs, <type> *%rhs) | ||||||
5912 | // store <type> %2, <type>* %lhs | ||||||
5913 | CGM.getOpenMPRuntime().emitSingleReductionCombiner( | ||||||
5914 | CGF, ReductionOp, PrivateRef, cast<DeclRefExpr>(LHS), | ||||||
5915 | cast<DeclRefExpr>(RHS)); | ||||||
5916 | CGF.FinishFunction(); | ||||||
5917 | return Fn; | ||||||
5918 | } | ||||||
5919 | |||||||
5920 | /// Emits reduction finalizer function: | ||||||
5921 | /// \code | ||||||
5922 | /// void @.red_fini(void* %arg) { | ||||||
5923 | /// %0 = bitcast void* %arg to <type>* | ||||||
5924 | /// <destroy>(<type>* %0) | ||||||
5925 | /// ret void | ||||||
5926 | /// } | ||||||
5927 | /// \endcode | ||||||
5928 | static llvm::Value *emitReduceFiniFunction(CodeGenModule &CGM, | ||||||
5929 | SourceLocation Loc, | ||||||
5930 | ReductionCodeGen &RCG, unsigned N) { | ||||||
5931 | if (!RCG.needCleanups(N)) | ||||||
5932 | return nullptr; | ||||||
5933 | ASTContext &C = CGM.getContext(); | ||||||
5934 | FunctionArgList Args; | ||||||
5935 | ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
5936 | ImplicitParamDecl::Other); | ||||||
5937 | Args.emplace_back(&Param); | ||||||
5938 | const auto &FnInfo = | ||||||
5939 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
5940 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | ||||||
5941 | std::string Name = CGM.getOpenMPRuntime().getName({"red_fini", ""}); | ||||||
5942 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | ||||||
5943 | Name, &CGM.getModule()); | ||||||
5944 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | ||||||
5945 | Fn->setDoesNotRecurse(); | ||||||
5946 | CodeGenFunction CGF(CGM); | ||||||
5947 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | ||||||
5948 | Address PrivateAddr = CGF.EmitLoadOfPointer( | ||||||
5949 | CGF.GetAddrOfLocalVar(&Param), | ||||||
5950 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | ||||||
5951 | llvm::Value *Size = nullptr; | ||||||
5952 | // If the size of the reduction item is non-constant, load it from global | ||||||
5953 | // threadprivate variable. | ||||||
5954 | if (RCG.getSizes(N).second) { | ||||||
5955 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | ||||||
5956 | CGF, CGM.getContext().getSizeType(), | ||||||
5957 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | ||||||
5958 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | ||||||
5959 | CGM.getContext().getSizeType(), Loc); | ||||||
5960 | } | ||||||
5961 | RCG.emitAggregateType(CGF, N, Size); | ||||||
5962 | // Emit the finalizer body: | ||||||
5963 | // <destroy>(<type>* %0) | ||||||
5964 | RCG.emitCleanups(CGF, N, PrivateAddr); | ||||||
5965 | CGF.FinishFunction(Loc); | ||||||
5966 | return Fn; | ||||||
5967 | } | ||||||
5968 | |||||||
5969 | llvm::Value *CGOpenMPRuntime::emitTaskReductionInit( | ||||||
5970 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs, | ||||||
5971 | ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) { | ||||||
5972 | if (!CGF.HaveInsertPoint() || Data.ReductionVars.empty()) | ||||||
5973 | return nullptr; | ||||||
5974 | |||||||
5975 | // Build typedef struct: | ||||||
5976 | // kmp_taskred_input { | ||||||
5977 | // void *reduce_shar; // shared reduction item | ||||||
5978 | // void *reduce_orig; // original reduction item used for initialization | ||||||
5979 | // size_t reduce_size; // size of data item | ||||||
5980 | // void *reduce_init; // data initialization routine | ||||||
5981 | // void *reduce_fini; // data finalization routine | ||||||
5982 | // void *reduce_comb; // data combiner routine | ||||||
5983 | // kmp_task_red_flags_t flags; // flags for additional info from compiler | ||||||
5984 | // } kmp_taskred_input_t; | ||||||
5985 | ASTContext &C = CGM.getContext(); | ||||||
5986 | RecordDecl *RD = C.buildImplicitRecord("kmp_taskred_input_t"); | ||||||
5987 | RD->startDefinition(); | ||||||
5988 | const FieldDecl *SharedFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
5989 | const FieldDecl *OrigFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
5990 | const FieldDecl *SizeFD = addFieldToRecordDecl(C, RD, C.getSizeType()); | ||||||
5991 | const FieldDecl *InitFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
5992 | const FieldDecl *FiniFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
5993 | const FieldDecl *CombFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | ||||||
5994 | const FieldDecl *FlagsFD = addFieldToRecordDecl( | ||||||
5995 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false)); | ||||||
5996 | RD->completeDefinition(); | ||||||
5997 | QualType RDType = C.getRecordType(RD); | ||||||
5998 | unsigned Size = Data.ReductionVars.size(); | ||||||
5999 | llvm::APInt ArraySize(/*numBits=*/64, Size); | ||||||
6000 | QualType ArrayRDType = C.getConstantArrayType( | ||||||
6001 | RDType, ArraySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0); | ||||||
6002 | // kmp_task_red_input_t .rd_input.[Size]; | ||||||
6003 | Address TaskRedInput = CGF.CreateMemTemp(ArrayRDType, ".rd_input."); | ||||||
6004 | ReductionCodeGen RCG(Data.ReductionVars, Data.ReductionOrigs, | ||||||
6005 | Data.ReductionCopies, Data.ReductionOps); | ||||||
6006 | for (unsigned Cnt = 0; Cnt < Size; ++Cnt) { | ||||||
6007 | // kmp_task_red_input_t &ElemLVal = .rd_input.[Cnt]; | ||||||
6008 | llvm::Value *Idxs[] = {llvm::ConstantInt::get(CGM.SizeTy, /*V=*/0), | ||||||
6009 | llvm::ConstantInt::get(CGM.SizeTy, Cnt)}; | ||||||
6010 | llvm::Value *GEP = CGF.EmitCheckedInBoundsGEP( | ||||||
6011 | TaskRedInput.getPointer(), Idxs, | ||||||
6012 | /*SignedIndices=*/false, /*IsSubtraction=*/false, Loc, | ||||||
6013 | ".rd_input.gep."); | ||||||
6014 | LValue ElemLVal = CGF.MakeNaturalAlignAddrLValue(GEP, RDType); | ||||||
6015 | // ElemLVal.reduce_shar = &Shareds[Cnt]; | ||||||
6016 | LValue SharedLVal = CGF.EmitLValueForField(ElemLVal, SharedFD); | ||||||
6017 | RCG.emitSharedOrigLValue(CGF, Cnt); | ||||||
6018 | llvm::Value *CastedShared = | ||||||
6019 | CGF.EmitCastToVoidPtr(RCG.getSharedLValue(Cnt).getPointer(CGF)); | ||||||
6020 | CGF.EmitStoreOfScalar(CastedShared, SharedLVal); | ||||||
6021 | // ElemLVal.reduce_orig = &Origs[Cnt]; | ||||||
6022 | LValue OrigLVal = CGF.EmitLValueForField(ElemLVal, OrigFD); | ||||||
6023 | llvm::Value *CastedOrig = | ||||||
6024 | CGF.EmitCastToVoidPtr(RCG.getOrigLValue(Cnt).getPointer(CGF)); | ||||||
6025 | CGF.EmitStoreOfScalar(CastedOrig, OrigLVal); | ||||||
6026 | RCG.emitAggregateType(CGF, Cnt); | ||||||
6027 | llvm::Value *SizeValInChars; | ||||||
6028 | llvm::Value *SizeVal; | ||||||
6029 | std::tie(SizeValInChars, SizeVal) = RCG.getSizes(Cnt); | ||||||
6030 | // We use delayed creation/initialization for VLAs and array sections. It is | ||||||
6031 | // required because runtime does not provide the way to pass the sizes of | ||||||
6032 | // VLAs/array sections to initializer/combiner/finalizer functions. Instead | ||||||
6033 | // threadprivate global variables are used to store these values and use | ||||||
6034 | // them in the functions. | ||||||
6035 | bool DelayedCreation = !!SizeVal; | ||||||
6036 | SizeValInChars = CGF.Builder.CreateIntCast(SizeValInChars, CGM.SizeTy, | ||||||
6037 | /*isSigned=*/false); | ||||||
6038 | LValue SizeLVal = CGF.EmitLValueForField(ElemLVal, SizeFD); | ||||||
6039 | CGF.EmitStoreOfScalar(SizeValInChars, SizeLVal); | ||||||
6040 | // ElemLVal.reduce_init = init; | ||||||
6041 | LValue InitLVal = CGF.EmitLValueForField(ElemLVal, InitFD); | ||||||
6042 | llvm::Value *InitAddr = | ||||||
6043 | CGF.EmitCastToVoidPtr(emitReduceInitFunction(CGM, Loc, RCG, Cnt)); | ||||||
6044 | CGF.EmitStoreOfScalar(InitAddr, InitLVal); | ||||||
6045 | // ElemLVal.reduce_fini = fini; | ||||||
6046 | LValue FiniLVal = CGF.EmitLValueForField(ElemLVal, FiniFD); | ||||||
6047 | llvm::Value *Fini = emitReduceFiniFunction(CGM, Loc, RCG, Cnt); | ||||||
6048 | llvm::Value *FiniAddr = Fini | ||||||
6049 | ? CGF.EmitCastToVoidPtr(Fini) | ||||||
6050 | : llvm::ConstantPointerNull::get(CGM.VoidPtrTy); | ||||||
6051 | CGF.EmitStoreOfScalar(FiniAddr, FiniLVal); | ||||||
6052 | // ElemLVal.reduce_comb = comb; | ||||||
6053 | LValue CombLVal = CGF.EmitLValueForField(ElemLVal, CombFD); | ||||||
6054 | llvm::Value *CombAddr = CGF.EmitCastToVoidPtr(emitReduceCombFunction( | ||||||
6055 | CGM, Loc, RCG, Cnt, Data.ReductionOps[Cnt], LHSExprs[Cnt], | ||||||
6056 | RHSExprs[Cnt], Data.ReductionCopies[Cnt])); | ||||||
6057 | CGF.EmitStoreOfScalar(CombAddr, CombLVal); | ||||||
6058 | // ElemLVal.flags = 0; | ||||||
6059 | LValue FlagsLVal = CGF.EmitLValueForField(ElemLVal, FlagsFD); | ||||||
6060 | if (DelayedCreation) { | ||||||
6061 | CGF.EmitStoreOfScalar( | ||||||
6062 | llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1, /*isSigned=*/true), | ||||||
6063 | FlagsLVal); | ||||||
6064 | } else | ||||||
6065 | CGF.EmitNullInitialization(FlagsLVal.getAddress(CGF), | ||||||
6066 | FlagsLVal.getType()); | ||||||
6067 | } | ||||||
6068 | if (Data.IsReductionWithTaskMod) { | ||||||
6069 | // Build call void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int | ||||||
6070 | // is_ws, int num, void *data); | ||||||
6071 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc); | ||||||
6072 | llvm::Value *GTid = CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), | ||||||
6073 | CGM.IntTy, /*isSigned=*/true); | ||||||
6074 | llvm::Value *Args[] = { | ||||||
6075 | IdentTLoc, GTid, | ||||||
6076 | llvm::ConstantInt::get(CGM.IntTy, Data.IsWorksharingReduction ? 1 : 0, | ||||||
6077 | /*isSigned=*/true), | ||||||
6078 | llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true), | ||||||
6079 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
6080 | TaskRedInput.getPointer(), CGM.VoidPtrTy)}; | ||||||
6081 | return CGF.EmitRuntimeCall( | ||||||
6082 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6083 | CGM.getModule(), OMPRTL___kmpc_taskred_modifier_init), | ||||||
6084 | Args); | ||||||
6085 | } | ||||||
6086 | // Build call void *__kmpc_taskred_init(int gtid, int num_data, void *data); | ||||||
6087 | llvm::Value *Args[] = { | ||||||
6088 | CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), CGM.IntTy, | ||||||
6089 | /*isSigned=*/true), | ||||||
6090 | llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true), | ||||||
6091 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TaskRedInput.getPointer(), | ||||||
6092 | CGM.VoidPtrTy)}; | ||||||
6093 | return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6094 | CGM.getModule(), OMPRTL___kmpc_taskred_init), | ||||||
6095 | Args); | ||||||
6096 | } | ||||||
6097 | |||||||
6098 | void CGOpenMPRuntime::emitTaskReductionFini(CodeGenFunction &CGF, | ||||||
6099 | SourceLocation Loc, | ||||||
6100 | bool IsWorksharingReduction) { | ||||||
6101 | // Build call void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int | ||||||
6102 | // is_ws, int num, void *data); | ||||||
6103 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc); | ||||||
6104 | llvm::Value *GTid = CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), | ||||||
6105 | CGM.IntTy, /*isSigned=*/true); | ||||||
6106 | llvm::Value *Args[] = {IdentTLoc, GTid, | ||||||
6107 | llvm::ConstantInt::get(CGM.IntTy, | ||||||
6108 | IsWorksharingReduction ? 1 : 0, | ||||||
6109 | /*isSigned=*/true)}; | ||||||
6110 | (void)CGF.EmitRuntimeCall( | ||||||
6111 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6112 | CGM.getModule(), OMPRTL___kmpc_task_reduction_modifier_fini), | ||||||
6113 | Args); | ||||||
6114 | } | ||||||
6115 | |||||||
6116 | void CGOpenMPRuntime::emitTaskReductionFixups(CodeGenFunction &CGF, | ||||||
6117 | SourceLocation Loc, | ||||||
6118 | ReductionCodeGen &RCG, | ||||||
6119 | unsigned N) { | ||||||
6120 | auto Sizes = RCG.getSizes(N); | ||||||
6121 | // Emit threadprivate global variable if the type is non-constant | ||||||
6122 | // (Sizes.second = nullptr). | ||||||
6123 | if (Sizes.second) { | ||||||
6124 | llvm::Value *SizeVal = CGF.Builder.CreateIntCast(Sizes.second, CGM.SizeTy, | ||||||
6125 | /*isSigned=*/false); | ||||||
6126 | Address SizeAddr = getAddrOfArtificialThreadPrivate( | ||||||
6127 | CGF, CGM.getContext().getSizeType(), | ||||||
6128 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | ||||||
6129 | CGF.Builder.CreateStore(SizeVal, SizeAddr, /*IsVolatile=*/false); | ||||||
6130 | } | ||||||
6131 | } | ||||||
6132 | |||||||
6133 | Address CGOpenMPRuntime::getTaskReductionItem(CodeGenFunction &CGF, | ||||||
6134 | SourceLocation Loc, | ||||||
6135 | llvm::Value *ReductionsPtr, | ||||||
6136 | LValue SharedLVal) { | ||||||
6137 | // Build call void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void | ||||||
6138 | // *d); | ||||||
6139 | llvm::Value *Args[] = {CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), | ||||||
6140 | CGM.IntTy, | ||||||
6141 | /*isSigned=*/true), | ||||||
6142 | ReductionsPtr, | ||||||
6143 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
6144 | SharedLVal.getPointer(CGF), CGM.VoidPtrTy)}; | ||||||
6145 | return Address( | ||||||
6146 | CGF.EmitRuntimeCall( | ||||||
6147 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6148 | CGM.getModule(), OMPRTL___kmpc_task_reduction_get_th_data), | ||||||
6149 | Args), | ||||||
6150 | SharedLVal.getAlignment()); | ||||||
6151 | } | ||||||
6152 | |||||||
6153 | void CGOpenMPRuntime::emitTaskwaitCall(CodeGenFunction &CGF, | ||||||
6154 | SourceLocation Loc) { | ||||||
6155 | if (!CGF.HaveInsertPoint()) | ||||||
6156 | return; | ||||||
6157 | |||||||
6158 | if (CGF.CGM.getLangOpts().OpenMPIRBuilder) { | ||||||
6159 | OMPBuilder.CreateTaskwait(CGF.Builder); | ||||||
6160 | } else { | ||||||
6161 | // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 | ||||||
6162 | // global_tid); | ||||||
6163 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | ||||||
6164 | // Ignore return result until untied tasks are supported. | ||||||
6165 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6166 | CGM.getModule(), OMPRTL___kmpc_omp_taskwait), | ||||||
6167 | Args); | ||||||
6168 | } | ||||||
6169 | |||||||
6170 | if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | ||||||
6171 | Region->emitUntiedSwitch(CGF); | ||||||
6172 | } | ||||||
6173 | |||||||
6174 | void CGOpenMPRuntime::emitInlinedDirective(CodeGenFunction &CGF, | ||||||
6175 | OpenMPDirectiveKind InnerKind, | ||||||
6176 | const RegionCodeGenTy &CodeGen, | ||||||
6177 | bool HasCancel) { | ||||||
6178 | if (!CGF.HaveInsertPoint()) | ||||||
6179 | return; | ||||||
6180 | InlinedOpenMPRegionRAII Region(CGF, CodeGen, InnerKind, HasCancel); | ||||||
6181 | CGF.CapturedStmtInfo->EmitBody(CGF, /*S=*/nullptr); | ||||||
6182 | } | ||||||
6183 | |||||||
6184 | namespace { | ||||||
6185 | enum RTCancelKind { | ||||||
6186 | CancelNoreq = 0, | ||||||
6187 | CancelParallel = 1, | ||||||
6188 | CancelLoop = 2, | ||||||
6189 | CancelSections = 3, | ||||||
6190 | CancelTaskgroup = 4 | ||||||
6191 | }; | ||||||
6192 | } // anonymous namespace | ||||||
6193 | |||||||
6194 | static RTCancelKind getCancellationKind(OpenMPDirectiveKind CancelRegion) { | ||||||
6195 | RTCancelKind CancelKind = CancelNoreq; | ||||||
6196 | if (CancelRegion == OMPD_parallel) | ||||||
6197 | CancelKind = CancelParallel; | ||||||
6198 | else if (CancelRegion == OMPD_for) | ||||||
6199 | CancelKind = CancelLoop; | ||||||
6200 | else if (CancelRegion == OMPD_sections) | ||||||
6201 | CancelKind = CancelSections; | ||||||
6202 | else { | ||||||
6203 | assert(CancelRegion == OMPD_taskgroup)((CancelRegion == OMPD_taskgroup) ? static_cast<void> ( 0) : __assert_fail ("CancelRegion == OMPD_taskgroup", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6203, __PRETTY_FUNCTION__)); | ||||||
6204 | CancelKind = CancelTaskgroup; | ||||||
6205 | } | ||||||
6206 | return CancelKind; | ||||||
6207 | } | ||||||
6208 | |||||||
6209 | void CGOpenMPRuntime::emitCancellationPointCall( | ||||||
6210 | CodeGenFunction &CGF, SourceLocation Loc, | ||||||
6211 | OpenMPDirectiveKind CancelRegion) { | ||||||
6212 | if (!CGF.HaveInsertPoint()) | ||||||
6213 | return; | ||||||
6214 | // Build call kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32 | ||||||
6215 | // global_tid, kmp_int32 cncl_kind); | ||||||
6216 | if (auto *OMPRegionInfo = | ||||||
6217 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | ||||||
6218 | // For 'cancellation point taskgroup', the task region info may not have a | ||||||
6219 | // cancel. This may instead happen in another adjacent task. | ||||||
6220 | if (CancelRegion == OMPD_taskgroup || OMPRegionInfo->hasCancel()) { | ||||||
6221 | llvm::Value *Args[] = { | ||||||
6222 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | ||||||
6223 | CGF.Builder.getInt32(getCancellationKind(CancelRegion))}; | ||||||
6224 | // Ignore return result until untied tasks are supported. | ||||||
6225 | llvm::Value *Result = CGF.EmitRuntimeCall( | ||||||
6226 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6227 | CGM.getModule(), OMPRTL___kmpc_cancellationpoint), | ||||||
6228 | Args); | ||||||
6229 | // if (__kmpc_cancellationpoint()) { | ||||||
6230 | // exit from construct; | ||||||
6231 | // } | ||||||
6232 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | ||||||
6233 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | ||||||
6234 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | ||||||
6235 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | ||||||
6236 | CGF.EmitBlock(ExitBB); | ||||||
6237 | // exit from construct; | ||||||
6238 | CodeGenFunction::JumpDest CancelDest = | ||||||
6239 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | ||||||
6240 | CGF.EmitBranchThroughCleanup(CancelDest); | ||||||
6241 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | ||||||
6242 | } | ||||||
6243 | } | ||||||
6244 | } | ||||||
6245 | |||||||
6246 | void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
6247 | const Expr *IfCond, | ||||||
6248 | OpenMPDirectiveKind CancelRegion) { | ||||||
6249 | if (!CGF.HaveInsertPoint()) | ||||||
6250 | return; | ||||||
6251 | // Build call kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid, | ||||||
6252 | // kmp_int32 cncl_kind); | ||||||
6253 | auto &M = CGM.getModule(); | ||||||
6254 | if (auto *OMPRegionInfo = | ||||||
6255 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | ||||||
6256 | auto &&ThenGen = [this, &M, Loc, CancelRegion, | ||||||
6257 | OMPRegionInfo](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
6258 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | ||||||
6259 | llvm::Value *Args[] = { | ||||||
6260 | RT.emitUpdateLocation(CGF, Loc), RT.getThreadID(CGF, Loc), | ||||||
6261 | CGF.Builder.getInt32(getCancellationKind(CancelRegion))}; | ||||||
6262 | // Ignore return result until untied tasks are supported. | ||||||
6263 | llvm::Value *Result = CGF.EmitRuntimeCall( | ||||||
6264 | OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_cancel), Args); | ||||||
6265 | // if (__kmpc_cancel()) { | ||||||
6266 | // exit from construct; | ||||||
6267 | // } | ||||||
6268 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | ||||||
6269 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | ||||||
6270 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | ||||||
6271 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | ||||||
6272 | CGF.EmitBlock(ExitBB); | ||||||
6273 | // exit from construct; | ||||||
6274 | CodeGenFunction::JumpDest CancelDest = | ||||||
6275 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | ||||||
6276 | CGF.EmitBranchThroughCleanup(CancelDest); | ||||||
6277 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | ||||||
6278 | }; | ||||||
6279 | if (IfCond) { | ||||||
6280 | emitIfClause(CGF, IfCond, ThenGen, | ||||||
6281 | [](CodeGenFunction &, PrePostActionTy &) {}); | ||||||
6282 | } else { | ||||||
6283 | RegionCodeGenTy ThenRCG(ThenGen); | ||||||
6284 | ThenRCG(CGF); | ||||||
6285 | } | ||||||
6286 | } | ||||||
6287 | } | ||||||
6288 | |||||||
6289 | namespace { | ||||||
6290 | /// Cleanup action for uses_allocators support. | ||||||
6291 | class OMPUsesAllocatorsActionTy final : public PrePostActionTy { | ||||||
6292 | ArrayRef<std::pair<const Expr *, const Expr *>> Allocators; | ||||||
6293 | |||||||
6294 | public: | ||||||
6295 | OMPUsesAllocatorsActionTy( | ||||||
6296 | ArrayRef<std::pair<const Expr *, const Expr *>> Allocators) | ||||||
6297 | : Allocators(Allocators) {} | ||||||
6298 | void Enter(CodeGenFunction &CGF) override { | ||||||
6299 | if (!CGF.HaveInsertPoint()) | ||||||
6300 | return; | ||||||
6301 | for (const auto &AllocatorData : Allocators) { | ||||||
6302 | CGF.CGM.getOpenMPRuntime().emitUsesAllocatorsInit( | ||||||
6303 | CGF, AllocatorData.first, AllocatorData.second); | ||||||
6304 | } | ||||||
6305 | } | ||||||
6306 | void Exit(CodeGenFunction &CGF) override { | ||||||
6307 | if (!CGF.HaveInsertPoint()) | ||||||
6308 | return; | ||||||
6309 | for (const auto &AllocatorData : Allocators) { | ||||||
6310 | CGF.CGM.getOpenMPRuntime().emitUsesAllocatorsFini(CGF, | ||||||
6311 | AllocatorData.first); | ||||||
6312 | } | ||||||
6313 | } | ||||||
6314 | }; | ||||||
6315 | } // namespace | ||||||
6316 | |||||||
6317 | void CGOpenMPRuntime::emitTargetOutlinedFunction( | ||||||
6318 | const OMPExecutableDirective &D, StringRef ParentName, | ||||||
6319 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | ||||||
6320 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | ||||||
6321 | assert(!ParentName.empty() && "Invalid target region parent name!")((!ParentName.empty() && "Invalid target region parent name!" ) ? static_cast<void> (0) : __assert_fail ("!ParentName.empty() && \"Invalid target region parent name!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6321, __PRETTY_FUNCTION__)); | ||||||
6322 | HasEmittedTargetRegion = true; | ||||||
6323 | SmallVector<std::pair<const Expr *, const Expr *>, 4> Allocators; | ||||||
6324 | for (const auto *C : D.getClausesOfKind<OMPUsesAllocatorsClause>()) { | ||||||
6325 | for (unsigned I = 0, E = C->getNumberOfAllocators(); I < E; ++I) { | ||||||
6326 | const OMPUsesAllocatorsClause::Data D = C->getAllocatorData(I); | ||||||
6327 | if (!D.AllocatorTraits) | ||||||
6328 | continue; | ||||||
6329 | Allocators.emplace_back(D.Allocator, D.AllocatorTraits); | ||||||
6330 | } | ||||||
6331 | } | ||||||
6332 | OMPUsesAllocatorsActionTy UsesAllocatorAction(Allocators); | ||||||
6333 | CodeGen.setAction(UsesAllocatorAction); | ||||||
6334 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, | ||||||
6335 | IsOffloadEntry, CodeGen); | ||||||
6336 | } | ||||||
6337 | |||||||
6338 | void CGOpenMPRuntime::emitUsesAllocatorsInit(CodeGenFunction &CGF, | ||||||
6339 | const Expr *Allocator, | ||||||
6340 | const Expr *AllocatorTraits) { | ||||||
6341 | llvm::Value *ThreadId = getThreadID(CGF, Allocator->getExprLoc()); | ||||||
6342 | ThreadId = CGF.Builder.CreateIntCast(ThreadId, CGF.IntTy, /*isSigned=*/true); | ||||||
6343 | // Use default memspace handle. | ||||||
6344 | llvm::Value *MemSpaceHandle = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | ||||||
6345 | llvm::Value *NumTraits = llvm::ConstantInt::get( | ||||||
6346 | CGF.IntTy, cast<ConstantArrayType>( | ||||||
6347 | AllocatorTraits->getType()->getAsArrayTypeUnsafe()) | ||||||
6348 | ->getSize() | ||||||
6349 | .getLimitedValue()); | ||||||
6350 | LValue AllocatorTraitsLVal = CGF.EmitLValue(AllocatorTraits); | ||||||
6351 | Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
6352 | AllocatorTraitsLVal.getAddress(CGF), CGF.VoidPtrPtrTy); | ||||||
6353 | AllocatorTraitsLVal = CGF.MakeAddrLValue(Addr, CGF.getContext().VoidPtrTy, | ||||||
6354 | AllocatorTraitsLVal.getBaseInfo(), | ||||||
6355 | AllocatorTraitsLVal.getTBAAInfo()); | ||||||
6356 | llvm::Value *Traits = | ||||||
6357 | CGF.EmitLoadOfScalar(AllocatorTraitsLVal, AllocatorTraits->getExprLoc()); | ||||||
6358 | |||||||
6359 | llvm::Value *AllocatorVal = | ||||||
6360 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
6361 | CGM.getModule(), OMPRTL___kmpc_init_allocator), | ||||||
6362 | {ThreadId, MemSpaceHandle, NumTraits, Traits}); | ||||||
6363 | // Store to allocator. | ||||||
6364 | CGF.EmitVarDecl(*cast<VarDecl>( | ||||||
6365 | cast<DeclRefExpr>(Allocator->IgnoreParenImpCasts())->getDecl())); | ||||||
6366 | LValue AllocatorLVal = CGF.EmitLValue(Allocator->IgnoreParenImpCasts()); | ||||||
6367 | AllocatorVal = | ||||||
6368 | CGF.EmitScalarConversion(AllocatorVal, CGF.getContext().VoidPtrTy, | ||||||
6369 | Allocator->getType(), Allocator->getExprLoc()); | ||||||
6370 | CGF.EmitStoreOfScalar(AllocatorVal, AllocatorLVal); | ||||||
6371 | } | ||||||
6372 | |||||||
6373 | void CGOpenMPRuntime::emitUsesAllocatorsFini(CodeGenFunction &CGF, | ||||||
6374 | const Expr *Allocator) { | ||||||
6375 | llvm::Value *ThreadId = getThreadID(CGF, Allocator->getExprLoc()); | ||||||
6376 | ThreadId = CGF.Builder.CreateIntCast(ThreadId, CGF.IntTy, /*isSigned=*/true); | ||||||
6377 | LValue AllocatorLVal = CGF.EmitLValue(Allocator->IgnoreParenImpCasts()); | ||||||
6378 | llvm::Value *AllocatorVal = | ||||||
6379 | CGF.EmitLoadOfScalar(AllocatorLVal, Allocator->getExprLoc()); | ||||||
6380 | AllocatorVal = CGF.EmitScalarConversion(AllocatorVal, Allocator->getType(), | ||||||
6381 | CGF.getContext().VoidPtrTy, | ||||||
6382 | Allocator->getExprLoc()); | ||||||
6383 | (void)CGF.EmitRuntimeCall( | ||||||
6384 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
6385 | OMPRTL___kmpc_destroy_allocator), | ||||||
6386 | {ThreadId, AllocatorVal}); | ||||||
6387 | } | ||||||
6388 | |||||||
6389 | void CGOpenMPRuntime::emitTargetOutlinedFunctionHelper( | ||||||
6390 | const OMPExecutableDirective &D, StringRef ParentName, | ||||||
6391 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | ||||||
6392 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | ||||||
6393 | // Create a unique name for the entry function using the source location | ||||||
6394 | // information of the current target region. The name will be something like: | ||||||
6395 | // | ||||||
6396 | // __omp_offloading_DD_FFFF_PP_lBB | ||||||
6397 | // | ||||||
6398 | // where DD_FFFF is an ID unique to the file (device and file IDs), PP is the | ||||||
6399 | // mangled name of the function that encloses the target region and BB is the | ||||||
6400 | // line number of the target region. | ||||||
6401 | |||||||
6402 | unsigned DeviceID; | ||||||
6403 | unsigned FileID; | ||||||
6404 | unsigned Line; | ||||||
6405 | getTargetEntryUniqueInfo(CGM.getContext(), D.getBeginLoc(), DeviceID, FileID, | ||||||
6406 | Line); | ||||||
6407 | SmallString<64> EntryFnName; | ||||||
6408 | { | ||||||
6409 | llvm::raw_svector_ostream OS(EntryFnName); | ||||||
6410 | OS << "__omp_offloading" << llvm::format("_%x", DeviceID) | ||||||
6411 | << llvm::format("_%x_", FileID) << ParentName << "_l" << Line; | ||||||
6412 | } | ||||||
6413 | |||||||
6414 | const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target); | ||||||
6415 | |||||||
6416 | CodeGenFunction CGF(CGM, true); | ||||||
6417 | CGOpenMPTargetRegionInfo CGInfo(CS, CodeGen, EntryFnName); | ||||||
6418 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
6419 | |||||||
6420 | OutlinedFn = CGF.GenerateOpenMPCapturedStmtFunction(CS, D.getBeginLoc()); | ||||||
6421 | |||||||
6422 | // If this target outline function is not an offload entry, we don't need to | ||||||
6423 | // register it. | ||||||
6424 | if (!IsOffloadEntry) | ||||||
6425 | return; | ||||||
6426 | |||||||
6427 | // The target region ID is used by the runtime library to identify the current | ||||||
6428 | // target region, so it only has to be unique and not necessarily point to | ||||||
6429 | // anything. It could be the pointer to the outlined function that implements | ||||||
6430 | // the target region, but we aren't using that so that the compiler doesn't | ||||||
6431 | // need to keep that, and could therefore inline the host function if proven | ||||||
6432 | // worthwhile during optimization. In the other hand, if emitting code for the | ||||||
6433 | // device, the ID has to be the function address so that it can retrieved from | ||||||
6434 | // the offloading entry and launched by the runtime library. We also mark the | ||||||
6435 | // outlined function to have external linkage in case we are emitting code for | ||||||
6436 | // the device, because these functions will be entry points to the device. | ||||||
6437 | |||||||
6438 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
6439 | OutlinedFnID = llvm::ConstantExpr::getBitCast(OutlinedFn, CGM.Int8PtrTy); | ||||||
6440 | OutlinedFn->setLinkage(llvm::GlobalValue::WeakAnyLinkage); | ||||||
6441 | OutlinedFn->setDSOLocal(false); | ||||||
6442 | } else { | ||||||
6443 | std::string Name = getName({EntryFnName, "region_id"}); | ||||||
6444 | OutlinedFnID = new llvm::GlobalVariable( | ||||||
6445 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | ||||||
6446 | llvm::GlobalValue::WeakAnyLinkage, | ||||||
6447 | llvm::Constant::getNullValue(CGM.Int8Ty), Name); | ||||||
6448 | } | ||||||
6449 | |||||||
6450 | // Register the information for the entry associated with this target region. | ||||||
6451 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | ||||||
6452 | DeviceID, FileID, ParentName, Line, OutlinedFn, OutlinedFnID, | ||||||
6453 | OffloadEntriesInfoManagerTy::OMPTargetRegionEntryTargetRegion); | ||||||
6454 | } | ||||||
6455 | |||||||
6456 | /// Checks if the expression is constant or does not have non-trivial function | ||||||
6457 | /// calls. | ||||||
6458 | static bool isTrivial(ASTContext &Ctx, const Expr * E) { | ||||||
6459 | // We can skip constant expressions. | ||||||
6460 | // We can skip expressions with trivial calls or simple expressions. | ||||||
6461 | return (E->isEvaluatable(Ctx, Expr::SE_AllowUndefinedBehavior) || | ||||||
6462 | !E->hasNonTrivialCall(Ctx)) && | ||||||
6463 | !E->HasSideEffects(Ctx, /*IncludePossibleEffects=*/true); | ||||||
6464 | } | ||||||
6465 | |||||||
6466 | const Stmt *CGOpenMPRuntime::getSingleCompoundChild(ASTContext &Ctx, | ||||||
6467 | const Stmt *Body) { | ||||||
6468 | const Stmt *Child = Body->IgnoreContainers(); | ||||||
6469 | while (const auto *C = dyn_cast_or_null<CompoundStmt>(Child)) { | ||||||
6470 | Child = nullptr; | ||||||
6471 | for (const Stmt *S : C->body()) { | ||||||
6472 | if (const auto *E = dyn_cast<Expr>(S)) { | ||||||
6473 | if (isTrivial(Ctx, E)) | ||||||
6474 | continue; | ||||||
6475 | } | ||||||
6476 | // Some of the statements can be ignored. | ||||||
6477 | if (isa<AsmStmt>(S) || isa<NullStmt>(S) || isa<OMPFlushDirective>(S) || | ||||||
6478 | isa<OMPBarrierDirective>(S) || isa<OMPTaskyieldDirective>(S)) | ||||||
6479 | continue; | ||||||
6480 | // Analyze declarations. | ||||||
6481 | if (const auto *DS = dyn_cast<DeclStmt>(S)) { | ||||||
6482 | if (llvm::all_of(DS->decls(), [&Ctx](const Decl *D) { | ||||||
6483 | if (isa<EmptyDecl>(D) || isa<DeclContext>(D) || | ||||||
6484 | isa<TypeDecl>(D) || isa<PragmaCommentDecl>(D) || | ||||||
6485 | isa<PragmaDetectMismatchDecl>(D) || isa<UsingDecl>(D) || | ||||||
6486 | isa<UsingDirectiveDecl>(D) || | ||||||
6487 | isa<OMPDeclareReductionDecl>(D) || | ||||||
6488 | isa<OMPThreadPrivateDecl>(D) || isa<OMPAllocateDecl>(D)) | ||||||
6489 | return true; | ||||||
6490 | const auto *VD = dyn_cast<VarDecl>(D); | ||||||
6491 | if (!VD) | ||||||
6492 | return false; | ||||||
6493 | return VD->isConstexpr() || | ||||||
6494 | ((VD->getType().isTrivialType(Ctx) || | ||||||
6495 | VD->getType()->isReferenceType()) && | ||||||
6496 | (!VD->hasInit() || isTrivial(Ctx, VD->getInit()))); | ||||||
6497 | })) | ||||||
6498 | continue; | ||||||
6499 | } | ||||||
6500 | // Found multiple children - cannot get the one child only. | ||||||
6501 | if (Child) | ||||||
6502 | return nullptr; | ||||||
6503 | Child = S; | ||||||
6504 | } | ||||||
6505 | if (Child) | ||||||
6506 | Child = Child->IgnoreContainers(); | ||||||
6507 | } | ||||||
6508 | return Child; | ||||||
6509 | } | ||||||
6510 | |||||||
6511 | /// Emit the number of teams for a target directive. Inspect the num_teams | ||||||
6512 | /// clause associated with a teams construct combined or closely nested | ||||||
6513 | /// with the target directive. | ||||||
6514 | /// | ||||||
6515 | /// Emit a team of size one for directives such as 'target parallel' that | ||||||
6516 | /// have no associated teams construct. | ||||||
6517 | /// | ||||||
6518 | /// Otherwise, return nullptr. | ||||||
6519 | static llvm::Value * | ||||||
6520 | emitNumTeamsForTargetDirective(CodeGenFunction &CGF, | ||||||
6521 | const OMPExecutableDirective &D) { | ||||||
6522 | assert(!CGF.getLangOpts().OpenMPIsDevice &&((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6524, __PRETTY_FUNCTION__)) | ||||||
6523 | "Clauses associated with the teams directive expected to be emitted "((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6524, __PRETTY_FUNCTION__)) | ||||||
6524 | "only for the host!")((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6524, __PRETTY_FUNCTION__)); | ||||||
6525 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | ||||||
6526 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6527, __PRETTY_FUNCTION__)) | ||||||
6527 | "Expected target-based executable directive.")((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6527, __PRETTY_FUNCTION__)); | ||||||
6528 | CGBuilderTy &Bld = CGF.Builder; | ||||||
6529 | switch (DirectiveKind) { | ||||||
6530 | case OMPD_target: { | ||||||
6531 | const auto *CS = D.getInnermostCapturedStmt(); | ||||||
6532 | const auto *Body = | ||||||
6533 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | ||||||
6534 | const Stmt *ChildStmt = | ||||||
6535 | CGOpenMPRuntime::getSingleCompoundChild(CGF.getContext(), Body); | ||||||
6536 | if (const auto *NestedDir = | ||||||
6537 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | ||||||
6538 | if (isOpenMPTeamsDirective(NestedDir->getDirectiveKind())) { | ||||||
6539 | if (NestedDir->hasClausesOfKind<OMPNumTeamsClause>()) { | ||||||
6540 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | ||||||
6541 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
6542 | const Expr *NumTeams = | ||||||
6543 | NestedDir->getSingleClause<OMPNumTeamsClause>()->getNumTeams(); | ||||||
6544 | llvm::Value *NumTeamsVal = | ||||||
6545 | CGF.EmitScalarExpr(NumTeams, | ||||||
6546 | /*IgnoreResultAssign*/ true); | ||||||
6547 | return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty, | ||||||
6548 | /*isSigned=*/true); | ||||||
6549 | } | ||||||
6550 | return Bld.getInt32(0); | ||||||
6551 | } | ||||||
6552 | if (isOpenMPParallelDirective(NestedDir->getDirectiveKind()) || | ||||||
6553 | isOpenMPSimdDirective(NestedDir->getDirectiveKind())) | ||||||
6554 | return Bld.getInt32(1); | ||||||
6555 | return Bld.getInt32(0); | ||||||
6556 | } | ||||||
6557 | return nullptr; | ||||||
6558 | } | ||||||
6559 | case OMPD_target_teams: | ||||||
6560 | case OMPD_target_teams_distribute: | ||||||
6561 | case OMPD_target_teams_distribute_simd: | ||||||
6562 | case OMPD_target_teams_distribute_parallel_for: | ||||||
6563 | case OMPD_target_teams_distribute_parallel_for_simd: { | ||||||
6564 | if (D.hasClausesOfKind<OMPNumTeamsClause>()) { | ||||||
6565 | CodeGenFunction::RunCleanupsScope NumTeamsScope(CGF); | ||||||
6566 | const Expr *NumTeams = | ||||||
6567 | D.getSingleClause<OMPNumTeamsClause>()->getNumTeams(); | ||||||
6568 | llvm::Value *NumTeamsVal = | ||||||
6569 | CGF.EmitScalarExpr(NumTeams, | ||||||
6570 | /*IgnoreResultAssign*/ true); | ||||||
6571 | return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty, | ||||||
6572 | /*isSigned=*/true); | ||||||
6573 | } | ||||||
6574 | return Bld.getInt32(0); | ||||||
6575 | } | ||||||
6576 | case OMPD_target_parallel: | ||||||
6577 | case OMPD_target_parallel_for: | ||||||
6578 | case OMPD_target_parallel_for_simd: | ||||||
6579 | case OMPD_target_simd: | ||||||
6580 | return Bld.getInt32(1); | ||||||
6581 | case OMPD_parallel: | ||||||
6582 | case OMPD_for: | ||||||
6583 | case OMPD_parallel_for: | ||||||
6584 | case OMPD_parallel_master: | ||||||
6585 | case OMPD_parallel_sections: | ||||||
6586 | case OMPD_for_simd: | ||||||
6587 | case OMPD_parallel_for_simd: | ||||||
6588 | case OMPD_cancel: | ||||||
6589 | case OMPD_cancellation_point: | ||||||
6590 | case OMPD_ordered: | ||||||
6591 | case OMPD_threadprivate: | ||||||
6592 | case OMPD_allocate: | ||||||
6593 | case OMPD_task: | ||||||
6594 | case OMPD_simd: | ||||||
6595 | case OMPD_sections: | ||||||
6596 | case OMPD_section: | ||||||
6597 | case OMPD_single: | ||||||
6598 | case OMPD_master: | ||||||
6599 | case OMPD_critical: | ||||||
6600 | case OMPD_taskyield: | ||||||
6601 | case OMPD_barrier: | ||||||
6602 | case OMPD_taskwait: | ||||||
6603 | case OMPD_taskgroup: | ||||||
6604 | case OMPD_atomic: | ||||||
6605 | case OMPD_flush: | ||||||
6606 | case OMPD_depobj: | ||||||
6607 | case OMPD_scan: | ||||||
6608 | case OMPD_teams: | ||||||
6609 | case OMPD_target_data: | ||||||
6610 | case OMPD_target_exit_data: | ||||||
6611 | case OMPD_target_enter_data: | ||||||
6612 | case OMPD_distribute: | ||||||
6613 | case OMPD_distribute_simd: | ||||||
6614 | case OMPD_distribute_parallel_for: | ||||||
6615 | case OMPD_distribute_parallel_for_simd: | ||||||
6616 | case OMPD_teams_distribute: | ||||||
6617 | case OMPD_teams_distribute_simd: | ||||||
6618 | case OMPD_teams_distribute_parallel_for: | ||||||
6619 | case OMPD_teams_distribute_parallel_for_simd: | ||||||
6620 | case OMPD_target_update: | ||||||
6621 | case OMPD_declare_simd: | ||||||
6622 | case OMPD_declare_variant: | ||||||
6623 | case OMPD_begin_declare_variant: | ||||||
6624 | case OMPD_end_declare_variant: | ||||||
6625 | case OMPD_declare_target: | ||||||
6626 | case OMPD_end_declare_target: | ||||||
6627 | case OMPD_declare_reduction: | ||||||
6628 | case OMPD_declare_mapper: | ||||||
6629 | case OMPD_taskloop: | ||||||
6630 | case OMPD_taskloop_simd: | ||||||
6631 | case OMPD_master_taskloop: | ||||||
6632 | case OMPD_master_taskloop_simd: | ||||||
6633 | case OMPD_parallel_master_taskloop: | ||||||
6634 | case OMPD_parallel_master_taskloop_simd: | ||||||
6635 | case OMPD_requires: | ||||||
6636 | case OMPD_unknown: | ||||||
6637 | break; | ||||||
6638 | default: | ||||||
6639 | break; | ||||||
6640 | } | ||||||
6641 | llvm_unreachable("Unexpected directive kind.")::llvm::llvm_unreachable_internal("Unexpected directive kind." , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6641); | ||||||
6642 | } | ||||||
6643 | |||||||
6644 | static llvm::Value *getNumThreads(CodeGenFunction &CGF, const CapturedStmt *CS, | ||||||
6645 | llvm::Value *DefaultThreadLimitVal) { | ||||||
6646 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | ||||||
6647 | CGF.getContext(), CS->getCapturedStmt()); | ||||||
6648 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | ||||||
6649 | if (isOpenMPParallelDirective(Dir->getDirectiveKind())) { | ||||||
6650 | llvm::Value *NumThreads = nullptr; | ||||||
6651 | llvm::Value *CondVal = nullptr; | ||||||
6652 | // Handle if clause. If if clause present, the number of threads is | ||||||
6653 | // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1. | ||||||
6654 | if (Dir->hasClausesOfKind<OMPIfClause>()) { | ||||||
6655 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | ||||||
6656 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
6657 | const OMPIfClause *IfClause = nullptr; | ||||||
6658 | for (const auto *C : Dir->getClausesOfKind<OMPIfClause>()) { | ||||||
6659 | if (C->getNameModifier() == OMPD_unknown || | ||||||
6660 | C->getNameModifier() == OMPD_parallel) { | ||||||
6661 | IfClause = C; | ||||||
6662 | break; | ||||||
6663 | } | ||||||
6664 | } | ||||||
6665 | if (IfClause) { | ||||||
6666 | const Expr *Cond = IfClause->getCondition(); | ||||||
6667 | bool Result; | ||||||
6668 | if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) { | ||||||
6669 | if (!Result) | ||||||
6670 | return CGF.Builder.getInt32(1); | ||||||
6671 | } else { | ||||||
6672 | CodeGenFunction::LexicalScope Scope(CGF, Cond->getSourceRange()); | ||||||
6673 | if (const auto *PreInit = | ||||||
6674 | cast_or_null<DeclStmt>(IfClause->getPreInitStmt())) { | ||||||
6675 | for (const auto *I : PreInit->decls()) { | ||||||
6676 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | ||||||
6677 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | ||||||
6678 | } else { | ||||||
6679 | CodeGenFunction::AutoVarEmission Emission = | ||||||
6680 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | ||||||
6681 | CGF.EmitAutoVarCleanups(Emission); | ||||||
6682 | } | ||||||
6683 | } | ||||||
6684 | } | ||||||
6685 | CondVal = CGF.EvaluateExprAsBool(Cond); | ||||||
6686 | } | ||||||
6687 | } | ||||||
6688 | } | ||||||
6689 | // Check the value of num_threads clause iff if clause was not specified | ||||||
6690 | // or is not evaluated to false. | ||||||
6691 | if (Dir->hasClausesOfKind<OMPNumThreadsClause>()) { | ||||||
6692 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | ||||||
6693 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
6694 | const auto *NumThreadsClause = | ||||||
6695 | Dir->getSingleClause<OMPNumThreadsClause>(); | ||||||
6696 | CodeGenFunction::LexicalScope Scope( | ||||||
6697 | CGF, NumThreadsClause->getNumThreads()->getSourceRange()); | ||||||
6698 | if (const auto *PreInit = | ||||||
6699 | cast_or_null<DeclStmt>(NumThreadsClause->getPreInitStmt())) { | ||||||
6700 | for (const auto *I : PreInit->decls()) { | ||||||
6701 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | ||||||
6702 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | ||||||
6703 | } else { | ||||||
6704 | CodeGenFunction::AutoVarEmission Emission = | ||||||
6705 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | ||||||
6706 | CGF.EmitAutoVarCleanups(Emission); | ||||||
6707 | } | ||||||
6708 | } | ||||||
6709 | } | ||||||
6710 | NumThreads = CGF.EmitScalarExpr(NumThreadsClause->getNumThreads()); | ||||||
6711 | NumThreads = CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, | ||||||
6712 | /*isSigned=*/false); | ||||||
6713 | if (DefaultThreadLimitVal) | ||||||
6714 | NumThreads = CGF.Builder.CreateSelect( | ||||||
6715 | CGF.Builder.CreateICmpULT(DefaultThreadLimitVal, NumThreads), | ||||||
6716 | DefaultThreadLimitVal, NumThreads); | ||||||
6717 | } else { | ||||||
6718 | NumThreads = DefaultThreadLimitVal ? DefaultThreadLimitVal | ||||||
6719 | : CGF.Builder.getInt32(0); | ||||||
6720 | } | ||||||
6721 | // Process condition of the if clause. | ||||||
6722 | if (CondVal) { | ||||||
6723 | NumThreads = CGF.Builder.CreateSelect(CondVal, NumThreads, | ||||||
6724 | CGF.Builder.getInt32(1)); | ||||||
6725 | } | ||||||
6726 | return NumThreads; | ||||||
6727 | } | ||||||
6728 | if (isOpenMPSimdDirective(Dir->getDirectiveKind())) | ||||||
6729 | return CGF.Builder.getInt32(1); | ||||||
6730 | return DefaultThreadLimitVal; | ||||||
6731 | } | ||||||
6732 | return DefaultThreadLimitVal ? DefaultThreadLimitVal | ||||||
6733 | : CGF.Builder.getInt32(0); | ||||||
6734 | } | ||||||
6735 | |||||||
6736 | /// Emit the number of threads for a target directive. Inspect the | ||||||
6737 | /// thread_limit clause associated with a teams construct combined or closely | ||||||
6738 | /// nested with the target directive. | ||||||
6739 | /// | ||||||
6740 | /// Emit the num_threads clause for directives such as 'target parallel' that | ||||||
6741 | /// have no associated teams construct. | ||||||
6742 | /// | ||||||
6743 | /// Otherwise, return nullptr. | ||||||
6744 | static llvm::Value * | ||||||
6745 | emitNumThreadsForTargetDirective(CodeGenFunction &CGF, | ||||||
6746 | const OMPExecutableDirective &D) { | ||||||
6747 | assert(!CGF.getLangOpts().OpenMPIsDevice &&((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6749, __PRETTY_FUNCTION__)) | ||||||
6748 | "Clauses associated with the teams directive expected to be emitted "((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6749, __PRETTY_FUNCTION__)) | ||||||
6749 | "only for the host!")((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6749, __PRETTY_FUNCTION__)); | ||||||
6750 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | ||||||
6751 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6752, __PRETTY_FUNCTION__)) | ||||||
6752 | "Expected target-based executable directive.")((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6752, __PRETTY_FUNCTION__)); | ||||||
6753 | CGBuilderTy &Bld = CGF.Builder; | ||||||
6754 | llvm::Value *ThreadLimitVal = nullptr; | ||||||
6755 | llvm::Value *NumThreadsVal = nullptr; | ||||||
6756 | switch (DirectiveKind) { | ||||||
6757 | case OMPD_target: { | ||||||
6758 | const CapturedStmt *CS = D.getInnermostCapturedStmt(); | ||||||
6759 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | ||||||
6760 | return NumThreads; | ||||||
6761 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | ||||||
6762 | CGF.getContext(), CS->getCapturedStmt()); | ||||||
6763 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | ||||||
6764 | if (Dir->hasClausesOfKind<OMPThreadLimitClause>()) { | ||||||
6765 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | ||||||
6766 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | ||||||
6767 | const auto *ThreadLimitClause = | ||||||
6768 | Dir->getSingleClause<OMPThreadLimitClause>(); | ||||||
6769 | CodeGenFunction::LexicalScope Scope( | ||||||
6770 | CGF, ThreadLimitClause->getThreadLimit()->getSourceRange()); | ||||||
6771 | if (const auto *PreInit = | ||||||
6772 | cast_or_null<DeclStmt>(ThreadLimitClause->getPreInitStmt())) { | ||||||
6773 | for (const auto *I : PreInit->decls()) { | ||||||
6774 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | ||||||
6775 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | ||||||
6776 | } else { | ||||||
6777 | CodeGenFunction::AutoVarEmission Emission = | ||||||
6778 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | ||||||
6779 | CGF.EmitAutoVarCleanups(Emission); | ||||||
6780 | } | ||||||
6781 | } | ||||||
6782 | } | ||||||
6783 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | ||||||
6784 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | ||||||
6785 | ThreadLimitVal = | ||||||
6786 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | ||||||
6787 | } | ||||||
6788 | if (isOpenMPTeamsDirective(Dir->getDirectiveKind()) && | ||||||
6789 | !isOpenMPDistributeDirective(Dir->getDirectiveKind())) { | ||||||
6790 | CS = Dir->getInnermostCapturedStmt(); | ||||||
6791 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | ||||||
6792 | CGF.getContext(), CS->getCapturedStmt()); | ||||||
6793 | Dir = dyn_cast_or_null<OMPExecutableDirective>(Child); | ||||||
6794 | } | ||||||
6795 | if (Dir && isOpenMPDistributeDirective(Dir->getDirectiveKind()) && | ||||||
6796 | !isOpenMPSimdDirective(Dir->getDirectiveKind())) { | ||||||
6797 | CS = Dir->getInnermostCapturedStmt(); | ||||||
6798 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | ||||||
6799 | return NumThreads; | ||||||
6800 | } | ||||||
6801 | if (Dir && isOpenMPSimdDirective(Dir->getDirectiveKind())) | ||||||
6802 | return Bld.getInt32(1); | ||||||
6803 | } | ||||||
6804 | return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0); | ||||||
6805 | } | ||||||
6806 | case OMPD_target_teams: { | ||||||
6807 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | ||||||
6808 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | ||||||
6809 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | ||||||
6810 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | ||||||
6811 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | ||||||
6812 | ThreadLimitVal = | ||||||
6813 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | ||||||
6814 | } | ||||||
6815 | const CapturedStmt *CS = D.getInnermostCapturedStmt(); | ||||||
6816 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | ||||||
6817 | return NumThreads; | ||||||
6818 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | ||||||
6819 | CGF.getContext(), CS->getCapturedStmt()); | ||||||
6820 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | ||||||
6821 | if (Dir->getDirectiveKind() == OMPD_distribute) { | ||||||
6822 | CS = Dir->getInnermostCapturedStmt(); | ||||||
6823 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | ||||||
6824 | return NumThreads; | ||||||
6825 | } | ||||||
6826 | } | ||||||
6827 | return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0); | ||||||
6828 | } | ||||||
6829 | case OMPD_target_teams_distribute: | ||||||
6830 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | ||||||
6831 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | ||||||
6832 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | ||||||
6833 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | ||||||
6834 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | ||||||
6835 | ThreadLimitVal = | ||||||
6836 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | ||||||
6837 | } | ||||||
6838 | return getNumThreads(CGF, D.getInnermostCapturedStmt(), ThreadLimitVal); | ||||||
6839 | case OMPD_target_parallel: | ||||||
6840 | case OMPD_target_parallel_for: | ||||||
6841 | case OMPD_target_parallel_for_simd: | ||||||
6842 | case OMPD_target_teams_distribute_parallel_for: | ||||||
6843 | case OMPD_target_teams_distribute_parallel_for_simd: { | ||||||
6844 | llvm::Value *CondVal = nullptr; | ||||||
6845 | // Handle if clause. If if clause present, the number of threads is | ||||||
6846 | // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1. | ||||||
6847 | if (D.hasClausesOfKind<OMPIfClause>()) { | ||||||
6848 | const OMPIfClause *IfClause = nullptr; | ||||||
6849 | for (const auto *C : D.getClausesOfKind<OMPIfClause>()) { | ||||||
6850 | if (C->getNameModifier() == OMPD_unknown || | ||||||
6851 | C->getNameModifier() == OMPD_parallel) { | ||||||
6852 | IfClause = C; | ||||||
6853 | break; | ||||||
6854 | } | ||||||
6855 | } | ||||||
6856 | if (IfClause) { | ||||||
6857 | const Expr *Cond = IfClause->getCondition(); | ||||||
6858 | bool Result; | ||||||
6859 | if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) { | ||||||
6860 | if (!Result) | ||||||
6861 | return Bld.getInt32(1); | ||||||
6862 | } else { | ||||||
6863 | CodeGenFunction::RunCleanupsScope Scope(CGF); | ||||||
6864 | CondVal = CGF.EvaluateExprAsBool(Cond); | ||||||
6865 | } | ||||||
6866 | } | ||||||
6867 | } | ||||||
6868 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | ||||||
6869 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | ||||||
6870 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | ||||||
6871 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | ||||||
6872 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | ||||||
6873 | ThreadLimitVal = | ||||||
6874 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false); | ||||||
6875 | } | ||||||
6876 | if (D.hasClausesOfKind<OMPNumThreadsClause>()) { | ||||||
6877 | CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF); | ||||||
6878 | const auto *NumThreadsClause = D.getSingleClause<OMPNumThreadsClause>(); | ||||||
6879 | llvm::Value *NumThreads = CGF.EmitScalarExpr( | ||||||
6880 | NumThreadsClause->getNumThreads(), /*IgnoreResultAssign=*/true); | ||||||
6881 | NumThreadsVal = | ||||||
6882 | Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned=*/false); | ||||||
6883 | ThreadLimitVal = ThreadLimitVal | ||||||
6884 | ? Bld.CreateSelect(Bld.CreateICmpULT(NumThreadsVal, | ||||||
6885 | ThreadLimitVal), | ||||||
6886 | NumThreadsVal, ThreadLimitVal) | ||||||
6887 | : NumThreadsVal; | ||||||
6888 | } | ||||||
6889 | if (!ThreadLimitVal) | ||||||
6890 | ThreadLimitVal = Bld.getInt32(0); | ||||||
6891 | if (CondVal) | ||||||
6892 | return Bld.CreateSelect(CondVal, ThreadLimitVal, Bld.getInt32(1)); | ||||||
6893 | return ThreadLimitVal; | ||||||
6894 | } | ||||||
6895 | case OMPD_target_teams_distribute_simd: | ||||||
6896 | case OMPD_target_simd: | ||||||
6897 | return Bld.getInt32(1); | ||||||
6898 | case OMPD_parallel: | ||||||
6899 | case OMPD_for: | ||||||
6900 | case OMPD_parallel_for: | ||||||
6901 | case OMPD_parallel_master: | ||||||
6902 | case OMPD_parallel_sections: | ||||||
6903 | case OMPD_for_simd: | ||||||
6904 | case OMPD_parallel_for_simd: | ||||||
6905 | case OMPD_cancel: | ||||||
6906 | case OMPD_cancellation_point: | ||||||
6907 | case OMPD_ordered: | ||||||
6908 | case OMPD_threadprivate: | ||||||
6909 | case OMPD_allocate: | ||||||
6910 | case OMPD_task: | ||||||
6911 | case OMPD_simd: | ||||||
6912 | case OMPD_sections: | ||||||
6913 | case OMPD_section: | ||||||
6914 | case OMPD_single: | ||||||
6915 | case OMPD_master: | ||||||
6916 | case OMPD_critical: | ||||||
6917 | case OMPD_taskyield: | ||||||
6918 | case OMPD_barrier: | ||||||
6919 | case OMPD_taskwait: | ||||||
6920 | case OMPD_taskgroup: | ||||||
6921 | case OMPD_atomic: | ||||||
6922 | case OMPD_flush: | ||||||
6923 | case OMPD_depobj: | ||||||
6924 | case OMPD_scan: | ||||||
6925 | case OMPD_teams: | ||||||
6926 | case OMPD_target_data: | ||||||
6927 | case OMPD_target_exit_data: | ||||||
6928 | case OMPD_target_enter_data: | ||||||
6929 | case OMPD_distribute: | ||||||
6930 | case OMPD_distribute_simd: | ||||||
6931 | case OMPD_distribute_parallel_for: | ||||||
6932 | case OMPD_distribute_parallel_for_simd: | ||||||
6933 | case OMPD_teams_distribute: | ||||||
6934 | case OMPD_teams_distribute_simd: | ||||||
6935 | case OMPD_teams_distribute_parallel_for: | ||||||
6936 | case OMPD_teams_distribute_parallel_for_simd: | ||||||
6937 | case OMPD_target_update: | ||||||
6938 | case OMPD_declare_simd: | ||||||
6939 | case OMPD_declare_variant: | ||||||
6940 | case OMPD_begin_declare_variant: | ||||||
6941 | case OMPD_end_declare_variant: | ||||||
6942 | case OMPD_declare_target: | ||||||
6943 | case OMPD_end_declare_target: | ||||||
6944 | case OMPD_declare_reduction: | ||||||
6945 | case OMPD_declare_mapper: | ||||||
6946 | case OMPD_taskloop: | ||||||
6947 | case OMPD_taskloop_simd: | ||||||
6948 | case OMPD_master_taskloop: | ||||||
6949 | case OMPD_master_taskloop_simd: | ||||||
6950 | case OMPD_parallel_master_taskloop: | ||||||
6951 | case OMPD_parallel_master_taskloop_simd: | ||||||
6952 | case OMPD_requires: | ||||||
6953 | case OMPD_unknown: | ||||||
6954 | break; | ||||||
6955 | default: | ||||||
6956 | break; | ||||||
6957 | } | ||||||
6958 | llvm_unreachable("Unsupported directive kind.")::llvm::llvm_unreachable_internal("Unsupported directive kind." , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6958); | ||||||
6959 | } | ||||||
6960 | |||||||
6961 | namespace { | ||||||
6962 | 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^=; | ||||||
6963 | |||||||
6964 | // Utility to handle information from clauses associated with a given | ||||||
6965 | // construct that use mappable expressions (e.g. 'map' clause, 'to' clause). | ||||||
6966 | // It provides a convenient interface to obtain the information and generate | ||||||
6967 | // code for that information. | ||||||
6968 | class MappableExprsHandler { | ||||||
6969 | public: | ||||||
6970 | /// Values for bit flags used to specify the mapping type for | ||||||
6971 | /// offloading. | ||||||
6972 | enum OpenMPOffloadMappingFlags : uint64_t { | ||||||
6973 | /// No flags | ||||||
6974 | OMP_MAP_NONE = 0x0, | ||||||
6975 | /// Allocate memory on the device and move data from host to device. | ||||||
6976 | OMP_MAP_TO = 0x01, | ||||||
6977 | /// Allocate memory on the device and move data from device to host. | ||||||
6978 | OMP_MAP_FROM = 0x02, | ||||||
6979 | /// Always perform the requested mapping action on the element, even | ||||||
6980 | /// if it was already mapped before. | ||||||
6981 | OMP_MAP_ALWAYS = 0x04, | ||||||
6982 | /// Delete the element from the device environment, ignoring the | ||||||
6983 | /// current reference count associated with the element. | ||||||
6984 | OMP_MAP_DELETE = 0x08, | ||||||
6985 | /// The element being mapped is a pointer-pointee pair; both the | ||||||
6986 | /// pointer and the pointee should be mapped. | ||||||
6987 | OMP_MAP_PTR_AND_OBJ = 0x10, | ||||||
6988 | /// This flags signals that the base address of an entry should be | ||||||
6989 | /// passed to the target kernel as an argument. | ||||||
6990 | OMP_MAP_TARGET_PARAM = 0x20, | ||||||
6991 | /// Signal that the runtime library has to return the device pointer | ||||||
6992 | /// in the current position for the data being mapped. Used when we have the | ||||||
6993 | /// use_device_ptr or use_device_addr clause. | ||||||
6994 | OMP_MAP_RETURN_PARAM = 0x40, | ||||||
6995 | /// This flag signals that the reference being passed is a pointer to | ||||||
6996 | /// private data. | ||||||
6997 | OMP_MAP_PRIVATE = 0x80, | ||||||
6998 | /// Pass the element to the device by value. | ||||||
6999 | OMP_MAP_LITERAL = 0x100, | ||||||
7000 | /// Implicit map | ||||||
7001 | OMP_MAP_IMPLICIT = 0x200, | ||||||
7002 | /// Close is a hint to the runtime to allocate memory close to | ||||||
7003 | /// the target device. | ||||||
7004 | OMP_MAP_CLOSE = 0x400, | ||||||
7005 | /// 0x800 is reserved for compatibility with XLC. | ||||||
7006 | /// Produce a runtime error if the data is not already allocated. | ||||||
7007 | OMP_MAP_PRESENT = 0x1000, | ||||||
7008 | /// The 16 MSBs of the flags indicate whether the entry is member of some | ||||||
7009 | /// struct/class. | ||||||
7010 | OMP_MAP_MEMBER_OF = 0xffff000000000000, | ||||||
7011 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ OMP_MAP_MEMBER_OF)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_MAP_MEMBER_OF, | ||||||
7012 | }; | ||||||
7013 | |||||||
7014 | /// Get the offset of the OMP_MAP_MEMBER_OF field. | ||||||
7015 | static unsigned getFlagMemberOffset() { | ||||||
7016 | unsigned Offset = 0; | ||||||
7017 | for (uint64_t Remain = OMP_MAP_MEMBER_OF; !(Remain & 1); | ||||||
7018 | Remain = Remain >> 1) | ||||||
7019 | Offset++; | ||||||
7020 | return Offset; | ||||||
7021 | } | ||||||
7022 | |||||||
7023 | /// Class that associates information with a base pointer to be passed to the | ||||||
7024 | /// runtime library. | ||||||
7025 | class BasePointerInfo { | ||||||
7026 | /// The base pointer. | ||||||
7027 | llvm::Value *Ptr = nullptr; | ||||||
7028 | /// The base declaration that refers to this device pointer, or null if | ||||||
7029 | /// there is none. | ||||||
7030 | const ValueDecl *DevPtrDecl = nullptr; | ||||||
7031 | |||||||
7032 | public: | ||||||
7033 | BasePointerInfo(llvm::Value *Ptr, const ValueDecl *DevPtrDecl = nullptr) | ||||||
7034 | : Ptr(Ptr), DevPtrDecl(DevPtrDecl) {} | ||||||
7035 | llvm::Value *operator*() const { return Ptr; } | ||||||
7036 | const ValueDecl *getDevicePtrDecl() const { return DevPtrDecl; } | ||||||
7037 | void setDevicePtrDecl(const ValueDecl *D) { DevPtrDecl = D; } | ||||||
7038 | }; | ||||||
7039 | |||||||
7040 | using MapBaseValuesArrayTy = SmallVector<BasePointerInfo, 4>; | ||||||
7041 | using MapValuesArrayTy = SmallVector<llvm::Value *, 4>; | ||||||
7042 | using MapFlagsArrayTy = SmallVector<OpenMPOffloadMappingFlags, 4>; | ||||||
7043 | using MapMappersArrayTy = SmallVector<const ValueDecl *, 4>; | ||||||
7044 | |||||||
7045 | /// This structure contains combined information generated for mappable | ||||||
7046 | /// clauses, including base pointers, pointers, sizes, map types, and | ||||||
7047 | /// user-defined mappers. | ||||||
7048 | struct MapCombinedInfoTy { | ||||||
7049 | MapBaseValuesArrayTy BasePointers; | ||||||
7050 | MapValuesArrayTy Pointers; | ||||||
7051 | MapValuesArrayTy Sizes; | ||||||
7052 | MapFlagsArrayTy Types; | ||||||
7053 | MapMappersArrayTy Mappers; | ||||||
7054 | |||||||
7055 | /// Append arrays in \a CurInfo. | ||||||
7056 | void append(MapCombinedInfoTy &CurInfo) { | ||||||
7057 | BasePointers.append(CurInfo.BasePointers.begin(), | ||||||
7058 | CurInfo.BasePointers.end()); | ||||||
7059 | Pointers.append(CurInfo.Pointers.begin(), CurInfo.Pointers.end()); | ||||||
7060 | Sizes.append(CurInfo.Sizes.begin(), CurInfo.Sizes.end()); | ||||||
7061 | Types.append(CurInfo.Types.begin(), CurInfo.Types.end()); | ||||||
7062 | Mappers.append(CurInfo.Mappers.begin(), CurInfo.Mappers.end()); | ||||||
7063 | } | ||||||
7064 | }; | ||||||
7065 | |||||||
7066 | /// Map between a struct and the its lowest & highest elements which have been | ||||||
7067 | /// mapped. | ||||||
7068 | /// [ValueDecl *] --> {LE(FieldIndex, Pointer), | ||||||
7069 | /// HE(FieldIndex, Pointer)} | ||||||
7070 | struct StructRangeInfoTy { | ||||||
7071 | std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> LowestElem = { | ||||||
7072 | 0, Address::invalid()}; | ||||||
7073 | std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> HighestElem = { | ||||||
7074 | 0, Address::invalid()}; | ||||||
7075 | Address Base = Address::invalid(); | ||||||
7076 | }; | ||||||
7077 | |||||||
7078 | private: | ||||||
7079 | /// Kind that defines how a device pointer has to be returned. | ||||||
7080 | struct MapInfo { | ||||||
7081 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | ||||||
7082 | OpenMPMapClauseKind MapType = OMPC_MAP_unknown; | ||||||
7083 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | ||||||
7084 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers; | ||||||
7085 | bool ReturnDevicePointer = false; | ||||||
7086 | bool IsImplicit = false; | ||||||
7087 | const ValueDecl *Mapper = nullptr; | ||||||
7088 | bool ForDeviceAddr = false; | ||||||
7089 | |||||||
7090 | MapInfo() = default; | ||||||
7091 | MapInfo( | ||||||
7092 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components, | ||||||
7093 | OpenMPMapClauseKind MapType, | ||||||
7094 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | ||||||
7095 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, | ||||||
7096 | bool ReturnDevicePointer, bool IsImplicit, | ||||||
7097 | const ValueDecl *Mapper = nullptr, bool ForDeviceAddr = false) | ||||||
7098 | : Components(Components), MapType(MapType), MapModifiers(MapModifiers), | ||||||
7099 | MotionModifiers(MotionModifiers), | ||||||
7100 | ReturnDevicePointer(ReturnDevicePointer), IsImplicit(IsImplicit), | ||||||
7101 | Mapper(Mapper), ForDeviceAddr(ForDeviceAddr) {} | ||||||
7102 | }; | ||||||
7103 | |||||||
7104 | /// If use_device_ptr or use_device_addr is used on a decl which is a struct | ||||||
7105 | /// member and there is no map information about it, then emission of that | ||||||
7106 | /// entry is deferred until the whole struct has been processed. | ||||||
7107 | struct DeferredDevicePtrEntryTy { | ||||||
7108 | const Expr *IE = nullptr; | ||||||
7109 | const ValueDecl *VD = nullptr; | ||||||
7110 | bool ForDeviceAddr = false; | ||||||
7111 | |||||||
7112 | DeferredDevicePtrEntryTy(const Expr *IE, const ValueDecl *VD, | ||||||
7113 | bool ForDeviceAddr) | ||||||
7114 | : IE(IE), VD(VD), ForDeviceAddr(ForDeviceAddr) {} | ||||||
7115 | }; | ||||||
7116 | |||||||
7117 | /// The target directive from where the mappable clauses were extracted. It | ||||||
7118 | /// is either a executable directive or a user-defined mapper directive. | ||||||
7119 | llvm::PointerUnion<const OMPExecutableDirective *, | ||||||
7120 | const OMPDeclareMapperDecl *> | ||||||
7121 | CurDir; | ||||||
7122 | |||||||
7123 | /// Function the directive is being generated for. | ||||||
7124 | CodeGenFunction &CGF; | ||||||
7125 | |||||||
7126 | /// Set of all first private variables in the current directive. | ||||||
7127 | /// bool data is set to true if the variable is implicitly marked as | ||||||
7128 | /// firstprivate, false otherwise. | ||||||
7129 | llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, bool> FirstPrivateDecls; | ||||||
7130 | |||||||
7131 | /// Map between device pointer declarations and their expression components. | ||||||
7132 | /// The key value for declarations in 'this' is null. | ||||||
7133 | llvm::DenseMap< | ||||||
7134 | const ValueDecl *, | ||||||
7135 | SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>> | ||||||
7136 | DevPointersMap; | ||||||
7137 | |||||||
7138 | llvm::Value *getExprTypeSize(const Expr *E) const { | ||||||
7139 | QualType ExprTy = E->getType().getCanonicalType(); | ||||||
7140 | |||||||
7141 | // Calculate the size for array shaping expression. | ||||||
7142 | if (const auto *OAE = dyn_cast<OMPArrayShapingExpr>(E)) { | ||||||
7143 | llvm::Value *Size = | ||||||
7144 | CGF.getTypeSize(OAE->getBase()->getType()->getPointeeType()); | ||||||
7145 | for (const Expr *SE : OAE->getDimensions()) { | ||||||
7146 | llvm::Value *Sz = CGF.EmitScalarExpr(SE); | ||||||
7147 | Sz = CGF.EmitScalarConversion(Sz, SE->getType(), | ||||||
7148 | CGF.getContext().getSizeType(), | ||||||
7149 | SE->getExprLoc()); | ||||||
7150 | Size = CGF.Builder.CreateNUWMul(Size, Sz); | ||||||
7151 | } | ||||||
7152 | return Size; | ||||||
7153 | } | ||||||
7154 | |||||||
7155 | // Reference types are ignored for mapping purposes. | ||||||
7156 | if (const auto *RefTy = ExprTy->getAs<ReferenceType>()) | ||||||
7157 | ExprTy = RefTy->getPointeeType().getCanonicalType(); | ||||||
7158 | |||||||
7159 | // Given that an array section is considered a built-in type, we need to | ||||||
7160 | // do the calculation based on the length of the section instead of relying | ||||||
7161 | // on CGF.getTypeSize(E->getType()). | ||||||
7162 | if (const auto *OAE = dyn_cast<OMPArraySectionExpr>(E)) { | ||||||
7163 | QualType BaseTy = OMPArraySectionExpr::getBaseOriginalType( | ||||||
7164 | OAE->getBase()->IgnoreParenImpCasts()) | ||||||
7165 | .getCanonicalType(); | ||||||
7166 | |||||||
7167 | // If there is no length associated with the expression and lower bound is | ||||||
7168 | // not specified too, that means we are using the whole length of the | ||||||
7169 | // base. | ||||||
7170 | if (!OAE->getLength() && OAE->getColonLocFirst().isValid() && | ||||||
7171 | !OAE->getLowerBound()) | ||||||
7172 | return CGF.getTypeSize(BaseTy); | ||||||
7173 | |||||||
7174 | llvm::Value *ElemSize; | ||||||
7175 | if (const auto *PTy = BaseTy->getAs<PointerType>()) { | ||||||
7176 | ElemSize = CGF.getTypeSize(PTy->getPointeeType().getCanonicalType()); | ||||||
7177 | } else { | ||||||
7178 | const auto *ATy = cast<ArrayType>(BaseTy.getTypePtr()); | ||||||
7179 | assert(ATy && "Expecting array type if not a pointer type.")((ATy && "Expecting array type if not a pointer type." ) ? static_cast<void> (0) : __assert_fail ("ATy && \"Expecting array type if not a pointer type.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7179, __PRETTY_FUNCTION__)); | ||||||
7180 | ElemSize = CGF.getTypeSize(ATy->getElementType().getCanonicalType()); | ||||||
7181 | } | ||||||
7182 | |||||||
7183 | // If we don't have a length at this point, that is because we have an | ||||||
7184 | // array section with a single element. | ||||||
7185 | if (!OAE->getLength() && OAE->getColonLocFirst().isInvalid()) | ||||||
7186 | return ElemSize; | ||||||
7187 | |||||||
7188 | if (const Expr *LenExpr = OAE->getLength()) { | ||||||
7189 | llvm::Value *LengthVal = CGF.EmitScalarExpr(LenExpr); | ||||||
7190 | LengthVal = CGF.EmitScalarConversion(LengthVal, LenExpr->getType(), | ||||||
7191 | CGF.getContext().getSizeType(), | ||||||
7192 | LenExpr->getExprLoc()); | ||||||
7193 | return CGF.Builder.CreateNUWMul(LengthVal, ElemSize); | ||||||
7194 | } | ||||||
7195 | assert(!OAE->getLength() && OAE->getColonLocFirst().isValid() &&((!OAE->getLength() && OAE->getColonLocFirst(). isValid() && OAE->getLowerBound() && "expected array_section[lb:]." ) ? static_cast<void> (0) : __assert_fail ("!OAE->getLength() && OAE->getColonLocFirst().isValid() && OAE->getLowerBound() && \"expected array_section[lb:].\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7196, __PRETTY_FUNCTION__)) | ||||||
7196 | OAE->getLowerBound() && "expected array_section[lb:].")((!OAE->getLength() && OAE->getColonLocFirst(). isValid() && OAE->getLowerBound() && "expected array_section[lb:]." ) ? static_cast<void> (0) : __assert_fail ("!OAE->getLength() && OAE->getColonLocFirst().isValid() && OAE->getLowerBound() && \"expected array_section[lb:].\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7196, __PRETTY_FUNCTION__)); | ||||||
7197 | // Size = sizetype - lb * elemtype; | ||||||
7198 | llvm::Value *LengthVal = CGF.getTypeSize(BaseTy); | ||||||
7199 | llvm::Value *LBVal = CGF.EmitScalarExpr(OAE->getLowerBound()); | ||||||
7200 | LBVal = CGF.EmitScalarConversion(LBVal, OAE->getLowerBound()->getType(), | ||||||
7201 | CGF.getContext().getSizeType(), | ||||||
7202 | OAE->getLowerBound()->getExprLoc()); | ||||||
7203 | LBVal = CGF.Builder.CreateNUWMul(LBVal, ElemSize); | ||||||
7204 | llvm::Value *Cmp = CGF.Builder.CreateICmpUGT(LengthVal, LBVal); | ||||||
7205 | llvm::Value *TrueVal = CGF.Builder.CreateNUWSub(LengthVal, LBVal); | ||||||
7206 | LengthVal = CGF.Builder.CreateSelect( | ||||||
7207 | Cmp, TrueVal, llvm::ConstantInt::get(CGF.SizeTy, 0)); | ||||||
7208 | return LengthVal; | ||||||
7209 | } | ||||||
7210 | return CGF.getTypeSize(ExprTy); | ||||||
7211 | } | ||||||
7212 | |||||||
7213 | /// Return the corresponding bits for a given map clause modifier. Add | ||||||
7214 | /// a flag marking the map as a pointer if requested. Add a flag marking the | ||||||
7215 | /// map as the first one of a series of maps that relate to the same map | ||||||
7216 | /// expression. | ||||||
7217 | OpenMPOffloadMappingFlags getMapTypeBits( | ||||||
7218 | OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers, | ||||||
7219 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, bool IsImplicit, | ||||||
7220 | bool AddPtrFlag, bool AddIsTargetParamFlag) const { | ||||||
7221 | OpenMPOffloadMappingFlags Bits = | ||||||
7222 | IsImplicit ? OMP_MAP_IMPLICIT : OMP_MAP_NONE; | ||||||
7223 | switch (MapType) { | ||||||
7224 | case OMPC_MAP_alloc: | ||||||
7225 | case OMPC_MAP_release: | ||||||
7226 | // alloc and release is the default behavior in the runtime library, i.e. | ||||||
7227 | // if we don't pass any bits alloc/release that is what the runtime is | ||||||
7228 | // going to do. Therefore, we don't need to signal anything for these two | ||||||
7229 | // type modifiers. | ||||||
7230 | break; | ||||||
7231 | case OMPC_MAP_to: | ||||||
7232 | Bits |= OMP_MAP_TO; | ||||||
7233 | break; | ||||||
7234 | case OMPC_MAP_from: | ||||||
7235 | Bits |= OMP_MAP_FROM; | ||||||
7236 | break; | ||||||
7237 | case OMPC_MAP_tofrom: | ||||||
7238 | Bits |= OMP_MAP_TO | OMP_MAP_FROM; | ||||||
7239 | break; | ||||||
7240 | case OMPC_MAP_delete: | ||||||
7241 | Bits |= OMP_MAP_DELETE; | ||||||
7242 | break; | ||||||
7243 | case OMPC_MAP_unknown: | ||||||
7244 | llvm_unreachable("Unexpected map type!")::llvm::llvm_unreachable_internal("Unexpected map type!", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7244); | ||||||
7245 | } | ||||||
7246 | if (AddPtrFlag) | ||||||
7247 | Bits |= OMP_MAP_PTR_AND_OBJ; | ||||||
7248 | if (AddIsTargetParamFlag) | ||||||
7249 | Bits |= OMP_MAP_TARGET_PARAM; | ||||||
7250 | if (llvm::find(MapModifiers, OMPC_MAP_MODIFIER_always) | ||||||
7251 | != MapModifiers.end()) | ||||||
7252 | Bits |= OMP_MAP_ALWAYS; | ||||||
7253 | if (llvm::find(MapModifiers, OMPC_MAP_MODIFIER_close) | ||||||
7254 | != MapModifiers.end()) | ||||||
7255 | Bits |= OMP_MAP_CLOSE; | ||||||
7256 | if (llvm::find(MapModifiers, OMPC_MAP_MODIFIER_present) | ||||||
7257 | != MapModifiers.end()) | ||||||
7258 | Bits |= OMP_MAP_PRESENT; | ||||||
7259 | if (llvm::find(MotionModifiers, OMPC_MOTION_MODIFIER_present) | ||||||
7260 | != MotionModifiers.end()) | ||||||
7261 | Bits |= OMP_MAP_PRESENT; | ||||||
7262 | return Bits; | ||||||
7263 | } | ||||||
7264 | |||||||
7265 | /// Return true if the provided expression is a final array section. A | ||||||
7266 | /// final array section, is one whose length can't be proved to be one. | ||||||
7267 | bool isFinalArraySectionExpression(const Expr *E) const { | ||||||
7268 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(E); | ||||||
7269 | |||||||
7270 | // It is not an array section and therefore not a unity-size one. | ||||||
7271 | if (!OASE) | ||||||
7272 | return false; | ||||||
7273 | |||||||
7274 | // An array section with no colon always refer to a single element. | ||||||
7275 | if (OASE->getColonLocFirst().isInvalid()) | ||||||
7276 | return false; | ||||||
7277 | |||||||
7278 | const Expr *Length = OASE->getLength(); | ||||||
7279 | |||||||
7280 | // If we don't have a length we have to check if the array has size 1 | ||||||
7281 | // for this dimension. Also, we should always expect a length if the | ||||||
7282 | // base type is pointer. | ||||||
7283 | if (!Length) { | ||||||
7284 | QualType BaseQTy = OMPArraySectionExpr::getBaseOriginalType( | ||||||
7285 | OASE->getBase()->IgnoreParenImpCasts()) | ||||||
7286 | .getCanonicalType(); | ||||||
7287 | if (const auto *ATy = dyn_cast<ConstantArrayType>(BaseQTy.getTypePtr())) | ||||||
7288 | return ATy->getSize().getSExtValue() != 1; | ||||||
7289 | // If we don't have a constant dimension length, we have to consider | ||||||
7290 | // the current section as having any size, so it is not necessarily | ||||||
7291 | // unitary. If it happen to be unity size, that's user fault. | ||||||
7292 | return true; | ||||||
7293 | } | ||||||
7294 | |||||||
7295 | // Check if the length evaluates to 1. | ||||||
7296 | Expr::EvalResult Result; | ||||||
7297 | if (!Length->EvaluateAsInt(Result, CGF.getContext())) | ||||||
7298 | return true; // Can have more that size 1. | ||||||
7299 | |||||||
7300 | llvm::APSInt ConstLength = Result.Val.getInt(); | ||||||
7301 | return ConstLength.getSExtValue() != 1; | ||||||
7302 | } | ||||||
7303 | |||||||
7304 | /// Generate the base pointers, section pointers, sizes, map type bits, and | ||||||
7305 | /// user-defined mappers (all included in \a CombinedInfo) for the provided | ||||||
7306 | /// map type, map or motion modifiers, and expression components. | ||||||
7307 | /// \a IsFirstComponent should be set to true if the provided set of | ||||||
7308 | /// components is the first associated with a capture. | ||||||
7309 | void generateInfoForComponentList( | ||||||
7310 | OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers, | ||||||
7311 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, | ||||||
7312 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components, | ||||||
7313 | MapCombinedInfoTy &CombinedInfo, StructRangeInfoTy &PartialStruct, | ||||||
7314 | bool IsFirstComponentList, bool IsImplicit, | ||||||
7315 | const ValueDecl *Mapper = nullptr, bool ForDeviceAddr = false, | ||||||
7316 | ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> | ||||||
7317 | OverlappedElements = llvm::None) const { | ||||||
7318 | // The following summarizes what has to be generated for each map and the | ||||||
7319 | // types below. The generated information is expressed in this order: | ||||||
7320 | // base pointer, section pointer, size, flags | ||||||
7321 | // (to add to the ones that come from the map type and modifier). | ||||||
7322 | // | ||||||
7323 | // double d; | ||||||
7324 | // int i[100]; | ||||||
7325 | // float *p; | ||||||
7326 | // | ||||||
7327 | // struct S1 { | ||||||
7328 | // int i; | ||||||
7329 | // float f[50]; | ||||||
7330 | // } | ||||||
7331 | // struct S2 { | ||||||
7332 | // int i; | ||||||
7333 | // float f[50]; | ||||||
7334 | // S1 s; | ||||||
7335 | // double *p; | ||||||
7336 | // struct S2 *ps; | ||||||
7337 | // } | ||||||
7338 | // S2 s; | ||||||
7339 | // S2 *ps; | ||||||
7340 | // | ||||||
7341 | // map(d) | ||||||
7342 | // &d, &d, sizeof(double), TARGET_PARAM | TO | FROM | ||||||
7343 | // | ||||||
7344 | // map(i) | ||||||
7345 | // &i, &i, 100*sizeof(int), TARGET_PARAM | TO | FROM | ||||||
7346 | // | ||||||
7347 | // map(i[1:23]) | ||||||
7348 | // &i(=&i[0]), &i[1], 23*sizeof(int), TARGET_PARAM | TO | FROM | ||||||
7349 | // | ||||||
7350 | // map(p) | ||||||
7351 | // &p, &p, sizeof(float*), TARGET_PARAM | TO | FROM | ||||||
7352 | // | ||||||
7353 | // map(p[1:24]) | ||||||
7354 | // &p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM | PTR_AND_OBJ | ||||||
7355 | // in unified shared memory mode or for local pointers | ||||||
7356 | // p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM | ||||||
7357 | // | ||||||
7358 | // map(s) | ||||||
7359 | // &s, &s, sizeof(S2), TARGET_PARAM | TO | FROM | ||||||
7360 | // | ||||||
7361 | // map(s.i) | ||||||
7362 | // &s, &(s.i), sizeof(int), TARGET_PARAM | TO | FROM | ||||||
7363 | // | ||||||
7364 | // map(s.s.f) | ||||||
7365 | // &s, &(s.s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM | ||||||
7366 | // | ||||||
7367 | // map(s.p) | ||||||
7368 | // &s, &(s.p), sizeof(double*), TARGET_PARAM | TO | FROM | ||||||
7369 | // | ||||||
7370 | // map(to: s.p[:22]) | ||||||
7371 | // &s, &(s.p), sizeof(double*), TARGET_PARAM (*) | ||||||
7372 | // &s, &(s.p), sizeof(double*), MEMBER_OF(1) (**) | ||||||
7373 | // &(s.p), &(s.p[0]), 22*sizeof(double), | ||||||
7374 | // MEMBER_OF(1) | PTR_AND_OBJ | TO (***) | ||||||
7375 | // (*) alloc space for struct members, only this is a target parameter | ||||||
7376 | // (**) map the pointer (nothing to be mapped in this example) (the compiler | ||||||
7377 | // optimizes this entry out, same in the examples below) | ||||||
7378 | // (***) map the pointee (map: to) | ||||||
7379 | // | ||||||
7380 | // map(s.ps) | ||||||
7381 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | TO | FROM | ||||||
7382 | // | ||||||
7383 | // map(from: s.ps->s.i) | ||||||
7384 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | ||||||
7385 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7386 | // &(s.ps), &(s.ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM | ||||||
7387 | // | ||||||
7388 | // map(to: s.ps->ps) | ||||||
7389 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | ||||||
7390 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7391 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | TO | ||||||
7392 | // | ||||||
7393 | // map(s.ps->ps->ps) | ||||||
7394 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | ||||||
7395 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7396 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | ||||||
7397 | // &(s.ps->ps), &(s.ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM | ||||||
7398 | // | ||||||
7399 | // map(to: s.ps->ps->s.f[:22]) | ||||||
7400 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | ||||||
7401 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7402 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | ||||||
7403 | // &(s.ps->ps), &(s.ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO | ||||||
7404 | // | ||||||
7405 | // map(ps) | ||||||
7406 | // &ps, &ps, sizeof(S2*), TARGET_PARAM | TO | FROM | ||||||
7407 | // | ||||||
7408 | // map(ps->i) | ||||||
7409 | // ps, &(ps->i), sizeof(int), TARGET_PARAM | TO | FROM | ||||||
7410 | // | ||||||
7411 | // map(ps->s.f) | ||||||
7412 | // ps, &(ps->s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM | ||||||
7413 | // | ||||||
7414 | // map(from: ps->p) | ||||||
7415 | // ps, &(ps->p), sizeof(double*), TARGET_PARAM | FROM | ||||||
7416 | // | ||||||
7417 | // map(to: ps->p[:22]) | ||||||
7418 | // ps, &(ps->p), sizeof(double*), TARGET_PARAM | ||||||
7419 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(1) | ||||||
7420 | // &(ps->p), &(ps->p[0]), 22*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | TO | ||||||
7421 | // | ||||||
7422 | // map(ps->ps) | ||||||
7423 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | TO | FROM | ||||||
7424 | // | ||||||
7425 | // map(from: ps->ps->s.i) | ||||||
7426 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | ||||||
7427 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7428 | // &(ps->ps), &(ps->ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM | ||||||
7429 | // | ||||||
7430 | // map(from: ps->ps->ps) | ||||||
7431 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | ||||||
7432 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7433 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | FROM | ||||||
7434 | // | ||||||
7435 | // map(ps->ps->ps->ps) | ||||||
7436 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | ||||||
7437 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7438 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | ||||||
7439 | // &(ps->ps->ps), &(ps->ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM | ||||||
7440 | // | ||||||
7441 | // map(to: ps->ps->ps->s.f[:22]) | ||||||
7442 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | ||||||
7443 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | ||||||
7444 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | ||||||
7445 | // &(ps->ps->ps), &(ps->ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO | ||||||
7446 | // | ||||||
7447 | // map(to: s.f[:22]) map(from: s.p[:33]) | ||||||
7448 | // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1) + | ||||||
7449 | // sizeof(double*) (**), TARGET_PARAM | ||||||
7450 | // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | TO | ||||||
7451 | // &s, &(s.p), sizeof(double*), MEMBER_OF(1) | ||||||
7452 | // &(s.p), &(s.p[0]), 33*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | FROM | ||||||
7453 | // (*) allocate contiguous space needed to fit all mapped members even if | ||||||
7454 | // we allocate space for members not mapped (in this example, | ||||||
7455 | // s.f[22..49] and s.s are not mapped, yet we must allocate space for | ||||||
7456 | // them as well because they fall between &s.f[0] and &s.p) | ||||||
7457 | // | ||||||
7458 | // map(from: s.f[:22]) map(to: ps->p[:33]) | ||||||
7459 | // &s, &(s.f[0]), 22*sizeof(float), TARGET_PARAM | FROM | ||||||
7460 | // ps, &(ps->p), sizeof(S2*), TARGET_PARAM | ||||||
7461 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(2) (*) | ||||||
7462 | // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(2) | PTR_AND_OBJ | TO | ||||||
7463 | // (*) the struct this entry pertains to is the 2nd element in the list of | ||||||
7464 | // arguments, hence MEMBER_OF(2) | ||||||
7465 | // | ||||||
7466 | // map(from: s.f[:22], s.s) map(to: ps->p[:33]) | ||||||
7467 | // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1), TARGET_PARAM | ||||||
7468 | // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | FROM | ||||||
7469 | // &s, &(s.s), sizeof(struct S1), MEMBER_OF(1) | FROM | ||||||
7470 | // ps, &(ps->p), sizeof(S2*), TARGET_PARAM | ||||||
7471 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(4) (*) | ||||||
7472 | // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(4) | PTR_AND_OBJ | TO | ||||||
7473 | // (*) the struct this entry pertains to is the 4th element in the list | ||||||
7474 | // of arguments, hence MEMBER_OF(4) | ||||||
7475 | |||||||
7476 | // Track if the map information being generated is the first for a capture. | ||||||
7477 | bool IsCaptureFirstInfo = IsFirstComponentList; | ||||||
7478 | // When the variable is on a declare target link or in a to clause with | ||||||
7479 | // unified memory, a reference is needed to hold the host/device address | ||||||
7480 | // of the variable. | ||||||
7481 | bool RequiresReference = false; | ||||||
7482 | |||||||
7483 | // Scan the components from the base to the complete expression. | ||||||
7484 | auto CI = Components.rbegin(); | ||||||
7485 | auto CE = Components.rend(); | ||||||
7486 | auto I = CI; | ||||||
7487 | |||||||
7488 | // Track if the map information being generated is the first for a list of | ||||||
7489 | // components. | ||||||
7490 | bool IsExpressionFirstInfo = true; | ||||||
7491 | bool FirstPointerInComplexData = false; | ||||||
7492 | Address BP = Address::invalid(); | ||||||
7493 | const Expr *AssocExpr = I->getAssociatedExpression(); | ||||||
7494 | const auto *AE = dyn_cast<ArraySubscriptExpr>(AssocExpr); | ||||||
| |||||||
7495 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr); | ||||||
7496 | const auto *OAShE = dyn_cast<OMPArrayShapingExpr>(AssocExpr); | ||||||
7497 | |||||||
7498 | if (isa<MemberExpr>(AssocExpr)) { | ||||||
7499 | // The base is the 'this' pointer. The content of the pointer is going | ||||||
7500 | // to be the base of the field being mapped. | ||||||
7501 | BP = CGF.LoadCXXThisAddress(); | ||||||
7502 | } else if ((AE
| ||||||
7503 | (OASE
| ||||||
7504 | isa<CXXThisExpr>(OASE->getBase()->IgnoreParenImpCasts()))) { | ||||||
7505 | BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(CGF); | ||||||
7506 | } else if (OAShE
| ||||||
7507 | isa<CXXThisExpr>(OAShE->getBase()->IgnoreParenCasts())) { | ||||||
7508 | BP = Address( | ||||||
7509 | CGF.EmitScalarExpr(OAShE->getBase()), | ||||||
7510 | CGF.getContext().getTypeAlignInChars(OAShE->getBase()->getType())); | ||||||
7511 | } else { | ||||||
7512 | // The base is the reference to the variable. | ||||||
7513 | // BP = &Var. | ||||||
7514 | BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(CGF); | ||||||
7515 | if (const auto *VD
| ||||||
7516 | dyn_cast_or_null<VarDecl>(I->getAssociatedDeclaration())) { | ||||||
7517 | if (llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | ||||||
7518 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) { | ||||||
7519 | if ((*Res == OMPDeclareTargetDeclAttr::MT_Link) || | ||||||
7520 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | ||||||
7521 | CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory())) { | ||||||
7522 | RequiresReference = true; | ||||||
7523 | BP = CGF.CGM.getOpenMPRuntime().getAddrOfDeclareTargetVar(VD); | ||||||
7524 | } | ||||||
7525 | } | ||||||
7526 | } | ||||||
7527 | |||||||
7528 | // If the variable is a pointer and is being dereferenced (i.e. is not | ||||||
7529 | // the last component), the base has to be the pointer itself, not its | ||||||
7530 | // reference. References are ignored for mapping purposes. | ||||||
7531 | QualType Ty = | ||||||
7532 | I->getAssociatedDeclaration()->getType().getNonReferenceType(); | ||||||
7533 | if (Ty->isAnyPointerType() && std::next(I) != CE) { | ||||||
7534 | // No need to generate individual map information for the pointer, it | ||||||
7535 | // can be associated with the combined storage if shared memory mode is | ||||||
7536 | // active or the base declaration is not global variable. | ||||||
7537 | const auto *VD = dyn_cast<VarDecl>(I->getAssociatedDeclaration()); | ||||||
7538 | if (CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory() || | ||||||
7539 | !VD || VD->hasLocalStorage()) | ||||||
7540 | BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>()); | ||||||
7541 | else | ||||||
7542 | FirstPointerInComplexData = true; | ||||||
7543 | ++I; | ||||||
7544 | } | ||||||
7545 | } | ||||||
7546 | |||||||
7547 | // Track whether a component of the list should be marked as MEMBER_OF some | ||||||
7548 | // combined entry (for partial structs). Only the first PTR_AND_OBJ entry | ||||||
7549 | // in a component list should be marked as MEMBER_OF, all subsequent entries | ||||||
7550 | // do not belong to the base struct. E.g. | ||||||
7551 | // struct S2 s; | ||||||
7552 | // s.ps->ps->ps->f[:] | ||||||
7553 | // (1) (2) (3) (4) | ||||||
7554 | // ps(1) is a member pointer, ps(2) is a pointee of ps(1), so it is a | ||||||
7555 | // PTR_AND_OBJ entry; the PTR is ps(1), so MEMBER_OF the base struct. ps(3) | ||||||
7556 | // is the pointee of ps(2) which is not member of struct s, so it should not | ||||||
7557 | // be marked as such (it is still PTR_AND_OBJ). | ||||||
7558 | // The variable is initialized to false so that PTR_AND_OBJ entries which | ||||||
7559 | // are not struct members are not considered (e.g. array of pointers to | ||||||
7560 | // data). | ||||||
7561 | bool ShouldBeMemberOf = false; | ||||||
7562 | |||||||
7563 | // Variable keeping track of whether or not we have encountered a component | ||||||
7564 | // in the component list which is a member expression. Useful when we have a | ||||||
7565 | // pointer or a final array section, in which case it is the previous | ||||||
7566 | // component in the list which tells us whether we have a member expression. | ||||||
7567 | // E.g. X.f[:] | ||||||
7568 | // While processing the final array section "[:]" it is "f" which tells us | ||||||
7569 | // whether we are dealing with a member of a declared struct. | ||||||
7570 | const MemberExpr *EncounteredME = nullptr; | ||||||
7571 | |||||||
7572 | for (; I != CE; ++I) { | ||||||
7573 | // If the current component is member of a struct (parent struct) mark it. | ||||||
7574 | if (!EncounteredME
| ||||||
7575 | EncounteredME = dyn_cast<MemberExpr>(I->getAssociatedExpression()); | ||||||
7576 | // If we encounter a PTR_AND_OBJ entry from now on it should be marked | ||||||
7577 | // as MEMBER_OF the parent struct. | ||||||
7578 | if (EncounteredME
| ||||||
7579 | ShouldBeMemberOf = true; | ||||||
7580 | // Do not emit as complex pointer if this is actually not array-like | ||||||
7581 | // expression. | ||||||
7582 | if (FirstPointerInComplexData) { | ||||||
7583 | QualType Ty = std::prev(I) | ||||||
7584 | ->getAssociatedDeclaration() | ||||||
7585 | ->getType() | ||||||
7586 | .getNonReferenceType(); | ||||||
7587 | BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>()); | ||||||
7588 | FirstPointerInComplexData = false; | ||||||
7589 | } | ||||||
7590 | } | ||||||
7591 | } | ||||||
7592 | |||||||
7593 | auto Next = std::next(I); | ||||||
7594 | |||||||
7595 | // We need to generate the addresses and sizes if this is the last | ||||||
7596 | // component, if the component is a pointer or if it is an array section | ||||||
7597 | // whose length can't be proved to be one. If this is a pointer, it | ||||||
7598 | // becomes the base address for the following components. | ||||||
7599 | |||||||
7600 | // A final array section, is one whose length can't be proved to be one. | ||||||
7601 | bool IsFinalArraySection = | ||||||
7602 | isFinalArraySectionExpression(I->getAssociatedExpression()); | ||||||
7603 | |||||||
7604 | // Get information on whether the element is a pointer. Have to do a | ||||||
7605 | // special treatment for array sections given that they are built-in | ||||||
7606 | // types. | ||||||
7607 | const auto *OASE = | ||||||
7608 | dyn_cast<OMPArraySectionExpr>(I->getAssociatedExpression()); | ||||||
7609 | const auto *OAShE = | ||||||
7610 | dyn_cast<OMPArrayShapingExpr>(I->getAssociatedExpression()); | ||||||
7611 | const auto *UO = dyn_cast<UnaryOperator>(I->getAssociatedExpression()); | ||||||
7612 | const auto *BO = dyn_cast<BinaryOperator>(I->getAssociatedExpression()); | ||||||
7613 | bool IsPointer = | ||||||
7614 | OAShE
| ||||||
7615 | (OASE
| ||||||
7616 | .getCanonicalType() | ||||||
7617 | ->isAnyPointerType()) || | ||||||
7618 | I->getAssociatedExpression()->getType()->isAnyPointerType(); | ||||||
7619 | bool IsNonDerefPointer = IsPointer
| ||||||
7620 | |||||||
7621 | if (Next == CE || IsNonDerefPointer || IsFinalArraySection) { | ||||||
7622 | // If this is not the last component, we expect the pointer to be | ||||||
7623 | // associated with an array expression or member expression. | ||||||
7624 | assert((Next == CE ||(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7625 | isa<MemberExpr>(Next->getAssociatedExpression()) ||(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7626 | isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) ||(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7627 | isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) ||(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7628 | isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) ||(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7629 | isa<UnaryOperator>(Next->getAssociatedExpression()) ||(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7630 | isa<BinaryOperator>(Next->getAssociatedExpression())) &&(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)) | ||||||
7631 | "Unexpected expression")(((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") ? static_cast<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\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7631, __PRETTY_FUNCTION__)); | ||||||
7632 | |||||||
7633 | Address LB = Address::invalid(); | ||||||
7634 | if (OAShE
| ||||||
7635 | LB = Address(CGF.EmitScalarExpr(OAShE->getBase()), | ||||||
7636 | CGF.getContext().getTypeAlignInChars( | ||||||
7637 | OAShE->getBase()->getType())); | ||||||
7638 | } else { | ||||||
7639 | LB = CGF.EmitOMPSharedLValue(I->getAssociatedExpression()) | ||||||
7640 | .getAddress(CGF); | ||||||
7641 | } | ||||||
7642 | |||||||
7643 | // If this component is a pointer inside the base struct then we don't | ||||||
7644 | // need to create any entry for it - it will be combined with the object | ||||||
7645 | // it is pointing to into a single PTR_AND_OBJ entry. | ||||||
7646 | bool IsMemberPointerOrAddr = | ||||||
7647 | (IsPointer
| ||||||
7648 | (dyn_cast<MemberExpr>(I->getAssociatedExpression()) == | ||||||
7649 | EncounteredME); | ||||||
7650 | if (!OverlappedElements.empty()) { | ||||||
7651 | // Handle base element with the info for overlapped elements. | ||||||
7652 | assert(!PartialStruct.Base.isValid() && "The base element is set.")((!PartialStruct.Base.isValid() && "The base element is set." ) ? static_cast<void> (0) : __assert_fail ("!PartialStruct.Base.isValid() && \"The base element is set.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7652, __PRETTY_FUNCTION__)); | ||||||
7653 | assert(Next == CE &&((Next == CE && "Expected last element for the overlapped elements." ) ? static_cast<void> (0) : __assert_fail ("Next == CE && \"Expected last element for the overlapped elements.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7654, __PRETTY_FUNCTION__)) | ||||||
7654 | "Expected last element for the overlapped elements.")((Next == CE && "Expected last element for the overlapped elements." ) ? static_cast<void> (0) : __assert_fail ("Next == CE && \"Expected last element for the overlapped elements.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7654, __PRETTY_FUNCTION__)); | ||||||
7655 | assert(!IsPointer &&((!IsPointer && "Unexpected base element with the pointer type." ) ? static_cast<void> (0) : __assert_fail ("!IsPointer && \"Unexpected base element with the pointer type.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7656, __PRETTY_FUNCTION__)) | ||||||
7656 | "Unexpected base element with the pointer type.")((!IsPointer && "Unexpected base element with the pointer type." ) ? static_cast<void> (0) : __assert_fail ("!IsPointer && \"Unexpected base element with the pointer type.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7656, __PRETTY_FUNCTION__)); | ||||||
7657 | // Mark the whole struct as the struct that requires allocation on the | ||||||
7658 | // device. | ||||||
7659 | PartialStruct.LowestElem = {0, LB}; | ||||||
7660 | CharUnits TypeSize = CGF.getContext().getTypeSizeInChars( | ||||||
7661 | I->getAssociatedExpression()->getType()); | ||||||
7662 | Address HB = CGF.Builder.CreateConstGEP( | ||||||
7663 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(LB, | ||||||
7664 | CGF.VoidPtrTy), | ||||||
7665 | TypeSize.getQuantity() - 1); | ||||||
7666 | PartialStruct.HighestElem = { | ||||||
7667 | std::numeric_limits<decltype( | ||||||
7668 | PartialStruct.HighestElem.first)>::max(), | ||||||
7669 | HB}; | ||||||
7670 | PartialStruct.Base = BP; | ||||||
7671 | // Emit data for non-overlapped data. | ||||||
7672 | OpenMPOffloadMappingFlags Flags = | ||||||
7673 | OMP_MAP_MEMBER_OF | | ||||||
7674 | getMapTypeBits(MapType, MapModifiers, MotionModifiers, IsImplicit, | ||||||
7675 | /*AddPtrFlag=*/false, | ||||||
7676 | /*AddIsTargetParamFlag=*/false); | ||||||
7677 | LB = BP; | ||||||
7678 | llvm::Value *Size = nullptr; | ||||||
7679 | // Do bitcopy of all non-overlapped structure elements. | ||||||
7680 | for (OMPClauseMappableExprCommon::MappableExprComponentListRef | ||||||
7681 | Component : OverlappedElements) { | ||||||
7682 | Address ComponentLB = Address::invalid(); | ||||||
7683 | for (const OMPClauseMappableExprCommon::MappableComponent &MC : | ||||||
7684 | Component) { | ||||||
7685 | if (MC.getAssociatedDeclaration()) { | ||||||
7686 | ComponentLB = | ||||||
7687 | CGF.EmitOMPSharedLValue(MC.getAssociatedExpression()) | ||||||
7688 | .getAddress(CGF); | ||||||
7689 | Size = CGF.Builder.CreatePtrDiff( | ||||||
7690 | CGF.EmitCastToVoidPtr(ComponentLB.getPointer()), | ||||||
7691 | CGF.EmitCastToVoidPtr(LB.getPointer())); | ||||||
7692 | break; | ||||||
7693 | } | ||||||
7694 | } | ||||||
7695 | CombinedInfo.BasePointers.push_back(BP.getPointer()); | ||||||
7696 | CombinedInfo.Pointers.push_back(LB.getPointer()); | ||||||
7697 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | ||||||
7698 | Size, CGF.Int64Ty, /*isSigned=*/true)); | ||||||
7699 | CombinedInfo.Types.push_back(Flags); | ||||||
7700 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
7701 | LB = CGF.Builder.CreateConstGEP(ComponentLB, 1); | ||||||
7702 | } | ||||||
7703 | CombinedInfo.BasePointers.push_back(BP.getPointer()); | ||||||
7704 | CombinedInfo.Pointers.push_back(LB.getPointer()); | ||||||
7705 | Size = CGF.Builder.CreatePtrDiff( | ||||||
7706 | CGF.EmitCastToVoidPtr( | ||||||
7707 | CGF.Builder.CreateConstGEP(HB, 1).getPointer()), | ||||||
7708 | CGF.EmitCastToVoidPtr(LB.getPointer())); | ||||||
7709 | CombinedInfo.Sizes.push_back( | ||||||
7710 | CGF.Builder.CreateIntCast(Size, CGF.Int64Ty, /*isSigned=*/true)); | ||||||
7711 | CombinedInfo.Types.push_back(Flags); | ||||||
7712 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
7713 | break; | ||||||
7714 | } | ||||||
7715 | llvm::Value *Size = getExprTypeSize(I->getAssociatedExpression()); | ||||||
7716 | if (!IsMemberPointerOrAddr) { | ||||||
7717 | CombinedInfo.BasePointers.push_back(BP.getPointer()); | ||||||
7718 | CombinedInfo.Pointers.push_back(LB.getPointer()); | ||||||
7719 | CombinedInfo.Sizes.push_back( | ||||||
7720 | CGF.Builder.CreateIntCast(Size, CGF.Int64Ty, /*isSigned=*/true)); | ||||||
7721 | |||||||
7722 | // If Mapper is valid, the last component inherits the mapper. | ||||||
7723 | bool HasMapper = Mapper && Next == CE; | ||||||
7724 | CombinedInfo.Mappers.push_back(HasMapper ? Mapper : nullptr); | ||||||
7725 | |||||||
7726 | // We need to add a pointer flag for each map that comes from the | ||||||
7727 | // same expression except for the first one. We also need to signal | ||||||
7728 | // this map is the first one that relates with the current capture | ||||||
7729 | // (there is a set of entries for each capture). | ||||||
7730 | OpenMPOffloadMappingFlags Flags = | ||||||
7731 | getMapTypeBits(MapType, MapModifiers, MotionModifiers, IsImplicit, | ||||||
7732 | !IsExpressionFirstInfo || RequiresReference || | ||||||
7733 | FirstPointerInComplexData, | ||||||
7734 | IsCaptureFirstInfo && !RequiresReference); | ||||||
7735 | |||||||
7736 | if (!IsExpressionFirstInfo) { | ||||||
7737 | // If we have a PTR_AND_OBJ pair where the OBJ is a pointer as well, | ||||||
7738 | // then we reset the TO/FROM/ALWAYS/DELETE/CLOSE flags. | ||||||
7739 | if (IsPointer) | ||||||
7740 | Flags &= ~(OMP_MAP_TO | OMP_MAP_FROM | OMP_MAP_ALWAYS | | ||||||
7741 | OMP_MAP_DELETE | OMP_MAP_CLOSE); | ||||||
7742 | |||||||
7743 | if (ShouldBeMemberOf) { | ||||||
7744 | // Set placeholder value MEMBER_OF=FFFF to indicate that the flag | ||||||
7745 | // should be later updated with the correct value of MEMBER_OF. | ||||||
7746 | Flags |= OMP_MAP_MEMBER_OF; | ||||||
7747 | // From now on, all subsequent PTR_AND_OBJ entries should not be | ||||||
7748 | // marked as MEMBER_OF. | ||||||
7749 | ShouldBeMemberOf = false; | ||||||
7750 | } | ||||||
7751 | } | ||||||
7752 | |||||||
7753 | CombinedInfo.Types.push_back(Flags); | ||||||
7754 | } | ||||||
7755 | |||||||
7756 | // If we have encountered a member expression so far, keep track of the | ||||||
7757 | // mapped member. If the parent is "*this", then the value declaration | ||||||
7758 | // is nullptr. | ||||||
7759 | if (EncounteredME) { | ||||||
7760 | const auto *FD = cast<FieldDecl>(EncounteredME->getMemberDecl()); | ||||||
7761 | unsigned FieldIndex = FD->getFieldIndex(); | ||||||
7762 | |||||||
7763 | // Update info about the lowest and highest elements for this struct | ||||||
7764 | if (!PartialStruct.Base.isValid()) { | ||||||
7765 | PartialStruct.LowestElem = {FieldIndex, LB}; | ||||||
7766 | if (IsFinalArraySection) { | ||||||
7767 | Address HB = | ||||||
7768 | CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false) | ||||||
7769 | .getAddress(CGF); | ||||||
7770 | PartialStruct.HighestElem = {FieldIndex, HB}; | ||||||
7771 | } else { | ||||||
7772 | PartialStruct.HighestElem = {FieldIndex, LB}; | ||||||
7773 | } | ||||||
7774 | PartialStruct.Base = BP; | ||||||
7775 | } else if (FieldIndex < PartialStruct.LowestElem.first) { | ||||||
7776 | PartialStruct.LowestElem = {FieldIndex, LB}; | ||||||
7777 | } else if (FieldIndex > PartialStruct.HighestElem.first) { | ||||||
7778 | PartialStruct.HighestElem = {FieldIndex, LB}; | ||||||
7779 | } | ||||||
7780 | } | ||||||
7781 | |||||||
7782 | // If we have a final array section, we are done with this expression. | ||||||
7783 | if (IsFinalArraySection) | ||||||
7784 | break; | ||||||
7785 | |||||||
7786 | // The pointer becomes the base for the next element. | ||||||
7787 | if (Next != CE) | ||||||
7788 | BP = LB; | ||||||
7789 | |||||||
7790 | IsExpressionFirstInfo = false; | ||||||
7791 | IsCaptureFirstInfo = false; | ||||||
7792 | FirstPointerInComplexData = false; | ||||||
7793 | } | ||||||
7794 | } | ||||||
7795 | } | ||||||
7796 | |||||||
7797 | /// Return the adjusted map modifiers if the declaration a capture refers to | ||||||
7798 | /// appears in a first-private clause. This is expected to be used only with | ||||||
7799 | /// directives that start with 'target'. | ||||||
7800 | MappableExprsHandler::OpenMPOffloadMappingFlags | ||||||
7801 | getMapModifiersForPrivateClauses(const CapturedStmt::Capture &Cap) const { | ||||||
7802 | assert(Cap.capturesVariable() && "Expected capture by reference only!")((Cap.capturesVariable() && "Expected capture by reference only!" ) ? static_cast<void> (0) : __assert_fail ("Cap.capturesVariable() && \"Expected capture by reference only!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7802, __PRETTY_FUNCTION__)); | ||||||
7803 | |||||||
7804 | // A first private variable captured by reference will use only the | ||||||
7805 | // 'private ptr' and 'map to' flag. Return the right flags if the captured | ||||||
7806 | // declaration is known as first-private in this handler. | ||||||
7807 | if (FirstPrivateDecls.count(Cap.getCapturedVar())) { | ||||||
7808 | if (Cap.getCapturedVar()->getType().isConstant(CGF.getContext()) && | ||||||
7809 | Cap.getCaptureKind() == CapturedStmt::VCK_ByRef) | ||||||
7810 | return MappableExprsHandler::OMP_MAP_ALWAYS | | ||||||
7811 | MappableExprsHandler::OMP_MAP_TO; | ||||||
7812 | if (Cap.getCapturedVar()->getType()->isAnyPointerType()) | ||||||
7813 | return MappableExprsHandler::OMP_MAP_TO | | ||||||
7814 | MappableExprsHandler::OMP_MAP_PTR_AND_OBJ; | ||||||
7815 | return MappableExprsHandler::OMP_MAP_PRIVATE | | ||||||
7816 | MappableExprsHandler::OMP_MAP_TO; | ||||||
7817 | } | ||||||
7818 | return MappableExprsHandler::OMP_MAP_TO | | ||||||
7819 | MappableExprsHandler::OMP_MAP_FROM; | ||||||
7820 | } | ||||||
7821 | |||||||
7822 | static OpenMPOffloadMappingFlags getMemberOfFlag(unsigned Position) { | ||||||
7823 | // Rotate by getFlagMemberOffset() bits. | ||||||
7824 | return static_cast<OpenMPOffloadMappingFlags>(((uint64_t)Position + 1) | ||||||
7825 | << getFlagMemberOffset()); | ||||||
7826 | } | ||||||
7827 | |||||||
7828 | static void setCorrectMemberOfFlag(OpenMPOffloadMappingFlags &Flags, | ||||||
7829 | OpenMPOffloadMappingFlags MemberOfFlag) { | ||||||
7830 | // If the entry is PTR_AND_OBJ but has not been marked with the special | ||||||
7831 | // placeholder value 0xFFFF in the MEMBER_OF field, then it should not be | ||||||
7832 | // marked as MEMBER_OF. | ||||||
7833 | if ((Flags & OMP_MAP_PTR_AND_OBJ) && | ||||||
7834 | ((Flags & OMP_MAP_MEMBER_OF) != OMP_MAP_MEMBER_OF)) | ||||||
7835 | return; | ||||||
7836 | |||||||
7837 | // Reset the placeholder value to prepare the flag for the assignment of the | ||||||
7838 | // proper MEMBER_OF value. | ||||||
7839 | Flags &= ~OMP_MAP_MEMBER_OF; | ||||||
7840 | Flags |= MemberOfFlag; | ||||||
7841 | } | ||||||
7842 | |||||||
7843 | void getPlainLayout(const CXXRecordDecl *RD, | ||||||
7844 | llvm::SmallVectorImpl<const FieldDecl *> &Layout, | ||||||
7845 | bool AsBase) const { | ||||||
7846 | const CGRecordLayout &RL = CGF.getTypes().getCGRecordLayout(RD); | ||||||
7847 | |||||||
7848 | llvm::StructType *St = | ||||||
7849 | AsBase ? RL.getBaseSubobjectLLVMType() : RL.getLLVMType(); | ||||||
7850 | |||||||
7851 | unsigned NumElements = St->getNumElements(); | ||||||
7852 | llvm::SmallVector< | ||||||
7853 | llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *>, 4> | ||||||
7854 | RecordLayout(NumElements); | ||||||
7855 | |||||||
7856 | // Fill bases. | ||||||
7857 | for (const auto &I : RD->bases()) { | ||||||
7858 | if (I.isVirtual()) | ||||||
7859 | continue; | ||||||
7860 | const auto *Base = I.getType()->getAsCXXRecordDecl(); | ||||||
7861 | // Ignore empty bases. | ||||||
7862 | if (Base->isEmpty() || CGF.getContext() | ||||||
7863 | .getASTRecordLayout(Base) | ||||||
7864 | .getNonVirtualSize() | ||||||
7865 | .isZero()) | ||||||
7866 | continue; | ||||||
7867 | |||||||
7868 | unsigned FieldIndex = RL.getNonVirtualBaseLLVMFieldNo(Base); | ||||||
7869 | RecordLayout[FieldIndex] = Base; | ||||||
7870 | } | ||||||
7871 | // Fill in virtual bases. | ||||||
7872 | for (const auto &I : RD->vbases()) { | ||||||
7873 | const auto *Base = I.getType()->getAsCXXRecordDecl(); | ||||||
7874 | // Ignore empty bases. | ||||||
7875 | if (Base->isEmpty()) | ||||||
7876 | continue; | ||||||
7877 | unsigned FieldIndex = RL.getVirtualBaseIndex(Base); | ||||||
7878 | if (RecordLayout[FieldIndex]) | ||||||
7879 | continue; | ||||||
7880 | RecordLayout[FieldIndex] = Base; | ||||||
7881 | } | ||||||
7882 | // Fill in all the fields. | ||||||
7883 | assert(!RD->isUnion() && "Unexpected union.")((!RD->isUnion() && "Unexpected union.") ? static_cast <void> (0) : __assert_fail ("!RD->isUnion() && \"Unexpected union.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7883, __PRETTY_FUNCTION__)); | ||||||
7884 | for (const auto *Field : RD->fields()) { | ||||||
7885 | // Fill in non-bitfields. (Bitfields always use a zero pattern, which we | ||||||
7886 | // will fill in later.) | ||||||
7887 | if (!Field->isBitField() && !Field->isZeroSize(CGF.getContext())) { | ||||||
7888 | unsigned FieldIndex = RL.getLLVMFieldNo(Field); | ||||||
7889 | RecordLayout[FieldIndex] = Field; | ||||||
7890 | } | ||||||
7891 | } | ||||||
7892 | for (const llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *> | ||||||
7893 | &Data : RecordLayout) { | ||||||
7894 | if (Data.isNull()) | ||||||
7895 | continue; | ||||||
7896 | if (const auto *Base = Data.dyn_cast<const CXXRecordDecl *>()) | ||||||
7897 | getPlainLayout(Base, Layout, /*AsBase=*/true); | ||||||
7898 | else | ||||||
7899 | Layout.push_back(Data.get<const FieldDecl *>()); | ||||||
7900 | } | ||||||
7901 | } | ||||||
7902 | |||||||
7903 | public: | ||||||
7904 | MappableExprsHandler(const OMPExecutableDirective &Dir, CodeGenFunction &CGF) | ||||||
7905 | : CurDir(&Dir), CGF(CGF) { | ||||||
7906 | // Extract firstprivate clause information. | ||||||
7907 | for (const auto *C : Dir.getClausesOfKind<OMPFirstprivateClause>()) | ||||||
7908 | for (const auto *D : C->varlists()) | ||||||
7909 | FirstPrivateDecls.try_emplace( | ||||||
7910 | cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl()), C->isImplicit()); | ||||||
7911 | // Extract implicit firstprivates from uses_allocators clauses. | ||||||
7912 | for (const auto *C : Dir.getClausesOfKind<OMPUsesAllocatorsClause>()) { | ||||||
7913 | for (unsigned I = 0, E = C->getNumberOfAllocators(); I < E; ++I) { | ||||||
7914 | OMPUsesAllocatorsClause::Data D = C->getAllocatorData(I); | ||||||
7915 | if (const auto *DRE = dyn_cast_or_null<DeclRefExpr>(D.AllocatorTraits)) | ||||||
7916 | FirstPrivateDecls.try_emplace(cast<VarDecl>(DRE->getDecl()), | ||||||
7917 | /*Implicit=*/true); | ||||||
7918 | else if (const auto *VD = dyn_cast<VarDecl>( | ||||||
7919 | cast<DeclRefExpr>(D.Allocator->IgnoreParenImpCasts()) | ||||||
7920 | ->getDecl())) | ||||||
7921 | FirstPrivateDecls.try_emplace(VD, /*Implicit=*/true); | ||||||
7922 | } | ||||||
7923 | } | ||||||
7924 | // Extract device pointer clause information. | ||||||
7925 | for (const auto *C : Dir.getClausesOfKind<OMPIsDevicePtrClause>()) | ||||||
7926 | for (auto L : C->component_lists()) | ||||||
7927 | DevPointersMap[std::get<0>(L)].push_back(std::get<1>(L)); | ||||||
7928 | } | ||||||
7929 | |||||||
7930 | /// Constructor for the declare mapper directive. | ||||||
7931 | MappableExprsHandler(const OMPDeclareMapperDecl &Dir, CodeGenFunction &CGF) | ||||||
7932 | : CurDir(&Dir), CGF(CGF) {} | ||||||
7933 | |||||||
7934 | /// Generate code for the combined entry if we have a partially mapped struct | ||||||
7935 | /// and take care of the mapping flags of the arguments corresponding to | ||||||
7936 | /// individual struct members. | ||||||
7937 | void emitCombinedEntry(MapCombinedInfoTy &CombinedInfo, | ||||||
7938 | MapFlagsArrayTy &CurTypes, | ||||||
7939 | const StructRangeInfoTy &PartialStruct, | ||||||
7940 | bool NotTargetParams = false) const { | ||||||
7941 | // Base is the base of the struct | ||||||
7942 | CombinedInfo.BasePointers.push_back(PartialStruct.Base.getPointer()); | ||||||
7943 | // Pointer is the address of the lowest element | ||||||
7944 | llvm::Value *LB = PartialStruct.LowestElem.second.getPointer(); | ||||||
7945 | CombinedInfo.Pointers.push_back(LB); | ||||||
7946 | // There should not be a mapper for a combined entry. | ||||||
7947 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
7948 | // Size is (addr of {highest+1} element) - (addr of lowest element) | ||||||
7949 | llvm::Value *HB = PartialStruct.HighestElem.second.getPointer(); | ||||||
7950 | llvm::Value *HAddr = CGF.Builder.CreateConstGEP1_32(HB, /*Idx0=*/1); | ||||||
7951 | llvm::Value *CLAddr = CGF.Builder.CreatePointerCast(LB, CGF.VoidPtrTy); | ||||||
7952 | llvm::Value *CHAddr = CGF.Builder.CreatePointerCast(HAddr, CGF.VoidPtrTy); | ||||||
7953 | llvm::Value *Diff = CGF.Builder.CreatePtrDiff(CHAddr, CLAddr); | ||||||
7954 | llvm::Value *Size = CGF.Builder.CreateIntCast(Diff, CGF.Int64Ty, | ||||||
7955 | /*isSigned=*/false); | ||||||
7956 | CombinedInfo.Sizes.push_back(Size); | ||||||
7957 | // Map type is always TARGET_PARAM, if generate info for captures. | ||||||
7958 | CombinedInfo.Types.push_back(NotTargetParams ? OMP_MAP_NONE | ||||||
7959 | : OMP_MAP_TARGET_PARAM); | ||||||
7960 | // If any element has the present modifier, then make sure the runtime | ||||||
7961 | // doesn't attempt to allocate the struct. | ||||||
7962 | if (CurTypes.end() != | ||||||
7963 | llvm::find_if(CurTypes, [](OpenMPOffloadMappingFlags Type) { | ||||||
7964 | return Type & OMP_MAP_PRESENT; | ||||||
7965 | })) | ||||||
7966 | CombinedInfo.Types.back() |= OMP_MAP_PRESENT; | ||||||
7967 | // Remove TARGET_PARAM flag from the first element | ||||||
7968 | (*CurTypes.begin()) &= ~OMP_MAP_TARGET_PARAM; | ||||||
7969 | |||||||
7970 | // All other current entries will be MEMBER_OF the combined entry | ||||||
7971 | // (except for PTR_AND_OBJ entries which do not have a placeholder value | ||||||
7972 | // 0xFFFF in the MEMBER_OF field). | ||||||
7973 | OpenMPOffloadMappingFlags MemberOfFlag = | ||||||
7974 | getMemberOfFlag(CombinedInfo.BasePointers.size() - 1); | ||||||
7975 | for (auto &M : CurTypes) | ||||||
7976 | setCorrectMemberOfFlag(M, MemberOfFlag); | ||||||
7977 | } | ||||||
7978 | |||||||
7979 | /// Generate all the base pointers, section pointers, sizes, map types, and | ||||||
7980 | /// mappers for the extracted mappable expressions (all included in \a | ||||||
7981 | /// CombinedInfo). Also, for each item that relates with a device pointer, a | ||||||
7982 | /// pair of the relevant declaration and index where it occurs is appended to | ||||||
7983 | /// the device pointers info array. | ||||||
7984 | void generateAllInfo( | ||||||
7985 | MapCombinedInfoTy &CombinedInfo, bool NotTargetParams = false, | ||||||
7986 | const llvm::DenseSet<CanonicalDeclPtr<const Decl>> &SkipVarSet = | ||||||
7987 | llvm::DenseSet<CanonicalDeclPtr<const Decl>>()) const { | ||||||
7988 | // We have to process the component lists that relate with the same | ||||||
7989 | // declaration in a single chunk so that we can generate the map flags | ||||||
7990 | // correctly. Therefore, we organize all lists in a map. | ||||||
7991 | llvm::MapVector<const ValueDecl *, SmallVector<MapInfo, 8>> Info; | ||||||
7992 | |||||||
7993 | // Helper function to fill the information map for the different supported | ||||||
7994 | // clauses. | ||||||
7995 | auto &&InfoGen = | ||||||
7996 | [&Info, &SkipVarSet]( | ||||||
7997 | const ValueDecl *D, | ||||||
7998 | OMPClauseMappableExprCommon::MappableExprComponentListRef L, | ||||||
7999 | OpenMPMapClauseKind MapType, | ||||||
8000 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | ||||||
8001 | ArrayRef<OpenMPMotionModifierKind> MotionModifiers, | ||||||
8002 | bool ReturnDevicePointer, bool IsImplicit, const ValueDecl *Mapper, | ||||||
8003 | bool ForDeviceAddr = false) { | ||||||
8004 | const ValueDecl *VD = | ||||||
8005 | D ? cast<ValueDecl>(D->getCanonicalDecl()) : nullptr; | ||||||
8006 | if (SkipVarSet.count(VD)) | ||||||
8007 | return; | ||||||
8008 | Info[VD].emplace_back(L, MapType, MapModifiers, MotionModifiers, | ||||||
8009 | ReturnDevicePointer, IsImplicit, Mapper, | ||||||
8010 | ForDeviceAddr); | ||||||
8011 | }; | ||||||
8012 | |||||||
8013 | assert(CurDir.is<const OMPExecutableDirective *>() &&((CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? static_cast<void> ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8014, __PRETTY_FUNCTION__)) | ||||||
8014 | "Expect a executable directive")((CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? static_cast<void> ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8014, __PRETTY_FUNCTION__)); | ||||||
8015 | const auto *CurExecDir = CurDir.get<const OMPExecutableDirective *>(); | ||||||
8016 | for (const auto *C : CurExecDir->getClausesOfKind<OMPMapClause>()) | ||||||
8017 | for (const auto L : C->component_lists()) { | ||||||
8018 | InfoGen(std::get<0>(L), std::get<1>(L), C->getMapType(), | ||||||
8019 | C->getMapTypeModifiers(), llvm::None, | ||||||
8020 | /*ReturnDevicePointer=*/false, C->isImplicit(), std::get<2>(L)); | ||||||
8021 | } | ||||||
8022 | for (const auto *C : CurExecDir->getClausesOfKind<OMPToClause>()) | ||||||
8023 | for (const auto L : C->component_lists()) { | ||||||
8024 | InfoGen(std::get<0>(L), std::get<1>(L), OMPC_MAP_to, llvm::None, | ||||||
8025 | C->getMotionModifiers(), /*ReturnDevicePointer=*/false, | ||||||
8026 | C->isImplicit(), std::get<2>(L)); | ||||||
8027 | } | ||||||
8028 | for (const auto *C : CurExecDir->getClausesOfKind<OMPFromClause>()) | ||||||
8029 | for (const auto L : C->component_lists()) { | ||||||
8030 | InfoGen(std::get<0>(L), std::get<1>(L), OMPC_MAP_from, llvm::None, | ||||||
8031 | C->getMotionModifiers(), /*ReturnDevicePointer=*/false, | ||||||
8032 | C->isImplicit(), std::get<2>(L)); | ||||||
8033 | } | ||||||
8034 | |||||||
8035 | // Look at the use_device_ptr clause information and mark the existing map | ||||||
8036 | // entries as such. If there is no map information for an entry in the | ||||||
8037 | // use_device_ptr list, we create one with map type 'alloc' and zero size | ||||||
8038 | // section. It is the user fault if that was not mapped before. If there is | ||||||
8039 | // no map information and the pointer is a struct member, then we defer the | ||||||
8040 | // emission of that entry until the whole struct has been processed. | ||||||
8041 | llvm::MapVector<const ValueDecl *, SmallVector<DeferredDevicePtrEntryTy, 4>> | ||||||
8042 | DeferredInfo; | ||||||
8043 | MapCombinedInfoTy UseDevicePtrCombinedInfo; | ||||||
8044 | |||||||
8045 | for (const auto *C : | ||||||
8046 | CurExecDir->getClausesOfKind<OMPUseDevicePtrClause>()) { | ||||||
8047 | for (const auto L : C->component_lists()) { | ||||||
8048 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components = | ||||||
8049 | std::get<1>(L); | ||||||
8050 | assert(!Components.empty() &&((!Components.empty() && "Not expecting empty list of components!" ) ? static_cast<void> (0) : __assert_fail ("!Components.empty() && \"Not expecting empty list of components!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8051, __PRETTY_FUNCTION__)) | ||||||
8051 | "Not expecting empty list of components!")((!Components.empty() && "Not expecting empty list of components!" ) ? static_cast<void> (0) : __assert_fail ("!Components.empty() && \"Not expecting empty list of components!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8051, __PRETTY_FUNCTION__)); | ||||||
8052 | const ValueDecl *VD = Components.back().getAssociatedDeclaration(); | ||||||
8053 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); | ||||||
8054 | const Expr *IE = Components.back().getAssociatedExpression(); | ||||||
8055 | // If the first component is a member expression, we have to look into | ||||||
8056 | // 'this', which maps to null in the map of map information. Otherwise | ||||||
8057 | // look directly for the information. | ||||||
8058 | auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD); | ||||||
8059 | |||||||
8060 | // We potentially have map information for this declaration already. | ||||||
8061 | // Look for the first set of components that refer to it. | ||||||
8062 | if (It != Info.end()) { | ||||||
8063 | auto *CI = llvm::find_if(It->second, [VD](const MapInfo &MI) { | ||||||
8064 | return MI.Components.back().getAssociatedDeclaration() == VD; | ||||||
8065 | }); | ||||||
8066 | // If we found a map entry, signal that the pointer has to be returned | ||||||
8067 | // and move on to the next declaration. | ||||||
8068 | // Exclude cases where the base pointer is mapped as array subscript, | ||||||
8069 | // array section or array shaping. The base address is passed as a | ||||||
8070 | // pointer to base in this case and cannot be used as a base for | ||||||
8071 | // use_device_ptr list item. | ||||||
8072 | if (CI != It->second.end()) { | ||||||
8073 | auto PrevCI = std::next(CI->Components.rbegin()); | ||||||
8074 | const auto *VarD = dyn_cast<VarDecl>(VD); | ||||||
8075 | if (CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory() || | ||||||
8076 | isa<MemberExpr>(IE) || | ||||||
8077 | !VD->getType().getNonReferenceType()->isPointerType() || | ||||||
8078 | PrevCI == CI->Components.rend() || | ||||||
8079 | isa<MemberExpr>(PrevCI->getAssociatedExpression()) || !VarD || | ||||||
8080 | VarD->hasLocalStorage()) { | ||||||
8081 | CI->ReturnDevicePointer = true; | ||||||
8082 | continue; | ||||||
8083 | } | ||||||
8084 | } | ||||||
8085 | } | ||||||
8086 | |||||||
8087 | // We didn't find any match in our map information - generate a zero | ||||||
8088 | // size array section - if the pointer is a struct member we defer this | ||||||
8089 | // action until the whole struct has been processed. | ||||||
8090 | if (isa<MemberExpr>(IE)) { | ||||||
8091 | // Insert the pointer into Info to be processed by | ||||||
8092 | // generateInfoForComponentList. Because it is a member pointer | ||||||
8093 | // without a pointee, no entry will be generated for it, therefore | ||||||
8094 | // we need to generate one after the whole struct has been processed. | ||||||
8095 | // Nonetheless, generateInfoForComponentList must be called to take | ||||||
8096 | // the pointer into account for the calculation of the range of the | ||||||
8097 | // partial struct. | ||||||
8098 | InfoGen(nullptr, Components, OMPC_MAP_unknown, llvm::None, llvm::None, | ||||||
8099 | /*ReturnDevicePointer=*/false, C->isImplicit(), nullptr); | ||||||
8100 | DeferredInfo[nullptr].emplace_back(IE, VD, /*ForDeviceAddr=*/false); | ||||||
8101 | } else { | ||||||
8102 | llvm::Value *Ptr = | ||||||
8103 | CGF.EmitLoadOfScalar(CGF.EmitLValue(IE), IE->getExprLoc()); | ||||||
8104 | UseDevicePtrCombinedInfo.BasePointers.emplace_back(Ptr, VD); | ||||||
8105 | UseDevicePtrCombinedInfo.Pointers.push_back(Ptr); | ||||||
8106 | UseDevicePtrCombinedInfo.Sizes.push_back( | ||||||
8107 | llvm::Constant::getNullValue(CGF.Int64Ty)); | ||||||
8108 | UseDevicePtrCombinedInfo.Types.push_back( | ||||||
8109 | OMP_MAP_RETURN_PARAM | | ||||||
8110 | (NotTargetParams ? OMP_MAP_NONE : OMP_MAP_TARGET_PARAM)); | ||||||
8111 | UseDevicePtrCombinedInfo.Mappers.push_back(nullptr); | ||||||
8112 | } | ||||||
8113 | } | ||||||
8114 | } | ||||||
8115 | |||||||
8116 | // Look at the use_device_addr clause information and mark the existing map | ||||||
8117 | // entries as such. If there is no map information for an entry in the | ||||||
8118 | // use_device_addr list, we create one with map type 'alloc' and zero size | ||||||
8119 | // section. It is the user fault if that was not mapped before. If there is | ||||||
8120 | // no map information and the pointer is a struct member, then we defer the | ||||||
8121 | // emission of that entry until the whole struct has been processed. | ||||||
8122 | llvm::SmallDenseSet<CanonicalDeclPtr<const Decl>, 4> Processed; | ||||||
8123 | for (const auto *C : | ||||||
8124 | CurExecDir->getClausesOfKind<OMPUseDeviceAddrClause>()) { | ||||||
8125 | for (const auto L : C->component_lists()) { | ||||||
8126 | assert(!std::get<1>(L).empty() &&((!std::get<1>(L).empty() && "Not expecting empty list of components!" ) ? static_cast<void> (0) : __assert_fail ("!std::get<1>(L).empty() && \"Not expecting empty list of components!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8127, __PRETTY_FUNCTION__)) | ||||||
8127 | "Not expecting empty list of components!")((!std::get<1>(L).empty() && "Not expecting empty list of components!" ) ? static_cast<void> (0) : __assert_fail ("!std::get<1>(L).empty() && \"Not expecting empty list of components!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8127, __PRETTY_FUNCTION__)); | ||||||
8128 | const ValueDecl *VD = std::get<1>(L).back().getAssociatedDeclaration(); | ||||||
8129 | if (!Processed.insert(VD).second) | ||||||
8130 | continue; | ||||||
8131 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); | ||||||
8132 | const Expr *IE = std::get<1>(L).back().getAssociatedExpression(); | ||||||
8133 | // If the first component is a member expression, we have to look into | ||||||
8134 | // 'this', which maps to null in the map of map information. Otherwise | ||||||
8135 | // look directly for the information. | ||||||
8136 | auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD); | ||||||
8137 | |||||||
8138 | // We potentially have map information for this declaration already. | ||||||
8139 | // Look for the first set of components that refer to it. | ||||||
8140 | if (It != Info.end()) { | ||||||
8141 | auto *CI = llvm::find_if(It->second, [VD](const MapInfo &MI) { | ||||||
8142 | return MI.Components.back().getAssociatedDeclaration() == VD; | ||||||
8143 | }); | ||||||
8144 | // If we found a map entry, signal that the pointer has to be returned | ||||||
8145 | // and move on to the next declaration. | ||||||
8146 | if (CI != It->second.end()) { | ||||||
8147 | CI->ReturnDevicePointer = true; | ||||||
8148 | continue; | ||||||
8149 | } | ||||||
8150 | } | ||||||
8151 | |||||||
8152 | // We didn't find any match in our map information - generate a zero | ||||||
8153 | // size array section - if the pointer is a struct member we defer this | ||||||
8154 | // action until the whole struct has been processed. | ||||||
8155 | if (isa<MemberExpr>(IE)) { | ||||||
8156 | // Insert the pointer into Info to be processed by | ||||||
8157 | // generateInfoForComponentList. Because it is a member pointer | ||||||
8158 | // without a pointee, no entry will be generated for it, therefore | ||||||
8159 | // we need to generate one after the whole struct has been processed. | ||||||
8160 | // Nonetheless, generateInfoForComponentList must be called to take | ||||||
8161 | // the pointer into account for the calculation of the range of the | ||||||
8162 | // partial struct. | ||||||
8163 | InfoGen(nullptr, std::get<1>(L), OMPC_MAP_unknown, llvm::None, | ||||||
8164 | llvm::None, /*ReturnDevicePointer=*/false, C->isImplicit(), | ||||||
8165 | nullptr, /*ForDeviceAddr=*/true); | ||||||
8166 | DeferredInfo[nullptr].emplace_back(IE, VD, /*ForDeviceAddr=*/true); | ||||||
8167 | } else { | ||||||
8168 | llvm::Value *Ptr; | ||||||
8169 | if (IE->isGLValue()) | ||||||
8170 | Ptr = CGF.EmitLValue(IE).getPointer(CGF); | ||||||
8171 | else | ||||||
8172 | Ptr = CGF.EmitScalarExpr(IE); | ||||||
8173 | CombinedInfo.BasePointers.emplace_back(Ptr, VD); | ||||||
8174 | CombinedInfo.Pointers.push_back(Ptr); | ||||||
8175 | CombinedInfo.Sizes.push_back( | ||||||
8176 | llvm::Constant::getNullValue(CGF.Int64Ty)); | ||||||
8177 | CombinedInfo.Types.push_back( | ||||||
8178 | OMP_MAP_RETURN_PARAM | | ||||||
8179 | (NotTargetParams ? OMP_MAP_NONE : OMP_MAP_TARGET_PARAM)); | ||||||
8180 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
8181 | } | ||||||
8182 | } | ||||||
8183 | } | ||||||
8184 | |||||||
8185 | for (const auto &M : Info) { | ||||||
8186 | // We need to know when we generate information for the first component | ||||||
8187 | // associated with a capture, because the mapping flags depend on it. | ||||||
8188 | bool IsFirstComponentList = !NotTargetParams; | ||||||
8189 | |||||||
8190 | // Temporary generated information. | ||||||
8191 | MapCombinedInfoTy CurInfo; | ||||||
8192 | StructRangeInfoTy PartialStruct; | ||||||
8193 | |||||||
8194 | for (const MapInfo &L : M.second) { | ||||||
8195 | assert(!L.Components.empty() &&((!L.Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8196, __PRETTY_FUNCTION__)) | ||||||
8196 | "Not expecting declaration with no component lists.")((!L.Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8196, __PRETTY_FUNCTION__)); | ||||||
8197 | |||||||
8198 | // Remember the current base pointer index. | ||||||
8199 | unsigned CurrentBasePointersIdx = CurInfo.BasePointers.size(); | ||||||
8200 | generateInfoForComponentList(L.MapType, L.MapModifiers, | ||||||
8201 | L.MotionModifiers, L.Components, CurInfo, | ||||||
8202 | PartialStruct, IsFirstComponentList, | ||||||
8203 | L.IsImplicit, L.Mapper, L.ForDeviceAddr); | ||||||
8204 | |||||||
8205 | // If this entry relates with a device pointer, set the relevant | ||||||
8206 | // declaration and add the 'return pointer' flag. | ||||||
8207 | if (L.ReturnDevicePointer) { | ||||||
8208 | assert(CurInfo.BasePointers.size() > CurrentBasePointersIdx &&((CurInfo.BasePointers.size() > CurrentBasePointersIdx && "Unexpected number of mapped base pointers.") ? static_cast< void> (0) : __assert_fail ("CurInfo.BasePointers.size() > CurrentBasePointersIdx && \"Unexpected number of mapped base pointers.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8209, __PRETTY_FUNCTION__)) | ||||||
8209 | "Unexpected number of mapped base pointers.")((CurInfo.BasePointers.size() > CurrentBasePointersIdx && "Unexpected number of mapped base pointers.") ? static_cast< void> (0) : __assert_fail ("CurInfo.BasePointers.size() > CurrentBasePointersIdx && \"Unexpected number of mapped base pointers.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8209, __PRETTY_FUNCTION__)); | ||||||
8210 | |||||||
8211 | const ValueDecl *RelevantVD = | ||||||
8212 | L.Components.back().getAssociatedDeclaration(); | ||||||
8213 | assert(RelevantVD &&((RelevantVD && "No relevant declaration related with device pointer??" ) ? static_cast<void> (0) : __assert_fail ("RelevantVD && \"No relevant declaration related with device pointer??\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8214, __PRETTY_FUNCTION__)) | ||||||
8214 | "No relevant declaration related with device pointer??")((RelevantVD && "No relevant declaration related with device pointer??" ) ? static_cast<void> (0) : __assert_fail ("RelevantVD && \"No relevant declaration related with device pointer??\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8214, __PRETTY_FUNCTION__)); | ||||||
8215 | |||||||
8216 | CurInfo.BasePointers[CurrentBasePointersIdx].setDevicePtrDecl( | ||||||
8217 | RelevantVD); | ||||||
8218 | CurInfo.Types[CurrentBasePointersIdx] |= OMP_MAP_RETURN_PARAM; | ||||||
8219 | } | ||||||
8220 | IsFirstComponentList = false; | ||||||
8221 | } | ||||||
8222 | |||||||
8223 | // Append any pending zero-length pointers which are struct members and | ||||||
8224 | // used with use_device_ptr or use_device_addr. | ||||||
8225 | auto CI = DeferredInfo.find(M.first); | ||||||
8226 | if (CI != DeferredInfo.end()) { | ||||||
8227 | for (const DeferredDevicePtrEntryTy &L : CI->second) { | ||||||
8228 | llvm::Value *BasePtr; | ||||||
8229 | llvm::Value *Ptr; | ||||||
8230 | if (L.ForDeviceAddr) { | ||||||
8231 | if (L.IE->isGLValue()) | ||||||
8232 | Ptr = this->CGF.EmitLValue(L.IE).getPointer(CGF); | ||||||
8233 | else | ||||||
8234 | Ptr = this->CGF.EmitScalarExpr(L.IE); | ||||||
8235 | BasePtr = Ptr; | ||||||
8236 | // Entry is RETURN_PARAM. Also, set the placeholder value | ||||||
8237 | // MEMBER_OF=FFFF so that the entry is later updated with the | ||||||
8238 | // correct value of MEMBER_OF. | ||||||
8239 | CurInfo.Types.push_back(OMP_MAP_RETURN_PARAM | OMP_MAP_MEMBER_OF); | ||||||
8240 | } else { | ||||||
8241 | BasePtr = this->CGF.EmitLValue(L.IE).getPointer(CGF); | ||||||
8242 | Ptr = this->CGF.EmitLoadOfScalar(this->CGF.EmitLValue(L.IE), | ||||||
8243 | L.IE->getExprLoc()); | ||||||
8244 | // Entry is PTR_AND_OBJ and RETURN_PARAM. Also, set the placeholder | ||||||
8245 | // value MEMBER_OF=FFFF so that the entry is later updated with the | ||||||
8246 | // correct value of MEMBER_OF. | ||||||
8247 | CurInfo.Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_RETURN_PARAM | | ||||||
8248 | OMP_MAP_MEMBER_OF); | ||||||
8249 | } | ||||||
8250 | CurInfo.BasePointers.emplace_back(BasePtr, L.VD); | ||||||
8251 | CurInfo.Pointers.push_back(Ptr); | ||||||
8252 | CurInfo.Sizes.push_back( | ||||||
8253 | llvm::Constant::getNullValue(this->CGF.Int64Ty)); | ||||||
8254 | CurInfo.Mappers.push_back(nullptr); | ||||||
8255 | } | ||||||
8256 | } | ||||||
8257 | |||||||
8258 | // If there is an entry in PartialStruct it means we have a struct with | ||||||
8259 | // individual members mapped. Emit an extra combined entry. | ||||||
8260 | if (PartialStruct.Base.isValid()) | ||||||
8261 | emitCombinedEntry(CombinedInfo, CurInfo.Types, PartialStruct, | ||||||
8262 | NotTargetParams); | ||||||
8263 | |||||||
8264 | // We need to append the results of this capture to what we already have. | ||||||
8265 | CombinedInfo.append(CurInfo); | ||||||
8266 | } | ||||||
8267 | // Append data for use_device_ptr clauses. | ||||||
8268 | CombinedInfo.append(UseDevicePtrCombinedInfo); | ||||||
8269 | } | ||||||
8270 | |||||||
8271 | /// Generate all the base pointers, section pointers, sizes, map types, and | ||||||
8272 | /// mappers for the extracted map clauses of user-defined mapper (all included | ||||||
8273 | /// in \a CombinedInfo). | ||||||
8274 | void generateAllInfoForMapper(MapCombinedInfoTy &CombinedInfo) const { | ||||||
8275 | assert(CurDir.is<const OMPDeclareMapperDecl *>() &&((CurDir.is<const OMPDeclareMapperDecl *>() && "Expect a declare mapper directive" ) ? static_cast<void> (0) : __assert_fail ("CurDir.is<const OMPDeclareMapperDecl *>() && \"Expect a declare mapper directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8276, __PRETTY_FUNCTION__)) | ||||||
8276 | "Expect a declare mapper directive")((CurDir.is<const OMPDeclareMapperDecl *>() && "Expect a declare mapper directive" ) ? static_cast<void> (0) : __assert_fail ("CurDir.is<const OMPDeclareMapperDecl *>() && \"Expect a declare mapper directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8276, __PRETTY_FUNCTION__)); | ||||||
8277 | const auto *CurMapperDir = CurDir.get<const OMPDeclareMapperDecl *>(); | ||||||
8278 | // We have to process the component lists that relate with the same | ||||||
8279 | // declaration in a single chunk so that we can generate the map flags | ||||||
8280 | // correctly. Therefore, we organize all lists in a map. | ||||||
8281 | llvm::MapVector<const ValueDecl *, SmallVector<MapInfo, 8>> Info; | ||||||
8282 | |||||||
8283 | // Fill the information map for map clauses. | ||||||
8284 | for (const auto *C : CurMapperDir->clauselists()) { | ||||||
8285 | const auto *MC = cast<OMPMapClause>(C); | ||||||
8286 | for (const auto L : MC->component_lists()) { | ||||||
8287 | const ValueDecl *VD = | ||||||
8288 | std::get<0>(L) ? cast<ValueDecl>(std::get<0>(L)->getCanonicalDecl()) | ||||||
8289 | : nullptr; | ||||||
8290 | // Get the corresponding user-defined mapper. | ||||||
8291 | Info[VD].emplace_back(std::get<1>(L), MC->getMapType(), | ||||||
8292 | MC->getMapTypeModifiers(), llvm::None, | ||||||
8293 | /*ReturnDevicePointer=*/false, MC->isImplicit(), | ||||||
8294 | std::get<2>(L)); | ||||||
8295 | } | ||||||
8296 | } | ||||||
8297 | |||||||
8298 | for (const auto &M : Info) { | ||||||
8299 | // We need to know when we generate information for the first component | ||||||
8300 | // associated with a capture, because the mapping flags depend on it. | ||||||
8301 | bool IsFirstComponentList = true; | ||||||
8302 | |||||||
8303 | // Temporary generated information. | ||||||
8304 | MapCombinedInfoTy CurInfo; | ||||||
8305 | StructRangeInfoTy PartialStruct; | ||||||
8306 | |||||||
8307 | for (const MapInfo &L : M.second) { | ||||||
8308 | assert(!L.Components.empty() &&((!L.Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8309, __PRETTY_FUNCTION__)) | ||||||
8309 | "Not expecting declaration with no component lists.")((!L.Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8309, __PRETTY_FUNCTION__)); | ||||||
8310 | generateInfoForComponentList(L.MapType, L.MapModifiers, | ||||||
8311 | L.MotionModifiers, L.Components, CurInfo, | ||||||
8312 | PartialStruct, IsFirstComponentList, | ||||||
8313 | L.IsImplicit, L.Mapper, L.ForDeviceAddr); | ||||||
8314 | IsFirstComponentList = false; | ||||||
8315 | } | ||||||
8316 | |||||||
8317 | // If there is an entry in PartialStruct it means we have a struct with | ||||||
8318 | // individual members mapped. Emit an extra combined entry. | ||||||
8319 | if (PartialStruct.Base.isValid()) | ||||||
8320 | emitCombinedEntry(CombinedInfo, CurInfo.Types, PartialStruct); | ||||||
8321 | |||||||
8322 | // We need to append the results of this capture to what we already have. | ||||||
8323 | CombinedInfo.append(CurInfo); | ||||||
8324 | } | ||||||
8325 | } | ||||||
8326 | |||||||
8327 | /// Emit capture info for lambdas for variables captured by reference. | ||||||
8328 | void generateInfoForLambdaCaptures( | ||||||
8329 | const ValueDecl *VD, llvm::Value *Arg, MapCombinedInfoTy &CombinedInfo, | ||||||
8330 | llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers) const { | ||||||
8331 | const auto *RD = VD->getType() | ||||||
8332 | .getCanonicalType() | ||||||
8333 | .getNonReferenceType() | ||||||
8334 | ->getAsCXXRecordDecl(); | ||||||
8335 | if (!RD || !RD->isLambda()) | ||||||
8336 | return; | ||||||
8337 | Address VDAddr = Address(Arg, CGF.getContext().getDeclAlign(VD)); | ||||||
8338 | LValue VDLVal = CGF.MakeAddrLValue( | ||||||
8339 | VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); | ||||||
8340 | llvm::DenseMap<const VarDecl *, FieldDecl *> Captures; | ||||||
8341 | FieldDecl *ThisCapture = nullptr; | ||||||
8342 | RD->getCaptureFields(Captures, ThisCapture); | ||||||
8343 | if (ThisCapture) { | ||||||
8344 | LValue ThisLVal = | ||||||
8345 | CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); | ||||||
8346 | LValue ThisLValVal = CGF.EmitLValueForField(VDLVal, ThisCapture); | ||||||
8347 | LambdaPointers.try_emplace(ThisLVal.getPointer(CGF), | ||||||
8348 | VDLVal.getPointer(CGF)); | ||||||
8349 | CombinedInfo.BasePointers.push_back(ThisLVal.getPointer(CGF)); | ||||||
8350 | CombinedInfo.Pointers.push_back(ThisLValVal.getPointer(CGF)); | ||||||
8351 | CombinedInfo.Sizes.push_back( | ||||||
8352 | CGF.Builder.CreateIntCast(CGF.getTypeSize(CGF.getContext().VoidPtrTy), | ||||||
8353 | CGF.Int64Ty, /*isSigned=*/true)); | ||||||
8354 | CombinedInfo.Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | ||||||
8355 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT); | ||||||
8356 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
8357 | } | ||||||
8358 | for (const LambdaCapture &LC : RD->captures()) { | ||||||
8359 | if (!LC.capturesVariable()) | ||||||
8360 | continue; | ||||||
8361 | const VarDecl *VD = LC.getCapturedVar(); | ||||||
8362 | if (LC.getCaptureKind() != LCK_ByRef && !VD->getType()->isPointerType()) | ||||||
8363 | continue; | ||||||
8364 | auto It = Captures.find(VD); | ||||||
8365 | assert(It != Captures.end() && "Found lambda capture without field.")((It != Captures.end() && "Found lambda capture without field." ) ? static_cast<void> (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8365, __PRETTY_FUNCTION__)); | ||||||
8366 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); | ||||||
8367 | if (LC.getCaptureKind() == LCK_ByRef) { | ||||||
8368 | LValue VarLValVal = CGF.EmitLValueForField(VDLVal, It->second); | ||||||
8369 | LambdaPointers.try_emplace(VarLVal.getPointer(CGF), | ||||||
8370 | VDLVal.getPointer(CGF)); | ||||||
8371 | CombinedInfo.BasePointers.push_back(VarLVal.getPointer(CGF)); | ||||||
8372 | CombinedInfo.Pointers.push_back(VarLValVal.getPointer(CGF)); | ||||||
8373 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | ||||||
8374 | CGF.getTypeSize( | ||||||
8375 | VD->getType().getCanonicalType().getNonReferenceType()), | ||||||
8376 | CGF.Int64Ty, /*isSigned=*/true)); | ||||||
8377 | } else { | ||||||
8378 | RValue VarRVal = CGF.EmitLoadOfLValue(VarLVal, RD->getLocation()); | ||||||
8379 | LambdaPointers.try_emplace(VarLVal.getPointer(CGF), | ||||||
8380 | VDLVal.getPointer(CGF)); | ||||||
8381 | CombinedInfo.BasePointers.push_back(VarLVal.getPointer(CGF)); | ||||||
8382 | CombinedInfo.Pointers.push_back(VarRVal.getScalarVal()); | ||||||
8383 | CombinedInfo.Sizes.push_back(llvm::ConstantInt::get(CGF.Int64Ty, 0)); | ||||||
8384 | } | ||||||
8385 | CombinedInfo.Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | ||||||
8386 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT); | ||||||
8387 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
8388 | } | ||||||
8389 | } | ||||||
8390 | |||||||
8391 | /// Set correct indices for lambdas captures. | ||||||
8392 | void adjustMemberOfForLambdaCaptures( | ||||||
8393 | const llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers, | ||||||
8394 | MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers, | ||||||
8395 | MapFlagsArrayTy &Types) const { | ||||||
8396 | for (unsigned I = 0, E = Types.size(); I < E; ++I) { | ||||||
8397 | // Set correct member_of idx for all implicit lambda captures. | ||||||
8398 | if (Types[I] != (OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | ||||||
8399 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT)) | ||||||
8400 | continue; | ||||||
8401 | llvm::Value *BasePtr = LambdaPointers.lookup(*BasePointers[I]); | ||||||
8402 | assert(BasePtr && "Unable to find base lambda address.")((BasePtr && "Unable to find base lambda address.") ? static_cast<void> (0) : __assert_fail ("BasePtr && \"Unable to find base lambda address.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8402, __PRETTY_FUNCTION__)); | ||||||
8403 | int TgtIdx = -1; | ||||||
8404 | for (unsigned J = I; J > 0; --J) { | ||||||
8405 | unsigned Idx = J - 1; | ||||||
8406 | if (Pointers[Idx] != BasePtr) | ||||||
8407 | continue; | ||||||
8408 | TgtIdx = Idx; | ||||||
8409 | break; | ||||||
8410 | } | ||||||
8411 | assert(TgtIdx != -1 && "Unable to find parent lambda.")((TgtIdx != -1 && "Unable to find parent lambda.") ? static_cast <void> (0) : __assert_fail ("TgtIdx != -1 && \"Unable to find parent lambda.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8411, __PRETTY_FUNCTION__)); | ||||||
8412 | // All other current entries will be MEMBER_OF the combined entry | ||||||
8413 | // (except for PTR_AND_OBJ entries which do not have a placeholder value | ||||||
8414 | // 0xFFFF in the MEMBER_OF field). | ||||||
8415 | OpenMPOffloadMappingFlags MemberOfFlag = getMemberOfFlag(TgtIdx); | ||||||
8416 | setCorrectMemberOfFlag(Types[I], MemberOfFlag); | ||||||
8417 | } | ||||||
8418 | } | ||||||
8419 | |||||||
8420 | /// Generate the base pointers, section pointers, sizes, map types, and | ||||||
8421 | /// mappers associated to a given capture (all included in \a CombinedInfo). | ||||||
8422 | void generateInfoForCapture(const CapturedStmt::Capture *Cap, | ||||||
8423 | llvm::Value *Arg, MapCombinedInfoTy &CombinedInfo, | ||||||
8424 | StructRangeInfoTy &PartialStruct) const { | ||||||
8425 | assert(!Cap->capturesVariableArrayType() &&((!Cap->capturesVariableArrayType() && "Not expecting to generate map info for a variable array type!" ) ? static_cast<void> (0) : __assert_fail ("!Cap->capturesVariableArrayType() && \"Not expecting to generate map info for a variable array type!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8426, __PRETTY_FUNCTION__)) | ||||||
8426 | "Not expecting to generate map info for a variable array type!")((!Cap->capturesVariableArrayType() && "Not expecting to generate map info for a variable array type!" ) ? static_cast<void> (0) : __assert_fail ("!Cap->capturesVariableArrayType() && \"Not expecting to generate map info for a variable array type!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8426, __PRETTY_FUNCTION__)); | ||||||
8427 | |||||||
8428 | // We need to know when we generating information for the first component | ||||||
8429 | const ValueDecl *VD = Cap->capturesThis() | ||||||
8430 | ? nullptr | ||||||
8431 | : Cap->getCapturedVar()->getCanonicalDecl(); | ||||||
8432 | |||||||
8433 | // If this declaration appears in a is_device_ptr clause we just have to | ||||||
8434 | // pass the pointer by value. If it is a reference to a declaration, we just | ||||||
8435 | // pass its value. | ||||||
8436 | if (DevPointersMap.count(VD)) { | ||||||
8437 | CombinedInfo.BasePointers.emplace_back(Arg, VD); | ||||||
8438 | CombinedInfo.Pointers.push_back(Arg); | ||||||
8439 | CombinedInfo.Sizes.push_back( | ||||||
8440 | CGF.Builder.CreateIntCast(CGF.getTypeSize(CGF.getContext().VoidPtrTy), | ||||||
8441 | CGF.Int64Ty, /*isSigned=*/true)); | ||||||
8442 | CombinedInfo.Types.push_back(OMP_MAP_LITERAL | OMP_MAP_TARGET_PARAM); | ||||||
8443 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
8444 | return; | ||||||
8445 | } | ||||||
8446 | |||||||
8447 | using MapData = | ||||||
8448 | std::tuple<OMPClauseMappableExprCommon::MappableExprComponentListRef, | ||||||
8449 | OpenMPMapClauseKind, ArrayRef<OpenMPMapModifierKind>, bool, | ||||||
8450 | const ValueDecl *>; | ||||||
8451 | SmallVector<MapData, 4> DeclComponentLists; | ||||||
8452 | assert(CurDir.is<const OMPExecutableDirective *>() &&((CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? static_cast<void> ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8453, __PRETTY_FUNCTION__)) | ||||||
8453 | "Expect a executable directive")((CurDir.is<const OMPExecutableDirective *>() && "Expect a executable directive") ? static_cast<void> ( 0) : __assert_fail ("CurDir.is<const OMPExecutableDirective *>() && \"Expect a executable directive\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8453, __PRETTY_FUNCTION__)); | ||||||
8454 | const auto *CurExecDir = CurDir.get<const OMPExecutableDirective *>(); | ||||||
8455 | for (const auto *C : CurExecDir->getClausesOfKind<OMPMapClause>()) { | ||||||
8456 | for (const auto L : C->decl_component_lists(VD)) { | ||||||
8457 | const ValueDecl *VDecl, *Mapper; | ||||||
8458 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | ||||||
8459 | std::tie(VDecl, Components, Mapper) = L; | ||||||
8460 | assert(VDecl == VD && "We got information for the wrong declaration??")((VDecl == VD && "We got information for the wrong declaration??" ) ? static_cast<void> (0) : __assert_fail ("VDecl == VD && \"We got information for the wrong declaration??\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8460, __PRETTY_FUNCTION__)); | ||||||
8461 | assert(!Components.empty() &&((!Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8462, __PRETTY_FUNCTION__)) | ||||||
8462 | "Not expecting declaration with no component lists.")((!Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8462, __PRETTY_FUNCTION__)); | ||||||
8463 | DeclComponentLists.emplace_back(Components, C->getMapType(), | ||||||
8464 | C->getMapTypeModifiers(), | ||||||
8465 | C->isImplicit(), Mapper); | ||||||
8466 | } | ||||||
8467 | } | ||||||
8468 | |||||||
8469 | // Find overlapping elements (including the offset from the base element). | ||||||
8470 | llvm::SmallDenseMap< | ||||||
8471 | const MapData *, | ||||||
8472 | llvm::SmallVector< | ||||||
8473 | OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>, | ||||||
8474 | 4> | ||||||
8475 | OverlappedData; | ||||||
8476 | size_t Count = 0; | ||||||
8477 | for (const MapData &L : DeclComponentLists) { | ||||||
8478 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | ||||||
8479 | OpenMPMapClauseKind MapType; | ||||||
8480 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | ||||||
8481 | bool IsImplicit; | ||||||
8482 | const ValueDecl *Mapper; | ||||||
8483 | std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper) = L; | ||||||
8484 | ++Count; | ||||||
8485 | for (const MapData &L1 : makeArrayRef(DeclComponentLists).slice(Count)) { | ||||||
8486 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components1; | ||||||
8487 | std::tie(Components1, MapType, MapModifiers, IsImplicit, Mapper) = L1; | ||||||
8488 | auto CI = Components.rbegin(); | ||||||
8489 | auto CE = Components.rend(); | ||||||
8490 | auto SI = Components1.rbegin(); | ||||||
8491 | auto SE = Components1.rend(); | ||||||
8492 | for (; CI != CE && SI != SE; ++CI, ++SI) { | ||||||
8493 | if (CI->getAssociatedExpression()->getStmtClass() != | ||||||
8494 | SI->getAssociatedExpression()->getStmtClass()) | ||||||
8495 | break; | ||||||
8496 | // Are we dealing with different variables/fields? | ||||||
8497 | if (CI->getAssociatedDeclaration() != SI->getAssociatedDeclaration()) | ||||||
8498 | break; | ||||||
8499 | } | ||||||
8500 | // Found overlapping if, at least for one component, reached the head of | ||||||
8501 | // the components list. | ||||||
8502 | if (CI == CE || SI == SE) { | ||||||
8503 | assert((CI != CE || SI != SE) &&(((CI != CE || SI != SE) && "Unexpected full match of the mapping components." ) ? static_cast<void> (0) : __assert_fail ("(CI != CE || SI != SE) && \"Unexpected full match of the mapping components.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8504, __PRETTY_FUNCTION__)) | ||||||
8504 | "Unexpected full match of the mapping components.")(((CI != CE || SI != SE) && "Unexpected full match of the mapping components." ) ? static_cast<void> (0) : __assert_fail ("(CI != CE || SI != SE) && \"Unexpected full match of the mapping components.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8504, __PRETTY_FUNCTION__)); | ||||||
8505 | const MapData &BaseData = CI == CE ? L : L1; | ||||||
8506 | OMPClauseMappableExprCommon::MappableExprComponentListRef SubData = | ||||||
8507 | SI == SE ? Components : Components1; | ||||||
8508 | auto &OverlappedElements = OverlappedData.FindAndConstruct(&BaseData); | ||||||
8509 | OverlappedElements.getSecond().push_back(SubData); | ||||||
8510 | } | ||||||
8511 | } | ||||||
8512 | } | ||||||
8513 | // Sort the overlapped elements for each item. | ||||||
8514 | llvm::SmallVector<const FieldDecl *, 4> Layout; | ||||||
8515 | if (!OverlappedData.empty()) { | ||||||
8516 | if (const auto *CRD = | ||||||
8517 | VD->getType().getCanonicalType()->getAsCXXRecordDecl()) | ||||||
8518 | getPlainLayout(CRD, Layout, /*AsBase=*/false); | ||||||
8519 | else { | ||||||
8520 | const auto *RD = VD->getType().getCanonicalType()->getAsRecordDecl(); | ||||||
8521 | Layout.append(RD->field_begin(), RD->field_end()); | ||||||
8522 | } | ||||||
8523 | } | ||||||
8524 | for (auto &Pair : OverlappedData) { | ||||||
8525 | llvm::sort( | ||||||
8526 | Pair.getSecond(), | ||||||
8527 | [&Layout]( | ||||||
8528 | OMPClauseMappableExprCommon::MappableExprComponentListRef First, | ||||||
8529 | OMPClauseMappableExprCommon::MappableExprComponentListRef | ||||||
8530 | Second) { | ||||||
8531 | auto CI = First.rbegin(); | ||||||
8532 | auto CE = First.rend(); | ||||||
8533 | auto SI = Second.rbegin(); | ||||||
8534 | auto SE = Second.rend(); | ||||||
8535 | for (; CI != CE && SI != SE; ++CI, ++SI) { | ||||||
8536 | if (CI->getAssociatedExpression()->getStmtClass() != | ||||||
8537 | SI->getAssociatedExpression()->getStmtClass()) | ||||||
8538 | break; | ||||||
8539 | // Are we dealing with different variables/fields? | ||||||
8540 | if (CI->getAssociatedDeclaration() != | ||||||
8541 | SI->getAssociatedDeclaration()) | ||||||
8542 | break; | ||||||
8543 | } | ||||||
8544 | |||||||
8545 | // Lists contain the same elements. | ||||||
8546 | if (CI == CE && SI == SE) | ||||||
8547 | return false; | ||||||
8548 | |||||||
8549 | // List with less elements is less than list with more elements. | ||||||
8550 | if (CI == CE || SI == SE) | ||||||
8551 | return CI == CE; | ||||||
8552 | |||||||
8553 | const auto *FD1 = cast<FieldDecl>(CI->getAssociatedDeclaration()); | ||||||
8554 | const auto *FD2 = cast<FieldDecl>(SI->getAssociatedDeclaration()); | ||||||
8555 | if (FD1->getParent() == FD2->getParent()) | ||||||
8556 | return FD1->getFieldIndex() < FD2->getFieldIndex(); | ||||||
8557 | const auto It = | ||||||
8558 | llvm::find_if(Layout, [FD1, FD2](const FieldDecl *FD) { | ||||||
8559 | return FD == FD1 || FD == FD2; | ||||||
8560 | }); | ||||||
8561 | return *It == FD1; | ||||||
8562 | }); | ||||||
8563 | } | ||||||
8564 | |||||||
8565 | // Associated with a capture, because the mapping flags depend on it. | ||||||
8566 | // Go through all of the elements with the overlapped elements. | ||||||
8567 | for (const auto &Pair : OverlappedData) { | ||||||
8568 | const MapData &L = *Pair.getFirst(); | ||||||
8569 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | ||||||
8570 | OpenMPMapClauseKind MapType; | ||||||
8571 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | ||||||
8572 | bool IsImplicit; | ||||||
8573 | const ValueDecl *Mapper; | ||||||
8574 | std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper) = L; | ||||||
8575 | ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> | ||||||
8576 | OverlappedComponents = Pair.getSecond(); | ||||||
8577 | bool IsFirstComponentList = true; | ||||||
8578 | generateInfoForComponentList( | ||||||
8579 | MapType, MapModifiers, llvm::None, Components, CombinedInfo, | ||||||
8580 | PartialStruct, IsFirstComponentList, IsImplicit, Mapper, | ||||||
8581 | /*ForDeviceAddr=*/false, OverlappedComponents); | ||||||
8582 | } | ||||||
8583 | // Go through other elements without overlapped elements. | ||||||
8584 | bool IsFirstComponentList = OverlappedData.empty(); | ||||||
8585 | for (const MapData &L : DeclComponentLists) { | ||||||
8586 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | ||||||
8587 | OpenMPMapClauseKind MapType; | ||||||
8588 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | ||||||
8589 | bool IsImplicit; | ||||||
8590 | const ValueDecl *Mapper; | ||||||
8591 | std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper) = L; | ||||||
8592 | auto It = OverlappedData.find(&L); | ||||||
8593 | if (It == OverlappedData.end()) | ||||||
8594 | generateInfoForComponentList(MapType, MapModifiers, llvm::None, | ||||||
8595 | Components, CombinedInfo, PartialStruct, | ||||||
8596 | IsFirstComponentList, IsImplicit, Mapper); | ||||||
8597 | IsFirstComponentList = false; | ||||||
8598 | } | ||||||
8599 | } | ||||||
8600 | |||||||
8601 | /// Generate the default map information for a given capture \a CI, | ||||||
8602 | /// record field declaration \a RI and captured value \a CV. | ||||||
8603 | void generateDefaultMapInfo(const CapturedStmt::Capture &CI, | ||||||
8604 | const FieldDecl &RI, llvm::Value *CV, | ||||||
8605 | MapCombinedInfoTy &CombinedInfo) const { | ||||||
8606 | bool IsImplicit = true; | ||||||
8607 | // Do the default mapping. | ||||||
8608 | if (CI.capturesThis()) { | ||||||
8609 | CombinedInfo.BasePointers.push_back(CV); | ||||||
8610 | CombinedInfo.Pointers.push_back(CV); | ||||||
8611 | const auto *PtrTy = cast<PointerType>(RI.getType().getTypePtr()); | ||||||
8612 | CombinedInfo.Sizes.push_back( | ||||||
8613 | CGF.Builder.CreateIntCast(CGF.getTypeSize(PtrTy->getPointeeType()), | ||||||
8614 | CGF.Int64Ty, /*isSigned=*/true)); | ||||||
8615 | // Default map type. | ||||||
8616 | CombinedInfo.Types.push_back(OMP_MAP_TO | OMP_MAP_FROM); | ||||||
8617 | } else if (CI.capturesVariableByCopy()) { | ||||||
8618 | CombinedInfo.BasePointers.push_back(CV); | ||||||
8619 | CombinedInfo.Pointers.push_back(CV); | ||||||
8620 | if (!RI.getType()->isAnyPointerType()) { | ||||||
8621 | // We have to signal to the runtime captures passed by value that are | ||||||
8622 | // not pointers. | ||||||
8623 | CombinedInfo.Types.push_back(OMP_MAP_LITERAL); | ||||||
8624 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | ||||||
8625 | CGF.getTypeSize(RI.getType()), CGF.Int64Ty, /*isSigned=*/true)); | ||||||
8626 | } else { | ||||||
8627 | // Pointers are implicitly mapped with a zero size and no flags | ||||||
8628 | // (other than first map that is added for all implicit maps). | ||||||
8629 | CombinedInfo.Types.push_back(OMP_MAP_NONE); | ||||||
8630 | CombinedInfo.Sizes.push_back(llvm::Constant::getNullValue(CGF.Int64Ty)); | ||||||
8631 | } | ||||||
8632 | const VarDecl *VD = CI.getCapturedVar(); | ||||||
8633 | auto I = FirstPrivateDecls.find(VD); | ||||||
8634 | if (I != FirstPrivateDecls.end()) | ||||||
8635 | IsImplicit = I->getSecond(); | ||||||
8636 | } else { | ||||||
8637 | assert(CI.capturesVariable() && "Expected captured reference.")((CI.capturesVariable() && "Expected captured reference." ) ? static_cast<void> (0) : __assert_fail ("CI.capturesVariable() && \"Expected captured reference.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8637, __PRETTY_FUNCTION__)); | ||||||
8638 | const auto *PtrTy = cast<ReferenceType>(RI.getType().getTypePtr()); | ||||||
8639 | QualType ElementType = PtrTy->getPointeeType(); | ||||||
8640 | CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | ||||||
8641 | CGF.getTypeSize(ElementType), CGF.Int64Ty, /*isSigned=*/true)); | ||||||
8642 | // The default map type for a scalar/complex type is 'to' because by | ||||||
8643 | // default the value doesn't have to be retrieved. For an aggregate | ||||||
8644 | // type, the default is 'tofrom'. | ||||||
8645 | CombinedInfo.Types.push_back(getMapModifiersForPrivateClauses(CI)); | ||||||
8646 | const VarDecl *VD = CI.getCapturedVar(); | ||||||
8647 | auto I = FirstPrivateDecls.find(VD); | ||||||
8648 | if (I != FirstPrivateDecls.end() && | ||||||
8649 | VD->getType().isConstant(CGF.getContext())) { | ||||||
8650 | llvm::Constant *Addr = | ||||||
8651 | CGF.CGM.getOpenMPRuntime().registerTargetFirstprivateCopy(CGF, VD); | ||||||
8652 | // Copy the value of the original variable to the new global copy. | ||||||
8653 | CGF.Builder.CreateMemCpy( | ||||||
8654 | CGF.MakeNaturalAlignAddrLValue(Addr, ElementType).getAddress(CGF), | ||||||
8655 | Address(CV, CGF.getContext().getTypeAlignInChars(ElementType)), | ||||||
8656 | CombinedInfo.Sizes.back(), /*IsVolatile=*/false); | ||||||
8657 | // Use new global variable as the base pointers. | ||||||
8658 | CombinedInfo.BasePointers.push_back(Addr); | ||||||
8659 | CombinedInfo.Pointers.push_back(Addr); | ||||||
8660 | } else { | ||||||
8661 | CombinedInfo.BasePointers.push_back(CV); | ||||||
8662 | if (I != FirstPrivateDecls.end() && ElementType->isAnyPointerType()) { | ||||||
8663 | Address PtrAddr = CGF.EmitLoadOfReference(CGF.MakeAddrLValue( | ||||||
8664 | CV, ElementType, CGF.getContext().getDeclAlign(VD), | ||||||
8665 | AlignmentSource::Decl)); | ||||||
8666 | CombinedInfo.Pointers.push_back(PtrAddr.getPointer()); | ||||||
8667 | } else { | ||||||
8668 | CombinedInfo.Pointers.push_back(CV); | ||||||
8669 | } | ||||||
8670 | } | ||||||
8671 | if (I != FirstPrivateDecls.end()) | ||||||
8672 | IsImplicit = I->getSecond(); | ||||||
8673 | } | ||||||
8674 | // Every default map produces a single argument which is a target parameter. | ||||||
8675 | CombinedInfo.Types.back() |= OMP_MAP_TARGET_PARAM; | ||||||
8676 | |||||||
8677 | // Add flag stating this is an implicit map. | ||||||
8678 | if (IsImplicit) | ||||||
8679 | CombinedInfo.Types.back() |= OMP_MAP_IMPLICIT; | ||||||
8680 | |||||||
8681 | // No user-defined mapper for default mapping. | ||||||
8682 | CombinedInfo.Mappers.push_back(nullptr); | ||||||
8683 | } | ||||||
8684 | }; | ||||||
8685 | } // anonymous namespace | ||||||
8686 | |||||||
8687 | /// Emit the arrays used to pass the captures and map information to the | ||||||
8688 | /// offloading runtime library. If there is no map or capture information, | ||||||
8689 | /// return nullptr by reference. | ||||||
8690 | static void | ||||||
8691 | emitOffloadingArrays(CodeGenFunction &CGF, | ||||||
8692 | MappableExprsHandler::MapCombinedInfoTy &CombinedInfo, | ||||||
8693 | CGOpenMPRuntime::TargetDataInfo &Info) { | ||||||
8694 | CodeGenModule &CGM = CGF.CGM; | ||||||
8695 | ASTContext &Ctx = CGF.getContext(); | ||||||
8696 | |||||||
8697 | // Reset the array information. | ||||||
8698 | Info.clearArrayInfo(); | ||||||
8699 | Info.NumberOfPtrs = CombinedInfo.BasePointers.size(); | ||||||
8700 | |||||||
8701 | if (Info.NumberOfPtrs) { | ||||||
8702 | // Detect if we have any capture size requiring runtime evaluation of the | ||||||
8703 | // size so that a constant array could be eventually used. | ||||||
8704 | bool hasRuntimeEvaluationCaptureSize = false; | ||||||
8705 | for (llvm::Value *S : CombinedInfo.Sizes) | ||||||
8706 | if (!isa<llvm::Constant>(S)) { | ||||||
8707 | hasRuntimeEvaluationCaptureSize = true; | ||||||
8708 | break; | ||||||
8709 | } | ||||||
8710 | |||||||
8711 | llvm::APInt PointerNumAP(32, Info.NumberOfPtrs, /*isSigned=*/true); | ||||||
8712 | QualType PointerArrayType = Ctx.getConstantArrayType( | ||||||
8713 | Ctx.VoidPtrTy, PointerNumAP, nullptr, ArrayType::Normal, | ||||||
8714 | /*IndexTypeQuals=*/0); | ||||||
8715 | |||||||
8716 | Info.BasePointersArray = | ||||||
8717 | CGF.CreateMemTemp(PointerArrayType, ".offload_baseptrs").getPointer(); | ||||||
8718 | Info.PointersArray = | ||||||
8719 | CGF.CreateMemTemp(PointerArrayType, ".offload_ptrs").getPointer(); | ||||||
8720 | Address MappersArray = | ||||||
8721 | CGF.CreateMemTemp(PointerArrayType, ".offload_mappers"); | ||||||
8722 | Info.MappersArray = MappersArray.getPointer(); | ||||||
8723 | |||||||
8724 | // If we don't have any VLA types or other types that require runtime | ||||||
8725 | // evaluation, we can use a constant array for the map sizes, otherwise we | ||||||
8726 | // need to fill up the arrays as we do for the pointers. | ||||||
8727 | QualType Int64Ty = | ||||||
8728 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | ||||||
8729 | if (hasRuntimeEvaluationCaptureSize) { | ||||||
8730 | QualType SizeArrayType = Ctx.getConstantArrayType( | ||||||
8731 | Int64Ty, PointerNumAP, nullptr, ArrayType::Normal, | ||||||
8732 | /*IndexTypeQuals=*/0); | ||||||
8733 | Info.SizesArray = | ||||||
8734 | CGF.CreateMemTemp(SizeArrayType, ".offload_sizes").getPointer(); | ||||||
8735 | } else { | ||||||
8736 | // We expect all the sizes to be constant, so we collect them to create | ||||||
8737 | // a constant array. | ||||||
8738 | SmallVector<llvm::Constant *, 16> ConstSizes; | ||||||
8739 | for (llvm::Value *S : CombinedInfo.Sizes) | ||||||
8740 | ConstSizes.push_back(cast<llvm::Constant>(S)); | ||||||
8741 | |||||||
8742 | auto *SizesArrayInit = llvm::ConstantArray::get( | ||||||
8743 | llvm::ArrayType::get(CGM.Int64Ty, ConstSizes.size()), ConstSizes); | ||||||
8744 | std::string Name = CGM.getOpenMPRuntime().getName({"offload_sizes"}); | ||||||
8745 | auto *SizesArrayGbl = new llvm::GlobalVariable( | ||||||
8746 | CGM.getModule(), SizesArrayInit->getType(), | ||||||
8747 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, | ||||||
8748 | SizesArrayInit, Name); | ||||||
8749 | SizesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | ||||||
8750 | Info.SizesArray = SizesArrayGbl; | ||||||
8751 | } | ||||||
8752 | |||||||
8753 | // The map types are always constant so we don't need to generate code to | ||||||
8754 | // fill arrays. Instead, we create an array constant. | ||||||
8755 | SmallVector<uint64_t, 4> Mapping(CombinedInfo.Types.size(), 0); | ||||||
8756 | llvm::copy(CombinedInfo.Types, Mapping.begin()); | ||||||
8757 | llvm::Constant *MapTypesArrayInit = | ||||||
8758 | llvm::ConstantDataArray::get(CGF.Builder.getContext(), Mapping); | ||||||
8759 | std::string MaptypesName = | ||||||
8760 | CGM.getOpenMPRuntime().getName({"offload_maptypes"}); | ||||||
8761 | auto *MapTypesArrayGbl = new llvm::GlobalVariable( | ||||||
8762 | CGM.getModule(), MapTypesArrayInit->getType(), | ||||||
8763 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, | ||||||
8764 | MapTypesArrayInit, MaptypesName); | ||||||
8765 | MapTypesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | ||||||
8766 | Info.MapTypesArray = MapTypesArrayGbl; | ||||||
8767 | |||||||
8768 | // If there's a present map type modifier, it must not be applied to the end | ||||||
8769 | // of a region, so generate a separate map type array in that case. | ||||||
8770 | if (Info.separateBeginEndCalls()) { | ||||||
8771 | bool EndMapTypesDiffer = false; | ||||||
8772 | for (uint64_t &Type : Mapping) { | ||||||
8773 | if (Type & MappableExprsHandler::OMP_MAP_PRESENT) { | ||||||
8774 | Type &= ~MappableExprsHandler::OMP_MAP_PRESENT; | ||||||
8775 | EndMapTypesDiffer = true; | ||||||
8776 | } | ||||||
8777 | } | ||||||
8778 | if (EndMapTypesDiffer) { | ||||||
8779 | MapTypesArrayInit = | ||||||
8780 | llvm::ConstantDataArray::get(CGF.Builder.getContext(), Mapping); | ||||||
8781 | MaptypesName = CGM.getOpenMPRuntime().getName({"offload_maptypes"}); | ||||||
8782 | MapTypesArrayGbl = new llvm::GlobalVariable( | ||||||
8783 | CGM.getModule(), MapTypesArrayInit->getType(), | ||||||
8784 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, | ||||||
8785 | MapTypesArrayInit, MaptypesName); | ||||||
8786 | MapTypesArrayGbl->setUnnamedAddr( | ||||||
8787 | llvm::GlobalValue::UnnamedAddr::Global); | ||||||
8788 | Info.MapTypesArrayEnd = MapTypesArrayGbl; | ||||||
8789 | } | ||||||
8790 | } | ||||||
8791 | |||||||
8792 | for (unsigned I = 0; I < Info.NumberOfPtrs; ++I) { | ||||||
8793 | llvm::Value *BPVal = *CombinedInfo.BasePointers[I]; | ||||||
8794 | llvm::Value *BP = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8795 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | ||||||
8796 | Info.BasePointersArray, 0, I); | ||||||
8797 | BP = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
8798 | BP, BPVal->getType()->getPointerTo(/*AddrSpace=*/0)); | ||||||
8799 | Address BPAddr(BP, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy)); | ||||||
8800 | CGF.Builder.CreateStore(BPVal, BPAddr); | ||||||
8801 | |||||||
8802 | if (Info.requiresDevicePointerInfo()) | ||||||
8803 | if (const ValueDecl *DevVD = | ||||||
8804 | CombinedInfo.BasePointers[I].getDevicePtrDecl()) | ||||||
8805 | Info.CaptureDeviceAddrMap.try_emplace(DevVD, BPAddr); | ||||||
8806 | |||||||
8807 | llvm::Value *PVal = CombinedInfo.Pointers[I]; | ||||||
8808 | llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8809 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | ||||||
8810 | Info.PointersArray, 0, I); | ||||||
8811 | P = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
8812 | P, PVal->getType()->getPointerTo(/*AddrSpace=*/0)); | ||||||
8813 | Address PAddr(P, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy)); | ||||||
8814 | CGF.Builder.CreateStore(PVal, PAddr); | ||||||
8815 | |||||||
8816 | if (hasRuntimeEvaluationCaptureSize) { | ||||||
8817 | llvm::Value *S = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8818 | llvm::ArrayType::get(CGM.Int64Ty, Info.NumberOfPtrs), | ||||||
8819 | Info.SizesArray, | ||||||
8820 | /*Idx0=*/0, | ||||||
8821 | /*Idx1=*/I); | ||||||
8822 | Address SAddr(S, Ctx.getTypeAlignInChars(Int64Ty)); | ||||||
8823 | CGF.Builder.CreateStore(CGF.Builder.CreateIntCast(CombinedInfo.Sizes[I], | ||||||
8824 | CGM.Int64Ty, | ||||||
8825 | /*isSigned=*/true), | ||||||
8826 | SAddr); | ||||||
8827 | } | ||||||
8828 | |||||||
8829 | // Fill up the mapper array. | ||||||
8830 | llvm::Value *MFunc = llvm::ConstantPointerNull::get(CGM.VoidPtrTy); | ||||||
8831 | if (CombinedInfo.Mappers[I]) { | ||||||
8832 | MFunc = CGM.getOpenMPRuntime().getOrCreateUserDefinedMapperFunc( | ||||||
8833 | cast<OMPDeclareMapperDecl>(CombinedInfo.Mappers[I])); | ||||||
8834 | MFunc = CGF.Builder.CreatePointerCast(MFunc, CGM.VoidPtrTy); | ||||||
8835 | Info.HasMapper = true; | ||||||
8836 | } | ||||||
8837 | Address MAddr = CGF.Builder.CreateConstArrayGEP(MappersArray, I); | ||||||
8838 | CGF.Builder.CreateStore(MFunc, MAddr); | ||||||
8839 | } | ||||||
8840 | } | ||||||
8841 | } | ||||||
8842 | |||||||
8843 | /// Emit the arguments to be passed to the runtime library based on the | ||||||
8844 | /// arrays of base pointers, pointers, sizes, map types, and mappers. If | ||||||
8845 | /// ForEndCall, emit map types to be passed for the end of the region instead of | ||||||
8846 | /// the beginning. | ||||||
8847 | static void emitOffloadingArraysArgument( | ||||||
8848 | CodeGenFunction &CGF, llvm::Value *&BasePointersArrayArg, | ||||||
8849 | llvm::Value *&PointersArrayArg, llvm::Value *&SizesArrayArg, | ||||||
8850 | llvm::Value *&MapTypesArrayArg, llvm::Value *&MappersArrayArg, | ||||||
8851 | CGOpenMPRuntime::TargetDataInfo &Info, bool ForEndCall = false) { | ||||||
8852 | assert((!ForEndCall || Info.separateBeginEndCalls()) &&(((!ForEndCall || Info.separateBeginEndCalls()) && "expected region end call to runtime only when end call is separate" ) ? static_cast<void> (0) : __assert_fail ("(!ForEndCall || Info.separateBeginEndCalls()) && \"expected region end call to runtime only when end call is separate\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8853, __PRETTY_FUNCTION__)) | ||||||
8853 | "expected region end call to runtime only when end call is separate")(((!ForEndCall || Info.separateBeginEndCalls()) && "expected region end call to runtime only when end call is separate" ) ? static_cast<void> (0) : __assert_fail ("(!ForEndCall || Info.separateBeginEndCalls()) && \"expected region end call to runtime only when end call is separate\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8853, __PRETTY_FUNCTION__)); | ||||||
8854 | CodeGenModule &CGM = CGF.CGM; | ||||||
8855 | if (Info.NumberOfPtrs) { | ||||||
8856 | BasePointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8857 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | ||||||
8858 | Info.BasePointersArray, | ||||||
8859 | /*Idx0=*/0, /*Idx1=*/0); | ||||||
8860 | PointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8861 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | ||||||
8862 | Info.PointersArray, | ||||||
8863 | /*Idx0=*/0, | ||||||
8864 | /*Idx1=*/0); | ||||||
8865 | SizesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8866 | llvm::ArrayType::get(CGM.Int64Ty, Info.NumberOfPtrs), Info.SizesArray, | ||||||
8867 | /*Idx0=*/0, /*Idx1=*/0); | ||||||
8868 | MapTypesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | ||||||
8869 | llvm::ArrayType::get(CGM.Int64Ty, Info.NumberOfPtrs), | ||||||
8870 | ForEndCall && Info.MapTypesArrayEnd ? Info.MapTypesArrayEnd | ||||||
8871 | : Info.MapTypesArray, | ||||||
8872 | /*Idx0=*/0, | ||||||
8873 | /*Idx1=*/0); | ||||||
8874 | MappersArrayArg = | ||||||
8875 | Info.HasMapper | ||||||
8876 | ? CGF.Builder.CreatePointerCast(Info.MappersArray, CGM.VoidPtrPtrTy) | ||||||
8877 | : llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy); | ||||||
8878 | } else { | ||||||
8879 | BasePointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy); | ||||||
8880 | PointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy); | ||||||
8881 | SizesArrayArg = llvm::ConstantPointerNull::get(CGM.Int64Ty->getPointerTo()); | ||||||
8882 | MapTypesArrayArg = | ||||||
8883 | llvm::ConstantPointerNull::get(CGM.Int64Ty->getPointerTo()); | ||||||
8884 | MappersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy); | ||||||
8885 | } | ||||||
8886 | } | ||||||
8887 | |||||||
8888 | /// Check for inner distribute directive. | ||||||
8889 | static const OMPExecutableDirective * | ||||||
8890 | getNestedDistributeDirective(ASTContext &Ctx, const OMPExecutableDirective &D) { | ||||||
8891 | const auto *CS = D.getInnermostCapturedStmt(); | ||||||
8892 | const auto *Body = | ||||||
8893 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | ||||||
8894 | const Stmt *ChildStmt = | ||||||
8895 | CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body); | ||||||
8896 | |||||||
8897 | if (const auto *NestedDir = | ||||||
8898 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | ||||||
8899 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); | ||||||
8900 | switch (D.getDirectiveKind()) { | ||||||
8901 | case OMPD_target: | ||||||
8902 | if (isOpenMPDistributeDirective(DKind)) | ||||||
8903 | return NestedDir; | ||||||
8904 | if (DKind == OMPD_teams) { | ||||||
8905 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | ||||||
8906 | /*IgnoreCaptured=*/true); | ||||||
8907 | if (!Body) | ||||||
8908 | return nullptr; | ||||||
8909 | ChildStmt = CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body); | ||||||
8910 | if (const auto *NND = | ||||||
8911 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | ||||||
8912 | DKind = NND->getDirectiveKind(); | ||||||
8913 | if (isOpenMPDistributeDirective(DKind)) | ||||||
8914 | return NND; | ||||||
8915 | } | ||||||
8916 | } | ||||||
8917 | return nullptr; | ||||||
8918 | case OMPD_target_teams: | ||||||
8919 | if (isOpenMPDistributeDirective(DKind)) | ||||||
8920 | return NestedDir; | ||||||
8921 | return nullptr; | ||||||
8922 | case OMPD_target_parallel: | ||||||
8923 | case OMPD_target_simd: | ||||||
8924 | case OMPD_target_parallel_for: | ||||||
8925 | case OMPD_target_parallel_for_simd: | ||||||
8926 | return nullptr; | ||||||
8927 | case OMPD_target_teams_distribute: | ||||||
8928 | case OMPD_target_teams_distribute_simd: | ||||||
8929 | case OMPD_target_teams_distribute_parallel_for: | ||||||
8930 | case OMPD_target_teams_distribute_parallel_for_simd: | ||||||
8931 | case OMPD_parallel: | ||||||
8932 | case OMPD_for: | ||||||
8933 | case OMPD_parallel_for: | ||||||
8934 | case OMPD_parallel_master: | ||||||
8935 | case OMPD_parallel_sections: | ||||||
8936 | case OMPD_for_simd: | ||||||
8937 | case OMPD_parallel_for_simd: | ||||||
8938 | case OMPD_cancel: | ||||||
8939 | case OMPD_cancellation_point: | ||||||
8940 | case OMPD_ordered: | ||||||
8941 | case OMPD_threadprivate: | ||||||
8942 | case OMPD_allocate: | ||||||
8943 | case OMPD_task: | ||||||
8944 | case OMPD_simd: | ||||||
8945 | case OMPD_sections: | ||||||
8946 | case OMPD_section: | ||||||
8947 | case OMPD_single: | ||||||
8948 | case OMPD_master: | ||||||
8949 | case OMPD_critical: | ||||||
8950 | case OMPD_taskyield: | ||||||
8951 | case OMPD_barrier: | ||||||
8952 | case OMPD_taskwait: | ||||||
8953 | case OMPD_taskgroup: | ||||||
8954 | case OMPD_atomic: | ||||||
8955 | case OMPD_flush: | ||||||
8956 | case OMPD_depobj: | ||||||
8957 | case OMPD_scan: | ||||||
8958 | case OMPD_teams: | ||||||
8959 | case OMPD_target_data: | ||||||
8960 | case OMPD_target_exit_data: | ||||||
8961 | case OMPD_target_enter_data: | ||||||
8962 | case OMPD_distribute: | ||||||
8963 | case OMPD_distribute_simd: | ||||||
8964 | case OMPD_distribute_parallel_for: | ||||||
8965 | case OMPD_distribute_parallel_for_simd: | ||||||
8966 | case OMPD_teams_distribute: | ||||||
8967 | case OMPD_teams_distribute_simd: | ||||||
8968 | case OMPD_teams_distribute_parallel_for: | ||||||
8969 | case OMPD_teams_distribute_parallel_for_simd: | ||||||
8970 | case OMPD_target_update: | ||||||
8971 | case OMPD_declare_simd: | ||||||
8972 | case OMPD_declare_variant: | ||||||
8973 | case OMPD_begin_declare_variant: | ||||||
8974 | case OMPD_end_declare_variant: | ||||||
8975 | case OMPD_declare_target: | ||||||
8976 | case OMPD_end_declare_target: | ||||||
8977 | case OMPD_declare_reduction: | ||||||
8978 | case OMPD_declare_mapper: | ||||||
8979 | case OMPD_taskloop: | ||||||
8980 | case OMPD_taskloop_simd: | ||||||
8981 | case OMPD_master_taskloop: | ||||||
8982 | case OMPD_master_taskloop_simd: | ||||||
8983 | case OMPD_parallel_master_taskloop: | ||||||
8984 | case OMPD_parallel_master_taskloop_simd: | ||||||
8985 | case OMPD_requires: | ||||||
8986 | case OMPD_unknown: | ||||||
8987 | default: | ||||||
8988 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8988); | ||||||
8989 | } | ||||||
8990 | } | ||||||
8991 | |||||||
8992 | return nullptr; | ||||||
8993 | } | ||||||
8994 | |||||||
8995 | /// Emit the user-defined mapper function. The code generation follows the | ||||||
8996 | /// pattern in the example below. | ||||||
8997 | /// \code | ||||||
8998 | /// void .omp_mapper.<type_name>.<mapper_id>.(void *rt_mapper_handle, | ||||||
8999 | /// void *base, void *begin, | ||||||
9000 | /// int64_t size, int64_t type) { | ||||||
9001 | /// // Allocate space for an array section first. | ||||||
9002 | /// if (size > 1 && !maptype.IsDelete) | ||||||
9003 | /// __tgt_push_mapper_component(rt_mapper_handle, base, begin, | ||||||
9004 | /// size*sizeof(Ty), clearToFrom(type)); | ||||||
9005 | /// // Map members. | ||||||
9006 | /// for (unsigned i = 0; i < size; i++) { | ||||||
9007 | /// // For each component specified by this mapper: | ||||||
9008 | /// for (auto c : all_components) { | ||||||
9009 | /// if (c.hasMapper()) | ||||||
9010 | /// (*c.Mapper())(rt_mapper_handle, c.arg_base, c.arg_begin, c.arg_size, | ||||||
9011 | /// c.arg_type); | ||||||
9012 | /// else | ||||||
9013 | /// __tgt_push_mapper_component(rt_mapper_handle, c.arg_base, | ||||||
9014 | /// c.arg_begin, c.arg_size, c.arg_type); | ||||||
9015 | /// } | ||||||
9016 | /// } | ||||||
9017 | /// // Delete the array section. | ||||||
9018 | /// if (size > 1 && maptype.IsDelete) | ||||||
9019 | /// __tgt_push_mapper_component(rt_mapper_handle, base, begin, | ||||||
9020 | /// size*sizeof(Ty), clearToFrom(type)); | ||||||
9021 | /// } | ||||||
9022 | /// \endcode | ||||||
9023 | void CGOpenMPRuntime::emitUserDefinedMapper(const OMPDeclareMapperDecl *D, | ||||||
9024 | CodeGenFunction *CGF) { | ||||||
9025 | if (UDMMap.count(D) > 0) | ||||||
9026 | return; | ||||||
9027 | ASTContext &C = CGM.getContext(); | ||||||
9028 | QualType Ty = D->getType(); | ||||||
9029 | QualType PtrTy = C.getPointerType(Ty).withRestrict(); | ||||||
9030 | QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true); | ||||||
9031 | auto *MapperVarDecl = | ||||||
9032 | cast<VarDecl>(cast<DeclRefExpr>(D->getMapperVarRef())->getDecl()); | ||||||
9033 | SourceLocation Loc = D->getLocation(); | ||||||
9034 | CharUnits ElementSize = C.getTypeSizeInChars(Ty); | ||||||
9035 | |||||||
9036 | // Prepare mapper function arguments and attributes. | ||||||
9037 | ImplicitParamDecl HandleArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
9038 | C.VoidPtrTy, ImplicitParamDecl::Other); | ||||||
9039 | ImplicitParamDecl BaseArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | ||||||
9040 | ImplicitParamDecl::Other); | ||||||
9041 | ImplicitParamDecl BeginArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | ||||||
9042 | C.VoidPtrTy, ImplicitParamDecl::Other); | ||||||
9043 | ImplicitParamDecl SizeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, Int64Ty, | ||||||
9044 | ImplicitParamDecl::Other); | ||||||
9045 | ImplicitParamDecl TypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, Int64Ty, | ||||||
9046 | ImplicitParamDecl::Other); | ||||||
9047 | FunctionArgList Args; | ||||||
9048 | Args.push_back(&HandleArg); | ||||||
9049 | Args.push_back(&BaseArg); | ||||||
9050 | Args.push_back(&BeginArg); | ||||||
9051 | Args.push_back(&SizeArg); | ||||||
9052 | Args.push_back(&TypeArg); | ||||||
9053 | const CGFunctionInfo &FnInfo = | ||||||
9054 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | ||||||
9055 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | ||||||
9056 | SmallString<64> TyStr; | ||||||
9057 | llvm::raw_svector_ostream Out(TyStr); | ||||||
9058 | CGM.getCXXABI().getMangleContext().mangleTypeName(Ty, Out); | ||||||
9059 | std::string Name = getName({"omp_mapper", TyStr, D->getName()}); | ||||||
9060 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | ||||||
9061 | Name, &CGM.getModule()); | ||||||
9062 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | ||||||
9063 | Fn->removeFnAttr(llvm::Attribute::OptimizeNone); | ||||||
9064 | // Start the mapper function code generation. | ||||||
9065 | CodeGenFunction MapperCGF(CGM); | ||||||
9066 | MapperCGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | ||||||
9067 | // Compute the starting and end addreses of array elements. | ||||||
9068 | llvm::Value *Size = MapperCGF.EmitLoadOfScalar( | ||||||
9069 | MapperCGF.GetAddrOfLocalVar(&SizeArg), /*Volatile=*/false, | ||||||
9070 | C.getPointerType(Int64Ty), Loc); | ||||||
9071 | // Convert the size in bytes into the number of array elements. | ||||||
9072 | Size = MapperCGF.Builder.CreateExactUDiv( | ||||||
9073 | Size, MapperCGF.Builder.getInt64(ElementSize.getQuantity())); | ||||||
9074 | llvm::Value *PtrBegin = MapperCGF.Builder.CreateBitCast( | ||||||
9075 | MapperCGF.GetAddrOfLocalVar(&BeginArg).getPointer(), | ||||||
9076 | CGM.getTypes().ConvertTypeForMem(C.getPointerType(PtrTy))); | ||||||
9077 | llvm::Value *PtrEnd = MapperCGF.Builder.CreateGEP(PtrBegin, Size); | ||||||
9078 | llvm::Value *MapType = MapperCGF.EmitLoadOfScalar( | ||||||
9079 | MapperCGF.GetAddrOfLocalVar(&TypeArg), /*Volatile=*/false, | ||||||
9080 | C.getPointerType(Int64Ty), Loc); | ||||||
9081 | // Prepare common arguments for array initiation and deletion. | ||||||
9082 | llvm::Value *Handle = MapperCGF.EmitLoadOfScalar( | ||||||
9083 | MapperCGF.GetAddrOfLocalVar(&HandleArg), | ||||||
9084 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | ||||||
9085 | llvm::Value *BaseIn = MapperCGF.EmitLoadOfScalar( | ||||||
9086 | MapperCGF.GetAddrOfLocalVar(&BaseArg), | ||||||
9087 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | ||||||
9088 | llvm::Value *BeginIn = MapperCGF.EmitLoadOfScalar( | ||||||
9089 | MapperCGF.GetAddrOfLocalVar(&BeginArg), | ||||||
9090 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | ||||||
9091 | |||||||
9092 | // Emit array initiation if this is an array section and \p MapType indicates | ||||||
9093 | // that memory allocation is required. | ||||||
9094 | llvm::BasicBlock *HeadBB = MapperCGF.createBasicBlock("omp.arraymap.head"); | ||||||
9095 | emitUDMapperArrayInitOrDel(MapperCGF, Handle, BaseIn, BeginIn, Size, MapType, | ||||||
9096 | ElementSize, HeadBB, /*IsInit=*/true); | ||||||
9097 | |||||||
9098 | // Emit a for loop to iterate through SizeArg of elements and map all of them. | ||||||
9099 | |||||||
9100 | // Emit the loop header block. | ||||||
9101 | MapperCGF.EmitBlock(HeadBB); | ||||||
9102 | llvm::BasicBlock *BodyBB = MapperCGF.createBasicBlock("omp.arraymap.body"); | ||||||
9103 | llvm::BasicBlock *DoneBB = MapperCGF.createBasicBlock("omp.done"); | ||||||
9104 | // Evaluate whether the initial condition is satisfied. | ||||||
9105 | llvm::Value *IsEmpty = | ||||||
9106 | MapperCGF.Builder.CreateICmpEQ(PtrBegin, PtrEnd, "omp.arraymap.isempty"); | ||||||
9107 | MapperCGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | ||||||
9108 | llvm::BasicBlock *EntryBB = MapperCGF.Builder.GetInsertBlock(); | ||||||
9109 | |||||||
9110 | // Emit the loop body block. | ||||||
9111 | MapperCGF.EmitBlock(BodyBB); | ||||||
9112 | llvm::BasicBlock *LastBB = BodyBB; | ||||||
9113 | llvm::PHINode *PtrPHI = MapperCGF.Builder.CreatePHI( | ||||||
9114 | PtrBegin->getType(), 2, "omp.arraymap.ptrcurrent"); | ||||||
9115 | PtrPHI->addIncoming(PtrBegin, EntryBB); | ||||||
9116 | Address PtrCurrent = | ||||||
9117 | Address(PtrPHI, MapperCGF.GetAddrOfLocalVar(&BeginArg) | ||||||
9118 | .getAlignment() | ||||||
9119 | .alignmentOfArrayElement(ElementSize)); | ||||||
9120 | // Privatize the declared variable of mapper to be the current array element. | ||||||
9121 | CodeGenFunction::OMPPrivateScope Scope(MapperCGF); | ||||||
9122 | Scope.addPrivate(MapperVarDecl, [&MapperCGF, PtrCurrent, PtrTy]() { | ||||||
9123 | return MapperCGF | ||||||
9124 | .EmitLoadOfPointerLValue(PtrCurrent, PtrTy->castAs<PointerType>()) | ||||||
9125 | .getAddress(MapperCGF); | ||||||
9126 | }); | ||||||
9127 | (void)Scope.Privatize(); | ||||||
9128 | |||||||
9129 | // Get map clause information. Fill up the arrays with all mapped variables. | ||||||
9130 | MappableExprsHandler::MapCombinedInfoTy Info; | ||||||
9131 | MappableExprsHandler MEHandler(*D, MapperCGF); | ||||||
9132 | MEHandler.generateAllInfoForMapper(Info); | ||||||
9133 | |||||||
9134 | // Call the runtime API __tgt_mapper_num_components to get the number of | ||||||
9135 | // pre-existing components. | ||||||
9136 | llvm::Value *OffloadingArgs[] = {Handle}; | ||||||
9137 | llvm::Value *PreviousSize = MapperCGF.EmitRuntimeCall( | ||||||
9138 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
9139 | OMPRTL___tgt_mapper_num_components), | ||||||
9140 | OffloadingArgs); | ||||||
9141 | llvm::Value *ShiftedPreviousSize = MapperCGF.Builder.CreateShl( | ||||||
9142 | PreviousSize, | ||||||
9143 | MapperCGF.Builder.getInt64(MappableExprsHandler::getFlagMemberOffset())); | ||||||
9144 | |||||||
9145 | // Fill up the runtime mapper handle for all components. | ||||||
9146 | for (unsigned I = 0; I < Info.BasePointers.size(); ++I) { | ||||||
9147 | llvm::Value *CurBaseArg = MapperCGF.Builder.CreateBitCast( | ||||||
9148 | *Info.BasePointers[I], CGM.getTypes().ConvertTypeForMem(C.VoidPtrTy)); | ||||||
9149 | llvm::Value *CurBeginArg = MapperCGF.Builder.CreateBitCast( | ||||||
9150 | Info.Pointers[I], CGM.getTypes().ConvertTypeForMem(C.VoidPtrTy)); | ||||||
9151 | llvm::Value *CurSizeArg = Info.Sizes[I]; | ||||||
9152 | |||||||
9153 | // Extract the MEMBER_OF field from the map type. | ||||||
9154 | llvm::BasicBlock *MemberBB = MapperCGF.createBasicBlock("omp.member"); | ||||||
9155 | MapperCGF.EmitBlock(MemberBB); | ||||||
9156 | llvm::Value *OriMapType = MapperCGF.Builder.getInt64(Info.Types[I]); | ||||||
9157 | llvm::Value *Member = MapperCGF.Builder.CreateAnd( | ||||||
9158 | OriMapType, | ||||||
9159 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_MEMBER_OF)); | ||||||
9160 | llvm::BasicBlock *MemberCombineBB = | ||||||
9161 | MapperCGF.createBasicBlock("omp.member.combine"); | ||||||
9162 | llvm::BasicBlock *TypeBB = MapperCGF.createBasicBlock("omp.type"); | ||||||
9163 | llvm::Value *IsMember = MapperCGF.Builder.CreateIsNull(Member); | ||||||
9164 | MapperCGF.Builder.CreateCondBr(IsMember, TypeBB, MemberCombineBB); | ||||||
9165 | // Add the number of pre-existing components to the MEMBER_OF field if it | ||||||
9166 | // is valid. | ||||||
9167 | MapperCGF.EmitBlock(MemberCombineBB); | ||||||
9168 | llvm::Value *CombinedMember = | ||||||
9169 | MapperCGF.Builder.CreateNUWAdd(OriMapType, ShiftedPreviousSize); | ||||||
9170 | // Do nothing if it is not a member of previous components. | ||||||
9171 | MapperCGF.EmitBlock(TypeBB); | ||||||
9172 | llvm::PHINode *MemberMapType = | ||||||
9173 | MapperCGF.Builder.CreatePHI(CGM.Int64Ty, 4, "omp.membermaptype"); | ||||||
9174 | MemberMapType->addIncoming(OriMapType, MemberBB); | ||||||
9175 | MemberMapType->addIncoming(CombinedMember, MemberCombineBB); | ||||||
9176 | |||||||
9177 | // Combine the map type inherited from user-defined mapper with that | ||||||
9178 | // specified in the program. According to the OMP_MAP_TO and OMP_MAP_FROM | ||||||
9179 | // bits of the \a MapType, which is the input argument of the mapper | ||||||
9180 | // function, the following code will set the OMP_MAP_TO and OMP_MAP_FROM | ||||||
9181 | // bits of MemberMapType. | ||||||
9182 | // [OpenMP 5.0], 1.2.6. map-type decay. | ||||||
9183 | // | alloc | to | from | tofrom | release | delete | ||||||
9184 | // ---------------------------------------------------------- | ||||||
9185 | // alloc | alloc | alloc | alloc | alloc | release | delete | ||||||
9186 | // to | alloc | to | alloc | to | release | delete | ||||||
9187 | // from | alloc | alloc | from | from | release | delete | ||||||
9188 | // tofrom | alloc | to | from | tofrom | release | delete | ||||||
9189 | llvm::Value *LeftToFrom = MapperCGF.Builder.CreateAnd( | ||||||
9190 | MapType, | ||||||
9191 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_TO | | ||||||
9192 | MappableExprsHandler::OMP_MAP_FROM)); | ||||||
9193 | llvm::BasicBlock *AllocBB = MapperCGF.createBasicBlock("omp.type.alloc"); | ||||||
9194 | llvm::BasicBlock *AllocElseBB = | ||||||
9195 | MapperCGF.createBasicBlock("omp.type.alloc.else"); | ||||||
9196 | llvm::BasicBlock *ToBB = MapperCGF.createBasicBlock("omp.type.to"); | ||||||
9197 | llvm::BasicBlock *ToElseBB = MapperCGF.createBasicBlock("omp.type.to.else"); | ||||||
9198 | llvm::BasicBlock *FromBB = MapperCGF.createBasicBlock("omp.type.from"); | ||||||
9199 | llvm::BasicBlock *EndBB = MapperCGF.createBasicBlock("omp.type.end"); | ||||||
9200 | llvm::Value *IsAlloc = MapperCGF.Builder.CreateIsNull(LeftToFrom); | ||||||
9201 | MapperCGF.Builder.CreateCondBr(IsAlloc, AllocBB, AllocElseBB); | ||||||
9202 | // In case of alloc, clear OMP_MAP_TO and OMP_MAP_FROM. | ||||||
9203 | MapperCGF.EmitBlock(AllocBB); | ||||||
9204 | llvm::Value *AllocMapType = MapperCGF.Builder.CreateAnd( | ||||||
9205 | MemberMapType, | ||||||
9206 | MapperCGF.Builder.getInt64(~(MappableExprsHandler::OMP_MAP_TO | | ||||||
9207 | MappableExprsHandler::OMP_MAP_FROM))); | ||||||
9208 | MapperCGF.Builder.CreateBr(EndBB); | ||||||
9209 | MapperCGF.EmitBlock(AllocElseBB); | ||||||
9210 | llvm::Value *IsTo = MapperCGF.Builder.CreateICmpEQ( | ||||||
9211 | LeftToFrom, | ||||||
9212 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_TO)); | ||||||
9213 | MapperCGF.Builder.CreateCondBr(IsTo, ToBB, ToElseBB); | ||||||
9214 | // In case of to, clear OMP_MAP_FROM. | ||||||
9215 | MapperCGF.EmitBlock(ToBB); | ||||||
9216 | llvm::Value *ToMapType = MapperCGF.Builder.CreateAnd( | ||||||
9217 | MemberMapType, | ||||||
9218 | MapperCGF.Builder.getInt64(~MappableExprsHandler::OMP_MAP_FROM)); | ||||||
9219 | MapperCGF.Builder.CreateBr(EndBB); | ||||||
9220 | MapperCGF.EmitBlock(ToElseBB); | ||||||
9221 | llvm::Value *IsFrom = MapperCGF.Builder.CreateICmpEQ( | ||||||
9222 | LeftToFrom, | ||||||
9223 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_FROM)); | ||||||
9224 | MapperCGF.Builder.CreateCondBr(IsFrom, FromBB, EndBB); | ||||||
9225 | // In case of from, clear OMP_MAP_TO. | ||||||
9226 | MapperCGF.EmitBlock(FromBB); | ||||||
9227 | llvm::Value *FromMapType = MapperCGF.Builder.CreateAnd( | ||||||
9228 | MemberMapType, | ||||||
9229 | MapperCGF.Builder.getInt64(~MappableExprsHandler::OMP_MAP_TO)); | ||||||
9230 | // In case of tofrom, do nothing. | ||||||
9231 | MapperCGF.EmitBlock(EndBB); | ||||||
9232 | LastBB = EndBB; | ||||||
9233 | llvm::PHINode *CurMapType = | ||||||
9234 | MapperCGF.Builder.CreatePHI(CGM.Int64Ty, 4, "omp.maptype"); | ||||||
9235 | CurMapType->addIncoming(AllocMapType, AllocBB); | ||||||
9236 | CurMapType->addIncoming(ToMapType, ToBB); | ||||||
9237 | CurMapType->addIncoming(FromMapType, FromBB); | ||||||
9238 | CurMapType->addIncoming(MemberMapType, ToElseBB); | ||||||
9239 | |||||||
9240 | llvm::Value *OffloadingArgs[] = {Handle, CurBaseArg, CurBeginArg, | ||||||
9241 | CurSizeArg, CurMapType}; | ||||||
9242 | if (Info.Mappers[I]) { | ||||||
9243 | // Call the corresponding mapper function. | ||||||
9244 | llvm::Function *MapperFunc = getOrCreateUserDefinedMapperFunc( | ||||||
9245 | cast<OMPDeclareMapperDecl>(Info.Mappers[I])); | ||||||
9246 | assert(MapperFunc && "Expect a valid mapper function is available.")((MapperFunc && "Expect a valid mapper function is available." ) ? static_cast<void> (0) : __assert_fail ("MapperFunc && \"Expect a valid mapper function is available.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9246, __PRETTY_FUNCTION__)); | ||||||
9247 | MapperCGF.EmitNounwindRuntimeCall(MapperFunc, OffloadingArgs); | ||||||
9248 | } else { | ||||||
9249 | // Call the runtime API __tgt_push_mapper_component to fill up the runtime | ||||||
9250 | // data structure. | ||||||
9251 | MapperCGF.EmitRuntimeCall( | ||||||
9252 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
9253 | CGM.getModule(), OMPRTL___tgt_push_mapper_component), | ||||||
9254 | OffloadingArgs); | ||||||
9255 | } | ||||||
9256 | } | ||||||
9257 | |||||||
9258 | // Update the pointer to point to the next element that needs to be mapped, | ||||||
9259 | // and check whether we have mapped all elements. | ||||||
9260 | llvm::Value *PtrNext = MapperCGF.Builder.CreateConstGEP1_32( | ||||||
9261 | PtrPHI, /*Idx0=*/1, "omp.arraymap.next"); | ||||||
9262 | PtrPHI->addIncoming(PtrNext, LastBB); | ||||||
9263 | llvm::Value *IsDone = | ||||||
9264 | MapperCGF.Builder.CreateICmpEQ(PtrNext, PtrEnd, "omp.arraymap.isdone"); | ||||||
9265 | llvm::BasicBlock *ExitBB = MapperCGF.createBasicBlock("omp.arraymap.exit"); | ||||||
9266 | MapperCGF.Builder.CreateCondBr(IsDone, ExitBB, BodyBB); | ||||||
9267 | |||||||
9268 | MapperCGF.EmitBlock(ExitBB); | ||||||
9269 | // Emit array deletion if this is an array section and \p MapType indicates | ||||||
9270 | // that deletion is required. | ||||||
9271 | emitUDMapperArrayInitOrDel(MapperCGF, Handle, BaseIn, BeginIn, Size, MapType, | ||||||
9272 | ElementSize, DoneBB, /*IsInit=*/false); | ||||||
9273 | |||||||
9274 | // Emit the function exit block. | ||||||
9275 | MapperCGF.EmitBlock(DoneBB, /*IsFinished=*/true); | ||||||
9276 | MapperCGF.FinishFunction(); | ||||||
9277 | UDMMap.try_emplace(D, Fn); | ||||||
9278 | if (CGF) { | ||||||
9279 | auto &Decls = FunctionUDMMap.FindAndConstruct(CGF->CurFn); | ||||||
9280 | Decls.second.push_back(D); | ||||||
9281 | } | ||||||
9282 | } | ||||||
9283 | |||||||
9284 | /// Emit the array initialization or deletion portion for user-defined mapper | ||||||
9285 | /// code generation. First, it evaluates whether an array section is mapped and | ||||||
9286 | /// whether the \a MapType instructs to delete this section. If \a IsInit is | ||||||
9287 | /// true, and \a MapType indicates to not delete this array, array | ||||||
9288 | /// initialization code is generated. If \a IsInit is false, and \a MapType | ||||||
9289 | /// indicates to not this array, array deletion code is generated. | ||||||
9290 | void CGOpenMPRuntime::emitUDMapperArrayInitOrDel( | ||||||
9291 | CodeGenFunction &MapperCGF, llvm::Value *Handle, llvm::Value *Base, | ||||||
9292 | llvm::Value *Begin, llvm::Value *Size, llvm::Value *MapType, | ||||||
9293 | CharUnits ElementSize, llvm::BasicBlock *ExitBB, bool IsInit) { | ||||||
9294 | StringRef Prefix = IsInit ? ".init" : ".del"; | ||||||
9295 | |||||||
9296 | // Evaluate if this is an array section. | ||||||
9297 | llvm::BasicBlock *IsDeleteBB = | ||||||
9298 | MapperCGF.createBasicBlock(getName({"omp.array", Prefix, ".evaldelete"})); | ||||||
9299 | llvm::BasicBlock *BodyBB = | ||||||
9300 | MapperCGF.createBasicBlock(getName({"omp.array", Prefix})); | ||||||
9301 | llvm::Value *IsArray = MapperCGF.Builder.CreateICmpSGE( | ||||||
9302 | Size, MapperCGF.Builder.getInt64(1), "omp.arrayinit.isarray"); | ||||||
9303 | MapperCGF.Builder.CreateCondBr(IsArray, IsDeleteBB, ExitBB); | ||||||
9304 | |||||||
9305 | // Evaluate if we are going to delete this section. | ||||||
9306 | MapperCGF.EmitBlock(IsDeleteBB); | ||||||
9307 | llvm::Value *DeleteBit = MapperCGF.Builder.CreateAnd( | ||||||
9308 | MapType, | ||||||
9309 | MapperCGF.Builder.getInt64(MappableExprsHandler::OMP_MAP_DELETE)); | ||||||
9310 | llvm::Value *DeleteCond; | ||||||
9311 | if (IsInit) { | ||||||
9312 | DeleteCond = MapperCGF.Builder.CreateIsNull( | ||||||
9313 | DeleteBit, getName({"omp.array", Prefix, ".delete"})); | ||||||
9314 | } else { | ||||||
9315 | DeleteCond = MapperCGF.Builder.CreateIsNotNull( | ||||||
9316 | DeleteBit, getName({"omp.array", Prefix, ".delete"})); | ||||||
9317 | } | ||||||
9318 | MapperCGF.Builder.CreateCondBr(DeleteCond, BodyBB, ExitBB); | ||||||
9319 | |||||||
9320 | MapperCGF.EmitBlock(BodyBB); | ||||||
9321 | // Get the array size by multiplying element size and element number (i.e., \p | ||||||
9322 | // Size). | ||||||
9323 | llvm::Value *ArraySize = MapperCGF.Builder.CreateNUWMul( | ||||||
9324 | Size, MapperCGF.Builder.getInt64(ElementSize.getQuantity())); | ||||||
9325 | // Remove OMP_MAP_TO and OMP_MAP_FROM from the map type, so that it achieves | ||||||
9326 | // memory allocation/deletion purpose only. | ||||||
9327 | llvm::Value *MapTypeArg = MapperCGF.Builder.CreateAnd( | ||||||
9328 | MapType, | ||||||
9329 | MapperCGF.Builder.getInt64(~(MappableExprsHandler::OMP_MAP_TO | | ||||||
9330 | MappableExprsHandler::OMP_MAP_FROM))); | ||||||
9331 | // Call the runtime API __tgt_push_mapper_component to fill up the runtime | ||||||
9332 | // data structure. | ||||||
9333 | llvm::Value *OffloadingArgs[] = {Handle, Base, Begin, ArraySize, MapTypeArg}; | ||||||
9334 | MapperCGF.EmitRuntimeCall( | ||||||
9335 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
9336 | OMPRTL___tgt_push_mapper_component), | ||||||
9337 | OffloadingArgs); | ||||||
9338 | } | ||||||
9339 | |||||||
9340 | llvm::Function *CGOpenMPRuntime::getOrCreateUserDefinedMapperFunc( | ||||||
9341 | const OMPDeclareMapperDecl *D) { | ||||||
9342 | auto I = UDMMap.find(D); | ||||||
9343 | if (I != UDMMap.end()) | ||||||
9344 | return I->second; | ||||||
9345 | emitUserDefinedMapper(D); | ||||||
9346 | return UDMMap.lookup(D); | ||||||
9347 | } | ||||||
9348 | |||||||
9349 | void CGOpenMPRuntime::emitTargetNumIterationsCall( | ||||||
9350 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | ||||||
9351 | llvm::Value *DeviceID, | ||||||
9352 | llvm::function_ref<llvm::Value *(CodeGenFunction &CGF, | ||||||
9353 | const OMPLoopDirective &D)> | ||||||
9354 | SizeEmitter) { | ||||||
9355 | OpenMPDirectiveKind Kind = D.getDirectiveKind(); | ||||||
9356 | const OMPExecutableDirective *TD = &D; | ||||||
9357 | // Get nested teams distribute kind directive, if any. | ||||||
9358 | if (!isOpenMPDistributeDirective(Kind) || !isOpenMPTeamsDirective(Kind)) | ||||||
9359 | TD = getNestedDistributeDirective(CGM.getContext(), D); | ||||||
9360 | if (!TD) | ||||||
9361 | return; | ||||||
9362 | const auto *LD = cast<OMPLoopDirective>(TD); | ||||||
9363 | auto &&CodeGen = [LD, DeviceID, SizeEmitter, this](CodeGenFunction &CGF, | ||||||
9364 | PrePostActionTy &) { | ||||||
9365 | if (llvm::Value *NumIterations = SizeEmitter(CGF, *LD)) { | ||||||
9366 | llvm::Value *Args[] = {DeviceID, NumIterations}; | ||||||
9367 | CGF.EmitRuntimeCall( | ||||||
9368 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
9369 | CGM.getModule(), OMPRTL___kmpc_push_target_tripcount), | ||||||
9370 | Args); | ||||||
9371 | } | ||||||
9372 | }; | ||||||
9373 | emitInlinedDirective(CGF, OMPD_unknown, CodeGen); | ||||||
9374 | } | ||||||
9375 | |||||||
9376 | void CGOpenMPRuntime::emitTargetCall( | ||||||
9377 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | ||||||
9378 | llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, const Expr *IfCond, | ||||||
9379 | llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device, | ||||||
9380 | llvm::function_ref<llvm::Value *(CodeGenFunction &CGF, | ||||||
9381 | const OMPLoopDirective &D)> | ||||||
9382 | SizeEmitter) { | ||||||
9383 | if (!CGF.HaveInsertPoint()) | ||||||
9384 | return; | ||||||
9385 | |||||||
9386 | assert(OutlinedFn && "Invalid outlined function!")((OutlinedFn && "Invalid outlined function!") ? static_cast <void> (0) : __assert_fail ("OutlinedFn && \"Invalid outlined function!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9386, __PRETTY_FUNCTION__)); | ||||||
9387 | |||||||
9388 | const bool RequiresOuterTask = D.hasClausesOfKind<OMPDependClause>(); | ||||||
9389 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; | ||||||
9390 | const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target); | ||||||
9391 | auto &&ArgsCodegen = [&CS, &CapturedVars](CodeGenFunction &CGF, | ||||||
9392 | PrePostActionTy &) { | ||||||
9393 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | ||||||
9394 | }; | ||||||
9395 | emitInlinedDirective(CGF, OMPD_unknown, ArgsCodegen); | ||||||
9396 | |||||||
9397 | CodeGenFunction::OMPTargetDataInfo InputInfo; | ||||||
9398 | llvm::Value *MapTypesArray = nullptr; | ||||||
9399 | // Fill up the pointer arrays and transfer execution to the device. | ||||||
9400 | auto &&ThenGen = [this, Device, OutlinedFn, OutlinedFnID, &D, &InputInfo, | ||||||
9401 | &MapTypesArray, &CS, RequiresOuterTask, &CapturedVars, | ||||||
9402 | SizeEmitter](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
9403 | if (Device.getInt() == OMPC_DEVICE_ancestor) { | ||||||
9404 | // Reverse offloading is not supported, so just execute on the host. | ||||||
9405 | if (RequiresOuterTask) { | ||||||
9406 | CapturedVars.clear(); | ||||||
9407 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | ||||||
9408 | } | ||||||
9409 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars); | ||||||
9410 | return; | ||||||
9411 | } | ||||||
9412 | |||||||
9413 | // On top of the arrays that were filled up, the target offloading call | ||||||
9414 | // takes as arguments the device id as well as the host pointer. The host | ||||||
9415 | // pointer is used by the runtime library to identify the current target | ||||||
9416 | // region, so it only has to be unique and not necessarily point to | ||||||
9417 | // anything. It could be the pointer to the outlined function that | ||||||
9418 | // implements the target region, but we aren't using that so that the | ||||||
9419 | // compiler doesn't need to keep that, and could therefore inline the host | ||||||
9420 | // function if proven worthwhile during optimization. | ||||||
9421 | |||||||
9422 | // From this point on, we need to have an ID of the target region defined. | ||||||
9423 | assert(OutlinedFnID && "Invalid outlined function ID!")((OutlinedFnID && "Invalid outlined function ID!") ? static_cast <void> (0) : __assert_fail ("OutlinedFnID && \"Invalid outlined function ID!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9423, __PRETTY_FUNCTION__)); | ||||||
9424 | |||||||
9425 | // Emit device ID if any. | ||||||
9426 | llvm::Value *DeviceID; | ||||||
9427 | if (Device.getPointer()) { | ||||||
9428 | assert((Device.getInt() == OMPC_DEVICE_unknown ||(((Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && "Expected device_num modifier." ) ? static_cast<void> (0) : __assert_fail ("(Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && \"Expected device_num modifier.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9430, __PRETTY_FUNCTION__)) | ||||||
9429 | Device.getInt() == OMPC_DEVICE_device_num) &&(((Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && "Expected device_num modifier." ) ? static_cast<void> (0) : __assert_fail ("(Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && \"Expected device_num modifier.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9430, __PRETTY_FUNCTION__)) | ||||||
9430 | "Expected device_num modifier.")(((Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && "Expected device_num modifier." ) ? static_cast<void> (0) : __assert_fail ("(Device.getInt() == OMPC_DEVICE_unknown || Device.getInt() == OMPC_DEVICE_device_num) && \"Expected device_num modifier.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9430, __PRETTY_FUNCTION__)); | ||||||
9431 | llvm::Value *DevVal = CGF.EmitScalarExpr(Device.getPointer()); | ||||||
9432 | DeviceID = | ||||||
9433 | CGF.Builder.CreateIntCast(DevVal, CGF.Int64Ty, /*isSigned=*/true); | ||||||
9434 | } else { | ||||||
9435 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | ||||||
9436 | } | ||||||
9437 | |||||||
9438 | // Emit the number of elements in the offloading arrays. | ||||||
9439 | llvm::Value *PointerNum = | ||||||
9440 | CGF.Builder.getInt32(InputInfo.NumberOfTargetItems); | ||||||
9441 | |||||||
9442 | // Return value of the runtime offloading call. | ||||||
9443 | llvm::Value *Return; | ||||||
9444 | |||||||
9445 | llvm::Value *NumTeams = emitNumTeamsForTargetDirective(CGF, D); | ||||||
9446 | llvm::Value *NumThreads = emitNumThreadsForTargetDirective(CGF, D); | ||||||
9447 | |||||||
9448 | // Emit tripcount for the target loop-based directive. | ||||||
9449 | emitTargetNumIterationsCall(CGF, D, DeviceID, SizeEmitter); | ||||||
9450 | |||||||
9451 | bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>(); | ||||||
9452 | // The target region is an outlined function launched by the runtime | ||||||
9453 | // via calls __tgt_target() or __tgt_target_teams(). | ||||||
9454 | // | ||||||
9455 | // __tgt_target() launches a target region with one team and one thread, | ||||||
9456 | // executing a serial region. This master thread may in turn launch | ||||||
9457 | // more threads within its team upon encountering a parallel region, | ||||||
9458 | // however, no additional teams can be launched on the device. | ||||||
9459 | // | ||||||
9460 | // __tgt_target_teams() launches a target region with one or more teams, | ||||||
9461 | // each with one or more threads. This call is required for target | ||||||
9462 | // constructs such as: | ||||||
9463 | // 'target teams' | ||||||
9464 | // 'target' / 'teams' | ||||||
9465 | // 'target teams distribute parallel for' | ||||||
9466 | // 'target parallel' | ||||||
9467 | // and so on. | ||||||
9468 | // | ||||||
9469 | // Note that on the host and CPU targets, the runtime implementation of | ||||||
9470 | // these calls simply call the outlined function without forking threads. | ||||||
9471 | // The outlined functions themselves have runtime calls to | ||||||
9472 | // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by | ||||||
9473 | // the compiler in emitTeamsCall() and emitParallelCall(). | ||||||
9474 | // | ||||||
9475 | // In contrast, on the NVPTX target, the implementation of | ||||||
9476 | // __tgt_target_teams() launches a GPU kernel with the requested number | ||||||
9477 | // of teams and threads so no additional calls to the runtime are required. | ||||||
9478 | if (NumTeams) { | ||||||
9479 | // If we have NumTeams defined this means that we have an enclosed teams | ||||||
9480 | // region. Therefore we also expect to have NumThreads defined. These two | ||||||
9481 | // values should be defined in the presence of a teams directive, | ||||||
9482 | // regardless of having any clauses associated. If the user is using teams | ||||||
9483 | // but no clauses, these two values will be the default that should be | ||||||
9484 | // passed to the runtime library - a 32-bit integer with the value zero. | ||||||
9485 | assert(NumThreads && "Thread limit expression should be available along "((NumThreads && "Thread limit expression should be available along " "with number of teams.") ? static_cast<void> (0) : __assert_fail ("NumThreads && \"Thread limit expression should be available along \" \"with number of teams.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9486, __PRETTY_FUNCTION__)) | ||||||
9486 | "with number of teams.")((NumThreads && "Thread limit expression should be available along " "with number of teams.") ? static_cast<void> (0) : __assert_fail ("NumThreads && \"Thread limit expression should be available along \" \"with number of teams.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9486, __PRETTY_FUNCTION__)); | ||||||
9487 | llvm::Value *OffloadingArgs[] = {DeviceID, | ||||||
9488 | OutlinedFnID, | ||||||
9489 | PointerNum, | ||||||
9490 | InputInfo.BasePointersArray.getPointer(), | ||||||
9491 | InputInfo.PointersArray.getPointer(), | ||||||
9492 | InputInfo.SizesArray.getPointer(), | ||||||
9493 | MapTypesArray, | ||||||
9494 | InputInfo.MappersArray.getPointer(), | ||||||
9495 | NumTeams, | ||||||
9496 | NumThreads}; | ||||||
9497 | Return = CGF.EmitRuntimeCall( | ||||||
9498 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
9499 | CGM.getModule(), HasNowait | ||||||
9500 | ? OMPRTL___tgt_target_teams_nowait_mapper | ||||||
9501 | : OMPRTL___tgt_target_teams_mapper), | ||||||
9502 | OffloadingArgs); | ||||||
9503 | } else { | ||||||
9504 | llvm::Value *OffloadingArgs[] = {DeviceID, | ||||||
9505 | OutlinedFnID, | ||||||
9506 | PointerNum, | ||||||
9507 | InputInfo.BasePointersArray.getPointer(), | ||||||
9508 | InputInfo.PointersArray.getPointer(), | ||||||
9509 | InputInfo.SizesArray.getPointer(), | ||||||
9510 | MapTypesArray, | ||||||
9511 | InputInfo.MappersArray.getPointer()}; | ||||||
9512 | Return = CGF.EmitRuntimeCall( | ||||||
9513 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
9514 | CGM.getModule(), HasNowait ? OMPRTL___tgt_target_nowait_mapper | ||||||
9515 | : OMPRTL___tgt_target_mapper), | ||||||
9516 | OffloadingArgs); | ||||||
9517 | } | ||||||
9518 | |||||||
9519 | // Check the error code and execute the host version if required. | ||||||
9520 | llvm::BasicBlock *OffloadFailedBlock = | ||||||
9521 | CGF.createBasicBlock("omp_offload.failed"); | ||||||
9522 | llvm::BasicBlock *OffloadContBlock = | ||||||
9523 | CGF.createBasicBlock("omp_offload.cont"); | ||||||
9524 | llvm::Value *Failed = CGF.Builder.CreateIsNotNull(Return); | ||||||
9525 | CGF.Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock); | ||||||
9526 | |||||||
9527 | CGF.EmitBlock(OffloadFailedBlock); | ||||||
9528 | if (RequiresOuterTask) { | ||||||
9529 | CapturedVars.clear(); | ||||||
9530 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | ||||||
9531 | } | ||||||
9532 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars); | ||||||
9533 | CGF.EmitBranch(OffloadContBlock); | ||||||
9534 | |||||||
9535 | CGF.EmitBlock(OffloadContBlock, /*IsFinished=*/true); | ||||||
9536 | }; | ||||||
9537 | |||||||
9538 | // Notify that the host version must be executed. | ||||||
9539 | auto &&ElseGen = [this, &D, OutlinedFn, &CS, &CapturedVars, | ||||||
9540 | RequiresOuterTask](CodeGenFunction &CGF, | ||||||
9541 | PrePostActionTy &) { | ||||||
9542 | if (RequiresOuterTask) { | ||||||
9543 | CapturedVars.clear(); | ||||||
9544 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | ||||||
9545 | } | ||||||
9546 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars); | ||||||
9547 | }; | ||||||
9548 | |||||||
9549 | auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray, | ||||||
9550 | &CapturedVars, RequiresOuterTask, | ||||||
9551 | &CS](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
9552 | // Fill up the arrays with all the captured variables. | ||||||
9553 | MappableExprsHandler::MapCombinedInfoTy CombinedInfo; | ||||||
9554 | |||||||
9555 | // Get mappable expression information. | ||||||
9556 | MappableExprsHandler MEHandler(D, CGF); | ||||||
9557 | llvm::DenseMap<llvm::Value *, llvm::Value *> LambdaPointers; | ||||||
9558 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> MappedVarSet; | ||||||
9559 | |||||||
9560 | auto RI = CS.getCapturedRecordDecl()->field_begin(); | ||||||
9561 | auto CV = CapturedVars.begin(); | ||||||
9562 | for (CapturedStmt::const_capture_iterator CI = CS.capture_begin(), | ||||||
9563 | CE = CS.capture_end(); | ||||||
9564 | CI != CE; ++CI, ++RI, ++CV) { | ||||||
9565 | MappableExprsHandler::MapCombinedInfoTy CurInfo; | ||||||
9566 | MappableExprsHandler::StructRangeInfoTy PartialStruct; | ||||||
9567 | |||||||
9568 | // VLA sizes are passed to the outlined region by copy and do not have map | ||||||
9569 | // information associated. | ||||||
9570 | if (CI->capturesVariableArrayType()) { | ||||||
9571 | CurInfo.BasePointers.push_back(*CV); | ||||||
9572 | CurInfo.Pointers.push_back(*CV); | ||||||
9573 | CurInfo.Sizes.push_back(CGF.Builder.CreateIntCast( | ||||||
9574 | CGF.getTypeSize(RI->getType()), CGF.Int64Ty, /*isSigned=*/true)); | ||||||
9575 | // Copy to the device as an argument. No need to retrieve it. | ||||||
9576 | CurInfo.Types.push_back(MappableExprsHandler::OMP_MAP_LITERAL | | ||||||
9577 | MappableExprsHandler::OMP_MAP_TARGET_PARAM | | ||||||
9578 | MappableExprsHandler::OMP_MAP_IMPLICIT); | ||||||
9579 | CurInfo.Mappers.push_back(nullptr); | ||||||
9580 | } else { | ||||||
9581 | // If we have any information in the map clause, we use it, otherwise we | ||||||
9582 | // just do a default mapping. | ||||||
9583 | MEHandler.generateInfoForCapture(CI, *CV, CurInfo, PartialStruct); | ||||||
9584 | if (!CI->capturesThis()) | ||||||
9585 | MappedVarSet.insert(CI->getCapturedVar()); | ||||||
9586 | else | ||||||
9587 | MappedVarSet.insert(nullptr); | ||||||
9588 | if (CurInfo.BasePointers.empty()) | ||||||
9589 | MEHandler.generateDefaultMapInfo(*CI, **RI, *CV, CurInfo); | ||||||
9590 | // Generate correct mapping for variables captured by reference in | ||||||
9591 | // lambdas. | ||||||
9592 | if (CI->capturesVariable()) | ||||||
9593 | MEHandler.generateInfoForLambdaCaptures(CI->getCapturedVar(), *CV, | ||||||
9594 | CurInfo, LambdaPointers); | ||||||
9595 | } | ||||||
9596 | // We expect to have at least an element of information for this capture. | ||||||
9597 | assert(!CurInfo.BasePointers.empty() &&((!CurInfo.BasePointers.empty() && "Non-existing map pointer for capture!" ) ? static_cast<void> (0) : __assert_fail ("!CurInfo.BasePointers.empty() && \"Non-existing map pointer for capture!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9598, __PRETTY_FUNCTION__)) | ||||||
9598 | "Non-existing map pointer for capture!")((!CurInfo.BasePointers.empty() && "Non-existing map pointer for capture!" ) ? static_cast<void> (0) : __assert_fail ("!CurInfo.BasePointers.empty() && \"Non-existing map pointer for capture!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9598, __PRETTY_FUNCTION__)); | ||||||
9599 | assert(CurInfo.BasePointers.size() == CurInfo.Pointers.size() &&((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!") ? static_cast<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!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9603, __PRETTY_FUNCTION__)) | ||||||
9600 | CurInfo.BasePointers.size() == CurInfo.Sizes.size() &&((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!") ? static_cast<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!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9603, __PRETTY_FUNCTION__)) | ||||||
9601 | CurInfo.BasePointers.size() == CurInfo.Types.size() &&((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!") ? static_cast<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!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9603, __PRETTY_FUNCTION__)) | ||||||
9602 | CurInfo.BasePointers.size() == CurInfo.Mappers.size() &&((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!") ? static_cast<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!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9603, __PRETTY_FUNCTION__)) | ||||||
9603 | "Inconsistent map information sizes!")((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!") ? static_cast<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!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9603, __PRETTY_FUNCTION__)); | ||||||
9604 | |||||||
9605 | // If there is an entry in PartialStruct it means we have a struct with | ||||||
9606 | // individual members mapped. Emit an extra combined entry. | ||||||
9607 | if (PartialStruct.Base.isValid()) | ||||||
9608 | MEHandler.emitCombinedEntry(CombinedInfo, CurInfo.Types, PartialStruct); | ||||||
9609 | |||||||
9610 | // We need to append the results of this capture to what we already have. | ||||||
9611 | CombinedInfo.append(CurInfo); | ||||||
9612 | } | ||||||
9613 | // Adjust MEMBER_OF flags for the lambdas captures. | ||||||
9614 | MEHandler.adjustMemberOfForLambdaCaptures( | ||||||
9615 | LambdaPointers, CombinedInfo.BasePointers, CombinedInfo.Pointers, | ||||||
9616 | CombinedInfo.Types); | ||||||
9617 | // Map any list items in a map clause that were not captures because they | ||||||
9618 | // weren't referenced within the construct. | ||||||
9619 | MEHandler.generateAllInfo(CombinedInfo, /*NotTargetParams=*/true, | ||||||
9620 | MappedVarSet); | ||||||
9621 | |||||||
9622 | TargetDataInfo Info; | ||||||
9623 | // Fill up the arrays and create the arguments. | ||||||
9624 | emitOffloadingArrays(CGF, CombinedInfo, Info); | ||||||
9625 | emitOffloadingArraysArgument(CGF, Info.BasePointersArray, | ||||||
9626 | Info.PointersArray, Info.SizesArray, | ||||||
9627 | Info.MapTypesArray, Info.MappersArray, Info); | ||||||
9628 | InputInfo.NumberOfTargetItems = Info.NumberOfPtrs; | ||||||
9629 | InputInfo.BasePointersArray = | ||||||
9630 | Address(Info.BasePointersArray, CGM.getPointerAlign()); | ||||||
9631 | InputInfo.PointersArray = | ||||||
9632 | Address(Info.PointersArray, CGM.getPointerAlign()); | ||||||
9633 | InputInfo.SizesArray = Address(Info.SizesArray, CGM.getPointerAlign()); | ||||||
9634 | InputInfo.MappersArray = Address(Info.MappersArray, CGM.getPointerAlign()); | ||||||
9635 | MapTypesArray = Info.MapTypesArray; | ||||||
9636 | if (RequiresOuterTask) | ||||||
9637 | CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo); | ||||||
9638 | else | ||||||
9639 | emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen); | ||||||
9640 | }; | ||||||
9641 | |||||||
9642 | auto &&TargetElseGen = [this, &ElseGen, &D, RequiresOuterTask]( | ||||||
9643 | CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
9644 | if (RequiresOuterTask) { | ||||||
9645 | CodeGenFunction::OMPTargetDataInfo InputInfo; | ||||||
9646 | CGF.EmitOMPTargetTaskBasedDirective(D, ElseGen, InputInfo); | ||||||
9647 | } else { | ||||||
9648 | emitInlinedDirective(CGF, D.getDirectiveKind(), ElseGen); | ||||||
9649 | } | ||||||
9650 | }; | ||||||
9651 | |||||||
9652 | // If we have a target function ID it means that we need to support | ||||||
9653 | // offloading, otherwise, just execute on the host. We need to execute on host | ||||||
9654 | // regardless of the conditional in the if clause if, e.g., the user do not | ||||||
9655 | // specify target triples. | ||||||
9656 | if (OutlinedFnID) { | ||||||
9657 | if (IfCond) { | ||||||
9658 | emitIfClause(CGF, IfCond, TargetThenGen, TargetElseGen); | ||||||
9659 | } else { | ||||||
9660 | RegionCodeGenTy ThenRCG(TargetThenGen); | ||||||
9661 | ThenRCG(CGF); | ||||||
9662 | } | ||||||
9663 | } else { | ||||||
9664 | RegionCodeGenTy ElseRCG(TargetElseGen); | ||||||
9665 | ElseRCG(CGF); | ||||||
9666 | } | ||||||
9667 | } | ||||||
9668 | |||||||
9669 | void CGOpenMPRuntime::scanForTargetRegionsFunctions(const Stmt *S, | ||||||
9670 | StringRef ParentName) { | ||||||
9671 | if (!S) | ||||||
9672 | return; | ||||||
9673 | |||||||
9674 | // Codegen OMP target directives that offload compute to the device. | ||||||
9675 | bool RequiresDeviceCodegen = | ||||||
9676 | isa<OMPExecutableDirective>(S) && | ||||||
9677 | isOpenMPTargetExecutionDirective( | ||||||
9678 | cast<OMPExecutableDirective>(S)->getDirectiveKind()); | ||||||
9679 | |||||||
9680 | if (RequiresDeviceCodegen) { | ||||||
9681 | const auto &E = *cast<OMPExecutableDirective>(S); | ||||||
9682 | unsigned DeviceID; | ||||||
9683 | unsigned FileID; | ||||||
9684 | unsigned Line; | ||||||
9685 | getTargetEntryUniqueInfo(CGM.getContext(), E.getBeginLoc(), DeviceID, | ||||||
9686 | FileID, Line); | ||||||
9687 | |||||||
9688 | // Is this a target region that should not be emitted as an entry point? If | ||||||
9689 | // so just signal we are done with this target region. | ||||||
9690 | if (!OffloadEntriesInfoManager.hasTargetRegionEntryInfo(DeviceID, FileID, | ||||||
9691 | ParentName, Line)) | ||||||
9692 | return; | ||||||
9693 | |||||||
9694 | switch (E.getDirectiveKind()) { | ||||||
9695 | case OMPD_target: | ||||||
9696 | CodeGenFunction::EmitOMPTargetDeviceFunction(CGM, ParentName, | ||||||
9697 | cast<OMPTargetDirective>(E)); | ||||||
9698 | break; | ||||||
9699 | case OMPD_target_parallel: | ||||||
9700 | CodeGenFunction::EmitOMPTargetParallelDeviceFunction( | ||||||
9701 | CGM, ParentName, cast<OMPTargetParallelDirective>(E)); | ||||||
9702 | break; | ||||||
9703 | case OMPD_target_teams: | ||||||
9704 | CodeGenFunction::EmitOMPTargetTeamsDeviceFunction( | ||||||
9705 | CGM, ParentName, cast<OMPTargetTeamsDirective>(E)); | ||||||
9706 | break; | ||||||
9707 | case OMPD_target_teams_distribute: | ||||||
9708 | CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction( | ||||||
9709 | CGM, ParentName, cast<OMPTargetTeamsDistributeDirective>(E)); | ||||||
9710 | break; | ||||||
9711 | case OMPD_target_teams_distribute_simd: | ||||||
9712 | CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction( | ||||||
9713 | CGM, ParentName, cast<OMPTargetTeamsDistributeSimdDirective>(E)); | ||||||
9714 | break; | ||||||
9715 | case OMPD_target_parallel_for: | ||||||
9716 | CodeGenFunction::EmitOMPTargetParallelForDeviceFunction( | ||||||
9717 | CGM, ParentName, cast<OMPTargetParallelForDirective>(E)); | ||||||
9718 | break; | ||||||
9719 | case OMPD_target_parallel_for_simd: | ||||||
9720 | CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction( | ||||||
9721 | CGM, ParentName, cast<OMPTargetParallelForSimdDirective>(E)); | ||||||
9722 | break; | ||||||
9723 | case OMPD_target_simd: | ||||||
9724 | CodeGenFunction::EmitOMPTargetSimdDeviceFunction( | ||||||
9725 | CGM, ParentName, cast<OMPTargetSimdDirective>(E)); | ||||||
9726 | break; | ||||||
9727 | case OMPD_target_teams_distribute_parallel_for: | ||||||
9728 | CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction( | ||||||
9729 | CGM, ParentName, | ||||||
9730 | cast<OMPTargetTeamsDistributeParallelForDirective>(E)); | ||||||
9731 | break; | ||||||
9732 | case OMPD_target_teams_distribute_parallel_for_simd: | ||||||
9733 | CodeGenFunction:: | ||||||
9734 | EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction( | ||||||
9735 | CGM, ParentName, | ||||||
9736 | cast<OMPTargetTeamsDistributeParallelForSimdDirective>(E)); | ||||||
9737 | break; | ||||||
9738 | case OMPD_parallel: | ||||||
9739 | case OMPD_for: | ||||||
9740 | case OMPD_parallel_for: | ||||||
9741 | case OMPD_parallel_master: | ||||||
9742 | case OMPD_parallel_sections: | ||||||
9743 | case OMPD_for_simd: | ||||||
9744 | case OMPD_parallel_for_simd: | ||||||
9745 | case OMPD_cancel: | ||||||
9746 | case OMPD_cancellation_point: | ||||||
9747 | case OMPD_ordered: | ||||||
9748 | case OMPD_threadprivate: | ||||||
9749 | case OMPD_allocate: | ||||||
9750 | case OMPD_task: | ||||||
9751 | case OMPD_simd: | ||||||
9752 | case OMPD_sections: | ||||||
9753 | case OMPD_section: | ||||||
9754 | case OMPD_single: | ||||||
9755 | case OMPD_master: | ||||||
9756 | case OMPD_critical: | ||||||
9757 | case OMPD_taskyield: | ||||||
9758 | case OMPD_barrier: | ||||||
9759 | case OMPD_taskwait: | ||||||
9760 | case OMPD_taskgroup: | ||||||
9761 | case OMPD_atomic: | ||||||
9762 | case OMPD_flush: | ||||||
9763 | case OMPD_depobj: | ||||||
9764 | case OMPD_scan: | ||||||
9765 | case OMPD_teams: | ||||||
9766 | case OMPD_target_data: | ||||||
9767 | case OMPD_target_exit_data: | ||||||
9768 | case OMPD_target_enter_data: | ||||||
9769 | case OMPD_distribute: | ||||||
9770 | case OMPD_distribute_simd: | ||||||
9771 | case OMPD_distribute_parallel_for: | ||||||
9772 | case OMPD_distribute_parallel_for_simd: | ||||||
9773 | case OMPD_teams_distribute: | ||||||
9774 | case OMPD_teams_distribute_simd: | ||||||
9775 | case OMPD_teams_distribute_parallel_for: | ||||||
9776 | case OMPD_teams_distribute_parallel_for_simd: | ||||||
9777 | case OMPD_target_update: | ||||||
9778 | case OMPD_declare_simd: | ||||||
9779 | case OMPD_declare_variant: | ||||||
9780 | case OMPD_begin_declare_variant: | ||||||
9781 | case OMPD_end_declare_variant: | ||||||
9782 | case OMPD_declare_target: | ||||||
9783 | case OMPD_end_declare_target: | ||||||
9784 | case OMPD_declare_reduction: | ||||||
9785 | case OMPD_declare_mapper: | ||||||
9786 | case OMPD_taskloop: | ||||||
9787 | case OMPD_taskloop_simd: | ||||||
9788 | case OMPD_master_taskloop: | ||||||
9789 | case OMPD_master_taskloop_simd: | ||||||
9790 | case OMPD_parallel_master_taskloop: | ||||||
9791 | case OMPD_parallel_master_taskloop_simd: | ||||||
9792 | case OMPD_requires: | ||||||
9793 | case OMPD_unknown: | ||||||
9794 | default: | ||||||
9795 | llvm_unreachable("Unknown target directive for OpenMP device codegen.")::llvm::llvm_unreachable_internal("Unknown target directive for OpenMP device codegen." , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9795); | ||||||
9796 | } | ||||||
9797 | return; | ||||||
9798 | } | ||||||
9799 | |||||||
9800 | if (const auto *E = dyn_cast<OMPExecutableDirective>(S)) { | ||||||
9801 | if (!E->hasAssociatedStmt() || !E->getAssociatedStmt()) | ||||||
9802 | return; | ||||||
9803 | |||||||
9804 | scanForTargetRegionsFunctions(E->getRawStmt(), ParentName); | ||||||
9805 | return; | ||||||
9806 | } | ||||||
9807 | |||||||
9808 | // If this is a lambda function, look into its body. | ||||||
9809 | if (const auto *L = dyn_cast<LambdaExpr>(S)) | ||||||
9810 | S = L->getBody(); | ||||||
9811 | |||||||
9812 | // Keep looking for target regions recursively. | ||||||
9813 | for (const Stmt *II : S->children()) | ||||||
9814 | scanForTargetRegionsFunctions(II, ParentName); | ||||||
9815 | } | ||||||
9816 | |||||||
9817 | bool CGOpenMPRuntime::emitTargetFunctions(GlobalDecl GD) { | ||||||
9818 | // If emitting code for the host, we do not process FD here. Instead we do | ||||||
9819 | // the normal code generation. | ||||||
9820 | if (!CGM.getLangOpts().OpenMPIsDevice) { | ||||||
9821 | if (const auto *FD = dyn_cast<FunctionDecl>(GD.getDecl())) { | ||||||
9822 | Optional<OMPDeclareTargetDeclAttr::DevTypeTy> DevTy = | ||||||
9823 | OMPDeclareTargetDeclAttr::getDeviceType(FD); | ||||||
9824 | // Do not emit device_type(nohost) functions for the host. | ||||||
9825 | if (DevTy && *DevTy == OMPDeclareTargetDeclAttr::DT_NoHost) | ||||||
9826 | return true; | ||||||
9827 | } | ||||||
9828 | return false; | ||||||
9829 | } | ||||||
9830 | |||||||
9831 | const ValueDecl *VD = cast<ValueDecl>(GD.getDecl()); | ||||||
9832 | // Try to detect target regions in the function. | ||||||
9833 | if (const auto *FD = dyn_cast<FunctionDecl>(VD)) { | ||||||
9834 | StringRef Name = CGM.getMangledName(GD); | ||||||
9835 | scanForTargetRegionsFunctions(FD->getBody(), Name); | ||||||
9836 | Optional<OMPDeclareTargetDeclAttr::DevTypeTy> DevTy = | ||||||
9837 | OMPDeclareTargetDeclAttr::getDeviceType(FD); | ||||||
9838 | // Do not emit device_type(nohost) functions for the host. | ||||||
9839 | if (DevTy && *DevTy == OMPDeclareTargetDeclAttr::DT_Host) | ||||||
9840 | return true; | ||||||
9841 | } | ||||||
9842 | |||||||
9843 | // Do not to emit function if it is not marked as declare target. | ||||||
9844 | return !OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD) && | ||||||
9845 | AlreadyEmittedTargetDecls.count(VD) == 0; | ||||||
9846 | } | ||||||
9847 | |||||||
9848 | bool CGOpenMPRuntime::emitTargetGlobalVariable(GlobalDecl GD) { | ||||||
9849 | if (!CGM.getLangOpts().OpenMPIsDevice) | ||||||
9850 | return false; | ||||||
9851 | |||||||
9852 | // Check if there are Ctors/Dtors in this declaration and look for target | ||||||
9853 | // regions in it. We use the complete variant to produce the kernel name | ||||||
9854 | // mangling. | ||||||
9855 | QualType RDTy = cast<VarDecl>(GD.getDecl())->getType(); | ||||||
9856 | if (const auto *RD = RDTy->getBaseElementTypeUnsafe()->getAsCXXRecordDecl()) { | ||||||
9857 | for (const CXXConstructorDecl *Ctor : RD->ctors()) { | ||||||
9858 | StringRef ParentName = | ||||||
9859 | CGM.getMangledName(GlobalDecl(Ctor, Ctor_Complete)); | ||||||
9860 | scanForTargetRegionsFunctions(Ctor->getBody(), ParentName); | ||||||
9861 | } | ||||||
9862 | if (const CXXDestructorDecl *Dtor = RD->getDestructor()) { | ||||||
9863 | StringRef ParentName = | ||||||
9864 | CGM.getMangledName(GlobalDecl(Dtor, Dtor_Complete)); | ||||||
9865 | scanForTargetRegionsFunctions(Dtor->getBody(), ParentName); | ||||||
9866 | } | ||||||
9867 | } | ||||||
9868 | |||||||
9869 | // Do not to emit variable if it is not marked as declare target. | ||||||
9870 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | ||||||
9871 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration( | ||||||
9872 | cast<VarDecl>(GD.getDecl())); | ||||||
9873 | if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link || | ||||||
9874 | (*Res == OMPDeclareTargetDeclAttr::MT_To && | ||||||
9875 | HasRequiresUnifiedSharedMemory)) { | ||||||
9876 | DeferredGlobalVariables.insert(cast<VarDecl>(GD.getDecl())); | ||||||
9877 | return true; | ||||||
9878 | } | ||||||
9879 | return false; | ||||||
9880 | } | ||||||
9881 | |||||||
9882 | llvm::Constant * | ||||||
9883 | CGOpenMPRuntime::registerTargetFirstprivateCopy(CodeGenFunction &CGF, | ||||||
9884 | const VarDecl *VD) { | ||||||
9885 | assert(VD->getType().isConstant(CGM.getContext()) &&((VD->getType().isConstant(CGM.getContext()) && "Expected constant variable." ) ? static_cast<void> (0) : __assert_fail ("VD->getType().isConstant(CGM.getContext()) && \"Expected constant variable.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9886, __PRETTY_FUNCTION__)) | ||||||
9886 | "Expected constant variable.")((VD->getType().isConstant(CGM.getContext()) && "Expected constant variable." ) ? static_cast<void> (0) : __assert_fail ("VD->getType().isConstant(CGM.getContext()) && \"Expected constant variable.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9886, __PRETTY_FUNCTION__)); | ||||||
9887 | StringRef VarName; | ||||||
9888 | llvm::Constant *Addr; | ||||||
9889 | llvm::GlobalValue::LinkageTypes Linkage; | ||||||
9890 | QualType Ty = VD->getType(); | ||||||
9891 | SmallString<128> Buffer; | ||||||
9892 | { | ||||||
9893 | unsigned DeviceID; | ||||||
9894 | unsigned FileID; | ||||||
9895 | unsigned Line; | ||||||
9896 | getTargetEntryUniqueInfo(CGM.getContext(), VD->getLocation(), DeviceID, | ||||||
9897 | FileID, Line); | ||||||
9898 | llvm::raw_svector_ostream OS(Buffer); | ||||||
9899 | OS << "__omp_offloading_firstprivate_" << llvm::format("_%x", DeviceID) | ||||||
9900 | << llvm::format("_%x_", FileID) << VD->getName() << "_l" << Line; | ||||||
9901 | VarName = OS.str(); | ||||||
9902 | } | ||||||
9903 | Linkage = llvm::GlobalValue::InternalLinkage; | ||||||
9904 | Addr = | ||||||
9905 | getOrCreateInternalVariable(CGM.getTypes().ConvertTypeForMem(Ty), VarName, | ||||||
9906 | getDefaultFirstprivateAddressSpace()); | ||||||
9907 | cast<llvm::GlobalValue>(Addr)->setLinkage(Linkage); | ||||||
9908 | CharUnits VarSize = CGM.getContext().getTypeSizeInChars(Ty); | ||||||
9909 | CGM.addCompilerUsedGlobal(cast<llvm::GlobalValue>(Addr)); | ||||||
9910 | OffloadEntriesInfoManager.registerDeviceGlobalVarEntryInfo( | ||||||
9911 | VarName, Addr, VarSize, | ||||||
9912 | OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo, Linkage); | ||||||
9913 | return Addr; | ||||||
9914 | } | ||||||
9915 | |||||||
9916 | void CGOpenMPRuntime::registerTargetGlobalVariable(const VarDecl *VD, | ||||||
9917 | llvm::Constant *Addr) { | ||||||
9918 | if (CGM.getLangOpts().OMPTargetTriples.empty() && | ||||||
9919 | !CGM.getLangOpts().OpenMPIsDevice) | ||||||
9920 | return; | ||||||
9921 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | ||||||
9922 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | ||||||
9923 | if (!Res) { | ||||||
9924 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
9925 | // Register non-target variables being emitted in device code (debug info | ||||||
9926 | // may cause this). | ||||||
9927 | StringRef VarName = CGM.getMangledName(VD); | ||||||
9928 | EmittedNonTargetVariables.try_emplace(VarName, Addr); | ||||||
9929 | } | ||||||
9930 | return; | ||||||
9931 | } | ||||||
9932 | // Register declare target variables. | ||||||
9933 | OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind Flags; | ||||||
9934 | StringRef VarName; | ||||||
9935 | CharUnits VarSize; | ||||||
9936 | llvm::GlobalValue::LinkageTypes Linkage; | ||||||
9937 | |||||||
9938 | if (*Res == OMPDeclareTargetDeclAttr::MT_To && | ||||||
9939 | !HasRequiresUnifiedSharedMemory) { | ||||||
9940 | Flags = OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo; | ||||||
9941 | VarName = CGM.getMangledName(VD); | ||||||
9942 | if (VD->hasDefinition(CGM.getContext()) != VarDecl::DeclarationOnly) { | ||||||
9943 | VarSize = CGM.getContext().getTypeSizeInChars(VD->getType()); | ||||||
9944 | assert(!VarSize.isZero() && "Expected non-zero size of the variable")((!VarSize.isZero() && "Expected non-zero size of the variable" ) ? static_cast<void> (0) : __assert_fail ("!VarSize.isZero() && \"Expected non-zero size of the variable\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9944, __PRETTY_FUNCTION__)); | ||||||
9945 | } else { | ||||||
9946 | VarSize = CharUnits::Zero(); | ||||||
9947 | } | ||||||
9948 | Linkage = CGM.getLLVMLinkageVarDefinition(VD, /*IsConstant=*/false); | ||||||
9949 | // Temp solution to prevent optimizations of the internal variables. | ||||||
9950 | if (CGM.getLangOpts().OpenMPIsDevice && !VD->isExternallyVisible()) { | ||||||
9951 | std::string RefName = getName({VarName, "ref"}); | ||||||
9952 | if (!CGM.GetGlobalValue(RefName)) { | ||||||
9953 | llvm::Constant *AddrRef = | ||||||
9954 | getOrCreateInternalVariable(Addr->getType(), RefName); | ||||||
9955 | auto *GVAddrRef = cast<llvm::GlobalVariable>(AddrRef); | ||||||
9956 | GVAddrRef->setConstant(/*Val=*/true); | ||||||
9957 | GVAddrRef->setLinkage(llvm::GlobalValue::InternalLinkage); | ||||||
9958 | GVAddrRef->setInitializer(Addr); | ||||||
9959 | CGM.addCompilerUsedGlobal(GVAddrRef); | ||||||
9960 | } | ||||||
9961 | } | ||||||
9962 | } else { | ||||||
9963 | assert(((*Res == OMPDeclareTargetDeclAttr::MT_Link) ||((((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? static_cast<void> (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9966, __PRETTY_FUNCTION__)) | ||||||
9964 | (*Res == OMPDeclareTargetDeclAttr::MT_To &&((((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? static_cast<void> (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9966, __PRETTY_FUNCTION__)) | ||||||
9965 | HasRequiresUnifiedSharedMemory)) &&((((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? static_cast<void> (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9966, __PRETTY_FUNCTION__)) | ||||||
9966 | "Declare target attribute must link or to with unified memory.")((((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Declare target attribute must link or to with unified memory." ) ? static_cast<void> (0) : __assert_fail ("((*Res == OMPDeclareTargetDeclAttr::MT_Link) || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Declare target attribute must link or to with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9966, __PRETTY_FUNCTION__)); | ||||||
9967 | if (*Res == OMPDeclareTargetDeclAttr::MT_Link) | ||||||
9968 | Flags = OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryLink; | ||||||
9969 | else | ||||||
9970 | Flags = OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo; | ||||||
9971 | |||||||
9972 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
9973 | VarName = Addr->getName(); | ||||||
9974 | Addr = nullptr; | ||||||
9975 | } else { | ||||||
9976 | VarName = getAddrOfDeclareTargetVar(VD).getName(); | ||||||
9977 | Addr = cast<llvm::Constant>(getAddrOfDeclareTargetVar(VD).getPointer()); | ||||||
9978 | } | ||||||
9979 | VarSize = CGM.getPointerSize(); | ||||||
9980 | Linkage = llvm::GlobalValue::WeakAnyLinkage; | ||||||
9981 | } | ||||||
9982 | |||||||
9983 | OffloadEntriesInfoManager.registerDeviceGlobalVarEntryInfo( | ||||||
9984 | VarName, Addr, VarSize, Flags, Linkage); | ||||||
9985 | } | ||||||
9986 | |||||||
9987 | bool CGOpenMPRuntime::emitTargetGlobal(GlobalDecl GD) { | ||||||
9988 | if (isa<FunctionDecl>(GD.getDecl()) || | ||||||
9989 | isa<OMPDeclareReductionDecl>(GD.getDecl())) | ||||||
9990 | return emitTargetFunctions(GD); | ||||||
9991 | |||||||
9992 | return emitTargetGlobalVariable(GD); | ||||||
9993 | } | ||||||
9994 | |||||||
9995 | void CGOpenMPRuntime::emitDeferredTargetDecls() const { | ||||||
9996 | for (const VarDecl *VD : DeferredGlobalVariables) { | ||||||
9997 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | ||||||
9998 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | ||||||
9999 | if (!Res) | ||||||
10000 | continue; | ||||||
10001 | if (*Res == OMPDeclareTargetDeclAttr::MT_To && | ||||||
10002 | !HasRequiresUnifiedSharedMemory) { | ||||||
10003 | CGM.EmitGlobal(VD); | ||||||
10004 | } else { | ||||||
10005 | assert((*Res == OMPDeclareTargetDeclAttr::MT_Link ||(((*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory.") ? static_cast <void> (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10008, __PRETTY_FUNCTION__)) | ||||||
10006 | (*Res == OMPDeclareTargetDeclAttr::MT_To &&(((*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory.") ? static_cast <void> (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10008, __PRETTY_FUNCTION__)) | ||||||
10007 | HasRequiresUnifiedSharedMemory)) &&(((*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory.") ? static_cast <void> (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10008, __PRETTY_FUNCTION__)) | ||||||
10008 | "Expected link clause or to clause with unified memory.")(((*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr ::MT_To && HasRequiresUnifiedSharedMemory)) && "Expected link clause or to clause with unified memory.") ? static_cast <void> (0) : __assert_fail ("(*Res == OMPDeclareTargetDeclAttr::MT_Link || (*Res == OMPDeclareTargetDeclAttr::MT_To && HasRequiresUnifiedSharedMemory)) && \"Expected link clause or to clause with unified memory.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10008, __PRETTY_FUNCTION__)); | ||||||
10009 | (void)CGM.getOpenMPRuntime().getAddrOfDeclareTargetVar(VD); | ||||||
10010 | } | ||||||
10011 | } | ||||||
10012 | } | ||||||
10013 | |||||||
10014 | void CGOpenMPRuntime::adjustTargetSpecificDataForLambdas( | ||||||
10015 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { | ||||||
10016 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && " Expected target-based directive.") ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10017, __PRETTY_FUNCTION__)) | ||||||
10017 | " Expected target-based directive.")((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && " Expected target-based directive.") ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10017, __PRETTY_FUNCTION__)); | ||||||
10018 | } | ||||||
10019 | |||||||
10020 | void CGOpenMPRuntime::processRequiresDirective(const OMPRequiresDecl *D) { | ||||||
10021 | for (const OMPClause *Clause : D->clauselists()) { | ||||||
10022 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { | ||||||
10023 | HasRequiresUnifiedSharedMemory = true; | ||||||
10024 | } else if (const auto *AC = | ||||||
10025 | dyn_cast<OMPAtomicDefaultMemOrderClause>(Clause)) { | ||||||
10026 | switch (AC->getAtomicDefaultMemOrderKind()) { | ||||||
10027 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_acq_rel: | ||||||
10028 | RequiresAtomicOrdering = llvm::AtomicOrdering::AcquireRelease; | ||||||
10029 | break; | ||||||
10030 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_seq_cst: | ||||||
10031 | RequiresAtomicOrdering = llvm::AtomicOrdering::SequentiallyConsistent; | ||||||
10032 | break; | ||||||
10033 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_relaxed: | ||||||
10034 | RequiresAtomicOrdering = llvm::AtomicOrdering::Monotonic; | ||||||
10035 | break; | ||||||
10036 | case OMPC_ATOMIC_DEFAULT_MEM_ORDER_unknown: | ||||||
10037 | break; | ||||||
10038 | } | ||||||
10039 | } | ||||||
10040 | } | ||||||
10041 | } | ||||||
10042 | |||||||
10043 | llvm::AtomicOrdering CGOpenMPRuntime::getDefaultMemoryOrdering() const { | ||||||
10044 | return RequiresAtomicOrdering; | ||||||
10045 | } | ||||||
10046 | |||||||
10047 | bool CGOpenMPRuntime::hasAllocateAttributeForGlobalVar(const VarDecl *VD, | ||||||
10048 | LangAS &AS) { | ||||||
10049 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) | ||||||
10050 | return false; | ||||||
10051 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); | ||||||
10052 | switch(A->getAllocatorType()) { | ||||||
10053 | case OMPAllocateDeclAttr::OMPNullMemAlloc: | ||||||
10054 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: | ||||||
10055 | // Not supported, fallback to the default mem space. | ||||||
10056 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: | ||||||
10057 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: | ||||||
10058 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: | ||||||
10059 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: | ||||||
10060 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: | ||||||
10061 | case OMPAllocateDeclAttr::OMPConstMemAlloc: | ||||||
10062 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: | ||||||
10063 | AS = LangAS::Default; | ||||||
10064 | return true; | ||||||
10065 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: | ||||||
10066 | llvm_unreachable("Expected predefined allocator for the variables with the "::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10067) | ||||||
10067 | "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10067); | ||||||
10068 | } | ||||||
10069 | return false; | ||||||
10070 | } | ||||||
10071 | |||||||
10072 | bool CGOpenMPRuntime::hasRequiresUnifiedSharedMemory() const { | ||||||
10073 | return HasRequiresUnifiedSharedMemory; | ||||||
10074 | } | ||||||
10075 | |||||||
10076 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII::DisableAutoDeclareTargetRAII( | ||||||
10077 | CodeGenModule &CGM) | ||||||
10078 | : CGM(CGM) { | ||||||
10079 | if (CGM.getLangOpts().OpenMPIsDevice) { | ||||||
10080 | SavedShouldMarkAsGlobal = CGM.getOpenMPRuntime().ShouldMarkAsGlobal; | ||||||
10081 | CGM.getOpenMPRuntime().ShouldMarkAsGlobal = false; | ||||||
10082 | } | ||||||
10083 | } | ||||||
10084 | |||||||
10085 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII::~DisableAutoDeclareTargetRAII() { | ||||||
10086 | if (CGM.getLangOpts().OpenMPIsDevice) | ||||||
10087 | CGM.getOpenMPRuntime().ShouldMarkAsGlobal = SavedShouldMarkAsGlobal; | ||||||
10088 | } | ||||||
10089 | |||||||
10090 | bool CGOpenMPRuntime::markAsGlobalTarget(GlobalDecl GD) { | ||||||
10091 | if (!CGM.getLangOpts().OpenMPIsDevice || !ShouldMarkAsGlobal) | ||||||
10092 | return true; | ||||||
10093 | |||||||
10094 | const auto *D = cast<FunctionDecl>(GD.getDecl()); | ||||||
10095 | // Do not to emit function if it is marked as declare target as it was already | ||||||
10096 | // emitted. | ||||||
10097 | if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(D)) { | ||||||
10098 | if (D->hasBody() && AlreadyEmittedTargetDecls.count(D) == 0) { | ||||||
10099 | if (auto *F = dyn_cast_or_null<llvm::Function>( | ||||||
10100 | CGM.GetGlobalValue(CGM.getMangledName(GD)))) | ||||||
10101 | return !F->isDeclaration(); | ||||||
10102 | return false; | ||||||
10103 | } | ||||||
10104 | return true; | ||||||
10105 | } | ||||||
10106 | |||||||
10107 | return !AlreadyEmittedTargetDecls.insert(D).second; | ||||||
10108 | } | ||||||
10109 | |||||||
10110 | llvm::Function *CGOpenMPRuntime::emitRequiresDirectiveRegFun() { | ||||||
10111 | // If we don't have entries or if we are emitting code for the device, we | ||||||
10112 | // don't need to do anything. | ||||||
10113 | if (CGM.getLangOpts().OMPTargetTriples.empty() || | ||||||
10114 | CGM.getLangOpts().OpenMPSimd || CGM.getLangOpts().OpenMPIsDevice || | ||||||
10115 | (OffloadEntriesInfoManager.empty() && | ||||||
10116 | !HasEmittedDeclareTargetRegion && | ||||||
10117 | !HasEmittedTargetRegion)) | ||||||
10118 | return nullptr; | ||||||
10119 | |||||||
10120 | // Create and register the function that handles the requires directives. | ||||||
10121 | ASTContext &C = CGM.getContext(); | ||||||
10122 | |||||||
10123 | llvm::Function *RequiresRegFn; | ||||||
10124 | { | ||||||
10125 | CodeGenFunction CGF(CGM); | ||||||
10126 | const auto &FI = CGM.getTypes().arrangeNullaryFunction(); | ||||||
10127 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | ||||||
10128 | std::string ReqName = getName({"omp_offloading", "requires_reg"}); | ||||||
10129 | RequiresRegFn = CGM.CreateGlobalInitOrCleanUpFunction(FTy, ReqName, FI); | ||||||
10130 | CGF.StartFunction(GlobalDecl(), C.VoidTy, RequiresRegFn, FI, {}); | ||||||
10131 | OpenMPOffloadingRequiresDirFlags Flags = OMP_REQ_NONE; | ||||||
10132 | // TODO: check for other requires clauses. | ||||||
10133 | // The requires directive takes effect only when a target region is | ||||||
10134 | // present in the compilation unit. Otherwise it is ignored and not | ||||||
10135 | // passed to the runtime. This avoids the runtime from throwing an error | ||||||
10136 | // for mismatching requires clauses across compilation units that don't | ||||||
10137 | // contain at least 1 target region. | ||||||
10138 | assert((HasEmittedTargetRegion ||(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10141, __PRETTY_FUNCTION__)) | ||||||
10139 | HasEmittedDeclareTargetRegion ||(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10141, __PRETTY_FUNCTION__)) | ||||||
10140 | !OffloadEntriesInfoManager.empty()) &&(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10141, __PRETTY_FUNCTION__)) | ||||||
10141 | "Target or declare target region expected.")(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10141, __PRETTY_FUNCTION__)); | ||||||
10142 | if (HasRequiresUnifiedSharedMemory) | ||||||
10143 | Flags = OMP_REQ_UNIFIED_SHARED_MEMORY; | ||||||
10144 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
10145 | CGM.getModule(), OMPRTL___tgt_register_requires), | ||||||
10146 | llvm::ConstantInt::get(CGM.Int64Ty, Flags)); | ||||||
10147 | CGF.FinishFunction(); | ||||||
10148 | } | ||||||
10149 | return RequiresRegFn; | ||||||
10150 | } | ||||||
10151 | |||||||
10152 | void CGOpenMPRuntime::emitTeamsCall(CodeGenFunction &CGF, | ||||||
10153 | const OMPExecutableDirective &D, | ||||||
10154 | SourceLocation Loc, | ||||||
10155 | llvm::Function *OutlinedFn, | ||||||
10156 | ArrayRef<llvm::Value *> CapturedVars) { | ||||||
10157 | if (!CGF.HaveInsertPoint()) | ||||||
10158 | return; | ||||||
10159 | |||||||
10160 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | ||||||
10161 | CodeGenFunction::RunCleanupsScope Scope(CGF); | ||||||
10162 | |||||||
10163 | // Build call __kmpc_fork_teams(loc, n, microtask, var1, .., varn); | ||||||
10164 | llvm::Value *Args[] = { | ||||||
10165 | RTLoc, | ||||||
10166 | CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars | ||||||
10167 | CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy())}; | ||||||
10168 | llvm::SmallVector<llvm::Value *, 16> RealArgs; | ||||||
10169 | RealArgs.append(std::begin(Args), std::end(Args)); | ||||||
10170 | RealArgs.append(CapturedVars.begin(), CapturedVars.end()); | ||||||
10171 | |||||||
10172 | llvm::FunctionCallee RTLFn = OMPBuilder.getOrCreateRuntimeFunction( | ||||||
10173 | CGM.getModule(), OMPRTL___kmpc_fork_teams); | ||||||
10174 | CGF.EmitRuntimeCall(RTLFn, RealArgs); | ||||||
10175 | } | ||||||
10176 | |||||||
10177 | void CGOpenMPRuntime::emitNumTeamsClause(CodeGenFunction &CGF, | ||||||
10178 | const Expr *NumTeams, | ||||||
10179 | const Expr *ThreadLimit, | ||||||
10180 | SourceLocation Loc) { | ||||||
10181 | if (!CGF.HaveInsertPoint()) | ||||||
10182 | return; | ||||||
10183 | |||||||
10184 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | ||||||
10185 | |||||||
10186 | llvm::Value *NumTeamsVal = | ||||||
10187 | NumTeams | ||||||
10188 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(NumTeams), | ||||||
10189 | CGF.CGM.Int32Ty, /* isSigned = */ true) | ||||||
10190 | : CGF.Builder.getInt32(0); | ||||||
10191 | |||||||
10192 | llvm::Value *ThreadLimitVal = | ||||||
10193 | ThreadLimit | ||||||
10194 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(ThreadLimit), | ||||||
10195 | CGF.CGM.Int32Ty, /* isSigned = */ true) | ||||||
10196 | : CGF.Builder.getInt32(0); | ||||||
10197 | |||||||
10198 | // Build call __kmpc_push_num_teamss(&loc, global_tid, num_teams, thread_limit) | ||||||
10199 | llvm::Value *PushNumTeamsArgs[] = {RTLoc, getThreadID(CGF, Loc), NumTeamsVal, | ||||||
10200 | ThreadLimitVal}; | ||||||
10201 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
10202 | CGM.getModule(), OMPRTL___kmpc_push_num_teams), | ||||||
10203 | PushNumTeamsArgs); | ||||||
10204 | } | ||||||
10205 | |||||||
10206 | void CGOpenMPRuntime::emitTargetDataCalls( | ||||||
10207 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | ||||||
10208 | const Expr *Device, const RegionCodeGenTy &CodeGen, TargetDataInfo &Info) { | ||||||
10209 | if (!CGF.HaveInsertPoint()) | ||||||
10210 | return; | ||||||
10211 | |||||||
10212 | // Action used to replace the default codegen action and turn privatization | ||||||
10213 | // off. | ||||||
10214 | PrePostActionTy NoPrivAction; | ||||||
10215 | |||||||
10216 | // Generate the code for the opening of the data environment. Capture all the | ||||||
10217 | // arguments of the runtime call by reference because they are used in the | ||||||
10218 | // closing of the region. | ||||||
10219 | auto &&BeginThenGen = [this, &D, Device, &Info, | ||||||
10220 | &CodeGen](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
10221 | // Fill up the arrays with all the mapped variables. | ||||||
10222 | MappableExprsHandler::MapCombinedInfoTy CombinedInfo; | ||||||
10223 | |||||||
10224 | // Get map clause information. | ||||||
10225 | MappableExprsHandler MEHandler(D, CGF); | ||||||
10226 | MEHandler.generateAllInfo(CombinedInfo); | ||||||
10227 | |||||||
10228 | // Fill up the arrays and create the arguments. | ||||||
10229 | emitOffloadingArrays(CGF, CombinedInfo, Info); | ||||||
10230 | |||||||
10231 | llvm::Value *BasePointersArrayArg = nullptr; | ||||||
10232 | llvm::Value *PointersArrayArg = nullptr; | ||||||
10233 | llvm::Value *SizesArrayArg = nullptr; | ||||||
10234 | llvm::Value *MapTypesArrayArg = nullptr; | ||||||
10235 | llvm::Value *MappersArrayArg = nullptr; | ||||||
10236 | emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg, | ||||||
10237 | SizesArrayArg, MapTypesArrayArg, | ||||||
10238 | MappersArrayArg, Info, /*ForEndCall=*/false); | ||||||
10239 | |||||||
10240 | // Emit device ID if any. | ||||||
10241 | llvm::Value *DeviceID = nullptr; | ||||||
10242 | if (Device) { | ||||||
10243 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | ||||||
10244 | CGF.Int64Ty, /*isSigned=*/true); | ||||||
10245 | } else { | ||||||
10246 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | ||||||
10247 | } | ||||||
10248 | |||||||
10249 | // Emit the number of elements in the offloading arrays. | ||||||
10250 | llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs); | ||||||
10251 | |||||||
10252 | llvm::Value *OffloadingArgs[] = { | ||||||
10253 | DeviceID, PointerNum, BasePointersArrayArg, PointersArrayArg, | ||||||
10254 | SizesArrayArg, MapTypesArrayArg, MappersArrayArg}; | ||||||
10255 | CGF.EmitRuntimeCall( | ||||||
10256 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
10257 | CGM.getModule(), OMPRTL___tgt_target_data_begin_mapper), | ||||||
10258 | OffloadingArgs); | ||||||
10259 | |||||||
10260 | // If device pointer privatization is required, emit the body of the region | ||||||
10261 | // here. It will have to be duplicated: with and without privatization. | ||||||
10262 | if (!Info.CaptureDeviceAddrMap.empty()) | ||||||
10263 | CodeGen(CGF); | ||||||
10264 | }; | ||||||
10265 | |||||||
10266 | // Generate code for the closing of the data region. | ||||||
10267 | auto &&EndThenGen = [this, Device, &Info](CodeGenFunction &CGF, | ||||||
10268 | PrePostActionTy &) { | ||||||
10269 | assert(Info.isValid() && "Invalid data environment closing arguments.")((Info.isValid() && "Invalid data environment closing arguments." ) ? static_cast<void> (0) : __assert_fail ("Info.isValid() && \"Invalid data environment closing arguments.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10269, __PRETTY_FUNCTION__)); | ||||||
10270 | |||||||
10271 | llvm::Value *BasePointersArrayArg = nullptr; | ||||||
10272 | llvm::Value *PointersArrayArg = nullptr; | ||||||
10273 | llvm::Value *SizesArrayArg = nullptr; | ||||||
10274 | llvm::Value *MapTypesArrayArg = nullptr; | ||||||
10275 | llvm::Value *MappersArrayArg = nullptr; | ||||||
10276 | emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg, | ||||||
10277 | SizesArrayArg, MapTypesArrayArg, | ||||||
10278 | MappersArrayArg, Info, /*ForEndCall=*/true); | ||||||
10279 | |||||||
10280 | // Emit device ID if any. | ||||||
10281 | llvm::Value *DeviceID = nullptr; | ||||||
10282 | if (Device) { | ||||||
10283 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | ||||||
10284 | CGF.Int64Ty, /*isSigned=*/true); | ||||||
10285 | } else { | ||||||
10286 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | ||||||
10287 | } | ||||||
10288 | |||||||
10289 | // Emit the number of elements in the offloading arrays. | ||||||
10290 | llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs); | ||||||
10291 | |||||||
10292 | llvm::Value *OffloadingArgs[] = { | ||||||
10293 | DeviceID, PointerNum, BasePointersArrayArg, PointersArrayArg, | ||||||
10294 | SizesArrayArg, MapTypesArrayArg, MappersArrayArg}; | ||||||
10295 | CGF.EmitRuntimeCall( | ||||||
10296 | OMPBuilder.getOrCreateRuntimeFunction( | ||||||
10297 | CGM.getModule(), OMPRTL___tgt_target_data_end_mapper), | ||||||
10298 | OffloadingArgs); | ||||||
10299 | }; | ||||||
10300 | |||||||
10301 | // If we need device pointer privatization, we need to emit the body of the | ||||||
10302 | // region with no privatization in the 'else' branch of the conditional. | ||||||
10303 | // Otherwise, we don't have to do anything. | ||||||
10304 | auto &&BeginElseGen = [&Info, &CodeGen, &NoPrivAction](CodeGenFunction &CGF, | ||||||
10305 | PrePostActionTy &) { | ||||||
10306 | if (!Info.CaptureDeviceAddrMap.empty()) { | ||||||
10307 | CodeGen.setAction(NoPrivAction); | ||||||
10308 | CodeGen(CGF); | ||||||
10309 | } | ||||||
10310 | }; | ||||||
10311 | |||||||
10312 | // We don't have to do anything to close the region if the if clause evaluates | ||||||
10313 | // to false. | ||||||
10314 | auto &&EndElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {}; | ||||||
10315 | |||||||
10316 | if (IfCond) { | ||||||
10317 | emitIfClause(CGF, IfCond, BeginThenGen, BeginElseGen); | ||||||
10318 | } else { | ||||||
10319 | RegionCodeGenTy RCG(BeginThenGen); | ||||||
10320 | RCG(CGF); | ||||||
10321 | } | ||||||
10322 | |||||||
10323 | // If we don't require privatization of device pointers, we emit the body in | ||||||
10324 | // between the runtime calls. This avoids duplicating the body code. | ||||||
10325 | if (Info.CaptureDeviceAddrMap.empty()) { | ||||||
10326 | CodeGen.setAction(NoPrivAction); | ||||||
10327 | CodeGen(CGF); | ||||||
10328 | } | ||||||
10329 | |||||||
10330 | if (IfCond) { | ||||||
10331 | emitIfClause(CGF, IfCond, EndThenGen, EndElseGen); | ||||||
10332 | } else { | ||||||
10333 | RegionCodeGenTy RCG(EndThenGen); | ||||||
10334 | RCG(CGF); | ||||||
10335 | } | ||||||
10336 | } | ||||||
10337 | |||||||
10338 | void CGOpenMPRuntime::emitTargetDataStandAloneCall( | ||||||
10339 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | ||||||
10340 | const Expr *Device) { | ||||||
10341 | if (!CGF.HaveInsertPoint()) | ||||||
10342 | return; | ||||||
10343 | |||||||
10344 | assert((isa<OMPTargetEnterDataDirective>(D) ||(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10347, __PRETTY_FUNCTION__)) | ||||||
10345 | isa<OMPTargetExitDataDirective>(D) ||(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10347, __PRETTY_FUNCTION__)) | ||||||
10346 | isa<OMPTargetUpdateDirective>(D)) &&(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10347, __PRETTY_FUNCTION__)) | ||||||
10347 | "Expecting either target enter, exit data, or update directives.")(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10347, __PRETTY_FUNCTION__)); | ||||||
10348 | |||||||
10349 | CodeGenFunction::OMPTargetDataInfo InputInfo; | ||||||
10350 | llvm::Value *MapTypesArray = nullptr; | ||||||
10351 | // Generate the code for the opening of the data environment. | ||||||
10352 | auto &&ThenGen = [this, &D, Device, &InputInfo, | ||||||
10353 | &MapTypesArray](CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
10354 | // Emit device ID if any. | ||||||
10355 | llvm::Value *DeviceID = nullptr; | ||||||
10356 | if (Device) { | ||||||
10357 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | ||||||
10358 | CGF.Int64Ty, /*isSigned=*/true); | ||||||
10359 | } else { | ||||||
10360 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | ||||||
10361 | } | ||||||
10362 | |||||||
10363 | // Emit the number of elements in the offloading arrays. | ||||||
10364 | llvm::Constant *PointerNum = | ||||||
10365 | CGF.Builder.getInt32(InputInfo.NumberOfTargetItems); | ||||||
10366 | |||||||
10367 | llvm::Value *OffloadingArgs[] = {DeviceID, | ||||||
10368 | PointerNum, | ||||||
10369 | InputInfo.BasePointersArray.getPointer(), | ||||||
10370 | InputInfo.PointersArray.getPointer(), | ||||||
10371 | InputInfo.SizesArray.getPointer(), | ||||||
10372 | MapTypesArray, | ||||||
10373 | InputInfo.MappersArray.getPointer()}; | ||||||
10374 | |||||||
10375 | // Select the right runtime function call for each standalone | ||||||
10376 | // directive. | ||||||
10377 | const bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>(); | ||||||
10378 | RuntimeFunction RTLFn; | ||||||
10379 | switch (D.getDirectiveKind()) { | ||||||
10380 | case OMPD_target_enter_data: | ||||||
10381 | RTLFn = HasNowait ? OMPRTL___tgt_target_data_begin_nowait_mapper | ||||||
10382 | : OMPRTL___tgt_target_data_begin_mapper; | ||||||
10383 | break; | ||||||
10384 | case OMPD_target_exit_data: | ||||||
10385 | RTLFn = HasNowait ? OMPRTL___tgt_target_data_end_nowait_mapper | ||||||
10386 | : OMPRTL___tgt_target_data_end_mapper; | ||||||
10387 | break; | ||||||
10388 | case OMPD_target_update: | ||||||
10389 | RTLFn = HasNowait ? OMPRTL___tgt_target_data_update_nowait_mapper | ||||||
10390 | : OMPRTL___tgt_target_data_update_mapper; | ||||||
10391 | break; | ||||||
10392 | case OMPD_parallel: | ||||||
10393 | case OMPD_for: | ||||||
10394 | case OMPD_parallel_for: | ||||||
10395 | case OMPD_parallel_master: | ||||||
10396 | case OMPD_parallel_sections: | ||||||
10397 | case OMPD_for_simd: | ||||||
10398 | case OMPD_parallel_for_simd: | ||||||
10399 | case OMPD_cancel: | ||||||
10400 | case OMPD_cancellation_point: | ||||||
10401 | case OMPD_ordered: | ||||||
10402 | case OMPD_threadprivate: | ||||||
10403 | case OMPD_allocate: | ||||||
10404 | case OMPD_task: | ||||||
10405 | case OMPD_simd: | ||||||
10406 | case OMPD_sections: | ||||||
10407 | case OMPD_section: | ||||||
10408 | case OMPD_single: | ||||||
10409 | case OMPD_master: | ||||||
10410 | case OMPD_critical: | ||||||
10411 | case OMPD_taskyield: | ||||||
10412 | case OMPD_barrier: | ||||||
10413 | case OMPD_taskwait: | ||||||
10414 | case OMPD_taskgroup: | ||||||
10415 | case OMPD_atomic: | ||||||
10416 | case OMPD_flush: | ||||||
10417 | case OMPD_depobj: | ||||||
10418 | case OMPD_scan: | ||||||
10419 | case OMPD_teams: | ||||||
10420 | case OMPD_target_data: | ||||||
10421 | case OMPD_distribute: | ||||||
10422 | case OMPD_distribute_simd: | ||||||
10423 | case OMPD_distribute_parallel_for: | ||||||
10424 | case OMPD_distribute_parallel_for_simd: | ||||||
10425 | case OMPD_teams_distribute: | ||||||
10426 | case OMPD_teams_distribute_simd: | ||||||
10427 | case OMPD_teams_distribute_parallel_for: | ||||||
10428 | case OMPD_teams_distribute_parallel_for_simd: | ||||||
10429 | case OMPD_declare_simd: | ||||||
10430 | case OMPD_declare_variant: | ||||||
10431 | case OMPD_begin_declare_variant: | ||||||
10432 | case OMPD_end_declare_variant: | ||||||
10433 | case OMPD_declare_target: | ||||||
10434 | case OMPD_end_declare_target: | ||||||
10435 | case OMPD_declare_reduction: | ||||||
10436 | case OMPD_declare_mapper: | ||||||
10437 | case OMPD_taskloop: | ||||||
10438 | case OMPD_taskloop_simd: | ||||||
10439 | case OMPD_master_taskloop: | ||||||
10440 | case OMPD_master_taskloop_simd: | ||||||
10441 | case OMPD_parallel_master_taskloop: | ||||||
10442 | case OMPD_parallel_master_taskloop_simd: | ||||||
10443 | case OMPD_target: | ||||||
10444 | case OMPD_target_simd: | ||||||
10445 | case OMPD_target_teams_distribute: | ||||||
10446 | case OMPD_target_teams_distribute_simd: | ||||||
10447 | case OMPD_target_teams_distribute_parallel_for: | ||||||
10448 | case OMPD_target_teams_distribute_parallel_for_simd: | ||||||
10449 | case OMPD_target_teams: | ||||||
10450 | case OMPD_target_parallel: | ||||||
10451 | case OMPD_target_parallel_for: | ||||||
10452 | case OMPD_target_parallel_for_simd: | ||||||
10453 | case OMPD_requires: | ||||||
10454 | case OMPD_unknown: | ||||||
10455 | default: | ||||||
10456 | llvm_unreachable("Unexpected standalone target data directive.")::llvm::llvm_unreachable_internal("Unexpected standalone target data directive." , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10456); | ||||||
10457 | break; | ||||||
10458 | } | ||||||
10459 | CGF.EmitRuntimeCall( | ||||||
10460 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), RTLFn), | ||||||
10461 | OffloadingArgs); | ||||||
10462 | }; | ||||||
10463 | |||||||
10464 | auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray]( | ||||||
10465 | CodeGenFunction &CGF, PrePostActionTy &) { | ||||||
10466 | // Fill up the arrays with all the mapped variables. | ||||||
10467 | MappableExprsHandler::MapCombinedInfoTy CombinedInfo; | ||||||
10468 | |||||||
10469 | // Get map clause information. | ||||||
10470 | MappableExprsHandler MEHandler(D, CGF); | ||||||
10471 | MEHandler.generateAllInfo(CombinedInfo); | ||||||
10472 | |||||||
10473 | TargetDataInfo Info; | ||||||
10474 | // Fill up the arrays and create the arguments. | ||||||
10475 | emitOffloadingArrays(CGF, CombinedInfo, Info); | ||||||
10476 | emitOffloadingArraysArgument(CGF, Info.BasePointersArray, | ||||||
10477 | Info.PointersArray, Info.SizesArray, | ||||||
10478 | Info.MapTypesArray, Info.MappersArray, Info); | ||||||
10479 | InputInfo.NumberOfTargetItems = Info.NumberOfPtrs; | ||||||
10480 | InputInfo.BasePointersArray = | ||||||
10481 | Address(Info.BasePointersArray, CGM.getPointerAlign()); | ||||||
10482 | InputInfo.PointersArray = | ||||||
10483 | Address(Info.PointersArray, CGM.getPointerAlign()); | ||||||
10484 | InputInfo.SizesArray = | ||||||
10485 | Address(Info.SizesArray, CGM.getPointerAlign()); | ||||||
10486 | InputInfo.MappersArray = Address(Info.MappersArray, CGM.getPointerAlign()); | ||||||
10487 | MapTypesArray = Info.MapTypesArray; | ||||||
10488 | if (D.hasClausesOfKind<OMPDependClause>()) | ||||||
10489 | CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo); | ||||||
10490 | else | ||||||
10491 | emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen); | ||||||
10492 | }; | ||||||
10493 | |||||||
10494 | if (IfCond) { | ||||||
10495 | emitIfClause(CGF, IfCond, TargetThenGen, | ||||||
10496 | [](CodeGenFunction &CGF, PrePostActionTy &) {}); | ||||||
10497 | } else { | ||||||
10498 | RegionCodeGenTy ThenRCG(TargetThenGen); | ||||||
10499 | ThenRCG(CGF); | ||||||
10500 | } | ||||||
10501 | } | ||||||
10502 | |||||||
10503 | namespace { | ||||||
10504 | /// Kind of parameter in a function with 'declare simd' directive. | ||||||
10505 | enum ParamKindTy { LinearWithVarStride, Linear, Uniform, Vector }; | ||||||
10506 | /// Attribute set of the parameter. | ||||||
10507 | struct ParamAttrTy { | ||||||
10508 | ParamKindTy Kind = Vector; | ||||||
10509 | llvm::APSInt StrideOrArg; | ||||||
10510 | llvm::APSInt Alignment; | ||||||
10511 | }; | ||||||
10512 | } // namespace | ||||||
10513 | |||||||
10514 | static unsigned evaluateCDTSize(const FunctionDecl *FD, | ||||||
10515 | ArrayRef<ParamAttrTy> ParamAttrs) { | ||||||
10516 | // Every vector variant of a SIMD-enabled function has a vector length (VLEN). | ||||||
10517 | // If OpenMP clause "simdlen" is used, the VLEN is the value of the argument | ||||||
10518 | // of that clause. The VLEN value must be power of 2. | ||||||
10519 | // In other case the notion of the function`s "characteristic data type" (CDT) | ||||||
10520 | // is used to compute the vector length. | ||||||
10521 | // CDT is defined in the following order: | ||||||
10522 | // a) For non-void function, the CDT is the return type. | ||||||
10523 | // b) If the function has any non-uniform, non-linear parameters, then the | ||||||
10524 | // CDT is the type of the first such parameter. | ||||||
10525 | // c) If the CDT determined by a) or b) above is struct, union, or class | ||||||
10526 | // type which is pass-by-value (except for the type that maps to the | ||||||
10527 | // built-in complex data type), the characteristic data type is int. | ||||||
10528 | // d) If none of the above three cases is applicable, the CDT is int. | ||||||
10529 | // The VLEN is then determined based on the CDT and the size of vector | ||||||
10530 | // register of that ISA for which current vector version is generated. The | ||||||
10531 | // VLEN is computed using the formula below: | ||||||
10532 | // VLEN = sizeof(vector_register) / sizeof(CDT), | ||||||
10533 | // where vector register size specified in section 3.2.1 Registers and the | ||||||
10534 | // Stack Frame of original AMD64 ABI document. | ||||||
10535 | QualType RetType = FD->getReturnType(); | ||||||
10536 | if (RetType.isNull()) | ||||||
10537 | return 0; | ||||||
10538 | ASTContext &C = FD->getASTContext(); | ||||||
10539 | QualType CDT; | ||||||
10540 | if (!RetType.isNull() && !RetType->isVoidType()) { | ||||||
10541 | CDT = RetType; | ||||||
10542 | } else { | ||||||
10543 | unsigned Offset = 0; | ||||||
10544 | if (const auto *MD = dyn_cast<CXXMethodDecl>(FD)) { | ||||||
10545 | if (ParamAttrs[Offset].Kind == Vector) | ||||||
10546 | CDT = C.getPointerType(C.getRecordType(MD->getParent())); | ||||||
10547 | ++Offset; | ||||||
10548 | } | ||||||
10549 | if (CDT.isNull()) { | ||||||
10550 | for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) { | ||||||
10551 | if (ParamAttrs[I + Offset].Kind == Vector) { | ||||||
10552 | CDT = FD->getParamDecl(I)->getType(); | ||||||
10553 | break; | ||||||
10554 | } | ||||||
10555 | } | ||||||
10556 | } | ||||||
10557 | } | ||||||
10558 | if (CDT.isNull()) | ||||||
10559 | CDT = C.IntTy; | ||||||
10560 | CDT = CDT->getCanonicalTypeUnqualified(); | ||||||
10561 | if (CDT->isRecordType() || CDT->isUnionType()) | ||||||
10562 | CDT = C.IntTy; | ||||||
10563 | return C.getTypeSize(CDT); | ||||||
10564 | } | ||||||
10565 | |||||||
10566 | static void | ||||||
10567 | emitX86DeclareSimdFunction(const FunctionDecl *FD, llvm::Function *Fn, | ||||||
10568 | const llvm::APSInt &VLENVal, | ||||||
10569 | ArrayRef<ParamAttrTy> ParamAttrs, | ||||||
10570 | OMPDeclareSimdDeclAttr::BranchStateTy State) { | ||||||
10571 | struct ISADataTy { | ||||||
10572 | char ISA; | ||||||
10573 | unsigned VecRegSize; | ||||||
10574 | }; | ||||||
10575 | ISADataTy ISAData[] = { | ||||||
10576 | { | ||||||
10577 | 'b', 128 | ||||||
10578 | }, // SSE | ||||||
10579 | { | ||||||
10580 | 'c', 256 | ||||||
10581 | }, // AVX | ||||||
10582 | { | ||||||
10583 | 'd', 256 | ||||||
10584 | }, // AVX2 | ||||||
10585 | { | ||||||
10586 | 'e', 512 | ||||||
10587 | }, // AVX512 | ||||||
10588 | }; | ||||||
10589 | llvm::SmallVector<char, 2> Masked; | ||||||
10590 | switch (State) { | ||||||
10591 | case OMPDeclareSimdDeclAttr::BS_Undefined: | ||||||
10592 | Masked.push_back('N'); | ||||||
10593 | Masked.push_back('M'); | ||||||
10594 | break; | ||||||
10595 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | ||||||
10596 | Masked.push_back('N'); | ||||||
10597 | break; | ||||||
10598 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | ||||||
10599 | Masked.push_back('M'); | ||||||
10600 | break; | ||||||
10601 | } | ||||||
10602 | for (char Mask : Masked) { | ||||||
10603 | for (const ISADataTy &Data : ISAData) { | ||||||
10604 | SmallString<256> Buffer; | ||||||
10605 | llvm::raw_svector_ostream Out(Buffer); | ||||||
10606 | Out << "_ZGV" << Data.ISA << Mask; | ||||||
10607 | if (!VLENVal) { | ||||||
10608 | unsigned NumElts = evaluateCDTSize(FD, ParamAttrs); | ||||||
10609 | assert(NumElts && "Non-zero simdlen/cdtsize expected")((NumElts && "Non-zero simdlen/cdtsize expected") ? static_cast <void> (0) : __assert_fail ("NumElts && \"Non-zero simdlen/cdtsize expected\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10609, __PRETTY_FUNCTION__)); | ||||||
10610 | Out << llvm::APSInt::getUnsigned(Data.VecRegSize / NumElts); | ||||||
10611 | } else { | ||||||
10612 | Out << VLENVal; | ||||||
10613 | } | ||||||
10614 | for (const ParamAttrTy &ParamAttr : ParamAttrs) { | ||||||
10615 | switch (ParamAttr.Kind){ | ||||||
10616 | case LinearWithVarStride: | ||||||
10617 | Out << 's' << ParamAttr.StrideOrArg; | ||||||
10618 | break; | ||||||
10619 | case Linear: | ||||||
10620 | Out << 'l'; | ||||||
10621 | if (ParamAttr.StrideOrArg != 1) | ||||||
10622 | Out << ParamAttr.StrideOrArg; | ||||||
10623 | break; | ||||||
10624 | case Uniform: | ||||||
10625 | Out << 'u'; | ||||||
10626 | break; | ||||||
10627 | case Vector: | ||||||
10628 | Out << 'v'; | ||||||
10629 | break; | ||||||
10630 | } | ||||||
10631 | if (!!ParamAttr.Alignment) | ||||||
10632 | Out << 'a' << ParamAttr.Alignment; | ||||||
10633 | } | ||||||
10634 | Out << '_' << Fn->getName(); | ||||||
10635 | Fn->addFnAttr(Out.str()); | ||||||
10636 | } | ||||||
10637 | } | ||||||
10638 | } | ||||||
10639 | |||||||
10640 | // This are the Functions that are needed to mangle the name of the | ||||||
10641 | // vector functions generated by the compiler, according to the rules | ||||||
10642 | // defined in the "Vector Function ABI specifications for AArch64", | ||||||
10643 | // available at | ||||||
10644 | // https://developer.arm.com/products/software-development-tools/hpc/arm-compiler-for-hpc/vector-function-abi. | ||||||
10645 | |||||||
10646 | /// Maps To Vector (MTV), as defined in 3.1.1 of the AAVFABI. | ||||||
10647 | /// | ||||||
10648 | /// TODO: Need to implement the behavior for reference marked with a | ||||||
10649 | /// var or no linear modifiers (1.b in the section). For this, we | ||||||
10650 | /// need to extend ParamKindTy to support the linear modifiers. | ||||||
10651 | static bool getAArch64MTV(QualType QT, ParamKindTy Kind) { | ||||||
10652 | QT = QT.getCanonicalType(); | ||||||
10653 | |||||||
10654 | if (QT->isVoidType()) | ||||||
10655 | return false; | ||||||
10656 | |||||||
10657 | if (Kind == ParamKindTy::Uniform) | ||||||
10658 | return false; | ||||||
10659 | |||||||
10660 | if (Kind == ParamKindTy::Linear) | ||||||
10661 | return false; | ||||||
10662 | |||||||
10663 | // TODO: Handle linear references with modifiers | ||||||
10664 | |||||||
10665 | if (Kind == ParamKindTy::LinearWithVarStride) | ||||||
10666 | return false; | ||||||
10667 | |||||||
10668 | return true; | ||||||
10669 | } | ||||||
10670 | |||||||
10671 | /// Pass By Value (PBV), as defined in 3.1.2 of the AAVFABI. | ||||||
10672 | static bool getAArch64PBV(QualType QT, ASTContext &C) { | ||||||
10673 | QT = QT.getCanonicalType(); | ||||||
10674 | unsigned Size = C.getTypeSize(QT); | ||||||
10675 | |||||||
10676 | // Only scalars and complex within 16 bytes wide set PVB to true. | ||||||
10677 | if (Size != 8 && Size != 16 && Size != 32 && Size != 64 && Size != 128) | ||||||
10678 | return false; | ||||||
10679 | |||||||
10680 | if (QT->isFloatingType()) | ||||||
10681 | return true; | ||||||
10682 | |||||||
10683 | if (QT->isIntegerType()) | ||||||
10684 | return true; | ||||||
10685 | |||||||
10686 | if (QT->isPointerType()) | ||||||
10687 | return true; | ||||||
10688 | |||||||
10689 | // TODO: Add support for complex types (section 3.1.2, item 2). | ||||||
10690 | |||||||
10691 | return false; | ||||||
10692 | } | ||||||
10693 | |||||||
10694 | /// Computes the lane size (LS) of a return type or of an input parameter, | ||||||
10695 | /// as defined by `LS(P)` in 3.2.1 of the AAVFABI. | ||||||
10696 | /// TODO: Add support for references, section 3.2.1, item 1. | ||||||
10697 | static unsigned getAArch64LS(QualType QT, ParamKindTy Kind, ASTContext &C) { | ||||||
10698 | if (!getAArch64MTV(QT, Kind) && QT.getCanonicalType()->isPointerType()) { | ||||||
10699 | QualType PTy = QT.getCanonicalType()->getPointeeType(); | ||||||
10700 | if (getAArch64PBV(PTy, C)) | ||||||
10701 | return C.getTypeSize(PTy); | ||||||
10702 | } | ||||||
10703 | if (getAArch64PBV(QT, C)) | ||||||
10704 | return C.getTypeSize(QT); | ||||||
10705 | |||||||
10706 | return C.getTypeSize(C.getUIntPtrType()); | ||||||
10707 | } | ||||||
10708 | |||||||
10709 | // Get Narrowest Data Size (NDS) and Widest Data Size (WDS) from the | ||||||
10710 | // signature of the scalar function, as defined in 3.2.2 of the | ||||||
10711 | // AAVFABI. | ||||||
10712 | static std::tuple<unsigned, unsigned, bool> | ||||||
10713 | getNDSWDS(const FunctionDecl *FD, ArrayRef<ParamAttrTy> ParamAttrs) { | ||||||
10714 | QualType RetType = FD->getReturnType().getCanonicalType(); | ||||||
10715 | |||||||
10716 | ASTContext &C = FD->getASTContext(); | ||||||
10717 | |||||||
10718 | bool OutputBecomesInput = false; | ||||||
10719 | |||||||
10720 | llvm::SmallVector<unsigned, 8> Sizes; | ||||||
10721 | if (!RetType->isVoidType()) { | ||||||
10722 | Sizes.push_back(getAArch64LS(RetType, ParamKindTy::Vector, C)); | ||||||
10723 | if (!getAArch64PBV(RetType, C) && getAArch64MTV(RetType, {})) | ||||||
10724 | OutputBecomesInput = true; | ||||||
10725 | } | ||||||
10726 | for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) { | ||||||
10727 | QualType QT = FD->getParamDecl(I)->getType().getCanonicalType(); | ||||||
10728 | Sizes.push_back(getAArch64LS(QT, ParamAttrs[I].Kind, C)); | ||||||
10729 | } | ||||||
10730 | |||||||
10731 | assert(!Sizes.empty() && "Unable to determine NDS and WDS.")((!Sizes.empty() && "Unable to determine NDS and WDS." ) ? static_cast<void> (0) : __assert_fail ("!Sizes.empty() && \"Unable to determine NDS and WDS.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10731, __PRETTY_FUNCTION__)); | ||||||
10732 | // The LS of a function parameter / return value can only be a power | ||||||
10733 | // of 2, starting from 8 bits, up to 128. | ||||||
10734 | assert(std::all_of(Sizes.begin(), Sizes.end(),((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10739, __PRETTY_FUNCTION__)) | ||||||
10735 | [](unsigned Size) {((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10739, __PRETTY_FUNCTION__)) | ||||||
10736 | return Size == 8 || Size == 16 || Size == 32 ||((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10739, __PRETTY_FUNCTION__)) | ||||||
10737 | Size == 64 || Size == 128;((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10739, __PRETTY_FUNCTION__)) | ||||||
10738 | }) &&((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10739, __PRETTY_FUNCTION__)) | ||||||
10739 | "Invalid size")((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10739, __PRETTY_FUNCTION__)); | ||||||
10740 | |||||||
10741 | return std::make_tuple(*std::min_element(std::begin(Sizes), std::end(Sizes)), | ||||||
10742 | *std::max_element(std::begin(Sizes), std::end(Sizes)), | ||||||
10743 | OutputBecomesInput); | ||||||
10744 | } | ||||||
10745 | |||||||
10746 | /// Mangle the parameter part of the vector function name according to | ||||||
10747 | /// their OpenMP classification. The mangling function is defined in | ||||||
10748 | /// section 3.5 of the AAVFABI. | ||||||
10749 | static std::string mangleVectorParameters(ArrayRef<ParamAttrTy> ParamAttrs) { | ||||||
10750 | SmallString<256> Buffer; | ||||||
10751 | llvm::raw_svector_ostream Out(Buffer); | ||||||
10752 | for (const auto &ParamAttr : ParamAttrs) { | ||||||
10753 | switch (ParamAttr.Kind) { | ||||||
10754 | case LinearWithVarStride: | ||||||
10755 | Out << "ls" << ParamAttr.StrideOrArg; | ||||||
10756 | break; | ||||||
10757 | case Linear: | ||||||
10758 | Out << 'l'; | ||||||
10759 | // Don't print the step value if it is not present or if it is | ||||||
10760 | // equal to 1. | ||||||
10761 | if (ParamAttr.StrideOrArg != 1) | ||||||
10762 | Out << ParamAttr.StrideOrArg; | ||||||
10763 | break; | ||||||
10764 | case Uniform: | ||||||
10765 | Out << 'u'; | ||||||
10766 | break; | ||||||
10767 | case Vector: | ||||||
10768 | Out << 'v'; | ||||||
10769 | break; | ||||||
10770 | } | ||||||
10771 | |||||||
10772 | if (!!ParamAttr.Alignment) | ||||||
10773 | Out << 'a' << ParamAttr.Alignment; | ||||||
10774 | } | ||||||
10775 | |||||||
10776 | return std::string(Out.str()); | ||||||
10777 | } | ||||||
10778 | |||||||
10779 | // Function used to add the attribute. The parameter `VLEN` is | ||||||
10780 | // templated to allow the use of "x" when targeting scalable functions | ||||||
10781 | // for SVE. | ||||||
10782 | template <typename T> | ||||||
10783 | static void addAArch64VectorName(T VLEN, StringRef LMask, StringRef Prefix, | ||||||
10784 | char ISA, StringRef ParSeq, | ||||||
10785 | StringRef MangledName, bool OutputBecomesInput, | ||||||
10786 | llvm::Function *Fn) { | ||||||
10787 | SmallString<256> Buffer; | ||||||
10788 | llvm::raw_svector_ostream Out(Buffer); | ||||||
10789 | Out << Prefix << ISA << LMask << VLEN; | ||||||
10790 | if (OutputBecomesInput) | ||||||
10791 | Out << "v"; | ||||||
10792 | Out << ParSeq << "_" << MangledName; | ||||||
10793 | Fn->addFnAttr(Out.str()); | ||||||
10794 | } | ||||||
10795 | |||||||
10796 | // Helper function to generate the Advanced SIMD names depending on | ||||||
10797 | // the value of the NDS when simdlen is not present. | ||||||
10798 | static void addAArch64AdvSIMDNDSNames(unsigned NDS, StringRef Mask, | ||||||
10799 | StringRef Prefix, char ISA, | ||||||
10800 | StringRef ParSeq, StringRef MangledName, | ||||||
10801 | bool OutputBecomesInput, | ||||||
10802 | llvm::Function *Fn) { | ||||||
10803 | switch (NDS) { | ||||||
10804 | case 8: | ||||||
10805 | addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10806 | OutputBecomesInput, Fn); | ||||||
10807 | addAArch64VectorName(16, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10808 | OutputBecomesInput, Fn); | ||||||
10809 | break; | ||||||
10810 | case 16: | ||||||
10811 | addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10812 | OutputBecomesInput, Fn); | ||||||
10813 | addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10814 | OutputBecomesInput, Fn); | ||||||
10815 | break; | ||||||
10816 | case 32: | ||||||
10817 | addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10818 | OutputBecomesInput, Fn); | ||||||
10819 | addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10820 | OutputBecomesInput, Fn); | ||||||
10821 | break; | ||||||
10822 | case 64: | ||||||
10823 | case 128: | ||||||
10824 | addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName, | ||||||
10825 | OutputBecomesInput, Fn); | ||||||
10826 | break; | ||||||
10827 | default: | ||||||
10828 | llvm_unreachable("Scalar type is too wide.")::llvm::llvm_unreachable_internal("Scalar type is too wide.", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10828); | ||||||
10829 | } | ||||||
10830 | } | ||||||
10831 | |||||||
10832 | /// Emit vector function attributes for AArch64, as defined in the AAVFABI. | ||||||
10833 | static void emitAArch64DeclareSimdFunction( | ||||||
10834 | CodeGenModule &CGM, const FunctionDecl *FD, unsigned UserVLEN, | ||||||
10835 | ArrayRef<ParamAttrTy> ParamAttrs, | ||||||
10836 | OMPDeclareSimdDeclAttr::BranchStateTy State, StringRef MangledName, | ||||||
10837 | char ISA, unsigned VecRegSize, llvm::Function *Fn, SourceLocation SLoc) { | ||||||
10838 | |||||||
10839 | // Get basic data for building the vector signature. | ||||||
10840 | const auto Data = getNDSWDS(FD, ParamAttrs); | ||||||
10841 | const unsigned NDS = std::get<0>(Data); | ||||||
10842 | const unsigned WDS = std::get<1>(Data); | ||||||
10843 | const bool OutputBecomesInput = std::get<2>(Data); | ||||||
10844 | |||||||
10845 | // Check the values provided via `simdlen` by the user. | ||||||
10846 | // 1. A `simdlen(1)` doesn't produce vector signatures, | ||||||
10847 | if (UserVLEN == 1) { | ||||||
10848 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
10849 | DiagnosticsEngine::Warning, | ||||||
10850 | "The clause simdlen(1) has no effect when targeting aarch64."); | ||||||
10851 | CGM.getDiags().Report(SLoc, DiagID); | ||||||
10852 | return; | ||||||
10853 | } | ||||||
10854 | |||||||
10855 | // 2. Section 3.3.1, item 1: user input must be a power of 2 for | ||||||
10856 | // Advanced SIMD output. | ||||||
10857 | if (ISA == 'n' && UserVLEN && !llvm::isPowerOf2_32(UserVLEN)) { | ||||||
10858 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
10859 | DiagnosticsEngine::Warning, "The value specified in simdlen must be a " | ||||||
10860 | "power of 2 when targeting Advanced SIMD."); | ||||||
10861 | CGM.getDiags().Report(SLoc, DiagID); | ||||||
10862 | return; | ||||||
10863 | } | ||||||
10864 | |||||||
10865 | // 3. Section 3.4.1. SVE fixed lengh must obey the architectural | ||||||
10866 | // limits. | ||||||
10867 | if (ISA == 's' && UserVLEN != 0) { | ||||||
10868 | if ((UserVLEN * WDS > 2048) || (UserVLEN * WDS % 128 != 0)) { | ||||||
10869 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | ||||||
10870 | DiagnosticsEngine::Warning, "The clause simdlen must fit the %0-bit " | ||||||
10871 | "lanes in the architectural constraints " | ||||||
10872 | "for SVE (min is 128-bit, max is " | ||||||
10873 | "2048-bit, by steps of 128-bit)"); | ||||||
10874 | CGM.getDiags().Report(SLoc, DiagID) << WDS; | ||||||
10875 | return; | ||||||
10876 | } | ||||||
10877 | } | ||||||
10878 | |||||||
10879 | // Sort out parameter sequence. | ||||||
10880 | const std::string ParSeq = mangleVectorParameters(ParamAttrs); | ||||||
10881 | StringRef Prefix = "_ZGV"; | ||||||
10882 | // Generate simdlen from user input (if any). | ||||||
10883 | if (UserVLEN) { | ||||||
10884 | if (ISA == 's') { | ||||||
10885 | // SVE generates only a masked function. | ||||||
10886 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | ||||||
10887 | OutputBecomesInput, Fn); | ||||||
10888 | } else { | ||||||
10889 | assert(ISA == 'n' && "Expected ISA either 's' or 'n'.")((ISA == 'n' && "Expected ISA either 's' or 'n'.") ? static_cast <void> (0) : __assert_fail ("ISA == 'n' && \"Expected ISA either 's' or 'n'.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10889, __PRETTY_FUNCTION__)); | ||||||
10890 | // Advanced SIMD generates one or two functions, depending on | ||||||
10891 | // the `[not]inbranch` clause. | ||||||
10892 | switch (State) { | ||||||
10893 | case OMPDeclareSimdDeclAttr::BS_Undefined: | ||||||
10894 | addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName, | ||||||
10895 | OutputBecomesInput, Fn); | ||||||
10896 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | ||||||
10897 | OutputBecomesInput, Fn); | ||||||
10898 | break; | ||||||
10899 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | ||||||
10900 | addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName, | ||||||
10901 | OutputBecomesInput, Fn); | ||||||
10902 | break; | ||||||
10903 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | ||||||
10904 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | ||||||
10905 | OutputBecomesInput, Fn); | ||||||
10906 | break; | ||||||
10907 | } | ||||||
10908 | } | ||||||
10909 | } else { | ||||||
10910 | // If no user simdlen is provided, follow the AAVFABI rules for | ||||||
10911 | // generating the vector length. | ||||||
10912 | if (ISA == 's') { | ||||||
10913 | // SVE, section 3.4.1, item 1. | ||||||
10914 | addAArch64VectorName("x", "M", Prefix, ISA, ParSeq, MangledName, | ||||||
10915 | OutputBecomesInput, Fn); | ||||||
10916 | } else { | ||||||
10917 | assert(ISA == 'n' && "Expected ISA either 's' or 'n'.")((ISA == 'n' && "Expected ISA either 's' or 'n'.") ? static_cast <void> (0) : __assert_fail ("ISA == 'n' && \"Expected ISA either 's' or 'n'.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10917, __PRETTY_FUNCTION__)); | ||||||
10918 | // Advanced SIMD, Section 3.3.1 of the AAVFABI, generates one or | ||||||
10919 | // two vector names depending on the use of the clause | ||||||
10920 | // `[not]inbranch`. | ||||||
10921 | switch (State) { | ||||||
10922 | case OMPDeclareSimdDeclAttr::BS_Undefined: | ||||||
10923 | addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName, | ||||||
10924 | OutputBecomesInput, Fn); | ||||||
10925 | addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName, | ||||||
10926 | OutputBecomesInput, Fn); | ||||||
10927 | break; | ||||||
10928 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | ||||||
10929 | addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName, | ||||||
10930 | OutputBecomesInput, Fn); | ||||||
10931 | break; | ||||||
10932 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | ||||||
10933 | addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName, | ||||||
10934 | OutputBecomesInput, Fn); | ||||||
10935 | break; | ||||||
10936 | } | ||||||
10937 | } | ||||||
10938 | } | ||||||
10939 | } | ||||||
10940 | |||||||
10941 | void CGOpenMPRuntime::emitDeclareSimdFunction(const FunctionDecl *FD, | ||||||
10942 | llvm::Function *Fn) { | ||||||
10943 | ASTContext &C = CGM.getContext(); | ||||||
10944 | FD = FD->getMostRecentDecl(); | ||||||
10945 | // Map params to their positions in function decl. | ||||||
10946 | llvm::DenseMap<const Decl *, unsigned> ParamPositions; | ||||||
10947 | if (isa<CXXMethodDecl>(FD)) | ||||||
10948 | ParamPositions.try_emplace(FD, 0); | ||||||
10949 | unsigned ParamPos = ParamPositions.size(); | ||||||
10950 | for (const ParmVarDecl *P : FD->parameters()) { | ||||||
10951 | ParamPositions.try_emplace(P->getCanonicalDecl(), ParamPos); | ||||||
10952 | ++ParamPos; | ||||||
10953 | } | ||||||
10954 | while (FD) { | ||||||
10955 | for (const auto *Attr : FD->specific_attrs<OMPDeclareSimdDeclAttr>()) { | ||||||
10956 | llvm::SmallVector<ParamAttrTy, 8> ParamAttrs(ParamPositions.size()); | ||||||
10957 | // Mark uniform parameters. | ||||||
10958 | for (const Expr *E : Attr->uniforms()) { | ||||||
10959 | E = E->IgnoreParenImpCasts(); | ||||||
10960 | unsigned Pos; | ||||||
10961 | if (isa<CXXThisExpr>(E)) { | ||||||
10962 | Pos = ParamPositions[FD]; | ||||||
10963 | } else { | ||||||
10964 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | ||||||
10965 | ->getCanonicalDecl(); | ||||||
10966 | Pos = ParamPositions[PVD]; | ||||||
10967 | } | ||||||
10968 | ParamAttrs[Pos].Kind = Uniform; | ||||||
10969 | } | ||||||
10970 | // Get alignment info. | ||||||
10971 | auto NI = Attr->alignments_begin(); | ||||||
10972 | for (const Expr *E : Attr->aligneds()) { | ||||||
10973 | E = E->IgnoreParenImpCasts(); | ||||||
10974 | unsigned Pos; | ||||||
10975 | QualType ParmTy; | ||||||
10976 | if (isa<CXXThisExpr>(E)) { | ||||||
10977 | Pos = ParamPositions[FD]; | ||||||
10978 | ParmTy = E->getType(); | ||||||
10979 | } else { | ||||||
10980 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | ||||||
10981 | ->getCanonicalDecl(); | ||||||
10982 | Pos = ParamPositions[PVD]; | ||||||
10983 | ParmTy = PVD->getType(); | ||||||
10984 | } | ||||||
10985 | ParamAttrs[Pos].Alignment = | ||||||
10986 | (*NI) | ||||||
10987 | ? (*NI)->EvaluateKnownConstInt(C) | ||||||
10988 | : llvm::APSInt::getUnsigned( | ||||||
10989 | C.toCharUnitsFromBits(C.getOpenMPDefaultSimdAlign(ParmTy)) | ||||||
10990 | .getQuantity()); | ||||||
10991 | ++NI; | ||||||
10992 | } | ||||||
10993 | // Mark linear parameters. | ||||||
10994 | auto SI = Attr->steps_begin(); | ||||||
10995 | auto MI = Attr->modifiers_begin(); | ||||||
10996 | for (const Expr *E : Attr->linears()) { | ||||||
10997 | E = E->IgnoreParenImpCasts(); | ||||||
10998 | unsigned Pos; | ||||||
10999 | // Rescaling factor needed to compute the linear parameter | ||||||
11000 | // value in the mangled name. | ||||||
11001 | unsigned PtrRescalingFactor = 1; | ||||||
11002 | if (isa<CXXThisExpr>(E)) { | ||||||
11003 | Pos = ParamPositions[FD]; | ||||||
11004 | } else { | ||||||
11005 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | ||||||
11006 | ->getCanonicalDecl(); | ||||||
11007 | Pos = ParamPositions[PVD]; | ||||||
11008 | if (auto *P = dyn_cast<PointerType>(PVD->getType())) | ||||||
11009 | PtrRescalingFactor = CGM.getContext() | ||||||
11010 | .getTypeSizeInChars(P->getPointeeType()) | ||||||
11011 | .getQuantity(); | ||||||
11012 | } | ||||||
11013 | ParamAttrTy &ParamAttr = ParamAttrs[Pos]; | ||||||
11014 | ParamAttr.Kind = Linear; | ||||||
11015 | // Assuming a stride of 1, for `linear` without modifiers. | ||||||
11016 | ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(1); | ||||||
11017 | if (*SI) { | ||||||
11018 | Expr::EvalResult Result; | ||||||
11019 | if (!(*SI)->EvaluateAsInt(Result, C, Expr::SE_AllowSideEffects)) { | ||||||
11020 | if (const auto *DRE = | ||||||
11021 | cast<DeclRefExpr>((*SI)->IgnoreParenImpCasts())) { | ||||||
11022 | if (const auto *StridePVD = cast<ParmVarDecl>(DRE->getDecl())) { | ||||||
11023 | ParamAttr.Kind = LinearWithVarStride; | ||||||
11024 | ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned( | ||||||
11025 | ParamPositions[StridePVD->getCanonicalDecl()]); | ||||||
11026 | } | ||||||
11027 | } | ||||||
11028 | } else { | ||||||
11029 | ParamAttr.StrideOrArg = Result.Val.getInt(); | ||||||
11030 | } | ||||||
11031 | } | ||||||
11032 | // If we are using a linear clause on a pointer, we need to | ||||||
11033 | // rescale the value of linear_step with the byte size of the | ||||||
11034 | // pointee type. | ||||||
11035 | if (Linear == ParamAttr.Kind) | ||||||
11036 | ParamAttr.StrideOrArg = ParamAttr.StrideOrArg * PtrRescalingFactor; | ||||||
11037 | ++SI; | ||||||
11038 | ++MI; | ||||||
11039 | } | ||||||
11040 | llvm::APSInt VLENVal; | ||||||
11041 | SourceLocation ExprLoc; | ||||||
11042 | const Expr *VLENExpr = Attr->getSimdlen(); | ||||||
11043 | if (VLENExpr) { | ||||||
11044 | VLENVal = VLENExpr->EvaluateKnownConstInt(C); | ||||||
11045 | ExprLoc = VLENExpr->getExprLoc(); | ||||||
11046 | } | ||||||
11047 | OMPDeclareSimdDeclAttr::BranchStateTy State = Attr->getBranchState(); | ||||||
11048 | if (CGM.getTriple().isX86()) { | ||||||
11049 | emitX86DeclareSimdFunction(FD, Fn, VLENVal, ParamAttrs, State); | ||||||
11050 | } else if (CGM.getTriple().getArch() == llvm::Triple::aarch64) { | ||||||
11051 | unsigned VLEN = VLENVal.getExtValue(); | ||||||
11052 | StringRef MangledName = Fn->getName(); | ||||||
11053 | if (CGM.getTarget().hasFeature("sve")) | ||||||
11054 | emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State, | ||||||
11055 | MangledName, 's', 128, Fn, ExprLoc); | ||||||
11056 | if (CGM.getTarget().hasFeature("neon")) | ||||||
11057 | emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State, | ||||||
11058 | MangledName, 'n', 128, Fn, ExprLoc); | ||||||
11059 | } | ||||||
11060 | } | ||||||
11061 | FD = FD->getPreviousDecl(); | ||||||
11062 | } | ||||||
11063 | } | ||||||
11064 | |||||||
11065 | namespace { | ||||||
11066 | /// Cleanup action for doacross support. | ||||||
11067 | class DoacrossCleanupTy final : public EHScopeStack::Cleanup { | ||||||
11068 | public: | ||||||
11069 | static const int DoacrossFinArgs = 2; | ||||||
11070 | |||||||
11071 | private: | ||||||
11072 | llvm::FunctionCallee RTLFn; | ||||||
11073 | llvm::Value *Args[DoacrossFinArgs]; | ||||||
11074 | |||||||
11075 | public: | ||||||
11076 | DoacrossCleanupTy(llvm::FunctionCallee RTLFn, | ||||||
11077 | ArrayRef<llvm::Value *> CallArgs) | ||||||
11078 | : RTLFn(RTLFn) { | ||||||
11079 | assert(CallArgs.size() == DoacrossFinArgs)((CallArgs.size() == DoacrossFinArgs) ? static_cast<void> (0) : __assert_fail ("CallArgs.size() == DoacrossFinArgs", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11079, __PRETTY_FUNCTION__)); | ||||||
11080 | std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args)); | ||||||
11081 | } | ||||||
11082 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | ||||||
11083 | if (!CGF.HaveInsertPoint()) | ||||||
11084 | return; | ||||||
11085 | CGF.EmitRuntimeCall(RTLFn, Args); | ||||||
11086 | } | ||||||
11087 | }; | ||||||
11088 | } // namespace | ||||||
11089 | |||||||
11090 | void CGOpenMPRuntime::emitDoacrossInit(CodeGenFunction &CGF, | ||||||
11091 | const OMPLoopDirective &D, | ||||||
11092 | ArrayRef<Expr *> NumIterations) { | ||||||
11093 | if (!CGF.HaveInsertPoint()) | ||||||
11094 | return; | ||||||
11095 | |||||||
11096 | ASTContext &C = CGM.getContext(); | ||||||
11097 | QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true); | ||||||
11098 | RecordDecl *RD; | ||||||
11099 | if (KmpDimTy.isNull()) { | ||||||
11100 | // Build struct kmp_dim { // loop bounds info casted to kmp_int64 | ||||||
11101 | // kmp_int64 lo; // lower | ||||||
11102 | // kmp_int64 up; // upper | ||||||
11103 | // kmp_int64 st; // stride | ||||||
11104 | // }; | ||||||
11105 | RD = C.buildImplicitRecord("kmp_dim"); | ||||||
11106 | RD->startDefinition(); | ||||||
11107 | addFieldToRecordDecl(C, RD, Int64Ty); | ||||||
11108 | addFieldToRecordDecl(C, RD, Int64Ty); | ||||||
11109 | addFieldToRecordDecl(C, RD, Int64Ty); | ||||||
11110 | RD->completeDefinition(); | ||||||
11111 | KmpDimTy = C.getRecordType(RD); | ||||||
11112 | } else { | ||||||
11113 | RD = cast<RecordDecl>(KmpDimTy->getAsTagDecl()); | ||||||
11114 | } | ||||||
11115 | llvm::APInt Size(/*numBits=*/32, NumIterations.size()); | ||||||
11116 | QualType ArrayTy = | ||||||
11117 | C.getConstantArrayType(KmpDimTy, Size, nullptr, ArrayType::Normal, 0); | ||||||
11118 | |||||||
11119 | Address DimsAddr = CGF.CreateMemTemp(ArrayTy, "dims"); | ||||||
11120 | CGF.EmitNullInitialization(DimsAddr, ArrayTy); | ||||||
11121 | enum { LowerFD = 0, UpperFD, StrideFD }; | ||||||
11122 | // Fill dims with data. | ||||||
11123 | for (unsigned I = 0, E = NumIterations.size(); I < E; ++I) { | ||||||
11124 | LValue DimsLVal = CGF.MakeAddrLValue( | ||||||
11125 | CGF.Builder.CreateConstArrayGEP(DimsAddr, I), KmpDimTy); | ||||||
11126 | // dims.upper = num_iterations; | ||||||
11127 | LValue UpperLVal = CGF.EmitLValueForField( | ||||||
11128 | DimsLVal, *std::next(RD->field_begin(), UpperFD)); | ||||||
11129 | llvm::Value *NumIterVal = CGF.EmitScalarConversion( | ||||||
11130 | CGF.EmitScalarExpr(NumIterations[I]), NumIterations[I]->getType(), | ||||||
11131 | Int64Ty, NumIterations[I]->getExprLoc()); | ||||||
11132 | CGF.EmitStoreOfScalar(NumIterVal, UpperLVal); | ||||||
11133 | // dims.stride = 1; | ||||||
11134 | LValue StrideLVal = CGF.EmitLValueForField( | ||||||
11135 | DimsLVal, *std::next(RD->field_begin(), StrideFD)); | ||||||
11136 | CGF.EmitStoreOfScalar(llvm::ConstantInt::getSigned(CGM.Int64Ty, /*V=*/1), | ||||||
11137 | StrideLVal); | ||||||
11138 | } | ||||||
11139 | |||||||
11140 | // Build call void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, | ||||||
11141 | // kmp_int32 num_dims, struct kmp_dim * dims); | ||||||
11142 | llvm::Value *Args[] = { | ||||||
11143 | emitUpdateLocation(CGF, D.getBeginLoc()), | ||||||
11144 | getThreadID(CGF, D.getBeginLoc()), | ||||||
11145 | llvm::ConstantInt::getSigned(CGM.Int32Ty, NumIterations.size()), | ||||||
11146 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
11147 | CGF.Builder.CreateConstArrayGEP(DimsAddr, 0).getPointer(), | ||||||
11148 | CGM.VoidPtrTy)}; | ||||||
11149 | |||||||
11150 | llvm::FunctionCallee RTLFn = OMPBuilder.getOrCreateRuntimeFunction( | ||||||
11151 | CGM.getModule(), OMPRTL___kmpc_doacross_init); | ||||||
11152 | CGF.EmitRuntimeCall(RTLFn, Args); | ||||||
11153 | llvm::Value *FiniArgs[DoacrossCleanupTy::DoacrossFinArgs] = { | ||||||
11154 | emitUpdateLocation(CGF, D.getEndLoc()), getThreadID(CGF, D.getEndLoc())}; | ||||||
11155 | llvm::FunctionCallee FiniRTLFn = OMPBuilder.getOrCreateRuntimeFunction( | ||||||
11156 | CGM.getModule(), OMPRTL___kmpc_doacross_fini); | ||||||
11157 | CGF.EHStack.pushCleanup<DoacrossCleanupTy>(NormalAndEHCleanup, FiniRTLFn, | ||||||
11158 | llvm::makeArrayRef(FiniArgs)); | ||||||
11159 | } | ||||||
11160 | |||||||
11161 | void CGOpenMPRuntime::emitDoacrossOrdered(CodeGenFunction &CGF, | ||||||
11162 | const OMPDependClause *C) { | ||||||
11163 | QualType Int64Ty = | ||||||
11164 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | ||||||
11165 | llvm::APInt Size(/*numBits=*/32, C->getNumLoops()); | ||||||
11166 | QualType ArrayTy = CGM.getContext().getConstantArrayType( | ||||||
11167 | Int64Ty, Size, nullptr, ArrayType::Normal, 0); | ||||||
11168 | Address CntAddr = CGF.CreateMemTemp(ArrayTy, ".cnt.addr"); | ||||||
11169 | for (unsigned I = 0, E = C->getNumLoops(); I < E; ++I) { | ||||||
11170 | const Expr *CounterVal = C->getLoopData(I); | ||||||
11171 | assert(CounterVal)((CounterVal) ? static_cast<void> (0) : __assert_fail ( "CounterVal", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11171, __PRETTY_FUNCTION__)); | ||||||
11172 | llvm::Value *CntVal = CGF.EmitScalarConversion( | ||||||
11173 | CGF.EmitScalarExpr(CounterVal), CounterVal->getType(), Int64Ty, | ||||||
11174 | CounterVal->getExprLoc()); | ||||||
11175 | CGF.EmitStoreOfScalar(CntVal, CGF.Builder.CreateConstArrayGEP(CntAddr, I), | ||||||
11176 | /*Volatile=*/false, Int64Ty); | ||||||
11177 | } | ||||||
11178 | llvm::Value *Args[] = { | ||||||
11179 | emitUpdateLocation(CGF, C->getBeginLoc()), | ||||||
11180 | getThreadID(CGF, C->getBeginLoc()), | ||||||
11181 | CGF.Builder.CreateConstArrayGEP(CntAddr, 0).getPointer()}; | ||||||
11182 | llvm::FunctionCallee RTLFn; | ||||||
11183 | if (C->getDependencyKind() == OMPC_DEPEND_source) { | ||||||
11184 | RTLFn = OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
11185 | OMPRTL___kmpc_doacross_post); | ||||||
11186 | } else { | ||||||
11187 | assert(C->getDependencyKind() == OMPC_DEPEND_sink)((C->getDependencyKind() == OMPC_DEPEND_sink) ? static_cast <void> (0) : __assert_fail ("C->getDependencyKind() == OMPC_DEPEND_sink" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11187, __PRETTY_FUNCTION__)); | ||||||
11188 | RTLFn = OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), | ||||||
11189 | OMPRTL___kmpc_doacross_wait); | ||||||
11190 | } | ||||||
11191 | CGF.EmitRuntimeCall(RTLFn, Args); | ||||||
11192 | } | ||||||
11193 | |||||||
11194 | void CGOpenMPRuntime::emitCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
11195 | llvm::FunctionCallee Callee, | ||||||
11196 | ArrayRef<llvm::Value *> Args) const { | ||||||
11197 | assert(Loc.isValid() && "Outlined function call location must be valid.")((Loc.isValid() && "Outlined function call location must be valid." ) ? static_cast<void> (0) : __assert_fail ("Loc.isValid() && \"Outlined function call location must be valid.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11197, __PRETTY_FUNCTION__)); | ||||||
11198 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | ||||||
11199 | |||||||
11200 | if (auto *Fn = dyn_cast<llvm::Function>(Callee.getCallee())) { | ||||||
11201 | if (Fn->doesNotThrow()) { | ||||||
11202 | CGF.EmitNounwindRuntimeCall(Fn, Args); | ||||||
11203 | return; | ||||||
11204 | } | ||||||
11205 | } | ||||||
11206 | CGF.EmitRuntimeCall(Callee, Args); | ||||||
11207 | } | ||||||
11208 | |||||||
11209 | void CGOpenMPRuntime::emitOutlinedFunctionCall( | ||||||
11210 | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, | ||||||
11211 | ArrayRef<llvm::Value *> Args) const { | ||||||
11212 | emitCall(CGF, Loc, OutlinedFn, Args); | ||||||
11213 | } | ||||||
11214 | |||||||
11215 | void CGOpenMPRuntime::emitFunctionProlog(CodeGenFunction &CGF, const Decl *D) { | ||||||
11216 | if (const auto *FD = dyn_cast<FunctionDecl>(D)) | ||||||
11217 | if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(FD)) | ||||||
11218 | HasEmittedDeclareTargetRegion = true; | ||||||
11219 | } | ||||||
11220 | |||||||
11221 | Address CGOpenMPRuntime::getParameterAddress(CodeGenFunction &CGF, | ||||||
11222 | const VarDecl *NativeParam, | ||||||
11223 | const VarDecl *TargetParam) const { | ||||||
11224 | return CGF.GetAddrOfLocalVar(NativeParam); | ||||||
11225 | } | ||||||
11226 | |||||||
11227 | namespace { | ||||||
11228 | /// Cleanup action for allocate support. | ||||||
11229 | class OMPAllocateCleanupTy final : public EHScopeStack::Cleanup { | ||||||
11230 | public: | ||||||
11231 | static const int CleanupArgs = 3; | ||||||
11232 | |||||||
11233 | private: | ||||||
11234 | llvm::FunctionCallee RTLFn; | ||||||
11235 | llvm::Value *Args[CleanupArgs]; | ||||||
11236 | |||||||
11237 | public: | ||||||
11238 | OMPAllocateCleanupTy(llvm::FunctionCallee RTLFn, | ||||||
11239 | ArrayRef<llvm::Value *> CallArgs) | ||||||
11240 | : RTLFn(RTLFn) { | ||||||
11241 | assert(CallArgs.size() == CleanupArgs &&((CallArgs.size() == CleanupArgs && "Size of arguments does not match." ) ? static_cast<void> (0) : __assert_fail ("CallArgs.size() == CleanupArgs && \"Size of arguments does not match.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11242, __PRETTY_FUNCTION__)) | ||||||
11242 | "Size of arguments does not match.")((CallArgs.size() == CleanupArgs && "Size of arguments does not match." ) ? static_cast<void> (0) : __assert_fail ("CallArgs.size() == CleanupArgs && \"Size of arguments does not match.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11242, __PRETTY_FUNCTION__)); | ||||||
11243 | std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args)); | ||||||
11244 | } | ||||||
11245 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | ||||||
11246 | if (!CGF.HaveInsertPoint()) | ||||||
11247 | return; | ||||||
11248 | CGF.EmitRuntimeCall(RTLFn, Args); | ||||||
11249 | } | ||||||
11250 | }; | ||||||
11251 | } // namespace | ||||||
11252 | |||||||
11253 | Address CGOpenMPRuntime::getAddressOfLocalVariable(CodeGenFunction &CGF, | ||||||
11254 | const VarDecl *VD) { | ||||||
11255 | if (!VD) | ||||||
11256 | return Address::invalid(); | ||||||
11257 | const VarDecl *CVD = VD->getCanonicalDecl(); | ||||||
11258 | if (CVD->hasAttr<OMPAllocateDeclAttr>()) { | ||||||
11259 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); | ||||||
11260 | // Use the default allocation. | ||||||
11261 | if ((AA->getAllocatorType() == OMPAllocateDeclAttr::OMPDefaultMemAlloc || | ||||||
11262 | AA->getAllocatorType() == OMPAllocateDeclAttr::OMPNullMemAlloc) && | ||||||
11263 | !AA->getAllocator()) | ||||||
11264 | return Address::invalid(); | ||||||
11265 | llvm::Value *Size; | ||||||
11266 | CharUnits Align = CGM.getContext().getDeclAlign(CVD); | ||||||
11267 | if (CVD->getType()->isVariablyModifiedType()) { | ||||||
11268 | Size = CGF.getTypeSize(CVD->getType()); | ||||||
11269 | // Align the size: ((size + align - 1) / align) * align | ||||||
11270 | Size = CGF.Builder.CreateNUWAdd( | ||||||
11271 | Size, CGM.getSize(Align - CharUnits::fromQuantity(1))); | ||||||
11272 | Size = CGF.Builder.CreateUDiv(Size, CGM.getSize(Align)); | ||||||
11273 | Size = CGF.Builder.CreateNUWMul(Size, CGM.getSize(Align)); | ||||||
11274 | } else { | ||||||
11275 | CharUnits Sz = CGM.getContext().getTypeSizeInChars(CVD->getType()); | ||||||
11276 | Size = CGM.getSize(Sz.alignTo(Align)); | ||||||
11277 | } | ||||||
11278 | llvm::Value *ThreadID = getThreadID(CGF, CVD->getBeginLoc()); | ||||||
11279 | assert(AA->getAllocator() &&((AA->getAllocator() && "Expected allocator expression for non-default allocator." ) ? static_cast<void> (0) : __assert_fail ("AA->getAllocator() && \"Expected allocator expression for non-default allocator.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11280, __PRETTY_FUNCTION__)) | ||||||
11280 | "Expected allocator expression for non-default allocator.")((AA->getAllocator() && "Expected allocator expression for non-default allocator." ) ? static_cast<void> (0) : __assert_fail ("AA->getAllocator() && \"Expected allocator expression for non-default allocator.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11280, __PRETTY_FUNCTION__)); | ||||||
11281 | llvm::Value *Allocator = CGF.EmitScalarExpr(AA->getAllocator()); | ||||||
11282 | // According to the standard, the original allocator type is a enum | ||||||
11283 | // (integer). Convert to pointer type, if required. | ||||||
11284 | if (Allocator->getType()->isIntegerTy()) | ||||||
11285 | Allocator = CGF.Builder.CreateIntToPtr(Allocator, CGM.VoidPtrTy); | ||||||
11286 | else if (Allocator->getType()->isPointerTy()) | ||||||
11287 | Allocator = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
11288 | Allocator, CGM.VoidPtrTy); | ||||||
11289 | llvm::Value *Args[] = {ThreadID, Size, Allocator}; | ||||||
11290 | |||||||
11291 | llvm::Value *Addr = | ||||||
11292 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( | ||||||
11293 | CGM.getModule(), OMPRTL___kmpc_alloc), | ||||||
11294 | Args, getName({CVD->getName(), ".void.addr"})); | ||||||
11295 | llvm::Value *FiniArgs[OMPAllocateCleanupTy::CleanupArgs] = {ThreadID, Addr, | ||||||
11296 | Allocator}; | ||||||
11297 | llvm::FunctionCallee FiniRTLFn = OMPBuilder.getOrCreateRuntimeFunction( | ||||||
11298 | CGM.getModule(), OMPRTL___kmpc_free); | ||||||
11299 | |||||||
11300 | CGF.EHStack.pushCleanup<OMPAllocateCleanupTy>(NormalAndEHCleanup, FiniRTLFn, | ||||||
11301 | llvm::makeArrayRef(FiniArgs)); | ||||||
11302 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
11303 | Addr, | ||||||
11304 | CGF.ConvertTypeForMem(CGM.getContext().getPointerType(CVD->getType())), | ||||||
11305 | getName({CVD->getName(), ".addr"})); | ||||||
11306 | return Address(Addr, Align); | ||||||
11307 | } | ||||||
11308 | if (UntiedLocalVarsStack.empty()) | ||||||
11309 | return Address::invalid(); | ||||||
11310 | const UntiedLocalVarsAddressesMap &UntiedData = UntiedLocalVarsStack.back(); | ||||||
11311 | auto It = UntiedData.find(VD); | ||||||
11312 | if (It == UntiedData.end()) | ||||||
11313 | return Address::invalid(); | ||||||
11314 | |||||||
11315 | return It->second; | ||||||
11316 | } | ||||||
11317 | |||||||
11318 | CGOpenMPRuntime::NontemporalDeclsRAII::NontemporalDeclsRAII( | ||||||
11319 | CodeGenModule &CGM, const OMPLoopDirective &S) | ||||||
11320 | : CGM(CGM), NeedToPush(S.hasClausesOfKind<OMPNontemporalClause>()) { | ||||||
11321 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")((CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? static_cast<void> (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11321, __PRETTY_FUNCTION__)); | ||||||
11322 | if (!NeedToPush) | ||||||
11323 | return; | ||||||
11324 | NontemporalDeclsSet &DS = | ||||||
11325 | CGM.getOpenMPRuntime().NontemporalDeclsStack.emplace_back(); | ||||||
11326 | for (const auto *C : S.getClausesOfKind<OMPNontemporalClause>()) { | ||||||
11327 | for (const Stmt *Ref : C->private_refs()) { | ||||||
11328 | const auto *SimpleRefExpr = cast<Expr>(Ref)->IgnoreParenImpCasts(); | ||||||
11329 | const ValueDecl *VD; | ||||||
11330 | if (const auto *DRE = dyn_cast<DeclRefExpr>(SimpleRefExpr)) { | ||||||
11331 | VD = DRE->getDecl(); | ||||||
11332 | } else { | ||||||
11333 | const auto *ME = cast<MemberExpr>(SimpleRefExpr); | ||||||
11334 | assert((ME->isImplicitCXXThis() ||(((ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME-> getBase()->IgnoreParenImpCasts())) && "Expected member of current class." ) ? static_cast<void> (0) : __assert_fail ("(ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) && \"Expected member of current class.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11336, __PRETTY_FUNCTION__)) | ||||||
11335 | isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) &&(((ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME-> getBase()->IgnoreParenImpCasts())) && "Expected member of current class." ) ? static_cast<void> (0) : __assert_fail ("(ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) && \"Expected member of current class.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11336, __PRETTY_FUNCTION__)) | ||||||
11336 | "Expected member of current class.")(((ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME-> getBase()->IgnoreParenImpCasts())) && "Expected member of current class." ) ? static_cast<void> (0) : __assert_fail ("(ME->isImplicitCXXThis() || isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) && \"Expected member of current class.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11336, __PRETTY_FUNCTION__)); | ||||||
11337 | VD = ME->getMemberDecl(); | ||||||
11338 | } | ||||||
11339 | DS.insert(VD); | ||||||
11340 | } | ||||||
11341 | } | ||||||
11342 | } | ||||||
11343 | |||||||
11344 | CGOpenMPRuntime::NontemporalDeclsRAII::~NontemporalDeclsRAII() { | ||||||
11345 | if (!NeedToPush) | ||||||
11346 | return; | ||||||
11347 | CGM.getOpenMPRuntime().NontemporalDeclsStack.pop_back(); | ||||||
11348 | } | ||||||
11349 | |||||||
11350 | CGOpenMPRuntime::UntiedTaskLocalDeclsRAII::UntiedTaskLocalDeclsRAII( | ||||||
11351 | CodeGenModule &CGM, | ||||||
11352 | const llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, Address> &LocalVars) | ||||||
11353 | : CGM(CGM), NeedToPush(!LocalVars.empty()) { | ||||||
11354 | if (!NeedToPush) | ||||||
11355 | return; | ||||||
11356 | CGM.getOpenMPRuntime().UntiedLocalVarsStack.push_back(LocalVars); | ||||||
11357 | } | ||||||
11358 | |||||||
11359 | CGOpenMPRuntime::UntiedTaskLocalDeclsRAII::~UntiedTaskLocalDeclsRAII() { | ||||||
11360 | if (!NeedToPush) | ||||||
11361 | return; | ||||||
11362 | CGM.getOpenMPRuntime().UntiedLocalVarsStack.pop_back(); | ||||||
11363 | } | ||||||
11364 | |||||||
11365 | bool CGOpenMPRuntime::isNontemporalDecl(const ValueDecl *VD) const { | ||||||
11366 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")((CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? static_cast<void> (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11366, __PRETTY_FUNCTION__)); | ||||||
11367 | |||||||
11368 | return llvm::any_of( | ||||||
11369 | CGM.getOpenMPRuntime().NontemporalDeclsStack, | ||||||
11370 | [VD](const NontemporalDeclsSet &Set) { return Set.count(VD) > 0; }); | ||||||
11371 | } | ||||||
11372 | |||||||
11373 | void CGOpenMPRuntime::LastprivateConditionalRAII::tryToDisableInnerAnalysis( | ||||||
11374 | const OMPExecutableDirective &S, | ||||||
11375 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> &NeedToAddForLPCsAsDisabled) | ||||||
11376 | const { | ||||||
11377 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> NeedToCheckForLPCs; | ||||||
11378 | // Vars in target/task regions must be excluded completely. | ||||||
11379 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()) || | ||||||
11380 | isOpenMPTaskingDirective(S.getDirectiveKind())) { | ||||||
11381 | SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; | ||||||
11382 | getOpenMPCaptureRegions(CaptureRegions, S.getDirectiveKind()); | ||||||
11383 | const CapturedStmt *CS = S.getCapturedStmt(CaptureRegions.front()); | ||||||
11384 | for (const CapturedStmt::Capture &Cap : CS->captures()) { | ||||||
11385 | if (Cap.capturesVariable() || Cap.capturesVariableByCopy()) | ||||||
11386 | NeedToCheckForLPCs.insert(Cap.getCapturedVar()); | ||||||
11387 | } | ||||||
11388 | } | ||||||
11389 | // Exclude vars in private clauses. | ||||||
11390 | for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) { | ||||||
11391 | for (const Expr *Ref : C->varlists()) { | ||||||
11392 | if (!Ref->getType()->isScalarType()) | ||||||
11393 | continue; | ||||||
11394 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | ||||||
11395 | if (!DRE) | ||||||
11396 | continue; | ||||||
11397 | NeedToCheckForLPCs.insert(DRE->getDecl()); | ||||||
11398 | } | ||||||
11399 | } | ||||||
11400 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { | ||||||
11401 | for (const Expr *Ref : C->varlists()) { | ||||||
11402 | if (!Ref->getType()->isScalarType()) | ||||||
11403 | continue; | ||||||
11404 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | ||||||
11405 | if (!DRE) | ||||||
11406 | continue; | ||||||
11407 | NeedToCheckForLPCs.insert(DRE->getDecl()); | ||||||
11408 | } | ||||||
11409 | } | ||||||
11410 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { | ||||||
11411 | for (const Expr *Ref : C->varlists()) { | ||||||
11412 | if (!Ref->getType()->isScalarType()) | ||||||
11413 | continue; | ||||||
11414 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | ||||||
11415 | if (!DRE) | ||||||
11416 | continue; | ||||||
11417 | NeedToCheckForLPCs.insert(DRE->getDecl()); | ||||||
11418 | } | ||||||
11419 | } | ||||||
11420 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { | ||||||
11421 | for (const Expr *Ref : C->varlists()) { | ||||||
11422 | if (!Ref->getType()->isScalarType()) | ||||||
11423 | continue; | ||||||
11424 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | ||||||
11425 | if (!DRE) | ||||||
11426 | continue; | ||||||
11427 | NeedToCheckForLPCs.insert(DRE->getDecl()); | ||||||
11428 | } | ||||||
11429 | } | ||||||
11430 | for (const auto *C : S.getClausesOfKind<OMPLinearClause>()) { | ||||||
11431 | for (const Expr *Ref : C->varlists()) { | ||||||
11432 | if (!Ref->getType()->isScalarType()) | ||||||
11433 | continue; | ||||||
11434 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); | ||||||
11435 | if (!DRE) | ||||||
11436 | continue; | ||||||
11437 | NeedToCheckForLPCs.insert(DRE->getDecl()); | ||||||
11438 | } | ||||||
11439 | } | ||||||
11440 | for (const Decl *VD : NeedToCheckForLPCs) { | ||||||
11441 | for (const LastprivateConditionalData &Data : | ||||||
11442 | llvm::reverse(CGM.getOpenMPRuntime().LastprivateConditionalStack)) { | ||||||
11443 | if (Data.DeclToUniqueName.count(VD) > 0) { | ||||||
11444 | if (!Data.Disabled) | ||||||
11445 | NeedToAddForLPCsAsDisabled.insert(VD); | ||||||
11446 | break; | ||||||
11447 | } | ||||||
11448 | } | ||||||
11449 | } | ||||||
11450 | } | ||||||
11451 | |||||||
11452 | CGOpenMPRuntime::LastprivateConditionalRAII::LastprivateConditionalRAII( | ||||||
11453 | CodeGenFunction &CGF, const OMPExecutableDirective &S, LValue IVLVal) | ||||||
11454 | : CGM(CGF.CGM), | ||||||
11455 | Action((CGM.getLangOpts().OpenMP >= 50 && | ||||||
11456 | llvm::any_of(S.getClausesOfKind<OMPLastprivateClause>(), | ||||||
11457 | [](const OMPLastprivateClause *C) { | ||||||
11458 | return C->getKind() == | ||||||
11459 | OMPC_LASTPRIVATE_conditional; | ||||||
11460 | })) | ||||||
11461 | ? ActionToDo::PushAsLastprivateConditional | ||||||
11462 | : ActionToDo::DoNotPush) { | ||||||
11463 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")((CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? static_cast<void> (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11463, __PRETTY_FUNCTION__)); | ||||||
11464 | if (CGM.getLangOpts().OpenMP < 50 || Action == ActionToDo::DoNotPush) | ||||||
11465 | return; | ||||||
11466 | assert(Action == ActionToDo::PushAsLastprivateConditional &&((Action == ActionToDo::PushAsLastprivateConditional && "Expected a push action.") ? static_cast<void> (0) : __assert_fail ("Action == ActionToDo::PushAsLastprivateConditional && \"Expected a push action.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11467, __PRETTY_FUNCTION__)) | ||||||
11467 | "Expected a push action.")((Action == ActionToDo::PushAsLastprivateConditional && "Expected a push action.") ? static_cast<void> (0) : __assert_fail ("Action == ActionToDo::PushAsLastprivateConditional && \"Expected a push action.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11467, __PRETTY_FUNCTION__)); | ||||||
11468 | LastprivateConditionalData &Data = | ||||||
11469 | CGM.getOpenMPRuntime().LastprivateConditionalStack.emplace_back(); | ||||||
11470 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { | ||||||
11471 | if (C->getKind() != OMPC_LASTPRIVATE_conditional) | ||||||
11472 | continue; | ||||||
11473 | |||||||
11474 | for (const Expr *Ref : C->varlists()) { | ||||||
11475 | Data.DeclToUniqueName.insert(std::make_pair( | ||||||
11476 | cast<DeclRefExpr>(Ref->IgnoreParenImpCasts())->getDecl(), | ||||||
11477 | SmallString<16>(generateUniqueName(CGM, "pl_cond", Ref)))); | ||||||
11478 | } | ||||||
11479 | } | ||||||
11480 | Data.IVLVal = IVLVal; | ||||||
11481 | Data.Fn = CGF.CurFn; | ||||||
11482 | } | ||||||
11483 | |||||||
11484 | CGOpenMPRuntime::LastprivateConditionalRAII::LastprivateConditionalRAII( | ||||||
11485 | CodeGenFunction &CGF, const OMPExecutableDirective &S) | ||||||
11486 | : CGM(CGF.CGM), Action(ActionToDo::DoNotPush) { | ||||||
11487 | assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.")((CGM.getLangOpts().OpenMP && "Not in OpenMP mode.") ? static_cast<void> (0) : __assert_fail ("CGM.getLangOpts().OpenMP && \"Not in OpenMP mode.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11487, __PRETTY_FUNCTION__)); | ||||||
11488 | if (CGM.getLangOpts().OpenMP < 50) | ||||||
11489 | return; | ||||||
11490 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> NeedToAddForLPCsAsDisabled; | ||||||
11491 | tryToDisableInnerAnalysis(S, NeedToAddForLPCsAsDisabled); | ||||||
11492 | if (!NeedToAddForLPCsAsDisabled.empty()) { | ||||||
11493 | Action = ActionToDo::DisableLastprivateConditional; | ||||||
11494 | LastprivateConditionalData &Data = | ||||||
11495 | CGM.getOpenMPRuntime().LastprivateConditionalStack.emplace_back(); | ||||||
11496 | for (const Decl *VD : NeedToAddForLPCsAsDisabled) | ||||||
11497 | Data.DeclToUniqueName.insert(std::make_pair(VD, SmallString<16>())); | ||||||
11498 | Data.Fn = CGF.CurFn; | ||||||
11499 | Data.Disabled = true; | ||||||
11500 | } | ||||||
11501 | } | ||||||
11502 | |||||||
11503 | CGOpenMPRuntime::LastprivateConditionalRAII | ||||||
11504 | CGOpenMPRuntime::LastprivateConditionalRAII::disable( | ||||||
11505 | CodeGenFunction &CGF, const OMPExecutableDirective &S) { | ||||||
11506 | return LastprivateConditionalRAII(CGF, S); | ||||||
11507 | } | ||||||
11508 | |||||||
11509 | CGOpenMPRuntime::LastprivateConditionalRAII::~LastprivateConditionalRAII() { | ||||||
11510 | if (CGM.getLangOpts().OpenMP < 50) | ||||||
11511 | return; | ||||||
11512 | if (Action == ActionToDo::DisableLastprivateConditional) { | ||||||
11513 | assert(CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled &&((CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && "Expected list of disabled private vars.") ? static_cast <void> (0) : __assert_fail ("CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of disabled private vars.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11514, __PRETTY_FUNCTION__)) | ||||||
11514 | "Expected list of disabled private vars.")((CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && "Expected list of disabled private vars.") ? static_cast <void> (0) : __assert_fail ("CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of disabled private vars.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11514, __PRETTY_FUNCTION__)); | ||||||
11515 | CGM.getOpenMPRuntime().LastprivateConditionalStack.pop_back(); | ||||||
11516 | } | ||||||
11517 | if (Action == ActionToDo::PushAsLastprivateConditional) { | ||||||
11518 | assert(((!CGM.getOpenMPRuntime().LastprivateConditionalStack.back(). Disabled && "Expected list of lastprivate conditional vars." ) ? static_cast<void> (0) : __assert_fail ("!CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of lastprivate conditional vars.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11520, __PRETTY_FUNCTION__)) | ||||||
11519 | !CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled &&((!CGM.getOpenMPRuntime().LastprivateConditionalStack.back(). Disabled && "Expected list of lastprivate conditional vars." ) ? static_cast<void> (0) : __assert_fail ("!CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of lastprivate conditional vars.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11520, __PRETTY_FUNCTION__)) | ||||||
11520 | "Expected list of lastprivate conditional vars.")((!CGM.getOpenMPRuntime().LastprivateConditionalStack.back(). Disabled && "Expected list of lastprivate conditional vars." ) ? static_cast<void> (0) : __assert_fail ("!CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled && \"Expected list of lastprivate conditional vars.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11520, __PRETTY_FUNCTION__)); | ||||||
11521 | CGM.getOpenMPRuntime().LastprivateConditionalStack.pop_back(); | ||||||
11522 | } | ||||||
11523 | } | ||||||
11524 | |||||||
11525 | Address CGOpenMPRuntime::emitLastprivateConditionalInit(CodeGenFunction &CGF, | ||||||
11526 | const VarDecl *VD) { | ||||||
11527 | ASTContext &C = CGM.getContext(); | ||||||
11528 | auto I = LastprivateConditionalToTypes.find(CGF.CurFn); | ||||||
11529 | if (I == LastprivateConditionalToTypes.end()) | ||||||
11530 | I = LastprivateConditionalToTypes.try_emplace(CGF.CurFn).first; | ||||||
11531 | QualType NewType; | ||||||
11532 | const FieldDecl *VDField; | ||||||
11533 | const FieldDecl *FiredField; | ||||||
11534 | LValue BaseLVal; | ||||||
11535 | auto VI = I->getSecond().find(VD); | ||||||
11536 | if (VI == I->getSecond().end()) { | ||||||
11537 | RecordDecl *RD = C.buildImplicitRecord("lasprivate.conditional"); | ||||||
11538 | RD->startDefinition(); | ||||||
11539 | VDField = addFieldToRecordDecl(C, RD, VD->getType().getNonReferenceType()); | ||||||
11540 | FiredField = addFieldToRecordDecl(C, RD, C.CharTy); | ||||||
11541 | RD->completeDefinition(); | ||||||
11542 | NewType = C.getRecordType(RD); | ||||||
11543 | Address Addr = CGF.CreateMemTemp(NewType, C.getDeclAlign(VD), VD->getName()); | ||||||
11544 | BaseLVal = CGF.MakeAddrLValue(Addr, NewType, AlignmentSource::Decl); | ||||||
11545 | I->getSecond().try_emplace(VD, NewType, VDField, FiredField, BaseLVal); | ||||||
11546 | } else { | ||||||
11547 | NewType = std::get<0>(VI->getSecond()); | ||||||
11548 | VDField = std::get<1>(VI->getSecond()); | ||||||
11549 | FiredField = std::get<2>(VI->getSecond()); | ||||||
11550 | BaseLVal = std::get<3>(VI->getSecond()); | ||||||
11551 | } | ||||||
11552 | LValue FiredLVal = | ||||||
11553 | CGF.EmitLValueForField(BaseLVal, FiredField); | ||||||
11554 | CGF.EmitStoreOfScalar( | ||||||
11555 | llvm::ConstantInt::getNullValue(CGF.ConvertTypeForMem(C.CharTy)), | ||||||
11556 | FiredLVal); | ||||||
11557 | return CGF.EmitLValueForField(BaseLVal, VDField).getAddress(CGF); | ||||||
11558 | } | ||||||
11559 | |||||||
11560 | namespace { | ||||||
11561 | /// Checks if the lastprivate conditional variable is referenced in LHS. | ||||||
11562 | class LastprivateConditionalRefChecker final | ||||||
11563 | : public ConstStmtVisitor<LastprivateConditionalRefChecker, bool> { | ||||||
11564 | ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM; | ||||||
11565 | const Expr *FoundE = nullptr; | ||||||
11566 | const Decl *FoundD = nullptr; | ||||||
11567 | StringRef UniqueDeclName; | ||||||
11568 | LValue IVLVal; | ||||||
11569 | llvm::Function *FoundFn = nullptr; | ||||||
11570 | SourceLocation Loc; | ||||||
11571 | |||||||
11572 | public: | ||||||
11573 | bool VisitDeclRefExpr(const DeclRefExpr *E) { | ||||||
11574 | for (const CGOpenMPRuntime::LastprivateConditionalData &D : | ||||||
11575 | llvm::reverse(LPM)) { | ||||||
11576 | auto It = D.DeclToUniqueName.find(E->getDecl()); | ||||||
11577 | if (It == D.DeclToUniqueName.end()) | ||||||
11578 | continue; | ||||||
11579 | if (D.Disabled) | ||||||
11580 | return false; | ||||||
11581 | FoundE = E; | ||||||
11582 | FoundD = E->getDecl()->getCanonicalDecl(); | ||||||
11583 | UniqueDeclName = It->second; | ||||||
11584 | IVLVal = D.IVLVal; | ||||||
11585 | FoundFn = D.Fn; | ||||||
11586 | break; | ||||||
11587 | } | ||||||
11588 | return FoundE == E; | ||||||
11589 | } | ||||||
11590 | bool VisitMemberExpr(const MemberExpr *E) { | ||||||
11591 | if (!CodeGenFunction::IsWrappedCXXThis(E->getBase())) | ||||||
11592 | return false; | ||||||
11593 | for (const CGOpenMPRuntime::LastprivateConditionalData &D : | ||||||
11594 | llvm::reverse(LPM)) { | ||||||
11595 | auto It = D.DeclToUniqueName.find(E->getMemberDecl()); | ||||||
11596 | if (It == D.DeclToUniqueName.end()) | ||||||
11597 | continue; | ||||||
11598 | if (D.Disabled) | ||||||
11599 | return false; | ||||||
11600 | FoundE = E; | ||||||
11601 | FoundD = E->getMemberDecl()->getCanonicalDecl(); | ||||||
11602 | UniqueDeclName = It->second; | ||||||
11603 | IVLVal = D.IVLVal; | ||||||
11604 | FoundFn = D.Fn; | ||||||
11605 | break; | ||||||
11606 | } | ||||||
11607 | return FoundE == E; | ||||||
11608 | } | ||||||
11609 | bool VisitStmt(const Stmt *S) { | ||||||
11610 | for (const Stmt *Child : S->children()) { | ||||||
11611 | if (!Child) | ||||||
11612 | continue; | ||||||
11613 | if (const auto *E = dyn_cast<Expr>(Child)) | ||||||
11614 | if (!E->isGLValue()) | ||||||
11615 | continue; | ||||||
11616 | if (Visit(Child)) | ||||||
11617 | return true; | ||||||
11618 | } | ||||||
11619 | return false; | ||||||
11620 | } | ||||||
11621 | explicit LastprivateConditionalRefChecker( | ||||||
11622 | ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM) | ||||||
11623 | : LPM(LPM) {} | ||||||
11624 | std::tuple<const Expr *, const Decl *, StringRef, LValue, llvm::Function *> | ||||||
11625 | getFoundData() const { | ||||||
11626 | return std::make_tuple(FoundE, FoundD, UniqueDeclName, IVLVal, FoundFn); | ||||||
11627 | } | ||||||
11628 | }; | ||||||
11629 | } // namespace | ||||||
11630 | |||||||
11631 | void CGOpenMPRuntime::emitLastprivateConditionalUpdate(CodeGenFunction &CGF, | ||||||
11632 | LValue IVLVal, | ||||||
11633 | StringRef UniqueDeclName, | ||||||
11634 | LValue LVal, | ||||||
11635 | SourceLocation Loc) { | ||||||
11636 | // Last updated loop counter for the lastprivate conditional var. | ||||||
11637 | // int<xx> last_iv = 0; | ||||||
11638 | llvm::Type *LLIVTy = CGF.ConvertTypeForMem(IVLVal.getType()); | ||||||
11639 | llvm::Constant *LastIV = | ||||||
11640 | getOrCreateInternalVariable(LLIVTy, getName({UniqueDeclName, "iv"})); | ||||||
11641 | cast<llvm::GlobalVariable>(LastIV)->setAlignment( | ||||||
11642 | IVLVal.getAlignment().getAsAlign()); | ||||||
11643 | LValue LastIVLVal = CGF.MakeNaturalAlignAddrLValue(LastIV, IVLVal.getType()); | ||||||
11644 | |||||||
11645 | // Last value of the lastprivate conditional. | ||||||
11646 | // decltype(priv_a) last_a; | ||||||
11647 | llvm::Constant *Last = getOrCreateInternalVariable( | ||||||
11648 | CGF.ConvertTypeForMem(LVal.getType()), UniqueDeclName); | ||||||
11649 | cast<llvm::GlobalVariable>(Last)->setAlignment( | ||||||
11650 | LVal.getAlignment().getAsAlign()); | ||||||
11651 | LValue LastLVal = | ||||||
11652 | CGF.MakeAddrLValue(Last, LVal.getType(), LVal.getAlignment()); | ||||||
11653 | |||||||
11654 | // Global loop counter. Required to handle inner parallel-for regions. | ||||||
11655 | // iv | ||||||
11656 | llvm::Value *IVVal = CGF.EmitLoadOfScalar(IVLVal, Loc); | ||||||
11657 | |||||||
11658 | // #pragma omp critical(a) | ||||||
11659 | // if (last_iv <= iv) { | ||||||
11660 | // last_iv = iv; | ||||||
11661 | // last_a = priv_a; | ||||||
11662 | // } | ||||||
11663 | auto &&CodeGen = [&LastIVLVal, &IVLVal, IVVal, &LVal, &LastLVal, | ||||||
11664 | Loc](CodeGenFunction &CGF, PrePostActionTy &Action) { | ||||||
11665 | Action.Enter(CGF); | ||||||
11666 | llvm::Value *LastIVVal = CGF.EmitLoadOfScalar(LastIVLVal, Loc); | ||||||
11667 | // (last_iv <= iv) ? Check if the variable is updated and store new | ||||||
11668 | // value in global var. | ||||||
11669 | llvm::Value *CmpRes; | ||||||
11670 | if (IVLVal.getType()->isSignedIntegerType()) { | ||||||
11671 | CmpRes = CGF.Builder.CreateICmpSLE(LastIVVal, IVVal); | ||||||
11672 | } else { | ||||||
11673 | assert(IVLVal.getType()->isUnsignedIntegerType() &&((IVLVal.getType()->isUnsignedIntegerType() && "Loop iteration variable must be integer." ) ? static_cast<void> (0) : __assert_fail ("IVLVal.getType()->isUnsignedIntegerType() && \"Loop iteration variable must be integer.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11674, __PRETTY_FUNCTION__)) | ||||||
11674 | "Loop iteration variable must be integer.")((IVLVal.getType()->isUnsignedIntegerType() && "Loop iteration variable must be integer." ) ? static_cast<void> (0) : __assert_fail ("IVLVal.getType()->isUnsignedIntegerType() && \"Loop iteration variable must be integer.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11674, __PRETTY_FUNCTION__)); | ||||||
11675 | CmpRes = CGF.Builder.CreateICmpULE(LastIVVal, IVVal); | ||||||
11676 | } | ||||||
11677 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("lp_cond_then"); | ||||||
11678 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock("lp_cond_exit"); | ||||||
11679 | CGF.Builder.CreateCondBr(CmpRes, ThenBB, ExitBB); | ||||||
11680 | // { | ||||||
11681 | CGF.EmitBlock(ThenBB); | ||||||
11682 | |||||||
11683 | // last_iv = iv; | ||||||
11684 | CGF.EmitStoreOfScalar(IVVal, LastIVLVal); | ||||||
11685 | |||||||
11686 | // last_a = priv_a; | ||||||
11687 | switch (CGF.getEvaluationKind(LVal.getType())) { | ||||||
11688 | case TEK_Scalar: { | ||||||
11689 | llvm::Value *PrivVal = CGF.EmitLoadOfScalar(LVal, Loc); | ||||||
11690 | CGF.EmitStoreOfScalar(PrivVal, LastLVal); | ||||||
11691 | break; | ||||||
11692 | } | ||||||
11693 | case TEK_Complex: { | ||||||
11694 | CodeGenFunction::ComplexPairTy PrivVal = CGF.EmitLoadOfComplex(LVal, Loc); | ||||||
11695 | CGF.EmitStoreOfComplex(PrivVal, LastLVal, /*isInit=*/false); | ||||||
11696 | break; | ||||||
11697 | } | ||||||
11698 | case TEK_Aggregate: | ||||||
11699 | llvm_unreachable(::llvm::llvm_unreachable_internal("Aggregates are not supported in lastprivate conditional." , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11700) | ||||||
11700 | "Aggregates are not supported in lastprivate conditional.")::llvm::llvm_unreachable_internal("Aggregates are not supported in lastprivate conditional." , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11700); | ||||||
11701 | } | ||||||
11702 | // } | ||||||
11703 | CGF.EmitBranch(ExitBB); | ||||||
11704 | // There is no need to emit line number for unconditional branch. | ||||||
11705 | (void)ApplyDebugLocation::CreateEmpty(CGF); | ||||||
11706 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); | ||||||
11707 | }; | ||||||
11708 | |||||||
11709 | if (CGM.getLangOpts().OpenMPSimd) { | ||||||
11710 | // Do not emit as a critical region as no parallel region could be emitted. | ||||||
11711 | RegionCodeGenTy ThenRCG(CodeGen); | ||||||
11712 | ThenRCG(CGF); | ||||||
11713 | } else { | ||||||
11714 | emitCriticalRegion(CGF, UniqueDeclName, CodeGen, Loc); | ||||||
11715 | } | ||||||
11716 | } | ||||||
11717 | |||||||
11718 | void CGOpenMPRuntime::checkAndEmitLastprivateConditional(CodeGenFunction &CGF, | ||||||
11719 | const Expr *LHS) { | ||||||
11720 | if (CGF.getLangOpts().OpenMP < 50 || LastprivateConditionalStack.empty()) | ||||||
11721 | return; | ||||||
11722 | LastprivateConditionalRefChecker Checker(LastprivateConditionalStack); | ||||||
11723 | if (!Checker.Visit(LHS)) | ||||||
11724 | return; | ||||||
11725 | const Expr *FoundE; | ||||||
11726 | const Decl *FoundD; | ||||||
11727 | StringRef UniqueDeclName; | ||||||
11728 | LValue IVLVal; | ||||||
11729 | llvm::Function *FoundFn; | ||||||
11730 | std::tie(FoundE, FoundD, UniqueDeclName, IVLVal, FoundFn) = | ||||||
11731 | Checker.getFoundData(); | ||||||
11732 | if (FoundFn != CGF.CurFn) { | ||||||
11733 | // Special codegen for inner parallel regions. | ||||||
11734 | // ((struct.lastprivate.conditional*)&priv_a)->Fired = 1; | ||||||
11735 | auto It = LastprivateConditionalToTypes[FoundFn].find(FoundD); | ||||||
11736 | assert(It != LastprivateConditionalToTypes[FoundFn].end() &&((It != LastprivateConditionalToTypes[FoundFn].end() && "Lastprivate conditional is not found in outer region.") ? static_cast <void> (0) : __assert_fail ("It != LastprivateConditionalToTypes[FoundFn].end() && \"Lastprivate conditional is not found in outer region.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11737, __PRETTY_FUNCTION__)) | ||||||
11737 | "Lastprivate conditional is not found in outer region.")((It != LastprivateConditionalToTypes[FoundFn].end() && "Lastprivate conditional is not found in outer region.") ? static_cast <void> (0) : __assert_fail ("It != LastprivateConditionalToTypes[FoundFn].end() && \"Lastprivate conditional is not found in outer region.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11737, __PRETTY_FUNCTION__)); | ||||||
11738 | QualType StructTy = std::get<0>(It->getSecond()); | ||||||
11739 | const FieldDecl* FiredDecl = std::get<2>(It->getSecond()); | ||||||
11740 | LValue PrivLVal = CGF.EmitLValue(FoundE); | ||||||
11741 | Address StructAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | ||||||
11742 | PrivLVal.getAddress(CGF), | ||||||
11743 | CGF.ConvertTypeForMem(CGF.getContext().getPointerType(StructTy))); | ||||||
11744 | LValue BaseLVal = | ||||||
11745 | CGF.MakeAddrLValue(StructAddr, StructTy, AlignmentSource::Decl); | ||||||
11746 | LValue FiredLVal = CGF.EmitLValueForField(BaseLVal, FiredDecl); | ||||||
11747 | CGF.EmitAtomicStore(RValue::get(llvm::ConstantInt::get( | ||||||
11748 | CGF.ConvertTypeForMem(FiredDecl->getType()), 1)), | ||||||
11749 | FiredLVal, llvm::AtomicOrdering::Unordered, | ||||||
11750 | /*IsVolatile=*/true, /*isInit=*/false); | ||||||
11751 | return; | ||||||
11752 | } | ||||||
11753 | |||||||
11754 | // Private address of the lastprivate conditional in the current context. | ||||||
11755 | // priv_a | ||||||
11756 | LValue LVal = CGF.EmitLValue(FoundE); | ||||||
11757 | emitLastprivateConditionalUpdate(CGF, IVLVal, UniqueDeclName, LVal, | ||||||
11758 | FoundE->getExprLoc()); | ||||||
11759 | } | ||||||
11760 | |||||||
11761 | void CGOpenMPRuntime::checkAndEmitSharedLastprivateConditional( | ||||||
11762 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | ||||||
11763 | const llvm::DenseSet<CanonicalDeclPtr<const VarDecl>> &IgnoredDecls) { | ||||||
11764 | if (CGF.getLangOpts().OpenMP < 50 || LastprivateConditionalStack.empty()) | ||||||
11765 | return; | ||||||
11766 | auto Range = llvm::reverse(LastprivateConditionalStack); | ||||||
11767 | auto It = llvm::find_if( | ||||||
11768 | Range, [](const LastprivateConditionalData &D) { return !D.Disabled; }); | ||||||
11769 | if (It == Range.end() || It->Fn != CGF.CurFn) | ||||||
11770 | return; | ||||||
11771 | auto LPCI = LastprivateConditionalToTypes.find(It->Fn); | ||||||
11772 | assert(LPCI != LastprivateConditionalToTypes.end() &&((LPCI != LastprivateConditionalToTypes.end() && "Lastprivates must be registered already." ) ? static_cast<void> (0) : __assert_fail ("LPCI != LastprivateConditionalToTypes.end() && \"Lastprivates must be registered already.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11773, __PRETTY_FUNCTION__)) | ||||||
11773 | "Lastprivates must be registered already.")((LPCI != LastprivateConditionalToTypes.end() && "Lastprivates must be registered already." ) ? static_cast<void> (0) : __assert_fail ("LPCI != LastprivateConditionalToTypes.end() && \"Lastprivates must be registered already.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11773, __PRETTY_FUNCTION__)); | ||||||
11774 | SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; | ||||||
11775 | getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind()); | ||||||
11776 | const CapturedStmt *CS = D.getCapturedStmt(CaptureRegions.back()); | ||||||
11777 | for (const auto &Pair : It->DeclToUniqueName) { | ||||||
11778 | const auto *VD = cast<VarDecl>(Pair.first->getCanonicalDecl()); | ||||||
11779 | if (!CS->capturesVariable(VD) || IgnoredDecls.count(VD) > 0) | ||||||
11780 | continue; | ||||||
11781 | auto I = LPCI->getSecond().find(Pair.first); | ||||||
11782 | assert(I != LPCI->getSecond().end() &&((I != LPCI->getSecond().end() && "Lastprivate must be rehistered already." ) ? static_cast<void> (0) : __assert_fail ("I != LPCI->getSecond().end() && \"Lastprivate must be rehistered already.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11783, __PRETTY_FUNCTION__)) | ||||||
11783 | "Lastprivate must be rehistered already.")((I != LPCI->getSecond().end() && "Lastprivate must be rehistered already." ) ? static_cast<void> (0) : __assert_fail ("I != LPCI->getSecond().end() && \"Lastprivate must be rehistered already.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11783, __PRETTY_FUNCTION__)); | ||||||
11784 | // bool Cmp = priv_a.Fired != 0; | ||||||
11785 | LValue BaseLVal = std::get<3>(I->getSecond()); | ||||||
11786 | LValue FiredLVal = | ||||||
11787 | CGF.EmitLValueForField(BaseLVal, std::get<2>(I->getSecond())); | ||||||
11788 | llvm::Value *Res = CGF.EmitLoadOfScalar(FiredLVal, D.getBeginLoc()); | ||||||
11789 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Res); | ||||||
11790 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("lpc.then"); | ||||||
11791 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("lpc.done"); | ||||||
11792 | // if (Cmp) { | ||||||
11793 | CGF.Builder.CreateCondBr(Cmp, ThenBB, DoneBB); | ||||||
11794 | CGF.EmitBlock(ThenBB); | ||||||
11795 | Address Addr = CGF.GetAddrOfLocalVar(VD); | ||||||
11796 | LValue LVal; | ||||||
11797 | if (VD->getType()->isReferenceType()) | ||||||
11798 | LVal = CGF.EmitLoadOfReferenceLValue(Addr, VD->getType(), | ||||||
11799 | AlignmentSource::Decl); | ||||||
11800 | else | ||||||
11801 | LVal = CGF.MakeAddrLValue(Addr, VD->getType().getNonReferenceType(), | ||||||
11802 | AlignmentSource::Decl); | ||||||
11803 | emitLastprivateConditionalUpdate(CGF, It->IVLVal, Pair.second, LVal, | ||||||
11804 | D.getBeginLoc()); | ||||||
11805 | auto AL = ApplyDebugLocation::CreateArtificial(CGF); | ||||||
11806 | CGF.EmitBlock(DoneBB, /*IsFinal=*/true); | ||||||
11807 | // } | ||||||
11808 | } | ||||||
11809 | } | ||||||
11810 | |||||||
11811 | void CGOpenMPRuntime::emitLastprivateConditionalFinalUpdate( | ||||||
11812 | CodeGenFunction &CGF, LValue PrivLVal, const VarDecl *VD, | ||||||
11813 | SourceLocation Loc) { | ||||||
11814 | if (CGF.getLangOpts().OpenMP < 50) | ||||||
11815 | return; | ||||||
11816 | auto It = LastprivateConditionalStack.back().DeclToUniqueName.find(VD); | ||||||
11817 | assert(It != LastprivateConditionalStack.back().DeclToUniqueName.end() &&((It != LastprivateConditionalStack.back().DeclToUniqueName.end () && "Unknown lastprivate conditional variable.") ? static_cast <void> (0) : __assert_fail ("It != LastprivateConditionalStack.back().DeclToUniqueName.end() && \"Unknown lastprivate conditional variable.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11818, __PRETTY_FUNCTION__)) | ||||||
11818 | "Unknown lastprivate conditional variable.")((It != LastprivateConditionalStack.back().DeclToUniqueName.end () && "Unknown lastprivate conditional variable.") ? static_cast <void> (0) : __assert_fail ("It != LastprivateConditionalStack.back().DeclToUniqueName.end() && \"Unknown lastprivate conditional variable.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11818, __PRETTY_FUNCTION__)); | ||||||
11819 | StringRef UniqueName = It->second; | ||||||
11820 | llvm::GlobalVariable *GV = CGM.getModule().getNamedGlobal(UniqueName); | ||||||
11821 | // The variable was not updated in the region - exit. | ||||||
11822 | if (!GV) | ||||||
11823 | return; | ||||||
11824 | LValue LPLVal = CGF.MakeAddrLValue( | ||||||
11825 | GV, PrivLVal.getType().getNonReferenceType(), PrivLVal.getAlignment()); | ||||||
11826 | llvm::Value *Res = CGF.EmitLoadOfScalar(LPLVal, Loc); | ||||||
11827 | CGF.EmitStoreOfScalar(Res, PrivLVal); | ||||||
11828 | } | ||||||
11829 | |||||||
11830 | llvm::Function *CGOpenMPSIMDRuntime::emitParallelOutlinedFunction( | ||||||
11831 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | ||||||
11832 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | ||||||
11833 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11833); | ||||||
11834 | } | ||||||
11835 | |||||||
11836 | llvm::Function *CGOpenMPSIMDRuntime::emitTeamsOutlinedFunction( | ||||||
11837 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | ||||||
11838 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | ||||||
11839 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11839); | ||||||
11840 | } | ||||||
11841 | |||||||
11842 | llvm::Function *CGOpenMPSIMDRuntime::emitTaskOutlinedFunction( | ||||||
11843 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | ||||||
11844 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, | ||||||
11845 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, | ||||||
11846 | bool Tied, unsigned &NumberOfParts) { | ||||||
11847 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11847); | ||||||
11848 | } | ||||||
11849 | |||||||
11850 | void CGOpenMPSIMDRuntime::emitParallelCall(CodeGenFunction &CGF, | ||||||
11851 | SourceLocation Loc, | ||||||
11852 | llvm::Function *OutlinedFn, | ||||||
11853 | ArrayRef<llvm::Value *> CapturedVars, | ||||||
11854 | const Expr *IfCond) { | ||||||
11855 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11855); | ||||||
11856 | } | ||||||
11857 | |||||||
11858 | void CGOpenMPSIMDRuntime::emitCriticalRegion( | ||||||
11859 | CodeGenFunction &CGF, StringRef CriticalName, | ||||||
11860 | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, | ||||||
11861 | const Expr *Hint) { | ||||||
11862 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11862); | ||||||
11863 | } | ||||||
11864 | |||||||
11865 | void CGOpenMPSIMDRuntime::emitMasterRegion(CodeGenFunction &CGF, | ||||||
11866 | const RegionCodeGenTy &MasterOpGen, | ||||||
11867 | SourceLocation Loc) { | ||||||
11868 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11868); | ||||||
11869 | } | ||||||
11870 | |||||||
11871 | void CGOpenMPSIMDRuntime::emitTaskyieldCall(CodeGenFunction &CGF, | ||||||
11872 | SourceLocation Loc) { | ||||||
11873 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11873); | ||||||
11874 | } | ||||||
11875 | |||||||
11876 | void CGOpenMPSIMDRuntime::emitTaskgroupRegion( | ||||||
11877 | CodeGenFunction &CGF, const RegionCodeGenTy &TaskgroupOpGen, | ||||||
11878 | SourceLocation Loc) { | ||||||
11879 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11879); | ||||||
11880 | } | ||||||
11881 | |||||||
11882 | void CGOpenMPSIMDRuntime::emitSingleRegion( | ||||||
11883 | CodeGenFunction &CGF, const RegionCodeGenTy &SingleOpGen, | ||||||
11884 | SourceLocation Loc, ArrayRef<const Expr *> CopyprivateVars, | ||||||
11885 | ArrayRef<const Expr *> DestExprs, ArrayRef<const Expr *> SrcExprs, | ||||||
11886 | ArrayRef<const Expr *> AssignmentOps) { | ||||||
11887 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11887); | ||||||
11888 | } | ||||||
11889 | |||||||
11890 | void CGOpenMPSIMDRuntime::emitOrderedRegion(CodeGenFunction &CGF, | ||||||
11891 | const RegionCodeGenTy &OrderedOpGen, | ||||||
11892 | SourceLocation Loc, | ||||||
11893 | bool IsThreads) { | ||||||
11894 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11894); | ||||||
11895 | } | ||||||
11896 | |||||||
11897 | void CGOpenMPSIMDRuntime::emitBarrierCall(CodeGenFunction &CGF, | ||||||
11898 | SourceLocation Loc, | ||||||
11899 | OpenMPDirectiveKind Kind, | ||||||
11900 | bool EmitChecks, | ||||||
11901 | bool ForceSimpleCall) { | ||||||
11902 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11902); | ||||||
11903 | } | ||||||
11904 | |||||||
11905 | void CGOpenMPSIMDRuntime::emitForDispatchInit( | ||||||
11906 | CodeGenFunction &CGF, SourceLocation Loc, | ||||||
11907 | const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned, | ||||||
11908 | bool Ordered, const DispatchRTInput &DispatchValues) { | ||||||
11909 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11909); | ||||||
11910 | } | ||||||
11911 | |||||||
11912 | void CGOpenMPSIMDRuntime::emitForStaticInit( | ||||||
11913 | CodeGenFunction &CGF, SourceLocation Loc, OpenMPDirectiveKind DKind, | ||||||
11914 | const OpenMPScheduleTy &ScheduleKind, const StaticRTInput &Values) { | ||||||
11915 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11915); | ||||||
11916 | } | ||||||
11917 | |||||||
11918 | void CGOpenMPSIMDRuntime::emitDistributeStaticInit( | ||||||
11919 | CodeGenFunction &CGF, SourceLocation Loc, | ||||||
11920 | OpenMPDistScheduleClauseKind SchedKind, const StaticRTInput &Values) { | ||||||
11921 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11921); | ||||||
11922 | } | ||||||
11923 | |||||||
11924 | void CGOpenMPSIMDRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF, | ||||||
11925 | SourceLocation Loc, | ||||||
11926 | unsigned IVSize, | ||||||
11927 | bool IVSigned) { | ||||||
11928 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11928); | ||||||
11929 | } | ||||||
11930 | |||||||
11931 | void CGOpenMPSIMDRuntime::emitForStaticFinish(CodeGenFunction &CGF, | ||||||
11932 | SourceLocation Loc, | ||||||
11933 | OpenMPDirectiveKind DKind) { | ||||||
11934 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11934); | ||||||
11935 | } | ||||||
11936 | |||||||
11937 | llvm::Value *CGOpenMPSIMDRuntime::emitForNext(CodeGenFunction &CGF, | ||||||
11938 | SourceLocation Loc, | ||||||
11939 | unsigned IVSize, bool IVSigned, | ||||||
11940 | Address IL, Address LB, | ||||||
11941 | Address UB, Address ST) { | ||||||
11942 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11942); | ||||||
11943 | } | ||||||
11944 | |||||||
11945 | void CGOpenMPSIMDRuntime::emitNumThreadsClause(CodeGenFunction &CGF, | ||||||
11946 | llvm::Value *NumThreads, | ||||||
11947 | SourceLocation Loc) { | ||||||
11948 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11948); | ||||||
11949 | } | ||||||
11950 | |||||||
11951 | void CGOpenMPSIMDRuntime::emitProcBindClause(CodeGenFunction &CGF, | ||||||
11952 | ProcBindKind ProcBind, | ||||||
11953 | SourceLocation Loc) { | ||||||
11954 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11954); | ||||||
11955 | } | ||||||
11956 | |||||||
11957 | Address CGOpenMPSIMDRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF, | ||||||
11958 | const VarDecl *VD, | ||||||
11959 | Address VDAddr, | ||||||
11960 | SourceLocation Loc) { | ||||||
11961 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11961); | ||||||
11962 | } | ||||||
11963 | |||||||
11964 | llvm::Function *CGOpenMPSIMDRuntime::emitThreadPrivateVarDefinition( | ||||||
11965 | const VarDecl *VD, Address VDAddr, SourceLocation Loc, bool PerformInit, | ||||||
11966 | CodeGenFunction *CGF) { | ||||||
11967 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11967); | ||||||
11968 | } | ||||||
11969 | |||||||
11970 | Address CGOpenMPSIMDRuntime::getAddrOfArtificialThreadPrivate( | ||||||
11971 | CodeGenFunction &CGF, QualType VarType, StringRef Name) { | ||||||
11972 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11972); | ||||||
11973 | } | ||||||
11974 | |||||||
11975 | void CGOpenMPSIMDRuntime::emitFlush(CodeGenFunction &CGF, | ||||||
11976 | ArrayRef<const Expr *> Vars, | ||||||
11977 | SourceLocation Loc, | ||||||
11978 | llvm::AtomicOrdering AO) { | ||||||
11979 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11979); | ||||||
11980 | } | ||||||
11981 | |||||||
11982 | void CGOpenMPSIMDRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, | ||||||
11983 | const OMPExecutableDirective &D, | ||||||
11984 | llvm::Function *TaskFunction, | ||||||
11985 | QualType SharedsTy, Address Shareds, | ||||||
11986 | const Expr *IfCond, | ||||||
11987 | const OMPTaskDataTy &Data) { | ||||||
11988 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11988); | ||||||
11989 | } | ||||||
11990 | |||||||
11991 | void CGOpenMPSIMDRuntime::emitTaskLoopCall( | ||||||
11992 | CodeGenFunction &CGF, SourceLocation Loc, const OMPLoopDirective &D, | ||||||
11993 | llvm::Function *TaskFunction, QualType SharedsTy, Address Shareds, | ||||||
11994 | const Expr *IfCond, const OMPTaskDataTy &Data) { | ||||||
11995 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 11995); | ||||||
11996 | } | ||||||
11997 | |||||||
11998 | void CGOpenMPSIMDRuntime::emitReduction( | ||||||
11999 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, | ||||||
12000 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, | ||||||
12001 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { | ||||||
12002 | assert(Options.SimpleReduction && "Only simple reduction is expected.")((Options.SimpleReduction && "Only simple reduction is expected." ) ? static_cast<void> (0) : __assert_fail ("Options.SimpleReduction && \"Only simple reduction is expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12002, __PRETTY_FUNCTION__)); | ||||||
12003 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, | ||||||
12004 | ReductionOps, Options); | ||||||
12005 | } | ||||||
12006 | |||||||
12007 | llvm::Value *CGOpenMPSIMDRuntime::emitTaskReductionInit( | ||||||
12008 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs, | ||||||
12009 | ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) { | ||||||
12010 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12010); | ||||||
12011 | } | ||||||
12012 | |||||||
12013 | void CGOpenMPSIMDRuntime::emitTaskReductionFini(CodeGenFunction &CGF, | ||||||
12014 | SourceLocation Loc, | ||||||
12015 | bool IsWorksharingReduction) { | ||||||
12016 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12016); | ||||||
12017 | } | ||||||
12018 | |||||||
12019 | void CGOpenMPSIMDRuntime::emitTaskReductionFixups(CodeGenFunction &CGF, | ||||||
12020 | SourceLocation Loc, | ||||||
12021 | ReductionCodeGen &RCG, | ||||||
12022 | unsigned N) { | ||||||
12023 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12023); | ||||||
12024 | } | ||||||
12025 | |||||||
12026 | Address CGOpenMPSIMDRuntime::getTaskReductionItem(CodeGenFunction &CGF, | ||||||
12027 | SourceLocation Loc, | ||||||
12028 | llvm::Value *ReductionsPtr, | ||||||
12029 | LValue SharedLVal) { | ||||||
12030 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12030); | ||||||
12031 | } | ||||||
12032 | |||||||
12033 | void CGOpenMPSIMDRuntime::emitTaskwaitCall(CodeGenFunction &CGF, | ||||||
12034 | SourceLocation Loc) { | ||||||
12035 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12035); | ||||||
12036 | } | ||||||
12037 | |||||||
12038 | void CGOpenMPSIMDRuntime::emitCancellationPointCall( | ||||||
12039 | CodeGenFunction &CGF, SourceLocation Loc, | ||||||
12040 | OpenMPDirectiveKind CancelRegion) { | ||||||
12041 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12041); | ||||||
12042 | } | ||||||
12043 | |||||||
12044 | void CGOpenMPSIMDRuntime::emitCancelCall(CodeGenFunction &CGF, | ||||||
12045 | SourceLocation Loc, const Expr *IfCond, | ||||||
12046 | OpenMPDirectiveKind CancelRegion) { | ||||||
12047 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12047); | ||||||
12048 | } | ||||||
12049 | |||||||
12050 | void CGOpenMPSIMDRuntime::emitTargetOutlinedFunction( | ||||||
12051 | const OMPExecutableDirective &D, StringRef ParentName, | ||||||
12052 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | ||||||
12053 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | ||||||
12054 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12054); | ||||||
12055 | } | ||||||
12056 | |||||||
12057 | void CGOpenMPSIMDRuntime::emitTargetCall( | ||||||
12058 | CodeGenFunction &CGF, const OMPExecutableDirective &D, | ||||||
12059 | llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, const Expr *IfCond, | ||||||
12060 | llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device, | ||||||
12061 | llvm::function_ref<llvm::Value *(CodeGenFunction &CGF, | ||||||
12062 | const OMPLoopDirective &D)> | ||||||
12063 | SizeEmitter) { | ||||||
12064 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12064); | ||||||
12065 | } | ||||||
12066 | |||||||
12067 | bool CGOpenMPSIMDRuntime::emitTargetFunctions(GlobalDecl GD) { | ||||||
12068 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12068); | ||||||
12069 | } | ||||||
12070 | |||||||
12071 | bool CGOpenMPSIMDRuntime::emitTargetGlobalVariable(GlobalDecl GD) { | ||||||
12072 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12072); | ||||||
12073 | } | ||||||
12074 | |||||||
12075 | bool CGOpenMPSIMDRuntime::emitTargetGlobal(GlobalDecl GD) { | ||||||
12076 | return false; | ||||||
12077 | } | ||||||
12078 | |||||||
12079 | void CGOpenMPSIMDRuntime::emitTeamsCall(CodeGenFunction &CGF, | ||||||
12080 | const OMPExecutableDirective &D, | ||||||
12081 | SourceLocation Loc, | ||||||
12082 | llvm::Function *OutlinedFn, | ||||||
12083 | ArrayRef<llvm::Value *> CapturedVars) { | ||||||
12084 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12084); | ||||||
12085 | } | ||||||
12086 | |||||||
12087 | void CGOpenMPSIMDRuntime::emitNumTeamsClause(CodeGenFunction &CGF, | ||||||
12088 | const Expr *NumTeams, | ||||||
12089 | const Expr *ThreadLimit, | ||||||
12090 | SourceLocation Loc) { | ||||||
12091 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12091); | ||||||
12092 | } | ||||||
12093 | |||||||
12094 | void CGOpenMPSIMDRuntime::emitTargetDataCalls( | ||||||
12095 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | ||||||
12096 | const Expr *Device, const RegionCodeGenTy &CodeGen, TargetDataInfo &Info) { | ||||||
12097 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12097); | ||||||
12098 | } | ||||||
12099 | |||||||
12100 | void CGOpenMPSIMDRuntime::emitTargetDataStandAloneCall( | ||||||
12101 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | ||||||
12102 | const Expr *Device) { | ||||||
12103 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12103); | ||||||
12104 | } | ||||||
12105 | |||||||
12106 | void CGOpenMPSIMDRuntime::emitDoacrossInit(CodeGenFunction &CGF, | ||||||
12107 | const OMPLoopDirective &D, | ||||||
12108 | ArrayRef<Expr *> NumIterations) { | ||||||
12109 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12109); | ||||||
12110 | } | ||||||
12111 | |||||||
12112 | void CGOpenMPSIMDRuntime::emitDoacrossOrdered(CodeGenFunction &CGF, | ||||||
12113 | const OMPDependClause *C) { | ||||||
12114 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12114); | ||||||
12115 | } | ||||||
12116 | |||||||
12117 | const VarDecl * | ||||||
12118 | CGOpenMPSIMDRuntime::translateParameter(const FieldDecl *FD, | ||||||
12119 | const VarDecl *NativeParam) const { | ||||||
12120 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12120); | ||||||
12121 | } | ||||||
12122 | |||||||
12123 | Address | ||||||
12124 | CGOpenMPSIMDRuntime::getParameterAddress(CodeGenFunction &CGF, | ||||||
12125 | const VarDecl *NativeParam, | ||||||
12126 | const VarDecl *TargetParam) const { | ||||||
12127 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 12127); | ||||||
12128 | } |
1 | // Iterators -*- C++ -*- |
2 | |
3 | // Copyright (C) 2001-2016 Free Software Foundation, Inc. |
4 | // |
5 | // This file is part of the GNU ISO C++ Library. This library is free |
6 | // software; you can redistribute it and/or modify it under the |
7 | // terms of the GNU General Public License as published by the |
8 | // Free Software Foundation; either version 3, or (at your option) |
9 | // any later version. |
10 | |
11 | // This library is distributed in the hope that it will be useful, |
12 | // but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
14 | // GNU General Public License for more details. |
15 | |
16 | // Under Section 7 of GPL version 3, you are granted additional |
17 | // permissions described in the GCC Runtime Library Exception, version |
18 | // 3.1, as published by the Free Software Foundation. |
19 | |
20 | // You should have received a copy of the GNU General Public License and |
21 | // a copy of the GCC Runtime Library Exception along with this program; |
22 | // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
23 | // <http://www.gnu.org/licenses/>. |
24 | |
25 | /* |
26 | * |
27 | * Copyright (c) 1994 |
28 | * Hewlett-Packard Company |
29 | * |
30 | * Permission to use, copy, modify, distribute and sell this software |
31 | * and its documentation for any purpose is hereby granted without fee, |
32 | * provided that the above copyright notice appear in all copies and |
33 | * that both that copyright notice and this permission notice appear |
34 | * in supporting documentation. Hewlett-Packard Company makes no |
35 | * representations about the suitability of this software for any |
36 | * purpose. It is provided "as is" without express or implied warranty. |
37 | * |
38 | * |
39 | * Copyright (c) 1996-1998 |
40 | * Silicon Graphics Computer Systems, Inc. |
41 | * |
42 | * Permission to use, copy, modify, distribute and sell this software |
43 | * and its documentation for any purpose is hereby granted without fee, |
44 | * provided that the above copyright notice appear in all copies and |
45 | * that both that copyright notice and this permission notice appear |
46 | * in supporting documentation. Silicon Graphics makes no |
47 | * representations about the suitability of this software for any |
48 | * purpose. It is provided "as is" without express or implied warranty. |
49 | */ |
50 | |
51 | /** @file bits/stl_iterator.h |
52 | * This is an internal header file, included by other library headers. |
53 | * Do not attempt to use it directly. @headername{iterator} |
54 | * |
55 | * This file implements reverse_iterator, back_insert_iterator, |
56 | * front_insert_iterator, insert_iterator, __normal_iterator, and their |
57 | * supporting functions and overloaded operators. |
58 | */ |
59 | |
60 | #ifndef _STL_ITERATOR_H1 |
61 | #define _STL_ITERATOR_H1 1 |
62 | |
63 | #include <bits/cpp_type_traits.h> |
64 | #include <ext/type_traits.h> |
65 | #include <bits/move.h> |
66 | #include <bits/ptr_traits.h> |
67 | |
68 | namespace std _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
69 | { |
70 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
71 | |
72 | /** |
73 | * @addtogroup iterators |
74 | * @{ |
75 | */ |
76 | |
77 | // 24.4.1 Reverse iterators |
78 | /** |
79 | * Bidirectional and random access iterators have corresponding reverse |
80 | * %iterator adaptors that iterate through the data structure in the |
81 | * opposite direction. They have the same signatures as the corresponding |
82 | * iterators. The fundamental relation between a reverse %iterator and its |
83 | * corresponding %iterator @c i is established by the identity: |
84 | * @code |
85 | * &*(reverse_iterator(i)) == &*(i - 1) |
86 | * @endcode |
87 | * |
88 | * <em>This mapping is dictated by the fact that while there is always a |
89 | * pointer past the end of an array, there might not be a valid pointer |
90 | * before the beginning of an array.</em> [24.4.1]/1,2 |
91 | * |
92 | * Reverse iterators can be tricky and surprising at first. Their |
93 | * semantics make sense, however, and the trickiness is a side effect of |
94 | * the requirement that the iterators must be safe. |
95 | */ |
96 | template<typename _Iterator> |
97 | class reverse_iterator |
98 | : public iterator<typename iterator_traits<_Iterator>::iterator_category, |
99 | typename iterator_traits<_Iterator>::value_type, |
100 | typename iterator_traits<_Iterator>::difference_type, |
101 | typename iterator_traits<_Iterator>::pointer, |
102 | typename iterator_traits<_Iterator>::reference> |
103 | { |
104 | protected: |
105 | _Iterator current; |
106 | |
107 | typedef iterator_traits<_Iterator> __traits_type; |
108 | |
109 | public: |
110 | typedef _Iterator iterator_type; |
111 | typedef typename __traits_type::difference_type difference_type; |
112 | typedef typename __traits_type::pointer pointer; |
113 | typedef typename __traits_type::reference reference; |
114 | |
115 | /** |
116 | * The default constructor value-initializes member @p current. |
117 | * If it is a pointer, that means it is zero-initialized. |
118 | */ |
119 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
120 | // 235 No specification of default ctor for reverse_iterator |
121 | reverse_iterator() : current() { } |
122 | |
123 | /** |
124 | * This %iterator will move in the opposite direction that @p x does. |
125 | */ |
126 | explicit |
127 | reverse_iterator(iterator_type __x) : current(__x) { } |
128 | |
129 | /** |
130 | * The copy constructor is normal. |
131 | */ |
132 | reverse_iterator(const reverse_iterator& __x) |
133 | : current(__x.current) { } |
134 | |
135 | /** |
136 | * A %reverse_iterator across other types can be copied if the |
137 | * underlying %iterator can be converted to the type of @c current. |
138 | */ |
139 | template<typename _Iter> |
140 | reverse_iterator(const reverse_iterator<_Iter>& __x) |
141 | : current(__x.base()) { } |
142 | |
143 | /** |
144 | * @return @c current, the %iterator used for underlying work. |
145 | */ |
146 | iterator_type |
147 | base() const |
148 | { return current; } |
149 | |
150 | /** |
151 | * @return A reference to the value at @c --current |
152 | * |
153 | * This requires that @c --current is dereferenceable. |
154 | * |
155 | * @warning This implementation requires that for an iterator of the |
156 | * underlying iterator type, @c x, a reference obtained by |
157 | * @c *x remains valid after @c x has been modified or |
158 | * destroyed. This is a bug: http://gcc.gnu.org/PR51823 |
159 | */ |
160 | reference |
161 | operator*() const |
162 | { |
163 | _Iterator __tmp = current; |
164 | return *--__tmp; |
165 | } |
166 | |
167 | /** |
168 | * @return A pointer to the value at @c --current |
169 | * |
170 | * This requires that @c --current is dereferenceable. |
171 | */ |
172 | pointer |
173 | operator->() const |
174 | { return &(operator*()); } |
175 | |
176 | /** |
177 | * @return @c *this |
178 | * |
179 | * Decrements the underlying iterator. |
180 | */ |
181 | reverse_iterator& |
182 | operator++() |
183 | { |
184 | --current; |
185 | return *this; |
186 | } |
187 | |
188 | /** |
189 | * @return The original value of @c *this |
190 | * |
191 | * Decrements the underlying iterator. |
192 | */ |
193 | reverse_iterator |
194 | operator++(int) |
195 | { |
196 | reverse_iterator __tmp = *this; |
197 | --current; |
198 | return __tmp; |
199 | } |
200 | |
201 | /** |
202 | * @return @c *this |
203 | * |
204 | * Increments the underlying iterator. |
205 | */ |
206 | reverse_iterator& |
207 | operator--() |
208 | { |
209 | ++current; |
210 | return *this; |
211 | } |
212 | |
213 | /** |
214 | * @return A reverse_iterator with the previous value of @c *this |
215 | * |
216 | * Increments the underlying iterator. |
217 | */ |
218 | reverse_iterator |
219 | operator--(int) |
220 | { |
221 | reverse_iterator __tmp = *this; |
222 | ++current; |
223 | return __tmp; |
224 | } |
225 | |
226 | /** |
227 | * @return A reverse_iterator that refers to @c current - @a __n |
228 | * |
229 | * The underlying iterator must be a Random Access Iterator. |
230 | */ |
231 | reverse_iterator |
232 | operator+(difference_type __n) const |
233 | { return reverse_iterator(current - __n); } |
234 | |
235 | /** |
236 | * @return *this |
237 | * |
238 | * Moves the underlying iterator backwards @a __n steps. |
239 | * The underlying iterator must be a Random Access Iterator. |
240 | */ |
241 | reverse_iterator& |
242 | operator+=(difference_type __n) |
243 | { |
244 | current -= __n; |
245 | return *this; |
246 | } |
247 | |
248 | /** |
249 | * @return A reverse_iterator that refers to @c current - @a __n |
250 | * |
251 | * The underlying iterator must be a Random Access Iterator. |
252 | */ |
253 | reverse_iterator |
254 | operator-(difference_type __n) const |
255 | { return reverse_iterator(current + __n); } |
256 | |
257 | /** |
258 | * @return *this |
259 | * |
260 | * Moves the underlying iterator forwards @a __n steps. |
261 | * The underlying iterator must be a Random Access Iterator. |
262 | */ |
263 | reverse_iterator& |
264 | operator-=(difference_type __n) |
265 | { |
266 | current += __n; |
267 | return *this; |
268 | } |
269 | |
270 | /** |
271 | * @return The value at @c current - @a __n - 1 |
272 | * |
273 | * The underlying iterator must be a Random Access Iterator. |
274 | */ |
275 | reference |
276 | operator[](difference_type __n) const |
277 | { return *(*this + __n); } |
278 | }; |
279 | |
280 | //@{ |
281 | /** |
282 | * @param __x A %reverse_iterator. |
283 | * @param __y A %reverse_iterator. |
284 | * @return A simple bool. |
285 | * |
286 | * Reverse iterators forward many operations to their underlying base() |
287 | * iterators. Others are implemented in terms of one another. |
288 | * |
289 | */ |
290 | template<typename _Iterator> |
291 | inline bool |
292 | operator==(const reverse_iterator<_Iterator>& __x, |
293 | const reverse_iterator<_Iterator>& __y) |
294 | { return __x.base() == __y.base(); } |
295 | |
296 | template<typename _Iterator> |
297 | inline bool |
298 | operator<(const reverse_iterator<_Iterator>& __x, |
299 | const reverse_iterator<_Iterator>& __y) |
300 | { return __y.base() < __x.base(); } |
301 | |
302 | template<typename _Iterator> |
303 | inline bool |
304 | operator!=(const reverse_iterator<_Iterator>& __x, |
305 | const reverse_iterator<_Iterator>& __y) |
306 | { return !(__x == __y); } |
307 | |
308 | template<typename _Iterator> |
309 | inline bool |
310 | operator>(const reverse_iterator<_Iterator>& __x, |
311 | const reverse_iterator<_Iterator>& __y) |
312 | { return __y < __x; } |
313 | |
314 | template<typename _Iterator> |
315 | inline bool |
316 | operator<=(const reverse_iterator<_Iterator>& __x, |
317 | const reverse_iterator<_Iterator>& __y) |
318 | { return !(__y < __x); } |
319 | |
320 | template<typename _Iterator> |
321 | inline bool |
322 | operator>=(const reverse_iterator<_Iterator>& __x, |
323 | const reverse_iterator<_Iterator>& __y) |
324 | { return !(__x < __y); } |
325 | |
326 | template<typename _Iterator> |
327 | #if __cplusplus201402L < 201103L |
328 | inline typename reverse_iterator<_Iterator>::difference_type |
329 | operator-(const reverse_iterator<_Iterator>& __x, |
330 | const reverse_iterator<_Iterator>& __y) |
331 | #else |
332 | inline auto |
333 | operator-(const reverse_iterator<_Iterator>& __x, |
334 | const reverse_iterator<_Iterator>& __y) |
335 | -> decltype(__x.base() - __y.base()) |
336 | #endif |
337 | { return __y.base() - __x.base(); } |
338 | |
339 | template<typename _Iterator> |
340 | inline reverse_iterator<_Iterator> |
341 | operator+(typename reverse_iterator<_Iterator>::difference_type __n, |
342 | const reverse_iterator<_Iterator>& __x) |
343 | { return reverse_iterator<_Iterator>(__x.base() - __n); } |
344 | |
345 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
346 | // DR 280. Comparison of reverse_iterator to const reverse_iterator. |
347 | template<typename _IteratorL, typename _IteratorR> |
348 | inline bool |
349 | operator==(const reverse_iterator<_IteratorL>& __x, |
350 | const reverse_iterator<_IteratorR>& __y) |
351 | { return __x.base() == __y.base(); } |
352 | |
353 | template<typename _IteratorL, typename _IteratorR> |
354 | inline bool |
355 | operator<(const reverse_iterator<_IteratorL>& __x, |
356 | const reverse_iterator<_IteratorR>& __y) |
357 | { return __y.base() < __x.base(); } |
358 | |
359 | template<typename _IteratorL, typename _IteratorR> |
360 | inline bool |
361 | operator!=(const reverse_iterator<_IteratorL>& __x, |
362 | const reverse_iterator<_IteratorR>& __y) |
363 | { return !(__x == __y); } |
364 | |
365 | template<typename _IteratorL, typename _IteratorR> |
366 | inline bool |
367 | operator>(const reverse_iterator<_IteratorL>& __x, |
368 | const reverse_iterator<_IteratorR>& __y) |
369 | { return __y < __x; } |
370 | |
371 | template<typename _IteratorL, typename _IteratorR> |
372 | inline bool |
373 | operator<=(const reverse_iterator<_IteratorL>& __x, |
374 | const reverse_iterator<_IteratorR>& __y) |
375 | { return !(__y < __x); } |
376 | |
377 | template<typename _IteratorL, typename _IteratorR> |
378 | inline bool |
379 | operator>=(const reverse_iterator<_IteratorL>& __x, |
380 | const reverse_iterator<_IteratorR>& __y) |
381 | { return !(__x < __y); } |
382 | |
383 | template<typename _IteratorL, typename _IteratorR> |
384 | #if __cplusplus201402L >= 201103L |
385 | // DR 685. |
386 | inline auto |
387 | operator-(const reverse_iterator<_IteratorL>& __x, |
388 | const reverse_iterator<_IteratorR>& __y) |
389 | -> decltype(__y.base() - __x.base()) |
390 | #else |
391 | inline typename reverse_iterator<_IteratorL>::difference_type |
392 | operator-(const reverse_iterator<_IteratorL>& __x, |
393 | const reverse_iterator<_IteratorR>& __y) |
394 | #endif |
395 | { return __y.base() - __x.base(); } |
396 | //@} |
397 | |
398 | #if __cplusplus201402L >= 201103L |
399 | // Same as C++14 make_reverse_iterator but used in C++03 mode too. |
400 | template<typename _Iterator> |
401 | inline reverse_iterator<_Iterator> |
402 | __make_reverse_iterator(_Iterator __i) |
403 | { return reverse_iterator<_Iterator>(__i); } |
404 | |
405 | # if __cplusplus201402L > 201103L |
406 | # define __cpp_lib_make_reverse_iterator201402 201402 |
407 | |
408 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
409 | // DR 2285. make_reverse_iterator |
410 | /// Generator function for reverse_iterator. |
411 | template<typename _Iterator> |
412 | inline reverse_iterator<_Iterator> |
413 | make_reverse_iterator(_Iterator __i) |
414 | { return reverse_iterator<_Iterator>(__i); } |
415 | # endif |
416 | #endif |
417 | |
418 | #if __cplusplus201402L >= 201103L |
419 | template<typename _Iterator> |
420 | auto |
421 | __niter_base(reverse_iterator<_Iterator> __it) |
422 | -> decltype(__make_reverse_iterator(__niter_base(__it.base()))) |
423 | { return __make_reverse_iterator(__niter_base(__it.base())); } |
424 | |
425 | template<typename _Iterator> |
426 | struct __is_move_iterator<reverse_iterator<_Iterator> > |
427 | : __is_move_iterator<_Iterator> |
428 | { }; |
429 | |
430 | template<typename _Iterator> |
431 | auto |
432 | __miter_base(reverse_iterator<_Iterator> __it) |
433 | -> decltype(__make_reverse_iterator(__miter_base(__it.base()))) |
434 | { return __make_reverse_iterator(__miter_base(__it.base())); } |
435 | #endif |
436 | |
437 | // 24.4.2.2.1 back_insert_iterator |
438 | /** |
439 | * @brief Turns assignment into insertion. |
440 | * |
441 | * These are output iterators, constructed from a container-of-T. |
442 | * Assigning a T to the iterator appends it to the container using |
443 | * push_back. |
444 | * |
445 | * Tip: Using the back_inserter function to create these iterators can |
446 | * save typing. |
447 | */ |
448 | template<typename _Container> |
449 | class back_insert_iterator |
450 | : public iterator<output_iterator_tag, void, void, void, void> |
451 | { |
452 | protected: |
453 | _Container* container; |
454 | |
455 | public: |
456 | /// A nested typedef for the type of whatever container you used. |
457 | typedef _Container container_type; |
458 | |
459 | /// The only way to create this %iterator is with a container. |
460 | explicit |
461 | back_insert_iterator(_Container& __x) |
462 | : container(std::__addressof(__x)) { } |
463 | |
464 | /** |
465 | * @param __value An instance of whatever type |
466 | * container_type::const_reference is; presumably a |
467 | * reference-to-const T for container<T>. |
468 | * @return This %iterator, for chained operations. |
469 | * |
470 | * This kind of %iterator doesn't really have a @a position in the |
471 | * container (you can think of the position as being permanently at |
472 | * the end, if you like). Assigning a value to the %iterator will |
473 | * always append the value to the end of the container. |
474 | */ |
475 | #if __cplusplus201402L < 201103L |
476 | back_insert_iterator& |
477 | operator=(typename _Container::const_reference __value) |
478 | { |
479 | container->push_back(__value); |
480 | return *this; |
481 | } |
482 | #else |
483 | back_insert_iterator& |
484 | operator=(const typename _Container::value_type& __value) |
485 | { |
486 | container->push_back(__value); |
487 | return *this; |
488 | } |
489 | |
490 | back_insert_iterator& |
491 | operator=(typename _Container::value_type&& __value) |
492 | { |
493 | container->push_back(std::move(__value)); |
494 | return *this; |
495 | } |
496 | #endif |
497 | |
498 | /// Simply returns *this. |
499 | back_insert_iterator& |
500 | operator*() |
501 | { return *this; } |
502 | |
503 | /// Simply returns *this. (This %iterator does not @a move.) |
504 | back_insert_iterator& |
505 | operator++() |
506 | { return *this; } |
507 | |
508 | /// Simply returns *this. (This %iterator does not @a move.) |
509 | back_insert_iterator |
510 | operator++(int) |
511 | { return *this; } |
512 | }; |
513 | |
514 | /** |
515 | * @param __x A container of arbitrary type. |
516 | * @return An instance of back_insert_iterator working on @p __x. |
517 | * |
518 | * This wrapper function helps in creating back_insert_iterator instances. |
519 | * Typing the name of the %iterator requires knowing the precise full |
520 | * type of the container, which can be tedious and impedes generic |
521 | * programming. Using this function lets you take advantage of automatic |
522 | * template parameter deduction, making the compiler match the correct |
523 | * types for you. |
524 | */ |
525 | template<typename _Container> |
526 | inline back_insert_iterator<_Container> |
527 | back_inserter(_Container& __x) |
528 | { return back_insert_iterator<_Container>(__x); } |
529 | |
530 | /** |
531 | * @brief Turns assignment into insertion. |
532 | * |
533 | * These are output iterators, constructed from a container-of-T. |
534 | * Assigning a T to the iterator prepends it to the container using |
535 | * push_front. |
536 | * |
537 | * Tip: Using the front_inserter function to create these iterators can |
538 | * save typing. |
539 | */ |
540 | template<typename _Container> |
541 | class front_insert_iterator |
542 | : public iterator<output_iterator_tag, void, void, void, void> |
543 | { |
544 | protected: |
545 | _Container* container; |
546 | |
547 | public: |
548 | /// A nested typedef for the type of whatever container you used. |
549 | typedef _Container container_type; |
550 | |
551 | /// The only way to create this %iterator is with a container. |
552 | explicit front_insert_iterator(_Container& __x) |
553 | : container(std::__addressof(__x)) { } |
554 | |
555 | /** |
556 | * @param __value An instance of whatever type |
557 | * container_type::const_reference is; presumably a |
558 | * reference-to-const T for container<T>. |
559 | * @return This %iterator, for chained operations. |
560 | * |
561 | * This kind of %iterator doesn't really have a @a position in the |
562 | * container (you can think of the position as being permanently at |
563 | * the front, if you like). Assigning a value to the %iterator will |
564 | * always prepend the value to the front of the container. |
565 | */ |
566 | #if __cplusplus201402L < 201103L |
567 | front_insert_iterator& |
568 | operator=(typename _Container::const_reference __value) |
569 | { |
570 | container->push_front(__value); |
571 | return *this; |
572 | } |
573 | #else |
574 | front_insert_iterator& |
575 | operator=(const typename _Container::value_type& __value) |
576 | { |
577 | container->push_front(__value); |
578 | return *this; |
579 | } |
580 | |
581 | front_insert_iterator& |
582 | operator=(typename _Container::value_type&& __value) |
583 | { |
584 | container->push_front(std::move(__value)); |
585 | return *this; |
586 | } |
587 | #endif |
588 | |
589 | /// Simply returns *this. |
590 | front_insert_iterator& |
591 | operator*() |
592 | { return *this; } |
593 | |
594 | /// Simply returns *this. (This %iterator does not @a move.) |
595 | front_insert_iterator& |
596 | operator++() |
597 | { return *this; } |
598 | |
599 | /// Simply returns *this. (This %iterator does not @a move.) |
600 | front_insert_iterator |
601 | operator++(int) |
602 | { return *this; } |
603 | }; |
604 | |
605 | /** |
606 | * @param __x A container of arbitrary type. |
607 | * @return An instance of front_insert_iterator working on @p x. |
608 | * |
609 | * This wrapper function helps in creating front_insert_iterator instances. |
610 | * Typing the name of the %iterator requires knowing the precise full |
611 | * type of the container, which can be tedious and impedes generic |
612 | * programming. Using this function lets you take advantage of automatic |
613 | * template parameter deduction, making the compiler match the correct |
614 | * types for you. |
615 | */ |
616 | template<typename _Container> |
617 | inline front_insert_iterator<_Container> |
618 | front_inserter(_Container& __x) |
619 | { return front_insert_iterator<_Container>(__x); } |
620 | |
621 | /** |
622 | * @brief Turns assignment into insertion. |
623 | * |
624 | * These are output iterators, constructed from a container-of-T. |
625 | * Assigning a T to the iterator inserts it in the container at the |
626 | * %iterator's position, rather than overwriting the value at that |
627 | * position. |
628 | * |
629 | * (Sequences will actually insert a @e copy of the value before the |
630 | * %iterator's position.) |
631 | * |
632 | * Tip: Using the inserter function to create these iterators can |
633 | * save typing. |
634 | */ |
635 | template<typename _Container> |
636 | class insert_iterator |
637 | : public iterator<output_iterator_tag, void, void, void, void> |
638 | { |
639 | protected: |
640 | _Container* container; |
641 | typename _Container::iterator iter; |
642 | |
643 | public: |
644 | /// A nested typedef for the type of whatever container you used. |
645 | typedef _Container container_type; |
646 | |
647 | /** |
648 | * The only way to create this %iterator is with a container and an |
649 | * initial position (a normal %iterator into the container). |
650 | */ |
651 | insert_iterator(_Container& __x, typename _Container::iterator __i) |
652 | : container(std::__addressof(__x)), iter(__i) {} |
653 | |
654 | /** |
655 | * @param __value An instance of whatever type |
656 | * container_type::const_reference is; presumably a |
657 | * reference-to-const T for container<T>. |
658 | * @return This %iterator, for chained operations. |
659 | * |
660 | * This kind of %iterator maintains its own position in the |
661 | * container. Assigning a value to the %iterator will insert the |
662 | * value into the container at the place before the %iterator. |
663 | * |
664 | * The position is maintained such that subsequent assignments will |
665 | * insert values immediately after one another. For example, |
666 | * @code |
667 | * // vector v contains A and Z |
668 | * |
669 | * insert_iterator i (v, ++v.begin()); |
670 | * i = 1; |
671 | * i = 2; |
672 | * i = 3; |
673 | * |
674 | * // vector v contains A, 1, 2, 3, and Z |
675 | * @endcode |
676 | */ |
677 | #if __cplusplus201402L < 201103L |
678 | insert_iterator& |
679 | operator=(typename _Container::const_reference __value) |
680 | { |
681 | iter = container->insert(iter, __value); |
682 | ++iter; |
683 | return *this; |
684 | } |
685 | #else |
686 | insert_iterator& |
687 | operator=(const typename _Container::value_type& __value) |
688 | { |
689 | iter = container->insert(iter, __value); |
690 | ++iter; |
691 | return *this; |
692 | } |
693 | |
694 | insert_iterator& |
695 | operator=(typename _Container::value_type&& __value) |
696 | { |
697 | iter = container->insert(iter, std::move(__value)); |
698 | ++iter; |
699 | return *this; |
700 | } |
701 | #endif |
702 | |
703 | /// Simply returns *this. |
704 | insert_iterator& |
705 | operator*() |
706 | { return *this; } |
707 | |
708 | /// Simply returns *this. (This %iterator does not @a move.) |
709 | insert_iterator& |
710 | operator++() |
711 | { return *this; } |
712 | |
713 | /// Simply returns *this. (This %iterator does not @a move.) |
714 | insert_iterator& |
715 | operator++(int) |
716 | { return *this; } |
717 | }; |
718 | |
719 | /** |
720 | * @param __x A container of arbitrary type. |
721 | * @return An instance of insert_iterator working on @p __x. |
722 | * |
723 | * This wrapper function helps in creating insert_iterator instances. |
724 | * Typing the name of the %iterator requires knowing the precise full |
725 | * type of the container, which can be tedious and impedes generic |
726 | * programming. Using this function lets you take advantage of automatic |
727 | * template parameter deduction, making the compiler match the correct |
728 | * types for you. |
729 | */ |
730 | template<typename _Container, typename _Iterator> |
731 | inline insert_iterator<_Container> |
732 | inserter(_Container& __x, _Iterator __i) |
733 | { |
734 | return insert_iterator<_Container>(__x, |
735 | typename _Container::iterator(__i)); |
736 | } |
737 | |
738 | // @} group iterators |
739 | |
740 | _GLIBCXX_END_NAMESPACE_VERSION |
741 | } // namespace |
742 | |
743 | namespace __gnu_cxx _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
744 | { |
745 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
746 | |
747 | // This iterator adapter is @a normal in the sense that it does not |
748 | // change the semantics of any of the operators of its iterator |
749 | // parameter. Its primary purpose is to convert an iterator that is |
750 | // not a class, e.g. a pointer, into an iterator that is a class. |
751 | // The _Container parameter exists solely so that different containers |
752 | // using this template can instantiate different types, even if the |
753 | // _Iterator parameter is the same. |
754 | using std::iterator_traits; |
755 | using std::iterator; |
756 | template<typename _Iterator, typename _Container> |
757 | class __normal_iterator |
758 | { |
759 | protected: |
760 | _Iterator _M_current; |
761 | |
762 | typedef iterator_traits<_Iterator> __traits_type; |
763 | |
764 | public: |
765 | typedef _Iterator iterator_type; |
766 | typedef typename __traits_type::iterator_category iterator_category; |
767 | typedef typename __traits_type::value_type value_type; |
768 | typedef typename __traits_type::difference_type difference_type; |
769 | typedef typename __traits_type::reference reference; |
770 | typedef typename __traits_type::pointer pointer; |
771 | |
772 | _GLIBCXX_CONSTEXPRconstexpr __normal_iterator() _GLIBCXX_NOEXCEPTnoexcept |
773 | : _M_current(_Iterator()) { } |
774 | |
775 | explicit |
776 | __normal_iterator(const _Iterator& __i) _GLIBCXX_NOEXCEPTnoexcept |
777 | : _M_current(__i) { } |
778 | |
779 | // Allow iterator to const_iterator conversion |
780 | template<typename _Iter> |
781 | __normal_iterator(const __normal_iterator<_Iter, |
782 | typename __enable_if< |
783 | (std::__are_same<_Iter, typename _Container::pointer>::__value), |
784 | _Container>::__type>& __i) _GLIBCXX_NOEXCEPTnoexcept |
785 | : _M_current(__i.base()) { } |
786 | |
787 | // Forward iterator requirements |
788 | reference |
789 | operator*() const _GLIBCXX_NOEXCEPTnoexcept |
790 | { return *_M_current; } |
791 | |
792 | pointer |
793 | operator->() const _GLIBCXX_NOEXCEPTnoexcept |
794 | { return _M_current; } |
795 | |
796 | __normal_iterator& |
797 | operator++() _GLIBCXX_NOEXCEPTnoexcept |
798 | { |
799 | ++_M_current; |
800 | return *this; |
801 | } |
802 | |
803 | __normal_iterator |
804 | operator++(int) _GLIBCXX_NOEXCEPTnoexcept |
805 | { return __normal_iterator(_M_current++); } |
806 | |
807 | // Bidirectional iterator requirements |
808 | __normal_iterator& |
809 | operator--() _GLIBCXX_NOEXCEPTnoexcept |
810 | { |
811 | --_M_current; |
812 | return *this; |
813 | } |
814 | |
815 | __normal_iterator |
816 | operator--(int) _GLIBCXX_NOEXCEPTnoexcept |
817 | { return __normal_iterator(_M_current--); } |
818 | |
819 | // Random access iterator requirements |
820 | reference |
821 | operator[](difference_type __n) const _GLIBCXX_NOEXCEPTnoexcept |
822 | { return _M_current[__n]; } |
823 | |
824 | __normal_iterator& |
825 | operator+=(difference_type __n) _GLIBCXX_NOEXCEPTnoexcept |
826 | { _M_current += __n; return *this; } |
827 | |
828 | __normal_iterator |
829 | operator+(difference_type __n) const _GLIBCXX_NOEXCEPTnoexcept |
830 | { return __normal_iterator(_M_current + __n); } |
831 | |
832 | __normal_iterator& |
833 | operator-=(difference_type __n) _GLIBCXX_NOEXCEPTnoexcept |
834 | { _M_current -= __n; return *this; } |
835 | |
836 | __normal_iterator |
837 | operator-(difference_type __n) const _GLIBCXX_NOEXCEPTnoexcept |
838 | { return __normal_iterator(_M_current - __n); } |
839 | |
840 | const _Iterator& |
841 | base() const _GLIBCXX_NOEXCEPTnoexcept |
842 | { return _M_current; } |
843 | }; |
844 | |
845 | // Note: In what follows, the left- and right-hand-side iterators are |
846 | // allowed to vary in types (conceptually in cv-qualification) so that |
847 | // comparison between cv-qualified and non-cv-qualified iterators be |
848 | // valid. However, the greedy and unfriendly operators in std::rel_ops |
849 | // will make overload resolution ambiguous (when in scope) if we don't |
850 | // provide overloads whose operands are of the same type. Can someone |
851 | // remind me what generic programming is about? -- Gaby |
852 | |
853 | // Forward iterator requirements |
854 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
855 | inline bool |
856 | operator==(const __normal_iterator<_IteratorL, _Container>& __lhs, |
857 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
858 | _GLIBCXX_NOEXCEPTnoexcept |
859 | { return __lhs.base() == __rhs.base(); } |
860 | |
861 | template<typename _Iterator, typename _Container> |
862 | inline bool |
863 | operator==(const __normal_iterator<_Iterator, _Container>& __lhs, |
864 | const __normal_iterator<_Iterator, _Container>& __rhs) |
865 | _GLIBCXX_NOEXCEPTnoexcept |
866 | { return __lhs.base() == __rhs.base(); } |
867 | |
868 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
869 | inline bool |
870 | operator!=(const __normal_iterator<_IteratorL, _Container>& __lhs, |
871 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
872 | _GLIBCXX_NOEXCEPTnoexcept |
873 | { return __lhs.base() != __rhs.base(); } |
874 | |
875 | template<typename _Iterator, typename _Container> |
876 | inline bool |
877 | operator!=(const __normal_iterator<_Iterator, _Container>& __lhs, |
878 | const __normal_iterator<_Iterator, _Container>& __rhs) |
879 | _GLIBCXX_NOEXCEPTnoexcept |
880 | { return __lhs.base() != __rhs.base(); } |
881 | |
882 | // Random access iterator requirements |
883 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
884 | inline bool |
885 | operator<(const __normal_iterator<_IteratorL, _Container>& __lhs, |
886 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
887 | _GLIBCXX_NOEXCEPTnoexcept |
888 | { return __lhs.base() < __rhs.base(); } |
889 | |
890 | template<typename _Iterator, typename _Container> |
891 | inline bool |
892 | operator<(const __normal_iterator<_Iterator, _Container>& __lhs, |
893 | const __normal_iterator<_Iterator, _Container>& __rhs) |
894 | _GLIBCXX_NOEXCEPTnoexcept |
895 | { return __lhs.base() < __rhs.base(); } |
896 | |
897 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
898 | inline bool |
899 | operator>(const __normal_iterator<_IteratorL, _Container>& __lhs, |
900 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
901 | _GLIBCXX_NOEXCEPTnoexcept |
902 | { return __lhs.base() > __rhs.base(); } |
903 | |
904 | template<typename _Iterator, typename _Container> |
905 | inline bool |
906 | operator>(const __normal_iterator<_Iterator, _Container>& __lhs, |
907 | const __normal_iterator<_Iterator, _Container>& __rhs) |
908 | _GLIBCXX_NOEXCEPTnoexcept |
909 | { return __lhs.base() > __rhs.base(); } |
910 | |
911 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
912 | inline bool |
913 | operator<=(const __normal_iterator<_IteratorL, _Container>& __lhs, |
914 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
915 | _GLIBCXX_NOEXCEPTnoexcept |
916 | { return __lhs.base() <= __rhs.base(); } |
917 | |
918 | template<typename _Iterator, typename _Container> |
919 | inline bool |
920 | operator<=(const __normal_iterator<_Iterator, _Container>& __lhs, |
921 | const __normal_iterator<_Iterator, _Container>& __rhs) |
922 | _GLIBCXX_NOEXCEPTnoexcept |
923 | { return __lhs.base() <= __rhs.base(); } |
924 | |
925 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
926 | inline bool |
927 | operator>=(const __normal_iterator<_IteratorL, _Container>& __lhs, |
928 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
929 | _GLIBCXX_NOEXCEPTnoexcept |
930 | { return __lhs.base() >= __rhs.base(); } |
931 | |
932 | template<typename _Iterator, typename _Container> |
933 | inline bool |
934 | operator>=(const __normal_iterator<_Iterator, _Container>& __lhs, |
935 | const __normal_iterator<_Iterator, _Container>& __rhs) |
936 | _GLIBCXX_NOEXCEPTnoexcept |
937 | { return __lhs.base() >= __rhs.base(); } |
938 | |
939 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
940 | // According to the resolution of DR179 not only the various comparison |
941 | // operators but also operator- must accept mixed iterator/const_iterator |
942 | // parameters. |
943 | template<typename _IteratorL, typename _IteratorR, typename _Container> |
944 | #if __cplusplus201402L >= 201103L |
945 | // DR 685. |
946 | inline auto |
947 | operator-(const __normal_iterator<_IteratorL, _Container>& __lhs, |
948 | const __normal_iterator<_IteratorR, _Container>& __rhs) noexcept |
949 | -> decltype(__lhs.base() - __rhs.base()) |
950 | #else |
951 | inline typename __normal_iterator<_IteratorL, _Container>::difference_type |
952 | operator-(const __normal_iterator<_IteratorL, _Container>& __lhs, |
953 | const __normal_iterator<_IteratorR, _Container>& __rhs) |
954 | #endif |
955 | { return __lhs.base() - __rhs.base(); } |
956 | |
957 | template<typename _Iterator, typename _Container> |
958 | inline typename __normal_iterator<_Iterator, _Container>::difference_type |
959 | operator-(const __normal_iterator<_Iterator, _Container>& __lhs, |
960 | const __normal_iterator<_Iterator, _Container>& __rhs) |
961 | _GLIBCXX_NOEXCEPTnoexcept |
962 | { return __lhs.base() - __rhs.base(); } |
963 | |
964 | template<typename _Iterator, typename _Container> |
965 | inline __normal_iterator<_Iterator, _Container> |
966 | operator+(typename __normal_iterator<_Iterator, _Container>::difference_type |
967 | __n, const __normal_iterator<_Iterator, _Container>& __i) |
968 | _GLIBCXX_NOEXCEPTnoexcept |
969 | { return __normal_iterator<_Iterator, _Container>(__i.base() + __n); } |
970 | |
971 | _GLIBCXX_END_NAMESPACE_VERSION |
972 | } // namespace |
973 | |
974 | namespace std _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
975 | { |
976 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
977 | |
978 | template<typename _Iterator, typename _Container> |
979 | _Iterator |
980 | __niter_base(__gnu_cxx::__normal_iterator<_Iterator, _Container> __it) |
981 | { return __it.base(); } |
982 | |
983 | _GLIBCXX_END_NAMESPACE_VERSION |
984 | } // namespace |
985 | |
986 | #if __cplusplus201402L >= 201103L |
987 | |
988 | namespace std _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
989 | { |
990 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
991 | |
992 | /** |
993 | * @addtogroup iterators |
994 | * @{ |
995 | */ |
996 | |
997 | // 24.4.3 Move iterators |
998 | /** |
999 | * Class template move_iterator is an iterator adapter with the same |
1000 | * behavior as the underlying iterator except that its dereference |
1001 | * operator implicitly converts the value returned by the underlying |
1002 | * iterator's dereference operator to an rvalue reference. Some |
1003 | * generic algorithms can be called with move iterators to replace |
1004 | * copying with moving. |
1005 | */ |
1006 | template<typename _Iterator> |
1007 | class move_iterator |
1008 | { |
1009 | protected: |
1010 | _Iterator _M_current; |
1011 | |
1012 | typedef iterator_traits<_Iterator> __traits_type; |
1013 | typedef typename __traits_type::reference __base_ref; |
1014 | |
1015 | public: |
1016 | typedef _Iterator iterator_type; |
1017 | typedef typename __traits_type::iterator_category iterator_category; |
1018 | typedef typename __traits_type::value_type value_type; |
1019 | typedef typename __traits_type::difference_type difference_type; |
1020 | // NB: DR 680. |
1021 | typedef _Iterator pointer; |
1022 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
1023 | // 2106. move_iterator wrapping iterators returning prvalues |
1024 | typedef typename conditional<is_reference<__base_ref>::value, |
1025 | typename remove_reference<__base_ref>::type&&, |
1026 | __base_ref>::type reference; |
1027 | |
1028 | move_iterator() |
1029 | : _M_current() { } |
1030 | |
1031 | explicit |
1032 | move_iterator(iterator_type __i) |
1033 | : _M_current(__i) { } |
1034 | |
1035 | template<typename _Iter> |
1036 | move_iterator(const move_iterator<_Iter>& __i) |
1037 | : _M_current(__i.base()) { } |
1038 | |
1039 | iterator_type |
1040 | base() const |
1041 | { return _M_current; } |
1042 | |
1043 | reference |
1044 | operator*() const |
1045 | { return static_cast<reference>(*_M_current); } |
1046 | |
1047 | pointer |
1048 | operator->() const |
1049 | { return _M_current; } |
1050 | |
1051 | move_iterator& |
1052 | operator++() |
1053 | { |
1054 | ++_M_current; |
1055 | return *this; |
1056 | } |
1057 | |
1058 | move_iterator |
1059 | operator++(int) |
1060 | { |
1061 | move_iterator __tmp = *this; |
1062 | ++_M_current; |
1063 | return __tmp; |
1064 | } |
1065 | |
1066 | move_iterator& |
1067 | operator--() |
1068 | { |
1069 | --_M_current; |
1070 | return *this; |
1071 | } |
1072 | |
1073 | move_iterator |
1074 | operator--(int) |
1075 | { |
1076 | move_iterator __tmp = *this; |
1077 | --_M_current; |
1078 | return __tmp; |
1079 | } |
1080 | |
1081 | move_iterator |
1082 | operator+(difference_type __n) const |
1083 | { return move_iterator(_M_current + __n); } |
1084 | |
1085 | move_iterator& |
1086 | operator+=(difference_type __n) |
1087 | { |
1088 | _M_current += __n; |
1089 | return *this; |
1090 | } |
1091 | |
1092 | move_iterator |
1093 | operator-(difference_type __n) const |
1094 | { return move_iterator(_M_current - __n); } |
1095 | |
1096 | move_iterator& |
1097 | operator-=(difference_type __n) |
1098 | { |
1099 | _M_current -= __n; |
1100 | return *this; |
1101 | } |
1102 | |
1103 | reference |
1104 | operator[](difference_type __n) const |
1105 | { return std::move(_M_current[__n]); } |
1106 | }; |
1107 | |
1108 | // Note: See __normal_iterator operators note from Gaby to understand |
1109 | // why there are always 2 versions for most of the move_iterator |
1110 | // operators. |
1111 | template<typename _IteratorL, typename _IteratorR> |
1112 | inline bool |
1113 | operator==(const move_iterator<_IteratorL>& __x, |
1114 | const move_iterator<_IteratorR>& __y) |
1115 | { return __x.base() == __y.base(); } |
1116 | |
1117 | template<typename _Iterator> |
1118 | inline bool |
1119 | operator==(const move_iterator<_Iterator>& __x, |
1120 | const move_iterator<_Iterator>& __y) |
1121 | { return __x.base() == __y.base(); } |
1122 | |
1123 | template<typename _IteratorL, typename _IteratorR> |
1124 | inline bool |
1125 | operator!=(const move_iterator<_IteratorL>& __x, |
1126 | const move_iterator<_IteratorR>& __y) |
1127 | { return !(__x == __y); } |
1128 | |
1129 | template<typename _Iterator> |
1130 | inline bool |
1131 | operator!=(const move_iterator<_Iterator>& __x, |
1132 | const move_iterator<_Iterator>& __y) |
1133 | { return !(__x == __y); } |
1134 | |
1135 | template<typename _IteratorL, typename _IteratorR> |
1136 | inline bool |
1137 | operator<(const move_iterator<_IteratorL>& __x, |
1138 | const move_iterator<_IteratorR>& __y) |
1139 | { return __x.base() < __y.base(); } |
1140 | |
1141 | template<typename _Iterator> |
1142 | inline bool |
1143 | operator<(const move_iterator<_Iterator>& __x, |
1144 | const move_iterator<_Iterator>& __y) |
1145 | { return __x.base() < __y.base(); } |
1146 | |
1147 | template<typename _IteratorL, typename _IteratorR> |
1148 | inline bool |
1149 | operator<=(const move_iterator<_IteratorL>& __x, |
1150 | const move_iterator<_IteratorR>& __y) |
1151 | { return !(__y < __x); } |
1152 | |
1153 | template<typename _Iterator> |
1154 | inline bool |
1155 | operator<=(const move_iterator<_Iterator>& __x, |
1156 | const move_iterator<_Iterator>& __y) |
1157 | { return !(__y < __x); } |
1158 | |
1159 | template<typename _IteratorL, typename _IteratorR> |
1160 | inline bool |
1161 | operator>(const move_iterator<_IteratorL>& __x, |
1162 | const move_iterator<_IteratorR>& __y) |
1163 | { return __y < __x; } |
1164 | |
1165 | template<typename _Iterator> |
1166 | inline bool |
1167 | operator>(const move_iterator<_Iterator>& __x, |
1168 | const move_iterator<_Iterator>& __y) |
1169 | { return __y < __x; } |
1170 | |
1171 | template<typename _IteratorL, typename _IteratorR> |
1172 | inline bool |
1173 | operator>=(const move_iterator<_IteratorL>& __x, |
1174 | const move_iterator<_IteratorR>& __y) |
1175 | { return !(__x < __y); } |
1176 | |
1177 | template<typename _Iterator> |
1178 | inline bool |
1179 | operator>=(const move_iterator<_Iterator>& __x, |
1180 | const move_iterator<_Iterator>& __y) |
1181 | { return !(__x < __y); } |
1182 | |
1183 | // DR 685. |
1184 | template<typename _IteratorL, typename _IteratorR> |
1185 | inline auto |
1186 | operator-(const move_iterator<_IteratorL>& __x, |
1187 | const move_iterator<_IteratorR>& __y) |
1188 | -> decltype(__x.base() - __y.base()) |
1189 | { return __x.base() - __y.base(); } |
1190 | |
1191 | template<typename _Iterator> |
1192 | inline auto |
1193 | operator-(const move_iterator<_Iterator>& __x, |
1194 | const move_iterator<_Iterator>& __y) |
1195 | -> decltype(__x.base() - __y.base()) |
1196 | { return __x.base() - __y.base(); } |
1197 | |
1198 | template<typename _Iterator> |
1199 | inline move_iterator<_Iterator> |
1200 | operator+(typename move_iterator<_Iterator>::difference_type __n, |
1201 | const move_iterator<_Iterator>& __x) |
1202 | { return __x + __n; } |
1203 | |
1204 | template<typename _Iterator> |
1205 | inline move_iterator<_Iterator> |
1206 | make_move_iterator(_Iterator __i) |
1207 | { return move_iterator<_Iterator>(__i); } |
1208 | |
1209 | template<typename _Iterator, typename _ReturnType |
1210 | = typename conditional<__move_if_noexcept_cond |
1211 | <typename iterator_traits<_Iterator>::value_type>::value, |
1212 | _Iterator, move_iterator<_Iterator>>::type> |
1213 | inline _ReturnType |
1214 | __make_move_if_noexcept_iterator(_Iterator __i) |
1215 | { return _ReturnType(__i); } |
1216 | |
1217 | // Overload for pointers that matches std::move_if_noexcept more closely, |
1218 | // returning a constant iterator when we don't want to move. |
1219 | template<typename _Tp, typename _ReturnType |
1220 | = typename conditional<__move_if_noexcept_cond<_Tp>::value, |
1221 | const _Tp*, move_iterator<_Tp*>>::type> |
1222 | inline _ReturnType |
1223 | __make_move_if_noexcept_iterator(_Tp* __i) |
1224 | { return _ReturnType(__i); } |
1225 | |
1226 | // @} group iterators |
1227 | |
1228 | template<typename _Iterator> |
1229 | auto |
1230 | __niter_base(move_iterator<_Iterator> __it) |
1231 | -> decltype(make_move_iterator(__niter_base(__it.base()))) |
1232 | { return make_move_iterator(__niter_base(__it.base())); } |
1233 | |
1234 | template<typename _Iterator> |
1235 | struct __is_move_iterator<move_iterator<_Iterator> > |
1236 | { |
1237 | enum { __value = 1 }; |
1238 | typedef __true_type __type; |
1239 | }; |
1240 | |
1241 | template<typename _Iterator> |
1242 | auto |
1243 | __miter_base(move_iterator<_Iterator> __it) |
1244 | -> decltype(__miter_base(__it.base())) |
1245 | { return __miter_base(__it.base()); } |
1246 | |
1247 | _GLIBCXX_END_NAMESPACE_VERSION |
1248 | } // namespace |
1249 | |
1250 | #define _GLIBCXX_MAKE_MOVE_ITERATOR(_Iter)std::make_move_iterator(_Iter) std::make_move_iterator(_Iter) |
1251 | #define _GLIBCXX_MAKE_MOVE_IF_NOEXCEPT_ITERATOR(_Iter)std::__make_move_if_noexcept_iterator(_Iter) \ |
1252 | std::__make_move_if_noexcept_iterator(_Iter) |
1253 | #else |
1254 | #define _GLIBCXX_MAKE_MOVE_ITERATOR(_Iter)std::make_move_iterator(_Iter) (_Iter) |
1255 | #define _GLIBCXX_MAKE_MOVE_IF_NOEXCEPT_ITERATOR(_Iter)std::__make_move_if_noexcept_iterator(_Iter) (_Iter) |
1256 | #endif // C++11 |
1257 | |
1258 | #ifdef _GLIBCXX_DEBUG |
1259 | # include <debug/stl_iterator.h> |
1260 | #endif |
1261 | |
1262 | #endif |
1 | //===- llvm/IRBuilder.h - Builder for LLVM Instructions ---------*- C++ -*-===// | |||
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 file defines the IRBuilder class, which is used as a convenient way | |||
10 | // to create LLVM instructions with a consistent and simplified interface. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #ifndef LLVM_IR_IRBUILDER_H | |||
15 | #define LLVM_IR_IRBUILDER_H | |||
16 | ||||
17 | #include "llvm-c/Types.h" | |||
18 | #include "llvm/ADT/ArrayRef.h" | |||
19 | #include "llvm/ADT/None.h" | |||
20 | #include "llvm/ADT/StringRef.h" | |||
21 | #include "llvm/ADT/Twine.h" | |||
22 | #include "llvm/IR/BasicBlock.h" | |||
23 | #include "llvm/IR/Constant.h" | |||
24 | #include "llvm/IR/ConstantFolder.h" | |||
25 | #include "llvm/IR/Constants.h" | |||
26 | #include "llvm/IR/DataLayout.h" | |||
27 | #include "llvm/IR/DebugLoc.h" | |||
28 | #include "llvm/IR/DerivedTypes.h" | |||
29 | #include "llvm/IR/Function.h" | |||
30 | #include "llvm/IR/GlobalVariable.h" | |||
31 | #include "llvm/IR/InstrTypes.h" | |||
32 | #include "llvm/IR/Instruction.h" | |||
33 | #include "llvm/IR/Instructions.h" | |||
34 | #include "llvm/IR/IntrinsicInst.h" | |||
35 | #include "llvm/IR/LLVMContext.h" | |||
36 | #include "llvm/IR/Module.h" | |||
37 | #include "llvm/IR/Operator.h" | |||
38 | #include "llvm/IR/Type.h" | |||
39 | #include "llvm/IR/Value.h" | |||
40 | #include "llvm/IR/ValueHandle.h" | |||
41 | #include "llvm/Support/AtomicOrdering.h" | |||
42 | #include "llvm/Support/CBindingWrapping.h" | |||
43 | #include "llvm/Support/Casting.h" | |||
44 | #include <cassert> | |||
45 | #include <cstddef> | |||
46 | #include <cstdint> | |||
47 | #include <functional> | |||
48 | #include <utility> | |||
49 | ||||
50 | namespace llvm { | |||
51 | ||||
52 | class APInt; | |||
53 | class MDNode; | |||
54 | class Use; | |||
55 | ||||
56 | /// This provides the default implementation of the IRBuilder | |||
57 | /// 'InsertHelper' method that is called whenever an instruction is created by | |||
58 | /// IRBuilder and needs to be inserted. | |||
59 | /// | |||
60 | /// By default, this inserts the instruction at the insertion point. | |||
61 | class IRBuilderDefaultInserter { | |||
62 | public: | |||
63 | virtual ~IRBuilderDefaultInserter(); | |||
64 | ||||
65 | virtual void InsertHelper(Instruction *I, const Twine &Name, | |||
66 | BasicBlock *BB, | |||
67 | BasicBlock::iterator InsertPt) const { | |||
68 | if (BB) BB->getInstList().insert(InsertPt, I); | |||
69 | I->setName(Name); | |||
70 | } | |||
71 | }; | |||
72 | ||||
73 | /// Provides an 'InsertHelper' that calls a user-provided callback after | |||
74 | /// performing the default insertion. | |||
75 | class IRBuilderCallbackInserter : public IRBuilderDefaultInserter { | |||
76 | std::function<void(Instruction *)> Callback; | |||
77 | ||||
78 | public: | |||
79 | virtual ~IRBuilderCallbackInserter(); | |||
80 | ||||
81 | IRBuilderCallbackInserter(std::function<void(Instruction *)> Callback) | |||
82 | : Callback(std::move(Callback)) {} | |||
83 | ||||
84 | void InsertHelper(Instruction *I, const Twine &Name, | |||
85 | BasicBlock *BB, | |||
86 | BasicBlock::iterator InsertPt) const override { | |||
87 | IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt); | |||
88 | Callback(I); | |||
89 | } | |||
90 | }; | |||
91 | ||||
92 | /// Common base class shared among various IRBuilders. | |||
93 | class IRBuilderBase { | |||
94 | DebugLoc CurDbgLocation; | |||
95 | ||||
96 | protected: | |||
97 | BasicBlock *BB; | |||
98 | BasicBlock::iterator InsertPt; | |||
99 | LLVMContext &Context; | |||
100 | const IRBuilderFolder &Folder; | |||
101 | const IRBuilderDefaultInserter &Inserter; | |||
102 | ||||
103 | MDNode *DefaultFPMathTag; | |||
104 | FastMathFlags FMF; | |||
105 | ||||
106 | bool IsFPConstrained; | |||
107 | fp::ExceptionBehavior DefaultConstrainedExcept; | |||
108 | RoundingMode DefaultConstrainedRounding; | |||
109 | ||||
110 | ArrayRef<OperandBundleDef> DefaultOperandBundles; | |||
111 | ||||
112 | public: | |||
113 | IRBuilderBase(LLVMContext &context, const IRBuilderFolder &Folder, | |||
114 | const IRBuilderDefaultInserter &Inserter, | |||
115 | MDNode *FPMathTag, ArrayRef<OperandBundleDef> OpBundles) | |||
116 | : Context(context), Folder(Folder), Inserter(Inserter), | |||
117 | DefaultFPMathTag(FPMathTag), IsFPConstrained(false), | |||
118 | DefaultConstrainedExcept(fp::ebStrict), | |||
119 | DefaultConstrainedRounding(RoundingMode::Dynamic), | |||
120 | DefaultOperandBundles(OpBundles) { | |||
121 | ClearInsertionPoint(); | |||
122 | } | |||
123 | ||||
124 | /// Insert and return the specified instruction. | |||
125 | template<typename InstTy> | |||
126 | InstTy *Insert(InstTy *I, const Twine &Name = "") const { | |||
127 | Inserter.InsertHelper(I, Name, BB, InsertPt); | |||
128 | SetInstDebugLocation(I); | |||
129 | return I; | |||
130 | } | |||
131 | ||||
132 | /// No-op overload to handle constants. | |||
133 | Constant *Insert(Constant *C, const Twine& = "") const { | |||
134 | return C; | |||
135 | } | |||
136 | ||||
137 | Value *Insert(Value *V, const Twine &Name = "") const { | |||
138 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
139 | return Insert(I, Name); | |||
140 | assert(isa<Constant>(V))((isa<Constant>(V)) ? static_cast<void> (0) : __assert_fail ("isa<Constant>(V)", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 140, __PRETTY_FUNCTION__)); | |||
141 | return V; | |||
142 | } | |||
143 | ||||
144 | //===--------------------------------------------------------------------===// | |||
145 | // Builder configuration methods | |||
146 | //===--------------------------------------------------------------------===// | |||
147 | ||||
148 | /// Clear the insertion point: created instructions will not be | |||
149 | /// inserted into a block. | |||
150 | void ClearInsertionPoint() { | |||
151 | BB = nullptr; | |||
152 | InsertPt = BasicBlock::iterator(); | |||
153 | } | |||
154 | ||||
155 | BasicBlock *GetInsertBlock() const { return BB; } | |||
156 | BasicBlock::iterator GetInsertPoint() const { return InsertPt; } | |||
157 | LLVMContext &getContext() const { return Context; } | |||
158 | ||||
159 | /// This specifies that created instructions should be appended to the | |||
160 | /// end of the specified block. | |||
161 | void SetInsertPoint(BasicBlock *TheBB) { | |||
162 | BB = TheBB; | |||
163 | InsertPt = BB->end(); | |||
164 | } | |||
165 | ||||
166 | /// This specifies that created instructions should be inserted before | |||
167 | /// the specified instruction. | |||
168 | void SetInsertPoint(Instruction *I) { | |||
169 | BB = I->getParent(); | |||
170 | InsertPt = I->getIterator(); | |||
171 | assert(InsertPt != BB->end() && "Can't read debug loc from end()")((InsertPt != BB->end() && "Can't read debug loc from end()" ) ? static_cast<void> (0) : __assert_fail ("InsertPt != BB->end() && \"Can't read debug loc from end()\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 171, __PRETTY_FUNCTION__)); | |||
172 | SetCurrentDebugLocation(I->getDebugLoc()); | |||
173 | } | |||
174 | ||||
175 | /// This specifies that created instructions should be inserted at the | |||
176 | /// specified point. | |||
177 | void SetInsertPoint(BasicBlock *TheBB, BasicBlock::iterator IP) { | |||
178 | BB = TheBB; | |||
179 | InsertPt = IP; | |||
180 | if (IP != TheBB->end()) | |||
181 | SetCurrentDebugLocation(IP->getDebugLoc()); | |||
182 | } | |||
183 | ||||
184 | /// Set location information used by debugging information. | |||
185 | void SetCurrentDebugLocation(DebugLoc L) { CurDbgLocation = std::move(L); } | |||
186 | ||||
187 | /// Get location information used by debugging information. | |||
188 | const DebugLoc &getCurrentDebugLocation() const { return CurDbgLocation; } | |||
189 | ||||
190 | /// If this builder has a current debug location, set it on the | |||
191 | /// specified instruction. | |||
192 | void SetInstDebugLocation(Instruction *I) const { | |||
193 | if (CurDbgLocation) | |||
194 | I->setDebugLoc(CurDbgLocation); | |||
195 | } | |||
196 | ||||
197 | /// Get the return type of the current function that we're emitting | |||
198 | /// into. | |||
199 | Type *getCurrentFunctionReturnType() const; | |||
200 | ||||
201 | /// InsertPoint - A saved insertion point. | |||
202 | class InsertPoint { | |||
203 | BasicBlock *Block = nullptr; | |||
204 | BasicBlock::iterator Point; | |||
205 | ||||
206 | public: | |||
207 | /// Creates a new insertion point which doesn't point to anything. | |||
208 | InsertPoint() = default; | |||
209 | ||||
210 | /// Creates a new insertion point at the given location. | |||
211 | InsertPoint(BasicBlock *InsertBlock, BasicBlock::iterator InsertPoint) | |||
212 | : Block(InsertBlock), Point(InsertPoint) {} | |||
213 | ||||
214 | /// Returns true if this insert point is set. | |||
215 | bool isSet() const { return (Block != nullptr); } | |||
216 | ||||
217 | BasicBlock *getBlock() const { return Block; } | |||
218 | BasicBlock::iterator getPoint() const { return Point; } | |||
219 | }; | |||
220 | ||||
221 | /// Returns the current insert point. | |||
222 | InsertPoint saveIP() const { | |||
223 | return InsertPoint(GetInsertBlock(), GetInsertPoint()); | |||
224 | } | |||
225 | ||||
226 | /// Returns the current insert point, clearing it in the process. | |||
227 | InsertPoint saveAndClearIP() { | |||
228 | InsertPoint IP(GetInsertBlock(), GetInsertPoint()); | |||
229 | ClearInsertionPoint(); | |||
230 | return IP; | |||
231 | } | |||
232 | ||||
233 | /// Sets the current insert point to a previously-saved location. | |||
234 | void restoreIP(InsertPoint IP) { | |||
235 | if (IP.isSet()) | |||
236 | SetInsertPoint(IP.getBlock(), IP.getPoint()); | |||
237 | else | |||
238 | ClearInsertionPoint(); | |||
239 | } | |||
240 | ||||
241 | /// Get the floating point math metadata being used. | |||
242 | MDNode *getDefaultFPMathTag() const { return DefaultFPMathTag; } | |||
243 | ||||
244 | /// Get the flags to be applied to created floating point ops | |||
245 | FastMathFlags getFastMathFlags() const { return FMF; } | |||
246 | ||||
247 | FastMathFlags &getFastMathFlags() { return FMF; } | |||
248 | ||||
249 | /// Clear the fast-math flags. | |||
250 | void clearFastMathFlags() { FMF.clear(); } | |||
251 | ||||
252 | /// Set the floating point math metadata to be used. | |||
253 | void setDefaultFPMathTag(MDNode *FPMathTag) { DefaultFPMathTag = FPMathTag; } | |||
254 | ||||
255 | /// Set the fast-math flags to be used with generated fp-math operators | |||
256 | void setFastMathFlags(FastMathFlags NewFMF) { FMF = NewFMF; } | |||
257 | ||||
258 | /// Enable/Disable use of constrained floating point math. When | |||
259 | /// enabled the CreateF<op>() calls instead create constrained | |||
260 | /// floating point intrinsic calls. Fast math flags are unaffected | |||
261 | /// by this setting. | |||
262 | void setIsFPConstrained(bool IsCon) { IsFPConstrained = IsCon; } | |||
263 | ||||
264 | /// Query for the use of constrained floating point math | |||
265 | bool getIsFPConstrained() { return IsFPConstrained; } | |||
266 | ||||
267 | /// Set the exception handling to be used with constrained floating point | |||
268 | void setDefaultConstrainedExcept(fp::ExceptionBehavior NewExcept) { | |||
269 | DefaultConstrainedExcept = NewExcept; | |||
270 | } | |||
271 | ||||
272 | /// Set the rounding mode handling to be used with constrained floating point | |||
273 | void setDefaultConstrainedRounding(RoundingMode NewRounding) { | |||
274 | DefaultConstrainedRounding = NewRounding; | |||
275 | } | |||
276 | ||||
277 | /// Get the exception handling used with constrained floating point | |||
278 | fp::ExceptionBehavior getDefaultConstrainedExcept() { | |||
279 | return DefaultConstrainedExcept; | |||
280 | } | |||
281 | ||||
282 | /// Get the rounding mode handling used with constrained floating point | |||
283 | RoundingMode getDefaultConstrainedRounding() { | |||
284 | return DefaultConstrainedRounding; | |||
285 | } | |||
286 | ||||
287 | void setConstrainedFPFunctionAttr() { | |||
288 | assert(BB && "Must have a basic block to set any function attributes!")((BB && "Must have a basic block to set any function attributes!" ) ? static_cast<void> (0) : __assert_fail ("BB && \"Must have a basic block to set any function attributes!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 288, __PRETTY_FUNCTION__)); | |||
289 | ||||
290 | Function *F = BB->getParent(); | |||
291 | if (!F->hasFnAttribute(Attribute::StrictFP)) { | |||
292 | F->addFnAttr(Attribute::StrictFP); | |||
293 | } | |||
294 | } | |||
295 | ||||
296 | void setConstrainedFPCallAttr(CallInst *I) { | |||
297 | I->addAttribute(AttributeList::FunctionIndex, Attribute::StrictFP); | |||
298 | } | |||
299 | ||||
300 | void setDefaultOperandBundles(ArrayRef<OperandBundleDef> OpBundles) { | |||
301 | DefaultOperandBundles = OpBundles; | |||
302 | } | |||
303 | ||||
304 | //===--------------------------------------------------------------------===// | |||
305 | // RAII helpers. | |||
306 | //===--------------------------------------------------------------------===// | |||
307 | ||||
308 | // RAII object that stores the current insertion point and restores it | |||
309 | // when the object is destroyed. This includes the debug location. | |||
310 | class InsertPointGuard { | |||
311 | IRBuilderBase &Builder; | |||
312 | AssertingVH<BasicBlock> Block; | |||
313 | BasicBlock::iterator Point; | |||
314 | DebugLoc DbgLoc; | |||
315 | ||||
316 | public: | |||
317 | InsertPointGuard(IRBuilderBase &B) | |||
318 | : Builder(B), Block(B.GetInsertBlock()), Point(B.GetInsertPoint()), | |||
319 | DbgLoc(B.getCurrentDebugLocation()) {} | |||
320 | ||||
321 | InsertPointGuard(const InsertPointGuard &) = delete; | |||
322 | InsertPointGuard &operator=(const InsertPointGuard &) = delete; | |||
323 | ||||
324 | ~InsertPointGuard() { | |||
325 | Builder.restoreIP(InsertPoint(Block, Point)); | |||
326 | Builder.SetCurrentDebugLocation(DbgLoc); | |||
327 | } | |||
328 | }; | |||
329 | ||||
330 | // RAII object that stores the current fast math settings and restores | |||
331 | // them when the object is destroyed. | |||
332 | class FastMathFlagGuard { | |||
333 | IRBuilderBase &Builder; | |||
334 | FastMathFlags FMF; | |||
335 | MDNode *FPMathTag; | |||
336 | bool IsFPConstrained; | |||
337 | fp::ExceptionBehavior DefaultConstrainedExcept; | |||
338 | RoundingMode DefaultConstrainedRounding; | |||
339 | ||||
340 | public: | |||
341 | FastMathFlagGuard(IRBuilderBase &B) | |||
342 | : Builder(B), FMF(B.FMF), FPMathTag(B.DefaultFPMathTag), | |||
343 | IsFPConstrained(B.IsFPConstrained), | |||
344 | DefaultConstrainedExcept(B.DefaultConstrainedExcept), | |||
345 | DefaultConstrainedRounding(B.DefaultConstrainedRounding) {} | |||
346 | ||||
347 | FastMathFlagGuard(const FastMathFlagGuard &) = delete; | |||
348 | FastMathFlagGuard &operator=(const FastMathFlagGuard &) = delete; | |||
349 | ||||
350 | ~FastMathFlagGuard() { | |||
351 | Builder.FMF = FMF; | |||
352 | Builder.DefaultFPMathTag = FPMathTag; | |||
353 | Builder.IsFPConstrained = IsFPConstrained; | |||
354 | Builder.DefaultConstrainedExcept = DefaultConstrainedExcept; | |||
355 | Builder.DefaultConstrainedRounding = DefaultConstrainedRounding; | |||
356 | } | |||
357 | }; | |||
358 | ||||
359 | // RAII object that stores the current default operand bundles and restores | |||
360 | // them when the object is destroyed. | |||
361 | class OperandBundlesGuard { | |||
362 | IRBuilderBase &Builder; | |||
363 | ArrayRef<OperandBundleDef> DefaultOperandBundles; | |||
364 | ||||
365 | public: | |||
366 | OperandBundlesGuard(IRBuilderBase &B) | |||
367 | : Builder(B), DefaultOperandBundles(B.DefaultOperandBundles) {} | |||
368 | ||||
369 | OperandBundlesGuard(const OperandBundlesGuard &) = delete; | |||
370 | OperandBundlesGuard &operator=(const OperandBundlesGuard &) = delete; | |||
371 | ||||
372 | ~OperandBundlesGuard() { | |||
373 | Builder.DefaultOperandBundles = DefaultOperandBundles; | |||
374 | } | |||
375 | }; | |||
376 | ||||
377 | ||||
378 | //===--------------------------------------------------------------------===// | |||
379 | // Miscellaneous creation methods. | |||
380 | //===--------------------------------------------------------------------===// | |||
381 | ||||
382 | /// Make a new global variable with initializer type i8* | |||
383 | /// | |||
384 | /// Make a new global variable with an initializer that has array of i8 type | |||
385 | /// filled in with the null terminated string value specified. The new global | |||
386 | /// variable will be marked mergable with any others of the same contents. If | |||
387 | /// Name is specified, it is the name of the global variable created. | |||
388 | /// | |||
389 | /// If no module is given via \p M, it is take from the insertion point basic | |||
390 | /// block. | |||
391 | GlobalVariable *CreateGlobalString(StringRef Str, const Twine &Name = "", | |||
392 | unsigned AddressSpace = 0, | |||
393 | Module *M = nullptr); | |||
394 | ||||
395 | /// Get a constant value representing either true or false. | |||
396 | ConstantInt *getInt1(bool V) { | |||
397 | return ConstantInt::get(getInt1Ty(), V); | |||
398 | } | |||
399 | ||||
400 | /// Get the constant value for i1 true. | |||
401 | ConstantInt *getTrue() { | |||
402 | return ConstantInt::getTrue(Context); | |||
403 | } | |||
404 | ||||
405 | /// Get the constant value for i1 false. | |||
406 | ConstantInt *getFalse() { | |||
407 | return ConstantInt::getFalse(Context); | |||
408 | } | |||
409 | ||||
410 | /// Get a constant 8-bit value. | |||
411 | ConstantInt *getInt8(uint8_t C) { | |||
412 | return ConstantInt::get(getInt8Ty(), C); | |||
413 | } | |||
414 | ||||
415 | /// Get a constant 16-bit value. | |||
416 | ConstantInt *getInt16(uint16_t C) { | |||
417 | return ConstantInt::get(getInt16Ty(), C); | |||
418 | } | |||
419 | ||||
420 | /// Get a constant 32-bit value. | |||
421 | ConstantInt *getInt32(uint32_t C) { | |||
422 | return ConstantInt::get(getInt32Ty(), C); | |||
423 | } | |||
424 | ||||
425 | /// Get a constant 64-bit value. | |||
426 | ConstantInt *getInt64(uint64_t C) { | |||
427 | return ConstantInt::get(getInt64Ty(), C); | |||
428 | } | |||
429 | ||||
430 | /// Get a constant N-bit value, zero extended or truncated from | |||
431 | /// a 64-bit value. | |||
432 | ConstantInt *getIntN(unsigned N, uint64_t C) { | |||
433 | return ConstantInt::get(getIntNTy(N), C); | |||
434 | } | |||
435 | ||||
436 | /// Get a constant integer value. | |||
437 | ConstantInt *getInt(const APInt &AI) { | |||
438 | return ConstantInt::get(Context, AI); | |||
439 | } | |||
440 | ||||
441 | //===--------------------------------------------------------------------===// | |||
442 | // Type creation methods | |||
443 | //===--------------------------------------------------------------------===// | |||
444 | ||||
445 | /// Fetch the type representing a single bit | |||
446 | IntegerType *getInt1Ty() { | |||
447 | return Type::getInt1Ty(Context); | |||
448 | } | |||
449 | ||||
450 | /// Fetch the type representing an 8-bit integer. | |||
451 | IntegerType *getInt8Ty() { | |||
452 | return Type::getInt8Ty(Context); | |||
453 | } | |||
454 | ||||
455 | /// Fetch the type representing a 16-bit integer. | |||
456 | IntegerType *getInt16Ty() { | |||
457 | return Type::getInt16Ty(Context); | |||
458 | } | |||
459 | ||||
460 | /// Fetch the type representing a 32-bit integer. | |||
461 | IntegerType *getInt32Ty() { | |||
462 | return Type::getInt32Ty(Context); | |||
463 | } | |||
464 | ||||
465 | /// Fetch the type representing a 64-bit integer. | |||
466 | IntegerType *getInt64Ty() { | |||
467 | return Type::getInt64Ty(Context); | |||
468 | } | |||
469 | ||||
470 | /// Fetch the type representing a 128-bit integer. | |||
471 | IntegerType *getInt128Ty() { return Type::getInt128Ty(Context); } | |||
472 | ||||
473 | /// Fetch the type representing an N-bit integer. | |||
474 | IntegerType *getIntNTy(unsigned N) { | |||
475 | return Type::getIntNTy(Context, N); | |||
476 | } | |||
477 | ||||
478 | /// Fetch the type representing a 16-bit floating point value. | |||
479 | Type *getHalfTy() { | |||
480 | return Type::getHalfTy(Context); | |||
481 | } | |||
482 | ||||
483 | /// Fetch the type representing a 16-bit brain floating point value. | |||
484 | Type *getBFloatTy() { | |||
485 | return Type::getBFloatTy(Context); | |||
486 | } | |||
487 | ||||
488 | /// Fetch the type representing a 32-bit floating point value. | |||
489 | Type *getFloatTy() { | |||
490 | return Type::getFloatTy(Context); | |||
491 | } | |||
492 | ||||
493 | /// Fetch the type representing a 64-bit floating point value. | |||
494 | Type *getDoubleTy() { | |||
495 | return Type::getDoubleTy(Context); | |||
496 | } | |||
497 | ||||
498 | /// Fetch the type representing void. | |||
499 | Type *getVoidTy() { | |||
500 | return Type::getVoidTy(Context); | |||
501 | } | |||
502 | ||||
503 | /// Fetch the type representing a pointer to an 8-bit integer value. | |||
504 | PointerType *getInt8PtrTy(unsigned AddrSpace = 0) { | |||
505 | return Type::getInt8PtrTy(Context, AddrSpace); | |||
506 | } | |||
507 | ||||
508 | /// Fetch the type representing a pointer to an integer value. | |||
509 | IntegerType *getIntPtrTy(const DataLayout &DL, unsigned AddrSpace = 0) { | |||
510 | return DL.getIntPtrType(Context, AddrSpace); | |||
511 | } | |||
512 | ||||
513 | //===--------------------------------------------------------------------===// | |||
514 | // Intrinsic creation methods | |||
515 | //===--------------------------------------------------------------------===// | |||
516 | ||||
517 | /// Create and insert a memset to the specified pointer and the | |||
518 | /// specified value. | |||
519 | /// | |||
520 | /// If the pointer isn't an i8*, it will be converted. If a TBAA tag is | |||
521 | /// specified, it will be added to the instruction. Likewise with alias.scope | |||
522 | /// and noalias tags. | |||
523 | CallInst *CreateMemSet(Value *Ptr, Value *Val, uint64_t Size, | |||
524 | MaybeAlign Align, bool isVolatile = false, | |||
525 | MDNode *TBAATag = nullptr, MDNode *ScopeTag = nullptr, | |||
526 | MDNode *NoAliasTag = nullptr) { | |||
527 | return CreateMemSet(Ptr, Val, getInt64(Size), Align, isVolatile, | |||
528 | TBAATag, ScopeTag, NoAliasTag); | |||
529 | } | |||
530 | ||||
531 | CallInst *CreateMemSet(Value *Ptr, Value *Val, Value *Size, MaybeAlign Align, | |||
532 | bool isVolatile = false, MDNode *TBAATag = nullptr, | |||
533 | MDNode *ScopeTag = nullptr, | |||
534 | MDNode *NoAliasTag = nullptr); | |||
535 | ||||
536 | /// Create and insert an element unordered-atomic memset of the region of | |||
537 | /// memory starting at the given pointer to the given value. | |||
538 | /// | |||
539 | /// If the pointer isn't an i8*, it will be converted. If a TBAA tag is | |||
540 | /// specified, it will be added to the instruction. Likewise with alias.scope | |||
541 | /// and noalias tags. | |||
542 | CallInst *CreateElementUnorderedAtomicMemSet(Value *Ptr, Value *Val, | |||
543 | uint64_t Size, Align Alignment, | |||
544 | uint32_t ElementSize, | |||
545 | MDNode *TBAATag = nullptr, | |||
546 | MDNode *ScopeTag = nullptr, | |||
547 | MDNode *NoAliasTag = nullptr) { | |||
548 | return CreateElementUnorderedAtomicMemSet(Ptr, Val, getInt64(Size), | |||
549 | Align(Alignment), ElementSize, | |||
550 | TBAATag, ScopeTag, NoAliasTag); | |||
551 | } | |||
552 | ||||
553 | CallInst *CreateElementUnorderedAtomicMemSet(Value *Ptr, Value *Val, | |||
554 | Value *Size, Align Alignment, | |||
555 | uint32_t ElementSize, | |||
556 | MDNode *TBAATag = nullptr, | |||
557 | MDNode *ScopeTag = nullptr, | |||
558 | MDNode *NoAliasTag = nullptr); | |||
559 | ||||
560 | /// Create and insert a memcpy between the specified pointers. | |||
561 | /// | |||
562 | /// If the pointers aren't i8*, they will be converted. If a TBAA tag is | |||
563 | /// specified, it will be added to the instruction. Likewise with alias.scope | |||
564 | /// and noalias tags. | |||
565 | CallInst *CreateMemCpy(Value *Dst, MaybeAlign DstAlign, Value *Src, | |||
566 | MaybeAlign SrcAlign, uint64_t Size, | |||
567 | bool isVolatile = false, MDNode *TBAATag = nullptr, | |||
568 | MDNode *TBAAStructTag = nullptr, | |||
569 | MDNode *ScopeTag = nullptr, | |||
570 | MDNode *NoAliasTag = nullptr) { | |||
571 | return CreateMemCpy(Dst, DstAlign, Src, SrcAlign, getInt64(Size), | |||
572 | isVolatile, TBAATag, TBAAStructTag, ScopeTag, | |||
573 | NoAliasTag); | |||
574 | } | |||
575 | ||||
576 | CallInst *CreateMemCpy(Value *Dst, MaybeAlign DstAlign, Value *Src, | |||
577 | MaybeAlign SrcAlign, Value *Size, | |||
578 | bool isVolatile = false, MDNode *TBAATag = nullptr, | |||
579 | MDNode *TBAAStructTag = nullptr, | |||
580 | MDNode *ScopeTag = nullptr, | |||
581 | MDNode *NoAliasTag = nullptr); | |||
582 | ||||
583 | CallInst *CreateMemCpyInline(Value *Dst, MaybeAlign DstAlign, Value *Src, | |||
584 | MaybeAlign SrcAlign, Value *Size); | |||
585 | ||||
586 | /// Create and insert an element unordered-atomic memcpy between the | |||
587 | /// specified pointers. | |||
588 | /// | |||
589 | /// DstAlign/SrcAlign are the alignments of the Dst/Src pointers, respectively. | |||
590 | /// | |||
591 | /// If the pointers aren't i8*, they will be converted. If a TBAA tag is | |||
592 | /// specified, it will be added to the instruction. Likewise with alias.scope | |||
593 | /// and noalias tags. | |||
594 | CallInst *CreateElementUnorderedAtomicMemCpy( | |||
595 | Value *Dst, Align DstAlign, Value *Src, Align SrcAlign, Value *Size, | |||
596 | uint32_t ElementSize, MDNode *TBAATag = nullptr, | |||
597 | MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, | |||
598 | MDNode *NoAliasTag = nullptr); | |||
599 | ||||
600 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateElementUnorderedAtomicMemCpy(CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
601 | Value *Dst, unsigned DstAlign, Value *Src,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
602 | unsigned SrcAlign, uint64_t Size,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
603 | uint32_t ElementSize, MDNode *TBAATag = nullptr,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
604 | MDNode *TBAAStructTag = nullptr,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
605 | MDNode *ScopeTag = nullptr,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
606 | MDNode *NoAliasTag = nullptr),CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
607 | "Use the version that takes Align instead")CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
608 | return CreateElementUnorderedAtomicMemCpy( | |||
609 | Dst, Align(DstAlign), Src, Align(SrcAlign), getInt64(Size), ElementSize, | |||
610 | TBAATag, TBAAStructTag, ScopeTag, NoAliasTag); | |||
611 | } | |||
612 | ||||
613 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateElementUnorderedAtomicMemCpy(CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
614 | Value *Dst, unsigned DstAlign, Value *Src,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
615 | unsigned SrcAlign, Value *Size,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
616 | uint32_t ElementSize, MDNode *TBAATag = nullptr,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
617 | MDNode *TBAAStructTag = nullptr,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
618 | MDNode *ScopeTag = nullptr,CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
619 | MDNode *NoAliasTag = nullptr),CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
620 | "Use the version that takes Align instead")CallInst *CreateElementUnorderedAtomicMemCpy( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
621 | return CreateElementUnorderedAtomicMemCpy( | |||
622 | Dst, Align(DstAlign), Src, Align(SrcAlign), Size, ElementSize, TBAATag, | |||
623 | TBAAStructTag, ScopeTag, NoAliasTag); | |||
624 | } | |||
625 | ||||
626 | CallInst *CreateMemMove(Value *Dst, MaybeAlign DstAlign, Value *Src, | |||
627 | MaybeAlign SrcAlign, uint64_t Size, | |||
628 | bool isVolatile = false, MDNode *TBAATag = nullptr, | |||
629 | MDNode *ScopeTag = nullptr, | |||
630 | MDNode *NoAliasTag = nullptr) { | |||
631 | return CreateMemMove(Dst, DstAlign, Src, SrcAlign, getInt64(Size), | |||
632 | isVolatile, TBAATag, ScopeTag, NoAliasTag); | |||
633 | } | |||
634 | ||||
635 | CallInst *CreateMemMove(Value *Dst, MaybeAlign DstAlign, Value *Src, | |||
636 | MaybeAlign SrcAlign, Value *Size, | |||
637 | bool isVolatile = false, MDNode *TBAATag = nullptr, | |||
638 | MDNode *ScopeTag = nullptr, | |||
639 | MDNode *NoAliasTag = nullptr); | |||
640 | ||||
641 | /// \brief Create and insert an element unordered-atomic memmove between the | |||
642 | /// specified pointers. | |||
643 | /// | |||
644 | /// DstAlign/SrcAlign are the alignments of the Dst/Src pointers, | |||
645 | /// respectively. | |||
646 | /// | |||
647 | /// If the pointers aren't i8*, they will be converted. If a TBAA tag is | |||
648 | /// specified, it will be added to the instruction. Likewise with alias.scope | |||
649 | /// and noalias tags. | |||
650 | CallInst *CreateElementUnorderedAtomicMemMove( | |||
651 | Value *Dst, Align DstAlign, Value *Src, Align SrcAlign, Value *Size, | |||
652 | uint32_t ElementSize, MDNode *TBAATag = nullptr, | |||
653 | MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, | |||
654 | MDNode *NoAliasTag = nullptr); | |||
655 | ||||
656 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateElementUnorderedAtomicMemMove(CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
657 | Value *Dst, unsigned DstAlign, Value *Src,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
658 | unsigned SrcAlign, uint64_t Size,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
659 | uint32_t ElementSize, MDNode *TBAATag = nullptr,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
660 | MDNode *TBAAStructTag = nullptr,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
661 | MDNode *ScopeTag = nullptr,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
662 | MDNode *NoAliasTag = nullptr),CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
663 | "Use the version that takes Align instead")CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
664 | return CreateElementUnorderedAtomicMemMove( | |||
665 | Dst, Align(DstAlign), Src, Align(SrcAlign), getInt64(Size), ElementSize, | |||
666 | TBAATag, TBAAStructTag, ScopeTag, NoAliasTag); | |||
667 | } | |||
668 | ||||
669 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateElementUnorderedAtomicMemMove(CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
670 | Value *Dst, unsigned DstAlign, Value *Src,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
671 | unsigned SrcAlign, Value *Size,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
672 | uint32_t ElementSize, MDNode *TBAATag = nullptr,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
673 | MDNode *TBAAStructTag = nullptr,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
674 | MDNode *ScopeTag = nullptr,CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
675 | MDNode *NoAliasTag = nullptr),CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
676 | "Use the version that takes Align instead")CallInst *CreateElementUnorderedAtomicMemMove( Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, Value *Size, uint32_t ElementSize, MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr ) __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
677 | return CreateElementUnorderedAtomicMemMove( | |||
678 | Dst, Align(DstAlign), Src, Align(SrcAlign), Size, ElementSize, TBAATag, | |||
679 | TBAAStructTag, ScopeTag, NoAliasTag); | |||
680 | } | |||
681 | ||||
682 | /// Create a vector fadd reduction intrinsic of the source vector. | |||
683 | /// The first parameter is a scalar accumulator value for ordered reductions. | |||
684 | CallInst *CreateFAddReduce(Value *Acc, Value *Src); | |||
685 | ||||
686 | /// Create a vector fmul reduction intrinsic of the source vector. | |||
687 | /// The first parameter is a scalar accumulator value for ordered reductions. | |||
688 | CallInst *CreateFMulReduce(Value *Acc, Value *Src); | |||
689 | ||||
690 | /// Create a vector int add reduction intrinsic of the source vector. | |||
691 | CallInst *CreateAddReduce(Value *Src); | |||
692 | ||||
693 | /// Create a vector int mul reduction intrinsic of the source vector. | |||
694 | CallInst *CreateMulReduce(Value *Src); | |||
695 | ||||
696 | /// Create a vector int AND reduction intrinsic of the source vector. | |||
697 | CallInst *CreateAndReduce(Value *Src); | |||
698 | ||||
699 | /// Create a vector int OR reduction intrinsic of the source vector. | |||
700 | CallInst *CreateOrReduce(Value *Src); | |||
701 | ||||
702 | /// Create a vector int XOR reduction intrinsic of the source vector. | |||
703 | CallInst *CreateXorReduce(Value *Src); | |||
704 | ||||
705 | /// Create a vector integer max reduction intrinsic of the source | |||
706 | /// vector. | |||
707 | CallInst *CreateIntMaxReduce(Value *Src, bool IsSigned = false); | |||
708 | ||||
709 | /// Create a vector integer min reduction intrinsic of the source | |||
710 | /// vector. | |||
711 | CallInst *CreateIntMinReduce(Value *Src, bool IsSigned = false); | |||
712 | ||||
713 | /// Create a vector float max reduction intrinsic of the source | |||
714 | /// vector. | |||
715 | CallInst *CreateFPMaxReduce(Value *Src, bool NoNaN = false); | |||
716 | ||||
717 | /// Create a vector float min reduction intrinsic of the source | |||
718 | /// vector. | |||
719 | CallInst *CreateFPMinReduce(Value *Src, bool NoNaN = false); | |||
720 | ||||
721 | /// Create a lifetime.start intrinsic. | |||
722 | /// | |||
723 | /// If the pointer isn't i8* it will be converted. | |||
724 | CallInst *CreateLifetimeStart(Value *Ptr, ConstantInt *Size = nullptr); | |||
725 | ||||
726 | /// Create a lifetime.end intrinsic. | |||
727 | /// | |||
728 | /// If the pointer isn't i8* it will be converted. | |||
729 | CallInst *CreateLifetimeEnd(Value *Ptr, ConstantInt *Size = nullptr); | |||
730 | ||||
731 | /// Create a call to invariant.start intrinsic. | |||
732 | /// | |||
733 | /// If the pointer isn't i8* it will be converted. | |||
734 | CallInst *CreateInvariantStart(Value *Ptr, ConstantInt *Size = nullptr); | |||
735 | ||||
736 | /// Create a call to Masked Load intrinsic | |||
737 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateMaskedLoad(Value *Ptr, unsigned Alignment, Value *Mask, Value *PassThru = nullptr, const Twine &Name = "" ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
738 | CallInst *CreateMaskedLoad(Value *Ptr, unsigned Alignment, Value *Mask,CallInst *CreateMaskedLoad(Value *Ptr, unsigned Alignment, Value *Mask, Value *PassThru = nullptr, const Twine &Name = "" ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
739 | Value *PassThru = nullptr,CallInst *CreateMaskedLoad(Value *Ptr, unsigned Alignment, Value *Mask, Value *PassThru = nullptr, const Twine &Name = "" ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
740 | const Twine &Name = ""),CallInst *CreateMaskedLoad(Value *Ptr, unsigned Alignment, Value *Mask, Value *PassThru = nullptr, const Twine &Name = "" ) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
741 | "Use the version that takes Align instead")CallInst *CreateMaskedLoad(Value *Ptr, unsigned Alignment, Value *Mask, Value *PassThru = nullptr, const Twine &Name = "" ) __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
742 | return CreateMaskedLoad(Ptr, assumeAligned(Alignment), Mask, PassThru, | |||
743 | Name); | |||
744 | } | |||
745 | CallInst *CreateMaskedLoad(Value *Ptr, Align Alignment, Value *Mask, | |||
746 | Value *PassThru = nullptr, const Twine &Name = ""); | |||
747 | ||||
748 | /// Create a call to Masked Store intrinsic | |||
749 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateMaskedStore(Value *Val, Value *Ptr,CallInst *CreateMaskedStore(Value *Val, Value *Ptr, unsigned Alignment , Value *Mask) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
750 | unsigned Alignment,CallInst *CreateMaskedStore(Value *Val, Value *Ptr, unsigned Alignment , Value *Mask) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
751 | Value *Mask),CallInst *CreateMaskedStore(Value *Val, Value *Ptr, unsigned Alignment , Value *Mask) __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
752 | "Use the version that takes Align instead")CallInst *CreateMaskedStore(Value *Val, Value *Ptr, unsigned Alignment , Value *Mask) __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
753 | return CreateMaskedStore(Val, Ptr, assumeAligned(Alignment), Mask); | |||
754 | } | |||
755 | ||||
756 | CallInst *CreateMaskedStore(Value *Val, Value *Ptr, Align Alignment, | |||
757 | Value *Mask); | |||
758 | ||||
759 | /// Create a call to Masked Gather intrinsic | |||
760 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment, Value *Mask = nullptr, Value *PassThru = nullptr, const Twine &Name = "") __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
761 | CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment,CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment, Value *Mask = nullptr, Value *PassThru = nullptr, const Twine &Name = "") __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
762 | Value *Mask = nullptr,CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment, Value *Mask = nullptr, Value *PassThru = nullptr, const Twine &Name = "") __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
763 | Value *PassThru = nullptr,CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment, Value *Mask = nullptr, Value *PassThru = nullptr, const Twine &Name = "") __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
764 | const Twine &Name = ""),CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment, Value *Mask = nullptr, Value *PassThru = nullptr, const Twine &Name = "") __attribute__((deprecated("Use the version that takes Align instead" ))) | |||
765 | "Use the version that takes Align instead")CallInst *CreateMaskedGather(Value *Ptrs, unsigned Alignment, Value *Mask = nullptr, Value *PassThru = nullptr, const Twine &Name = "") __attribute__((deprecated("Use the version that takes Align instead" ))) { | |||
766 | return CreateMaskedGather(Ptrs, Align(Alignment), Mask, PassThru, Name); | |||
767 | } | |||
768 | ||||
769 | /// Create a call to Masked Gather intrinsic | |||
770 | CallInst *CreateMaskedGather(Value *Ptrs, Align Alignment, | |||
771 | Value *Mask = nullptr, Value *PassThru = nullptr, | |||
772 | const Twine &Name = ""); | |||
773 | ||||
774 | /// Create a call to Masked Scatter intrinsic | |||
775 | LLVM_ATTRIBUTE_DEPRECATED(CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, unsigned Alignment, Value *Mask = nullptr) __attribute__((deprecated( "Use the version that takes Align instead"))) | |||
776 | CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, unsigned Alignment,CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, unsigned Alignment, Value *Mask = nullptr) __attribute__((deprecated( "Use the version that takes Align instead"))) | |||
777 | Value *Mask = nullptr),CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, unsigned Alignment, Value *Mask = nullptr) __attribute__((deprecated( "Use the version that takes Align instead"))) | |||
778 | "Use the version that takes Align instead")CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, unsigned Alignment, Value *Mask = nullptr) __attribute__((deprecated( "Use the version that takes Align instead"))) { | |||
779 | return CreateMaskedScatter(Val, Ptrs, Align(Alignment), Mask); | |||
780 | } | |||
781 | ||||
782 | /// Create a call to Masked Scatter intrinsic | |||
783 | CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, Align Alignment, | |||
784 | Value *Mask = nullptr); | |||
785 | ||||
786 | /// Create an assume intrinsic call that allows the optimizer to | |||
787 | /// assume that the provided condition will be true. | |||
788 | CallInst *CreateAssumption(Value *Cond); | |||
789 | ||||
790 | /// Create a call to the experimental.gc.statepoint intrinsic to | |||
791 | /// start a new statepoint sequence. | |||
792 | CallInst *CreateGCStatepointCall(uint64_t ID, uint32_t NumPatchBytes, | |||
793 | Value *ActualCallee, | |||
794 | ArrayRef<Value *> CallArgs, | |||
795 | Optional<ArrayRef<Value *>> DeoptArgs, | |||
796 | ArrayRef<Value *> GCArgs, | |||
797 | const Twine &Name = ""); | |||
798 | ||||
799 | /// Create a call to the experimental.gc.statepoint intrinsic to | |||
800 | /// start a new statepoint sequence. | |||
801 | CallInst *CreateGCStatepointCall(uint64_t ID, uint32_t NumPatchBytes, | |||
802 | Value *ActualCallee, uint32_t Flags, | |||
803 | ArrayRef<Use> CallArgs, | |||
804 | Optional<ArrayRef<Use>> TransitionArgs, | |||
805 | Optional<ArrayRef<Use>> DeoptArgs, | |||
806 | ArrayRef<Value *> GCArgs, | |||
807 | const Twine &Name = ""); | |||
808 | ||||
809 | /// Conveninence function for the common case when CallArgs are filled | |||
810 | /// in using makeArrayRef(CS.arg_begin(), CS.arg_end()); Use needs to be | |||
811 | /// .get()'ed to get the Value pointer. | |||
812 | CallInst *CreateGCStatepointCall(uint64_t ID, uint32_t NumPatchBytes, | |||
813 | Value *ActualCallee, ArrayRef<Use> CallArgs, | |||
814 | Optional<ArrayRef<Value *>> DeoptArgs, | |||
815 | ArrayRef<Value *> GCArgs, | |||
816 | const Twine &Name = ""); | |||
817 | ||||
818 | /// Create an invoke to the experimental.gc.statepoint intrinsic to | |||
819 | /// start a new statepoint sequence. | |||
820 | InvokeInst * | |||
821 | CreateGCStatepointInvoke(uint64_t ID, uint32_t NumPatchBytes, | |||
822 | Value *ActualInvokee, BasicBlock *NormalDest, | |||
823 | BasicBlock *UnwindDest, ArrayRef<Value *> InvokeArgs, | |||
824 | Optional<ArrayRef<Value *>> DeoptArgs, | |||
825 | ArrayRef<Value *> GCArgs, const Twine &Name = ""); | |||
826 | ||||
827 | /// Create an invoke to the experimental.gc.statepoint intrinsic to | |||
828 | /// start a new statepoint sequence. | |||
829 | InvokeInst *CreateGCStatepointInvoke( | |||
830 | uint64_t ID, uint32_t NumPatchBytes, Value *ActualInvokee, | |||
831 | BasicBlock *NormalDest, BasicBlock *UnwindDest, uint32_t Flags, | |||
832 | ArrayRef<Use> InvokeArgs, Optional<ArrayRef<Use>> TransitionArgs, | |||
833 | Optional<ArrayRef<Use>> DeoptArgs, ArrayRef<Value *> GCArgs, | |||
834 | const Twine &Name = ""); | |||
835 | ||||
836 | // Convenience function for the common case when CallArgs are filled in using | |||
837 | // makeArrayRef(CS.arg_begin(), CS.arg_end()); Use needs to be .get()'ed to | |||
838 | // get the Value *. | |||
839 | InvokeInst * | |||
840 | CreateGCStatepointInvoke(uint64_t ID, uint32_t NumPatchBytes, | |||
841 | Value *ActualInvokee, BasicBlock *NormalDest, | |||
842 | BasicBlock *UnwindDest, ArrayRef<Use> InvokeArgs, | |||
843 | Optional<ArrayRef<Value *>> DeoptArgs, | |||
844 | ArrayRef<Value *> GCArgs, const Twine &Name = ""); | |||
845 | ||||
846 | /// Create a call to the experimental.gc.result intrinsic to extract | |||
847 | /// the result from a call wrapped in a statepoint. | |||
848 | CallInst *CreateGCResult(Instruction *Statepoint, | |||
849 | Type *ResultType, | |||
850 | const Twine &Name = ""); | |||
851 | ||||
852 | /// Create a call to the experimental.gc.relocate intrinsics to | |||
853 | /// project the relocated value of one pointer from the statepoint. | |||
854 | CallInst *CreateGCRelocate(Instruction *Statepoint, | |||
855 | int BaseOffset, | |||
856 | int DerivedOffset, | |||
857 | Type *ResultType, | |||
858 | const Twine &Name = ""); | |||
859 | ||||
860 | /// Create a call to intrinsic \p ID with 1 operand which is mangled on its | |||
861 | /// type. | |||
862 | CallInst *CreateUnaryIntrinsic(Intrinsic::ID ID, Value *V, | |||
863 | Instruction *FMFSource = nullptr, | |||
864 | const Twine &Name = ""); | |||
865 | ||||
866 | /// Create a call to intrinsic \p ID with 2 operands which is mangled on the | |||
867 | /// first type. | |||
868 | CallInst *CreateBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, | |||
869 | Instruction *FMFSource = nullptr, | |||
870 | const Twine &Name = ""); | |||
871 | ||||
872 | /// Create a call to intrinsic \p ID with \p args, mangled using \p Types. If | |||
873 | /// \p FMFSource is provided, copy fast-math-flags from that instruction to | |||
874 | /// the intrinsic. | |||
875 | CallInst *CreateIntrinsic(Intrinsic::ID ID, ArrayRef<Type *> Types, | |||
876 | ArrayRef<Value *> Args, | |||
877 | Instruction *FMFSource = nullptr, | |||
878 | const Twine &Name = ""); | |||
879 | ||||
880 | /// Create call to the minnum intrinsic. | |||
881 | CallInst *CreateMinNum(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
882 | return CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS, nullptr, Name); | |||
883 | } | |||
884 | ||||
885 | /// Create call to the maxnum intrinsic. | |||
886 | CallInst *CreateMaxNum(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
887 | return CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS, nullptr, Name); | |||
888 | } | |||
889 | ||||
890 | /// Create call to the minimum intrinsic. | |||
891 | CallInst *CreateMinimum(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
892 | return CreateBinaryIntrinsic(Intrinsic::minimum, LHS, RHS, nullptr, Name); | |||
893 | } | |||
894 | ||||
895 | /// Create call to the maximum intrinsic. | |||
896 | CallInst *CreateMaximum(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
897 | return CreateBinaryIntrinsic(Intrinsic::maximum, LHS, RHS, nullptr, Name); | |||
898 | } | |||
899 | ||||
900 | private: | |||
901 | /// Create a call to a masked intrinsic with given Id. | |||
902 | CallInst *CreateMaskedIntrinsic(Intrinsic::ID Id, ArrayRef<Value *> Ops, | |||
903 | ArrayRef<Type *> OverloadedTypes, | |||
904 | const Twine &Name = ""); | |||
905 | ||||
906 | Value *getCastedInt8PtrValue(Value *Ptr); | |||
907 | ||||
908 | //===--------------------------------------------------------------------===// | |||
909 | // Instruction creation methods: Terminators | |||
910 | //===--------------------------------------------------------------------===// | |||
911 | ||||
912 | private: | |||
913 | /// Helper to add branch weight and unpredictable metadata onto an | |||
914 | /// instruction. | |||
915 | /// \returns The annotated instruction. | |||
916 | template <typename InstTy> | |||
917 | InstTy *addBranchMetadata(InstTy *I, MDNode *Weights, MDNode *Unpredictable) { | |||
918 | if (Weights) | |||
919 | I->setMetadata(LLVMContext::MD_prof, Weights); | |||
920 | if (Unpredictable) | |||
921 | I->setMetadata(LLVMContext::MD_unpredictable, Unpredictable); | |||
922 | return I; | |||
923 | } | |||
924 | ||||
925 | public: | |||
926 | /// Create a 'ret void' instruction. | |||
927 | ReturnInst *CreateRetVoid() { | |||
928 | return Insert(ReturnInst::Create(Context)); | |||
929 | } | |||
930 | ||||
931 | /// Create a 'ret <val>' instruction. | |||
932 | ReturnInst *CreateRet(Value *V) { | |||
933 | return Insert(ReturnInst::Create(Context, V)); | |||
934 | } | |||
935 | ||||
936 | /// Create a sequence of N insertvalue instructions, | |||
937 | /// with one Value from the retVals array each, that build a aggregate | |||
938 | /// return value one value at a time, and a ret instruction to return | |||
939 | /// the resulting aggregate value. | |||
940 | /// | |||
941 | /// This is a convenience function for code that uses aggregate return values | |||
942 | /// as a vehicle for having multiple return values. | |||
943 | ReturnInst *CreateAggregateRet(Value *const *retVals, unsigned N) { | |||
944 | Value *V = UndefValue::get(getCurrentFunctionReturnType()); | |||
945 | for (unsigned i = 0; i != N; ++i) | |||
946 | V = CreateInsertValue(V, retVals[i], i, "mrv"); | |||
947 | return Insert(ReturnInst::Create(Context, V)); | |||
948 | } | |||
949 | ||||
950 | /// Create an unconditional 'br label X' instruction. | |||
951 | BranchInst *CreateBr(BasicBlock *Dest) { | |||
952 | return Insert(BranchInst::Create(Dest)); | |||
953 | } | |||
954 | ||||
955 | /// Create a conditional 'br Cond, TrueDest, FalseDest' | |||
956 | /// instruction. | |||
957 | BranchInst *CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, | |||
958 | MDNode *BranchWeights = nullptr, | |||
959 | MDNode *Unpredictable = nullptr) { | |||
960 | return Insert(addBranchMetadata(BranchInst::Create(True, False, Cond), | |||
961 | BranchWeights, Unpredictable)); | |||
962 | } | |||
963 | ||||
964 | /// Create a conditional 'br Cond, TrueDest, FalseDest' | |||
965 | /// instruction. Copy branch meta data if available. | |||
966 | BranchInst *CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, | |||
967 | Instruction *MDSrc) { | |||
968 | BranchInst *Br = BranchInst::Create(True, False, Cond); | |||
969 | if (MDSrc) { | |||
970 | unsigned WL[4] = {LLVMContext::MD_prof, LLVMContext::MD_unpredictable, | |||
971 | LLVMContext::MD_make_implicit, LLVMContext::MD_dbg}; | |||
972 | Br->copyMetadata(*MDSrc, makeArrayRef(&WL[0], 4)); | |||
973 | } | |||
974 | return Insert(Br); | |||
975 | } | |||
976 | ||||
977 | /// Create a switch instruction with the specified value, default dest, | |||
978 | /// and with a hint for the number of cases that will be added (for efficient | |||
979 | /// allocation). | |||
980 | SwitchInst *CreateSwitch(Value *V, BasicBlock *Dest, unsigned NumCases = 10, | |||
981 | MDNode *BranchWeights = nullptr, | |||
982 | MDNode *Unpredictable = nullptr) { | |||
983 | return Insert(addBranchMetadata(SwitchInst::Create(V, Dest, NumCases), | |||
984 | BranchWeights, Unpredictable)); | |||
985 | } | |||
986 | ||||
987 | /// Create an indirect branch instruction with the specified address | |||
988 | /// operand, with an optional hint for the number of destinations that will be | |||
989 | /// added (for efficient allocation). | |||
990 | IndirectBrInst *CreateIndirectBr(Value *Addr, unsigned NumDests = 10) { | |||
991 | return Insert(IndirectBrInst::Create(Addr, NumDests)); | |||
992 | } | |||
993 | ||||
994 | /// Create an invoke instruction. | |||
995 | InvokeInst *CreateInvoke(FunctionType *Ty, Value *Callee, | |||
996 | BasicBlock *NormalDest, BasicBlock *UnwindDest, | |||
997 | ArrayRef<Value *> Args, | |||
998 | ArrayRef<OperandBundleDef> OpBundles, | |||
999 | const Twine &Name = "") { | |||
1000 | return Insert( | |||
1001 | InvokeInst::Create(Ty, Callee, NormalDest, UnwindDest, Args, OpBundles), | |||
1002 | Name); | |||
1003 | } | |||
1004 | InvokeInst *CreateInvoke(FunctionType *Ty, Value *Callee, | |||
1005 | BasicBlock *NormalDest, BasicBlock *UnwindDest, | |||
1006 | ArrayRef<Value *> Args = None, | |||
1007 | const Twine &Name = "") { | |||
1008 | return Insert(InvokeInst::Create(Ty, Callee, NormalDest, UnwindDest, Args), | |||
1009 | Name); | |||
1010 | } | |||
1011 | ||||
1012 | InvokeInst *CreateInvoke(FunctionCallee Callee, BasicBlock *NormalDest, | |||
1013 | BasicBlock *UnwindDest, ArrayRef<Value *> Args, | |||
1014 | ArrayRef<OperandBundleDef> OpBundles, | |||
1015 | const Twine &Name = "") { | |||
1016 | return CreateInvoke(Callee.getFunctionType(), Callee.getCallee(), | |||
1017 | NormalDest, UnwindDest, Args, OpBundles, Name); | |||
1018 | } | |||
1019 | ||||
1020 | InvokeInst *CreateInvoke(FunctionCallee Callee, BasicBlock *NormalDest, | |||
1021 | BasicBlock *UnwindDest, | |||
1022 | ArrayRef<Value *> Args = None, | |||
1023 | const Twine &Name = "") { | |||
1024 | return CreateInvoke(Callee.getFunctionType(), Callee.getCallee(), | |||
1025 | NormalDest, UnwindDest, Args, Name); | |||
1026 | } | |||
1027 | ||||
1028 | /// \brief Create a callbr instruction. | |||
1029 | CallBrInst *CreateCallBr(FunctionType *Ty, Value *Callee, | |||
1030 | BasicBlock *DefaultDest, | |||
1031 | ArrayRef<BasicBlock *> IndirectDests, | |||
1032 | ArrayRef<Value *> Args = None, | |||
1033 | const Twine &Name = "") { | |||
1034 | return Insert(CallBrInst::Create(Ty, Callee, DefaultDest, IndirectDests, | |||
1035 | Args), Name); | |||
1036 | } | |||
1037 | CallBrInst *CreateCallBr(FunctionType *Ty, Value *Callee, | |||
1038 | BasicBlock *DefaultDest, | |||
1039 | ArrayRef<BasicBlock *> IndirectDests, | |||
1040 | ArrayRef<Value *> Args, | |||
1041 | ArrayRef<OperandBundleDef> OpBundles, | |||
1042 | const Twine &Name = "") { | |||
1043 | return Insert( | |||
1044 | CallBrInst::Create(Ty, Callee, DefaultDest, IndirectDests, Args, | |||
1045 | OpBundles), Name); | |||
1046 | } | |||
1047 | ||||
1048 | CallBrInst *CreateCallBr(FunctionCallee Callee, BasicBlock *DefaultDest, | |||
1049 | ArrayRef<BasicBlock *> IndirectDests, | |||
1050 | ArrayRef<Value *> Args = None, | |||
1051 | const Twine &Name = "") { | |||
1052 | return CreateCallBr(Callee.getFunctionType(), Callee.getCallee(), | |||
1053 | DefaultDest, IndirectDests, Args, Name); | |||
1054 | } | |||
1055 | CallBrInst *CreateCallBr(FunctionCallee Callee, BasicBlock *DefaultDest, | |||
1056 | ArrayRef<BasicBlock *> IndirectDests, | |||
1057 | ArrayRef<Value *> Args, | |||
1058 | ArrayRef<OperandBundleDef> OpBundles, | |||
1059 | const Twine &Name = "") { | |||
1060 | return CreateCallBr(Callee.getFunctionType(), Callee.getCallee(), | |||
1061 | DefaultDest, IndirectDests, Args, Name); | |||
1062 | } | |||
1063 | ||||
1064 | ResumeInst *CreateResume(Value *Exn) { | |||
1065 | return Insert(ResumeInst::Create(Exn)); | |||
1066 | } | |||
1067 | ||||
1068 | CleanupReturnInst *CreateCleanupRet(CleanupPadInst *CleanupPad, | |||
1069 | BasicBlock *UnwindBB = nullptr) { | |||
1070 | return Insert(CleanupReturnInst::Create(CleanupPad, UnwindBB)); | |||
1071 | } | |||
1072 | ||||
1073 | CatchSwitchInst *CreateCatchSwitch(Value *ParentPad, BasicBlock *UnwindBB, | |||
1074 | unsigned NumHandlers, | |||
1075 | const Twine &Name = "") { | |||
1076 | return Insert(CatchSwitchInst::Create(ParentPad, UnwindBB, NumHandlers), | |||
1077 | Name); | |||
1078 | } | |||
1079 | ||||
1080 | CatchPadInst *CreateCatchPad(Value *ParentPad, ArrayRef<Value *> Args, | |||
1081 | const Twine &Name = "") { | |||
1082 | return Insert(CatchPadInst::Create(ParentPad, Args), Name); | |||
1083 | } | |||
1084 | ||||
1085 | CleanupPadInst *CreateCleanupPad(Value *ParentPad, | |||
1086 | ArrayRef<Value *> Args = None, | |||
1087 | const Twine &Name = "") { | |||
1088 | return Insert(CleanupPadInst::Create(ParentPad, Args), Name); | |||
1089 | } | |||
1090 | ||||
1091 | CatchReturnInst *CreateCatchRet(CatchPadInst *CatchPad, BasicBlock *BB) { | |||
1092 | return Insert(CatchReturnInst::Create(CatchPad, BB)); | |||
1093 | } | |||
1094 | ||||
1095 | UnreachableInst *CreateUnreachable() { | |||
1096 | return Insert(new UnreachableInst(Context)); | |||
1097 | } | |||
1098 | ||||
1099 | //===--------------------------------------------------------------------===// | |||
1100 | // Instruction creation methods: Binary Operators | |||
1101 | //===--------------------------------------------------------------------===// | |||
1102 | private: | |||
1103 | BinaryOperator *CreateInsertNUWNSWBinOp(BinaryOperator::BinaryOps Opc, | |||
1104 | Value *LHS, Value *RHS, | |||
1105 | const Twine &Name, | |||
1106 | bool HasNUW, bool HasNSW) { | |||
1107 | BinaryOperator *BO = Insert(BinaryOperator::Create(Opc, LHS, RHS), Name); | |||
1108 | if (HasNUW) BO->setHasNoUnsignedWrap(); | |||
1109 | if (HasNSW) BO->setHasNoSignedWrap(); | |||
1110 | return BO; | |||
1111 | } | |||
1112 | ||||
1113 | Instruction *setFPAttrs(Instruction *I, MDNode *FPMD, | |||
1114 | FastMathFlags FMF) const { | |||
1115 | if (!FPMD) | |||
1116 | FPMD = DefaultFPMathTag; | |||
1117 | if (FPMD) | |||
1118 | I->setMetadata(LLVMContext::MD_fpmath, FPMD); | |||
1119 | I->setFastMathFlags(FMF); | |||
1120 | return I; | |||
1121 | } | |||
1122 | ||||
1123 | Value *foldConstant(Instruction::BinaryOps Opc, Value *L, | |||
1124 | Value *R, const Twine &Name) const { | |||
1125 | auto *LC = dyn_cast<Constant>(L); | |||
1126 | auto *RC = dyn_cast<Constant>(R); | |||
1127 | return (LC && RC) ? Insert(Folder.CreateBinOp(Opc, LC, RC), Name) : nullptr; | |||
1128 | } | |||
1129 | ||||
1130 | Value *getConstrainedFPRounding(Optional<RoundingMode> Rounding) { | |||
1131 | RoundingMode UseRounding = DefaultConstrainedRounding; | |||
1132 | ||||
1133 | if (Rounding.hasValue()) | |||
1134 | UseRounding = Rounding.getValue(); | |||
1135 | ||||
1136 | Optional<StringRef> RoundingStr = RoundingModeToStr(UseRounding); | |||
1137 | assert(RoundingStr.hasValue() && "Garbage strict rounding mode!")((RoundingStr.hasValue() && "Garbage strict rounding mode!" ) ? static_cast<void> (0) : __assert_fail ("RoundingStr.hasValue() && \"Garbage strict rounding mode!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1137, __PRETTY_FUNCTION__)); | |||
1138 | auto *RoundingMDS = MDString::get(Context, RoundingStr.getValue()); | |||
1139 | ||||
1140 | return MetadataAsValue::get(Context, RoundingMDS); | |||
1141 | } | |||
1142 | ||||
1143 | Value *getConstrainedFPExcept(Optional<fp::ExceptionBehavior> Except) { | |||
1144 | fp::ExceptionBehavior UseExcept = DefaultConstrainedExcept; | |||
1145 | ||||
1146 | if (Except.hasValue()) | |||
1147 | UseExcept = Except.getValue(); | |||
1148 | ||||
1149 | Optional<StringRef> ExceptStr = ExceptionBehaviorToStr(UseExcept); | |||
1150 | assert(ExceptStr.hasValue() && "Garbage strict exception behavior!")((ExceptStr.hasValue() && "Garbage strict exception behavior!" ) ? static_cast<void> (0) : __assert_fail ("ExceptStr.hasValue() && \"Garbage strict exception behavior!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1150, __PRETTY_FUNCTION__)); | |||
1151 | auto *ExceptMDS = MDString::get(Context, ExceptStr.getValue()); | |||
1152 | ||||
1153 | return MetadataAsValue::get(Context, ExceptMDS); | |||
1154 | } | |||
1155 | ||||
1156 | Value *getConstrainedFPPredicate(CmpInst::Predicate Predicate) { | |||
1157 | assert(CmpInst::isFPPredicate(Predicate) &&((CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst ::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && "Invalid constrained FP comparison predicate!") ? static_cast <void> (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1160, __PRETTY_FUNCTION__)) | |||
1158 | Predicate != CmpInst::FCMP_FALSE &&((CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst ::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && "Invalid constrained FP comparison predicate!") ? static_cast <void> (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1160, __PRETTY_FUNCTION__)) | |||
1159 | Predicate != CmpInst::FCMP_TRUE &&((CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst ::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && "Invalid constrained FP comparison predicate!") ? static_cast <void> (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1160, __PRETTY_FUNCTION__)) | |||
1160 | "Invalid constrained FP comparison predicate!")((CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst ::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && "Invalid constrained FP comparison predicate!") ? static_cast <void> (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1160, __PRETTY_FUNCTION__)); | |||
1161 | ||||
1162 | StringRef PredicateStr = CmpInst::getPredicateName(Predicate); | |||
1163 | auto *PredicateMDS = MDString::get(Context, PredicateStr); | |||
1164 | ||||
1165 | return MetadataAsValue::get(Context, PredicateMDS); | |||
1166 | } | |||
1167 | ||||
1168 | public: | |||
1169 | Value *CreateAdd(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1170 | bool HasNUW = false, bool HasNSW = false) { | |||
1171 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1172 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1173 | return Insert(Folder.CreateAdd(LC, RC, HasNUW, HasNSW), Name); | |||
1174 | return CreateInsertNUWNSWBinOp(Instruction::Add, LHS, RHS, Name, | |||
1175 | HasNUW, HasNSW); | |||
1176 | } | |||
1177 | ||||
1178 | Value *CreateNSWAdd(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1179 | return CreateAdd(LHS, RHS, Name, false, true); | |||
1180 | } | |||
1181 | ||||
1182 | Value *CreateNUWAdd(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1183 | return CreateAdd(LHS, RHS, Name, true, false); | |||
1184 | } | |||
1185 | ||||
1186 | Value *CreateSub(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1187 | bool HasNUW = false, bool HasNSW = false) { | |||
1188 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1189 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1190 | return Insert(Folder.CreateSub(LC, RC, HasNUW, HasNSW), Name); | |||
1191 | return CreateInsertNUWNSWBinOp(Instruction::Sub, LHS, RHS, Name, | |||
1192 | HasNUW, HasNSW); | |||
1193 | } | |||
1194 | ||||
1195 | Value *CreateNSWSub(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1196 | return CreateSub(LHS, RHS, Name, false, true); | |||
1197 | } | |||
1198 | ||||
1199 | Value *CreateNUWSub(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1200 | return CreateSub(LHS, RHS, Name, true, false); | |||
1201 | } | |||
1202 | ||||
1203 | Value *CreateMul(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1204 | bool HasNUW = false, bool HasNSW = false) { | |||
1205 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1206 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1207 | return Insert(Folder.CreateMul(LC, RC, HasNUW, HasNSW), Name); | |||
1208 | return CreateInsertNUWNSWBinOp(Instruction::Mul, LHS, RHS, Name, | |||
1209 | HasNUW, HasNSW); | |||
1210 | } | |||
1211 | ||||
1212 | Value *CreateNSWMul(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1213 | return CreateMul(LHS, RHS, Name, false, true); | |||
1214 | } | |||
1215 | ||||
1216 | Value *CreateNUWMul(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1217 | return CreateMul(LHS, RHS, Name, true, false); | |||
1218 | } | |||
1219 | ||||
1220 | Value *CreateUDiv(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1221 | bool isExact = false) { | |||
1222 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1223 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1224 | return Insert(Folder.CreateUDiv(LC, RC, isExact), Name); | |||
1225 | if (!isExact) | |||
1226 | return Insert(BinaryOperator::CreateUDiv(LHS, RHS), Name); | |||
1227 | return Insert(BinaryOperator::CreateExactUDiv(LHS, RHS), Name); | |||
1228 | } | |||
1229 | ||||
1230 | Value *CreateExactUDiv(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1231 | return CreateUDiv(LHS, RHS, Name, true); | |||
1232 | } | |||
1233 | ||||
1234 | Value *CreateSDiv(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1235 | bool isExact = false) { | |||
1236 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1237 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1238 | return Insert(Folder.CreateSDiv(LC, RC, isExact), Name); | |||
1239 | if (!isExact) | |||
1240 | return Insert(BinaryOperator::CreateSDiv(LHS, RHS), Name); | |||
1241 | return Insert(BinaryOperator::CreateExactSDiv(LHS, RHS), Name); | |||
1242 | } | |||
1243 | ||||
1244 | Value *CreateExactSDiv(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1245 | return CreateSDiv(LHS, RHS, Name, true); | |||
1246 | } | |||
1247 | ||||
1248 | Value *CreateURem(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1249 | if (Value *V = foldConstant(Instruction::URem, LHS, RHS, Name)) return V; | |||
1250 | return Insert(BinaryOperator::CreateURem(LHS, RHS), Name); | |||
1251 | } | |||
1252 | ||||
1253 | Value *CreateSRem(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1254 | if (Value *V = foldConstant(Instruction::SRem, LHS, RHS, Name)) return V; | |||
1255 | return Insert(BinaryOperator::CreateSRem(LHS, RHS), Name); | |||
1256 | } | |||
1257 | ||||
1258 | Value *CreateShl(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1259 | bool HasNUW = false, bool HasNSW = false) { | |||
1260 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1261 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1262 | return Insert(Folder.CreateShl(LC, RC, HasNUW, HasNSW), Name); | |||
1263 | return CreateInsertNUWNSWBinOp(Instruction::Shl, LHS, RHS, Name, | |||
1264 | HasNUW, HasNSW); | |||
1265 | } | |||
1266 | ||||
1267 | Value *CreateShl(Value *LHS, const APInt &RHS, const Twine &Name = "", | |||
1268 | bool HasNUW = false, bool HasNSW = false) { | |||
1269 | return CreateShl(LHS, ConstantInt::get(LHS->getType(), RHS), Name, | |||
1270 | HasNUW, HasNSW); | |||
1271 | } | |||
1272 | ||||
1273 | Value *CreateShl(Value *LHS, uint64_t RHS, const Twine &Name = "", | |||
1274 | bool HasNUW = false, bool HasNSW = false) { | |||
1275 | return CreateShl(LHS, ConstantInt::get(LHS->getType(), RHS), Name, | |||
1276 | HasNUW, HasNSW); | |||
1277 | } | |||
1278 | ||||
1279 | Value *CreateLShr(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1280 | bool isExact = false) { | |||
1281 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1282 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1283 | return Insert(Folder.CreateLShr(LC, RC, isExact), Name); | |||
1284 | if (!isExact) | |||
1285 | return Insert(BinaryOperator::CreateLShr(LHS, RHS), Name); | |||
1286 | return Insert(BinaryOperator::CreateExactLShr(LHS, RHS), Name); | |||
1287 | } | |||
1288 | ||||
1289 | Value *CreateLShr(Value *LHS, const APInt &RHS, const Twine &Name = "", | |||
1290 | bool isExact = false) { | |||
1291 | return CreateLShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); | |||
1292 | } | |||
1293 | ||||
1294 | Value *CreateLShr(Value *LHS, uint64_t RHS, const Twine &Name = "", | |||
1295 | bool isExact = false) { | |||
1296 | return CreateLShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); | |||
1297 | } | |||
1298 | ||||
1299 | Value *CreateAShr(Value *LHS, Value *RHS, const Twine &Name = "", | |||
1300 | bool isExact = false) { | |||
1301 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1302 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
1303 | return Insert(Folder.CreateAShr(LC, RC, isExact), Name); | |||
1304 | if (!isExact) | |||
1305 | return Insert(BinaryOperator::CreateAShr(LHS, RHS), Name); | |||
1306 | return Insert(BinaryOperator::CreateExactAShr(LHS, RHS), Name); | |||
1307 | } | |||
1308 | ||||
1309 | Value *CreateAShr(Value *LHS, const APInt &RHS, const Twine &Name = "", | |||
1310 | bool isExact = false) { | |||
1311 | return CreateAShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); | |||
1312 | } | |||
1313 | ||||
1314 | Value *CreateAShr(Value *LHS, uint64_t RHS, const Twine &Name = "", | |||
1315 | bool isExact = false) { | |||
1316 | return CreateAShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); | |||
1317 | } | |||
1318 | ||||
1319 | Value *CreateAnd(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1320 | if (auto *RC = dyn_cast<Constant>(RHS)) { | |||
1321 | if (isa<ConstantInt>(RC) && cast<ConstantInt>(RC)->isMinusOne()) | |||
1322 | return LHS; // LHS & -1 -> LHS | |||
1323 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1324 | return Insert(Folder.CreateAnd(LC, RC), Name); | |||
1325 | } | |||
1326 | return Insert(BinaryOperator::CreateAnd(LHS, RHS), Name); | |||
1327 | } | |||
1328 | ||||
1329 | Value *CreateAnd(Value *LHS, const APInt &RHS, const Twine &Name = "") { | |||
1330 | return CreateAnd(LHS, ConstantInt::get(LHS->getType(), RHS), Name); | |||
1331 | } | |||
1332 | ||||
1333 | Value *CreateAnd(Value *LHS, uint64_t RHS, const Twine &Name = "") { | |||
1334 | return CreateAnd(LHS, ConstantInt::get(LHS->getType(), RHS), Name); | |||
1335 | } | |||
1336 | ||||
1337 | Value *CreateAnd(ArrayRef<Value*> Ops) { | |||
1338 | assert(!Ops.empty())((!Ops.empty()) ? static_cast<void> (0) : __assert_fail ("!Ops.empty()", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1338, __PRETTY_FUNCTION__)); | |||
1339 | Value *Accum = Ops[0]; | |||
1340 | for (unsigned i = 1; i < Ops.size(); i++) | |||
1341 | Accum = CreateAnd(Accum, Ops[i]); | |||
1342 | return Accum; | |||
1343 | } | |||
1344 | ||||
1345 | Value *CreateOr(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1346 | if (auto *RC = dyn_cast<Constant>(RHS)) { | |||
1347 | if (RC->isNullValue()) | |||
1348 | return LHS; // LHS | 0 -> LHS | |||
1349 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
1350 | return Insert(Folder.CreateOr(LC, RC), Name); | |||
1351 | } | |||
1352 | return Insert(BinaryOperator::CreateOr(LHS, RHS), Name); | |||
1353 | } | |||
1354 | ||||
1355 | Value *CreateOr(Value *LHS, const APInt &RHS, const Twine &Name = "") { | |||
1356 | return CreateOr(LHS, ConstantInt::get(LHS->getType(), RHS), Name); | |||
1357 | } | |||
1358 | ||||
1359 | Value *CreateOr(Value *LHS, uint64_t RHS, const Twine &Name = "") { | |||
1360 | return CreateOr(LHS, ConstantInt::get(LHS->getType(), RHS), Name); | |||
1361 | } | |||
1362 | ||||
1363 | Value *CreateOr(ArrayRef<Value*> Ops) { | |||
1364 | assert(!Ops.empty())((!Ops.empty()) ? static_cast<void> (0) : __assert_fail ("!Ops.empty()", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1364, __PRETTY_FUNCTION__)); | |||
1365 | Value *Accum = Ops[0]; | |||
1366 | for (unsigned i = 1; i < Ops.size(); i++) | |||
1367 | Accum = CreateOr(Accum, Ops[i]); | |||
1368 | return Accum; | |||
1369 | } | |||
1370 | ||||
1371 | Value *CreateXor(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
1372 | if (Value *V = foldConstant(Instruction::Xor, LHS, RHS, Name)) return V; | |||
1373 | return Insert(BinaryOperator::CreateXor(LHS, RHS), Name); | |||
1374 | } | |||
1375 | ||||
1376 | Value *CreateXor(Value *LHS, const APInt &RHS, const Twine &Name = "") { | |||
1377 | return CreateXor(LHS, ConstantInt::get(LHS->getType(), RHS), Name); | |||
1378 | } | |||
1379 | ||||
1380 | Value *CreateXor(Value *LHS, uint64_t RHS, const Twine &Name = "") { | |||
1381 | return CreateXor(LHS, ConstantInt::get(LHS->getType(), RHS), Name); | |||
1382 | } | |||
1383 | ||||
1384 | Value *CreateFAdd(Value *L, Value *R, const Twine &Name = "", | |||
1385 | MDNode *FPMD = nullptr) { | |||
1386 | if (IsFPConstrained) | |||
1387 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fadd, | |||
1388 | L, R, nullptr, Name, FPMD); | |||
1389 | ||||
1390 | if (Value *V = foldConstant(Instruction::FAdd, L, R, Name)) return V; | |||
1391 | Instruction *I = setFPAttrs(BinaryOperator::CreateFAdd(L, R), FPMD, FMF); | |||
1392 | return Insert(I, Name); | |||
1393 | } | |||
1394 | ||||
1395 | /// Copy fast-math-flags from an instruction rather than using the builder's | |||
1396 | /// default FMF. | |||
1397 | Value *CreateFAddFMF(Value *L, Value *R, Instruction *FMFSource, | |||
1398 | const Twine &Name = "") { | |||
1399 | if (IsFPConstrained) | |||
1400 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fadd, | |||
1401 | L, R, FMFSource, Name); | |||
1402 | ||||
1403 | if (Value *V = foldConstant(Instruction::FAdd, L, R, Name)) return V; | |||
1404 | Instruction *I = setFPAttrs(BinaryOperator::CreateFAdd(L, R), nullptr, | |||
1405 | FMFSource->getFastMathFlags()); | |||
1406 | return Insert(I, Name); | |||
1407 | } | |||
1408 | ||||
1409 | Value *CreateFSub(Value *L, Value *R, const Twine &Name = "", | |||
1410 | MDNode *FPMD = nullptr) { | |||
1411 | if (IsFPConstrained) | |||
1412 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fsub, | |||
1413 | L, R, nullptr, Name, FPMD); | |||
1414 | ||||
1415 | if (Value *V = foldConstant(Instruction::FSub, L, R, Name)) return V; | |||
1416 | Instruction *I = setFPAttrs(BinaryOperator::CreateFSub(L, R), FPMD, FMF); | |||
1417 | return Insert(I, Name); | |||
1418 | } | |||
1419 | ||||
1420 | /// Copy fast-math-flags from an instruction rather than using the builder's | |||
1421 | /// default FMF. | |||
1422 | Value *CreateFSubFMF(Value *L, Value *R, Instruction *FMFSource, | |||
1423 | const Twine &Name = "") { | |||
1424 | if (IsFPConstrained) | |||
1425 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fsub, | |||
1426 | L, R, FMFSource, Name); | |||
1427 | ||||
1428 | if (Value *V = foldConstant(Instruction::FSub, L, R, Name)) return V; | |||
1429 | Instruction *I = setFPAttrs(BinaryOperator::CreateFSub(L, R), nullptr, | |||
1430 | FMFSource->getFastMathFlags()); | |||
1431 | return Insert(I, Name); | |||
1432 | } | |||
1433 | ||||
1434 | Value *CreateFMul(Value *L, Value *R, const Twine &Name = "", | |||
1435 | MDNode *FPMD = nullptr) { | |||
1436 | if (IsFPConstrained) | |||
1437 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fmul, | |||
1438 | L, R, nullptr, Name, FPMD); | |||
1439 | ||||
1440 | if (Value *V = foldConstant(Instruction::FMul, L, R, Name)) return V; | |||
1441 | Instruction *I = setFPAttrs(BinaryOperator::CreateFMul(L, R), FPMD, FMF); | |||
1442 | return Insert(I, Name); | |||
1443 | } | |||
1444 | ||||
1445 | /// Copy fast-math-flags from an instruction rather than using the builder's | |||
1446 | /// default FMF. | |||
1447 | Value *CreateFMulFMF(Value *L, Value *R, Instruction *FMFSource, | |||
1448 | const Twine &Name = "") { | |||
1449 | if (IsFPConstrained) | |||
1450 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fmul, | |||
1451 | L, R, FMFSource, Name); | |||
1452 | ||||
1453 | if (Value *V = foldConstant(Instruction::FMul, L, R, Name)) return V; | |||
1454 | Instruction *I = setFPAttrs(BinaryOperator::CreateFMul(L, R), nullptr, | |||
1455 | FMFSource->getFastMathFlags()); | |||
1456 | return Insert(I, Name); | |||
1457 | } | |||
1458 | ||||
1459 | Value *CreateFDiv(Value *L, Value *R, const Twine &Name = "", | |||
1460 | MDNode *FPMD = nullptr) { | |||
1461 | if (IsFPConstrained) | |||
1462 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fdiv, | |||
1463 | L, R, nullptr, Name, FPMD); | |||
1464 | ||||
1465 | if (Value *V = foldConstant(Instruction::FDiv, L, R, Name)) return V; | |||
1466 | Instruction *I = setFPAttrs(BinaryOperator::CreateFDiv(L, R), FPMD, FMF); | |||
1467 | return Insert(I, Name); | |||
1468 | } | |||
1469 | ||||
1470 | /// Copy fast-math-flags from an instruction rather than using the builder's | |||
1471 | /// default FMF. | |||
1472 | Value *CreateFDivFMF(Value *L, Value *R, Instruction *FMFSource, | |||
1473 | const Twine &Name = "") { | |||
1474 | if (IsFPConstrained) | |||
1475 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fdiv, | |||
1476 | L, R, FMFSource, Name); | |||
1477 | ||||
1478 | if (Value *V = foldConstant(Instruction::FDiv, L, R, Name)) return V; | |||
1479 | Instruction *I = setFPAttrs(BinaryOperator::CreateFDiv(L, R), nullptr, | |||
1480 | FMFSource->getFastMathFlags()); | |||
1481 | return Insert(I, Name); | |||
1482 | } | |||
1483 | ||||
1484 | Value *CreateFRem(Value *L, Value *R, const Twine &Name = "", | |||
1485 | MDNode *FPMD = nullptr) { | |||
1486 | if (IsFPConstrained) | |||
1487 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_frem, | |||
1488 | L, R, nullptr, Name, FPMD); | |||
1489 | ||||
1490 | if (Value *V = foldConstant(Instruction::FRem, L, R, Name)) return V; | |||
1491 | Instruction *I = setFPAttrs(BinaryOperator::CreateFRem(L, R), FPMD, FMF); | |||
1492 | return Insert(I, Name); | |||
1493 | } | |||
1494 | ||||
1495 | /// Copy fast-math-flags from an instruction rather than using the builder's | |||
1496 | /// default FMF. | |||
1497 | Value *CreateFRemFMF(Value *L, Value *R, Instruction *FMFSource, | |||
1498 | const Twine &Name = "") { | |||
1499 | if (IsFPConstrained) | |||
1500 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_frem, | |||
1501 | L, R, FMFSource, Name); | |||
1502 | ||||
1503 | if (Value *V = foldConstant(Instruction::FRem, L, R, Name)) return V; | |||
1504 | Instruction *I = setFPAttrs(BinaryOperator::CreateFRem(L, R), nullptr, | |||
1505 | FMFSource->getFastMathFlags()); | |||
1506 | return Insert(I, Name); | |||
1507 | } | |||
1508 | ||||
1509 | Value *CreateBinOp(Instruction::BinaryOps Opc, | |||
1510 | Value *LHS, Value *RHS, const Twine &Name = "", | |||
1511 | MDNode *FPMathTag = nullptr) { | |||
1512 | if (Value *V = foldConstant(Opc, LHS, RHS, Name)) return V; | |||
1513 | Instruction *BinOp = BinaryOperator::Create(Opc, LHS, RHS); | |||
1514 | if (isa<FPMathOperator>(BinOp)) | |||
1515 | setFPAttrs(BinOp, FPMathTag, FMF); | |||
1516 | return Insert(BinOp, Name); | |||
1517 | } | |||
1518 | ||||
1519 | CallInst *CreateConstrainedFPBinOp( | |||
1520 | Intrinsic::ID ID, Value *L, Value *R, Instruction *FMFSource = nullptr, | |||
1521 | const Twine &Name = "", MDNode *FPMathTag = nullptr, | |||
1522 | Optional<RoundingMode> Rounding = None, | |||
1523 | Optional<fp::ExceptionBehavior> Except = None); | |||
1524 | ||||
1525 | Value *CreateNeg(Value *V, const Twine &Name = "", | |||
1526 | bool HasNUW = false, bool HasNSW = false) { | |||
1527 | if (auto *VC = dyn_cast<Constant>(V)) | |||
1528 | return Insert(Folder.CreateNeg(VC, HasNUW, HasNSW), Name); | |||
1529 | BinaryOperator *BO = Insert(BinaryOperator::CreateNeg(V), Name); | |||
1530 | if (HasNUW) BO->setHasNoUnsignedWrap(); | |||
1531 | if (HasNSW) BO->setHasNoSignedWrap(); | |||
1532 | return BO; | |||
1533 | } | |||
1534 | ||||
1535 | Value *CreateNSWNeg(Value *V, const Twine &Name = "") { | |||
1536 | return CreateNeg(V, Name, false, true); | |||
1537 | } | |||
1538 | ||||
1539 | Value *CreateNUWNeg(Value *V, const Twine &Name = "") { | |||
1540 | return CreateNeg(V, Name, true, false); | |||
1541 | } | |||
1542 | ||||
1543 | Value *CreateFNeg(Value *V, const Twine &Name = "", | |||
1544 | MDNode *FPMathTag = nullptr) { | |||
1545 | if (auto *VC = dyn_cast<Constant>(V)) | |||
1546 | return Insert(Folder.CreateFNeg(VC), Name); | |||
1547 | return Insert(setFPAttrs(UnaryOperator::CreateFNeg(V), FPMathTag, FMF), | |||
1548 | Name); | |||
1549 | } | |||
1550 | ||||
1551 | /// Copy fast-math-flags from an instruction rather than using the builder's | |||
1552 | /// default FMF. | |||
1553 | Value *CreateFNegFMF(Value *V, Instruction *FMFSource, | |||
1554 | const Twine &Name = "") { | |||
1555 | if (auto *VC = dyn_cast<Constant>(V)) | |||
1556 | return Insert(Folder.CreateFNeg(VC), Name); | |||
1557 | return Insert(setFPAttrs(UnaryOperator::CreateFNeg(V), nullptr, | |||
1558 | FMFSource->getFastMathFlags()), | |||
1559 | Name); | |||
1560 | } | |||
1561 | ||||
1562 | Value *CreateNot(Value *V, const Twine &Name = "") { | |||
1563 | if (auto *VC = dyn_cast<Constant>(V)) | |||
1564 | return Insert(Folder.CreateNot(VC), Name); | |||
1565 | return Insert(BinaryOperator::CreateNot(V), Name); | |||
1566 | } | |||
1567 | ||||
1568 | Value *CreateUnOp(Instruction::UnaryOps Opc, | |||
1569 | Value *V, const Twine &Name = "", | |||
1570 | MDNode *FPMathTag = nullptr) { | |||
1571 | if (auto *VC = dyn_cast<Constant>(V)) | |||
1572 | return Insert(Folder.CreateUnOp(Opc, VC), Name); | |||
1573 | Instruction *UnOp = UnaryOperator::Create(Opc, V); | |||
1574 | if (isa<FPMathOperator>(UnOp)) | |||
1575 | setFPAttrs(UnOp, FPMathTag, FMF); | |||
1576 | return Insert(UnOp, Name); | |||
1577 | } | |||
1578 | ||||
1579 | /// Create either a UnaryOperator or BinaryOperator depending on \p Opc. | |||
1580 | /// Correct number of operands must be passed accordingly. | |||
1581 | Value *CreateNAryOp(unsigned Opc, ArrayRef<Value *> Ops, | |||
1582 | const Twine &Name = "", MDNode *FPMathTag = nullptr); | |||
1583 | ||||
1584 | //===--------------------------------------------------------------------===// | |||
1585 | // Instruction creation methods: Memory Instructions | |||
1586 | //===--------------------------------------------------------------------===// | |||
1587 | ||||
1588 | AllocaInst *CreateAlloca(Type *Ty, unsigned AddrSpace, | |||
1589 | Value *ArraySize = nullptr, const Twine &Name = "") { | |||
1590 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
1591 | Align AllocaAlign = DL.getPrefTypeAlign(Ty); | |||
1592 | return Insert(new AllocaInst(Ty, AddrSpace, ArraySize, AllocaAlign), Name); | |||
1593 | } | |||
1594 | ||||
1595 | AllocaInst *CreateAlloca(Type *Ty, Value *ArraySize = nullptr, | |||
1596 | const Twine &Name = "") { | |||
1597 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
1598 | Align AllocaAlign = DL.getPrefTypeAlign(Ty); | |||
1599 | unsigned AddrSpace = DL.getAllocaAddrSpace(); | |||
1600 | return Insert(new AllocaInst(Ty, AddrSpace, ArraySize, AllocaAlign), Name); | |||
1601 | } | |||
1602 | ||||
1603 | /// Provided to resolve 'CreateLoad(Ty, Ptr, "...")' correctly, instead of | |||
1604 | /// converting the string to 'bool' for the isVolatile parameter. | |||
1605 | LoadInst *CreateLoad(Type *Ty, Value *Ptr, const char *Name) { | |||
1606 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(), Name); | |||
1607 | } | |||
1608 | ||||
1609 | LoadInst *CreateLoad(Type *Ty, Value *Ptr, const Twine &Name = "") { | |||
1610 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(), Name); | |||
1611 | } | |||
1612 | ||||
1613 | LoadInst *CreateLoad(Type *Ty, Value *Ptr, bool isVolatile, | |||
1614 | const Twine &Name = "") { | |||
1615 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(), isVolatile, Name); | |||
1616 | } | |||
1617 | ||||
1618 | // Deprecated [opaque pointer types] | |||
1619 | LoadInst *CreateLoad(Value *Ptr, const char *Name) { | |||
1620 | return CreateLoad(Ptr->getType()->getPointerElementType(), Ptr, Name); | |||
1621 | } | |||
1622 | ||||
1623 | // Deprecated [opaque pointer types] | |||
1624 | LoadInst *CreateLoad(Value *Ptr, const Twine &Name = "") { | |||
1625 | return CreateLoad(Ptr->getType()->getPointerElementType(), Ptr, Name); | |||
1626 | } | |||
1627 | ||||
1628 | // Deprecated [opaque pointer types] | |||
1629 | LoadInst *CreateLoad(Value *Ptr, bool isVolatile, const Twine &Name = "") { | |||
1630 | return CreateLoad(Ptr->getType()->getPointerElementType(), Ptr, isVolatile, | |||
1631 | Name); | |||
1632 | } | |||
1633 | ||||
1634 | StoreInst *CreateStore(Value *Val, Value *Ptr, bool isVolatile = false) { | |||
1635 | return CreateAlignedStore(Val, Ptr, MaybeAlign(), isVolatile); | |||
1636 | } | |||
1637 | ||||
1638 | LLVM_ATTRIBUTE_DEPRECATED(LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const char *Name) __attribute__((deprecated("Use the version that takes NaybeAlign instead" ))) | |||
1639 | unsigned Align,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const char *Name) __attribute__((deprecated("Use the version that takes NaybeAlign instead" ))) | |||
1640 | const char *Name),LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const char *Name) __attribute__((deprecated("Use the version that takes NaybeAlign instead" ))) | |||
1641 | "Use the version that takes NaybeAlign instead")LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const char *Name) __attribute__((deprecated("Use the version that takes NaybeAlign instead" ))) { | |||
1642 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(Align), Name); | |||
1643 | } | |||
1644 | LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, | |||
1645 | const char *Name) { | |||
1646 | return CreateAlignedLoad(Ty, Ptr, Align, /*isVolatile*/false, Name); | |||
1647 | } | |||
1648 | ||||
1649 | LLVM_ATTRIBUTE_DEPRECATED(LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1650 | unsigned Align,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1651 | const Twine &Name = ""),LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1652 | "Use the version that takes MaybeAlign instead")LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) { | |||
1653 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(Align), Name); | |||
1654 | } | |||
1655 | LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, | |||
1656 | const Twine &Name = "") { | |||
1657 | return CreateAlignedLoad(Ty, Ptr, Align, /*isVolatile*/false, Name); | |||
1658 | } | |||
1659 | ||||
1660 | LLVM_ATTRIBUTE_DEPRECATED(LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , bool isVolatile, const Twine &Name = "") __attribute__( (deprecated("Use the version that takes MaybeAlign instead")) ) | |||
1661 | unsigned Align,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , bool isVolatile, const Twine &Name = "") __attribute__( (deprecated("Use the version that takes MaybeAlign instead")) ) | |||
1662 | bool isVolatile,LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , bool isVolatile, const Twine &Name = "") __attribute__( (deprecated("Use the version that takes MaybeAlign instead")) ) | |||
1663 | const Twine &Name = ""),LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , bool isVolatile, const Twine &Name = "") __attribute__( (deprecated("Use the version that takes MaybeAlign instead")) ) | |||
1664 | "Use the version that takes MaybeAlign instead")LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align , bool isVolatile, const Twine &Name = "") __attribute__( (deprecated("Use the version that takes MaybeAlign instead")) ) { | |||
1665 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(Align), isVolatile, Name); | |||
1666 | } | |||
1667 | LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, | |||
1668 | bool isVolatile, const Twine &Name = "") { | |||
1669 | if (!Align) { | |||
1670 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
1671 | Align = DL.getABITypeAlign(Ty); | |||
1672 | } | |||
1673 | return Insert(new LoadInst(Ty, Ptr, Twine(), isVolatile, *Align), Name); | |||
1674 | } | |||
1675 | ||||
1676 | // Deprecated [opaque pointer types] | |||
1677 | LLVM_ATTRIBUTE_DEPRECATED(LoadInst *CreateAlignedLoad(Value *Ptr,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const char *Name) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1678 | unsigned Align,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const char *Name) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1679 | const char *Name),LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const char *Name) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1680 | "Use the version that takes MaybeAlign instead")LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const char *Name) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) { | |||
1681 | return CreateAlignedLoad(Ptr->getType()->getPointerElementType(), Ptr, | |||
1682 | MaybeAlign(Align), Name); | |||
1683 | } | |||
1684 | // Deprecated [opaque pointer types] | |||
1685 | LLVM_ATTRIBUTE_DEPRECATED(LoadInst *CreateAlignedLoad(Value *Ptr,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1686 | unsigned Align,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1687 | const Twine &Name = ""),LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1688 | "Use the version that takes MaybeAlign instead")LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) { | |||
1689 | return CreateAlignedLoad(Ptr->getType()->getPointerElementType(), Ptr, | |||
1690 | MaybeAlign(Align), Name); | |||
1691 | } | |||
1692 | // Deprecated [opaque pointer types] | |||
1693 | LLVM_ATTRIBUTE_DEPRECATED(LoadInst *CreateAlignedLoad(Value *Ptr,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, bool isVolatile , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1694 | unsigned Align,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, bool isVolatile , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1695 | bool isVolatile,LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, bool isVolatile , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1696 | const Twine &Name = ""),LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, bool isVolatile , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1697 | "Use the version that takes MaybeAlign instead")LoadInst *CreateAlignedLoad(Value *Ptr, unsigned Align, bool isVolatile , const Twine &Name = "") __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) { | |||
1698 | return CreateAlignedLoad(Ptr->getType()->getPointerElementType(), Ptr, | |||
1699 | MaybeAlign(Align), isVolatile, Name); | |||
1700 | } | |||
1701 | // Deprecated [opaque pointer types] | |||
1702 | LoadInst *CreateAlignedLoad(Value *Ptr, MaybeAlign Align, const char *Name) { | |||
1703 | return CreateAlignedLoad(Ptr->getType()->getPointerElementType(), Ptr, | |||
1704 | Align, Name); | |||
1705 | } | |||
1706 | // Deprecated [opaque pointer types] | |||
1707 | LoadInst *CreateAlignedLoad(Value *Ptr, MaybeAlign Align, | |||
1708 | const Twine &Name = "") { | |||
1709 | return CreateAlignedLoad(Ptr->getType()->getPointerElementType(), Ptr, | |||
1710 | Align, Name); | |||
1711 | } | |||
1712 | // Deprecated [opaque pointer types] | |||
1713 | LoadInst *CreateAlignedLoad(Value *Ptr, MaybeAlign Align, bool isVolatile, | |||
1714 | const Twine &Name = "") { | |||
1715 | return CreateAlignedLoad(Ptr->getType()->getPointerElementType(), Ptr, | |||
1716 | Align, isVolatile, Name); | |||
1717 | } | |||
1718 | ||||
1719 | LLVM_ATTRIBUTE_DEPRECATED(StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, unsigned Align, bool isVolatile = false) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1720 | StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, unsigned Align,StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, unsigned Align, bool isVolatile = false) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1721 | bool isVolatile = false),StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, unsigned Align, bool isVolatile = false) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) | |||
1722 | "Use the version that takes MaybeAlign instead")StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, unsigned Align, bool isVolatile = false) __attribute__((deprecated("Use the version that takes MaybeAlign instead" ))) { | |||
1723 | return CreateAlignedStore(Val, Ptr, MaybeAlign(Align), isVolatile); | |||
1724 | } | |||
1725 | StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, MaybeAlign Align, | |||
1726 | bool isVolatile = false) { | |||
1727 | if (!Align) { | |||
1728 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
1729 | Align = DL.getABITypeAlign(Val->getType()); | |||
1730 | } | |||
1731 | return Insert(new StoreInst(Val, Ptr, isVolatile, *Align)); | |||
1732 | } | |||
1733 | FenceInst *CreateFence(AtomicOrdering Ordering, | |||
1734 | SyncScope::ID SSID = SyncScope::System, | |||
1735 | const Twine &Name = "") { | |||
1736 | return Insert(new FenceInst(Context, Ordering, SSID), Name); | |||
1737 | } | |||
1738 | ||||
1739 | AtomicCmpXchgInst *CreateAtomicCmpXchg( | |||
1740 | Value *Ptr, Value *Cmp, Value *New, AtomicOrdering SuccessOrdering, | |||
1741 | AtomicOrdering FailureOrdering, SyncScope::ID SSID = SyncScope::System) { | |||
1742 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
1743 | Align Alignment(DL.getTypeStoreSize(New->getType())); | |||
1744 | return Insert(new AtomicCmpXchgInst( | |||
1745 | Ptr, Cmp, New, Alignment, SuccessOrdering, FailureOrdering, SSID)); | |||
1746 | } | |||
1747 | ||||
1748 | AtomicRMWInst *CreateAtomicRMW(AtomicRMWInst::BinOp Op, Value *Ptr, Value *Val, | |||
1749 | AtomicOrdering Ordering, | |||
1750 | SyncScope::ID SSID = SyncScope::System) { | |||
1751 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
1752 | Align Alignment(DL.getTypeStoreSize(Val->getType())); | |||
1753 | return Insert(new AtomicRMWInst(Op, Ptr, Val, Alignment, Ordering, SSID)); | |||
1754 | } | |||
1755 | ||||
1756 | Value *CreateGEP(Value *Ptr, ArrayRef<Value *> IdxList, | |||
1757 | const Twine &Name = "") { | |||
1758 | return CreateGEP(nullptr, Ptr, IdxList, Name); | |||
1759 | } | |||
1760 | ||||
1761 | Value *CreateGEP(Type *Ty, Value *Ptr, ArrayRef<Value *> IdxList, | |||
1762 | const Twine &Name = "") { | |||
1763 | if (auto *PC = dyn_cast<Constant>(Ptr)) { | |||
1764 | // Every index must be constant. | |||
1765 | size_t i, e; | |||
1766 | for (i = 0, e = IdxList.size(); i != e; ++i) | |||
1767 | if (!isa<Constant>(IdxList[i])) | |||
1768 | break; | |||
1769 | if (i == e) | |||
1770 | return Insert(Folder.CreateGetElementPtr(Ty, PC, IdxList), Name); | |||
1771 | } | |||
1772 | return Insert(GetElementPtrInst::Create(Ty, Ptr, IdxList), Name); | |||
1773 | } | |||
1774 | ||||
1775 | Value *CreateInBoundsGEP(Value *Ptr, ArrayRef<Value *> IdxList, | |||
1776 | const Twine &Name = "") { | |||
1777 | return CreateInBoundsGEP(nullptr, Ptr, IdxList, Name); | |||
1778 | } | |||
1779 | ||||
1780 | Value *CreateInBoundsGEP(Type *Ty, Value *Ptr, ArrayRef<Value *> IdxList, | |||
1781 | const Twine &Name = "") { | |||
1782 | if (auto *PC = dyn_cast<Constant>(Ptr)) { | |||
1783 | // Every index must be constant. | |||
1784 | size_t i, e; | |||
1785 | for (i = 0, e = IdxList.size(); i != e; ++i) | |||
1786 | if (!isa<Constant>(IdxList[i])) | |||
1787 | break; | |||
1788 | if (i == e) | |||
1789 | return Insert(Folder.CreateInBoundsGetElementPtr(Ty, PC, IdxList), | |||
1790 | Name); | |||
1791 | } | |||
1792 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, IdxList), Name); | |||
1793 | } | |||
1794 | ||||
1795 | Value *CreateGEP(Value *Ptr, Value *Idx, const Twine &Name = "") { | |||
1796 | return CreateGEP(nullptr, Ptr, Idx, Name); | |||
1797 | } | |||
1798 | ||||
1799 | Value *CreateGEP(Type *Ty, Value *Ptr, Value *Idx, const Twine &Name = "") { | |||
1800 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1801 | if (auto *IC = dyn_cast<Constant>(Idx)) | |||
1802 | return Insert(Folder.CreateGetElementPtr(Ty, PC, IC), Name); | |||
1803 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idx), Name); | |||
1804 | } | |||
1805 | ||||
1806 | Value *CreateInBoundsGEP(Type *Ty, Value *Ptr, Value *Idx, | |||
1807 | const Twine &Name = "") { | |||
1808 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1809 | if (auto *IC = dyn_cast<Constant>(Idx)) | |||
1810 | return Insert(Folder.CreateInBoundsGetElementPtr(Ty, PC, IC), Name); | |||
1811 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idx), Name); | |||
1812 | } | |||
1813 | ||||
1814 | Value *CreateConstGEP1_32(Value *Ptr, unsigned Idx0, const Twine &Name = "") { | |||
1815 | return CreateConstGEP1_32(nullptr, Ptr, Idx0, Name); | |||
1816 | } | |||
1817 | ||||
1818 | Value *CreateConstGEP1_32(Type *Ty, Value *Ptr, unsigned Idx0, | |||
1819 | const Twine &Name = "") { | |||
1820 | Value *Idx = ConstantInt::get(Type::getInt32Ty(Context), Idx0); | |||
1821 | ||||
1822 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1823 | return Insert(Folder.CreateGetElementPtr(Ty, PC, Idx), Name); | |||
1824 | ||||
1825 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idx), Name); | |||
1826 | } | |||
1827 | ||||
1828 | Value *CreateConstInBoundsGEP1_32(Type *Ty, Value *Ptr, unsigned Idx0, | |||
1829 | const Twine &Name = "") { | |||
1830 | Value *Idx = ConstantInt::get(Type::getInt32Ty(Context), Idx0); | |||
1831 | ||||
1832 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1833 | return Insert(Folder.CreateInBoundsGetElementPtr(Ty, PC, Idx), Name); | |||
1834 | ||||
1835 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idx), Name); | |||
1836 | } | |||
1837 | ||||
1838 | Value *CreateConstGEP2_32(Type *Ty, Value *Ptr, unsigned Idx0, unsigned Idx1, | |||
1839 | const Twine &Name = "") { | |||
1840 | Value *Idxs[] = { | |||
1841 | ConstantInt::get(Type::getInt32Ty(Context), Idx0), | |||
1842 | ConstantInt::get(Type::getInt32Ty(Context), Idx1) | |||
1843 | }; | |||
1844 | ||||
1845 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1846 | return Insert(Folder.CreateGetElementPtr(Ty, PC, Idxs), Name); | |||
1847 | ||||
1848 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idxs), Name); | |||
1849 | } | |||
1850 | ||||
1851 | Value *CreateConstInBoundsGEP2_32(Type *Ty, Value *Ptr, unsigned Idx0, | |||
1852 | unsigned Idx1, const Twine &Name = "") { | |||
1853 | Value *Idxs[] = { | |||
1854 | ConstantInt::get(Type::getInt32Ty(Context), Idx0), | |||
1855 | ConstantInt::get(Type::getInt32Ty(Context), Idx1) | |||
1856 | }; | |||
1857 | ||||
1858 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1859 | return Insert(Folder.CreateInBoundsGetElementPtr(Ty, PC, Idxs), Name); | |||
1860 | ||||
1861 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idxs), Name); | |||
1862 | } | |||
1863 | ||||
1864 | Value *CreateConstGEP1_64(Type *Ty, Value *Ptr, uint64_t Idx0, | |||
1865 | const Twine &Name = "") { | |||
1866 | Value *Idx = ConstantInt::get(Type::getInt64Ty(Context), Idx0); | |||
1867 | ||||
1868 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1869 | return Insert(Folder.CreateGetElementPtr(Ty, PC, Idx), Name); | |||
1870 | ||||
1871 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idx), Name); | |||
1872 | } | |||
1873 | ||||
1874 | Value *CreateConstGEP1_64(Value *Ptr, uint64_t Idx0, const Twine &Name = "") { | |||
1875 | return CreateConstGEP1_64(nullptr, Ptr, Idx0, Name); | |||
1876 | } | |||
1877 | ||||
1878 | Value *CreateConstInBoundsGEP1_64(Type *Ty, Value *Ptr, uint64_t Idx0, | |||
1879 | const Twine &Name = "") { | |||
1880 | Value *Idx = ConstantInt::get(Type::getInt64Ty(Context), Idx0); | |||
1881 | ||||
1882 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1883 | return Insert(Folder.CreateInBoundsGetElementPtr(Ty, PC, Idx), Name); | |||
1884 | ||||
1885 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idx), Name); | |||
1886 | } | |||
1887 | ||||
1888 | Value *CreateConstInBoundsGEP1_64(Value *Ptr, uint64_t Idx0, | |||
1889 | const Twine &Name = "") { | |||
1890 | return CreateConstInBoundsGEP1_64(nullptr, Ptr, Idx0, Name); | |||
1891 | } | |||
1892 | ||||
1893 | Value *CreateConstGEP2_64(Type *Ty, Value *Ptr, uint64_t Idx0, uint64_t Idx1, | |||
1894 | const Twine &Name = "") { | |||
1895 | Value *Idxs[] = { | |||
1896 | ConstantInt::get(Type::getInt64Ty(Context), Idx0), | |||
1897 | ConstantInt::get(Type::getInt64Ty(Context), Idx1) | |||
1898 | }; | |||
1899 | ||||
1900 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1901 | return Insert(Folder.CreateGetElementPtr(Ty, PC, Idxs), Name); | |||
1902 | ||||
1903 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idxs), Name); | |||
1904 | } | |||
1905 | ||||
1906 | Value *CreateConstGEP2_64(Value *Ptr, uint64_t Idx0, uint64_t Idx1, | |||
1907 | const Twine &Name = "") { | |||
1908 | return CreateConstGEP2_64(nullptr, Ptr, Idx0, Idx1, Name); | |||
1909 | } | |||
1910 | ||||
1911 | Value *CreateConstInBoundsGEP2_64(Type *Ty, Value *Ptr, uint64_t Idx0, | |||
1912 | uint64_t Idx1, const Twine &Name = "") { | |||
1913 | Value *Idxs[] = { | |||
1914 | ConstantInt::get(Type::getInt64Ty(Context), Idx0), | |||
1915 | ConstantInt::get(Type::getInt64Ty(Context), Idx1) | |||
1916 | }; | |||
1917 | ||||
1918 | if (auto *PC = dyn_cast<Constant>(Ptr)) | |||
1919 | return Insert(Folder.CreateInBoundsGetElementPtr(Ty, PC, Idxs), Name); | |||
1920 | ||||
1921 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idxs), Name); | |||
1922 | } | |||
1923 | ||||
1924 | Value *CreateConstInBoundsGEP2_64(Value *Ptr, uint64_t Idx0, uint64_t Idx1, | |||
1925 | const Twine &Name = "") { | |||
1926 | return CreateConstInBoundsGEP2_64(nullptr, Ptr, Idx0, Idx1, Name); | |||
1927 | } | |||
1928 | ||||
1929 | Value *CreateStructGEP(Type *Ty, Value *Ptr, unsigned Idx, | |||
1930 | const Twine &Name = "") { | |||
1931 | return CreateConstInBoundsGEP2_32(Ty, Ptr, 0, Idx, Name); | |||
1932 | } | |||
1933 | ||||
1934 | Value *CreateStructGEP(Value *Ptr, unsigned Idx, const Twine &Name = "") { | |||
1935 | return CreateConstInBoundsGEP2_32(nullptr, Ptr, 0, Idx, Name); | |||
1936 | } | |||
1937 | ||||
1938 | /// Same as CreateGlobalString, but return a pointer with "i8*" type | |||
1939 | /// instead of a pointer to array of i8. | |||
1940 | /// | |||
1941 | /// If no module is given via \p M, it is take from the insertion point basic | |||
1942 | /// block. | |||
1943 | Constant *CreateGlobalStringPtr(StringRef Str, const Twine &Name = "", | |||
1944 | unsigned AddressSpace = 0, | |||
1945 | Module *M = nullptr) { | |||
1946 | GlobalVariable *GV = CreateGlobalString(Str, Name, AddressSpace, M); | |||
1947 | Constant *Zero = ConstantInt::get(Type::getInt32Ty(Context), 0); | |||
1948 | Constant *Indices[] = {Zero, Zero}; | |||
1949 | return ConstantExpr::getInBoundsGetElementPtr(GV->getValueType(), GV, | |||
1950 | Indices); | |||
1951 | } | |||
1952 | ||||
1953 | //===--------------------------------------------------------------------===// | |||
1954 | // Instruction creation methods: Cast/Conversion Operators | |||
1955 | //===--------------------------------------------------------------------===// | |||
1956 | ||||
1957 | Value *CreateTrunc(Value *V, Type *DestTy, const Twine &Name = "") { | |||
1958 | return CreateCast(Instruction::Trunc, V, DestTy, Name); | |||
1959 | } | |||
1960 | ||||
1961 | Value *CreateZExt(Value *V, Type *DestTy, const Twine &Name = "") { | |||
1962 | return CreateCast(Instruction::ZExt, V, DestTy, Name); | |||
1963 | } | |||
1964 | ||||
1965 | Value *CreateSExt(Value *V, Type *DestTy, const Twine &Name = "") { | |||
1966 | return CreateCast(Instruction::SExt, V, DestTy, Name); | |||
1967 | } | |||
1968 | ||||
1969 | /// Create a ZExt or Trunc from the integer value V to DestTy. Return | |||
1970 | /// the value untouched if the type of V is already DestTy. | |||
1971 | Value *CreateZExtOrTrunc(Value *V, Type *DestTy, | |||
1972 | const Twine &Name = "") { | |||
1973 | assert(V->getType()->isIntOrIntVectorTy() &&((V->getType()->isIntOrIntVectorTy() && DestTy-> isIntOrIntVectorTy() && "Can only zero extend/truncate integers!" ) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only zero extend/truncate integers!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1975, __PRETTY_FUNCTION__)) | |||
1974 | DestTy->isIntOrIntVectorTy() &&((V->getType()->isIntOrIntVectorTy() && DestTy-> isIntOrIntVectorTy() && "Can only zero extend/truncate integers!" ) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only zero extend/truncate integers!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1975, __PRETTY_FUNCTION__)) | |||
1975 | "Can only zero extend/truncate integers!")((V->getType()->isIntOrIntVectorTy() && DestTy-> isIntOrIntVectorTy() && "Can only zero extend/truncate integers!" ) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only zero extend/truncate integers!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1975, __PRETTY_FUNCTION__)); | |||
1976 | Type *VTy = V->getType(); | |||
1977 | if (VTy->getScalarSizeInBits() < DestTy->getScalarSizeInBits()) | |||
1978 | return CreateZExt(V, DestTy, Name); | |||
1979 | if (VTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits()) | |||
1980 | return CreateTrunc(V, DestTy, Name); | |||
1981 | return V; | |||
1982 | } | |||
1983 | ||||
1984 | /// Create a SExt or Trunc from the integer value V to DestTy. Return | |||
1985 | /// the value untouched if the type of V is already DestTy. | |||
1986 | Value *CreateSExtOrTrunc(Value *V, Type *DestTy, | |||
1987 | const Twine &Name = "") { | |||
1988 | assert(V->getType()->isIntOrIntVectorTy() &&((V->getType()->isIntOrIntVectorTy() && DestTy-> isIntOrIntVectorTy() && "Can only sign extend/truncate integers!" ) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only sign extend/truncate integers!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1990, __PRETTY_FUNCTION__)) | |||
1989 | DestTy->isIntOrIntVectorTy() &&((V->getType()->isIntOrIntVectorTy() && DestTy-> isIntOrIntVectorTy() && "Can only sign extend/truncate integers!" ) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only sign extend/truncate integers!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1990, __PRETTY_FUNCTION__)) | |||
1990 | "Can only sign extend/truncate integers!")((V->getType()->isIntOrIntVectorTy() && DestTy-> isIntOrIntVectorTy() && "Can only sign extend/truncate integers!" ) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only sign extend/truncate integers!\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/include/llvm/IR/IRBuilder.h" , 1990, __PRETTY_FUNCTION__)); | |||
1991 | Type *VTy = V->getType(); | |||
1992 | if (VTy->getScalarSizeInBits() < DestTy->getScalarSizeInBits()) | |||
1993 | return CreateSExt(V, DestTy, Name); | |||
1994 | if (VTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits()) | |||
1995 | return CreateTrunc(V, DestTy, Name); | |||
1996 | return V; | |||
1997 | } | |||
1998 | ||||
1999 | Value *CreateFPToUI(Value *V, Type *DestTy, const Twine &Name = "") { | |||
2000 | if (IsFPConstrained) | |||
2001 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_fptoui, | |||
2002 | V, DestTy, nullptr, Name); | |||
2003 | return CreateCast(Instruction::FPToUI, V, DestTy, Name); | |||
2004 | } | |||
2005 | ||||
2006 | Value *CreateFPToSI(Value *V, Type *DestTy, const Twine &Name = "") { | |||
2007 | if (IsFPConstrained) | |||
2008 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_fptosi, | |||
2009 | V, DestTy, nullptr, Name); | |||
2010 | return CreateCast(Instruction::FPToSI, V, DestTy, Name); | |||
2011 | } | |||
2012 | ||||
2013 | Value *CreateUIToFP(Value *V, Type *DestTy, const Twine &Name = ""){ | |||
2014 | if (IsFPConstrained) | |||
2015 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_uitofp, | |||
2016 | V, DestTy, nullptr, Name); | |||
2017 | return CreateCast(Instruction::UIToFP, V, DestTy, Name); | |||
2018 | } | |||
2019 | ||||
2020 | Value *CreateSIToFP(Value *V, Type *DestTy, const Twine &Name = ""){ | |||
2021 | if (IsFPConstrained) | |||
2022 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_sitofp, | |||
2023 | V, DestTy, nullptr, Name); | |||
2024 | return CreateCast(Instruction::SIToFP, V, DestTy, Name); | |||
2025 | } | |||
2026 | ||||
2027 | Value *CreateFPTrunc(Value *V, Type *DestTy, | |||
2028 | const Twine &Name = "") { | |||
2029 | if (IsFPConstrained) | |||
2030 | return CreateConstrainedFPCast( | |||
2031 | Intrinsic::experimental_constrained_fptrunc, V, DestTy, nullptr, | |||
2032 | Name); | |||
2033 | return CreateCast(Instruction::FPTrunc, V, DestTy, Name); | |||
2034 | } | |||
2035 | ||||
2036 | Value *CreateFPExt(Value *V, Type *DestTy, const Twine &Name = "") { | |||
2037 | if (IsFPConstrained) | |||
2038 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_fpext, | |||
2039 | V, DestTy, nullptr, Name); | |||
2040 | return CreateCast(Instruction::FPExt, V, DestTy, Name); | |||
2041 | } | |||
2042 | ||||
2043 | Value *CreatePtrToInt(Value *V, Type *DestTy, | |||
2044 | const Twine &Name = "") { | |||
2045 | return CreateCast(Instruction::PtrToInt, V, DestTy, Name); | |||
2046 | } | |||
2047 | ||||
2048 | Value *CreateIntToPtr(Value *V, Type *DestTy, | |||
2049 | const Twine &Name = "") { | |||
2050 | return CreateCast(Instruction::IntToPtr, V, DestTy, Name); | |||
2051 | } | |||
2052 | ||||
2053 | Value *CreateBitCast(Value *V, Type *DestTy, | |||
2054 | const Twine &Name = "") { | |||
2055 | return CreateCast(Instruction::BitCast, V, DestTy, Name); | |||
2056 | } | |||
2057 | ||||
2058 | Value *CreateAddrSpaceCast(Value *V, Type *DestTy, | |||
2059 | const Twine &Name = "") { | |||
2060 | return CreateCast(Instruction::AddrSpaceCast, V, DestTy, Name); | |||
2061 | } | |||
2062 | ||||
2063 | Value *CreateZExtOrBitCast(Value *V, Type *DestTy, | |||
2064 | const Twine &Name = "") { | |||
2065 | if (V->getType() == DestTy) | |||
2066 | return V; | |||
2067 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2068 | return Insert(Folder.CreateZExtOrBitCast(VC, DestTy), Name); | |||
2069 | return Insert(CastInst::CreateZExtOrBitCast(V, DestTy), Name); | |||
2070 | } | |||
2071 | ||||
2072 | Value *CreateSExtOrBitCast(Value *V, Type *DestTy, | |||
2073 | const Twine &Name = "") { | |||
2074 | if (V->getType() == DestTy) | |||
2075 | return V; | |||
2076 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2077 | return Insert(Folder.CreateSExtOrBitCast(VC, DestTy), Name); | |||
2078 | return Insert(CastInst::CreateSExtOrBitCast(V, DestTy), Name); | |||
2079 | } | |||
2080 | ||||
2081 | Value *CreateTruncOrBitCast(Value *V, Type *DestTy, | |||
2082 | const Twine &Name = "") { | |||
2083 | if (V->getType() == DestTy) | |||
2084 | return V; | |||
2085 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2086 | return Insert(Folder.CreateTruncOrBitCast(VC, DestTy), Name); | |||
2087 | return Insert(CastInst::CreateTruncOrBitCast(V, DestTy), Name); | |||
2088 | } | |||
2089 | ||||
2090 | Value *CreateCast(Instruction::CastOps Op, Value *V, Type *DestTy, | |||
2091 | const Twine &Name = "") { | |||
2092 | if (V->getType() == DestTy) | |||
2093 | return V; | |||
2094 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2095 | return Insert(Folder.CreateCast(Op, VC, DestTy), Name); | |||
2096 | return Insert(CastInst::Create(Op, V, DestTy), Name); | |||
2097 | } | |||
2098 | ||||
2099 | Value *CreatePointerCast(Value *V, Type *DestTy, | |||
2100 | const Twine &Name = "") { | |||
2101 | if (V->getType() == DestTy) | |||
2102 | return V; | |||
2103 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2104 | return Insert(Folder.CreatePointerCast(VC, DestTy), Name); | |||
2105 | return Insert(CastInst::CreatePointerCast(V, DestTy), Name); | |||
2106 | } | |||
2107 | ||||
2108 | Value *CreatePointerBitCastOrAddrSpaceCast(Value *V, Type *DestTy, | |||
2109 | const Twine &Name = "") { | |||
2110 | if (V->getType() == DestTy) | |||
2111 | return V; | |||
2112 | ||||
2113 | if (auto *VC = dyn_cast<Constant>(V)) { | |||
2114 | return Insert(Folder.CreatePointerBitCastOrAddrSpaceCast(VC, DestTy), | |||
2115 | Name); | |||
2116 | } | |||
2117 | ||||
2118 | return Insert(CastInst::CreatePointerBitCastOrAddrSpaceCast(V, DestTy), | |||
2119 | Name); | |||
2120 | } | |||
2121 | ||||
2122 | Value *CreateIntCast(Value *V, Type *DestTy, bool isSigned, | |||
2123 | const Twine &Name = "") { | |||
2124 | if (V->getType() == DestTy) | |||
| ||||
2125 | return V; | |||
2126 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2127 | return Insert(Folder.CreateIntCast(VC, DestTy, isSigned), Name); | |||
2128 | return Insert(CastInst::CreateIntegerCast(V, DestTy, isSigned), Name); | |||
2129 | } | |||
2130 | ||||
2131 | Value *CreateBitOrPointerCast(Value *V, Type *DestTy, | |||
2132 | const Twine &Name = "") { | |||
2133 | if (V->getType() == DestTy) | |||
2134 | return V; | |||
2135 | if (V->getType()->isPtrOrPtrVectorTy() && DestTy->isIntOrIntVectorTy()) | |||
2136 | return CreatePtrToInt(V, DestTy, Name); | |||
2137 | if (V->getType()->isIntOrIntVectorTy() && DestTy->isPtrOrPtrVectorTy()) | |||
2138 | return CreateIntToPtr(V, DestTy, Name); | |||
2139 | ||||
2140 | return CreateBitCast(V, DestTy, Name); | |||
2141 | } | |||
2142 | ||||
2143 | Value *CreateFPCast(Value *V, Type *DestTy, const Twine &Name = "") { | |||
2144 | if (V->getType() == DestTy) | |||
2145 | return V; | |||
2146 | if (auto *VC = dyn_cast<Constant>(V)) | |||
2147 | return Insert(Folder.CreateFPCast(VC, DestTy), Name); | |||
2148 | return Insert(CastInst::CreateFPCast(V, DestTy), Name); | |||
2149 | } | |||
2150 | ||||
2151 | CallInst *CreateConstrainedFPCast( | |||
2152 | Intrinsic::ID ID, Value *V, Type *DestTy, | |||
2153 | Instruction *FMFSource = nullptr, const Twine &Name = "", | |||
2154 | MDNode *FPMathTag = nullptr, | |||
2155 | Optional<RoundingMode> Rounding = None, | |||
2156 | Optional<fp::ExceptionBehavior> Except = None); | |||
2157 | ||||
2158 | // Provided to resolve 'CreateIntCast(Ptr, Ptr, "...")', giving a | |||
2159 | // compile time error, instead of converting the string to bool for the | |||
2160 | // isSigned parameter. | |||
2161 | Value *CreateIntCast(Value *, Type *, const char *) = delete; | |||
2162 | ||||
2163 | //===--------------------------------------------------------------------===// | |||
2164 | // Instruction creation methods: Compare Instructions | |||
2165 | //===--------------------------------------------------------------------===// | |||
2166 | ||||
2167 | Value *CreateICmpEQ(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2168 | return CreateICmp(ICmpInst::ICMP_EQ, LHS, RHS, Name); | |||
2169 | } | |||
2170 | ||||
2171 | Value *CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2172 | return CreateICmp(ICmpInst::ICMP_NE, LHS, RHS, Name); | |||
2173 | } | |||
2174 | ||||
2175 | Value *CreateICmpUGT(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2176 | return CreateICmp(ICmpInst::ICMP_UGT, LHS, RHS, Name); | |||
2177 | } | |||
2178 | ||||
2179 | Value *CreateICmpUGE(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2180 | return CreateICmp(ICmpInst::ICMP_UGE, LHS, RHS, Name); | |||
2181 | } | |||
2182 | ||||
2183 | Value *CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2184 | return CreateICmp(ICmpInst::ICMP_ULT, LHS, RHS, Name); | |||
2185 | } | |||
2186 | ||||
2187 | Value *CreateICmpULE(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2188 | return CreateICmp(ICmpInst::ICMP_ULE, LHS, RHS, Name); | |||
2189 | } | |||
2190 | ||||
2191 | Value *CreateICmpSGT(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2192 | return CreateICmp(ICmpInst::ICMP_SGT, LHS, RHS, Name); | |||
2193 | } | |||
2194 | ||||
2195 | Value *CreateICmpSGE(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2196 | return CreateICmp(ICmpInst::ICMP_SGE, LHS, RHS, Name); | |||
2197 | } | |||
2198 | ||||
2199 | Value *CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2200 | return CreateICmp(ICmpInst::ICMP_SLT, LHS, RHS, Name); | |||
2201 | } | |||
2202 | ||||
2203 | Value *CreateICmpSLE(Value *LHS, Value *RHS, const Twine &Name = "") { | |||
2204 | return CreateICmp(ICmpInst::ICMP_SLE, LHS, RHS, Name); | |||
2205 | } | |||
2206 | ||||
2207 | Value *CreateFCmpOEQ(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2208 | MDNode *FPMathTag = nullptr) { | |||
2209 | return CreateFCmp(FCmpInst::FCMP_OEQ, LHS, RHS, Name, FPMathTag); | |||
2210 | } | |||
2211 | ||||
2212 | Value *CreateFCmpOGT(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2213 | MDNode *FPMathTag = nullptr) { | |||
2214 | return CreateFCmp(FCmpInst::FCMP_OGT, LHS, RHS, Name, FPMathTag); | |||
2215 | } | |||
2216 | ||||
2217 | Value *CreateFCmpOGE(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2218 | MDNode *FPMathTag = nullptr) { | |||
2219 | return CreateFCmp(FCmpInst::FCMP_OGE, LHS, RHS, Name, FPMathTag); | |||
2220 | } | |||
2221 | ||||
2222 | Value *CreateFCmpOLT(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2223 | MDNode *FPMathTag = nullptr) { | |||
2224 | return CreateFCmp(FCmpInst::FCMP_OLT, LHS, RHS, Name, FPMathTag); | |||
2225 | } | |||
2226 | ||||
2227 | Value *CreateFCmpOLE(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2228 | MDNode *FPMathTag = nullptr) { | |||
2229 | return CreateFCmp(FCmpInst::FCMP_OLE, LHS, RHS, Name, FPMathTag); | |||
2230 | } | |||
2231 | ||||
2232 | Value *CreateFCmpONE(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2233 | MDNode *FPMathTag = nullptr) { | |||
2234 | return CreateFCmp(FCmpInst::FCMP_ONE, LHS, RHS, Name, FPMathTag); | |||
2235 | } | |||
2236 | ||||
2237 | Value *CreateFCmpORD(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2238 | MDNode *FPMathTag = nullptr) { | |||
2239 | return CreateFCmp(FCmpInst::FCMP_ORD, LHS, RHS, Name, FPMathTag); | |||
2240 | } | |||
2241 | ||||
2242 | Value *CreateFCmpUNO(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2243 | MDNode *FPMathTag = nullptr) { | |||
2244 | return CreateFCmp(FCmpInst::FCMP_UNO, LHS, RHS, Name, FPMathTag); | |||
2245 | } | |||
2246 | ||||
2247 | Value *CreateFCmpUEQ(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2248 | MDNode *FPMathTag = nullptr) { | |||
2249 | return CreateFCmp(FCmpInst::FCMP_UEQ, LHS, RHS, Name, FPMathTag); | |||
2250 | } | |||
2251 | ||||
2252 | Value *CreateFCmpUGT(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2253 | MDNode *FPMathTag = nullptr) { | |||
2254 | return CreateFCmp(FCmpInst::FCMP_UGT, LHS, RHS, Name, FPMathTag); | |||
2255 | } | |||
2256 | ||||
2257 | Value *CreateFCmpUGE(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2258 | MDNode *FPMathTag = nullptr) { | |||
2259 | return CreateFCmp(FCmpInst::FCMP_UGE, LHS, RHS, Name, FPMathTag); | |||
2260 | } | |||
2261 | ||||
2262 | Value *CreateFCmpULT(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2263 | MDNode *FPMathTag = nullptr) { | |||
2264 | return CreateFCmp(FCmpInst::FCMP_ULT, LHS, RHS, Name, FPMathTag); | |||
2265 | } | |||
2266 | ||||
2267 | Value *CreateFCmpULE(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2268 | MDNode *FPMathTag = nullptr) { | |||
2269 | return CreateFCmp(FCmpInst::FCMP_ULE, LHS, RHS, Name, FPMathTag); | |||
2270 | } | |||
2271 | ||||
2272 | Value *CreateFCmpUNE(Value *LHS, Value *RHS, const Twine &Name = "", | |||
2273 | MDNode *FPMathTag = nullptr) { | |||
2274 | return CreateFCmp(FCmpInst::FCMP_UNE, LHS, RHS, Name, FPMathTag); | |||
2275 | } | |||
2276 | ||||
2277 | Value *CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, | |||
2278 | const Twine &Name = "") { | |||
2279 | if (auto *LC = dyn_cast<Constant>(LHS)) | |||
2280 | if (auto *RC = dyn_cast<Constant>(RHS)) | |||
2281 | return Insert(Folder.CreateICmp(P, LC, RC), Name); | |||
2282 | return Insert(new ICmpInst(P, LHS, RHS), Name); | |||
2283 | } | |||
2284 | ||||
2285 | // Create a quiet floating-point comparison (i.e. one that raises an FP | |||
2286 | // exception only in the case where an input is a signaling NaN). | |||
2287 | // Note that this differs from CreateFCmpS only if IsFPConstrained is true. | |||
2288 | Value *CreateFCmp(CmpInst::Predicate P, Value *LHS, Value *RHS, | |||
2289 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { | |||
2290 | return CreateFCmpHelper(P, LHS, RHS, Name, FPMathTag, false); | |||
2291 | } | |||
2292 | ||||
2293 | Value *CreateCmp(CmpInst::Predicate Pred, Value *LHS, Value *RHS, | |||
2294 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { | |||
2295 | return CmpInst::isFPPredicate(Pred) | |||
2296 | ? CreateFCmp(Pred, LHS, RHS, Name, FPMathTag) | |||
2297 | : CreateICmp(Pred, LHS, RHS, Name); | |||
2298 | } | |||
2299 | ||||
2300 | // Create a signaling floating-point comparison (i.e. one that raises an FP | |||
2301 | // exception whenever an input is any NaN, signaling or quiet). | |||
2302 | // Note that this differs from CreateFCmp only if IsFPConstrained is true. | |||
2303 | Value *CreateFCmpS(CmpInst::Predicate P, Value *LHS, Value *RHS, | |||
2304 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { | |||
2305 | return CreateFCmpHelper(P, LHS, RHS, Name, FPMathTag, true); | |||
2306 | } | |||
2307 | ||||
2308 | private: | |||
2309 | // Helper routine to create either a signaling or a quiet FP comparison. | |||
2310 | Value *CreateFCmpHelper(CmpInst::Predicate P, Value *LHS, Value *RHS, | |||
2311 | const Twine &Name, MDNode *FPMathTag, | |||
2312 | bool IsSignaling); | |||
2313 | ||||
2314 | public: | |||
2315 | CallInst *CreateConstrainedFPCmp( | |||
2316 | Intrinsic::ID ID, CmpInst::Predicate P, Value *L, Value *R, | |||
2317 | const Twine &Name = "", Optional<fp::ExceptionBehavior> Except = None); | |||
2318 | ||||
2319 | //===--------------------------------------------------------------------===// | |||
2320 | // Instruction creation methods: Other Instructions | |||
2321 | //===--------------------------------------------------------------------===// | |||
2322 | ||||
2323 | PHINode *CreatePHI(Type *Ty, unsigned NumReservedValues, | |||
2324 | const Twine &Name = "") { | |||
2325 | PHINode *Phi = PHINode::Create(Ty, NumReservedValues); | |||
2326 | if (isa<FPMathOperator>(Phi)) | |||
2327 | setFPAttrs(Phi, nullptr /* MDNode* */, FMF); | |||
2328 | return Insert(Phi, Name); | |||
2329 | } | |||
2330 | ||||
2331 | CallInst *CreateCall(FunctionType *FTy, Value *Callee, | |||
2332 | ArrayRef<Value *> Args = None, const Twine &Name = "", | |||
2333 | MDNode *FPMathTag = nullptr) { | |||
2334 | CallInst *CI = CallInst::Create(FTy, Callee, Args, DefaultOperandBundles); | |||
2335 | if (IsFPConstrained) | |||
2336 | setConstrainedFPCallAttr(CI); | |||
2337 | if (isa<FPMathOperator>(CI)) | |||
2338 | setFPAttrs(CI, FPMathTag, FMF); | |||
2339 | return Insert(CI, Name); | |||
2340 | } | |||
2341 | ||||
2342 | CallInst *CreateCall(FunctionType *FTy, Value *Callee, ArrayRef<Value *> Args, | |||
2343 | ArrayRef<OperandBundleDef> OpBundles, | |||
2344 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { | |||
2345 | CallInst *CI = CallInst::Create(FTy, Callee, Args, OpBundles); | |||
2346 | if (IsFPConstrained) | |||
2347 | setConstrainedFPCallAttr(CI); | |||
2348 | if (isa<FPMathOperator>(CI)) | |||
2349 | setFPAttrs(CI, FPMathTag, FMF); | |||
2350 | return Insert(CI, Name); | |||
2351 | } | |||
2352 | ||||
2353 | CallInst *CreateCall(FunctionCallee Callee, ArrayRef<Value *> Args = None, | |||
2354 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { | |||
2355 | return CreateCall(Callee.getFunctionType(), Callee.getCallee(), Args, Name, | |||
2356 | FPMathTag); | |||
2357 | } | |||
2358 | ||||
2359 | CallInst *CreateCall(FunctionCallee Callee, ArrayRef<Value *> Args, | |||
2360 | ArrayRef<OperandBundleDef> OpBundles, | |||
2361 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { | |||
2362 | return CreateCall(Callee.getFunctionType(), Callee.getCallee(), Args, | |||
2363 | OpBundles, Name, FPMathTag); | |||
2364 | } | |||
2365 | ||||
2366 | CallInst *CreateConstrainedFPCall( | |||
2367 | Function *Callee, ArrayRef<Value *> Args, const Twine &Name = "", | |||
2368 | Optional<RoundingMode> Rounding = None, | |||
2369 | Optional<fp::ExceptionBehavior> Except = None); | |||
2370 | ||||
2371 | Value *CreateSelect(Value *C, Value *True, Value *False, | |||
2372 | const Twine &Name = "", Instruction *MDFrom = nullptr); | |||
2373 | ||||
2374 | VAArgInst *CreateVAArg(Value *List, Type *Ty, const Twine &Name = "") { | |||
2375 | return Insert(new VAArgInst(List, Ty), Name); | |||
2376 | } | |||
2377 | ||||
2378 | Value *CreateExtractElement(Value *Vec, Value *Idx, | |||
2379 | const Twine &Name = "") { | |||
2380 | if (auto *VC = dyn_cast<Constant>(Vec)) | |||
2381 | if (auto *IC = dyn_cast<Constant>(Idx)) | |||
2382 | return Insert(Folder.CreateExtractElement(VC, IC), Name); | |||
2383 | return Insert(ExtractElementInst::Create(Vec, Idx), Name); | |||
2384 | } | |||
2385 | ||||
2386 | Value *CreateExtractElement(Value *Vec, uint64_t Idx, | |||
2387 | const Twine &Name = "") { | |||
2388 | return CreateExtractElement(Vec, getInt64(Idx), Name); | |||
2389 | } | |||
2390 | ||||
2391 | Value *CreateInsertElement(Value *Vec, Value *NewElt, Value *Idx, | |||
2392 | const Twine &Name = "") { | |||
2393 | if (auto *VC = dyn_cast<Constant>(Vec)) | |||
2394 | if (auto *NC = dyn_cast<Constant>(NewElt)) | |||
2395 | if (auto *IC = dyn_cast<Constant>(Idx)) | |||
2396 | return Insert(Folder.CreateInsertElement(VC, NC, IC), Name); | |||
2397 | return Insert(InsertElementInst::Create(Vec, NewElt, Idx), Name); | |||
2398 | } | |||
2399 | ||||
2400 | Value *CreateInsertElement(Value *Vec, Value *NewElt, uint64_t Idx, | |||
2401 | const Twine &Name = "") { | |||
2402 | return CreateInsertElement(Vec, NewElt, getInt64(Idx), Name); | |||
2403 | } | |||
2404 | ||||
2405 | Value *CreateShuffleVector(Value *V1, Value *V2, Value *Mask, | |||
2406 | const Twine &Name = "") { | |||
2407 | SmallVector<int, 16> IntMask; | |||
2408 | ShuffleVectorInst::getShuffleMask(cast<Constant>(Mask), IntMask); | |||
2409 | return CreateShuffleVector(V1, V2, IntMask, Name); | |||
2410 | } | |||
2411 | ||||
2412 | LLVM_ATTRIBUTE_DEPRECATED(Value *CreateShuffleVector(Value *V1, Value *V2,Value *CreateShuffleVector(Value *V1, Value *V2, ArrayRef< uint32_t> Mask, const Twine &Name = "") __attribute__( (deprecated("Pass indices as 'int' instead"))) | |||
2413 | ArrayRef<uint32_t> Mask,Value *CreateShuffleVector(Value *V1, Value *V2, ArrayRef< uint32_t> Mask, const Twine &Name = "") __attribute__( (deprecated("Pass indices as 'int' instead"))) | |||
2414 | const Twine &Name = ""),Value *CreateShuffleVector(Value *V1, Value *V2, ArrayRef< uint32_t> Mask, const Twine &Name = "") __attribute__( (deprecated("Pass indices as 'int' instead"))) | |||
2415 | "Pass indices as 'int' instead")Value *CreateShuffleVector(Value *V1, Value *V2, ArrayRef< uint32_t> Mask, const Twine &Name = "") __attribute__( (deprecated("Pass indices as 'int' instead"))) { | |||
2416 | SmallVector<int, 16> IntMask; | |||
2417 | IntMask.assign(Mask.begin(), Mask.end()); | |||
2418 | return CreateShuffleVector(V1, V2, IntMask, Name); | |||
2419 | } | |||
2420 | ||||
2421 | /// See class ShuffleVectorInst for a description of the mask representation. | |||
2422 | Value *CreateShuffleVector(Value *V1, Value *V2, ArrayRef<int> Mask, | |||
2423 | const Twine &Name = "") { | |||
2424 | if (auto *V1C = dyn_cast<Constant>(V1)) | |||
2425 | if (auto *V2C = dyn_cast<Constant>(V2)) | |||
2426 | return Insert(Folder.CreateShuffleVector(V1C, V2C, Mask), Name); | |||
2427 | return Insert(new ShuffleVectorInst(V1, V2, Mask), Name); | |||
2428 | } | |||
2429 | ||||
2430 | Value *CreateExtractValue(Value *Agg, | |||
2431 | ArrayRef<unsigned> Idxs, | |||
2432 | const Twine &Name = "") { | |||
2433 | if (auto *AggC = dyn_cast<Constant>(Agg)) | |||
2434 | return Insert(Folder.CreateExtractValue(AggC, Idxs), Name); | |||
2435 | return Insert(ExtractValueInst::Create(Agg, Idxs), Name); | |||
2436 | } | |||
2437 | ||||
2438 | Value *CreateInsertValue(Value *Agg, Value *Val, | |||
2439 | ArrayRef<unsigned> Idxs, | |||
2440 | const Twine &Name = "") { | |||
2441 | if (auto *AggC = dyn_cast<Constant>(Agg)) | |||
2442 | if (auto *ValC = dyn_cast<Constant>(Val)) | |||
2443 | return Insert(Folder.CreateInsertValue(AggC, ValC, Idxs), Name); | |||
2444 | return Insert(InsertValueInst::Create(Agg, Val, Idxs), Name); | |||
2445 | } | |||
2446 | ||||
2447 | LandingPadInst *CreateLandingPad(Type *Ty, unsigned NumClauses, | |||
2448 | const Twine &Name = "") { | |||
2449 | return Insert(LandingPadInst::Create(Ty, NumClauses), Name); | |||
2450 | } | |||
2451 | ||||
2452 | Value *CreateFreeze(Value *V, const Twine &Name = "") { | |||
2453 | return Insert(new FreezeInst(V), Name); | |||
2454 | } | |||
2455 | ||||
2456 | //===--------------------------------------------------------------------===// | |||
2457 | // Utility creation methods | |||
2458 | //===--------------------------------------------------------------------===// | |||
2459 | ||||
2460 | /// Return an i1 value testing if \p Arg is null. | |||
2461 | Value *CreateIsNull(Value *Arg, const Twine &Name = "") { | |||
2462 | return CreateICmpEQ(Arg, Constant::getNullValue(Arg->getType()), | |||
2463 | Name); | |||
2464 | } | |||
2465 | ||||
2466 | /// Return an i1 value testing if \p Arg is not null. | |||
2467 | Value *CreateIsNotNull(Value *Arg, const Twine &Name = "") { | |||
2468 | return CreateICmpNE(Arg, Constant::getNullValue(Arg->getType()), | |||
2469 | Name); | |||
2470 | } | |||
2471 | ||||
2472 | /// Return the i64 difference between two pointer values, dividing out | |||
2473 | /// the size of the pointed-to objects. | |||
2474 | /// | |||
2475 | /// This is intended to implement C-style pointer subtraction. As such, the | |||
2476 | /// pointers must be appropriately aligned for their element types and | |||
2477 | /// pointing into the same object. | |||
2478 | Value *CreatePtrDiff(Value *LHS, Value *RHS, const Twine &Name = ""); | |||
2479 | ||||
2480 | /// Create a launder.invariant.group intrinsic call. If Ptr type is | |||
2481 | /// different from pointer to i8, it's casted to pointer to i8 in the same | |||
2482 | /// address space before call and casted back to Ptr type after call. | |||
2483 | Value *CreateLaunderInvariantGroup(Value *Ptr); | |||
2484 | ||||
2485 | /// \brief Create a strip.invariant.group intrinsic call. If Ptr type is | |||
2486 | /// different from pointer to i8, it's casted to pointer to i8 in the same | |||
2487 | /// address space before call and casted back to Ptr type after call. | |||
2488 | Value *CreateStripInvariantGroup(Value *Ptr); | |||
2489 | ||||
2490 | /// Return a vector value that contains \arg V broadcasted to \p | |||
2491 | /// NumElts elements. | |||
2492 | Value *CreateVectorSplat(unsigned NumElts, Value *V, const Twine &Name = ""); | |||
2493 | ||||
2494 | /// Return a vector value that contains \arg V broadcasted to \p | |||
2495 | /// EC elements. | |||
2496 | Value *CreateVectorSplat(ElementCount EC, Value *V, const Twine &Name = ""); | |||
2497 | ||||
2498 | /// Return a value that has been extracted from a larger integer type. | |||
2499 | Value *CreateExtractInteger(const DataLayout &DL, Value *From, | |||
2500 | IntegerType *ExtractedTy, uint64_t Offset, | |||
2501 | const Twine &Name); | |||
2502 | ||||
2503 | Value *CreatePreserveArrayAccessIndex(Type *ElTy, Value *Base, | |||
2504 | unsigned Dimension, unsigned LastIndex, | |||
2505 | MDNode *DbgInfo); | |||
2506 | ||||
2507 | Value *CreatePreserveUnionAccessIndex(Value *Base, unsigned FieldIndex, | |||
2508 | MDNode *DbgInfo); | |||
2509 | ||||
2510 | Value *CreatePreserveStructAccessIndex(Type *ElTy, Value *Base, | |||
2511 | unsigned Index, unsigned FieldIndex, | |||
2512 | MDNode *DbgInfo); | |||
2513 | ||||
2514 | private: | |||
2515 | /// Helper function that creates an assume intrinsic call that | |||
2516 | /// represents an alignment assumption on the provided Ptr, Mask, Type | |||
2517 | /// and Offset. It may be sometimes useful to do some other logic | |||
2518 | /// based on this alignment check, thus it can be stored into 'TheCheck'. | |||
2519 | CallInst *CreateAlignmentAssumptionHelper(const DataLayout &DL, | |||
2520 | Value *PtrValue, Value *Mask, | |||
2521 | Type *IntPtrTy, Value *OffsetValue, | |||
2522 | Value **TheCheck); | |||
2523 | ||||
2524 | public: | |||
2525 | /// Create an assume intrinsic call that represents an alignment | |||
2526 | /// assumption on the provided pointer. | |||
2527 | /// | |||
2528 | /// An optional offset can be provided, and if it is provided, the offset | |||
2529 | /// must be subtracted from the provided pointer to get the pointer with the | |||
2530 | /// specified alignment. | |||
2531 | /// | |||
2532 | /// It may be sometimes useful to do some other logic | |||
2533 | /// based on this alignment check, thus it can be stored into 'TheCheck'. | |||
2534 | CallInst *CreateAlignmentAssumption(const DataLayout &DL, Value *PtrValue, | |||
2535 | unsigned Alignment, | |||
2536 | Value *OffsetValue = nullptr, | |||
2537 | Value **TheCheck = nullptr); | |||
2538 | ||||
2539 | /// Create an assume intrinsic call that represents an alignment | |||
2540 | /// assumption on the provided pointer. | |||
2541 | /// | |||
2542 | /// An optional offset can be provided, and if it is provided, the offset | |||
2543 | /// must be subtracted from the provided pointer to get the pointer with the | |||
2544 | /// specified alignment. | |||
2545 | /// | |||
2546 | /// It may be sometimes useful to do some other logic | |||
2547 | /// based on this alignment check, thus it can be stored into 'TheCheck'. | |||
2548 | /// | |||
2549 | /// This overload handles the condition where the Alignment is dependent | |||
2550 | /// on an existing value rather than a static value. | |||
2551 | CallInst *CreateAlignmentAssumption(const DataLayout &DL, Value *PtrValue, | |||
2552 | Value *Alignment, | |||
2553 | Value *OffsetValue = nullptr, | |||
2554 | Value **TheCheck = nullptr); | |||
2555 | }; | |||
2556 | ||||
2557 | /// This provides a uniform API for creating instructions and inserting | |||
2558 | /// them into a basic block: either at the end of a BasicBlock, or at a specific | |||
2559 | /// iterator location in a block. | |||
2560 | /// | |||
2561 | /// Note that the builder does not expose the full generality of LLVM | |||
2562 | /// instructions. For access to extra instruction properties, use the mutators | |||
2563 | /// (e.g. setVolatile) on the instructions after they have been | |||
2564 | /// created. Convenience state exists to specify fast-math flags and fp-math | |||
2565 | /// tags. | |||
2566 | /// | |||
2567 | /// The first template argument specifies a class to use for creating constants. | |||
2568 | /// This defaults to creating minimally folded constants. The second template | |||
2569 | /// argument allows clients to specify custom insertion hooks that are called on | |||
2570 | /// every newly created insertion. | |||
2571 | template <typename FolderTy = ConstantFolder, | |||
2572 | typename InserterTy = IRBuilderDefaultInserter> | |||
2573 | class IRBuilder : public IRBuilderBase { | |||
2574 | private: | |||
2575 | FolderTy Folder; | |||
2576 | InserterTy Inserter; | |||
2577 | ||||
2578 | public: | |||
2579 | IRBuilder(LLVMContext &C, FolderTy Folder, InserterTy Inserter = InserterTy(), | |||
2580 | MDNode *FPMathTag = nullptr, | |||
2581 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2582 | : IRBuilderBase(C, this->Folder, this->Inserter, FPMathTag, OpBundles), | |||
2583 | Folder(Folder), Inserter(Inserter) {} | |||
2584 | ||||
2585 | explicit IRBuilder(LLVMContext &C, MDNode *FPMathTag = nullptr, | |||
2586 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2587 | : IRBuilderBase(C, this->Folder, this->Inserter, FPMathTag, OpBundles) {} | |||
2588 | ||||
2589 | explicit IRBuilder(BasicBlock *TheBB, FolderTy Folder, | |||
2590 | MDNode *FPMathTag = nullptr, | |||
2591 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2592 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, | |||
2593 | FPMathTag, OpBundles), Folder(Folder) { | |||
2594 | SetInsertPoint(TheBB); | |||
2595 | } | |||
2596 | ||||
2597 | explicit IRBuilder(BasicBlock *TheBB, MDNode *FPMathTag = nullptr, | |||
2598 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2599 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, | |||
2600 | FPMathTag, OpBundles) { | |||
2601 | SetInsertPoint(TheBB); | |||
2602 | } | |||
2603 | ||||
2604 | explicit IRBuilder(Instruction *IP, MDNode *FPMathTag = nullptr, | |||
2605 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2606 | : IRBuilderBase(IP->getContext(), this->Folder, this->Inserter, | |||
2607 | FPMathTag, OpBundles) { | |||
2608 | SetInsertPoint(IP); | |||
2609 | } | |||
2610 | ||||
2611 | IRBuilder(BasicBlock *TheBB, BasicBlock::iterator IP, FolderTy Folder, | |||
2612 | MDNode *FPMathTag = nullptr, | |||
2613 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2614 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, | |||
2615 | FPMathTag, OpBundles), Folder(Folder) { | |||
2616 | SetInsertPoint(TheBB, IP); | |||
2617 | } | |||
2618 | ||||
2619 | IRBuilder(BasicBlock *TheBB, BasicBlock::iterator IP, | |||
2620 | MDNode *FPMathTag = nullptr, | |||
2621 | ArrayRef<OperandBundleDef> OpBundles = None) | |||
2622 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, | |||
2623 | FPMathTag, OpBundles) { | |||
2624 | SetInsertPoint(TheBB, IP); | |||
2625 | } | |||
2626 | ||||
2627 | /// Avoid copying the full IRBuilder. Prefer using InsertPointGuard | |||
2628 | /// or FastMathFlagGuard instead. | |||
2629 | IRBuilder(const IRBuilder &) = delete; | |||
2630 | ||||
2631 | InserterTy &getInserter() { return Inserter; } | |||
2632 | }; | |||
2633 | ||||
2634 | // Create wrappers for C Binding types (see CBindingWrapping.h). | |||
2635 | DEFINE_SIMPLE_CONVERSION_FUNCTIONS(IRBuilder<>, LLVMBuilderRef)inline IRBuilder<> *unwrap(LLVMBuilderRef P) { return reinterpret_cast <IRBuilder<>*>(P); } inline LLVMBuilderRef wrap(const IRBuilder<> *P) { return reinterpret_cast<LLVMBuilderRef >(const_cast<IRBuilder<>*>(P)); } | |||
2636 | ||||
2637 | } // end namespace llvm | |||
2638 | ||||
2639 | #endif // LLVM_IR_IRBUILDER_H |