1 //===----- CGOpenMPRuntime.cpp - Interface to OpenMP Runtimes -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This provides a class for OpenMP runtime code generation.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "CGOpenMPRuntime.h"
14 #include "CGCXXABI.h"
15 #include "CGCleanup.h"
16 #include "CGRecordLayout.h"
17 #include "CodeGenFunction.h"
18 #include "TargetInfo.h"
19 #include "clang/AST/APValue.h"
20 #include "clang/AST/Attr.h"
21 #include "clang/AST/Decl.h"
22 #include "clang/AST/OpenMPClause.h"
23 #include "clang/AST/StmtOpenMP.h"
24 #include "clang/AST/StmtVisitor.h"
25 #include "clang/Basic/BitmaskEnum.h"
26 #include "clang/Basic/FileManager.h"
27 #include "clang/Basic/OpenMPKinds.h"
28 #include "clang/Basic/SourceManager.h"
29 #include "clang/CodeGen/ConstantInitBuilder.h"
30 #include "llvm/ADT/ArrayRef.h"
31 #include "llvm/ADT/SetOperations.h"
32 #include "llvm/ADT/SmallBitVector.h"
33 #include "llvm/ADT/StringExtras.h"
34 #include "llvm/Bitcode/BitcodeReader.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/GlobalValue.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Value.h"
40 #include "llvm/Support/AtomicOrdering.h"
41 #include "llvm/Support/Format.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include <cassert>
44 #include <numeric>
45 #include <optional>
46
47 using namespace clang;
48 using namespace CodeGen;
49 using namespace llvm::omp;
50
51 namespace {
52 /// Base class for handling code generation inside OpenMP regions.
53 class CGOpenMPRegionInfo : public CodeGenFunction::CGCapturedStmtInfo {
54 public:
55 /// Kinds of OpenMP regions used in codegen.
56 enum CGOpenMPRegionKind {
57 /// Region with outlined function for standalone 'parallel'
58 /// directive.
59 ParallelOutlinedRegion,
60 /// Region with outlined function for standalone 'task' directive.
61 TaskOutlinedRegion,
62 /// Region for constructs that do not require function outlining,
63 /// like 'for', 'sections', 'atomic' etc. directives.
64 InlinedRegion,
65 /// Region with outlined function for standalone 'target' directive.
66 TargetRegion,
67 };
68
CGOpenMPRegionInfo(const CapturedStmt & CS,const CGOpenMPRegionKind RegionKind,const RegionCodeGenTy & CodeGen,OpenMPDirectiveKind Kind,bool HasCancel)69 CGOpenMPRegionInfo(const CapturedStmt &CS,
70 const CGOpenMPRegionKind RegionKind,
71 const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind,
72 bool HasCancel)
73 : CGCapturedStmtInfo(CS, CR_OpenMP), RegionKind(RegionKind),
74 CodeGen(CodeGen), Kind(Kind), HasCancel(HasCancel) {}
75
CGOpenMPRegionInfo(const CGOpenMPRegionKind RegionKind,const RegionCodeGenTy & CodeGen,OpenMPDirectiveKind Kind,bool HasCancel)76 CGOpenMPRegionInfo(const CGOpenMPRegionKind RegionKind,
77 const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind,
78 bool HasCancel)
79 : CGCapturedStmtInfo(CR_OpenMP), RegionKind(RegionKind), CodeGen(CodeGen),
80 Kind(Kind), HasCancel(HasCancel) {}
81
82 /// Get a variable or parameter for storing global thread id
83 /// inside OpenMP construct.
84 virtual const VarDecl *getThreadIDVariable() const = 0;
85
86 /// Emit the captured statement body.
87 void EmitBody(CodeGenFunction &CGF, const Stmt *S) override;
88
89 /// Get an LValue for the current ThreadID variable.
90 /// \return LValue for thread id variable. This LValue always has type int32*.
91 virtual LValue getThreadIDVariableLValue(CodeGenFunction &CGF);
92
emitUntiedSwitch(CodeGenFunction &)93 virtual void emitUntiedSwitch(CodeGenFunction & /*CGF*/) {}
94
getRegionKind() const95 CGOpenMPRegionKind getRegionKind() const { return RegionKind; }
96
getDirectiveKind() const97 OpenMPDirectiveKind getDirectiveKind() const { return Kind; }
98
hasCancel() const99 bool hasCancel() const { return HasCancel; }
100
classof(const CGCapturedStmtInfo * Info)101 static bool classof(const CGCapturedStmtInfo *Info) {
102 return Info->getKind() == CR_OpenMP;
103 }
104
105 ~CGOpenMPRegionInfo() override = default;
106
107 protected:
108 CGOpenMPRegionKind RegionKind;
109 RegionCodeGenTy CodeGen;
110 OpenMPDirectiveKind Kind;
111 bool HasCancel;
112 };
113
114 /// API for captured statement code generation in OpenMP constructs.
115 class CGOpenMPOutlinedRegionInfo final : public CGOpenMPRegionInfo {
116 public:
CGOpenMPOutlinedRegionInfo(const CapturedStmt & CS,const VarDecl * ThreadIDVar,const RegionCodeGenTy & CodeGen,OpenMPDirectiveKind Kind,bool HasCancel,StringRef HelperName)117 CGOpenMPOutlinedRegionInfo(const CapturedStmt &CS, const VarDecl *ThreadIDVar,
118 const RegionCodeGenTy &CodeGen,
119 OpenMPDirectiveKind Kind, bool HasCancel,
120 StringRef HelperName)
121 : CGOpenMPRegionInfo(CS, ParallelOutlinedRegion, CodeGen, Kind,
122 HasCancel),
123 ThreadIDVar(ThreadIDVar), HelperName(HelperName) {
124 assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.");
125 }
126
127 /// Get a variable or parameter for storing global thread id
128 /// inside OpenMP construct.
getThreadIDVariable() const129 const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; }
130
131 /// Get the name of the capture helper.
getHelperName() const132 StringRef getHelperName() const override { return HelperName; }
133
classof(const CGCapturedStmtInfo * Info)134 static bool classof(const CGCapturedStmtInfo *Info) {
135 return CGOpenMPRegionInfo::classof(Info) &&
136 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() ==
137 ParallelOutlinedRegion;
138 }
139
140 private:
141 /// A variable or parameter storing global thread id for OpenMP
142 /// constructs.
143 const VarDecl *ThreadIDVar;
144 StringRef HelperName;
145 };
146
147 /// API for captured statement code generation in OpenMP constructs.
148 class CGOpenMPTaskOutlinedRegionInfo final : public CGOpenMPRegionInfo {
149 public:
150 class UntiedTaskActionTy final : public PrePostActionTy {
151 bool Untied;
152 const VarDecl *PartIDVar;
153 const RegionCodeGenTy UntiedCodeGen;
154 llvm::SwitchInst *UntiedSwitch = nullptr;
155
156 public:
UntiedTaskActionTy(bool Tied,const VarDecl * PartIDVar,const RegionCodeGenTy & UntiedCodeGen)157 UntiedTaskActionTy(bool Tied, const VarDecl *PartIDVar,
158 const RegionCodeGenTy &UntiedCodeGen)
159 : Untied(!Tied), PartIDVar(PartIDVar), UntiedCodeGen(UntiedCodeGen) {}
Enter(CodeGenFunction & CGF)160 void Enter(CodeGenFunction &CGF) override {
161 if (Untied) {
162 // Emit task switching point.
163 LValue PartIdLVal = CGF.EmitLoadOfPointerLValue(
164 CGF.GetAddrOfLocalVar(PartIDVar),
165 PartIDVar->getType()->castAs<PointerType>());
166 llvm::Value *Res =
167 CGF.EmitLoadOfScalar(PartIdLVal, PartIDVar->getLocation());
168 llvm::BasicBlock *DoneBB = CGF.createBasicBlock(".untied.done.");
169 UntiedSwitch = CGF.Builder.CreateSwitch(Res, DoneBB);
170 CGF.EmitBlock(DoneBB);
171 CGF.EmitBranchThroughCleanup(CGF.ReturnBlock);
172 CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp."));
173 UntiedSwitch->addCase(CGF.Builder.getInt32(0),
174 CGF.Builder.GetInsertBlock());
175 emitUntiedSwitch(CGF);
176 }
177 }
emitUntiedSwitch(CodeGenFunction & CGF) const178 void emitUntiedSwitch(CodeGenFunction &CGF) const {
179 if (Untied) {
180 LValue PartIdLVal = CGF.EmitLoadOfPointerLValue(
181 CGF.GetAddrOfLocalVar(PartIDVar),
182 PartIDVar->getType()->castAs<PointerType>());
183 CGF.EmitStoreOfScalar(CGF.Builder.getInt32(UntiedSwitch->getNumCases()),
184 PartIdLVal);
185 UntiedCodeGen(CGF);
186 CodeGenFunction::JumpDest CurPoint =
187 CGF.getJumpDestInCurrentScope(".untied.next.");
188 CGF.EmitBranch(CGF.ReturnBlock.getBlock());
189 CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp."));
190 UntiedSwitch->addCase(CGF.Builder.getInt32(UntiedSwitch->getNumCases()),
191 CGF.Builder.GetInsertBlock());
192 CGF.EmitBranchThroughCleanup(CurPoint);
193 CGF.EmitBlock(CurPoint.getBlock());
194 }
195 }
getNumberOfParts() const196 unsigned getNumberOfParts() const { return UntiedSwitch->getNumCases(); }
197 };
CGOpenMPTaskOutlinedRegionInfo(const CapturedStmt & CS,const VarDecl * ThreadIDVar,const RegionCodeGenTy & CodeGen,OpenMPDirectiveKind Kind,bool HasCancel,const UntiedTaskActionTy & Action)198 CGOpenMPTaskOutlinedRegionInfo(const CapturedStmt &CS,
199 const VarDecl *ThreadIDVar,
200 const RegionCodeGenTy &CodeGen,
201 OpenMPDirectiveKind Kind, bool HasCancel,
202 const UntiedTaskActionTy &Action)
203 : CGOpenMPRegionInfo(CS, TaskOutlinedRegion, CodeGen, Kind, HasCancel),
204 ThreadIDVar(ThreadIDVar), Action(Action) {
205 assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.");
206 }
207
208 /// Get a variable or parameter for storing global thread id
209 /// inside OpenMP construct.
getThreadIDVariable() const210 const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; }
211
212 /// Get an LValue for the current ThreadID variable.
213 LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override;
214
215 /// Get the name of the capture helper.
getHelperName() const216 StringRef getHelperName() const override { return ".omp_outlined."; }
217
emitUntiedSwitch(CodeGenFunction & CGF)218 void emitUntiedSwitch(CodeGenFunction &CGF) override {
219 Action.emitUntiedSwitch(CGF);
220 }
221
classof(const CGCapturedStmtInfo * Info)222 static bool classof(const CGCapturedStmtInfo *Info) {
223 return CGOpenMPRegionInfo::classof(Info) &&
224 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() ==
225 TaskOutlinedRegion;
226 }
227
228 private:
229 /// A variable or parameter storing global thread id for OpenMP
230 /// constructs.
231 const VarDecl *ThreadIDVar;
232 /// Action for emitting code for untied tasks.
233 const UntiedTaskActionTy &Action;
234 };
235
236 /// API for inlined captured statement code generation in OpenMP
237 /// constructs.
238 class CGOpenMPInlinedRegionInfo : public CGOpenMPRegionInfo {
239 public:
CGOpenMPInlinedRegionInfo(CodeGenFunction::CGCapturedStmtInfo * OldCSI,const RegionCodeGenTy & CodeGen,OpenMPDirectiveKind Kind,bool HasCancel)240 CGOpenMPInlinedRegionInfo(CodeGenFunction::CGCapturedStmtInfo *OldCSI,
241 const RegionCodeGenTy &CodeGen,
242 OpenMPDirectiveKind Kind, bool HasCancel)
243 : CGOpenMPRegionInfo(InlinedRegion, CodeGen, Kind, HasCancel),
244 OldCSI(OldCSI),
245 OuterRegionInfo(dyn_cast_or_null<CGOpenMPRegionInfo>(OldCSI)) {}
246
247 // Retrieve the value of the context parameter.
getContextValue() const248 llvm::Value *getContextValue() const override {
249 if (OuterRegionInfo)
250 return OuterRegionInfo->getContextValue();
251 llvm_unreachable("No context value for inlined OpenMP region");
252 }
253
setContextValue(llvm::Value * V)254 void setContextValue(llvm::Value *V) override {
255 if (OuterRegionInfo) {
256 OuterRegionInfo->setContextValue(V);
257 return;
258 }
259 llvm_unreachable("No context value for inlined OpenMP region");
260 }
261
262 /// Lookup the captured field decl for a variable.
lookup(const VarDecl * VD) const263 const FieldDecl *lookup(const VarDecl *VD) const override {
264 if (OuterRegionInfo)
265 return OuterRegionInfo->lookup(VD);
266 // If there is no outer outlined region,no need to lookup in a list of
267 // captured variables, we can use the original one.
268 return nullptr;
269 }
270
getThisFieldDecl() const271 FieldDecl *getThisFieldDecl() const override {
272 if (OuterRegionInfo)
273 return OuterRegionInfo->getThisFieldDecl();
274 return nullptr;
275 }
276
277 /// Get a variable or parameter for storing global thread id
278 /// inside OpenMP construct.
getThreadIDVariable() const279 const VarDecl *getThreadIDVariable() const override {
280 if (OuterRegionInfo)
281 return OuterRegionInfo->getThreadIDVariable();
282 return nullptr;
283 }
284
285 /// Get an LValue for the current ThreadID variable.
getThreadIDVariableLValue(CodeGenFunction & CGF)286 LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override {
287 if (OuterRegionInfo)
288 return OuterRegionInfo->getThreadIDVariableLValue(CGF);
289 llvm_unreachable("No LValue for inlined OpenMP construct");
290 }
291
292 /// Get the name of the capture helper.
getHelperName() const293 StringRef getHelperName() const override {
294 if (auto *OuterRegionInfo = getOldCSI())
295 return OuterRegionInfo->getHelperName();
296 llvm_unreachable("No helper name for inlined OpenMP construct");
297 }
298
emitUntiedSwitch(CodeGenFunction & CGF)299 void emitUntiedSwitch(CodeGenFunction &CGF) override {
300 if (OuterRegionInfo)
301 OuterRegionInfo->emitUntiedSwitch(CGF);
302 }
303
getOldCSI() const304 CodeGenFunction::CGCapturedStmtInfo *getOldCSI() const { return OldCSI; }
305
classof(const CGCapturedStmtInfo * Info)306 static bool classof(const CGCapturedStmtInfo *Info) {
307 return CGOpenMPRegionInfo::classof(Info) &&
308 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == InlinedRegion;
309 }
310
311 ~CGOpenMPInlinedRegionInfo() override = default;
312
313 private:
314 /// CodeGen info about outer OpenMP region.
315 CodeGenFunction::CGCapturedStmtInfo *OldCSI;
316 CGOpenMPRegionInfo *OuterRegionInfo;
317 };
318
319 /// API for captured statement code generation in OpenMP target
320 /// constructs. For this captures, implicit parameters are used instead of the
321 /// captured fields. The name of the target region has to be unique in a given
322 /// application so it is provided by the client, because only the client has
323 /// the information to generate that.
324 class CGOpenMPTargetRegionInfo final : public CGOpenMPRegionInfo {
325 public:
CGOpenMPTargetRegionInfo(const CapturedStmt & CS,const RegionCodeGenTy & CodeGen,StringRef HelperName)326 CGOpenMPTargetRegionInfo(const CapturedStmt &CS,
327 const RegionCodeGenTy &CodeGen, StringRef HelperName)
328 : CGOpenMPRegionInfo(CS, TargetRegion, CodeGen, OMPD_target,
329 /*HasCancel=*/false),
330 HelperName(HelperName) {}
331
332 /// This is unused for target regions because each starts executing
333 /// with a single thread.
getThreadIDVariable() const334 const VarDecl *getThreadIDVariable() const override { return nullptr; }
335
336 /// Get the name of the capture helper.
getHelperName() const337 StringRef getHelperName() const override { return HelperName; }
338
classof(const CGCapturedStmtInfo * Info)339 static bool classof(const CGCapturedStmtInfo *Info) {
340 return CGOpenMPRegionInfo::classof(Info) &&
341 cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == TargetRegion;
342 }
343
344 private:
345 StringRef HelperName;
346 };
347
EmptyCodeGen(CodeGenFunction &,PrePostActionTy &)348 static void EmptyCodeGen(CodeGenFunction &, PrePostActionTy &) {
349 llvm_unreachable("No codegen for expressions");
350 }
351 /// API for generation of expressions captured in a innermost OpenMP
352 /// region.
353 class CGOpenMPInnerExprInfo final : public CGOpenMPInlinedRegionInfo {
354 public:
CGOpenMPInnerExprInfo(CodeGenFunction & CGF,const CapturedStmt & CS)355 CGOpenMPInnerExprInfo(CodeGenFunction &CGF, const CapturedStmt &CS)
356 : CGOpenMPInlinedRegionInfo(CGF.CapturedStmtInfo, EmptyCodeGen,
357 OMPD_unknown,
358 /*HasCancel=*/false),
359 PrivScope(CGF) {
360 // Make sure the globals captured in the provided statement are local by
361 // using the privatization logic. We assume the same variable is not
362 // captured more than once.
363 for (const auto &C : CS.captures()) {
364 if (!C.capturesVariable() && !C.capturesVariableByCopy())
365 continue;
366
367 const VarDecl *VD = C.getCapturedVar();
368 if (VD->isLocalVarDeclOrParm())
369 continue;
370
371 DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD),
372 /*RefersToEnclosingVariableOrCapture=*/false,
373 VD->getType().getNonReferenceType(), VK_LValue,
374 C.getLocation());
375 PrivScope.addPrivate(VD, CGF.EmitLValue(&DRE).getAddress(CGF));
376 }
377 (void)PrivScope.Privatize();
378 }
379
380 /// Lookup the captured field decl for a variable.
lookup(const VarDecl * VD) const381 const FieldDecl *lookup(const VarDecl *VD) const override {
382 if (const FieldDecl *FD = CGOpenMPInlinedRegionInfo::lookup(VD))
383 return FD;
384 return nullptr;
385 }
386
387 /// Emit the captured statement body.
EmitBody(CodeGenFunction & CGF,const Stmt * S)388 void EmitBody(CodeGenFunction &CGF, const Stmt *S) override {
389 llvm_unreachable("No body for expressions");
390 }
391
392 /// Get a variable or parameter for storing global thread id
393 /// inside OpenMP construct.
getThreadIDVariable() const394 const VarDecl *getThreadIDVariable() const override {
395 llvm_unreachable("No thread id for expressions");
396 }
397
398 /// Get the name of the capture helper.
getHelperName() const399 StringRef getHelperName() const override {
400 llvm_unreachable("No helper name for expressions");
401 }
402
classof(const CGCapturedStmtInfo * Info)403 static bool classof(const CGCapturedStmtInfo *Info) { return false; }
404
405 private:
406 /// Private scope to capture global variables.
407 CodeGenFunction::OMPPrivateScope PrivScope;
408 };
409
410 /// RAII for emitting code of OpenMP constructs.
411 class InlinedOpenMPRegionRAII {
412 CodeGenFunction &CGF;
413 llvm::DenseMap<const ValueDecl *, FieldDecl *> LambdaCaptureFields;
414 FieldDecl *LambdaThisCaptureField = nullptr;
415 const CodeGen::CGBlockInfo *BlockInfo = nullptr;
416 bool NoInheritance = false;
417
418 public:
419 /// Constructs region for combined constructs.
420 /// \param CodeGen Code generation sequence for combined directives. Includes
421 /// a list of functions used for code generation of implicitly inlined
422 /// regions.
InlinedOpenMPRegionRAII(CodeGenFunction & CGF,const RegionCodeGenTy & CodeGen,OpenMPDirectiveKind Kind,bool HasCancel,bool NoInheritance=true)423 InlinedOpenMPRegionRAII(CodeGenFunction &CGF, const RegionCodeGenTy &CodeGen,
424 OpenMPDirectiveKind Kind, bool HasCancel,
425 bool NoInheritance = true)
426 : CGF(CGF), NoInheritance(NoInheritance) {
427 // Start emission for the construct.
428 CGF.CapturedStmtInfo = new CGOpenMPInlinedRegionInfo(
429 CGF.CapturedStmtInfo, CodeGen, Kind, HasCancel);
430 if (NoInheritance) {
431 std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields);
432 LambdaThisCaptureField = CGF.LambdaThisCaptureField;
433 CGF.LambdaThisCaptureField = nullptr;
434 BlockInfo = CGF.BlockInfo;
435 CGF.BlockInfo = nullptr;
436 }
437 }
438
~InlinedOpenMPRegionRAII()439 ~InlinedOpenMPRegionRAII() {
440 // Restore original CapturedStmtInfo only if we're done with code emission.
441 auto *OldCSI =
442 cast<CGOpenMPInlinedRegionInfo>(CGF.CapturedStmtInfo)->getOldCSI();
443 delete CGF.CapturedStmtInfo;
444 CGF.CapturedStmtInfo = OldCSI;
445 if (NoInheritance) {
446 std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields);
447 CGF.LambdaThisCaptureField = LambdaThisCaptureField;
448 CGF.BlockInfo = BlockInfo;
449 }
450 }
451 };
452
453 /// Values for bit flags used in the ident_t to describe the fields.
454 /// All enumeric elements are named and described in accordance with the code
455 /// from https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp.h
456 enum OpenMPLocationFlags : unsigned {
457 /// Use trampoline for internal microtask.
458 OMP_IDENT_IMD = 0x01,
459 /// Use c-style ident structure.
460 OMP_IDENT_KMPC = 0x02,
461 /// Atomic reduction option for kmpc_reduce.
462 OMP_ATOMIC_REDUCE = 0x10,
463 /// Explicit 'barrier' directive.
464 OMP_IDENT_BARRIER_EXPL = 0x20,
465 /// Implicit barrier in code.
466 OMP_IDENT_BARRIER_IMPL = 0x40,
467 /// Implicit barrier in 'for' directive.
468 OMP_IDENT_BARRIER_IMPL_FOR = 0x40,
469 /// Implicit barrier in 'sections' directive.
470 OMP_IDENT_BARRIER_IMPL_SECTIONS = 0xC0,
471 /// Implicit barrier in 'single' directive.
472 OMP_IDENT_BARRIER_IMPL_SINGLE = 0x140,
473 /// Call of __kmp_for_static_init for static loop.
474 OMP_IDENT_WORK_LOOP = 0x200,
475 /// Call of __kmp_for_static_init for sections.
476 OMP_IDENT_WORK_SECTIONS = 0x400,
477 /// Call of __kmp_for_static_init for distribute.
478 OMP_IDENT_WORK_DISTRIBUTE = 0x800,
479 LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_IDENT_WORK_DISTRIBUTE)
480 };
481
482 namespace {
483 LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE();
484 /// Values for bit flags for marking which requires clauses have been used.
485 enum OpenMPOffloadingRequiresDirFlags : int64_t {
486 /// flag undefined.
487 OMP_REQ_UNDEFINED = 0x000,
488 /// no requires clause present.
489 OMP_REQ_NONE = 0x001,
490 /// reverse_offload clause.
491 OMP_REQ_REVERSE_OFFLOAD = 0x002,
492 /// unified_address clause.
493 OMP_REQ_UNIFIED_ADDRESS = 0x004,
494 /// unified_shared_memory clause.
495 OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008,
496 /// dynamic_allocators clause.
497 OMP_REQ_DYNAMIC_ALLOCATORS = 0x010,
498 LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_REQ_DYNAMIC_ALLOCATORS)
499 };
500
501 enum OpenMPOffloadingReservedDeviceIDs {
502 /// Device ID if the device was not defined, runtime should get it
503 /// from environment variables in the spec.
504 OMP_DEVICEID_UNDEF = -1,
505 };
506 } // anonymous namespace
507
508 /// Describes ident structure that describes a source location.
509 /// All descriptions are taken from
510 /// https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp.h
511 /// Original structure:
512 /// typedef struct ident {
513 /// kmp_int32 reserved_1; /**< might be used in Fortran;
514 /// see above */
515 /// kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags;
516 /// KMP_IDENT_KMPC identifies this union
517 /// member */
518 /// kmp_int32 reserved_2; /**< not really used in Fortran any more;
519 /// see above */
520 ///#if USE_ITT_BUILD
521 /// /* but currently used for storing
522 /// region-specific ITT */
523 /// /* contextual information. */
524 ///#endif /* USE_ITT_BUILD */
525 /// kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for
526 /// C++ */
527 /// char const *psource; /**< String describing the source location.
528 /// The string is composed of semi-colon separated
529 // fields which describe the source file,
530 /// the function and a pair of line numbers that
531 /// delimit the construct.
532 /// */
533 /// } ident_t;
534 enum IdentFieldIndex {
535 /// might be used in Fortran
536 IdentField_Reserved_1,
537 /// OMP_IDENT_xxx flags; OMP_IDENT_KMPC identifies this union member.
538 IdentField_Flags,
539 /// Not really used in Fortran any more
540 IdentField_Reserved_2,
541 /// Source[4] in Fortran, do not use for C++
542 IdentField_Reserved_3,
543 /// String describing the source location. The string is composed of
544 /// semi-colon separated fields which describe the source file, the function
545 /// and a pair of line numbers that delimit the construct.
546 IdentField_PSource
547 };
548
549 /// Schedule types for 'omp for' loops (these enumerators are taken from
550 /// the enum sched_type in kmp.h).
551 enum OpenMPSchedType {
552 /// Lower bound for default (unordered) versions.
553 OMP_sch_lower = 32,
554 OMP_sch_static_chunked = 33,
555 OMP_sch_static = 34,
556 OMP_sch_dynamic_chunked = 35,
557 OMP_sch_guided_chunked = 36,
558 OMP_sch_runtime = 37,
559 OMP_sch_auto = 38,
560 /// static with chunk adjustment (e.g., simd)
561 OMP_sch_static_balanced_chunked = 45,
562 /// Lower bound for 'ordered' versions.
563 OMP_ord_lower = 64,
564 OMP_ord_static_chunked = 65,
565 OMP_ord_static = 66,
566 OMP_ord_dynamic_chunked = 67,
567 OMP_ord_guided_chunked = 68,
568 OMP_ord_runtime = 69,
569 OMP_ord_auto = 70,
570 OMP_sch_default = OMP_sch_static,
571 /// dist_schedule types
572 OMP_dist_sch_static_chunked = 91,
573 OMP_dist_sch_static = 92,
574 /// Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers.
575 /// Set if the monotonic schedule modifier was present.
576 OMP_sch_modifier_monotonic = (1 << 29),
577 /// Set if the nonmonotonic schedule modifier was present.
578 OMP_sch_modifier_nonmonotonic = (1 << 30),
579 };
580
581 /// A basic class for pre|post-action for advanced codegen sequence for OpenMP
582 /// region.
583 class CleanupTy final : public EHScopeStack::Cleanup {
584 PrePostActionTy *Action;
585
586 public:
CleanupTy(PrePostActionTy * Action)587 explicit CleanupTy(PrePostActionTy *Action) : Action(Action) {}
Emit(CodeGenFunction & CGF,Flags)588 void Emit(CodeGenFunction &CGF, Flags /*flags*/) override {
589 if (!CGF.HaveInsertPoint())
590 return;
591 Action->Exit(CGF);
592 }
593 };
594
595 } // anonymous namespace
596
operator ()(CodeGenFunction & CGF) const597 void RegionCodeGenTy::operator()(CodeGenFunction &CGF) const {
598 CodeGenFunction::RunCleanupsScope Scope(CGF);
599 if (PrePostAction) {
600 CGF.EHStack.pushCleanup<CleanupTy>(NormalAndEHCleanup, PrePostAction);
601 Callback(CodeGen, CGF, *PrePostAction);
602 } else {
603 PrePostActionTy Action;
604 Callback(CodeGen, CGF, Action);
605 }
606 }
607
608 /// Check if the combiner is a call to UDR combiner and if it is so return the
609 /// UDR decl used for reduction.
610 static const OMPDeclareReductionDecl *
getReductionInit(const Expr * ReductionOp)611 getReductionInit(const Expr *ReductionOp) {
612 if (const auto *CE = dyn_cast<CallExpr>(ReductionOp))
613 if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee()))
614 if (const auto *DRE =
615 dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts()))
616 if (const auto *DRD = dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl()))
617 return DRD;
618 return nullptr;
619 }
620
emitInitWithReductionInitializer(CodeGenFunction & CGF,const OMPDeclareReductionDecl * DRD,const Expr * InitOp,Address Private,Address Original,QualType Ty)621 static void emitInitWithReductionInitializer(CodeGenFunction &CGF,
622 const OMPDeclareReductionDecl *DRD,
623 const Expr *InitOp,
624 Address Private, Address Original,
625 QualType Ty) {
626 if (DRD->getInitializer()) {
627 std::pair<llvm::Function *, llvm::Function *> Reduction =
628 CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD);
629 const auto *CE = cast<CallExpr>(InitOp);
630 const auto *OVE = cast<OpaqueValueExpr>(CE->getCallee());
631 const Expr *LHS = CE->getArg(/*Arg=*/0)->IgnoreParenImpCasts();
632 const Expr *RHS = CE->getArg(/*Arg=*/1)->IgnoreParenImpCasts();
633 const auto *LHSDRE =
634 cast<DeclRefExpr>(cast<UnaryOperator>(LHS)->getSubExpr());
635 const auto *RHSDRE =
636 cast<DeclRefExpr>(cast<UnaryOperator>(RHS)->getSubExpr());
637 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
638 PrivateScope.addPrivate(cast<VarDecl>(LHSDRE->getDecl()), Private);
639 PrivateScope.addPrivate(cast<VarDecl>(RHSDRE->getDecl()), Original);
640 (void)PrivateScope.Privatize();
641 RValue Func = RValue::get(Reduction.second);
642 CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func);
643 CGF.EmitIgnoredExpr(InitOp);
644 } else {
645 llvm::Constant *Init = CGF.CGM.EmitNullConstant(Ty);
646 std::string Name = CGF.CGM.getOpenMPRuntime().getName({"init"});
647 auto *GV = new llvm::GlobalVariable(
648 CGF.CGM.getModule(), Init->getType(), /*isConstant=*/true,
649 llvm::GlobalValue::PrivateLinkage, Init, Name);
650 LValue LV = CGF.MakeNaturalAlignAddrLValue(GV, Ty);
651 RValue InitRVal;
652 switch (CGF.getEvaluationKind(Ty)) {
653 case TEK_Scalar:
654 InitRVal = CGF.EmitLoadOfLValue(LV, DRD->getLocation());
655 break;
656 case TEK_Complex:
657 InitRVal =
658 RValue::getComplex(CGF.EmitLoadOfComplex(LV, DRD->getLocation()));
659 break;
660 case TEK_Aggregate: {
661 OpaqueValueExpr OVE(DRD->getLocation(), Ty, VK_LValue);
662 CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, LV);
663 CGF.EmitAnyExprToMem(&OVE, Private, Ty.getQualifiers(),
664 /*IsInitializer=*/false);
665 return;
666 }
667 }
668 OpaqueValueExpr OVE(DRD->getLocation(), Ty, VK_PRValue);
669 CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, InitRVal);
670 CGF.EmitAnyExprToMem(&OVE, Private, Ty.getQualifiers(),
671 /*IsInitializer=*/false);
672 }
673 }
674
675 /// Emit initialization of arrays of complex types.
676 /// \param DestAddr Address of the array.
677 /// \param Type Type of array.
678 /// \param Init Initial expression of array.
679 /// \param SrcAddr Address of the original array.
EmitOMPAggregateInit(CodeGenFunction & CGF,Address DestAddr,QualType Type,bool EmitDeclareReductionInit,const Expr * Init,const OMPDeclareReductionDecl * DRD,Address SrcAddr=Address::invalid ())680 static void EmitOMPAggregateInit(CodeGenFunction &CGF, Address DestAddr,
681 QualType Type, bool EmitDeclareReductionInit,
682 const Expr *Init,
683 const OMPDeclareReductionDecl *DRD,
684 Address SrcAddr = Address::invalid()) {
685 // Perform element-by-element initialization.
686 QualType ElementTy;
687
688 // Drill down to the base element type on both arrays.
689 const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe();
690 llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, DestAddr);
691 if (DRD)
692 SrcAddr =
693 CGF.Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType());
694
695 llvm::Value *SrcBegin = nullptr;
696 if (DRD)
697 SrcBegin = SrcAddr.getPointer();
698 llvm::Value *DestBegin = DestAddr.getPointer();
699 // Cast from pointer to array type to pointer to single element.
700 llvm::Value *DestEnd =
701 CGF.Builder.CreateGEP(DestAddr.getElementType(), DestBegin, NumElements);
702 // The basic structure here is a while-do loop.
703 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arrayinit.body");
704 llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arrayinit.done");
705 llvm::Value *IsEmpty =
706 CGF.Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arrayinit.isempty");
707 CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
708
709 // Enter the loop body, making that address the current address.
710 llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock();
711 CGF.EmitBlock(BodyBB);
712
713 CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy);
714
715 llvm::PHINode *SrcElementPHI = nullptr;
716 Address SrcElementCurrent = Address::invalid();
717 if (DRD) {
718 SrcElementPHI = CGF.Builder.CreatePHI(SrcBegin->getType(), 2,
719 "omp.arraycpy.srcElementPast");
720 SrcElementPHI->addIncoming(SrcBegin, EntryBB);
721 SrcElementCurrent =
722 Address(SrcElementPHI, SrcAddr.getElementType(),
723 SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize));
724 }
725 llvm::PHINode *DestElementPHI = CGF.Builder.CreatePHI(
726 DestBegin->getType(), 2, "omp.arraycpy.destElementPast");
727 DestElementPHI->addIncoming(DestBegin, EntryBB);
728 Address DestElementCurrent =
729 Address(DestElementPHI, DestAddr.getElementType(),
730 DestAddr.getAlignment().alignmentOfArrayElement(ElementSize));
731
732 // Emit copy.
733 {
734 CodeGenFunction::RunCleanupsScope InitScope(CGF);
735 if (EmitDeclareReductionInit) {
736 emitInitWithReductionInitializer(CGF, DRD, Init, DestElementCurrent,
737 SrcElementCurrent, ElementTy);
738 } else
739 CGF.EmitAnyExprToMem(Init, DestElementCurrent, ElementTy.getQualifiers(),
740 /*IsInitializer=*/false);
741 }
742
743 if (DRD) {
744 // Shift the address forward by one element.
745 llvm::Value *SrcElementNext = CGF.Builder.CreateConstGEP1_32(
746 SrcAddr.getElementType(), SrcElementPHI, /*Idx0=*/1,
747 "omp.arraycpy.dest.element");
748 SrcElementPHI->addIncoming(SrcElementNext, CGF.Builder.GetInsertBlock());
749 }
750
751 // Shift the address forward by one element.
752 llvm::Value *DestElementNext = CGF.Builder.CreateConstGEP1_32(
753 DestAddr.getElementType(), DestElementPHI, /*Idx0=*/1,
754 "omp.arraycpy.dest.element");
755 // Check whether we've reached the end.
756 llvm::Value *Done =
757 CGF.Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done");
758 CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB);
759 DestElementPHI->addIncoming(DestElementNext, CGF.Builder.GetInsertBlock());
760
761 // Done.
762 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
763 }
764
emitSharedLValue(CodeGenFunction & CGF,const Expr * E)765 LValue ReductionCodeGen::emitSharedLValue(CodeGenFunction &CGF, const Expr *E) {
766 return CGF.EmitOMPSharedLValue(E);
767 }
768
emitSharedLValueUB(CodeGenFunction & CGF,const Expr * E)769 LValue ReductionCodeGen::emitSharedLValueUB(CodeGenFunction &CGF,
770 const Expr *E) {
771 if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(E))
772 return CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false);
773 return LValue();
774 }
775
emitAggregateInitialization(CodeGenFunction & CGF,unsigned N,Address PrivateAddr,Address SharedAddr,const OMPDeclareReductionDecl * DRD)776 void ReductionCodeGen::emitAggregateInitialization(
777 CodeGenFunction &CGF, unsigned N, Address PrivateAddr, Address SharedAddr,
778 const OMPDeclareReductionDecl *DRD) {
779 // Emit VarDecl with copy init for arrays.
780 // Get the address of the original variable captured in current
781 // captured region.
782 const auto *PrivateVD =
783 cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl());
784 bool EmitDeclareReductionInit =
785 DRD && (DRD->getInitializer() || !PrivateVD->hasInit());
786 EmitOMPAggregateInit(CGF, PrivateAddr, PrivateVD->getType(),
787 EmitDeclareReductionInit,
788 EmitDeclareReductionInit ? ClausesData[N].ReductionOp
789 : PrivateVD->getInit(),
790 DRD, SharedAddr);
791 }
792
ReductionCodeGen(ArrayRef<const Expr * > Shareds,ArrayRef<const Expr * > Origs,ArrayRef<const Expr * > Privates,ArrayRef<const Expr * > ReductionOps)793 ReductionCodeGen::ReductionCodeGen(ArrayRef<const Expr *> Shareds,
794 ArrayRef<const Expr *> Origs,
795 ArrayRef<const Expr *> Privates,
796 ArrayRef<const Expr *> ReductionOps) {
797 ClausesData.reserve(Shareds.size());
798 SharedAddresses.reserve(Shareds.size());
799 Sizes.reserve(Shareds.size());
800 BaseDecls.reserve(Shareds.size());
801 const auto *IOrig = Origs.begin();
802 const auto *IPriv = Privates.begin();
803 const auto *IRed = ReductionOps.begin();
804 for (const Expr *Ref : Shareds) {
805 ClausesData.emplace_back(Ref, *IOrig, *IPriv, *IRed);
806 std::advance(IOrig, 1);
807 std::advance(IPriv, 1);
808 std::advance(IRed, 1);
809 }
810 }
811
emitSharedOrigLValue(CodeGenFunction & CGF,unsigned N)812 void ReductionCodeGen::emitSharedOrigLValue(CodeGenFunction &CGF, unsigned N) {
813 assert(SharedAddresses.size() == N && OrigAddresses.size() == N &&
814 "Number of generated lvalues must be exactly N.");
815 LValue First = emitSharedLValue(CGF, ClausesData[N].Shared);
816 LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Shared);
817 SharedAddresses.emplace_back(First, Second);
818 if (ClausesData[N].Shared == ClausesData[N].Ref) {
819 OrigAddresses.emplace_back(First, Second);
820 } else {
821 LValue First = emitSharedLValue(CGF, ClausesData[N].Ref);
822 LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Ref);
823 OrigAddresses.emplace_back(First, Second);
824 }
825 }
826
emitAggregateType(CodeGenFunction & CGF,unsigned N)827 void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N) {
828 QualType PrivateType = getPrivateType(N);
829 bool AsArraySection = isa<OMPArraySectionExpr>(ClausesData[N].Ref);
830 if (!PrivateType->isVariablyModifiedType()) {
831 Sizes.emplace_back(
832 CGF.getTypeSize(OrigAddresses[N].first.getType().getNonReferenceType()),
833 nullptr);
834 return;
835 }
836 llvm::Value *Size;
837 llvm::Value *SizeInChars;
838 auto *ElemType = OrigAddresses[N].first.getAddress(CGF).getElementType();
839 auto *ElemSizeOf = llvm::ConstantExpr::getSizeOf(ElemType);
840 if (AsArraySection) {
841 Size = CGF.Builder.CreatePtrDiff(ElemType,
842 OrigAddresses[N].second.getPointer(CGF),
843 OrigAddresses[N].first.getPointer(CGF));
844 Size = CGF.Builder.CreateNUWAdd(
845 Size, llvm::ConstantInt::get(Size->getType(), /*V=*/1));
846 SizeInChars = CGF.Builder.CreateNUWMul(Size, ElemSizeOf);
847 } else {
848 SizeInChars =
849 CGF.getTypeSize(OrigAddresses[N].first.getType().getNonReferenceType());
850 Size = CGF.Builder.CreateExactUDiv(SizeInChars, ElemSizeOf);
851 }
852 Sizes.emplace_back(SizeInChars, Size);
853 CodeGenFunction::OpaqueValueMapping OpaqueMap(
854 CGF,
855 cast<OpaqueValueExpr>(
856 CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()),
857 RValue::get(Size));
858 CGF.EmitVariablyModifiedType(PrivateType);
859 }
860
emitAggregateType(CodeGenFunction & CGF,unsigned N,llvm::Value * Size)861 void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N,
862 llvm::Value *Size) {
863 QualType PrivateType = getPrivateType(N);
864 if (!PrivateType->isVariablyModifiedType()) {
865 assert(!Size && !Sizes[N].second &&
866 "Size should be nullptr for non-variably modified reduction "
867 "items.");
868 return;
869 }
870 CodeGenFunction::OpaqueValueMapping OpaqueMap(
871 CGF,
872 cast<OpaqueValueExpr>(
873 CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()),
874 RValue::get(Size));
875 CGF.EmitVariablyModifiedType(PrivateType);
876 }
877
emitInitialization(CodeGenFunction & CGF,unsigned N,Address PrivateAddr,Address SharedAddr,llvm::function_ref<bool (CodeGenFunction &)> DefaultInit)878 void ReductionCodeGen::emitInitialization(
879 CodeGenFunction &CGF, unsigned N, Address PrivateAddr, Address SharedAddr,
880 llvm::function_ref<bool(CodeGenFunction &)> DefaultInit) {
881 assert(SharedAddresses.size() > N && "No variable was generated");
882 const auto *PrivateVD =
883 cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl());
884 const OMPDeclareReductionDecl *DRD =
885 getReductionInit(ClausesData[N].ReductionOp);
886 if (CGF.getContext().getAsArrayType(PrivateVD->getType())) {
887 if (DRD && DRD->getInitializer())
888 (void)DefaultInit(CGF);
889 emitAggregateInitialization(CGF, N, PrivateAddr, SharedAddr, DRD);
890 } else if (DRD && (DRD->getInitializer() || !PrivateVD->hasInit())) {
891 (void)DefaultInit(CGF);
892 QualType SharedType = SharedAddresses[N].first.getType();
893 emitInitWithReductionInitializer(CGF, DRD, ClausesData[N].ReductionOp,
894 PrivateAddr, SharedAddr, SharedType);
895 } else if (!DefaultInit(CGF) && PrivateVD->hasInit() &&
896 !CGF.isTrivialInitializer(PrivateVD->getInit())) {
897 CGF.EmitAnyExprToMem(PrivateVD->getInit(), PrivateAddr,
898 PrivateVD->getType().getQualifiers(),
899 /*IsInitializer=*/false);
900 }
901 }
902
needCleanups(unsigned N)903 bool ReductionCodeGen::needCleanups(unsigned N) {
904 QualType PrivateType = getPrivateType(N);
905 QualType::DestructionKind DTorKind = PrivateType.isDestructedType();
906 return DTorKind != QualType::DK_none;
907 }
908
emitCleanups(CodeGenFunction & CGF,unsigned N,Address PrivateAddr)909 void ReductionCodeGen::emitCleanups(CodeGenFunction &CGF, unsigned N,
910 Address PrivateAddr) {
911 QualType PrivateType = getPrivateType(N);
912 QualType::DestructionKind DTorKind = PrivateType.isDestructedType();
913 if (needCleanups(N)) {
914 PrivateAddr = CGF.Builder.CreateElementBitCast(
915 PrivateAddr, CGF.ConvertTypeForMem(PrivateType));
916 CGF.pushDestroy(DTorKind, PrivateAddr, PrivateType);
917 }
918 }
919
loadToBegin(CodeGenFunction & CGF,QualType BaseTy,QualType ElTy,LValue BaseLV)920 static LValue loadToBegin(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy,
921 LValue BaseLV) {
922 BaseTy = BaseTy.getNonReferenceType();
923 while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) &&
924 !CGF.getContext().hasSameType(BaseTy, ElTy)) {
925 if (const auto *PtrTy = BaseTy->getAs<PointerType>()) {
926 BaseLV = CGF.EmitLoadOfPointerLValue(BaseLV.getAddress(CGF), PtrTy);
927 } else {
928 LValue RefLVal = CGF.MakeAddrLValue(BaseLV.getAddress(CGF), BaseTy);
929 BaseLV = CGF.EmitLoadOfReferenceLValue(RefLVal);
930 }
931 BaseTy = BaseTy->getPointeeType();
932 }
933 return CGF.MakeAddrLValue(
934 CGF.Builder.CreateElementBitCast(BaseLV.getAddress(CGF),
935 CGF.ConvertTypeForMem(ElTy)),
936 BaseLV.getType(), BaseLV.getBaseInfo(),
937 CGF.CGM.getTBAAInfoForSubobject(BaseLV, BaseLV.getType()));
938 }
939
castToBase(CodeGenFunction & CGF,QualType BaseTy,QualType ElTy,Address OriginalBaseAddress,llvm::Value * Addr)940 static Address castToBase(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy,
941 Address OriginalBaseAddress, llvm::Value *Addr) {
942 Address Tmp = Address::invalid();
943 Address TopTmp = Address::invalid();
944 Address MostTopTmp = Address::invalid();
945 BaseTy = BaseTy.getNonReferenceType();
946 while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) &&
947 !CGF.getContext().hasSameType(BaseTy, ElTy)) {
948 Tmp = CGF.CreateMemTemp(BaseTy);
949 if (TopTmp.isValid())
950 CGF.Builder.CreateStore(Tmp.getPointer(), TopTmp);
951 else
952 MostTopTmp = Tmp;
953 TopTmp = Tmp;
954 BaseTy = BaseTy->getPointeeType();
955 }
956
957 if (Tmp.isValid()) {
958 Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
959 Addr, Tmp.getElementType());
960 CGF.Builder.CreateStore(Addr, Tmp);
961 return MostTopTmp;
962 }
963
964 Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
965 Addr, OriginalBaseAddress.getType());
966 return OriginalBaseAddress.withPointer(Addr);
967 }
968
getBaseDecl(const Expr * Ref,const DeclRefExpr * & DE)969 static const VarDecl *getBaseDecl(const Expr *Ref, const DeclRefExpr *&DE) {
970 const VarDecl *OrigVD = nullptr;
971 if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(Ref)) {
972 const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
973 while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base))
974 Base = TempOASE->getBase()->IgnoreParenImpCasts();
975 while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
976 Base = TempASE->getBase()->IgnoreParenImpCasts();
977 DE = cast<DeclRefExpr>(Base);
978 OrigVD = cast<VarDecl>(DE->getDecl());
979 } else if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Ref)) {
980 const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
981 while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
982 Base = TempASE->getBase()->IgnoreParenImpCasts();
983 DE = cast<DeclRefExpr>(Base);
984 OrigVD = cast<VarDecl>(DE->getDecl());
985 }
986 return OrigVD;
987 }
988
adjustPrivateAddress(CodeGenFunction & CGF,unsigned N,Address PrivateAddr)989 Address ReductionCodeGen::adjustPrivateAddress(CodeGenFunction &CGF, unsigned N,
990 Address PrivateAddr) {
991 const DeclRefExpr *DE;
992 if (const VarDecl *OrigVD = ::getBaseDecl(ClausesData[N].Ref, DE)) {
993 BaseDecls.emplace_back(OrigVD);
994 LValue OriginalBaseLValue = CGF.EmitLValue(DE);
995 LValue BaseLValue =
996 loadToBegin(CGF, OrigVD->getType(), SharedAddresses[N].first.getType(),
997 OriginalBaseLValue);
998 Address SharedAddr = SharedAddresses[N].first.getAddress(CGF);
999 llvm::Value *Adjustment = CGF.Builder.CreatePtrDiff(
1000 SharedAddr.getElementType(), BaseLValue.getPointer(CGF),
1001 SharedAddr.getPointer());
1002 llvm::Value *PrivatePointer =
1003 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
1004 PrivateAddr.getPointer(), SharedAddr.getType());
1005 llvm::Value *Ptr = CGF.Builder.CreateGEP(
1006 SharedAddr.getElementType(), PrivatePointer, Adjustment);
1007 return castToBase(CGF, OrigVD->getType(),
1008 SharedAddresses[N].first.getType(),
1009 OriginalBaseLValue.getAddress(CGF), Ptr);
1010 }
1011 BaseDecls.emplace_back(
1012 cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Ref)->getDecl()));
1013 return PrivateAddr;
1014 }
1015
usesReductionInitializer(unsigned N) const1016 bool ReductionCodeGen::usesReductionInitializer(unsigned N) const {
1017 const OMPDeclareReductionDecl *DRD =
1018 getReductionInit(ClausesData[N].ReductionOp);
1019 return DRD && DRD->getInitializer();
1020 }
1021
getThreadIDVariableLValue(CodeGenFunction & CGF)1022 LValue CGOpenMPRegionInfo::getThreadIDVariableLValue(CodeGenFunction &CGF) {
1023 return CGF.EmitLoadOfPointerLValue(
1024 CGF.GetAddrOfLocalVar(getThreadIDVariable()),
1025 getThreadIDVariable()->getType()->castAs<PointerType>());
1026 }
1027
EmitBody(CodeGenFunction & CGF,const Stmt * S)1028 void CGOpenMPRegionInfo::EmitBody(CodeGenFunction &CGF, const Stmt *S) {
1029 if (!CGF.HaveInsertPoint())
1030 return;
1031 // 1.2.2 OpenMP Language Terminology
1032 // Structured block - An executable statement with a single entry at the
1033 // top and a single exit at the bottom.
1034 // The point of exit cannot be a branch out of the structured block.
1035 // longjmp() and throw() must not violate the entry/exit criteria.
1036 CGF.EHStack.pushTerminate();
1037 if (S)
1038 CGF.incrementProfileCounter(S);
1039 CodeGen(CGF);
1040 CGF.EHStack.popTerminate();
1041 }
1042
getThreadIDVariableLValue(CodeGenFunction & CGF)1043 LValue CGOpenMPTaskOutlinedRegionInfo::getThreadIDVariableLValue(
1044 CodeGenFunction &CGF) {
1045 return CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(getThreadIDVariable()),
1046 getThreadIDVariable()->getType(),
1047 AlignmentSource::Decl);
1048 }
1049
addFieldToRecordDecl(ASTContext & C,DeclContext * DC,QualType FieldTy)1050 static FieldDecl *addFieldToRecordDecl(ASTContext &C, DeclContext *DC,
1051 QualType FieldTy) {
1052 auto *Field = FieldDecl::Create(
1053 C, DC, SourceLocation(), SourceLocation(), /*Id=*/nullptr, FieldTy,
1054 C.getTrivialTypeSourceInfo(FieldTy, SourceLocation()),
1055 /*BW=*/nullptr, /*Mutable=*/false, /*InitStyle=*/ICIS_NoInit);
1056 Field->setAccess(AS_public);
1057 DC->addDecl(Field);
1058 return Field;
1059 }
1060
CGOpenMPRuntime(CodeGenModule & CGM)1061 CGOpenMPRuntime::CGOpenMPRuntime(CodeGenModule &CGM)
1062 : CGM(CGM), OMPBuilder(CGM.getModule()), OffloadEntriesInfoManager() {
1063 KmpCriticalNameTy = llvm::ArrayType::get(CGM.Int32Ty, /*NumElements*/ 8);
1064 llvm::OpenMPIRBuilderConfig Config(CGM.getLangOpts().OpenMPIsDevice, false,
1065 hasRequiresUnifiedSharedMemory(),
1066 CGM.getLangOpts().OpenMPOffloadMandatory);
1067 // Initialize Types used in OpenMPIRBuilder from OMPKinds.def
1068 OMPBuilder.initialize();
1069 OMPBuilder.setConfig(Config);
1070 OffloadEntriesInfoManager.setConfig(Config);
1071 loadOffloadInfoMetadata();
1072 }
1073
clear()1074 void CGOpenMPRuntime::clear() {
1075 InternalVars.clear();
1076 // Clean non-target variable declarations possibly used only in debug info.
1077 for (const auto &Data : EmittedNonTargetVariables) {
1078 if (!Data.getValue().pointsToAliveValue())
1079 continue;
1080 auto *GV = dyn_cast<llvm::GlobalVariable>(Data.getValue());
1081 if (!GV)
1082 continue;
1083 if (!GV->isDeclaration() || GV->getNumUses() > 0)
1084 continue;
1085 GV->eraseFromParent();
1086 }
1087 }
1088
getName(ArrayRef<StringRef> Parts) const1089 std::string CGOpenMPRuntime::getName(ArrayRef<StringRef> Parts) const {
1090 return OMPBuilder.createPlatformSpecificName(Parts);
1091 }
1092
1093 static llvm::Function *
emitCombinerOrInitializer(CodeGenModule & CGM,QualType Ty,const Expr * CombinerInitializer,const VarDecl * In,const VarDecl * Out,bool IsCombiner)1094 emitCombinerOrInitializer(CodeGenModule &CGM, QualType Ty,
1095 const Expr *CombinerInitializer, const VarDecl *In,
1096 const VarDecl *Out, bool IsCombiner) {
1097 // void .omp_combiner.(Ty *in, Ty *out);
1098 ASTContext &C = CGM.getContext();
1099 QualType PtrTy = C.getPointerType(Ty).withRestrict();
1100 FunctionArgList Args;
1101 ImplicitParamDecl OmpOutParm(C, /*DC=*/nullptr, Out->getLocation(),
1102 /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other);
1103 ImplicitParamDecl OmpInParm(C, /*DC=*/nullptr, In->getLocation(),
1104 /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other);
1105 Args.push_back(&OmpOutParm);
1106 Args.push_back(&OmpInParm);
1107 const CGFunctionInfo &FnInfo =
1108 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
1109 llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo);
1110 std::string Name = CGM.getOpenMPRuntime().getName(
1111 {IsCombiner ? "omp_combiner" : "omp_initializer", ""});
1112 auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage,
1113 Name, &CGM.getModule());
1114 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo);
1115 if (CGM.getLangOpts().Optimize) {
1116 Fn->removeFnAttr(llvm::Attribute::NoInline);
1117 Fn->removeFnAttr(llvm::Attribute::OptimizeNone);
1118 Fn->addFnAttr(llvm::Attribute::AlwaysInline);
1119 }
1120 CodeGenFunction CGF(CGM);
1121 // Map "T omp_in;" variable to "*omp_in_parm" value in all expressions.
1122 // Map "T omp_out;" variable to "*omp_out_parm" value in all expressions.
1123 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, In->getLocation(),
1124 Out->getLocation());
1125 CodeGenFunction::OMPPrivateScope Scope(CGF);
1126 Address AddrIn = CGF.GetAddrOfLocalVar(&OmpInParm);
1127 Scope.addPrivate(
1128 In, CGF.EmitLoadOfPointerLValue(AddrIn, PtrTy->castAs<PointerType>())
1129 .getAddress(CGF));
1130 Address AddrOut = CGF.GetAddrOfLocalVar(&OmpOutParm);
1131 Scope.addPrivate(
1132 Out, CGF.EmitLoadOfPointerLValue(AddrOut, PtrTy->castAs<PointerType>())
1133 .getAddress(CGF));
1134 (void)Scope.Privatize();
1135 if (!IsCombiner && Out->hasInit() &&
1136 !CGF.isTrivialInitializer(Out->getInit())) {
1137 CGF.EmitAnyExprToMem(Out->getInit(), CGF.GetAddrOfLocalVar(Out),
1138 Out->getType().getQualifiers(),
1139 /*IsInitializer=*/true);
1140 }
1141 if (CombinerInitializer)
1142 CGF.EmitIgnoredExpr(CombinerInitializer);
1143 Scope.ForceCleanup();
1144 CGF.FinishFunction();
1145 return Fn;
1146 }
1147
emitUserDefinedReduction(CodeGenFunction * CGF,const OMPDeclareReductionDecl * D)1148 void CGOpenMPRuntime::emitUserDefinedReduction(
1149 CodeGenFunction *CGF, const OMPDeclareReductionDecl *D) {
1150 if (UDRMap.count(D) > 0)
1151 return;
1152 llvm::Function *Combiner = emitCombinerOrInitializer(
1153 CGM, D->getType(), D->getCombiner(),
1154 cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerIn())->getDecl()),
1155 cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerOut())->getDecl()),
1156 /*IsCombiner=*/true);
1157 llvm::Function *Initializer = nullptr;
1158 if (const Expr *Init = D->getInitializer()) {
1159 Initializer = emitCombinerOrInitializer(
1160 CGM, D->getType(),
1161 D->getInitializerKind() == OMPDeclareReductionDecl::CallInit ? Init
1162 : nullptr,
1163 cast<VarDecl>(cast<DeclRefExpr>(D->getInitOrig())->getDecl()),
1164 cast<VarDecl>(cast<DeclRefExpr>(D->getInitPriv())->getDecl()),
1165 /*IsCombiner=*/false);
1166 }
1167 UDRMap.try_emplace(D, Combiner, Initializer);
1168 if (CGF) {
1169 auto &Decls = FunctionUDRMap.FindAndConstruct(CGF->CurFn);
1170 Decls.second.push_back(D);
1171 }
1172 }
1173
1174 std::pair<llvm::Function *, llvm::Function *>
getUserDefinedReduction(const OMPDeclareReductionDecl * D)1175 CGOpenMPRuntime::getUserDefinedReduction(const OMPDeclareReductionDecl *D) {
1176 auto I = UDRMap.find(D);
1177 if (I != UDRMap.end())
1178 return I->second;
1179 emitUserDefinedReduction(/*CGF=*/nullptr, D);
1180 return UDRMap.lookup(D);
1181 }
1182
1183 namespace {
1184 // Temporary RAII solution to perform a push/pop stack event on the OpenMP IR
1185 // Builder if one is present.
1186 struct PushAndPopStackRAII {
PushAndPopStackRAII__anon7bb087080311::PushAndPopStackRAII1187 PushAndPopStackRAII(llvm::OpenMPIRBuilder *OMPBuilder, CodeGenFunction &CGF,
1188 bool HasCancel, llvm::omp::Directive Kind)
1189 : OMPBuilder(OMPBuilder) {
1190 if (!OMPBuilder)
1191 return;
1192
1193 // The following callback is the crucial part of clangs cleanup process.
1194 //
1195 // NOTE:
1196 // Once the OpenMPIRBuilder is used to create parallel regions (and
1197 // similar), the cancellation destination (Dest below) is determined via
1198 // IP. That means if we have variables to finalize we split the block at IP,
1199 // use the new block (=BB) as destination to build a JumpDest (via
1200 // getJumpDestInCurrentScope(BB)) which then is fed to
1201 // EmitBranchThroughCleanup. Furthermore, there will not be the need
1202 // to push & pop an FinalizationInfo object.
1203 // The FiniCB will still be needed but at the point where the
1204 // OpenMPIRBuilder is asked to construct a parallel (or similar) construct.
1205 auto FiniCB = [&CGF](llvm::OpenMPIRBuilder::InsertPointTy IP) {
1206 assert(IP.getBlock()->end() == IP.getPoint() &&
1207 "Clang CG should cause non-terminated block!");
1208 CGBuilderTy::InsertPointGuard IPG(CGF.Builder);
1209 CGF.Builder.restoreIP(IP);
1210 CodeGenFunction::JumpDest Dest =
1211 CGF.getOMPCancelDestination(OMPD_parallel);
1212 CGF.EmitBranchThroughCleanup(Dest);
1213 };
1214
1215 // TODO: Remove this once we emit parallel regions through the
1216 // OpenMPIRBuilder as it can do this setup internally.
1217 llvm::OpenMPIRBuilder::FinalizationInfo FI({FiniCB, Kind, HasCancel});
1218 OMPBuilder->pushFinalizationCB(std::move(FI));
1219 }
~PushAndPopStackRAII__anon7bb087080311::PushAndPopStackRAII1220 ~PushAndPopStackRAII() {
1221 if (OMPBuilder)
1222 OMPBuilder->popFinalizationCB();
1223 }
1224 llvm::OpenMPIRBuilder *OMPBuilder;
1225 };
1226 } // namespace
1227
emitParallelOrTeamsOutlinedFunction(CodeGenModule & CGM,const OMPExecutableDirective & D,const CapturedStmt * CS,const VarDecl * ThreadIDVar,OpenMPDirectiveKind InnermostKind,const StringRef OutlinedHelperName,const RegionCodeGenTy & CodeGen)1228 static llvm::Function *emitParallelOrTeamsOutlinedFunction(
1229 CodeGenModule &CGM, const OMPExecutableDirective &D, const CapturedStmt *CS,
1230 const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
1231 const StringRef OutlinedHelperName, const RegionCodeGenTy &CodeGen) {
1232 assert(ThreadIDVar->getType()->isPointerType() &&
1233 "thread id variable must be of type kmp_int32 *");
1234 CodeGenFunction CGF(CGM, true);
1235 bool HasCancel = false;
1236 if (const auto *OPD = dyn_cast<OMPParallelDirective>(&D))
1237 HasCancel = OPD->hasCancel();
1238 else if (const auto *OPD = dyn_cast<OMPTargetParallelDirective>(&D))
1239 HasCancel = OPD->hasCancel();
1240 else if (const auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&D))
1241 HasCancel = OPSD->hasCancel();
1242 else if (const auto *OPFD = dyn_cast<OMPParallelForDirective>(&D))
1243 HasCancel = OPFD->hasCancel();
1244 else if (const auto *OPFD = dyn_cast<OMPTargetParallelForDirective>(&D))
1245 HasCancel = OPFD->hasCancel();
1246 else if (const auto *OPFD = dyn_cast<OMPDistributeParallelForDirective>(&D))
1247 HasCancel = OPFD->hasCancel();
1248 else if (const auto *OPFD =
1249 dyn_cast<OMPTeamsDistributeParallelForDirective>(&D))
1250 HasCancel = OPFD->hasCancel();
1251 else if (const auto *OPFD =
1252 dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&D))
1253 HasCancel = OPFD->hasCancel();
1254
1255 // TODO: Temporarily inform the OpenMPIRBuilder, if any, about the new
1256 // parallel region to make cancellation barriers work properly.
1257 llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder();
1258 PushAndPopStackRAII PSR(&OMPBuilder, CGF, HasCancel, InnermostKind);
1259 CGOpenMPOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, InnermostKind,
1260 HasCancel, OutlinedHelperName);
1261 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
1262 return CGF.GenerateOpenMPCapturedStmtFunction(*CS, D.getBeginLoc());
1263 }
1264
emitParallelOutlinedFunction(const OMPExecutableDirective & D,const VarDecl * ThreadIDVar,OpenMPDirectiveKind InnermostKind,const RegionCodeGenTy & CodeGen)1265 llvm::Function *CGOpenMPRuntime::emitParallelOutlinedFunction(
1266 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
1267 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
1268 const CapturedStmt *CS = D.getCapturedStmt(OMPD_parallel);
1269 return emitParallelOrTeamsOutlinedFunction(
1270 CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen);
1271 }
1272
emitTeamsOutlinedFunction(const OMPExecutableDirective & D,const VarDecl * ThreadIDVar,OpenMPDirectiveKind InnermostKind,const RegionCodeGenTy & CodeGen)1273 llvm::Function *CGOpenMPRuntime::emitTeamsOutlinedFunction(
1274 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
1275 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
1276 const CapturedStmt *CS = D.getCapturedStmt(OMPD_teams);
1277 return emitParallelOrTeamsOutlinedFunction(
1278 CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen);
1279 }
1280
emitTaskOutlinedFunction(const OMPExecutableDirective & D,const VarDecl * ThreadIDVar,const VarDecl * PartIDVar,const VarDecl * TaskTVar,OpenMPDirectiveKind InnermostKind,const RegionCodeGenTy & CodeGen,bool Tied,unsigned & NumberOfParts)1281 llvm::Function *CGOpenMPRuntime::emitTaskOutlinedFunction(
1282 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
1283 const VarDecl *PartIDVar, const VarDecl *TaskTVar,
1284 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
1285 bool Tied, unsigned &NumberOfParts) {
1286 auto &&UntiedCodeGen = [this, &D, TaskTVar](CodeGenFunction &CGF,
1287 PrePostActionTy &) {
1288 llvm::Value *ThreadID = getThreadID(CGF, D.getBeginLoc());
1289 llvm::Value *UpLoc = emitUpdateLocation(CGF, D.getBeginLoc());
1290 llvm::Value *TaskArgs[] = {
1291 UpLoc, ThreadID,
1292 CGF.EmitLoadOfPointerLValue(CGF.GetAddrOfLocalVar(TaskTVar),
1293 TaskTVar->getType()->castAs<PointerType>())
1294 .getPointer(CGF)};
1295 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1296 CGM.getModule(), OMPRTL___kmpc_omp_task),
1297 TaskArgs);
1298 };
1299 CGOpenMPTaskOutlinedRegionInfo::UntiedTaskActionTy Action(Tied, PartIDVar,
1300 UntiedCodeGen);
1301 CodeGen.setAction(Action);
1302 assert(!ThreadIDVar->getType()->isPointerType() &&
1303 "thread id variable must be of type kmp_int32 for tasks");
1304 const OpenMPDirectiveKind Region =
1305 isOpenMPTaskLoopDirective(D.getDirectiveKind()) ? OMPD_taskloop
1306 : OMPD_task;
1307 const CapturedStmt *CS = D.getCapturedStmt(Region);
1308 bool HasCancel = false;
1309 if (const auto *TD = dyn_cast<OMPTaskDirective>(&D))
1310 HasCancel = TD->hasCancel();
1311 else if (const auto *TD = dyn_cast<OMPTaskLoopDirective>(&D))
1312 HasCancel = TD->hasCancel();
1313 else if (const auto *TD = dyn_cast<OMPMasterTaskLoopDirective>(&D))
1314 HasCancel = TD->hasCancel();
1315 else if (const auto *TD = dyn_cast<OMPParallelMasterTaskLoopDirective>(&D))
1316 HasCancel = TD->hasCancel();
1317
1318 CodeGenFunction CGF(CGM, true);
1319 CGOpenMPTaskOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen,
1320 InnermostKind, HasCancel, Action);
1321 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
1322 llvm::Function *Res = CGF.GenerateCapturedStmtFunction(*CS);
1323 if (!Tied)
1324 NumberOfParts = Action.getNumberOfParts();
1325 return Res;
1326 }
1327
setLocThreadIdInsertPt(CodeGenFunction & CGF,bool AtCurrentPoint)1328 void CGOpenMPRuntime::setLocThreadIdInsertPt(CodeGenFunction &CGF,
1329 bool AtCurrentPoint) {
1330 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
1331 assert(!Elem.second.ServiceInsertPt && "Insert point is set already.");
1332
1333 llvm::Value *Undef = llvm::UndefValue::get(CGF.Int32Ty);
1334 if (AtCurrentPoint) {
1335 Elem.second.ServiceInsertPt = new llvm::BitCastInst(
1336 Undef, CGF.Int32Ty, "svcpt", CGF.Builder.GetInsertBlock());
1337 } else {
1338 Elem.second.ServiceInsertPt =
1339 new llvm::BitCastInst(Undef, CGF.Int32Ty, "svcpt");
1340 Elem.second.ServiceInsertPt->insertAfter(CGF.AllocaInsertPt);
1341 }
1342 }
1343
clearLocThreadIdInsertPt(CodeGenFunction & CGF)1344 void CGOpenMPRuntime::clearLocThreadIdInsertPt(CodeGenFunction &CGF) {
1345 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
1346 if (Elem.second.ServiceInsertPt) {
1347 llvm::Instruction *Ptr = Elem.second.ServiceInsertPt;
1348 Elem.second.ServiceInsertPt = nullptr;
1349 Ptr->eraseFromParent();
1350 }
1351 }
1352
getIdentStringFromSourceLocation(CodeGenFunction & CGF,SourceLocation Loc,SmallString<128> & Buffer)1353 static StringRef getIdentStringFromSourceLocation(CodeGenFunction &CGF,
1354 SourceLocation Loc,
1355 SmallString<128> &Buffer) {
1356 llvm::raw_svector_ostream OS(Buffer);
1357 // Build debug location
1358 PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc);
1359 OS << ";" << PLoc.getFilename() << ";";
1360 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl))
1361 OS << FD->getQualifiedNameAsString();
1362 OS << ";" << PLoc.getLine() << ";" << PLoc.getColumn() << ";;";
1363 return OS.str();
1364 }
1365
emitUpdateLocation(CodeGenFunction & CGF,SourceLocation Loc,unsigned Flags,bool EmitLoc)1366 llvm::Value *CGOpenMPRuntime::emitUpdateLocation(CodeGenFunction &CGF,
1367 SourceLocation Loc,
1368 unsigned Flags, bool EmitLoc) {
1369 uint32_t SrcLocStrSize;
1370 llvm::Constant *SrcLocStr;
1371 if ((!EmitLoc &&
1372 CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo) ||
1373 Loc.isInvalid()) {
1374 SrcLocStr = OMPBuilder.getOrCreateDefaultSrcLocStr(SrcLocStrSize);
1375 } else {
1376 std::string FunctionName;
1377 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl))
1378 FunctionName = FD->getQualifiedNameAsString();
1379 PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc);
1380 const char *FileName = PLoc.getFilename();
1381 unsigned Line = PLoc.getLine();
1382 unsigned Column = PLoc.getColumn();
1383 SrcLocStr = OMPBuilder.getOrCreateSrcLocStr(FunctionName, FileName, Line,
1384 Column, SrcLocStrSize);
1385 }
1386 unsigned Reserved2Flags = getDefaultLocationReserved2Flags();
1387 return OMPBuilder.getOrCreateIdent(
1388 SrcLocStr, SrcLocStrSize, llvm::omp::IdentFlag(Flags), Reserved2Flags);
1389 }
1390
getThreadID(CodeGenFunction & CGF,SourceLocation Loc)1391 llvm::Value *CGOpenMPRuntime::getThreadID(CodeGenFunction &CGF,
1392 SourceLocation Loc) {
1393 assert(CGF.CurFn && "No function in current CodeGenFunction.");
1394 // If the OpenMPIRBuilder is used we need to use it for all thread id calls as
1395 // the clang invariants used below might be broken.
1396 if (CGM.getLangOpts().OpenMPIRBuilder) {
1397 SmallString<128> Buffer;
1398 OMPBuilder.updateToLocation(CGF.Builder.saveIP());
1399 uint32_t SrcLocStrSize;
1400 auto *SrcLocStr = OMPBuilder.getOrCreateSrcLocStr(
1401 getIdentStringFromSourceLocation(CGF, Loc, Buffer), SrcLocStrSize);
1402 return OMPBuilder.getOrCreateThreadID(
1403 OMPBuilder.getOrCreateIdent(SrcLocStr, SrcLocStrSize));
1404 }
1405
1406 llvm::Value *ThreadID = nullptr;
1407 // Check whether we've already cached a load of the thread id in this
1408 // function.
1409 auto I = OpenMPLocThreadIDMap.find(CGF.CurFn);
1410 if (I != OpenMPLocThreadIDMap.end()) {
1411 ThreadID = I->second.ThreadID;
1412 if (ThreadID != nullptr)
1413 return ThreadID;
1414 }
1415 // If exceptions are enabled, do not use parameter to avoid possible crash.
1416 if (auto *OMPRegionInfo =
1417 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
1418 if (OMPRegionInfo->getThreadIDVariable()) {
1419 // Check if this an outlined function with thread id passed as argument.
1420 LValue LVal = OMPRegionInfo->getThreadIDVariableLValue(CGF);
1421 llvm::BasicBlock *TopBlock = CGF.AllocaInsertPt->getParent();
1422 if (!CGF.EHStack.requiresLandingPad() || !CGF.getLangOpts().Exceptions ||
1423 !CGF.getLangOpts().CXXExceptions ||
1424 CGF.Builder.GetInsertBlock() == TopBlock ||
1425 !isa<llvm::Instruction>(LVal.getPointer(CGF)) ||
1426 cast<llvm::Instruction>(LVal.getPointer(CGF))->getParent() ==
1427 TopBlock ||
1428 cast<llvm::Instruction>(LVal.getPointer(CGF))->getParent() ==
1429 CGF.Builder.GetInsertBlock()) {
1430 ThreadID = CGF.EmitLoadOfScalar(LVal, Loc);
1431 // If value loaded in entry block, cache it and use it everywhere in
1432 // function.
1433 if (CGF.Builder.GetInsertBlock() == TopBlock) {
1434 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
1435 Elem.second.ThreadID = ThreadID;
1436 }
1437 return ThreadID;
1438 }
1439 }
1440 }
1441
1442 // This is not an outlined function region - need to call __kmpc_int32
1443 // kmpc_global_thread_num(ident_t *loc).
1444 // Generate thread id value and cache this value for use across the
1445 // function.
1446 auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn);
1447 if (!Elem.second.ServiceInsertPt)
1448 setLocThreadIdInsertPt(CGF);
1449 CGBuilderTy::InsertPointGuard IPG(CGF.Builder);
1450 CGF.Builder.SetInsertPoint(Elem.second.ServiceInsertPt);
1451 llvm::CallInst *Call = CGF.Builder.CreateCall(
1452 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
1453 OMPRTL___kmpc_global_thread_num),
1454 emitUpdateLocation(CGF, Loc));
1455 Call->setCallingConv(CGF.getRuntimeCC());
1456 Elem.second.ThreadID = Call;
1457 return Call;
1458 }
1459
functionFinished(CodeGenFunction & CGF)1460 void CGOpenMPRuntime::functionFinished(CodeGenFunction &CGF) {
1461 assert(CGF.CurFn && "No function in current CodeGenFunction.");
1462 if (OpenMPLocThreadIDMap.count(CGF.CurFn)) {
1463 clearLocThreadIdInsertPt(CGF);
1464 OpenMPLocThreadIDMap.erase(CGF.CurFn);
1465 }
1466 if (FunctionUDRMap.count(CGF.CurFn) > 0) {
1467 for(const auto *D : FunctionUDRMap[CGF.CurFn])
1468 UDRMap.erase(D);
1469 FunctionUDRMap.erase(CGF.CurFn);
1470 }
1471 auto I = FunctionUDMMap.find(CGF.CurFn);
1472 if (I != FunctionUDMMap.end()) {
1473 for(const auto *D : I->second)
1474 UDMMap.erase(D);
1475 FunctionUDMMap.erase(I);
1476 }
1477 LastprivateConditionalToTypes.erase(CGF.CurFn);
1478 FunctionToUntiedTaskStackMap.erase(CGF.CurFn);
1479 }
1480
getIdentTyPointerTy()1481 llvm::Type *CGOpenMPRuntime::getIdentTyPointerTy() {
1482 return OMPBuilder.IdentPtr;
1483 }
1484
getKmpc_MicroPointerTy()1485 llvm::Type *CGOpenMPRuntime::getKmpc_MicroPointerTy() {
1486 if (!Kmpc_MicroTy) {
1487 // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
1488 llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty),
1489 llvm::PointerType::getUnqual(CGM.Int32Ty)};
1490 Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true);
1491 }
1492 return llvm::PointerType::getUnqual(Kmpc_MicroTy);
1493 }
1494
1495 llvm::FunctionCallee
createForStaticInitFunction(unsigned IVSize,bool IVSigned,bool IsGPUDistribute)1496 CGOpenMPRuntime::createForStaticInitFunction(unsigned IVSize, bool IVSigned,
1497 bool IsGPUDistribute) {
1498 assert((IVSize == 32 || IVSize == 64) &&
1499 "IV size is not compatible with the omp runtime");
1500 StringRef Name;
1501 if (IsGPUDistribute)
1502 Name = IVSize == 32 ? (IVSigned ? "__kmpc_distribute_static_init_4"
1503 : "__kmpc_distribute_static_init_4u")
1504 : (IVSigned ? "__kmpc_distribute_static_init_8"
1505 : "__kmpc_distribute_static_init_8u");
1506 else
1507 Name = IVSize == 32 ? (IVSigned ? "__kmpc_for_static_init_4"
1508 : "__kmpc_for_static_init_4u")
1509 : (IVSigned ? "__kmpc_for_static_init_8"
1510 : "__kmpc_for_static_init_8u");
1511
1512 llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty;
1513 auto *PtrTy = llvm::PointerType::getUnqual(ITy);
1514 llvm::Type *TypeParams[] = {
1515 getIdentTyPointerTy(), // loc
1516 CGM.Int32Ty, // tid
1517 CGM.Int32Ty, // schedtype
1518 llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter
1519 PtrTy, // p_lower
1520 PtrTy, // p_upper
1521 PtrTy, // p_stride
1522 ITy, // incr
1523 ITy // chunk
1524 };
1525 auto *FnTy =
1526 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1527 return CGM.CreateRuntimeFunction(FnTy, Name);
1528 }
1529
1530 llvm::FunctionCallee
createDispatchInitFunction(unsigned IVSize,bool IVSigned)1531 CGOpenMPRuntime::createDispatchInitFunction(unsigned IVSize, bool IVSigned) {
1532 assert((IVSize == 32 || IVSize == 64) &&
1533 "IV size is not compatible with the omp runtime");
1534 StringRef Name =
1535 IVSize == 32
1536 ? (IVSigned ? "__kmpc_dispatch_init_4" : "__kmpc_dispatch_init_4u")
1537 : (IVSigned ? "__kmpc_dispatch_init_8" : "__kmpc_dispatch_init_8u");
1538 llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty;
1539 llvm::Type *TypeParams[] = { getIdentTyPointerTy(), // loc
1540 CGM.Int32Ty, // tid
1541 CGM.Int32Ty, // schedtype
1542 ITy, // lower
1543 ITy, // upper
1544 ITy, // stride
1545 ITy // chunk
1546 };
1547 auto *FnTy =
1548 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1549 return CGM.CreateRuntimeFunction(FnTy, Name);
1550 }
1551
1552 llvm::FunctionCallee
createDispatchFiniFunction(unsigned IVSize,bool IVSigned)1553 CGOpenMPRuntime::createDispatchFiniFunction(unsigned IVSize, bool IVSigned) {
1554 assert((IVSize == 32 || IVSize == 64) &&
1555 "IV size is not compatible with the omp runtime");
1556 StringRef Name =
1557 IVSize == 32
1558 ? (IVSigned ? "__kmpc_dispatch_fini_4" : "__kmpc_dispatch_fini_4u")
1559 : (IVSigned ? "__kmpc_dispatch_fini_8" : "__kmpc_dispatch_fini_8u");
1560 llvm::Type *TypeParams[] = {
1561 getIdentTyPointerTy(), // loc
1562 CGM.Int32Ty, // tid
1563 };
1564 auto *FnTy =
1565 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1566 return CGM.CreateRuntimeFunction(FnTy, Name);
1567 }
1568
1569 llvm::FunctionCallee
createDispatchNextFunction(unsigned IVSize,bool IVSigned)1570 CGOpenMPRuntime::createDispatchNextFunction(unsigned IVSize, bool IVSigned) {
1571 assert((IVSize == 32 || IVSize == 64) &&
1572 "IV size is not compatible with the omp runtime");
1573 StringRef Name =
1574 IVSize == 32
1575 ? (IVSigned ? "__kmpc_dispatch_next_4" : "__kmpc_dispatch_next_4u")
1576 : (IVSigned ? "__kmpc_dispatch_next_8" : "__kmpc_dispatch_next_8u");
1577 llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty;
1578 auto *PtrTy = llvm::PointerType::getUnqual(ITy);
1579 llvm::Type *TypeParams[] = {
1580 getIdentTyPointerTy(), // loc
1581 CGM.Int32Ty, // tid
1582 llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter
1583 PtrTy, // p_lower
1584 PtrTy, // p_upper
1585 PtrTy // p_stride
1586 };
1587 auto *FnTy =
1588 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1589 return CGM.CreateRuntimeFunction(FnTy, Name);
1590 }
1591
1592 /// Obtain information that uniquely identifies a target entry. This
1593 /// consists of the file and device IDs as well as line number associated with
1594 /// the relevant entry source location.
1595 static llvm::TargetRegionEntryInfo
getTargetEntryUniqueInfo(ASTContext & C,SourceLocation Loc,StringRef ParentName="")1596 getTargetEntryUniqueInfo(ASTContext &C, SourceLocation Loc,
1597 StringRef ParentName = "") {
1598 SourceManager &SM = C.getSourceManager();
1599
1600 // The loc should be always valid and have a file ID (the user cannot use
1601 // #pragma directives in macros)
1602
1603 assert(Loc.isValid() && "Source location is expected to be always valid.");
1604
1605 PresumedLoc PLoc = SM.getPresumedLoc(Loc);
1606 assert(PLoc.isValid() && "Source location is expected to be always valid.");
1607
1608 llvm::sys::fs::UniqueID ID;
1609 if (auto EC = llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID)) {
1610 PLoc = SM.getPresumedLoc(Loc, /*UseLineDirectives=*/false);
1611 assert(PLoc.isValid() && "Source location is expected to be always valid.");
1612 if (auto EC = llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID))
1613 SM.getDiagnostics().Report(diag::err_cannot_open_file)
1614 << PLoc.getFilename() << EC.message();
1615 }
1616
1617 return llvm::TargetRegionEntryInfo(ParentName, ID.getDevice(), ID.getFile(),
1618 PLoc.getLine());
1619 }
1620
getAddrOfDeclareTargetVar(const VarDecl * VD)1621 Address CGOpenMPRuntime::getAddrOfDeclareTargetVar(const VarDecl *VD) {
1622 if (CGM.getLangOpts().OpenMPSimd)
1623 return Address::invalid();
1624 std::optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res =
1625 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD);
1626 if (Res && (*Res == OMPDeclareTargetDeclAttr::MT_Link ||
1627 ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
1628 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
1629 HasRequiresUnifiedSharedMemory))) {
1630 SmallString<64> PtrName;
1631 {
1632 llvm::raw_svector_ostream OS(PtrName);
1633 OS << CGM.getMangledName(GlobalDecl(VD));
1634 if (!VD->isExternallyVisible()) {
1635 auto EntryInfo = getTargetEntryUniqueInfo(
1636 CGM.getContext(), VD->getCanonicalDecl()->getBeginLoc());
1637 OS << llvm::format("_%x", EntryInfo.FileID);
1638 }
1639 OS << "_decl_tgt_ref_ptr";
1640 }
1641 llvm::Value *Ptr = CGM.getModule().getNamedValue(PtrName);
1642 QualType PtrTy = CGM.getContext().getPointerType(VD->getType());
1643 llvm::Type *LlvmPtrTy = CGM.getTypes().ConvertTypeForMem(PtrTy);
1644 if (!Ptr) {
1645 Ptr = OMPBuilder.getOrCreateInternalVariable(LlvmPtrTy, PtrName);
1646
1647 auto *GV = cast<llvm::GlobalVariable>(Ptr);
1648 GV->setLinkage(llvm::GlobalValue::WeakAnyLinkage);
1649
1650 if (!CGM.getLangOpts().OpenMPIsDevice)
1651 GV->setInitializer(CGM.GetAddrOfGlobal(VD));
1652 registerTargetGlobalVariable(VD, cast<llvm::Constant>(Ptr));
1653 }
1654 return Address(Ptr, LlvmPtrTy, CGM.getContext().getDeclAlign(VD));
1655 }
1656 return Address::invalid();
1657 }
1658
1659 llvm::Constant *
getOrCreateThreadPrivateCache(const VarDecl * VD)1660 CGOpenMPRuntime::getOrCreateThreadPrivateCache(const VarDecl *VD) {
1661 assert(!CGM.getLangOpts().OpenMPUseTLS ||
1662 !CGM.getContext().getTargetInfo().isTLSSupported());
1663 // Lookup the entry, lazily creating it if necessary.
1664 std::string Suffix = getName({"cache", ""});
1665 return OMPBuilder.getOrCreateInternalVariable(
1666 CGM.Int8PtrPtrTy, Twine(CGM.getMangledName(VD)).concat(Suffix).str());
1667 }
1668
getAddrOfThreadPrivate(CodeGenFunction & CGF,const VarDecl * VD,Address VDAddr,SourceLocation Loc)1669 Address CGOpenMPRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF,
1670 const VarDecl *VD,
1671 Address VDAddr,
1672 SourceLocation Loc) {
1673 if (CGM.getLangOpts().OpenMPUseTLS &&
1674 CGM.getContext().getTargetInfo().isTLSSupported())
1675 return VDAddr;
1676
1677 llvm::Type *VarTy = VDAddr.getElementType();
1678 llvm::Value *Args[] = {
1679 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
1680 CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.Int8PtrTy),
1681 CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)),
1682 getOrCreateThreadPrivateCache(VD)};
1683 return Address(
1684 CGF.EmitRuntimeCall(
1685 OMPBuilder.getOrCreateRuntimeFunction(
1686 CGM.getModule(), OMPRTL___kmpc_threadprivate_cached),
1687 Args),
1688 CGF.Int8Ty, VDAddr.getAlignment());
1689 }
1690
emitThreadPrivateVarInit(CodeGenFunction & CGF,Address VDAddr,llvm::Value * Ctor,llvm::Value * CopyCtor,llvm::Value * Dtor,SourceLocation Loc)1691 void CGOpenMPRuntime::emitThreadPrivateVarInit(
1692 CodeGenFunction &CGF, Address VDAddr, llvm::Value *Ctor,
1693 llvm::Value *CopyCtor, llvm::Value *Dtor, SourceLocation Loc) {
1694 // Call kmp_int32 __kmpc_global_thread_num(&loc) to init OpenMP runtime
1695 // library.
1696 llvm::Value *OMPLoc = emitUpdateLocation(CGF, Loc);
1697 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1698 CGM.getModule(), OMPRTL___kmpc_global_thread_num),
1699 OMPLoc);
1700 // Call __kmpc_threadprivate_register(&loc, &var, ctor, cctor/*NULL*/, dtor)
1701 // to register constructor/destructor for variable.
1702 llvm::Value *Args[] = {
1703 OMPLoc, CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.VoidPtrTy),
1704 Ctor, CopyCtor, Dtor};
1705 CGF.EmitRuntimeCall(
1706 OMPBuilder.getOrCreateRuntimeFunction(
1707 CGM.getModule(), OMPRTL___kmpc_threadprivate_register),
1708 Args);
1709 }
1710
emitThreadPrivateVarDefinition(const VarDecl * VD,Address VDAddr,SourceLocation Loc,bool PerformInit,CodeGenFunction * CGF)1711 llvm::Function *CGOpenMPRuntime::emitThreadPrivateVarDefinition(
1712 const VarDecl *VD, Address VDAddr, SourceLocation Loc,
1713 bool PerformInit, CodeGenFunction *CGF) {
1714 if (CGM.getLangOpts().OpenMPUseTLS &&
1715 CGM.getContext().getTargetInfo().isTLSSupported())
1716 return nullptr;
1717
1718 VD = VD->getDefinition(CGM.getContext());
1719 if (VD && ThreadPrivateWithDefinition.insert(CGM.getMangledName(VD)).second) {
1720 QualType ASTTy = VD->getType();
1721
1722 llvm::Value *Ctor = nullptr, *CopyCtor = nullptr, *Dtor = nullptr;
1723 const Expr *Init = VD->getAnyInitializer();
1724 if (CGM.getLangOpts().CPlusPlus && PerformInit) {
1725 // Generate function that re-emits the declaration's initializer into the
1726 // threadprivate copy of the variable VD
1727 CodeGenFunction CtorCGF(CGM);
1728 FunctionArgList Args;
1729 ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc,
1730 /*Id=*/nullptr, CGM.getContext().VoidPtrTy,
1731 ImplicitParamDecl::Other);
1732 Args.push_back(&Dst);
1733
1734 const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(
1735 CGM.getContext().VoidPtrTy, Args);
1736 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI);
1737 std::string Name = getName({"__kmpc_global_ctor_", ""});
1738 llvm::Function *Fn =
1739 CGM.CreateGlobalInitOrCleanUpFunction(FTy, Name, FI, Loc);
1740 CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidPtrTy, Fn, FI,
1741 Args, Loc, Loc);
1742 llvm::Value *ArgVal = CtorCGF.EmitLoadOfScalar(
1743 CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false,
1744 CGM.getContext().VoidPtrTy, Dst.getLocation());
1745 Address Arg(ArgVal, CtorCGF.Int8Ty, VDAddr.getAlignment());
1746 Arg = CtorCGF.Builder.CreateElementBitCast(
1747 Arg, CtorCGF.ConvertTypeForMem(ASTTy));
1748 CtorCGF.EmitAnyExprToMem(Init, Arg, Init->getType().getQualifiers(),
1749 /*IsInitializer=*/true);
1750 ArgVal = CtorCGF.EmitLoadOfScalar(
1751 CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false,
1752 CGM.getContext().VoidPtrTy, Dst.getLocation());
1753 CtorCGF.Builder.CreateStore(ArgVal, CtorCGF.ReturnValue);
1754 CtorCGF.FinishFunction();
1755 Ctor = Fn;
1756 }
1757 if (VD->getType().isDestructedType() != QualType::DK_none) {
1758 // Generate function that emits destructor call for the threadprivate copy
1759 // of the variable VD
1760 CodeGenFunction DtorCGF(CGM);
1761 FunctionArgList Args;
1762 ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc,
1763 /*Id=*/nullptr, CGM.getContext().VoidPtrTy,
1764 ImplicitParamDecl::Other);
1765 Args.push_back(&Dst);
1766
1767 const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(
1768 CGM.getContext().VoidTy, Args);
1769 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI);
1770 std::string Name = getName({"__kmpc_global_dtor_", ""});
1771 llvm::Function *Fn =
1772 CGM.CreateGlobalInitOrCleanUpFunction(FTy, Name, FI, Loc);
1773 auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF);
1774 DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, Args,
1775 Loc, Loc);
1776 // Create a scope with an artificial location for the body of this function.
1777 auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF);
1778 llvm::Value *ArgVal = DtorCGF.EmitLoadOfScalar(
1779 DtorCGF.GetAddrOfLocalVar(&Dst),
1780 /*Volatile=*/false, CGM.getContext().VoidPtrTy, Dst.getLocation());
1781 DtorCGF.emitDestroy(
1782 Address(ArgVal, DtorCGF.Int8Ty, VDAddr.getAlignment()), ASTTy,
1783 DtorCGF.getDestroyer(ASTTy.isDestructedType()),
1784 DtorCGF.needsEHCleanup(ASTTy.isDestructedType()));
1785 DtorCGF.FinishFunction();
1786 Dtor = Fn;
1787 }
1788 // Do not emit init function if it is not required.
1789 if (!Ctor && !Dtor)
1790 return nullptr;
1791
1792 llvm::Type *CopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy};
1793 auto *CopyCtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CopyCtorTyArgs,
1794 /*isVarArg=*/false)
1795 ->getPointerTo();
1796 // Copying constructor for the threadprivate variable.
1797 // Must be NULL - reserved by runtime, but currently it requires that this
1798 // parameter is always NULL. Otherwise it fires assertion.
1799 CopyCtor = llvm::Constant::getNullValue(CopyCtorTy);
1800 if (Ctor == nullptr) {
1801 auto *CtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy,
1802 /*isVarArg=*/false)
1803 ->getPointerTo();
1804 Ctor = llvm::Constant::getNullValue(CtorTy);
1805 }
1806 if (Dtor == nullptr) {
1807 auto *DtorTy = llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy,
1808 /*isVarArg=*/false)
1809 ->getPointerTo();
1810 Dtor = llvm::Constant::getNullValue(DtorTy);
1811 }
1812 if (!CGF) {
1813 auto *InitFunctionTy =
1814 llvm::FunctionType::get(CGM.VoidTy, /*isVarArg*/ false);
1815 std::string Name = getName({"__omp_threadprivate_init_", ""});
1816 llvm::Function *InitFunction = CGM.CreateGlobalInitOrCleanUpFunction(
1817 InitFunctionTy, Name, CGM.getTypes().arrangeNullaryFunction());
1818 CodeGenFunction InitCGF(CGM);
1819 FunctionArgList ArgList;
1820 InitCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, InitFunction,
1821 CGM.getTypes().arrangeNullaryFunction(), ArgList,
1822 Loc, Loc);
1823 emitThreadPrivateVarInit(InitCGF, VDAddr, Ctor, CopyCtor, Dtor, Loc);
1824 InitCGF.FinishFunction();
1825 return InitFunction;
1826 }
1827 emitThreadPrivateVarInit(*CGF, VDAddr, Ctor, CopyCtor, Dtor, Loc);
1828 }
1829 return nullptr;
1830 }
1831
emitDeclareTargetVarDefinition(const VarDecl * VD,llvm::GlobalVariable * Addr,bool PerformInit)1832 bool CGOpenMPRuntime::emitDeclareTargetVarDefinition(const VarDecl *VD,
1833 llvm::GlobalVariable *Addr,
1834 bool PerformInit) {
1835 if (CGM.getLangOpts().OMPTargetTriples.empty() &&
1836 !CGM.getLangOpts().OpenMPIsDevice)
1837 return false;
1838 std::optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res =
1839 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD);
1840 if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link ||
1841 ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
1842 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
1843 HasRequiresUnifiedSharedMemory))
1844 return CGM.getLangOpts().OpenMPIsDevice;
1845 VD = VD->getDefinition(CGM.getContext());
1846 assert(VD && "Unknown VarDecl");
1847
1848 if (!DeclareTargetWithDefinition.insert(CGM.getMangledName(VD)).second)
1849 return CGM.getLangOpts().OpenMPIsDevice;
1850
1851 QualType ASTTy = VD->getType();
1852 SourceLocation Loc = VD->getCanonicalDecl()->getBeginLoc();
1853
1854 // Produce the unique prefix to identify the new target regions. We use
1855 // the source location of the variable declaration which we know to not
1856 // conflict with any target region.
1857 auto EntryInfo =
1858 getTargetEntryUniqueInfo(CGM.getContext(), Loc, VD->getName());
1859 SmallString<128> Buffer, Out;
1860 OffloadEntriesInfoManager.getTargetRegionEntryFnName(Buffer, EntryInfo);
1861
1862 const Expr *Init = VD->getAnyInitializer();
1863 if (CGM.getLangOpts().CPlusPlus && PerformInit) {
1864 llvm::Constant *Ctor;
1865 llvm::Constant *ID;
1866 if (CGM.getLangOpts().OpenMPIsDevice) {
1867 // Generate function that re-emits the declaration's initializer into
1868 // the threadprivate copy of the variable VD
1869 CodeGenFunction CtorCGF(CGM);
1870
1871 const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction();
1872 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI);
1873 llvm::Function *Fn = CGM.CreateGlobalInitOrCleanUpFunction(
1874 FTy, Twine(Buffer, "_ctor"), FI, Loc, false,
1875 llvm::GlobalValue::WeakODRLinkage);
1876 Fn->setVisibility(llvm::GlobalValue::ProtectedVisibility);
1877 if (CGM.getTriple().isAMDGCN())
1878 Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL);
1879 auto NL = ApplyDebugLocation::CreateEmpty(CtorCGF);
1880 CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI,
1881 FunctionArgList(), Loc, Loc);
1882 auto AL = ApplyDebugLocation::CreateArtificial(CtorCGF);
1883 llvm::Constant *AddrInAS0 = Addr;
1884 if (Addr->getAddressSpace() != 0)
1885 AddrInAS0 = llvm::ConstantExpr::getAddrSpaceCast(
1886 Addr, llvm::PointerType::getWithSamePointeeType(
1887 cast<llvm::PointerType>(Addr->getType()), 0));
1888 CtorCGF.EmitAnyExprToMem(Init,
1889 Address(AddrInAS0, Addr->getValueType(),
1890 CGM.getContext().getDeclAlign(VD)),
1891 Init->getType().getQualifiers(),
1892 /*IsInitializer=*/true);
1893 CtorCGF.FinishFunction();
1894 Ctor = Fn;
1895 ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy);
1896 } else {
1897 Ctor = new llvm::GlobalVariable(
1898 CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
1899 llvm::GlobalValue::PrivateLinkage,
1900 llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_ctor"));
1901 ID = Ctor;
1902 }
1903
1904 // Register the information for the entry associated with the constructor.
1905 Out.clear();
1906 auto CtorEntryInfo = EntryInfo;
1907 CtorEntryInfo.ParentName = Twine(Buffer, "_ctor").toStringRef(Out);
1908 OffloadEntriesInfoManager.registerTargetRegionEntryInfo(
1909 CtorEntryInfo, Ctor, ID,
1910 llvm::OffloadEntriesInfoManager::OMPTargetRegionEntryCtor);
1911 }
1912 if (VD->getType().isDestructedType() != QualType::DK_none) {
1913 llvm::Constant *Dtor;
1914 llvm::Constant *ID;
1915 if (CGM.getLangOpts().OpenMPIsDevice) {
1916 // Generate function that emits destructor call for the threadprivate
1917 // copy of the variable VD
1918 CodeGenFunction DtorCGF(CGM);
1919
1920 const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction();
1921 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI);
1922 llvm::Function *Fn = CGM.CreateGlobalInitOrCleanUpFunction(
1923 FTy, Twine(Buffer, "_dtor"), FI, Loc, false,
1924 llvm::GlobalValue::WeakODRLinkage);
1925 Fn->setVisibility(llvm::GlobalValue::ProtectedVisibility);
1926 if (CGM.getTriple().isAMDGCN())
1927 Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL);
1928 auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF);
1929 DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI,
1930 FunctionArgList(), Loc, Loc);
1931 // Create a scope with an artificial location for the body of this
1932 // function.
1933 auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF);
1934 llvm::Constant *AddrInAS0 = Addr;
1935 if (Addr->getAddressSpace() != 0)
1936 AddrInAS0 = llvm::ConstantExpr::getAddrSpaceCast(
1937 Addr, llvm::PointerType::getWithSamePointeeType(
1938 cast<llvm::PointerType>(Addr->getType()), 0));
1939 DtorCGF.emitDestroy(Address(AddrInAS0, Addr->getValueType(),
1940 CGM.getContext().getDeclAlign(VD)),
1941 ASTTy, DtorCGF.getDestroyer(ASTTy.isDestructedType()),
1942 DtorCGF.needsEHCleanup(ASTTy.isDestructedType()));
1943 DtorCGF.FinishFunction();
1944 Dtor = Fn;
1945 ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy);
1946 } else {
1947 Dtor = new llvm::GlobalVariable(
1948 CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
1949 llvm::GlobalValue::PrivateLinkage,
1950 llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_dtor"));
1951 ID = Dtor;
1952 }
1953 // Register the information for the entry associated with the destructor.
1954 Out.clear();
1955 auto DtorEntryInfo = EntryInfo;
1956 DtorEntryInfo.ParentName = Twine(Buffer, "_dtor").toStringRef(Out);
1957 OffloadEntriesInfoManager.registerTargetRegionEntryInfo(
1958 DtorEntryInfo, Dtor, ID,
1959 llvm::OffloadEntriesInfoManager::OMPTargetRegionEntryDtor);
1960 }
1961 return CGM.getLangOpts().OpenMPIsDevice;
1962 }
1963
getAddrOfArtificialThreadPrivate(CodeGenFunction & CGF,QualType VarType,StringRef Name)1964 Address CGOpenMPRuntime::getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF,
1965 QualType VarType,
1966 StringRef Name) {
1967 std::string Suffix = getName({"artificial", ""});
1968 llvm::Type *VarLVType = CGF.ConvertTypeForMem(VarType);
1969 llvm::GlobalVariable *GAddr = OMPBuilder.getOrCreateInternalVariable(
1970 VarLVType, Twine(Name).concat(Suffix).str());
1971 if (CGM.getLangOpts().OpenMP && CGM.getLangOpts().OpenMPUseTLS &&
1972 CGM.getTarget().isTLSSupported()) {
1973 GAddr->setThreadLocal(/*Val=*/true);
1974 return Address(GAddr, GAddr->getValueType(),
1975 CGM.getContext().getTypeAlignInChars(VarType));
1976 }
1977 std::string CacheSuffix = getName({"cache", ""});
1978 llvm::Value *Args[] = {
1979 emitUpdateLocation(CGF, SourceLocation()),
1980 getThreadID(CGF, SourceLocation()),
1981 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(GAddr, CGM.VoidPtrTy),
1982 CGF.Builder.CreateIntCast(CGF.getTypeSize(VarType), CGM.SizeTy,
1983 /*isSigned=*/false),
1984 OMPBuilder.getOrCreateInternalVariable(
1985 CGM.VoidPtrPtrTy,
1986 Twine(Name).concat(Suffix).concat(CacheSuffix).str())};
1987 return Address(
1988 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
1989 CGF.EmitRuntimeCall(
1990 OMPBuilder.getOrCreateRuntimeFunction(
1991 CGM.getModule(), OMPRTL___kmpc_threadprivate_cached),
1992 Args),
1993 VarLVType->getPointerTo(/*AddrSpace=*/0)),
1994 VarLVType, CGM.getContext().getTypeAlignInChars(VarType));
1995 }
1996
emitIfClause(CodeGenFunction & CGF,const Expr * Cond,const RegionCodeGenTy & ThenGen,const RegionCodeGenTy & ElseGen)1997 void CGOpenMPRuntime::emitIfClause(CodeGenFunction &CGF, const Expr *Cond,
1998 const RegionCodeGenTy &ThenGen,
1999 const RegionCodeGenTy &ElseGen) {
2000 CodeGenFunction::LexicalScope ConditionScope(CGF, Cond->getSourceRange());
2001
2002 // If the condition constant folds and can be elided, try to avoid emitting
2003 // the condition and the dead arm of the if/else.
2004 bool CondConstant;
2005 if (CGF.ConstantFoldsToSimpleInteger(Cond, CondConstant)) {
2006 if (CondConstant)
2007 ThenGen(CGF);
2008 else
2009 ElseGen(CGF);
2010 return;
2011 }
2012
2013 // Otherwise, the condition did not fold, or we couldn't elide it. Just
2014 // emit the conditional branch.
2015 llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("omp_if.then");
2016 llvm::BasicBlock *ElseBlock = CGF.createBasicBlock("omp_if.else");
2017 llvm::BasicBlock *ContBlock = CGF.createBasicBlock("omp_if.end");
2018 CGF.EmitBranchOnBoolExpr(Cond, ThenBlock, ElseBlock, /*TrueCount=*/0);
2019
2020 // Emit the 'then' code.
2021 CGF.EmitBlock(ThenBlock);
2022 ThenGen(CGF);
2023 CGF.EmitBranch(ContBlock);
2024 // Emit the 'else' code if present.
2025 // There is no need to emit line number for unconditional branch.
2026 (void)ApplyDebugLocation::CreateEmpty(CGF);
2027 CGF.EmitBlock(ElseBlock);
2028 ElseGen(CGF);
2029 // There is no need to emit line number for unconditional branch.
2030 (void)ApplyDebugLocation::CreateEmpty(CGF);
2031 CGF.EmitBranch(ContBlock);
2032 // Emit the continuation block for code after the if.
2033 CGF.EmitBlock(ContBlock, /*IsFinished=*/true);
2034 }
2035
emitParallelCall(CodeGenFunction & CGF,SourceLocation Loc,llvm::Function * OutlinedFn,ArrayRef<llvm::Value * > CapturedVars,const Expr * IfCond,llvm::Value * NumThreads)2036 void CGOpenMPRuntime::emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc,
2037 llvm::Function *OutlinedFn,
2038 ArrayRef<llvm::Value *> CapturedVars,
2039 const Expr *IfCond,
2040 llvm::Value *NumThreads) {
2041 if (!CGF.HaveInsertPoint())
2042 return;
2043 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2044 auto &M = CGM.getModule();
2045 auto &&ThenGen = [&M, OutlinedFn, CapturedVars, RTLoc,
2046 this](CodeGenFunction &CGF, PrePostActionTy &) {
2047 // Build call __kmpc_fork_call(loc, n, microtask, var1, .., varn);
2048 CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime();
2049 llvm::Value *Args[] = {
2050 RTLoc,
2051 CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars
2052 CGF.Builder.CreateBitCast(OutlinedFn, RT.getKmpc_MicroPointerTy())};
2053 llvm::SmallVector<llvm::Value *, 16> RealArgs;
2054 RealArgs.append(std::begin(Args), std::end(Args));
2055 RealArgs.append(CapturedVars.begin(), CapturedVars.end());
2056
2057 llvm::FunctionCallee RTLFn =
2058 OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_fork_call);
2059 CGF.EmitRuntimeCall(RTLFn, RealArgs);
2060 };
2061 auto &&ElseGen = [&M, OutlinedFn, CapturedVars, RTLoc, Loc,
2062 this](CodeGenFunction &CGF, PrePostActionTy &) {
2063 CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime();
2064 llvm::Value *ThreadID = RT.getThreadID(CGF, Loc);
2065 // Build calls:
2066 // __kmpc_serialized_parallel(&Loc, GTid);
2067 llvm::Value *Args[] = {RTLoc, ThreadID};
2068 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2069 M, OMPRTL___kmpc_serialized_parallel),
2070 Args);
2071
2072 // OutlinedFn(>id, &zero_bound, CapturedStruct);
2073 Address ThreadIDAddr = RT.emitThreadIDAddress(CGF, Loc);
2074 Address ZeroAddrBound =
2075 CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
2076 /*Name=*/".bound.zero.addr");
2077 CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddrBound);
2078 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
2079 // ThreadId for serialized parallels is 0.
2080 OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
2081 OutlinedFnArgs.push_back(ZeroAddrBound.getPointer());
2082 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
2083
2084 // Ensure we do not inline the function. This is trivially true for the ones
2085 // passed to __kmpc_fork_call but the ones called in serialized regions
2086 // could be inlined. This is not a perfect but it is closer to the invariant
2087 // we want, namely, every data environment starts with a new function.
2088 // TODO: We should pass the if condition to the runtime function and do the
2089 // handling there. Much cleaner code.
2090 OutlinedFn->removeFnAttr(llvm::Attribute::AlwaysInline);
2091 OutlinedFn->addFnAttr(llvm::Attribute::NoInline);
2092 RT.emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
2093
2094 // __kmpc_end_serialized_parallel(&Loc, GTid);
2095 llvm::Value *EndArgs[] = {RT.emitUpdateLocation(CGF, Loc), ThreadID};
2096 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2097 M, OMPRTL___kmpc_end_serialized_parallel),
2098 EndArgs);
2099 };
2100 if (IfCond) {
2101 emitIfClause(CGF, IfCond, ThenGen, ElseGen);
2102 } else {
2103 RegionCodeGenTy ThenRCG(ThenGen);
2104 ThenRCG(CGF);
2105 }
2106 }
2107
2108 // If we're inside an (outlined) parallel region, use the region info's
2109 // thread-ID variable (it is passed in a first argument of the outlined function
2110 // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in
2111 // regular serial code region, get thread ID by calling kmp_int32
2112 // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and
2113 // return the address of that temp.
emitThreadIDAddress(CodeGenFunction & CGF,SourceLocation Loc)2114 Address CGOpenMPRuntime::emitThreadIDAddress(CodeGenFunction &CGF,
2115 SourceLocation Loc) {
2116 if (auto *OMPRegionInfo =
2117 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
2118 if (OMPRegionInfo->getThreadIDVariable())
2119 return OMPRegionInfo->getThreadIDVariableLValue(CGF).getAddress(CGF);
2120
2121 llvm::Value *ThreadID = getThreadID(CGF, Loc);
2122 QualType Int32Ty =
2123 CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true);
2124 Address ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp.");
2125 CGF.EmitStoreOfScalar(ThreadID,
2126 CGF.MakeAddrLValue(ThreadIDTemp, Int32Ty));
2127
2128 return ThreadIDTemp;
2129 }
2130
getCriticalRegionLock(StringRef CriticalName)2131 llvm::Value *CGOpenMPRuntime::getCriticalRegionLock(StringRef CriticalName) {
2132 std::string Prefix = Twine("gomp_critical_user_", CriticalName).str();
2133 std::string Name = getName({Prefix, "var"});
2134 return OMPBuilder.getOrCreateInternalVariable(KmpCriticalNameTy, Name);
2135 }
2136
2137 namespace {
2138 /// Common pre(post)-action for different OpenMP constructs.
2139 class CommonActionTy final : public PrePostActionTy {
2140 llvm::FunctionCallee EnterCallee;
2141 ArrayRef<llvm::Value *> EnterArgs;
2142 llvm::FunctionCallee ExitCallee;
2143 ArrayRef<llvm::Value *> ExitArgs;
2144 bool Conditional;
2145 llvm::BasicBlock *ContBlock = nullptr;
2146
2147 public:
CommonActionTy(llvm::FunctionCallee EnterCallee,ArrayRef<llvm::Value * > EnterArgs,llvm::FunctionCallee ExitCallee,ArrayRef<llvm::Value * > ExitArgs,bool Conditional=false)2148 CommonActionTy(llvm::FunctionCallee EnterCallee,
2149 ArrayRef<llvm::Value *> EnterArgs,
2150 llvm::FunctionCallee ExitCallee,
2151 ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false)
2152 : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
2153 ExitArgs(ExitArgs), Conditional(Conditional) {}
Enter(CodeGenFunction & CGF)2154 void Enter(CodeGenFunction &CGF) override {
2155 llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
2156 if (Conditional) {
2157 llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
2158 auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
2159 ContBlock = CGF.createBasicBlock("omp_if.end");
2160 // Generate the branch (If-stmt)
2161 CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
2162 CGF.EmitBlock(ThenBlock);
2163 }
2164 }
Done(CodeGenFunction & CGF)2165 void Done(CodeGenFunction &CGF) {
2166 // Emit the rest of blocks/branches
2167 CGF.EmitBranch(ContBlock);
2168 CGF.EmitBlock(ContBlock, true);
2169 }
Exit(CodeGenFunction & CGF)2170 void Exit(CodeGenFunction &CGF) override {
2171 CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
2172 }
2173 };
2174 } // anonymous namespace
2175
emitCriticalRegion(CodeGenFunction & CGF,StringRef CriticalName,const RegionCodeGenTy & CriticalOpGen,SourceLocation Loc,const Expr * Hint)2176 void CGOpenMPRuntime::emitCriticalRegion(CodeGenFunction &CGF,
2177 StringRef CriticalName,
2178 const RegionCodeGenTy &CriticalOpGen,
2179 SourceLocation Loc, const Expr *Hint) {
2180 // __kmpc_critical[_with_hint](ident_t *, gtid, Lock[, hint]);
2181 // CriticalOpGen();
2182 // __kmpc_end_critical(ident_t *, gtid, Lock);
2183 // Prepare arguments and build a call to __kmpc_critical
2184 if (!CGF.HaveInsertPoint())
2185 return;
2186 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2187 getCriticalRegionLock(CriticalName)};
2188 llvm::SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args),
2189 std::end(Args));
2190 if (Hint) {
2191 EnterArgs.push_back(CGF.Builder.CreateIntCast(
2192 CGF.EmitScalarExpr(Hint), CGM.Int32Ty, /*isSigned=*/false));
2193 }
2194 CommonActionTy Action(
2195 OMPBuilder.getOrCreateRuntimeFunction(
2196 CGM.getModule(),
2197 Hint ? OMPRTL___kmpc_critical_with_hint : OMPRTL___kmpc_critical),
2198 EnterArgs,
2199 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
2200 OMPRTL___kmpc_end_critical),
2201 Args);
2202 CriticalOpGen.setAction(Action);
2203 emitInlinedDirective(CGF, OMPD_critical, CriticalOpGen);
2204 }
2205
emitMasterRegion(CodeGenFunction & CGF,const RegionCodeGenTy & MasterOpGen,SourceLocation Loc)2206 void CGOpenMPRuntime::emitMasterRegion(CodeGenFunction &CGF,
2207 const RegionCodeGenTy &MasterOpGen,
2208 SourceLocation Loc) {
2209 if (!CGF.HaveInsertPoint())
2210 return;
2211 // if(__kmpc_master(ident_t *, gtid)) {
2212 // MasterOpGen();
2213 // __kmpc_end_master(ident_t *, gtid);
2214 // }
2215 // Prepare arguments and build a call to __kmpc_master
2216 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2217 CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction(
2218 CGM.getModule(), OMPRTL___kmpc_master),
2219 Args,
2220 OMPBuilder.getOrCreateRuntimeFunction(
2221 CGM.getModule(), OMPRTL___kmpc_end_master),
2222 Args,
2223 /*Conditional=*/true);
2224 MasterOpGen.setAction(Action);
2225 emitInlinedDirective(CGF, OMPD_master, MasterOpGen);
2226 Action.Done(CGF);
2227 }
2228
emitMaskedRegion(CodeGenFunction & CGF,const RegionCodeGenTy & MaskedOpGen,SourceLocation Loc,const Expr * Filter)2229 void CGOpenMPRuntime::emitMaskedRegion(CodeGenFunction &CGF,
2230 const RegionCodeGenTy &MaskedOpGen,
2231 SourceLocation Loc, const Expr *Filter) {
2232 if (!CGF.HaveInsertPoint())
2233 return;
2234 // if(__kmpc_masked(ident_t *, gtid, filter)) {
2235 // MaskedOpGen();
2236 // __kmpc_end_masked(iden_t *, gtid);
2237 // }
2238 // Prepare arguments and build a call to __kmpc_masked
2239 llvm::Value *FilterVal = Filter
2240 ? CGF.EmitScalarExpr(Filter, CGF.Int32Ty)
2241 : llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/0);
2242 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2243 FilterVal};
2244 llvm::Value *ArgsEnd[] = {emitUpdateLocation(CGF, Loc),
2245 getThreadID(CGF, Loc)};
2246 CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction(
2247 CGM.getModule(), OMPRTL___kmpc_masked),
2248 Args,
2249 OMPBuilder.getOrCreateRuntimeFunction(
2250 CGM.getModule(), OMPRTL___kmpc_end_masked),
2251 ArgsEnd,
2252 /*Conditional=*/true);
2253 MaskedOpGen.setAction(Action);
2254 emitInlinedDirective(CGF, OMPD_masked, MaskedOpGen);
2255 Action.Done(CGF);
2256 }
2257
emitTaskyieldCall(CodeGenFunction & CGF,SourceLocation Loc)2258 void CGOpenMPRuntime::emitTaskyieldCall(CodeGenFunction &CGF,
2259 SourceLocation Loc) {
2260 if (!CGF.HaveInsertPoint())
2261 return;
2262 if (CGF.CGM.getLangOpts().OpenMPIRBuilder) {
2263 OMPBuilder.createTaskyield(CGF.Builder);
2264 } else {
2265 // Build call __kmpc_omp_taskyield(loc, thread_id, 0);
2266 llvm::Value *Args[] = {
2267 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2268 llvm::ConstantInt::get(CGM.IntTy, /*V=*/0, /*isSigned=*/true)};
2269 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2270 CGM.getModule(), OMPRTL___kmpc_omp_taskyield),
2271 Args);
2272 }
2273
2274 if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
2275 Region->emitUntiedSwitch(CGF);
2276 }
2277
emitTaskgroupRegion(CodeGenFunction & CGF,const RegionCodeGenTy & TaskgroupOpGen,SourceLocation Loc)2278 void CGOpenMPRuntime::emitTaskgroupRegion(CodeGenFunction &CGF,
2279 const RegionCodeGenTy &TaskgroupOpGen,
2280 SourceLocation Loc) {
2281 if (!CGF.HaveInsertPoint())
2282 return;
2283 // __kmpc_taskgroup(ident_t *, gtid);
2284 // TaskgroupOpGen();
2285 // __kmpc_end_taskgroup(ident_t *, gtid);
2286 // Prepare arguments and build a call to __kmpc_taskgroup
2287 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2288 CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction(
2289 CGM.getModule(), OMPRTL___kmpc_taskgroup),
2290 Args,
2291 OMPBuilder.getOrCreateRuntimeFunction(
2292 CGM.getModule(), OMPRTL___kmpc_end_taskgroup),
2293 Args);
2294 TaskgroupOpGen.setAction(Action);
2295 emitInlinedDirective(CGF, OMPD_taskgroup, TaskgroupOpGen);
2296 }
2297
2298 /// Given an array of pointers to variables, project the address of a
2299 /// given variable.
emitAddrOfVarFromArray(CodeGenFunction & CGF,Address Array,unsigned Index,const VarDecl * Var)2300 static Address emitAddrOfVarFromArray(CodeGenFunction &CGF, Address Array,
2301 unsigned Index, const VarDecl *Var) {
2302 // Pull out the pointer to the variable.
2303 Address PtrAddr = CGF.Builder.CreateConstArrayGEP(Array, Index);
2304 llvm::Value *Ptr = CGF.Builder.CreateLoad(PtrAddr);
2305
2306 llvm::Type *ElemTy = CGF.ConvertTypeForMem(Var->getType());
2307 return Address(
2308 CGF.Builder.CreateBitCast(
2309 Ptr, ElemTy->getPointerTo(Ptr->getType()->getPointerAddressSpace())),
2310 ElemTy, CGF.getContext().getDeclAlign(Var));
2311 }
2312
emitCopyprivateCopyFunction(CodeGenModule & CGM,llvm::Type * ArgsElemType,ArrayRef<const Expr * > CopyprivateVars,ArrayRef<const Expr * > DestExprs,ArrayRef<const Expr * > SrcExprs,ArrayRef<const Expr * > AssignmentOps,SourceLocation Loc)2313 static llvm::Value *emitCopyprivateCopyFunction(
2314 CodeGenModule &CGM, llvm::Type *ArgsElemType,
2315 ArrayRef<const Expr *> CopyprivateVars, ArrayRef<const Expr *> DestExprs,
2316 ArrayRef<const Expr *> SrcExprs, ArrayRef<const Expr *> AssignmentOps,
2317 SourceLocation Loc) {
2318 ASTContext &C = CGM.getContext();
2319 // void copy_func(void *LHSArg, void *RHSArg);
2320 FunctionArgList Args;
2321 ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
2322 ImplicitParamDecl::Other);
2323 ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
2324 ImplicitParamDecl::Other);
2325 Args.push_back(&LHSArg);
2326 Args.push_back(&RHSArg);
2327 const auto &CGFI =
2328 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2329 std::string Name =
2330 CGM.getOpenMPRuntime().getName({"omp", "copyprivate", "copy_func"});
2331 auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
2332 llvm::GlobalValue::InternalLinkage, Name,
2333 &CGM.getModule());
2334 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2335 Fn->setDoesNotRecurse();
2336 CodeGenFunction CGF(CGM);
2337 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2338 // Dest = (void*[n])(LHSArg);
2339 // Src = (void*[n])(RHSArg);
2340 Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2341 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)),
2342 ArgsElemType->getPointerTo()),
2343 ArgsElemType, CGF.getPointerAlign());
2344 Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2345 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)),
2346 ArgsElemType->getPointerTo()),
2347 ArgsElemType, CGF.getPointerAlign());
2348 // *(Type0*)Dst[0] = *(Type0*)Src[0];
2349 // *(Type1*)Dst[1] = *(Type1*)Src[1];
2350 // ...
2351 // *(Typen*)Dst[n] = *(Typen*)Src[n];
2352 for (unsigned I = 0, E = AssignmentOps.size(); I < E; ++I) {
2353 const auto *DestVar =
2354 cast<VarDecl>(cast<DeclRefExpr>(DestExprs[I])->getDecl());
2355 Address DestAddr = emitAddrOfVarFromArray(CGF, LHS, I, DestVar);
2356
2357 const auto *SrcVar =
2358 cast<VarDecl>(cast<DeclRefExpr>(SrcExprs[I])->getDecl());
2359 Address SrcAddr = emitAddrOfVarFromArray(CGF, RHS, I, SrcVar);
2360
2361 const auto *VD = cast<DeclRefExpr>(CopyprivateVars[I])->getDecl();
2362 QualType Type = VD->getType();
2363 CGF.EmitOMPCopy(Type, DestAddr, SrcAddr, DestVar, SrcVar, AssignmentOps[I]);
2364 }
2365 CGF.FinishFunction();
2366 return Fn;
2367 }
2368
emitSingleRegion(CodeGenFunction & CGF,const RegionCodeGenTy & SingleOpGen,SourceLocation Loc,ArrayRef<const Expr * > CopyprivateVars,ArrayRef<const Expr * > SrcExprs,ArrayRef<const Expr * > DstExprs,ArrayRef<const Expr * > AssignmentOps)2369 void CGOpenMPRuntime::emitSingleRegion(CodeGenFunction &CGF,
2370 const RegionCodeGenTy &SingleOpGen,
2371 SourceLocation Loc,
2372 ArrayRef<const Expr *> CopyprivateVars,
2373 ArrayRef<const Expr *> SrcExprs,
2374 ArrayRef<const Expr *> DstExprs,
2375 ArrayRef<const Expr *> AssignmentOps) {
2376 if (!CGF.HaveInsertPoint())
2377 return;
2378 assert(CopyprivateVars.size() == SrcExprs.size() &&
2379 CopyprivateVars.size() == DstExprs.size() &&
2380 CopyprivateVars.size() == AssignmentOps.size());
2381 ASTContext &C = CGM.getContext();
2382 // int32 did_it = 0;
2383 // if(__kmpc_single(ident_t *, gtid)) {
2384 // SingleOpGen();
2385 // __kmpc_end_single(ident_t *, gtid);
2386 // did_it = 1;
2387 // }
2388 // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>,
2389 // <copy_func>, did_it);
2390
2391 Address DidIt = Address::invalid();
2392 if (!CopyprivateVars.empty()) {
2393 // int32 did_it = 0;
2394 QualType KmpInt32Ty =
2395 C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
2396 DidIt = CGF.CreateMemTemp(KmpInt32Ty, ".omp.copyprivate.did_it");
2397 CGF.Builder.CreateStore(CGF.Builder.getInt32(0), DidIt);
2398 }
2399 // Prepare arguments and build a call to __kmpc_single
2400 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2401 CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction(
2402 CGM.getModule(), OMPRTL___kmpc_single),
2403 Args,
2404 OMPBuilder.getOrCreateRuntimeFunction(
2405 CGM.getModule(), OMPRTL___kmpc_end_single),
2406 Args,
2407 /*Conditional=*/true);
2408 SingleOpGen.setAction(Action);
2409 emitInlinedDirective(CGF, OMPD_single, SingleOpGen);
2410 if (DidIt.isValid()) {
2411 // did_it = 1;
2412 CGF.Builder.CreateStore(CGF.Builder.getInt32(1), DidIt);
2413 }
2414 Action.Done(CGF);
2415 // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>,
2416 // <copy_func>, did_it);
2417 if (DidIt.isValid()) {
2418 llvm::APInt ArraySize(/*unsigned int numBits=*/32, CopyprivateVars.size());
2419 QualType CopyprivateArrayTy = C.getConstantArrayType(
2420 C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
2421 /*IndexTypeQuals=*/0);
2422 // Create a list of all private variables for copyprivate.
2423 Address CopyprivateList =
2424 CGF.CreateMemTemp(CopyprivateArrayTy, ".omp.copyprivate.cpr_list");
2425 for (unsigned I = 0, E = CopyprivateVars.size(); I < E; ++I) {
2426 Address Elem = CGF.Builder.CreateConstArrayGEP(CopyprivateList, I);
2427 CGF.Builder.CreateStore(
2428 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2429 CGF.EmitLValue(CopyprivateVars[I]).getPointer(CGF),
2430 CGF.VoidPtrTy),
2431 Elem);
2432 }
2433 // Build function that copies private values from single region to all other
2434 // threads in the corresponding parallel region.
2435 llvm::Value *CpyFn = emitCopyprivateCopyFunction(
2436 CGM, CGF.ConvertTypeForMem(CopyprivateArrayTy), CopyprivateVars,
2437 SrcExprs, DstExprs, AssignmentOps, Loc);
2438 llvm::Value *BufSize = CGF.getTypeSize(CopyprivateArrayTy);
2439 Address CL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2440 CopyprivateList, CGF.VoidPtrTy, CGF.Int8Ty);
2441 llvm::Value *DidItVal = CGF.Builder.CreateLoad(DidIt);
2442 llvm::Value *Args[] = {
2443 emitUpdateLocation(CGF, Loc), // ident_t *<loc>
2444 getThreadID(CGF, Loc), // i32 <gtid>
2445 BufSize, // size_t <buf_size>
2446 CL.getPointer(), // void *<copyprivate list>
2447 CpyFn, // void (*) (void *, void *) <copy_func>
2448 DidItVal // i32 did_it
2449 };
2450 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2451 CGM.getModule(), OMPRTL___kmpc_copyprivate),
2452 Args);
2453 }
2454 }
2455
emitOrderedRegion(CodeGenFunction & CGF,const RegionCodeGenTy & OrderedOpGen,SourceLocation Loc,bool IsThreads)2456 void CGOpenMPRuntime::emitOrderedRegion(CodeGenFunction &CGF,
2457 const RegionCodeGenTy &OrderedOpGen,
2458 SourceLocation Loc, bool IsThreads) {
2459 if (!CGF.HaveInsertPoint())
2460 return;
2461 // __kmpc_ordered(ident_t *, gtid);
2462 // OrderedOpGen();
2463 // __kmpc_end_ordered(ident_t *, gtid);
2464 // Prepare arguments and build a call to __kmpc_ordered
2465 if (IsThreads) {
2466 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2467 CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction(
2468 CGM.getModule(), OMPRTL___kmpc_ordered),
2469 Args,
2470 OMPBuilder.getOrCreateRuntimeFunction(
2471 CGM.getModule(), OMPRTL___kmpc_end_ordered),
2472 Args);
2473 OrderedOpGen.setAction(Action);
2474 emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen);
2475 return;
2476 }
2477 emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen);
2478 }
2479
getDefaultFlagsForBarriers(OpenMPDirectiveKind Kind)2480 unsigned CGOpenMPRuntime::getDefaultFlagsForBarriers(OpenMPDirectiveKind Kind) {
2481 unsigned Flags;
2482 if (Kind == OMPD_for)
2483 Flags = OMP_IDENT_BARRIER_IMPL_FOR;
2484 else if (Kind == OMPD_sections)
2485 Flags = OMP_IDENT_BARRIER_IMPL_SECTIONS;
2486 else if (Kind == OMPD_single)
2487 Flags = OMP_IDENT_BARRIER_IMPL_SINGLE;
2488 else if (Kind == OMPD_barrier)
2489 Flags = OMP_IDENT_BARRIER_EXPL;
2490 else
2491 Flags = OMP_IDENT_BARRIER_IMPL;
2492 return Flags;
2493 }
2494
getDefaultScheduleAndChunk(CodeGenFunction & CGF,const OMPLoopDirective & S,OpenMPScheduleClauseKind & ScheduleKind,const Expr * & ChunkExpr) const2495 void CGOpenMPRuntime::getDefaultScheduleAndChunk(
2496 CodeGenFunction &CGF, const OMPLoopDirective &S,
2497 OpenMPScheduleClauseKind &ScheduleKind, const Expr *&ChunkExpr) const {
2498 // Check if the loop directive is actually a doacross loop directive. In this
2499 // case choose static, 1 schedule.
2500 if (llvm::any_of(
2501 S.getClausesOfKind<OMPOrderedClause>(),
2502 [](const OMPOrderedClause *C) { return C->getNumForLoops(); })) {
2503 ScheduleKind = OMPC_SCHEDULE_static;
2504 // Chunk size is 1 in this case.
2505 llvm::APInt ChunkSize(32, 1);
2506 ChunkExpr = IntegerLiteral::Create(
2507 CGF.getContext(), ChunkSize,
2508 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
2509 SourceLocation());
2510 }
2511 }
2512
emitBarrierCall(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind Kind,bool EmitChecks,bool ForceSimpleCall)2513 void CGOpenMPRuntime::emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc,
2514 OpenMPDirectiveKind Kind, bool EmitChecks,
2515 bool ForceSimpleCall) {
2516 // Check if we should use the OMPBuilder
2517 auto *OMPRegionInfo =
2518 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo);
2519 if (CGF.CGM.getLangOpts().OpenMPIRBuilder) {
2520 CGF.Builder.restoreIP(OMPBuilder.createBarrier(
2521 CGF.Builder, Kind, ForceSimpleCall, EmitChecks));
2522 return;
2523 }
2524
2525 if (!CGF.HaveInsertPoint())
2526 return;
2527 // Build call __kmpc_cancel_barrier(loc, thread_id);
2528 // Build call __kmpc_barrier(loc, thread_id);
2529 unsigned Flags = getDefaultFlagsForBarriers(Kind);
2530 // Build call __kmpc_cancel_barrier(loc, thread_id) or __kmpc_barrier(loc,
2531 // thread_id);
2532 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
2533 getThreadID(CGF, Loc)};
2534 if (OMPRegionInfo) {
2535 if (!ForceSimpleCall && OMPRegionInfo->hasCancel()) {
2536 llvm::Value *Result = CGF.EmitRuntimeCall(
2537 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
2538 OMPRTL___kmpc_cancel_barrier),
2539 Args);
2540 if (EmitChecks) {
2541 // if (__kmpc_cancel_barrier()) {
2542 // exit from construct;
2543 // }
2544 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit");
2545 llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue");
2546 llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result);
2547 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
2548 CGF.EmitBlock(ExitBB);
2549 // exit from construct;
2550 CodeGenFunction::JumpDest CancelDestination =
2551 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
2552 CGF.EmitBranchThroughCleanup(CancelDestination);
2553 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
2554 }
2555 return;
2556 }
2557 }
2558 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2559 CGM.getModule(), OMPRTL___kmpc_barrier),
2560 Args);
2561 }
2562
emitErrorCall(CodeGenFunction & CGF,SourceLocation Loc,Expr * ME,bool IsFatal)2563 void CGOpenMPRuntime::emitErrorCall(CodeGenFunction &CGF, SourceLocation Loc,
2564 Expr *ME, bool IsFatal) {
2565 llvm::Value *MVL =
2566 ME ? CGF.EmitStringLiteralLValue(cast<StringLiteral>(ME)).getPointer(CGF)
2567 : llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
2568 // Build call void __kmpc_error(ident_t *loc, int severity, const char
2569 // *message)
2570 llvm::Value *Args[] = {
2571 emitUpdateLocation(CGF, Loc, /*Flags=*/0, /*GenLoc=*/true),
2572 llvm::ConstantInt::get(CGM.Int32Ty, IsFatal ? 2 : 1),
2573 CGF.Builder.CreatePointerCast(MVL, CGM.Int8PtrTy)};
2574 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2575 CGM.getModule(), OMPRTL___kmpc_error),
2576 Args);
2577 }
2578
2579 /// Map the OpenMP loop schedule to the runtime enumeration.
getRuntimeSchedule(OpenMPScheduleClauseKind ScheduleKind,bool Chunked,bool Ordered)2580 static OpenMPSchedType getRuntimeSchedule(OpenMPScheduleClauseKind ScheduleKind,
2581 bool Chunked, bool Ordered) {
2582 switch (ScheduleKind) {
2583 case OMPC_SCHEDULE_static:
2584 return Chunked ? (Ordered ? OMP_ord_static_chunked : OMP_sch_static_chunked)
2585 : (Ordered ? OMP_ord_static : OMP_sch_static);
2586 case OMPC_SCHEDULE_dynamic:
2587 return Ordered ? OMP_ord_dynamic_chunked : OMP_sch_dynamic_chunked;
2588 case OMPC_SCHEDULE_guided:
2589 return Ordered ? OMP_ord_guided_chunked : OMP_sch_guided_chunked;
2590 case OMPC_SCHEDULE_runtime:
2591 return Ordered ? OMP_ord_runtime : OMP_sch_runtime;
2592 case OMPC_SCHEDULE_auto:
2593 return Ordered ? OMP_ord_auto : OMP_sch_auto;
2594 case OMPC_SCHEDULE_unknown:
2595 assert(!Chunked && "chunk was specified but schedule kind not known");
2596 return Ordered ? OMP_ord_static : OMP_sch_static;
2597 }
2598 llvm_unreachable("Unexpected runtime schedule");
2599 }
2600
2601 /// Map the OpenMP distribute schedule to the runtime enumeration.
2602 static OpenMPSchedType
getRuntimeSchedule(OpenMPDistScheduleClauseKind ScheduleKind,bool Chunked)2603 getRuntimeSchedule(OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) {
2604 // only static is allowed for dist_schedule
2605 return Chunked ? OMP_dist_sch_static_chunked : OMP_dist_sch_static;
2606 }
2607
isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind,bool Chunked) const2608 bool CGOpenMPRuntime::isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind,
2609 bool Chunked) const {
2610 OpenMPSchedType Schedule =
2611 getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false);
2612 return Schedule == OMP_sch_static;
2613 }
2614
isStaticNonchunked(OpenMPDistScheduleClauseKind ScheduleKind,bool Chunked) const2615 bool CGOpenMPRuntime::isStaticNonchunked(
2616 OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const {
2617 OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked);
2618 return Schedule == OMP_dist_sch_static;
2619 }
2620
isStaticChunked(OpenMPScheduleClauseKind ScheduleKind,bool Chunked) const2621 bool CGOpenMPRuntime::isStaticChunked(OpenMPScheduleClauseKind ScheduleKind,
2622 bool Chunked) const {
2623 OpenMPSchedType Schedule =
2624 getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false);
2625 return Schedule == OMP_sch_static_chunked;
2626 }
2627
isStaticChunked(OpenMPDistScheduleClauseKind ScheduleKind,bool Chunked) const2628 bool CGOpenMPRuntime::isStaticChunked(
2629 OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const {
2630 OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked);
2631 return Schedule == OMP_dist_sch_static_chunked;
2632 }
2633
isDynamic(OpenMPScheduleClauseKind ScheduleKind) const2634 bool CGOpenMPRuntime::isDynamic(OpenMPScheduleClauseKind ScheduleKind) const {
2635 OpenMPSchedType Schedule =
2636 getRuntimeSchedule(ScheduleKind, /*Chunked=*/false, /*Ordered=*/false);
2637 assert(Schedule != OMP_sch_static_chunked && "cannot be chunked here");
2638 return Schedule != OMP_sch_static;
2639 }
2640
addMonoNonMonoModifier(CodeGenModule & CGM,OpenMPSchedType Schedule,OpenMPScheduleClauseModifier M1,OpenMPScheduleClauseModifier M2)2641 static int addMonoNonMonoModifier(CodeGenModule &CGM, OpenMPSchedType Schedule,
2642 OpenMPScheduleClauseModifier M1,
2643 OpenMPScheduleClauseModifier M2) {
2644 int Modifier = 0;
2645 switch (M1) {
2646 case OMPC_SCHEDULE_MODIFIER_monotonic:
2647 Modifier = OMP_sch_modifier_monotonic;
2648 break;
2649 case OMPC_SCHEDULE_MODIFIER_nonmonotonic:
2650 Modifier = OMP_sch_modifier_nonmonotonic;
2651 break;
2652 case OMPC_SCHEDULE_MODIFIER_simd:
2653 if (Schedule == OMP_sch_static_chunked)
2654 Schedule = OMP_sch_static_balanced_chunked;
2655 break;
2656 case OMPC_SCHEDULE_MODIFIER_last:
2657 case OMPC_SCHEDULE_MODIFIER_unknown:
2658 break;
2659 }
2660 switch (M2) {
2661 case OMPC_SCHEDULE_MODIFIER_monotonic:
2662 Modifier = OMP_sch_modifier_monotonic;
2663 break;
2664 case OMPC_SCHEDULE_MODIFIER_nonmonotonic:
2665 Modifier = OMP_sch_modifier_nonmonotonic;
2666 break;
2667 case OMPC_SCHEDULE_MODIFIER_simd:
2668 if (Schedule == OMP_sch_static_chunked)
2669 Schedule = OMP_sch_static_balanced_chunked;
2670 break;
2671 case OMPC_SCHEDULE_MODIFIER_last:
2672 case OMPC_SCHEDULE_MODIFIER_unknown:
2673 break;
2674 }
2675 // OpenMP 5.0, 2.9.2 Worksharing-Loop Construct, Desription.
2676 // If the static schedule kind is specified or if the ordered clause is
2677 // specified, and if the nonmonotonic modifier is not specified, the effect is
2678 // as if the monotonic modifier is specified. Otherwise, unless the monotonic
2679 // modifier is specified, the effect is as if the nonmonotonic modifier is
2680 // specified.
2681 if (CGM.getLangOpts().OpenMP >= 50 && Modifier == 0) {
2682 if (!(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static ||
2683 Schedule == OMP_sch_static_balanced_chunked ||
2684 Schedule == OMP_ord_static_chunked || Schedule == OMP_ord_static ||
2685 Schedule == OMP_dist_sch_static_chunked ||
2686 Schedule == OMP_dist_sch_static))
2687 Modifier = OMP_sch_modifier_nonmonotonic;
2688 }
2689 return Schedule | Modifier;
2690 }
2691
emitForDispatchInit(CodeGenFunction & CGF,SourceLocation Loc,const OpenMPScheduleTy & ScheduleKind,unsigned IVSize,bool IVSigned,bool Ordered,const DispatchRTInput & DispatchValues)2692 void CGOpenMPRuntime::emitForDispatchInit(
2693 CodeGenFunction &CGF, SourceLocation Loc,
2694 const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned,
2695 bool Ordered, const DispatchRTInput &DispatchValues) {
2696 if (!CGF.HaveInsertPoint())
2697 return;
2698 OpenMPSchedType Schedule = getRuntimeSchedule(
2699 ScheduleKind.Schedule, DispatchValues.Chunk != nullptr, Ordered);
2700 assert(Ordered ||
2701 (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked &&
2702 Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked &&
2703 Schedule != OMP_sch_static_balanced_chunked));
2704 // Call __kmpc_dispatch_init(
2705 // ident_t *loc, kmp_int32 tid, kmp_int32 schedule,
2706 // kmp_int[32|64] lower, kmp_int[32|64] upper,
2707 // kmp_int[32|64] stride, kmp_int[32|64] chunk);
2708
2709 // If the Chunk was not specified in the clause - use default value 1.
2710 llvm::Value *Chunk = DispatchValues.Chunk ? DispatchValues.Chunk
2711 : CGF.Builder.getIntN(IVSize, 1);
2712 llvm::Value *Args[] = {
2713 emitUpdateLocation(CGF, Loc),
2714 getThreadID(CGF, Loc),
2715 CGF.Builder.getInt32(addMonoNonMonoModifier(
2716 CGM, Schedule, ScheduleKind.M1, ScheduleKind.M2)), // Schedule type
2717 DispatchValues.LB, // Lower
2718 DispatchValues.UB, // Upper
2719 CGF.Builder.getIntN(IVSize, 1), // Stride
2720 Chunk // Chunk
2721 };
2722 CGF.EmitRuntimeCall(createDispatchInitFunction(IVSize, IVSigned), Args);
2723 }
2724
emitForStaticInitCall(CodeGenFunction & CGF,llvm::Value * UpdateLocation,llvm::Value * ThreadId,llvm::FunctionCallee ForStaticInitFunction,OpenMPSchedType Schedule,OpenMPScheduleClauseModifier M1,OpenMPScheduleClauseModifier M2,const CGOpenMPRuntime::StaticRTInput & Values)2725 static void emitForStaticInitCall(
2726 CodeGenFunction &CGF, llvm::Value *UpdateLocation, llvm::Value *ThreadId,
2727 llvm::FunctionCallee ForStaticInitFunction, OpenMPSchedType Schedule,
2728 OpenMPScheduleClauseModifier M1, OpenMPScheduleClauseModifier M2,
2729 const CGOpenMPRuntime::StaticRTInput &Values) {
2730 if (!CGF.HaveInsertPoint())
2731 return;
2732
2733 assert(!Values.Ordered);
2734 assert(Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked ||
2735 Schedule == OMP_sch_static_balanced_chunked ||
2736 Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked ||
2737 Schedule == OMP_dist_sch_static ||
2738 Schedule == OMP_dist_sch_static_chunked);
2739
2740 // Call __kmpc_for_static_init(
2741 // ident_t *loc, kmp_int32 tid, kmp_int32 schedtype,
2742 // kmp_int32 *p_lastiter, kmp_int[32|64] *p_lower,
2743 // kmp_int[32|64] *p_upper, kmp_int[32|64] *p_stride,
2744 // kmp_int[32|64] incr, kmp_int[32|64] chunk);
2745 llvm::Value *Chunk = Values.Chunk;
2746 if (Chunk == nullptr) {
2747 assert((Schedule == OMP_sch_static || Schedule == OMP_ord_static ||
2748 Schedule == OMP_dist_sch_static) &&
2749 "expected static non-chunked schedule");
2750 // If the Chunk was not specified in the clause - use default value 1.
2751 Chunk = CGF.Builder.getIntN(Values.IVSize, 1);
2752 } else {
2753 assert((Schedule == OMP_sch_static_chunked ||
2754 Schedule == OMP_sch_static_balanced_chunked ||
2755 Schedule == OMP_ord_static_chunked ||
2756 Schedule == OMP_dist_sch_static_chunked) &&
2757 "expected static chunked schedule");
2758 }
2759 llvm::Value *Args[] = {
2760 UpdateLocation,
2761 ThreadId,
2762 CGF.Builder.getInt32(addMonoNonMonoModifier(CGF.CGM, Schedule, M1,
2763 M2)), // Schedule type
2764 Values.IL.getPointer(), // &isLastIter
2765 Values.LB.getPointer(), // &LB
2766 Values.UB.getPointer(), // &UB
2767 Values.ST.getPointer(), // &Stride
2768 CGF.Builder.getIntN(Values.IVSize, 1), // Incr
2769 Chunk // Chunk
2770 };
2771 CGF.EmitRuntimeCall(ForStaticInitFunction, Args);
2772 }
2773
emitForStaticInit(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind DKind,const OpenMPScheduleTy & ScheduleKind,const StaticRTInput & Values)2774 void CGOpenMPRuntime::emitForStaticInit(CodeGenFunction &CGF,
2775 SourceLocation Loc,
2776 OpenMPDirectiveKind DKind,
2777 const OpenMPScheduleTy &ScheduleKind,
2778 const StaticRTInput &Values) {
2779 OpenMPSchedType ScheduleNum = getRuntimeSchedule(
2780 ScheduleKind.Schedule, Values.Chunk != nullptr, Values.Ordered);
2781 assert(isOpenMPWorksharingDirective(DKind) &&
2782 "Expected loop-based or sections-based directive.");
2783 llvm::Value *UpdatedLocation = emitUpdateLocation(CGF, Loc,
2784 isOpenMPLoopDirective(DKind)
2785 ? OMP_IDENT_WORK_LOOP
2786 : OMP_IDENT_WORK_SECTIONS);
2787 llvm::Value *ThreadId = getThreadID(CGF, Loc);
2788 llvm::FunctionCallee StaticInitFunction =
2789 createForStaticInitFunction(Values.IVSize, Values.IVSigned, false);
2790 auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc);
2791 emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction,
2792 ScheduleNum, ScheduleKind.M1, ScheduleKind.M2, Values);
2793 }
2794
emitDistributeStaticInit(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDistScheduleClauseKind SchedKind,const CGOpenMPRuntime::StaticRTInput & Values)2795 void CGOpenMPRuntime::emitDistributeStaticInit(
2796 CodeGenFunction &CGF, SourceLocation Loc,
2797 OpenMPDistScheduleClauseKind SchedKind,
2798 const CGOpenMPRuntime::StaticRTInput &Values) {
2799 OpenMPSchedType ScheduleNum =
2800 getRuntimeSchedule(SchedKind, Values.Chunk != nullptr);
2801 llvm::Value *UpdatedLocation =
2802 emitUpdateLocation(CGF, Loc, OMP_IDENT_WORK_DISTRIBUTE);
2803 llvm::Value *ThreadId = getThreadID(CGF, Loc);
2804 llvm::FunctionCallee StaticInitFunction;
2805 bool isGPUDistribute =
2806 CGM.getLangOpts().OpenMPIsDevice &&
2807 (CGM.getTriple().isAMDGCN() || CGM.getTriple().isNVPTX());
2808 StaticInitFunction = createForStaticInitFunction(
2809 Values.IVSize, Values.IVSigned, isGPUDistribute);
2810
2811 emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction,
2812 ScheduleNum, OMPC_SCHEDULE_MODIFIER_unknown,
2813 OMPC_SCHEDULE_MODIFIER_unknown, Values);
2814 }
2815
emitForStaticFinish(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind DKind)2816 void CGOpenMPRuntime::emitForStaticFinish(CodeGenFunction &CGF,
2817 SourceLocation Loc,
2818 OpenMPDirectiveKind DKind) {
2819 if (!CGF.HaveInsertPoint())
2820 return;
2821 // Call __kmpc_for_static_fini(ident_t *loc, kmp_int32 tid);
2822 llvm::Value *Args[] = {
2823 emitUpdateLocation(CGF, Loc,
2824 isOpenMPDistributeDirective(DKind)
2825 ? OMP_IDENT_WORK_DISTRIBUTE
2826 : isOpenMPLoopDirective(DKind)
2827 ? OMP_IDENT_WORK_LOOP
2828 : OMP_IDENT_WORK_SECTIONS),
2829 getThreadID(CGF, Loc)};
2830 auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc);
2831 if (isOpenMPDistributeDirective(DKind) && CGM.getLangOpts().OpenMPIsDevice &&
2832 (CGM.getTriple().isAMDGCN() || CGM.getTriple().isNVPTX()))
2833 CGF.EmitRuntimeCall(
2834 OMPBuilder.getOrCreateRuntimeFunction(
2835 CGM.getModule(), OMPRTL___kmpc_distribute_static_fini),
2836 Args);
2837 else
2838 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2839 CGM.getModule(), OMPRTL___kmpc_for_static_fini),
2840 Args);
2841 }
2842
emitForOrderedIterationEnd(CodeGenFunction & CGF,SourceLocation Loc,unsigned IVSize,bool IVSigned)2843 void CGOpenMPRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF,
2844 SourceLocation Loc,
2845 unsigned IVSize,
2846 bool IVSigned) {
2847 if (!CGF.HaveInsertPoint())
2848 return;
2849 // Call __kmpc_for_dynamic_fini_(4|8)[u](ident_t *loc, kmp_int32 tid);
2850 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)};
2851 CGF.EmitRuntimeCall(createDispatchFiniFunction(IVSize, IVSigned), Args);
2852 }
2853
emitForNext(CodeGenFunction & CGF,SourceLocation Loc,unsigned IVSize,bool IVSigned,Address IL,Address LB,Address UB,Address ST)2854 llvm::Value *CGOpenMPRuntime::emitForNext(CodeGenFunction &CGF,
2855 SourceLocation Loc, unsigned IVSize,
2856 bool IVSigned, Address IL,
2857 Address LB, Address UB,
2858 Address ST) {
2859 // Call __kmpc_dispatch_next(
2860 // ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter,
2861 // kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper,
2862 // kmp_int[32|64] *p_stride);
2863 llvm::Value *Args[] = {
2864 emitUpdateLocation(CGF, Loc),
2865 getThreadID(CGF, Loc),
2866 IL.getPointer(), // &isLastIter
2867 LB.getPointer(), // &Lower
2868 UB.getPointer(), // &Upper
2869 ST.getPointer() // &Stride
2870 };
2871 llvm::Value *Call =
2872 CGF.EmitRuntimeCall(createDispatchNextFunction(IVSize, IVSigned), Args);
2873 return CGF.EmitScalarConversion(
2874 Call, CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/1),
2875 CGF.getContext().BoolTy, Loc);
2876 }
2877
emitNumThreadsClause(CodeGenFunction & CGF,llvm::Value * NumThreads,SourceLocation Loc)2878 void CGOpenMPRuntime::emitNumThreadsClause(CodeGenFunction &CGF,
2879 llvm::Value *NumThreads,
2880 SourceLocation Loc) {
2881 if (!CGF.HaveInsertPoint())
2882 return;
2883 // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads)
2884 llvm::Value *Args[] = {
2885 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2886 CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)};
2887 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2888 CGM.getModule(), OMPRTL___kmpc_push_num_threads),
2889 Args);
2890 }
2891
emitProcBindClause(CodeGenFunction & CGF,ProcBindKind ProcBind,SourceLocation Loc)2892 void CGOpenMPRuntime::emitProcBindClause(CodeGenFunction &CGF,
2893 ProcBindKind ProcBind,
2894 SourceLocation Loc) {
2895 if (!CGF.HaveInsertPoint())
2896 return;
2897 assert(ProcBind != OMP_PROC_BIND_unknown && "Unsupported proc_bind value.");
2898 // Build call __kmpc_push_proc_bind(&loc, global_tid, proc_bind)
2899 llvm::Value *Args[] = {
2900 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
2901 llvm::ConstantInt::get(CGM.IntTy, unsigned(ProcBind), /*isSigned=*/true)};
2902 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2903 CGM.getModule(), OMPRTL___kmpc_push_proc_bind),
2904 Args);
2905 }
2906
emitFlush(CodeGenFunction & CGF,ArrayRef<const Expr * >,SourceLocation Loc,llvm::AtomicOrdering AO)2907 void CGOpenMPRuntime::emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>,
2908 SourceLocation Loc, llvm::AtomicOrdering AO) {
2909 if (CGF.CGM.getLangOpts().OpenMPIRBuilder) {
2910 OMPBuilder.createFlush(CGF.Builder);
2911 } else {
2912 if (!CGF.HaveInsertPoint())
2913 return;
2914 // Build call void __kmpc_flush(ident_t *loc)
2915 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
2916 CGM.getModule(), OMPRTL___kmpc_flush),
2917 emitUpdateLocation(CGF, Loc));
2918 }
2919 }
2920
2921 namespace {
2922 /// Indexes of fields for type kmp_task_t.
2923 enum KmpTaskTFields {
2924 /// List of shared variables.
2925 KmpTaskTShareds,
2926 /// Task routine.
2927 KmpTaskTRoutine,
2928 /// Partition id for the untied tasks.
2929 KmpTaskTPartId,
2930 /// Function with call of destructors for private variables.
2931 Data1,
2932 /// Task priority.
2933 Data2,
2934 /// (Taskloops only) Lower bound.
2935 KmpTaskTLowerBound,
2936 /// (Taskloops only) Upper bound.
2937 KmpTaskTUpperBound,
2938 /// (Taskloops only) Stride.
2939 KmpTaskTStride,
2940 /// (Taskloops only) Is last iteration flag.
2941 KmpTaskTLastIter,
2942 /// (Taskloops only) Reduction data.
2943 KmpTaskTReductions,
2944 };
2945 } // anonymous namespace
2946
createOffloadEntriesAndInfoMetadata()2947 void CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata() {
2948 // If we are in simd mode or there are no entries, we don't need to do
2949 // anything.
2950 if (CGM.getLangOpts().OpenMPSimd || OffloadEntriesInfoManager.empty())
2951 return;
2952
2953 llvm::OpenMPIRBuilder::EmitMetadataErrorReportFunctionTy &&ErrorReportFn =
2954 [this](llvm::OpenMPIRBuilder::EmitMetadataErrorKind Kind,
2955 const llvm::TargetRegionEntryInfo &EntryInfo) -> void {
2956 SourceLocation Loc;
2957 if (Kind != llvm::OpenMPIRBuilder::EMIT_MD_GLOBAL_VAR_LINK_ERROR) {
2958 for (auto I = CGM.getContext().getSourceManager().fileinfo_begin(),
2959 E = CGM.getContext().getSourceManager().fileinfo_end();
2960 I != E; ++I) {
2961 if (I->getFirst()->getUniqueID().getDevice() == EntryInfo.DeviceID &&
2962 I->getFirst()->getUniqueID().getFile() == EntryInfo.FileID) {
2963 Loc = CGM.getContext().getSourceManager().translateFileLineCol(
2964 I->getFirst(), EntryInfo.Line, 1);
2965 break;
2966 }
2967 }
2968 }
2969 switch (Kind) {
2970 case llvm::OpenMPIRBuilder::EMIT_MD_TARGET_REGION_ERROR: {
2971 unsigned DiagID = CGM.getDiags().getCustomDiagID(
2972 DiagnosticsEngine::Error, "Offloading entry for target region in "
2973 "%0 is incorrect: either the "
2974 "address or the ID is invalid.");
2975 CGM.getDiags().Report(Loc, DiagID) << EntryInfo.ParentName;
2976 } break;
2977 case llvm::OpenMPIRBuilder::EMIT_MD_DECLARE_TARGET_ERROR: {
2978 unsigned DiagID = CGM.getDiags().getCustomDiagID(
2979 DiagnosticsEngine::Error, "Offloading entry for declare target "
2980 "variable %0 is incorrect: the "
2981 "address is invalid.");
2982 CGM.getDiags().Report(Loc, DiagID) << EntryInfo.ParentName;
2983 } break;
2984 case llvm::OpenMPIRBuilder::EMIT_MD_GLOBAL_VAR_LINK_ERROR: {
2985 unsigned DiagID = CGM.getDiags().getCustomDiagID(
2986 DiagnosticsEngine::Error,
2987 "Offloading entry for declare target variable is incorrect: the "
2988 "address is invalid.");
2989 CGM.getDiags().Report(DiagID);
2990 } break;
2991 }
2992 };
2993
2994 OMPBuilder.createOffloadEntriesAndInfoMetadata(OffloadEntriesInfoManager,
2995 ErrorReportFn);
2996 }
2997
2998 /// Loads all the offload entries information from the host IR
2999 /// metadata.
loadOffloadInfoMetadata()3000 void CGOpenMPRuntime::loadOffloadInfoMetadata() {
3001 // If we are in target mode, load the metadata from the host IR. This code has
3002 // to match the metadaata creation in createOffloadEntriesAndInfoMetadata().
3003
3004 if (!CGM.getLangOpts().OpenMPIsDevice)
3005 return;
3006
3007 if (CGM.getLangOpts().OMPHostIRFile.empty())
3008 return;
3009
3010 auto Buf = llvm::MemoryBuffer::getFile(CGM.getLangOpts().OMPHostIRFile);
3011 if (auto EC = Buf.getError()) {
3012 CGM.getDiags().Report(diag::err_cannot_open_file)
3013 << CGM.getLangOpts().OMPHostIRFile << EC.message();
3014 return;
3015 }
3016
3017 llvm::LLVMContext C;
3018 auto ME = expectedToErrorOrAndEmitErrors(
3019 C, llvm::parseBitcodeFile(Buf.get()->getMemBufferRef(), C));
3020
3021 if (auto EC = ME.getError()) {
3022 unsigned DiagID = CGM.getDiags().getCustomDiagID(
3023 DiagnosticsEngine::Error, "Unable to parse host IR file '%0':'%1'");
3024 CGM.getDiags().Report(DiagID)
3025 << CGM.getLangOpts().OMPHostIRFile << EC.message();
3026 return;
3027 }
3028
3029 OMPBuilder.loadOffloadInfoMetadata(*ME.get(), OffloadEntriesInfoManager);
3030 }
3031
emitKmpRoutineEntryT(QualType KmpInt32Ty)3032 void CGOpenMPRuntime::emitKmpRoutineEntryT(QualType KmpInt32Ty) {
3033 if (!KmpRoutineEntryPtrTy) {
3034 // Build typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); type.
3035 ASTContext &C = CGM.getContext();
3036 QualType KmpRoutineEntryTyArgs[] = {KmpInt32Ty, C.VoidPtrTy};
3037 FunctionProtoType::ExtProtoInfo EPI;
3038 KmpRoutineEntryPtrQTy = C.getPointerType(
3039 C.getFunctionType(KmpInt32Ty, KmpRoutineEntryTyArgs, EPI));
3040 KmpRoutineEntryPtrTy = CGM.getTypes().ConvertType(KmpRoutineEntryPtrQTy);
3041 }
3042 }
3043
3044 namespace {
3045 struct PrivateHelpersTy {
PrivateHelpersTy__anon7bb087080c11::PrivateHelpersTy3046 PrivateHelpersTy(const Expr *OriginalRef, const VarDecl *Original,
3047 const VarDecl *PrivateCopy, const VarDecl *PrivateElemInit)
3048 : OriginalRef(OriginalRef), Original(Original), PrivateCopy(PrivateCopy),
3049 PrivateElemInit(PrivateElemInit) {}
PrivateHelpersTy__anon7bb087080c11::PrivateHelpersTy3050 PrivateHelpersTy(const VarDecl *Original) : Original(Original) {}
3051 const Expr *OriginalRef = nullptr;
3052 const VarDecl *Original = nullptr;
3053 const VarDecl *PrivateCopy = nullptr;
3054 const VarDecl *PrivateElemInit = nullptr;
isLocalPrivate__anon7bb087080c11::PrivateHelpersTy3055 bool isLocalPrivate() const {
3056 return !OriginalRef && !PrivateCopy && !PrivateElemInit;
3057 }
3058 };
3059 typedef std::pair<CharUnits /*Align*/, PrivateHelpersTy> PrivateDataTy;
3060 } // anonymous namespace
3061
isAllocatableDecl(const VarDecl * VD)3062 static bool isAllocatableDecl(const VarDecl *VD) {
3063 const VarDecl *CVD = VD->getCanonicalDecl();
3064 if (!CVD->hasAttr<OMPAllocateDeclAttr>())
3065 return false;
3066 const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>();
3067 // Use the default allocation.
3068 return !(AA->getAllocatorType() == OMPAllocateDeclAttr::OMPDefaultMemAlloc &&
3069 !AA->getAllocator());
3070 }
3071
3072 static RecordDecl *
createPrivatesRecordDecl(CodeGenModule & CGM,ArrayRef<PrivateDataTy> Privates)3073 createPrivatesRecordDecl(CodeGenModule &CGM, ArrayRef<PrivateDataTy> Privates) {
3074 if (!Privates.empty()) {
3075 ASTContext &C = CGM.getContext();
3076 // Build struct .kmp_privates_t. {
3077 // /* private vars */
3078 // };
3079 RecordDecl *RD = C.buildImplicitRecord(".kmp_privates.t");
3080 RD->startDefinition();
3081 for (const auto &Pair : Privates) {
3082 const VarDecl *VD = Pair.second.Original;
3083 QualType Type = VD->getType().getNonReferenceType();
3084 // If the private variable is a local variable with lvalue ref type,
3085 // allocate the pointer instead of the pointee type.
3086 if (Pair.second.isLocalPrivate()) {
3087 if (VD->getType()->isLValueReferenceType())
3088 Type = C.getPointerType(Type);
3089 if (isAllocatableDecl(VD))
3090 Type = C.getPointerType(Type);
3091 }
3092 FieldDecl *FD = addFieldToRecordDecl(C, RD, Type);
3093 if (VD->hasAttrs()) {
3094 for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
3095 E(VD->getAttrs().end());
3096 I != E; ++I)
3097 FD->addAttr(*I);
3098 }
3099 }
3100 RD->completeDefinition();
3101 return RD;
3102 }
3103 return nullptr;
3104 }
3105
3106 static RecordDecl *
createKmpTaskTRecordDecl(CodeGenModule & CGM,OpenMPDirectiveKind Kind,QualType KmpInt32Ty,QualType KmpRoutineEntryPointerQTy)3107 createKmpTaskTRecordDecl(CodeGenModule &CGM, OpenMPDirectiveKind Kind,
3108 QualType KmpInt32Ty,
3109 QualType KmpRoutineEntryPointerQTy) {
3110 ASTContext &C = CGM.getContext();
3111 // Build struct kmp_task_t {
3112 // void * shareds;
3113 // kmp_routine_entry_t routine;
3114 // kmp_int32 part_id;
3115 // kmp_cmplrdata_t data1;
3116 // kmp_cmplrdata_t data2;
3117 // For taskloops additional fields:
3118 // kmp_uint64 lb;
3119 // kmp_uint64 ub;
3120 // kmp_int64 st;
3121 // kmp_int32 liter;
3122 // void * reductions;
3123 // };
3124 RecordDecl *UD = C.buildImplicitRecord("kmp_cmplrdata_t", TTK_Union);
3125 UD->startDefinition();
3126 addFieldToRecordDecl(C, UD, KmpInt32Ty);
3127 addFieldToRecordDecl(C, UD, KmpRoutineEntryPointerQTy);
3128 UD->completeDefinition();
3129 QualType KmpCmplrdataTy = C.getRecordType(UD);
3130 RecordDecl *RD = C.buildImplicitRecord("kmp_task_t");
3131 RD->startDefinition();
3132 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3133 addFieldToRecordDecl(C, RD, KmpRoutineEntryPointerQTy);
3134 addFieldToRecordDecl(C, RD, KmpInt32Ty);
3135 addFieldToRecordDecl(C, RD, KmpCmplrdataTy);
3136 addFieldToRecordDecl(C, RD, KmpCmplrdataTy);
3137 if (isOpenMPTaskLoopDirective(Kind)) {
3138 QualType KmpUInt64Ty =
3139 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0);
3140 QualType KmpInt64Ty =
3141 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
3142 addFieldToRecordDecl(C, RD, KmpUInt64Ty);
3143 addFieldToRecordDecl(C, RD, KmpUInt64Ty);
3144 addFieldToRecordDecl(C, RD, KmpInt64Ty);
3145 addFieldToRecordDecl(C, RD, KmpInt32Ty);
3146 addFieldToRecordDecl(C, RD, C.VoidPtrTy);
3147 }
3148 RD->completeDefinition();
3149 return RD;
3150 }
3151
3152 static RecordDecl *
createKmpTaskTWithPrivatesRecordDecl(CodeGenModule & CGM,QualType KmpTaskTQTy,ArrayRef<PrivateDataTy> Privates)3153 createKmpTaskTWithPrivatesRecordDecl(CodeGenModule &CGM, QualType KmpTaskTQTy,
3154 ArrayRef<PrivateDataTy> Privates) {
3155 ASTContext &C = CGM.getContext();
3156 // Build struct kmp_task_t_with_privates {
3157 // kmp_task_t task_data;
3158 // .kmp_privates_t. privates;
3159 // };
3160 RecordDecl *RD = C.buildImplicitRecord("kmp_task_t_with_privates");
3161 RD->startDefinition();
3162 addFieldToRecordDecl(C, RD, KmpTaskTQTy);
3163 if (const RecordDecl *PrivateRD = createPrivatesRecordDecl(CGM, Privates))
3164 addFieldToRecordDecl(C, RD, C.getRecordType(PrivateRD));
3165 RD->completeDefinition();
3166 return RD;
3167 }
3168
3169 /// Emit a proxy function which accepts kmp_task_t as the second
3170 /// argument.
3171 /// \code
3172 /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) {
3173 /// TaskFunction(gtid, tt->part_id, &tt->privates, task_privates_map, tt,
3174 /// For taskloops:
3175 /// tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter,
3176 /// tt->reductions, tt->shareds);
3177 /// return 0;
3178 /// }
3179 /// \endcode
3180 static llvm::Function *
emitProxyTaskFunction(CodeGenModule & CGM,SourceLocation Loc,OpenMPDirectiveKind Kind,QualType KmpInt32Ty,QualType KmpTaskTWithPrivatesPtrQTy,QualType KmpTaskTWithPrivatesQTy,QualType KmpTaskTQTy,QualType SharedsPtrTy,llvm::Function * TaskFunction,llvm::Value * TaskPrivatesMap)3181 emitProxyTaskFunction(CodeGenModule &CGM, SourceLocation Loc,
3182 OpenMPDirectiveKind Kind, QualType KmpInt32Ty,
3183 QualType KmpTaskTWithPrivatesPtrQTy,
3184 QualType KmpTaskTWithPrivatesQTy, QualType KmpTaskTQTy,
3185 QualType SharedsPtrTy, llvm::Function *TaskFunction,
3186 llvm::Value *TaskPrivatesMap) {
3187 ASTContext &C = CGM.getContext();
3188 FunctionArgList Args;
3189 ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty,
3190 ImplicitParamDecl::Other);
3191 ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3192 KmpTaskTWithPrivatesPtrQTy.withRestrict(),
3193 ImplicitParamDecl::Other);
3194 Args.push_back(&GtidArg);
3195 Args.push_back(&TaskTypeArg);
3196 const auto &TaskEntryFnInfo =
3197 CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args);
3198 llvm::FunctionType *TaskEntryTy =
3199 CGM.getTypes().GetFunctionType(TaskEntryFnInfo);
3200 std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_entry", ""});
3201 auto *TaskEntry = llvm::Function::Create(
3202 TaskEntryTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule());
3203 CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskEntry, TaskEntryFnInfo);
3204 TaskEntry->setDoesNotRecurse();
3205 CodeGenFunction CGF(CGM);
3206 CGF.StartFunction(GlobalDecl(), KmpInt32Ty, TaskEntry, TaskEntryFnInfo, Args,
3207 Loc, Loc);
3208
3209 // TaskFunction(gtid, tt->task_data.part_id, &tt->privates, task_privates_map,
3210 // tt,
3211 // For taskloops:
3212 // tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter,
3213 // tt->task_data.shareds);
3214 llvm::Value *GtidParam = CGF.EmitLoadOfScalar(
3215 CGF.GetAddrOfLocalVar(&GtidArg), /*Volatile=*/false, KmpInt32Ty, Loc);
3216 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3217 CGF.GetAddrOfLocalVar(&TaskTypeArg),
3218 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3219 const auto *KmpTaskTWithPrivatesQTyRD =
3220 cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl());
3221 LValue Base =
3222 CGF.EmitLValueForField(TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3223 const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl());
3224 auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId);
3225 LValue PartIdLVal = CGF.EmitLValueForField(Base, *PartIdFI);
3226 llvm::Value *PartidParam = PartIdLVal.getPointer(CGF);
3227
3228 auto SharedsFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds);
3229 LValue SharedsLVal = CGF.EmitLValueForField(Base, *SharedsFI);
3230 llvm::Value *SharedsParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3231 CGF.EmitLoadOfScalar(SharedsLVal, Loc),
3232 CGF.ConvertTypeForMem(SharedsPtrTy));
3233
3234 auto PrivatesFI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1);
3235 llvm::Value *PrivatesParam;
3236 if (PrivatesFI != KmpTaskTWithPrivatesQTyRD->field_end()) {
3237 LValue PrivatesLVal = CGF.EmitLValueForField(TDBase, *PrivatesFI);
3238 PrivatesParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3239 PrivatesLVal.getPointer(CGF), CGF.VoidPtrTy);
3240 } else {
3241 PrivatesParam = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
3242 }
3243
3244 llvm::Value *CommonArgs[] = {
3245 GtidParam, PartidParam, PrivatesParam, TaskPrivatesMap,
3246 CGF.Builder
3247 .CreatePointerBitCastOrAddrSpaceCast(TDBase.getAddress(CGF),
3248 CGF.VoidPtrTy, CGF.Int8Ty)
3249 .getPointer()};
3250 SmallVector<llvm::Value *, 16> CallArgs(std::begin(CommonArgs),
3251 std::end(CommonArgs));
3252 if (isOpenMPTaskLoopDirective(Kind)) {
3253 auto LBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound);
3254 LValue LBLVal = CGF.EmitLValueForField(Base, *LBFI);
3255 llvm::Value *LBParam = CGF.EmitLoadOfScalar(LBLVal, Loc);
3256 auto UBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound);
3257 LValue UBLVal = CGF.EmitLValueForField(Base, *UBFI);
3258 llvm::Value *UBParam = CGF.EmitLoadOfScalar(UBLVal, Loc);
3259 auto StFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTStride);
3260 LValue StLVal = CGF.EmitLValueForField(Base, *StFI);
3261 llvm::Value *StParam = CGF.EmitLoadOfScalar(StLVal, Loc);
3262 auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter);
3263 LValue LILVal = CGF.EmitLValueForField(Base, *LIFI);
3264 llvm::Value *LIParam = CGF.EmitLoadOfScalar(LILVal, Loc);
3265 auto RFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTReductions);
3266 LValue RLVal = CGF.EmitLValueForField(Base, *RFI);
3267 llvm::Value *RParam = CGF.EmitLoadOfScalar(RLVal, Loc);
3268 CallArgs.push_back(LBParam);
3269 CallArgs.push_back(UBParam);
3270 CallArgs.push_back(StParam);
3271 CallArgs.push_back(LIParam);
3272 CallArgs.push_back(RParam);
3273 }
3274 CallArgs.push_back(SharedsParam);
3275
3276 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskFunction,
3277 CallArgs);
3278 CGF.EmitStoreThroughLValue(RValue::get(CGF.Builder.getInt32(/*C=*/0)),
3279 CGF.MakeAddrLValue(CGF.ReturnValue, KmpInt32Ty));
3280 CGF.FinishFunction();
3281 return TaskEntry;
3282 }
3283
emitDestructorsFunction(CodeGenModule & CGM,SourceLocation Loc,QualType KmpInt32Ty,QualType KmpTaskTWithPrivatesPtrQTy,QualType KmpTaskTWithPrivatesQTy)3284 static llvm::Value *emitDestructorsFunction(CodeGenModule &CGM,
3285 SourceLocation Loc,
3286 QualType KmpInt32Ty,
3287 QualType KmpTaskTWithPrivatesPtrQTy,
3288 QualType KmpTaskTWithPrivatesQTy) {
3289 ASTContext &C = CGM.getContext();
3290 FunctionArgList Args;
3291 ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty,
3292 ImplicitParamDecl::Other);
3293 ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3294 KmpTaskTWithPrivatesPtrQTy.withRestrict(),
3295 ImplicitParamDecl::Other);
3296 Args.push_back(&GtidArg);
3297 Args.push_back(&TaskTypeArg);
3298 const auto &DestructorFnInfo =
3299 CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args);
3300 llvm::FunctionType *DestructorFnTy =
3301 CGM.getTypes().GetFunctionType(DestructorFnInfo);
3302 std::string Name =
3303 CGM.getOpenMPRuntime().getName({"omp_task_destructor", ""});
3304 auto *DestructorFn =
3305 llvm::Function::Create(DestructorFnTy, llvm::GlobalValue::InternalLinkage,
3306 Name, &CGM.getModule());
3307 CGM.SetInternalFunctionAttributes(GlobalDecl(), DestructorFn,
3308 DestructorFnInfo);
3309 DestructorFn->setDoesNotRecurse();
3310 CodeGenFunction CGF(CGM);
3311 CGF.StartFunction(GlobalDecl(), KmpInt32Ty, DestructorFn, DestructorFnInfo,
3312 Args, Loc, Loc);
3313
3314 LValue Base = CGF.EmitLoadOfPointerLValue(
3315 CGF.GetAddrOfLocalVar(&TaskTypeArg),
3316 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3317 const auto *KmpTaskTWithPrivatesQTyRD =
3318 cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl());
3319 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3320 Base = CGF.EmitLValueForField(Base, *FI);
3321 for (const auto *Field :
3322 cast<RecordDecl>(FI->getType()->getAsTagDecl())->fields()) {
3323 if (QualType::DestructionKind DtorKind =
3324 Field->getType().isDestructedType()) {
3325 LValue FieldLValue = CGF.EmitLValueForField(Base, Field);
3326 CGF.pushDestroy(DtorKind, FieldLValue.getAddress(CGF), Field->getType());
3327 }
3328 }
3329 CGF.FinishFunction();
3330 return DestructorFn;
3331 }
3332
3333 /// Emit a privates mapping function for correct handling of private and
3334 /// firstprivate variables.
3335 /// \code
3336 /// void .omp_task_privates_map.(const .privates. *noalias privs, <ty1>
3337 /// **noalias priv1,..., <tyn> **noalias privn) {
3338 /// *priv1 = &.privates.priv1;
3339 /// ...;
3340 /// *privn = &.privates.privn;
3341 /// }
3342 /// \endcode
3343 static llvm::Value *
emitTaskPrivateMappingFunction(CodeGenModule & CGM,SourceLocation Loc,const OMPTaskDataTy & Data,QualType PrivatesQTy,ArrayRef<PrivateDataTy> Privates)3344 emitTaskPrivateMappingFunction(CodeGenModule &CGM, SourceLocation Loc,
3345 const OMPTaskDataTy &Data, QualType PrivatesQTy,
3346 ArrayRef<PrivateDataTy> Privates) {
3347 ASTContext &C = CGM.getContext();
3348 FunctionArgList Args;
3349 ImplicitParamDecl TaskPrivatesArg(
3350 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3351 C.getPointerType(PrivatesQTy).withConst().withRestrict(),
3352 ImplicitParamDecl::Other);
3353 Args.push_back(&TaskPrivatesArg);
3354 llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, unsigned> PrivateVarsPos;
3355 unsigned Counter = 1;
3356 for (const Expr *E : Data.PrivateVars) {
3357 Args.push_back(ImplicitParamDecl::Create(
3358 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3359 C.getPointerType(C.getPointerType(E->getType()))
3360 .withConst()
3361 .withRestrict(),
3362 ImplicitParamDecl::Other));
3363 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3364 PrivateVarsPos[VD] = Counter;
3365 ++Counter;
3366 }
3367 for (const Expr *E : Data.FirstprivateVars) {
3368 Args.push_back(ImplicitParamDecl::Create(
3369 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3370 C.getPointerType(C.getPointerType(E->getType()))
3371 .withConst()
3372 .withRestrict(),
3373 ImplicitParamDecl::Other));
3374 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3375 PrivateVarsPos[VD] = Counter;
3376 ++Counter;
3377 }
3378 for (const Expr *E : Data.LastprivateVars) {
3379 Args.push_back(ImplicitParamDecl::Create(
3380 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3381 C.getPointerType(C.getPointerType(E->getType()))
3382 .withConst()
3383 .withRestrict(),
3384 ImplicitParamDecl::Other));
3385 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3386 PrivateVarsPos[VD] = Counter;
3387 ++Counter;
3388 }
3389 for (const VarDecl *VD : Data.PrivateLocals) {
3390 QualType Ty = VD->getType().getNonReferenceType();
3391 if (VD->getType()->isLValueReferenceType())
3392 Ty = C.getPointerType(Ty);
3393 if (isAllocatableDecl(VD))
3394 Ty = C.getPointerType(Ty);
3395 Args.push_back(ImplicitParamDecl::Create(
3396 C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3397 C.getPointerType(C.getPointerType(Ty)).withConst().withRestrict(),
3398 ImplicitParamDecl::Other));
3399 PrivateVarsPos[VD] = Counter;
3400 ++Counter;
3401 }
3402 const auto &TaskPrivatesMapFnInfo =
3403 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3404 llvm::FunctionType *TaskPrivatesMapTy =
3405 CGM.getTypes().GetFunctionType(TaskPrivatesMapFnInfo);
3406 std::string Name =
3407 CGM.getOpenMPRuntime().getName({"omp_task_privates_map", ""});
3408 auto *TaskPrivatesMap = llvm::Function::Create(
3409 TaskPrivatesMapTy, llvm::GlobalValue::InternalLinkage, Name,
3410 &CGM.getModule());
3411 CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskPrivatesMap,
3412 TaskPrivatesMapFnInfo);
3413 if (CGM.getLangOpts().Optimize) {
3414 TaskPrivatesMap->removeFnAttr(llvm::Attribute::NoInline);
3415 TaskPrivatesMap->removeFnAttr(llvm::Attribute::OptimizeNone);
3416 TaskPrivatesMap->addFnAttr(llvm::Attribute::AlwaysInline);
3417 }
3418 CodeGenFunction CGF(CGM);
3419 CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskPrivatesMap,
3420 TaskPrivatesMapFnInfo, Args, Loc, Loc);
3421
3422 // *privi = &.privates.privi;
3423 LValue Base = CGF.EmitLoadOfPointerLValue(
3424 CGF.GetAddrOfLocalVar(&TaskPrivatesArg),
3425 TaskPrivatesArg.getType()->castAs<PointerType>());
3426 const auto *PrivatesQTyRD = cast<RecordDecl>(PrivatesQTy->getAsTagDecl());
3427 Counter = 0;
3428 for (const FieldDecl *Field : PrivatesQTyRD->fields()) {
3429 LValue FieldLVal = CGF.EmitLValueForField(Base, Field);
3430 const VarDecl *VD = Args[PrivateVarsPos[Privates[Counter].second.Original]];
3431 LValue RefLVal =
3432 CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
3433 LValue RefLoadLVal = CGF.EmitLoadOfPointerLValue(
3434 RefLVal.getAddress(CGF), RefLVal.getType()->castAs<PointerType>());
3435 CGF.EmitStoreOfScalar(FieldLVal.getPointer(CGF), RefLoadLVal);
3436 ++Counter;
3437 }
3438 CGF.FinishFunction();
3439 return TaskPrivatesMap;
3440 }
3441
3442 /// Emit initialization for private variables in task-based directives.
emitPrivatesInit(CodeGenFunction & CGF,const OMPExecutableDirective & D,Address KmpTaskSharedsPtr,LValue TDBase,const RecordDecl * KmpTaskTWithPrivatesQTyRD,QualType SharedsTy,QualType SharedsPtrTy,const OMPTaskDataTy & Data,ArrayRef<PrivateDataTy> Privates,bool ForDup)3443 static void emitPrivatesInit(CodeGenFunction &CGF,
3444 const OMPExecutableDirective &D,
3445 Address KmpTaskSharedsPtr, LValue TDBase,
3446 const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3447 QualType SharedsTy, QualType SharedsPtrTy,
3448 const OMPTaskDataTy &Data,
3449 ArrayRef<PrivateDataTy> Privates, bool ForDup) {
3450 ASTContext &C = CGF.getContext();
3451 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3452 LValue PrivatesBase = CGF.EmitLValueForField(TDBase, *FI);
3453 OpenMPDirectiveKind Kind = isOpenMPTaskLoopDirective(D.getDirectiveKind())
3454 ? OMPD_taskloop
3455 : OMPD_task;
3456 const CapturedStmt &CS = *D.getCapturedStmt(Kind);
3457 CodeGenFunction::CGCapturedStmtInfo CapturesInfo(CS);
3458 LValue SrcBase;
3459 bool IsTargetTask =
3460 isOpenMPTargetDataManagementDirective(D.getDirectiveKind()) ||
3461 isOpenMPTargetExecutionDirective(D.getDirectiveKind());
3462 // For target-based directives skip 4 firstprivate arrays BasePointersArray,
3463 // PointersArray, SizesArray, and MappersArray. The original variables for
3464 // these arrays are not captured and we get their addresses explicitly.
3465 if ((!IsTargetTask && !Data.FirstprivateVars.empty() && ForDup) ||
3466 (IsTargetTask && KmpTaskSharedsPtr.isValid())) {
3467 SrcBase = CGF.MakeAddrLValue(
3468 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3469 KmpTaskSharedsPtr, CGF.ConvertTypeForMem(SharedsPtrTy),
3470 CGF.ConvertTypeForMem(SharedsTy)),
3471 SharedsTy);
3472 }
3473 FI = cast<RecordDecl>(FI->getType()->getAsTagDecl())->field_begin();
3474 for (const PrivateDataTy &Pair : Privates) {
3475 // Do not initialize private locals.
3476 if (Pair.second.isLocalPrivate()) {
3477 ++FI;
3478 continue;
3479 }
3480 const VarDecl *VD = Pair.second.PrivateCopy;
3481 const Expr *Init = VD->getAnyInitializer();
3482 if (Init && (!ForDup || (isa<CXXConstructExpr>(Init) &&
3483 !CGF.isTrivialInitializer(Init)))) {
3484 LValue PrivateLValue = CGF.EmitLValueForField(PrivatesBase, *FI);
3485 if (const VarDecl *Elem = Pair.second.PrivateElemInit) {
3486 const VarDecl *OriginalVD = Pair.second.Original;
3487 // Check if the variable is the target-based BasePointersArray,
3488 // PointersArray, SizesArray, or MappersArray.
3489 LValue SharedRefLValue;
3490 QualType Type = PrivateLValue.getType();
3491 const FieldDecl *SharedField = CapturesInfo.lookup(OriginalVD);
3492 if (IsTargetTask && !SharedField) {
3493 assert(isa<ImplicitParamDecl>(OriginalVD) &&
3494 isa<CapturedDecl>(OriginalVD->getDeclContext()) &&
3495 cast<CapturedDecl>(OriginalVD->getDeclContext())
3496 ->getNumParams() == 0 &&
3497 isa<TranslationUnitDecl>(
3498 cast<CapturedDecl>(OriginalVD->getDeclContext())
3499 ->getDeclContext()) &&
3500 "Expected artificial target data variable.");
3501 SharedRefLValue =
3502 CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(OriginalVD), Type);
3503 } else if (ForDup) {
3504 SharedRefLValue = CGF.EmitLValueForField(SrcBase, SharedField);
3505 SharedRefLValue = CGF.MakeAddrLValue(
3506 SharedRefLValue.getAddress(CGF).withAlignment(
3507 C.getDeclAlign(OriginalVD)),
3508 SharedRefLValue.getType(), LValueBaseInfo(AlignmentSource::Decl),
3509 SharedRefLValue.getTBAAInfo());
3510 } else if (CGF.LambdaCaptureFields.count(
3511 Pair.second.Original->getCanonicalDecl()) > 0 ||
3512 isa_and_nonnull<BlockDecl>(CGF.CurCodeDecl)) {
3513 SharedRefLValue = CGF.EmitLValue(Pair.second.OriginalRef);
3514 } else {
3515 // Processing for implicitly captured variables.
3516 InlinedOpenMPRegionRAII Region(
3517 CGF, [](CodeGenFunction &, PrePostActionTy &) {}, OMPD_unknown,
3518 /*HasCancel=*/false, /*NoInheritance=*/true);
3519 SharedRefLValue = CGF.EmitLValue(Pair.second.OriginalRef);
3520 }
3521 if (Type->isArrayType()) {
3522 // Initialize firstprivate array.
3523 if (!isa<CXXConstructExpr>(Init) || CGF.isTrivialInitializer(Init)) {
3524 // Perform simple memcpy.
3525 CGF.EmitAggregateAssign(PrivateLValue, SharedRefLValue, Type);
3526 } else {
3527 // Initialize firstprivate array using element-by-element
3528 // initialization.
3529 CGF.EmitOMPAggregateAssign(
3530 PrivateLValue.getAddress(CGF), SharedRefLValue.getAddress(CGF),
3531 Type,
3532 [&CGF, Elem, Init, &CapturesInfo](Address DestElement,
3533 Address SrcElement) {
3534 // Clean up any temporaries needed by the initialization.
3535 CodeGenFunction::OMPPrivateScope InitScope(CGF);
3536 InitScope.addPrivate(Elem, SrcElement);
3537 (void)InitScope.Privatize();
3538 // Emit initialization for single element.
3539 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(
3540 CGF, &CapturesInfo);
3541 CGF.EmitAnyExprToMem(Init, DestElement,
3542 Init->getType().getQualifiers(),
3543 /*IsInitializer=*/false);
3544 });
3545 }
3546 } else {
3547 CodeGenFunction::OMPPrivateScope InitScope(CGF);
3548 InitScope.addPrivate(Elem, SharedRefLValue.getAddress(CGF));
3549 (void)InitScope.Privatize();
3550 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CapturesInfo);
3551 CGF.EmitExprAsInit(Init, VD, PrivateLValue,
3552 /*capturedByInit=*/false);
3553 }
3554 } else {
3555 CGF.EmitExprAsInit(Init, VD, PrivateLValue, /*capturedByInit=*/false);
3556 }
3557 }
3558 ++FI;
3559 }
3560 }
3561
3562 /// Check if duplication function is required for taskloops.
checkInitIsRequired(CodeGenFunction & CGF,ArrayRef<PrivateDataTy> Privates)3563 static bool checkInitIsRequired(CodeGenFunction &CGF,
3564 ArrayRef<PrivateDataTy> Privates) {
3565 bool InitRequired = false;
3566 for (const PrivateDataTy &Pair : Privates) {
3567 if (Pair.second.isLocalPrivate())
3568 continue;
3569 const VarDecl *VD = Pair.second.PrivateCopy;
3570 const Expr *Init = VD->getAnyInitializer();
3571 InitRequired = InitRequired || (isa_and_nonnull<CXXConstructExpr>(Init) &&
3572 !CGF.isTrivialInitializer(Init));
3573 if (InitRequired)
3574 break;
3575 }
3576 return InitRequired;
3577 }
3578
3579
3580 /// Emit task_dup function (for initialization of
3581 /// private/firstprivate/lastprivate vars and last_iter flag)
3582 /// \code
3583 /// void __task_dup_entry(kmp_task_t *task_dst, const kmp_task_t *task_src, int
3584 /// lastpriv) {
3585 /// // setup lastprivate flag
3586 /// task_dst->last = lastpriv;
3587 /// // could be constructor calls here...
3588 /// }
3589 /// \endcode
3590 static llvm::Value *
emitTaskDupFunction(CodeGenModule & CGM,SourceLocation Loc,const OMPExecutableDirective & D,QualType KmpTaskTWithPrivatesPtrQTy,const RecordDecl * KmpTaskTWithPrivatesQTyRD,const RecordDecl * KmpTaskTQTyRD,QualType SharedsTy,QualType SharedsPtrTy,const OMPTaskDataTy & Data,ArrayRef<PrivateDataTy> Privates,bool WithLastIter)3591 emitTaskDupFunction(CodeGenModule &CGM, SourceLocation Loc,
3592 const OMPExecutableDirective &D,
3593 QualType KmpTaskTWithPrivatesPtrQTy,
3594 const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3595 const RecordDecl *KmpTaskTQTyRD, QualType SharedsTy,
3596 QualType SharedsPtrTy, const OMPTaskDataTy &Data,
3597 ArrayRef<PrivateDataTy> Privates, bool WithLastIter) {
3598 ASTContext &C = CGM.getContext();
3599 FunctionArgList Args;
3600 ImplicitParamDecl DstArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3601 KmpTaskTWithPrivatesPtrQTy,
3602 ImplicitParamDecl::Other);
3603 ImplicitParamDecl SrcArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3604 KmpTaskTWithPrivatesPtrQTy,
3605 ImplicitParamDecl::Other);
3606 ImplicitParamDecl LastprivArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
3607 ImplicitParamDecl::Other);
3608 Args.push_back(&DstArg);
3609 Args.push_back(&SrcArg);
3610 Args.push_back(&LastprivArg);
3611 const auto &TaskDupFnInfo =
3612 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3613 llvm::FunctionType *TaskDupTy = CGM.getTypes().GetFunctionType(TaskDupFnInfo);
3614 std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_dup", ""});
3615 auto *TaskDup = llvm::Function::Create(
3616 TaskDupTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule());
3617 CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskDup, TaskDupFnInfo);
3618 TaskDup->setDoesNotRecurse();
3619 CodeGenFunction CGF(CGM);
3620 CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskDup, TaskDupFnInfo, Args, Loc,
3621 Loc);
3622
3623 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3624 CGF.GetAddrOfLocalVar(&DstArg),
3625 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3626 // task_dst->liter = lastpriv;
3627 if (WithLastIter) {
3628 auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter);
3629 LValue Base = CGF.EmitLValueForField(
3630 TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3631 LValue LILVal = CGF.EmitLValueForField(Base, *LIFI);
3632 llvm::Value *Lastpriv = CGF.EmitLoadOfScalar(
3633 CGF.GetAddrOfLocalVar(&LastprivArg), /*Volatile=*/false, C.IntTy, Loc);
3634 CGF.EmitStoreOfScalar(Lastpriv, LILVal);
3635 }
3636
3637 // Emit initial values for private copies (if any).
3638 assert(!Privates.empty());
3639 Address KmpTaskSharedsPtr = Address::invalid();
3640 if (!Data.FirstprivateVars.empty()) {
3641 LValue TDBase = CGF.EmitLoadOfPointerLValue(
3642 CGF.GetAddrOfLocalVar(&SrcArg),
3643 KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>());
3644 LValue Base = CGF.EmitLValueForField(
3645 TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin());
3646 KmpTaskSharedsPtr = Address(
3647 CGF.EmitLoadOfScalar(CGF.EmitLValueForField(
3648 Base, *std::next(KmpTaskTQTyRD->field_begin(),
3649 KmpTaskTShareds)),
3650 Loc),
3651 CGF.Int8Ty, CGM.getNaturalTypeAlignment(SharedsTy));
3652 }
3653 emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, TDBase, KmpTaskTWithPrivatesQTyRD,
3654 SharedsTy, SharedsPtrTy, Data, Privates, /*ForDup=*/true);
3655 CGF.FinishFunction();
3656 return TaskDup;
3657 }
3658
3659 /// Checks if destructor function is required to be generated.
3660 /// \return true if cleanups are required, false otherwise.
3661 static bool
checkDestructorsRequired(const RecordDecl * KmpTaskTWithPrivatesQTyRD,ArrayRef<PrivateDataTy> Privates)3662 checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD,
3663 ArrayRef<PrivateDataTy> Privates) {
3664 for (const PrivateDataTy &P : Privates) {
3665 if (P.second.isLocalPrivate())
3666 continue;
3667 QualType Ty = P.second.Original->getType().getNonReferenceType();
3668 if (Ty.isDestructedType())
3669 return true;
3670 }
3671 return false;
3672 }
3673
3674 namespace {
3675 /// Loop generator for OpenMP iterator expression.
3676 class OMPIteratorGeneratorScope final
3677 : public CodeGenFunction::OMPPrivateScope {
3678 CodeGenFunction &CGF;
3679 const OMPIteratorExpr *E = nullptr;
3680 SmallVector<CodeGenFunction::JumpDest, 4> ContDests;
3681 SmallVector<CodeGenFunction::JumpDest, 4> ExitDests;
3682 OMPIteratorGeneratorScope() = delete;
3683 OMPIteratorGeneratorScope(OMPIteratorGeneratorScope &) = delete;
3684
3685 public:
OMPIteratorGeneratorScope(CodeGenFunction & CGF,const OMPIteratorExpr * E)3686 OMPIteratorGeneratorScope(CodeGenFunction &CGF, const OMPIteratorExpr *E)
3687 : CodeGenFunction::OMPPrivateScope(CGF), CGF(CGF), E(E) {
3688 if (!E)
3689 return;
3690 SmallVector<llvm::Value *, 4> Uppers;
3691 for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) {
3692 Uppers.push_back(CGF.EmitScalarExpr(E->getHelper(I).Upper));
3693 const auto *VD = cast<VarDecl>(E->getIteratorDecl(I));
3694 addPrivate(VD, CGF.CreateMemTemp(VD->getType(), VD->getName()));
3695 const OMPIteratorHelperData &HelperData = E->getHelper(I);
3696 addPrivate(
3697 HelperData.CounterVD,
3698 CGF.CreateMemTemp(HelperData.CounterVD->getType(), "counter.addr"));
3699 }
3700 Privatize();
3701
3702 for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) {
3703 const OMPIteratorHelperData &HelperData = E->getHelper(I);
3704 LValue CLVal =
3705 CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(HelperData.CounterVD),
3706 HelperData.CounterVD->getType());
3707 // Counter = 0;
3708 CGF.EmitStoreOfScalar(
3709 llvm::ConstantInt::get(CLVal.getAddress(CGF).getElementType(), 0),
3710 CLVal);
3711 CodeGenFunction::JumpDest &ContDest =
3712 ContDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.cont"));
3713 CodeGenFunction::JumpDest &ExitDest =
3714 ExitDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.exit"));
3715 // N = <number-of_iterations>;
3716 llvm::Value *N = Uppers[I];
3717 // cont:
3718 // if (Counter < N) goto body; else goto exit;
3719 CGF.EmitBlock(ContDest.getBlock());
3720 auto *CVal =
3721 CGF.EmitLoadOfScalar(CLVal, HelperData.CounterVD->getLocation());
3722 llvm::Value *Cmp =
3723 HelperData.CounterVD->getType()->isSignedIntegerOrEnumerationType()
3724 ? CGF.Builder.CreateICmpSLT(CVal, N)
3725 : CGF.Builder.CreateICmpULT(CVal, N);
3726 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("iter.body");
3727 CGF.Builder.CreateCondBr(Cmp, BodyBB, ExitDest.getBlock());
3728 // body:
3729 CGF.EmitBlock(BodyBB);
3730 // Iteri = Begini + Counter * Stepi;
3731 CGF.EmitIgnoredExpr(HelperData.Update);
3732 }
3733 }
~OMPIteratorGeneratorScope()3734 ~OMPIteratorGeneratorScope() {
3735 if (!E)
3736 return;
3737 for (unsigned I = E->numOfIterators(); I > 0; --I) {
3738 // Counter = Counter + 1;
3739 const OMPIteratorHelperData &HelperData = E->getHelper(I - 1);
3740 CGF.EmitIgnoredExpr(HelperData.CounterUpdate);
3741 // goto cont;
3742 CGF.EmitBranchThroughCleanup(ContDests[I - 1]);
3743 // exit:
3744 CGF.EmitBlock(ExitDests[I - 1].getBlock(), /*IsFinished=*/I == 1);
3745 }
3746 }
3747 };
3748 } // namespace
3749
3750 static std::pair<llvm::Value *, llvm::Value *>
getPointerAndSize(CodeGenFunction & CGF,const Expr * E)3751 getPointerAndSize(CodeGenFunction &CGF, const Expr *E) {
3752 const auto *OASE = dyn_cast<OMPArrayShapingExpr>(E);
3753 llvm::Value *Addr;
3754 if (OASE) {
3755 const Expr *Base = OASE->getBase();
3756 Addr = CGF.EmitScalarExpr(Base);
3757 } else {
3758 Addr = CGF.EmitLValue(E).getPointer(CGF);
3759 }
3760 llvm::Value *SizeVal;
3761 QualType Ty = E->getType();
3762 if (OASE) {
3763 SizeVal = CGF.getTypeSize(OASE->getBase()->getType()->getPointeeType());
3764 for (const Expr *SE : OASE->getDimensions()) {
3765 llvm::Value *Sz = CGF.EmitScalarExpr(SE);
3766 Sz = CGF.EmitScalarConversion(
3767 Sz, SE->getType(), CGF.getContext().getSizeType(), SE->getExprLoc());
3768 SizeVal = CGF.Builder.CreateNUWMul(SizeVal, Sz);
3769 }
3770 } else if (const auto *ASE =
3771 dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) {
3772 LValue UpAddrLVal =
3773 CGF.EmitOMPArraySectionExpr(ASE, /*IsLowerBound=*/false);
3774 Address UpAddrAddress = UpAddrLVal.getAddress(CGF);
3775 llvm::Value *UpAddr = CGF.Builder.CreateConstGEP1_32(
3776 UpAddrAddress.getElementType(), UpAddrAddress.getPointer(), /*Idx0=*/1);
3777 llvm::Value *LowIntPtr = CGF.Builder.CreatePtrToInt(Addr, CGF.SizeTy);
3778 llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGF.SizeTy);
3779 SizeVal = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr);
3780 } else {
3781 SizeVal = CGF.getTypeSize(Ty);
3782 }
3783 return std::make_pair(Addr, SizeVal);
3784 }
3785
3786 /// Builds kmp_depend_info, if it is not built yet, and builds flags type.
getKmpAffinityType(ASTContext & C,QualType & KmpTaskAffinityInfoTy)3787 static void getKmpAffinityType(ASTContext &C, QualType &KmpTaskAffinityInfoTy) {
3788 QualType FlagsTy = C.getIntTypeForBitwidth(32, /*Signed=*/false);
3789 if (KmpTaskAffinityInfoTy.isNull()) {
3790 RecordDecl *KmpAffinityInfoRD =
3791 C.buildImplicitRecord("kmp_task_affinity_info_t");
3792 KmpAffinityInfoRD->startDefinition();
3793 addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getIntPtrType());
3794 addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getSizeType());
3795 addFieldToRecordDecl(C, KmpAffinityInfoRD, FlagsTy);
3796 KmpAffinityInfoRD->completeDefinition();
3797 KmpTaskAffinityInfoTy = C.getRecordType(KmpAffinityInfoRD);
3798 }
3799 }
3800
3801 CGOpenMPRuntime::TaskResultTy
emitTaskInit(CodeGenFunction & CGF,SourceLocation Loc,const OMPExecutableDirective & D,llvm::Function * TaskFunction,QualType SharedsTy,Address Shareds,const OMPTaskDataTy & Data)3802 CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc,
3803 const OMPExecutableDirective &D,
3804 llvm::Function *TaskFunction, QualType SharedsTy,
3805 Address Shareds, const OMPTaskDataTy &Data) {
3806 ASTContext &C = CGM.getContext();
3807 llvm::SmallVector<PrivateDataTy, 4> Privates;
3808 // Aggregate privates and sort them by the alignment.
3809 const auto *I = Data.PrivateCopies.begin();
3810 for (const Expr *E : Data.PrivateVars) {
3811 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3812 Privates.emplace_back(
3813 C.getDeclAlign(VD),
3814 PrivateHelpersTy(E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3815 /*PrivateElemInit=*/nullptr));
3816 ++I;
3817 }
3818 I = Data.FirstprivateCopies.begin();
3819 const auto *IElemInitRef = Data.FirstprivateInits.begin();
3820 for (const Expr *E : Data.FirstprivateVars) {
3821 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3822 Privates.emplace_back(
3823 C.getDeclAlign(VD),
3824 PrivateHelpersTy(
3825 E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3826 cast<VarDecl>(cast<DeclRefExpr>(*IElemInitRef)->getDecl())));
3827 ++I;
3828 ++IElemInitRef;
3829 }
3830 I = Data.LastprivateCopies.begin();
3831 for (const Expr *E : Data.LastprivateVars) {
3832 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3833 Privates.emplace_back(
3834 C.getDeclAlign(VD),
3835 PrivateHelpersTy(E, VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()),
3836 /*PrivateElemInit=*/nullptr));
3837 ++I;
3838 }
3839 for (const VarDecl *VD : Data.PrivateLocals) {
3840 if (isAllocatableDecl(VD))
3841 Privates.emplace_back(CGM.getPointerAlign(), PrivateHelpersTy(VD));
3842 else
3843 Privates.emplace_back(C.getDeclAlign(VD), PrivateHelpersTy(VD));
3844 }
3845 llvm::stable_sort(Privates,
3846 [](const PrivateDataTy &L, const PrivateDataTy &R) {
3847 return L.first > R.first;
3848 });
3849 QualType KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
3850 // Build type kmp_routine_entry_t (if not built yet).
3851 emitKmpRoutineEntryT(KmpInt32Ty);
3852 // Build type kmp_task_t (if not built yet).
3853 if (isOpenMPTaskLoopDirective(D.getDirectiveKind())) {
3854 if (SavedKmpTaskloopTQTy.isNull()) {
3855 SavedKmpTaskloopTQTy = C.getRecordType(createKmpTaskTRecordDecl(
3856 CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy));
3857 }
3858 KmpTaskTQTy = SavedKmpTaskloopTQTy;
3859 } else {
3860 assert((D.getDirectiveKind() == OMPD_task ||
3861 isOpenMPTargetExecutionDirective(D.getDirectiveKind()) ||
3862 isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) &&
3863 "Expected taskloop, task or target directive");
3864 if (SavedKmpTaskTQTy.isNull()) {
3865 SavedKmpTaskTQTy = C.getRecordType(createKmpTaskTRecordDecl(
3866 CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy));
3867 }
3868 KmpTaskTQTy = SavedKmpTaskTQTy;
3869 }
3870 const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl());
3871 // Build particular struct kmp_task_t for the given task.
3872 const RecordDecl *KmpTaskTWithPrivatesQTyRD =
3873 createKmpTaskTWithPrivatesRecordDecl(CGM, KmpTaskTQTy, Privates);
3874 QualType KmpTaskTWithPrivatesQTy = C.getRecordType(KmpTaskTWithPrivatesQTyRD);
3875 QualType KmpTaskTWithPrivatesPtrQTy =
3876 C.getPointerType(KmpTaskTWithPrivatesQTy);
3877 llvm::Type *KmpTaskTWithPrivatesTy = CGF.ConvertType(KmpTaskTWithPrivatesQTy);
3878 llvm::Type *KmpTaskTWithPrivatesPtrTy =
3879 KmpTaskTWithPrivatesTy->getPointerTo();
3880 llvm::Value *KmpTaskTWithPrivatesTySize =
3881 CGF.getTypeSize(KmpTaskTWithPrivatesQTy);
3882 QualType SharedsPtrTy = C.getPointerType(SharedsTy);
3883
3884 // Emit initial values for private copies (if any).
3885 llvm::Value *TaskPrivatesMap = nullptr;
3886 llvm::Type *TaskPrivatesMapTy =
3887 std::next(TaskFunction->arg_begin(), 3)->getType();
3888 if (!Privates.empty()) {
3889 auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin());
3890 TaskPrivatesMap =
3891 emitTaskPrivateMappingFunction(CGM, Loc, Data, FI->getType(), Privates);
3892 TaskPrivatesMap = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3893 TaskPrivatesMap, TaskPrivatesMapTy);
3894 } else {
3895 TaskPrivatesMap = llvm::ConstantPointerNull::get(
3896 cast<llvm::PointerType>(TaskPrivatesMapTy));
3897 }
3898 // Build a proxy function kmp_int32 .omp_task_entry.(kmp_int32 gtid,
3899 // kmp_task_t *tt);
3900 llvm::Function *TaskEntry = emitProxyTaskFunction(
3901 CGM, Loc, D.getDirectiveKind(), KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy,
3902 KmpTaskTWithPrivatesQTy, KmpTaskTQTy, SharedsPtrTy, TaskFunction,
3903 TaskPrivatesMap);
3904
3905 // Build call kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid,
3906 // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
3907 // kmp_routine_entry_t *task_entry);
3908 // Task flags. Format is taken from
3909 // https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp.h,
3910 // description of kmp_tasking_flags struct.
3911 enum {
3912 TiedFlag = 0x1,
3913 FinalFlag = 0x2,
3914 DestructorsFlag = 0x8,
3915 PriorityFlag = 0x20,
3916 DetachableFlag = 0x40,
3917 };
3918 unsigned Flags = Data.Tied ? TiedFlag : 0;
3919 bool NeedsCleanup = false;
3920 if (!Privates.empty()) {
3921 NeedsCleanup =
3922 checkDestructorsRequired(KmpTaskTWithPrivatesQTyRD, Privates);
3923 if (NeedsCleanup)
3924 Flags = Flags | DestructorsFlag;
3925 }
3926 if (Data.Priority.getInt())
3927 Flags = Flags | PriorityFlag;
3928 if (D.hasClausesOfKind<OMPDetachClause>())
3929 Flags = Flags | DetachableFlag;
3930 llvm::Value *TaskFlags =
3931 Data.Final.getPointer()
3932 ? CGF.Builder.CreateSelect(Data.Final.getPointer(),
3933 CGF.Builder.getInt32(FinalFlag),
3934 CGF.Builder.getInt32(/*C=*/0))
3935 : CGF.Builder.getInt32(Data.Final.getInt() ? FinalFlag : 0);
3936 TaskFlags = CGF.Builder.CreateOr(TaskFlags, CGF.Builder.getInt32(Flags));
3937 llvm::Value *SharedsSize = CGM.getSize(C.getTypeSizeInChars(SharedsTy));
3938 SmallVector<llvm::Value *, 8> AllocArgs = {emitUpdateLocation(CGF, Loc),
3939 getThreadID(CGF, Loc), TaskFlags, KmpTaskTWithPrivatesTySize,
3940 SharedsSize, CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3941 TaskEntry, KmpRoutineEntryPtrTy)};
3942 llvm::Value *NewTask;
3943 if (D.hasClausesOfKind<OMPNowaitClause>()) {
3944 // Check if we have any device clause associated with the directive.
3945 const Expr *Device = nullptr;
3946 if (auto *C = D.getSingleClause<OMPDeviceClause>())
3947 Device = C->getDevice();
3948 // Emit device ID if any otherwise use default value.
3949 llvm::Value *DeviceID;
3950 if (Device)
3951 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
3952 CGF.Int64Ty, /*isSigned=*/true);
3953 else
3954 DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF);
3955 AllocArgs.push_back(DeviceID);
3956 NewTask = CGF.EmitRuntimeCall(
3957 OMPBuilder.getOrCreateRuntimeFunction(
3958 CGM.getModule(), OMPRTL___kmpc_omp_target_task_alloc),
3959 AllocArgs);
3960 } else {
3961 NewTask =
3962 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3963 CGM.getModule(), OMPRTL___kmpc_omp_task_alloc),
3964 AllocArgs);
3965 }
3966 // Emit detach clause initialization.
3967 // evt = (typeof(evt))__kmpc_task_allow_completion_event(loc, tid,
3968 // task_descriptor);
3969 if (const auto *DC = D.getSingleClause<OMPDetachClause>()) {
3970 const Expr *Evt = DC->getEventHandler()->IgnoreParenImpCasts();
3971 LValue EvtLVal = CGF.EmitLValue(Evt);
3972
3973 // Build kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3974 // int gtid, kmp_task_t *task);
3975 llvm::Value *Loc = emitUpdateLocation(CGF, DC->getBeginLoc());
3976 llvm::Value *Tid = getThreadID(CGF, DC->getBeginLoc());
3977 Tid = CGF.Builder.CreateIntCast(Tid, CGF.IntTy, /*isSigned=*/false);
3978 llvm::Value *EvtVal = CGF.EmitRuntimeCall(
3979 OMPBuilder.getOrCreateRuntimeFunction(
3980 CGM.getModule(), OMPRTL___kmpc_task_allow_completion_event),
3981 {Loc, Tid, NewTask});
3982 EvtVal = CGF.EmitScalarConversion(EvtVal, C.VoidPtrTy, Evt->getType(),
3983 Evt->getExprLoc());
3984 CGF.EmitStoreOfScalar(EvtVal, EvtLVal);
3985 }
3986 // Process affinity clauses.
3987 if (D.hasClausesOfKind<OMPAffinityClause>()) {
3988 // Process list of affinity data.
3989 ASTContext &C = CGM.getContext();
3990 Address AffinitiesArray = Address::invalid();
3991 // Calculate number of elements to form the array of affinity data.
3992 llvm::Value *NumOfElements = nullptr;
3993 unsigned NumAffinities = 0;
3994 for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) {
3995 if (const Expr *Modifier = C->getModifier()) {
3996 const auto *IE = cast<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts());
3997 for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) {
3998 llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper);
3999 Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false);
4000 NumOfElements =
4001 NumOfElements ? CGF.Builder.CreateNUWMul(NumOfElements, Sz) : Sz;
4002 }
4003 } else {
4004 NumAffinities += C->varlist_size();
4005 }
4006 }
4007 getKmpAffinityType(CGM.getContext(), KmpTaskAffinityInfoTy);
4008 // Fields ids in kmp_task_affinity_info record.
4009 enum RTLAffinityInfoFieldsTy { BaseAddr, Len, Flags };
4010
4011 QualType KmpTaskAffinityInfoArrayTy;
4012 if (NumOfElements) {
4013 NumOfElements = CGF.Builder.CreateNUWAdd(
4014 llvm::ConstantInt::get(CGF.SizeTy, NumAffinities), NumOfElements);
4015 auto *OVE = new (C) OpaqueValueExpr(
4016 Loc,
4017 C.getIntTypeForBitwidth(C.getTypeSize(C.getSizeType()), /*Signed=*/0),
4018 VK_PRValue);
4019 CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, OVE,
4020 RValue::get(NumOfElements));
4021 KmpTaskAffinityInfoArrayTy =
4022 C.getVariableArrayType(KmpTaskAffinityInfoTy, OVE, ArrayType::Normal,
4023 /*IndexTypeQuals=*/0, SourceRange(Loc, Loc));
4024 // Properly emit variable-sized array.
4025 auto *PD = ImplicitParamDecl::Create(C, KmpTaskAffinityInfoArrayTy,
4026 ImplicitParamDecl::Other);
4027 CGF.EmitVarDecl(*PD);
4028 AffinitiesArray = CGF.GetAddrOfLocalVar(PD);
4029 NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty,
4030 /*isSigned=*/false);
4031 } else {
4032 KmpTaskAffinityInfoArrayTy = C.getConstantArrayType(
4033 KmpTaskAffinityInfoTy,
4034 llvm::APInt(C.getTypeSize(C.getSizeType()), NumAffinities), nullptr,
4035 ArrayType::Normal, /*IndexTypeQuals=*/0);
4036 AffinitiesArray =
4037 CGF.CreateMemTemp(KmpTaskAffinityInfoArrayTy, ".affs.arr.addr");
4038 AffinitiesArray = CGF.Builder.CreateConstArrayGEP(AffinitiesArray, 0);
4039 NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumAffinities,
4040 /*isSigned=*/false);
4041 }
4042
4043 const auto *KmpAffinityInfoRD = KmpTaskAffinityInfoTy->getAsRecordDecl();
4044 // Fill array by elements without iterators.
4045 unsigned Pos = 0;
4046 bool HasIterator = false;
4047 for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) {
4048 if (C->getModifier()) {
4049 HasIterator = true;
4050 continue;
4051 }
4052 for (const Expr *E : C->varlists()) {
4053 llvm::Value *Addr;
4054 llvm::Value *Size;
4055 std::tie(Addr, Size) = getPointerAndSize(CGF, E);
4056 LValue Base =
4057 CGF.MakeAddrLValue(CGF.Builder.CreateConstGEP(AffinitiesArray, Pos),
4058 KmpTaskAffinityInfoTy);
4059 // affs[i].base_addr = &<Affinities[i].second>;
4060 LValue BaseAddrLVal = CGF.EmitLValueForField(
4061 Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr));
4062 CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy),
4063 BaseAddrLVal);
4064 // affs[i].len = sizeof(<Affinities[i].second>);
4065 LValue LenLVal = CGF.EmitLValueForField(
4066 Base, *std::next(KmpAffinityInfoRD->field_begin(), Len));
4067 CGF.EmitStoreOfScalar(Size, LenLVal);
4068 ++Pos;
4069 }
4070 }
4071 LValue PosLVal;
4072 if (HasIterator) {
4073 PosLVal = CGF.MakeAddrLValue(
4074 CGF.CreateMemTemp(C.getSizeType(), "affs.counter.addr"),
4075 C.getSizeType());
4076 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal);
4077 }
4078 // Process elements with iterators.
4079 for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) {
4080 const Expr *Modifier = C->getModifier();
4081 if (!Modifier)
4082 continue;
4083 OMPIteratorGeneratorScope IteratorScope(
4084 CGF, cast_or_null<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts()));
4085 for (const Expr *E : C->varlists()) {
4086 llvm::Value *Addr;
4087 llvm::Value *Size;
4088 std::tie(Addr, Size) = getPointerAndSize(CGF, E);
4089 llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc());
4090 LValue Base = CGF.MakeAddrLValue(
4091 CGF.Builder.CreateGEP(AffinitiesArray, Idx), KmpTaskAffinityInfoTy);
4092 // affs[i].base_addr = &<Affinities[i].second>;
4093 LValue BaseAddrLVal = CGF.EmitLValueForField(
4094 Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr));
4095 CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy),
4096 BaseAddrLVal);
4097 // affs[i].len = sizeof(<Affinities[i].second>);
4098 LValue LenLVal = CGF.EmitLValueForField(
4099 Base, *std::next(KmpAffinityInfoRD->field_begin(), Len));
4100 CGF.EmitStoreOfScalar(Size, LenLVal);
4101 Idx = CGF.Builder.CreateNUWAdd(
4102 Idx, llvm::ConstantInt::get(Idx->getType(), 1));
4103 CGF.EmitStoreOfScalar(Idx, PosLVal);
4104 }
4105 }
4106 // Call to kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref,
4107 // kmp_int32 gtid, kmp_task_t *new_task, kmp_int32
4108 // naffins, kmp_task_affinity_info_t *affin_list);
4109 llvm::Value *LocRef = emitUpdateLocation(CGF, Loc);
4110 llvm::Value *GTid = getThreadID(CGF, Loc);
4111 llvm::Value *AffinListPtr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4112 AffinitiesArray.getPointer(), CGM.VoidPtrTy);
4113 // FIXME: Emit the function and ignore its result for now unless the
4114 // runtime function is properly implemented.
4115 (void)CGF.EmitRuntimeCall(
4116 OMPBuilder.getOrCreateRuntimeFunction(
4117 CGM.getModule(), OMPRTL___kmpc_omp_reg_task_with_affinity),
4118 {LocRef, GTid, NewTask, NumOfElements, AffinListPtr});
4119 }
4120 llvm::Value *NewTaskNewTaskTTy =
4121 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4122 NewTask, KmpTaskTWithPrivatesPtrTy);
4123 LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy,
4124 KmpTaskTWithPrivatesQTy);
4125 LValue TDBase =
4126 CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin());
4127 // Fill the data in the resulting kmp_task_t record.
4128 // Copy shareds if there are any.
4129 Address KmpTaskSharedsPtr = Address::invalid();
4130 if (!SharedsTy->getAsStructureType()->getDecl()->field_empty()) {
4131 KmpTaskSharedsPtr = Address(
4132 CGF.EmitLoadOfScalar(
4133 CGF.EmitLValueForField(
4134 TDBase,
4135 *std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds)),
4136 Loc),
4137 CGF.Int8Ty, CGM.getNaturalTypeAlignment(SharedsTy));
4138 LValue Dest = CGF.MakeAddrLValue(KmpTaskSharedsPtr, SharedsTy);
4139 LValue Src = CGF.MakeAddrLValue(Shareds, SharedsTy);
4140 CGF.EmitAggregateCopy(Dest, Src, SharedsTy, AggValueSlot::DoesNotOverlap);
4141 }
4142 // Emit initial values for private copies (if any).
4143 TaskResultTy Result;
4144 if (!Privates.empty()) {
4145 emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, Base, KmpTaskTWithPrivatesQTyRD,
4146 SharedsTy, SharedsPtrTy, Data, Privates,
4147 /*ForDup=*/false);
4148 if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) &&
4149 (!Data.LastprivateVars.empty() || checkInitIsRequired(CGF, Privates))) {
4150 Result.TaskDupFn = emitTaskDupFunction(
4151 CGM, Loc, D, KmpTaskTWithPrivatesPtrQTy, KmpTaskTWithPrivatesQTyRD,
4152 KmpTaskTQTyRD, SharedsTy, SharedsPtrTy, Data, Privates,
4153 /*WithLastIter=*/!Data.LastprivateVars.empty());
4154 }
4155 }
4156 // Fields of union "kmp_cmplrdata_t" for destructors and priority.
4157 enum { Priority = 0, Destructors = 1 };
4158 // Provide pointer to function with destructors for privates.
4159 auto FI = std::next(KmpTaskTQTyRD->field_begin(), Data1);
4160 const RecordDecl *KmpCmplrdataUD =
4161 (*FI)->getType()->getAsUnionType()->getDecl();
4162 if (NeedsCleanup) {
4163 llvm::Value *DestructorFn = emitDestructorsFunction(
4164 CGM, Loc, KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy,
4165 KmpTaskTWithPrivatesQTy);
4166 LValue Data1LV = CGF.EmitLValueForField(TDBase, *FI);
4167 LValue DestructorsLV = CGF.EmitLValueForField(
4168 Data1LV, *std::next(KmpCmplrdataUD->field_begin(), Destructors));
4169 CGF.EmitStoreOfScalar(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4170 DestructorFn, KmpRoutineEntryPtrTy),
4171 DestructorsLV);
4172 }
4173 // Set priority.
4174 if (Data.Priority.getInt()) {
4175 LValue Data2LV = CGF.EmitLValueForField(
4176 TDBase, *std::next(KmpTaskTQTyRD->field_begin(), Data2));
4177 LValue PriorityLV = CGF.EmitLValueForField(
4178 Data2LV, *std::next(KmpCmplrdataUD->field_begin(), Priority));
4179 CGF.EmitStoreOfScalar(Data.Priority.getPointer(), PriorityLV);
4180 }
4181 Result.NewTask = NewTask;
4182 Result.TaskEntry = TaskEntry;
4183 Result.NewTaskNewTaskTTy = NewTaskNewTaskTTy;
4184 Result.TDBase = TDBase;
4185 Result.KmpTaskTQTyRD = KmpTaskTQTyRD;
4186 return Result;
4187 }
4188
4189 /// Translates internal dependency kind into the runtime kind.
translateDependencyKind(OpenMPDependClauseKind K)4190 static RTLDependenceKindTy translateDependencyKind(OpenMPDependClauseKind K) {
4191 RTLDependenceKindTy DepKind;
4192 switch (K) {
4193 case OMPC_DEPEND_in:
4194 DepKind = RTLDependenceKindTy::DepIn;
4195 break;
4196 // Out and InOut dependencies must use the same code.
4197 case OMPC_DEPEND_out:
4198 case OMPC_DEPEND_inout:
4199 DepKind = RTLDependenceKindTy::DepInOut;
4200 break;
4201 case OMPC_DEPEND_mutexinoutset:
4202 DepKind = RTLDependenceKindTy::DepMutexInOutSet;
4203 break;
4204 case OMPC_DEPEND_inoutset:
4205 DepKind = RTLDependenceKindTy::DepInOutSet;
4206 break;
4207 case OMPC_DEPEND_outallmemory:
4208 DepKind = RTLDependenceKindTy::DepOmpAllMem;
4209 break;
4210 case OMPC_DEPEND_source:
4211 case OMPC_DEPEND_sink:
4212 case OMPC_DEPEND_depobj:
4213 case OMPC_DEPEND_inoutallmemory:
4214 case OMPC_DEPEND_unknown:
4215 llvm_unreachable("Unknown task dependence type");
4216 }
4217 return DepKind;
4218 }
4219
4220 /// Builds kmp_depend_info, if it is not built yet, and builds flags type.
getDependTypes(ASTContext & C,QualType & KmpDependInfoTy,QualType & FlagsTy)4221 static void getDependTypes(ASTContext &C, QualType &KmpDependInfoTy,
4222 QualType &FlagsTy) {
4223 FlagsTy = C.getIntTypeForBitwidth(C.getTypeSize(C.BoolTy), /*Signed=*/false);
4224 if (KmpDependInfoTy.isNull()) {
4225 RecordDecl *KmpDependInfoRD = C.buildImplicitRecord("kmp_depend_info");
4226 KmpDependInfoRD->startDefinition();
4227 addFieldToRecordDecl(C, KmpDependInfoRD, C.getIntPtrType());
4228 addFieldToRecordDecl(C, KmpDependInfoRD, C.getSizeType());
4229 addFieldToRecordDecl(C, KmpDependInfoRD, FlagsTy);
4230 KmpDependInfoRD->completeDefinition();
4231 KmpDependInfoTy = C.getRecordType(KmpDependInfoRD);
4232 }
4233 }
4234
4235 std::pair<llvm::Value *, LValue>
getDepobjElements(CodeGenFunction & CGF,LValue DepobjLVal,SourceLocation Loc)4236 CGOpenMPRuntime::getDepobjElements(CodeGenFunction &CGF, LValue DepobjLVal,
4237 SourceLocation Loc) {
4238 ASTContext &C = CGM.getContext();
4239 QualType FlagsTy;
4240 getDependTypes(C, KmpDependInfoTy, FlagsTy);
4241 RecordDecl *KmpDependInfoRD =
4242 cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
4243 QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy);
4244 LValue Base = CGF.EmitLoadOfPointerLValue(
4245 CGF.Builder.CreateElementBitCast(
4246 DepobjLVal.getAddress(CGF),
4247 CGF.ConvertTypeForMem(KmpDependInfoPtrTy)),
4248 KmpDependInfoPtrTy->castAs<PointerType>());
4249 Address DepObjAddr = CGF.Builder.CreateGEP(
4250 Base.getAddress(CGF),
4251 llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true));
4252 LValue NumDepsBase = CGF.MakeAddrLValue(
4253 DepObjAddr, KmpDependInfoTy, Base.getBaseInfo(), Base.getTBAAInfo());
4254 // NumDeps = deps[i].base_addr;
4255 LValue BaseAddrLVal = CGF.EmitLValueForField(
4256 NumDepsBase,
4257 *std::next(KmpDependInfoRD->field_begin(),
4258 static_cast<unsigned int>(RTLDependInfoFields::BaseAddr)));
4259 llvm::Value *NumDeps = CGF.EmitLoadOfScalar(BaseAddrLVal, Loc);
4260 return std::make_pair(NumDeps, Base);
4261 }
4262
emitDependData(CodeGenFunction & CGF,QualType & KmpDependInfoTy,llvm::PointerUnion<unsigned *,LValue * > Pos,const OMPTaskDataTy::DependData & Data,Address DependenciesArray)4263 static void emitDependData(CodeGenFunction &CGF, QualType &KmpDependInfoTy,
4264 llvm::PointerUnion<unsigned *, LValue *> Pos,
4265 const OMPTaskDataTy::DependData &Data,
4266 Address DependenciesArray) {
4267 CodeGenModule &CGM = CGF.CGM;
4268 ASTContext &C = CGM.getContext();
4269 QualType FlagsTy;
4270 getDependTypes(C, KmpDependInfoTy, FlagsTy);
4271 RecordDecl *KmpDependInfoRD =
4272 cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
4273 llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy);
4274
4275 OMPIteratorGeneratorScope IteratorScope(
4276 CGF, cast_or_null<OMPIteratorExpr>(
4277 Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts()
4278 : nullptr));
4279 for (const Expr *E : Data.DepExprs) {
4280 llvm::Value *Addr;
4281 llvm::Value *Size;
4282
4283 // The expression will be a nullptr in the 'omp_all_memory' case.
4284 if (E) {
4285 std::tie(Addr, Size) = getPointerAndSize(CGF, E);
4286 Addr = CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy);
4287 } else {
4288 Addr = llvm::ConstantInt::get(CGF.IntPtrTy, 0);
4289 Size = llvm::ConstantInt::get(CGF.SizeTy, 0);
4290 }
4291 LValue Base;
4292 if (unsigned *P = Pos.dyn_cast<unsigned *>()) {
4293 Base = CGF.MakeAddrLValue(
4294 CGF.Builder.CreateConstGEP(DependenciesArray, *P), KmpDependInfoTy);
4295 } else {
4296 assert(E && "Expected a non-null expression");
4297 LValue &PosLVal = *Pos.get<LValue *>();
4298 llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc());
4299 Base = CGF.MakeAddrLValue(
4300 CGF.Builder.CreateGEP(DependenciesArray, Idx), KmpDependInfoTy);
4301 }
4302 // deps[i].base_addr = &<Dependencies[i].second>;
4303 LValue BaseAddrLVal = CGF.EmitLValueForField(
4304 Base,
4305 *std::next(KmpDependInfoRD->field_begin(),
4306 static_cast<unsigned int>(RTLDependInfoFields::BaseAddr)));
4307 CGF.EmitStoreOfScalar(Addr, BaseAddrLVal);
4308 // deps[i].len = sizeof(<Dependencies[i].second>);
4309 LValue LenLVal = CGF.EmitLValueForField(
4310 Base, *std::next(KmpDependInfoRD->field_begin(),
4311 static_cast<unsigned int>(RTLDependInfoFields::Len)));
4312 CGF.EmitStoreOfScalar(Size, LenLVal);
4313 // deps[i].flags = <Dependencies[i].first>;
4314 RTLDependenceKindTy DepKind = translateDependencyKind(Data.DepKind);
4315 LValue FlagsLVal = CGF.EmitLValueForField(
4316 Base,
4317 *std::next(KmpDependInfoRD->field_begin(),
4318 static_cast<unsigned int>(RTLDependInfoFields::Flags)));
4319 CGF.EmitStoreOfScalar(
4320 llvm::ConstantInt::get(LLVMFlagsTy, static_cast<unsigned int>(DepKind)),
4321 FlagsLVal);
4322 if (unsigned *P = Pos.dyn_cast<unsigned *>()) {
4323 ++(*P);
4324 } else {
4325 LValue &PosLVal = *Pos.get<LValue *>();
4326 llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc());
4327 Idx = CGF.Builder.CreateNUWAdd(Idx,
4328 llvm::ConstantInt::get(Idx->getType(), 1));
4329 CGF.EmitStoreOfScalar(Idx, PosLVal);
4330 }
4331 }
4332 }
4333
emitDepobjElementsSizes(CodeGenFunction & CGF,QualType & KmpDependInfoTy,const OMPTaskDataTy::DependData & Data)4334 SmallVector<llvm::Value *, 4> CGOpenMPRuntime::emitDepobjElementsSizes(
4335 CodeGenFunction &CGF, QualType &KmpDependInfoTy,
4336 const OMPTaskDataTy::DependData &Data) {
4337 assert(Data.DepKind == OMPC_DEPEND_depobj &&
4338 "Expected depobj dependency kind.");
4339 SmallVector<llvm::Value *, 4> Sizes;
4340 SmallVector<LValue, 4> SizeLVals;
4341 ASTContext &C = CGF.getContext();
4342 {
4343 OMPIteratorGeneratorScope IteratorScope(
4344 CGF, cast_or_null<OMPIteratorExpr>(
4345 Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts()
4346 : nullptr));
4347 for (const Expr *E : Data.DepExprs) {
4348 llvm::Value *NumDeps;
4349 LValue Base;
4350 LValue DepobjLVal = CGF.EmitLValue(E->IgnoreParenImpCasts());
4351 std::tie(NumDeps, Base) =
4352 getDepobjElements(CGF, DepobjLVal, E->getExprLoc());
4353 LValue NumLVal = CGF.MakeAddrLValue(
4354 CGF.CreateMemTemp(C.getUIntPtrType(), "depobj.size.addr"),
4355 C.getUIntPtrType());
4356 CGF.Builder.CreateStore(llvm::ConstantInt::get(CGF.IntPtrTy, 0),
4357 NumLVal.getAddress(CGF));
4358 llvm::Value *PrevVal = CGF.EmitLoadOfScalar(NumLVal, E->getExprLoc());
4359 llvm::Value *Add = CGF.Builder.CreateNUWAdd(PrevVal, NumDeps);
4360 CGF.EmitStoreOfScalar(Add, NumLVal);
4361 SizeLVals.push_back(NumLVal);
4362 }
4363 }
4364 for (unsigned I = 0, E = SizeLVals.size(); I < E; ++I) {
4365 llvm::Value *Size =
4366 CGF.EmitLoadOfScalar(SizeLVals[I], Data.DepExprs[I]->getExprLoc());
4367 Sizes.push_back(Size);
4368 }
4369 return Sizes;
4370 }
4371
emitDepobjElements(CodeGenFunction & CGF,QualType & KmpDependInfoTy,LValue PosLVal,const OMPTaskDataTy::DependData & Data,Address DependenciesArray)4372 void CGOpenMPRuntime::emitDepobjElements(CodeGenFunction &CGF,
4373 QualType &KmpDependInfoTy,
4374 LValue PosLVal,
4375 const OMPTaskDataTy::DependData &Data,
4376 Address DependenciesArray) {
4377 assert(Data.DepKind == OMPC_DEPEND_depobj &&
4378 "Expected depobj dependency kind.");
4379 llvm::Value *ElSize = CGF.getTypeSize(KmpDependInfoTy);
4380 {
4381 OMPIteratorGeneratorScope IteratorScope(
4382 CGF, cast_or_null<OMPIteratorExpr>(
4383 Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts()
4384 : nullptr));
4385 for (unsigned I = 0, End = Data.DepExprs.size(); I < End; ++I) {
4386 const Expr *E = Data.DepExprs[I];
4387 llvm::Value *NumDeps;
4388 LValue Base;
4389 LValue DepobjLVal = CGF.EmitLValue(E->IgnoreParenImpCasts());
4390 std::tie(NumDeps, Base) =
4391 getDepobjElements(CGF, DepobjLVal, E->getExprLoc());
4392
4393 // memcopy dependency data.
4394 llvm::Value *Size = CGF.Builder.CreateNUWMul(
4395 ElSize,
4396 CGF.Builder.CreateIntCast(NumDeps, CGF.SizeTy, /*isSigned=*/false));
4397 llvm::Value *Pos = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc());
4398 Address DepAddr = CGF.Builder.CreateGEP(DependenciesArray, Pos);
4399 CGF.Builder.CreateMemCpy(DepAddr, Base.getAddress(CGF), Size);
4400
4401 // Increase pos.
4402 // pos += size;
4403 llvm::Value *Add = CGF.Builder.CreateNUWAdd(Pos, NumDeps);
4404 CGF.EmitStoreOfScalar(Add, PosLVal);
4405 }
4406 }
4407 }
4408
emitDependClause(CodeGenFunction & CGF,ArrayRef<OMPTaskDataTy::DependData> Dependencies,SourceLocation Loc)4409 std::pair<llvm::Value *, Address> CGOpenMPRuntime::emitDependClause(
4410 CodeGenFunction &CGF, ArrayRef<OMPTaskDataTy::DependData> Dependencies,
4411 SourceLocation Loc) {
4412 if (llvm::all_of(Dependencies, [](const OMPTaskDataTy::DependData &D) {
4413 return D.DepExprs.empty();
4414 }))
4415 return std::make_pair(nullptr, Address::invalid());
4416 // Process list of dependencies.
4417 ASTContext &C = CGM.getContext();
4418 Address DependenciesArray = Address::invalid();
4419 llvm::Value *NumOfElements = nullptr;
4420 unsigned NumDependencies = std::accumulate(
4421 Dependencies.begin(), Dependencies.end(), 0,
4422 [](unsigned V, const OMPTaskDataTy::DependData &D) {
4423 return D.DepKind == OMPC_DEPEND_depobj
4424 ? V
4425 : (V + (D.IteratorExpr ? 0 : D.DepExprs.size()));
4426 });
4427 QualType FlagsTy;
4428 getDependTypes(C, KmpDependInfoTy, FlagsTy);
4429 bool HasDepobjDeps = false;
4430 bool HasRegularWithIterators = false;
4431 llvm::Value *NumOfDepobjElements = llvm::ConstantInt::get(CGF.IntPtrTy, 0);
4432 llvm::Value *NumOfRegularWithIterators =
4433 llvm::ConstantInt::get(CGF.IntPtrTy, 0);
4434 // Calculate number of depobj dependencies and regular deps with the
4435 // iterators.
4436 for (const OMPTaskDataTy::DependData &D : Dependencies) {
4437 if (D.DepKind == OMPC_DEPEND_depobj) {
4438 SmallVector<llvm::Value *, 4> Sizes =
4439 emitDepobjElementsSizes(CGF, KmpDependInfoTy, D);
4440 for (llvm::Value *Size : Sizes) {
4441 NumOfDepobjElements =
4442 CGF.Builder.CreateNUWAdd(NumOfDepobjElements, Size);
4443 }
4444 HasDepobjDeps = true;
4445 continue;
4446 }
4447 // Include number of iterations, if any.
4448
4449 if (const auto *IE = cast_or_null<OMPIteratorExpr>(D.IteratorExpr)) {
4450 for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) {
4451 llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper);
4452 Sz = CGF.Builder.CreateIntCast(Sz, CGF.IntPtrTy, /*isSigned=*/false);
4453 llvm::Value *NumClauseDeps = CGF.Builder.CreateNUWMul(
4454 Sz, llvm::ConstantInt::get(CGF.IntPtrTy, D.DepExprs.size()));
4455 NumOfRegularWithIterators =
4456 CGF.Builder.CreateNUWAdd(NumOfRegularWithIterators, NumClauseDeps);
4457 }
4458 HasRegularWithIterators = true;
4459 continue;
4460 }
4461 }
4462
4463 QualType KmpDependInfoArrayTy;
4464 if (HasDepobjDeps || HasRegularWithIterators) {
4465 NumOfElements = llvm::ConstantInt::get(CGM.IntPtrTy, NumDependencies,
4466 /*isSigned=*/false);
4467 if (HasDepobjDeps) {
4468 NumOfElements =
4469 CGF.Builder.CreateNUWAdd(NumOfDepobjElements, NumOfElements);
4470 }
4471 if (HasRegularWithIterators) {
4472 NumOfElements =
4473 CGF.Builder.CreateNUWAdd(NumOfRegularWithIterators, NumOfElements);
4474 }
4475 auto *OVE = new (C) OpaqueValueExpr(
4476 Loc, C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0),
4477 VK_PRValue);
4478 CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, OVE,
4479 RValue::get(NumOfElements));
4480 KmpDependInfoArrayTy =
4481 C.getVariableArrayType(KmpDependInfoTy, OVE, ArrayType::Normal,
4482 /*IndexTypeQuals=*/0, SourceRange(Loc, Loc));
4483 // CGF.EmitVariablyModifiedType(KmpDependInfoArrayTy);
4484 // Properly emit variable-sized array.
4485 auto *PD = ImplicitParamDecl::Create(C, KmpDependInfoArrayTy,
4486 ImplicitParamDecl::Other);
4487 CGF.EmitVarDecl(*PD);
4488 DependenciesArray = CGF.GetAddrOfLocalVar(PD);
4489 NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty,
4490 /*isSigned=*/false);
4491 } else {
4492 KmpDependInfoArrayTy = C.getConstantArrayType(
4493 KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies), nullptr,
4494 ArrayType::Normal, /*IndexTypeQuals=*/0);
4495 DependenciesArray =
4496 CGF.CreateMemTemp(KmpDependInfoArrayTy, ".dep.arr.addr");
4497 DependenciesArray = CGF.Builder.CreateConstArrayGEP(DependenciesArray, 0);
4498 NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumDependencies,
4499 /*isSigned=*/false);
4500 }
4501 unsigned Pos = 0;
4502 for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) {
4503 if (Dependencies[I].DepKind == OMPC_DEPEND_depobj ||
4504 Dependencies[I].IteratorExpr)
4505 continue;
4506 emitDependData(CGF, KmpDependInfoTy, &Pos, Dependencies[I],
4507 DependenciesArray);
4508 }
4509 // Copy regular dependencies with iterators.
4510 LValue PosLVal = CGF.MakeAddrLValue(
4511 CGF.CreateMemTemp(C.getSizeType(), "dep.counter.addr"), C.getSizeType());
4512 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal);
4513 for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) {
4514 if (Dependencies[I].DepKind == OMPC_DEPEND_depobj ||
4515 !Dependencies[I].IteratorExpr)
4516 continue;
4517 emitDependData(CGF, KmpDependInfoTy, &PosLVal, Dependencies[I],
4518 DependenciesArray);
4519 }
4520 // Copy final depobj arrays without iterators.
4521 if (HasDepobjDeps) {
4522 for (unsigned I = 0, End = Dependencies.size(); I < End; ++I) {
4523 if (Dependencies[I].DepKind != OMPC_DEPEND_depobj)
4524 continue;
4525 emitDepobjElements(CGF, KmpDependInfoTy, PosLVal, Dependencies[I],
4526 DependenciesArray);
4527 }
4528 }
4529 DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4530 DependenciesArray, CGF.VoidPtrTy, CGF.Int8Ty);
4531 return std::make_pair(NumOfElements, DependenciesArray);
4532 }
4533
emitDepobjDependClause(CodeGenFunction & CGF,const OMPTaskDataTy::DependData & Dependencies,SourceLocation Loc)4534 Address CGOpenMPRuntime::emitDepobjDependClause(
4535 CodeGenFunction &CGF, const OMPTaskDataTy::DependData &Dependencies,
4536 SourceLocation Loc) {
4537 if (Dependencies.DepExprs.empty())
4538 return Address::invalid();
4539 // Process list of dependencies.
4540 ASTContext &C = CGM.getContext();
4541 Address DependenciesArray = Address::invalid();
4542 unsigned NumDependencies = Dependencies.DepExprs.size();
4543 QualType FlagsTy;
4544 getDependTypes(C, KmpDependInfoTy, FlagsTy);
4545 RecordDecl *KmpDependInfoRD =
4546 cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
4547
4548 llvm::Value *Size;
4549 // Define type kmp_depend_info[<Dependencies.size()>];
4550 // For depobj reserve one extra element to store the number of elements.
4551 // It is required to handle depobj(x) update(in) construct.
4552 // kmp_depend_info[<Dependencies.size()>] deps;
4553 llvm::Value *NumDepsVal;
4554 CharUnits Align = C.getTypeAlignInChars(KmpDependInfoTy);
4555 if (const auto *IE =
4556 cast_or_null<OMPIteratorExpr>(Dependencies.IteratorExpr)) {
4557 NumDepsVal = llvm::ConstantInt::get(CGF.SizeTy, 1);
4558 for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) {
4559 llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper);
4560 Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false);
4561 NumDepsVal = CGF.Builder.CreateNUWMul(NumDepsVal, Sz);
4562 }
4563 Size = CGF.Builder.CreateNUWAdd(llvm::ConstantInt::get(CGF.SizeTy, 1),
4564 NumDepsVal);
4565 CharUnits SizeInBytes =
4566 C.getTypeSizeInChars(KmpDependInfoTy).alignTo(Align);
4567 llvm::Value *RecSize = CGM.getSize(SizeInBytes);
4568 Size = CGF.Builder.CreateNUWMul(Size, RecSize);
4569 NumDepsVal =
4570 CGF.Builder.CreateIntCast(NumDepsVal, CGF.IntPtrTy, /*isSigned=*/false);
4571 } else {
4572 QualType KmpDependInfoArrayTy = C.getConstantArrayType(
4573 KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies + 1),
4574 nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0);
4575 CharUnits Sz = C.getTypeSizeInChars(KmpDependInfoArrayTy);
4576 Size = CGM.getSize(Sz.alignTo(Align));
4577 NumDepsVal = llvm::ConstantInt::get(CGF.IntPtrTy, NumDependencies);
4578 }
4579 // Need to allocate on the dynamic memory.
4580 llvm::Value *ThreadID = getThreadID(CGF, Loc);
4581 // Use default allocator.
4582 llvm::Value *Allocator = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4583 llvm::Value *Args[] = {ThreadID, Size, Allocator};
4584
4585 llvm::Value *Addr =
4586 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
4587 CGM.getModule(), OMPRTL___kmpc_alloc),
4588 Args, ".dep.arr.addr");
4589 llvm::Type *KmpDependInfoLlvmTy = CGF.ConvertTypeForMem(KmpDependInfoTy);
4590 Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4591 Addr, KmpDependInfoLlvmTy->getPointerTo());
4592 DependenciesArray = Address(Addr, KmpDependInfoLlvmTy, Align);
4593 // Write number of elements in the first element of array for depobj.
4594 LValue Base = CGF.MakeAddrLValue(DependenciesArray, KmpDependInfoTy);
4595 // deps[i].base_addr = NumDependencies;
4596 LValue BaseAddrLVal = CGF.EmitLValueForField(
4597 Base,
4598 *std::next(KmpDependInfoRD->field_begin(),
4599 static_cast<unsigned int>(RTLDependInfoFields::BaseAddr)));
4600 CGF.EmitStoreOfScalar(NumDepsVal, BaseAddrLVal);
4601 llvm::PointerUnion<unsigned *, LValue *> Pos;
4602 unsigned Idx = 1;
4603 LValue PosLVal;
4604 if (Dependencies.IteratorExpr) {
4605 PosLVal = CGF.MakeAddrLValue(
4606 CGF.CreateMemTemp(C.getSizeType(), "iterator.counter.addr"),
4607 C.getSizeType());
4608 CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Idx), PosLVal,
4609 /*IsInit=*/true);
4610 Pos = &PosLVal;
4611 } else {
4612 Pos = &Idx;
4613 }
4614 emitDependData(CGF, KmpDependInfoTy, Pos, Dependencies, DependenciesArray);
4615 DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4616 CGF.Builder.CreateConstGEP(DependenciesArray, 1), CGF.VoidPtrTy,
4617 CGF.Int8Ty);
4618 return DependenciesArray;
4619 }
4620
emitDestroyClause(CodeGenFunction & CGF,LValue DepobjLVal,SourceLocation Loc)4621 void CGOpenMPRuntime::emitDestroyClause(CodeGenFunction &CGF, LValue DepobjLVal,
4622 SourceLocation Loc) {
4623 ASTContext &C = CGM.getContext();
4624 QualType FlagsTy;
4625 getDependTypes(C, KmpDependInfoTy, FlagsTy);
4626 LValue Base = CGF.EmitLoadOfPointerLValue(
4627 DepobjLVal.getAddress(CGF), C.VoidPtrTy.castAs<PointerType>());
4628 QualType KmpDependInfoPtrTy = C.getPointerType(KmpDependInfoTy);
4629 Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4630 Base.getAddress(CGF), CGF.ConvertTypeForMem(KmpDependInfoPtrTy),
4631 CGF.ConvertTypeForMem(KmpDependInfoTy));
4632 llvm::Value *DepObjAddr = CGF.Builder.CreateGEP(
4633 Addr.getElementType(), Addr.getPointer(),
4634 llvm::ConstantInt::get(CGF.IntPtrTy, -1, /*isSigned=*/true));
4635 DepObjAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(DepObjAddr,
4636 CGF.VoidPtrTy);
4637 llvm::Value *ThreadID = getThreadID(CGF, Loc);
4638 // Use default allocator.
4639 llvm::Value *Allocator = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4640 llvm::Value *Args[] = {ThreadID, DepObjAddr, Allocator};
4641
4642 // _kmpc_free(gtid, addr, nullptr);
4643 (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
4644 CGM.getModule(), OMPRTL___kmpc_free),
4645 Args);
4646 }
4647
emitUpdateClause(CodeGenFunction & CGF,LValue DepobjLVal,OpenMPDependClauseKind NewDepKind,SourceLocation Loc)4648 void CGOpenMPRuntime::emitUpdateClause(CodeGenFunction &CGF, LValue DepobjLVal,
4649 OpenMPDependClauseKind NewDepKind,
4650 SourceLocation Loc) {
4651 ASTContext &C = CGM.getContext();
4652 QualType FlagsTy;
4653 getDependTypes(C, KmpDependInfoTy, FlagsTy);
4654 RecordDecl *KmpDependInfoRD =
4655 cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
4656 llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy);
4657 llvm::Value *NumDeps;
4658 LValue Base;
4659 std::tie(NumDeps, Base) = getDepobjElements(CGF, DepobjLVal, Loc);
4660
4661 Address Begin = Base.getAddress(CGF);
4662 // Cast from pointer to array type to pointer to single element.
4663 llvm::Value *End = CGF.Builder.CreateGEP(
4664 Begin.getElementType(), Begin.getPointer(), NumDeps);
4665 // The basic structure here is a while-do loop.
4666 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.body");
4667 llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.done");
4668 llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock();
4669 CGF.EmitBlock(BodyBB);
4670 llvm::PHINode *ElementPHI =
4671 CGF.Builder.CreatePHI(Begin.getType(), 2, "omp.elementPast");
4672 ElementPHI->addIncoming(Begin.getPointer(), EntryBB);
4673 Begin = Begin.withPointer(ElementPHI);
4674 Base = CGF.MakeAddrLValue(Begin, KmpDependInfoTy, Base.getBaseInfo(),
4675 Base.getTBAAInfo());
4676 // deps[i].flags = NewDepKind;
4677 RTLDependenceKindTy DepKind = translateDependencyKind(NewDepKind);
4678 LValue FlagsLVal = CGF.EmitLValueForField(
4679 Base, *std::next(KmpDependInfoRD->field_begin(),
4680 static_cast<unsigned int>(RTLDependInfoFields::Flags)));
4681 CGF.EmitStoreOfScalar(
4682 llvm::ConstantInt::get(LLVMFlagsTy, static_cast<unsigned int>(DepKind)),
4683 FlagsLVal);
4684
4685 // Shift the address forward by one element.
4686 Address ElementNext =
4687 CGF.Builder.CreateConstGEP(Begin, /*Index=*/1, "omp.elementNext");
4688 ElementPHI->addIncoming(ElementNext.getPointer(),
4689 CGF.Builder.GetInsertBlock());
4690 llvm::Value *IsEmpty =
4691 CGF.Builder.CreateICmpEQ(ElementNext.getPointer(), End, "omp.isempty");
4692 CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
4693 // Done.
4694 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
4695 }
4696
emitTaskCall(CodeGenFunction & CGF,SourceLocation Loc,const OMPExecutableDirective & D,llvm::Function * TaskFunction,QualType SharedsTy,Address Shareds,const Expr * IfCond,const OMPTaskDataTy & Data)4697 void CGOpenMPRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc,
4698 const OMPExecutableDirective &D,
4699 llvm::Function *TaskFunction,
4700 QualType SharedsTy, Address Shareds,
4701 const Expr *IfCond,
4702 const OMPTaskDataTy &Data) {
4703 if (!CGF.HaveInsertPoint())
4704 return;
4705
4706 TaskResultTy Result =
4707 emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data);
4708 llvm::Value *NewTask = Result.NewTask;
4709 llvm::Function *TaskEntry = Result.TaskEntry;
4710 llvm::Value *NewTaskNewTaskTTy = Result.NewTaskNewTaskTTy;
4711 LValue TDBase = Result.TDBase;
4712 const RecordDecl *KmpTaskTQTyRD = Result.KmpTaskTQTyRD;
4713 // Process list of dependences.
4714 Address DependenciesArray = Address::invalid();
4715 llvm::Value *NumOfElements;
4716 std::tie(NumOfElements, DependenciesArray) =
4717 emitDependClause(CGF, Data.Dependences, Loc);
4718
4719 // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc()
4720 // libcall.
4721 // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid,
4722 // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,
4723 // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) if dependence
4724 // list is not empty
4725 llvm::Value *ThreadID = getThreadID(CGF, Loc);
4726 llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc);
4727 llvm::Value *TaskArgs[] = { UpLoc, ThreadID, NewTask };
4728 llvm::Value *DepTaskArgs[7];
4729 if (!Data.Dependences.empty()) {
4730 DepTaskArgs[0] = UpLoc;
4731 DepTaskArgs[1] = ThreadID;
4732 DepTaskArgs[2] = NewTask;
4733 DepTaskArgs[3] = NumOfElements;
4734 DepTaskArgs[4] = DependenciesArray.getPointer();
4735 DepTaskArgs[5] = CGF.Builder.getInt32(0);
4736 DepTaskArgs[6] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4737 }
4738 auto &&ThenCodeGen = [this, &Data, TDBase, KmpTaskTQTyRD, &TaskArgs,
4739 &DepTaskArgs](CodeGenFunction &CGF, PrePostActionTy &) {
4740 if (!Data.Tied) {
4741 auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId);
4742 LValue PartIdLVal = CGF.EmitLValueForField(TDBase, *PartIdFI);
4743 CGF.EmitStoreOfScalar(CGF.Builder.getInt32(0), PartIdLVal);
4744 }
4745 if (!Data.Dependences.empty()) {
4746 CGF.EmitRuntimeCall(
4747 OMPBuilder.getOrCreateRuntimeFunction(
4748 CGM.getModule(), OMPRTL___kmpc_omp_task_with_deps),
4749 DepTaskArgs);
4750 } else {
4751 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
4752 CGM.getModule(), OMPRTL___kmpc_omp_task),
4753 TaskArgs);
4754 }
4755 // Check if parent region is untied and build return for untied task;
4756 if (auto *Region =
4757 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
4758 Region->emitUntiedSwitch(CGF);
4759 };
4760
4761 llvm::Value *DepWaitTaskArgs[7];
4762 if (!Data.Dependences.empty()) {
4763 DepWaitTaskArgs[0] = UpLoc;
4764 DepWaitTaskArgs[1] = ThreadID;
4765 DepWaitTaskArgs[2] = NumOfElements;
4766 DepWaitTaskArgs[3] = DependenciesArray.getPointer();
4767 DepWaitTaskArgs[4] = CGF.Builder.getInt32(0);
4768 DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
4769 DepWaitTaskArgs[6] =
4770 llvm::ConstantInt::get(CGF.Int32Ty, Data.HasNowaitClause);
4771 }
4772 auto &M = CGM.getModule();
4773 auto &&ElseCodeGen = [this, &M, &TaskArgs, ThreadID, NewTaskNewTaskTTy,
4774 TaskEntry, &Data, &DepWaitTaskArgs,
4775 Loc](CodeGenFunction &CGF, PrePostActionTy &) {
4776 CodeGenFunction::RunCleanupsScope LocalScope(CGF);
4777 // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid,
4778 // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32
4779 // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info
4780 // is specified.
4781 if (!Data.Dependences.empty())
4782 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
4783 M, OMPRTL___kmpc_omp_taskwait_deps_51),
4784 DepWaitTaskArgs);
4785 // Call proxy_task_entry(gtid, new_task);
4786 auto &&CodeGen = [TaskEntry, ThreadID, NewTaskNewTaskTTy,
4787 Loc](CodeGenFunction &CGF, PrePostActionTy &Action) {
4788 Action.Enter(CGF);
4789 llvm::Value *OutlinedFnArgs[] = {ThreadID, NewTaskNewTaskTTy};
4790 CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskEntry,
4791 OutlinedFnArgs);
4792 };
4793
4794 // Build void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid,
4795 // kmp_task_t *new_task);
4796 // Build void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid,
4797 // kmp_task_t *new_task);
4798 RegionCodeGenTy RCG(CodeGen);
4799 CommonActionTy Action(OMPBuilder.getOrCreateRuntimeFunction(
4800 M, OMPRTL___kmpc_omp_task_begin_if0),
4801 TaskArgs,
4802 OMPBuilder.getOrCreateRuntimeFunction(
4803 M, OMPRTL___kmpc_omp_task_complete_if0),
4804 TaskArgs);
4805 RCG.setAction(Action);
4806 RCG(CGF);
4807 };
4808
4809 if (IfCond) {
4810 emitIfClause(CGF, IfCond, ThenCodeGen, ElseCodeGen);
4811 } else {
4812 RegionCodeGenTy ThenRCG(ThenCodeGen);
4813 ThenRCG(CGF);
4814 }
4815 }
4816
emitTaskLoopCall(CodeGenFunction & CGF,SourceLocation Loc,const OMPLoopDirective & D,llvm::Function * TaskFunction,QualType SharedsTy,Address Shareds,const Expr * IfCond,const OMPTaskDataTy & Data)4817 void CGOpenMPRuntime::emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc,
4818 const OMPLoopDirective &D,
4819 llvm::Function *TaskFunction,
4820 QualType SharedsTy, Address Shareds,
4821 const Expr *IfCond,
4822 const OMPTaskDataTy &Data) {
4823 if (!CGF.HaveInsertPoint())
4824 return;
4825 TaskResultTy Result =
4826 emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data);
4827 // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc()
4828 // libcall.
4829 // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int
4830 // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int
4831 // sched, kmp_uint64 grainsize, void *task_dup);
4832 llvm::Value *ThreadID = getThreadID(CGF, Loc);
4833 llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc);
4834 llvm::Value *IfVal;
4835 if (IfCond) {
4836 IfVal = CGF.Builder.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.IntTy,
4837 /*isSigned=*/true);
4838 } else {
4839 IfVal = llvm::ConstantInt::getSigned(CGF.IntTy, /*V=*/1);
4840 }
4841
4842 LValue LBLVal = CGF.EmitLValueForField(
4843 Result.TDBase,
4844 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound));
4845 const auto *LBVar =
4846 cast<VarDecl>(cast<DeclRefExpr>(D.getLowerBoundVariable())->getDecl());
4847 CGF.EmitAnyExprToMem(LBVar->getInit(), LBLVal.getAddress(CGF),
4848 LBLVal.getQuals(),
4849 /*IsInitializer=*/true);
4850 LValue UBLVal = CGF.EmitLValueForField(
4851 Result.TDBase,
4852 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound));
4853 const auto *UBVar =
4854 cast<VarDecl>(cast<DeclRefExpr>(D.getUpperBoundVariable())->getDecl());
4855 CGF.EmitAnyExprToMem(UBVar->getInit(), UBLVal.getAddress(CGF),
4856 UBLVal.getQuals(),
4857 /*IsInitializer=*/true);
4858 LValue StLVal = CGF.EmitLValueForField(
4859 Result.TDBase,
4860 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTStride));
4861 const auto *StVar =
4862 cast<VarDecl>(cast<DeclRefExpr>(D.getStrideVariable())->getDecl());
4863 CGF.EmitAnyExprToMem(StVar->getInit(), StLVal.getAddress(CGF),
4864 StLVal.getQuals(),
4865 /*IsInitializer=*/true);
4866 // Store reductions address.
4867 LValue RedLVal = CGF.EmitLValueForField(
4868 Result.TDBase,
4869 *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTReductions));
4870 if (Data.Reductions) {
4871 CGF.EmitStoreOfScalar(Data.Reductions, RedLVal);
4872 } else {
4873 CGF.EmitNullInitialization(RedLVal.getAddress(CGF),
4874 CGF.getContext().VoidPtrTy);
4875 }
4876 enum { NoSchedule = 0, Grainsize = 1, NumTasks = 2 };
4877 llvm::Value *TaskArgs[] = {
4878 UpLoc,
4879 ThreadID,
4880 Result.NewTask,
4881 IfVal,
4882 LBLVal.getPointer(CGF),
4883 UBLVal.getPointer(CGF),
4884 CGF.EmitLoadOfScalar(StLVal, Loc),
4885 llvm::ConstantInt::getSigned(
4886 CGF.IntTy, 1), // Always 1 because taskgroup emitted by the compiler
4887 llvm::ConstantInt::getSigned(
4888 CGF.IntTy, Data.Schedule.getPointer()
4889 ? Data.Schedule.getInt() ? NumTasks : Grainsize
4890 : NoSchedule),
4891 Data.Schedule.getPointer()
4892 ? CGF.Builder.CreateIntCast(Data.Schedule.getPointer(), CGF.Int64Ty,
4893 /*isSigned=*/false)
4894 : llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/0),
4895 Result.TaskDupFn ? CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4896 Result.TaskDupFn, CGF.VoidPtrTy)
4897 : llvm::ConstantPointerNull::get(CGF.VoidPtrTy)};
4898 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
4899 CGM.getModule(), OMPRTL___kmpc_taskloop),
4900 TaskArgs);
4901 }
4902
4903 /// Emit reduction operation for each element of array (required for
4904 /// array sections) LHS op = RHS.
4905 /// \param Type Type of array.
4906 /// \param LHSVar Variable on the left side of the reduction operation
4907 /// (references element of array in original variable).
4908 /// \param RHSVar Variable on the right side of the reduction operation
4909 /// (references element of array in original variable).
4910 /// \param RedOpGen Generator of reduction operation with use of LHSVar and
4911 /// RHSVar.
EmitOMPAggregateReduction(CodeGenFunction & CGF,QualType Type,const VarDecl * LHSVar,const VarDecl * RHSVar,const llvm::function_ref<void (CodeGenFunction & CGF,const Expr *,const Expr *,const Expr *)> & RedOpGen,const Expr * XExpr=nullptr,const Expr * EExpr=nullptr,const Expr * UpExpr=nullptr)4912 static void EmitOMPAggregateReduction(
4913 CodeGenFunction &CGF, QualType Type, const VarDecl *LHSVar,
4914 const VarDecl *RHSVar,
4915 const llvm::function_ref<void(CodeGenFunction &CGF, const Expr *,
4916 const Expr *, const Expr *)> &RedOpGen,
4917 const Expr *XExpr = nullptr, const Expr *EExpr = nullptr,
4918 const Expr *UpExpr = nullptr) {
4919 // Perform element-by-element initialization.
4920 QualType ElementTy;
4921 Address LHSAddr = CGF.GetAddrOfLocalVar(LHSVar);
4922 Address RHSAddr = CGF.GetAddrOfLocalVar(RHSVar);
4923
4924 // Drill down to the base element type on both arrays.
4925 const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe();
4926 llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, LHSAddr);
4927
4928 llvm::Value *RHSBegin = RHSAddr.getPointer();
4929 llvm::Value *LHSBegin = LHSAddr.getPointer();
4930 // Cast from pointer to array type to pointer to single element.
4931 llvm::Value *LHSEnd =
4932 CGF.Builder.CreateGEP(LHSAddr.getElementType(), LHSBegin, NumElements);
4933 // The basic structure here is a while-do loop.
4934 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arraycpy.body");
4935 llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arraycpy.done");
4936 llvm::Value *IsEmpty =
4937 CGF.Builder.CreateICmpEQ(LHSBegin, LHSEnd, "omp.arraycpy.isempty");
4938 CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
4939
4940 // Enter the loop body, making that address the current address.
4941 llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock();
4942 CGF.EmitBlock(BodyBB);
4943
4944 CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy);
4945
4946 llvm::PHINode *RHSElementPHI = CGF.Builder.CreatePHI(
4947 RHSBegin->getType(), 2, "omp.arraycpy.srcElementPast");
4948 RHSElementPHI->addIncoming(RHSBegin, EntryBB);
4949 Address RHSElementCurrent(
4950 RHSElementPHI, RHSAddr.getElementType(),
4951 RHSAddr.getAlignment().alignmentOfArrayElement(ElementSize));
4952
4953 llvm::PHINode *LHSElementPHI = CGF.Builder.CreatePHI(
4954 LHSBegin->getType(), 2, "omp.arraycpy.destElementPast");
4955 LHSElementPHI->addIncoming(LHSBegin, EntryBB);
4956 Address LHSElementCurrent(
4957 LHSElementPHI, LHSAddr.getElementType(),
4958 LHSAddr.getAlignment().alignmentOfArrayElement(ElementSize));
4959
4960 // Emit copy.
4961 CodeGenFunction::OMPPrivateScope Scope(CGF);
4962 Scope.addPrivate(LHSVar, LHSElementCurrent);
4963 Scope.addPrivate(RHSVar, RHSElementCurrent);
4964 Scope.Privatize();
4965 RedOpGen(CGF, XExpr, EExpr, UpExpr);
4966 Scope.ForceCleanup();
4967
4968 // Shift the address forward by one element.
4969 llvm::Value *LHSElementNext = CGF.Builder.CreateConstGEP1_32(
4970 LHSAddr.getElementType(), LHSElementPHI, /*Idx0=*/1,
4971 "omp.arraycpy.dest.element");
4972 llvm::Value *RHSElementNext = CGF.Builder.CreateConstGEP1_32(
4973 RHSAddr.getElementType(), RHSElementPHI, /*Idx0=*/1,
4974 "omp.arraycpy.src.element");
4975 // Check whether we've reached the end.
4976 llvm::Value *Done =
4977 CGF.Builder.CreateICmpEQ(LHSElementNext, LHSEnd, "omp.arraycpy.done");
4978 CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB);
4979 LHSElementPHI->addIncoming(LHSElementNext, CGF.Builder.GetInsertBlock());
4980 RHSElementPHI->addIncoming(RHSElementNext, CGF.Builder.GetInsertBlock());
4981
4982 // Done.
4983 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
4984 }
4985
4986 /// Emit reduction combiner. If the combiner is a simple expression emit it as
4987 /// is, otherwise consider it as combiner of UDR decl and emit it as a call of
4988 /// UDR combiner function.
emitReductionCombiner(CodeGenFunction & CGF,const Expr * ReductionOp)4989 static void emitReductionCombiner(CodeGenFunction &CGF,
4990 const Expr *ReductionOp) {
4991 if (const auto *CE = dyn_cast<CallExpr>(ReductionOp))
4992 if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee()))
4993 if (const auto *DRE =
4994 dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts()))
4995 if (const auto *DRD =
4996 dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) {
4997 std::pair<llvm::Function *, llvm::Function *> Reduction =
4998 CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD);
4999 RValue Func = RValue::get(Reduction.first);
5000 CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func);
5001 CGF.EmitIgnoredExpr(ReductionOp);
5002 return;
5003 }
5004 CGF.EmitIgnoredExpr(ReductionOp);
5005 }
5006
emitReductionFunction(SourceLocation Loc,llvm::Type * ArgsElemType,ArrayRef<const Expr * > Privates,ArrayRef<const Expr * > LHSExprs,ArrayRef<const Expr * > RHSExprs,ArrayRef<const Expr * > ReductionOps)5007 llvm::Function *CGOpenMPRuntime::emitReductionFunction(
5008 SourceLocation Loc, llvm::Type *ArgsElemType,
5009 ArrayRef<const Expr *> Privates, ArrayRef<const Expr *> LHSExprs,
5010 ArrayRef<const Expr *> RHSExprs, ArrayRef<const Expr *> ReductionOps) {
5011 ASTContext &C = CGM.getContext();
5012
5013 // void reduction_func(void *LHSArg, void *RHSArg);
5014 FunctionArgList Args;
5015 ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
5016 ImplicitParamDecl::Other);
5017 ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
5018 ImplicitParamDecl::Other);
5019 Args.push_back(&LHSArg);
5020 Args.push_back(&RHSArg);
5021 const auto &CGFI =
5022 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
5023 std::string Name = getName({"omp", "reduction", "reduction_func"});
5024 auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
5025 llvm::GlobalValue::InternalLinkage, Name,
5026 &CGM.getModule());
5027 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
5028 Fn->setDoesNotRecurse();
5029 CodeGenFunction CGF(CGM);
5030 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
5031
5032 // Dst = (void*[n])(LHSArg);
5033 // Src = (void*[n])(RHSArg);
5034 Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
5035 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)),
5036 ArgsElemType->getPointerTo()),
5037 ArgsElemType, CGF.getPointerAlign());
5038 Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
5039 CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)),
5040 ArgsElemType->getPointerTo()),
5041 ArgsElemType, CGF.getPointerAlign());
5042
5043 // ...
5044 // *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]);
5045 // ...
5046 CodeGenFunction::OMPPrivateScope Scope(CGF);
5047 const auto *IPriv = Privates.begin();
5048 unsigned Idx = 0;
5049 for (unsigned I = 0, E = ReductionOps.size(); I < E; ++I, ++IPriv, ++Idx) {
5050 const auto *RHSVar =
5051 cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl());
5052 Scope.addPrivate(RHSVar, emitAddrOfVarFromArray(CGF, RHS, Idx, RHSVar));
5053 const auto *LHSVar =
5054 cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl());
5055 Scope.addPrivate(LHSVar, emitAddrOfVarFromArray(CGF, LHS, Idx, LHSVar));
5056 QualType PrivTy = (*IPriv)->getType();
5057 if (PrivTy->isVariablyModifiedType()) {
5058 // Get array size and emit VLA type.
5059 ++Idx;
5060 Address Elem = CGF.Builder.CreateConstArrayGEP(LHS, Idx);
5061 llvm::Value *Ptr = CGF.Builder.CreateLoad(Elem);
5062 const VariableArrayType *VLA =
5063 CGF.getContext().getAsVariableArrayType(PrivTy);
5064 const auto *OVE = cast<OpaqueValueExpr>(VLA->getSizeExpr());
5065 CodeGenFunction::OpaqueValueMapping OpaqueMap(
5066 CGF, OVE, RValue::get(CGF.Builder.CreatePtrToInt(Ptr, CGF.SizeTy)));
5067 CGF.EmitVariablyModifiedType(PrivTy);
5068 }
5069 }
5070 Scope.Privatize();
5071 IPriv = Privates.begin();
5072 const auto *ILHS = LHSExprs.begin();
5073 const auto *IRHS = RHSExprs.begin();
5074 for (const Expr *E : ReductionOps) {
5075 if ((*IPriv)->getType()->isArrayType()) {
5076 // Emit reduction for array section.
5077 const auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
5078 const auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
5079 EmitOMPAggregateReduction(
5080 CGF, (*IPriv)->getType(), LHSVar, RHSVar,
5081 [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) {
5082 emitReductionCombiner(CGF, E);
5083 });
5084 } else {
5085 // Emit reduction for array subscript or single variable.
5086 emitReductionCombiner(CGF, E);
5087 }
5088 ++IPriv;
5089 ++ILHS;
5090 ++IRHS;
5091 }
5092 Scope.ForceCleanup();
5093 CGF.FinishFunction();
5094 return Fn;
5095 }
5096
emitSingleReductionCombiner(CodeGenFunction & CGF,const Expr * ReductionOp,const Expr * PrivateRef,const DeclRefExpr * LHS,const DeclRefExpr * RHS)5097 void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF,
5098 const Expr *ReductionOp,
5099 const Expr *PrivateRef,
5100 const DeclRefExpr *LHS,
5101 const DeclRefExpr *RHS) {
5102 if (PrivateRef->getType()->isArrayType()) {
5103 // Emit reduction for array section.
5104 const auto *LHSVar = cast<VarDecl>(LHS->getDecl());
5105 const auto *RHSVar = cast<VarDecl>(RHS->getDecl());
5106 EmitOMPAggregateReduction(
5107 CGF, PrivateRef->getType(), LHSVar, RHSVar,
5108 [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) {
5109 emitReductionCombiner(CGF, ReductionOp);
5110 });
5111 } else {
5112 // Emit reduction for array subscript or single variable.
5113 emitReductionCombiner(CGF, ReductionOp);
5114 }
5115 }
5116
emitReduction(CodeGenFunction & CGF,SourceLocation Loc,ArrayRef<const Expr * > Privates,ArrayRef<const Expr * > LHSExprs,ArrayRef<const Expr * > RHSExprs,ArrayRef<const Expr * > ReductionOps,ReductionOptionsTy Options)5117 void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
5118 ArrayRef<const Expr *> Privates,
5119 ArrayRef<const Expr *> LHSExprs,
5120 ArrayRef<const Expr *> RHSExprs,
5121 ArrayRef<const Expr *> ReductionOps,
5122 ReductionOptionsTy Options) {
5123 if (!CGF.HaveInsertPoint())
5124 return;
5125
5126 bool WithNowait = Options.WithNowait;
5127 bool SimpleReduction = Options.SimpleReduction;
5128
5129 // Next code should be emitted for reduction:
5130 //
5131 // static kmp_critical_name lock = { 0 };
5132 //
5133 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
5134 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
5135 // ...
5136 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
5137 // *(Type<n>-1*)rhs[<n>-1]);
5138 // }
5139 //
5140 // ...
5141 // void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]};
5142 // switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
5143 // RedList, reduce_func, &<lock>)) {
5144 // case 1:
5145 // ...
5146 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
5147 // ...
5148 // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
5149 // break;
5150 // case 2:
5151 // ...
5152 // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
5153 // ...
5154 // [__kmpc_end_reduce(<loc>, <gtid>, &<lock>);]
5155 // break;
5156 // default:;
5157 // }
5158 //
5159 // if SimpleReduction is true, only the next code is generated:
5160 // ...
5161 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
5162 // ...
5163
5164 ASTContext &C = CGM.getContext();
5165
5166 if (SimpleReduction) {
5167 CodeGenFunction::RunCleanupsScope Scope(CGF);
5168 const auto *IPriv = Privates.begin();
5169 const auto *ILHS = LHSExprs.begin();
5170 const auto *IRHS = RHSExprs.begin();
5171 for (const Expr *E : ReductionOps) {
5172 emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
5173 cast<DeclRefExpr>(*IRHS));
5174 ++IPriv;
5175 ++ILHS;
5176 ++IRHS;
5177 }
5178 return;
5179 }
5180
5181 // 1. Build a list of reduction variables.
5182 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
5183 auto Size = RHSExprs.size();
5184 for (const Expr *E : Privates) {
5185 if (E->getType()->isVariablyModifiedType())
5186 // Reserve place for array size.
5187 ++Size;
5188 }
5189 llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
5190 QualType ReductionArrayTy =
5191 C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
5192 /*IndexTypeQuals=*/0);
5193 Address ReductionList =
5194 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
5195 const auto *IPriv = Privates.begin();
5196 unsigned Idx = 0;
5197 for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
5198 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
5199 CGF.Builder.CreateStore(
5200 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
5201 CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy),
5202 Elem);
5203 if ((*IPriv)->getType()->isVariablyModifiedType()) {
5204 // Store array size.
5205 ++Idx;
5206 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
5207 llvm::Value *Size = CGF.Builder.CreateIntCast(
5208 CGF.getVLASize(
5209 CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
5210 .NumElts,
5211 CGF.SizeTy, /*isSigned=*/false);
5212 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
5213 Elem);
5214 }
5215 }
5216
5217 // 2. Emit reduce_func().
5218 llvm::Function *ReductionFn =
5219 emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy),
5220 Privates, LHSExprs, RHSExprs, ReductionOps);
5221
5222 // 3. Create static kmp_critical_name lock = { 0 };
5223 std::string Name = getName({"reduction"});
5224 llvm::Value *Lock = getCriticalRegionLock(Name);
5225
5226 // 4. Build res = __kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList),
5227 // RedList, reduce_func, &<lock>);
5228 llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE);
5229 llvm::Value *ThreadId = getThreadID(CGF, Loc);
5230 llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
5231 llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
5232 ReductionList.getPointer(), CGF.VoidPtrTy);
5233 llvm::Value *Args[] = {
5234 IdentTLoc, // ident_t *<loc>
5235 ThreadId, // i32 <gtid>
5236 CGF.Builder.getInt32(RHSExprs.size()), // i32 <n>
5237 ReductionArrayTySize, // size_type sizeof(RedList)
5238 RL, // void *RedList
5239 ReductionFn, // void (*) (void *, void *) <reduce_func>
5240 Lock // kmp_critical_name *&<lock>
5241 };
5242 llvm::Value *Res = CGF.EmitRuntimeCall(
5243 OMPBuilder.getOrCreateRuntimeFunction(
5244 CGM.getModule(),
5245 WithNowait ? OMPRTL___kmpc_reduce_nowait : OMPRTL___kmpc_reduce),
5246 Args);
5247
5248 // 5. Build switch(res)
5249 llvm::BasicBlock *DefaultBB = CGF.createBasicBlock(".omp.reduction.default");
5250 llvm::SwitchInst *SwInst =
5251 CGF.Builder.CreateSwitch(Res, DefaultBB, /*NumCases=*/2);
5252
5253 // 6. Build case 1:
5254 // ...
5255 // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]);
5256 // ...
5257 // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
5258 // break;
5259 llvm::BasicBlock *Case1BB = CGF.createBasicBlock(".omp.reduction.case1");
5260 SwInst->addCase(CGF.Builder.getInt32(1), Case1BB);
5261 CGF.EmitBlock(Case1BB);
5262
5263 // Add emission of __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
5264 llvm::Value *EndArgs[] = {
5265 IdentTLoc, // ident_t *<loc>
5266 ThreadId, // i32 <gtid>
5267 Lock // kmp_critical_name *&<lock>
5268 };
5269 auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps](
5270 CodeGenFunction &CGF, PrePostActionTy &Action) {
5271 CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime();
5272 const auto *IPriv = Privates.begin();
5273 const auto *ILHS = LHSExprs.begin();
5274 const auto *IRHS = RHSExprs.begin();
5275 for (const Expr *E : ReductionOps) {
5276 RT.emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
5277 cast<DeclRefExpr>(*IRHS));
5278 ++IPriv;
5279 ++ILHS;
5280 ++IRHS;
5281 }
5282 };
5283 RegionCodeGenTy RCG(CodeGen);
5284 CommonActionTy Action(
5285 nullptr, std::nullopt,
5286 OMPBuilder.getOrCreateRuntimeFunction(
5287 CGM.getModule(), WithNowait ? OMPRTL___kmpc_end_reduce_nowait
5288 : OMPRTL___kmpc_end_reduce),
5289 EndArgs);
5290 RCG.setAction(Action);
5291 RCG(CGF);
5292
5293 CGF.EmitBranch(DefaultBB);
5294
5295 // 7. Build case 2:
5296 // ...
5297 // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]));
5298 // ...
5299 // break;
5300 llvm::BasicBlock *Case2BB = CGF.createBasicBlock(".omp.reduction.case2");
5301 SwInst->addCase(CGF.Builder.getInt32(2), Case2BB);
5302 CGF.EmitBlock(Case2BB);
5303
5304 auto &&AtomicCodeGen = [Loc, Privates, LHSExprs, RHSExprs, ReductionOps](
5305 CodeGenFunction &CGF, PrePostActionTy &Action) {
5306 const auto *ILHS = LHSExprs.begin();
5307 const auto *IRHS = RHSExprs.begin();
5308 const auto *IPriv = Privates.begin();
5309 for (const Expr *E : ReductionOps) {
5310 const Expr *XExpr = nullptr;
5311 const Expr *EExpr = nullptr;
5312 const Expr *UpExpr = nullptr;
5313 BinaryOperatorKind BO = BO_Comma;
5314 if (const auto *BO = dyn_cast<BinaryOperator>(E)) {
5315 if (BO->getOpcode() == BO_Assign) {
5316 XExpr = BO->getLHS();
5317 UpExpr = BO->getRHS();
5318 }
5319 }
5320 // Try to emit update expression as a simple atomic.
5321 const Expr *RHSExpr = UpExpr;
5322 if (RHSExpr) {
5323 // Analyze RHS part of the whole expression.
5324 if (const auto *ACO = dyn_cast<AbstractConditionalOperator>(
5325 RHSExpr->IgnoreParenImpCasts())) {
5326 // If this is a conditional operator, analyze its condition for
5327 // min/max reduction operator.
5328 RHSExpr = ACO->getCond();
5329 }
5330 if (const auto *BORHS =
5331 dyn_cast<BinaryOperator>(RHSExpr->IgnoreParenImpCasts())) {
5332 EExpr = BORHS->getRHS();
5333 BO = BORHS->getOpcode();
5334 }
5335 }
5336 if (XExpr) {
5337 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
5338 auto &&AtomicRedGen = [BO, VD,
5339 Loc](CodeGenFunction &CGF, const Expr *XExpr,
5340 const Expr *EExpr, const Expr *UpExpr) {
5341 LValue X = CGF.EmitLValue(XExpr);
5342 RValue E;
5343 if (EExpr)
5344 E = CGF.EmitAnyExpr(EExpr);
5345 CGF.EmitOMPAtomicSimpleUpdateExpr(
5346 X, E, BO, /*IsXLHSInRHSPart=*/true,
5347 llvm::AtomicOrdering::Monotonic, Loc,
5348 [&CGF, UpExpr, VD, Loc](RValue XRValue) {
5349 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
5350 Address LHSTemp = CGF.CreateMemTemp(VD->getType());
5351 CGF.emitOMPSimpleStore(
5352 CGF.MakeAddrLValue(LHSTemp, VD->getType()), XRValue,
5353 VD->getType().getNonReferenceType(), Loc);
5354 PrivateScope.addPrivate(VD, LHSTemp);
5355 (void)PrivateScope.Privatize();
5356 return CGF.EmitAnyExpr(UpExpr);
5357 });
5358 };
5359 if ((*IPriv)->getType()->isArrayType()) {
5360 // Emit atomic reduction for array section.
5361 const auto *RHSVar =
5362 cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
5363 EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), VD, RHSVar,
5364 AtomicRedGen, XExpr, EExpr, UpExpr);
5365 } else {
5366 // Emit atomic reduction for array subscript or single variable.
5367 AtomicRedGen(CGF, XExpr, EExpr, UpExpr);
5368 }
5369 } else {
5370 // Emit as a critical region.
5371 auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *,
5372 const Expr *, const Expr *) {
5373 CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime();
5374 std::string Name = RT.getName({"atomic_reduction"});
5375 RT.emitCriticalRegion(
5376 CGF, Name,
5377 [=](CodeGenFunction &CGF, PrePostActionTy &Action) {
5378 Action.Enter(CGF);
5379 emitReductionCombiner(CGF, E);
5380 },
5381 Loc);
5382 };
5383 if ((*IPriv)->getType()->isArrayType()) {
5384 const auto *LHSVar =
5385 cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
5386 const auto *RHSVar =
5387 cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
5388 EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), LHSVar, RHSVar,
5389 CritRedGen);
5390 } else {
5391 CritRedGen(CGF, nullptr, nullptr, nullptr);
5392 }
5393 }
5394 ++ILHS;
5395 ++IRHS;
5396 ++IPriv;
5397 }
5398 };
5399 RegionCodeGenTy AtomicRCG(AtomicCodeGen);
5400 if (!WithNowait) {
5401 // Add emission of __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
5402 llvm::Value *EndArgs[] = {
5403 IdentTLoc, // ident_t *<loc>
5404 ThreadId, // i32 <gtid>
5405 Lock // kmp_critical_name *&<lock>
5406 };
5407 CommonActionTy Action(nullptr, std::nullopt,
5408 OMPBuilder.getOrCreateRuntimeFunction(
5409 CGM.getModule(), OMPRTL___kmpc_end_reduce),
5410 EndArgs);
5411 AtomicRCG.setAction(Action);
5412 AtomicRCG(CGF);
5413 } else {
5414 AtomicRCG(CGF);
5415 }
5416
5417 CGF.EmitBranch(DefaultBB);
5418 CGF.EmitBlock(DefaultBB, /*IsFinished=*/true);
5419 }
5420
5421 /// Generates unique name for artificial threadprivate variables.
5422 /// Format is: <Prefix> "." <Decl_mangled_name> "_" "<Decl_start_loc_raw_enc>"
generateUniqueName(CodeGenModule & CGM,StringRef Prefix,const Expr * Ref)5423 static std::string generateUniqueName(CodeGenModule &CGM, StringRef Prefix,
5424 const Expr *Ref) {
5425 SmallString<256> Buffer;
5426 llvm::raw_svector_ostream Out(Buffer);
5427 const clang::DeclRefExpr *DE;
5428 const VarDecl *D = ::getBaseDecl(Ref, DE);
5429 if (!D)
5430 D = cast<VarDecl>(cast<DeclRefExpr>(Ref)->getDecl());
5431 D = D->getCanonicalDecl();
5432 std::string Name = CGM.getOpenMPRuntime().getName(
5433 {D->isLocalVarDeclOrParm() ? D->getName() : CGM.getMangledName(D)});
5434 Out << Prefix << Name << "_"
5435 << D->getCanonicalDecl()->getBeginLoc().getRawEncoding();
5436 return std::string(Out.str());
5437 }
5438
5439 /// Emits reduction initializer function:
5440 /// \code
5441 /// void @.red_init(void* %arg, void* %orig) {
5442 /// %0 = bitcast void* %arg to <type>*
5443 /// store <type> <init>, <type>* %0
5444 /// ret void
5445 /// }
5446 /// \endcode
emitReduceInitFunction(CodeGenModule & CGM,SourceLocation Loc,ReductionCodeGen & RCG,unsigned N)5447 static llvm::Value *emitReduceInitFunction(CodeGenModule &CGM,
5448 SourceLocation Loc,
5449 ReductionCodeGen &RCG, unsigned N) {
5450 ASTContext &C = CGM.getContext();
5451 QualType VoidPtrTy = C.VoidPtrTy;
5452 VoidPtrTy.addRestrict();
5453 FunctionArgList Args;
5454 ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, VoidPtrTy,
5455 ImplicitParamDecl::Other);
5456 ImplicitParamDecl ParamOrig(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, VoidPtrTy,
5457 ImplicitParamDecl::Other);
5458 Args.emplace_back(&Param);
5459 Args.emplace_back(&ParamOrig);
5460 const auto &FnInfo =
5461 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
5462 llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo);
5463 std::string Name = CGM.getOpenMPRuntime().getName({"red_init", ""});
5464 auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage,
5465 Name, &CGM.getModule());
5466 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo);
5467 Fn->setDoesNotRecurse();
5468 CodeGenFunction CGF(CGM);
5469 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc);
5470 QualType PrivateType = RCG.getPrivateType(N);
5471 Address PrivateAddr = CGF.EmitLoadOfPointer(
5472 CGF.Builder.CreateElementBitCast(
5473 CGF.GetAddrOfLocalVar(&Param),
5474 CGF.ConvertTypeForMem(PrivateType)->getPointerTo()),
5475 C.getPointerType(PrivateType)->castAs<PointerType>());
5476 llvm::Value *Size = nullptr;
5477 // If the size of the reduction item is non-constant, load it from global
5478 // threadprivate variable.
5479 if (RCG.getSizes(N).second) {
5480 Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate(
5481 CGF, CGM.getContext().getSizeType(),
5482 generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N)));
5483 Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false,
5484 CGM.getContext().getSizeType(), Loc);
5485 }
5486 RCG.emitAggregateType(CGF, N, Size);
5487 Address OrigAddr = Address::invalid();
5488 // If initializer uses initializer from declare reduction construct, emit a
5489 // pointer to the address of the original reduction item (reuired by reduction
5490 // initializer)
5491 if (RCG.usesReductionInitializer(N)) {
5492 Address SharedAddr = CGF.GetAddrOfLocalVar(&ParamOrig);
5493 OrigAddr = CGF.EmitLoadOfPointer(
5494 SharedAddr,
5495 CGM.getContext().VoidPtrTy.castAs<PointerType>()->getTypePtr());
5496 }
5497 // Emit the initializer:
5498 // %0 = bitcast void* %arg to <type>*
5499 // store <type> <init>, <type>* %0
5500 RCG.emitInitialization(CGF, N, PrivateAddr, OrigAddr,
5501 [](CodeGenFunction &) { return false; });
5502 CGF.FinishFunction();
5503 return Fn;
5504 }
5505
5506 /// Emits reduction combiner function:
5507 /// \code
5508 /// void @.red_comb(void* %arg0, void* %arg1) {
5509 /// %lhs = bitcast void* %arg0 to <type>*
5510 /// %rhs = bitcast void* %arg1 to <type>*
5511 /// %2 = <ReductionOp>(<type>* %lhs, <type>* %rhs)
5512 /// store <type> %2, <type>* %lhs
5513 /// ret void
5514 /// }
5515 /// \endcode
emitReduceCombFunction(CodeGenModule & CGM,SourceLocation Loc,ReductionCodeGen & RCG,unsigned N,const Expr * ReductionOp,const Expr * LHS,const Expr * RHS,const Expr * PrivateRef)5516 static llvm::Value *emitReduceCombFunction(CodeGenModule &CGM,
5517 SourceLocation Loc,
5518 ReductionCodeGen &RCG, unsigned N,
5519 const Expr *ReductionOp,
5520 const Expr *LHS, const Expr *RHS,
5521 const Expr *PrivateRef) {
5522 ASTContext &C = CGM.getContext();
5523 const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(LHS)->getDecl());
5524 const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(RHS)->getDecl());
5525 FunctionArgList Args;
5526 ImplicitParamDecl ParamInOut(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
5527 C.VoidPtrTy, ImplicitParamDecl::Other);
5528 ImplicitParamDecl ParamIn(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
5529 ImplicitParamDecl::Other);
5530 Args.emplace_back(&ParamInOut);
5531 Args.emplace_back(&ParamIn);
5532 const auto &FnInfo =
5533 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
5534 llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo);
5535 std::string Name = CGM.getOpenMPRuntime().getName({"red_comb", ""});
5536 auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage,
5537 Name, &CGM.getModule());
5538 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo);
5539 Fn->setDoesNotRecurse();
5540 CodeGenFunction CGF(CGM);
5541 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc);
5542 llvm::Value *Size = nullptr;
5543 // If the size of the reduction item is non-constant, load it from global
5544 // threadprivate variable.
5545 if (RCG.getSizes(N).second) {
5546 Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate(
5547 CGF, CGM.getContext().getSizeType(),
5548 generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N)));
5549 Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false,
5550 CGM.getContext().getSizeType(), Loc);
5551 }
5552 RCG.emitAggregateType(CGF, N, Size);
5553 // Remap lhs and rhs variables to the addresses of the function arguments.
5554 // %lhs = bitcast void* %arg0 to <type>*
5555 // %rhs = bitcast void* %arg1 to <type>*
5556 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
5557 PrivateScope.addPrivate(
5558 LHSVD,
5559 // Pull out the pointer to the variable.
5560 CGF.EmitLoadOfPointer(
5561 CGF.Builder.CreateElementBitCast(
5562 CGF.GetAddrOfLocalVar(&ParamInOut),
5563 CGF.ConvertTypeForMem(LHSVD->getType())->getPointerTo()),
5564 C.getPointerType(LHSVD->getType())->castAs<PointerType>()));
5565 PrivateScope.addPrivate(
5566 RHSVD,
5567 // Pull out the pointer to the variable.
5568 CGF.EmitLoadOfPointer(
5569 CGF.Builder.CreateElementBitCast(
5570 CGF.GetAddrOfLocalVar(&ParamIn),
5571 CGF.ConvertTypeForMem(RHSVD->getType())->getPointerTo()),
5572 C.getPointerType(RHSVD->getType())->castAs<PointerType>()));
5573 PrivateScope.Privatize();
5574 // Emit the combiner body:
5575 // %2 = <ReductionOp>(<type> *%lhs, <type> *%rhs)
5576 // store <type> %2, <type>* %lhs
5577 CGM.getOpenMPRuntime().emitSingleReductionCombiner(
5578 CGF, ReductionOp, PrivateRef, cast<DeclRefExpr>(LHS),
5579 cast<DeclRefExpr>(RHS));
5580 CGF.FinishFunction();
5581 return Fn;
5582 }
5583
5584 /// Emits reduction finalizer function:
5585 /// \code
5586 /// void @.red_fini(void* %arg) {
5587 /// %0 = bitcast void* %arg to <type>*
5588 /// <destroy>(<type>* %0)
5589 /// ret void
5590 /// }
5591 /// \endcode
emitReduceFiniFunction(CodeGenModule & CGM,SourceLocation Loc,ReductionCodeGen & RCG,unsigned N)5592 static llvm::Value *emitReduceFiniFunction(CodeGenModule &CGM,
5593 SourceLocation Loc,
5594 ReductionCodeGen &RCG, unsigned N) {
5595 if (!RCG.needCleanups(N))
5596 return nullptr;
5597 ASTContext &C = CGM.getContext();
5598 FunctionArgList Args;
5599 ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
5600 ImplicitParamDecl::Other);
5601 Args.emplace_back(&Param);
5602 const auto &FnInfo =
5603 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
5604 llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo);
5605 std::string Name = CGM.getOpenMPRuntime().getName({"red_fini", ""});
5606 auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage,
5607 Name, &CGM.getModule());
5608 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo);
5609 Fn->setDoesNotRecurse();
5610 CodeGenFunction CGF(CGM);
5611 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc);
5612 Address PrivateAddr = CGF.EmitLoadOfPointer(
5613 CGF.GetAddrOfLocalVar(&Param), C.VoidPtrTy.castAs<PointerType>());
5614 llvm::Value *Size = nullptr;
5615 // If the size of the reduction item is non-constant, load it from global
5616 // threadprivate variable.
5617 if (RCG.getSizes(N).second) {
5618 Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate(
5619 CGF, CGM.getContext().getSizeType(),
5620 generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N)));
5621 Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false,
5622 CGM.getContext().getSizeType(), Loc);
5623 }
5624 RCG.emitAggregateType(CGF, N, Size);
5625 // Emit the finalizer body:
5626 // <destroy>(<type>* %0)
5627 RCG.emitCleanups(CGF, N, PrivateAddr);
5628 CGF.FinishFunction(Loc);
5629 return Fn;
5630 }
5631
emitTaskReductionInit(CodeGenFunction & CGF,SourceLocation Loc,ArrayRef<const Expr * > LHSExprs,ArrayRef<const Expr * > RHSExprs,const OMPTaskDataTy & Data)5632 llvm::Value *CGOpenMPRuntime::emitTaskReductionInit(
5633 CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs,
5634 ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) {
5635 if (!CGF.HaveInsertPoint() || Data.ReductionVars.empty())
5636 return nullptr;
5637
5638 // Build typedef struct:
5639 // kmp_taskred_input {
5640 // void *reduce_shar; // shared reduction item
5641 // void *reduce_orig; // original reduction item used for initialization
5642 // size_t reduce_size; // size of data item
5643 // void *reduce_init; // data initialization routine
5644 // void *reduce_fini; // data finalization routine
5645 // void *reduce_comb; // data combiner routine
5646 // kmp_task_red_flags_t flags; // flags for additional info from compiler
5647 // } kmp_taskred_input_t;
5648 ASTContext &C = CGM.getContext();
5649 RecordDecl *RD = C.buildImplicitRecord("kmp_taskred_input_t");
5650 RD->startDefinition();
5651 const FieldDecl *SharedFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy);
5652 const FieldDecl *OrigFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy);
5653 const FieldDecl *SizeFD = addFieldToRecordDecl(C, RD, C.getSizeType());
5654 const FieldDecl *InitFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy);
5655 const FieldDecl *FiniFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy);
5656 const FieldDecl *CombFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy);
5657 const FieldDecl *FlagsFD = addFieldToRecordDecl(
5658 C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false));
5659 RD->completeDefinition();
5660 QualType RDType = C.getRecordType(RD);
5661 unsigned Size = Data.ReductionVars.size();
5662 llvm::APInt ArraySize(/*numBits=*/64, Size);
5663 QualType ArrayRDType = C.getConstantArrayType(
5664 RDType, ArraySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0);
5665 // kmp_task_red_input_t .rd_input.[Size];
5666 Address TaskRedInput = CGF.CreateMemTemp(ArrayRDType, ".rd_input.");
5667 ReductionCodeGen RCG(Data.ReductionVars, Data.ReductionOrigs,
5668 Data.ReductionCopies, Data.ReductionOps);
5669 for (unsigned Cnt = 0; Cnt < Size; ++Cnt) {
5670 // kmp_task_red_input_t &ElemLVal = .rd_input.[Cnt];
5671 llvm::Value *Idxs[] = {llvm::ConstantInt::get(CGM.SizeTy, /*V=*/0),
5672 llvm::ConstantInt::get(CGM.SizeTy, Cnt)};
5673 llvm::Value *GEP = CGF.EmitCheckedInBoundsGEP(
5674 TaskRedInput.getElementType(), TaskRedInput.getPointer(), Idxs,
5675 /*SignedIndices=*/false, /*IsSubtraction=*/false, Loc,
5676 ".rd_input.gep.");
5677 LValue ElemLVal = CGF.MakeNaturalAlignAddrLValue(GEP, RDType);
5678 // ElemLVal.reduce_shar = &Shareds[Cnt];
5679 LValue SharedLVal = CGF.EmitLValueForField(ElemLVal, SharedFD);
5680 RCG.emitSharedOrigLValue(CGF, Cnt);
5681 llvm::Value *CastedShared =
5682 CGF.EmitCastToVoidPtr(RCG.getSharedLValue(Cnt).getPointer(CGF));
5683 CGF.EmitStoreOfScalar(CastedShared, SharedLVal);
5684 // ElemLVal.reduce_orig = &Origs[Cnt];
5685 LValue OrigLVal = CGF.EmitLValueForField(ElemLVal, OrigFD);
5686 llvm::Value *CastedOrig =
5687 CGF.EmitCastToVoidPtr(RCG.getOrigLValue(Cnt).getPointer(CGF));
5688 CGF.EmitStoreOfScalar(CastedOrig, OrigLVal);
5689 RCG.emitAggregateType(CGF, Cnt);
5690 llvm::Value *SizeValInChars;
5691 llvm::Value *SizeVal;
5692 std::tie(SizeValInChars, SizeVal) = RCG.getSizes(Cnt);
5693 // We use delayed creation/initialization for VLAs and array sections. It is
5694 // required because runtime does not provide the way to pass the sizes of
5695 // VLAs/array sections to initializer/combiner/finalizer functions. Instead
5696 // threadprivate global variables are used to store these values and use
5697 // them in the functions.
5698 bool DelayedCreation = !!SizeVal;
5699 SizeValInChars = CGF.Builder.CreateIntCast(SizeValInChars, CGM.SizeTy,
5700 /*isSigned=*/false);
5701 LValue SizeLVal = CGF.EmitLValueForField(ElemLVal, SizeFD);
5702 CGF.EmitStoreOfScalar(SizeValInChars, SizeLVal);
5703 // ElemLVal.reduce_init = init;
5704 LValue InitLVal = CGF.EmitLValueForField(ElemLVal, InitFD);
5705 llvm::Value *InitAddr =
5706 CGF.EmitCastToVoidPtr(emitReduceInitFunction(CGM, Loc, RCG, Cnt));
5707 CGF.EmitStoreOfScalar(InitAddr, InitLVal);
5708 // ElemLVal.reduce_fini = fini;
5709 LValue FiniLVal = CGF.EmitLValueForField(ElemLVal, FiniFD);
5710 llvm::Value *Fini = emitReduceFiniFunction(CGM, Loc, RCG, Cnt);
5711 llvm::Value *FiniAddr = Fini
5712 ? CGF.EmitCastToVoidPtr(Fini)
5713 : llvm::ConstantPointerNull::get(CGM.VoidPtrTy);
5714 CGF.EmitStoreOfScalar(FiniAddr, FiniLVal);
5715 // ElemLVal.reduce_comb = comb;
5716 LValue CombLVal = CGF.EmitLValueForField(ElemLVal, CombFD);
5717 llvm::Value *CombAddr = CGF.EmitCastToVoidPtr(emitReduceCombFunction(
5718 CGM, Loc, RCG, Cnt, Data.ReductionOps[Cnt], LHSExprs[Cnt],
5719 RHSExprs[Cnt], Data.ReductionCopies[Cnt]));
5720 CGF.EmitStoreOfScalar(CombAddr, CombLVal);
5721 // ElemLVal.flags = 0;
5722 LValue FlagsLVal = CGF.EmitLValueForField(ElemLVal, FlagsFD);
5723 if (DelayedCreation) {
5724 CGF.EmitStoreOfScalar(
5725 llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1, /*isSigned=*/true),
5726 FlagsLVal);
5727 } else
5728 CGF.EmitNullInitialization(FlagsLVal.getAddress(CGF),
5729 FlagsLVal.getType());
5730 }
5731 if (Data.IsReductionWithTaskMod) {
5732 // Build call void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int
5733 // is_ws, int num, void *data);
5734 llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc);
5735 llvm::Value *GTid = CGF.Builder.CreateIntCast(getThreadID(CGF, Loc),
5736 CGM.IntTy, /*isSigned=*/true);
5737 llvm::Value *Args[] = {
5738 IdentTLoc, GTid,
5739 llvm::ConstantInt::get(CGM.IntTy, Data.IsWorksharingReduction ? 1 : 0,
5740 /*isSigned=*/true),
5741 llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true),
5742 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
5743 TaskRedInput.getPointer(), CGM.VoidPtrTy)};
5744 return CGF.EmitRuntimeCall(
5745 OMPBuilder.getOrCreateRuntimeFunction(
5746 CGM.getModule(), OMPRTL___kmpc_taskred_modifier_init),
5747 Args);
5748 }
5749 // Build call void *__kmpc_taskred_init(int gtid, int num_data, void *data);
5750 llvm::Value *Args[] = {
5751 CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), CGM.IntTy,
5752 /*isSigned=*/true),
5753 llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true),
5754 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TaskRedInput.getPointer(),
5755 CGM.VoidPtrTy)};
5756 return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
5757 CGM.getModule(), OMPRTL___kmpc_taskred_init),
5758 Args);
5759 }
5760
emitTaskReductionFini(CodeGenFunction & CGF,SourceLocation Loc,bool IsWorksharingReduction)5761 void CGOpenMPRuntime::emitTaskReductionFini(CodeGenFunction &CGF,
5762 SourceLocation Loc,
5763 bool IsWorksharingReduction) {
5764 // Build call void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int
5765 // is_ws, int num, void *data);
5766 llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc);
5767 llvm::Value *GTid = CGF.Builder.CreateIntCast(getThreadID(CGF, Loc),
5768 CGM.IntTy, /*isSigned=*/true);
5769 llvm::Value *Args[] = {IdentTLoc, GTid,
5770 llvm::ConstantInt::get(CGM.IntTy,
5771 IsWorksharingReduction ? 1 : 0,
5772 /*isSigned=*/true)};
5773 (void)CGF.EmitRuntimeCall(
5774 OMPBuilder.getOrCreateRuntimeFunction(
5775 CGM.getModule(), OMPRTL___kmpc_task_reduction_modifier_fini),
5776 Args);
5777 }
5778
emitTaskReductionFixups(CodeGenFunction & CGF,SourceLocation Loc,ReductionCodeGen & RCG,unsigned N)5779 void CGOpenMPRuntime::emitTaskReductionFixups(CodeGenFunction &CGF,
5780 SourceLocation Loc,
5781 ReductionCodeGen &RCG,
5782 unsigned N) {
5783 auto Sizes = RCG.getSizes(N);
5784 // Emit threadprivate global variable if the type is non-constant
5785 // (Sizes.second = nullptr).
5786 if (Sizes.second) {
5787 llvm::Value *SizeVal = CGF.Builder.CreateIntCast(Sizes.second, CGM.SizeTy,
5788 /*isSigned=*/false);
5789 Address SizeAddr = getAddrOfArtificialThreadPrivate(
5790 CGF, CGM.getContext().getSizeType(),
5791 generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N)));
5792 CGF.Builder.CreateStore(SizeVal, SizeAddr, /*IsVolatile=*/false);
5793 }
5794 }
5795
getTaskReductionItem(CodeGenFunction & CGF,SourceLocation Loc,llvm::Value * ReductionsPtr,LValue SharedLVal)5796 Address CGOpenMPRuntime::getTaskReductionItem(CodeGenFunction &CGF,
5797 SourceLocation Loc,
5798 llvm::Value *ReductionsPtr,
5799 LValue SharedLVal) {
5800 // Build call void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void
5801 // *d);
5802 llvm::Value *Args[] = {CGF.Builder.CreateIntCast(getThreadID(CGF, Loc),
5803 CGM.IntTy,
5804 /*isSigned=*/true),
5805 ReductionsPtr,
5806 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
5807 SharedLVal.getPointer(CGF), CGM.VoidPtrTy)};
5808 return Address(
5809 CGF.EmitRuntimeCall(
5810 OMPBuilder.getOrCreateRuntimeFunction(
5811 CGM.getModule(), OMPRTL___kmpc_task_reduction_get_th_data),
5812 Args),
5813 CGF.Int8Ty, SharedLVal.getAlignment());
5814 }
5815
emitTaskwaitCall(CodeGenFunction & CGF,SourceLocation Loc,const OMPTaskDataTy & Data)5816 void CGOpenMPRuntime::emitTaskwaitCall(CodeGenFunction &CGF, SourceLocation Loc,
5817 const OMPTaskDataTy &Data) {
5818 if (!CGF.HaveInsertPoint())
5819 return;
5820
5821 if (CGF.CGM.getLangOpts().OpenMPIRBuilder && Data.Dependences.empty()) {
5822 // TODO: Need to support taskwait with dependences in the OpenMPIRBuilder.
5823 OMPBuilder.createTaskwait(CGF.Builder);
5824 } else {
5825 llvm::Value *ThreadID = getThreadID(CGF, Loc);
5826 llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc);
5827 auto &M = CGM.getModule();
5828 Address DependenciesArray = Address::invalid();
5829 llvm::Value *NumOfElements;
5830 std::tie(NumOfElements, DependenciesArray) =
5831 emitDependClause(CGF, Data.Dependences, Loc);
5832 if (!Data.Dependences.empty()) {
5833 llvm::Value *DepWaitTaskArgs[7];
5834 DepWaitTaskArgs[0] = UpLoc;
5835 DepWaitTaskArgs[1] = ThreadID;
5836 DepWaitTaskArgs[2] = NumOfElements;
5837 DepWaitTaskArgs[3] = DependenciesArray.getPointer();
5838 DepWaitTaskArgs[4] = CGF.Builder.getInt32(0);
5839 DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
5840 DepWaitTaskArgs[6] =
5841 llvm::ConstantInt::get(CGF.Int32Ty, Data.HasNowaitClause);
5842
5843 CodeGenFunction::RunCleanupsScope LocalScope(CGF);
5844
5845 // Build void __kmpc_omp_taskwait_deps_51(ident_t *, kmp_int32 gtid,
5846 // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32
5847 // ndeps_noalias, kmp_depend_info_t *noalias_dep_list,
5848 // kmp_int32 has_no_wait); if dependence info is specified.
5849 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
5850 M, OMPRTL___kmpc_omp_taskwait_deps_51),
5851 DepWaitTaskArgs);
5852
5853 } else {
5854
5855 // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
5856 // global_tid);
5857 llvm::Value *Args[] = {UpLoc, ThreadID};
5858 // Ignore return result until untied tasks are supported.
5859 CGF.EmitRuntimeCall(
5860 OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_omp_taskwait),
5861 Args);
5862 }
5863 }
5864
5865 if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
5866 Region->emitUntiedSwitch(CGF);
5867 }
5868
emitInlinedDirective(CodeGenFunction & CGF,OpenMPDirectiveKind InnerKind,const RegionCodeGenTy & CodeGen,bool HasCancel)5869 void CGOpenMPRuntime::emitInlinedDirective(CodeGenFunction &CGF,
5870 OpenMPDirectiveKind InnerKind,
5871 const RegionCodeGenTy &CodeGen,
5872 bool HasCancel) {
5873 if (!CGF.HaveInsertPoint())
5874 return;
5875 InlinedOpenMPRegionRAII Region(CGF, CodeGen, InnerKind, HasCancel,
5876 InnerKind != OMPD_critical &&
5877 InnerKind != OMPD_master &&
5878 InnerKind != OMPD_masked);
5879 CGF.CapturedStmtInfo->EmitBody(CGF, /*S=*/nullptr);
5880 }
5881
5882 namespace {
5883 enum RTCancelKind {
5884 CancelNoreq = 0,
5885 CancelParallel = 1,
5886 CancelLoop = 2,
5887 CancelSections = 3,
5888 CancelTaskgroup = 4
5889 };
5890 } // anonymous namespace
5891
getCancellationKind(OpenMPDirectiveKind CancelRegion)5892 static RTCancelKind getCancellationKind(OpenMPDirectiveKind CancelRegion) {
5893 RTCancelKind CancelKind = CancelNoreq;
5894 if (CancelRegion == OMPD_parallel)
5895 CancelKind = CancelParallel;
5896 else if (CancelRegion == OMPD_for)
5897 CancelKind = CancelLoop;
5898 else if (CancelRegion == OMPD_sections)
5899 CancelKind = CancelSections;
5900 else {
5901 assert(CancelRegion == OMPD_taskgroup);
5902 CancelKind = CancelTaskgroup;
5903 }
5904 return CancelKind;
5905 }
5906
emitCancellationPointCall(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind CancelRegion)5907 void CGOpenMPRuntime::emitCancellationPointCall(
5908 CodeGenFunction &CGF, SourceLocation Loc,
5909 OpenMPDirectiveKind CancelRegion) {
5910 if (!CGF.HaveInsertPoint())
5911 return;
5912 // Build call kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32
5913 // global_tid, kmp_int32 cncl_kind);
5914 if (auto *OMPRegionInfo =
5915 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
5916 // For 'cancellation point taskgroup', the task region info may not have a
5917 // cancel. This may instead happen in another adjacent task.
5918 if (CancelRegion == OMPD_taskgroup || OMPRegionInfo->hasCancel()) {
5919 llvm::Value *Args[] = {
5920 emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc),
5921 CGF.Builder.getInt32(getCancellationKind(CancelRegion))};
5922 // Ignore return result until untied tasks are supported.
5923 llvm::Value *Result = CGF.EmitRuntimeCall(
5924 OMPBuilder.getOrCreateRuntimeFunction(
5925 CGM.getModule(), OMPRTL___kmpc_cancellationpoint),
5926 Args);
5927 // if (__kmpc_cancellationpoint()) {
5928 // call i32 @__kmpc_cancel_barrier( // for parallel cancellation only
5929 // exit from construct;
5930 // }
5931 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit");
5932 llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue");
5933 llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result);
5934 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
5935 CGF.EmitBlock(ExitBB);
5936 if (CancelRegion == OMPD_parallel)
5937 emitBarrierCall(CGF, Loc, OMPD_unknown, /*EmitChecks=*/false);
5938 // exit from construct;
5939 CodeGenFunction::JumpDest CancelDest =
5940 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
5941 CGF.EmitBranchThroughCleanup(CancelDest);
5942 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
5943 }
5944 }
5945 }
5946
emitCancelCall(CodeGenFunction & CGF,SourceLocation Loc,const Expr * IfCond,OpenMPDirectiveKind CancelRegion)5947 void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc,
5948 const Expr *IfCond,
5949 OpenMPDirectiveKind CancelRegion) {
5950 if (!CGF.HaveInsertPoint())
5951 return;
5952 // Build call kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid,
5953 // kmp_int32 cncl_kind);
5954 auto &M = CGM.getModule();
5955 if (auto *OMPRegionInfo =
5956 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) {
5957 auto &&ThenGen = [this, &M, Loc, CancelRegion,
5958 OMPRegionInfo](CodeGenFunction &CGF, PrePostActionTy &) {
5959 CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime();
5960 llvm::Value *Args[] = {
5961 RT.emitUpdateLocation(CGF, Loc), RT.getThreadID(CGF, Loc),
5962 CGF.Builder.getInt32(getCancellationKind(CancelRegion))};
5963 // Ignore return result until untied tasks are supported.
5964 llvm::Value *Result = CGF.EmitRuntimeCall(
5965 OMPBuilder.getOrCreateRuntimeFunction(M, OMPRTL___kmpc_cancel), Args);
5966 // if (__kmpc_cancel()) {
5967 // call i32 @__kmpc_cancel_barrier( // for parallel cancellation only
5968 // exit from construct;
5969 // }
5970 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit");
5971 llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue");
5972 llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result);
5973 CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB);
5974 CGF.EmitBlock(ExitBB);
5975 if (CancelRegion == OMPD_parallel)
5976 RT.emitBarrierCall(CGF, Loc, OMPD_unknown, /*EmitChecks=*/false);
5977 // exit from construct;
5978 CodeGenFunction::JumpDest CancelDest =
5979 CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind());
5980 CGF.EmitBranchThroughCleanup(CancelDest);
5981 CGF.EmitBlock(ContBB, /*IsFinished=*/true);
5982 };
5983 if (IfCond) {
5984 emitIfClause(CGF, IfCond, ThenGen,
5985 [](CodeGenFunction &, PrePostActionTy &) {});
5986 } else {
5987 RegionCodeGenTy ThenRCG(ThenGen);
5988 ThenRCG(CGF);
5989 }
5990 }
5991 }
5992
5993 namespace {
5994 /// Cleanup action for uses_allocators support.
5995 class OMPUsesAllocatorsActionTy final : public PrePostActionTy {
5996 ArrayRef<std::pair<const Expr *, const Expr *>> Allocators;
5997
5998 public:
OMPUsesAllocatorsActionTy(ArrayRef<std::pair<const Expr *,const Expr * >> Allocators)5999 OMPUsesAllocatorsActionTy(
6000 ArrayRef<std::pair<const Expr *, const Expr *>> Allocators)
6001 : Allocators(Allocators) {}
Enter(CodeGenFunction & CGF)6002 void Enter(CodeGenFunction &CGF) override {
6003 if (!CGF.HaveInsertPoint())
6004 return;
6005 for (const auto &AllocatorData : Allocators) {
6006 CGF.CGM.getOpenMPRuntime().emitUsesAllocatorsInit(
6007 CGF, AllocatorData.first, AllocatorData.second);
6008 }
6009 }
Exit(CodeGenFunction & CGF)6010 void Exit(CodeGenFunction &CGF) override {
6011 if (!CGF.HaveInsertPoint())
6012 return;
6013 for (const auto &AllocatorData : Allocators) {
6014 CGF.CGM.getOpenMPRuntime().emitUsesAllocatorsFini(CGF,
6015 AllocatorData.first);
6016 }
6017 }
6018 };
6019 } // namespace
6020
emitTargetOutlinedFunction(const OMPExecutableDirective & D,StringRef ParentName,llvm::Function * & OutlinedFn,llvm::Constant * & OutlinedFnID,bool IsOffloadEntry,const RegionCodeGenTy & CodeGen)6021 void CGOpenMPRuntime::emitTargetOutlinedFunction(
6022 const OMPExecutableDirective &D, StringRef ParentName,
6023 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
6024 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
6025 assert(!ParentName.empty() && "Invalid target entry parent name!");
6026 HasEmittedTargetRegion = true;
6027 SmallVector<std::pair<const Expr *, const Expr *>, 4> Allocators;
6028 for (const auto *C : D.getClausesOfKind<OMPUsesAllocatorsClause>()) {
6029 for (unsigned I = 0, E = C->getNumberOfAllocators(); I < E; ++I) {
6030 const OMPUsesAllocatorsClause::Data D = C->getAllocatorData(I);
6031 if (!D.AllocatorTraits)
6032 continue;
6033 Allocators.emplace_back(D.Allocator, D.AllocatorTraits);
6034 }
6035 }
6036 OMPUsesAllocatorsActionTy UsesAllocatorAction(Allocators);
6037 CodeGen.setAction(UsesAllocatorAction);
6038 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
6039 IsOffloadEntry, CodeGen);
6040 }
6041
emitUsesAllocatorsInit(CodeGenFunction & CGF,const Expr * Allocator,const Expr * AllocatorTraits)6042 void CGOpenMPRuntime::emitUsesAllocatorsInit(CodeGenFunction &CGF,
6043 const Expr *Allocator,
6044 const Expr *AllocatorTraits) {
6045 llvm::Value *ThreadId = getThreadID(CGF, Allocator->getExprLoc());
6046 ThreadId = CGF.Builder.CreateIntCast(ThreadId, CGF.IntTy, /*isSigned=*/true);
6047 // Use default memspace handle.
6048 llvm::Value *MemSpaceHandle = llvm::ConstantPointerNull::get(CGF.VoidPtrTy);
6049 llvm::Value *NumTraits = llvm::ConstantInt::get(
6050 CGF.IntTy, cast<ConstantArrayType>(
6051 AllocatorTraits->getType()->getAsArrayTypeUnsafe())
6052 ->getSize()
6053 .getLimitedValue());
6054 LValue AllocatorTraitsLVal = CGF.EmitLValue(AllocatorTraits);
6055 Address Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
6056 AllocatorTraitsLVal.getAddress(CGF), CGF.VoidPtrPtrTy, CGF.VoidPtrTy);
6057 AllocatorTraitsLVal = CGF.MakeAddrLValue(Addr, CGF.getContext().VoidPtrTy,
6058 AllocatorTraitsLVal.getBaseInfo(),
6059 AllocatorTraitsLVal.getTBAAInfo());
6060 llvm::Value *Traits =
6061 CGF.EmitLoadOfScalar(AllocatorTraitsLVal, AllocatorTraits->getExprLoc());
6062
6063 llvm::Value *AllocatorVal =
6064 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
6065 CGM.getModule(), OMPRTL___kmpc_init_allocator),
6066 {ThreadId, MemSpaceHandle, NumTraits, Traits});
6067 // Store to allocator.
6068 CGF.EmitVarDecl(*cast<VarDecl>(
6069 cast<DeclRefExpr>(Allocator->IgnoreParenImpCasts())->getDecl()));
6070 LValue AllocatorLVal = CGF.EmitLValue(Allocator->IgnoreParenImpCasts());
6071 AllocatorVal =
6072 CGF.EmitScalarConversion(AllocatorVal, CGF.getContext().VoidPtrTy,
6073 Allocator->getType(), Allocator->getExprLoc());
6074 CGF.EmitStoreOfScalar(AllocatorVal, AllocatorLVal);
6075 }
6076
emitUsesAllocatorsFini(CodeGenFunction & CGF,const Expr * Allocator)6077 void CGOpenMPRuntime::emitUsesAllocatorsFini(CodeGenFunction &CGF,
6078 const Expr *Allocator) {
6079 llvm::Value *ThreadId = getThreadID(CGF, Allocator->getExprLoc());
6080 ThreadId = CGF.Builder.CreateIntCast(ThreadId, CGF.IntTy, /*isSigned=*/true);
6081 LValue AllocatorLVal = CGF.EmitLValue(Allocator->IgnoreParenImpCasts());
6082 llvm::Value *AllocatorVal =
6083 CGF.EmitLoadOfScalar(AllocatorLVal, Allocator->getExprLoc());
6084 AllocatorVal = CGF.EmitScalarConversion(AllocatorVal, Allocator->getType(),
6085 CGF.getContext().VoidPtrTy,
6086 Allocator->getExprLoc());
6087 (void)CGF.EmitRuntimeCall(
6088 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
6089 OMPRTL___kmpc_destroy_allocator),
6090 {ThreadId, AllocatorVal});
6091 }
6092
emitTargetOutlinedFunctionHelper(const OMPExecutableDirective & D,StringRef ParentName,llvm::Function * & OutlinedFn,llvm::Constant * & OutlinedFnID,bool IsOffloadEntry,const RegionCodeGenTy & CodeGen)6093 void CGOpenMPRuntime::emitTargetOutlinedFunctionHelper(
6094 const OMPExecutableDirective &D, StringRef ParentName,
6095 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
6096 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
6097
6098 auto EntryInfo =
6099 getTargetEntryUniqueInfo(CGM.getContext(), D.getBeginLoc(), ParentName);
6100
6101 CodeGenFunction CGF(CGM, true);
6102 llvm::OpenMPIRBuilder::FunctionGenCallback &&GenerateOutlinedFunction =
6103 [&CGF, &D, &CodeGen](StringRef EntryFnName) {
6104 const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target);
6105
6106 CGOpenMPTargetRegionInfo CGInfo(CS, CodeGen, EntryFnName);
6107 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
6108 return CGF.GenerateOpenMPCapturedStmtFunction(CS, D.getBeginLoc());
6109 };
6110
6111 // Get NumTeams and ThreadLimit attributes
6112 int32_t DefaultValTeams = -1;
6113 int32_t DefaultValThreads = -1;
6114 getNumTeamsExprForTargetDirective(CGF, D, DefaultValTeams);
6115 getNumThreadsExprForTargetDirective(CGF, D, DefaultValThreads);
6116
6117 OMPBuilder.emitTargetRegionFunction(OffloadEntriesInfoManager, EntryInfo,
6118 GenerateOutlinedFunction, DefaultValTeams,
6119 DefaultValThreads, IsOffloadEntry,
6120 OutlinedFn, OutlinedFnID);
6121
6122 if (OutlinedFn != nullptr)
6123 CGM.getTargetCodeGenInfo().setTargetAttributes(nullptr, OutlinedFn, CGM);
6124 }
6125
6126 /// Checks if the expression is constant or does not have non-trivial function
6127 /// calls.
isTrivial(ASTContext & Ctx,const Expr * E)6128 static bool isTrivial(ASTContext &Ctx, const Expr * E) {
6129 // We can skip constant expressions.
6130 // We can skip expressions with trivial calls or simple expressions.
6131 return (E->isEvaluatable(Ctx, Expr::SE_AllowUndefinedBehavior) ||
6132 !E->hasNonTrivialCall(Ctx)) &&
6133 !E->HasSideEffects(Ctx, /*IncludePossibleEffects=*/true);
6134 }
6135
getSingleCompoundChild(ASTContext & Ctx,const Stmt * Body)6136 const Stmt *CGOpenMPRuntime::getSingleCompoundChild(ASTContext &Ctx,
6137 const Stmt *Body) {
6138 const Stmt *Child = Body->IgnoreContainers();
6139 while (const auto *C = dyn_cast_or_null<CompoundStmt>(Child)) {
6140 Child = nullptr;
6141 for (const Stmt *S : C->body()) {
6142 if (const auto *E = dyn_cast<Expr>(S)) {
6143 if (isTrivial(Ctx, E))
6144 continue;
6145 }
6146 // Some of the statements can be ignored.
6147 if (isa<AsmStmt>(S) || isa<NullStmt>(S) || isa<OMPFlushDirective>(S) ||
6148 isa<OMPBarrierDirective>(S) || isa<OMPTaskyieldDirective>(S))
6149 continue;
6150 // Analyze declarations.
6151 if (const auto *DS = dyn_cast<DeclStmt>(S)) {
6152 if (llvm::all_of(DS->decls(), [](const Decl *D) {
6153 if (isa<EmptyDecl>(D) || isa<DeclContext>(D) ||
6154 isa<TypeDecl>(D) || isa<PragmaCommentDecl>(D) ||
6155 isa<PragmaDetectMismatchDecl>(D) || isa<UsingDecl>(D) ||
6156 isa<UsingDirectiveDecl>(D) ||
6157 isa<OMPDeclareReductionDecl>(D) ||
6158 isa<OMPThreadPrivateDecl>(D) || isa<OMPAllocateDecl>(D))
6159 return true;
6160 const auto *VD = dyn_cast<VarDecl>(D);
6161 if (!VD)
6162 return false;
6163 return VD->hasGlobalStorage() || !VD->isUsed();
6164 }))
6165 continue;
6166 }
6167 // Found multiple children - cannot get the one child only.
6168 if (Child)
6169 return nullptr;
6170 Child = S;
6171 }
6172 if (Child)
6173 Child = Child->IgnoreContainers();
6174 }
6175 return Child;
6176 }
6177
getNumTeamsExprForTargetDirective(CodeGenFunction & CGF,const OMPExecutableDirective & D,int32_t & DefaultVal)6178 const Expr *CGOpenMPRuntime::getNumTeamsExprForTargetDirective(
6179 CodeGenFunction &CGF, const OMPExecutableDirective &D,
6180 int32_t &DefaultVal) {
6181
6182 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
6183 assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&
6184 "Expected target-based executable directive.");
6185 switch (DirectiveKind) {
6186 case OMPD_target: {
6187 const auto *CS = D.getInnermostCapturedStmt();
6188 const auto *Body =
6189 CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
6190 const Stmt *ChildStmt =
6191 CGOpenMPRuntime::getSingleCompoundChild(CGF.getContext(), Body);
6192 if (const auto *NestedDir =
6193 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
6194 if (isOpenMPTeamsDirective(NestedDir->getDirectiveKind())) {
6195 if (NestedDir->hasClausesOfKind<OMPNumTeamsClause>()) {
6196 const Expr *NumTeams =
6197 NestedDir->getSingleClause<OMPNumTeamsClause>()->getNumTeams();
6198 if (NumTeams->isIntegerConstantExpr(CGF.getContext()))
6199 if (auto Constant =
6200 NumTeams->getIntegerConstantExpr(CGF.getContext()))
6201 DefaultVal = Constant->getExtValue();
6202 return NumTeams;
6203 }
6204 DefaultVal = 0;
6205 return nullptr;
6206 }
6207 if (isOpenMPParallelDirective(NestedDir->getDirectiveKind()) ||
6208 isOpenMPSimdDirective(NestedDir->getDirectiveKind())) {
6209 DefaultVal = 1;
6210 return nullptr;
6211 }
6212 DefaultVal = 1;
6213 return nullptr;
6214 }
6215 // A value of -1 is used to check if we need to emit no teams region
6216 DefaultVal = -1;
6217 return nullptr;
6218 }
6219 case OMPD_target_teams:
6220 case OMPD_target_teams_distribute:
6221 case OMPD_target_teams_distribute_simd:
6222 case OMPD_target_teams_distribute_parallel_for:
6223 case OMPD_target_teams_distribute_parallel_for_simd: {
6224 if (D.hasClausesOfKind<OMPNumTeamsClause>()) {
6225 const Expr *NumTeams =
6226 D.getSingleClause<OMPNumTeamsClause>()->getNumTeams();
6227 if (NumTeams->isIntegerConstantExpr(CGF.getContext()))
6228 if (auto Constant = NumTeams->getIntegerConstantExpr(CGF.getContext()))
6229 DefaultVal = Constant->getExtValue();
6230 return NumTeams;
6231 }
6232 DefaultVal = 0;
6233 return nullptr;
6234 }
6235 case OMPD_target_parallel:
6236 case OMPD_target_parallel_for:
6237 case OMPD_target_parallel_for_simd:
6238 case OMPD_target_simd:
6239 DefaultVal = 1;
6240 return nullptr;
6241 case OMPD_parallel:
6242 case OMPD_for:
6243 case OMPD_parallel_for:
6244 case OMPD_parallel_master:
6245 case OMPD_parallel_sections:
6246 case OMPD_for_simd:
6247 case OMPD_parallel_for_simd:
6248 case OMPD_cancel:
6249 case OMPD_cancellation_point:
6250 case OMPD_ordered:
6251 case OMPD_threadprivate:
6252 case OMPD_allocate:
6253 case OMPD_task:
6254 case OMPD_simd:
6255 case OMPD_tile:
6256 case OMPD_unroll:
6257 case OMPD_sections:
6258 case OMPD_section:
6259 case OMPD_single:
6260 case OMPD_master:
6261 case OMPD_critical:
6262 case OMPD_taskyield:
6263 case OMPD_barrier:
6264 case OMPD_taskwait:
6265 case OMPD_taskgroup:
6266 case OMPD_atomic:
6267 case OMPD_flush:
6268 case OMPD_depobj:
6269 case OMPD_scan:
6270 case OMPD_teams:
6271 case OMPD_target_data:
6272 case OMPD_target_exit_data:
6273 case OMPD_target_enter_data:
6274 case OMPD_distribute:
6275 case OMPD_distribute_simd:
6276 case OMPD_distribute_parallel_for:
6277 case OMPD_distribute_parallel_for_simd:
6278 case OMPD_teams_distribute:
6279 case OMPD_teams_distribute_simd:
6280 case OMPD_teams_distribute_parallel_for:
6281 case OMPD_teams_distribute_parallel_for_simd:
6282 case OMPD_target_update:
6283 case OMPD_declare_simd:
6284 case OMPD_declare_variant:
6285 case OMPD_begin_declare_variant:
6286 case OMPD_end_declare_variant:
6287 case OMPD_declare_target:
6288 case OMPD_end_declare_target:
6289 case OMPD_declare_reduction:
6290 case OMPD_declare_mapper:
6291 case OMPD_taskloop:
6292 case OMPD_taskloop_simd:
6293 case OMPD_master_taskloop:
6294 case OMPD_master_taskloop_simd:
6295 case OMPD_parallel_master_taskloop:
6296 case OMPD_parallel_master_taskloop_simd:
6297 case OMPD_requires:
6298 case OMPD_metadirective:
6299 case OMPD_unknown:
6300 break;
6301 default:
6302 break;
6303 }
6304 llvm_unreachable("Unexpected directive kind.");
6305 }
6306
emitNumTeamsForTargetDirective(CodeGenFunction & CGF,const OMPExecutableDirective & D)6307 llvm::Value *CGOpenMPRuntime::emitNumTeamsForTargetDirective(
6308 CodeGenFunction &CGF, const OMPExecutableDirective &D) {
6309 assert(!CGF.getLangOpts().OpenMPIsDevice &&
6310 "Clauses associated with the teams directive expected to be emitted "
6311 "only for the host!");
6312 CGBuilderTy &Bld = CGF.Builder;
6313 int32_t DefaultNT = -1;
6314 const Expr *NumTeams = getNumTeamsExprForTargetDirective(CGF, D, DefaultNT);
6315 if (NumTeams != nullptr) {
6316 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
6317
6318 switch (DirectiveKind) {
6319 case OMPD_target: {
6320 const auto *CS = D.getInnermostCapturedStmt();
6321 CGOpenMPInnerExprInfo CGInfo(CGF, *CS);
6322 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
6323 llvm::Value *NumTeamsVal = CGF.EmitScalarExpr(NumTeams,
6324 /*IgnoreResultAssign*/ true);
6325 return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty,
6326 /*isSigned=*/true);
6327 }
6328 case OMPD_target_teams:
6329 case OMPD_target_teams_distribute:
6330 case OMPD_target_teams_distribute_simd:
6331 case OMPD_target_teams_distribute_parallel_for:
6332 case OMPD_target_teams_distribute_parallel_for_simd: {
6333 CodeGenFunction::RunCleanupsScope NumTeamsScope(CGF);
6334 llvm::Value *NumTeamsVal = CGF.EmitScalarExpr(NumTeams,
6335 /*IgnoreResultAssign*/ true);
6336 return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty,
6337 /*isSigned=*/true);
6338 }
6339 default:
6340 break;
6341 }
6342 }
6343
6344 return llvm::ConstantInt::get(CGF.Int32Ty, DefaultNT);
6345 }
6346
getNumThreads(CodeGenFunction & CGF,const CapturedStmt * CS,llvm::Value * DefaultThreadLimitVal)6347 static llvm::Value *getNumThreads(CodeGenFunction &CGF, const CapturedStmt *CS,
6348 llvm::Value *DefaultThreadLimitVal) {
6349 const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild(
6350 CGF.getContext(), CS->getCapturedStmt());
6351 if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) {
6352 if (isOpenMPParallelDirective(Dir->getDirectiveKind())) {
6353 llvm::Value *NumThreads = nullptr;
6354 llvm::Value *CondVal = nullptr;
6355 // Handle if clause. If if clause present, the number of threads is
6356 // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1.
6357 if (Dir->hasClausesOfKind<OMPIfClause>()) {
6358 CGOpenMPInnerExprInfo CGInfo(CGF, *CS);
6359 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
6360 const OMPIfClause *IfClause = nullptr;
6361 for (const auto *C : Dir->getClausesOfKind<OMPIfClause>()) {
6362 if (C->getNameModifier() == OMPD_unknown ||
6363 C->getNameModifier() == OMPD_parallel) {
6364 IfClause = C;
6365 break;
6366 }
6367 }
6368 if (IfClause) {
6369 const Expr *Cond = IfClause->getCondition();
6370 bool Result;
6371 if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) {
6372 if (!Result)
6373 return CGF.Builder.getInt32(1);
6374 } else {
6375 CodeGenFunction::LexicalScope Scope(CGF, Cond->getSourceRange());
6376 if (const auto *PreInit =
6377 cast_or_null<DeclStmt>(IfClause->getPreInitStmt())) {
6378 for (const auto *I : PreInit->decls()) {
6379 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
6380 CGF.EmitVarDecl(cast<VarDecl>(*I));
6381 } else {
6382 CodeGenFunction::AutoVarEmission Emission =
6383 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
6384 CGF.EmitAutoVarCleanups(Emission);
6385 }
6386 }
6387 }
6388 CondVal = CGF.EvaluateExprAsBool(Cond);
6389 }
6390 }
6391 }
6392 // Check the value of num_threads clause iff if clause was not specified
6393 // or is not evaluated to false.
6394 if (Dir->hasClausesOfKind<OMPNumThreadsClause>()) {
6395 CGOpenMPInnerExprInfo CGInfo(CGF, *CS);
6396 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
6397 const auto *NumThreadsClause =
6398 Dir->getSingleClause<OMPNumThreadsClause>();
6399 CodeGenFunction::LexicalScope Scope(
6400 CGF, NumThreadsClause->getNumThreads()->getSourceRange());
6401 if (const auto *PreInit =
6402 cast_or_null<DeclStmt>(NumThreadsClause->getPreInitStmt())) {
6403 for (const auto *I : PreInit->decls()) {
6404 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
6405 CGF.EmitVarDecl(cast<VarDecl>(*I));
6406 } else {
6407 CodeGenFunction::AutoVarEmission Emission =
6408 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
6409 CGF.EmitAutoVarCleanups(Emission);
6410 }
6411 }
6412 }
6413 NumThreads = CGF.EmitScalarExpr(NumThreadsClause->getNumThreads());
6414 NumThreads = CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty,
6415 /*isSigned=*/false);
6416 if (DefaultThreadLimitVal)
6417 NumThreads = CGF.Builder.CreateSelect(
6418 CGF.Builder.CreateICmpULT(DefaultThreadLimitVal, NumThreads),
6419 DefaultThreadLimitVal, NumThreads);
6420 } else {
6421 NumThreads = DefaultThreadLimitVal ? DefaultThreadLimitVal
6422 : CGF.Builder.getInt32(0);
6423 }
6424 // Process condition of the if clause.
6425 if (CondVal) {
6426 NumThreads = CGF.Builder.CreateSelect(CondVal, NumThreads,
6427 CGF.Builder.getInt32(1));
6428 }
6429 return NumThreads;
6430 }
6431 if (isOpenMPSimdDirective(Dir->getDirectiveKind()))
6432 return CGF.Builder.getInt32(1);
6433 }
6434 return DefaultThreadLimitVal;
6435 }
6436
getNumThreadsExprForTargetDirective(CodeGenFunction & CGF,const OMPExecutableDirective & D,int32_t & DefaultVal)6437 const Expr *CGOpenMPRuntime::getNumThreadsExprForTargetDirective(
6438 CodeGenFunction &CGF, const OMPExecutableDirective &D,
6439 int32_t &DefaultVal) {
6440 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
6441 assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&
6442 "Expected target-based executable directive.");
6443
6444 switch (DirectiveKind) {
6445 case OMPD_target:
6446 // Teams have no clause thread_limit
6447 return nullptr;
6448 case OMPD_target_teams:
6449 case OMPD_target_teams_distribute:
6450 if (D.hasClausesOfKind<OMPThreadLimitClause>()) {
6451 const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
6452 const Expr *ThreadLimit = ThreadLimitClause->getThreadLimit();
6453 if (ThreadLimit->isIntegerConstantExpr(CGF.getContext()))
6454 if (auto Constant =
6455 ThreadLimit->getIntegerConstantExpr(CGF.getContext()))
6456 DefaultVal = Constant->getExtValue();
6457 return ThreadLimit;
6458 }
6459 return nullptr;
6460 case OMPD_target_parallel:
6461 case OMPD_target_parallel_for:
6462 case OMPD_target_parallel_for_simd:
6463 case OMPD_target_teams_distribute_parallel_for:
6464 case OMPD_target_teams_distribute_parallel_for_simd: {
6465 Expr *ThreadLimit = nullptr;
6466 Expr *NumThreads = nullptr;
6467 if (D.hasClausesOfKind<OMPThreadLimitClause>()) {
6468 const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
6469 ThreadLimit = ThreadLimitClause->getThreadLimit();
6470 if (ThreadLimit->isIntegerConstantExpr(CGF.getContext()))
6471 if (auto Constant =
6472 ThreadLimit->getIntegerConstantExpr(CGF.getContext()))
6473 DefaultVal = Constant->getExtValue();
6474 }
6475 if (D.hasClausesOfKind<OMPNumThreadsClause>()) {
6476 const auto *NumThreadsClause = D.getSingleClause<OMPNumThreadsClause>();
6477 NumThreads = NumThreadsClause->getNumThreads();
6478 if (NumThreads->isIntegerConstantExpr(CGF.getContext())) {
6479 if (auto Constant =
6480 NumThreads->getIntegerConstantExpr(CGF.getContext())) {
6481 if (Constant->getExtValue() < DefaultVal) {
6482 DefaultVal = Constant->getExtValue();
6483 ThreadLimit = NumThreads;
6484 }
6485 }
6486 }
6487 }
6488 return ThreadLimit;
6489 }
6490 case OMPD_target_teams_distribute_simd:
6491 case OMPD_target_simd:
6492 DefaultVal = 1;
6493 return nullptr;
6494 case OMPD_parallel:
6495 case OMPD_for:
6496 case OMPD_parallel_for:
6497 case OMPD_parallel_master:
6498 case OMPD_parallel_sections:
6499 case OMPD_for_simd:
6500 case OMPD_parallel_for_simd:
6501 case OMPD_cancel:
6502 case OMPD_cancellation_point:
6503 case OMPD_ordered:
6504 case OMPD_threadprivate:
6505 case OMPD_allocate:
6506 case OMPD_task:
6507 case OMPD_simd:
6508 case OMPD_tile:
6509 case OMPD_unroll:
6510 case OMPD_sections:
6511 case OMPD_section:
6512 case OMPD_single:
6513 case OMPD_master:
6514 case OMPD_critical:
6515 case OMPD_taskyield:
6516 case OMPD_barrier:
6517 case OMPD_taskwait:
6518 case OMPD_taskgroup:
6519 case OMPD_atomic:
6520 case OMPD_flush:
6521 case OMPD_depobj:
6522 case OMPD_scan:
6523 case OMPD_teams:
6524 case OMPD_target_data:
6525 case OMPD_target_exit_data:
6526 case OMPD_target_enter_data:
6527 case OMPD_distribute:
6528 case OMPD_distribute_simd:
6529 case OMPD_distribute_parallel_for:
6530 case OMPD_distribute_parallel_for_simd:
6531 case OMPD_teams_distribute:
6532 case OMPD_teams_distribute_simd:
6533 case OMPD_teams_distribute_parallel_for:
6534 case OMPD_teams_distribute_parallel_for_simd:
6535 case OMPD_target_update:
6536 case OMPD_declare_simd:
6537 case OMPD_declare_variant:
6538 case OMPD_begin_declare_variant:
6539 case OMPD_end_declare_variant:
6540 case OMPD_declare_target:
6541 case OMPD_end_declare_target:
6542 case OMPD_declare_reduction:
6543 case OMPD_declare_mapper:
6544 case OMPD_taskloop:
6545 case OMPD_taskloop_simd:
6546 case OMPD_master_taskloop:
6547 case OMPD_master_taskloop_simd:
6548 case OMPD_parallel_master_taskloop:
6549 case OMPD_parallel_master_taskloop_simd:
6550 case OMPD_requires:
6551 case OMPD_unknown:
6552 break;
6553 default:
6554 break;
6555 }
6556 llvm_unreachable("Unsupported directive kind.");
6557 }
6558
emitNumThreadsForTargetDirective(CodeGenFunction & CGF,const OMPExecutableDirective & D)6559 llvm::Value *CGOpenMPRuntime::emitNumThreadsForTargetDirective(
6560 CodeGenFunction &CGF, const OMPExecutableDirective &D) {
6561 assert(!CGF.getLangOpts().OpenMPIsDevice &&
6562 "Clauses associated with the teams directive expected to be emitted "
6563 "only for the host!");
6564 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
6565 assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&
6566 "Expected target-based executable directive.");
6567 CGBuilderTy &Bld = CGF.Builder;
6568 llvm::Value *ThreadLimitVal = nullptr;
6569 llvm::Value *NumThreadsVal = nullptr;
6570 switch (DirectiveKind) {
6571 case OMPD_target: {
6572 const CapturedStmt *CS = D.getInnermostCapturedStmt();
6573 if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal))
6574 return NumThreads;
6575 const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild(
6576 CGF.getContext(), CS->getCapturedStmt());
6577 // TODO: The standard is not clear how to resolve two thread limit clauses,
6578 // let's pick the teams one if it's present, otherwise the target one.
6579 const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
6580 if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) {
6581 if (const auto *TLC = Dir->getSingleClause<OMPThreadLimitClause>()) {
6582 ThreadLimitClause = TLC;
6583 CGOpenMPInnerExprInfo CGInfo(CGF, *CS);
6584 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
6585 CodeGenFunction::LexicalScope Scope(
6586 CGF, ThreadLimitClause->getThreadLimit()->getSourceRange());
6587 if (const auto *PreInit =
6588 cast_or_null<DeclStmt>(ThreadLimitClause->getPreInitStmt())) {
6589 for (const auto *I : PreInit->decls()) {
6590 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
6591 CGF.EmitVarDecl(cast<VarDecl>(*I));
6592 } else {
6593 CodeGenFunction::AutoVarEmission Emission =
6594 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
6595 CGF.EmitAutoVarCleanups(Emission);
6596 }
6597 }
6598 }
6599 }
6600 }
6601 if (ThreadLimitClause) {
6602 llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
6603 ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
6604 ThreadLimitVal =
6605 Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false);
6606 }
6607 if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) {
6608 if (isOpenMPTeamsDirective(Dir->getDirectiveKind()) &&
6609 !isOpenMPDistributeDirective(Dir->getDirectiveKind())) {
6610 CS = Dir->getInnermostCapturedStmt();
6611 const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild(
6612 CGF.getContext(), CS->getCapturedStmt());
6613 Dir = dyn_cast_or_null<OMPExecutableDirective>(Child);
6614 }
6615 if (Dir && isOpenMPDistributeDirective(Dir->getDirectiveKind()) &&
6616 !isOpenMPSimdDirective(Dir->getDirectiveKind())) {
6617 CS = Dir->getInnermostCapturedStmt();
6618 if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal))
6619 return NumThreads;
6620 }
6621 if (Dir && isOpenMPSimdDirective(Dir->getDirectiveKind()))
6622 return Bld.getInt32(1);
6623 }
6624 return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0);
6625 }
6626 case OMPD_target_teams: {
6627 if (D.hasClausesOfKind<OMPThreadLimitClause>()) {
6628 CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF);
6629 const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
6630 llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
6631 ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
6632 ThreadLimitVal =
6633 Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false);
6634 }
6635 const CapturedStmt *CS = D.getInnermostCapturedStmt();
6636 if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal))
6637 return NumThreads;
6638 const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild(
6639 CGF.getContext(), CS->getCapturedStmt());
6640 if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) {
6641 if (Dir->getDirectiveKind() == OMPD_distribute) {
6642 CS = Dir->getInnermostCapturedStmt();
6643 if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal))
6644 return NumThreads;
6645 }
6646 }
6647 return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0);
6648 }
6649 case OMPD_target_teams_distribute:
6650 if (D.hasClausesOfKind<OMPThreadLimitClause>()) {
6651 CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF);
6652 const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
6653 llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
6654 ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
6655 ThreadLimitVal =
6656 Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false);
6657 }
6658 if (llvm::Value *NumThreads =
6659 getNumThreads(CGF, D.getInnermostCapturedStmt(), ThreadLimitVal))
6660 return NumThreads;
6661 return Bld.getInt32(0);
6662 case OMPD_target_parallel:
6663 case OMPD_target_parallel_for:
6664 case OMPD_target_parallel_for_simd:
6665 case OMPD_target_teams_distribute_parallel_for:
6666 case OMPD_target_teams_distribute_parallel_for_simd: {
6667 llvm::Value *CondVal = nullptr;
6668 // Handle if clause. If if clause present, the number of threads is
6669 // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1.
6670 if (D.hasClausesOfKind<OMPIfClause>()) {
6671 const OMPIfClause *IfClause = nullptr;
6672 for (const auto *C : D.getClausesOfKind<OMPIfClause>()) {
6673 if (C->getNameModifier() == OMPD_unknown ||
6674 C->getNameModifier() == OMPD_parallel) {
6675 IfClause = C;
6676 break;
6677 }
6678 }
6679 if (IfClause) {
6680 const Expr *Cond = IfClause->getCondition();
6681 bool Result;
6682 if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) {
6683 if (!Result)
6684 return Bld.getInt32(1);
6685 } else {
6686 CodeGenFunction::RunCleanupsScope Scope(CGF);
6687 CondVal = CGF.EvaluateExprAsBool(Cond);
6688 }
6689 }
6690 }
6691 if (D.hasClausesOfKind<OMPThreadLimitClause>()) {
6692 CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF);
6693 const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
6694 llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
6695 ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
6696 ThreadLimitVal =
6697 Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*isSigned=*/false);
6698 }
6699 if (D.hasClausesOfKind<OMPNumThreadsClause>()) {
6700 CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
6701 const auto *NumThreadsClause = D.getSingleClause<OMPNumThreadsClause>();
6702 llvm::Value *NumThreads = CGF.EmitScalarExpr(
6703 NumThreadsClause->getNumThreads(), /*IgnoreResultAssign=*/true);
6704 NumThreadsVal =
6705 Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned=*/false);
6706 ThreadLimitVal = ThreadLimitVal
6707 ? Bld.CreateSelect(Bld.CreateICmpULT(NumThreadsVal,
6708 ThreadLimitVal),
6709 NumThreadsVal, ThreadLimitVal)
6710 : NumThreadsVal;
6711 }
6712 if (!ThreadLimitVal)
6713 ThreadLimitVal = Bld.getInt32(0);
6714 if (CondVal)
6715 return Bld.CreateSelect(CondVal, ThreadLimitVal, Bld.getInt32(1));
6716 return ThreadLimitVal;
6717 }
6718 case OMPD_target_teams_distribute_simd:
6719 case OMPD_target_simd:
6720 return Bld.getInt32(1);
6721 case OMPD_parallel:
6722 case OMPD_for:
6723 case OMPD_parallel_for:
6724 case OMPD_parallel_master:
6725 case OMPD_parallel_sections:
6726 case OMPD_for_simd:
6727 case OMPD_parallel_for_simd:
6728 case OMPD_cancel:
6729 case OMPD_cancellation_point:
6730 case OMPD_ordered:
6731 case OMPD_threadprivate:
6732 case OMPD_allocate:
6733 case OMPD_task:
6734 case OMPD_simd:
6735 case OMPD_tile:
6736 case OMPD_unroll:
6737 case OMPD_sections:
6738 case OMPD_section:
6739 case OMPD_single:
6740 case OMPD_master:
6741 case OMPD_critical:
6742 case OMPD_taskyield:
6743 case OMPD_barrier:
6744 case OMPD_taskwait:
6745 case OMPD_taskgroup:
6746 case OMPD_atomic:
6747 case OMPD_flush:
6748 case OMPD_depobj:
6749 case OMPD_scan:
6750 case OMPD_teams:
6751 case OMPD_target_data:
6752 case OMPD_target_exit_data:
6753 case OMPD_target_enter_data:
6754 case OMPD_distribute:
6755 case OMPD_distribute_simd:
6756 case OMPD_distribute_parallel_for:
6757 case OMPD_distribute_parallel_for_simd:
6758 case OMPD_teams_distribute:
6759 case OMPD_teams_distribute_simd:
6760 case OMPD_teams_distribute_parallel_for:
6761 case OMPD_teams_distribute_parallel_for_simd:
6762 case OMPD_target_update:
6763 case OMPD_declare_simd:
6764 case OMPD_declare_variant:
6765 case OMPD_begin_declare_variant:
6766 case OMPD_end_declare_variant:
6767 case OMPD_declare_target:
6768 case OMPD_end_declare_target:
6769 case OMPD_declare_reduction:
6770 case OMPD_declare_mapper:
6771 case OMPD_taskloop:
6772 case OMPD_taskloop_simd:
6773 case OMPD_master_taskloop:
6774 case OMPD_master_taskloop_simd:
6775 case OMPD_parallel_master_taskloop:
6776 case OMPD_parallel_master_taskloop_simd:
6777 case OMPD_requires:
6778 case OMPD_metadirective:
6779 case OMPD_unknown:
6780 break;
6781 default:
6782 break;
6783 }
6784 llvm_unreachable("Unsupported directive kind.");
6785 }
6786
6787 namespace {
6788 LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE();
6789
6790 // Utility to handle information from clauses associated with a given
6791 // construct that use mappable expressions (e.g. 'map' clause, 'to' clause).
6792 // It provides a convenient interface to obtain the information and generate
6793 // code for that information.
6794 class MappableExprsHandler {
6795 public:
6796 /// Get the offset of the OMP_MAP_MEMBER_OF field.
getFlagMemberOffset()6797 static unsigned getFlagMemberOffset() {
6798 unsigned Offset = 0;
6799 for (uint64_t Remain =
6800 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
6801 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF);
6802 !(Remain & 1); Remain = Remain >> 1)
6803 Offset++;
6804 return Offset;
6805 }
6806
6807 /// Class that holds debugging information for a data mapping to be passed to
6808 /// the runtime library.
6809 class MappingExprInfo {
6810 /// The variable declaration used for the data mapping.
6811 const ValueDecl *MapDecl = nullptr;
6812 /// The original expression used in the map clause, or null if there is
6813 /// none.
6814 const Expr *MapExpr = nullptr;
6815
6816 public:
MappingExprInfo(const ValueDecl * MapDecl,const Expr * MapExpr=nullptr)6817 MappingExprInfo(const ValueDecl *MapDecl, const Expr *MapExpr = nullptr)
6818 : MapDecl(MapDecl), MapExpr(MapExpr) {}
6819
getMapDecl() const6820 const ValueDecl *getMapDecl() const { return MapDecl; }
getMapExpr() const6821 const Expr *getMapExpr() const { return MapExpr; }
6822 };
6823
6824 /// Class that associates information with a base pointer to be passed to the
6825 /// runtime library.
6826 class BasePointerInfo {
6827 /// The base pointer.
6828 llvm::Value *Ptr = nullptr;
6829 /// The base declaration that refers to this device pointer, or null if
6830 /// there is none.
6831 const ValueDecl *DevPtrDecl = nullptr;
6832
6833 public:
BasePointerInfo(llvm::Value * Ptr,const ValueDecl * DevPtrDecl=nullptr)6834 BasePointerInfo(llvm::Value *Ptr, const ValueDecl *DevPtrDecl = nullptr)
6835 : Ptr(Ptr), DevPtrDecl(DevPtrDecl) {}
operator *() const6836 llvm::Value *operator*() const { return Ptr; }
getDevicePtrDecl() const6837 const ValueDecl *getDevicePtrDecl() const { return DevPtrDecl; }
setDevicePtrDecl(const ValueDecl * D)6838 void setDevicePtrDecl(const ValueDecl *D) { DevPtrDecl = D; }
6839 };
6840
6841 using MapExprsArrayTy = SmallVector<MappingExprInfo, 4>;
6842 using MapBaseValuesArrayTy = SmallVector<BasePointerInfo, 4>;
6843 using MapValuesArrayTy = SmallVector<llvm::Value *, 4>;
6844 using MapFlagsArrayTy = SmallVector<OpenMPOffloadMappingFlags, 4>;
6845 using MapMappersArrayTy = SmallVector<const ValueDecl *, 4>;
6846 using MapDimArrayTy = SmallVector<uint64_t, 4>;
6847 using MapNonContiguousArrayTy = SmallVector<MapValuesArrayTy, 4>;
6848
6849 /// This structure contains combined information generated for mappable
6850 /// clauses, including base pointers, pointers, sizes, map types, user-defined
6851 /// mappers, and non-contiguous information.
6852 struct MapCombinedInfoTy {
6853 struct StructNonContiguousInfo {
6854 bool IsNonContiguous = false;
6855 MapDimArrayTy Dims;
6856 MapNonContiguousArrayTy Offsets;
6857 MapNonContiguousArrayTy Counts;
6858 MapNonContiguousArrayTy Strides;
6859 };
6860 MapExprsArrayTy Exprs;
6861 MapBaseValuesArrayTy BasePointers;
6862 MapValuesArrayTy Pointers;
6863 MapValuesArrayTy Sizes;
6864 MapFlagsArrayTy Types;
6865 MapMappersArrayTy Mappers;
6866 StructNonContiguousInfo NonContigInfo;
6867
6868 /// Append arrays in \a CurInfo.
append__anon7bb087082811::MappableExprsHandler::MapCombinedInfoTy6869 void append(MapCombinedInfoTy &CurInfo) {
6870 Exprs.append(CurInfo.Exprs.begin(), CurInfo.Exprs.end());
6871 BasePointers.append(CurInfo.BasePointers.begin(),
6872 CurInfo.BasePointers.end());
6873 Pointers.append(CurInfo.Pointers.begin(), CurInfo.Pointers.end());
6874 Sizes.append(CurInfo.Sizes.begin(), CurInfo.Sizes.end());
6875 Types.append(CurInfo.Types.begin(), CurInfo.Types.end());
6876 Mappers.append(CurInfo.Mappers.begin(), CurInfo.Mappers.end());
6877 NonContigInfo.Dims.append(CurInfo.NonContigInfo.Dims.begin(),
6878 CurInfo.NonContigInfo.Dims.end());
6879 NonContigInfo.Offsets.append(CurInfo.NonContigInfo.Offsets.begin(),
6880 CurInfo.NonContigInfo.Offsets.end());
6881 NonContigInfo.Counts.append(CurInfo.NonContigInfo.Counts.begin(),
6882 CurInfo.NonContigInfo.Counts.end());
6883 NonContigInfo.Strides.append(CurInfo.NonContigInfo.Strides.begin(),
6884 CurInfo.NonContigInfo.Strides.end());
6885 }
6886 };
6887
6888 /// Map between a struct and the its lowest & highest elements which have been
6889 /// mapped.
6890 /// [ValueDecl *] --> {LE(FieldIndex, Pointer),
6891 /// HE(FieldIndex, Pointer)}
6892 struct StructRangeInfoTy {
6893 MapCombinedInfoTy PreliminaryMapData;
6894 std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> LowestElem = {
6895 0, Address::invalid()};
6896 std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> HighestElem = {
6897 0, Address::invalid()};
6898 Address Base = Address::invalid();
6899 Address LB = Address::invalid();
6900 bool IsArraySection = false;
6901 bool HasCompleteRecord = false;
6902 };
6903
6904 private:
6905 /// Kind that defines how a device pointer has to be returned.
6906 struct MapInfo {
6907 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
6908 OpenMPMapClauseKind MapType = OMPC_MAP_unknown;
6909 ArrayRef<OpenMPMapModifierKind> MapModifiers;
6910 ArrayRef<OpenMPMotionModifierKind> MotionModifiers;
6911 bool ReturnDevicePointer = false;
6912 bool IsImplicit = false;
6913 const ValueDecl *Mapper = nullptr;
6914 const Expr *VarRef = nullptr;
6915 bool ForDeviceAddr = false;
6916
6917 MapInfo() = default;
MapInfo__anon7bb087082811::MappableExprsHandler::MapInfo6918 MapInfo(
6919 OMPClauseMappableExprCommon::MappableExprComponentListRef Components,
6920 OpenMPMapClauseKind MapType,
6921 ArrayRef<OpenMPMapModifierKind> MapModifiers,
6922 ArrayRef<OpenMPMotionModifierKind> MotionModifiers,
6923 bool ReturnDevicePointer, bool IsImplicit,
6924 const ValueDecl *Mapper = nullptr, const Expr *VarRef = nullptr,
6925 bool ForDeviceAddr = false)
6926 : Components(Components), MapType(MapType), MapModifiers(MapModifiers),
6927 MotionModifiers(MotionModifiers),
6928 ReturnDevicePointer(ReturnDevicePointer), IsImplicit(IsImplicit),
6929 Mapper(Mapper), VarRef(VarRef), ForDeviceAddr(ForDeviceAddr) {}
6930 };
6931
6932 /// If use_device_ptr or use_device_addr is used on a decl which is a struct
6933 /// member and there is no map information about it, then emission of that
6934 /// entry is deferred until the whole struct has been processed.
6935 struct DeferredDevicePtrEntryTy {
6936 const Expr *IE = nullptr;
6937 const ValueDecl *VD = nullptr;
6938 bool ForDeviceAddr = false;
6939
DeferredDevicePtrEntryTy__anon7bb087082811::MappableExprsHandler::DeferredDevicePtrEntryTy6940 DeferredDevicePtrEntryTy(const Expr *IE, const ValueDecl *VD,
6941 bool ForDeviceAddr)
6942 : IE(IE), VD(VD), ForDeviceAddr(ForDeviceAddr) {}
6943 };
6944
6945 /// The target directive from where the mappable clauses were extracted. It
6946 /// is either a executable directive or a user-defined mapper directive.
6947 llvm::PointerUnion<const OMPExecutableDirective *,
6948 const OMPDeclareMapperDecl *>
6949 CurDir;
6950
6951 /// Function the directive is being generated for.
6952 CodeGenFunction &CGF;
6953
6954 /// Set of all first private variables in the current directive.
6955 /// bool data is set to true if the variable is implicitly marked as
6956 /// firstprivate, false otherwise.
6957 llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, bool> FirstPrivateDecls;
6958
6959 /// Map between device pointer declarations and their expression components.
6960 /// The key value for declarations in 'this' is null.
6961 llvm::DenseMap<
6962 const ValueDecl *,
6963 SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>>
6964 DevPointersMap;
6965
6966 /// Map between device addr declarations and their expression components.
6967 /// The key value for declarations in 'this' is null.
6968 llvm::DenseMap<
6969 const ValueDecl *,
6970 SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>>
6971 HasDevAddrsMap;
6972
6973 /// Map between lambda declarations and their map type.
6974 llvm::DenseMap<const ValueDecl *, const OMPMapClause *> LambdasMap;
6975
getExprTypeSize(const Expr * E) const6976 llvm::Value *getExprTypeSize(const Expr *E) const {
6977 QualType ExprTy = E->getType().getCanonicalType();
6978
6979 // Calculate the size for array shaping expression.
6980 if (const auto *OAE = dyn_cast<OMPArrayShapingExpr>(E)) {
6981 llvm::Value *Size =
6982 CGF.getTypeSize(OAE->getBase()->getType()->getPointeeType());
6983 for (const Expr *SE : OAE->getDimensions()) {
6984 llvm::Value *Sz = CGF.EmitScalarExpr(SE);
6985 Sz = CGF.EmitScalarConversion(Sz, SE->getType(),
6986 CGF.getContext().getSizeType(),
6987 SE->getExprLoc());
6988 Size = CGF.Builder.CreateNUWMul(Size, Sz);
6989 }
6990 return Size;
6991 }
6992
6993 // Reference types are ignored for mapping purposes.
6994 if (const auto *RefTy = ExprTy->getAs<ReferenceType>())
6995 ExprTy = RefTy->getPointeeType().getCanonicalType();
6996
6997 // Given that an array section is considered a built-in type, we need to
6998 // do the calculation based on the length of the section instead of relying
6999 // on CGF.getTypeSize(E->getType()).
7000 if (const auto *OAE = dyn_cast<OMPArraySectionExpr>(E)) {
7001 QualType BaseTy = OMPArraySectionExpr::getBaseOriginalType(
7002 OAE->getBase()->IgnoreParenImpCasts())
7003 .getCanonicalType();
7004
7005 // If there is no length associated with the expression and lower bound is
7006 // not specified too, that means we are using the whole length of the
7007 // base.
7008 if (!OAE->getLength() && OAE->getColonLocFirst().isValid() &&
7009 !OAE->getLowerBound())
7010 return CGF.getTypeSize(BaseTy);
7011
7012 llvm::Value *ElemSize;
7013 if (const auto *PTy = BaseTy->getAs<PointerType>()) {
7014 ElemSize = CGF.getTypeSize(PTy->getPointeeType().getCanonicalType());
7015 } else {
7016 const auto *ATy = cast<ArrayType>(BaseTy.getTypePtr());
7017 assert(ATy && "Expecting array type if not a pointer type.");
7018 ElemSize = CGF.getTypeSize(ATy->getElementType().getCanonicalType());
7019 }
7020
7021 // If we don't have a length at this point, that is because we have an
7022 // array section with a single element.
7023 if (!OAE->getLength() && OAE->getColonLocFirst().isInvalid())
7024 return ElemSize;
7025
7026 if (const Expr *LenExpr = OAE->getLength()) {
7027 llvm::Value *LengthVal = CGF.EmitScalarExpr(LenExpr);
7028 LengthVal = CGF.EmitScalarConversion(LengthVal, LenExpr->getType(),
7029 CGF.getContext().getSizeType(),
7030 LenExpr->getExprLoc());
7031 return CGF.Builder.CreateNUWMul(LengthVal, ElemSize);
7032 }
7033 assert(!OAE->getLength() && OAE->getColonLocFirst().isValid() &&
7034 OAE->getLowerBound() && "expected array_section[lb:].");
7035 // Size = sizetype - lb * elemtype;
7036 llvm::Value *LengthVal = CGF.getTypeSize(BaseTy);
7037 llvm::Value *LBVal = CGF.EmitScalarExpr(OAE->getLowerBound());
7038 LBVal = CGF.EmitScalarConversion(LBVal, OAE->getLowerBound()->getType(),
7039 CGF.getContext().getSizeType(),
7040 OAE->getLowerBound()->getExprLoc());
7041 LBVal = CGF.Builder.CreateNUWMul(LBVal, ElemSize);
7042 llvm::Value *Cmp = CGF.Builder.CreateICmpUGT(LengthVal, LBVal);
7043 llvm::Value *TrueVal = CGF.Builder.CreateNUWSub(LengthVal, LBVal);
7044 LengthVal = CGF.Builder.CreateSelect(
7045 Cmp, TrueVal, llvm::ConstantInt::get(CGF.SizeTy, 0));
7046 return LengthVal;
7047 }
7048 return CGF.getTypeSize(ExprTy);
7049 }
7050
7051 /// Return the corresponding bits for a given map clause modifier. Add
7052 /// a flag marking the map as a pointer if requested. Add a flag marking the
7053 /// map as the first one of a series of maps that relate to the same map
7054 /// expression.
getMapTypeBits(OpenMPMapClauseKind MapType,ArrayRef<OpenMPMapModifierKind> MapModifiers,ArrayRef<OpenMPMotionModifierKind> MotionModifiers,bool IsImplicit,bool AddPtrFlag,bool AddIsTargetParamFlag,bool IsNonContiguous) const7055 OpenMPOffloadMappingFlags getMapTypeBits(
7056 OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers,
7057 ArrayRef<OpenMPMotionModifierKind> MotionModifiers, bool IsImplicit,
7058 bool AddPtrFlag, bool AddIsTargetParamFlag, bool IsNonContiguous) const {
7059 OpenMPOffloadMappingFlags Bits =
7060 IsImplicit ? OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT
7061 : OpenMPOffloadMappingFlags::OMP_MAP_NONE;
7062 switch (MapType) {
7063 case OMPC_MAP_alloc:
7064 case OMPC_MAP_release:
7065 // alloc and release is the default behavior in the runtime library, i.e.
7066 // if we don't pass any bits alloc/release that is what the runtime is
7067 // going to do. Therefore, we don't need to signal anything for these two
7068 // type modifiers.
7069 break;
7070 case OMPC_MAP_to:
7071 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_TO;
7072 break;
7073 case OMPC_MAP_from:
7074 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_FROM;
7075 break;
7076 case OMPC_MAP_tofrom:
7077 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_TO |
7078 OpenMPOffloadMappingFlags::OMP_MAP_FROM;
7079 break;
7080 case OMPC_MAP_delete:
7081 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_DELETE;
7082 break;
7083 case OMPC_MAP_unknown:
7084 llvm_unreachable("Unexpected map type!");
7085 }
7086 if (AddPtrFlag)
7087 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ;
7088 if (AddIsTargetParamFlag)
7089 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
7090 if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_always))
7091 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_ALWAYS;
7092 if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_close))
7093 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_CLOSE;
7094 if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_present) ||
7095 llvm::is_contained(MotionModifiers, OMPC_MOTION_MODIFIER_present))
7096 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_PRESENT;
7097 if (llvm::is_contained(MapModifiers, OMPC_MAP_MODIFIER_ompx_hold))
7098 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_OMPX_HOLD;
7099 if (IsNonContiguous)
7100 Bits |= OpenMPOffloadMappingFlags::OMP_MAP_NON_CONTIG;
7101 return Bits;
7102 }
7103
7104 /// Return true if the provided expression is a final array section. A
7105 /// final array section, is one whose length can't be proved to be one.
isFinalArraySectionExpression(const Expr * E) const7106 bool isFinalArraySectionExpression(const Expr *E) const {
7107 const auto *OASE = dyn_cast<OMPArraySectionExpr>(E);
7108
7109 // It is not an array section and therefore not a unity-size one.
7110 if (!OASE)
7111 return false;
7112
7113 // An array section with no colon always refer to a single element.
7114 if (OASE->getColonLocFirst().isInvalid())
7115 return false;
7116
7117 const Expr *Length = OASE->getLength();
7118
7119 // If we don't have a length we have to check if the array has size 1
7120 // for this dimension. Also, we should always expect a length if the
7121 // base type is pointer.
7122 if (!Length) {
7123 QualType BaseQTy = OMPArraySectionExpr::getBaseOriginalType(
7124 OASE->getBase()->IgnoreParenImpCasts())
7125 .getCanonicalType();
7126 if (const auto *ATy = dyn_cast<ConstantArrayType>(BaseQTy.getTypePtr()))
7127 return ATy->getSize().getSExtValue() != 1;
7128 // If we don't have a constant dimension length, we have to consider
7129 // the current section as having any size, so it is not necessarily
7130 // unitary. If it happen to be unity size, that's user fault.
7131 return true;
7132 }
7133
7134 // Check if the length evaluates to 1.
7135 Expr::EvalResult Result;
7136 if (!Length->EvaluateAsInt(Result, CGF.getContext()))
7137 return true; // Can have more that size 1.
7138
7139 llvm::APSInt ConstLength = Result.Val.getInt();
7140 return ConstLength.getSExtValue() != 1;
7141 }
7142
7143 /// Generate the base pointers, section pointers, sizes, map type bits, and
7144 /// user-defined mappers (all included in \a CombinedInfo) for the provided
7145 /// map type, map or motion modifiers, and expression components.
7146 /// \a IsFirstComponent should be set to true if the provided set of
7147 /// components is the first associated with a capture.
generateInfoForComponentList(OpenMPMapClauseKind MapType,ArrayRef<OpenMPMapModifierKind> MapModifiers,ArrayRef<OpenMPMotionModifierKind> MotionModifiers,OMPClauseMappableExprCommon::MappableExprComponentListRef Components,MapCombinedInfoTy & CombinedInfo,StructRangeInfoTy & PartialStruct,bool IsFirstComponentList,bool IsImplicit,const ValueDecl * Mapper=nullptr,bool ForDeviceAddr=false,const ValueDecl * BaseDecl=nullptr,const Expr * MapExpr=nullptr,ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> OverlappedElements=std::nullopt) const7148 void generateInfoForComponentList(
7149 OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers,
7150 ArrayRef<OpenMPMotionModifierKind> MotionModifiers,
7151 OMPClauseMappableExprCommon::MappableExprComponentListRef Components,
7152 MapCombinedInfoTy &CombinedInfo, StructRangeInfoTy &PartialStruct,
7153 bool IsFirstComponentList, bool IsImplicit,
7154 const ValueDecl *Mapper = nullptr, bool ForDeviceAddr = false,
7155 const ValueDecl *BaseDecl = nullptr, const Expr *MapExpr = nullptr,
7156 ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef>
7157 OverlappedElements = std::nullopt) const {
7158 // The following summarizes what has to be generated for each map and the
7159 // types below. The generated information is expressed in this order:
7160 // base pointer, section pointer, size, flags
7161 // (to add to the ones that come from the map type and modifier).
7162 //
7163 // double d;
7164 // int i[100];
7165 // float *p;
7166 //
7167 // struct S1 {
7168 // int i;
7169 // float f[50];
7170 // }
7171 // struct S2 {
7172 // int i;
7173 // float f[50];
7174 // S1 s;
7175 // double *p;
7176 // struct S2 *ps;
7177 // int &ref;
7178 // }
7179 // S2 s;
7180 // S2 *ps;
7181 //
7182 // map(d)
7183 // &d, &d, sizeof(double), TARGET_PARAM | TO | FROM
7184 //
7185 // map(i)
7186 // &i, &i, 100*sizeof(int), TARGET_PARAM | TO | FROM
7187 //
7188 // map(i[1:23])
7189 // &i(=&i[0]), &i[1], 23*sizeof(int), TARGET_PARAM | TO | FROM
7190 //
7191 // map(p)
7192 // &p, &p, sizeof(float*), TARGET_PARAM | TO | FROM
7193 //
7194 // map(p[1:24])
7195 // &p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM | PTR_AND_OBJ
7196 // in unified shared memory mode or for local pointers
7197 // p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM
7198 //
7199 // map(s)
7200 // &s, &s, sizeof(S2), TARGET_PARAM | TO | FROM
7201 //
7202 // map(s.i)
7203 // &s, &(s.i), sizeof(int), TARGET_PARAM | TO | FROM
7204 //
7205 // map(s.s.f)
7206 // &s, &(s.s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM
7207 //
7208 // map(s.p)
7209 // &s, &(s.p), sizeof(double*), TARGET_PARAM | TO | FROM
7210 //
7211 // map(to: s.p[:22])
7212 // &s, &(s.p), sizeof(double*), TARGET_PARAM (*)
7213 // &s, &(s.p), sizeof(double*), MEMBER_OF(1) (**)
7214 // &(s.p), &(s.p[0]), 22*sizeof(double),
7215 // MEMBER_OF(1) | PTR_AND_OBJ | TO (***)
7216 // (*) alloc space for struct members, only this is a target parameter
7217 // (**) map the pointer (nothing to be mapped in this example) (the compiler
7218 // optimizes this entry out, same in the examples below)
7219 // (***) map the pointee (map: to)
7220 //
7221 // map(to: s.ref)
7222 // &s, &(s.ref), sizeof(int*), TARGET_PARAM (*)
7223 // &s, &(s.ref), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | TO (***)
7224 // (*) alloc space for struct members, only this is a target parameter
7225 // (**) map the pointer (nothing to be mapped in this example) (the compiler
7226 // optimizes this entry out, same in the examples below)
7227 // (***) map the pointee (map: to)
7228 //
7229 // map(s.ps)
7230 // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | TO | FROM
7231 //
7232 // map(from: s.ps->s.i)
7233 // &s, &(s.ps), sizeof(S2*), TARGET_PARAM
7234 // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1)
7235 // &(s.ps), &(s.ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM
7236 //
7237 // map(to: s.ps->ps)
7238 // &s, &(s.ps), sizeof(S2*), TARGET_PARAM
7239 // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1)
7240 // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | TO
7241 //
7242 // map(s.ps->ps->ps)
7243 // &s, &(s.ps), sizeof(S2*), TARGET_PARAM
7244 // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1)
7245 // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ
7246 // &(s.ps->ps), &(s.ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM
7247 //
7248 // map(to: s.ps->ps->s.f[:22])
7249 // &s, &(s.ps), sizeof(S2*), TARGET_PARAM
7250 // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1)
7251 // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ
7252 // &(s.ps->ps), &(s.ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO
7253 //
7254 // map(ps)
7255 // &ps, &ps, sizeof(S2*), TARGET_PARAM | TO | FROM
7256 //
7257 // map(ps->i)
7258 // ps, &(ps->i), sizeof(int), TARGET_PARAM | TO | FROM
7259 //
7260 // map(ps->s.f)
7261 // ps, &(ps->s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM
7262 //
7263 // map(from: ps->p)
7264 // ps, &(ps->p), sizeof(double*), TARGET_PARAM | FROM
7265 //
7266 // map(to: ps->p[:22])
7267 // ps, &(ps->p), sizeof(double*), TARGET_PARAM
7268 // ps, &(ps->p), sizeof(double*), MEMBER_OF(1)
7269 // &(ps->p), &(ps->p[0]), 22*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | TO
7270 //
7271 // map(ps->ps)
7272 // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | TO | FROM
7273 //
7274 // map(from: ps->ps->s.i)
7275 // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM
7276 // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1)
7277 // &(ps->ps), &(ps->ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM
7278 //
7279 // map(from: ps->ps->ps)
7280 // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM
7281 // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1)
7282 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | FROM
7283 //
7284 // map(ps->ps->ps->ps)
7285 // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM
7286 // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1)
7287 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ
7288 // &(ps->ps->ps), &(ps->ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM
7289 //
7290 // map(to: ps->ps->ps->s.f[:22])
7291 // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM
7292 // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1)
7293 // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ
7294 // &(ps->ps->ps), &(ps->ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO
7295 //
7296 // map(to: s.f[:22]) map(from: s.p[:33])
7297 // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1) +
7298 // sizeof(double*) (**), TARGET_PARAM
7299 // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | TO
7300 // &s, &(s.p), sizeof(double*), MEMBER_OF(1)
7301 // &(s.p), &(s.p[0]), 33*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | FROM
7302 // (*) allocate contiguous space needed to fit all mapped members even if
7303 // we allocate space for members not mapped (in this example,
7304 // s.f[22..49] and s.s are not mapped, yet we must allocate space for
7305 // them as well because they fall between &s.f[0] and &s.p)
7306 //
7307 // map(from: s.f[:22]) map(to: ps->p[:33])
7308 // &s, &(s.f[0]), 22*sizeof(float), TARGET_PARAM | FROM
7309 // ps, &(ps->p), sizeof(S2*), TARGET_PARAM
7310 // ps, &(ps->p), sizeof(double*), MEMBER_OF(2) (*)
7311 // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(2) | PTR_AND_OBJ | TO
7312 // (*) the struct this entry pertains to is the 2nd element in the list of
7313 // arguments, hence MEMBER_OF(2)
7314 //
7315 // map(from: s.f[:22], s.s) map(to: ps->p[:33])
7316 // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1), TARGET_PARAM
7317 // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | FROM
7318 // &s, &(s.s), sizeof(struct S1), MEMBER_OF(1) | FROM
7319 // ps, &(ps->p), sizeof(S2*), TARGET_PARAM
7320 // ps, &(ps->p), sizeof(double*), MEMBER_OF(4) (*)
7321 // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(4) | PTR_AND_OBJ | TO
7322 // (*) the struct this entry pertains to is the 4th element in the list
7323 // of arguments, hence MEMBER_OF(4)
7324
7325 // Track if the map information being generated is the first for a capture.
7326 bool IsCaptureFirstInfo = IsFirstComponentList;
7327 // When the variable is on a declare target link or in a to clause with
7328 // unified memory, a reference is needed to hold the host/device address
7329 // of the variable.
7330 bool RequiresReference = false;
7331
7332 // Scan the components from the base to the complete expression.
7333 auto CI = Components.rbegin();
7334 auto CE = Components.rend();
7335 auto I = CI;
7336
7337 // Track if the map information being generated is the first for a list of
7338 // components.
7339 bool IsExpressionFirstInfo = true;
7340 bool FirstPointerInComplexData = false;
7341 Address BP = Address::invalid();
7342 const Expr *AssocExpr = I->getAssociatedExpression();
7343 const auto *AE = dyn_cast<ArraySubscriptExpr>(AssocExpr);
7344 const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr);
7345 const auto *OAShE = dyn_cast<OMPArrayShapingExpr>(AssocExpr);
7346
7347 if (isa<MemberExpr>(AssocExpr)) {
7348 // The base is the 'this' pointer. The content of the pointer is going
7349 // to be the base of the field being mapped.
7350 BP = CGF.LoadCXXThisAddress();
7351 } else if ((AE && isa<CXXThisExpr>(AE->getBase()->IgnoreParenImpCasts())) ||
7352 (OASE &&
7353 isa<CXXThisExpr>(OASE->getBase()->IgnoreParenImpCasts()))) {
7354 BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(CGF);
7355 } else if (OAShE &&
7356 isa<CXXThisExpr>(OAShE->getBase()->IgnoreParenCasts())) {
7357 BP = Address(
7358 CGF.EmitScalarExpr(OAShE->getBase()),
7359 CGF.ConvertTypeForMem(OAShE->getBase()->getType()->getPointeeType()),
7360 CGF.getContext().getTypeAlignInChars(OAShE->getBase()->getType()));
7361 } else {
7362 // The base is the reference to the variable.
7363 // BP = &Var.
7364 BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(CGF);
7365 if (const auto *VD =
7366 dyn_cast_or_null<VarDecl>(I->getAssociatedDeclaration())) {
7367 if (std::optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res =
7368 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) {
7369 if ((*Res == OMPDeclareTargetDeclAttr::MT_Link) ||
7370 ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
7371 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
7372 CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory())) {
7373 RequiresReference = true;
7374 BP = CGF.CGM.getOpenMPRuntime().getAddrOfDeclareTargetVar(VD);
7375 }
7376 }
7377 }
7378
7379 // If the variable is a pointer and is being dereferenced (i.e. is not
7380 // the last component), the base has to be the pointer itself, not its
7381 // reference. References are ignored for mapping purposes.
7382 QualType Ty =
7383 I->getAssociatedDeclaration()->getType().getNonReferenceType();
7384 if (Ty->isAnyPointerType() && std::next(I) != CE) {
7385 // No need to generate individual map information for the pointer, it
7386 // can be associated with the combined storage if shared memory mode is
7387 // active or the base declaration is not global variable.
7388 const auto *VD = dyn_cast<VarDecl>(I->getAssociatedDeclaration());
7389 if (CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory() ||
7390 !VD || VD->hasLocalStorage())
7391 BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>());
7392 else
7393 FirstPointerInComplexData = true;
7394 ++I;
7395 }
7396 }
7397
7398 // Track whether a component of the list should be marked as MEMBER_OF some
7399 // combined entry (for partial structs). Only the first PTR_AND_OBJ entry
7400 // in a component list should be marked as MEMBER_OF, all subsequent entries
7401 // do not belong to the base struct. E.g.
7402 // struct S2 s;
7403 // s.ps->ps->ps->f[:]
7404 // (1) (2) (3) (4)
7405 // ps(1) is a member pointer, ps(2) is a pointee of ps(1), so it is a
7406 // PTR_AND_OBJ entry; the PTR is ps(1), so MEMBER_OF the base struct. ps(3)
7407 // is the pointee of ps(2) which is not member of struct s, so it should not
7408 // be marked as such (it is still PTR_AND_OBJ).
7409 // The variable is initialized to false so that PTR_AND_OBJ entries which
7410 // are not struct members are not considered (e.g. array of pointers to
7411 // data).
7412 bool ShouldBeMemberOf = false;
7413
7414 // Variable keeping track of whether or not we have encountered a component
7415 // in the component list which is a member expression. Useful when we have a
7416 // pointer or a final array section, in which case it is the previous
7417 // component in the list which tells us whether we have a member expression.
7418 // E.g. X.f[:]
7419 // While processing the final array section "[:]" it is "f" which tells us
7420 // whether we are dealing with a member of a declared struct.
7421 const MemberExpr *EncounteredME = nullptr;
7422
7423 // Track for the total number of dimension. Start from one for the dummy
7424 // dimension.
7425 uint64_t DimSize = 1;
7426
7427 bool IsNonContiguous = CombinedInfo.NonContigInfo.IsNonContiguous;
7428 bool IsPrevMemberReference = false;
7429
7430 for (; I != CE; ++I) {
7431 // If the current component is member of a struct (parent struct) mark it.
7432 if (!EncounteredME) {
7433 EncounteredME = dyn_cast<MemberExpr>(I->getAssociatedExpression());
7434 // If we encounter a PTR_AND_OBJ entry from now on it should be marked
7435 // as MEMBER_OF the parent struct.
7436 if (EncounteredME) {
7437 ShouldBeMemberOf = true;
7438 // Do not emit as complex pointer if this is actually not array-like
7439 // expression.
7440 if (FirstPointerInComplexData) {
7441 QualType Ty = std::prev(I)
7442 ->getAssociatedDeclaration()
7443 ->getType()
7444 .getNonReferenceType();
7445 BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>());
7446 FirstPointerInComplexData = false;
7447 }
7448 }
7449 }
7450
7451 auto Next = std::next(I);
7452
7453 // We need to generate the addresses and sizes if this is the last
7454 // component, if the component is a pointer or if it is an array section
7455 // whose length can't be proved to be one. If this is a pointer, it
7456 // becomes the base address for the following components.
7457
7458 // A final array section, is one whose length can't be proved to be one.
7459 // If the map item is non-contiguous then we don't treat any array section
7460 // as final array section.
7461 bool IsFinalArraySection =
7462 !IsNonContiguous &&
7463 isFinalArraySectionExpression(I->getAssociatedExpression());
7464
7465 // If we have a declaration for the mapping use that, otherwise use
7466 // the base declaration of the map clause.
7467 const ValueDecl *MapDecl = (I->getAssociatedDeclaration())
7468 ? I->getAssociatedDeclaration()
7469 : BaseDecl;
7470 MapExpr = (I->getAssociatedExpression()) ? I->getAssociatedExpression()
7471 : MapExpr;
7472
7473 // Get information on whether the element is a pointer. Have to do a
7474 // special treatment for array sections given that they are built-in
7475 // types.
7476 const auto *OASE =
7477 dyn_cast<OMPArraySectionExpr>(I->getAssociatedExpression());
7478 const auto *OAShE =
7479 dyn_cast<OMPArrayShapingExpr>(I->getAssociatedExpression());
7480 const auto *UO = dyn_cast<UnaryOperator>(I->getAssociatedExpression());
7481 const auto *BO = dyn_cast<BinaryOperator>(I->getAssociatedExpression());
7482 bool IsPointer =
7483 OAShE ||
7484 (OASE && OMPArraySectionExpr::getBaseOriginalType(OASE)
7485 .getCanonicalType()
7486 ->isAnyPointerType()) ||
7487 I->getAssociatedExpression()->getType()->isAnyPointerType();
7488 bool IsMemberReference = isa<MemberExpr>(I->getAssociatedExpression()) &&
7489 MapDecl &&
7490 MapDecl->getType()->isLValueReferenceType();
7491 bool IsNonDerefPointer = IsPointer && !UO && !BO && !IsNonContiguous;
7492
7493 if (OASE)
7494 ++DimSize;
7495
7496 if (Next == CE || IsMemberReference || IsNonDerefPointer ||
7497 IsFinalArraySection) {
7498 // If this is not the last component, we expect the pointer to be
7499 // associated with an array expression or member expression.
7500 assert((Next == CE ||
7501 isa<MemberExpr>(Next->getAssociatedExpression()) ||
7502 isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) ||
7503 isa<OMPArraySectionExpr>(Next->getAssociatedExpression()) ||
7504 isa<OMPArrayShapingExpr>(Next->getAssociatedExpression()) ||
7505 isa<UnaryOperator>(Next->getAssociatedExpression()) ||
7506 isa<BinaryOperator>(Next->getAssociatedExpression())) &&
7507 "Unexpected expression");
7508
7509 Address LB = Address::invalid();
7510 Address LowestElem = Address::invalid();
7511 auto &&EmitMemberExprBase = [](CodeGenFunction &CGF,
7512 const MemberExpr *E) {
7513 const Expr *BaseExpr = E->getBase();
7514 // If this is s.x, emit s as an lvalue. If it is s->x, emit s as a
7515 // scalar.
7516 LValue BaseLV;
7517 if (E->isArrow()) {
7518 LValueBaseInfo BaseInfo;
7519 TBAAAccessInfo TBAAInfo;
7520 Address Addr =
7521 CGF.EmitPointerWithAlignment(BaseExpr, &BaseInfo, &TBAAInfo);
7522 QualType PtrTy = BaseExpr->getType()->getPointeeType();
7523 BaseLV = CGF.MakeAddrLValue(Addr, PtrTy, BaseInfo, TBAAInfo);
7524 } else {
7525 BaseLV = CGF.EmitOMPSharedLValue(BaseExpr);
7526 }
7527 return BaseLV;
7528 };
7529 if (OAShE) {
7530 LowestElem = LB =
7531 Address(CGF.EmitScalarExpr(OAShE->getBase()),
7532 CGF.ConvertTypeForMem(
7533 OAShE->getBase()->getType()->getPointeeType()),
7534 CGF.getContext().getTypeAlignInChars(
7535 OAShE->getBase()->getType()));
7536 } else if (IsMemberReference) {
7537 const auto *ME = cast<MemberExpr>(I->getAssociatedExpression());
7538 LValue BaseLVal = EmitMemberExprBase(CGF, ME);
7539 LowestElem = CGF.EmitLValueForFieldInitialization(
7540 BaseLVal, cast<FieldDecl>(MapDecl))
7541 .getAddress(CGF);
7542 LB = CGF.EmitLoadOfReferenceLValue(LowestElem, MapDecl->getType())
7543 .getAddress(CGF);
7544 } else {
7545 LowestElem = LB =
7546 CGF.EmitOMPSharedLValue(I->getAssociatedExpression())
7547 .getAddress(CGF);
7548 }
7549
7550 // If this component is a pointer inside the base struct then we don't
7551 // need to create any entry for it - it will be combined with the object
7552 // it is pointing to into a single PTR_AND_OBJ entry.
7553 bool IsMemberPointerOrAddr =
7554 EncounteredME &&
7555 (((IsPointer || ForDeviceAddr) &&
7556 I->getAssociatedExpression() == EncounteredME) ||
7557 (IsPrevMemberReference && !IsPointer) ||
7558 (IsMemberReference && Next != CE &&
7559 !Next->getAssociatedExpression()->getType()->isPointerType()));
7560 if (!OverlappedElements.empty() && Next == CE) {
7561 // Handle base element with the info for overlapped elements.
7562 assert(!PartialStruct.Base.isValid() && "The base element is set.");
7563 assert(!IsPointer &&
7564 "Unexpected base element with the pointer type.");
7565 // Mark the whole struct as the struct that requires allocation on the
7566 // device.
7567 PartialStruct.LowestElem = {0, LowestElem};
7568 CharUnits TypeSize = CGF.getContext().getTypeSizeInChars(
7569 I->getAssociatedExpression()->getType());
7570 Address HB = CGF.Builder.CreateConstGEP(
7571 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
7572 LowestElem, CGF.VoidPtrTy, CGF.Int8Ty),
7573 TypeSize.getQuantity() - 1);
7574 PartialStruct.HighestElem = {
7575 std::numeric_limits<decltype(
7576 PartialStruct.HighestElem.first)>::max(),
7577 HB};
7578 PartialStruct.Base = BP;
7579 PartialStruct.LB = LB;
7580 assert(
7581 PartialStruct.PreliminaryMapData.BasePointers.empty() &&
7582 "Overlapped elements must be used only once for the variable.");
7583 std::swap(PartialStruct.PreliminaryMapData, CombinedInfo);
7584 // Emit data for non-overlapped data.
7585 OpenMPOffloadMappingFlags Flags =
7586 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF |
7587 getMapTypeBits(MapType, MapModifiers, MotionModifiers, IsImplicit,
7588 /*AddPtrFlag=*/false,
7589 /*AddIsTargetParamFlag=*/false, IsNonContiguous);
7590 llvm::Value *Size = nullptr;
7591 // Do bitcopy of all non-overlapped structure elements.
7592 for (OMPClauseMappableExprCommon::MappableExprComponentListRef
7593 Component : OverlappedElements) {
7594 Address ComponentLB = Address::invalid();
7595 for (const OMPClauseMappableExprCommon::MappableComponent &MC :
7596 Component) {
7597 if (const ValueDecl *VD = MC.getAssociatedDeclaration()) {
7598 const auto *FD = dyn_cast<FieldDecl>(VD);
7599 if (FD && FD->getType()->isLValueReferenceType()) {
7600 const auto *ME =
7601 cast<MemberExpr>(MC.getAssociatedExpression());
7602 LValue BaseLVal = EmitMemberExprBase(CGF, ME);
7603 ComponentLB =
7604 CGF.EmitLValueForFieldInitialization(BaseLVal, FD)
7605 .getAddress(CGF);
7606 } else {
7607 ComponentLB =
7608 CGF.EmitOMPSharedLValue(MC.getAssociatedExpression())
7609 .getAddress(CGF);
7610 }
7611 Size = CGF.Builder.CreatePtrDiff(
7612 CGF.Int8Ty, CGF.EmitCastToVoidPtr(ComponentLB.getPointer()),
7613 CGF.EmitCastToVoidPtr(LB.getPointer()));
7614 break;
7615 }
7616 }
7617 assert(Size && "Failed to determine structure size");
7618 CombinedInfo.Exprs.emplace_back(MapDecl, MapExpr);
7619 CombinedInfo.BasePointers.push_back(BP.getPointer());
7620 CombinedInfo.Pointers.push_back(LB.getPointer());
7621 CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast(
7622 Size, CGF.Int64Ty, /*isSigned=*/true));
7623 CombinedInfo.Types.push_back(Flags);
7624 CombinedInfo.Mappers.push_back(nullptr);
7625 CombinedInfo.NonContigInfo.Dims.push_back(IsNonContiguous ? DimSize
7626 : 1);
7627 LB = CGF.Builder.CreateConstGEP(ComponentLB, 1);
7628 }
7629 CombinedInfo.Exprs.emplace_back(MapDecl, MapExpr);
7630 CombinedInfo.BasePointers.push_back(BP.getPointer());
7631 CombinedInfo.Pointers.push_back(LB.getPointer());
7632 Size = CGF.Builder.CreatePtrDiff(
7633 CGF.Int8Ty, CGF.Builder.CreateConstGEP(HB, 1).getPointer(),
7634 CGF.EmitCastToVoidPtr(LB.getPointer()));
7635 CombinedInfo.Sizes.push_back(
7636 CGF.Builder.CreateIntCast(Size, CGF.Int64Ty, /*isSigned=*/true));
7637 CombinedInfo.Types.push_back(Flags);
7638 CombinedInfo.Mappers.push_back(nullptr);
7639 CombinedInfo.NonContigInfo.Dims.push_back(IsNonContiguous ? DimSize
7640 : 1);
7641 break;
7642 }
7643 llvm::Value *Size = getExprTypeSize(I->getAssociatedExpression());
7644 if (!IsMemberPointerOrAddr ||
7645 (Next == CE && MapType != OMPC_MAP_unknown)) {
7646 CombinedInfo.Exprs.emplace_back(MapDecl, MapExpr);
7647 CombinedInfo.BasePointers.push_back(BP.getPointer());
7648 CombinedInfo.Pointers.push_back(LB.getPointer());
7649 CombinedInfo.Sizes.push_back(
7650 CGF.Builder.CreateIntCast(Size, CGF.Int64Ty, /*isSigned=*/true));
7651 CombinedInfo.NonContigInfo.Dims.push_back(IsNonContiguous ? DimSize
7652 : 1);
7653
7654 // If Mapper is valid, the last component inherits the mapper.
7655 bool HasMapper = Mapper && Next == CE;
7656 CombinedInfo.Mappers.push_back(HasMapper ? Mapper : nullptr);
7657
7658 // We need to add a pointer flag for each map that comes from the
7659 // same expression except for the first one. We also need to signal
7660 // this map is the first one that relates with the current capture
7661 // (there is a set of entries for each capture).
7662 OpenMPOffloadMappingFlags Flags = getMapTypeBits(
7663 MapType, MapModifiers, MotionModifiers, IsImplicit,
7664 !IsExpressionFirstInfo || RequiresReference ||
7665 FirstPointerInComplexData || IsMemberReference,
7666 IsCaptureFirstInfo && !RequiresReference, IsNonContiguous);
7667
7668 if (!IsExpressionFirstInfo || IsMemberReference) {
7669 // If we have a PTR_AND_OBJ pair where the OBJ is a pointer as well,
7670 // then we reset the TO/FROM/ALWAYS/DELETE/CLOSE flags.
7671 if (IsPointer || (IsMemberReference && Next != CE))
7672 Flags &= ~(OpenMPOffloadMappingFlags::OMP_MAP_TO |
7673 OpenMPOffloadMappingFlags::OMP_MAP_FROM |
7674 OpenMPOffloadMappingFlags::OMP_MAP_ALWAYS |
7675 OpenMPOffloadMappingFlags::OMP_MAP_DELETE |
7676 OpenMPOffloadMappingFlags::OMP_MAP_CLOSE);
7677
7678 if (ShouldBeMemberOf) {
7679 // Set placeholder value MEMBER_OF=FFFF to indicate that the flag
7680 // should be later updated with the correct value of MEMBER_OF.
7681 Flags |= OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF;
7682 // From now on, all subsequent PTR_AND_OBJ entries should not be
7683 // marked as MEMBER_OF.
7684 ShouldBeMemberOf = false;
7685 }
7686 }
7687
7688 CombinedInfo.Types.push_back(Flags);
7689 }
7690
7691 // If we have encountered a member expression so far, keep track of the
7692 // mapped member. If the parent is "*this", then the value declaration
7693 // is nullptr.
7694 if (EncounteredME) {
7695 const auto *FD = cast<FieldDecl>(EncounteredME->getMemberDecl());
7696 unsigned FieldIndex = FD->getFieldIndex();
7697
7698 // Update info about the lowest and highest elements for this struct
7699 if (!PartialStruct.Base.isValid()) {
7700 PartialStruct.LowestElem = {FieldIndex, LowestElem};
7701 if (IsFinalArraySection) {
7702 Address HB =
7703 CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false)
7704 .getAddress(CGF);
7705 PartialStruct.HighestElem = {FieldIndex, HB};
7706 } else {
7707 PartialStruct.HighestElem = {FieldIndex, LowestElem};
7708 }
7709 PartialStruct.Base = BP;
7710 PartialStruct.LB = BP;
7711 } else if (FieldIndex < PartialStruct.LowestElem.first) {
7712 PartialStruct.LowestElem = {FieldIndex, LowestElem};
7713 } else if (FieldIndex > PartialStruct.HighestElem.first) {
7714 PartialStruct.HighestElem = {FieldIndex, LowestElem};
7715 }
7716 }
7717
7718 // Need to emit combined struct for array sections.
7719 if (IsFinalArraySection || IsNonContiguous)
7720 PartialStruct.IsArraySection = true;
7721
7722 // If we have a final array section, we are done with this expression.
7723 if (IsFinalArraySection)
7724 break;
7725
7726 // The pointer becomes the base for the next element.
7727 if (Next != CE)
7728 BP = IsMemberReference ? LowestElem : LB;
7729
7730 IsExpressionFirstInfo = false;
7731 IsCaptureFirstInfo = false;
7732 FirstPointerInComplexData = false;
7733 IsPrevMemberReference = IsMemberReference;
7734 } else if (FirstPointerInComplexData) {
7735 QualType Ty = Components.rbegin()
7736 ->getAssociatedDeclaration()
7737 ->getType()
7738 .getNonReferenceType();
7739 BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>());
7740 FirstPointerInComplexData = false;
7741 }
7742 }
7743 // If ran into the whole component - allocate the space for the whole
7744 // record.
7745 if (!EncounteredME)
7746 PartialStruct.HasCompleteRecord = true;
7747
7748 if (!IsNonContiguous)
7749 return;
7750
7751 const ASTContext &Context = CGF.getContext();
7752
7753 // For supporting stride in array section, we need to initialize the first
7754 // dimension size as 1, first offset as 0, and first count as 1
7755 MapValuesArrayTy CurOffsets = {llvm::ConstantInt::get(CGF.CGM.Int64Ty, 0)};
7756 MapValuesArrayTy CurCounts = {llvm::ConstantInt::get(CGF.CGM.Int64Ty, 1)};
7757 MapValuesArrayTy CurStrides;
7758 MapValuesArrayTy DimSizes{llvm::ConstantInt::get(CGF.CGM.Int64Ty, 1)};
7759 uint64_t ElementTypeSize;
7760
7761 // Collect Size information for each dimension and get the element size as
7762 // the first Stride. For example, for `int arr[10][10]`, the DimSizes
7763 // should be [10, 10] and the first stride is 4 btyes.
7764 for (const OMPClauseMappableExprCommon::MappableComponent &Component :
7765 Components) {
7766 const Expr *AssocExpr = Component.getAssociatedExpression();
7767 const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr);
7768
7769 if (!OASE)
7770 continue;
7771
7772 QualType Ty = OMPArraySectionExpr::getBaseOriginalType(OASE->getBase());
7773 auto *CAT = Context.getAsConstantArrayType(Ty);
7774 auto *VAT = Context.getAsVariableArrayType(Ty);
7775
7776 // We need all the dimension size except for the last dimension.
7777 assert((VAT || CAT || &Component == &*Components.begin()) &&
7778 "Should be either ConstantArray or VariableArray if not the "
7779 "first Component");
7780
7781 // Get element size if CurStrides is empty.
7782 if (CurStrides.empty()) {
7783 const Type *ElementType = nullptr;
7784 if (CAT)
7785 ElementType = CAT->getElementType().getTypePtr();
7786 else if (VAT)
7787 ElementType = VAT->getElementType().getTypePtr();
7788 else
7789 assert(&Component == &*Components.begin() &&
7790 "Only expect pointer (non CAT or VAT) when this is the "
7791 "first Component");
7792 // If ElementType is null, then it means the base is a pointer
7793 // (neither CAT nor VAT) and we'll attempt to get ElementType again
7794 // for next iteration.
7795 if (ElementType) {
7796 // For the case that having pointer as base, we need to remove one
7797 // level of indirection.
7798 if (&Component != &*Components.begin())
7799 ElementType = ElementType->getPointeeOrArrayElementType();
7800 ElementTypeSize =
7801 Context.getTypeSizeInChars(ElementType).getQuantity();
7802 CurStrides.push_back(
7803 llvm::ConstantInt::get(CGF.Int64Ty, ElementTypeSize));
7804 }
7805 }
7806 // Get dimension value except for the last dimension since we don't need
7807 // it.
7808 if (DimSizes.size() < Components.size() - 1) {
7809 if (CAT)
7810 DimSizes.push_back(llvm::ConstantInt::get(
7811 CGF.Int64Ty, CAT->getSize().getZExtValue()));
7812 else if (VAT)
7813 DimSizes.push_back(CGF.Builder.CreateIntCast(
7814 CGF.EmitScalarExpr(VAT->getSizeExpr()), CGF.Int64Ty,
7815 /*IsSigned=*/false));
7816 }
7817 }
7818
7819 // Skip the dummy dimension since we have already have its information.
7820 auto *DI = DimSizes.begin() + 1;
7821 // Product of dimension.
7822 llvm::Value *DimProd =
7823 llvm::ConstantInt::get(CGF.CGM.Int64Ty, ElementTypeSize);
7824
7825 // Collect info for non-contiguous. Notice that offset, count, and stride
7826 // are only meaningful for array-section, so we insert a null for anything
7827 // other than array-section.
7828 // Also, the size of offset, count, and stride are not the same as
7829 // pointers, base_pointers, sizes, or dims. Instead, the size of offset,
7830 // count, and stride are the same as the number of non-contiguous
7831 // declaration in target update to/from clause.
7832 for (const OMPClauseMappableExprCommon::MappableComponent &Component :
7833 Components) {
7834 const Expr *AssocExpr = Component.getAssociatedExpression();
7835
7836 if (const auto *AE = dyn_cast<ArraySubscriptExpr>(AssocExpr)) {
7837 llvm::Value *Offset = CGF.Builder.CreateIntCast(
7838 CGF.EmitScalarExpr(AE->getIdx()), CGF.Int64Ty,
7839 /*isSigned=*/false);
7840 CurOffsets.push_back(Offset);
7841 CurCounts.push_back(llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/1));
7842 CurStrides.push_back(CurStrides.back());
7843 continue;
7844 }
7845
7846 const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr);
7847
7848 if (!OASE)
7849 continue;
7850
7851 // Offset
7852 const Expr *OffsetExpr = OASE->getLowerBound();
7853 llvm::Value *Offset = nullptr;
7854 if (!OffsetExpr) {
7855 // If offset is absent, then we just set it to zero.
7856 Offset = llvm::ConstantInt::get(CGF.Int64Ty, 0);
7857 } else {
7858 Offset = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(OffsetExpr),
7859 CGF.Int64Ty,
7860 /*isSigned=*/false);
7861 }
7862 CurOffsets.push_back(Offset);
7863
7864 // Count
7865 const Expr *CountExpr = OASE->getLength();
7866 llvm::Value *Count = nullptr;
7867 if (!CountExpr) {
7868 // In Clang, once a high dimension is an array section, we construct all
7869 // the lower dimension as array section, however, for case like
7870 // arr[0:2][2], Clang construct the inner dimension as an array section
7871 // but it actually is not in an array section form according to spec.
7872 if (!OASE->getColonLocFirst().isValid() &&
7873 !OASE->getColonLocSecond().isValid()) {
7874 Count = llvm::ConstantInt::get(CGF.Int64Ty, 1);
7875 } else {
7876 // OpenMP 5.0, 2.1.5 Array Sections, Description.
7877 // When the length is absent it defaults to ⌈(size −
7878 // lower-bound)/stride⌉, where size is the size of the array
7879 // dimension.
7880 const Expr *StrideExpr = OASE->getStride();
7881 llvm::Value *Stride =
7882 StrideExpr
7883 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(StrideExpr),
7884 CGF.Int64Ty, /*isSigned=*/false)
7885 : nullptr;
7886 if (Stride)
7887 Count = CGF.Builder.CreateUDiv(
7888 CGF.Builder.CreateNUWSub(*DI, Offset), Stride);
7889 else
7890 Count = CGF.Builder.CreateNUWSub(*DI, Offset);
7891 }
7892 } else {
7893 Count = CGF.EmitScalarExpr(CountExpr);
7894 }
7895 Count = CGF.Builder.CreateIntCast(Count, CGF.Int64Ty, /*isSigned=*/false);
7896 CurCounts.push_back(Count);
7897
7898 // Stride_n' = Stride_n * (D_0 * D_1 ... * D_n-1) * Unit size
7899 // Take `int arr[5][5][5]` and `arr[0:2:2][1:2:1][0:2:2]` as an example:
7900 // Offset Count Stride
7901 // D0 0 1 4 (int) <- dummy dimension
7902 // D1 0 2 8 (2 * (1) * 4)
7903 // D2 1 2 20 (1 * (1 * 5) * 4)
7904 // D3 0 2 200 (2 * (1 * 5 * 4) * 4)
7905 const Expr *StrideExpr = OASE->getStride();
7906 llvm::Value *Stride =
7907 StrideExpr
7908 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(StrideExpr),
7909 CGF.Int64Ty, /*isSigned=*/false)
7910 : nullptr;
7911 DimProd = CGF.Builder.CreateNUWMul(DimProd, *(DI - 1));
7912 if (Stride)
7913 CurStrides.push_back(CGF.Builder.CreateNUWMul(DimProd, Stride));
7914 else
7915 CurStrides.push_back(DimProd);
7916 if (DI != DimSizes.end())
7917 ++DI;
7918 }
7919
7920 CombinedInfo.NonContigInfo.Offsets.push_back(CurOffsets);
7921 CombinedInfo.NonContigInfo.Counts.push_back(CurCounts);
7922 CombinedInfo.NonContigInfo.Strides.push_back(CurStrides);
7923 }
7924
7925 /// Return the adjusted map modifiers if the declaration a capture refers to
7926 /// appears in a first-private clause. This is expected to be used only with
7927 /// directives that start with 'target'.
7928 OpenMPOffloadMappingFlags
getMapModifiersForPrivateClauses(const CapturedStmt::Capture & Cap) const7929 getMapModifiersForPrivateClauses(const CapturedStmt::Capture &Cap) const {
7930 assert(Cap.capturesVariable() && "Expected capture by reference only!");
7931
7932 // A first private variable captured by reference will use only the
7933 // 'private ptr' and 'map to' flag. Return the right flags if the captured
7934 // declaration is known as first-private in this handler.
7935 if (FirstPrivateDecls.count(Cap.getCapturedVar())) {
7936 if (Cap.getCapturedVar()->getType()->isAnyPointerType())
7937 return OpenMPOffloadMappingFlags::OMP_MAP_TO |
7938 OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ;
7939 return OpenMPOffloadMappingFlags::OMP_MAP_PRIVATE |
7940 OpenMPOffloadMappingFlags::OMP_MAP_TO;
7941 }
7942 auto I = LambdasMap.find(Cap.getCapturedVar()->getCanonicalDecl());
7943 if (I != LambdasMap.end())
7944 // for map(to: lambda): using user specified map type.
7945 return getMapTypeBits(
7946 I->getSecond()->getMapType(), I->getSecond()->getMapTypeModifiers(),
7947 /*MotionModifiers=*/std::nullopt, I->getSecond()->isImplicit(),
7948 /*AddPtrFlag=*/false,
7949 /*AddIsTargetParamFlag=*/false,
7950 /*isNonContiguous=*/false);
7951 return OpenMPOffloadMappingFlags::OMP_MAP_TO |
7952 OpenMPOffloadMappingFlags::OMP_MAP_FROM;
7953 }
7954
getMemberOfFlag(unsigned Position)7955 static OpenMPOffloadMappingFlags getMemberOfFlag(unsigned Position) {
7956 // Rotate by getFlagMemberOffset() bits.
7957 return static_cast<OpenMPOffloadMappingFlags>(((uint64_t)Position + 1)
7958 << getFlagMemberOffset());
7959 }
7960
setCorrectMemberOfFlag(OpenMPOffloadMappingFlags & Flags,OpenMPOffloadMappingFlags MemberOfFlag)7961 static void setCorrectMemberOfFlag(OpenMPOffloadMappingFlags &Flags,
7962 OpenMPOffloadMappingFlags MemberOfFlag) {
7963 // If the entry is PTR_AND_OBJ but has not been marked with the special
7964 // placeholder value 0xFFFF in the MEMBER_OF field, then it should not be
7965 // marked as MEMBER_OF.
7966 if (static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
7967 Flags & OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ) &&
7968 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
7969 (Flags & OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF) !=
7970 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF))
7971 return;
7972
7973 // Reset the placeholder value to prepare the flag for the assignment of the
7974 // proper MEMBER_OF value.
7975 Flags &= ~OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF;
7976 Flags |= MemberOfFlag;
7977 }
7978
getPlainLayout(const CXXRecordDecl * RD,llvm::SmallVectorImpl<const FieldDecl * > & Layout,bool AsBase) const7979 void getPlainLayout(const CXXRecordDecl *RD,
7980 llvm::SmallVectorImpl<const FieldDecl *> &Layout,
7981 bool AsBase) const {
7982 const CGRecordLayout &RL = CGF.getTypes().getCGRecordLayout(RD);
7983
7984 llvm::StructType *St =
7985 AsBase ? RL.getBaseSubobjectLLVMType() : RL.getLLVMType();
7986
7987 unsigned NumElements = St->getNumElements();
7988 llvm::SmallVector<
7989 llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *>, 4>
7990 RecordLayout(NumElements);
7991
7992 // Fill bases.
7993 for (const auto &I : RD->bases()) {
7994 if (I.isVirtual())
7995 continue;
7996 const auto *Base = I.getType()->getAsCXXRecordDecl();
7997 // Ignore empty bases.
7998 if (Base->isEmpty() || CGF.getContext()
7999 .getASTRecordLayout(Base)
8000 .getNonVirtualSize()
8001 .isZero())
8002 continue;
8003
8004 unsigned FieldIndex = RL.getNonVirtualBaseLLVMFieldNo(Base);
8005 RecordLayout[FieldIndex] = Base;
8006 }
8007 // Fill in virtual bases.
8008 for (const auto &I : RD->vbases()) {
8009 const auto *Base = I.getType()->getAsCXXRecordDecl();
8010 // Ignore empty bases.
8011 if (Base->isEmpty())
8012 continue;
8013 unsigned FieldIndex = RL.getVirtualBaseIndex(Base);
8014 if (RecordLayout[FieldIndex])
8015 continue;
8016 RecordLayout[FieldIndex] = Base;
8017 }
8018 // Fill in all the fields.
8019 assert(!RD->isUnion() && "Unexpected union.");
8020 for (const auto *Field : RD->fields()) {
8021 // Fill in non-bitfields. (Bitfields always use a zero pattern, which we
8022 // will fill in later.)
8023 if (!Field->isBitField() && !Field->isZeroSize(CGF.getContext())) {
8024 unsigned FieldIndex = RL.getLLVMFieldNo(Field);
8025 RecordLayout[FieldIndex] = Field;
8026 }
8027 }
8028 for (const llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *>
8029 &Data : RecordLayout) {
8030 if (Data.isNull())
8031 continue;
8032 if (const auto *Base = Data.dyn_cast<const CXXRecordDecl *>())
8033 getPlainLayout(Base, Layout, /*AsBase=*/true);
8034 else
8035 Layout.push_back(Data.get<const FieldDecl *>());
8036 }
8037 }
8038
8039 /// Generate all the base pointers, section pointers, sizes, map types, and
8040 /// mappers for the extracted mappable expressions (all included in \a
8041 /// CombinedInfo). Also, for each item that relates with a device pointer, a
8042 /// pair of the relevant declaration and index where it occurs is appended to
8043 /// the device pointers info array.
generateAllInfoForClauses(ArrayRef<const OMPClause * > Clauses,MapCombinedInfoTy & CombinedInfo,const llvm::DenseSet<CanonicalDeclPtr<const Decl>> & SkipVarSet=llvm::DenseSet<CanonicalDeclPtr<const Decl>> ()) const8044 void generateAllInfoForClauses(
8045 ArrayRef<const OMPClause *> Clauses, MapCombinedInfoTy &CombinedInfo,
8046 const llvm::DenseSet<CanonicalDeclPtr<const Decl>> &SkipVarSet =
8047 llvm::DenseSet<CanonicalDeclPtr<const Decl>>()) const {
8048 // We have to process the component lists that relate with the same
8049 // declaration in a single chunk so that we can generate the map flags
8050 // correctly. Therefore, we organize all lists in a map.
8051 enum MapKind { Present, Allocs, Other, Total };
8052 llvm::MapVector<CanonicalDeclPtr<const Decl>,
8053 SmallVector<SmallVector<MapInfo, 8>, 4>>
8054 Info;
8055
8056 // Helper function to fill the information map for the different supported
8057 // clauses.
8058 auto &&InfoGen =
8059 [&Info, &SkipVarSet](
8060 const ValueDecl *D, MapKind Kind,
8061 OMPClauseMappableExprCommon::MappableExprComponentListRef L,
8062 OpenMPMapClauseKind MapType,
8063 ArrayRef<OpenMPMapModifierKind> MapModifiers,
8064 ArrayRef<OpenMPMotionModifierKind> MotionModifiers,
8065 bool ReturnDevicePointer, bool IsImplicit, const ValueDecl *Mapper,
8066 const Expr *VarRef = nullptr, bool ForDeviceAddr = false) {
8067 if (SkipVarSet.contains(D))
8068 return;
8069 auto It = Info.find(D);
8070 if (It == Info.end())
8071 It = Info
8072 .insert(std::make_pair(
8073 D, SmallVector<SmallVector<MapInfo, 8>, 4>(Total)))
8074 .first;
8075 It->second[Kind].emplace_back(
8076 L, MapType, MapModifiers, MotionModifiers, ReturnDevicePointer,
8077 IsImplicit, Mapper, VarRef, ForDeviceAddr);
8078 };
8079
8080 for (const auto *Cl : Clauses) {
8081 const auto *C = dyn_cast<OMPMapClause>(Cl);
8082 if (!C)
8083 continue;
8084 MapKind Kind = Other;
8085 if (llvm::is_contained(C->getMapTypeModifiers(),
8086 OMPC_MAP_MODIFIER_present))
8087 Kind = Present;
8088 else if (C->getMapType() == OMPC_MAP_alloc)
8089 Kind = Allocs;
8090 const auto *EI = C->getVarRefs().begin();
8091 for (const auto L : C->component_lists()) {
8092 const Expr *E = (C->getMapLoc().isValid()) ? *EI : nullptr;
8093 InfoGen(std::get<0>(L), Kind, std::get<1>(L), C->getMapType(),
8094 C->getMapTypeModifiers(), std::nullopt,
8095 /*ReturnDevicePointer=*/false, C->isImplicit(), std::get<2>(L),
8096 E);
8097 ++EI;
8098 }
8099 }
8100 for (const auto *Cl : Clauses) {
8101 const auto *C = dyn_cast<OMPToClause>(Cl);
8102 if (!C)
8103 continue;
8104 MapKind Kind = Other;
8105 if (llvm::is_contained(C->getMotionModifiers(),
8106 OMPC_MOTION_MODIFIER_present))
8107 Kind = Present;
8108 const auto *EI = C->getVarRefs().begin();
8109 for (const auto L : C->component_lists()) {
8110 InfoGen(std::get<0>(L), Kind, std::get<1>(L), OMPC_MAP_to, std::nullopt,
8111 C->getMotionModifiers(), /*ReturnDevicePointer=*/false,
8112 C->isImplicit(), std::get<2>(L), *EI);
8113 ++EI;
8114 }
8115 }
8116 for (const auto *Cl : Clauses) {
8117 const auto *C = dyn_cast<OMPFromClause>(Cl);
8118 if (!C)
8119 continue;
8120 MapKind Kind = Other;
8121 if (llvm::is_contained(C->getMotionModifiers(),
8122 OMPC_MOTION_MODIFIER_present))
8123 Kind = Present;
8124 const auto *EI = C->getVarRefs().begin();
8125 for (const auto L : C->component_lists()) {
8126 InfoGen(std::get<0>(L), Kind, std::get<1>(L), OMPC_MAP_from,
8127 std::nullopt, C->getMotionModifiers(),
8128 /*ReturnDevicePointer=*/false, C->isImplicit(), std::get<2>(L),
8129 *EI);
8130 ++EI;
8131 }
8132 }
8133
8134 // Look at the use_device_ptr and use_device_addr clauses information and
8135 // mark the existing map entries as such. If there is no map information for
8136 // an entry in the use_device_ptr and use_device_addr list, we create one
8137 // with map type 'alloc' and zero size section. It is the user fault if that
8138 // was not mapped before. If there is no map information and the pointer is
8139 // a struct member, then we defer the emission of that entry until the whole
8140 // struct has been processed.
8141 llvm::MapVector<CanonicalDeclPtr<const Decl>,
8142 SmallVector<DeferredDevicePtrEntryTy, 4>>
8143 DeferredInfo;
8144 MapCombinedInfoTy UseDeviceDataCombinedInfo;
8145
8146 auto &&UseDeviceDataCombinedInfoGen =
8147 [&UseDeviceDataCombinedInfo](const ValueDecl *VD, llvm::Value *Ptr,
8148 CodeGenFunction &CGF) {
8149 UseDeviceDataCombinedInfo.Exprs.push_back(VD);
8150 UseDeviceDataCombinedInfo.BasePointers.emplace_back(Ptr, VD);
8151 UseDeviceDataCombinedInfo.Pointers.push_back(Ptr);
8152 UseDeviceDataCombinedInfo.Sizes.push_back(
8153 llvm::Constant::getNullValue(CGF.Int64Ty));
8154 UseDeviceDataCombinedInfo.Types.push_back(
8155 OpenMPOffloadMappingFlags::OMP_MAP_RETURN_PARAM);
8156 UseDeviceDataCombinedInfo.Mappers.push_back(nullptr);
8157 };
8158
8159 auto &&MapInfoGen =
8160 [&DeferredInfo, &UseDeviceDataCombinedInfoGen,
8161 &InfoGen](CodeGenFunction &CGF, const Expr *IE, const ValueDecl *VD,
8162 OMPClauseMappableExprCommon::MappableExprComponentListRef
8163 Components,
8164 bool IsImplicit, bool IsDevAddr) {
8165 // We didn't find any match in our map information - generate a zero
8166 // size array section - if the pointer is a struct member we defer
8167 // this action until the whole struct has been processed.
8168 if (isa<MemberExpr>(IE)) {
8169 // Insert the pointer into Info to be processed by
8170 // generateInfoForComponentList. Because it is a member pointer
8171 // without a pointee, no entry will be generated for it, therefore
8172 // we need to generate one after the whole struct has been
8173 // processed. Nonetheless, generateInfoForComponentList must be
8174 // called to take the pointer into account for the calculation of
8175 // the range of the partial struct.
8176 InfoGen(nullptr, Other, Components, OMPC_MAP_unknown, std::nullopt,
8177 std::nullopt, /*ReturnDevicePointer=*/false, IsImplicit,
8178 nullptr, nullptr, IsDevAddr);
8179 DeferredInfo[nullptr].emplace_back(IE, VD, IsDevAddr);
8180 } else {
8181 llvm::Value *Ptr;
8182 if (IsDevAddr) {
8183 if (IE->isGLValue())
8184 Ptr = CGF.EmitLValue(IE).getPointer(CGF);
8185 else
8186 Ptr = CGF.EmitScalarExpr(IE);
8187 } else {
8188 Ptr = CGF.EmitLoadOfScalar(CGF.EmitLValue(IE), IE->getExprLoc());
8189 }
8190 UseDeviceDataCombinedInfoGen(VD, Ptr, CGF);
8191 }
8192 };
8193
8194 auto &&IsMapInfoExist = [&Info](CodeGenFunction &CGF, const ValueDecl *VD,
8195 const Expr *IE, bool IsDevAddr) -> bool {
8196 // We potentially have map information for this declaration already.
8197 // Look for the first set of components that refer to it. If found,
8198 // return true.
8199 // If the first component is a member expression, we have to look into
8200 // 'this', which maps to null in the map of map information. Otherwise
8201 // look directly for the information.
8202 auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD);
8203 if (It != Info.end()) {
8204 bool Found = false;
8205 for (auto &Data : It->second) {
8206 auto *CI = llvm::find_if(Data, [VD](const MapInfo &MI) {
8207 return MI.Components.back().getAssociatedDeclaration() == VD;
8208 });
8209 // If we found a map entry, signal that the pointer has to be
8210 // returned and move on to the next declaration. Exclude cases where
8211 // the base pointer is mapped as array subscript, array section or
8212 // array shaping. The base address is passed as a pointer to base in
8213 // this case and cannot be used as a base for use_device_ptr list
8214 // item.
8215 if (CI != Data.end()) {
8216 if (IsDevAddr) {
8217 CI->ReturnDevicePointer = true;
8218 Found = true;
8219 break;
8220 } else {
8221 auto PrevCI = std::next(CI->Components.rbegin());
8222 const auto *VarD = dyn_cast<VarDecl>(VD);
8223 if (CGF.CGM.getOpenMPRuntime().hasRequiresUnifiedSharedMemory() ||
8224 isa<MemberExpr>(IE) ||
8225 !VD->getType().getNonReferenceType()->isPointerType() ||
8226 PrevCI == CI->Components.rend() ||
8227 isa<MemberExpr>(PrevCI->getAssociatedExpression()) || !VarD ||
8228 VarD->hasLocalStorage()) {
8229 CI->ReturnDevicePointer = true;
8230 Found = true;
8231 break;
8232 }
8233 }
8234 }
8235 }
8236 return Found;
8237 }
8238 return false;
8239 };
8240
8241 // Look at the use_device_ptr clause information and mark the existing map
8242 // entries as such. If there is no map information for an entry in the
8243 // use_device_ptr list, we create one with map type 'alloc' and zero size
8244 // section. It is the user fault if that was not mapped before. If there is
8245 // no map information and the pointer is a struct member, then we defer the
8246 // emission of that entry until the whole struct has been processed.
8247 for (const auto *Cl : Clauses) {
8248 const auto *C = dyn_cast<OMPUseDevicePtrClause>(Cl);
8249 if (!C)
8250 continue;
8251 for (const auto L : C->component_lists()) {
8252 OMPClauseMappableExprCommon::MappableExprComponentListRef Components =
8253 std::get<1>(L);
8254 assert(!Components.empty() &&
8255 "Not expecting empty list of components!");
8256 const ValueDecl *VD = Components.back().getAssociatedDeclaration();
8257 VD = cast<ValueDecl>(VD->getCanonicalDecl());
8258 const Expr *IE = Components.back().getAssociatedExpression();
8259 if (IsMapInfoExist(CGF, VD, IE, /*IsDevAddr=*/false))
8260 continue;
8261 MapInfoGen(CGF, IE, VD, Components, C->isImplicit(),
8262 /*IsDevAddr=*/false);
8263 }
8264 }
8265
8266 llvm::SmallDenseSet<CanonicalDeclPtr<const Decl>, 4> Processed;
8267 for (const auto *Cl : Clauses) {
8268 const auto *C = dyn_cast<OMPUseDeviceAddrClause>(Cl);
8269 if (!C)
8270 continue;
8271 for (const auto L : C->component_lists()) {
8272 OMPClauseMappableExprCommon::MappableExprComponentListRef Components =
8273 std::get<1>(L);
8274 assert(!std::get<1>(L).empty() &&
8275 "Not expecting empty list of components!");
8276 const ValueDecl *VD = std::get<1>(L).back().getAssociatedDeclaration();
8277 if (!Processed.insert(VD).second)
8278 continue;
8279 VD = cast<ValueDecl>(VD->getCanonicalDecl());
8280 const Expr *IE = std::get<1>(L).back().getAssociatedExpression();
8281 if (IsMapInfoExist(CGF, VD, IE, /*IsDevAddr=*/true))
8282 continue;
8283 MapInfoGen(CGF, IE, VD, Components, C->isImplicit(),
8284 /*IsDevAddr=*/true);
8285 }
8286 }
8287
8288 for (const auto &Data : Info) {
8289 StructRangeInfoTy PartialStruct;
8290 // Temporary generated information.
8291 MapCombinedInfoTy CurInfo;
8292 const Decl *D = Data.first;
8293 const ValueDecl *VD = cast_or_null<ValueDecl>(D);
8294 for (const auto &M : Data.second) {
8295 for (const MapInfo &L : M) {
8296 assert(!L.Components.empty() &&
8297 "Not expecting declaration with no component lists.");
8298
8299 // Remember the current base pointer index.
8300 unsigned CurrentBasePointersIdx = CurInfo.BasePointers.size();
8301 CurInfo.NonContigInfo.IsNonContiguous =
8302 L.Components.back().isNonContiguous();
8303 generateInfoForComponentList(
8304 L.MapType, L.MapModifiers, L.MotionModifiers, L.Components,
8305 CurInfo, PartialStruct, /*IsFirstComponentList=*/false,
8306 L.IsImplicit, L.Mapper, L.ForDeviceAddr, VD, L.VarRef);
8307
8308 // If this entry relates with a device pointer, set the relevant
8309 // declaration and add the 'return pointer' flag.
8310 if (L.ReturnDevicePointer) {
8311 assert(CurInfo.BasePointers.size() > CurrentBasePointersIdx &&
8312 "Unexpected number of mapped base pointers.");
8313
8314 const ValueDecl *RelevantVD =
8315 L.Components.back().getAssociatedDeclaration();
8316 assert(RelevantVD &&
8317 "No relevant declaration related with device pointer??");
8318
8319 CurInfo.BasePointers[CurrentBasePointersIdx].setDevicePtrDecl(
8320 RelevantVD);
8321 CurInfo.Types[CurrentBasePointersIdx] |=
8322 OpenMPOffloadMappingFlags::OMP_MAP_RETURN_PARAM;
8323 }
8324 }
8325 }
8326
8327 // Append any pending zero-length pointers which are struct members and
8328 // used with use_device_ptr or use_device_addr.
8329 auto CI = DeferredInfo.find(Data.first);
8330 if (CI != DeferredInfo.end()) {
8331 for (const DeferredDevicePtrEntryTy &L : CI->second) {
8332 llvm::Value *BasePtr;
8333 llvm::Value *Ptr;
8334 if (L.ForDeviceAddr) {
8335 if (L.IE->isGLValue())
8336 Ptr = this->CGF.EmitLValue(L.IE).getPointer(CGF);
8337 else
8338 Ptr = this->CGF.EmitScalarExpr(L.IE);
8339 BasePtr = Ptr;
8340 // Entry is RETURN_PARAM. Also, set the placeholder value
8341 // MEMBER_OF=FFFF so that the entry is later updated with the
8342 // correct value of MEMBER_OF.
8343 CurInfo.Types.push_back(
8344 OpenMPOffloadMappingFlags::OMP_MAP_RETURN_PARAM |
8345 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF);
8346 } else {
8347 BasePtr = this->CGF.EmitLValue(L.IE).getPointer(CGF);
8348 Ptr = this->CGF.EmitLoadOfScalar(this->CGF.EmitLValue(L.IE),
8349 L.IE->getExprLoc());
8350 // Entry is PTR_AND_OBJ and RETURN_PARAM. Also, set the
8351 // placeholder value MEMBER_OF=FFFF so that the entry is later
8352 // updated with the correct value of MEMBER_OF.
8353 CurInfo.Types.push_back(
8354 OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ |
8355 OpenMPOffloadMappingFlags::OMP_MAP_RETURN_PARAM |
8356 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF);
8357 }
8358 CurInfo.Exprs.push_back(L.VD);
8359 CurInfo.BasePointers.emplace_back(BasePtr, L.VD);
8360 CurInfo.Pointers.push_back(Ptr);
8361 CurInfo.Sizes.push_back(
8362 llvm::Constant::getNullValue(this->CGF.Int64Ty));
8363 CurInfo.Mappers.push_back(nullptr);
8364 }
8365 }
8366 // If there is an entry in PartialStruct it means we have a struct with
8367 // individual members mapped. Emit an extra combined entry.
8368 if (PartialStruct.Base.isValid()) {
8369 CurInfo.NonContigInfo.Dims.push_back(0);
8370 emitCombinedEntry(CombinedInfo, CurInfo.Types, PartialStruct, VD);
8371 }
8372
8373 // We need to append the results of this capture to what we already
8374 // have.
8375 CombinedInfo.append(CurInfo);
8376 }
8377 // Append data for use_device_ptr clauses.
8378 CombinedInfo.append(UseDeviceDataCombinedInfo);
8379 }
8380
8381 public:
MappableExprsHandler(const OMPExecutableDirective & Dir,CodeGenFunction & CGF)8382 MappableExprsHandler(const OMPExecutableDirective &Dir, CodeGenFunction &CGF)
8383 : CurDir(&Dir), CGF(CGF) {
8384 // Extract firstprivate clause information.
8385 for (const auto *C : Dir.getClausesOfKind<OMPFirstprivateClause>())
8386 for (const auto *D : C->varlists())
8387 FirstPrivateDecls.try_emplace(
8388 cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl()), C->isImplicit());
8389 // Extract implicit firstprivates from uses_allocators clauses.
8390 for (const auto *C : Dir.getClausesOfKind<OMPUsesAllocatorsClause>()) {
8391 for (unsigned I = 0, E = C->getNumberOfAllocators(); I < E; ++I) {
8392 OMPUsesAllocatorsClause::Data D = C->getAllocatorData(I);
8393 if (const auto *DRE = dyn_cast_or_null<DeclRefExpr>(D.AllocatorTraits))
8394 FirstPrivateDecls.try_emplace(cast<VarDecl>(DRE->getDecl()),
8395 /*Implicit=*/true);
8396 else if (const auto *VD = dyn_cast<VarDecl>(
8397 cast<DeclRefExpr>(D.Allocator->IgnoreParenImpCasts())
8398 ->getDecl()))
8399 FirstPrivateDecls.try_emplace(VD, /*Implicit=*/true);
8400 }
8401 }
8402 // Extract device pointer clause information.
8403 for (const auto *C : Dir.getClausesOfKind<OMPIsDevicePtrClause>())
8404 for (auto L : C->component_lists())
8405 DevPointersMap[std::get<0>(L)].push_back(std::get<1>(L));
8406 // Extract device addr clause information.
8407 for (const auto *C : Dir.getClausesOfKind<OMPHasDeviceAddrClause>())
8408 for (auto L : C->component_lists())
8409 HasDevAddrsMap[std::get<0>(L)].push_back(std::get<1>(L));
8410 // Extract map information.
8411 for (const auto *C : Dir.getClausesOfKind<OMPMapClause>()) {
8412 if (C->getMapType() != OMPC_MAP_to)
8413 continue;
8414 for (auto L : C->component_lists()) {
8415 const ValueDecl *VD = std::get<0>(L);
8416 const auto *RD = VD ? VD->getType()
8417 .getCanonicalType()
8418 .getNonReferenceType()
8419 ->getAsCXXRecordDecl()
8420 : nullptr;
8421 if (RD && RD->isLambda())
8422 LambdasMap.try_emplace(std::get<0>(L), C);
8423 }
8424 }
8425 }
8426
8427 /// Constructor for the declare mapper directive.
MappableExprsHandler(const OMPDeclareMapperDecl & Dir,CodeGenFunction & CGF)8428 MappableExprsHandler(const OMPDeclareMapperDecl &Dir, CodeGenFunction &CGF)
8429 : CurDir(&Dir), CGF(CGF) {}
8430
8431 /// Generate code for the combined entry if we have a partially mapped struct
8432 /// and take care of the mapping flags of the arguments corresponding to
8433 /// individual struct members.
emitCombinedEntry(MapCombinedInfoTy & CombinedInfo,MapFlagsArrayTy & CurTypes,const StructRangeInfoTy & PartialStruct,const ValueDecl * VD=nullptr,bool NotTargetParams=true) const8434 void emitCombinedEntry(MapCombinedInfoTy &CombinedInfo,
8435 MapFlagsArrayTy &CurTypes,
8436 const StructRangeInfoTy &PartialStruct,
8437 const ValueDecl *VD = nullptr,
8438 bool NotTargetParams = true) const {
8439 if (CurTypes.size() == 1 &&
8440 ((CurTypes.back() & OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF) !=
8441 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF) &&
8442 !PartialStruct.IsArraySection)
8443 return;
8444 Address LBAddr = PartialStruct.LowestElem.second;
8445 Address HBAddr = PartialStruct.HighestElem.second;
8446 if (PartialStruct.HasCompleteRecord) {
8447 LBAddr = PartialStruct.LB;
8448 HBAddr = PartialStruct.LB;
8449 }
8450 CombinedInfo.Exprs.push_back(VD);
8451 // Base is the base of the struct
8452 CombinedInfo.BasePointers.push_back(PartialStruct.Base.getPointer());
8453 // Pointer is the address of the lowest element
8454 llvm::Value *LB = LBAddr.getPointer();
8455 const CXXMethodDecl *MD =
8456 CGF.CurFuncDecl ? dyn_cast<CXXMethodDecl>(CGF.CurFuncDecl) : nullptr;
8457 const CXXRecordDecl *RD = MD ? MD->getParent() : nullptr;
8458 bool HasBaseClass = RD ? RD->getNumBases() > 0 : false;
8459 // There should not be a mapper for a combined entry.
8460 if (HasBaseClass) {
8461 // OpenMP 5.2 148:21:
8462 // If the target construct is within a class non-static member function,
8463 // and a variable is an accessible data member of the object for which the
8464 // non-static data member function is invoked, the variable is treated as
8465 // if the this[:1] expression had appeared in a map clause with a map-type
8466 // of tofrom.
8467 // Emit this[:1]
8468 CombinedInfo.Pointers.push_back(PartialStruct.Base.getPointer());
8469 QualType Ty = MD->getThisType()->getPointeeType();
8470 llvm::Value *Size =
8471 CGF.Builder.CreateIntCast(CGF.getTypeSize(Ty), CGF.Int64Ty,
8472 /*isSigned=*/true);
8473 CombinedInfo.Sizes.push_back(Size);
8474 } else {
8475 CombinedInfo.Pointers.push_back(LB);
8476 // Size is (addr of {highest+1} element) - (addr of lowest element)
8477 llvm::Value *HB = HBAddr.getPointer();
8478 llvm::Value *HAddr = CGF.Builder.CreateConstGEP1_32(
8479 HBAddr.getElementType(), HB, /*Idx0=*/1);
8480 llvm::Value *CLAddr = CGF.Builder.CreatePointerCast(LB, CGF.VoidPtrTy);
8481 llvm::Value *CHAddr = CGF.Builder.CreatePointerCast(HAddr, CGF.VoidPtrTy);
8482 llvm::Value *Diff = CGF.Builder.CreatePtrDiff(CGF.Int8Ty, CHAddr, CLAddr);
8483 llvm::Value *Size = CGF.Builder.CreateIntCast(Diff, CGF.Int64Ty,
8484 /*isSigned=*/false);
8485 CombinedInfo.Sizes.push_back(Size);
8486 }
8487 CombinedInfo.Mappers.push_back(nullptr);
8488 // Map type is always TARGET_PARAM, if generate info for captures.
8489 CombinedInfo.Types.push_back(
8490 NotTargetParams ? OpenMPOffloadMappingFlags::OMP_MAP_NONE
8491 : OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM);
8492 // If any element has the present modifier, then make sure the runtime
8493 // doesn't attempt to allocate the struct.
8494 if (CurTypes.end() !=
8495 llvm::find_if(CurTypes, [](OpenMPOffloadMappingFlags Type) {
8496 return static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
8497 Type & OpenMPOffloadMappingFlags::OMP_MAP_PRESENT);
8498 }))
8499 CombinedInfo.Types.back() |= OpenMPOffloadMappingFlags::OMP_MAP_PRESENT;
8500 // Remove TARGET_PARAM flag from the first element
8501 (*CurTypes.begin()) &= ~OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
8502 // If any element has the ompx_hold modifier, then make sure the runtime
8503 // uses the hold reference count for the struct as a whole so that it won't
8504 // be unmapped by an extra dynamic reference count decrement. Add it to all
8505 // elements as well so the runtime knows which reference count to check
8506 // when determining whether it's time for device-to-host transfers of
8507 // individual elements.
8508 if (CurTypes.end() !=
8509 llvm::find_if(CurTypes, [](OpenMPOffloadMappingFlags Type) {
8510 return static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
8511 Type & OpenMPOffloadMappingFlags::OMP_MAP_OMPX_HOLD);
8512 })) {
8513 CombinedInfo.Types.back() |= OpenMPOffloadMappingFlags::OMP_MAP_OMPX_HOLD;
8514 for (auto &M : CurTypes)
8515 M |= OpenMPOffloadMappingFlags::OMP_MAP_OMPX_HOLD;
8516 }
8517
8518 // All other current entries will be MEMBER_OF the combined entry
8519 // (except for PTR_AND_OBJ entries which do not have a placeholder value
8520 // 0xFFFF in the MEMBER_OF field).
8521 OpenMPOffloadMappingFlags MemberOfFlag =
8522 getMemberOfFlag(CombinedInfo.BasePointers.size() - 1);
8523 for (auto &M : CurTypes)
8524 setCorrectMemberOfFlag(M, MemberOfFlag);
8525 }
8526
8527 /// Generate all the base pointers, section pointers, sizes, map types, and
8528 /// mappers for the extracted mappable expressions (all included in \a
8529 /// CombinedInfo). Also, for each item that relates with a device pointer, a
8530 /// pair of the relevant declaration and index where it occurs is appended to
8531 /// the device pointers info array.
generateAllInfo(MapCombinedInfoTy & CombinedInfo,const llvm::DenseSet<CanonicalDeclPtr<const Decl>> & SkipVarSet=llvm::DenseSet<CanonicalDeclPtr<const Decl>> ()) const8532 void generateAllInfo(
8533 MapCombinedInfoTy &CombinedInfo,
8534 const llvm::DenseSet<CanonicalDeclPtr<const Decl>> &SkipVarSet =
8535 llvm::DenseSet<CanonicalDeclPtr<const Decl>>()) const {
8536 assert(CurDir.is<const OMPExecutableDirective *>() &&
8537 "Expect a executable directive");
8538 const auto *CurExecDir = CurDir.get<const OMPExecutableDirective *>();
8539 generateAllInfoForClauses(CurExecDir->clauses(), CombinedInfo, SkipVarSet);
8540 }
8541
8542 /// Generate all the base pointers, section pointers, sizes, map types, and
8543 /// mappers for the extracted map clauses of user-defined mapper (all included
8544 /// in \a CombinedInfo).
generateAllInfoForMapper(MapCombinedInfoTy & CombinedInfo) const8545 void generateAllInfoForMapper(MapCombinedInfoTy &CombinedInfo) const {
8546 assert(CurDir.is<const OMPDeclareMapperDecl *>() &&
8547 "Expect a declare mapper directive");
8548 const auto *CurMapperDir = CurDir.get<const OMPDeclareMapperDecl *>();
8549 generateAllInfoForClauses(CurMapperDir->clauses(), CombinedInfo);
8550 }
8551
8552 /// Emit capture info for lambdas for variables captured by reference.
generateInfoForLambdaCaptures(const ValueDecl * VD,llvm::Value * Arg,MapCombinedInfoTy & CombinedInfo,llvm::DenseMap<llvm::Value *,llvm::Value * > & LambdaPointers) const8553 void generateInfoForLambdaCaptures(
8554 const ValueDecl *VD, llvm::Value *Arg, MapCombinedInfoTy &CombinedInfo,
8555 llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers) const {
8556 QualType VDType = VD->getType().getCanonicalType().getNonReferenceType();
8557 const auto *RD = VDType->getAsCXXRecordDecl();
8558 if (!RD || !RD->isLambda())
8559 return;
8560 Address VDAddr(Arg, CGF.ConvertTypeForMem(VDType),
8561 CGF.getContext().getDeclAlign(VD));
8562 LValue VDLVal = CGF.MakeAddrLValue(VDAddr, VDType);
8563 llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures;
8564 FieldDecl *ThisCapture = nullptr;
8565 RD->getCaptureFields(Captures, ThisCapture);
8566 if (ThisCapture) {
8567 LValue ThisLVal =
8568 CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
8569 LValue ThisLValVal = CGF.EmitLValueForField(VDLVal, ThisCapture);
8570 LambdaPointers.try_emplace(ThisLVal.getPointer(CGF),
8571 VDLVal.getPointer(CGF));
8572 CombinedInfo.Exprs.push_back(VD);
8573 CombinedInfo.BasePointers.push_back(ThisLVal.getPointer(CGF));
8574 CombinedInfo.Pointers.push_back(ThisLValVal.getPointer(CGF));
8575 CombinedInfo.Sizes.push_back(
8576 CGF.Builder.CreateIntCast(CGF.getTypeSize(CGF.getContext().VoidPtrTy),
8577 CGF.Int64Ty, /*isSigned=*/true));
8578 CombinedInfo.Types.push_back(
8579 OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ |
8580 OpenMPOffloadMappingFlags::OMP_MAP_LITERAL |
8581 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF |
8582 OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT);
8583 CombinedInfo.Mappers.push_back(nullptr);
8584 }
8585 for (const LambdaCapture &LC : RD->captures()) {
8586 if (!LC.capturesVariable())
8587 continue;
8588 const VarDecl *VD = cast<VarDecl>(LC.getCapturedVar());
8589 if (LC.getCaptureKind() != LCK_ByRef && !VD->getType()->isPointerType())
8590 continue;
8591 auto It = Captures.find(VD);
8592 assert(It != Captures.end() && "Found lambda capture without field.");
8593 LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
8594 if (LC.getCaptureKind() == LCK_ByRef) {
8595 LValue VarLValVal = CGF.EmitLValueForField(VDLVal, It->second);
8596 LambdaPointers.try_emplace(VarLVal.getPointer(CGF),
8597 VDLVal.getPointer(CGF));
8598 CombinedInfo.Exprs.push_back(VD);
8599 CombinedInfo.BasePointers.push_back(VarLVal.getPointer(CGF));
8600 CombinedInfo.Pointers.push_back(VarLValVal.getPointer(CGF));
8601 CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast(
8602 CGF.getTypeSize(
8603 VD->getType().getCanonicalType().getNonReferenceType()),
8604 CGF.Int64Ty, /*isSigned=*/true));
8605 } else {
8606 RValue VarRVal = CGF.EmitLoadOfLValue(VarLVal, RD->getLocation());
8607 LambdaPointers.try_emplace(VarLVal.getPointer(CGF),
8608 VDLVal.getPointer(CGF));
8609 CombinedInfo.Exprs.push_back(VD);
8610 CombinedInfo.BasePointers.push_back(VarLVal.getPointer(CGF));
8611 CombinedInfo.Pointers.push_back(VarRVal.getScalarVal());
8612 CombinedInfo.Sizes.push_back(llvm::ConstantInt::get(CGF.Int64Ty, 0));
8613 }
8614 CombinedInfo.Types.push_back(
8615 OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ |
8616 OpenMPOffloadMappingFlags::OMP_MAP_LITERAL |
8617 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF |
8618 OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT);
8619 CombinedInfo.Mappers.push_back(nullptr);
8620 }
8621 }
8622
8623 /// Set correct indices for lambdas captures.
adjustMemberOfForLambdaCaptures(const llvm::DenseMap<llvm::Value *,llvm::Value * > & LambdaPointers,MapBaseValuesArrayTy & BasePointers,MapValuesArrayTy & Pointers,MapFlagsArrayTy & Types) const8624 void adjustMemberOfForLambdaCaptures(
8625 const llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers,
8626 MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers,
8627 MapFlagsArrayTy &Types) const {
8628 for (unsigned I = 0, E = Types.size(); I < E; ++I) {
8629 // Set correct member_of idx for all implicit lambda captures.
8630 if (Types[I] != (OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ |
8631 OpenMPOffloadMappingFlags::OMP_MAP_LITERAL |
8632 OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF |
8633 OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT))
8634 continue;
8635 llvm::Value *BasePtr = LambdaPointers.lookup(*BasePointers[I]);
8636 assert(BasePtr && "Unable to find base lambda address.");
8637 int TgtIdx = -1;
8638 for (unsigned J = I; J > 0; --J) {
8639 unsigned Idx = J - 1;
8640 if (Pointers[Idx] != BasePtr)
8641 continue;
8642 TgtIdx = Idx;
8643 break;
8644 }
8645 assert(TgtIdx != -1 && "Unable to find parent lambda.");
8646 // All other current entries will be MEMBER_OF the combined entry
8647 // (except for PTR_AND_OBJ entries which do not have a placeholder value
8648 // 0xFFFF in the MEMBER_OF field).
8649 OpenMPOffloadMappingFlags MemberOfFlag = getMemberOfFlag(TgtIdx);
8650 setCorrectMemberOfFlag(Types[I], MemberOfFlag);
8651 }
8652 }
8653
8654 /// Generate the base pointers, section pointers, sizes, map types, and
8655 /// mappers associated to a given capture (all included in \a CombinedInfo).
generateInfoForCapture(const CapturedStmt::Capture * Cap,llvm::Value * Arg,MapCombinedInfoTy & CombinedInfo,StructRangeInfoTy & PartialStruct) const8656 void generateInfoForCapture(const CapturedStmt::Capture *Cap,
8657 llvm::Value *Arg, MapCombinedInfoTy &CombinedInfo,
8658 StructRangeInfoTy &PartialStruct) const {
8659 assert(!Cap->capturesVariableArrayType() &&
8660 "Not expecting to generate map info for a variable array type!");
8661
8662 // We need to know when we generating information for the first component
8663 const ValueDecl *VD = Cap->capturesThis()
8664 ? nullptr
8665 : Cap->getCapturedVar()->getCanonicalDecl();
8666
8667 // for map(to: lambda): skip here, processing it in
8668 // generateDefaultMapInfo
8669 if (LambdasMap.count(VD))
8670 return;
8671
8672 // If this declaration appears in a is_device_ptr clause we just have to
8673 // pass the pointer by value. If it is a reference to a declaration, we just
8674 // pass its value.
8675 if (VD && (DevPointersMap.count(VD) || HasDevAddrsMap.count(VD))) {
8676 CombinedInfo.Exprs.push_back(VD);
8677 CombinedInfo.BasePointers.emplace_back(Arg, VD);
8678 CombinedInfo.Pointers.push_back(Arg);
8679 CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast(
8680 CGF.getTypeSize(CGF.getContext().VoidPtrTy), CGF.Int64Ty,
8681 /*isSigned=*/true));
8682 CombinedInfo.Types.push_back(
8683 (Cap->capturesVariable()
8684 ? OpenMPOffloadMappingFlags::OMP_MAP_TO
8685 : OpenMPOffloadMappingFlags::OMP_MAP_LITERAL) |
8686 OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM);
8687 CombinedInfo.Mappers.push_back(nullptr);
8688 return;
8689 }
8690
8691 using MapData =
8692 std::tuple<OMPClauseMappableExprCommon::MappableExprComponentListRef,
8693 OpenMPMapClauseKind, ArrayRef<OpenMPMapModifierKind>, bool,
8694 const ValueDecl *, const Expr *>;
8695 SmallVector<MapData, 4> DeclComponentLists;
8696 // For member fields list in is_device_ptr, store it in
8697 // DeclComponentLists for generating components info.
8698 static const OpenMPMapModifierKind Unknown = OMPC_MAP_MODIFIER_unknown;
8699 auto It = DevPointersMap.find(VD);
8700 if (It != DevPointersMap.end())
8701 for (const auto &MCL : It->second)
8702 DeclComponentLists.emplace_back(MCL, OMPC_MAP_to, Unknown,
8703 /*IsImpicit = */ true, nullptr,
8704 nullptr);
8705 auto I = HasDevAddrsMap.find(VD);
8706 if (I != HasDevAddrsMap.end())
8707 for (const auto &MCL : I->second)
8708 DeclComponentLists.emplace_back(MCL, OMPC_MAP_tofrom, Unknown,
8709 /*IsImpicit = */ true, nullptr,
8710 nullptr);
8711 assert(CurDir.is<const OMPExecutableDirective *>() &&
8712 "Expect a executable directive");
8713 const auto *CurExecDir = CurDir.get<const OMPExecutableDirective *>();
8714 for (const auto *C : CurExecDir->getClausesOfKind<OMPMapClause>()) {
8715 const auto *EI = C->getVarRefs().begin();
8716 for (const auto L : C->decl_component_lists(VD)) {
8717 const ValueDecl *VDecl, *Mapper;
8718 // The Expression is not correct if the mapping is implicit
8719 const Expr *E = (C->getMapLoc().isValid()) ? *EI : nullptr;
8720 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
8721 std::tie(VDecl, Components, Mapper) = L;
8722 assert(VDecl == VD && "We got information for the wrong declaration??");
8723 assert(!Components.empty() &&
8724 "Not expecting declaration with no component lists.");
8725 DeclComponentLists.emplace_back(Components, C->getMapType(),
8726 C->getMapTypeModifiers(),
8727 C->isImplicit(), Mapper, E);
8728 ++EI;
8729 }
8730 }
8731 llvm::stable_sort(DeclComponentLists, [](const MapData &LHS,
8732 const MapData &RHS) {
8733 ArrayRef<OpenMPMapModifierKind> MapModifiers = std::get<2>(LHS);
8734 OpenMPMapClauseKind MapType = std::get<1>(RHS);
8735 bool HasPresent =
8736 llvm::is_contained(MapModifiers, clang::OMPC_MAP_MODIFIER_present);
8737 bool HasAllocs = MapType == OMPC_MAP_alloc;
8738 MapModifiers = std::get<2>(RHS);
8739 MapType = std::get<1>(LHS);
8740 bool HasPresentR =
8741 llvm::is_contained(MapModifiers, clang::OMPC_MAP_MODIFIER_present);
8742 bool HasAllocsR = MapType == OMPC_MAP_alloc;
8743 return (HasPresent && !HasPresentR) || (HasAllocs && !HasAllocsR);
8744 });
8745
8746 // Find overlapping elements (including the offset from the base element).
8747 llvm::SmallDenseMap<
8748 const MapData *,
8749 llvm::SmallVector<
8750 OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>,
8751 4>
8752 OverlappedData;
8753 size_t Count = 0;
8754 for (const MapData &L : DeclComponentLists) {
8755 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
8756 OpenMPMapClauseKind MapType;
8757 ArrayRef<OpenMPMapModifierKind> MapModifiers;
8758 bool IsImplicit;
8759 const ValueDecl *Mapper;
8760 const Expr *VarRef;
8761 std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper, VarRef) =
8762 L;
8763 ++Count;
8764 for (const MapData &L1 : ArrayRef(DeclComponentLists).slice(Count)) {
8765 OMPClauseMappableExprCommon::MappableExprComponentListRef Components1;
8766 std::tie(Components1, MapType, MapModifiers, IsImplicit, Mapper,
8767 VarRef) = L1;
8768 auto CI = Components.rbegin();
8769 auto CE = Components.rend();
8770 auto SI = Components1.rbegin();
8771 auto SE = Components1.rend();
8772 for (; CI != CE && SI != SE; ++CI, ++SI) {
8773 if (CI->getAssociatedExpression()->getStmtClass() !=
8774 SI->getAssociatedExpression()->getStmtClass())
8775 break;
8776 // Are we dealing with different variables/fields?
8777 if (CI->getAssociatedDeclaration() != SI->getAssociatedDeclaration())
8778 break;
8779 }
8780 // Found overlapping if, at least for one component, reached the head
8781 // of the components list.
8782 if (CI == CE || SI == SE) {
8783 // Ignore it if it is the same component.
8784 if (CI == CE && SI == SE)
8785 continue;
8786 const auto It = (SI == SE) ? CI : SI;
8787 // If one component is a pointer and another one is a kind of
8788 // dereference of this pointer (array subscript, section, dereference,
8789 // etc.), it is not an overlapping.
8790 // Same, if one component is a base and another component is a
8791 // dereferenced pointer memberexpr with the same base.
8792 if (!isa<MemberExpr>(It->getAssociatedExpression()) ||
8793 (std::prev(It)->getAssociatedDeclaration() &&
8794 std::prev(It)
8795 ->getAssociatedDeclaration()
8796 ->getType()
8797 ->isPointerType()) ||
8798 (It->getAssociatedDeclaration() &&
8799 It->getAssociatedDeclaration()->getType()->isPointerType() &&
8800 std::next(It) != CE && std::next(It) != SE))
8801 continue;
8802 const MapData &BaseData = CI == CE ? L : L1;
8803 OMPClauseMappableExprCommon::MappableExprComponentListRef SubData =
8804 SI == SE ? Components : Components1;
8805 auto &OverlappedElements = OverlappedData.FindAndConstruct(&BaseData);
8806 OverlappedElements.getSecond().push_back(SubData);
8807 }
8808 }
8809 }
8810 // Sort the overlapped elements for each item.
8811 llvm::SmallVector<const FieldDecl *, 4> Layout;
8812 if (!OverlappedData.empty()) {
8813 const Type *BaseType = VD->getType().getCanonicalType().getTypePtr();
8814 const Type *OrigType = BaseType->getPointeeOrArrayElementType();
8815 while (BaseType != OrigType) {
8816 BaseType = OrigType->getCanonicalTypeInternal().getTypePtr();
8817 OrigType = BaseType->getPointeeOrArrayElementType();
8818 }
8819
8820 if (const auto *CRD = BaseType->getAsCXXRecordDecl())
8821 getPlainLayout(CRD, Layout, /*AsBase=*/false);
8822 else {
8823 const auto *RD = BaseType->getAsRecordDecl();
8824 Layout.append(RD->field_begin(), RD->field_end());
8825 }
8826 }
8827 for (auto &Pair : OverlappedData) {
8828 llvm::stable_sort(
8829 Pair.getSecond(),
8830 [&Layout](
8831 OMPClauseMappableExprCommon::MappableExprComponentListRef First,
8832 OMPClauseMappableExprCommon::MappableExprComponentListRef
8833 Second) {
8834 auto CI = First.rbegin();
8835 auto CE = First.rend();
8836 auto SI = Second.rbegin();
8837 auto SE = Second.rend();
8838 for (; CI != CE && SI != SE; ++CI, ++SI) {
8839 if (CI->getAssociatedExpression()->getStmtClass() !=
8840 SI->getAssociatedExpression()->getStmtClass())
8841 break;
8842 // Are we dealing with different variables/fields?
8843 if (CI->getAssociatedDeclaration() !=
8844 SI->getAssociatedDeclaration())
8845 break;
8846 }
8847
8848 // Lists contain the same elements.
8849 if (CI == CE && SI == SE)
8850 return false;
8851
8852 // List with less elements is less than list with more elements.
8853 if (CI == CE || SI == SE)
8854 return CI == CE;
8855
8856 const auto *FD1 = cast<FieldDecl>(CI->getAssociatedDeclaration());
8857 const auto *FD2 = cast<FieldDecl>(SI->getAssociatedDeclaration());
8858 if (FD1->getParent() == FD2->getParent())
8859 return FD1->getFieldIndex() < FD2->getFieldIndex();
8860 const auto *It =
8861 llvm::find_if(Layout, [FD1, FD2](const FieldDecl *FD) {
8862 return FD == FD1 || FD == FD2;
8863 });
8864 return *It == FD1;
8865 });
8866 }
8867
8868 // Associated with a capture, because the mapping flags depend on it.
8869 // Go through all of the elements with the overlapped elements.
8870 bool IsFirstComponentList = true;
8871 for (const auto &Pair : OverlappedData) {
8872 const MapData &L = *Pair.getFirst();
8873 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
8874 OpenMPMapClauseKind MapType;
8875 ArrayRef<OpenMPMapModifierKind> MapModifiers;
8876 bool IsImplicit;
8877 const ValueDecl *Mapper;
8878 const Expr *VarRef;
8879 std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper, VarRef) =
8880 L;
8881 ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef>
8882 OverlappedComponents = Pair.getSecond();
8883 generateInfoForComponentList(
8884 MapType, MapModifiers, std::nullopt, Components, CombinedInfo,
8885 PartialStruct, IsFirstComponentList, IsImplicit, Mapper,
8886 /*ForDeviceAddr=*/false, VD, VarRef, OverlappedComponents);
8887 IsFirstComponentList = false;
8888 }
8889 // Go through other elements without overlapped elements.
8890 for (const MapData &L : DeclComponentLists) {
8891 OMPClauseMappableExprCommon::MappableExprComponentListRef Components;
8892 OpenMPMapClauseKind MapType;
8893 ArrayRef<OpenMPMapModifierKind> MapModifiers;
8894 bool IsImplicit;
8895 const ValueDecl *Mapper;
8896 const Expr *VarRef;
8897 std::tie(Components, MapType, MapModifiers, IsImplicit, Mapper, VarRef) =
8898 L;
8899 auto It = OverlappedData.find(&L);
8900 if (It == OverlappedData.end())
8901 generateInfoForComponentList(MapType, MapModifiers, std::nullopt,
8902 Components, CombinedInfo, PartialStruct,
8903 IsFirstComponentList, IsImplicit, Mapper,
8904 /*ForDeviceAddr=*/false, VD, VarRef);
8905 IsFirstComponentList = false;
8906 }
8907 }
8908
8909 /// Generate the default map information for a given capture \a CI,
8910 /// record field declaration \a RI and captured value \a CV.
generateDefaultMapInfo(const CapturedStmt::Capture & CI,const FieldDecl & RI,llvm::Value * CV,MapCombinedInfoTy & CombinedInfo) const8911 void generateDefaultMapInfo(const CapturedStmt::Capture &CI,
8912 const FieldDecl &RI, llvm::Value *CV,
8913 MapCombinedInfoTy &CombinedInfo) const {
8914 bool IsImplicit = true;
8915 // Do the default mapping.
8916 if (CI.capturesThis()) {
8917 CombinedInfo.Exprs.push_back(nullptr);
8918 CombinedInfo.BasePointers.push_back(CV);
8919 CombinedInfo.Pointers.push_back(CV);
8920 const auto *PtrTy = cast<PointerType>(RI.getType().getTypePtr());
8921 CombinedInfo.Sizes.push_back(
8922 CGF.Builder.CreateIntCast(CGF.getTypeSize(PtrTy->getPointeeType()),
8923 CGF.Int64Ty, /*isSigned=*/true));
8924 // Default map type.
8925 CombinedInfo.Types.push_back(OpenMPOffloadMappingFlags::OMP_MAP_TO |
8926 OpenMPOffloadMappingFlags::OMP_MAP_FROM);
8927 } else if (CI.capturesVariableByCopy()) {
8928 const VarDecl *VD = CI.getCapturedVar();
8929 CombinedInfo.Exprs.push_back(VD->getCanonicalDecl());
8930 CombinedInfo.BasePointers.push_back(CV);
8931 CombinedInfo.Pointers.push_back(CV);
8932 if (!RI.getType()->isAnyPointerType()) {
8933 // We have to signal to the runtime captures passed by value that are
8934 // not pointers.
8935 CombinedInfo.Types.push_back(
8936 OpenMPOffloadMappingFlags::OMP_MAP_LITERAL);
8937 CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast(
8938 CGF.getTypeSize(RI.getType()), CGF.Int64Ty, /*isSigned=*/true));
8939 } else {
8940 // Pointers are implicitly mapped with a zero size and no flags
8941 // (other than first map that is added for all implicit maps).
8942 CombinedInfo.Types.push_back(OpenMPOffloadMappingFlags::OMP_MAP_NONE);
8943 CombinedInfo.Sizes.push_back(llvm::Constant::getNullValue(CGF.Int64Ty));
8944 }
8945 auto I = FirstPrivateDecls.find(VD);
8946 if (I != FirstPrivateDecls.end())
8947 IsImplicit = I->getSecond();
8948 } else {
8949 assert(CI.capturesVariable() && "Expected captured reference.");
8950 const auto *PtrTy = cast<ReferenceType>(RI.getType().getTypePtr());
8951 QualType ElementType = PtrTy->getPointeeType();
8952 CombinedInfo.Sizes.push_back(CGF.Builder.CreateIntCast(
8953 CGF.getTypeSize(ElementType), CGF.Int64Ty, /*isSigned=*/true));
8954 // The default map type for a scalar/complex type is 'to' because by
8955 // default the value doesn't have to be retrieved. For an aggregate
8956 // type, the default is 'tofrom'.
8957 CombinedInfo.Types.push_back(getMapModifiersForPrivateClauses(CI));
8958 const VarDecl *VD = CI.getCapturedVar();
8959 auto I = FirstPrivateDecls.find(VD);
8960 CombinedInfo.Exprs.push_back(VD->getCanonicalDecl());
8961 CombinedInfo.BasePointers.push_back(CV);
8962 if (I != FirstPrivateDecls.end() && ElementType->isAnyPointerType()) {
8963 Address PtrAddr = CGF.EmitLoadOfReference(CGF.MakeAddrLValue(
8964 CV, ElementType, CGF.getContext().getDeclAlign(VD),
8965 AlignmentSource::Decl));
8966 CombinedInfo.Pointers.push_back(PtrAddr.getPointer());
8967 } else {
8968 CombinedInfo.Pointers.push_back(CV);
8969 }
8970 if (I != FirstPrivateDecls.end())
8971 IsImplicit = I->getSecond();
8972 }
8973 // Every default map produces a single argument which is a target parameter.
8974 CombinedInfo.Types.back() |=
8975 OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
8976
8977 // Add flag stating this is an implicit map.
8978 if (IsImplicit)
8979 CombinedInfo.Types.back() |= OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT;
8980
8981 // No user-defined mapper for default mapping.
8982 CombinedInfo.Mappers.push_back(nullptr);
8983 }
8984 };
8985 } // anonymous namespace
8986
emitNonContiguousDescriptor(CodeGenFunction & CGF,MappableExprsHandler::MapCombinedInfoTy & CombinedInfo,CGOpenMPRuntime::TargetDataInfo & Info)8987 static void emitNonContiguousDescriptor(
8988 CodeGenFunction &CGF, MappableExprsHandler::MapCombinedInfoTy &CombinedInfo,
8989 CGOpenMPRuntime::TargetDataInfo &Info) {
8990 CodeGenModule &CGM = CGF.CGM;
8991 MappableExprsHandler::MapCombinedInfoTy::StructNonContiguousInfo
8992 &NonContigInfo = CombinedInfo.NonContigInfo;
8993
8994 // Build an array of struct descriptor_dim and then assign it to
8995 // offload_args.
8996 //
8997 // struct descriptor_dim {
8998 // uint64_t offset;
8999 // uint64_t count;
9000 // uint64_t stride
9001 // };
9002 ASTContext &C = CGF.getContext();
9003 QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0);
9004 RecordDecl *RD;
9005 RD = C.buildImplicitRecord("descriptor_dim");
9006 RD->startDefinition();
9007 addFieldToRecordDecl(C, RD, Int64Ty);
9008 addFieldToRecordDecl(C, RD, Int64Ty);
9009 addFieldToRecordDecl(C, RD, Int64Ty);
9010 RD->completeDefinition();
9011 QualType DimTy = C.getRecordType(RD);
9012
9013 enum { OffsetFD = 0, CountFD, StrideFD };
9014 // We need two index variable here since the size of "Dims" is the same as the
9015 // size of Components, however, the size of offset, count, and stride is equal
9016 // to the size of base declaration that is non-contiguous.
9017 for (unsigned I = 0, L = 0, E = NonContigInfo.Dims.size(); I < E; ++I) {
9018 // Skip emitting ir if dimension size is 1 since it cannot be
9019 // non-contiguous.
9020 if (NonContigInfo.Dims[I] == 1)
9021 continue;
9022 llvm::APInt Size(/*numBits=*/32, NonContigInfo.Dims[I]);
9023 QualType ArrayTy =
9024 C.getConstantArrayType(DimTy, Size, nullptr, ArrayType::Normal, 0);
9025 Address DimsAddr = CGF.CreateMemTemp(ArrayTy, "dims");
9026 for (unsigned II = 0, EE = NonContigInfo.Dims[I]; II < EE; ++II) {
9027 unsigned RevIdx = EE - II - 1;
9028 LValue DimsLVal = CGF.MakeAddrLValue(
9029 CGF.Builder.CreateConstArrayGEP(DimsAddr, II), DimTy);
9030 // Offset
9031 LValue OffsetLVal = CGF.EmitLValueForField(
9032 DimsLVal, *std::next(RD->field_begin(), OffsetFD));
9033 CGF.EmitStoreOfScalar(NonContigInfo.Offsets[L][RevIdx], OffsetLVal);
9034 // Count
9035 LValue CountLVal = CGF.EmitLValueForField(
9036 DimsLVal, *std::next(RD->field_begin(), CountFD));
9037 CGF.EmitStoreOfScalar(NonContigInfo.Counts[L][RevIdx], CountLVal);
9038 // Stride
9039 LValue StrideLVal = CGF.EmitLValueForField(
9040 DimsLVal, *std::next(RD->field_begin(), StrideFD));
9041 CGF.EmitStoreOfScalar(NonContigInfo.Strides[L][RevIdx], StrideLVal);
9042 }
9043 // args[I] = &dims
9044 Address DAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
9045 DimsAddr, CGM.Int8PtrTy, CGM.Int8Ty);
9046 llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32(
9047 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
9048 Info.RTArgs.PointersArray, 0, I);
9049 Address PAddr(P, CGM.VoidPtrTy, CGF.getPointerAlign());
9050 CGF.Builder.CreateStore(DAddr.getPointer(), PAddr);
9051 ++L;
9052 }
9053 }
9054
9055 // Try to extract the base declaration from a `this->x` expression if possible.
getDeclFromThisExpr(const Expr * E)9056 static ValueDecl *getDeclFromThisExpr(const Expr *E) {
9057 if (!E)
9058 return nullptr;
9059
9060 if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(E->IgnoreParenCasts()))
9061 if (const MemberExpr *ME =
9062 dyn_cast<MemberExpr>(OASE->getBase()->IgnoreParenImpCasts()))
9063 return ME->getMemberDecl();
9064 return nullptr;
9065 }
9066
9067 /// Emit a string constant containing the names of the values mapped to the
9068 /// offloading runtime library.
9069 llvm::Constant *
emitMappingInformation(CodeGenFunction & CGF,llvm::OpenMPIRBuilder & OMPBuilder,MappableExprsHandler::MappingExprInfo & MapExprs)9070 emitMappingInformation(CodeGenFunction &CGF, llvm::OpenMPIRBuilder &OMPBuilder,
9071 MappableExprsHandler::MappingExprInfo &MapExprs) {
9072
9073 uint32_t SrcLocStrSize;
9074 if (!MapExprs.getMapDecl() && !MapExprs.getMapExpr())
9075 return OMPBuilder.getOrCreateDefaultSrcLocStr(SrcLocStrSize);
9076
9077 SourceLocation Loc;
9078 if (!MapExprs.getMapDecl() && MapExprs.getMapExpr()) {
9079 if (const ValueDecl *VD = getDeclFromThisExpr(MapExprs.getMapExpr()))
9080 Loc = VD->getLocation();
9081 else
9082 Loc = MapExprs.getMapExpr()->getExprLoc();
9083 } else {
9084 Loc = MapExprs.getMapDecl()->getLocation();
9085 }
9086
9087 std::string ExprName;
9088 if (MapExprs.getMapExpr()) {
9089 PrintingPolicy P(CGF.getContext().getLangOpts());
9090 llvm::raw_string_ostream OS(ExprName);
9091 MapExprs.getMapExpr()->printPretty(OS, nullptr, P);
9092 OS.flush();
9093 } else {
9094 ExprName = MapExprs.getMapDecl()->getNameAsString();
9095 }
9096
9097 PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc);
9098 return OMPBuilder.getOrCreateSrcLocStr(PLoc.getFilename(), ExprName,
9099 PLoc.getLine(), PLoc.getColumn(),
9100 SrcLocStrSize);
9101 }
9102
9103 /// Emit the arrays used to pass the captures and map information to the
9104 /// offloading runtime library. If there is no map or capture information,
9105 /// return nullptr by reference.
emitOffloadingArrays(CodeGenFunction & CGF,MappableExprsHandler::MapCombinedInfoTy & CombinedInfo,CGOpenMPRuntime::TargetDataInfo & Info,llvm::OpenMPIRBuilder & OMPBuilder,bool IsNonContiguous=false)9106 static void emitOffloadingArrays(
9107 CodeGenFunction &CGF, MappableExprsHandler::MapCombinedInfoTy &CombinedInfo,
9108 CGOpenMPRuntime::TargetDataInfo &Info, llvm::OpenMPIRBuilder &OMPBuilder,
9109 bool IsNonContiguous = false) {
9110 CodeGenModule &CGM = CGF.CGM;
9111 ASTContext &Ctx = CGF.getContext();
9112
9113 // Reset the array information.
9114 Info.clearArrayInfo();
9115 Info.NumberOfPtrs = CombinedInfo.BasePointers.size();
9116
9117 if (Info.NumberOfPtrs) {
9118 // Detect if we have any capture size requiring runtime evaluation of the
9119 // size so that a constant array could be eventually used.
9120
9121 llvm::APInt PointerNumAP(32, Info.NumberOfPtrs, /*isSigned=*/true);
9122 QualType PointerArrayType = Ctx.getConstantArrayType(
9123 Ctx.VoidPtrTy, PointerNumAP, nullptr, ArrayType::Normal,
9124 /*IndexTypeQuals=*/0);
9125
9126 Info.RTArgs.BasePointersArray =
9127 CGF.CreateMemTemp(PointerArrayType, ".offload_baseptrs").getPointer();
9128 Info.RTArgs.PointersArray =
9129 CGF.CreateMemTemp(PointerArrayType, ".offload_ptrs").getPointer();
9130 Address MappersArray =
9131 CGF.CreateMemTemp(PointerArrayType, ".offload_mappers");
9132 Info.RTArgs.MappersArray = MappersArray.getPointer();
9133
9134 // If we don't have any VLA types or other types that require runtime
9135 // evaluation, we can use a constant array for the map sizes, otherwise we
9136 // need to fill up the arrays as we do for the pointers.
9137 QualType Int64Ty =
9138 Ctx.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
9139 SmallVector<llvm::Constant *> ConstSizes(
9140 CombinedInfo.Sizes.size(), llvm::ConstantInt::get(CGF.Int64Ty, 0));
9141 llvm::SmallBitVector RuntimeSizes(CombinedInfo.Sizes.size());
9142 for (unsigned I = 0, E = CombinedInfo.Sizes.size(); I < E; ++I) {
9143 if (auto *CI = dyn_cast<llvm::Constant>(CombinedInfo.Sizes[I])) {
9144 if (!isa<llvm::ConstantExpr>(CI) && !isa<llvm::GlobalValue>(CI)) {
9145 if (IsNonContiguous &&
9146 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9147 CombinedInfo.Types[I] &
9148 OpenMPOffloadMappingFlags::OMP_MAP_NON_CONTIG))
9149 ConstSizes[I] = llvm::ConstantInt::get(
9150 CGF.Int64Ty, CombinedInfo.NonContigInfo.Dims[I]);
9151 else
9152 ConstSizes[I] = CI;
9153 continue;
9154 }
9155 }
9156 RuntimeSizes.set(I);
9157 }
9158
9159 if (RuntimeSizes.all()) {
9160 QualType SizeArrayType = Ctx.getConstantArrayType(
9161 Int64Ty, PointerNumAP, nullptr, ArrayType::Normal,
9162 /*IndexTypeQuals=*/0);
9163 Info.RTArgs.SizesArray =
9164 CGF.CreateMemTemp(SizeArrayType, ".offload_sizes").getPointer();
9165 } else {
9166 auto *SizesArrayInit = llvm::ConstantArray::get(
9167 llvm::ArrayType::get(CGM.Int64Ty, ConstSizes.size()), ConstSizes);
9168 std::string Name = CGM.getOpenMPRuntime().getName({"offload_sizes"});
9169 auto *SizesArrayGbl = new llvm::GlobalVariable(
9170 CGM.getModule(), SizesArrayInit->getType(), /*isConstant=*/true,
9171 llvm::GlobalValue::PrivateLinkage, SizesArrayInit, Name);
9172 SizesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
9173 if (RuntimeSizes.any()) {
9174 QualType SizeArrayType = Ctx.getConstantArrayType(
9175 Int64Ty, PointerNumAP, nullptr, ArrayType::Normal,
9176 /*IndexTypeQuals=*/0);
9177 Address Buffer = CGF.CreateMemTemp(SizeArrayType, ".offload_sizes");
9178 llvm::Value *GblConstPtr =
9179 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
9180 SizesArrayGbl, CGM.Int64Ty->getPointerTo());
9181 CGF.Builder.CreateMemCpy(
9182 Buffer,
9183 Address(GblConstPtr, CGM.Int64Ty,
9184 CGM.getNaturalTypeAlignment(Ctx.getIntTypeForBitwidth(
9185 /*DestWidth=*/64, /*Signed=*/false))),
9186 CGF.getTypeSize(SizeArrayType));
9187 Info.RTArgs.SizesArray = Buffer.getPointer();
9188 } else {
9189 Info.RTArgs.SizesArray = SizesArrayGbl;
9190 }
9191 }
9192
9193 // The map types are always constant so we don't need to generate code to
9194 // fill arrays. Instead, we create an array constant.
9195 SmallVector<uint64_t, 4> Mapping;
9196 for (auto mapFlag : CombinedInfo.Types)
9197 Mapping.push_back(
9198 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9199 mapFlag));
9200 std::string MaptypesName =
9201 CGM.getOpenMPRuntime().getName({"offload_maptypes"});
9202 auto *MapTypesArrayGbl =
9203 OMPBuilder.createOffloadMaptypes(Mapping, MaptypesName);
9204 Info.RTArgs.MapTypesArray = MapTypesArrayGbl;
9205
9206 // The information types are only built if there is debug information
9207 // requested.
9208 if (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo) {
9209 Info.RTArgs.MapNamesArray = llvm::Constant::getNullValue(
9210 llvm::Type::getInt8Ty(CGF.Builder.getContext())->getPointerTo());
9211 } else {
9212 auto fillInfoMap = [&](MappableExprsHandler::MappingExprInfo &MapExpr) {
9213 return emitMappingInformation(CGF, OMPBuilder, MapExpr);
9214 };
9215 SmallVector<llvm::Constant *, 4> InfoMap(CombinedInfo.Exprs.size());
9216 llvm::transform(CombinedInfo.Exprs, InfoMap.begin(), fillInfoMap);
9217 std::string MapnamesName =
9218 CGM.getOpenMPRuntime().getName({"offload_mapnames"});
9219 auto *MapNamesArrayGbl =
9220 OMPBuilder.createOffloadMapnames(InfoMap, MapnamesName);
9221 Info.RTArgs.MapNamesArray = MapNamesArrayGbl;
9222 }
9223
9224 // If there's a present map type modifier, it must not be applied to the end
9225 // of a region, so generate a separate map type array in that case.
9226 if (Info.separateBeginEndCalls()) {
9227 bool EndMapTypesDiffer = false;
9228 for (uint64_t &Type : Mapping) {
9229 if (Type &
9230 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9231 OpenMPOffloadMappingFlags::OMP_MAP_PRESENT)) {
9232 Type &=
9233 ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9234 OpenMPOffloadMappingFlags::OMP_MAP_PRESENT);
9235 EndMapTypesDiffer = true;
9236 }
9237 }
9238 if (EndMapTypesDiffer) {
9239 MapTypesArrayGbl =
9240 OMPBuilder.createOffloadMaptypes(Mapping, MaptypesName);
9241 Info.RTArgs.MapTypesArrayEnd = MapTypesArrayGbl;
9242 }
9243 }
9244
9245 for (unsigned I = 0; I < Info.NumberOfPtrs; ++I) {
9246 llvm::Value *BPVal = *CombinedInfo.BasePointers[I];
9247 llvm::Value *BP = CGF.Builder.CreateConstInBoundsGEP2_32(
9248 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
9249 Info.RTArgs.BasePointersArray, 0, I);
9250 BP = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
9251 BP, BPVal->getType()->getPointerTo(/*AddrSpace=*/0));
9252 Address BPAddr(BP, BPVal->getType(),
9253 Ctx.getTypeAlignInChars(Ctx.VoidPtrTy));
9254 CGF.Builder.CreateStore(BPVal, BPAddr);
9255
9256 if (Info.requiresDevicePointerInfo())
9257 if (const ValueDecl *DevVD =
9258 CombinedInfo.BasePointers[I].getDevicePtrDecl())
9259 Info.CaptureDeviceAddrMap.try_emplace(DevVD, BPAddr);
9260
9261 llvm::Value *PVal = CombinedInfo.Pointers[I];
9262 llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32(
9263 llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs),
9264 Info.RTArgs.PointersArray, 0, I);
9265 P = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
9266 P, PVal->getType()->getPointerTo(/*AddrSpace=*/0));
9267 Address PAddr(P, PVal->getType(), Ctx.getTypeAlignInChars(Ctx.VoidPtrTy));
9268 CGF.Builder.CreateStore(PVal, PAddr);
9269
9270 if (RuntimeSizes.test(I)) {
9271 llvm::Value *S = CGF.Builder.CreateConstInBoundsGEP2_32(
9272 llvm::ArrayType::get(CGM.Int64Ty, Info.NumberOfPtrs),
9273 Info.RTArgs.SizesArray,
9274 /*Idx0=*/0,
9275 /*Idx1=*/I);
9276 Address SAddr(S, CGM.Int64Ty, Ctx.getTypeAlignInChars(Int64Ty));
9277 CGF.Builder.CreateStore(CGF.Builder.CreateIntCast(CombinedInfo.Sizes[I],
9278 CGM.Int64Ty,
9279 /*isSigned=*/true),
9280 SAddr);
9281 }
9282
9283 // Fill up the mapper array.
9284 llvm::Value *MFunc = llvm::ConstantPointerNull::get(CGM.VoidPtrTy);
9285 if (CombinedInfo.Mappers[I]) {
9286 MFunc = CGM.getOpenMPRuntime().getOrCreateUserDefinedMapperFunc(
9287 cast<OMPDeclareMapperDecl>(CombinedInfo.Mappers[I]));
9288 MFunc = CGF.Builder.CreatePointerCast(MFunc, CGM.VoidPtrTy);
9289 Info.HasMapper = true;
9290 }
9291 Address MAddr = CGF.Builder.CreateConstArrayGEP(MappersArray, I);
9292 CGF.Builder.CreateStore(MFunc, MAddr);
9293 }
9294 }
9295
9296 if (!IsNonContiguous || CombinedInfo.NonContigInfo.Offsets.empty() ||
9297 Info.NumberOfPtrs == 0)
9298 return;
9299
9300 emitNonContiguousDescriptor(CGF, CombinedInfo, Info);
9301 }
9302
9303 /// Check for inner distribute directive.
9304 static const OMPExecutableDirective *
getNestedDistributeDirective(ASTContext & Ctx,const OMPExecutableDirective & D)9305 getNestedDistributeDirective(ASTContext &Ctx, const OMPExecutableDirective &D) {
9306 const auto *CS = D.getInnermostCapturedStmt();
9307 const auto *Body =
9308 CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
9309 const Stmt *ChildStmt =
9310 CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body);
9311
9312 if (const auto *NestedDir =
9313 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
9314 OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
9315 switch (D.getDirectiveKind()) {
9316 case OMPD_target:
9317 if (isOpenMPDistributeDirective(DKind))
9318 return NestedDir;
9319 if (DKind == OMPD_teams) {
9320 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
9321 /*IgnoreCaptured=*/true);
9322 if (!Body)
9323 return nullptr;
9324 ChildStmt = CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body);
9325 if (const auto *NND =
9326 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
9327 DKind = NND->getDirectiveKind();
9328 if (isOpenMPDistributeDirective(DKind))
9329 return NND;
9330 }
9331 }
9332 return nullptr;
9333 case OMPD_target_teams:
9334 if (isOpenMPDistributeDirective(DKind))
9335 return NestedDir;
9336 return nullptr;
9337 case OMPD_target_parallel:
9338 case OMPD_target_simd:
9339 case OMPD_target_parallel_for:
9340 case OMPD_target_parallel_for_simd:
9341 return nullptr;
9342 case OMPD_target_teams_distribute:
9343 case OMPD_target_teams_distribute_simd:
9344 case OMPD_target_teams_distribute_parallel_for:
9345 case OMPD_target_teams_distribute_parallel_for_simd:
9346 case OMPD_parallel:
9347 case OMPD_for:
9348 case OMPD_parallel_for:
9349 case OMPD_parallel_master:
9350 case OMPD_parallel_sections:
9351 case OMPD_for_simd:
9352 case OMPD_parallel_for_simd:
9353 case OMPD_cancel:
9354 case OMPD_cancellation_point:
9355 case OMPD_ordered:
9356 case OMPD_threadprivate:
9357 case OMPD_allocate:
9358 case OMPD_task:
9359 case OMPD_simd:
9360 case OMPD_tile:
9361 case OMPD_unroll:
9362 case OMPD_sections:
9363 case OMPD_section:
9364 case OMPD_single:
9365 case OMPD_master:
9366 case OMPD_critical:
9367 case OMPD_taskyield:
9368 case OMPD_barrier:
9369 case OMPD_taskwait:
9370 case OMPD_taskgroup:
9371 case OMPD_atomic:
9372 case OMPD_flush:
9373 case OMPD_depobj:
9374 case OMPD_scan:
9375 case OMPD_teams:
9376 case OMPD_target_data:
9377 case OMPD_target_exit_data:
9378 case OMPD_target_enter_data:
9379 case OMPD_distribute:
9380 case OMPD_distribute_simd:
9381 case OMPD_distribute_parallel_for:
9382 case OMPD_distribute_parallel_for_simd:
9383 case OMPD_teams_distribute:
9384 case OMPD_teams_distribute_simd:
9385 case OMPD_teams_distribute_parallel_for:
9386 case OMPD_teams_distribute_parallel_for_simd:
9387 case OMPD_target_update:
9388 case OMPD_declare_simd:
9389 case OMPD_declare_variant:
9390 case OMPD_begin_declare_variant:
9391 case OMPD_end_declare_variant:
9392 case OMPD_declare_target:
9393 case OMPD_end_declare_target:
9394 case OMPD_declare_reduction:
9395 case OMPD_declare_mapper:
9396 case OMPD_taskloop:
9397 case OMPD_taskloop_simd:
9398 case OMPD_master_taskloop:
9399 case OMPD_master_taskloop_simd:
9400 case OMPD_parallel_master_taskloop:
9401 case OMPD_parallel_master_taskloop_simd:
9402 case OMPD_requires:
9403 case OMPD_metadirective:
9404 case OMPD_unknown:
9405 default:
9406 llvm_unreachable("Unexpected directive.");
9407 }
9408 }
9409
9410 return nullptr;
9411 }
9412
9413 /// Emit the user-defined mapper function. The code generation follows the
9414 /// pattern in the example below.
9415 /// \code
9416 /// void .omp_mapper.<type_name>.<mapper_id>.(void *rt_mapper_handle,
9417 /// void *base, void *begin,
9418 /// int64_t size, int64_t type,
9419 /// void *name = nullptr) {
9420 /// // Allocate space for an array section first or add a base/begin for
9421 /// // pointer dereference.
9422 /// if ((size > 1 || (base != begin && maptype.IsPtrAndObj)) &&
9423 /// !maptype.IsDelete)
9424 /// __tgt_push_mapper_component(rt_mapper_handle, base, begin,
9425 /// size*sizeof(Ty), clearToFromMember(type));
9426 /// // Map members.
9427 /// for (unsigned i = 0; i < size; i++) {
9428 /// // For each component specified by this mapper:
9429 /// for (auto c : begin[i]->all_components) {
9430 /// if (c.hasMapper())
9431 /// (*c.Mapper())(rt_mapper_handle, c.arg_base, c.arg_begin, c.arg_size,
9432 /// c.arg_type, c.arg_name);
9433 /// else
9434 /// __tgt_push_mapper_component(rt_mapper_handle, c.arg_base,
9435 /// c.arg_begin, c.arg_size, c.arg_type,
9436 /// c.arg_name);
9437 /// }
9438 /// }
9439 /// // Delete the array section.
9440 /// if (size > 1 && maptype.IsDelete)
9441 /// __tgt_push_mapper_component(rt_mapper_handle, base, begin,
9442 /// size*sizeof(Ty), clearToFromMember(type));
9443 /// }
9444 /// \endcode
emitUserDefinedMapper(const OMPDeclareMapperDecl * D,CodeGenFunction * CGF)9445 void CGOpenMPRuntime::emitUserDefinedMapper(const OMPDeclareMapperDecl *D,
9446 CodeGenFunction *CGF) {
9447 if (UDMMap.count(D) > 0)
9448 return;
9449 ASTContext &C = CGM.getContext();
9450 QualType Ty = D->getType();
9451 QualType PtrTy = C.getPointerType(Ty).withRestrict();
9452 QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true);
9453 auto *MapperVarDecl =
9454 cast<VarDecl>(cast<DeclRefExpr>(D->getMapperVarRef())->getDecl());
9455 SourceLocation Loc = D->getLocation();
9456 CharUnits ElementSize = C.getTypeSizeInChars(Ty);
9457 llvm::Type *ElemTy = CGM.getTypes().ConvertTypeForMem(Ty);
9458
9459 // Prepare mapper function arguments and attributes.
9460 ImplicitParamDecl HandleArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
9461 C.VoidPtrTy, ImplicitParamDecl::Other);
9462 ImplicitParamDecl BaseArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
9463 ImplicitParamDecl::Other);
9464 ImplicitParamDecl BeginArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
9465 C.VoidPtrTy, ImplicitParamDecl::Other);
9466 ImplicitParamDecl SizeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, Int64Ty,
9467 ImplicitParamDecl::Other);
9468 ImplicitParamDecl TypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, Int64Ty,
9469 ImplicitParamDecl::Other);
9470 ImplicitParamDecl NameArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy,
9471 ImplicitParamDecl::Other);
9472 FunctionArgList Args;
9473 Args.push_back(&HandleArg);
9474 Args.push_back(&BaseArg);
9475 Args.push_back(&BeginArg);
9476 Args.push_back(&SizeArg);
9477 Args.push_back(&TypeArg);
9478 Args.push_back(&NameArg);
9479 const CGFunctionInfo &FnInfo =
9480 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
9481 llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo);
9482 SmallString<64> TyStr;
9483 llvm::raw_svector_ostream Out(TyStr);
9484 CGM.getCXXABI().getMangleContext().mangleTypeName(Ty, Out);
9485 std::string Name = getName({"omp_mapper", TyStr, D->getName()});
9486 auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage,
9487 Name, &CGM.getModule());
9488 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo);
9489 Fn->removeFnAttr(llvm::Attribute::OptimizeNone);
9490 // Start the mapper function code generation.
9491 CodeGenFunction MapperCGF(CGM);
9492 MapperCGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc);
9493 // Compute the starting and end addresses of array elements.
9494 llvm::Value *Size = MapperCGF.EmitLoadOfScalar(
9495 MapperCGF.GetAddrOfLocalVar(&SizeArg), /*Volatile=*/false,
9496 C.getPointerType(Int64Ty), Loc);
9497 // Prepare common arguments for array initiation and deletion.
9498 llvm::Value *Handle = MapperCGF.EmitLoadOfScalar(
9499 MapperCGF.GetAddrOfLocalVar(&HandleArg),
9500 /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
9501 llvm::Value *BaseIn = MapperCGF.EmitLoadOfScalar(
9502 MapperCGF.GetAddrOfLocalVar(&BaseArg),
9503 /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
9504 llvm::Value *BeginIn = MapperCGF.EmitLoadOfScalar(
9505 MapperCGF.GetAddrOfLocalVar(&BeginArg),
9506 /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
9507 // Convert the size in bytes into the number of array elements.
9508 Size = MapperCGF.Builder.CreateExactUDiv(
9509 Size, MapperCGF.Builder.getInt64(ElementSize.getQuantity()));
9510 llvm::Value *PtrBegin = MapperCGF.Builder.CreateBitCast(
9511 BeginIn, CGM.getTypes().ConvertTypeForMem(PtrTy));
9512 llvm::Value *PtrEnd = MapperCGF.Builder.CreateGEP(ElemTy, PtrBegin, Size);
9513 llvm::Value *MapType = MapperCGF.EmitLoadOfScalar(
9514 MapperCGF.GetAddrOfLocalVar(&TypeArg), /*Volatile=*/false,
9515 C.getPointerType(Int64Ty), Loc);
9516 llvm::Value *MapName = MapperCGF.EmitLoadOfScalar(
9517 MapperCGF.GetAddrOfLocalVar(&NameArg),
9518 /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
9519
9520 // Emit array initiation if this is an array section and \p MapType indicates
9521 // that memory allocation is required.
9522 llvm::BasicBlock *HeadBB = MapperCGF.createBasicBlock("omp.arraymap.head");
9523 emitUDMapperArrayInitOrDel(MapperCGF, Handle, BaseIn, BeginIn, Size, MapType,
9524 MapName, ElementSize, HeadBB, /*IsInit=*/true);
9525
9526 // Emit a for loop to iterate through SizeArg of elements and map all of them.
9527
9528 // Emit the loop header block.
9529 MapperCGF.EmitBlock(HeadBB);
9530 llvm::BasicBlock *BodyBB = MapperCGF.createBasicBlock("omp.arraymap.body");
9531 llvm::BasicBlock *DoneBB = MapperCGF.createBasicBlock("omp.done");
9532 // Evaluate whether the initial condition is satisfied.
9533 llvm::Value *IsEmpty =
9534 MapperCGF.Builder.CreateICmpEQ(PtrBegin, PtrEnd, "omp.arraymap.isempty");
9535 MapperCGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
9536 llvm::BasicBlock *EntryBB = MapperCGF.Builder.GetInsertBlock();
9537
9538 // Emit the loop body block.
9539 MapperCGF.EmitBlock(BodyBB);
9540 llvm::BasicBlock *LastBB = BodyBB;
9541 llvm::PHINode *PtrPHI = MapperCGF.Builder.CreatePHI(
9542 PtrBegin->getType(), 2, "omp.arraymap.ptrcurrent");
9543 PtrPHI->addIncoming(PtrBegin, EntryBB);
9544 Address PtrCurrent(PtrPHI, ElemTy,
9545 MapperCGF.GetAddrOfLocalVar(&BeginArg)
9546 .getAlignment()
9547 .alignmentOfArrayElement(ElementSize));
9548 // Privatize the declared variable of mapper to be the current array element.
9549 CodeGenFunction::OMPPrivateScope Scope(MapperCGF);
9550 Scope.addPrivate(MapperVarDecl, PtrCurrent);
9551 (void)Scope.Privatize();
9552
9553 // Get map clause information. Fill up the arrays with all mapped variables.
9554 MappableExprsHandler::MapCombinedInfoTy Info;
9555 MappableExprsHandler MEHandler(*D, MapperCGF);
9556 MEHandler.generateAllInfoForMapper(Info);
9557
9558 // Call the runtime API __tgt_mapper_num_components to get the number of
9559 // pre-existing components.
9560 llvm::Value *OffloadingArgs[] = {Handle};
9561 llvm::Value *PreviousSize = MapperCGF.EmitRuntimeCall(
9562 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
9563 OMPRTL___tgt_mapper_num_components),
9564 OffloadingArgs);
9565 llvm::Value *ShiftedPreviousSize = MapperCGF.Builder.CreateShl(
9566 PreviousSize,
9567 MapperCGF.Builder.getInt64(MappableExprsHandler::getFlagMemberOffset()));
9568
9569 // Fill up the runtime mapper handle for all components.
9570 for (unsigned I = 0; I < Info.BasePointers.size(); ++I) {
9571 llvm::Value *CurBaseArg = MapperCGF.Builder.CreateBitCast(
9572 *Info.BasePointers[I], CGM.getTypes().ConvertTypeForMem(C.VoidPtrTy));
9573 llvm::Value *CurBeginArg = MapperCGF.Builder.CreateBitCast(
9574 Info.Pointers[I], CGM.getTypes().ConvertTypeForMem(C.VoidPtrTy));
9575 llvm::Value *CurSizeArg = Info.Sizes[I];
9576 llvm::Value *CurNameArg =
9577 (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo)
9578 ? llvm::ConstantPointerNull::get(CGM.VoidPtrTy)
9579 : emitMappingInformation(MapperCGF, OMPBuilder, Info.Exprs[I]);
9580
9581 // Extract the MEMBER_OF field from the map type.
9582 llvm::Value *OriMapType = MapperCGF.Builder.getInt64(
9583 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9584 Info.Types[I]));
9585 llvm::Value *MemberMapType =
9586 MapperCGF.Builder.CreateNUWAdd(OriMapType, ShiftedPreviousSize);
9587
9588 // Combine the map type inherited from user-defined mapper with that
9589 // specified in the program. According to the OMP_MAP_TO and OMP_MAP_FROM
9590 // bits of the \a MapType, which is the input argument of the mapper
9591 // function, the following code will set the OMP_MAP_TO and OMP_MAP_FROM
9592 // bits of MemberMapType.
9593 // [OpenMP 5.0], 1.2.6. map-type decay.
9594 // | alloc | to | from | tofrom | release | delete
9595 // ----------------------------------------------------------
9596 // alloc | alloc | alloc | alloc | alloc | release | delete
9597 // to | alloc | to | alloc | to | release | delete
9598 // from | alloc | alloc | from | from | release | delete
9599 // tofrom | alloc | to | from | tofrom | release | delete
9600 llvm::Value *LeftToFrom = MapperCGF.Builder.CreateAnd(
9601 MapType,
9602 MapperCGF.Builder.getInt64(
9603 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9604 OpenMPOffloadMappingFlags::OMP_MAP_TO |
9605 OpenMPOffloadMappingFlags::OMP_MAP_FROM)));
9606 llvm::BasicBlock *AllocBB = MapperCGF.createBasicBlock("omp.type.alloc");
9607 llvm::BasicBlock *AllocElseBB =
9608 MapperCGF.createBasicBlock("omp.type.alloc.else");
9609 llvm::BasicBlock *ToBB = MapperCGF.createBasicBlock("omp.type.to");
9610 llvm::BasicBlock *ToElseBB = MapperCGF.createBasicBlock("omp.type.to.else");
9611 llvm::BasicBlock *FromBB = MapperCGF.createBasicBlock("omp.type.from");
9612 llvm::BasicBlock *EndBB = MapperCGF.createBasicBlock("omp.type.end");
9613 llvm::Value *IsAlloc = MapperCGF.Builder.CreateIsNull(LeftToFrom);
9614 MapperCGF.Builder.CreateCondBr(IsAlloc, AllocBB, AllocElseBB);
9615 // In case of alloc, clear OMP_MAP_TO and OMP_MAP_FROM.
9616 MapperCGF.EmitBlock(AllocBB);
9617 llvm::Value *AllocMapType = MapperCGF.Builder.CreateAnd(
9618 MemberMapType,
9619 MapperCGF.Builder.getInt64(
9620 ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9621 OpenMPOffloadMappingFlags::OMP_MAP_TO |
9622 OpenMPOffloadMappingFlags::OMP_MAP_FROM)));
9623 MapperCGF.Builder.CreateBr(EndBB);
9624 MapperCGF.EmitBlock(AllocElseBB);
9625 llvm::Value *IsTo = MapperCGF.Builder.CreateICmpEQ(
9626 LeftToFrom,
9627 MapperCGF.Builder.getInt64(
9628 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9629 OpenMPOffloadMappingFlags::OMP_MAP_TO)));
9630 MapperCGF.Builder.CreateCondBr(IsTo, ToBB, ToElseBB);
9631 // In case of to, clear OMP_MAP_FROM.
9632 MapperCGF.EmitBlock(ToBB);
9633 llvm::Value *ToMapType = MapperCGF.Builder.CreateAnd(
9634 MemberMapType,
9635 MapperCGF.Builder.getInt64(
9636 ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9637 OpenMPOffloadMappingFlags::OMP_MAP_FROM)));
9638 MapperCGF.Builder.CreateBr(EndBB);
9639 MapperCGF.EmitBlock(ToElseBB);
9640 llvm::Value *IsFrom = MapperCGF.Builder.CreateICmpEQ(
9641 LeftToFrom,
9642 MapperCGF.Builder.getInt64(
9643 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9644 OpenMPOffloadMappingFlags::OMP_MAP_FROM)));
9645 MapperCGF.Builder.CreateCondBr(IsFrom, FromBB, EndBB);
9646 // In case of from, clear OMP_MAP_TO.
9647 MapperCGF.EmitBlock(FromBB);
9648 llvm::Value *FromMapType = MapperCGF.Builder.CreateAnd(
9649 MemberMapType,
9650 MapperCGF.Builder.getInt64(
9651 ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9652 OpenMPOffloadMappingFlags::OMP_MAP_TO)));
9653 // In case of tofrom, do nothing.
9654 MapperCGF.EmitBlock(EndBB);
9655 LastBB = EndBB;
9656 llvm::PHINode *CurMapType =
9657 MapperCGF.Builder.CreatePHI(CGM.Int64Ty, 4, "omp.maptype");
9658 CurMapType->addIncoming(AllocMapType, AllocBB);
9659 CurMapType->addIncoming(ToMapType, ToBB);
9660 CurMapType->addIncoming(FromMapType, FromBB);
9661 CurMapType->addIncoming(MemberMapType, ToElseBB);
9662
9663 llvm::Value *OffloadingArgs[] = {Handle, CurBaseArg, CurBeginArg,
9664 CurSizeArg, CurMapType, CurNameArg};
9665 if (Info.Mappers[I]) {
9666 // Call the corresponding mapper function.
9667 llvm::Function *MapperFunc = getOrCreateUserDefinedMapperFunc(
9668 cast<OMPDeclareMapperDecl>(Info.Mappers[I]));
9669 assert(MapperFunc && "Expect a valid mapper function is available.");
9670 MapperCGF.EmitNounwindRuntimeCall(MapperFunc, OffloadingArgs);
9671 } else {
9672 // Call the runtime API __tgt_push_mapper_component to fill up the runtime
9673 // data structure.
9674 MapperCGF.EmitRuntimeCall(
9675 OMPBuilder.getOrCreateRuntimeFunction(
9676 CGM.getModule(), OMPRTL___tgt_push_mapper_component),
9677 OffloadingArgs);
9678 }
9679 }
9680
9681 // Update the pointer to point to the next element that needs to be mapped,
9682 // and check whether we have mapped all elements.
9683 llvm::Value *PtrNext = MapperCGF.Builder.CreateConstGEP1_32(
9684 ElemTy, PtrPHI, /*Idx0=*/1, "omp.arraymap.next");
9685 PtrPHI->addIncoming(PtrNext, LastBB);
9686 llvm::Value *IsDone =
9687 MapperCGF.Builder.CreateICmpEQ(PtrNext, PtrEnd, "omp.arraymap.isdone");
9688 llvm::BasicBlock *ExitBB = MapperCGF.createBasicBlock("omp.arraymap.exit");
9689 MapperCGF.Builder.CreateCondBr(IsDone, ExitBB, BodyBB);
9690
9691 MapperCGF.EmitBlock(ExitBB);
9692 // Emit array deletion if this is an array section and \p MapType indicates
9693 // that deletion is required.
9694 emitUDMapperArrayInitOrDel(MapperCGF, Handle, BaseIn, BeginIn, Size, MapType,
9695 MapName, ElementSize, DoneBB, /*IsInit=*/false);
9696
9697 // Emit the function exit block.
9698 MapperCGF.EmitBlock(DoneBB, /*IsFinished=*/true);
9699 MapperCGF.FinishFunction();
9700 UDMMap.try_emplace(D, Fn);
9701 if (CGF) {
9702 auto &Decls = FunctionUDMMap.FindAndConstruct(CGF->CurFn);
9703 Decls.second.push_back(D);
9704 }
9705 }
9706
9707 /// Emit the array initialization or deletion portion for user-defined mapper
9708 /// code generation. First, it evaluates whether an array section is mapped and
9709 /// whether the \a MapType instructs to delete this section. If \a IsInit is
9710 /// true, and \a MapType indicates to not delete this array, array
9711 /// initialization code is generated. If \a IsInit is false, and \a MapType
9712 /// indicates to not this array, array deletion code is generated.
emitUDMapperArrayInitOrDel(CodeGenFunction & MapperCGF,llvm::Value * Handle,llvm::Value * Base,llvm::Value * Begin,llvm::Value * Size,llvm::Value * MapType,llvm::Value * MapName,CharUnits ElementSize,llvm::BasicBlock * ExitBB,bool IsInit)9713 void CGOpenMPRuntime::emitUDMapperArrayInitOrDel(
9714 CodeGenFunction &MapperCGF, llvm::Value *Handle, llvm::Value *Base,
9715 llvm::Value *Begin, llvm::Value *Size, llvm::Value *MapType,
9716 llvm::Value *MapName, CharUnits ElementSize, llvm::BasicBlock *ExitBB,
9717 bool IsInit) {
9718 StringRef Prefix = IsInit ? ".init" : ".del";
9719
9720 // Evaluate if this is an array section.
9721 llvm::BasicBlock *BodyBB =
9722 MapperCGF.createBasicBlock(getName({"omp.array", Prefix}));
9723 llvm::Value *IsArray = MapperCGF.Builder.CreateICmpSGT(
9724 Size, MapperCGF.Builder.getInt64(1), "omp.arrayinit.isarray");
9725 llvm::Value *DeleteBit = MapperCGF.Builder.CreateAnd(
9726 MapType,
9727 MapperCGF.Builder.getInt64(
9728 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9729 OpenMPOffloadMappingFlags::OMP_MAP_DELETE)));
9730 llvm::Value *DeleteCond;
9731 llvm::Value *Cond;
9732 if (IsInit) {
9733 // base != begin?
9734 llvm::Value *BaseIsBegin = MapperCGF.Builder.CreateICmpNE(Base, Begin);
9735 // IsPtrAndObj?
9736 llvm::Value *PtrAndObjBit = MapperCGF.Builder.CreateAnd(
9737 MapType,
9738 MapperCGF.Builder.getInt64(
9739 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9740 OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ)));
9741 PtrAndObjBit = MapperCGF.Builder.CreateIsNotNull(PtrAndObjBit);
9742 BaseIsBegin = MapperCGF.Builder.CreateAnd(BaseIsBegin, PtrAndObjBit);
9743 Cond = MapperCGF.Builder.CreateOr(IsArray, BaseIsBegin);
9744 DeleteCond = MapperCGF.Builder.CreateIsNull(
9745 DeleteBit, getName({"omp.array", Prefix, ".delete"}));
9746 } else {
9747 Cond = IsArray;
9748 DeleteCond = MapperCGF.Builder.CreateIsNotNull(
9749 DeleteBit, getName({"omp.array", Prefix, ".delete"}));
9750 }
9751 Cond = MapperCGF.Builder.CreateAnd(Cond, DeleteCond);
9752 MapperCGF.Builder.CreateCondBr(Cond, BodyBB, ExitBB);
9753
9754 MapperCGF.EmitBlock(BodyBB);
9755 // Get the array size by multiplying element size and element number (i.e., \p
9756 // Size).
9757 llvm::Value *ArraySize = MapperCGF.Builder.CreateNUWMul(
9758 Size, MapperCGF.Builder.getInt64(ElementSize.getQuantity()));
9759 // Remove OMP_MAP_TO and OMP_MAP_FROM from the map type, so that it achieves
9760 // memory allocation/deletion purpose only.
9761 llvm::Value *MapTypeArg = MapperCGF.Builder.CreateAnd(
9762 MapType,
9763 MapperCGF.Builder.getInt64(
9764 ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9765 OpenMPOffloadMappingFlags::OMP_MAP_TO |
9766 OpenMPOffloadMappingFlags::OMP_MAP_FROM)));
9767 MapTypeArg = MapperCGF.Builder.CreateOr(
9768 MapTypeArg,
9769 MapperCGF.Builder.getInt64(
9770 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
9771 OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT)));
9772
9773 // Call the runtime API __tgt_push_mapper_component to fill up the runtime
9774 // data structure.
9775 llvm::Value *OffloadingArgs[] = {Handle, Base, Begin,
9776 ArraySize, MapTypeArg, MapName};
9777 MapperCGF.EmitRuntimeCall(
9778 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
9779 OMPRTL___tgt_push_mapper_component),
9780 OffloadingArgs);
9781 }
9782
getOrCreateUserDefinedMapperFunc(const OMPDeclareMapperDecl * D)9783 llvm::Function *CGOpenMPRuntime::getOrCreateUserDefinedMapperFunc(
9784 const OMPDeclareMapperDecl *D) {
9785 auto I = UDMMap.find(D);
9786 if (I != UDMMap.end())
9787 return I->second;
9788 emitUserDefinedMapper(D);
9789 return UDMMap.lookup(D);
9790 }
9791
emitTargetNumIterationsCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,llvm::function_ref<llvm::Value * (CodeGenFunction & CGF,const OMPLoopDirective & D)> SizeEmitter)9792 llvm::Value *CGOpenMPRuntime::emitTargetNumIterationsCall(
9793 CodeGenFunction &CGF, const OMPExecutableDirective &D,
9794 llvm::function_ref<llvm::Value *(CodeGenFunction &CGF,
9795 const OMPLoopDirective &D)>
9796 SizeEmitter) {
9797 OpenMPDirectiveKind Kind = D.getDirectiveKind();
9798 const OMPExecutableDirective *TD = &D;
9799 // Get nested teams distribute kind directive, if any.
9800 if (!isOpenMPDistributeDirective(Kind) || !isOpenMPTeamsDirective(Kind))
9801 TD = getNestedDistributeDirective(CGM.getContext(), D);
9802 if (!TD)
9803 return llvm::ConstantInt::get(CGF.Int64Ty, 0);
9804
9805 const auto *LD = cast<OMPLoopDirective>(TD);
9806 if (llvm::Value *NumIterations = SizeEmitter(CGF, *LD))
9807 return NumIterations;
9808 return llvm::ConstantInt::get(CGF.Int64Ty, 0);
9809 }
9810
emitTargetCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,llvm::Function * OutlinedFn,llvm::Value * OutlinedFnID,const Expr * IfCond,llvm::PointerIntPair<const Expr *,2,OpenMPDeviceClauseModifier> Device,llvm::function_ref<llvm::Value * (CodeGenFunction & CGF,const OMPLoopDirective & D)> SizeEmitter)9811 void CGOpenMPRuntime::emitTargetCall(
9812 CodeGenFunction &CGF, const OMPExecutableDirective &D,
9813 llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, const Expr *IfCond,
9814 llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device,
9815 llvm::function_ref<llvm::Value *(CodeGenFunction &CGF,
9816 const OMPLoopDirective &D)>
9817 SizeEmitter) {
9818 if (!CGF.HaveInsertPoint())
9819 return;
9820
9821 const bool OffloadingMandatory = !CGM.getLangOpts().OpenMPIsDevice &&
9822 CGM.getLangOpts().OpenMPOffloadMandatory;
9823
9824 assert((OffloadingMandatory || OutlinedFn) && "Invalid outlined function!");
9825
9826 const bool RequiresOuterTask = D.hasClausesOfKind<OMPDependClause>() ||
9827 D.hasClausesOfKind<OMPNowaitClause>() ||
9828 D.hasClausesOfKind<OMPInReductionClause>();
9829 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
9830 const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target);
9831 auto &&ArgsCodegen = [&CS, &CapturedVars](CodeGenFunction &CGF,
9832 PrePostActionTy &) {
9833 CGF.GenerateOpenMPCapturedVars(CS, CapturedVars);
9834 };
9835 emitInlinedDirective(CGF, OMPD_unknown, ArgsCodegen);
9836
9837 CodeGenFunction::OMPTargetDataInfo InputInfo;
9838 llvm::Value *MapTypesArray = nullptr;
9839 llvm::Value *MapNamesArray = nullptr;
9840 // Generate code for the host fallback function.
9841 auto &&FallbackGen = [this, OutlinedFn, &D, &CapturedVars, RequiresOuterTask,
9842 &CS, OffloadingMandatory](CodeGenFunction &CGF) {
9843 if (OffloadingMandatory) {
9844 CGF.Builder.CreateUnreachable();
9845 } else {
9846 if (RequiresOuterTask) {
9847 CapturedVars.clear();
9848 CGF.GenerateOpenMPCapturedVars(CS, CapturedVars);
9849 }
9850 emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars);
9851 }
9852 };
9853 // Fill up the pointer arrays and transfer execution to the device.
9854 auto &&ThenGen = [this, Device, OutlinedFnID, &D, &InputInfo, &MapTypesArray,
9855 &MapNamesArray, SizeEmitter,
9856 FallbackGen](CodeGenFunction &CGF, PrePostActionTy &) {
9857 if (Device.getInt() == OMPC_DEVICE_ancestor) {
9858 // Reverse offloading is not supported, so just execute on the host.
9859 FallbackGen(CGF);
9860 return;
9861 }
9862
9863 // On top of the arrays that were filled up, the target offloading call
9864 // takes as arguments the device id as well as the host pointer. The host
9865 // pointer is used by the runtime library to identify the current target
9866 // region, so it only has to be unique and not necessarily point to
9867 // anything. It could be the pointer to the outlined function that
9868 // implements the target region, but we aren't using that so that the
9869 // compiler doesn't need to keep that, and could therefore inline the host
9870 // function if proven worthwhile during optimization.
9871
9872 // From this point on, we need to have an ID of the target region defined.
9873 assert(OutlinedFnID && "Invalid outlined function ID!");
9874 (void)OutlinedFnID;
9875
9876 // Emit device ID if any.
9877 llvm::Value *DeviceID;
9878 if (Device.getPointer()) {
9879 assert((Device.getInt() == OMPC_DEVICE_unknown ||
9880 Device.getInt() == OMPC_DEVICE_device_num) &&
9881 "Expected device_num modifier.");
9882 llvm::Value *DevVal = CGF.EmitScalarExpr(Device.getPointer());
9883 DeviceID =
9884 CGF.Builder.CreateIntCast(DevVal, CGF.Int64Ty, /*isSigned=*/true);
9885 } else {
9886 DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF);
9887 }
9888
9889 // Emit the number of elements in the offloading arrays.
9890 llvm::Value *PointerNum =
9891 CGF.Builder.getInt32(InputInfo.NumberOfTargetItems);
9892
9893 // Return value of the runtime offloading call.
9894 llvm::Value *Return;
9895
9896 llvm::Value *NumTeams = emitNumTeamsForTargetDirective(CGF, D);
9897 llvm::Value *NumThreads = emitNumThreadsForTargetDirective(CGF, D);
9898
9899 // Source location for the ident struct
9900 llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc());
9901
9902 // Get tripcount for the target loop-based directive.
9903 llvm::Value *NumIterations =
9904 emitTargetNumIterationsCall(CGF, D, SizeEmitter);
9905
9906 llvm::Value *DynCGroupMem = CGF.Builder.getInt32(0);
9907 if (auto *DynMemClause = D.getSingleClause<OMPXDynCGroupMemClause>()) {
9908 CodeGenFunction::RunCleanupsScope DynCGroupMemScope(CGF);
9909 llvm::Value *DynCGroupMemVal = CGF.EmitScalarExpr(
9910 DynMemClause->getSize(), /*IgnoreResultAssign=*/true);
9911 DynCGroupMem = CGF.Builder.CreateIntCast(DynCGroupMemVal, CGF.Int32Ty,
9912 /*isSigned=*/false);
9913 }
9914
9915 llvm::Value *ZeroArray =
9916 llvm::Constant::getNullValue(llvm::ArrayType::get(CGF.CGM.Int32Ty, 3));
9917
9918 bool HasNoWait = D.hasClausesOfKind<OMPNowaitClause>();
9919 llvm::Value *Flags = CGF.Builder.getInt64(HasNoWait);
9920
9921 llvm::Value *NumTeams3D =
9922 CGF.Builder.CreateInsertValue(ZeroArray, NumTeams, {0});
9923 llvm::Value *NumThreads3D =
9924 CGF.Builder.CreateInsertValue(ZeroArray, NumThreads, {0});
9925
9926 // Arguments for the target kernel.
9927 SmallVector<llvm::Value *> KernelArgs{
9928 CGF.Builder.getInt32(/* Version */ 2),
9929 PointerNum,
9930 InputInfo.BasePointersArray.getPointer(),
9931 InputInfo.PointersArray.getPointer(),
9932 InputInfo.SizesArray.getPointer(),
9933 MapTypesArray,
9934 MapNamesArray,
9935 InputInfo.MappersArray.getPointer(),
9936 NumIterations,
9937 Flags,
9938 NumTeams3D,
9939 NumThreads3D,
9940 DynCGroupMem,
9941 };
9942
9943 // The target region is an outlined function launched by the runtime
9944 // via calls to __tgt_target_kernel().
9945 //
9946 // Note that on the host and CPU targets, the runtime implementation of
9947 // these calls simply call the outlined function without forking threads.
9948 // The outlined functions themselves have runtime calls to
9949 // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by
9950 // the compiler in emitTeamsCall() and emitParallelCall().
9951 //
9952 // In contrast, on the NVPTX target, the implementation of
9953 // __tgt_target_teams() launches a GPU kernel with the requested number
9954 // of teams and threads so no additional calls to the runtime are required.
9955 // Check the error code and execute the host version if required.
9956 CGF.Builder.restoreIP(OMPBuilder.emitTargetKernel(
9957 CGF.Builder, Return, RTLoc, DeviceID, NumTeams, NumThreads,
9958 OutlinedFnID, KernelArgs));
9959
9960 llvm::BasicBlock *OffloadFailedBlock =
9961 CGF.createBasicBlock("omp_offload.failed");
9962 llvm::BasicBlock *OffloadContBlock =
9963 CGF.createBasicBlock("omp_offload.cont");
9964 llvm::Value *Failed = CGF.Builder.CreateIsNotNull(Return);
9965 CGF.Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock);
9966
9967 CGF.EmitBlock(OffloadFailedBlock);
9968 FallbackGen(CGF);
9969
9970 CGF.EmitBranch(OffloadContBlock);
9971
9972 CGF.EmitBlock(OffloadContBlock, /*IsFinished=*/true);
9973 };
9974
9975 // Notify that the host version must be executed.
9976 auto &&ElseGen = [FallbackGen](CodeGenFunction &CGF, PrePostActionTy &) {
9977 FallbackGen(CGF);
9978 };
9979
9980 auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray,
9981 &MapNamesArray, &CapturedVars, RequiresOuterTask,
9982 &CS](CodeGenFunction &CGF, PrePostActionTy &) {
9983 // Fill up the arrays with all the captured variables.
9984 MappableExprsHandler::MapCombinedInfoTy CombinedInfo;
9985
9986 // Get mappable expression information.
9987 MappableExprsHandler MEHandler(D, CGF);
9988 llvm::DenseMap<llvm::Value *, llvm::Value *> LambdaPointers;
9989 llvm::DenseSet<CanonicalDeclPtr<const Decl>> MappedVarSet;
9990
9991 auto RI = CS.getCapturedRecordDecl()->field_begin();
9992 auto *CV = CapturedVars.begin();
9993 for (CapturedStmt::const_capture_iterator CI = CS.capture_begin(),
9994 CE = CS.capture_end();
9995 CI != CE; ++CI, ++RI, ++CV) {
9996 MappableExprsHandler::MapCombinedInfoTy CurInfo;
9997 MappableExprsHandler::StructRangeInfoTy PartialStruct;
9998
9999 // VLA sizes are passed to the outlined region by copy and do not have map
10000 // information associated.
10001 if (CI->capturesVariableArrayType()) {
10002 CurInfo.Exprs.push_back(nullptr);
10003 CurInfo.BasePointers.push_back(*CV);
10004 CurInfo.Pointers.push_back(*CV);
10005 CurInfo.Sizes.push_back(CGF.Builder.CreateIntCast(
10006 CGF.getTypeSize(RI->getType()), CGF.Int64Ty, /*isSigned=*/true));
10007 // Copy to the device as an argument. No need to retrieve it.
10008 CurInfo.Types.push_back(
10009 OpenMPOffloadMappingFlags::OMP_MAP_LITERAL |
10010 OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM |
10011 OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT);
10012 CurInfo.Mappers.push_back(nullptr);
10013 } else {
10014 // If we have any information in the map clause, we use it, otherwise we
10015 // just do a default mapping.
10016 MEHandler.generateInfoForCapture(CI, *CV, CurInfo, PartialStruct);
10017 if (!CI->capturesThis())
10018 MappedVarSet.insert(CI->getCapturedVar());
10019 else
10020 MappedVarSet.insert(nullptr);
10021 if (CurInfo.BasePointers.empty() && !PartialStruct.Base.isValid())
10022 MEHandler.generateDefaultMapInfo(*CI, **RI, *CV, CurInfo);
10023 // Generate correct mapping for variables captured by reference in
10024 // lambdas.
10025 if (CI->capturesVariable())
10026 MEHandler.generateInfoForLambdaCaptures(CI->getCapturedVar(), *CV,
10027 CurInfo, LambdaPointers);
10028 }
10029 // We expect to have at least an element of information for this capture.
10030 assert((!CurInfo.BasePointers.empty() || PartialStruct.Base.isValid()) &&
10031 "Non-existing map pointer for capture!");
10032 assert(CurInfo.BasePointers.size() == CurInfo.Pointers.size() &&
10033 CurInfo.BasePointers.size() == CurInfo.Sizes.size() &&
10034 CurInfo.BasePointers.size() == CurInfo.Types.size() &&
10035 CurInfo.BasePointers.size() == CurInfo.Mappers.size() &&
10036 "Inconsistent map information sizes!");
10037
10038 // If there is an entry in PartialStruct it means we have a struct with
10039 // individual members mapped. Emit an extra combined entry.
10040 if (PartialStruct.Base.isValid()) {
10041 CombinedInfo.append(PartialStruct.PreliminaryMapData);
10042 MEHandler.emitCombinedEntry(
10043 CombinedInfo, CurInfo.Types, PartialStruct, nullptr,
10044 !PartialStruct.PreliminaryMapData.BasePointers.empty());
10045 }
10046
10047 // We need to append the results of this capture to what we already have.
10048 CombinedInfo.append(CurInfo);
10049 }
10050 // Adjust MEMBER_OF flags for the lambdas captures.
10051 MEHandler.adjustMemberOfForLambdaCaptures(
10052 LambdaPointers, CombinedInfo.BasePointers, CombinedInfo.Pointers,
10053 CombinedInfo.Types);
10054 // Map any list items in a map clause that were not captures because they
10055 // weren't referenced within the construct.
10056 MEHandler.generateAllInfo(CombinedInfo, MappedVarSet);
10057
10058 CGOpenMPRuntime::TargetDataInfo Info;
10059 // Fill up the arrays and create the arguments.
10060 emitOffloadingArrays(CGF, CombinedInfo, Info, OMPBuilder);
10061 bool EmitDebug =
10062 CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo;
10063 OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, Info.RTArgs, Info,
10064 EmitDebug,
10065 /*ForEndCall=*/false);
10066
10067 InputInfo.NumberOfTargetItems = Info.NumberOfPtrs;
10068 InputInfo.BasePointersArray = Address(Info.RTArgs.BasePointersArray,
10069 CGF.VoidPtrTy, CGM.getPointerAlign());
10070 InputInfo.PointersArray = Address(Info.RTArgs.PointersArray, CGF.VoidPtrTy,
10071 CGM.getPointerAlign());
10072 InputInfo.SizesArray =
10073 Address(Info.RTArgs.SizesArray, CGF.Int64Ty, CGM.getPointerAlign());
10074 InputInfo.MappersArray =
10075 Address(Info.RTArgs.MappersArray, CGF.VoidPtrTy, CGM.getPointerAlign());
10076 MapTypesArray = Info.RTArgs.MapTypesArray;
10077 MapNamesArray = Info.RTArgs.MapNamesArray;
10078 if (RequiresOuterTask)
10079 CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo);
10080 else
10081 emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen);
10082 };
10083
10084 auto &&TargetElseGen = [this, &ElseGen, &D, RequiresOuterTask](
10085 CodeGenFunction &CGF, PrePostActionTy &) {
10086 if (RequiresOuterTask) {
10087 CodeGenFunction::OMPTargetDataInfo InputInfo;
10088 CGF.EmitOMPTargetTaskBasedDirective(D, ElseGen, InputInfo);
10089 } else {
10090 emitInlinedDirective(CGF, D.getDirectiveKind(), ElseGen);
10091 }
10092 };
10093
10094 // If we have a target function ID it means that we need to support
10095 // offloading, otherwise, just execute on the host. We need to execute on host
10096 // regardless of the conditional in the if clause if, e.g., the user do not
10097 // specify target triples.
10098 if (OutlinedFnID) {
10099 if (IfCond) {
10100 emitIfClause(CGF, IfCond, TargetThenGen, TargetElseGen);
10101 } else {
10102 RegionCodeGenTy ThenRCG(TargetThenGen);
10103 ThenRCG(CGF);
10104 }
10105 } else {
10106 RegionCodeGenTy ElseRCG(TargetElseGen);
10107 ElseRCG(CGF);
10108 }
10109 }
10110
scanForTargetRegionsFunctions(const Stmt * S,StringRef ParentName)10111 void CGOpenMPRuntime::scanForTargetRegionsFunctions(const Stmt *S,
10112 StringRef ParentName) {
10113 if (!S)
10114 return;
10115
10116 // Codegen OMP target directives that offload compute to the device.
10117 bool RequiresDeviceCodegen =
10118 isa<OMPExecutableDirective>(S) &&
10119 isOpenMPTargetExecutionDirective(
10120 cast<OMPExecutableDirective>(S)->getDirectiveKind());
10121
10122 if (RequiresDeviceCodegen) {
10123 const auto &E = *cast<OMPExecutableDirective>(S);
10124 auto EntryInfo =
10125 getTargetEntryUniqueInfo(CGM.getContext(), E.getBeginLoc(), ParentName);
10126
10127 // Is this a target region that should not be emitted as an entry point? If
10128 // so just signal we are done with this target region.
10129 if (!OffloadEntriesInfoManager.hasTargetRegionEntryInfo(EntryInfo))
10130 return;
10131
10132 switch (E.getDirectiveKind()) {
10133 case OMPD_target:
10134 CodeGenFunction::EmitOMPTargetDeviceFunction(CGM, ParentName,
10135 cast<OMPTargetDirective>(E));
10136 break;
10137 case OMPD_target_parallel:
10138 CodeGenFunction::EmitOMPTargetParallelDeviceFunction(
10139 CGM, ParentName, cast<OMPTargetParallelDirective>(E));
10140 break;
10141 case OMPD_target_teams:
10142 CodeGenFunction::EmitOMPTargetTeamsDeviceFunction(
10143 CGM, ParentName, cast<OMPTargetTeamsDirective>(E));
10144 break;
10145 case OMPD_target_teams_distribute:
10146 CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction(
10147 CGM, ParentName, cast<OMPTargetTeamsDistributeDirective>(E));
10148 break;
10149 case OMPD_target_teams_distribute_simd:
10150 CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction(
10151 CGM, ParentName, cast<OMPTargetTeamsDistributeSimdDirective>(E));
10152 break;
10153 case OMPD_target_parallel_for:
10154 CodeGenFunction::EmitOMPTargetParallelForDeviceFunction(
10155 CGM, ParentName, cast<OMPTargetParallelForDirective>(E));
10156 break;
10157 case OMPD_target_parallel_for_simd:
10158 CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction(
10159 CGM, ParentName, cast<OMPTargetParallelForSimdDirective>(E));
10160 break;
10161 case OMPD_target_simd:
10162 CodeGenFunction::EmitOMPTargetSimdDeviceFunction(
10163 CGM, ParentName, cast<OMPTargetSimdDirective>(E));
10164 break;
10165 case OMPD_target_teams_distribute_parallel_for:
10166 CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction(
10167 CGM, ParentName,
10168 cast<OMPTargetTeamsDistributeParallelForDirective>(E));
10169 break;
10170 case OMPD_target_teams_distribute_parallel_for_simd:
10171 CodeGenFunction::
10172 EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction(
10173 CGM, ParentName,
10174 cast<OMPTargetTeamsDistributeParallelForSimdDirective>(E));
10175 break;
10176 case OMPD_parallel:
10177 case OMPD_for:
10178 case OMPD_parallel_for:
10179 case OMPD_parallel_master:
10180 case OMPD_parallel_sections:
10181 case OMPD_for_simd:
10182 case OMPD_parallel_for_simd:
10183 case OMPD_cancel:
10184 case OMPD_cancellation_point:
10185 case OMPD_ordered:
10186 case OMPD_threadprivate:
10187 case OMPD_allocate:
10188 case OMPD_task:
10189 case OMPD_simd:
10190 case OMPD_tile:
10191 case OMPD_unroll:
10192 case OMPD_sections:
10193 case OMPD_section:
10194 case OMPD_single:
10195 case OMPD_master:
10196 case OMPD_critical:
10197 case OMPD_taskyield:
10198 case OMPD_barrier:
10199 case OMPD_taskwait:
10200 case OMPD_taskgroup:
10201 case OMPD_atomic:
10202 case OMPD_flush:
10203 case OMPD_depobj:
10204 case OMPD_scan:
10205 case OMPD_teams:
10206 case OMPD_target_data:
10207 case OMPD_target_exit_data:
10208 case OMPD_target_enter_data:
10209 case OMPD_distribute:
10210 case OMPD_distribute_simd:
10211 case OMPD_distribute_parallel_for:
10212 case OMPD_distribute_parallel_for_simd:
10213 case OMPD_teams_distribute:
10214 case OMPD_teams_distribute_simd:
10215 case OMPD_teams_distribute_parallel_for:
10216 case OMPD_teams_distribute_parallel_for_simd:
10217 case OMPD_target_update:
10218 case OMPD_declare_simd:
10219 case OMPD_declare_variant:
10220 case OMPD_begin_declare_variant:
10221 case OMPD_end_declare_variant:
10222 case OMPD_declare_target:
10223 case OMPD_end_declare_target:
10224 case OMPD_declare_reduction:
10225 case OMPD_declare_mapper:
10226 case OMPD_taskloop:
10227 case OMPD_taskloop_simd:
10228 case OMPD_master_taskloop:
10229 case OMPD_master_taskloop_simd:
10230 case OMPD_parallel_master_taskloop:
10231 case OMPD_parallel_master_taskloop_simd:
10232 case OMPD_requires:
10233 case OMPD_metadirective:
10234 case OMPD_unknown:
10235 default:
10236 llvm_unreachable("Unknown target directive for OpenMP device codegen.");
10237 }
10238 return;
10239 }
10240
10241 if (const auto *E = dyn_cast<OMPExecutableDirective>(S)) {
10242 if (!E->hasAssociatedStmt() || !E->getAssociatedStmt())
10243 return;
10244
10245 scanForTargetRegionsFunctions(E->getRawStmt(), ParentName);
10246 return;
10247 }
10248
10249 // If this is a lambda function, look into its body.
10250 if (const auto *L = dyn_cast<LambdaExpr>(S))
10251 S = L->getBody();
10252
10253 // Keep looking for target regions recursively.
10254 for (const Stmt *II : S->children())
10255 scanForTargetRegionsFunctions(II, ParentName);
10256 }
10257
isAssumedToBeNotEmitted(const ValueDecl * VD,bool IsDevice)10258 static bool isAssumedToBeNotEmitted(const ValueDecl *VD, bool IsDevice) {
10259 std::optional<OMPDeclareTargetDeclAttr::DevTypeTy> DevTy =
10260 OMPDeclareTargetDeclAttr::getDeviceType(VD);
10261 if (!DevTy)
10262 return false;
10263 // Do not emit device_type(nohost) functions for the host.
10264 if (!IsDevice && DevTy == OMPDeclareTargetDeclAttr::DT_NoHost)
10265 return true;
10266 // Do not emit device_type(host) functions for the device.
10267 if (IsDevice && DevTy == OMPDeclareTargetDeclAttr::DT_Host)
10268 return true;
10269 return false;
10270 }
10271
emitTargetFunctions(GlobalDecl GD)10272 bool CGOpenMPRuntime::emitTargetFunctions(GlobalDecl GD) {
10273 // If emitting code for the host, we do not process FD here. Instead we do
10274 // the normal code generation.
10275 if (!CGM.getLangOpts().OpenMPIsDevice) {
10276 if (const auto *FD = dyn_cast<FunctionDecl>(GD.getDecl()))
10277 if (isAssumedToBeNotEmitted(cast<ValueDecl>(FD),
10278 CGM.getLangOpts().OpenMPIsDevice))
10279 return true;
10280 return false;
10281 }
10282
10283 const ValueDecl *VD = cast<ValueDecl>(GD.getDecl());
10284 // Try to detect target regions in the function.
10285 if (const auto *FD = dyn_cast<FunctionDecl>(VD)) {
10286 StringRef Name = CGM.getMangledName(GD);
10287 scanForTargetRegionsFunctions(FD->getBody(), Name);
10288 if (isAssumedToBeNotEmitted(cast<ValueDecl>(FD),
10289 CGM.getLangOpts().OpenMPIsDevice))
10290 return true;
10291 }
10292
10293 // Do not to emit function if it is not marked as declare target.
10294 return !OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD) &&
10295 AlreadyEmittedTargetDecls.count(VD) == 0;
10296 }
10297
emitTargetGlobalVariable(GlobalDecl GD)10298 bool CGOpenMPRuntime::emitTargetGlobalVariable(GlobalDecl GD) {
10299 if (isAssumedToBeNotEmitted(cast<ValueDecl>(GD.getDecl()),
10300 CGM.getLangOpts().OpenMPIsDevice))
10301 return true;
10302
10303 if (!CGM.getLangOpts().OpenMPIsDevice)
10304 return false;
10305
10306 // Check if there are Ctors/Dtors in this declaration and look for target
10307 // regions in it. We use the complete variant to produce the kernel name
10308 // mangling.
10309 QualType RDTy = cast<VarDecl>(GD.getDecl())->getType();
10310 if (const auto *RD = RDTy->getBaseElementTypeUnsafe()->getAsCXXRecordDecl()) {
10311 for (const CXXConstructorDecl *Ctor : RD->ctors()) {
10312 StringRef ParentName =
10313 CGM.getMangledName(GlobalDecl(Ctor, Ctor_Complete));
10314 scanForTargetRegionsFunctions(Ctor->getBody(), ParentName);
10315 }
10316 if (const CXXDestructorDecl *Dtor = RD->getDestructor()) {
10317 StringRef ParentName =
10318 CGM.getMangledName(GlobalDecl(Dtor, Dtor_Complete));
10319 scanForTargetRegionsFunctions(Dtor->getBody(), ParentName);
10320 }
10321 }
10322
10323 // Do not to emit variable if it is not marked as declare target.
10324 std::optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res =
10325 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(
10326 cast<VarDecl>(GD.getDecl()));
10327 if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link ||
10328 ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
10329 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
10330 HasRequiresUnifiedSharedMemory)) {
10331 DeferredGlobalVariables.insert(cast<VarDecl>(GD.getDecl()));
10332 return true;
10333 }
10334 return false;
10335 }
10336
registerTargetGlobalVariable(const VarDecl * VD,llvm::Constant * Addr)10337 void CGOpenMPRuntime::registerTargetGlobalVariable(const VarDecl *VD,
10338 llvm::Constant *Addr) {
10339 if (CGM.getLangOpts().OMPTargetTriples.empty() &&
10340 !CGM.getLangOpts().OpenMPIsDevice)
10341 return;
10342
10343 // If we have host/nohost variables, they do not need to be registered.
10344 std::optional<OMPDeclareTargetDeclAttr::DevTypeTy> DevTy =
10345 OMPDeclareTargetDeclAttr::getDeviceType(VD);
10346 if (DevTy && *DevTy != OMPDeclareTargetDeclAttr::DT_Any)
10347 return;
10348
10349 std::optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res =
10350 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD);
10351 if (!Res) {
10352 if (CGM.getLangOpts().OpenMPIsDevice) {
10353 // Register non-target variables being emitted in device code (debug info
10354 // may cause this).
10355 StringRef VarName = CGM.getMangledName(VD);
10356 EmittedNonTargetVariables.try_emplace(VarName, Addr);
10357 }
10358 return;
10359 }
10360 // Register declare target variables.
10361 llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind Flags;
10362 StringRef VarName;
10363 int64_t VarSize;
10364 llvm::GlobalValue::LinkageTypes Linkage;
10365
10366 if ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
10367 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
10368 !HasRequiresUnifiedSharedMemory) {
10369 Flags = llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo;
10370 VarName = CGM.getMangledName(VD);
10371 if (VD->hasDefinition(CGM.getContext()) != VarDecl::DeclarationOnly) {
10372 VarSize =
10373 CGM.getContext().getTypeSizeInChars(VD->getType()).getQuantity();
10374 assert(VarSize != 0 && "Expected non-zero size of the variable");
10375 } else {
10376 VarSize = 0;
10377 }
10378 Linkage = CGM.getLLVMLinkageVarDefinition(VD, /*IsConstant=*/false);
10379 // Temp solution to prevent optimizations of the internal variables.
10380 if (CGM.getLangOpts().OpenMPIsDevice && !VD->isExternallyVisible()) {
10381 // Do not create a "ref-variable" if the original is not also available
10382 // on the host.
10383 if (!OffloadEntriesInfoManager.hasDeviceGlobalVarEntryInfo(VarName))
10384 return;
10385 std::string RefName = getName({VarName, "ref"});
10386 if (!CGM.GetGlobalValue(RefName)) {
10387 llvm::Constant *AddrRef =
10388 OMPBuilder.getOrCreateInternalVariable(Addr->getType(), RefName);
10389 auto *GVAddrRef = cast<llvm::GlobalVariable>(AddrRef);
10390 GVAddrRef->setConstant(/*Val=*/true);
10391 GVAddrRef->setLinkage(llvm::GlobalValue::InternalLinkage);
10392 GVAddrRef->setInitializer(Addr);
10393 CGM.addCompilerUsedGlobal(GVAddrRef);
10394 }
10395 }
10396 } else {
10397 assert(((*Res == OMPDeclareTargetDeclAttr::MT_Link) ||
10398 ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
10399 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
10400 HasRequiresUnifiedSharedMemory)) &&
10401 "Declare target attribute must link or to with unified memory.");
10402 if (*Res == OMPDeclareTargetDeclAttr::MT_Link)
10403 Flags = llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink;
10404 else
10405 Flags = llvm::OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo;
10406
10407 if (CGM.getLangOpts().OpenMPIsDevice) {
10408 VarName = Addr->getName();
10409 Addr = nullptr;
10410 } else {
10411 VarName = getAddrOfDeclareTargetVar(VD).getName();
10412 Addr = cast<llvm::Constant>(getAddrOfDeclareTargetVar(VD).getPointer());
10413 }
10414 VarSize = CGM.getPointerSize().getQuantity();
10415 Linkage = llvm::GlobalValue::WeakAnyLinkage;
10416 }
10417
10418 OffloadEntriesInfoManager.registerDeviceGlobalVarEntryInfo(
10419 VarName, Addr, VarSize, Flags, Linkage);
10420 }
10421
emitTargetGlobal(GlobalDecl GD)10422 bool CGOpenMPRuntime::emitTargetGlobal(GlobalDecl GD) {
10423 if (isa<FunctionDecl>(GD.getDecl()) ||
10424 isa<OMPDeclareReductionDecl>(GD.getDecl()))
10425 return emitTargetFunctions(GD);
10426
10427 return emitTargetGlobalVariable(GD);
10428 }
10429
emitDeferredTargetDecls() const10430 void CGOpenMPRuntime::emitDeferredTargetDecls() const {
10431 for (const VarDecl *VD : DeferredGlobalVariables) {
10432 std::optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res =
10433 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD);
10434 if (!Res)
10435 continue;
10436 if ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
10437 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
10438 !HasRequiresUnifiedSharedMemory) {
10439 CGM.EmitGlobal(VD);
10440 } else {
10441 assert((*Res == OMPDeclareTargetDeclAttr::MT_Link ||
10442 ((*Res == OMPDeclareTargetDeclAttr::MT_To ||
10443 *Res == OMPDeclareTargetDeclAttr::MT_Enter) &&
10444 HasRequiresUnifiedSharedMemory)) &&
10445 "Expected link clause or to clause with unified memory.");
10446 (void)CGM.getOpenMPRuntime().getAddrOfDeclareTargetVar(VD);
10447 }
10448 }
10449 }
10450
adjustTargetSpecificDataForLambdas(CodeGenFunction & CGF,const OMPExecutableDirective & D) const10451 void CGOpenMPRuntime::adjustTargetSpecificDataForLambdas(
10452 CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
10453 assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
10454 " Expected target-based directive.");
10455 }
10456
processRequiresDirective(const OMPRequiresDecl * D)10457 void CGOpenMPRuntime::processRequiresDirective(const OMPRequiresDecl *D) {
10458 for (const OMPClause *Clause : D->clauselists()) {
10459 if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
10460 HasRequiresUnifiedSharedMemory = true;
10461 OMPBuilder.Config.setHasRequiresUnifiedSharedMemory(true);
10462 } else if (const auto *AC =
10463 dyn_cast<OMPAtomicDefaultMemOrderClause>(Clause)) {
10464 switch (AC->getAtomicDefaultMemOrderKind()) {
10465 case OMPC_ATOMIC_DEFAULT_MEM_ORDER_acq_rel:
10466 RequiresAtomicOrdering = llvm::AtomicOrdering::AcquireRelease;
10467 break;
10468 case OMPC_ATOMIC_DEFAULT_MEM_ORDER_seq_cst:
10469 RequiresAtomicOrdering = llvm::AtomicOrdering::SequentiallyConsistent;
10470 break;
10471 case OMPC_ATOMIC_DEFAULT_MEM_ORDER_relaxed:
10472 RequiresAtomicOrdering = llvm::AtomicOrdering::Monotonic;
10473 break;
10474 case OMPC_ATOMIC_DEFAULT_MEM_ORDER_unknown:
10475 break;
10476 }
10477 }
10478 }
10479 }
10480
getDefaultMemoryOrdering() const10481 llvm::AtomicOrdering CGOpenMPRuntime::getDefaultMemoryOrdering() const {
10482 return RequiresAtomicOrdering;
10483 }
10484
hasAllocateAttributeForGlobalVar(const VarDecl * VD,LangAS & AS)10485 bool CGOpenMPRuntime::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
10486 LangAS &AS) {
10487 if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
10488 return false;
10489 const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
10490 switch(A->getAllocatorType()) {
10491 case OMPAllocateDeclAttr::OMPNullMemAlloc:
10492 case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
10493 // Not supported, fallback to the default mem space.
10494 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
10495 case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
10496 case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
10497 case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
10498 case OMPAllocateDeclAttr::OMPThreadMemAlloc:
10499 case OMPAllocateDeclAttr::OMPConstMemAlloc:
10500 case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
10501 AS = LangAS::Default;
10502 return true;
10503 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
10504 llvm_unreachable("Expected predefined allocator for the variables with the "
10505 "static storage.");
10506 }
10507 return false;
10508 }
10509
hasRequiresUnifiedSharedMemory() const10510 bool CGOpenMPRuntime::hasRequiresUnifiedSharedMemory() const {
10511 return HasRequiresUnifiedSharedMemory;
10512 }
10513
DisableAutoDeclareTargetRAII(CodeGenModule & CGM)10514 CGOpenMPRuntime::DisableAutoDeclareTargetRAII::DisableAutoDeclareTargetRAII(
10515 CodeGenModule &CGM)
10516 : CGM(CGM) {
10517 if (CGM.getLangOpts().OpenMPIsDevice) {
10518 SavedShouldMarkAsGlobal = CGM.getOpenMPRuntime().ShouldMarkAsGlobal;
10519 CGM.getOpenMPRuntime().ShouldMarkAsGlobal = false;
10520 }
10521 }
10522
~DisableAutoDeclareTargetRAII()10523 CGOpenMPRuntime::DisableAutoDeclareTargetRAII::~DisableAutoDeclareTargetRAII() {
10524 if (CGM.getLangOpts().OpenMPIsDevice)
10525 CGM.getOpenMPRuntime().ShouldMarkAsGlobal = SavedShouldMarkAsGlobal;
10526 }
10527
markAsGlobalTarget(GlobalDecl GD)10528 bool CGOpenMPRuntime::markAsGlobalTarget(GlobalDecl GD) {
10529 if (!CGM.getLangOpts().OpenMPIsDevice || !ShouldMarkAsGlobal)
10530 return true;
10531
10532 const auto *D = cast<FunctionDecl>(GD.getDecl());
10533 // Do not to emit function if it is marked as declare target as it was already
10534 // emitted.
10535 if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(D)) {
10536 if (D->hasBody() && AlreadyEmittedTargetDecls.count(D) == 0) {
10537 if (auto *F = dyn_cast_or_null<llvm::Function>(
10538 CGM.GetGlobalValue(CGM.getMangledName(GD))))
10539 return !F->isDeclaration();
10540 return false;
10541 }
10542 return true;
10543 }
10544
10545 return !AlreadyEmittedTargetDecls.insert(D).second;
10546 }
10547
emitRequiresDirectiveRegFun()10548 llvm::Function *CGOpenMPRuntime::emitRequiresDirectiveRegFun() {
10549 // If we don't have entries or if we are emitting code for the device, we
10550 // don't need to do anything.
10551 if (CGM.getLangOpts().OMPTargetTriples.empty() ||
10552 CGM.getLangOpts().OpenMPSimd || CGM.getLangOpts().OpenMPIsDevice ||
10553 (OffloadEntriesInfoManager.empty() &&
10554 !HasEmittedDeclareTargetRegion &&
10555 !HasEmittedTargetRegion))
10556 return nullptr;
10557
10558 // Create and register the function that handles the requires directives.
10559 ASTContext &C = CGM.getContext();
10560
10561 llvm::Function *RequiresRegFn;
10562 {
10563 CodeGenFunction CGF(CGM);
10564 const auto &FI = CGM.getTypes().arrangeNullaryFunction();
10565 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI);
10566 std::string ReqName = getName({"omp_offloading", "requires_reg"});
10567 RequiresRegFn = CGM.CreateGlobalInitOrCleanUpFunction(FTy, ReqName, FI);
10568 CGF.StartFunction(GlobalDecl(), C.VoidTy, RequiresRegFn, FI, {});
10569 OpenMPOffloadingRequiresDirFlags Flags = OMP_REQ_NONE;
10570 // TODO: check for other requires clauses.
10571 // The requires directive takes effect only when a target region is
10572 // present in the compilation unit. Otherwise it is ignored and not
10573 // passed to the runtime. This avoids the runtime from throwing an error
10574 // for mismatching requires clauses across compilation units that don't
10575 // contain at least 1 target region.
10576 assert((HasEmittedTargetRegion ||
10577 HasEmittedDeclareTargetRegion ||
10578 !OffloadEntriesInfoManager.empty()) &&
10579 "Target or declare target region expected.");
10580 if (HasRequiresUnifiedSharedMemory)
10581 Flags = OMP_REQ_UNIFIED_SHARED_MEMORY;
10582 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
10583 CGM.getModule(), OMPRTL___tgt_register_requires),
10584 llvm::ConstantInt::get(CGM.Int64Ty, Flags));
10585 CGF.FinishFunction();
10586 }
10587 return RequiresRegFn;
10588 }
10589
emitTeamsCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,SourceLocation Loc,llvm::Function * OutlinedFn,ArrayRef<llvm::Value * > CapturedVars)10590 void CGOpenMPRuntime::emitTeamsCall(CodeGenFunction &CGF,
10591 const OMPExecutableDirective &D,
10592 SourceLocation Loc,
10593 llvm::Function *OutlinedFn,
10594 ArrayRef<llvm::Value *> CapturedVars) {
10595 if (!CGF.HaveInsertPoint())
10596 return;
10597
10598 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
10599 CodeGenFunction::RunCleanupsScope Scope(CGF);
10600
10601 // Build call __kmpc_fork_teams(loc, n, microtask, var1, .., varn);
10602 llvm::Value *Args[] = {
10603 RTLoc,
10604 CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars
10605 CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy())};
10606 llvm::SmallVector<llvm::Value *, 16> RealArgs;
10607 RealArgs.append(std::begin(Args), std::end(Args));
10608 RealArgs.append(CapturedVars.begin(), CapturedVars.end());
10609
10610 llvm::FunctionCallee RTLFn = OMPBuilder.getOrCreateRuntimeFunction(
10611 CGM.getModule(), OMPRTL___kmpc_fork_teams);
10612 CGF.EmitRuntimeCall(RTLFn, RealArgs);
10613 }
10614
emitNumTeamsClause(CodeGenFunction & CGF,const Expr * NumTeams,const Expr * ThreadLimit,SourceLocation Loc)10615 void CGOpenMPRuntime::emitNumTeamsClause(CodeGenFunction &CGF,
10616 const Expr *NumTeams,
10617 const Expr *ThreadLimit,
10618 SourceLocation Loc) {
10619 if (!CGF.HaveInsertPoint())
10620 return;
10621
10622 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
10623
10624 llvm::Value *NumTeamsVal =
10625 NumTeams
10626 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(NumTeams),
10627 CGF.CGM.Int32Ty, /* isSigned = */ true)
10628 : CGF.Builder.getInt32(0);
10629
10630 llvm::Value *ThreadLimitVal =
10631 ThreadLimit
10632 ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(ThreadLimit),
10633 CGF.CGM.Int32Ty, /* isSigned = */ true)
10634 : CGF.Builder.getInt32(0);
10635
10636 // Build call __kmpc_push_num_teamss(&loc, global_tid, num_teams, thread_limit)
10637 llvm::Value *PushNumTeamsArgs[] = {RTLoc, getThreadID(CGF, Loc), NumTeamsVal,
10638 ThreadLimitVal};
10639 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
10640 CGM.getModule(), OMPRTL___kmpc_push_num_teams),
10641 PushNumTeamsArgs);
10642 }
10643
emitTargetDataCalls(CodeGenFunction & CGF,const OMPExecutableDirective & D,const Expr * IfCond,const Expr * Device,const RegionCodeGenTy & CodeGen,CGOpenMPRuntime::TargetDataInfo & Info)10644 void CGOpenMPRuntime::emitTargetDataCalls(
10645 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
10646 const Expr *Device, const RegionCodeGenTy &CodeGen,
10647 CGOpenMPRuntime::TargetDataInfo &Info) {
10648 if (!CGF.HaveInsertPoint())
10649 return;
10650
10651 // Action used to replace the default codegen action and turn privatization
10652 // off.
10653 PrePostActionTy NoPrivAction;
10654
10655 // Generate the code for the opening of the data environment. Capture all the
10656 // arguments of the runtime call by reference because they are used in the
10657 // closing of the region.
10658 auto &&BeginThenGen = [this, &D, Device, &Info,
10659 &CodeGen](CodeGenFunction &CGF, PrePostActionTy &) {
10660 // Fill up the arrays with all the mapped variables.
10661 MappableExprsHandler::MapCombinedInfoTy CombinedInfo;
10662
10663 // Get map clause information.
10664 MappableExprsHandler MEHandler(D, CGF);
10665 MEHandler.generateAllInfo(CombinedInfo);
10666
10667 // Fill up the arrays and create the arguments.
10668 emitOffloadingArrays(CGF, CombinedInfo, Info, OMPBuilder,
10669 /*IsNonContiguous=*/true);
10670
10671 llvm::OpenMPIRBuilder::TargetDataRTArgs RTArgs;
10672 bool EmitDebug =
10673 CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo;
10674 OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, RTArgs, Info,
10675 EmitDebug);
10676
10677 // Emit device ID if any.
10678 llvm::Value *DeviceID = nullptr;
10679 if (Device) {
10680 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
10681 CGF.Int64Ty, /*isSigned=*/true);
10682 } else {
10683 DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF);
10684 }
10685
10686 // Emit the number of elements in the offloading arrays.
10687 llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs);
10688 //
10689 // Source location for the ident struct
10690 llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc());
10691
10692 llvm::Value *OffloadingArgs[] = {RTLoc,
10693 DeviceID,
10694 PointerNum,
10695 RTArgs.BasePointersArray,
10696 RTArgs.PointersArray,
10697 RTArgs.SizesArray,
10698 RTArgs.MapTypesArray,
10699 RTArgs.MapNamesArray,
10700 RTArgs.MappersArray};
10701 CGF.EmitRuntimeCall(
10702 OMPBuilder.getOrCreateRuntimeFunction(
10703 CGM.getModule(), OMPRTL___tgt_target_data_begin_mapper),
10704 OffloadingArgs);
10705
10706 // If device pointer privatization is required, emit the body of the region
10707 // here. It will have to be duplicated: with and without privatization.
10708 if (!Info.CaptureDeviceAddrMap.empty())
10709 CodeGen(CGF);
10710 };
10711
10712 // Generate code for the closing of the data region.
10713 auto &&EndThenGen = [this, Device, &Info, &D](CodeGenFunction &CGF,
10714 PrePostActionTy &) {
10715 assert(Info.isValid() && "Invalid data environment closing arguments.");
10716
10717 llvm::OpenMPIRBuilder::TargetDataRTArgs RTArgs;
10718 bool EmitDebug =
10719 CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo;
10720 OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, RTArgs, Info,
10721 EmitDebug,
10722 /*ForEndCall=*/true);
10723
10724 // Emit device ID if any.
10725 llvm::Value *DeviceID = nullptr;
10726 if (Device) {
10727 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
10728 CGF.Int64Ty, /*isSigned=*/true);
10729 } else {
10730 DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF);
10731 }
10732
10733 // Emit the number of elements in the offloading arrays.
10734 llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs);
10735
10736 // Source location for the ident struct
10737 llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc());
10738
10739 llvm::Value *OffloadingArgs[] = {RTLoc,
10740 DeviceID,
10741 PointerNum,
10742 RTArgs.BasePointersArray,
10743 RTArgs.PointersArray,
10744 RTArgs.SizesArray,
10745 RTArgs.MapTypesArray,
10746 RTArgs.MapNamesArray,
10747 RTArgs.MappersArray};
10748 CGF.EmitRuntimeCall(
10749 OMPBuilder.getOrCreateRuntimeFunction(
10750 CGM.getModule(), OMPRTL___tgt_target_data_end_mapper),
10751 OffloadingArgs);
10752 };
10753
10754 // If we need device pointer privatization, we need to emit the body of the
10755 // region with no privatization in the 'else' branch of the conditional.
10756 // Otherwise, we don't have to do anything.
10757 auto &&BeginElseGen = [&Info, &CodeGen, &NoPrivAction](CodeGenFunction &CGF,
10758 PrePostActionTy &) {
10759 if (!Info.CaptureDeviceAddrMap.empty()) {
10760 CodeGen.setAction(NoPrivAction);
10761 CodeGen(CGF);
10762 }
10763 };
10764
10765 // We don't have to do anything to close the region if the if clause evaluates
10766 // to false.
10767 auto &&EndElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {};
10768
10769 if (IfCond) {
10770 emitIfClause(CGF, IfCond, BeginThenGen, BeginElseGen);
10771 } else {
10772 RegionCodeGenTy RCG(BeginThenGen);
10773 RCG(CGF);
10774 }
10775
10776 // If we don't require privatization of device pointers, we emit the body in
10777 // between the runtime calls. This avoids duplicating the body code.
10778 if (Info.CaptureDeviceAddrMap.empty()) {
10779 CodeGen.setAction(NoPrivAction);
10780 CodeGen(CGF);
10781 }
10782
10783 if (IfCond) {
10784 emitIfClause(CGF, IfCond, EndThenGen, EndElseGen);
10785 } else {
10786 RegionCodeGenTy RCG(EndThenGen);
10787 RCG(CGF);
10788 }
10789 }
10790
emitTargetDataStandAloneCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,const Expr * IfCond,const Expr * Device)10791 void CGOpenMPRuntime::emitTargetDataStandAloneCall(
10792 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
10793 const Expr *Device) {
10794 if (!CGF.HaveInsertPoint())
10795 return;
10796
10797 assert((isa<OMPTargetEnterDataDirective>(D) ||
10798 isa<OMPTargetExitDataDirective>(D) ||
10799 isa<OMPTargetUpdateDirective>(D)) &&
10800 "Expecting either target enter, exit data, or update directives.");
10801
10802 CodeGenFunction::OMPTargetDataInfo InputInfo;
10803 llvm::Value *MapTypesArray = nullptr;
10804 llvm::Value *MapNamesArray = nullptr;
10805 // Generate the code for the opening of the data environment.
10806 auto &&ThenGen = [this, &D, Device, &InputInfo, &MapTypesArray,
10807 &MapNamesArray](CodeGenFunction &CGF, PrePostActionTy &) {
10808 // Emit device ID if any.
10809 llvm::Value *DeviceID = nullptr;
10810 if (Device) {
10811 DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device),
10812 CGF.Int64Ty, /*isSigned=*/true);
10813 } else {
10814 DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF);
10815 }
10816
10817 // Emit the number of elements in the offloading arrays.
10818 llvm::Constant *PointerNum =
10819 CGF.Builder.getInt32(InputInfo.NumberOfTargetItems);
10820
10821 // Source location for the ident struct
10822 llvm::Value *RTLoc = emitUpdateLocation(CGF, D.getBeginLoc());
10823
10824 llvm::Value *OffloadingArgs[] = {RTLoc,
10825 DeviceID,
10826 PointerNum,
10827 InputInfo.BasePointersArray.getPointer(),
10828 InputInfo.PointersArray.getPointer(),
10829 InputInfo.SizesArray.getPointer(),
10830 MapTypesArray,
10831 MapNamesArray,
10832 InputInfo.MappersArray.getPointer()};
10833
10834 // Select the right runtime function call for each standalone
10835 // directive.
10836 const bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>();
10837 RuntimeFunction RTLFn;
10838 switch (D.getDirectiveKind()) {
10839 case OMPD_target_enter_data:
10840 RTLFn = HasNowait ? OMPRTL___tgt_target_data_begin_nowait_mapper
10841 : OMPRTL___tgt_target_data_begin_mapper;
10842 break;
10843 case OMPD_target_exit_data:
10844 RTLFn = HasNowait ? OMPRTL___tgt_target_data_end_nowait_mapper
10845 : OMPRTL___tgt_target_data_end_mapper;
10846 break;
10847 case OMPD_target_update:
10848 RTLFn = HasNowait ? OMPRTL___tgt_target_data_update_nowait_mapper
10849 : OMPRTL___tgt_target_data_update_mapper;
10850 break;
10851 case OMPD_parallel:
10852 case OMPD_for:
10853 case OMPD_parallel_for:
10854 case OMPD_parallel_master:
10855 case OMPD_parallel_sections:
10856 case OMPD_for_simd:
10857 case OMPD_parallel_for_simd:
10858 case OMPD_cancel:
10859 case OMPD_cancellation_point:
10860 case OMPD_ordered:
10861 case OMPD_threadprivate:
10862 case OMPD_allocate:
10863 case OMPD_task:
10864 case OMPD_simd:
10865 case OMPD_tile:
10866 case OMPD_unroll:
10867 case OMPD_sections:
10868 case OMPD_section:
10869 case OMPD_single:
10870 case OMPD_master:
10871 case OMPD_critical:
10872 case OMPD_taskyield:
10873 case OMPD_barrier:
10874 case OMPD_taskwait:
10875 case OMPD_taskgroup:
10876 case OMPD_atomic:
10877 case OMPD_flush:
10878 case OMPD_depobj:
10879 case OMPD_scan:
10880 case OMPD_teams:
10881 case OMPD_target_data:
10882 case OMPD_distribute:
10883 case OMPD_distribute_simd:
10884 case OMPD_distribute_parallel_for:
10885 case OMPD_distribute_parallel_for_simd:
10886 case OMPD_teams_distribute:
10887 case OMPD_teams_distribute_simd:
10888 case OMPD_teams_distribute_parallel_for:
10889 case OMPD_teams_distribute_parallel_for_simd:
10890 case OMPD_declare_simd:
10891 case OMPD_declare_variant:
10892 case OMPD_begin_declare_variant:
10893 case OMPD_end_declare_variant:
10894 case OMPD_declare_target:
10895 case OMPD_end_declare_target:
10896 case OMPD_declare_reduction:
10897 case OMPD_declare_mapper:
10898 case OMPD_taskloop:
10899 case OMPD_taskloop_simd:
10900 case OMPD_master_taskloop:
10901 case OMPD_master_taskloop_simd:
10902 case OMPD_parallel_master_taskloop:
10903 case OMPD_parallel_master_taskloop_simd:
10904 case OMPD_target:
10905 case OMPD_target_simd:
10906 case OMPD_target_teams_distribute:
10907 case OMPD_target_teams_distribute_simd:
10908 case OMPD_target_teams_distribute_parallel_for:
10909 case OMPD_target_teams_distribute_parallel_for_simd:
10910 case OMPD_target_teams:
10911 case OMPD_target_parallel:
10912 case OMPD_target_parallel_for:
10913 case OMPD_target_parallel_for_simd:
10914 case OMPD_requires:
10915 case OMPD_metadirective:
10916 case OMPD_unknown:
10917 default:
10918 llvm_unreachable("Unexpected standalone target data directive.");
10919 break;
10920 }
10921 CGF.EmitRuntimeCall(
10922 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), RTLFn),
10923 OffloadingArgs);
10924 };
10925
10926 auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray,
10927 &MapNamesArray](CodeGenFunction &CGF,
10928 PrePostActionTy &) {
10929 // Fill up the arrays with all the mapped variables.
10930 MappableExprsHandler::MapCombinedInfoTy CombinedInfo;
10931
10932 // Get map clause information.
10933 MappableExprsHandler MEHandler(D, CGF);
10934 MEHandler.generateAllInfo(CombinedInfo);
10935
10936 CGOpenMPRuntime::TargetDataInfo Info;
10937 // Fill up the arrays and create the arguments.
10938 emitOffloadingArrays(CGF, CombinedInfo, Info, OMPBuilder,
10939 /*IsNonContiguous=*/true);
10940 bool RequiresOuterTask = D.hasClausesOfKind<OMPDependClause>() ||
10941 D.hasClausesOfKind<OMPNowaitClause>();
10942 bool EmitDebug =
10943 CGF.CGM.getCodeGenOpts().getDebugInfo() != codegenoptions::NoDebugInfo;
10944 OMPBuilder.emitOffloadingArraysArgument(CGF.Builder, Info.RTArgs, Info,
10945 EmitDebug,
10946 /*ForEndCall=*/false);
10947 InputInfo.NumberOfTargetItems = Info.NumberOfPtrs;
10948 InputInfo.BasePointersArray = Address(Info.RTArgs.BasePointersArray,
10949 CGF.VoidPtrTy, CGM.getPointerAlign());
10950 InputInfo.PointersArray = Address(Info.RTArgs.PointersArray, CGF.VoidPtrTy,
10951 CGM.getPointerAlign());
10952 InputInfo.SizesArray =
10953 Address(Info.RTArgs.SizesArray, CGF.Int64Ty, CGM.getPointerAlign());
10954 InputInfo.MappersArray =
10955 Address(Info.RTArgs.MappersArray, CGF.VoidPtrTy, CGM.getPointerAlign());
10956 MapTypesArray = Info.RTArgs.MapTypesArray;
10957 MapNamesArray = Info.RTArgs.MapNamesArray;
10958 if (RequiresOuterTask)
10959 CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo);
10960 else
10961 emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen);
10962 };
10963
10964 if (IfCond) {
10965 emitIfClause(CGF, IfCond, TargetThenGen,
10966 [](CodeGenFunction &CGF, PrePostActionTy &) {});
10967 } else {
10968 RegionCodeGenTy ThenRCG(TargetThenGen);
10969 ThenRCG(CGF);
10970 }
10971 }
10972
10973 namespace {
10974 /// Kind of parameter in a function with 'declare simd' directive.
10975 enum ParamKindTy {
10976 Linear,
10977 LinearRef,
10978 LinearUVal,
10979 LinearVal,
10980 Uniform,
10981 Vector,
10982 };
10983 /// Attribute set of the parameter.
10984 struct ParamAttrTy {
10985 ParamKindTy Kind = Vector;
10986 llvm::APSInt StrideOrArg;
10987 llvm::APSInt Alignment;
10988 bool HasVarStride = false;
10989 };
10990 } // namespace
10991
evaluateCDTSize(const FunctionDecl * FD,ArrayRef<ParamAttrTy> ParamAttrs)10992 static unsigned evaluateCDTSize(const FunctionDecl *FD,
10993 ArrayRef<ParamAttrTy> ParamAttrs) {
10994 // Every vector variant of a SIMD-enabled function has a vector length (VLEN).
10995 // If OpenMP clause "simdlen" is used, the VLEN is the value of the argument
10996 // of that clause. The VLEN value must be power of 2.
10997 // In other case the notion of the function`s "characteristic data type" (CDT)
10998 // is used to compute the vector length.
10999 // CDT is defined in the following order:
11000 // a) For non-void function, the CDT is the return type.
11001 // b) If the function has any non-uniform, non-linear parameters, then the
11002 // CDT is the type of the first such parameter.
11003 // c) If the CDT determined by a) or b) above is struct, union, or class
11004 // type which is pass-by-value (except for the type that maps to the
11005 // built-in complex data type), the characteristic data type is int.
11006 // d) If none of the above three cases is applicable, the CDT is int.
11007 // The VLEN is then determined based on the CDT and the size of vector
11008 // register of that ISA for which current vector version is generated. The
11009 // VLEN is computed using the formula below:
11010 // VLEN = sizeof(vector_register) / sizeof(CDT),
11011 // where vector register size specified in section 3.2.1 Registers and the
11012 // Stack Frame of original AMD64 ABI document.
11013 QualType RetType = FD->getReturnType();
11014 if (RetType.isNull())
11015 return 0;
11016 ASTContext &C = FD->getASTContext();
11017 QualType CDT;
11018 if (!RetType.isNull() && !RetType->isVoidType()) {
11019 CDT = RetType;
11020 } else {
11021 unsigned Offset = 0;
11022 if (const auto *MD = dyn_cast<CXXMethodDecl>(FD)) {
11023 if (ParamAttrs[Offset].Kind == Vector)
11024 CDT = C.getPointerType(C.getRecordType(MD->getParent()));
11025 ++Offset;
11026 }
11027 if (CDT.isNull()) {
11028 for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) {
11029 if (ParamAttrs[I + Offset].Kind == Vector) {
11030 CDT = FD->getParamDecl(I)->getType();
11031 break;
11032 }
11033 }
11034 }
11035 }
11036 if (CDT.isNull())
11037 CDT = C.IntTy;
11038 CDT = CDT->getCanonicalTypeUnqualified();
11039 if (CDT->isRecordType() || CDT->isUnionType())
11040 CDT = C.IntTy;
11041 return C.getTypeSize(CDT);
11042 }
11043
11044 /// Mangle the parameter part of the vector function name according to
11045 /// their OpenMP classification. The mangling function is defined in
11046 /// section 4.5 of the AAVFABI(2021Q1).
mangleVectorParameters(ArrayRef<ParamAttrTy> ParamAttrs)11047 static std::string mangleVectorParameters(ArrayRef<ParamAttrTy> ParamAttrs) {
11048 SmallString<256> Buffer;
11049 llvm::raw_svector_ostream Out(Buffer);
11050 for (const auto &ParamAttr : ParamAttrs) {
11051 switch (ParamAttr.Kind) {
11052 case Linear:
11053 Out << 'l';
11054 break;
11055 case LinearRef:
11056 Out << 'R';
11057 break;
11058 case LinearUVal:
11059 Out << 'U';
11060 break;
11061 case LinearVal:
11062 Out << 'L';
11063 break;
11064 case Uniform:
11065 Out << 'u';
11066 break;
11067 case Vector:
11068 Out << 'v';
11069 break;
11070 }
11071 if (ParamAttr.HasVarStride)
11072 Out << "s" << ParamAttr.StrideOrArg;
11073 else if (ParamAttr.Kind == Linear || ParamAttr.Kind == LinearRef ||
11074 ParamAttr.Kind == LinearUVal || ParamAttr.Kind == LinearVal) {
11075 // Don't print the step value if it is not present or if it is
11076 // equal to 1.
11077 if (ParamAttr.StrideOrArg < 0)
11078 Out << 'n' << -ParamAttr.StrideOrArg;
11079 else if (ParamAttr.StrideOrArg != 1)
11080 Out << ParamAttr.StrideOrArg;
11081 }
11082
11083 if (!!ParamAttr.Alignment)
11084 Out << 'a' << ParamAttr.Alignment;
11085 }
11086
11087 return std::string(Out.str());
11088 }
11089
11090 static void
emitX86DeclareSimdFunction(const FunctionDecl * FD,llvm::Function * Fn,const llvm::APSInt & VLENVal,ArrayRef<ParamAttrTy> ParamAttrs,OMPDeclareSimdDeclAttr::BranchStateTy State)11091 emitX86DeclareSimdFunction(const FunctionDecl *FD, llvm::Function *Fn,
11092 const llvm::APSInt &VLENVal,
11093 ArrayRef<ParamAttrTy> ParamAttrs,
11094 OMPDeclareSimdDeclAttr::BranchStateTy State) {
11095 struct ISADataTy {
11096 char ISA;
11097 unsigned VecRegSize;
11098 };
11099 ISADataTy ISAData[] = {
11100 {
11101 'b', 128
11102 }, // SSE
11103 {
11104 'c', 256
11105 }, // AVX
11106 {
11107 'd', 256
11108 }, // AVX2
11109 {
11110 'e', 512
11111 }, // AVX512
11112 };
11113 llvm::SmallVector<char, 2> Masked;
11114 switch (State) {
11115 case OMPDeclareSimdDeclAttr::BS_Undefined:
11116 Masked.push_back('N');
11117 Masked.push_back('M');
11118 break;
11119 case OMPDeclareSimdDeclAttr::BS_Notinbranch:
11120 Masked.push_back('N');
11121 break;
11122 case OMPDeclareSimdDeclAttr::BS_Inbranch:
11123 Masked.push_back('M');
11124 break;
11125 }
11126 for (char Mask : Masked) {
11127 for (const ISADataTy &Data : ISAData) {
11128 SmallString<256> Buffer;
11129 llvm::raw_svector_ostream Out(Buffer);
11130 Out << "_ZGV" << Data.ISA << Mask;
11131 if (!VLENVal) {
11132 unsigned NumElts = evaluateCDTSize(FD, ParamAttrs);
11133 assert(NumElts && "Non-zero simdlen/cdtsize expected");
11134 Out << llvm::APSInt::getUnsigned(Data.VecRegSize / NumElts);
11135 } else {
11136 Out << VLENVal;
11137 }
11138 Out << mangleVectorParameters(ParamAttrs);
11139 Out << '_' << Fn->getName();
11140 Fn->addFnAttr(Out.str());
11141 }
11142 }
11143 }
11144
11145 // This are the Functions that are needed to mangle the name of the
11146 // vector functions generated by the compiler, according to the rules
11147 // defined in the "Vector Function ABI specifications for AArch64",
11148 // available at
11149 // https://developer.arm.com/products/software-development-tools/hpc/arm-compiler-for-hpc/vector-function-abi.
11150
11151 /// Maps To Vector (MTV), as defined in 4.1.1 of the AAVFABI (2021Q1).
getAArch64MTV(QualType QT,ParamKindTy Kind)11152 static bool getAArch64MTV(QualType QT, ParamKindTy Kind) {
11153 QT = QT.getCanonicalType();
11154
11155 if (QT->isVoidType())
11156 return false;
11157
11158 if (Kind == ParamKindTy::Uniform)
11159 return false;
11160
11161 if (Kind == ParamKindTy::LinearUVal || ParamKindTy::LinearRef)
11162 return false;
11163
11164 if ((Kind == ParamKindTy::Linear || Kind == ParamKindTy::LinearVal) &&
11165 !QT->isReferenceType())
11166 return false;
11167
11168 return true;
11169 }
11170
11171 /// Pass By Value (PBV), as defined in 3.1.2 of the AAVFABI.
getAArch64PBV(QualType QT,ASTContext & C)11172 static bool getAArch64PBV(QualType QT, ASTContext &C) {
11173 QT = QT.getCanonicalType();
11174 unsigned Size = C.getTypeSize(QT);
11175
11176 // Only scalars and complex within 16 bytes wide set PVB to true.
11177 if (Size != 8 && Size != 16 && Size != 32 && Size != 64 && Size != 128)
11178 return false;
11179
11180 if (QT->isFloatingType())
11181 return true;
11182
11183 if (QT->isIntegerType())
11184 return true;
11185
11186 if (QT->isPointerType())
11187 return true;
11188
11189 // TODO: Add support for complex types (section 3.1.2, item 2).
11190
11191 return false;
11192 }
11193
11194 /// Computes the lane size (LS) of a return type or of an input parameter,
11195 /// as defined by `LS(P)` in 3.2.1 of the AAVFABI.
11196 /// TODO: Add support for references, section 3.2.1, item 1.
getAArch64LS(QualType QT,ParamKindTy Kind,ASTContext & C)11197 static unsigned getAArch64LS(QualType QT, ParamKindTy Kind, ASTContext &C) {
11198 if (!getAArch64MTV(QT, Kind) && QT.getCanonicalType()->isPointerType()) {
11199 QualType PTy = QT.getCanonicalType()->getPointeeType();
11200 if (getAArch64PBV(PTy, C))
11201 return C.getTypeSize(PTy);
11202 }
11203 if (getAArch64PBV(QT, C))
11204 return C.getTypeSize(QT);
11205
11206 return C.getTypeSize(C.getUIntPtrType());
11207 }
11208
11209 // Get Narrowest Data Size (NDS) and Widest Data Size (WDS) from the
11210 // signature of the scalar function, as defined in 3.2.2 of the
11211 // AAVFABI.
11212 static std::tuple<unsigned, unsigned, bool>
getNDSWDS(const FunctionDecl * FD,ArrayRef<ParamAttrTy> ParamAttrs)11213 getNDSWDS(const FunctionDecl *FD, ArrayRef<ParamAttrTy> ParamAttrs) {
11214 QualType RetType = FD->getReturnType().getCanonicalType();
11215
11216 ASTContext &C = FD->getASTContext();
11217
11218 bool OutputBecomesInput = false;
11219
11220 llvm::SmallVector<unsigned, 8> Sizes;
11221 if (!RetType->isVoidType()) {
11222 Sizes.push_back(getAArch64LS(RetType, ParamKindTy::Vector, C));
11223 if (!getAArch64PBV(RetType, C) && getAArch64MTV(RetType, {}))
11224 OutputBecomesInput = true;
11225 }
11226 for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) {
11227 QualType QT = FD->getParamDecl(I)->getType().getCanonicalType();
11228 Sizes.push_back(getAArch64LS(QT, ParamAttrs[I].Kind, C));
11229 }
11230
11231 assert(!Sizes.empty() && "Unable to determine NDS and WDS.");
11232 // The LS of a function parameter / return value can only be a power
11233 // of 2, starting from 8 bits, up to 128.
11234 assert(llvm::all_of(Sizes,
11235 [](unsigned Size) {
11236 return Size == 8 || Size == 16 || Size == 32 ||
11237 Size == 64 || Size == 128;
11238 }) &&
11239 "Invalid size");
11240
11241 return std::make_tuple(*std::min_element(std::begin(Sizes), std::end(Sizes)),
11242 *std::max_element(std::begin(Sizes), std::end(Sizes)),
11243 OutputBecomesInput);
11244 }
11245
11246 // Function used to add the attribute. The parameter `VLEN` is
11247 // templated to allow the use of "x" when targeting scalable functions
11248 // for SVE.
11249 template <typename T>
addAArch64VectorName(T VLEN,StringRef LMask,StringRef Prefix,char ISA,StringRef ParSeq,StringRef MangledName,bool OutputBecomesInput,llvm::Function * Fn)11250 static void addAArch64VectorName(T VLEN, StringRef LMask, StringRef Prefix,
11251 char ISA, StringRef ParSeq,
11252 StringRef MangledName, bool OutputBecomesInput,
11253 llvm::Function *Fn) {
11254 SmallString<256> Buffer;
11255 llvm::raw_svector_ostream Out(Buffer);
11256 Out << Prefix << ISA << LMask << VLEN;
11257 if (OutputBecomesInput)
11258 Out << "v";
11259 Out << ParSeq << "_" << MangledName;
11260 Fn->addFnAttr(Out.str());
11261 }
11262
11263 // Helper function to generate the Advanced SIMD names depending on
11264 // the value of the NDS when simdlen is not present.
addAArch64AdvSIMDNDSNames(unsigned NDS,StringRef Mask,StringRef Prefix,char ISA,StringRef ParSeq,StringRef MangledName,bool OutputBecomesInput,llvm::Function * Fn)11265 static void addAArch64AdvSIMDNDSNames(unsigned NDS, StringRef Mask,
11266 StringRef Prefix, char ISA,
11267 StringRef ParSeq, StringRef MangledName,
11268 bool OutputBecomesInput,
11269 llvm::Function *Fn) {
11270 switch (NDS) {
11271 case 8:
11272 addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName,
11273 OutputBecomesInput, Fn);
11274 addAArch64VectorName(16, Mask, Prefix, ISA, ParSeq, MangledName,
11275 OutputBecomesInput, Fn);
11276 break;
11277 case 16:
11278 addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName,
11279 OutputBecomesInput, Fn);
11280 addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName,
11281 OutputBecomesInput, Fn);
11282 break;
11283 case 32:
11284 addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName,
11285 OutputBecomesInput, Fn);
11286 addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName,
11287 OutputBecomesInput, Fn);
11288 break;
11289 case 64:
11290 case 128:
11291 addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName,
11292 OutputBecomesInput, Fn);
11293 break;
11294 default:
11295 llvm_unreachable("Scalar type is too wide.");
11296 }
11297 }
11298
11299 /// Emit vector function attributes for AArch64, as defined in the AAVFABI.
emitAArch64DeclareSimdFunction(CodeGenModule & CGM,const FunctionDecl * FD,unsigned UserVLEN,ArrayRef<ParamAttrTy> ParamAttrs,OMPDeclareSimdDeclAttr::BranchStateTy State,StringRef MangledName,char ISA,unsigned VecRegSize,llvm::Function * Fn,SourceLocation SLoc)11300 static void emitAArch64DeclareSimdFunction(
11301 CodeGenModule &CGM, const FunctionDecl *FD, unsigned UserVLEN,
11302 ArrayRef<ParamAttrTy> ParamAttrs,
11303 OMPDeclareSimdDeclAttr::BranchStateTy State, StringRef MangledName,
11304 char ISA, unsigned VecRegSize, llvm::Function *Fn, SourceLocation SLoc) {
11305
11306 // Get basic data for building the vector signature.
11307 const auto Data = getNDSWDS(FD, ParamAttrs);
11308 const unsigned NDS = std::get<0>(Data);
11309 const unsigned WDS = std::get<1>(Data);
11310 const bool OutputBecomesInput = std::get<2>(Data);
11311
11312 // Check the values provided via `simdlen` by the user.
11313 // 1. A `simdlen(1)` doesn't produce vector signatures,
11314 if (UserVLEN == 1) {
11315 unsigned DiagID = CGM.getDiags().getCustomDiagID(
11316 DiagnosticsEngine::Warning,
11317 "The clause simdlen(1) has no effect when targeting aarch64.");
11318 CGM.getDiags().Report(SLoc, DiagID);
11319 return;
11320 }
11321
11322 // 2. Section 3.3.1, item 1: user input must be a power of 2 for
11323 // Advanced SIMD output.
11324 if (ISA == 'n' && UserVLEN && !llvm::isPowerOf2_32(UserVLEN)) {
11325 unsigned DiagID = CGM.getDiags().getCustomDiagID(
11326 DiagnosticsEngine::Warning, "The value specified in simdlen must be a "
11327 "power of 2 when targeting Advanced SIMD.");
11328 CGM.getDiags().Report(SLoc, DiagID);
11329 return;
11330 }
11331
11332 // 3. Section 3.4.1. SVE fixed lengh must obey the architectural
11333 // limits.
11334 if (ISA == 's' && UserVLEN != 0) {
11335 if ((UserVLEN * WDS > 2048) || (UserVLEN * WDS % 128 != 0)) {
11336 unsigned DiagID = CGM.getDiags().getCustomDiagID(
11337 DiagnosticsEngine::Warning, "The clause simdlen must fit the %0-bit "
11338 "lanes in the architectural constraints "
11339 "for SVE (min is 128-bit, max is "
11340 "2048-bit, by steps of 128-bit)");
11341 CGM.getDiags().Report(SLoc, DiagID) << WDS;
11342 return;
11343 }
11344 }
11345
11346 // Sort out parameter sequence.
11347 const std::string ParSeq = mangleVectorParameters(ParamAttrs);
11348 StringRef Prefix = "_ZGV";
11349 // Generate simdlen from user input (if any).
11350 if (UserVLEN) {
11351 if (ISA == 's') {
11352 // SVE generates only a masked function.
11353 addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName,
11354 OutputBecomesInput, Fn);
11355 } else {
11356 assert(ISA == 'n' && "Expected ISA either 's' or 'n'.");
11357 // Advanced SIMD generates one or two functions, depending on
11358 // the `[not]inbranch` clause.
11359 switch (State) {
11360 case OMPDeclareSimdDeclAttr::BS_Undefined:
11361 addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName,
11362 OutputBecomesInput, Fn);
11363 addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName,
11364 OutputBecomesInput, Fn);
11365 break;
11366 case OMPDeclareSimdDeclAttr::BS_Notinbranch:
11367 addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName,
11368 OutputBecomesInput, Fn);
11369 break;
11370 case OMPDeclareSimdDeclAttr::BS_Inbranch:
11371 addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName,
11372 OutputBecomesInput, Fn);
11373 break;
11374 }
11375 }
11376 } else {
11377 // If no user simdlen is provided, follow the AAVFABI rules for
11378 // generating the vector length.
11379 if (ISA == 's') {
11380 // SVE, section 3.4.1, item 1.
11381 addAArch64VectorName("x", "M", Prefix, ISA, ParSeq, MangledName,
11382 OutputBecomesInput, Fn);
11383 } else {
11384 assert(ISA == 'n' && "Expected ISA either 's' or 'n'.");
11385 // Advanced SIMD, Section 3.3.1 of the AAVFABI, generates one or
11386 // two vector names depending on the use of the clause
11387 // `[not]inbranch`.
11388 switch (State) {
11389 case OMPDeclareSimdDeclAttr::BS_Undefined:
11390 addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName,
11391 OutputBecomesInput, Fn);
11392 addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName,
11393 OutputBecomesInput, Fn);
11394 break;
11395 case OMPDeclareSimdDeclAttr::BS_Notinbranch:
11396 addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName,
11397 OutputBecomesInput, Fn);
11398 break;
11399 case OMPDeclareSimdDeclAttr::BS_Inbranch:
11400 addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName,
11401 OutputBecomesInput, Fn);
11402 break;
11403 }
11404 }
11405 }
11406 }
11407
emitDeclareSimdFunction(const FunctionDecl * FD,llvm::Function * Fn)11408 void CGOpenMPRuntime::emitDeclareSimdFunction(const FunctionDecl *FD,
11409 llvm::Function *Fn) {
11410 ASTContext &C = CGM.getContext();
11411 FD = FD->getMostRecentDecl();
11412 while (FD) {
11413 // Map params to their positions in function decl.
11414 llvm::DenseMap<const Decl *, unsigned> ParamPositions;
11415 if (isa<CXXMethodDecl>(FD))
11416 ParamPositions.try_emplace(FD, 0);
11417 unsigned ParamPos = ParamPositions.size();
11418 for (const ParmVarDecl *P : FD->parameters()) {
11419 ParamPositions.try_emplace(P->getCanonicalDecl(), ParamPos);
11420 ++ParamPos;
11421 }
11422 for (const auto *Attr : FD->specific_attrs<OMPDeclareSimdDeclAttr>()) {
11423 llvm::SmallVector<ParamAttrTy, 8> ParamAttrs(ParamPositions.size());
11424 // Mark uniform parameters.
11425 for (const Expr *E : Attr->uniforms()) {
11426 E = E->IgnoreParenImpCasts();
11427 unsigned Pos;
11428 if (isa<CXXThisExpr>(E)) {
11429 Pos = ParamPositions[FD];
11430 } else {
11431 const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
11432 ->getCanonicalDecl();
11433 auto It = ParamPositions.find(PVD);
11434 assert(It != ParamPositions.end() && "Function parameter not found");
11435 Pos = It->second;
11436 }
11437 ParamAttrs[Pos].Kind = Uniform;
11438 }
11439 // Get alignment info.
11440 auto *NI = Attr->alignments_begin();
11441 for (const Expr *E : Attr->aligneds()) {
11442 E = E->IgnoreParenImpCasts();
11443 unsigned Pos;
11444 QualType ParmTy;
11445 if (isa<CXXThisExpr>(E)) {
11446 Pos = ParamPositions[FD];
11447 ParmTy = E->getType();
11448 } else {
11449 const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
11450 ->getCanonicalDecl();
11451 auto It = ParamPositions.find(PVD);
11452 assert(It != ParamPositions.end() && "Function parameter not found");
11453 Pos = It->second;
11454 ParmTy = PVD->getType();
11455 }
11456 ParamAttrs[Pos].Alignment =
11457 (*NI)
11458 ? (*NI)->EvaluateKnownConstInt(C)
11459 : llvm::APSInt::getUnsigned(
11460 C.toCharUnitsFromBits(C.getOpenMPDefaultSimdAlign(ParmTy))
11461 .getQuantity());
11462 ++NI;
11463 }
11464 // Mark linear parameters.
11465 auto *SI = Attr->steps_begin();
11466 auto *MI = Attr->modifiers_begin();
11467 for (const Expr *E : Attr->linears()) {
11468 E = E->IgnoreParenImpCasts();
11469 unsigned Pos;
11470 bool IsReferenceType = false;
11471 // Rescaling factor needed to compute the linear parameter
11472 // value in the mangled name.
11473 unsigned PtrRescalingFactor = 1;
11474 if (isa<CXXThisExpr>(E)) {
11475 Pos = ParamPositions[FD];
11476 auto *P = cast<PointerType>(E->getType());
11477 PtrRescalingFactor = CGM.getContext()
11478 .getTypeSizeInChars(P->getPointeeType())
11479 .getQuantity();
11480 } else {
11481 const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl())
11482 ->getCanonicalDecl();
11483 auto It = ParamPositions.find(PVD);
11484 assert(It != ParamPositions.end() && "Function parameter not found");
11485 Pos = It->second;
11486 if (auto *P = dyn_cast<PointerType>(PVD->getType()))
11487 PtrRescalingFactor = CGM.getContext()
11488 .getTypeSizeInChars(P->getPointeeType())
11489 .getQuantity();
11490 else if (PVD->getType()->isReferenceType()) {
11491 IsReferenceType = true;
11492 PtrRescalingFactor =
11493 CGM.getContext()
11494 .getTypeSizeInChars(PVD->getType().getNonReferenceType())
11495 .getQuantity();
11496 }
11497 }
11498 ParamAttrTy &ParamAttr = ParamAttrs[Pos];
11499 if (*MI == OMPC_LINEAR_ref)
11500 ParamAttr.Kind = LinearRef;
11501 else if (*MI == OMPC_LINEAR_uval)
11502 ParamAttr.Kind = LinearUVal;
11503 else if (IsReferenceType)
11504 ParamAttr.Kind = LinearVal;
11505 else
11506 ParamAttr.Kind = Linear;
11507 // Assuming a stride of 1, for `linear` without modifiers.
11508 ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(1);
11509 if (*SI) {
11510 Expr::EvalResult Result;
11511 if (!(*SI)->EvaluateAsInt(Result, C, Expr::SE_AllowSideEffects)) {
11512 if (const auto *DRE =
11513 cast<DeclRefExpr>((*SI)->IgnoreParenImpCasts())) {
11514 if (const auto *StridePVD =
11515 dyn_cast<ParmVarDecl>(DRE->getDecl())) {
11516 ParamAttr.HasVarStride = true;
11517 auto It = ParamPositions.find(StridePVD->getCanonicalDecl());
11518 assert(It != ParamPositions.end() &&
11519 "Function parameter not found");
11520 ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned(It->second);
11521 }
11522 }
11523 } else {
11524 ParamAttr.StrideOrArg = Result.Val.getInt();
11525 }
11526 }
11527 // If we are using a linear clause on a pointer, we need to
11528 // rescale the value of linear_step with the byte size of the
11529 // pointee type.
11530 if (!ParamAttr.HasVarStride &&
11531 (ParamAttr.Kind == Linear || ParamAttr.Kind == LinearRef))
11532 ParamAttr.StrideOrArg = ParamAttr.StrideOrArg * PtrRescalingFactor;
11533 ++SI;
11534 ++MI;
11535 }
11536 llvm::APSInt VLENVal;
11537 SourceLocation ExprLoc;
11538 const Expr *VLENExpr = Attr->getSimdlen();
11539 if (VLENExpr) {
11540 VLENVal = VLENExpr->EvaluateKnownConstInt(C);
11541 ExprLoc = VLENExpr->getExprLoc();
11542 }
11543 OMPDeclareSimdDeclAttr::BranchStateTy State = Attr->getBranchState();
11544 if (CGM.getTriple().isX86()) {
11545 emitX86DeclareSimdFunction(FD, Fn, VLENVal, ParamAttrs, State);
11546 } else if (CGM.getTriple().getArch() == llvm::Triple::aarch64) {
11547 unsigned VLEN = VLENVal.getExtValue();
11548 StringRef MangledName = Fn->getName();
11549 if (CGM.getTarget().hasFeature("sve"))
11550 emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State,
11551 MangledName, 's', 128, Fn, ExprLoc);
11552 else if (CGM.getTarget().hasFeature("neon"))
11553 emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State,
11554 MangledName, 'n', 128, Fn, ExprLoc);
11555 }
11556 }
11557 FD = FD->getPreviousDecl();
11558 }
11559 }
11560
11561 namespace {
11562 /// Cleanup action for doacross support.
11563 class DoacrossCleanupTy final : public EHScopeStack::Cleanup {
11564 public:
11565 static const int DoacrossFinArgs = 2;
11566
11567 private:
11568 llvm::FunctionCallee RTLFn;
11569 llvm::Value *Args[DoacrossFinArgs];
11570
11571 public:
DoacrossCleanupTy(llvm::FunctionCallee RTLFn,ArrayRef<llvm::Value * > CallArgs)11572 DoacrossCleanupTy(llvm::FunctionCallee RTLFn,
11573 ArrayRef<llvm::Value *> CallArgs)
11574 : RTLFn(RTLFn) {
11575 assert(CallArgs.size() == DoacrossFinArgs);
11576 std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args));
11577 }
Emit(CodeGenFunction & CGF,Flags)11578 void Emit(CodeGenFunction &CGF, Flags /*flags*/) override {
11579 if (!CGF.HaveInsertPoint())
11580 return;
11581 CGF.EmitRuntimeCall(RTLFn, Args);
11582 }
11583 };
11584 } // namespace
11585
emitDoacrossInit(CodeGenFunction & CGF,const OMPLoopDirective & D,ArrayRef<Expr * > NumIterations)11586 void CGOpenMPRuntime::emitDoacrossInit(CodeGenFunction &CGF,
11587 const OMPLoopDirective &D,
11588 ArrayRef<Expr *> NumIterations) {
11589 if (!CGF.HaveInsertPoint())
11590 return;
11591
11592 ASTContext &C = CGM.getContext();
11593 QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true);
11594 RecordDecl *RD;
11595 if (KmpDimTy.isNull()) {
11596 // Build struct kmp_dim { // loop bounds info casted to kmp_int64
11597 // kmp_int64 lo; // lower
11598 // kmp_int64 up; // upper
11599 // kmp_int64 st; // stride
11600 // };
11601 RD = C.buildImplicitRecord("kmp_dim");
11602 RD->startDefinition();
11603 addFieldToRecordDecl(C, RD, Int64Ty);
11604 addFieldToRecordDecl(C, RD, Int64Ty);
11605 addFieldToRecordDecl(C, RD, Int64Ty);
11606 RD->completeDefinition();
11607 KmpDimTy = C.getRecordType(RD);
11608 } else {
11609 RD = cast<RecordDecl>(KmpDimTy->getAsTagDecl());
11610 }
11611 llvm::APInt Size(/*numBits=*/32, NumIterations.size());
11612 QualType ArrayTy =
11613 C.getConstantArrayType(KmpDimTy, Size, nullptr, ArrayType::Normal, 0);
11614
11615 Address DimsAddr = CGF.CreateMemTemp(ArrayTy, "dims");
11616 CGF.EmitNullInitialization(DimsAddr, ArrayTy);
11617 enum { LowerFD = 0, UpperFD, StrideFD };
11618 // Fill dims with data.
11619 for (unsigned I = 0, E = NumIterations.size(); I < E; ++I) {
11620 LValue DimsLVal = CGF.MakeAddrLValue(
11621 CGF.Builder.CreateConstArrayGEP(DimsAddr, I), KmpDimTy);
11622 // dims.upper = num_iterations;
11623 LValue UpperLVal = CGF.EmitLValueForField(
11624 DimsLVal, *std::next(RD->field_begin(), UpperFD));
11625 llvm::Value *NumIterVal = CGF.EmitScalarConversion(
11626 CGF.EmitScalarExpr(NumIterations[I]), NumIterations[I]->getType(),
11627 Int64Ty, NumIterations[I]->getExprLoc());
11628 CGF.EmitStoreOfScalar(NumIterVal, UpperLVal);
11629 // dims.stride = 1;
11630 LValue StrideLVal = CGF.EmitLValueForField(
11631 DimsLVal, *std::next(RD->field_begin(), StrideFD));
11632 CGF.EmitStoreOfScalar(llvm::ConstantInt::getSigned(CGM.Int64Ty, /*V=*/1),
11633 StrideLVal);
11634 }
11635
11636 // Build call void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
11637 // kmp_int32 num_dims, struct kmp_dim * dims);
11638 llvm::Value *Args[] = {
11639 emitUpdateLocation(CGF, D.getBeginLoc()),
11640 getThreadID(CGF, D.getBeginLoc()),
11641 llvm::ConstantInt::getSigned(CGM.Int32Ty, NumIterations.size()),
11642 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
11643 CGF.Builder.CreateConstArrayGEP(DimsAddr, 0).getPointer(),
11644 CGM.VoidPtrTy)};
11645
11646 llvm::FunctionCallee RTLFn = OMPBuilder.getOrCreateRuntimeFunction(
11647 CGM.getModule(), OMPRTL___kmpc_doacross_init);
11648 CGF.EmitRuntimeCall(RTLFn, Args);
11649 llvm::Value *FiniArgs[DoacrossCleanupTy::DoacrossFinArgs] = {
11650 emitUpdateLocation(CGF, D.getEndLoc()), getThreadID(CGF, D.getEndLoc())};
11651 llvm::FunctionCallee FiniRTLFn = OMPBuilder.getOrCreateRuntimeFunction(
11652 CGM.getModule(), OMPRTL___kmpc_doacross_fini);
11653 CGF.EHStack.pushCleanup<DoacrossCleanupTy>(NormalAndEHCleanup, FiniRTLFn,
11654 llvm::ArrayRef(FiniArgs));
11655 }
11656
emitDoacrossOrdered(CodeGenFunction & CGF,const OMPDependClause * C)11657 void CGOpenMPRuntime::emitDoacrossOrdered(CodeGenFunction &CGF,
11658 const OMPDependClause *C) {
11659 QualType Int64Ty =
11660 CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1);
11661 llvm::APInt Size(/*numBits=*/32, C->getNumLoops());
11662 QualType ArrayTy = CGM.getContext().getConstantArrayType(
11663 Int64Ty, Size, nullptr, ArrayType::Normal, 0);
11664 Address CntAddr = CGF.CreateMemTemp(ArrayTy, ".cnt.addr");
11665 for (unsigned I = 0, E = C->getNumLoops(); I < E; ++I) {
11666 const Expr *CounterVal = C->getLoopData(I);
11667 assert(CounterVal);
11668 llvm::Value *CntVal = CGF.EmitScalarConversion(
11669 CGF.EmitScalarExpr(CounterVal), CounterVal->getType(), Int64Ty,
11670 CounterVal->getExprLoc());
11671 CGF.EmitStoreOfScalar(CntVal, CGF.Builder.CreateConstArrayGEP(CntAddr, I),
11672 /*Volatile=*/false, Int64Ty);
11673 }
11674 llvm::Value *Args[] = {
11675 emitUpdateLocation(CGF, C->getBeginLoc()),
11676 getThreadID(CGF, C->getBeginLoc()),
11677 CGF.Builder.CreateConstArrayGEP(CntAddr, 0).getPointer()};
11678 llvm::FunctionCallee RTLFn;
11679 if (C->getDependencyKind() == OMPC_DEPEND_source) {
11680 RTLFn = OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
11681 OMPRTL___kmpc_doacross_post);
11682 } else {
11683 assert(C->getDependencyKind() == OMPC_DEPEND_sink);
11684 RTLFn = OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
11685 OMPRTL___kmpc_doacross_wait);
11686 }
11687 CGF.EmitRuntimeCall(RTLFn, Args);
11688 }
11689
emitCall(CodeGenFunction & CGF,SourceLocation Loc,llvm::FunctionCallee Callee,ArrayRef<llvm::Value * > Args) const11690 void CGOpenMPRuntime::emitCall(CodeGenFunction &CGF, SourceLocation Loc,
11691 llvm::FunctionCallee Callee,
11692 ArrayRef<llvm::Value *> Args) const {
11693 assert(Loc.isValid() && "Outlined function call location must be valid.");
11694 auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc);
11695
11696 if (auto *Fn = dyn_cast<llvm::Function>(Callee.getCallee())) {
11697 if (Fn->doesNotThrow()) {
11698 CGF.EmitNounwindRuntimeCall(Fn, Args);
11699 return;
11700 }
11701 }
11702 CGF.EmitRuntimeCall(Callee, Args);
11703 }
11704
emitOutlinedFunctionCall(CodeGenFunction & CGF,SourceLocation Loc,llvm::FunctionCallee OutlinedFn,ArrayRef<llvm::Value * > Args) const11705 void CGOpenMPRuntime::emitOutlinedFunctionCall(
11706 CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
11707 ArrayRef<llvm::Value *> Args) const {
11708 emitCall(CGF, Loc, OutlinedFn, Args);
11709 }
11710
emitFunctionProlog(CodeGenFunction & CGF,const Decl * D)11711 void CGOpenMPRuntime::emitFunctionProlog(CodeGenFunction &CGF, const Decl *D) {
11712 if (const auto *FD = dyn_cast<FunctionDecl>(D))
11713 if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(FD))
11714 HasEmittedDeclareTargetRegion = true;
11715 }
11716
getParameterAddress(CodeGenFunction & CGF,const VarDecl * NativeParam,const VarDecl * TargetParam) const11717 Address CGOpenMPRuntime::getParameterAddress(CodeGenFunction &CGF,
11718 const VarDecl *NativeParam,
11719 const VarDecl *TargetParam) const {
11720 return CGF.GetAddrOfLocalVar(NativeParam);
11721 }
11722
11723 /// Return allocator value from expression, or return a null allocator (default
11724 /// when no allocator specified).
getAllocatorVal(CodeGenFunction & CGF,const Expr * Allocator)11725 static llvm::Value *getAllocatorVal(CodeGenFunction &CGF,
11726 const Expr *Allocator) {
11727 llvm::Value *AllocVal;
11728 if (Allocator) {
11729 AllocVal = CGF.EmitScalarExpr(Allocator);
11730 // According to the standard, the original allocator type is a enum
11731 // (integer). Convert to pointer type, if required.
11732 AllocVal = CGF.EmitScalarConversion(AllocVal, Allocator->getType(),
11733 CGF.getContext().VoidPtrTy,
11734 Allocator->getExprLoc());
11735 } else {
11736 // If no allocator specified, it defaults to the null allocator.
11737 AllocVal = llvm::Constant::getNullValue(
11738 CGF.CGM.getTypes().ConvertType(CGF.getContext().VoidPtrTy));
11739 }
11740 return AllocVal;
11741 }
11742
11743 /// Return the alignment from an allocate directive if present.
getAlignmentValue(CodeGenModule & CGM,const VarDecl * VD)11744 static llvm::Value *getAlignmentValue(CodeGenModule &CGM, const VarDecl *VD) {
11745 std::optional<CharUnits> AllocateAlignment = CGM.getOMPAllocateAlignment(VD);
11746
11747 if (!AllocateAlignment)
11748 return nullptr;
11749
11750 return llvm::ConstantInt::get(CGM.SizeTy, AllocateAlignment->getQuantity());
11751 }
11752
getAddressOfLocalVariable(CodeGenFunction & CGF,const VarDecl * VD)11753 Address CGOpenMPRuntime::getAddressOfLocalVariable(CodeGenFunction &CGF,
11754 const VarDecl *VD) {
11755 if (!VD)
11756 return Address::invalid();
11757 Address UntiedAddr = Address::invalid();
11758 Address UntiedRealAddr = Address::invalid();
11759 auto It = FunctionToUntiedTaskStackMap.find(CGF.CurFn);
11760 if (It != FunctionToUntiedTaskStackMap.end()) {
11761 const UntiedLocalVarsAddressesMap &UntiedData =
11762 UntiedLocalVarsStack[It->second];
11763 auto I = UntiedData.find(VD);
11764 if (I != UntiedData.end()) {
11765 UntiedAddr = I->second.first;
11766 UntiedRealAddr = I->second.second;
11767 }
11768 }
11769 const VarDecl *CVD = VD->getCanonicalDecl();
11770 if (CVD->hasAttr<OMPAllocateDeclAttr>()) {
11771 // Use the default allocation.
11772 if (!isAllocatableDecl(VD))
11773 return UntiedAddr;
11774 llvm::Value *Size;
11775 CharUnits Align = CGM.getContext().getDeclAlign(CVD);
11776 if (CVD->getType()->isVariablyModifiedType()) {
11777 Size = CGF.getTypeSize(CVD->getType());
11778 // Align the size: ((size + align - 1) / align) * align
11779 Size = CGF.Builder.CreateNUWAdd(
11780 Size, CGM.getSize(Align - CharUnits::fromQuantity(1)));
11781 Size = CGF.Builder.CreateUDiv(Size, CGM.getSize(Align));
11782 Size = CGF.Builder.CreateNUWMul(Size, CGM.getSize(Align));
11783 } else {
11784 CharUnits Sz = CGM.getContext().getTypeSizeInChars(CVD->getType());
11785 Size = CGM.getSize(Sz.alignTo(Align));
11786 }
11787 llvm::Value *ThreadID = getThreadID(CGF, CVD->getBeginLoc());
11788 const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>();
11789 const Expr *Allocator = AA->getAllocator();
11790 llvm::Value *AllocVal = getAllocatorVal(CGF, Allocator);
11791 llvm::Value *Alignment = getAlignmentValue(CGM, CVD);
11792 SmallVector<llvm::Value *, 4> Args;
11793 Args.push_back(ThreadID);
11794 if (Alignment)
11795 Args.push_back(Alignment);
11796 Args.push_back(Size);
11797 Args.push_back(AllocVal);
11798 llvm::omp::RuntimeFunction FnID =
11799 Alignment ? OMPRTL___kmpc_aligned_alloc : OMPRTL___kmpc_alloc;
11800 llvm::Value *Addr = CGF.EmitRuntimeCall(
11801 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), FnID), Args,
11802 getName({CVD->getName(), ".void.addr"}));
11803 llvm::FunctionCallee FiniRTLFn = OMPBuilder.getOrCreateRuntimeFunction(
11804 CGM.getModule(), OMPRTL___kmpc_free);
11805 QualType Ty = CGM.getContext().getPointerType(CVD->getType());
11806 Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
11807 Addr, CGF.ConvertTypeForMem(Ty), getName({CVD->getName(), ".addr"}));
11808 if (UntiedAddr.isValid())
11809 CGF.EmitStoreOfScalar(Addr, UntiedAddr, /*Volatile=*/false, Ty);
11810
11811 // Cleanup action for allocate support.
11812 class OMPAllocateCleanupTy final : public EHScopeStack::Cleanup {
11813 llvm::FunctionCallee RTLFn;
11814 SourceLocation::UIntTy LocEncoding;
11815 Address Addr;
11816 const Expr *AllocExpr;
11817
11818 public:
11819 OMPAllocateCleanupTy(llvm::FunctionCallee RTLFn,
11820 SourceLocation::UIntTy LocEncoding, Address Addr,
11821 const Expr *AllocExpr)
11822 : RTLFn(RTLFn), LocEncoding(LocEncoding), Addr(Addr),
11823 AllocExpr(AllocExpr) {}
11824 void Emit(CodeGenFunction &CGF, Flags /*flags*/) override {
11825 if (!CGF.HaveInsertPoint())
11826 return;
11827 llvm::Value *Args[3];
11828 Args[0] = CGF.CGM.getOpenMPRuntime().getThreadID(
11829 CGF, SourceLocation::getFromRawEncoding(LocEncoding));
11830 Args[1] = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
11831 Addr.getPointer(), CGF.VoidPtrTy);
11832 llvm::Value *AllocVal = getAllocatorVal(CGF, AllocExpr);
11833 Args[2] = AllocVal;
11834 CGF.EmitRuntimeCall(RTLFn, Args);
11835 }
11836 };
11837 Address VDAddr =
11838 UntiedRealAddr.isValid()
11839 ? UntiedRealAddr
11840 : Address(Addr, CGF.ConvertTypeForMem(CVD->getType()), Align);
11841 CGF.EHStack.pushCleanup<OMPAllocateCleanupTy>(
11842 NormalAndEHCleanup, FiniRTLFn, CVD->getLocation().getRawEncoding(),
11843 VDAddr, Allocator);
11844 if (UntiedRealAddr.isValid())
11845 if (auto *Region =
11846 dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo))
11847 Region->emitUntiedSwitch(CGF);
11848 return VDAddr;
11849 }
11850 return UntiedAddr;
11851 }
11852
isLocalVarInUntiedTask(CodeGenFunction & CGF,const VarDecl * VD) const11853 bool CGOpenMPRuntime::isLocalVarInUntiedTask(CodeGenFunction &CGF,
11854 const VarDecl *VD) const {
11855 auto It = FunctionToUntiedTaskStackMap.find(CGF.CurFn);
11856 if (It == FunctionToUntiedTaskStackMap.end())
11857 return false;
11858 return UntiedLocalVarsStack[It->second].count(VD) > 0;
11859 }
11860
NontemporalDeclsRAII(CodeGenModule & CGM,const OMPLoopDirective & S)11861 CGOpenMPRuntime::NontemporalDeclsRAII::NontemporalDeclsRAII(
11862 CodeGenModule &CGM, const OMPLoopDirective &S)
11863 : CGM(CGM), NeedToPush(S.hasClausesOfKind<OMPNontemporalClause>()) {
11864 assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.");
11865 if (!NeedToPush)
11866 return;
11867 NontemporalDeclsSet &DS =
11868 CGM.getOpenMPRuntime().NontemporalDeclsStack.emplace_back();
11869 for (const auto *C : S.getClausesOfKind<OMPNontemporalClause>()) {
11870 for (const Stmt *Ref : C->private_refs()) {
11871 const auto *SimpleRefExpr = cast<Expr>(Ref)->IgnoreParenImpCasts();
11872 const ValueDecl *VD;
11873 if (const auto *DRE = dyn_cast<DeclRefExpr>(SimpleRefExpr)) {
11874 VD = DRE->getDecl();
11875 } else {
11876 const auto *ME = cast<MemberExpr>(SimpleRefExpr);
11877 assert((ME->isImplicitCXXThis() ||
11878 isa<CXXThisExpr>(ME->getBase()->IgnoreParenImpCasts())) &&
11879 "Expected member of current class.");
11880 VD = ME->getMemberDecl();
11881 }
11882 DS.insert(VD);
11883 }
11884 }
11885 }
11886
~NontemporalDeclsRAII()11887 CGOpenMPRuntime::NontemporalDeclsRAII::~NontemporalDeclsRAII() {
11888 if (!NeedToPush)
11889 return;
11890 CGM.getOpenMPRuntime().NontemporalDeclsStack.pop_back();
11891 }
11892
UntiedTaskLocalDeclsRAII(CodeGenFunction & CGF,const llvm::MapVector<CanonicalDeclPtr<const VarDecl>,std::pair<Address,Address>> & LocalVars)11893 CGOpenMPRuntime::UntiedTaskLocalDeclsRAII::UntiedTaskLocalDeclsRAII(
11894 CodeGenFunction &CGF,
11895 const llvm::MapVector<CanonicalDeclPtr<const VarDecl>,
11896 std::pair<Address, Address>> &LocalVars)
11897 : CGM(CGF.CGM), NeedToPush(!LocalVars.empty()) {
11898 if (!NeedToPush)
11899 return;
11900 CGM.getOpenMPRuntime().FunctionToUntiedTaskStackMap.try_emplace(
11901 CGF.CurFn, CGM.getOpenMPRuntime().UntiedLocalVarsStack.size());
11902 CGM.getOpenMPRuntime().UntiedLocalVarsStack.push_back(LocalVars);
11903 }
11904
~UntiedTaskLocalDeclsRAII()11905 CGOpenMPRuntime::UntiedTaskLocalDeclsRAII::~UntiedTaskLocalDeclsRAII() {
11906 if (!NeedToPush)
11907 return;
11908 CGM.getOpenMPRuntime().UntiedLocalVarsStack.pop_back();
11909 }
11910
isNontemporalDecl(const ValueDecl * VD) const11911 bool CGOpenMPRuntime::isNontemporalDecl(const ValueDecl *VD) const {
11912 assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.");
11913
11914 return llvm::any_of(
11915 CGM.getOpenMPRuntime().NontemporalDeclsStack,
11916 [VD](const NontemporalDeclsSet &Set) { return Set.contains(VD); });
11917 }
11918
tryToDisableInnerAnalysis(const OMPExecutableDirective & S,llvm::DenseSet<CanonicalDeclPtr<const Decl>> & NeedToAddForLPCsAsDisabled) const11919 void CGOpenMPRuntime::LastprivateConditionalRAII::tryToDisableInnerAnalysis(
11920 const OMPExecutableDirective &S,
11921 llvm::DenseSet<CanonicalDeclPtr<const Decl>> &NeedToAddForLPCsAsDisabled)
11922 const {
11923 llvm::DenseSet<CanonicalDeclPtr<const Decl>> NeedToCheckForLPCs;
11924 // Vars in target/task regions must be excluded completely.
11925 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()) ||
11926 isOpenMPTaskingDirective(S.getDirectiveKind())) {
11927 SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
11928 getOpenMPCaptureRegions(CaptureRegions, S.getDirectiveKind());
11929 const CapturedStmt *CS = S.getCapturedStmt(CaptureRegions.front());
11930 for (const CapturedStmt::Capture &Cap : CS->captures()) {
11931 if (Cap.capturesVariable() || Cap.capturesVariableByCopy())
11932 NeedToCheckForLPCs.insert(Cap.getCapturedVar());
11933 }
11934 }
11935 // Exclude vars in private clauses.
11936 for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) {
11937 for (const Expr *Ref : C->varlists()) {
11938 if (!Ref->getType()->isScalarType())
11939 continue;
11940 const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts());
11941 if (!DRE)
11942 continue;
11943 NeedToCheckForLPCs.insert(DRE->getDecl());
11944 }
11945 }
11946 for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) {
11947 for (const Expr *Ref : C->varlists()) {
11948 if (!Ref->getType()->isScalarType())
11949 continue;
11950 const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts());
11951 if (!DRE)
11952 continue;
11953 NeedToCheckForLPCs.insert(DRE->getDecl());
11954 }
11955 }
11956 for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) {
11957 for (const Expr *Ref : C->varlists()) {
11958 if (!Ref->getType()->isScalarType())
11959 continue;
11960 const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts());
11961 if (!DRE)
11962 continue;
11963 NeedToCheckForLPCs.insert(DRE->getDecl());
11964 }
11965 }
11966 for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) {
11967 for (const Expr *Ref : C->varlists()) {
11968 if (!Ref->getType()->isScalarType())
11969 continue;
11970 const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts());
11971 if (!DRE)
11972 continue;
11973 NeedToCheckForLPCs.insert(DRE->getDecl());
11974 }
11975 }
11976 for (const auto *C : S.getClausesOfKind<OMPLinearClause>()) {
11977 for (const Expr *Ref : C->varlists()) {
11978 if (!Ref->getType()->isScalarType())
11979 continue;
11980 const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts());
11981 if (!DRE)
11982 continue;
11983 NeedToCheckForLPCs.insert(DRE->getDecl());
11984 }
11985 }
11986 for (const Decl *VD : NeedToCheckForLPCs) {
11987 for (const LastprivateConditionalData &Data :
11988 llvm::reverse(CGM.getOpenMPRuntime().LastprivateConditionalStack)) {
11989 if (Data.DeclToUniqueName.count(VD) > 0) {
11990 if (!Data.Disabled)
11991 NeedToAddForLPCsAsDisabled.insert(VD);
11992 break;
11993 }
11994 }
11995 }
11996 }
11997
LastprivateConditionalRAII(CodeGenFunction & CGF,const OMPExecutableDirective & S,LValue IVLVal)11998 CGOpenMPRuntime::LastprivateConditionalRAII::LastprivateConditionalRAII(
11999 CodeGenFunction &CGF, const OMPExecutableDirective &S, LValue IVLVal)
12000 : CGM(CGF.CGM),
12001 Action((CGM.getLangOpts().OpenMP >= 50 &&
12002 llvm::any_of(S.getClausesOfKind<OMPLastprivateClause>(),
12003 [](const OMPLastprivateClause *C) {
12004 return C->getKind() ==
12005 OMPC_LASTPRIVATE_conditional;
12006 }))
12007 ? ActionToDo::PushAsLastprivateConditional
12008 : ActionToDo::DoNotPush) {
12009 assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.");
12010 if (CGM.getLangOpts().OpenMP < 50 || Action == ActionToDo::DoNotPush)
12011 return;
12012 assert(Action == ActionToDo::PushAsLastprivateConditional &&
12013 "Expected a push action.");
12014 LastprivateConditionalData &Data =
12015 CGM.getOpenMPRuntime().LastprivateConditionalStack.emplace_back();
12016 for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) {
12017 if (C->getKind() != OMPC_LASTPRIVATE_conditional)
12018 continue;
12019
12020 for (const Expr *Ref : C->varlists()) {
12021 Data.DeclToUniqueName.insert(std::make_pair(
12022 cast<DeclRefExpr>(Ref->IgnoreParenImpCasts())->getDecl(),
12023 SmallString<16>(generateUniqueName(CGM, "pl_cond", Ref))));
12024 }
12025 }
12026 Data.IVLVal = IVLVal;
12027 Data.Fn = CGF.CurFn;
12028 }
12029
LastprivateConditionalRAII(CodeGenFunction & CGF,const OMPExecutableDirective & S)12030 CGOpenMPRuntime::LastprivateConditionalRAII::LastprivateConditionalRAII(
12031 CodeGenFunction &CGF, const OMPExecutableDirective &S)
12032 : CGM(CGF.CGM), Action(ActionToDo::DoNotPush) {
12033 assert(CGM.getLangOpts().OpenMP && "Not in OpenMP mode.");
12034 if (CGM.getLangOpts().OpenMP < 50)
12035 return;
12036 llvm::DenseSet<CanonicalDeclPtr<const Decl>> NeedToAddForLPCsAsDisabled;
12037 tryToDisableInnerAnalysis(S, NeedToAddForLPCsAsDisabled);
12038 if (!NeedToAddForLPCsAsDisabled.empty()) {
12039 Action = ActionToDo::DisableLastprivateConditional;
12040 LastprivateConditionalData &Data =
12041 CGM.getOpenMPRuntime().LastprivateConditionalStack.emplace_back();
12042 for (const Decl *VD : NeedToAddForLPCsAsDisabled)
12043 Data.DeclToUniqueName.insert(std::make_pair(VD, SmallString<16>()));
12044 Data.Fn = CGF.CurFn;
12045 Data.Disabled = true;
12046 }
12047 }
12048
12049 CGOpenMPRuntime::LastprivateConditionalRAII
disable(CodeGenFunction & CGF,const OMPExecutableDirective & S)12050 CGOpenMPRuntime::LastprivateConditionalRAII::disable(
12051 CodeGenFunction &CGF, const OMPExecutableDirective &S) {
12052 return LastprivateConditionalRAII(CGF, S);
12053 }
12054
~LastprivateConditionalRAII()12055 CGOpenMPRuntime::LastprivateConditionalRAII::~LastprivateConditionalRAII() {
12056 if (CGM.getLangOpts().OpenMP < 50)
12057 return;
12058 if (Action == ActionToDo::DisableLastprivateConditional) {
12059 assert(CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled &&
12060 "Expected list of disabled private vars.");
12061 CGM.getOpenMPRuntime().LastprivateConditionalStack.pop_back();
12062 }
12063 if (Action == ActionToDo::PushAsLastprivateConditional) {
12064 assert(
12065 !CGM.getOpenMPRuntime().LastprivateConditionalStack.back().Disabled &&
12066 "Expected list of lastprivate conditional vars.");
12067 CGM.getOpenMPRuntime().LastprivateConditionalStack.pop_back();
12068 }
12069 }
12070
emitLastprivateConditionalInit(CodeGenFunction & CGF,const VarDecl * VD)12071 Address CGOpenMPRuntime::emitLastprivateConditionalInit(CodeGenFunction &CGF,
12072 const VarDecl *VD) {
12073 ASTContext &C = CGM.getContext();
12074 auto I = LastprivateConditionalToTypes.find(CGF.CurFn);
12075 if (I == LastprivateConditionalToTypes.end())
12076 I = LastprivateConditionalToTypes.try_emplace(CGF.CurFn).first;
12077 QualType NewType;
12078 const FieldDecl *VDField;
12079 const FieldDecl *FiredField;
12080 LValue BaseLVal;
12081 auto VI = I->getSecond().find(VD);
12082 if (VI == I->getSecond().end()) {
12083 RecordDecl *RD = C.buildImplicitRecord("lasprivate.conditional");
12084 RD->startDefinition();
12085 VDField = addFieldToRecordDecl(C, RD, VD->getType().getNonReferenceType());
12086 FiredField = addFieldToRecordDecl(C, RD, C.CharTy);
12087 RD->completeDefinition();
12088 NewType = C.getRecordType(RD);
12089 Address Addr = CGF.CreateMemTemp(NewType, C.getDeclAlign(VD), VD->getName());
12090 BaseLVal = CGF.MakeAddrLValue(Addr, NewType, AlignmentSource::Decl);
12091 I->getSecond().try_emplace(VD, NewType, VDField, FiredField, BaseLVal);
12092 } else {
12093 NewType = std::get<0>(VI->getSecond());
12094 VDField = std::get<1>(VI->getSecond());
12095 FiredField = std::get<2>(VI->getSecond());
12096 BaseLVal = std::get<3>(VI->getSecond());
12097 }
12098 LValue FiredLVal =
12099 CGF.EmitLValueForField(BaseLVal, FiredField);
12100 CGF.EmitStoreOfScalar(
12101 llvm::ConstantInt::getNullValue(CGF.ConvertTypeForMem(C.CharTy)),
12102 FiredLVal);
12103 return CGF.EmitLValueForField(BaseLVal, VDField).getAddress(CGF);
12104 }
12105
12106 namespace {
12107 /// Checks if the lastprivate conditional variable is referenced in LHS.
12108 class LastprivateConditionalRefChecker final
12109 : public ConstStmtVisitor<LastprivateConditionalRefChecker, bool> {
12110 ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM;
12111 const Expr *FoundE = nullptr;
12112 const Decl *FoundD = nullptr;
12113 StringRef UniqueDeclName;
12114 LValue IVLVal;
12115 llvm::Function *FoundFn = nullptr;
12116 SourceLocation Loc;
12117
12118 public:
VisitDeclRefExpr(const DeclRefExpr * E)12119 bool VisitDeclRefExpr(const DeclRefExpr *E) {
12120 for (const CGOpenMPRuntime::LastprivateConditionalData &D :
12121 llvm::reverse(LPM)) {
12122 auto It = D.DeclToUniqueName.find(E->getDecl());
12123 if (It == D.DeclToUniqueName.end())
12124 continue;
12125 if (D.Disabled)
12126 return false;
12127 FoundE = E;
12128 FoundD = E->getDecl()->getCanonicalDecl();
12129 UniqueDeclName = It->second;
12130 IVLVal = D.IVLVal;
12131 FoundFn = D.Fn;
12132 break;
12133 }
12134 return FoundE == E;
12135 }
VisitMemberExpr(const MemberExpr * E)12136 bool VisitMemberExpr(const MemberExpr *E) {
12137 if (!CodeGenFunction::IsWrappedCXXThis(E->getBase()))
12138 return false;
12139 for (const CGOpenMPRuntime::LastprivateConditionalData &D :
12140 llvm::reverse(LPM)) {
12141 auto It = D.DeclToUniqueName.find(E->getMemberDecl());
12142 if (It == D.DeclToUniqueName.end())
12143 continue;
12144 if (D.Disabled)
12145 return false;
12146 FoundE = E;
12147 FoundD = E->getMemberDecl()->getCanonicalDecl();
12148 UniqueDeclName = It->second;
12149 IVLVal = D.IVLVal;
12150 FoundFn = D.Fn;
12151 break;
12152 }
12153 return FoundE == E;
12154 }
VisitStmt(const Stmt * S)12155 bool VisitStmt(const Stmt *S) {
12156 for (const Stmt *Child : S->children()) {
12157 if (!Child)
12158 continue;
12159 if (const auto *E = dyn_cast<Expr>(Child))
12160 if (!E->isGLValue())
12161 continue;
12162 if (Visit(Child))
12163 return true;
12164 }
12165 return false;
12166 }
LastprivateConditionalRefChecker(ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM)12167 explicit LastprivateConditionalRefChecker(
12168 ArrayRef<CGOpenMPRuntime::LastprivateConditionalData> LPM)
12169 : LPM(LPM) {}
12170 std::tuple<const Expr *, const Decl *, StringRef, LValue, llvm::Function *>
getFoundData() const12171 getFoundData() const {
12172 return std::make_tuple(FoundE, FoundD, UniqueDeclName, IVLVal, FoundFn);
12173 }
12174 };
12175 } // namespace
12176
emitLastprivateConditionalUpdate(CodeGenFunction & CGF,LValue IVLVal,StringRef UniqueDeclName,LValue LVal,SourceLocation Loc)12177 void CGOpenMPRuntime::emitLastprivateConditionalUpdate(CodeGenFunction &CGF,
12178 LValue IVLVal,
12179 StringRef UniqueDeclName,
12180 LValue LVal,
12181 SourceLocation Loc) {
12182 // Last updated loop counter for the lastprivate conditional var.
12183 // int<xx> last_iv = 0;
12184 llvm::Type *LLIVTy = CGF.ConvertTypeForMem(IVLVal.getType());
12185 llvm::Constant *LastIV = OMPBuilder.getOrCreateInternalVariable(
12186 LLIVTy, getName({UniqueDeclName, "iv"}));
12187 cast<llvm::GlobalVariable>(LastIV)->setAlignment(
12188 IVLVal.getAlignment().getAsAlign());
12189 LValue LastIVLVal = CGF.MakeNaturalAlignAddrLValue(LastIV, IVLVal.getType());
12190
12191 // Last value of the lastprivate conditional.
12192 // decltype(priv_a) last_a;
12193 llvm::GlobalVariable *Last = OMPBuilder.getOrCreateInternalVariable(
12194 CGF.ConvertTypeForMem(LVal.getType()), UniqueDeclName);
12195 Last->setAlignment(LVal.getAlignment().getAsAlign());
12196 LValue LastLVal = CGF.MakeAddrLValue(
12197 Address(Last, Last->getValueType(), LVal.getAlignment()), LVal.getType());
12198
12199 // Global loop counter. Required to handle inner parallel-for regions.
12200 // iv
12201 llvm::Value *IVVal = CGF.EmitLoadOfScalar(IVLVal, Loc);
12202
12203 // #pragma omp critical(a)
12204 // if (last_iv <= iv) {
12205 // last_iv = iv;
12206 // last_a = priv_a;
12207 // }
12208 auto &&CodeGen = [&LastIVLVal, &IVLVal, IVVal, &LVal, &LastLVal,
12209 Loc](CodeGenFunction &CGF, PrePostActionTy &Action) {
12210 Action.Enter(CGF);
12211 llvm::Value *LastIVVal = CGF.EmitLoadOfScalar(LastIVLVal, Loc);
12212 // (last_iv <= iv) ? Check if the variable is updated and store new
12213 // value in global var.
12214 llvm::Value *CmpRes;
12215 if (IVLVal.getType()->isSignedIntegerType()) {
12216 CmpRes = CGF.Builder.CreateICmpSLE(LastIVVal, IVVal);
12217 } else {
12218 assert(IVLVal.getType()->isUnsignedIntegerType() &&
12219 "Loop iteration variable must be integer.");
12220 CmpRes = CGF.Builder.CreateICmpULE(LastIVVal, IVVal);
12221 }
12222 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("lp_cond_then");
12223 llvm::BasicBlock *ExitBB = CGF.createBasicBlock("lp_cond_exit");
12224 CGF.Builder.CreateCondBr(CmpRes, ThenBB, ExitBB);
12225 // {
12226 CGF.EmitBlock(ThenBB);
12227
12228 // last_iv = iv;
12229 CGF.EmitStoreOfScalar(IVVal, LastIVLVal);
12230
12231 // last_a = priv_a;
12232 switch (CGF.getEvaluationKind(LVal.getType())) {
12233 case TEK_Scalar: {
12234 llvm::Value *PrivVal = CGF.EmitLoadOfScalar(LVal, Loc);
12235 CGF.EmitStoreOfScalar(PrivVal, LastLVal);
12236 break;
12237 }
12238 case TEK_Complex: {
12239 CodeGenFunction::ComplexPairTy PrivVal = CGF.EmitLoadOfComplex(LVal, Loc);
12240 CGF.EmitStoreOfComplex(PrivVal, LastLVal, /*isInit=*/false);
12241 break;
12242 }
12243 case TEK_Aggregate:
12244 llvm_unreachable(
12245 "Aggregates are not supported in lastprivate conditional.");
12246 }
12247 // }
12248 CGF.EmitBranch(ExitBB);
12249 // There is no need to emit line number for unconditional branch.
12250 (void)ApplyDebugLocation::CreateEmpty(CGF);
12251 CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
12252 };
12253
12254 if (CGM.getLangOpts().OpenMPSimd) {
12255 // Do not emit as a critical region as no parallel region could be emitted.
12256 RegionCodeGenTy ThenRCG(CodeGen);
12257 ThenRCG(CGF);
12258 } else {
12259 emitCriticalRegion(CGF, UniqueDeclName, CodeGen, Loc);
12260 }
12261 }
12262
checkAndEmitLastprivateConditional(CodeGenFunction & CGF,const Expr * LHS)12263 void CGOpenMPRuntime::checkAndEmitLastprivateConditional(CodeGenFunction &CGF,
12264 const Expr *LHS) {
12265 if (CGF.getLangOpts().OpenMP < 50 || LastprivateConditionalStack.empty())
12266 return;
12267 LastprivateConditionalRefChecker Checker(LastprivateConditionalStack);
12268 if (!Checker.Visit(LHS))
12269 return;
12270 const Expr *FoundE;
12271 const Decl *FoundD;
12272 StringRef UniqueDeclName;
12273 LValue IVLVal;
12274 llvm::Function *FoundFn;
12275 std::tie(FoundE, FoundD, UniqueDeclName, IVLVal, FoundFn) =
12276 Checker.getFoundData();
12277 if (FoundFn != CGF.CurFn) {
12278 // Special codegen for inner parallel regions.
12279 // ((struct.lastprivate.conditional*)&priv_a)->Fired = 1;
12280 auto It = LastprivateConditionalToTypes[FoundFn].find(FoundD);
12281 assert(It != LastprivateConditionalToTypes[FoundFn].end() &&
12282 "Lastprivate conditional is not found in outer region.");
12283 QualType StructTy = std::get<0>(It->getSecond());
12284 const FieldDecl* FiredDecl = std::get<2>(It->getSecond());
12285 LValue PrivLVal = CGF.EmitLValue(FoundE);
12286 Address StructAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
12287 PrivLVal.getAddress(CGF),
12288 CGF.ConvertTypeForMem(CGF.getContext().getPointerType(StructTy)),
12289 CGF.ConvertTypeForMem(StructTy));
12290 LValue BaseLVal =
12291 CGF.MakeAddrLValue(StructAddr, StructTy, AlignmentSource::Decl);
12292 LValue FiredLVal = CGF.EmitLValueForField(BaseLVal, FiredDecl);
12293 CGF.EmitAtomicStore(RValue::get(llvm::ConstantInt::get(
12294 CGF.ConvertTypeForMem(FiredDecl->getType()), 1)),
12295 FiredLVal, llvm::AtomicOrdering::Unordered,
12296 /*IsVolatile=*/true, /*isInit=*/false);
12297 return;
12298 }
12299
12300 // Private address of the lastprivate conditional in the current context.
12301 // priv_a
12302 LValue LVal = CGF.EmitLValue(FoundE);
12303 emitLastprivateConditionalUpdate(CGF, IVLVal, UniqueDeclName, LVal,
12304 FoundE->getExprLoc());
12305 }
12306
checkAndEmitSharedLastprivateConditional(CodeGenFunction & CGF,const OMPExecutableDirective & D,const llvm::DenseSet<CanonicalDeclPtr<const VarDecl>> & IgnoredDecls)12307 void CGOpenMPRuntime::checkAndEmitSharedLastprivateConditional(
12308 CodeGenFunction &CGF, const OMPExecutableDirective &D,
12309 const llvm::DenseSet<CanonicalDeclPtr<const VarDecl>> &IgnoredDecls) {
12310 if (CGF.getLangOpts().OpenMP < 50 || LastprivateConditionalStack.empty())
12311 return;
12312 auto Range = llvm::reverse(LastprivateConditionalStack);
12313 auto It = llvm::find_if(
12314 Range, [](const LastprivateConditionalData &D) { return !D.Disabled; });
12315 if (It == Range.end() || It->Fn != CGF.CurFn)
12316 return;
12317 auto LPCI = LastprivateConditionalToTypes.find(It->Fn);
12318 assert(LPCI != LastprivateConditionalToTypes.end() &&
12319 "Lastprivates must be registered already.");
12320 SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
12321 getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind());
12322 const CapturedStmt *CS = D.getCapturedStmt(CaptureRegions.back());
12323 for (const auto &Pair : It->DeclToUniqueName) {
12324 const auto *VD = cast<VarDecl>(Pair.first->getCanonicalDecl());
12325 if (!CS->capturesVariable(VD) || IgnoredDecls.contains(VD))
12326 continue;
12327 auto I = LPCI->getSecond().find(Pair.first);
12328 assert(I != LPCI->getSecond().end() &&
12329 "Lastprivate must be rehistered already.");
12330 // bool Cmp = priv_a.Fired != 0;
12331 LValue BaseLVal = std::get<3>(I->getSecond());
12332 LValue FiredLVal =
12333 CGF.EmitLValueForField(BaseLVal, std::get<2>(I->getSecond()));
12334 llvm::Value *Res = CGF.EmitLoadOfScalar(FiredLVal, D.getBeginLoc());
12335 llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Res);
12336 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("lpc.then");
12337 llvm::BasicBlock *DoneBB = CGF.createBasicBlock("lpc.done");
12338 // if (Cmp) {
12339 CGF.Builder.CreateCondBr(Cmp, ThenBB, DoneBB);
12340 CGF.EmitBlock(ThenBB);
12341 Address Addr = CGF.GetAddrOfLocalVar(VD);
12342 LValue LVal;
12343 if (VD->getType()->isReferenceType())
12344 LVal = CGF.EmitLoadOfReferenceLValue(Addr, VD->getType(),
12345 AlignmentSource::Decl);
12346 else
12347 LVal = CGF.MakeAddrLValue(Addr, VD->getType().getNonReferenceType(),
12348 AlignmentSource::Decl);
12349 emitLastprivateConditionalUpdate(CGF, It->IVLVal, Pair.second, LVal,
12350 D.getBeginLoc());
12351 auto AL = ApplyDebugLocation::CreateArtificial(CGF);
12352 CGF.EmitBlock(DoneBB, /*IsFinal=*/true);
12353 // }
12354 }
12355 }
12356
emitLastprivateConditionalFinalUpdate(CodeGenFunction & CGF,LValue PrivLVal,const VarDecl * VD,SourceLocation Loc)12357 void CGOpenMPRuntime::emitLastprivateConditionalFinalUpdate(
12358 CodeGenFunction &CGF, LValue PrivLVal, const VarDecl *VD,
12359 SourceLocation Loc) {
12360 if (CGF.getLangOpts().OpenMP < 50)
12361 return;
12362 auto It = LastprivateConditionalStack.back().DeclToUniqueName.find(VD);
12363 assert(It != LastprivateConditionalStack.back().DeclToUniqueName.end() &&
12364 "Unknown lastprivate conditional variable.");
12365 StringRef UniqueName = It->second;
12366 llvm::GlobalVariable *GV = CGM.getModule().getNamedGlobal(UniqueName);
12367 // The variable was not updated in the region - exit.
12368 if (!GV)
12369 return;
12370 LValue LPLVal = CGF.MakeAddrLValue(
12371 Address(GV, GV->getValueType(), PrivLVal.getAlignment()),
12372 PrivLVal.getType().getNonReferenceType());
12373 llvm::Value *Res = CGF.EmitLoadOfScalar(LPLVal, Loc);
12374 CGF.EmitStoreOfScalar(Res, PrivLVal);
12375 }
12376
emitParallelOutlinedFunction(const OMPExecutableDirective & D,const VarDecl * ThreadIDVar,OpenMPDirectiveKind InnermostKind,const RegionCodeGenTy & CodeGen)12377 llvm::Function *CGOpenMPSIMDRuntime::emitParallelOutlinedFunction(
12378 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
12379 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
12380 llvm_unreachable("Not supported in SIMD-only mode");
12381 }
12382
emitTeamsOutlinedFunction(const OMPExecutableDirective & D,const VarDecl * ThreadIDVar,OpenMPDirectiveKind InnermostKind,const RegionCodeGenTy & CodeGen)12383 llvm::Function *CGOpenMPSIMDRuntime::emitTeamsOutlinedFunction(
12384 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
12385 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
12386 llvm_unreachable("Not supported in SIMD-only mode");
12387 }
12388
emitTaskOutlinedFunction(const OMPExecutableDirective & D,const VarDecl * ThreadIDVar,const VarDecl * PartIDVar,const VarDecl * TaskTVar,OpenMPDirectiveKind InnermostKind,const RegionCodeGenTy & CodeGen,bool Tied,unsigned & NumberOfParts)12389 llvm::Function *CGOpenMPSIMDRuntime::emitTaskOutlinedFunction(
12390 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
12391 const VarDecl *PartIDVar, const VarDecl *TaskTVar,
12392 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
12393 bool Tied, unsigned &NumberOfParts) {
12394 llvm_unreachable("Not supported in SIMD-only mode");
12395 }
12396
emitParallelCall(CodeGenFunction & CGF,SourceLocation Loc,llvm::Function * OutlinedFn,ArrayRef<llvm::Value * > CapturedVars,const Expr * IfCond,llvm::Value * NumThreads)12397 void CGOpenMPSIMDRuntime::emitParallelCall(CodeGenFunction &CGF,
12398 SourceLocation Loc,
12399 llvm::Function *OutlinedFn,
12400 ArrayRef<llvm::Value *> CapturedVars,
12401 const Expr *IfCond,
12402 llvm::Value *NumThreads) {
12403 llvm_unreachable("Not supported in SIMD-only mode");
12404 }
12405
emitCriticalRegion(CodeGenFunction & CGF,StringRef CriticalName,const RegionCodeGenTy & CriticalOpGen,SourceLocation Loc,const Expr * Hint)12406 void CGOpenMPSIMDRuntime::emitCriticalRegion(
12407 CodeGenFunction &CGF, StringRef CriticalName,
12408 const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
12409 const Expr *Hint) {
12410 llvm_unreachable("Not supported in SIMD-only mode");
12411 }
12412
emitMasterRegion(CodeGenFunction & CGF,const RegionCodeGenTy & MasterOpGen,SourceLocation Loc)12413 void CGOpenMPSIMDRuntime::emitMasterRegion(CodeGenFunction &CGF,
12414 const RegionCodeGenTy &MasterOpGen,
12415 SourceLocation Loc) {
12416 llvm_unreachable("Not supported in SIMD-only mode");
12417 }
12418
emitMaskedRegion(CodeGenFunction & CGF,const RegionCodeGenTy & MasterOpGen,SourceLocation Loc,const Expr * Filter)12419 void CGOpenMPSIMDRuntime::emitMaskedRegion(CodeGenFunction &CGF,
12420 const RegionCodeGenTy &MasterOpGen,
12421 SourceLocation Loc,
12422 const Expr *Filter) {
12423 llvm_unreachable("Not supported in SIMD-only mode");
12424 }
12425
emitTaskyieldCall(CodeGenFunction & CGF,SourceLocation Loc)12426 void CGOpenMPSIMDRuntime::emitTaskyieldCall(CodeGenFunction &CGF,
12427 SourceLocation Loc) {
12428 llvm_unreachable("Not supported in SIMD-only mode");
12429 }
12430
emitTaskgroupRegion(CodeGenFunction & CGF,const RegionCodeGenTy & TaskgroupOpGen,SourceLocation Loc)12431 void CGOpenMPSIMDRuntime::emitTaskgroupRegion(
12432 CodeGenFunction &CGF, const RegionCodeGenTy &TaskgroupOpGen,
12433 SourceLocation Loc) {
12434 llvm_unreachable("Not supported in SIMD-only mode");
12435 }
12436
emitSingleRegion(CodeGenFunction & CGF,const RegionCodeGenTy & SingleOpGen,SourceLocation Loc,ArrayRef<const Expr * > CopyprivateVars,ArrayRef<const Expr * > DestExprs,ArrayRef<const Expr * > SrcExprs,ArrayRef<const Expr * > AssignmentOps)12437 void CGOpenMPSIMDRuntime::emitSingleRegion(
12438 CodeGenFunction &CGF, const RegionCodeGenTy &SingleOpGen,
12439 SourceLocation Loc, ArrayRef<const Expr *> CopyprivateVars,
12440 ArrayRef<const Expr *> DestExprs, ArrayRef<const Expr *> SrcExprs,
12441 ArrayRef<const Expr *> AssignmentOps) {
12442 llvm_unreachable("Not supported in SIMD-only mode");
12443 }
12444
emitOrderedRegion(CodeGenFunction & CGF,const RegionCodeGenTy & OrderedOpGen,SourceLocation Loc,bool IsThreads)12445 void CGOpenMPSIMDRuntime::emitOrderedRegion(CodeGenFunction &CGF,
12446 const RegionCodeGenTy &OrderedOpGen,
12447 SourceLocation Loc,
12448 bool IsThreads) {
12449 llvm_unreachable("Not supported in SIMD-only mode");
12450 }
12451
emitBarrierCall(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind Kind,bool EmitChecks,bool ForceSimpleCall)12452 void CGOpenMPSIMDRuntime::emitBarrierCall(CodeGenFunction &CGF,
12453 SourceLocation Loc,
12454 OpenMPDirectiveKind Kind,
12455 bool EmitChecks,
12456 bool ForceSimpleCall) {
12457 llvm_unreachable("Not supported in SIMD-only mode");
12458 }
12459
emitForDispatchInit(CodeGenFunction & CGF,SourceLocation Loc,const OpenMPScheduleTy & ScheduleKind,unsigned IVSize,bool IVSigned,bool Ordered,const DispatchRTInput & DispatchValues)12460 void CGOpenMPSIMDRuntime::emitForDispatchInit(
12461 CodeGenFunction &CGF, SourceLocation Loc,
12462 const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned,
12463 bool Ordered, const DispatchRTInput &DispatchValues) {
12464 llvm_unreachable("Not supported in SIMD-only mode");
12465 }
12466
emitForStaticInit(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind DKind,const OpenMPScheduleTy & ScheduleKind,const StaticRTInput & Values)12467 void CGOpenMPSIMDRuntime::emitForStaticInit(
12468 CodeGenFunction &CGF, SourceLocation Loc, OpenMPDirectiveKind DKind,
12469 const OpenMPScheduleTy &ScheduleKind, const StaticRTInput &Values) {
12470 llvm_unreachable("Not supported in SIMD-only mode");
12471 }
12472
emitDistributeStaticInit(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDistScheduleClauseKind SchedKind,const StaticRTInput & Values)12473 void CGOpenMPSIMDRuntime::emitDistributeStaticInit(
12474 CodeGenFunction &CGF, SourceLocation Loc,
12475 OpenMPDistScheduleClauseKind SchedKind, const StaticRTInput &Values) {
12476 llvm_unreachable("Not supported in SIMD-only mode");
12477 }
12478
emitForOrderedIterationEnd(CodeGenFunction & CGF,SourceLocation Loc,unsigned IVSize,bool IVSigned)12479 void CGOpenMPSIMDRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF,
12480 SourceLocation Loc,
12481 unsigned IVSize,
12482 bool IVSigned) {
12483 llvm_unreachable("Not supported in SIMD-only mode");
12484 }
12485
emitForStaticFinish(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind DKind)12486 void CGOpenMPSIMDRuntime::emitForStaticFinish(CodeGenFunction &CGF,
12487 SourceLocation Loc,
12488 OpenMPDirectiveKind DKind) {
12489 llvm_unreachable("Not supported in SIMD-only mode");
12490 }
12491
emitForNext(CodeGenFunction & CGF,SourceLocation Loc,unsigned IVSize,bool IVSigned,Address IL,Address LB,Address UB,Address ST)12492 llvm::Value *CGOpenMPSIMDRuntime::emitForNext(CodeGenFunction &CGF,
12493 SourceLocation Loc,
12494 unsigned IVSize, bool IVSigned,
12495 Address IL, Address LB,
12496 Address UB, Address ST) {
12497 llvm_unreachable("Not supported in SIMD-only mode");
12498 }
12499
emitNumThreadsClause(CodeGenFunction & CGF,llvm::Value * NumThreads,SourceLocation Loc)12500 void CGOpenMPSIMDRuntime::emitNumThreadsClause(CodeGenFunction &CGF,
12501 llvm::Value *NumThreads,
12502 SourceLocation Loc) {
12503 llvm_unreachable("Not supported in SIMD-only mode");
12504 }
12505
emitProcBindClause(CodeGenFunction & CGF,ProcBindKind ProcBind,SourceLocation Loc)12506 void CGOpenMPSIMDRuntime::emitProcBindClause(CodeGenFunction &CGF,
12507 ProcBindKind ProcBind,
12508 SourceLocation Loc) {
12509 llvm_unreachable("Not supported in SIMD-only mode");
12510 }
12511
getAddrOfThreadPrivate(CodeGenFunction & CGF,const VarDecl * VD,Address VDAddr,SourceLocation Loc)12512 Address CGOpenMPSIMDRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF,
12513 const VarDecl *VD,
12514 Address VDAddr,
12515 SourceLocation Loc) {
12516 llvm_unreachable("Not supported in SIMD-only mode");
12517 }
12518
emitThreadPrivateVarDefinition(const VarDecl * VD,Address VDAddr,SourceLocation Loc,bool PerformInit,CodeGenFunction * CGF)12519 llvm::Function *CGOpenMPSIMDRuntime::emitThreadPrivateVarDefinition(
12520 const VarDecl *VD, Address VDAddr, SourceLocation Loc, bool PerformInit,
12521 CodeGenFunction *CGF) {
12522 llvm_unreachable("Not supported in SIMD-only mode");
12523 }
12524
getAddrOfArtificialThreadPrivate(CodeGenFunction & CGF,QualType VarType,StringRef Name)12525 Address CGOpenMPSIMDRuntime::getAddrOfArtificialThreadPrivate(
12526 CodeGenFunction &CGF, QualType VarType, StringRef Name) {
12527 llvm_unreachable("Not supported in SIMD-only mode");
12528 }
12529
emitFlush(CodeGenFunction & CGF,ArrayRef<const Expr * > Vars,SourceLocation Loc,llvm::AtomicOrdering AO)12530 void CGOpenMPSIMDRuntime::emitFlush(CodeGenFunction &CGF,
12531 ArrayRef<const Expr *> Vars,
12532 SourceLocation Loc,
12533 llvm::AtomicOrdering AO) {
12534 llvm_unreachable("Not supported in SIMD-only mode");
12535 }
12536
emitTaskCall(CodeGenFunction & CGF,SourceLocation Loc,const OMPExecutableDirective & D,llvm::Function * TaskFunction,QualType SharedsTy,Address Shareds,const Expr * IfCond,const OMPTaskDataTy & Data)12537 void CGOpenMPSIMDRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc,
12538 const OMPExecutableDirective &D,
12539 llvm::Function *TaskFunction,
12540 QualType SharedsTy, Address Shareds,
12541 const Expr *IfCond,
12542 const OMPTaskDataTy &Data) {
12543 llvm_unreachable("Not supported in SIMD-only mode");
12544 }
12545
emitTaskLoopCall(CodeGenFunction & CGF,SourceLocation Loc,const OMPLoopDirective & D,llvm::Function * TaskFunction,QualType SharedsTy,Address Shareds,const Expr * IfCond,const OMPTaskDataTy & Data)12546 void CGOpenMPSIMDRuntime::emitTaskLoopCall(
12547 CodeGenFunction &CGF, SourceLocation Loc, const OMPLoopDirective &D,
12548 llvm::Function *TaskFunction, QualType SharedsTy, Address Shareds,
12549 const Expr *IfCond, const OMPTaskDataTy &Data) {
12550 llvm_unreachable("Not supported in SIMD-only mode");
12551 }
12552
emitReduction(CodeGenFunction & CGF,SourceLocation Loc,ArrayRef<const Expr * > Privates,ArrayRef<const Expr * > LHSExprs,ArrayRef<const Expr * > RHSExprs,ArrayRef<const Expr * > ReductionOps,ReductionOptionsTy Options)12553 void CGOpenMPSIMDRuntime::emitReduction(
12554 CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
12555 ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
12556 ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
12557 assert(Options.SimpleReduction && "Only simple reduction is expected.");
12558 CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
12559 ReductionOps, Options);
12560 }
12561
emitTaskReductionInit(CodeGenFunction & CGF,SourceLocation Loc,ArrayRef<const Expr * > LHSExprs,ArrayRef<const Expr * > RHSExprs,const OMPTaskDataTy & Data)12562 llvm::Value *CGOpenMPSIMDRuntime::emitTaskReductionInit(
12563 CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs,
12564 ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) {
12565 llvm_unreachable("Not supported in SIMD-only mode");
12566 }
12567
emitTaskReductionFini(CodeGenFunction & CGF,SourceLocation Loc,bool IsWorksharingReduction)12568 void CGOpenMPSIMDRuntime::emitTaskReductionFini(CodeGenFunction &CGF,
12569 SourceLocation Loc,
12570 bool IsWorksharingReduction) {
12571 llvm_unreachable("Not supported in SIMD-only mode");
12572 }
12573
emitTaskReductionFixups(CodeGenFunction & CGF,SourceLocation Loc,ReductionCodeGen & RCG,unsigned N)12574 void CGOpenMPSIMDRuntime::emitTaskReductionFixups(CodeGenFunction &CGF,
12575 SourceLocation Loc,
12576 ReductionCodeGen &RCG,
12577 unsigned N) {
12578 llvm_unreachable("Not supported in SIMD-only mode");
12579 }
12580
getTaskReductionItem(CodeGenFunction & CGF,SourceLocation Loc,llvm::Value * ReductionsPtr,LValue SharedLVal)12581 Address CGOpenMPSIMDRuntime::getTaskReductionItem(CodeGenFunction &CGF,
12582 SourceLocation Loc,
12583 llvm::Value *ReductionsPtr,
12584 LValue SharedLVal) {
12585 llvm_unreachable("Not supported in SIMD-only mode");
12586 }
12587
emitTaskwaitCall(CodeGenFunction & CGF,SourceLocation Loc,const OMPTaskDataTy & Data)12588 void CGOpenMPSIMDRuntime::emitTaskwaitCall(CodeGenFunction &CGF,
12589 SourceLocation Loc,
12590 const OMPTaskDataTy &Data) {
12591 llvm_unreachable("Not supported in SIMD-only mode");
12592 }
12593
emitCancellationPointCall(CodeGenFunction & CGF,SourceLocation Loc,OpenMPDirectiveKind CancelRegion)12594 void CGOpenMPSIMDRuntime::emitCancellationPointCall(
12595 CodeGenFunction &CGF, SourceLocation Loc,
12596 OpenMPDirectiveKind CancelRegion) {
12597 llvm_unreachable("Not supported in SIMD-only mode");
12598 }
12599
emitCancelCall(CodeGenFunction & CGF,SourceLocation Loc,const Expr * IfCond,OpenMPDirectiveKind CancelRegion)12600 void CGOpenMPSIMDRuntime::emitCancelCall(CodeGenFunction &CGF,
12601 SourceLocation Loc, const Expr *IfCond,
12602 OpenMPDirectiveKind CancelRegion) {
12603 llvm_unreachable("Not supported in SIMD-only mode");
12604 }
12605
emitTargetOutlinedFunction(const OMPExecutableDirective & D,StringRef ParentName,llvm::Function * & OutlinedFn,llvm::Constant * & OutlinedFnID,bool IsOffloadEntry,const RegionCodeGenTy & CodeGen)12606 void CGOpenMPSIMDRuntime::emitTargetOutlinedFunction(
12607 const OMPExecutableDirective &D, StringRef ParentName,
12608 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
12609 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
12610 llvm_unreachable("Not supported in SIMD-only mode");
12611 }
12612
emitTargetCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,llvm::Function * OutlinedFn,llvm::Value * OutlinedFnID,const Expr * IfCond,llvm::PointerIntPair<const Expr *,2,OpenMPDeviceClauseModifier> Device,llvm::function_ref<llvm::Value * (CodeGenFunction & CGF,const OMPLoopDirective & D)> SizeEmitter)12613 void CGOpenMPSIMDRuntime::emitTargetCall(
12614 CodeGenFunction &CGF, const OMPExecutableDirective &D,
12615 llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, const Expr *IfCond,
12616 llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device,
12617 llvm::function_ref<llvm::Value *(CodeGenFunction &CGF,
12618 const OMPLoopDirective &D)>
12619 SizeEmitter) {
12620 llvm_unreachable("Not supported in SIMD-only mode");
12621 }
12622
emitTargetFunctions(GlobalDecl GD)12623 bool CGOpenMPSIMDRuntime::emitTargetFunctions(GlobalDecl GD) {
12624 llvm_unreachable("Not supported in SIMD-only mode");
12625 }
12626
emitTargetGlobalVariable(GlobalDecl GD)12627 bool CGOpenMPSIMDRuntime::emitTargetGlobalVariable(GlobalDecl GD) {
12628 llvm_unreachable("Not supported in SIMD-only mode");
12629 }
12630
emitTargetGlobal(GlobalDecl GD)12631 bool CGOpenMPSIMDRuntime::emitTargetGlobal(GlobalDecl GD) {
12632 return false;
12633 }
12634
emitTeamsCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,SourceLocation Loc,llvm::Function * OutlinedFn,ArrayRef<llvm::Value * > CapturedVars)12635 void CGOpenMPSIMDRuntime::emitTeamsCall(CodeGenFunction &CGF,
12636 const OMPExecutableDirective &D,
12637 SourceLocation Loc,
12638 llvm::Function *OutlinedFn,
12639 ArrayRef<llvm::Value *> CapturedVars) {
12640 llvm_unreachable("Not supported in SIMD-only mode");
12641 }
12642
emitNumTeamsClause(CodeGenFunction & CGF,const Expr * NumTeams,const Expr * ThreadLimit,SourceLocation Loc)12643 void CGOpenMPSIMDRuntime::emitNumTeamsClause(CodeGenFunction &CGF,
12644 const Expr *NumTeams,
12645 const Expr *ThreadLimit,
12646 SourceLocation Loc) {
12647 llvm_unreachable("Not supported in SIMD-only mode");
12648 }
12649
emitTargetDataCalls(CodeGenFunction & CGF,const OMPExecutableDirective & D,const Expr * IfCond,const Expr * Device,const RegionCodeGenTy & CodeGen,CGOpenMPRuntime::TargetDataInfo & Info)12650 void CGOpenMPSIMDRuntime::emitTargetDataCalls(
12651 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
12652 const Expr *Device, const RegionCodeGenTy &CodeGen,
12653 CGOpenMPRuntime::TargetDataInfo &Info) {
12654 llvm_unreachable("Not supported in SIMD-only mode");
12655 }
12656
emitTargetDataStandAloneCall(CodeGenFunction & CGF,const OMPExecutableDirective & D,const Expr * IfCond,const Expr * Device)12657 void CGOpenMPSIMDRuntime::emitTargetDataStandAloneCall(
12658 CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond,
12659 const Expr *Device) {
12660 llvm_unreachable("Not supported in SIMD-only mode");
12661 }
12662
emitDoacrossInit(CodeGenFunction & CGF,const OMPLoopDirective & D,ArrayRef<Expr * > NumIterations)12663 void CGOpenMPSIMDRuntime::emitDoacrossInit(CodeGenFunction &CGF,
12664 const OMPLoopDirective &D,
12665 ArrayRef<Expr *> NumIterations) {
12666 llvm_unreachable("Not supported in SIMD-only mode");
12667 }
12668
emitDoacrossOrdered(CodeGenFunction & CGF,const OMPDependClause * C)12669 void CGOpenMPSIMDRuntime::emitDoacrossOrdered(CodeGenFunction &CGF,
12670 const OMPDependClause *C) {
12671 llvm_unreachable("Not supported in SIMD-only mode");
12672 }
12673
12674 const VarDecl *
translateParameter(const FieldDecl * FD,const VarDecl * NativeParam) const12675 CGOpenMPSIMDRuntime::translateParameter(const FieldDecl *FD,
12676 const VarDecl *NativeParam) const {
12677 llvm_unreachable("Not supported in SIMD-only mode");
12678 }
12679
12680 Address
getParameterAddress(CodeGenFunction & CGF,const VarDecl * NativeParam,const VarDecl * TargetParam) const12681 CGOpenMPSIMDRuntime::getParameterAddress(CodeGenFunction &CGF,
12682 const VarDecl *NativeParam,
12683 const VarDecl *TargetParam) const {
12684 llvm_unreachable("Not supported in SIMD-only mode");
12685 }
12686