xref: /freebsd/contrib/llvm-project/clang/lib/Sema/SemaCUDA.cpp (revision f7c32ed617858bcd22f8d1b03199099d50125721)
1 //===--- SemaCUDA.cpp - Semantic Analysis for CUDA constructs -------------===//
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 /// \file
9 /// This file implements semantic analysis for CUDA constructs.
10 ///
11 //===----------------------------------------------------------------------===//
12 
13 #include "clang/AST/ASTContext.h"
14 #include "clang/AST/Decl.h"
15 #include "clang/AST/ExprCXX.h"
16 #include "clang/Basic/Cuda.h"
17 #include "clang/Basic/TargetInfo.h"
18 #include "clang/Lex/Preprocessor.h"
19 #include "clang/Sema/Lookup.h"
20 #include "clang/Sema/ScopeInfo.h"
21 #include "clang/Sema/Sema.h"
22 #include "clang/Sema/SemaDiagnostic.h"
23 #include "clang/Sema/SemaInternal.h"
24 #include "clang/Sema/Template.h"
25 #include "llvm/ADT/Optional.h"
26 #include "llvm/ADT/SmallVector.h"
27 using namespace clang;
28 
29 template <typename AttrT> static bool hasExplicitAttr(const VarDecl *D) {
30   if (!D)
31     return false;
32   if (auto *A = D->getAttr<AttrT>())
33     return !A->isImplicit();
34   return false;
35 }
36 
37 void Sema::PushForceCUDAHostDevice() {
38   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
39   ForceCUDAHostDeviceDepth++;
40 }
41 
42 bool Sema::PopForceCUDAHostDevice() {
43   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
44   if (ForceCUDAHostDeviceDepth == 0)
45     return false;
46   ForceCUDAHostDeviceDepth--;
47   return true;
48 }
49 
50 ExprResult Sema::ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc,
51                                          MultiExprArg ExecConfig,
52                                          SourceLocation GGGLoc) {
53   FunctionDecl *ConfigDecl = Context.getcudaConfigureCallDecl();
54   if (!ConfigDecl)
55     return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use)
56                      << getCudaConfigureFuncName());
57   QualType ConfigQTy = ConfigDecl->getType();
58 
59   DeclRefExpr *ConfigDR = new (Context)
60       DeclRefExpr(Context, ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc);
61   MarkFunctionReferenced(LLLLoc, ConfigDecl);
62 
63   return BuildCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr,
64                        /*IsExecConfig=*/true);
65 }
66 
67 Sema::CUDAFunctionTarget
68 Sema::IdentifyCUDATarget(const ParsedAttributesView &Attrs) {
69   bool HasHostAttr = false;
70   bool HasDeviceAttr = false;
71   bool HasGlobalAttr = false;
72   bool HasInvalidTargetAttr = false;
73   for (const ParsedAttr &AL : Attrs) {
74     switch (AL.getKind()) {
75     case ParsedAttr::AT_CUDAGlobal:
76       HasGlobalAttr = true;
77       break;
78     case ParsedAttr::AT_CUDAHost:
79       HasHostAttr = true;
80       break;
81     case ParsedAttr::AT_CUDADevice:
82       HasDeviceAttr = true;
83       break;
84     case ParsedAttr::AT_CUDAInvalidTarget:
85       HasInvalidTargetAttr = true;
86       break;
87     default:
88       break;
89     }
90   }
91 
92   if (HasInvalidTargetAttr)
93     return CFT_InvalidTarget;
94 
95   if (HasGlobalAttr)
96     return CFT_Global;
97 
98   if (HasHostAttr && HasDeviceAttr)
99     return CFT_HostDevice;
100 
101   if (HasDeviceAttr)
102     return CFT_Device;
103 
104   return CFT_Host;
105 }
106 
107 template <typename A>
108 static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
109   return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
110            return isa<A>(Attribute) &&
111                   !(IgnoreImplicitAttr && Attribute->isImplicit());
112          });
113 }
114 
115 /// IdentifyCUDATarget - Determine the CUDA compilation target for this function
116 Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D,
117                                                   bool IgnoreImplicitHDAttr) {
118   // Code that lives outside a function is run on the host.
119   if (D == nullptr)
120     return CFT_Host;
121 
122   if (D->hasAttr<CUDAInvalidTargetAttr>())
123     return CFT_InvalidTarget;
124 
125   if (D->hasAttr<CUDAGlobalAttr>())
126     return CFT_Global;
127 
128   if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) {
129     if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr))
130       return CFT_HostDevice;
131     return CFT_Device;
132   } else if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) {
133     return CFT_Host;
134   } else if ((D->isImplicit() || !D->isUserProvided()) &&
135              !IgnoreImplicitHDAttr) {
136     // Some implicit declarations (like intrinsic functions) are not marked.
137     // Set the most lenient target on them for maximal flexibility.
138     return CFT_HostDevice;
139   }
140 
141   return CFT_Host;
142 }
143 
144 /// IdentifyTarget - Determine the CUDA compilation target for this variable.
145 Sema::CUDAVariableTarget Sema::IdentifyCUDATarget(const VarDecl *Var) {
146   if (Var->hasAttr<HIPManagedAttr>())
147     return CVT_Unified;
148   if (Var->isConstexpr() && !hasExplicitAttr<CUDAConstantAttr>(Var))
149     return CVT_Both;
150   if (Var->getType().isConstQualified() && Var->hasAttr<CUDAConstantAttr>() &&
151       !hasExplicitAttr<CUDAConstantAttr>(Var))
152     return CVT_Both;
153   if (Var->hasAttr<CUDADeviceAttr>() || Var->hasAttr<CUDAConstantAttr>() ||
154       Var->hasAttr<CUDASharedAttr>() ||
155       Var->getType()->isCUDADeviceBuiltinSurfaceType() ||
156       Var->getType()->isCUDADeviceBuiltinTextureType())
157     return CVT_Device;
158   // Function-scope static variable without explicit device or constant
159   // attribute are emitted
160   //  - on both sides in host device functions
161   //  - on device side in device or global functions
162   if (auto *FD = dyn_cast<FunctionDecl>(Var->getDeclContext())) {
163     switch (IdentifyCUDATarget(FD)) {
164     case CFT_HostDevice:
165       return CVT_Both;
166     case CFT_Device:
167     case CFT_Global:
168       return CVT_Device;
169     default:
170       return CVT_Host;
171     }
172   }
173   return CVT_Host;
174 }
175 
176 // * CUDA Call preference table
177 //
178 // F - from,
179 // T - to
180 // Ph - preference in host mode
181 // Pd - preference in device mode
182 // H  - handled in (x)
183 // Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never.
184 //
185 // | F  | T  | Ph  | Pd  |  H  |
186 // |----+----+-----+-----+-----+
187 // | d  | d  | N   | N   | (c) |
188 // | d  | g  | --  | --  | (a) |
189 // | d  | h  | --  | --  | (e) |
190 // | d  | hd | HD  | HD  | (b) |
191 // | g  | d  | N   | N   | (c) |
192 // | g  | g  | --  | --  | (a) |
193 // | g  | h  | --  | --  | (e) |
194 // | g  | hd | HD  | HD  | (b) |
195 // | h  | d  | --  | --  | (e) |
196 // | h  | g  | N   | N   | (c) |
197 // | h  | h  | N   | N   | (c) |
198 // | h  | hd | HD  | HD  | (b) |
199 // | hd | d  | WS  | SS  | (d) |
200 // | hd | g  | SS  | --  |(d/a)|
201 // | hd | h  | SS  | WS  | (d) |
202 // | hd | hd | HD  | HD  | (b) |
203 
204 Sema::CUDAFunctionPreference
205 Sema::IdentifyCUDAPreference(const FunctionDecl *Caller,
206                              const FunctionDecl *Callee) {
207   assert(Callee && "Callee must be valid.");
208   CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller);
209   CUDAFunctionTarget CalleeTarget = IdentifyCUDATarget(Callee);
210 
211   // If one of the targets is invalid, the check always fails, no matter what
212   // the other target is.
213   if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget)
214     return CFP_Never;
215 
216   // (a) Can't call global from some contexts until we support CUDA's
217   // dynamic parallelism.
218   if (CalleeTarget == CFT_Global &&
219       (CallerTarget == CFT_Global || CallerTarget == CFT_Device))
220     return CFP_Never;
221 
222   // (b) Calling HostDevice is OK for everyone.
223   if (CalleeTarget == CFT_HostDevice)
224     return CFP_HostDevice;
225 
226   // (c) Best case scenarios
227   if (CalleeTarget == CallerTarget ||
228       (CallerTarget == CFT_Host && CalleeTarget == CFT_Global) ||
229       (CallerTarget == CFT_Global && CalleeTarget == CFT_Device))
230     return CFP_Native;
231 
232   // (d) HostDevice behavior depends on compilation mode.
233   if (CallerTarget == CFT_HostDevice) {
234     // It's OK to call a compilation-mode matching function from an HD one.
235     if ((getLangOpts().CUDAIsDevice && CalleeTarget == CFT_Device) ||
236         (!getLangOpts().CUDAIsDevice &&
237          (CalleeTarget == CFT_Host || CalleeTarget == CFT_Global)))
238       return CFP_SameSide;
239 
240     // Calls from HD to non-mode-matching functions (i.e., to host functions
241     // when compiling in device mode or to device functions when compiling in
242     // host mode) are allowed at the sema level, but eventually rejected if
243     // they're ever codegened.  TODO: Reject said calls earlier.
244     return CFP_WrongSide;
245   }
246 
247   // (e) Calling across device/host boundary is not something you should do.
248   if ((CallerTarget == CFT_Host && CalleeTarget == CFT_Device) ||
249       (CallerTarget == CFT_Device && CalleeTarget == CFT_Host) ||
250       (CallerTarget == CFT_Global && CalleeTarget == CFT_Host))
251     return CFP_Never;
252 
253   llvm_unreachable("All cases should've been handled by now.");
254 }
255 
256 template <typename AttrT> static bool hasImplicitAttr(const FunctionDecl *D) {
257   if (!D)
258     return false;
259   if (auto *A = D->getAttr<AttrT>())
260     return A->isImplicit();
261   return D->isImplicit();
262 }
263 
264 bool Sema::isCUDAImplicitHostDeviceFunction(const FunctionDecl *D) {
265   bool IsImplicitDevAttr = hasImplicitAttr<CUDADeviceAttr>(D);
266   bool IsImplicitHostAttr = hasImplicitAttr<CUDAHostAttr>(D);
267   return IsImplicitDevAttr && IsImplicitHostAttr;
268 }
269 
270 void Sema::EraseUnwantedCUDAMatches(
271     const FunctionDecl *Caller,
272     SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) {
273   if (Matches.size() <= 1)
274     return;
275 
276   using Pair = std::pair<DeclAccessPair, FunctionDecl*>;
277 
278   // Gets the CUDA function preference for a call from Caller to Match.
279   auto GetCFP = [&](const Pair &Match) {
280     return IdentifyCUDAPreference(Caller, Match.second);
281   };
282 
283   // Find the best call preference among the functions in Matches.
284   CUDAFunctionPreference BestCFP = GetCFP(*std::max_element(
285       Matches.begin(), Matches.end(),
286       [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); }));
287 
288   // Erase all functions with lower priority.
289   llvm::erase_if(Matches,
290                  [&](const Pair &Match) { return GetCFP(Match) < BestCFP; });
291 }
292 
293 /// When an implicitly-declared special member has to invoke more than one
294 /// base/field special member, conflicts may occur in the targets of these
295 /// members. For example, if one base's member __host__ and another's is
296 /// __device__, it's a conflict.
297 /// This function figures out if the given targets \param Target1 and
298 /// \param Target2 conflict, and if they do not it fills in
299 /// \param ResolvedTarget with a target that resolves for both calls.
300 /// \return true if there's a conflict, false otherwise.
301 static bool
302 resolveCalleeCUDATargetConflict(Sema::CUDAFunctionTarget Target1,
303                                 Sema::CUDAFunctionTarget Target2,
304                                 Sema::CUDAFunctionTarget *ResolvedTarget) {
305   // Only free functions and static member functions may be global.
306   assert(Target1 != Sema::CFT_Global);
307   assert(Target2 != Sema::CFT_Global);
308 
309   if (Target1 == Sema::CFT_HostDevice) {
310     *ResolvedTarget = Target2;
311   } else if (Target2 == Sema::CFT_HostDevice) {
312     *ResolvedTarget = Target1;
313   } else if (Target1 != Target2) {
314     return true;
315   } else {
316     *ResolvedTarget = Target1;
317   }
318 
319   return false;
320 }
321 
322 bool Sema::inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl,
323                                                    CXXSpecialMember CSM,
324                                                    CXXMethodDecl *MemberDecl,
325                                                    bool ConstRHS,
326                                                    bool Diagnose) {
327   // If the defaulted special member is defined lexically outside of its
328   // owning class, or the special member already has explicit device or host
329   // attributes, do not infer.
330   bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent();
331   bool HasH = MemberDecl->hasAttr<CUDAHostAttr>();
332   bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>();
333   bool HasExplicitAttr =
334       (HasD && !MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()) ||
335       (HasH && !MemberDecl->getAttr<CUDAHostAttr>()->isImplicit());
336   if (!InClass || HasExplicitAttr)
337     return false;
338 
339   llvm::Optional<CUDAFunctionTarget> InferredTarget;
340 
341   // We're going to invoke special member lookup; mark that these special
342   // members are called from this one, and not from its caller.
343   ContextRAII MethodContext(*this, MemberDecl);
344 
345   // Look for special members in base classes that should be invoked from here.
346   // Infer the target of this member base on the ones it should call.
347   // Skip direct and indirect virtual bases for abstract classes.
348   llvm::SmallVector<const CXXBaseSpecifier *, 16> Bases;
349   for (const auto &B : ClassDecl->bases()) {
350     if (!B.isVirtual()) {
351       Bases.push_back(&B);
352     }
353   }
354 
355   if (!ClassDecl->isAbstract()) {
356     for (const auto &VB : ClassDecl->vbases()) {
357       Bases.push_back(&VB);
358     }
359   }
360 
361   for (const auto *B : Bases) {
362     const RecordType *BaseType = B->getType()->getAs<RecordType>();
363     if (!BaseType) {
364       continue;
365     }
366 
367     CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());
368     Sema::SpecialMemberOverloadResult SMOR =
369         LookupSpecialMember(BaseClassDecl, CSM,
370                             /* ConstArg */ ConstRHS,
371                             /* VolatileArg */ false,
372                             /* RValueThis */ false,
373                             /* ConstThis */ false,
374                             /* VolatileThis */ false);
375 
376     if (!SMOR.getMethod())
377       continue;
378 
379     CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR.getMethod());
380     if (!InferredTarget.hasValue()) {
381       InferredTarget = BaseMethodTarget;
382     } else {
383       bool ResolutionError = resolveCalleeCUDATargetConflict(
384           InferredTarget.getValue(), BaseMethodTarget,
385           InferredTarget.getPointer());
386       if (ResolutionError) {
387         if (Diagnose) {
388           Diag(ClassDecl->getLocation(),
389                diag::note_implicit_member_target_infer_collision)
390               << (unsigned)CSM << InferredTarget.getValue() << BaseMethodTarget;
391         }
392         MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
393         return true;
394       }
395     }
396   }
397 
398   // Same as for bases, but now for special members of fields.
399   for (const auto *F : ClassDecl->fields()) {
400     if (F->isInvalidDecl()) {
401       continue;
402     }
403 
404     const RecordType *FieldType =
405         Context.getBaseElementType(F->getType())->getAs<RecordType>();
406     if (!FieldType) {
407       continue;
408     }
409 
410     CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());
411     Sema::SpecialMemberOverloadResult SMOR =
412         LookupSpecialMember(FieldRecDecl, CSM,
413                             /* ConstArg */ ConstRHS && !F->isMutable(),
414                             /* VolatileArg */ false,
415                             /* RValueThis */ false,
416                             /* ConstThis */ false,
417                             /* VolatileThis */ false);
418 
419     if (!SMOR.getMethod())
420       continue;
421 
422     CUDAFunctionTarget FieldMethodTarget =
423         IdentifyCUDATarget(SMOR.getMethod());
424     if (!InferredTarget.hasValue()) {
425       InferredTarget = FieldMethodTarget;
426     } else {
427       bool ResolutionError = resolveCalleeCUDATargetConflict(
428           InferredTarget.getValue(), FieldMethodTarget,
429           InferredTarget.getPointer());
430       if (ResolutionError) {
431         if (Diagnose) {
432           Diag(ClassDecl->getLocation(),
433                diag::note_implicit_member_target_infer_collision)
434               << (unsigned)CSM << InferredTarget.getValue()
435               << FieldMethodTarget;
436         }
437         MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
438         return true;
439       }
440     }
441   }
442 
443 
444   // If no target was inferred, mark this member as __host__ __device__;
445   // it's the least restrictive option that can be invoked from any target.
446   bool NeedsH = true, NeedsD = true;
447   if (InferredTarget.hasValue()) {
448     if (InferredTarget.getValue() == CFT_Device)
449       NeedsH = false;
450     else if (InferredTarget.getValue() == CFT_Host)
451       NeedsD = false;
452   }
453 
454   // We either setting attributes first time, or the inferred ones must match
455   // previously set ones.
456   if (NeedsD && !HasD)
457     MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
458   if (NeedsH && !HasH)
459     MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
460 
461   return false;
462 }
463 
464 bool Sema::isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD) {
465   if (!CD->isDefined() && CD->isTemplateInstantiation())
466     InstantiateFunctionDefinition(Loc, CD->getFirstDecl());
467 
468   // (E.2.3.1, CUDA 7.5) A constructor for a class type is considered
469   // empty at a point in the translation unit, if it is either a
470   // trivial constructor
471   if (CD->isTrivial())
472     return true;
473 
474   // ... or it satisfies all of the following conditions:
475   // The constructor function has been defined.
476   // The constructor function has no parameters,
477   // and the function body is an empty compound statement.
478   if (!(CD->hasTrivialBody() && CD->getNumParams() == 0))
479     return false;
480 
481   // Its class has no virtual functions and no virtual base classes.
482   if (CD->getParent()->isDynamicClass())
483     return false;
484 
485   // Union ctor does not call ctors of its data members.
486   if (CD->getParent()->isUnion())
487     return true;
488 
489   // The only form of initializer allowed is an empty constructor.
490   // This will recursively check all base classes and member initializers
491   if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {
492         if (const CXXConstructExpr *CE =
493                 dyn_cast<CXXConstructExpr>(CI->getInit()))
494           return isEmptyCudaConstructor(Loc, CE->getConstructor());
495         return false;
496       }))
497     return false;
498 
499   return true;
500 }
501 
502 bool Sema::isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *DD) {
503   // No destructor -> no problem.
504   if (!DD)
505     return true;
506 
507   if (!DD->isDefined() && DD->isTemplateInstantiation())
508     InstantiateFunctionDefinition(Loc, DD->getFirstDecl());
509 
510   // (E.2.3.1, CUDA 7.5) A destructor for a class type is considered
511   // empty at a point in the translation unit, if it is either a
512   // trivial constructor
513   if (DD->isTrivial())
514     return true;
515 
516   // ... or it satisfies all of the following conditions:
517   // The destructor function has been defined.
518   // and the function body is an empty compound statement.
519   if (!DD->hasTrivialBody())
520     return false;
521 
522   const CXXRecordDecl *ClassDecl = DD->getParent();
523 
524   // Its class has no virtual functions and no virtual base classes.
525   if (ClassDecl->isDynamicClass())
526     return false;
527 
528   // Union does not have base class and union dtor does not call dtors of its
529   // data members.
530   if (DD->getParent()->isUnion())
531     return true;
532 
533   // Only empty destructors are allowed. This will recursively check
534   // destructors for all base classes...
535   if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {
536         if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())
537           return isEmptyCudaDestructor(Loc, RD->getDestructor());
538         return true;
539       }))
540     return false;
541 
542   // ... and member fields.
543   if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {
544         if (CXXRecordDecl *RD = Field->getType()
545                                     ->getBaseElementTypeUnsafe()
546                                     ->getAsCXXRecordDecl())
547           return isEmptyCudaDestructor(Loc, RD->getDestructor());
548         return true;
549       }))
550     return false;
551 
552   return true;
553 }
554 
555 namespace {
556 enum CUDAInitializerCheckKind {
557   CICK_DeviceOrConstant, // Check initializer for device/constant variable
558   CICK_Shared,           // Check initializer for shared variable
559 };
560 
561 bool IsDependentVar(VarDecl *VD) {
562   if (VD->getType()->isDependentType())
563     return true;
564   if (const auto *Init = VD->getInit())
565     return Init->isValueDependent();
566   return false;
567 }
568 
569 // Check whether a variable has an allowed initializer for a CUDA device side
570 // variable with global storage. \p VD may be a host variable to be checked for
571 // potential promotion to device side variable.
572 //
573 // CUDA/HIP allows only empty constructors as initializers for global
574 // variables (see E.2.3.1, CUDA 7.5). The same restriction also applies to all
575 // __shared__ variables whether they are local or not (they all are implicitly
576 // static in CUDA). One exception is that CUDA allows constant initializers
577 // for __constant__ and __device__ variables.
578 bool HasAllowedCUDADeviceStaticInitializer(Sema &S, VarDecl *VD,
579                                            CUDAInitializerCheckKind CheckKind) {
580   assert(!VD->isInvalidDecl() && VD->hasGlobalStorage());
581   assert(!IsDependentVar(VD) && "do not check dependent var");
582   const Expr *Init = VD->getInit();
583   auto IsEmptyInit = [&](const Expr *Init) {
584     if (!Init)
585       return true;
586     if (const auto *CE = dyn_cast<CXXConstructExpr>(Init)) {
587       return S.isEmptyCudaConstructor(VD->getLocation(), CE->getConstructor());
588     }
589     return false;
590   };
591   auto IsConstantInit = [&](const Expr *Init) {
592     assert(Init);
593     return Init->isConstantInitializer(S.Context,
594                                        VD->getType()->isReferenceType());
595   };
596   auto HasEmptyDtor = [&](VarDecl *VD) {
597     if (const auto *RD = VD->getType()->getAsCXXRecordDecl())
598       return S.isEmptyCudaDestructor(VD->getLocation(), RD->getDestructor());
599     return true;
600   };
601   if (CheckKind == CICK_Shared)
602     return IsEmptyInit(Init) && HasEmptyDtor(VD);
603   return S.LangOpts.GPUAllowDeviceInit ||
604          ((IsEmptyInit(Init) || IsConstantInit(Init)) && HasEmptyDtor(VD));
605 }
606 } // namespace
607 
608 void Sema::checkAllowedCUDAInitializer(VarDecl *VD) {
609   // Do not check dependent variables since the ctor/dtor/initializer are not
610   // determined. Do it after instantiation.
611   if (VD->isInvalidDecl() || !VD->hasInit() || !VD->hasGlobalStorage() ||
612       IsDependentVar(VD))
613     return;
614   const Expr *Init = VD->getInit();
615   bool IsSharedVar = VD->hasAttr<CUDASharedAttr>();
616   bool IsDeviceOrConstantVar =
617       !IsSharedVar &&
618       (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>());
619   if (IsDeviceOrConstantVar || IsSharedVar) {
620     if (HasAllowedCUDADeviceStaticInitializer(
621             *this, VD, IsSharedVar ? CICK_Shared : CICK_DeviceOrConstant))
622       return;
623     Diag(VD->getLocation(),
624          IsSharedVar ? diag::err_shared_var_init : diag::err_dynamic_var_init)
625         << Init->getSourceRange();
626     VD->setInvalidDecl();
627   } else {
628     // This is a host-side global variable.  Check that the initializer is
629     // callable from the host side.
630     const FunctionDecl *InitFn = nullptr;
631     if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init)) {
632       InitFn = CE->getConstructor();
633     } else if (const CallExpr *CE = dyn_cast<CallExpr>(Init)) {
634       InitFn = CE->getDirectCallee();
635     }
636     if (InitFn) {
637       CUDAFunctionTarget InitFnTarget = IdentifyCUDATarget(InitFn);
638       if (InitFnTarget != CFT_Host && InitFnTarget != CFT_HostDevice) {
639         Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer)
640             << InitFnTarget << InitFn;
641         Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn;
642         VD->setInvalidDecl();
643       }
644     }
645   }
646 }
647 
648 // With -fcuda-host-device-constexpr, an unattributed constexpr function is
649 // treated as implicitly __host__ __device__, unless:
650 //  * it is a variadic function (device-side variadic functions are not
651 //    allowed), or
652 //  * a __device__ function with this signature was already declared, in which
653 //    case in which case we output an error, unless the __device__ decl is in a
654 //    system header, in which case we leave the constexpr function unattributed.
655 //
656 // In addition, all function decls are treated as __host__ __device__ when
657 // ForceCUDAHostDeviceDepth > 0 (corresponding to code within a
658 //   #pragma clang force_cuda_host_device_begin/end
659 // pair).
660 void Sema::maybeAddCUDAHostDeviceAttrs(FunctionDecl *NewD,
661                                        const LookupResult &Previous) {
662   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
663 
664   if (ForceCUDAHostDeviceDepth > 0) {
665     if (!NewD->hasAttr<CUDAHostAttr>())
666       NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
667     if (!NewD->hasAttr<CUDADeviceAttr>())
668       NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
669     return;
670   }
671 
672   if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() ||
673       NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() ||
674       NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>())
675     return;
676 
677   // Is D a __device__ function with the same signature as NewD, ignoring CUDA
678   // attributes?
679   auto IsMatchingDeviceFn = [&](NamedDecl *D) {
680     if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D))
681       D = Using->getTargetDecl();
682     FunctionDecl *OldD = D->getAsFunction();
683     return OldD && OldD->hasAttr<CUDADeviceAttr>() &&
684            !OldD->hasAttr<CUDAHostAttr>() &&
685            !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false,
686                        /* ConsiderCudaAttrs = */ false);
687   };
688   auto It = llvm::find_if(Previous, IsMatchingDeviceFn);
689   if (It != Previous.end()) {
690     // We found a __device__ function with the same name and signature as NewD
691     // (ignoring CUDA attrs).  This is an error unless that function is defined
692     // in a system header, in which case we simply return without making NewD
693     // host+device.
694     NamedDecl *Match = *It;
695     if (!getSourceManager().isInSystemHeader(Match->getLocation())) {
696       Diag(NewD->getLocation(),
697            diag::err_cuda_unattributed_constexpr_cannot_overload_device)
698           << NewD;
699       Diag(Match->getLocation(),
700            diag::note_cuda_conflicting_device_function_declared_here);
701     }
702     return;
703   }
704 
705   NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
706   NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
707 }
708 
709 // TODO: `__constant__` memory may be a limited resource for certain targets.
710 // A safeguard may be needed at the end of compilation pipeline if
711 // `__constant__` memory usage goes beyond limit.
712 void Sema::MaybeAddCUDAConstantAttr(VarDecl *VD) {
713   // Do not promote dependent variables since the cotr/dtor/initializer are
714   // not determined. Do it after instantiation.
715   if (getLangOpts().CUDAIsDevice && !VD->hasAttr<CUDAConstantAttr>() &&
716       !VD->hasAttr<CUDAConstantAttr>() && !VD->hasAttr<CUDASharedAttr>() &&
717       (VD->isFileVarDecl() || VD->isStaticDataMember()) &&
718       !IsDependentVar(VD) &&
719       (VD->isConstexpr() || (VD->getType().isConstQualified() &&
720                              HasAllowedCUDADeviceStaticInitializer(
721                                  *this, VD, CICK_DeviceOrConstant)))) {
722     VD->addAttr(CUDAConstantAttr::CreateImplicit(getASTContext()));
723   }
724 }
725 
726 Sema::SemaDiagnosticBuilder Sema::CUDADiagIfDeviceCode(SourceLocation Loc,
727                                                        unsigned DiagID) {
728   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
729   SemaDiagnosticBuilder::Kind DiagKind = [&] {
730     if (!isa<FunctionDecl>(CurContext))
731       return SemaDiagnosticBuilder::K_Nop;
732     switch (CurrentCUDATarget()) {
733     case CFT_Global:
734     case CFT_Device:
735       return SemaDiagnosticBuilder::K_Immediate;
736     case CFT_HostDevice:
737       // An HD function counts as host code if we're compiling for host, and
738       // device code if we're compiling for device.  Defer any errors in device
739       // mode until the function is known-emitted.
740       if (!getLangOpts().CUDAIsDevice)
741         return SemaDiagnosticBuilder::K_Nop;
742       if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID))
743         return SemaDiagnosticBuilder::K_Immediate;
744       return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
745               FunctionEmissionStatus::Emitted)
746                  ? SemaDiagnosticBuilder::K_ImmediateWithCallStack
747                  : SemaDiagnosticBuilder::K_Deferred;
748     default:
749       return SemaDiagnosticBuilder::K_Nop;
750     }
751   }();
752   return SemaDiagnosticBuilder(DiagKind, Loc, DiagID,
753                                dyn_cast<FunctionDecl>(CurContext), *this);
754 }
755 
756 Sema::SemaDiagnosticBuilder Sema::CUDADiagIfHostCode(SourceLocation Loc,
757                                                      unsigned DiagID) {
758   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
759   SemaDiagnosticBuilder::Kind DiagKind = [&] {
760     if (!isa<FunctionDecl>(CurContext))
761       return SemaDiagnosticBuilder::K_Nop;
762     switch (CurrentCUDATarget()) {
763     case CFT_Host:
764       return SemaDiagnosticBuilder::K_Immediate;
765     case CFT_HostDevice:
766       // An HD function counts as host code if we're compiling for host, and
767       // device code if we're compiling for device.  Defer any errors in device
768       // mode until the function is known-emitted.
769       if (getLangOpts().CUDAIsDevice)
770         return SemaDiagnosticBuilder::K_Nop;
771       if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID))
772         return SemaDiagnosticBuilder::K_Immediate;
773       return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
774               FunctionEmissionStatus::Emitted)
775                  ? SemaDiagnosticBuilder::K_ImmediateWithCallStack
776                  : SemaDiagnosticBuilder::K_Deferred;
777     default:
778       return SemaDiagnosticBuilder::K_Nop;
779     }
780   }();
781   return SemaDiagnosticBuilder(DiagKind, Loc, DiagID,
782                                dyn_cast<FunctionDecl>(CurContext), *this);
783 }
784 
785 bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) {
786   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
787   assert(Callee && "Callee may not be null.");
788 
789   auto &ExprEvalCtx = ExprEvalContexts.back();
790   if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated())
791     return true;
792 
793   // FIXME: Is bailing out early correct here?  Should we instead assume that
794   // the caller is a global initializer?
795   FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
796   if (!Caller)
797     return true;
798 
799   // If the caller is known-emitted, mark the callee as known-emitted.
800   // Otherwise, mark the call in our call graph so we can traverse it later.
801   bool CallerKnownEmitted =
802       getEmissionStatus(Caller) == FunctionEmissionStatus::Emitted;
803   SemaDiagnosticBuilder::Kind DiagKind = [this, Caller, Callee,
804                                           CallerKnownEmitted] {
805     switch (IdentifyCUDAPreference(Caller, Callee)) {
806     case CFP_Never:
807     case CFP_WrongSide:
808       assert(Caller && "Never/wrongSide calls require a non-null caller");
809       // If we know the caller will be emitted, we know this wrong-side call
810       // will be emitted, so it's an immediate error.  Otherwise, defer the
811       // error until we know the caller is emitted.
812       return CallerKnownEmitted
813                  ? SemaDiagnosticBuilder::K_ImmediateWithCallStack
814                  : SemaDiagnosticBuilder::K_Deferred;
815     default:
816       return SemaDiagnosticBuilder::K_Nop;
817     }
818   }();
819 
820   if (DiagKind == SemaDiagnosticBuilder::K_Nop)
821     return true;
822 
823   // Avoid emitting this error twice for the same location.  Using a hashtable
824   // like this is unfortunate, but because we must continue parsing as normal
825   // after encountering a deferred error, it's otherwise very tricky for us to
826   // ensure that we only emit this deferred error once.
827   if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second)
828     return true;
829 
830   SemaDiagnosticBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this)
831       << IdentifyCUDATarget(Callee) << /*function*/ 0 << Callee
832       << IdentifyCUDATarget(Caller);
833   if (!Callee->getBuiltinID())
834     SemaDiagnosticBuilder(DiagKind, Callee->getLocation(),
835                           diag::note_previous_decl, Caller, *this)
836         << Callee;
837   return DiagKind != SemaDiagnosticBuilder::K_Immediate &&
838          DiagKind != SemaDiagnosticBuilder::K_ImmediateWithCallStack;
839 }
840 
841 // Check the wrong-sided reference capture of lambda for CUDA/HIP.
842 // A lambda function may capture a stack variable by reference when it is
843 // defined and uses the capture by reference when the lambda is called. When
844 // the capture and use happen on different sides, the capture is invalid and
845 // should be diagnosed.
846 void Sema::CUDACheckLambdaCapture(CXXMethodDecl *Callee,
847                                   const sema::Capture &Capture) {
848   // In host compilation we only need to check lambda functions emitted on host
849   // side. In such lambda functions, a reference capture is invalid only
850   // if the lambda structure is populated by a device function or kernel then
851   // is passed to and called by a host function. However that is impossible,
852   // since a device function or kernel can only call a device function, also a
853   // kernel cannot pass a lambda back to a host function since we cannot
854   // define a kernel argument type which can hold the lambda before the lambda
855   // itself is defined.
856   if (!LangOpts.CUDAIsDevice)
857     return;
858 
859   // File-scope lambda can only do init captures for global variables, which
860   // results in passing by value for these global variables.
861   FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
862   if (!Caller)
863     return;
864 
865   // In device compilation, we only need to check lambda functions which are
866   // emitted on device side. For such lambdas, a reference capture is invalid
867   // only if the lambda structure is populated by a host function then passed
868   // to and called in a device function or kernel.
869   bool CalleeIsDevice = Callee->hasAttr<CUDADeviceAttr>();
870   bool CallerIsHost =
871       !Caller->hasAttr<CUDAGlobalAttr>() && !Caller->hasAttr<CUDADeviceAttr>();
872   bool ShouldCheck = CalleeIsDevice && CallerIsHost;
873   if (!ShouldCheck || !Capture.isReferenceCapture())
874     return;
875   auto DiagKind = SemaDiagnosticBuilder::K_Deferred;
876   if (Capture.isVariableCapture()) {
877     SemaDiagnosticBuilder(DiagKind, Capture.getLocation(),
878                           diag::err_capture_bad_target, Callee, *this)
879         << Capture.getVariable();
880   } else if (Capture.isThisCapture()) {
881     SemaDiagnosticBuilder(DiagKind, Capture.getLocation(),
882                           diag::err_capture_bad_target_this_ptr, Callee, *this);
883   }
884   return;
885 }
886 
887 void Sema::CUDASetLambdaAttrs(CXXMethodDecl *Method) {
888   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
889   if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>())
890     return;
891   Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
892   Method->addAttr(CUDAHostAttr::CreateImplicit(Context));
893 }
894 
895 void Sema::checkCUDATargetOverload(FunctionDecl *NewFD,
896                                    const LookupResult &Previous) {
897   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
898   CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD);
899   for (NamedDecl *OldND : Previous) {
900     FunctionDecl *OldFD = OldND->getAsFunction();
901     if (!OldFD)
902       continue;
903 
904     CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD);
905     // Don't allow HD and global functions to overload other functions with the
906     // same signature.  We allow overloading based on CUDA attributes so that
907     // functions can have different implementations on the host and device, but
908     // HD/global functions "exist" in some sense on both the host and device, so
909     // should have the same implementation on both sides.
910     if (NewTarget != OldTarget &&
911         ((NewTarget == CFT_HostDevice) || (OldTarget == CFT_HostDevice) ||
912          (NewTarget == CFT_Global) || (OldTarget == CFT_Global)) &&
913         !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false,
914                     /* ConsiderCudaAttrs = */ false)) {
915       Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)
916           << NewTarget << NewFD->getDeclName() << OldTarget << OldFD;
917       Diag(OldFD->getLocation(), diag::note_previous_declaration);
918       NewFD->setInvalidDecl();
919       break;
920     }
921   }
922 }
923 
924 template <typename AttrTy>
925 static void copyAttrIfPresent(Sema &S, FunctionDecl *FD,
926                               const FunctionDecl &TemplateFD) {
927   if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
928     AttrTy *Clone = Attribute->clone(S.Context);
929     Clone->setInherited(true);
930     FD->addAttr(Clone);
931   }
932 }
933 
934 void Sema::inheritCUDATargetAttrs(FunctionDecl *FD,
935                                   const FunctionTemplateDecl &TD) {
936   const FunctionDecl &TemplateFD = *TD.getTemplatedDecl();
937   copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD);
938   copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD);
939   copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD);
940 }
941 
942 std::string Sema::getCudaConfigureFuncName() const {
943   if (getLangOpts().HIP)
944     return getLangOpts().HIPUseNewLaunchAPI ? "__hipPushCallConfiguration"
945                                             : "hipConfigureCall";
946 
947   // New CUDA kernel launch sequence.
948   if (CudaFeatureEnabled(Context.getTargetInfo().getSDKVersion(),
949                          CudaFeature::CUDA_USES_NEW_LAUNCH))
950     return "__cudaPushCallConfiguration";
951 
952   // Legacy CUDA kernel configuration call
953   return "cudaConfigureCall";
954 }
955