xref: /freebsd-src/contrib/llvm-project/clang/lib/Sema/SemaCUDA.cpp (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
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/Lex/Preprocessor.h"
18 #include "clang/Sema/Lookup.h"
19 #include "clang/Sema/Sema.h"
20 #include "clang/Sema/SemaDiagnostic.h"
21 #include "clang/Sema/SemaInternal.h"
22 #include "clang/Sema/Template.h"
23 #include "llvm/ADT/Optional.h"
24 #include "llvm/ADT/SmallVector.h"
25 using namespace clang;
26 
27 void Sema::PushForceCUDAHostDevice() {
28   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
29   ForceCUDAHostDeviceDepth++;
30 }
31 
32 bool Sema::PopForceCUDAHostDevice() {
33   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
34   if (ForceCUDAHostDeviceDepth == 0)
35     return false;
36   ForceCUDAHostDeviceDepth--;
37   return true;
38 }
39 
40 ExprResult Sema::ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc,
41                                          MultiExprArg ExecConfig,
42                                          SourceLocation GGGLoc) {
43   FunctionDecl *ConfigDecl = Context.getcudaConfigureCallDecl();
44   if (!ConfigDecl)
45     return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use)
46                      << getCudaConfigureFuncName());
47   QualType ConfigQTy = ConfigDecl->getType();
48 
49   DeclRefExpr *ConfigDR = new (Context)
50       DeclRefExpr(Context, ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc);
51   MarkFunctionReferenced(LLLLoc, ConfigDecl);
52 
53   return BuildCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr,
54                        /*IsExecConfig=*/true);
55 }
56 
57 Sema::CUDAFunctionTarget
58 Sema::IdentifyCUDATarget(const ParsedAttributesView &Attrs) {
59   bool HasHostAttr = false;
60   bool HasDeviceAttr = false;
61   bool HasGlobalAttr = false;
62   bool HasInvalidTargetAttr = false;
63   for (const ParsedAttr &AL : Attrs) {
64     switch (AL.getKind()) {
65     case ParsedAttr::AT_CUDAGlobal:
66       HasGlobalAttr = true;
67       break;
68     case ParsedAttr::AT_CUDAHost:
69       HasHostAttr = true;
70       break;
71     case ParsedAttr::AT_CUDADevice:
72       HasDeviceAttr = true;
73       break;
74     case ParsedAttr::AT_CUDAInvalidTarget:
75       HasInvalidTargetAttr = true;
76       break;
77     default:
78       break;
79     }
80   }
81 
82   if (HasInvalidTargetAttr)
83     return CFT_InvalidTarget;
84 
85   if (HasGlobalAttr)
86     return CFT_Global;
87 
88   if (HasHostAttr && HasDeviceAttr)
89     return CFT_HostDevice;
90 
91   if (HasDeviceAttr)
92     return CFT_Device;
93 
94   return CFT_Host;
95 }
96 
97 template <typename A>
98 static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
99   return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
100            return isa<A>(Attribute) &&
101                   !(IgnoreImplicitAttr && Attribute->isImplicit());
102          });
103 }
104 
105 /// IdentifyCUDATarget - Determine the CUDA compilation target for this function
106 Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D,
107                                                   bool IgnoreImplicitHDAttr) {
108   // Code that lives outside a function is run on the host.
109   if (D == nullptr)
110     return CFT_Host;
111 
112   if (D->hasAttr<CUDAInvalidTargetAttr>())
113     return CFT_InvalidTarget;
114 
115   if (D->hasAttr<CUDAGlobalAttr>())
116     return CFT_Global;
117 
118   if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) {
119     if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr))
120       return CFT_HostDevice;
121     return CFT_Device;
122   } else if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) {
123     return CFT_Host;
124   } else if (D->isImplicit() && !IgnoreImplicitHDAttr) {
125     // Some implicit declarations (like intrinsic functions) are not marked.
126     // Set the most lenient target on them for maximal flexibility.
127     return CFT_HostDevice;
128   }
129 
130   return CFT_Host;
131 }
132 
133 // * CUDA Call preference table
134 //
135 // F - from,
136 // T - to
137 // Ph - preference in host mode
138 // Pd - preference in device mode
139 // H  - handled in (x)
140 // Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never.
141 //
142 // | F  | T  | Ph  | Pd  |  H  |
143 // |----+----+-----+-----+-----+
144 // | d  | d  | N   | N   | (c) |
145 // | d  | g  | --  | --  | (a) |
146 // | d  | h  | --  | --  | (e) |
147 // | d  | hd | HD  | HD  | (b) |
148 // | g  | d  | N   | N   | (c) |
149 // | g  | g  | --  | --  | (a) |
150 // | g  | h  | --  | --  | (e) |
151 // | g  | hd | HD  | HD  | (b) |
152 // | h  | d  | --  | --  | (e) |
153 // | h  | g  | N   | N   | (c) |
154 // | h  | h  | N   | N   | (c) |
155 // | h  | hd | HD  | HD  | (b) |
156 // | hd | d  | WS  | SS  | (d) |
157 // | hd | g  | SS  | --  |(d/a)|
158 // | hd | h  | SS  | WS  | (d) |
159 // | hd | hd | HD  | HD  | (b) |
160 
161 Sema::CUDAFunctionPreference
162 Sema::IdentifyCUDAPreference(const FunctionDecl *Caller,
163                              const FunctionDecl *Callee) {
164   assert(Callee && "Callee must be valid.");
165   CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller);
166   CUDAFunctionTarget CalleeTarget = IdentifyCUDATarget(Callee);
167 
168   // If one of the targets is invalid, the check always fails, no matter what
169   // the other target is.
170   if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget)
171     return CFP_Never;
172 
173   // (a) Can't call global from some contexts until we support CUDA's
174   // dynamic parallelism.
175   if (CalleeTarget == CFT_Global &&
176       (CallerTarget == CFT_Global || CallerTarget == CFT_Device))
177     return CFP_Never;
178 
179   // (b) Calling HostDevice is OK for everyone.
180   if (CalleeTarget == CFT_HostDevice)
181     return CFP_HostDevice;
182 
183   // (c) Best case scenarios
184   if (CalleeTarget == CallerTarget ||
185       (CallerTarget == CFT_Host && CalleeTarget == CFT_Global) ||
186       (CallerTarget == CFT_Global && CalleeTarget == CFT_Device))
187     return CFP_Native;
188 
189   // (d) HostDevice behavior depends on compilation mode.
190   if (CallerTarget == CFT_HostDevice) {
191     // It's OK to call a compilation-mode matching function from an HD one.
192     if ((getLangOpts().CUDAIsDevice && CalleeTarget == CFT_Device) ||
193         (!getLangOpts().CUDAIsDevice &&
194          (CalleeTarget == CFT_Host || CalleeTarget == CFT_Global)))
195       return CFP_SameSide;
196 
197     // Calls from HD to non-mode-matching functions (i.e., to host functions
198     // when compiling in device mode or to device functions when compiling in
199     // host mode) are allowed at the sema level, but eventually rejected if
200     // they're ever codegened.  TODO: Reject said calls earlier.
201     return CFP_WrongSide;
202   }
203 
204   // (e) Calling across device/host boundary is not something you should do.
205   if ((CallerTarget == CFT_Host && CalleeTarget == CFT_Device) ||
206       (CallerTarget == CFT_Device && CalleeTarget == CFT_Host) ||
207       (CallerTarget == CFT_Global && CalleeTarget == CFT_Host))
208     return CFP_Never;
209 
210   llvm_unreachable("All cases should've been handled by now.");
211 }
212 
213 void Sema::EraseUnwantedCUDAMatches(
214     const FunctionDecl *Caller,
215     SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) {
216   if (Matches.size() <= 1)
217     return;
218 
219   using Pair = std::pair<DeclAccessPair, FunctionDecl*>;
220 
221   // Gets the CUDA function preference for a call from Caller to Match.
222   auto GetCFP = [&](const Pair &Match) {
223     return IdentifyCUDAPreference(Caller, Match.second);
224   };
225 
226   // Find the best call preference among the functions in Matches.
227   CUDAFunctionPreference BestCFP = GetCFP(*std::max_element(
228       Matches.begin(), Matches.end(),
229       [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); }));
230 
231   // Erase all functions with lower priority.
232   llvm::erase_if(Matches,
233                  [&](const Pair &Match) { return GetCFP(Match) < BestCFP; });
234 }
235 
236 /// When an implicitly-declared special member has to invoke more than one
237 /// base/field special member, conflicts may occur in the targets of these
238 /// members. For example, if one base's member __host__ and another's is
239 /// __device__, it's a conflict.
240 /// This function figures out if the given targets \param Target1 and
241 /// \param Target2 conflict, and if they do not it fills in
242 /// \param ResolvedTarget with a target that resolves for both calls.
243 /// \return true if there's a conflict, false otherwise.
244 static bool
245 resolveCalleeCUDATargetConflict(Sema::CUDAFunctionTarget Target1,
246                                 Sema::CUDAFunctionTarget Target2,
247                                 Sema::CUDAFunctionTarget *ResolvedTarget) {
248   // Only free functions and static member functions may be global.
249   assert(Target1 != Sema::CFT_Global);
250   assert(Target2 != Sema::CFT_Global);
251 
252   if (Target1 == Sema::CFT_HostDevice) {
253     *ResolvedTarget = Target2;
254   } else if (Target2 == Sema::CFT_HostDevice) {
255     *ResolvedTarget = Target1;
256   } else if (Target1 != Target2) {
257     return true;
258   } else {
259     *ResolvedTarget = Target1;
260   }
261 
262   return false;
263 }
264 
265 bool Sema::inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl,
266                                                    CXXSpecialMember CSM,
267                                                    CXXMethodDecl *MemberDecl,
268                                                    bool ConstRHS,
269                                                    bool Diagnose) {
270   // If the defaulted special member is defined lexically outside of its
271   // owning class, or the special member already has explicit device or host
272   // attributes, do not infer.
273   bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent();
274   bool HasH = MemberDecl->hasAttr<CUDAHostAttr>();
275   bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>();
276   bool HasExplicitAttr =
277       (HasD && !MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()) ||
278       (HasH && !MemberDecl->getAttr<CUDAHostAttr>()->isImplicit());
279   if (!InClass || HasExplicitAttr)
280     return false;
281 
282   llvm::Optional<CUDAFunctionTarget> InferredTarget;
283 
284   // We're going to invoke special member lookup; mark that these special
285   // members are called from this one, and not from its caller.
286   ContextRAII MethodContext(*this, MemberDecl);
287 
288   // Look for special members in base classes that should be invoked from here.
289   // Infer the target of this member base on the ones it should call.
290   // Skip direct and indirect virtual bases for abstract classes.
291   llvm::SmallVector<const CXXBaseSpecifier *, 16> Bases;
292   for (const auto &B : ClassDecl->bases()) {
293     if (!B.isVirtual()) {
294       Bases.push_back(&B);
295     }
296   }
297 
298   if (!ClassDecl->isAbstract()) {
299     for (const auto &VB : ClassDecl->vbases()) {
300       Bases.push_back(&VB);
301     }
302   }
303 
304   for (const auto *B : Bases) {
305     const RecordType *BaseType = B->getType()->getAs<RecordType>();
306     if (!BaseType) {
307       continue;
308     }
309 
310     CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());
311     Sema::SpecialMemberOverloadResult SMOR =
312         LookupSpecialMember(BaseClassDecl, CSM,
313                             /* ConstArg */ ConstRHS,
314                             /* VolatileArg */ false,
315                             /* RValueThis */ false,
316                             /* ConstThis */ false,
317                             /* VolatileThis */ false);
318 
319     if (!SMOR.getMethod())
320       continue;
321 
322     CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR.getMethod());
323     if (!InferredTarget.hasValue()) {
324       InferredTarget = BaseMethodTarget;
325     } else {
326       bool ResolutionError = resolveCalleeCUDATargetConflict(
327           InferredTarget.getValue(), BaseMethodTarget,
328           InferredTarget.getPointer());
329       if (ResolutionError) {
330         if (Diagnose) {
331           Diag(ClassDecl->getLocation(),
332                diag::note_implicit_member_target_infer_collision)
333               << (unsigned)CSM << InferredTarget.getValue() << BaseMethodTarget;
334         }
335         MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
336         return true;
337       }
338     }
339   }
340 
341   // Same as for bases, but now for special members of fields.
342   for (const auto *F : ClassDecl->fields()) {
343     if (F->isInvalidDecl()) {
344       continue;
345     }
346 
347     const RecordType *FieldType =
348         Context.getBaseElementType(F->getType())->getAs<RecordType>();
349     if (!FieldType) {
350       continue;
351     }
352 
353     CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());
354     Sema::SpecialMemberOverloadResult SMOR =
355         LookupSpecialMember(FieldRecDecl, CSM,
356                             /* ConstArg */ ConstRHS && !F->isMutable(),
357                             /* VolatileArg */ false,
358                             /* RValueThis */ false,
359                             /* ConstThis */ false,
360                             /* VolatileThis */ false);
361 
362     if (!SMOR.getMethod())
363       continue;
364 
365     CUDAFunctionTarget FieldMethodTarget =
366         IdentifyCUDATarget(SMOR.getMethod());
367     if (!InferredTarget.hasValue()) {
368       InferredTarget = FieldMethodTarget;
369     } else {
370       bool ResolutionError = resolveCalleeCUDATargetConflict(
371           InferredTarget.getValue(), FieldMethodTarget,
372           InferredTarget.getPointer());
373       if (ResolutionError) {
374         if (Diagnose) {
375           Diag(ClassDecl->getLocation(),
376                diag::note_implicit_member_target_infer_collision)
377               << (unsigned)CSM << InferredTarget.getValue()
378               << FieldMethodTarget;
379         }
380         MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
381         return true;
382       }
383     }
384   }
385 
386 
387   // If no target was inferred, mark this member as __host__ __device__;
388   // it's the least restrictive option that can be invoked from any target.
389   bool NeedsH = true, NeedsD = true;
390   if (InferredTarget.hasValue()) {
391     if (InferredTarget.getValue() == CFT_Device)
392       NeedsH = false;
393     else if (InferredTarget.getValue() == CFT_Host)
394       NeedsD = false;
395   }
396 
397   // We either setting attributes first time, or the inferred ones must match
398   // previously set ones.
399   if (NeedsD && !HasD)
400     MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
401   if (NeedsH && !HasH)
402     MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
403 
404   return false;
405 }
406 
407 bool Sema::isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD) {
408   if (!CD->isDefined() && CD->isTemplateInstantiation())
409     InstantiateFunctionDefinition(Loc, CD->getFirstDecl());
410 
411   // (E.2.3.1, CUDA 7.5) A constructor for a class type is considered
412   // empty at a point in the translation unit, if it is either a
413   // trivial constructor
414   if (CD->isTrivial())
415     return true;
416 
417   // ... or it satisfies all of the following conditions:
418   // The constructor function has been defined.
419   // The constructor function has no parameters,
420   // and the function body is an empty compound statement.
421   if (!(CD->hasTrivialBody() && CD->getNumParams() == 0))
422     return false;
423 
424   // Its class has no virtual functions and no virtual base classes.
425   if (CD->getParent()->isDynamicClass())
426     return false;
427 
428   // The only form of initializer allowed is an empty constructor.
429   // This will recursively check all base classes and member initializers
430   if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {
431         if (const CXXConstructExpr *CE =
432                 dyn_cast<CXXConstructExpr>(CI->getInit()))
433           return isEmptyCudaConstructor(Loc, CE->getConstructor());
434         return false;
435       }))
436     return false;
437 
438   return true;
439 }
440 
441 bool Sema::isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *DD) {
442   // No destructor -> no problem.
443   if (!DD)
444     return true;
445 
446   if (!DD->isDefined() && DD->isTemplateInstantiation())
447     InstantiateFunctionDefinition(Loc, DD->getFirstDecl());
448 
449   // (E.2.3.1, CUDA 7.5) A destructor for a class type is considered
450   // empty at a point in the translation unit, if it is either a
451   // trivial constructor
452   if (DD->isTrivial())
453     return true;
454 
455   // ... or it satisfies all of the following conditions:
456   // The destructor function has been defined.
457   // and the function body is an empty compound statement.
458   if (!DD->hasTrivialBody())
459     return false;
460 
461   const CXXRecordDecl *ClassDecl = DD->getParent();
462 
463   // Its class has no virtual functions and no virtual base classes.
464   if (ClassDecl->isDynamicClass())
465     return false;
466 
467   // Only empty destructors are allowed. This will recursively check
468   // destructors for all base classes...
469   if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {
470         if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())
471           return isEmptyCudaDestructor(Loc, RD->getDestructor());
472         return true;
473       }))
474     return false;
475 
476   // ... and member fields.
477   if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {
478         if (CXXRecordDecl *RD = Field->getType()
479                                     ->getBaseElementTypeUnsafe()
480                                     ->getAsCXXRecordDecl())
481           return isEmptyCudaDestructor(Loc, RD->getDestructor());
482         return true;
483       }))
484     return false;
485 
486   return true;
487 }
488 
489 void Sema::checkAllowedCUDAInitializer(VarDecl *VD) {
490   if (VD->isInvalidDecl() || !VD->hasInit() || !VD->hasGlobalStorage())
491     return;
492   const Expr *Init = VD->getInit();
493   if (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>() ||
494       VD->hasAttr<CUDASharedAttr>()) {
495     assert(!VD->isStaticLocal() || VD->hasAttr<CUDASharedAttr>());
496     bool AllowedInit = false;
497     if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init))
498       AllowedInit =
499           isEmptyCudaConstructor(VD->getLocation(), CE->getConstructor());
500     // We'll allow constant initializers even if it's a non-empty
501     // constructor according to CUDA rules. This deviates from NVCC,
502     // but allows us to handle things like constexpr constructors.
503     if (!AllowedInit &&
504         (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>()))
505       AllowedInit = VD->getInit()->isConstantInitializer(
506           Context, VD->getType()->isReferenceType());
507 
508     // Also make sure that destructor, if there is one, is empty.
509     if (AllowedInit)
510       if (CXXRecordDecl *RD = VD->getType()->getAsCXXRecordDecl())
511         AllowedInit =
512             isEmptyCudaDestructor(VD->getLocation(), RD->getDestructor());
513 
514     if (!AllowedInit) {
515       Diag(VD->getLocation(), VD->hasAttr<CUDASharedAttr>()
516                                   ? diag::err_shared_var_init
517                                   : diag::err_dynamic_var_init)
518           << Init->getSourceRange();
519       VD->setInvalidDecl();
520     }
521   } else {
522     // This is a host-side global variable.  Check that the initializer is
523     // callable from the host side.
524     const FunctionDecl *InitFn = nullptr;
525     if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init)) {
526       InitFn = CE->getConstructor();
527     } else if (const CallExpr *CE = dyn_cast<CallExpr>(Init)) {
528       InitFn = CE->getDirectCallee();
529     }
530     if (InitFn) {
531       CUDAFunctionTarget InitFnTarget = IdentifyCUDATarget(InitFn);
532       if (InitFnTarget != CFT_Host && InitFnTarget != CFT_HostDevice) {
533         Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer)
534             << InitFnTarget << InitFn;
535         Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn;
536         VD->setInvalidDecl();
537       }
538     }
539   }
540 }
541 
542 // With -fcuda-host-device-constexpr, an unattributed constexpr function is
543 // treated as implicitly __host__ __device__, unless:
544 //  * it is a variadic function (device-side variadic functions are not
545 //    allowed), or
546 //  * a __device__ function with this signature was already declared, in which
547 //    case in which case we output an error, unless the __device__ decl is in a
548 //    system header, in which case we leave the constexpr function unattributed.
549 //
550 // In addition, all function decls are treated as __host__ __device__ when
551 // ForceCUDAHostDeviceDepth > 0 (corresponding to code within a
552 //   #pragma clang force_cuda_host_device_begin/end
553 // pair).
554 void Sema::maybeAddCUDAHostDeviceAttrs(FunctionDecl *NewD,
555                                        const LookupResult &Previous) {
556   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
557 
558   if (ForceCUDAHostDeviceDepth > 0) {
559     if (!NewD->hasAttr<CUDAHostAttr>())
560       NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
561     if (!NewD->hasAttr<CUDADeviceAttr>())
562       NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
563     return;
564   }
565 
566   if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() ||
567       NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() ||
568       NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>())
569     return;
570 
571   // Is D a __device__ function with the same signature as NewD, ignoring CUDA
572   // attributes?
573   auto IsMatchingDeviceFn = [&](NamedDecl *D) {
574     if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D))
575       D = Using->getTargetDecl();
576     FunctionDecl *OldD = D->getAsFunction();
577     return OldD && OldD->hasAttr<CUDADeviceAttr>() &&
578            !OldD->hasAttr<CUDAHostAttr>() &&
579            !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false,
580                        /* ConsiderCudaAttrs = */ false);
581   };
582   auto It = llvm::find_if(Previous, IsMatchingDeviceFn);
583   if (It != Previous.end()) {
584     // We found a __device__ function with the same name and signature as NewD
585     // (ignoring CUDA attrs).  This is an error unless that function is defined
586     // in a system header, in which case we simply return without making NewD
587     // host+device.
588     NamedDecl *Match = *It;
589     if (!getSourceManager().isInSystemHeader(Match->getLocation())) {
590       Diag(NewD->getLocation(),
591            diag::err_cuda_unattributed_constexpr_cannot_overload_device)
592           << NewD;
593       Diag(Match->getLocation(),
594            diag::note_cuda_conflicting_device_function_declared_here);
595     }
596     return;
597   }
598 
599   NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
600   NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
601 }
602 
603 Sema::DeviceDiagBuilder Sema::CUDADiagIfDeviceCode(SourceLocation Loc,
604                                                    unsigned DiagID) {
605   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
606   DeviceDiagBuilder::Kind DiagKind = [this] {
607     switch (CurrentCUDATarget()) {
608     case CFT_Global:
609     case CFT_Device:
610       return DeviceDiagBuilder::K_Immediate;
611     case CFT_HostDevice:
612       // An HD function counts as host code if we're compiling for host, and
613       // device code if we're compiling for device.  Defer any errors in device
614       // mode until the function is known-emitted.
615       if (getLangOpts().CUDAIsDevice) {
616         return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
617                 FunctionEmissionStatus::Emitted)
618                    ? DeviceDiagBuilder::K_ImmediateWithCallStack
619                    : DeviceDiagBuilder::K_Deferred;
620       }
621       return DeviceDiagBuilder::K_Nop;
622 
623     default:
624       return DeviceDiagBuilder::K_Nop;
625     }
626   }();
627   return DeviceDiagBuilder(DiagKind, Loc, DiagID,
628                            dyn_cast<FunctionDecl>(CurContext), *this);
629 }
630 
631 Sema::DeviceDiagBuilder Sema::CUDADiagIfHostCode(SourceLocation Loc,
632                                                  unsigned DiagID) {
633   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
634   DeviceDiagBuilder::Kind DiagKind = [this] {
635     switch (CurrentCUDATarget()) {
636     case CFT_Host:
637       return DeviceDiagBuilder::K_Immediate;
638     case CFT_HostDevice:
639       // An HD function counts as host code if we're compiling for host, and
640       // device code if we're compiling for device.  Defer any errors in device
641       // mode until the function is known-emitted.
642       if (getLangOpts().CUDAIsDevice)
643         return DeviceDiagBuilder::K_Nop;
644 
645       return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
646               FunctionEmissionStatus::Emitted)
647                  ? DeviceDiagBuilder::K_ImmediateWithCallStack
648                  : DeviceDiagBuilder::K_Deferred;
649     default:
650       return DeviceDiagBuilder::K_Nop;
651     }
652   }();
653   return DeviceDiagBuilder(DiagKind, Loc, DiagID,
654                            dyn_cast<FunctionDecl>(CurContext), *this);
655 }
656 
657 bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) {
658   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
659   assert(Callee && "Callee may not be null.");
660 
661   auto &ExprEvalCtx = ExprEvalContexts.back();
662   if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated())
663     return true;
664 
665   // FIXME: Is bailing out early correct here?  Should we instead assume that
666   // the caller is a global initializer?
667   FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
668   if (!Caller)
669     return true;
670 
671   // If the caller is known-emitted, mark the callee as known-emitted.
672   // Otherwise, mark the call in our call graph so we can traverse it later.
673   bool CallerKnownEmitted =
674       getEmissionStatus(Caller) == FunctionEmissionStatus::Emitted;
675   if (CallerKnownEmitted) {
676     // Host-side references to a __global__ function refer to the stub, so the
677     // function itself is never emitted and therefore should not be marked.
678     if (!shouldIgnoreInHostDeviceCheck(Callee))
679       markKnownEmitted(
680           *this, Caller, Callee, Loc, [](Sema &S, FunctionDecl *FD) {
681             return S.getEmissionStatus(FD) == FunctionEmissionStatus::Emitted;
682           });
683   } else {
684     // If we have
685     //   host fn calls kernel fn calls host+device,
686     // the HD function does not get instantiated on the host.  We model this by
687     // omitting at the call to the kernel from the callgraph.  This ensures
688     // that, when compiling for host, only HD functions actually called from the
689     // host get marked as known-emitted.
690     if (!shouldIgnoreInHostDeviceCheck(Callee))
691       DeviceCallGraph[Caller].insert({Callee, Loc});
692   }
693 
694   DeviceDiagBuilder::Kind DiagKind = [this, Caller, Callee,
695                                       CallerKnownEmitted] {
696     switch (IdentifyCUDAPreference(Caller, Callee)) {
697     case CFP_Never:
698       return DeviceDiagBuilder::K_Immediate;
699     case CFP_WrongSide:
700       assert(Caller && "WrongSide calls require a non-null caller");
701       // If we know the caller will be emitted, we know this wrong-side call
702       // will be emitted, so it's an immediate error.  Otherwise, defer the
703       // error until we know the caller is emitted.
704       return CallerKnownEmitted ? DeviceDiagBuilder::K_ImmediateWithCallStack
705                                 : DeviceDiagBuilder::K_Deferred;
706     default:
707       return DeviceDiagBuilder::K_Nop;
708     }
709   }();
710 
711   if (DiagKind == DeviceDiagBuilder::K_Nop)
712     return true;
713 
714   // Avoid emitting this error twice for the same location.  Using a hashtable
715   // like this is unfortunate, but because we must continue parsing as normal
716   // after encountering a deferred error, it's otherwise very tricky for us to
717   // ensure that we only emit this deferred error once.
718   if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second)
719     return true;
720 
721   DeviceDiagBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this)
722       << IdentifyCUDATarget(Callee) << Callee << IdentifyCUDATarget(Caller);
723   DeviceDiagBuilder(DiagKind, Callee->getLocation(), diag::note_previous_decl,
724                     Caller, *this)
725       << Callee;
726   return DiagKind != DeviceDiagBuilder::K_Immediate &&
727          DiagKind != DeviceDiagBuilder::K_ImmediateWithCallStack;
728 }
729 
730 void Sema::CUDASetLambdaAttrs(CXXMethodDecl *Method) {
731   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
732   if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>())
733     return;
734   FunctionDecl *CurFn = dyn_cast<FunctionDecl>(CurContext);
735   if (!CurFn)
736     return;
737   CUDAFunctionTarget Target = IdentifyCUDATarget(CurFn);
738   if (Target == CFT_Global || Target == CFT_Device) {
739     Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
740   } else if (Target == CFT_HostDevice) {
741     Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
742     Method->addAttr(CUDAHostAttr::CreateImplicit(Context));
743   }
744 }
745 
746 void Sema::checkCUDATargetOverload(FunctionDecl *NewFD,
747                                    const LookupResult &Previous) {
748   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
749   CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD);
750   for (NamedDecl *OldND : Previous) {
751     FunctionDecl *OldFD = OldND->getAsFunction();
752     if (!OldFD)
753       continue;
754 
755     CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD);
756     // Don't allow HD and global functions to overload other functions with the
757     // same signature.  We allow overloading based on CUDA attributes so that
758     // functions can have different implementations on the host and device, but
759     // HD/global functions "exist" in some sense on both the host and device, so
760     // should have the same implementation on both sides.
761     if (NewTarget != OldTarget &&
762         ((NewTarget == CFT_HostDevice) || (OldTarget == CFT_HostDevice) ||
763          (NewTarget == CFT_Global) || (OldTarget == CFT_Global)) &&
764         !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false,
765                     /* ConsiderCudaAttrs = */ false)) {
766       Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)
767           << NewTarget << NewFD->getDeclName() << OldTarget << OldFD;
768       Diag(OldFD->getLocation(), diag::note_previous_declaration);
769       NewFD->setInvalidDecl();
770       break;
771     }
772   }
773 }
774 
775 template <typename AttrTy>
776 static void copyAttrIfPresent(Sema &S, FunctionDecl *FD,
777                               const FunctionDecl &TemplateFD) {
778   if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
779     AttrTy *Clone = Attribute->clone(S.Context);
780     Clone->setInherited(true);
781     FD->addAttr(Clone);
782   }
783 }
784 
785 void Sema::inheritCUDATargetAttrs(FunctionDecl *FD,
786                                   const FunctionTemplateDecl &TD) {
787   const FunctionDecl &TemplateFD = *TD.getTemplatedDecl();
788   copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD);
789   copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD);
790   copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD);
791 }
792 
793 std::string Sema::getCudaConfigureFuncName() const {
794   if (getLangOpts().HIP)
795     return getLangOpts().HIPUseNewLaunchAPI ? "__hipPushCallConfiguration"
796                                             : "hipConfigureCall";
797 
798   // New CUDA kernel launch sequence.
799   if (CudaFeatureEnabled(Context.getTargetInfo().getSDKVersion(),
800                          CudaFeature::CUDA_USES_NEW_LAUNCH))
801     return "__cudaPushCallConfiguration";
802 
803   // Legacy CUDA kernel configuration call
804   return "cudaConfigureCall";
805 }
806