xref: /freebsd/contrib/llvm-project/clang/lib/Sema/SemaCUDA.cpp (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
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 void Sema::PushForceCUDAHostDevice() {
30   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
31   ForceCUDAHostDeviceDepth++;
32 }
33 
34 bool Sema::PopForceCUDAHostDevice() {
35   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
36   if (ForceCUDAHostDeviceDepth == 0)
37     return false;
38   ForceCUDAHostDeviceDepth--;
39   return true;
40 }
41 
42 ExprResult Sema::ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc,
43                                          MultiExprArg ExecConfig,
44                                          SourceLocation GGGLoc) {
45   FunctionDecl *ConfigDecl = Context.getcudaConfigureCallDecl();
46   if (!ConfigDecl)
47     return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use)
48                      << getCudaConfigureFuncName());
49   QualType ConfigQTy = ConfigDecl->getType();
50 
51   DeclRefExpr *ConfigDR = new (Context)
52       DeclRefExpr(Context, ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc);
53   MarkFunctionReferenced(LLLLoc, ConfigDecl);
54 
55   return BuildCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr,
56                        /*IsExecConfig=*/true);
57 }
58 
59 Sema::CUDAFunctionTarget
60 Sema::IdentifyCUDATarget(const ParsedAttributesView &Attrs) {
61   bool HasHostAttr = false;
62   bool HasDeviceAttr = false;
63   bool HasGlobalAttr = false;
64   bool HasInvalidTargetAttr = false;
65   for (const ParsedAttr &AL : Attrs) {
66     switch (AL.getKind()) {
67     case ParsedAttr::AT_CUDAGlobal:
68       HasGlobalAttr = true;
69       break;
70     case ParsedAttr::AT_CUDAHost:
71       HasHostAttr = true;
72       break;
73     case ParsedAttr::AT_CUDADevice:
74       HasDeviceAttr = true;
75       break;
76     case ParsedAttr::AT_CUDAInvalidTarget:
77       HasInvalidTargetAttr = true;
78       break;
79     default:
80       break;
81     }
82   }
83 
84   if (HasInvalidTargetAttr)
85     return CFT_InvalidTarget;
86 
87   if (HasGlobalAttr)
88     return CFT_Global;
89 
90   if (HasHostAttr && HasDeviceAttr)
91     return CFT_HostDevice;
92 
93   if (HasDeviceAttr)
94     return CFT_Device;
95 
96   return CFT_Host;
97 }
98 
99 template <typename A>
100 static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
101   return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
102            return isa<A>(Attribute) &&
103                   !(IgnoreImplicitAttr && Attribute->isImplicit());
104          });
105 }
106 
107 /// IdentifyCUDATarget - Determine the CUDA compilation target for this function
108 Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D,
109                                                   bool IgnoreImplicitHDAttr) {
110   // Code that lives outside a function is run on the host.
111   if (D == nullptr)
112     return CFT_Host;
113 
114   if (D->hasAttr<CUDAInvalidTargetAttr>())
115     return CFT_InvalidTarget;
116 
117   if (D->hasAttr<CUDAGlobalAttr>())
118     return CFT_Global;
119 
120   if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) {
121     if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr))
122       return CFT_HostDevice;
123     return CFT_Device;
124   } else if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) {
125     return CFT_Host;
126   } else if (D->isImplicit() && !IgnoreImplicitHDAttr) {
127     // Some implicit declarations (like intrinsic functions) are not marked.
128     // Set the most lenient target on them for maximal flexibility.
129     return CFT_HostDevice;
130   }
131 
132   return CFT_Host;
133 }
134 
135 // * CUDA Call preference table
136 //
137 // F - from,
138 // T - to
139 // Ph - preference in host mode
140 // Pd - preference in device mode
141 // H  - handled in (x)
142 // Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never.
143 //
144 // | F  | T  | Ph  | Pd  |  H  |
145 // |----+----+-----+-----+-----+
146 // | d  | d  | N   | N   | (c) |
147 // | d  | g  | --  | --  | (a) |
148 // | d  | h  | --  | --  | (e) |
149 // | d  | hd | HD  | HD  | (b) |
150 // | g  | d  | N   | N   | (c) |
151 // | g  | g  | --  | --  | (a) |
152 // | g  | h  | --  | --  | (e) |
153 // | g  | hd | HD  | HD  | (b) |
154 // | h  | d  | --  | --  | (e) |
155 // | h  | g  | N   | N   | (c) |
156 // | h  | h  | N   | N   | (c) |
157 // | h  | hd | HD  | HD  | (b) |
158 // | hd | d  | WS  | SS  | (d) |
159 // | hd | g  | SS  | --  |(d/a)|
160 // | hd | h  | SS  | WS  | (d) |
161 // | hd | hd | HD  | HD  | (b) |
162 
163 Sema::CUDAFunctionPreference
164 Sema::IdentifyCUDAPreference(const FunctionDecl *Caller,
165                              const FunctionDecl *Callee) {
166   assert(Callee && "Callee must be valid.");
167   CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller);
168   CUDAFunctionTarget CalleeTarget = IdentifyCUDATarget(Callee);
169 
170   // If one of the targets is invalid, the check always fails, no matter what
171   // the other target is.
172   if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget)
173     return CFP_Never;
174 
175   // (a) Can't call global from some contexts until we support CUDA's
176   // dynamic parallelism.
177   if (CalleeTarget == CFT_Global &&
178       (CallerTarget == CFT_Global || CallerTarget == CFT_Device))
179     return CFP_Never;
180 
181   // (b) Calling HostDevice is OK for everyone.
182   if (CalleeTarget == CFT_HostDevice)
183     return CFP_HostDevice;
184 
185   // (c) Best case scenarios
186   if (CalleeTarget == CallerTarget ||
187       (CallerTarget == CFT_Host && CalleeTarget == CFT_Global) ||
188       (CallerTarget == CFT_Global && CalleeTarget == CFT_Device))
189     return CFP_Native;
190 
191   // (d) HostDevice behavior depends on compilation mode.
192   if (CallerTarget == CFT_HostDevice) {
193     // It's OK to call a compilation-mode matching function from an HD one.
194     if ((getLangOpts().CUDAIsDevice && CalleeTarget == CFT_Device) ||
195         (!getLangOpts().CUDAIsDevice &&
196          (CalleeTarget == CFT_Host || CalleeTarget == CFT_Global)))
197       return CFP_SameSide;
198 
199     // Calls from HD to non-mode-matching functions (i.e., to host functions
200     // when compiling in device mode or to device functions when compiling in
201     // host mode) are allowed at the sema level, but eventually rejected if
202     // they're ever codegened.  TODO: Reject said calls earlier.
203     return CFP_WrongSide;
204   }
205 
206   // (e) Calling across device/host boundary is not something you should do.
207   if ((CallerTarget == CFT_Host && CalleeTarget == CFT_Device) ||
208       (CallerTarget == CFT_Device && CalleeTarget == CFT_Host) ||
209       (CallerTarget == CFT_Global && CalleeTarget == CFT_Host))
210     return CFP_Never;
211 
212   llvm_unreachable("All cases should've been handled by now.");
213 }
214 
215 template <typename AttrT> static bool hasImplicitAttr(const FunctionDecl *D) {
216   if (!D)
217     return false;
218   if (auto *A = D->getAttr<AttrT>())
219     return A->isImplicit();
220   return D->isImplicit();
221 }
222 
223 bool Sema::isCUDAImplicitHostDeviceFunction(const FunctionDecl *D) {
224   bool IsImplicitDevAttr = hasImplicitAttr<CUDADeviceAttr>(D);
225   bool IsImplicitHostAttr = hasImplicitAttr<CUDAHostAttr>(D);
226   return IsImplicitDevAttr && IsImplicitHostAttr;
227 }
228 
229 void Sema::EraseUnwantedCUDAMatches(
230     const FunctionDecl *Caller,
231     SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) {
232   if (Matches.size() <= 1)
233     return;
234 
235   using Pair = std::pair<DeclAccessPair, FunctionDecl*>;
236 
237   // Gets the CUDA function preference for a call from Caller to Match.
238   auto GetCFP = [&](const Pair &Match) {
239     return IdentifyCUDAPreference(Caller, Match.second);
240   };
241 
242   // Find the best call preference among the functions in Matches.
243   CUDAFunctionPreference BestCFP = GetCFP(*std::max_element(
244       Matches.begin(), Matches.end(),
245       [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); }));
246 
247   // Erase all functions with lower priority.
248   llvm::erase_if(Matches,
249                  [&](const Pair &Match) { return GetCFP(Match) < BestCFP; });
250 }
251 
252 /// When an implicitly-declared special member has to invoke more than one
253 /// base/field special member, conflicts may occur in the targets of these
254 /// members. For example, if one base's member __host__ and another's is
255 /// __device__, it's a conflict.
256 /// This function figures out if the given targets \param Target1 and
257 /// \param Target2 conflict, and if they do not it fills in
258 /// \param ResolvedTarget with a target that resolves for both calls.
259 /// \return true if there's a conflict, false otherwise.
260 static bool
261 resolveCalleeCUDATargetConflict(Sema::CUDAFunctionTarget Target1,
262                                 Sema::CUDAFunctionTarget Target2,
263                                 Sema::CUDAFunctionTarget *ResolvedTarget) {
264   // Only free functions and static member functions may be global.
265   assert(Target1 != Sema::CFT_Global);
266   assert(Target2 != Sema::CFT_Global);
267 
268   if (Target1 == Sema::CFT_HostDevice) {
269     *ResolvedTarget = Target2;
270   } else if (Target2 == Sema::CFT_HostDevice) {
271     *ResolvedTarget = Target1;
272   } else if (Target1 != Target2) {
273     return true;
274   } else {
275     *ResolvedTarget = Target1;
276   }
277 
278   return false;
279 }
280 
281 bool Sema::inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl,
282                                                    CXXSpecialMember CSM,
283                                                    CXXMethodDecl *MemberDecl,
284                                                    bool ConstRHS,
285                                                    bool Diagnose) {
286   // If the defaulted special member is defined lexically outside of its
287   // owning class, or the special member already has explicit device or host
288   // attributes, do not infer.
289   bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent();
290   bool HasH = MemberDecl->hasAttr<CUDAHostAttr>();
291   bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>();
292   bool HasExplicitAttr =
293       (HasD && !MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()) ||
294       (HasH && !MemberDecl->getAttr<CUDAHostAttr>()->isImplicit());
295   if (!InClass || HasExplicitAttr)
296     return false;
297 
298   llvm::Optional<CUDAFunctionTarget> InferredTarget;
299 
300   // We're going to invoke special member lookup; mark that these special
301   // members are called from this one, and not from its caller.
302   ContextRAII MethodContext(*this, MemberDecl);
303 
304   // Look for special members in base classes that should be invoked from here.
305   // Infer the target of this member base on the ones it should call.
306   // Skip direct and indirect virtual bases for abstract classes.
307   llvm::SmallVector<const CXXBaseSpecifier *, 16> Bases;
308   for (const auto &B : ClassDecl->bases()) {
309     if (!B.isVirtual()) {
310       Bases.push_back(&B);
311     }
312   }
313 
314   if (!ClassDecl->isAbstract()) {
315     for (const auto &VB : ClassDecl->vbases()) {
316       Bases.push_back(&VB);
317     }
318   }
319 
320   for (const auto *B : Bases) {
321     const RecordType *BaseType = B->getType()->getAs<RecordType>();
322     if (!BaseType) {
323       continue;
324     }
325 
326     CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());
327     Sema::SpecialMemberOverloadResult SMOR =
328         LookupSpecialMember(BaseClassDecl, CSM,
329                             /* ConstArg */ ConstRHS,
330                             /* VolatileArg */ false,
331                             /* RValueThis */ false,
332                             /* ConstThis */ false,
333                             /* VolatileThis */ false);
334 
335     if (!SMOR.getMethod())
336       continue;
337 
338     CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR.getMethod());
339     if (!InferredTarget.hasValue()) {
340       InferredTarget = BaseMethodTarget;
341     } else {
342       bool ResolutionError = resolveCalleeCUDATargetConflict(
343           InferredTarget.getValue(), BaseMethodTarget,
344           InferredTarget.getPointer());
345       if (ResolutionError) {
346         if (Diagnose) {
347           Diag(ClassDecl->getLocation(),
348                diag::note_implicit_member_target_infer_collision)
349               << (unsigned)CSM << InferredTarget.getValue() << BaseMethodTarget;
350         }
351         MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
352         return true;
353       }
354     }
355   }
356 
357   // Same as for bases, but now for special members of fields.
358   for (const auto *F : ClassDecl->fields()) {
359     if (F->isInvalidDecl()) {
360       continue;
361     }
362 
363     const RecordType *FieldType =
364         Context.getBaseElementType(F->getType())->getAs<RecordType>();
365     if (!FieldType) {
366       continue;
367     }
368 
369     CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());
370     Sema::SpecialMemberOverloadResult SMOR =
371         LookupSpecialMember(FieldRecDecl, CSM,
372                             /* ConstArg */ ConstRHS && !F->isMutable(),
373                             /* VolatileArg */ false,
374                             /* RValueThis */ false,
375                             /* ConstThis */ false,
376                             /* VolatileThis */ false);
377 
378     if (!SMOR.getMethod())
379       continue;
380 
381     CUDAFunctionTarget FieldMethodTarget =
382         IdentifyCUDATarget(SMOR.getMethod());
383     if (!InferredTarget.hasValue()) {
384       InferredTarget = FieldMethodTarget;
385     } else {
386       bool ResolutionError = resolveCalleeCUDATargetConflict(
387           InferredTarget.getValue(), FieldMethodTarget,
388           InferredTarget.getPointer());
389       if (ResolutionError) {
390         if (Diagnose) {
391           Diag(ClassDecl->getLocation(),
392                diag::note_implicit_member_target_infer_collision)
393               << (unsigned)CSM << InferredTarget.getValue()
394               << FieldMethodTarget;
395         }
396         MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
397         return true;
398       }
399     }
400   }
401 
402 
403   // If no target was inferred, mark this member as __host__ __device__;
404   // it's the least restrictive option that can be invoked from any target.
405   bool NeedsH = true, NeedsD = true;
406   if (InferredTarget.hasValue()) {
407     if (InferredTarget.getValue() == CFT_Device)
408       NeedsH = false;
409     else if (InferredTarget.getValue() == CFT_Host)
410       NeedsD = false;
411   }
412 
413   // We either setting attributes first time, or the inferred ones must match
414   // previously set ones.
415   if (NeedsD && !HasD)
416     MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
417   if (NeedsH && !HasH)
418     MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
419 
420   return false;
421 }
422 
423 bool Sema::isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD) {
424   if (!CD->isDefined() && CD->isTemplateInstantiation())
425     InstantiateFunctionDefinition(Loc, CD->getFirstDecl());
426 
427   // (E.2.3.1, CUDA 7.5) A constructor for a class type is considered
428   // empty at a point in the translation unit, if it is either a
429   // trivial constructor
430   if (CD->isTrivial())
431     return true;
432 
433   // ... or it satisfies all of the following conditions:
434   // The constructor function has been defined.
435   // The constructor function has no parameters,
436   // and the function body is an empty compound statement.
437   if (!(CD->hasTrivialBody() && CD->getNumParams() == 0))
438     return false;
439 
440   // Its class has no virtual functions and no virtual base classes.
441   if (CD->getParent()->isDynamicClass())
442     return false;
443 
444   // Union ctor does not call ctors of its data members.
445   if (CD->getParent()->isUnion())
446     return true;
447 
448   // The only form of initializer allowed is an empty constructor.
449   // This will recursively check all base classes and member initializers
450   if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {
451         if (const CXXConstructExpr *CE =
452                 dyn_cast<CXXConstructExpr>(CI->getInit()))
453           return isEmptyCudaConstructor(Loc, CE->getConstructor());
454         return false;
455       }))
456     return false;
457 
458   return true;
459 }
460 
461 bool Sema::isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *DD) {
462   // No destructor -> no problem.
463   if (!DD)
464     return true;
465 
466   if (!DD->isDefined() && DD->isTemplateInstantiation())
467     InstantiateFunctionDefinition(Loc, DD->getFirstDecl());
468 
469   // (E.2.3.1, CUDA 7.5) A destructor for a class type is considered
470   // empty at a point in the translation unit, if it is either a
471   // trivial constructor
472   if (DD->isTrivial())
473     return true;
474 
475   // ... or it satisfies all of the following conditions:
476   // The destructor function has been defined.
477   // and the function body is an empty compound statement.
478   if (!DD->hasTrivialBody())
479     return false;
480 
481   const CXXRecordDecl *ClassDecl = DD->getParent();
482 
483   // Its class has no virtual functions and no virtual base classes.
484   if (ClassDecl->isDynamicClass())
485     return false;
486 
487   // Union does not have base class and union dtor does not call dtors of its
488   // data members.
489   if (DD->getParent()->isUnion())
490     return true;
491 
492   // Only empty destructors are allowed. This will recursively check
493   // destructors for all base classes...
494   if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {
495         if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())
496           return isEmptyCudaDestructor(Loc, RD->getDestructor());
497         return true;
498       }))
499     return false;
500 
501   // ... and member fields.
502   if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {
503         if (CXXRecordDecl *RD = Field->getType()
504                                     ->getBaseElementTypeUnsafe()
505                                     ->getAsCXXRecordDecl())
506           return isEmptyCudaDestructor(Loc, RD->getDestructor());
507         return true;
508       }))
509     return false;
510 
511   return true;
512 }
513 
514 void Sema::checkAllowedCUDAInitializer(VarDecl *VD) {
515   if (VD->isInvalidDecl() || !VD->hasInit() || !VD->hasGlobalStorage())
516     return;
517   const Expr *Init = VD->getInit();
518   if (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>() ||
519       VD->hasAttr<CUDASharedAttr>()) {
520     if (LangOpts.GPUAllowDeviceInit)
521       return;
522     assert(!VD->isStaticLocal() || VD->hasAttr<CUDASharedAttr>());
523     bool AllowedInit = false;
524     if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init))
525       AllowedInit =
526           isEmptyCudaConstructor(VD->getLocation(), CE->getConstructor());
527     // We'll allow constant initializers even if it's a non-empty
528     // constructor according to CUDA rules. This deviates from NVCC,
529     // but allows us to handle things like constexpr constructors.
530     if (!AllowedInit &&
531         (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>())) {
532       auto *Init = VD->getInit();
533       AllowedInit =
534           ((VD->getType()->isDependentType() || Init->isValueDependent()) &&
535            VD->isConstexpr()) ||
536           Init->isConstantInitializer(Context,
537                                       VD->getType()->isReferenceType());
538     }
539 
540     // Also make sure that destructor, if there is one, is empty.
541     if (AllowedInit)
542       if (CXXRecordDecl *RD = VD->getType()->getAsCXXRecordDecl())
543         AllowedInit =
544             isEmptyCudaDestructor(VD->getLocation(), RD->getDestructor());
545 
546     if (!AllowedInit) {
547       Diag(VD->getLocation(), VD->hasAttr<CUDASharedAttr>()
548                                   ? diag::err_shared_var_init
549                                   : diag::err_dynamic_var_init)
550           << Init->getSourceRange();
551       VD->setInvalidDecl();
552     }
553   } else {
554     // This is a host-side global variable.  Check that the initializer is
555     // callable from the host side.
556     const FunctionDecl *InitFn = nullptr;
557     if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init)) {
558       InitFn = CE->getConstructor();
559     } else if (const CallExpr *CE = dyn_cast<CallExpr>(Init)) {
560       InitFn = CE->getDirectCallee();
561     }
562     if (InitFn) {
563       CUDAFunctionTarget InitFnTarget = IdentifyCUDATarget(InitFn);
564       if (InitFnTarget != CFT_Host && InitFnTarget != CFT_HostDevice) {
565         Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer)
566             << InitFnTarget << InitFn;
567         Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn;
568         VD->setInvalidDecl();
569       }
570     }
571   }
572 }
573 
574 // With -fcuda-host-device-constexpr, an unattributed constexpr function is
575 // treated as implicitly __host__ __device__, unless:
576 //  * it is a variadic function (device-side variadic functions are not
577 //    allowed), or
578 //  * a __device__ function with this signature was already declared, in which
579 //    case in which case we output an error, unless the __device__ decl is in a
580 //    system header, in which case we leave the constexpr function unattributed.
581 //
582 // In addition, all function decls are treated as __host__ __device__ when
583 // ForceCUDAHostDeviceDepth > 0 (corresponding to code within a
584 //   #pragma clang force_cuda_host_device_begin/end
585 // pair).
586 void Sema::maybeAddCUDAHostDeviceAttrs(FunctionDecl *NewD,
587                                        const LookupResult &Previous) {
588   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
589 
590   if (ForceCUDAHostDeviceDepth > 0) {
591     if (!NewD->hasAttr<CUDAHostAttr>())
592       NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
593     if (!NewD->hasAttr<CUDADeviceAttr>())
594       NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
595     return;
596   }
597 
598   if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() ||
599       NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() ||
600       NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>())
601     return;
602 
603   // Is D a __device__ function with the same signature as NewD, ignoring CUDA
604   // attributes?
605   auto IsMatchingDeviceFn = [&](NamedDecl *D) {
606     if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D))
607       D = Using->getTargetDecl();
608     FunctionDecl *OldD = D->getAsFunction();
609     return OldD && OldD->hasAttr<CUDADeviceAttr>() &&
610            !OldD->hasAttr<CUDAHostAttr>() &&
611            !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false,
612                        /* ConsiderCudaAttrs = */ false);
613   };
614   auto It = llvm::find_if(Previous, IsMatchingDeviceFn);
615   if (It != Previous.end()) {
616     // We found a __device__ function with the same name and signature as NewD
617     // (ignoring CUDA attrs).  This is an error unless that function is defined
618     // in a system header, in which case we simply return without making NewD
619     // host+device.
620     NamedDecl *Match = *It;
621     if (!getSourceManager().isInSystemHeader(Match->getLocation())) {
622       Diag(NewD->getLocation(),
623            diag::err_cuda_unattributed_constexpr_cannot_overload_device)
624           << NewD;
625       Diag(Match->getLocation(),
626            diag::note_cuda_conflicting_device_function_declared_here);
627     }
628     return;
629   }
630 
631   NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
632   NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
633 }
634 
635 void Sema::MaybeAddCUDAConstantAttr(VarDecl *VD) {
636   if (getLangOpts().CUDAIsDevice && VD->isConstexpr() &&
637       (VD->isFileVarDecl() || VD->isStaticDataMember())) {
638     VD->addAttr(CUDAConstantAttr::CreateImplicit(getASTContext()));
639   }
640 }
641 
642 Sema::DeviceDiagBuilder Sema::CUDADiagIfDeviceCode(SourceLocation Loc,
643                                                    unsigned DiagID) {
644   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
645   DeviceDiagBuilder::Kind DiagKind = [this] {
646     switch (CurrentCUDATarget()) {
647     case CFT_Global:
648     case CFT_Device:
649       return DeviceDiagBuilder::K_Immediate;
650     case CFT_HostDevice:
651       // An HD function counts as host code if we're compiling for host, and
652       // device code if we're compiling for device.  Defer any errors in device
653       // mode until the function is known-emitted.
654       if (getLangOpts().CUDAIsDevice) {
655         return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
656                 FunctionEmissionStatus::Emitted)
657                    ? DeviceDiagBuilder::K_ImmediateWithCallStack
658                    : DeviceDiagBuilder::K_Deferred;
659       }
660       return DeviceDiagBuilder::K_Nop;
661 
662     default:
663       return DeviceDiagBuilder::K_Nop;
664     }
665   }();
666   return DeviceDiagBuilder(DiagKind, Loc, DiagID,
667                            dyn_cast<FunctionDecl>(CurContext), *this);
668 }
669 
670 Sema::DeviceDiagBuilder Sema::CUDADiagIfHostCode(SourceLocation Loc,
671                                                  unsigned DiagID) {
672   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
673   DeviceDiagBuilder::Kind DiagKind = [this] {
674     switch (CurrentCUDATarget()) {
675     case CFT_Host:
676       return DeviceDiagBuilder::K_Immediate;
677     case CFT_HostDevice:
678       // An HD function counts as host code if we're compiling for host, and
679       // device code if we're compiling for device.  Defer any errors in device
680       // mode until the function is known-emitted.
681       if (getLangOpts().CUDAIsDevice)
682         return DeviceDiagBuilder::K_Nop;
683 
684       return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
685               FunctionEmissionStatus::Emitted)
686                  ? DeviceDiagBuilder::K_ImmediateWithCallStack
687                  : DeviceDiagBuilder::K_Deferred;
688     default:
689       return DeviceDiagBuilder::K_Nop;
690     }
691   }();
692   return DeviceDiagBuilder(DiagKind, Loc, DiagID,
693                            dyn_cast<FunctionDecl>(CurContext), *this);
694 }
695 
696 bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) {
697   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
698   assert(Callee && "Callee may not be null.");
699 
700   auto &ExprEvalCtx = ExprEvalContexts.back();
701   if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated())
702     return true;
703 
704   // FIXME: Is bailing out early correct here?  Should we instead assume that
705   // the caller is a global initializer?
706   FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
707   if (!Caller)
708     return true;
709 
710   // If the caller is known-emitted, mark the callee as known-emitted.
711   // Otherwise, mark the call in our call graph so we can traverse it later.
712   bool CallerKnownEmitted =
713       getEmissionStatus(Caller) == FunctionEmissionStatus::Emitted;
714   DeviceDiagBuilder::Kind DiagKind = [this, Caller, Callee,
715                                       CallerKnownEmitted] {
716     switch (IdentifyCUDAPreference(Caller, Callee)) {
717     case CFP_Never:
718       return DeviceDiagBuilder::K_Immediate;
719     case CFP_WrongSide:
720       assert(Caller && "WrongSide calls require a non-null caller");
721       // If we know the caller will be emitted, we know this wrong-side call
722       // will be emitted, so it's an immediate error.  Otherwise, defer the
723       // error until we know the caller is emitted.
724       return CallerKnownEmitted ? DeviceDiagBuilder::K_ImmediateWithCallStack
725                                 : DeviceDiagBuilder::K_Deferred;
726     default:
727       return DeviceDiagBuilder::K_Nop;
728     }
729   }();
730 
731   if (DiagKind == DeviceDiagBuilder::K_Nop)
732     return true;
733 
734   // Avoid emitting this error twice for the same location.  Using a hashtable
735   // like this is unfortunate, but because we must continue parsing as normal
736   // after encountering a deferred error, it's otherwise very tricky for us to
737   // ensure that we only emit this deferred error once.
738   if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second)
739     return true;
740 
741   DeviceDiagBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this)
742       << IdentifyCUDATarget(Callee) << Callee << IdentifyCUDATarget(Caller);
743   DeviceDiagBuilder(DiagKind, Callee->getLocation(), diag::note_previous_decl,
744                     Caller, *this)
745       << Callee;
746   return DiagKind != DeviceDiagBuilder::K_Immediate &&
747          DiagKind != DeviceDiagBuilder::K_ImmediateWithCallStack;
748 }
749 
750 // Check the wrong-sided reference capture of lambda for CUDA/HIP.
751 // A lambda function may capture a stack variable by reference when it is
752 // defined and uses the capture by reference when the lambda is called. When
753 // the capture and use happen on different sides, the capture is invalid and
754 // should be diagnosed.
755 void Sema::CUDACheckLambdaCapture(CXXMethodDecl *Callee,
756                                   const sema::Capture &Capture) {
757   // In host compilation we only need to check lambda functions emitted on host
758   // side. In such lambda functions, a reference capture is invalid only
759   // if the lambda structure is populated by a device function or kernel then
760   // is passed to and called by a host function. However that is impossible,
761   // since a device function or kernel can only call a device function, also a
762   // kernel cannot pass a lambda back to a host function since we cannot
763   // define a kernel argument type which can hold the lambda before the lambda
764   // itself is defined.
765   if (!LangOpts.CUDAIsDevice)
766     return;
767 
768   // File-scope lambda can only do init captures for global variables, which
769   // results in passing by value for these global variables.
770   FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
771   if (!Caller)
772     return;
773 
774   // In device compilation, we only need to check lambda functions which are
775   // emitted on device side. For such lambdas, a reference capture is invalid
776   // only if the lambda structure is populated by a host function then passed
777   // to and called in a device function or kernel.
778   bool CalleeIsDevice = Callee->hasAttr<CUDADeviceAttr>();
779   bool CallerIsHost =
780       !Caller->hasAttr<CUDAGlobalAttr>() && !Caller->hasAttr<CUDADeviceAttr>();
781   bool ShouldCheck = CalleeIsDevice && CallerIsHost;
782   if (!ShouldCheck || !Capture.isReferenceCapture())
783     return;
784   auto DiagKind = DeviceDiagBuilder::K_Deferred;
785   if (Capture.isVariableCapture()) {
786     DeviceDiagBuilder(DiagKind, Capture.getLocation(),
787                       diag::err_capture_bad_target, Callee, *this)
788         << Capture.getVariable();
789   } else if (Capture.isThisCapture()) {
790     DeviceDiagBuilder(DiagKind, Capture.getLocation(),
791                       diag::err_capture_bad_target_this_ptr, Callee, *this);
792   }
793   return;
794 }
795 
796 void Sema::CUDASetLambdaAttrs(CXXMethodDecl *Method) {
797   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
798   if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>())
799     return;
800   Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
801   Method->addAttr(CUDAHostAttr::CreateImplicit(Context));
802 }
803 
804 void Sema::checkCUDATargetOverload(FunctionDecl *NewFD,
805                                    const LookupResult &Previous) {
806   assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
807   CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD);
808   for (NamedDecl *OldND : Previous) {
809     FunctionDecl *OldFD = OldND->getAsFunction();
810     if (!OldFD)
811       continue;
812 
813     CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD);
814     // Don't allow HD and global functions to overload other functions with the
815     // same signature.  We allow overloading based on CUDA attributes so that
816     // functions can have different implementations on the host and device, but
817     // HD/global functions "exist" in some sense on both the host and device, so
818     // should have the same implementation on both sides.
819     if (NewTarget != OldTarget &&
820         ((NewTarget == CFT_HostDevice) || (OldTarget == CFT_HostDevice) ||
821          (NewTarget == CFT_Global) || (OldTarget == CFT_Global)) &&
822         !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false,
823                     /* ConsiderCudaAttrs = */ false)) {
824       Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)
825           << NewTarget << NewFD->getDeclName() << OldTarget << OldFD;
826       Diag(OldFD->getLocation(), diag::note_previous_declaration);
827       NewFD->setInvalidDecl();
828       break;
829     }
830   }
831 }
832 
833 template <typename AttrTy>
834 static void copyAttrIfPresent(Sema &S, FunctionDecl *FD,
835                               const FunctionDecl &TemplateFD) {
836   if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
837     AttrTy *Clone = Attribute->clone(S.Context);
838     Clone->setInherited(true);
839     FD->addAttr(Clone);
840   }
841 }
842 
843 void Sema::inheritCUDATargetAttrs(FunctionDecl *FD,
844                                   const FunctionTemplateDecl &TD) {
845   const FunctionDecl &TemplateFD = *TD.getTemplatedDecl();
846   copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD);
847   copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD);
848   copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD);
849 }
850 
851 std::string Sema::getCudaConfigureFuncName() const {
852   if (getLangOpts().HIP)
853     return getLangOpts().HIPUseNewLaunchAPI ? "__hipPushCallConfiguration"
854                                             : "hipConfigureCall";
855 
856   // New CUDA kernel launch sequence.
857   if (CudaFeatureEnabled(Context.getTargetInfo().getSDKVersion(),
858                          CudaFeature::CUDA_USES_NEW_LAUNCH))
859     return "__cudaPushCallConfiguration";
860 
861   // Legacy CUDA kernel configuration call
862   return "cudaConfigureCall";
863 }
864