1 //===--- SemaCUDA.cpp - Semantic Analysis for CUDA constructs -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// \brief This file implements semantic analysis for CUDA constructs.
12 //===----------------------------------------------------------------------===//
14 #include "clang/AST/ASTContext.h"
15 #include "clang/AST/Decl.h"
16 #include "clang/AST/ExprCXX.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;
27 void Sema::PushForceCUDAHostDevice() {
28 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
29 ForceCUDAHostDeviceDepth++;
32 bool Sema::PopForceCUDAHostDevice() {
33 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
34 if (ForceCUDAHostDeviceDepth == 0)
36 ForceCUDAHostDeviceDepth--;
40 ExprResult Sema::ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc,
41 MultiExprArg ExecConfig,
42 SourceLocation GGGLoc) {
43 FunctionDecl *ConfigDecl = Context.getcudaConfigureCallDecl();
45 return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use)
46 << "cudaConfigureCall");
47 QualType ConfigQTy = ConfigDecl->getType();
49 DeclRefExpr *ConfigDR = new (Context)
50 DeclRefExpr(ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc);
51 MarkFunctionReferenced(LLLLoc, ConfigDecl);
53 return ActOnCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr,
54 /*IsExecConfig=*/true);
57 Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const AttributeList *Attr) {
58 bool HasHostAttr = false;
59 bool HasDeviceAttr = false;
60 bool HasGlobalAttr = false;
61 bool HasInvalidTargetAttr = false;
63 switch(Attr->getKind()){
64 case AttributeList::AT_CUDAGlobal:
67 case AttributeList::AT_CUDAHost:
70 case AttributeList::AT_CUDADevice:
73 case AttributeList::AT_CUDAInvalidTarget:
74 HasInvalidTargetAttr = true;
79 Attr = Attr->getNext();
81 if (HasInvalidTargetAttr)
82 return CFT_InvalidTarget;
87 if (HasHostAttr && HasDeviceAttr)
88 return CFT_HostDevice;
97 static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
98 return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
99 return isa<A>(Attribute) &&
100 !(IgnoreImplicitAttr && Attribute->isImplicit());
104 /// IdentifyCUDATarget - Determine the CUDA compilation target for this function
105 Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D,
106 bool IgnoreImplicitHDAttr) {
107 // Code that lives outside a function is run on the host.
111 if (D->hasAttr<CUDAInvalidTargetAttr>())
112 return CFT_InvalidTarget;
114 if (D->hasAttr<CUDAGlobalAttr>())
117 if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) {
118 if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr))
119 return CFT_HostDevice;
121 } else if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) {
123 } else if (D->isImplicit() && !IgnoreImplicitHDAttr) {
124 // Some implicit declarations (like intrinsic functions) are not marked.
125 // Set the most lenient target on them for maximal flexibility.
126 return CFT_HostDevice;
132 // * CUDA Call preference table
136 // Ph - preference in host mode
137 // Pd - preference in device mode
138 // H - handled in (x)
139 // Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never.
141 // | F | T | Ph | Pd | H |
142 // |----+----+-----+-----+-----+
143 // | d | d | N | N | (c) |
144 // | d | g | -- | -- | (a) |
145 // | d | h | -- | -- | (e) |
146 // | d | hd | HD | HD | (b) |
147 // | g | d | N | N | (c) |
148 // | g | g | -- | -- | (a) |
149 // | g | h | -- | -- | (e) |
150 // | g | hd | HD | HD | (b) |
151 // | h | d | -- | -- | (e) |
152 // | h | g | N | N | (c) |
153 // | h | h | N | N | (c) |
154 // | h | hd | HD | HD | (b) |
155 // | hd | d | WS | SS | (d) |
156 // | hd | g | SS | -- |(d/a)|
157 // | hd | h | SS | WS | (d) |
158 // | hd | hd | HD | HD | (b) |
160 Sema::CUDAFunctionPreference
161 Sema::IdentifyCUDAPreference(const FunctionDecl *Caller,
162 const FunctionDecl *Callee) {
163 assert(Callee && "Callee must be valid.");
164 CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller);
165 CUDAFunctionTarget CalleeTarget = IdentifyCUDATarget(Callee);
167 // If one of the targets is invalid, the check always fails, no matter what
168 // the other target is.
169 if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget)
172 // (a) Can't call global from some contexts until we support CUDA's
173 // dynamic parallelism.
174 if (CalleeTarget == CFT_Global &&
175 (CallerTarget == CFT_Global || CallerTarget == CFT_Device))
178 // (b) Calling HostDevice is OK for everyone.
179 if (CalleeTarget == CFT_HostDevice)
180 return CFP_HostDevice;
182 // (c) Best case scenarios
183 if (CalleeTarget == CallerTarget ||
184 (CallerTarget == CFT_Host && CalleeTarget == CFT_Global) ||
185 (CallerTarget == CFT_Global && CalleeTarget == CFT_Device))
188 // (d) HostDevice behavior depends on compilation mode.
189 if (CallerTarget == CFT_HostDevice) {
190 // It's OK to call a compilation-mode matching function from an HD one.
191 if ((getLangOpts().CUDAIsDevice && CalleeTarget == CFT_Device) ||
192 (!getLangOpts().CUDAIsDevice &&
193 (CalleeTarget == CFT_Host || CalleeTarget == CFT_Global)))
196 // Calls from HD to non-mode-matching functions (i.e., to host functions
197 // when compiling in device mode or to device functions when compiling in
198 // host mode) are allowed at the sema level, but eventually rejected if
199 // they're ever codegened. TODO: Reject said calls earlier.
200 return CFP_WrongSide;
203 // (e) Calling across device/host boundary is not something you should do.
204 if ((CallerTarget == CFT_Host && CalleeTarget == CFT_Device) ||
205 (CallerTarget == CFT_Device && CalleeTarget == CFT_Host) ||
206 (CallerTarget == CFT_Global && CalleeTarget == CFT_Host))
209 llvm_unreachable("All cases should've been handled by now.");
212 void Sema::EraseUnwantedCUDAMatches(
213 const FunctionDecl *Caller,
214 SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) {
215 if (Matches.size() <= 1)
218 using Pair = std::pair<DeclAccessPair, FunctionDecl*>;
220 // Gets the CUDA function preference for a call from Caller to Match.
221 auto GetCFP = [&](const Pair &Match) {
222 return IdentifyCUDAPreference(Caller, Match.second);
225 // Find the best call preference among the functions in Matches.
226 CUDAFunctionPreference BestCFP = GetCFP(*std::max_element(
227 Matches.begin(), Matches.end(),
228 [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); }));
230 // Erase all functions with lower priority.
231 llvm::erase_if(Matches,
232 [&](const Pair &Match) { return GetCFP(Match) < BestCFP; });
235 /// When an implicitly-declared special member has to invoke more than one
236 /// base/field special member, conflicts may occur in the targets of these
237 /// members. For example, if one base's member __host__ and another's is
238 /// __device__, it's a conflict.
239 /// This function figures out if the given targets \param Target1 and
240 /// \param Target2 conflict, and if they do not it fills in
241 /// \param ResolvedTarget with a target that resolves for both calls.
242 /// \return true if there's a conflict, false otherwise.
244 resolveCalleeCUDATargetConflict(Sema::CUDAFunctionTarget Target1,
245 Sema::CUDAFunctionTarget Target2,
246 Sema::CUDAFunctionTarget *ResolvedTarget) {
247 // Only free functions and static member functions may be global.
248 assert(Target1 != Sema::CFT_Global);
249 assert(Target2 != Sema::CFT_Global);
251 if (Target1 == Sema::CFT_HostDevice) {
252 *ResolvedTarget = Target2;
253 } else if (Target2 == Sema::CFT_HostDevice) {
254 *ResolvedTarget = Target1;
255 } else if (Target1 != Target2) {
258 *ResolvedTarget = Target1;
264 bool Sema::inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl,
265 CXXSpecialMember CSM,
266 CXXMethodDecl *MemberDecl,
269 llvm::Optional<CUDAFunctionTarget> InferredTarget;
271 // We're going to invoke special member lookup; mark that these special
272 // members are called from this one, and not from its caller.
273 ContextRAII MethodContext(*this, MemberDecl);
275 // Look for special members in base classes that should be invoked from here.
276 // Infer the target of this member base on the ones it should call.
277 // Skip direct and indirect virtual bases for abstract classes.
278 llvm::SmallVector<const CXXBaseSpecifier *, 16> Bases;
279 for (const auto &B : ClassDecl->bases()) {
280 if (!B.isVirtual()) {
285 if (!ClassDecl->isAbstract()) {
286 for (const auto &VB : ClassDecl->vbases()) {
287 Bases.push_back(&VB);
291 for (const auto *B : Bases) {
292 const RecordType *BaseType = B->getType()->getAs<RecordType>();
297 CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());
298 Sema::SpecialMemberOverloadResult *SMOR =
299 LookupSpecialMember(BaseClassDecl, CSM,
300 /* ConstArg */ ConstRHS,
301 /* VolatileArg */ false,
302 /* RValueThis */ false,
303 /* ConstThis */ false,
304 /* VolatileThis */ false);
306 if (!SMOR || !SMOR->getMethod()) {
310 CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR->getMethod());
311 if (!InferredTarget.hasValue()) {
312 InferredTarget = BaseMethodTarget;
314 bool ResolutionError = resolveCalleeCUDATargetConflict(
315 InferredTarget.getValue(), BaseMethodTarget,
316 InferredTarget.getPointer());
317 if (ResolutionError) {
319 Diag(ClassDecl->getLocation(),
320 diag::note_implicit_member_target_infer_collision)
321 << (unsigned)CSM << InferredTarget.getValue() << BaseMethodTarget;
323 MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
329 // Same as for bases, but now for special members of fields.
330 for (const auto *F : ClassDecl->fields()) {
331 if (F->isInvalidDecl()) {
335 const RecordType *FieldType =
336 Context.getBaseElementType(F->getType())->getAs<RecordType>();
341 CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());
342 Sema::SpecialMemberOverloadResult *SMOR =
343 LookupSpecialMember(FieldRecDecl, CSM,
344 /* ConstArg */ ConstRHS && !F->isMutable(),
345 /* VolatileArg */ false,
346 /* RValueThis */ false,
347 /* ConstThis */ false,
348 /* VolatileThis */ false);
350 if (!SMOR || !SMOR->getMethod()) {
354 CUDAFunctionTarget FieldMethodTarget =
355 IdentifyCUDATarget(SMOR->getMethod());
356 if (!InferredTarget.hasValue()) {
357 InferredTarget = FieldMethodTarget;
359 bool ResolutionError = resolveCalleeCUDATargetConflict(
360 InferredTarget.getValue(), FieldMethodTarget,
361 InferredTarget.getPointer());
362 if (ResolutionError) {
364 Diag(ClassDecl->getLocation(),
365 diag::note_implicit_member_target_infer_collision)
366 << (unsigned)CSM << InferredTarget.getValue()
367 << FieldMethodTarget;
369 MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
375 if (InferredTarget.hasValue()) {
376 if (InferredTarget.getValue() == CFT_Device) {
377 MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
378 } else if (InferredTarget.getValue() == CFT_Host) {
379 MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
381 MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
382 MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
385 // If no target was inferred, mark this member as __host__ __device__;
386 // it's the least restrictive option that can be invoked from any target.
387 MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
388 MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
394 bool Sema::isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD) {
395 if (!CD->isDefined() && CD->isTemplateInstantiation())
396 InstantiateFunctionDefinition(Loc, CD->getFirstDecl());
398 // (E.2.3.1, CUDA 7.5) A constructor for a class type is considered
399 // empty at a point in the translation unit, if it is either a
400 // trivial constructor
404 // ... or it satisfies all of the following conditions:
405 // The constructor function has been defined.
406 // The constructor function has no parameters,
407 // and the function body is an empty compound statement.
408 if (!(CD->hasTrivialBody() && CD->getNumParams() == 0))
411 // Its class has no virtual functions and no virtual base classes.
412 if (CD->getParent()->isDynamicClass())
415 // The only form of initializer allowed is an empty constructor.
416 // This will recursively check all base classes and member initializers
417 if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {
418 if (const CXXConstructExpr *CE =
419 dyn_cast<CXXConstructExpr>(CI->getInit()))
420 return isEmptyCudaConstructor(Loc, CE->getConstructor());
428 bool Sema::isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *DD) {
429 // No destructor -> no problem.
433 if (!DD->isDefined() && DD->isTemplateInstantiation())
434 InstantiateFunctionDefinition(Loc, DD->getFirstDecl());
436 // (E.2.3.1, CUDA 7.5) A destructor for a class type is considered
437 // empty at a point in the translation unit, if it is either a
438 // trivial constructor
442 // ... or it satisfies all of the following conditions:
443 // The destructor function has been defined.
444 // and the function body is an empty compound statement.
445 if (!DD->hasTrivialBody())
448 const CXXRecordDecl *ClassDecl = DD->getParent();
450 // Its class has no virtual functions and no virtual base classes.
451 if (ClassDecl->isDynamicClass())
454 // Only empty destructors are allowed. This will recursively check
455 // destructors for all base classes...
456 if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {
457 if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())
458 return isEmptyCudaDestructor(Loc, RD->getDestructor());
463 // ... and member fields.
464 if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {
465 if (CXXRecordDecl *RD = Field->getType()
466 ->getBaseElementTypeUnsafe()
467 ->getAsCXXRecordDecl())
468 return isEmptyCudaDestructor(Loc, RD->getDestructor());
476 // With -fcuda-host-device-constexpr, an unattributed constexpr function is
477 // treated as implicitly __host__ __device__, unless:
478 // * it is a variadic function (device-side variadic functions are not
480 // * a __device__ function with this signature was already declared, in which
481 // case in which case we output an error, unless the __device__ decl is in a
482 // system header, in which case we leave the constexpr function unattributed.
484 // In addition, all function decls are treated as __host__ __device__ when
485 // ForceCUDAHostDeviceDepth > 0 (corresponding to code within a
486 // #pragma clang force_cuda_host_device_begin/end
488 void Sema::maybeAddCUDAHostDeviceAttrs(FunctionDecl *NewD,
489 const LookupResult &Previous) {
490 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
492 if (ForceCUDAHostDeviceDepth > 0) {
493 if (!NewD->hasAttr<CUDAHostAttr>())
494 NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
495 if (!NewD->hasAttr<CUDADeviceAttr>())
496 NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
500 if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() ||
501 NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() ||
502 NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>())
505 // Is D a __device__ function with the same signature as NewD, ignoring CUDA
507 auto IsMatchingDeviceFn = [&](NamedDecl *D) {
508 if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D))
509 D = Using->getTargetDecl();
510 FunctionDecl *OldD = D->getAsFunction();
511 return OldD && OldD->hasAttr<CUDADeviceAttr>() &&
512 !OldD->hasAttr<CUDAHostAttr>() &&
513 !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false,
514 /* ConsiderCudaAttrs = */ false);
516 auto It = llvm::find_if(Previous, IsMatchingDeviceFn);
517 if (It != Previous.end()) {
518 // We found a __device__ function with the same name and signature as NewD
519 // (ignoring CUDA attrs). This is an error unless that function is defined
520 // in a system header, in which case we simply return without making NewD
522 NamedDecl *Match = *It;
523 if (!getSourceManager().isInSystemHeader(Match->getLocation())) {
524 Diag(NewD->getLocation(),
525 diag::err_cuda_unattributed_constexpr_cannot_overload_device)
527 Diag(Match->getLocation(),
528 diag::note_cuda_conflicting_device_function_declared_here);
533 NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
534 NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
537 // In CUDA, there are some constructs which may appear in semantically-valid
538 // code, but trigger errors if we ever generate code for the function in which
539 // they appear. Essentially every construct you're not allowed to use on the
540 // device falls into this category, because you are allowed to use these
541 // constructs in a __host__ __device__ function, but only if that function is
542 // never codegen'ed on the device.
544 // To handle semantic checking for these constructs, we keep track of the set of
545 // functions we know will be emitted, either because we could tell a priori that
546 // they would be emitted, or because they were transitively called by a
547 // known-emitted function.
549 // We also keep a partial call graph of which not-known-emitted functions call
550 // which other not-known-emitted functions.
552 // When we see something which is illegal if the current function is emitted
553 // (usually by way of CUDADiagIfDeviceCode, CUDADiagIfHostCode, or
554 // CheckCUDACall), we first check if the current function is known-emitted. If
555 // so, we immediately output the diagnostic.
557 // Otherwise, we "defer" the diagnostic. It sits in Sema::CUDADeferredDiags
558 // until we discover that the function is known-emitted, at which point we take
559 // it out of this map and emit the diagnostic.
561 Sema::CUDADiagBuilder::CUDADiagBuilder(Kind K, SourceLocation Loc,
562 unsigned DiagID, FunctionDecl *Fn,
564 : S(S), Loc(Loc), DiagID(DiagID), Fn(Fn),
565 ShowCallStack(K == K_ImmediateWithCallStack || K == K_Deferred) {
570 case K_ImmediateWithCallStack:
571 ImmediateDiag.emplace(S.Diag(Loc, DiagID));
574 assert(Fn && "Must have a function to attach the deferred diag to.");
575 PartialDiag.emplace(S.PDiag(DiagID));
580 // Print notes showing how we can reach FD starting from an a priori
581 // known-callable function.
582 static void EmitCallStackNotes(Sema &S, FunctionDecl *FD) {
583 auto FnIt = S.CUDAKnownEmittedFns.find(FD);
584 while (FnIt != S.CUDAKnownEmittedFns.end()) {
585 DiagnosticBuilder Builder(
586 S.Diags.Report(FnIt->second.Loc, diag::note_called_by));
587 Builder << FnIt->second.FD;
588 Builder.setForceEmit();
590 FnIt = S.CUDAKnownEmittedFns.find(FnIt->second.FD);
594 Sema::CUDADiagBuilder::~CUDADiagBuilder() {
596 // Emit our diagnostic and, if it was a warning or error, output a callstack
597 // if Fn isn't a priori known-emitted.
598 bool IsWarningOrError = S.getDiagnostics().getDiagnosticLevel(
599 DiagID, Loc) >= DiagnosticsEngine::Warning;
600 ImmediateDiag.reset(); // Emit the immediate diag.
601 if (IsWarningOrError && ShowCallStack)
602 EmitCallStackNotes(S, Fn);
603 } else if (PartialDiag) {
604 assert(ShowCallStack && "Must always show call stack for deferred diags.");
605 S.CUDADeferredDiags[Fn].push_back({Loc, std::move(*PartialDiag)});
609 // Do we know that we will eventually codegen the given function?
610 static bool IsKnownEmitted(Sema &S, FunctionDecl *FD) {
611 // Templates are emitted when they're instantiated.
612 if (FD->isDependentContext())
615 // When compiling for device, host functions are never emitted. Similarly,
616 // when compiling for host, device and global functions are never emitted.
617 // (Technically, we do emit a host-side stub for global functions, but this
618 // doesn't count for our purposes here.)
619 Sema::CUDAFunctionTarget T = S.IdentifyCUDATarget(FD);
620 if (S.getLangOpts().CUDAIsDevice && T == Sema::CFT_Host)
622 if (!S.getLangOpts().CUDAIsDevice &&
623 (T == Sema::CFT_Device || T == Sema::CFT_Global))
626 // Check whether this function is externally visible -- if so, it's
629 // We have to check the GVA linkage of the function's *definition* -- if we
630 // only have a declaration, we don't know whether or not the function will be
631 // emitted, because (say) the definition could include "inline".
632 FunctionDecl *Def = FD->getDefinition();
634 // We may currently be parsing the body of FD, in which case
635 // FD->getDefinition() will be null, but we still want to treat FD as though
636 // it's a definition.
637 if (!Def && FD->willHaveBody())
641 !isDiscardableGVALinkage(S.getASTContext().GetGVALinkageForFunction(Def)))
644 // Otherwise, the function is known-emitted if it's in our set of
645 // known-emitted functions.
646 return S.CUDAKnownEmittedFns.count(FD) > 0;
649 Sema::CUDADiagBuilder Sema::CUDADiagIfDeviceCode(SourceLocation Loc,
651 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
652 CUDADiagBuilder::Kind DiagKind = [&] {
653 switch (CurrentCUDATarget()) {
656 return CUDADiagBuilder::K_Immediate;
658 // An HD function counts as host code if we're compiling for host, and
659 // device code if we're compiling for device. Defer any errors in device
660 // mode until the function is known-emitted.
661 if (getLangOpts().CUDAIsDevice) {
662 return IsKnownEmitted(*this, dyn_cast<FunctionDecl>(CurContext))
663 ? CUDADiagBuilder::K_ImmediateWithCallStack
664 : CUDADiagBuilder::K_Deferred;
666 return CUDADiagBuilder::K_Nop;
669 return CUDADiagBuilder::K_Nop;
672 return CUDADiagBuilder(DiagKind, Loc, DiagID,
673 dyn_cast<FunctionDecl>(CurContext), *this);
676 Sema::CUDADiagBuilder Sema::CUDADiagIfHostCode(SourceLocation Loc,
678 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
679 CUDADiagBuilder::Kind DiagKind = [&] {
680 switch (CurrentCUDATarget()) {
682 return CUDADiagBuilder::K_Immediate;
684 // An HD function counts as host code if we're compiling for host, and
685 // device code if we're compiling for device. Defer any errors in device
686 // mode until the function is known-emitted.
687 if (getLangOpts().CUDAIsDevice)
688 return CUDADiagBuilder::K_Nop;
690 return IsKnownEmitted(*this, dyn_cast<FunctionDecl>(CurContext))
691 ? CUDADiagBuilder::K_ImmediateWithCallStack
692 : CUDADiagBuilder::K_Deferred;
694 return CUDADiagBuilder::K_Nop;
697 return CUDADiagBuilder(DiagKind, Loc, DiagID,
698 dyn_cast<FunctionDecl>(CurContext), *this);
701 // Emit any deferred diagnostics for FD and erase them from the map in which
703 static void EmitDeferredDiags(Sema &S, FunctionDecl *FD) {
704 auto It = S.CUDADeferredDiags.find(FD);
705 if (It == S.CUDADeferredDiags.end())
707 bool HasWarningOrError = false;
708 for (PartialDiagnosticAt &PDAt : It->second) {
709 const SourceLocation &Loc = PDAt.first;
710 const PartialDiagnostic &PD = PDAt.second;
711 HasWarningOrError |= S.getDiagnostics().getDiagnosticLevel(
712 PD.getDiagID(), Loc) >= DiagnosticsEngine::Warning;
713 DiagnosticBuilder Builder(S.Diags.Report(Loc, PD.getDiagID()));
714 Builder.setForceEmit();
717 S.CUDADeferredDiags.erase(It);
719 // FIXME: Should this be called after every warning/error emitted in the loop
720 // above, instead of just once per function? That would be consistent with
721 // how we handle immediate errors, but it also seems like a bit much.
722 if (HasWarningOrError)
723 EmitCallStackNotes(S, FD);
726 // Indicate that this function (and thus everything it transtively calls) will
727 // be codegen'ed, and emit any deferred diagnostics on this function and its
728 // (transitive) callees.
729 static void MarkKnownEmitted(Sema &S, FunctionDecl *OrigCaller,
730 FunctionDecl *OrigCallee, SourceLocation OrigLoc) {
731 // Nothing to do if we already know that FD is emitted.
732 if (IsKnownEmitted(S, OrigCallee)) {
733 assert(!S.CUDACallGraph.count(OrigCallee));
737 // We've just discovered that OrigCallee is known-emitted. Walk our call
738 // graph to see what else we can now discover also must be emitted.
741 FunctionDecl *Caller;
742 FunctionDecl *Callee;
745 llvm::SmallVector<CallInfo, 4> Worklist = {{OrigCaller, OrigCallee, OrigLoc}};
746 llvm::SmallSet<CanonicalDeclPtr<FunctionDecl>, 4> Seen;
747 Seen.insert(OrigCallee);
748 while (!Worklist.empty()) {
749 CallInfo C = Worklist.pop_back_val();
750 assert(!IsKnownEmitted(S, C.Callee) &&
751 "Worklist should not contain known-emitted functions.");
752 S.CUDAKnownEmittedFns[C.Callee] = {C.Caller, C.Loc};
753 EmitDeferredDiags(S, C.Callee);
755 // If this is a template instantiation, explore its callgraph as well:
756 // Non-dependent calls are part of the template's callgraph, while dependent
757 // calls are part of to the instantiation's call graph.
758 if (auto *Templ = C.Callee->getPrimaryTemplate()) {
759 FunctionDecl *TemplFD = Templ->getAsFunction();
760 if (!Seen.count(TemplFD) && !S.CUDAKnownEmittedFns.count(TemplFD)) {
761 Seen.insert(TemplFD);
763 {/* Caller = */ C.Caller, /* Callee = */ TemplFD, C.Loc});
767 // Add all functions called by Callee to our worklist.
768 auto CGIt = S.CUDACallGraph.find(C.Callee);
769 if (CGIt == S.CUDACallGraph.end())
772 for (std::pair<CanonicalDeclPtr<FunctionDecl>, SourceLocation> FDLoc :
774 FunctionDecl *NewCallee = FDLoc.first;
775 SourceLocation CallLoc = FDLoc.second;
776 if (Seen.count(NewCallee) || IsKnownEmitted(S, NewCallee))
778 Seen.insert(NewCallee);
780 {/* Caller = */ C.Callee, /* Callee = */ NewCallee, CallLoc});
783 // C.Callee is now known-emitted, so we no longer need to maintain its list
784 // of callees in CUDACallGraph.
785 S.CUDACallGraph.erase(CGIt);
789 bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) {
790 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
791 assert(Callee && "Callee may not be null.");
792 // FIXME: Is bailing out early correct here? Should we instead assume that
793 // the caller is a global initializer?
794 FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
798 // If the caller is known-emitted, mark the callee as known-emitted.
799 // Otherwise, mark the call in our call graph so we can traverse it later.
800 bool CallerKnownEmitted = IsKnownEmitted(*this, Caller);
801 if (CallerKnownEmitted)
802 MarkKnownEmitted(*this, Caller, Callee, Loc);
805 // host fn calls kernel fn calls host+device,
806 // the HD function does not get instantiated on the host. We model this by
807 // omitting at the call to the kernel from the callgraph. This ensures
808 // that, when compiling for host, only HD functions actually called from the
809 // host get marked as known-emitted.
810 if (getLangOpts().CUDAIsDevice || IdentifyCUDATarget(Callee) != CFT_Global)
811 CUDACallGraph[Caller].insert({Callee, Loc});
814 CUDADiagBuilder::Kind DiagKind = [&] {
815 switch (IdentifyCUDAPreference(Caller, Callee)) {
817 return CUDADiagBuilder::K_Immediate;
819 assert(Caller && "WrongSide calls require a non-null caller");
820 // If we know the caller will be emitted, we know this wrong-side call
821 // will be emitted, so it's an immediate error. Otherwise, defer the
822 // error until we know the caller is emitted.
823 return CallerKnownEmitted ? CUDADiagBuilder::K_ImmediateWithCallStack
824 : CUDADiagBuilder::K_Deferred;
826 return CUDADiagBuilder::K_Nop;
830 if (DiagKind == CUDADiagBuilder::K_Nop)
833 // Avoid emitting this error twice for the same location. Using a hashtable
834 // like this is unfortunate, but because we must continue parsing as normal
835 // after encountering a deferred error, it's otherwise very tricky for us to
836 // ensure that we only emit this deferred error once.
837 if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second)
840 CUDADiagBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this)
841 << IdentifyCUDATarget(Callee) << Callee << IdentifyCUDATarget(Caller);
842 CUDADiagBuilder(DiagKind, Callee->getLocation(), diag::note_previous_decl,
845 return DiagKind != CUDADiagBuilder::K_Immediate &&
846 DiagKind != CUDADiagBuilder::K_ImmediateWithCallStack;
849 void Sema::CUDASetLambdaAttrs(CXXMethodDecl *Method) {
850 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
851 if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>())
853 FunctionDecl *CurFn = dyn_cast<FunctionDecl>(CurContext);
856 CUDAFunctionTarget Target = IdentifyCUDATarget(CurFn);
857 if (Target == CFT_Global || Target == CFT_Device) {
858 Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
859 } else if (Target == CFT_HostDevice) {
860 Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
861 Method->addAttr(CUDAHostAttr::CreateImplicit(Context));
865 void Sema::checkCUDATargetOverload(FunctionDecl *NewFD,
866 const LookupResult &Previous) {
867 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
868 CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD);
869 for (NamedDecl *OldND : Previous) {
870 FunctionDecl *OldFD = OldND->getAsFunction();
874 CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD);
875 // Don't allow HD and global functions to overload other functions with the
876 // same signature. We allow overloading based on CUDA attributes so that
877 // functions can have different implementations on the host and device, but
878 // HD/global functions "exist" in some sense on both the host and device, so
879 // should have the same implementation on both sides.
880 if (NewTarget != OldTarget &&
881 ((NewTarget == CFT_HostDevice) || (OldTarget == CFT_HostDevice) ||
882 (NewTarget == CFT_Global) || (OldTarget == CFT_Global)) &&
883 !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false,
884 /* ConsiderCudaAttrs = */ false)) {
885 Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)
886 << NewTarget << NewFD->getDeclName() << OldTarget << OldFD;
887 Diag(OldFD->getLocation(), diag::note_previous_declaration);
888 NewFD->setInvalidDecl();
894 template <typename AttrTy>
895 static void copyAttrIfPresent(Sema &S, FunctionDecl *FD,
896 const FunctionDecl &TemplateFD) {
897 if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
898 AttrTy *Clone = Attribute->clone(S.Context);
899 Clone->setInherited(true);
904 void Sema::inheritCUDATargetAttrs(FunctionDecl *FD,
905 const FunctionTemplateDecl &TD) {
906 const FunctionDecl &TemplateFD = *TD.getTemplatedDecl();
907 copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD);
908 copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD);
909 copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD);