clang 17.0.0git
SemaCUDA.cpp
Go to the documentation of this file.
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
14#include "clang/AST/Decl.h"
15#include "clang/AST/ExprCXX.h"
16#include "clang/Basic/Cuda.h"
19#include "clang/Sema/Lookup.h"
21#include "clang/Sema/Sema.h"
24#include "clang/Sema/Template.h"
25#include "llvm/ADT/SmallVector.h"
26#include <optional>
27using namespace clang;
28
29template <typename AttrT> static bool hasExplicitAttr(const VarDecl *D) {
30 if (!D)
31 return false;
32 if (auto *A = D->getAttr<AttrT>())
33 return !A->isImplicit();
34 return false;
35}
36
38 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
39 ForceCUDAHostDeviceDepth++;
40}
41
43 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
44 if (ForceCUDAHostDeviceDepth == 0)
45 return false;
46 ForceCUDAHostDeviceDepth--;
47 return true;
48}
49
51 MultiExprArg ExecConfig,
52 SourceLocation GGGLoc) {
54 if (!ConfigDecl)
55 return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use)
57 QualType ConfigQTy = ConfigDecl->getType();
58
59 DeclRefExpr *ConfigDR = new (Context)
60 DeclRefExpr(Context, ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc);
61 MarkFunctionReferenced(LLLLoc, ConfigDecl);
62
63 return BuildCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr,
64 /*IsExecConfig=*/true);
65}
66
69 bool HasHostAttr = false;
70 bool HasDeviceAttr = false;
71 bool HasGlobalAttr = false;
72 bool HasInvalidTargetAttr = false;
73 for (const ParsedAttr &AL : Attrs) {
74 switch (AL.getKind()) {
75 case ParsedAttr::AT_CUDAGlobal:
76 HasGlobalAttr = true;
77 break;
78 case ParsedAttr::AT_CUDAHost:
79 HasHostAttr = true;
80 break;
81 case ParsedAttr::AT_CUDADevice:
82 HasDeviceAttr = true;
83 break;
84 case ParsedAttr::AT_CUDAInvalidTarget:
85 HasInvalidTargetAttr = true;
86 break;
87 default:
88 break;
89 }
90 }
91
92 if (HasInvalidTargetAttr)
93 return CFT_InvalidTarget;
94
95 if (HasGlobalAttr)
96 return CFT_Global;
97
98 if (HasHostAttr && HasDeviceAttr)
99 return CFT_HostDevice;
100
101 if (HasDeviceAttr)
102 return CFT_Device;
103
104 return CFT_Host;
105}
106
107template <typename A>
108static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
109 return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
110 return isa<A>(Attribute) &&
111 !(IgnoreImplicitAttr && Attribute->isImplicit());
112 });
113}
114
115/// IdentifyCUDATarget - Determine the CUDA compilation target for this function
117 bool IgnoreImplicitHDAttr) {
118 // Code that lives outside a function is run on the host.
119 if (D == nullptr)
120 return CFT_Host;
121
122 if (D->hasAttr<CUDAInvalidTargetAttr>())
123 return CFT_InvalidTarget;
124
125 if (D->hasAttr<CUDAGlobalAttr>())
126 return CFT_Global;
127
128 if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) {
129 if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr))
130 return CFT_HostDevice;
131 return CFT_Device;
132 } else if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) {
133 return CFT_Host;
134 } else if ((D->isImplicit() || !D->isUserProvided()) &&
135 !IgnoreImplicitHDAttr) {
136 // Some implicit declarations (like intrinsic functions) are not marked.
137 // Set the most lenient target on them for maximal flexibility.
138 return CFT_HostDevice;
139 }
140
141 return CFT_Host;
142}
143
144/// IdentifyTarget - Determine the CUDA compilation target for this variable.
146 if (Var->hasAttr<HIPManagedAttr>())
147 return CVT_Unified;
148 // Only constexpr and const variabless with implicit constant attribute
149 // are emitted on both sides. Such variables are promoted to device side
150 // only if they have static constant intializers on device side.
151 if ((Var->isConstexpr() || Var->getType().isConstQualified()) &&
152 Var->hasAttr<CUDAConstantAttr>() &&
153 !hasExplicitAttr<CUDAConstantAttr>(Var))
154 return CVT_Both;
155 if (Var->hasAttr<CUDADeviceAttr>() || Var->hasAttr<CUDAConstantAttr>() ||
156 Var->hasAttr<CUDASharedAttr>() ||
159 return CVT_Device;
160 // Function-scope static variable without explicit device or constant
161 // attribute are emitted
162 // - on both sides in host device functions
163 // - on device side in device or global functions
164 if (auto *FD = dyn_cast<FunctionDecl>(Var->getDeclContext())) {
165 switch (IdentifyCUDATarget(FD)) {
166 case CFT_HostDevice:
167 return CVT_Both;
168 case CFT_Device:
169 case CFT_Global:
170 return CVT_Device;
171 default:
172 return CVT_Host;
173 }
174 }
175 return CVT_Host;
176}
177
178// * CUDA Call preference table
179//
180// F - from,
181// T - to
182// Ph - preference in host mode
183// Pd - preference in device mode
184// H - handled in (x)
185// Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never.
186//
187// | F | T | Ph | Pd | H |
188// |----+----+-----+-----+-----+
189// | d | d | N | N | (c) |
190// | d | g | -- | -- | (a) |
191// | d | h | -- | -- | (e) |
192// | d | hd | HD | HD | (b) |
193// | g | d | N | N | (c) |
194// | g | g | -- | -- | (a) |
195// | g | h | -- | -- | (e) |
196// | g | hd | HD | HD | (b) |
197// | h | d | -- | -- | (e) |
198// | h | g | N | N | (c) |
199// | h | h | N | N | (c) |
200// | h | hd | HD | HD | (b) |
201// | hd | d | WS | SS | (d) |
202// | hd | g | SS | -- |(d/a)|
203// | hd | h | SS | WS | (d) |
204// | hd | hd | HD | HD | (b) |
205
208 const FunctionDecl *Callee) {
209 assert(Callee && "Callee must be valid.");
210 CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller);
211 CUDAFunctionTarget CalleeTarget = IdentifyCUDATarget(Callee);
212
213 // If one of the targets is invalid, the check always fails, no matter what
214 // the other target is.
215 if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget)
216 return CFP_Never;
217
218 // (a) Can't call global from some contexts until we support CUDA's
219 // dynamic parallelism.
220 if (CalleeTarget == CFT_Global &&
221 (CallerTarget == CFT_Global || CallerTarget == CFT_Device))
222 return CFP_Never;
223
224 // (b) Calling HostDevice is OK for everyone.
225 if (CalleeTarget == CFT_HostDevice)
226 return CFP_HostDevice;
227
228 // (c) Best case scenarios
229 if (CalleeTarget == CallerTarget ||
230 (CallerTarget == CFT_Host && CalleeTarget == CFT_Global) ||
231 (CallerTarget == CFT_Global && CalleeTarget == CFT_Device))
232 return CFP_Native;
233
234 // (d) HostDevice behavior depends on compilation mode.
235 if (CallerTarget == CFT_HostDevice) {
236 // It's OK to call a compilation-mode matching function from an HD one.
237 if ((getLangOpts().CUDAIsDevice && CalleeTarget == CFT_Device) ||
238 (!getLangOpts().CUDAIsDevice &&
239 (CalleeTarget == CFT_Host || CalleeTarget == CFT_Global)))
240 return CFP_SameSide;
241
242 // Calls from HD to non-mode-matching functions (i.e., to host functions
243 // when compiling in device mode or to device functions when compiling in
244 // host mode) are allowed at the sema level, but eventually rejected if
245 // they're ever codegened. TODO: Reject said calls earlier.
246 return CFP_WrongSide;
247 }
248
249 // (e) Calling across device/host boundary is not something you should do.
250 if ((CallerTarget == CFT_Host && CalleeTarget == CFT_Device) ||
251 (CallerTarget == CFT_Device && CalleeTarget == CFT_Host) ||
252 (CallerTarget == CFT_Global && CalleeTarget == CFT_Host))
253 return CFP_Never;
254
255 llvm_unreachable("All cases should've been handled by now.");
256}
257
258template <typename AttrT> static bool hasImplicitAttr(const FunctionDecl *D) {
259 if (!D)
260 return false;
261 if (auto *A = D->getAttr<AttrT>())
262 return A->isImplicit();
263 return D->isImplicit();
264}
265
267 bool IsImplicitDevAttr = hasImplicitAttr<CUDADeviceAttr>(D);
268 bool IsImplicitHostAttr = hasImplicitAttr<CUDAHostAttr>(D);
269 return IsImplicitDevAttr && IsImplicitHostAttr;
270}
271
273 const FunctionDecl *Caller,
274 SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) {
275 if (Matches.size() <= 1)
276 return;
277
278 using Pair = std::pair<DeclAccessPair, FunctionDecl*>;
279
280 // Gets the CUDA function preference for a call from Caller to Match.
281 auto GetCFP = [&](const Pair &Match) {
282 return IdentifyCUDAPreference(Caller, Match.second);
283 };
284
285 // Find the best call preference among the functions in Matches.
286 CUDAFunctionPreference BestCFP = GetCFP(*std::max_element(
287 Matches.begin(), Matches.end(),
288 [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); }));
289
290 // Erase all functions with lower priority.
291 llvm::erase_if(Matches,
292 [&](const Pair &Match) { return GetCFP(Match) < BestCFP; });
293}
294
295/// When an implicitly-declared special member has to invoke more than one
296/// base/field special member, conflicts may occur in the targets of these
297/// members. For example, if one base's member __host__ and another's is
298/// __device__, it's a conflict.
299/// This function figures out if the given targets \param Target1 and
300/// \param Target2 conflict, and if they do not it fills in
301/// \param ResolvedTarget with a target that resolves for both calls.
302/// \return true if there's a conflict, false otherwise.
303static bool
306 Sema::CUDAFunctionTarget *ResolvedTarget) {
307 // Only free functions and static member functions may be global.
308 assert(Target1 != Sema::CFT_Global);
309 assert(Target2 != Sema::CFT_Global);
310
311 if (Target1 == Sema::CFT_HostDevice) {
312 *ResolvedTarget = Target2;
313 } else if (Target2 == Sema::CFT_HostDevice) {
314 *ResolvedTarget = Target1;
315 } else if (Target1 != Target2) {
316 return true;
317 } else {
318 *ResolvedTarget = Target1;
319 }
320
321 return false;
322}
323
326 CXXMethodDecl *MemberDecl,
327 bool ConstRHS,
328 bool Diagnose) {
329 // If the defaulted special member is defined lexically outside of its
330 // owning class, or the special member already has explicit device or host
331 // attributes, do not infer.
332 bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent();
333 bool HasH = MemberDecl->hasAttr<CUDAHostAttr>();
334 bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>();
335 bool HasExplicitAttr =
336 (HasD && !MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()) ||
337 (HasH && !MemberDecl->getAttr<CUDAHostAttr>()->isImplicit());
338 if (!InClass || HasExplicitAttr)
339 return false;
340
341 std::optional<CUDAFunctionTarget> InferredTarget;
342
343 // We're going to invoke special member lookup; mark that these special
344 // members are called from this one, and not from its caller.
345 ContextRAII MethodContext(*this, MemberDecl);
346
347 // Look for special members in base classes that should be invoked from here.
348 // Infer the target of this member base on the ones it should call.
349 // Skip direct and indirect virtual bases for abstract classes.
351 for (const auto &B : ClassDecl->bases()) {
352 if (!B.isVirtual()) {
353 Bases.push_back(&B);
354 }
355 }
356
357 if (!ClassDecl->isAbstract()) {
358 llvm::append_range(Bases, llvm::make_pointer_range(ClassDecl->vbases()));
359 }
360
361 for (const auto *B : Bases) {
362 const RecordType *BaseType = B->getType()->getAs<RecordType>();
363 if (!BaseType) {
364 continue;
365 }
366
367 CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());
369 LookupSpecialMember(BaseClassDecl, CSM,
370 /* ConstArg */ ConstRHS,
371 /* VolatileArg */ false,
372 /* RValueThis */ false,
373 /* ConstThis */ false,
374 /* VolatileThis */ false);
375
376 if (!SMOR.getMethod())
377 continue;
378
379 CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR.getMethod());
380 if (!InferredTarget) {
381 InferredTarget = BaseMethodTarget;
382 } else {
383 bool ResolutionError = resolveCalleeCUDATargetConflict(
384 *InferredTarget, BaseMethodTarget, &*InferredTarget);
385 if (ResolutionError) {
386 if (Diagnose) {
387 Diag(ClassDecl->getLocation(),
388 diag::note_implicit_member_target_infer_collision)
389 << (unsigned)CSM << *InferredTarget << BaseMethodTarget;
390 }
391 MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
392 return true;
393 }
394 }
395 }
396
397 // Same as for bases, but now for special members of fields.
398 for (const auto *F : ClassDecl->fields()) {
399 if (F->isInvalidDecl()) {
400 continue;
401 }
402
403 const RecordType *FieldType =
404 Context.getBaseElementType(F->getType())->getAs<RecordType>();
405 if (!FieldType) {
406 continue;
407 }
408
409 CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());
411 LookupSpecialMember(FieldRecDecl, CSM,
412 /* ConstArg */ ConstRHS && !F->isMutable(),
413 /* VolatileArg */ false,
414 /* RValueThis */ false,
415 /* ConstThis */ false,
416 /* VolatileThis */ false);
417
418 if (!SMOR.getMethod())
419 continue;
420
421 CUDAFunctionTarget FieldMethodTarget =
423 if (!InferredTarget) {
424 InferredTarget = FieldMethodTarget;
425 } else {
426 bool ResolutionError = resolveCalleeCUDATargetConflict(
427 *InferredTarget, FieldMethodTarget, &*InferredTarget);
428 if (ResolutionError) {
429 if (Diagnose) {
430 Diag(ClassDecl->getLocation(),
431 diag::note_implicit_member_target_infer_collision)
432 << (unsigned)CSM << *InferredTarget << FieldMethodTarget;
433 }
434 MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
435 return true;
436 }
437 }
438 }
439
440
441 // If no target was inferred, mark this member as __host__ __device__;
442 // it's the least restrictive option that can be invoked from any target.
443 bool NeedsH = true, NeedsD = true;
444 if (InferredTarget) {
445 if (*InferredTarget == CFT_Device)
446 NeedsH = false;
447 else if (*InferredTarget == CFT_Host)
448 NeedsD = false;
449 }
450
451 // We either setting attributes first time, or the inferred ones must match
452 // previously set ones.
453 if (NeedsD && !HasD)
454 MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
455 if (NeedsH && !HasH)
456 MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
457
458 return false;
459}
460
462 if (!CD->isDefined() && CD->isTemplateInstantiation())
464
465 // (E.2.3.1, CUDA 7.5) A constructor for a class type is considered
466 // empty at a point in the translation unit, if it is either a
467 // trivial constructor
468 if (CD->isTrivial())
469 return true;
470
471 // ... or it satisfies all of the following conditions:
472 // The constructor function has been defined.
473 // The constructor function has no parameters,
474 // and the function body is an empty compound statement.
475 if (!(CD->hasTrivialBody() && CD->getNumParams() == 0))
476 return false;
477
478 // Its class has no virtual functions and no virtual base classes.
479 if (CD->getParent()->isDynamicClass())
480 return false;
481
482 // Union ctor does not call ctors of its data members.
483 if (CD->getParent()->isUnion())
484 return true;
485
486 // The only form of initializer allowed is an empty constructor.
487 // This will recursively check all base classes and member initializers
488 if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {
489 if (const CXXConstructExpr *CE =
490 dyn_cast<CXXConstructExpr>(CI->getInit()))
491 return isEmptyCudaConstructor(Loc, CE->getConstructor());
492 return false;
493 }))
494 return false;
495
496 return true;
497}
498
500 // No destructor -> no problem.
501 if (!DD)
502 return true;
503
504 if (!DD->isDefined() && DD->isTemplateInstantiation())
506
507 // (E.2.3.1, CUDA 7.5) A destructor for a class type is considered
508 // empty at a point in the translation unit, if it is either a
509 // trivial constructor
510 if (DD->isTrivial())
511 return true;
512
513 // ... or it satisfies all of the following conditions:
514 // The destructor function has been defined.
515 // and the function body is an empty compound statement.
516 if (!DD->hasTrivialBody())
517 return false;
518
519 const CXXRecordDecl *ClassDecl = DD->getParent();
520
521 // Its class has no virtual functions and no virtual base classes.
522 if (ClassDecl->isDynamicClass())
523 return false;
524
525 // Union does not have base class and union dtor does not call dtors of its
526 // data members.
527 if (DD->getParent()->isUnion())
528 return true;
529
530 // Only empty destructors are allowed. This will recursively check
531 // destructors for all base classes...
532 if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {
533 if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())
534 return isEmptyCudaDestructor(Loc, RD->getDestructor());
535 return true;
536 }))
537 return false;
538
539 // ... and member fields.
540 if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {
541 if (CXXRecordDecl *RD = Field->getType()
542 ->getBaseElementTypeUnsafe()
543 ->getAsCXXRecordDecl())
544 return isEmptyCudaDestructor(Loc, RD->getDestructor());
545 return true;
546 }))
547 return false;
548
549 return true;
550}
551
552namespace {
553enum CUDAInitializerCheckKind {
554 CICK_DeviceOrConstant, // Check initializer for device/constant variable
555 CICK_Shared, // Check initializer for shared variable
556};
557
558bool IsDependentVar(VarDecl *VD) {
559 if (VD->getType()->isDependentType())
560 return true;
561 if (const auto *Init = VD->getInit())
562 return Init->isValueDependent();
563 return false;
564}
565
566// Check whether a variable has an allowed initializer for a CUDA device side
567// variable with global storage. \p VD may be a host variable to be checked for
568// potential promotion to device side variable.
569//
570// CUDA/HIP allows only empty constructors as initializers for global
571// variables (see E.2.3.1, CUDA 7.5). The same restriction also applies to all
572// __shared__ variables whether they are local or not (they all are implicitly
573// static in CUDA). One exception is that CUDA allows constant initializers
574// for __constant__ and __device__ variables.
575bool HasAllowedCUDADeviceStaticInitializer(Sema &S, VarDecl *VD,
576 CUDAInitializerCheckKind CheckKind) {
577 assert(!VD->isInvalidDecl() && VD->hasGlobalStorage());
578 assert(!IsDependentVar(VD) && "do not check dependent var");
579 const Expr *Init = VD->getInit();
580 auto IsEmptyInit = [&](const Expr *Init) {
581 if (!Init)
582 return true;
583 if (const auto *CE = dyn_cast<CXXConstructExpr>(Init)) {
584 return S.isEmptyCudaConstructor(VD->getLocation(), CE->getConstructor());
585 }
586 return false;
587 };
588 auto IsConstantInit = [&](const Expr *Init) {
589 assert(Init);
591 /*NoWronSidedVars=*/true);
592 return Init->isConstantInitializer(S.Context,
593 VD->getType()->isReferenceType());
594 };
595 auto HasEmptyDtor = [&](VarDecl *VD) {
596 if (const auto *RD = VD->getType()->getAsCXXRecordDecl())
597 return S.isEmptyCudaDestructor(VD->getLocation(), RD->getDestructor());
598 return true;
599 };
600 if (CheckKind == CICK_Shared)
601 return IsEmptyInit(Init) && HasEmptyDtor(VD);
602 return S.LangOpts.GPUAllowDeviceInit ||
603 ((IsEmptyInit(Init) || IsConstantInit(Init)) && HasEmptyDtor(VD));
604}
605} // namespace
606
608 // Do not check dependent variables since the ctor/dtor/initializer are not
609 // determined. Do it after instantiation.
610 if (VD->isInvalidDecl() || !VD->hasInit() || !VD->hasGlobalStorage() ||
611 IsDependentVar(VD))
612 return;
613 const Expr *Init = VD->getInit();
614 bool IsSharedVar = VD->hasAttr<CUDASharedAttr>();
615 bool IsDeviceOrConstantVar =
616 !IsSharedVar &&
617 (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>());
618 if (IsDeviceOrConstantVar || IsSharedVar) {
619 if (HasAllowedCUDADeviceStaticInitializer(
620 *this, VD, IsSharedVar ? CICK_Shared : CICK_DeviceOrConstant))
621 return;
622 Diag(VD->getLocation(),
623 IsSharedVar ? diag::err_shared_var_init : diag::err_dynamic_var_init)
624 << Init->getSourceRange();
625 VD->setInvalidDecl();
626 } else {
627 // This is a host-side global variable. Check that the initializer is
628 // callable from the host side.
629 const FunctionDecl *InitFn = nullptr;
630 if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init)) {
631 InitFn = CE->getConstructor();
632 } else if (const CallExpr *CE = dyn_cast<CallExpr>(Init)) {
633 InitFn = CE->getDirectCallee();
634 }
635 if (InitFn) {
636 CUDAFunctionTarget InitFnTarget = IdentifyCUDATarget(InitFn);
637 if (InitFnTarget != CFT_Host && InitFnTarget != CFT_HostDevice) {
638 Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer)
639 << InitFnTarget << InitFn;
640 Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn;
641 VD->setInvalidDecl();
642 }
643 }
644 }
645}
646
647// With -fcuda-host-device-constexpr, an unattributed constexpr function is
648// treated as implicitly __host__ __device__, unless:
649// * it is a variadic function (device-side variadic functions are not
650// allowed), or
651// * a __device__ function with this signature was already declared, in which
652// case in which case we output an error, unless the __device__ decl is in a
653// system header, in which case we leave the constexpr function unattributed.
654//
655// In addition, all function decls are treated as __host__ __device__ when
656// ForceCUDAHostDeviceDepth > 0 (corresponding to code within a
657// #pragma clang force_cuda_host_device_begin/end
658// pair).
660 const LookupResult &Previous) {
661 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
662
663 if (ForceCUDAHostDeviceDepth > 0) {
664 if (!NewD->hasAttr<CUDAHostAttr>())
665 NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
666 if (!NewD->hasAttr<CUDADeviceAttr>())
667 NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
668 return;
669 }
670
671 if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() ||
672 NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() ||
673 NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>())
674 return;
675
676 // Is D a __device__ function with the same signature as NewD, ignoring CUDA
677 // attributes?
678 auto IsMatchingDeviceFn = [&](NamedDecl *D) {
679 if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D))
680 D = Using->getTargetDecl();
681 FunctionDecl *OldD = D->getAsFunction();
682 return OldD && OldD->hasAttr<CUDADeviceAttr>() &&
683 !OldD->hasAttr<CUDAHostAttr>() &&
684 !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false,
685 /* ConsiderCudaAttrs = */ false);
686 };
687 auto It = llvm::find_if(Previous, IsMatchingDeviceFn);
688 if (It != Previous.end()) {
689 // We found a __device__ function with the same name and signature as NewD
690 // (ignoring CUDA attrs). This is an error unless that function is defined
691 // in a system header, in which case we simply return without making NewD
692 // host+device.
693 NamedDecl *Match = *It;
694 if (!getSourceManager().isInSystemHeader(Match->getLocation())) {
695 Diag(NewD->getLocation(),
696 diag::err_cuda_unattributed_constexpr_cannot_overload_device)
697 << NewD;
698 Diag(Match->getLocation(),
699 diag::note_cuda_conflicting_device_function_declared_here);
700 }
701 return;
702 }
703
704 NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
705 NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
706}
707
708// TODO: `__constant__` memory may be a limited resource for certain targets.
709// A safeguard may be needed at the end of compilation pipeline if
710// `__constant__` memory usage goes beyond limit.
712 // Do not promote dependent variables since the cotr/dtor/initializer are
713 // not determined. Do it after instantiation.
714 if (getLangOpts().CUDAIsDevice && !VD->hasAttr<CUDAConstantAttr>() &&
715 !VD->hasAttr<CUDASharedAttr>() &&
716 (VD->isFileVarDecl() || VD->isStaticDataMember()) &&
717 !IsDependentVar(VD) &&
718 ((VD->isConstexpr() || VD->getType().isConstQualified()) &&
719 HasAllowedCUDADeviceStaticInitializer(*this, VD,
720 CICK_DeviceOrConstant))) {
721 VD->addAttr(CUDAConstantAttr::CreateImplicit(getASTContext()));
722 }
723}
724
726 unsigned DiagID) {
727 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
728 FunctionDecl *CurFunContext = getCurFunctionDecl(/*AllowLambda=*/true);
729 SemaDiagnosticBuilder::Kind DiagKind = [&] {
730 if (!CurFunContext)
732 switch (CurrentCUDATarget()) {
733 case CFT_Global:
734 case CFT_Device:
736 case CFT_HostDevice:
737 // An HD function counts as host code if we're compiling for host, and
738 // device code if we're compiling for device. Defer any errors in device
739 // mode until the function is known-emitted.
740 if (!getLangOpts().CUDAIsDevice)
742 if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID))
744 return (getEmissionStatus(CurFunContext) ==
748 default:
750 }
751 }();
752 return SemaDiagnosticBuilder(DiagKind, Loc, DiagID, CurFunContext, *this);
753}
754
756 unsigned DiagID) {
757 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
758 FunctionDecl *CurFunContext = getCurFunctionDecl(/*AllowLambda=*/true);
759 SemaDiagnosticBuilder::Kind DiagKind = [&] {
760 if (!CurFunContext)
762 switch (CurrentCUDATarget()) {
763 case CFT_Host:
765 case CFT_HostDevice:
766 // An HD function counts as host code if we're compiling for host, and
767 // device code if we're compiling for device. Defer any errors in device
768 // mode until the function is known-emitted.
769 if (getLangOpts().CUDAIsDevice)
771 if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID))
773 return (getEmissionStatus(CurFunContext) ==
777 default:
779 }
780 }();
781 return SemaDiagnosticBuilder(DiagKind, Loc, DiagID, CurFunContext, *this);
782}
783
785 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
786 assert(Callee && "Callee may not be null.");
787
788 auto &ExprEvalCtx = ExprEvalContexts.back();
789 if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated())
790 return true;
791
792 // FIXME: Is bailing out early correct here? Should we instead assume that
793 // the caller is a global initializer?
794 FunctionDecl *Caller = getCurFunctionDecl(/*AllowLambda=*/true);
795 if (!Caller)
796 return true;
797
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 =
802 SemaDiagnosticBuilder::Kind DiagKind = [this, Caller, Callee,
803 CallerKnownEmitted] {
804 switch (IdentifyCUDAPreference(Caller, Callee)) {
805 case CFP_Never:
806 case CFP_WrongSide:
807 assert(Caller && "Never/wrongSide calls require a non-null caller");
808 // If we know the caller will be emitted, we know this wrong-side call
809 // will be emitted, so it's an immediate error. Otherwise, defer the
810 // error until we know the caller is emitted.
811 return CallerKnownEmitted
814 default:
816 }
817 }();
818
819 if (DiagKind == SemaDiagnosticBuilder::K_Nop) {
820 // For -fgpu-rdc, keep track of external kernels used by host functions.
821 if (LangOpts.CUDAIsDevice && LangOpts.GPURelocatableDeviceCode &&
822 Callee->hasAttr<CUDAGlobalAttr>() && !Callee->isDefined())
824 return true;
825 }
826
827 // Avoid emitting this error twice for the same location. Using a hashtable
828 // like this is unfortunate, but because we must continue parsing as normal
829 // after encountering a deferred error, it's otherwise very tricky for us to
830 // ensure that we only emit this deferred error once.
831 if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second)
832 return true;
833
834 SemaDiagnosticBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this)
835 << IdentifyCUDATarget(Callee) << /*function*/ 0 << Callee
836 << IdentifyCUDATarget(Caller);
837 if (!Callee->getBuiltinID())
838 SemaDiagnosticBuilder(DiagKind, Callee->getLocation(),
839 diag::note_previous_decl, Caller, *this)
840 << Callee;
841 return DiagKind != SemaDiagnosticBuilder::K_Immediate &&
843}
844
845// Check the wrong-sided reference capture of lambda for CUDA/HIP.
846// A lambda function may capture a stack variable by reference when it is
847// defined and uses the capture by reference when the lambda is called. When
848// the capture and use happen on different sides, the capture is invalid and
849// should be diagnosed.
851 const sema::Capture &Capture) {
852 // In host compilation we only need to check lambda functions emitted on host
853 // side. In such lambda functions, a reference capture is invalid only
854 // if the lambda structure is populated by a device function or kernel then
855 // is passed to and called by a host function. However that is impossible,
856 // since a device function or kernel can only call a device function, also a
857 // kernel cannot pass a lambda back to a host function since we cannot
858 // define a kernel argument type which can hold the lambda before the lambda
859 // itself is defined.
860 if (!LangOpts.CUDAIsDevice)
861 return;
862
863 // File-scope lambda can only do init captures for global variables, which
864 // results in passing by value for these global variables.
865 FunctionDecl *Caller = getCurFunctionDecl(/*AllowLambda=*/true);
866 if (!Caller)
867 return;
868
869 // In device compilation, we only need to check lambda functions which are
870 // emitted on device side. For such lambdas, a reference capture is invalid
871 // only if the lambda structure is populated by a host function then passed
872 // to and called in a device function or kernel.
873 bool CalleeIsDevice = Callee->hasAttr<CUDADeviceAttr>();
874 bool CallerIsHost =
875 !Caller->hasAttr<CUDAGlobalAttr>() && !Caller->hasAttr<CUDADeviceAttr>();
876 bool ShouldCheck = CalleeIsDevice && CallerIsHost;
877 if (!ShouldCheck || !Capture.isReferenceCapture())
878 return;
879 auto DiagKind = SemaDiagnosticBuilder::K_Deferred;
882 diag::err_capture_bad_target, Callee, *this)
883 << Capture.getVariable();
884 } else if (Capture.isThisCapture()) {
885 // Capture of this pointer is allowed since this pointer may be pointing to
886 // managed memory which is accessible on both device and host sides. It only
887 // results in invalid memory access if this pointer points to memory not
888 // accessible on device side.
890 diag::warn_maybe_capture_bad_target_this_ptr, Callee,
891 *this);
892 }
893}
894
896 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
897 if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>())
898 return;
899 Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
900 Method->addAttr(CUDAHostAttr::CreateImplicit(Context));
901}
902
904 const LookupResult &Previous) {
905 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
906 CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD);
907 for (NamedDecl *OldND : Previous) {
908 FunctionDecl *OldFD = OldND->getAsFunction();
909 if (!OldFD)
910 continue;
911
912 CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD);
913 // Don't allow HD and global functions to overload other functions with the
914 // same signature. We allow overloading based on CUDA attributes so that
915 // functions can have different implementations on the host and device, but
916 // HD/global functions "exist" in some sense on both the host and device, so
917 // should have the same implementation on both sides.
918 if (NewTarget != OldTarget &&
919 ((NewTarget == CFT_HostDevice) || (OldTarget == CFT_HostDevice) ||
920 (NewTarget == CFT_Global) || (OldTarget == CFT_Global)) &&
921 !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false,
922 /* ConsiderCudaAttrs = */ false)) {
923 Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)
924 << NewTarget << NewFD->getDeclName() << OldTarget << OldFD;
925 Diag(OldFD->getLocation(), diag::note_previous_declaration);
926 NewFD->setInvalidDecl();
927 break;
928 }
929 }
930}
931
932template <typename AttrTy>
934 const FunctionDecl &TemplateFD) {
935 if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
936 AttrTy *Clone = Attribute->clone(S.Context);
937 Clone->setInherited(true);
938 FD->addAttr(Clone);
939 }
940}
941
943 const FunctionTemplateDecl &TD) {
944 const FunctionDecl &TemplateFD = *TD.getTemplatedDecl();
945 copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD);
946 copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD);
947 copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD);
948}
949
951 if (getLangOpts().HIP)
952 return getLangOpts().HIPUseNewLaunchAPI ? "__hipPushCallConfiguration"
953 : "hipConfigureCall";
954
955 // New CUDA kernel launch sequence.
958 return "__cudaPushCallConfiguration";
959
960 // Legacy CUDA kernel configuration call
961 return "cudaConfigureCall";
962}
Defines the clang::ASTContext interface.
Defines the clang::Expr interface and subclasses for C++ expressions.
Defines the clang::Preprocessor interface.
static void copyAttrIfPresent(Sema &S, FunctionDecl *FD, const FunctionDecl &TemplateFD)
Definition: SemaCUDA.cpp:933
static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr)
Definition: SemaCUDA.cpp:108
static bool resolveCalleeCUDATargetConflict(Sema::CUDAFunctionTarget Target1, Sema::CUDAFunctionTarget Target2, Sema::CUDAFunctionTarget *ResolvedTarget)
When an implicitly-declared special member has to invoke more than one base/field special member,...
Definition: SemaCUDA.cpp:304
static bool hasImplicitAttr(const FunctionDecl *D)
Definition: SemaCUDA.cpp:258
static bool hasExplicitAttr(const VarDecl *D)
Definition: SemaCUDA.cpp:29
StateNode * Previous
QualType getBaseElementType(const ArrayType *VAT) const
Return the innermost element type of an array type.
llvm::DenseSet< const ValueDecl * > CUDAExternalDeviceDeclODRUsedByHost
Keep track of CUDA/HIP external kernels or device variables ODR-used by host code.
Definition: ASTContext.h:1149
FunctionDecl * getcudaConfigureCallDecl()
Definition: ASTContext.h:1393
const TargetInfo & getTargetInfo() const
Definition: ASTContext.h:744
Attr - This represents one attribute.
Definition: Attr.h:40
Represents a base class of a C++ class.
Definition: DeclCXX.h:146
Represents a call to a C++ constructor.
Definition: ExprCXX.h:1518
Represents a C++ constructor within a class.
Definition: DeclCXX.h:2474
Represents a C++ base or member initializer.
Definition: DeclCXX.h:2242
Represents a C++ destructor within a class.
Definition: DeclCXX.h:2738
Represents a static or instance method of a struct/union/class.
Definition: DeclCXX.h:2018
const CXXRecordDecl * getParent() const
Return the parent of this method declaration, which is the class in which this method is defined.
Definition: DeclCXX.h:2133
Represents a C++ struct/union/class.
Definition: DeclCXX.h:254
base_class_range bases()
Definition: DeclCXX.h:602
base_class_range vbases()
Definition: DeclCXX.h:619
bool isAbstract() const
Determine whether this class has a pure virtual function.
Definition: DeclCXX.h:1199
bool isDynamicClass() const
Definition: DeclCXX.h:568
CallExpr - Represents a function call (C99 6.5.2.2, C++ [expr.call]).
Definition: Expr.h:2812
DeclContext * getLexicalParent()
getLexicalParent - Returns the containing lexical DeclContext.
Definition: DeclBase.h:1949
A reference to a declared variable, function, enum, etc.
Definition: Expr.h:1238
T * getAttr() const
Definition: DeclBase.h:556
bool hasAttrs() const
Definition: DeclBase.h:502
void addAttr(Attr *A)
Definition: DeclBase.cpp:903
bool isImplicit() const
isImplicit - Indicates whether the declaration was implicitly generated by the implementation.
Definition: DeclBase.h:576
void setInvalidDecl(bool Invalid=true)
setInvalidDecl - Indicates the Decl had a semantic error.
Definition: DeclBase.cpp:132
FunctionDecl * getAsFunction() LLVM_READONLY
Returns the function itself, or the templated function if this is a function template.
Definition: DeclBase.cpp:227
bool isInvalidDecl() const
Definition: DeclBase.h:571
SourceLocation getLocation() const
Definition: DeclBase.h:432
DeclContext * getDeclContext()
Definition: DeclBase.h:441
AttrVec & getAttrs()
Definition: DeclBase.h:508
bool hasAttr() const
Definition: DeclBase.h:560
const IntrusiveRefCntPtr< DiagnosticIDs > & getDiagnosticIDs() const
Definition: Diagnostic.h:552
This represents one expression.
Definition: Expr.h:110
Represents a member of a struct/union/class.
Definition: Decl.h:2941
Represents a function declaration or definition.
Definition: Decl.h:1917
bool hasTrivialBody() const
Returns whether the function has a trivial body that does not require any specific codegen.
Definition: Decl.cpp:3065
bool isTrivial() const
Whether this function is "trivial" in some specialized C++ senses.
Definition: Decl.h:2272
bool isVariadic() const
Whether this function is variadic.
Definition: Decl.cpp:3022
bool isTemplateInstantiation() const
Determines if the given function was instantiated from a function template.
Definition: Decl.cpp:3883
bool isConstexpr() const
Whether this is a (C++11) constexpr function or constexpr constructor.
Definition: Decl.h:2365
bool isUserProvided() const
True if this method is user-declared and was not deleted or defaulted on its first declaration.
Definition: Decl.h:2305
unsigned getNumParams() const
Return the number of parameters this function must have based on its FunctionType.
Definition: Decl.cpp:3489
bool isDefined(const FunctionDecl *&Definition, bool CheckForPendingFriendDefinition=false) const
Returns true if the function has a definition that does not need to be instantiated.
Definition: Decl.cpp:3101
Declaration of a template function.
FunctionDecl * getTemplatedDecl() const
Get the underlying function declaration of the template.
Represents the results of name lookup.
Definition: Lookup.h:46
This represents a decl that may have a name.
Definition: Decl.h:247
DeclarationName getDeclName() const
Get the actual, stored name of the declaration, which may be a special name.
Definition: Decl.h:313
ParsedAttr - Represents a syntactic attribute.
Definition: ParsedAttr.h:123
A (possibly-)qualified type.
Definition: Type.h:736
bool isConstQualified() const
Determine whether this type is const-qualified.
Definition: Type.h:6721
field_range fields() const
Definition: Decl.h:4225
A helper class that allows the use of isa/cast/dyncast to detect TagType objects of structs/unions/cl...
Definition: Type.h:4835
RecordDecl * getDecl() const
Definition: Type.h:4845
decl_type * getFirstDecl()
Return the first declaration of this declaration or itself if this is the only declaration.
Definition: Redeclarable.h:216
Scope - A scope is a transient data structure that is used while parsing the program.
Definition: Scope.h:41
A RAII object to temporarily push a declaration context.
Definition: Sema.h:998
A generic diagnostic builder for errors which may or may not be deferred.
Definition: Sema.h:1760
@ K_Nop
Emit no diagnostics.
Definition: Sema.h:1764
@ K_Deferred
Create a deferred diagnostic, which is emitted only if the function it's attached to is codegen'ed.
Definition: Sema.h:1774
@ K_ImmediateWithCallStack
Emit the diagnostic immediately, and, if it's a warning or error, also emit a call stack showing how ...
Definition: Sema.h:1770
@ K_Immediate
Emit the diagnostic immediately (i.e., behave like Sema::Diag()).
Definition: Sema.h:1766
SpecialMemberOverloadResult - The overloading result for a special member function.
Definition: Sema.h:1412
CXXMethodDecl * getMethod() const
Definition: Sema.h:1428
Sema - This implements semantic analysis and AST building for C.
Definition: Sema.h:356
CUDAFunctionTarget IdentifyCUDATarget(const FunctionDecl *D, bool IgnoreImplicitHDAttr=false)
Determines whether the given function is a CUDA device/host/kernel/etc.
Definition: SemaCUDA.cpp:116
SemaDiagnosticBuilder Diag(SourceLocation Loc, unsigned DiagID, bool DeferHint=false)
Emit a diagnostic.
Definition: Sema.cpp:1884
bool IsOverload(FunctionDecl *New, FunctionDecl *Old, bool UseMemberUsingDeclRules, bool ConsiderCudaAttrs=true, bool ConsiderRequiresClauses=true)
CUDAFunctionPreference
Definition: Sema.h:13118
@ CFP_Never
Definition: Sema.h:13119
@ CFP_HostDevice
Definition: Sema.h:13123
@ CFP_SameSide
Definition: Sema.h:13124
@ CFP_Native
Definition: Sema.h:13126
@ CFP_WrongSide
Definition: Sema.h:13120
bool IsLastErrorImmediate
Is the last error level diagnostic immediate.
Definition: Sema.h:1864
void CUDACheckLambdaCapture(CXXMethodDecl *D, const sema::Capture &Capture)
Definition: SemaCUDA.cpp:850
void EraseUnwantedCUDAMatches(const FunctionDecl *Caller, SmallVectorImpl< std::pair< DeclAccessPair, FunctionDecl * > > &Matches)
Finds a function in Matches with highest calling priority from Caller context and erases all function...
Definition: SemaCUDA.cpp:272
ExprResult ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc, MultiExprArg ExecConfig, SourceLocation GGGLoc)
Definition: SemaCUDA.cpp:50
FunctionDecl * getCurFunctionDecl(bool AllowLambda=false) const
Returns a pointer to the innermost enclosing function, or nullptr if the current context is not insid...
Definition: Sema.cpp:1464
ASTContext & Context
Definition: Sema.h:407
ASTContext & getASTContext() const
Definition: Sema.h:1652
void checkAllowedCUDAInitializer(VarDecl *VD)
Definition: SemaCUDA.cpp:607
const LangOptions & getLangOpts() const
Definition: Sema.h:1645
SpecialMemberOverloadResult LookupSpecialMember(CXXRecordDecl *D, CXXSpecialMember SM, bool ConstArg, bool VolatileArg, bool RValueThis, bool ConstThis, bool VolatileThis)
SemaDiagnosticBuilder CUDADiagIfDeviceCode(SourceLocation Loc, unsigned DiagID)
Creates a SemaDiagnosticBuilder that emits the diagnostic if the current context is "used as device c...
Definition: SemaCUDA.cpp:725
bool CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee)
Check whether we're allowed to call Callee from the current context.
Definition: SemaCUDA.cpp:784
ExprResult BuildCallExpr(Scope *S, Expr *Fn, SourceLocation LParenLoc, MultiExprArg ArgExprs, SourceLocation RParenLoc, Expr *ExecConfig=nullptr, bool IsExecConfig=false, bool AllowRecovery=false)
BuildCallExpr - Handle a call to Fn with the specified array of arguments.
Definition: SemaExpr.cpp:6903
const LangOptions & LangOpts
Definition: Sema.h:405
void MaybeAddCUDAConstantAttr(VarDecl *VD)
May add implicit CUDAConstantAttr attribute to VD, depending on VD and current compilation settings.
Definition: SemaCUDA.cpp:711
CUDAFunctionPreference IdentifyCUDAPreference(const FunctionDecl *Caller, const FunctionDecl *Callee)
Identifies relative preference of a given Caller/Callee combination, based on their host/device attri...
Definition: SemaCUDA.cpp:207
void maybeAddCUDAHostDeviceAttrs(FunctionDecl *FD, const LookupResult &Previous)
May add implicit CUDAHostAttr and CUDADeviceAttr attributes to FD, depending on FD and the current co...
Definition: SemaCUDA.cpp:659
void PushForceCUDAHostDevice()
Increments our count of the number of times we've seen a pragma forcing functions to be host device.
Definition: SemaCUDA.cpp:37
SourceManager & getSourceManager() const
Definition: Sema.h:1650
llvm::DenseSet< FunctionDeclAndLoc > LocsWithCUDACallDiags
FunctionDecls and SourceLocations for which CheckCUDACall has emitted a (maybe deferred) "bad call" d...
Definition: Sema.h:13000
SemaDiagnosticBuilder CUDADiagIfHostCode(SourceLocation Loc, unsigned DiagID)
Creates a SemaDiagnosticBuilder that emits the diagnostic if the current context is "used as host cod...
Definition: SemaCUDA.cpp:755
bool inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl, CXXSpecialMember CSM, CXXMethodDecl *MemberDecl, bool ConstRHS, bool Diagnose)
Given a implicit special member, infer its CUDA target from the calls it needs to make to underlying ...
Definition: SemaCUDA.cpp:324
void CUDASetLambdaAttrs(CXXMethodDecl *Method)
Set device or host device attributes on the given lambda operator() method.
Definition: SemaCUDA.cpp:895
FunctionEmissionStatus getEmissionStatus(FunctionDecl *Decl, bool Final=false)
Definition: SemaDecl.cpp:19883
void InstantiateFunctionDefinition(SourceLocation PointOfInstantiation, FunctionDecl *Function, bool Recursive=false, bool DefinitionRequired=false, bool AtEndOfTU=false)
Instantiate the definition of the given function from its template.
CUDAVariableTarget
Definition: Sema.h:13100
@ CVT_Both
Emitted on host side only.
Definition: Sema.h:13103
@ CVT_Device
Definition: Sema.h:13101
@ CVT_Unified
Emitted on both sides with different addresses.
Definition: Sema.h:13104
@ CVT_Host
Emitted on device side with a shadow variable on host side.
Definition: Sema.h:13102
CUDAFunctionTarget
Definition: Sema.h:13083
@ CFT_Device
Definition: Sema.h:13084
@ CFT_HostDevice
Definition: Sema.h:13087
@ CFT_Global
Definition: Sema.h:13085
@ CFT_Host
Definition: Sema.h:13086
@ CFT_InvalidTarget
Definition: Sema.h:13088
void inheritCUDATargetAttrs(FunctionDecl *FD, const FunctionTemplateDecl &TD)
Copies target attributes from the template TD to the function FD.
Definition: SemaCUDA.cpp:942
CUDAFunctionTarget CurrentCUDATarget()
Gets the CUDA target for the current context.
Definition: Sema.h:13110
bool isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD)
Definition: SemaCUDA.cpp:461
SmallVector< ExpressionEvaluationContextRecord, 8 > ExprEvalContexts
A stack of expression evaluation contexts.
Definition: Sema.h:1393
DiagnosticsEngine & Diags
Definition: Sema.h:409
bool isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *CD)
Definition: SemaCUDA.cpp:499
void MarkFunctionReferenced(SourceLocation Loc, FunctionDecl *Func, bool MightBeOdrUse=true)
Mark a function referenced, and check whether it is odr-used (C++ [basic.def.odr]p2,...
Definition: SemaExpr.cpp:18265
static bool isCUDAImplicitHostDeviceFunction(const FunctionDecl *D)
Definition: SemaCUDA.cpp:266
bool PopForceCUDAHostDevice()
Decrements our count of the number of times we've seen a pragma forcing functions to be host device.
Definition: SemaCUDA.cpp:42
@ Diagnose
Diagnose issues that are non-constant or that are extensions.
CXXSpecialMember
Kinds of C++ special members.
Definition: Sema.h:1535
std::string getCudaConfigureFuncName() const
Returns the name of the launch configuration function.
Definition: SemaCUDA.cpp:950
void checkCUDATargetOverload(FunctionDecl *NewFD, const LookupResult &Previous)
Check whether NewFD is a valid overload for CUDA.
Definition: SemaCUDA.cpp:903
Encodes a location in the source.
bool isUnion() const
Definition: Decl.h:3644
const llvm::VersionTuple & getSDKVersion() const
Definition: TargetInfo.h:1674
CXXRecordDecl * getAsCXXRecordDecl() const
Retrieves the CXXRecordDecl that this type refers to, either because the type is a RecordType or beca...
Definition: Type.cpp:1783
bool isReferenceType() const
Definition: Type.h:6922
bool isCUDADeviceBuiltinSurfaceType() const
Check if the type is the CUDA device builtin surface type.
Definition: Type.cpp:4507
bool isDependentType() const
Whether this type is a dependent type, meaning that its definition somehow depends on a template para...
Definition: Type.h:2317
bool isCUDADeviceBuiltinTextureType() const
Check if the type is the CUDA device builtin texture type.
Definition: Type.cpp:4514
const T * getAs() const
Member-template getAs<specific type>'.
Definition: Type.h:7424
Represents a shadow declaration implicitly introduced into a scope by a (resolved) using-declaration ...
Definition: DeclCXX.h:3257
QualType getType() const
Definition: Decl.h:712
Represents a variable declaration or definition.
Definition: Decl.h:913
bool isConstexpr() const
Whether this variable is (C++11) constexpr.
Definition: Decl.h:1519
bool hasInit() const
Definition: Decl.cpp:2343
bool isStaticDataMember() const
Determines whether this is a static data member.
Definition: Decl.h:1240
bool hasGlobalStorage() const
Returns true for all variables that do not have local storage.
Definition: Decl.h:1183
bool isFileVarDecl() const
Returns true for file scoped variable declaration.
Definition: Decl.h:1299
const Expr * getInit() const
Definition: Decl.h:1325
ValueDecl * getVariable() const
Definition: ScopeInfo.h:646
bool isVariableCapture() const
Definition: ScopeInfo.h:621
SourceLocation getLocation() const
Retrieve the location at which this variable was captured.
Definition: ScopeInfo.h:657
bool isThisCapture() const
Definition: ScopeInfo.h:620
bool isReferenceCapture() const
Definition: ScopeInfo.h:626
Defines the clang::TargetInfo interface.
bool CudaFeatureEnabled(llvm::VersionTuple, CudaFeature)
Definition: Cuda.cpp:227
ExprResult ExprError()
Definition: Ownership.h:278
@ VK_LValue
An l-value expression is a reference to an object with independent storage.
Definition: Specifiers.h:127