1 //===--- SemaCUDA.cpp - Semantic Analysis for CUDA constructs -------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file 9 /// This file implements semantic analysis for CUDA constructs. 10 /// 11 //===----------------------------------------------------------------------===// 12 13 #include "clang/AST/ASTContext.h" 14 #include "clang/AST/Decl.h" 15 #include "clang/AST/ExprCXX.h" 16 #include "clang/Basic/Cuda.h" 17 #include "clang/Basic/TargetInfo.h" 18 #include "clang/Lex/Preprocessor.h" 19 #include "clang/Sema/Lookup.h" 20 #include "clang/Sema/ScopeInfo.h" 21 #include "clang/Sema/Sema.h" 22 #include "clang/Sema/SemaDiagnostic.h" 23 #include "clang/Sema/SemaInternal.h" 24 #include "clang/Sema/Template.h" 25 #include "llvm/ADT/Optional.h" 26 #include "llvm/ADT/SmallVector.h" 27 using namespace clang; 28 29 template <typename AttrT> static bool hasExplicitAttr(const VarDecl *D) { 30 if (!D) 31 return false; 32 if (auto *A = D->getAttr<AttrT>()) 33 return !A->isImplicit(); 34 return false; 35 } 36 37 void Sema::PushForceCUDAHostDevice() { 38 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 39 ForceCUDAHostDeviceDepth++; 40 } 41 42 bool Sema::PopForceCUDAHostDevice() { 43 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 44 if (ForceCUDAHostDeviceDepth == 0) 45 return false; 46 ForceCUDAHostDeviceDepth--; 47 return true; 48 } 49 50 ExprResult Sema::ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc, 51 MultiExprArg ExecConfig, 52 SourceLocation GGGLoc) { 53 FunctionDecl *ConfigDecl = Context.getcudaConfigureCallDecl(); 54 if (!ConfigDecl) 55 return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use) 56 << getCudaConfigureFuncName()); 57 QualType ConfigQTy = ConfigDecl->getType(); 58 59 DeclRefExpr *ConfigDR = new (Context) 60 DeclRefExpr(Context, ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc); 61 MarkFunctionReferenced(LLLLoc, ConfigDecl); 62 63 return BuildCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr, 64 /*IsExecConfig=*/true); 65 } 66 67 Sema::CUDAFunctionTarget 68 Sema::IdentifyCUDATarget(const ParsedAttributesView &Attrs) { 69 bool HasHostAttr = false; 70 bool HasDeviceAttr = false; 71 bool HasGlobalAttr = false; 72 bool HasInvalidTargetAttr = false; 73 for (const ParsedAttr &AL : Attrs) { 74 switch (AL.getKind()) { 75 case ParsedAttr::AT_CUDAGlobal: 76 HasGlobalAttr = true; 77 break; 78 case ParsedAttr::AT_CUDAHost: 79 HasHostAttr = true; 80 break; 81 case ParsedAttr::AT_CUDADevice: 82 HasDeviceAttr = true; 83 break; 84 case ParsedAttr::AT_CUDAInvalidTarget: 85 HasInvalidTargetAttr = true; 86 break; 87 default: 88 break; 89 } 90 } 91 92 if (HasInvalidTargetAttr) 93 return CFT_InvalidTarget; 94 95 if (HasGlobalAttr) 96 return CFT_Global; 97 98 if (HasHostAttr && HasDeviceAttr) 99 return CFT_HostDevice; 100 101 if (HasDeviceAttr) 102 return CFT_Device; 103 104 return CFT_Host; 105 } 106 107 template <typename A> 108 static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) { 109 return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) { 110 return isa<A>(Attribute) && 111 !(IgnoreImplicitAttr && Attribute->isImplicit()); 112 }); 113 } 114 115 /// IdentifyCUDATarget - Determine the CUDA compilation target for this function 116 Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D, 117 bool IgnoreImplicitHDAttr) { 118 // Code that lives outside a function is run on the host. 119 if (D == nullptr) 120 return CFT_Host; 121 122 if (D->hasAttr<CUDAInvalidTargetAttr>()) 123 return CFT_InvalidTarget; 124 125 if (D->hasAttr<CUDAGlobalAttr>()) 126 return CFT_Global; 127 128 if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) { 129 if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) 130 return CFT_HostDevice; 131 return CFT_Device; 132 } else if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)) { 133 return CFT_Host; 134 } else if ((D->isImplicit() || !D->isUserProvided()) && 135 !IgnoreImplicitHDAttr) { 136 // Some implicit declarations (like intrinsic functions) are not marked. 137 // Set the most lenient target on them for maximal flexibility. 138 return CFT_HostDevice; 139 } 140 141 return CFT_Host; 142 } 143 144 /// IdentifyTarget - Determine the CUDA compilation target for this variable. 145 Sema::CUDAVariableTarget Sema::IdentifyCUDATarget(const VarDecl *Var) { 146 if (Var->hasAttr<HIPManagedAttr>()) 147 return CVT_Unified; 148 if (Var->isConstexpr() && !hasExplicitAttr<CUDAConstantAttr>(Var)) 149 return CVT_Both; 150 if (Var->getType().isConstQualified() && Var->hasAttr<CUDAConstantAttr>() && 151 !hasExplicitAttr<CUDAConstantAttr>(Var)) 152 return CVT_Both; 153 if (Var->hasAttr<CUDADeviceAttr>() || Var->hasAttr<CUDAConstantAttr>() || 154 Var->hasAttr<CUDASharedAttr>() || 155 Var->getType()->isCUDADeviceBuiltinSurfaceType() || 156 Var->getType()->isCUDADeviceBuiltinTextureType()) 157 return CVT_Device; 158 // Function-scope static variable without explicit device or constant 159 // attribute are emitted 160 // - on both sides in host device functions 161 // - on device side in device or global functions 162 if (auto *FD = dyn_cast<FunctionDecl>(Var->getDeclContext())) { 163 switch (IdentifyCUDATarget(FD)) { 164 case CFT_HostDevice: 165 return CVT_Both; 166 case CFT_Device: 167 case CFT_Global: 168 return CVT_Device; 169 default: 170 return CVT_Host; 171 } 172 } 173 return CVT_Host; 174 } 175 176 // * CUDA Call preference table 177 // 178 // F - from, 179 // T - to 180 // Ph - preference in host mode 181 // Pd - preference in device mode 182 // H - handled in (x) 183 // Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never. 184 // 185 // | F | T | Ph | Pd | H | 186 // |----+----+-----+-----+-----+ 187 // | d | d | N | N | (c) | 188 // | d | g | -- | -- | (a) | 189 // | d | h | -- | -- | (e) | 190 // | d | hd | HD | HD | (b) | 191 // | g | d | N | N | (c) | 192 // | g | g | -- | -- | (a) | 193 // | g | h | -- | -- | (e) | 194 // | g | hd | HD | HD | (b) | 195 // | h | d | -- | -- | (e) | 196 // | h | g | N | N | (c) | 197 // | h | h | N | N | (c) | 198 // | h | hd | HD | HD | (b) | 199 // | hd | d | WS | SS | (d) | 200 // | hd | g | SS | -- |(d/a)| 201 // | hd | h | SS | WS | (d) | 202 // | hd | hd | HD | HD | (b) | 203 204 Sema::CUDAFunctionPreference 205 Sema::IdentifyCUDAPreference(const FunctionDecl *Caller, 206 const FunctionDecl *Callee) { 207 assert(Callee && "Callee must be valid."); 208 CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller); 209 CUDAFunctionTarget CalleeTarget = IdentifyCUDATarget(Callee); 210 211 // If one of the targets is invalid, the check always fails, no matter what 212 // the other target is. 213 if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget) 214 return CFP_Never; 215 216 // (a) Can't call global from some contexts until we support CUDA's 217 // dynamic parallelism. 218 if (CalleeTarget == CFT_Global && 219 (CallerTarget == CFT_Global || CallerTarget == CFT_Device)) 220 return CFP_Never; 221 222 // (b) Calling HostDevice is OK for everyone. 223 if (CalleeTarget == CFT_HostDevice) 224 return CFP_HostDevice; 225 226 // (c) Best case scenarios 227 if (CalleeTarget == CallerTarget || 228 (CallerTarget == CFT_Host && CalleeTarget == CFT_Global) || 229 (CallerTarget == CFT_Global && CalleeTarget == CFT_Device)) 230 return CFP_Native; 231 232 // (d) HostDevice behavior depends on compilation mode. 233 if (CallerTarget == CFT_HostDevice) { 234 // It's OK to call a compilation-mode matching function from an HD one. 235 if ((getLangOpts().CUDAIsDevice && CalleeTarget == CFT_Device) || 236 (!getLangOpts().CUDAIsDevice && 237 (CalleeTarget == CFT_Host || CalleeTarget == CFT_Global))) 238 return CFP_SameSide; 239 240 // Calls from HD to non-mode-matching functions (i.e., to host functions 241 // when compiling in device mode or to device functions when compiling in 242 // host mode) are allowed at the sema level, but eventually rejected if 243 // they're ever codegened. TODO: Reject said calls earlier. 244 return CFP_WrongSide; 245 } 246 247 // (e) Calling across device/host boundary is not something you should do. 248 if ((CallerTarget == CFT_Host && CalleeTarget == CFT_Device) || 249 (CallerTarget == CFT_Device && CalleeTarget == CFT_Host) || 250 (CallerTarget == CFT_Global && CalleeTarget == CFT_Host)) 251 return CFP_Never; 252 253 llvm_unreachable("All cases should've been handled by now."); 254 } 255 256 template <typename AttrT> static bool hasImplicitAttr(const FunctionDecl *D) { 257 if (!D) 258 return false; 259 if (auto *A = D->getAttr<AttrT>()) 260 return A->isImplicit(); 261 return D->isImplicit(); 262 } 263 264 bool Sema::isCUDAImplicitHostDeviceFunction(const FunctionDecl *D) { 265 bool IsImplicitDevAttr = hasImplicitAttr<CUDADeviceAttr>(D); 266 bool IsImplicitHostAttr = hasImplicitAttr<CUDAHostAttr>(D); 267 return IsImplicitDevAttr && IsImplicitHostAttr; 268 } 269 270 void Sema::EraseUnwantedCUDAMatches( 271 const FunctionDecl *Caller, 272 SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) { 273 if (Matches.size() <= 1) 274 return; 275 276 using Pair = std::pair<DeclAccessPair, FunctionDecl*>; 277 278 // Gets the CUDA function preference for a call from Caller to Match. 279 auto GetCFP = [&](const Pair &Match) { 280 return IdentifyCUDAPreference(Caller, Match.second); 281 }; 282 283 // Find the best call preference among the functions in Matches. 284 CUDAFunctionPreference BestCFP = GetCFP(*std::max_element( 285 Matches.begin(), Matches.end(), 286 [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); })); 287 288 // Erase all functions with lower priority. 289 llvm::erase_if(Matches, 290 [&](const Pair &Match) { return GetCFP(Match) < BestCFP; }); 291 } 292 293 /// When an implicitly-declared special member has to invoke more than one 294 /// base/field special member, conflicts may occur in the targets of these 295 /// members. For example, if one base's member __host__ and another's is 296 /// __device__, it's a conflict. 297 /// This function figures out if the given targets \param Target1 and 298 /// \param Target2 conflict, and if they do not it fills in 299 /// \param ResolvedTarget with a target that resolves for both calls. 300 /// \return true if there's a conflict, false otherwise. 301 static bool 302 resolveCalleeCUDATargetConflict(Sema::CUDAFunctionTarget Target1, 303 Sema::CUDAFunctionTarget Target2, 304 Sema::CUDAFunctionTarget *ResolvedTarget) { 305 // Only free functions and static member functions may be global. 306 assert(Target1 != Sema::CFT_Global); 307 assert(Target2 != Sema::CFT_Global); 308 309 if (Target1 == Sema::CFT_HostDevice) { 310 *ResolvedTarget = Target2; 311 } else if (Target2 == Sema::CFT_HostDevice) { 312 *ResolvedTarget = Target1; 313 } else if (Target1 != Target2) { 314 return true; 315 } else { 316 *ResolvedTarget = Target1; 317 } 318 319 return false; 320 } 321 322 bool Sema::inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl, 323 CXXSpecialMember CSM, 324 CXXMethodDecl *MemberDecl, 325 bool ConstRHS, 326 bool Diagnose) { 327 // If the defaulted special member is defined lexically outside of its 328 // owning class, or the special member already has explicit device or host 329 // attributes, do not infer. 330 bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent(); 331 bool HasH = MemberDecl->hasAttr<CUDAHostAttr>(); 332 bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>(); 333 bool HasExplicitAttr = 334 (HasD && !MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()) || 335 (HasH && !MemberDecl->getAttr<CUDAHostAttr>()->isImplicit()); 336 if (!InClass || HasExplicitAttr) 337 return false; 338 339 llvm::Optional<CUDAFunctionTarget> InferredTarget; 340 341 // We're going to invoke special member lookup; mark that these special 342 // members are called from this one, and not from its caller. 343 ContextRAII MethodContext(*this, MemberDecl); 344 345 // Look for special members in base classes that should be invoked from here. 346 // Infer the target of this member base on the ones it should call. 347 // Skip direct and indirect virtual bases for abstract classes. 348 llvm::SmallVector<const CXXBaseSpecifier *, 16> Bases; 349 for (const auto &B : ClassDecl->bases()) { 350 if (!B.isVirtual()) { 351 Bases.push_back(&B); 352 } 353 } 354 355 if (!ClassDecl->isAbstract()) { 356 for (const auto &VB : ClassDecl->vbases()) { 357 Bases.push_back(&VB); 358 } 359 } 360 361 for (const auto *B : Bases) { 362 const RecordType *BaseType = B->getType()->getAs<RecordType>(); 363 if (!BaseType) { 364 continue; 365 } 366 367 CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl()); 368 Sema::SpecialMemberOverloadResult SMOR = 369 LookupSpecialMember(BaseClassDecl, CSM, 370 /* ConstArg */ ConstRHS, 371 /* VolatileArg */ false, 372 /* RValueThis */ false, 373 /* ConstThis */ false, 374 /* VolatileThis */ false); 375 376 if (!SMOR.getMethod()) 377 continue; 378 379 CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR.getMethod()); 380 if (!InferredTarget.hasValue()) { 381 InferredTarget = BaseMethodTarget; 382 } else { 383 bool ResolutionError = resolveCalleeCUDATargetConflict( 384 InferredTarget.getValue(), BaseMethodTarget, 385 InferredTarget.getPointer()); 386 if (ResolutionError) { 387 if (Diagnose) { 388 Diag(ClassDecl->getLocation(), 389 diag::note_implicit_member_target_infer_collision) 390 << (unsigned)CSM << InferredTarget.getValue() << BaseMethodTarget; 391 } 392 MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context)); 393 return true; 394 } 395 } 396 } 397 398 // Same as for bases, but now for special members of fields. 399 for (const auto *F : ClassDecl->fields()) { 400 if (F->isInvalidDecl()) { 401 continue; 402 } 403 404 const RecordType *FieldType = 405 Context.getBaseElementType(F->getType())->getAs<RecordType>(); 406 if (!FieldType) { 407 continue; 408 } 409 410 CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl()); 411 Sema::SpecialMemberOverloadResult SMOR = 412 LookupSpecialMember(FieldRecDecl, CSM, 413 /* ConstArg */ ConstRHS && !F->isMutable(), 414 /* VolatileArg */ false, 415 /* RValueThis */ false, 416 /* ConstThis */ false, 417 /* VolatileThis */ false); 418 419 if (!SMOR.getMethod()) 420 continue; 421 422 CUDAFunctionTarget FieldMethodTarget = 423 IdentifyCUDATarget(SMOR.getMethod()); 424 if (!InferredTarget.hasValue()) { 425 InferredTarget = FieldMethodTarget; 426 } else { 427 bool ResolutionError = resolveCalleeCUDATargetConflict( 428 InferredTarget.getValue(), FieldMethodTarget, 429 InferredTarget.getPointer()); 430 if (ResolutionError) { 431 if (Diagnose) { 432 Diag(ClassDecl->getLocation(), 433 diag::note_implicit_member_target_infer_collision) 434 << (unsigned)CSM << InferredTarget.getValue() 435 << FieldMethodTarget; 436 } 437 MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context)); 438 return true; 439 } 440 } 441 } 442 443 444 // If no target was inferred, mark this member as __host__ __device__; 445 // it's the least restrictive option that can be invoked from any target. 446 bool NeedsH = true, NeedsD = true; 447 if (InferredTarget.hasValue()) { 448 if (InferredTarget.getValue() == CFT_Device) 449 NeedsH = false; 450 else if (InferredTarget.getValue() == CFT_Host) 451 NeedsD = false; 452 } 453 454 // We either setting attributes first time, or the inferred ones must match 455 // previously set ones. 456 if (NeedsD && !HasD) 457 MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context)); 458 if (NeedsH && !HasH) 459 MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context)); 460 461 return false; 462 } 463 464 bool Sema::isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD) { 465 if (!CD->isDefined() && CD->isTemplateInstantiation()) 466 InstantiateFunctionDefinition(Loc, CD->getFirstDecl()); 467 468 // (E.2.3.1, CUDA 7.5) A constructor for a class type is considered 469 // empty at a point in the translation unit, if it is either a 470 // trivial constructor 471 if (CD->isTrivial()) 472 return true; 473 474 // ... or it satisfies all of the following conditions: 475 // The constructor function has been defined. 476 // The constructor function has no parameters, 477 // and the function body is an empty compound statement. 478 if (!(CD->hasTrivialBody() && CD->getNumParams() == 0)) 479 return false; 480 481 // Its class has no virtual functions and no virtual base classes. 482 if (CD->getParent()->isDynamicClass()) 483 return false; 484 485 // Union ctor does not call ctors of its data members. 486 if (CD->getParent()->isUnion()) 487 return true; 488 489 // The only form of initializer allowed is an empty constructor. 490 // This will recursively check all base classes and member initializers 491 if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) { 492 if (const CXXConstructExpr *CE = 493 dyn_cast<CXXConstructExpr>(CI->getInit())) 494 return isEmptyCudaConstructor(Loc, CE->getConstructor()); 495 return false; 496 })) 497 return false; 498 499 return true; 500 } 501 502 bool Sema::isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *DD) { 503 // No destructor -> no problem. 504 if (!DD) 505 return true; 506 507 if (!DD->isDefined() && DD->isTemplateInstantiation()) 508 InstantiateFunctionDefinition(Loc, DD->getFirstDecl()); 509 510 // (E.2.3.1, CUDA 7.5) A destructor for a class type is considered 511 // empty at a point in the translation unit, if it is either a 512 // trivial constructor 513 if (DD->isTrivial()) 514 return true; 515 516 // ... or it satisfies all of the following conditions: 517 // The destructor function has been defined. 518 // and the function body is an empty compound statement. 519 if (!DD->hasTrivialBody()) 520 return false; 521 522 const CXXRecordDecl *ClassDecl = DD->getParent(); 523 524 // Its class has no virtual functions and no virtual base classes. 525 if (ClassDecl->isDynamicClass()) 526 return false; 527 528 // Union does not have base class and union dtor does not call dtors of its 529 // data members. 530 if (DD->getParent()->isUnion()) 531 return true; 532 533 // Only empty destructors are allowed. This will recursively check 534 // destructors for all base classes... 535 if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) { 536 if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl()) 537 return isEmptyCudaDestructor(Loc, RD->getDestructor()); 538 return true; 539 })) 540 return false; 541 542 // ... and member fields. 543 if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) { 544 if (CXXRecordDecl *RD = Field->getType() 545 ->getBaseElementTypeUnsafe() 546 ->getAsCXXRecordDecl()) 547 return isEmptyCudaDestructor(Loc, RD->getDestructor()); 548 return true; 549 })) 550 return false; 551 552 return true; 553 } 554 555 namespace { 556 enum CUDAInitializerCheckKind { 557 CICK_DeviceOrConstant, // Check initializer for device/constant variable 558 CICK_Shared, // Check initializer for shared variable 559 }; 560 561 bool IsDependentVar(VarDecl *VD) { 562 if (VD->getType()->isDependentType()) 563 return true; 564 if (const auto *Init = VD->getInit()) 565 return Init->isValueDependent(); 566 return false; 567 } 568 569 // Check whether a variable has an allowed initializer for a CUDA device side 570 // variable with global storage. \p VD may be a host variable to be checked for 571 // potential promotion to device side variable. 572 // 573 // CUDA/HIP allows only empty constructors as initializers for global 574 // variables (see E.2.3.1, CUDA 7.5). The same restriction also applies to all 575 // __shared__ variables whether they are local or not (they all are implicitly 576 // static in CUDA). One exception is that CUDA allows constant initializers 577 // for __constant__ and __device__ variables. 578 bool HasAllowedCUDADeviceStaticInitializer(Sema &S, VarDecl *VD, 579 CUDAInitializerCheckKind CheckKind) { 580 assert(!VD->isInvalidDecl() && VD->hasGlobalStorage()); 581 assert(!IsDependentVar(VD) && "do not check dependent var"); 582 const Expr *Init = VD->getInit(); 583 auto IsEmptyInit = [&](const Expr *Init) { 584 if (!Init) 585 return true; 586 if (const auto *CE = dyn_cast<CXXConstructExpr>(Init)) { 587 return S.isEmptyCudaConstructor(VD->getLocation(), CE->getConstructor()); 588 } 589 return false; 590 }; 591 auto IsConstantInit = [&](const Expr *Init) { 592 assert(Init); 593 return Init->isConstantInitializer(S.Context, 594 VD->getType()->isReferenceType()); 595 }; 596 auto HasEmptyDtor = [&](VarDecl *VD) { 597 if (const auto *RD = VD->getType()->getAsCXXRecordDecl()) 598 return S.isEmptyCudaDestructor(VD->getLocation(), RD->getDestructor()); 599 return true; 600 }; 601 if (CheckKind == CICK_Shared) 602 return IsEmptyInit(Init) && HasEmptyDtor(VD); 603 return S.LangOpts.GPUAllowDeviceInit || 604 ((IsEmptyInit(Init) || IsConstantInit(Init)) && HasEmptyDtor(VD)); 605 } 606 } // namespace 607 608 void Sema::checkAllowedCUDAInitializer(VarDecl *VD) { 609 // Do not check dependent variables since the ctor/dtor/initializer are not 610 // determined. Do it after instantiation. 611 if (VD->isInvalidDecl() || !VD->hasInit() || !VD->hasGlobalStorage() || 612 IsDependentVar(VD)) 613 return; 614 const Expr *Init = VD->getInit(); 615 bool IsSharedVar = VD->hasAttr<CUDASharedAttr>(); 616 bool IsDeviceOrConstantVar = 617 !IsSharedVar && 618 (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>()); 619 if (IsDeviceOrConstantVar || IsSharedVar) { 620 if (HasAllowedCUDADeviceStaticInitializer( 621 *this, VD, IsSharedVar ? CICK_Shared : CICK_DeviceOrConstant)) 622 return; 623 Diag(VD->getLocation(), 624 IsSharedVar ? diag::err_shared_var_init : diag::err_dynamic_var_init) 625 << Init->getSourceRange(); 626 VD->setInvalidDecl(); 627 } else { 628 // This is a host-side global variable. Check that the initializer is 629 // callable from the host side. 630 const FunctionDecl *InitFn = nullptr; 631 if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init)) { 632 InitFn = CE->getConstructor(); 633 } else if (const CallExpr *CE = dyn_cast<CallExpr>(Init)) { 634 InitFn = CE->getDirectCallee(); 635 } 636 if (InitFn) { 637 CUDAFunctionTarget InitFnTarget = IdentifyCUDATarget(InitFn); 638 if (InitFnTarget != CFT_Host && InitFnTarget != CFT_HostDevice) { 639 Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer) 640 << InitFnTarget << InitFn; 641 Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn; 642 VD->setInvalidDecl(); 643 } 644 } 645 } 646 } 647 648 // With -fcuda-host-device-constexpr, an unattributed constexpr function is 649 // treated as implicitly __host__ __device__, unless: 650 // * it is a variadic function (device-side variadic functions are not 651 // allowed), or 652 // * a __device__ function with this signature was already declared, in which 653 // case in which case we output an error, unless the __device__ decl is in a 654 // system header, in which case we leave the constexpr function unattributed. 655 // 656 // In addition, all function decls are treated as __host__ __device__ when 657 // ForceCUDAHostDeviceDepth > 0 (corresponding to code within a 658 // #pragma clang force_cuda_host_device_begin/end 659 // pair). 660 void Sema::maybeAddCUDAHostDeviceAttrs(FunctionDecl *NewD, 661 const LookupResult &Previous) { 662 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 663 664 if (ForceCUDAHostDeviceDepth > 0) { 665 if (!NewD->hasAttr<CUDAHostAttr>()) 666 NewD->addAttr(CUDAHostAttr::CreateImplicit(Context)); 667 if (!NewD->hasAttr<CUDADeviceAttr>()) 668 NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context)); 669 return; 670 } 671 672 if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() || 673 NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() || 674 NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>()) 675 return; 676 677 // Is D a __device__ function with the same signature as NewD, ignoring CUDA 678 // attributes? 679 auto IsMatchingDeviceFn = [&](NamedDecl *D) { 680 if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D)) 681 D = Using->getTargetDecl(); 682 FunctionDecl *OldD = D->getAsFunction(); 683 return OldD && OldD->hasAttr<CUDADeviceAttr>() && 684 !OldD->hasAttr<CUDAHostAttr>() && 685 !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false, 686 /* ConsiderCudaAttrs = */ false); 687 }; 688 auto It = llvm::find_if(Previous, IsMatchingDeviceFn); 689 if (It != Previous.end()) { 690 // We found a __device__ function with the same name and signature as NewD 691 // (ignoring CUDA attrs). This is an error unless that function is defined 692 // in a system header, in which case we simply return without making NewD 693 // host+device. 694 NamedDecl *Match = *It; 695 if (!getSourceManager().isInSystemHeader(Match->getLocation())) { 696 Diag(NewD->getLocation(), 697 diag::err_cuda_unattributed_constexpr_cannot_overload_device) 698 << NewD; 699 Diag(Match->getLocation(), 700 diag::note_cuda_conflicting_device_function_declared_here); 701 } 702 return; 703 } 704 705 NewD->addAttr(CUDAHostAttr::CreateImplicit(Context)); 706 NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context)); 707 } 708 709 // TODO: `__constant__` memory may be a limited resource for certain targets. 710 // A safeguard may be needed at the end of compilation pipeline if 711 // `__constant__` memory usage goes beyond limit. 712 void Sema::MaybeAddCUDAConstantAttr(VarDecl *VD) { 713 // Do not promote dependent variables since the cotr/dtor/initializer are 714 // not determined. Do it after instantiation. 715 if (getLangOpts().CUDAIsDevice && !VD->hasAttr<CUDAConstantAttr>() && 716 !VD->hasAttr<CUDAConstantAttr>() && !VD->hasAttr<CUDASharedAttr>() && 717 (VD->isFileVarDecl() || VD->isStaticDataMember()) && 718 !IsDependentVar(VD) && 719 (VD->isConstexpr() || (VD->getType().isConstQualified() && 720 HasAllowedCUDADeviceStaticInitializer( 721 *this, VD, CICK_DeviceOrConstant)))) { 722 VD->addAttr(CUDAConstantAttr::CreateImplicit(getASTContext())); 723 } 724 } 725 726 Sema::SemaDiagnosticBuilder Sema::CUDADiagIfDeviceCode(SourceLocation Loc, 727 unsigned DiagID) { 728 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 729 SemaDiagnosticBuilder::Kind DiagKind = [&] { 730 if (!isa<FunctionDecl>(CurContext)) 731 return SemaDiagnosticBuilder::K_Nop; 732 switch (CurrentCUDATarget()) { 733 case CFT_Global: 734 case CFT_Device: 735 return SemaDiagnosticBuilder::K_Immediate; 736 case CFT_HostDevice: 737 // An HD function counts as host code if we're compiling for host, and 738 // device code if we're compiling for device. Defer any errors in device 739 // mode until the function is known-emitted. 740 if (!getLangOpts().CUDAIsDevice) 741 return SemaDiagnosticBuilder::K_Nop; 742 if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID)) 743 return SemaDiagnosticBuilder::K_Immediate; 744 return (getEmissionStatus(cast<FunctionDecl>(CurContext)) == 745 FunctionEmissionStatus::Emitted) 746 ? SemaDiagnosticBuilder::K_ImmediateWithCallStack 747 : SemaDiagnosticBuilder::K_Deferred; 748 default: 749 return SemaDiagnosticBuilder::K_Nop; 750 } 751 }(); 752 return SemaDiagnosticBuilder(DiagKind, Loc, DiagID, 753 dyn_cast<FunctionDecl>(CurContext), *this); 754 } 755 756 Sema::SemaDiagnosticBuilder Sema::CUDADiagIfHostCode(SourceLocation Loc, 757 unsigned DiagID) { 758 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 759 SemaDiagnosticBuilder::Kind DiagKind = [&] { 760 if (!isa<FunctionDecl>(CurContext)) 761 return SemaDiagnosticBuilder::K_Nop; 762 switch (CurrentCUDATarget()) { 763 case CFT_Host: 764 return SemaDiagnosticBuilder::K_Immediate; 765 case CFT_HostDevice: 766 // An HD function counts as host code if we're compiling for host, and 767 // device code if we're compiling for device. Defer any errors in device 768 // mode until the function is known-emitted. 769 if (getLangOpts().CUDAIsDevice) 770 return SemaDiagnosticBuilder::K_Nop; 771 if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID)) 772 return SemaDiagnosticBuilder::K_Immediate; 773 return (getEmissionStatus(cast<FunctionDecl>(CurContext)) == 774 FunctionEmissionStatus::Emitted) 775 ? SemaDiagnosticBuilder::K_ImmediateWithCallStack 776 : SemaDiagnosticBuilder::K_Deferred; 777 default: 778 return SemaDiagnosticBuilder::K_Nop; 779 } 780 }(); 781 return SemaDiagnosticBuilder(DiagKind, Loc, DiagID, 782 dyn_cast<FunctionDecl>(CurContext), *this); 783 } 784 785 bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) { 786 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 787 assert(Callee && "Callee may not be null."); 788 789 auto &ExprEvalCtx = ExprEvalContexts.back(); 790 if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated()) 791 return true; 792 793 // FIXME: Is bailing out early correct here? Should we instead assume that 794 // the caller is a global initializer? 795 FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext); 796 if (!Caller) 797 return true; 798 799 // If the caller is known-emitted, mark the callee as known-emitted. 800 // Otherwise, mark the call in our call graph so we can traverse it later. 801 bool CallerKnownEmitted = 802 getEmissionStatus(Caller) == FunctionEmissionStatus::Emitted; 803 SemaDiagnosticBuilder::Kind DiagKind = [this, Caller, Callee, 804 CallerKnownEmitted] { 805 switch (IdentifyCUDAPreference(Caller, Callee)) { 806 case CFP_Never: 807 case CFP_WrongSide: 808 assert(Caller && "Never/wrongSide calls require a non-null caller"); 809 // If we know the caller will be emitted, we know this wrong-side call 810 // will be emitted, so it's an immediate error. Otherwise, defer the 811 // error until we know the caller is emitted. 812 return CallerKnownEmitted 813 ? SemaDiagnosticBuilder::K_ImmediateWithCallStack 814 : SemaDiagnosticBuilder::K_Deferred; 815 default: 816 return SemaDiagnosticBuilder::K_Nop; 817 } 818 }(); 819 820 if (DiagKind == SemaDiagnosticBuilder::K_Nop) 821 return true; 822 823 // Avoid emitting this error twice for the same location. Using a hashtable 824 // like this is unfortunate, but because we must continue parsing as normal 825 // after encountering a deferred error, it's otherwise very tricky for us to 826 // ensure that we only emit this deferred error once. 827 if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second) 828 return true; 829 830 SemaDiagnosticBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this) 831 << IdentifyCUDATarget(Callee) << /*function*/ 0 << Callee 832 << IdentifyCUDATarget(Caller); 833 if (!Callee->getBuiltinID()) 834 SemaDiagnosticBuilder(DiagKind, Callee->getLocation(), 835 diag::note_previous_decl, Caller, *this) 836 << Callee; 837 return DiagKind != SemaDiagnosticBuilder::K_Immediate && 838 DiagKind != SemaDiagnosticBuilder::K_ImmediateWithCallStack; 839 } 840 841 // Check the wrong-sided reference capture of lambda for CUDA/HIP. 842 // A lambda function may capture a stack variable by reference when it is 843 // defined and uses the capture by reference when the lambda is called. When 844 // the capture and use happen on different sides, the capture is invalid and 845 // should be diagnosed. 846 void Sema::CUDACheckLambdaCapture(CXXMethodDecl *Callee, 847 const sema::Capture &Capture) { 848 // In host compilation we only need to check lambda functions emitted on host 849 // side. In such lambda functions, a reference capture is invalid only 850 // if the lambda structure is populated by a device function or kernel then 851 // is passed to and called by a host function. However that is impossible, 852 // since a device function or kernel can only call a device function, also a 853 // kernel cannot pass a lambda back to a host function since we cannot 854 // define a kernel argument type which can hold the lambda before the lambda 855 // itself is defined. 856 if (!LangOpts.CUDAIsDevice) 857 return; 858 859 // File-scope lambda can only do init captures for global variables, which 860 // results in passing by value for these global variables. 861 FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext); 862 if (!Caller) 863 return; 864 865 // In device compilation, we only need to check lambda functions which are 866 // emitted on device side. For such lambdas, a reference capture is invalid 867 // only if the lambda structure is populated by a host function then passed 868 // to and called in a device function or kernel. 869 bool CalleeIsDevice = Callee->hasAttr<CUDADeviceAttr>(); 870 bool CallerIsHost = 871 !Caller->hasAttr<CUDAGlobalAttr>() && !Caller->hasAttr<CUDADeviceAttr>(); 872 bool ShouldCheck = CalleeIsDevice && CallerIsHost; 873 if (!ShouldCheck || !Capture.isReferenceCapture()) 874 return; 875 auto DiagKind = SemaDiagnosticBuilder::K_Deferred; 876 if (Capture.isVariableCapture()) { 877 SemaDiagnosticBuilder(DiagKind, Capture.getLocation(), 878 diag::err_capture_bad_target, Callee, *this) 879 << Capture.getVariable(); 880 } else if (Capture.isThisCapture()) { 881 SemaDiagnosticBuilder(DiagKind, Capture.getLocation(), 882 diag::err_capture_bad_target_this_ptr, Callee, *this); 883 } 884 return; 885 } 886 887 void Sema::CUDASetLambdaAttrs(CXXMethodDecl *Method) { 888 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 889 if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>()) 890 return; 891 Method->addAttr(CUDADeviceAttr::CreateImplicit(Context)); 892 Method->addAttr(CUDAHostAttr::CreateImplicit(Context)); 893 } 894 895 void Sema::checkCUDATargetOverload(FunctionDecl *NewFD, 896 const LookupResult &Previous) { 897 assert(getLangOpts().CUDA && "Should only be called during CUDA compilation"); 898 CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD); 899 for (NamedDecl *OldND : Previous) { 900 FunctionDecl *OldFD = OldND->getAsFunction(); 901 if (!OldFD) 902 continue; 903 904 CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD); 905 // Don't allow HD and global functions to overload other functions with the 906 // same signature. We allow overloading based on CUDA attributes so that 907 // functions can have different implementations on the host and device, but 908 // HD/global functions "exist" in some sense on both the host and device, so 909 // should have the same implementation on both sides. 910 if (NewTarget != OldTarget && 911 ((NewTarget == CFT_HostDevice) || (OldTarget == CFT_HostDevice) || 912 (NewTarget == CFT_Global) || (OldTarget == CFT_Global)) && 913 !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false, 914 /* ConsiderCudaAttrs = */ false)) { 915 Diag(NewFD->getLocation(), diag::err_cuda_ovl_target) 916 << NewTarget << NewFD->getDeclName() << OldTarget << OldFD; 917 Diag(OldFD->getLocation(), diag::note_previous_declaration); 918 NewFD->setInvalidDecl(); 919 break; 920 } 921 } 922 } 923 924 template <typename AttrTy> 925 static void copyAttrIfPresent(Sema &S, FunctionDecl *FD, 926 const FunctionDecl &TemplateFD) { 927 if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) { 928 AttrTy *Clone = Attribute->clone(S.Context); 929 Clone->setInherited(true); 930 FD->addAttr(Clone); 931 } 932 } 933 934 void Sema::inheritCUDATargetAttrs(FunctionDecl *FD, 935 const FunctionTemplateDecl &TD) { 936 const FunctionDecl &TemplateFD = *TD.getTemplatedDecl(); 937 copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD); 938 copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD); 939 copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD); 940 } 941 942 std::string Sema::getCudaConfigureFuncName() const { 943 if (getLangOpts().HIP) 944 return getLangOpts().HIPUseNewLaunchAPI ? "__hipPushCallConfiguration" 945 : "hipConfigureCall"; 946 947 // New CUDA kernel launch sequence. 948 if (CudaFeatureEnabled(Context.getTargetInfo().getSDKVersion(), 949 CudaFeature::CUDA_USES_NEW_LAUNCH)) 950 return "__cudaPushCallConfiguration"; 951 952 // Legacy CUDA kernel configuration call 953 return "cudaConfigureCall"; 954 } 955