Coverage Report

Created: 2022-01-18 06:27

/Users/buildslave/jenkins/workspace/coverage/llvm-project/clang/lib/Sema/SemaCUDA.cpp
Line
Count
Source (jump to first uncovered line)
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
81
template <typename AttrT> static bool hasExplicitAttr(const VarDecl *D) {
30
81
  if (!D)
31
0
    return false;
32
81
  if (auto *A = D->getAttr<AttrT>())
33
54
    return !A->isImplicit();
34
27
  return false;
35
81
}
36
37
7
void Sema::PushForceCUDAHostDevice() {
38
7
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
39
0
  ForceCUDAHostDeviceDepth++;
40
7
}
41
42
5
bool Sema::PopForceCUDAHostDevice() {
43
5
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
44
5
  if (ForceCUDAHostDeviceDepth == 0)
45
0
    return false;
46
5
  ForceCUDAHostDeviceDepth--;
47
5
  return true;
48
5
}
49
50
ExprResult Sema::ActOnCUDAExecConfigExpr(Scope *S, SourceLocation LLLLoc,
51
                                         MultiExprArg ExecConfig,
52
159
                                         SourceLocation GGGLoc) {
53
159
  FunctionDecl *ConfigDecl = Context.getcudaConfigureCallDecl();
54
159
  if (!ConfigDecl)
55
0
    return ExprError(Diag(LLLLoc, diag::err_undeclared_var_use)
56
0
                     << getCudaConfigureFuncName());
57
159
  QualType ConfigQTy = ConfigDecl->getType();
58
59
159
  DeclRefExpr *ConfigDR = new (Context)
60
159
      DeclRefExpr(Context, ConfigDecl, false, ConfigQTy, VK_LValue, LLLLoc);
61
159
  MarkFunctionReferenced(LLLLoc, ConfigDecl);
62
63
159
  return BuildCallExpr(S, ConfigDR, LLLLoc, ExecConfig, GGGLoc, nullptr,
64
159
                       /*IsExecConfig=*/true);
65
159
}
66
67
Sema::CUDAFunctionTarget
68
30
Sema::IdentifyCUDATarget(const ParsedAttributesView &Attrs) {
69
30
  bool HasHostAttr = false;
70
30
  bool HasDeviceAttr = false;
71
30
  bool HasGlobalAttr = false;
72
30
  bool HasInvalidTargetAttr = false;
73
31
  for (const ParsedAttr &AL : Attrs) {
74
31
    switch (AL.getKind()) {
75
11
    case ParsedAttr::AT_CUDAGlobal:
76
11
      HasGlobalAttr = true;
77
11
      break;
78
10
    case ParsedAttr::AT_CUDAHost:
79
10
      HasHostAttr = true;
80
10
      break;
81
10
    case ParsedAttr::AT_CUDADevice:
82
10
      HasDeviceAttr = true;
83
10
      break;
84
0
    case ParsedAttr::AT_CUDAInvalidTarget:
85
0
      HasInvalidTargetAttr = true;
86
0
      break;
87
0
    default:
88
0
      break;
89
31
    }
90
31
  }
91
92
30
  if (HasInvalidTargetAttr)
93
0
    return CFT_InvalidTarget;
94
95
30
  if (HasGlobalAttr)
96
11
    return CFT_Global;
97
98
19
  if (HasHostAttr && 
HasDeviceAttr10
)
99
6
    return CFT_HostDevice;
100
101
13
  if (HasDeviceAttr)
102
4
    return CFT_Device;
103
104
9
  return CFT_Host;
105
13
}
106
107
template <typename A>
108
103k
static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
109
103k
  return D->hasAttrs() && 
llvm::any_of(D->getAttrs(), [&](Attr *Attribute) 75.3k
{
110
101k
           return isa<A>(Attribute) &&
111
101k
                  
!(53.2k
IgnoreImplicitAttr53.2k
&&
Attribute->isImplicit()76
);
112
101k
         });
SemaCUDA.cpp:bool hasAttr<clang::CUDADeviceAttr>(clang::FunctionDecl const*, bool)::'lambda'(clang::Attr*)::operator()(clang::Attr*) const
Line
Count
Source
109
49.8k
  return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
110
49.8k
           return isa<A>(Attribute) &&
111
49.8k
                  
!(32.2k
IgnoreImplicitAttr32.2k
&&
Attribute->isImplicit()38
);
112
49.8k
         });
SemaCUDA.cpp:bool hasAttr<clang::CUDAHostAttr>(clang::FunctionDecl const*, bool)::'lambda'(clang::Attr*)::operator()(clang::Attr*) const
Line
Count
Source
109
51.4k
  return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
110
51.4k
           return isa<A>(Attribute) &&
111
51.4k
                  
!(21.0k
IgnoreImplicitAttr21.0k
&&
Attribute->isImplicit()38
);
112
51.4k
         });
113
103k
}
SemaCUDA.cpp:bool hasAttr<clang::CUDADeviceAttr>(clang::FunctionDecl const*, bool)
Line
Count
Source
108
51.7k
static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
109
51.7k
  return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
110
37.6k
           return isa<A>(Attribute) &&
111
37.6k
                  !(IgnoreImplicitAttr && Attribute->isImplicit());
112
37.6k
         });
113
51.7k
}
SemaCUDA.cpp:bool hasAttr<clang::CUDAHostAttr>(clang::FunctionDecl const*, bool)
Line
Count
Source
108
51.7k
static bool hasAttr(const FunctionDecl *D, bool IgnoreImplicitAttr) {
109
51.7k
  return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {
110
37.6k
           return isa<A>(Attribute) &&
111
37.6k
                  !(IgnoreImplicitAttr && Attribute->isImplicit());
112
37.6k
         });
113
51.7k
}
114
115
/// IdentifyCUDATarget - Determine the CUDA compilation target for this function
116
Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D,
117
54.5k
                                                  bool IgnoreImplicitHDAttr) {
118
  // Code that lives outside a function is run on the host.
119
54.5k
  if (D == nullptr)
120
185
    return CFT_Host;
121
122
54.3k
  if (D->hasAttr<CUDAInvalidTargetAttr>())
123
39
    return CFT_InvalidTarget;
124
125
54.3k
  if (D->hasAttr<CUDAGlobalAttr>())
126
2.57k
    return CFT_Global;
127
128
51.7k
  if (hasAttr<CUDADeviceAttr>(D, IgnoreImplicitHDAttr)) {
129
32.1k
    if (hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr))
130
16.1k
      return CFT_HostDevice;
131
16.0k
    return CFT_Device;
132
32.1k
  } else 
if (19.5k
hasAttr<CUDAHostAttr>(D, IgnoreImplicitHDAttr)19.5k
) {
133
4.88k
    return CFT_Host;
134
14.6k
  } else if ((D->isImplicit() || 
!D->isUserProvided()14.1k
) &&
135
14.6k
             
!IgnoreImplicitHDAttr578
) {
136
    // Some implicit declarations (like intrinsic functions) are not marked.
137
    // Set the most lenient target on them for maximal flexibility.
138
578
    return CFT_HostDevice;
139
578
  }
140
141
14.0k
  return CFT_Host;
142
51.7k
}
143
144
/// IdentifyTarget - Determine the CUDA compilation target for this variable.
145
501
Sema::CUDAVariableTarget Sema::IdentifyCUDATarget(const VarDecl *Var) {
146
501
  if (Var->hasAttr<HIPManagedAttr>())
147
8
    return CVT_Unified;
148
493
  if (Var->isConstexpr() && 
!hasExplicitAttr<CUDAConstantAttr>(Var)54
)
149
54
    return CVT_Both;
150
439
  if (Var->getType().isConstQualified() && 
Var->hasAttr<CUDAConstantAttr>()92
&&
151
439
      
!hasExplicitAttr<CUDAConstantAttr>(Var)27
)
152
27
    return CVT_Both;
153
412
  if (Var->hasAttr<CUDADeviceAttr>() || 
Var->hasAttr<CUDAConstantAttr>()283
||
154
412
      
Var->hasAttr<CUDASharedAttr>()219
||
155
412
      
Var->getType()->isCUDADeviceBuiltinSurfaceType()153
||
156
412
      
Var->getType()->isCUDADeviceBuiltinTextureType()153
)
157
259
    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
153
  if (auto *FD = dyn_cast<FunctionDecl>(Var->getDeclContext())) {
163
12
    switch (IdentifyCUDATarget(FD)) {
164
4
    case CFT_HostDevice:
165
4
      return CVT_Both;
166
6
    case CFT_Device:
167
6
    case CFT_Global:
168
6
      return CVT_Device;
169
2
    default:
170
2
      return CVT_Host;
171
12
    }
172
12
  }
173
141
  return CVT_Host;
174
153
}
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
16.6k
                             const FunctionDecl *Callee) {
207
16.6k
  assert(Callee && "Callee must be valid.");
208
0
  CUDAFunctionTarget CallerTarget = IdentifyCUDATarget(Caller);
209
16.6k
  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
16.6k
  if (CallerTarget == CFT_InvalidTarget || CalleeTarget == CFT_InvalidTarget)
214
7
    return CFP_Never;
215
216
  // (a) Can't call global from some contexts until we support CUDA's
217
  // dynamic parallelism.
218
16.6k
  if (CalleeTarget == CFT_Global &&
219
16.6k
      
(615
CallerTarget == CFT_Global615
||
CallerTarget == CFT_Device614
))
220
7
    return CFP_Never;
221
222
  // (b) Calling HostDevice is OK for everyone.
223
16.6k
  if (CalleeTarget == CFT_HostDevice)
224
9.51k
    return CFP_HostDevice;
225
226
  // (c) Best case scenarios
227
7.13k
  if (CalleeTarget == CallerTarget ||
228
7.13k
      
(2.35k
CallerTarget == CFT_Host2.35k
&&
CalleeTarget == CFT_Global682
) ||
229
7.13k
      
(1.76k
CallerTarget == CFT_Global1.76k
&&
CalleeTarget == CFT_Device328
))
230
5.69k
    return CFP_Native;
231
232
  // (d) HostDevice behavior depends on compilation mode.
233
1.43k
  if (CallerTarget == CFT_HostDevice) {
234
    // It's OK to call a compilation-mode matching function from an HD one.
235
1.23k
    if ((getLangOpts().CUDAIsDevice && 
CalleeTarget == CFT_Device593
) ||
236
1.23k
        
(959
!getLangOpts().CUDAIsDevice959
&&
237
959
         
(639
CalleeTarget == CFT_Host639
||
CalleeTarget == CFT_Global277
)))
238
651
      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
581
    return CFP_WrongSide;
245
1.23k
  }
246
247
  // (e) Calling across device/host boundary is not something you should do.
248
206
  if ((CallerTarget == CFT_Host && 
CalleeTarget == CFT_Device95
) ||
249
206
      
(111
CallerTarget == CFT_Device111
&&
CalleeTarget == CFT_Host109
) ||
250
206
      
(2
CallerTarget == CFT_Global2
&&
CalleeTarget == CFT_Host2
))
251
206
    return CFP_Never;
252
253
0
  llvm_unreachable("All cases should've been handled by now.");
254
0
}
255
256
0
template <typename AttrT> static bool hasImplicitAttr(const FunctionDecl *D) {
257
0
  if (!D)
258
0
    return false;
259
0
  if (auto *A = D->getAttr<AttrT>())
260
0
    return A->isImplicit();
261
0
  return D->isImplicit();
262
0
}
Unexecuted instantiation: SemaCUDA.cpp:bool hasImplicitAttr<clang::CUDADeviceAttr>(clang::FunctionDecl const*)
Unexecuted instantiation: SemaCUDA.cpp:bool hasImplicitAttr<clang::CUDAHostAttr>(clang::FunctionDecl const*)
263
264
0
bool Sema::isCUDAImplicitHostDeviceFunction(const FunctionDecl *D) {
265
0
  bool IsImplicitDevAttr = hasImplicitAttr<CUDADeviceAttr>(D);
266
0
  bool IsImplicitHostAttr = hasImplicitAttr<CUDAHostAttr>(D);
267
0
  return IsImplicitDevAttr && IsImplicitHostAttr;
268
0
}
269
270
void Sema::EraseUnwantedCUDAMatches(
271
    const FunctionDecl *Caller,
272
4
    SmallVectorImpl<std::pair<DeclAccessPair, FunctionDecl *>> &Matches) {
273
4
  if (Matches.size() <= 1)
274
0
    return;
275
276
4
  using Pair = std::pair<DeclAccessPair, FunctionDecl*>;
277
278
  // Gets the CUDA function preference for a call from Caller to Match.
279
20
  auto GetCFP = [&](const Pair &Match) {
280
20
    return IdentifyCUDAPreference(Caller, Match.second);
281
20
  };
282
283
  // Find the best call preference among the functions in Matches.
284
4
  CUDAFunctionPreference BestCFP = GetCFP(*std::max_element(
285
4
      Matches.begin(), Matches.end(),
286
4
      [&](const Pair &M1, const Pair &M2) { return GetCFP(M1) < GetCFP(M2); }));
287
288
  // Erase all functions with lower priority.
289
4
  llvm::erase_if(Matches,
290
8
                 [&](const Pair &Match) { return GetCFP(Match) < BestCFP; });
291
4
}
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
67
                                Sema::CUDAFunctionTarget *ResolvedTarget) {
305
  // Only free functions and static member functions may be global.
306
67
  assert(Target1 != Sema::CFT_Global);
307
0
  assert(Target2 != Sema::CFT_Global);
308
309
67
  if (Target1 == Sema::CFT_HostDevice) {
310
37
    *ResolvedTarget = Target2;
311
37
  } else 
if (30
Target2 == Sema::CFT_HostDevice30
) {
312
1
    *ResolvedTarget = Target1;
313
29
  } else if (Target1 != Target2) {
314
25
    return true;
315
25
  } else {
316
4
    *ResolvedTarget = Target1;
317
4
  }
318
319
42
  return false;
320
67
}
321
322
bool Sema::inferCUDATargetForImplicitSpecialMember(CXXRecordDecl *ClassDecl,
323
                                                   CXXSpecialMember CSM,
324
                                                   CXXMethodDecl *MemberDecl,
325
                                                   bool ConstRHS,
326
2.31k
                                                   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
2.31k
  bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent();
331
2.31k
  bool HasH = MemberDecl->hasAttr<CUDAHostAttr>();
332
2.31k
  bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>();
333
2.31k
  bool HasExplicitAttr =
334
2.31k
      (HasD && 
!MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()1.12k
) ||
335
2.31k
      
(2.31k
HasH2.31k
&&
!MemberDecl->getAttr<CUDAHostAttr>()->isImplicit()1.11k
);
336
2.31k
  if (!InClass || 
HasExplicitAttr2.31k
)
337
14
    return false;
338
339
2.30k
  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
2.30k
  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
2.30k
  llvm::SmallVector<const CXXBaseSpecifier *, 16> Bases;
349
2.30k
  for (const auto &B : ClassDecl->bases()) {
350
225
    if (!B.isVirtual()) {
351
225
      Bases.push_back(&B);
352
225
    }
353
225
  }
354
355
2.30k
  if (!ClassDecl->isAbstract()) {
356
2.30k
    for (const auto &VB : ClassDecl->vbases()) {
357
0
      Bases.push_back(&VB);
358
0
    }
359
2.30k
  }
360
361
2.30k
  for (const auto *B : Bases) {
362
225
    const RecordType *BaseType = B->getType()->getAs<RecordType>();
363
225
    if (!BaseType) {
364
0
      continue;
365
0
    }
366
367
225
    CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());
368
225
    Sema::SpecialMemberOverloadResult SMOR =
369
225
        LookupSpecialMember(BaseClassDecl, CSM,
370
225
                            /* ConstArg */ ConstRHS,
371
225
                            /* VolatileArg */ false,
372
225
                            /* RValueThis */ false,
373
225
                            /* ConstThis */ false,
374
225
                            /* VolatileThis */ false);
375
376
225
    if (!SMOR.getMethod())
377
5
      continue;
378
379
220
    CUDAFunctionTarget BaseMethodTarget = IdentifyCUDATarget(SMOR.getMethod());
380
220
    if (!InferredTarget.hasValue()) {
381
190
      InferredTarget = BaseMethodTarget;
382
190
    } else {
383
30
      bool ResolutionError = resolveCalleeCUDATargetConflict(
384
30
          InferredTarget.getValue(), BaseMethodTarget,
385
30
          InferredTarget.getPointer());
386
30
      if (ResolutionError) {
387
13
        if (Diagnose) {
388
3
          Diag(ClassDecl->getLocation(),
389
3
               diag::note_implicit_member_target_infer_collision)
390
3
              << (unsigned)CSM << InferredTarget.getValue() << BaseMethodTarget;
391
3
        }
392
13
        MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
393
13
        return true;
394
13
      }
395
30
    }
396
220
  }
397
398
  // Same as for bases, but now for special members of fields.
399
2.29k
  for (const auto *F : ClassDecl->fields()) {
400
1.62k
    if (F->isInvalidDecl()) {
401
0
      continue;
402
0
    }
403
404
1.62k
    const RecordType *FieldType =
405
1.62k
        Context.getBaseElementType(F->getType())->getAs<RecordType>();
406
1.62k
    if (!FieldType) {
407
1.46k
      continue;
408
1.46k
    }
409
410
166
    CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());
411
166
    Sema::SpecialMemberOverloadResult SMOR =
412
166
        LookupSpecialMember(FieldRecDecl, CSM,
413
166
                            /* ConstArg */ ConstRHS && 
!F->isMutable()66
,
414
166
                            /* VolatileArg */ false,
415
166
                            /* RValueThis */ false,
416
166
                            /* ConstThis */ false,
417
166
                            /* VolatileThis */ false);
418
419
166
    if (!SMOR.getMethod())
420
0
      continue;
421
422
166
    CUDAFunctionTarget FieldMethodTarget =
423
166
        IdentifyCUDATarget(SMOR.getMethod());
424
166
    if (!InferredTarget.hasValue()) {
425
129
      InferredTarget = FieldMethodTarget;
426
129
    } else {
427
37
      bool ResolutionError = resolveCalleeCUDATargetConflict(
428
37
          InferredTarget.getValue(), FieldMethodTarget,
429
37
          InferredTarget.getPointer());
430
37
      if (ResolutionError) {
431
12
        if (Diagnose) {
432
4
          Diag(ClassDecl->getLocation(),
433
4
               diag::note_implicit_member_target_infer_collision)
434
4
              << (unsigned)CSM << InferredTarget.getValue()
435
4
              << FieldMethodTarget;
436
4
        }
437
12
        MemberDecl->addAttr(CUDAInvalidTargetAttr::CreateImplicit(Context));
438
12
        return true;
439
12
      }
440
37
    }
441
166
  }
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
2.28k
  bool NeedsH = true, NeedsD = true;
447
2.28k
  if (InferredTarget.hasValue()) {
448
294
    if (InferredTarget.getValue() == CFT_Device)
449
58
      NeedsH = false;
450
236
    else if (InferredTarget.getValue() == CFT_Host)
451
65
      NeedsD = false;
452
294
  }
453
454
  // We either setting attributes first time, or the inferred ones must match
455
  // previously set ones.
456
2.28k
  if (NeedsD && 
!HasD2.21k
)
457
1.10k
    MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(Context));
458
2.28k
  if (NeedsH && 
!HasH2.22k
)
459
1.11k
    MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(Context));
460
461
2.28k
  return false;
462
2.29k
}
463
464
19
bool Sema::isEmptyCudaConstructor(SourceLocation Loc, CXXConstructorDecl *CD) {
465
19
  if (!CD->isDefined() && 
CD->isTemplateInstantiation()0
)
466
0
    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
19
  if (CD->isTrivial())
472
5
    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
14
  if (!(CD->hasTrivialBody() && 
CD->getNumParams() == 05
))
479
9
    return false;
480
481
  // Its class has no virtual functions and no virtual base classes.
482
5
  if (CD->getParent()->isDynamicClass())
483
0
    return false;
484
485
  // Union ctor does not call ctors of its data members.
486
5
  if (CD->getParent()->isUnion())
487
4
    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
1
  if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {
492
0
        if (const CXXConstructExpr *CE =
493
0
                dyn_cast<CXXConstructExpr>(CI->getInit()))
494
0
          return isEmptyCudaConstructor(Loc, CE->getConstructor());
495
0
        return false;
496
0
      }))
497
0
    return false;
498
499
1
  return true;
500
1
}
501
502
19
bool Sema::isEmptyCudaDestructor(SourceLocation Loc, CXXDestructorDecl *DD) {
503
  // No destructor -> no problem.
504
19
  if (!DD)
505
10
    return true;
506
507
9
  if (!DD->isDefined() && 
DD->isTemplateInstantiation()0
)
508
0
    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
9
  if (DD->isTrivial())
514
4
    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
5
  if (!DD->hasTrivialBody())
520
3
    return false;
521
522
2
  const CXXRecordDecl *ClassDecl = DD->getParent();
523
524
  // Its class has no virtual functions and no virtual base classes.
525
2
  if (ClassDecl->isDynamicClass())
526
0
    return false;
527
528
  // Union does not have base class and union dtor does not call dtors of its
529
  // data members.
530
2
  if (DD->getParent()->isUnion())
531
2
    return true;
532
533
  // Only empty destructors are allowed. This will recursively check
534
  // destructors for all base classes...
535
0
  if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {
536
0
        if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())
537
0
          return isEmptyCudaDestructor(Loc, RD->getDestructor());
538
0
        return true;
539
0
      }))
540
0
    return false;
541
542
  // ... and member fields.
543
0
  if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {
544
0
        if (CXXRecordDecl *RD = Field->getType()
545
0
                                    ->getBaseElementTypeUnsafe()
546
0
                                    ->getAsCXXRecordDecl())
547
0
          return isEmptyCudaDestructor(Loc, RD->getDestructor());
548
0
        return true;
549
0
      }))
550
0
    return false;
551
552
0
  return true;
553
0
}
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
503
bool IsDependentVar(VarDecl *VD) {
562
503
  if (VD->getType()->isDependentType())
563
22
    return true;
564
481
  if (const auto *Init = VD->getInit())
565
409
    return Init->isValueDependent();
566
72
  return false;
567
481
}
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
144
                                           CUDAInitializerCheckKind CheckKind) {
580
144
  assert(!VD->isInvalidDecl() && VD->hasGlobalStorage());
581
0
  assert(!IsDependentVar(VD) && "do not check dependent var");
582
0
  const Expr *Init = VD->getInit();
583
144
  auto IsEmptyInit = [&](const Expr *Init) {
584
144
    if (!Init)
585
0
      return true;
586
144
    if (const auto *CE = dyn_cast<CXXConstructExpr>(Init)) {
587
19
      return S.isEmptyCudaConstructor(VD->getLocation(), CE->getConstructor());
588
19
    }
589
125
    return false;
590
144
  };
591
144
  auto IsConstantInit = [&](const Expr *Init) {
592
133
    assert(Init);
593
0
    return Init->isConstantInitializer(S.Context,
594
133
                                       VD->getType()->isReferenceType());
595
133
  };
596
144
  auto HasEmptyDtor = [&](VarDecl *VD) {
597
130
    if (const auto *RD = VD->getType()->getAsCXXRecordDecl())
598
19
      return S.isEmptyCudaDestructor(VD->getLocation(), RD->getDestructor());
599
111
    return true;
600
130
  };
601
144
  if (CheckKind == CICK_Shared)
602
5
    return IsEmptyInit(Init) && 
HasEmptyDtor(VD)4
;
603
139
  return S.LangOpts.GPUAllowDeviceInit ||
604
139
         ((IsEmptyInit(Init) || 
IsConstantInit(Init)133
) &&
HasEmptyDtor(VD)126
);
605
144
}
606
} // namespace
607
608
1.28k
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
1.28k
  if (VD->isInvalidDecl() || 
!VD->hasInit()1.28k
||
!VD->hasGlobalStorage()778
||
612
1.28k
      
IsDependentVar(VD)213
)
613
1.09k
    return;
614
188
  const Expr *Init = VD->getInit();
615
188
  bool IsSharedVar = VD->hasAttr<CUDASharedAttr>();
616
188
  bool IsDeviceOrConstantVar =
617
188
      !IsSharedVar &&
618
188
      
(183
VD->hasAttr<CUDADeviceAttr>()183
||
VD->hasAttr<CUDAConstantAttr>()146
);
619
188
  if (IsDeviceOrConstantVar || 
IsSharedVar73
) {
620
120
    if (HasAllowedCUDADeviceStaticInitializer(
621
120
            *this, VD, IsSharedVar ? 
CICK_Shared5
:
CICK_DeviceOrConstant115
))
622
110
      return;
623
10
    Diag(VD->getLocation(),
624
10
         IsSharedVar ? 
diag::err_shared_var_init2
:
diag::err_dynamic_var_init8
)
625
10
        << Init->getSourceRange();
626
10
    VD->setInvalidDecl();
627
68
  } else {
628
    // This is a host-side global variable.  Check that the initializer is
629
    // callable from the host side.
630
68
    const FunctionDecl *InitFn = nullptr;
631
68
    if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init)) {
632
17
      InitFn = CE->getConstructor();
633
51
    } else if (const CallExpr *CE = dyn_cast<CallExpr>(Init)) {
634
3
      InitFn = CE->getDirectCallee();
635
3
    }
636
68
    if (InitFn) {
637
20
      CUDAFunctionTarget InitFnTarget = IdentifyCUDATarget(InitFn);
638
20
      if (InitFnTarget != CFT_Host && 
InitFnTarget != CFT_HostDevice12
) {
639
3
        Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer)
640
3
            << InitFnTarget << InitFn;
641
3
        Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn;
642
3
        VD->setInvalidDecl();
643
3
      }
644
20
    }
645
68
  }
646
188
}
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
3.01k
                                       const LookupResult &Previous) {
662
3.01k
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
663
664
3.01k
  if (ForceCUDAHostDeviceDepth > 0) {
665
7
    if (!NewD->hasAttr<CUDAHostAttr>())
666
5
      NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
667
7
    if (!NewD->hasAttr<CUDADeviceAttr>())
668
5
      NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
669
7
    return;
670
7
  }
671
672
3.00k
  if (!getLangOpts().CUDAHostDeviceConstexpr || 
!NewD->isConstexpr()2.98k
||
673
3.00k
      
NewD->isVariadic()33
||
NewD->hasAttr<CUDAHostAttr>()29
||
674
3.00k
      
NewD->hasAttr<CUDADeviceAttr>()25
||
NewD->hasAttr<CUDAGlobalAttr>()21
)
675
2.98k
    return;
676
677
  // Is D a __device__ function with the same signature as NewD, ignoring CUDA
678
  // attributes?
679
21
  auto IsMatchingDeviceFn = [&](NamedDecl *D) {
680
12
    if (UsingShadowDecl *Using = dyn_cast<UsingShadowDecl>(D))
681
4
      D = Using->getTargetDecl();
682
12
    FunctionDecl *OldD = D->getAsFunction();
683
12
    return OldD && OldD->hasAttr<CUDADeviceAttr>() &&
684
12
           !OldD->hasAttr<CUDAHostAttr>() &&
685
12
           !IsOverload(NewD, OldD, /* UseMemberUsingDeclRules = */ false,
686
12
                       /* ConsiderCudaAttrs = */ false);
687
12
  };
688
21
  auto It = llvm::find_if(Previous, IsMatchingDeviceFn);
689
21
  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
12
    NamedDecl *Match = *It;
695
12
    if (!getSourceManager().isInSystemHeader(Match->getLocation())) {
696
4
      Diag(NewD->getLocation(),
697
4
           diag::err_cuda_unattributed_constexpr_cannot_overload_device)
698
4
          << NewD;
699
4
      Diag(Match->getLocation(),
700
4
           diag::note_cuda_conflicting_device_function_declared_here);
701
4
    }
702
12
    return;
703
12
  }
704
705
9
  NewD->addAttr(CUDAHostAttr::CreateImplicit(Context));
706
9
  NewD->addAttr(CUDADeviceAttr::CreateImplicit(Context));
707
9
}
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
2.38M
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
2.38M
  if (getLangOpts().CUDAIsDevice && 
!VD->hasAttr<CUDAConstantAttr>()544
&&
716
2.38M
      
!VD->hasAttr<CUDAConstantAttr>()500
&&
!VD->hasAttr<CUDASharedAttr>()500
&&
717
2.38M
      
(464
VD->isFileVarDecl()464
||
VD->isStaticDataMember()318
) &&
718
2.38M
      
!IsDependentVar(VD)146
&&
719
2.38M
      
(138
VD->isConstexpr()138
||
(110
VD->getType().isConstQualified()110
&&
720
110
                             HasAllowedCUDADeviceStaticInitializer(
721
45
                                 *this, VD, CICK_DeviceOrConstant)))) {
722
45
    VD->addAttr(CUDAConstantAttr::CreateImplicit(getASTContext()));
723
45
  }
724
2.38M
}
725
726
Sema::SemaDiagnosticBuilder Sema::CUDADiagIfDeviceCode(SourceLocation Loc,
727
98
                                                       unsigned DiagID) {
728
98
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
729
98
  SemaDiagnosticBuilder::Kind DiagKind = [&] {
730
98
    if (!isa<FunctionDecl>(CurContext))
731
3
      return SemaDiagnosticBuilder::K_Nop;
732
95
    switch (CurrentCUDATarget()) {
733
8
    case CFT_Global:
734
52
    case CFT_Device:
735
52
      return SemaDiagnosticBuilder::K_Immediate;
736
34
    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
34
      if (!getLangOpts().CUDAIsDevice)
741
6
        return SemaDiagnosticBuilder::K_Nop;
742
28
      if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID))
743
7
        return SemaDiagnosticBuilder::K_Immediate;
744
21
      return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
745
21
              FunctionEmissionStatus::Emitted)
746
21
                 ? 
SemaDiagnosticBuilder::K_ImmediateWithCallStack5
747
21
                 : 
SemaDiagnosticBuilder::K_Deferred16
;
748
9
    default:
749
9
      return SemaDiagnosticBuilder::K_Nop;
750
95
    }
751
95
  }();
752
98
  return SemaDiagnosticBuilder(DiagKind, Loc, DiagID,
753
98
                               dyn_cast<FunctionDecl>(CurContext), *this);
754
98
}
755
756
Sema::SemaDiagnosticBuilder Sema::CUDADiagIfHostCode(SourceLocation Loc,
757
35
                                                     unsigned DiagID) {
758
35
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
759
35
  SemaDiagnosticBuilder::Kind DiagKind = [&] {
760
35
    if (!isa<FunctionDecl>(CurContext))
761
0
      return SemaDiagnosticBuilder::K_Nop;
762
35
    switch (CurrentCUDATarget()) {
763
6
    case CFT_Host:
764
6
      return SemaDiagnosticBuilder::K_Immediate;
765
1
    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
1
      if (getLangOpts().CUDAIsDevice)
770
0
        return SemaDiagnosticBuilder::K_Nop;
771
1
      if (IsLastErrorImmediate && Diags.getDiagnosticIDs()->isBuiltinNote(DiagID))
772
0
        return SemaDiagnosticBuilder::K_Immediate;
773
1
      return (getEmissionStatus(cast<FunctionDecl>(CurContext)) ==
774
1
              FunctionEmissionStatus::Emitted)
775
1
                 ? 
SemaDiagnosticBuilder::K_ImmediateWithCallStack0
776
1
                 : SemaDiagnosticBuilder::K_Deferred;
777
28
    default:
778
28
      return SemaDiagnosticBuilder::K_Nop;
779
35
    }
780
35
  }();
781
35
  return SemaDiagnosticBuilder(DiagKind, Loc, DiagID,
782
35
                               dyn_cast<FunctionDecl>(CurContext), *this);
783
35
}
784
785
10.0k
bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) {
786
10.0k
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
787
0
  assert(Callee && "Callee may not be null.");
788
789
0
  auto &ExprEvalCtx = ExprEvalContexts.back();
790
10.0k
  if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated())
791
27
    return true;
792
793
  // FIXME: Is bailing out early correct here?  Should we instead assume that
794
  // the caller is a global initializer?
795
10.0k
  FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
796
10.0k
  if (!Caller)
797
152
    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
9.91k
  bool CallerKnownEmitted =
802
9.91k
      getEmissionStatus(Caller) == FunctionEmissionStatus::Emitted;
803
9.91k
  SemaDiagnosticBuilder::Kind DiagKind = [this, Caller, Callee,
804
9.91k
                                          CallerKnownEmitted] {
805
9.91k
    switch (IdentifyCUDAPreference(Caller, Callee)) {
806
20
    case CFP_Never:
807
142
    case CFP_WrongSide:
808
142
      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
142
      return CallerKnownEmitted
813
142
                 ? 
SemaDiagnosticBuilder::K_ImmediateWithCallStack58
814
142
                 : 
SemaDiagnosticBuilder::K_Deferred84
;
815
9.77k
    default:
816
9.77k
      return SemaDiagnosticBuilder::K_Nop;
817
9.91k
    }
818
9.91k
  }();
819
820
9.91k
  if (DiagKind == SemaDiagnosticBuilder::K_Nop)
821
9.77k
    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
142
  if (!LocsWithCUDACallDiags.insert({Caller, Loc}).second)
828
58
    return true;
829
830
84
  SemaDiagnosticBuilder(DiagKind, Loc, diag::err_ref_bad_target, Caller, *this)
831
84
      << IdentifyCUDATarget(Callee) << /*function*/ 0 << Callee
832
84
      << IdentifyCUDATarget(Caller);
833
84
  if (!Callee->getBuiltinID())
834
84
    SemaDiagnosticBuilder(DiagKind, Callee->getLocation(),
835
84
                          diag::note_previous_decl, Caller, *this)
836
84
        << Callee;
837
84
  return DiagKind != SemaDiagnosticBuilder::K_Immediate &&
838
84
         DiagKind != SemaDiagnosticBuilder::K_ImmediateWithCallStack;
839
142
}
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
98
                                  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
98
  if (!LangOpts.CUDAIsDevice)
857
32
    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
66
  FunctionDecl *Caller = dyn_cast<FunctionDecl>(CurContext);
862
66
  if (!Caller)
863
0
    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
66
  bool CalleeIsDevice = Callee->hasAttr<CUDADeviceAttr>();
870
66
  bool CallerIsHost =
871
66
      !Caller->hasAttr<CUDAGlobalAttr>() && !Caller->hasAttr<CUDADeviceAttr>();
872
66
  bool ShouldCheck = CalleeIsDevice && CallerIsHost;
873
66
  if (!ShouldCheck || 
!Capture.isReferenceCapture()56
)
874
26
    return;
875
40
  auto DiagKind = SemaDiagnosticBuilder::K_Deferred;
876
40
  if (Capture.isVariableCapture()) {
877
19
    SemaDiagnosticBuilder(DiagKind, Capture.getLocation(),
878
19
                          diag::err_capture_bad_target, Callee, *this)
879
19
        << Capture.getVariable();
880
21
  } else if (Capture.isThisCapture()) {
881
    // Capture of this pointer is allowed since this pointer may be pointing to
882
    // managed memory which is accessible on both device and host sides. It only
883
    // results in invalid memory access if this pointer points to memory not
884
    // accessible on device side.
885
21
    SemaDiagnosticBuilder(DiagKind, Capture.getLocation(),
886
21
                          diag::warn_maybe_capture_bad_target_this_ptr, Callee,
887
21
                          *this);
888
21
  }
889
40
}
890
891
214
void Sema::CUDASetLambdaAttrs(CXXMethodDecl *Method) {
892
214
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
893
214
  if (Method->hasAttr<CUDAHostAttr>() || 
Method->hasAttr<CUDADeviceAttr>()174
)
894
90
    return;
895
124
  Method->addAttr(CUDADeviceAttr::CreateImplicit(Context));
896
124
  Method->addAttr(CUDAHostAttr::CreateImplicit(Context));
897
124
}
898
899
void Sema::checkCUDATargetOverload(FunctionDecl *NewFD,
900
4.38k
                                   const LookupResult &Previous) {
901
4.38k
  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
902
0
  CUDAFunctionTarget NewTarget = IdentifyCUDATarget(NewFD);
903
4.38k
  for (NamedDecl *OldND : Previous) {
904
2.13k
    FunctionDecl *OldFD = OldND->getAsFunction();
905
2.13k
    if (!OldFD)
906
82
      continue;
907
908
2.05k
    CUDAFunctionTarget OldTarget = IdentifyCUDATarget(OldFD);
909
    // Don't allow HD and global functions to overload other functions with the
910
    // same signature.  We allow overloading based on CUDA attributes so that
911
    // functions can have different implementations on the host and device, but
912
    // HD/global functions "exist" in some sense on both the host and device, so
913
    // should have the same implementation on both sides.
914
2.05k
    if (NewTarget != OldTarget &&
915
2.05k
        
(984
(NewTarget == CFT_HostDevice)984
||
(OldTarget == CFT_HostDevice)818
||
916
984
         
(NewTarget == CFT_Global)804
||
(OldTarget == CFT_Global)804
) &&
917
2.05k
        !IsOverload(NewFD, OldFD, /* UseMemberUsingDeclRules = */ false,
918
180
                    /* ConsiderCudaAttrs = */ false)) {
919
4
      Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)
920
4
          << NewTarget << NewFD->getDeclName() << OldTarget << OldFD;
921
4
      Diag(OldFD->getLocation(), diag::note_previous_declaration);
922
4
      NewFD->setInvalidDecl();
923
4
      break;
924
4
    }
925
2.05k
  }
926
4.38k
}
927
928
template <typename AttrTy>
929
static void copyAttrIfPresent(Sema &S, FunctionDecl *FD,
930
33
                              const FunctionDecl &TemplateFD) {
931
33
  if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
932
18
    AttrTy *Clone = Attribute->clone(S.Context);
933
18
    Clone->setInherited(true);
934
18
    FD->addAttr(Clone);
935
18
  }
936
33
}
SemaCUDA.cpp:void copyAttrIfPresent<clang::CUDAGlobalAttr>(clang::Sema&, clang::FunctionDecl*, clang::FunctionDecl const&)
Line
Count
Source
930
11
                              const FunctionDecl &TemplateFD) {
931
11
  if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
932
0
    AttrTy *Clone = Attribute->clone(S.Context);
933
0
    Clone->setInherited(true);
934
0
    FD->addAttr(Clone);
935
0
  }
936
11
}
SemaCUDA.cpp:void copyAttrIfPresent<clang::CUDAHostAttr>(clang::Sema&, clang::FunctionDecl*, clang::FunctionDecl const&)
Line
Count
Source
930
11
                              const FunctionDecl &TemplateFD) {
931
11
  if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
932
9
    AttrTy *Clone = Attribute->clone(S.Context);
933
9
    Clone->setInherited(true);
934
9
    FD->addAttr(Clone);
935
9
  }
936
11
}
SemaCUDA.cpp:void copyAttrIfPresent<clang::CUDADeviceAttr>(clang::Sema&, clang::FunctionDecl*, clang::FunctionDecl const&)
Line
Count
Source
930
11
                              const FunctionDecl &TemplateFD) {
931
11
  if (AttrTy *Attribute = TemplateFD.getAttr<AttrTy>()) {
932
9
    AttrTy *Clone = Attribute->clone(S.Context);
933
9
    Clone->setInherited(true);
934
9
    FD->addAttr(Clone);
935
9
  }
936
11
}
937
938
void Sema::inheritCUDATargetAttrs(FunctionDecl *FD,
939
11
                                  const FunctionTemplateDecl &TD) {
940
11
  const FunctionDecl &TemplateFD = *TD.getTemplatedDecl();
941
11
  copyAttrIfPresent<CUDAGlobalAttr>(*this, FD, TemplateFD);
942
11
  copyAttrIfPresent<CUDAHostAttr>(*this, FD, TemplateFD);
943
11
  copyAttrIfPresent<CUDADeviceAttr>(*this, FD, TemplateFD);
944
11
}
945
946
2.01k
std::string Sema::getCudaConfigureFuncName() const {
947
2.01k
  if (getLangOpts().HIP)
948
509
    return getLangOpts().HIPUseNewLaunchAPI ? 
"__hipPushCallConfiguration"6
949
509
                                            : 
"hipConfigureCall"503
;
950
951
  // New CUDA kernel launch sequence.
952
1.50k
  if (CudaFeatureEnabled(Context.getTargetInfo().getSDKVersion(),
953
1.50k
                         CudaFeature::CUDA_USES_NEW_LAUNCH))
954
71
    return "__cudaPushCallConfiguration";
955
956
  // Legacy CUDA kernel configuration call
957
1.43k
  return "cudaConfigureCall";
958
1.50k
}