clang  14.0.0git
CGCall.cpp
Go to the documentation of this file.
1 //===--- CGCall.cpp - Encapsulate calling convention details --------------===//
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 //
9 // These classes wrap the information about a call or function
10 // definition used to handle ABI compliancy.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CGCall.h"
15 #include "ABIInfo.h"
16 #include "CGBlocks.h"
17 #include "CGCXXABI.h"
18 #include "CGCleanup.h"
19 #include "CGRecordLayout.h"
20 #include "CodeGenFunction.h"
21 #include "CodeGenModule.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/Attr.h"
24 #include "clang/AST/Decl.h"
25 #include "clang/AST/DeclCXX.h"
26 #include "clang/AST/DeclObjC.h"
29 #include "clang/Basic/TargetInfo.h"
32 #include "llvm/ADT/StringExtras.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Assumptions.h"
35 #include "llvm/IR/Attributes.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/InlineAsm.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 using namespace clang;
43 using namespace CodeGen;
44 
45 /***/
46 
48  switch (CC) {
49  default: return llvm::CallingConv::C;
50  case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
51  case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
52  case CC_X86RegCall: return llvm::CallingConv::X86_RegCall;
53  case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
54  case CC_Win64: return llvm::CallingConv::Win64;
55  case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
56  case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
57  case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
58  case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
59  // TODO: Add support for __pascal to LLVM.
61  // TODO: Add support for __vectorcall to LLVM.
62  case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall;
63  case CC_AArch64VectorCall: return llvm::CallingConv::AArch64_VectorCall;
64  case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
66  case CC_PreserveMost: return llvm::CallingConv::PreserveMost;
67  case CC_PreserveAll: return llvm::CallingConv::PreserveAll;
68  case CC_Swift: return llvm::CallingConv::Swift;
69  case CC_SwiftAsync: return llvm::CallingConv::SwiftTail;
70  }
71 }
72 
73 /// Derives the 'this' type for codegen purposes, i.e. ignoring method CVR
74 /// qualification. Either or both of RD and MD may be null. A null RD indicates
75 /// that there is no meaningful 'this' type, and a null MD can occur when
76 /// calling a method pointer.
78  const CXXMethodDecl *MD) {
79  QualType RecTy;
80  if (RD)
81  RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
82  else
83  RecTy = Context.VoidTy;
84 
85  if (MD)
86  RecTy = Context.getAddrSpaceQualType(RecTy, MD->getMethodQualifiers().getAddressSpace());
87  return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
88 }
89 
90 /// Returns the canonical formal type of the given C++ method.
92  return MD->getType()->getCanonicalTypeUnqualified()
94 }
95 
96 /// Returns the "extra-canonicalized" return type, which discards
97 /// qualifiers on the return type. Codegen doesn't care about them,
98 /// and it makes ABI code a little easier to be able to assume that
99 /// all parameter and return types are top-level unqualified.
102 }
103 
104 /// Arrange the argument and result information for a value of the given
105 /// unprototyped freestanding function type.
106 const CGFunctionInfo &
108  // When translating an unprototyped function type, always use a
109  // variadic type.
110  return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
111  /*instanceMethod=*/false,
112  /*chainCall=*/false, None,
113  FTNP->getExtInfo(), {}, RequiredArgs(0));
114 }
115 
118  const FunctionProtoType *proto,
119  unsigned prefixArgs,
120  unsigned totalArgs) {
121  assert(proto->hasExtParameterInfos());
122  assert(paramInfos.size() <= prefixArgs);
123  assert(proto->getNumParams() + prefixArgs <= totalArgs);
124 
125  paramInfos.reserve(totalArgs);
126 
127  // Add default infos for any prefix args that don't already have infos.
128  paramInfos.resize(prefixArgs);
129 
130  // Add infos for the prototype.
131  for (const auto &ParamInfo : proto->getExtParameterInfos()) {
132  paramInfos.push_back(ParamInfo);
133  // pass_object_size params have no parameter info.
134  if (ParamInfo.hasPassObjectSize())
135  paramInfos.emplace_back();
136  }
137 
138  assert(paramInfos.size() <= totalArgs &&
139  "Did we forget to insert pass_object_size args?");
140  // Add default infos for the variadic and/or suffix arguments.
141  paramInfos.resize(totalArgs);
142 }
143 
144 /// Adds the formal parameters in FPT to the given prefix. If any parameter in
145 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
146 static void appendParameterTypes(const CodeGenTypes &CGT,
147  SmallVectorImpl<CanQualType> &prefix,
148  SmallVectorImpl<FunctionProtoType::ExtParameterInfo> &paramInfos,
150  // Fast path: don't touch param info if we don't need to.
151  if (!FPT->hasExtParameterInfos()) {
152  assert(paramInfos.empty() &&
153  "We have paramInfos, but the prototype doesn't?");
154  prefix.append(FPT->param_type_begin(), FPT->param_type_end());
155  return;
156  }
157 
158  unsigned PrefixSize = prefix.size();
159  // In the vast majority of cases, we'll have precisely FPT->getNumParams()
160  // parameters; the only thing that can change this is the presence of
161  // pass_object_size. So, we preallocate for the common case.
162  prefix.reserve(prefix.size() + FPT->getNumParams());
163 
164  auto ExtInfos = FPT->getExtParameterInfos();
165  assert(ExtInfos.size() == FPT->getNumParams());
166  for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
167  prefix.push_back(FPT->getParamType(I));
168  if (ExtInfos[I].hasPassObjectSize())
169  prefix.push_back(CGT.getContext().getSizeType());
170  }
171 
172  addExtParameterInfosForCall(paramInfos, FPT.getTypePtr(), PrefixSize,
173  prefix.size());
174 }
175 
176 /// Arrange the LLVM function layout for a value of the given function
177 /// type, on top of any implicit parameters already stored.
178 static const CGFunctionInfo &
179 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
180  SmallVectorImpl<CanQualType> &prefix,
183  RequiredArgs Required = RequiredArgs::forPrototypePlus(FTP, prefix.size());
184  // FIXME: Kill copy.
185  appendParameterTypes(CGT, prefix, paramInfos, FTP);
186  CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
187 
188  return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
189  /*chainCall=*/false, prefix,
190  FTP->getExtInfo(), paramInfos,
191  Required);
192 }
193 
194 /// Arrange the argument and result information for a value of the
195 /// given freestanding function type.
196 const CGFunctionInfo &
199  return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
200  FTP);
201 }
202 
204  bool IsWindows) {
205  // Set the appropriate calling convention for the Function.
206  if (D->hasAttr<StdCallAttr>())
207  return CC_X86StdCall;
208 
209  if (D->hasAttr<FastCallAttr>())
210  return CC_X86FastCall;
211 
212  if (D->hasAttr<RegCallAttr>())
213  return CC_X86RegCall;
214 
215  if (D->hasAttr<ThisCallAttr>())
216  return CC_X86ThisCall;
217 
218  if (D->hasAttr<VectorCallAttr>())
219  return CC_X86VectorCall;
220 
221  if (D->hasAttr<PascalAttr>())
222  return CC_X86Pascal;
223 
224  if (PcsAttr *PCS = D->getAttr<PcsAttr>())
225  return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
226 
227  if (D->hasAttr<AArch64VectorPcsAttr>())
228  return CC_AArch64VectorCall;
229 
230  if (D->hasAttr<IntelOclBiccAttr>())
231  return CC_IntelOclBicc;
232 
233  if (D->hasAttr<MSABIAttr>())
234  return IsWindows ? CC_C : CC_Win64;
235 
236  if (D->hasAttr<SysVABIAttr>())
237  return IsWindows ? CC_X86_64SysV : CC_C;
238 
239  if (D->hasAttr<PreserveMostAttr>())
240  return CC_PreserveMost;
241 
242  if (D->hasAttr<PreserveAllAttr>())
243  return CC_PreserveAll;
244 
245  return CC_C;
246 }
247 
248 /// Arrange the argument and result information for a call to an
249 /// unknown C++ non-static member function of the given abstract type.
250 /// (A null RD means we don't have any meaningful "this" argument type,
251 /// so fall back to a generic pointer type).
252 /// The member function must be an ordinary function, i.e. not a
253 /// constructor or destructor.
254 const CGFunctionInfo &
256  const FunctionProtoType *FTP,
257  const CXXMethodDecl *MD) {
259 
260  // Add the 'this' pointer.
261  argTypes.push_back(DeriveThisType(RD, MD));
262 
264  *this, true, argTypes,
266 }
267 
268 /// Set calling convention for CUDA/HIP kernel.
270  const FunctionDecl *FD) {
271  if (FD->hasAttr<CUDAGlobalAttr>()) {
272  const FunctionType *FT = FTy->getAs<FunctionType>();
274  FTy = FT->getCanonicalTypeUnqualified();
275  }
276 }
277 
278 /// Arrange the argument and result information for a declaration or
279 /// definition of the given C++ non-static member function. The
280 /// member function must be an ordinary function, i.e. not a
281 /// constructor or destructor.
282 const CGFunctionInfo &
284  assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
285  assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
286 
287  CanQualType FT = GetFormalType(MD).getAs<Type>();
288  setCUDAKernelCallingConvention(FT, CGM, MD);
289  auto prototype = FT.getAs<FunctionProtoType>();
290 
291  if (MD->isInstance()) {
292  // The abstract case is perfectly fine.
293  const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
294  return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
295  }
296 
297  return arrangeFreeFunctionType(prototype);
298 }
299 
301  const InheritedConstructor &Inherited, CXXCtorType Type) {
302  // Parameters are unnecessary if we're constructing a base class subobject
303  // and the inherited constructor lives in a virtual base.
304  return Type == Ctor_Complete ||
305  !Inherited.getShadowDecl()->constructsVirtualBase() ||
306  !Target.getCXXABI().hasConstructorVariants();
307 }
308 
309 const CGFunctionInfo &
311  auto *MD = cast<CXXMethodDecl>(GD.getDecl());
312 
315  argTypes.push_back(DeriveThisType(MD->getParent(), MD));
316 
317  bool PassParams = true;
318 
319  if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
320  // A base class inheriting constructor doesn't get forwarded arguments
321  // needed to construct a virtual base (or base class thereof).
322  if (auto Inherited = CD->getInheritedConstructor())
323  PassParams = inheritingCtorHasParams(Inherited, GD.getCtorType());
324  }
325 
327 
328  // Add the formal parameters.
329  if (PassParams)
330  appendParameterTypes(*this, argTypes, paramInfos, FTP);
331 
333  TheCXXABI.buildStructorSignature(GD, argTypes);
334  if (!paramInfos.empty()) {
335  // Note: prefix implies after the first param.
336  if (AddedArgs.Prefix)
337  paramInfos.insert(paramInfos.begin() + 1, AddedArgs.Prefix,
339  if (AddedArgs.Suffix)
340  paramInfos.append(AddedArgs.Suffix,
342  }
343 
344  RequiredArgs required =
345  (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
347 
348  FunctionType::ExtInfo extInfo = FTP->getExtInfo();
349  CanQualType resultType = TheCXXABI.HasThisReturn(GD)
350  ? argTypes.front()
351  : TheCXXABI.hasMostDerivedReturn(GD)
352  ? CGM.getContext().VoidPtrTy
353  : Context.VoidTy;
354  return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
355  /*chainCall=*/false, argTypes, extInfo,
356  paramInfos, required);
357 }
358 
362  for (auto &arg : args)
363  argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
364  return argTypes;
365 }
366 
370  for (auto &arg : args)
371  argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
372  return argTypes;
373 }
374 
377  unsigned prefixArgs, unsigned totalArgs) {
379  if (proto->hasExtParameterInfos()) {
380  addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
381  }
382  return result;
383 }
384 
385 /// Arrange a call to a C++ method, passing the given arguments.
386 ///
387 /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this`
388 /// parameter.
389 /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of
390 /// args.
391 /// PassProtoArgs indicates whether `args` has args for the parameters in the
392 /// given CXXConstructorDecl.
393 const CGFunctionInfo &
395  const CXXConstructorDecl *D,
396  CXXCtorType CtorKind,
397  unsigned ExtraPrefixArgs,
398  unsigned ExtraSuffixArgs,
399  bool PassProtoArgs) {
400  // FIXME: Kill copy.
402  for (const auto &Arg : args)
403  ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
404 
405  // +1 for implicit this, which should always be args[0].
406  unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs;
407 
409  RequiredArgs Required = PassProtoArgs
411  FPT, TotalPrefixArgs + ExtraSuffixArgs)
413 
414  GlobalDecl GD(D, CtorKind);
415  CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
416  ? ArgTypes.front()
417  : TheCXXABI.hasMostDerivedReturn(GD)
418  ? CGM.getContext().VoidPtrTy
419  : Context.VoidTy;
420 
421  FunctionType::ExtInfo Info = FPT->getExtInfo();
423  // If the prototype args are elided, we should only have ABI-specific args,
424  // which never have param info.
425  if (PassProtoArgs && FPT->hasExtParameterInfos()) {
426  // ABI-specific suffix arguments are treated the same as variadic arguments.
427  addExtParameterInfosForCall(ParamInfos, FPT.getTypePtr(), TotalPrefixArgs,
428  ArgTypes.size());
429  }
430  return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
431  /*chainCall=*/false, ArgTypes, Info,
432  ParamInfos, Required);
433 }
434 
435 /// Arrange the argument and result information for the declaration or
436 /// definition of the given function.
437 const CGFunctionInfo &
439  if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
440  if (MD->isInstance())
441  return arrangeCXXMethodDeclaration(MD);
442 
444 
445  assert(isa<FunctionType>(FTy));
446  setCUDAKernelCallingConvention(FTy, CGM, FD);
447 
448  // When declaring a function without a prototype, always use a
449  // non-variadic type.
452  noProto->getReturnType(), /*instanceMethod=*/false,
453  /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
454  }
455 
457 }
458 
459 /// Arrange the argument and result information for the declaration or
460 /// definition of an Objective-C method.
461 const CGFunctionInfo &
463  // It happens that this is the same as a call with no optional
464  // arguments, except also using the formal 'self' type.
466 }
467 
468 /// Arrange the argument and result information for the function type
469 /// through which to perform a send to the given Objective-C method,
470 /// using the given receiver type. The receiver type is not always
471 /// the 'self' type of the method or even an Objective-C pointer type.
472 /// This is *not* the right method for actually performing such a
473 /// message send, due to the possibility of optional arguments.
474 const CGFunctionInfo &
476  QualType receiverType) {
479  argTys.push_back(Context.getCanonicalParamType(receiverType));
480  argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
481  // FIXME: Kill copy?
482  for (const auto *I : MD->parameters()) {
483  argTys.push_back(Context.getCanonicalParamType(I->getType()));
485  I->hasAttr<NoEscapeAttr>());
486  extParamInfos.push_back(extParamInfo);
487  }
488 
489  FunctionType::ExtInfo einfo;
490  bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
491  einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
492 
493  if (getContext().getLangOpts().ObjCAutoRefCount &&
494  MD->hasAttr<NSReturnsRetainedAttr>())
495  einfo = einfo.withProducesResult(true);
496 
497  RequiredArgs required =
498  (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
499 
501  GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
502  /*chainCall=*/false, argTys, einfo, extParamInfos, required);
503 }
504 
505 const CGFunctionInfo &
507  const CallArgList &args) {
508  auto argTypes = getArgTypesForCall(Context, args);
509  FunctionType::ExtInfo einfo;
510 
512  GetReturnType(returnType), /*instanceMethod=*/false,
513  /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
514 }
515 
516 const CGFunctionInfo &
518  // FIXME: Do we need to handle ObjCMethodDecl?
519  const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
520 
521  if (isa<CXXConstructorDecl>(GD.getDecl()) ||
522  isa<CXXDestructorDecl>(GD.getDecl()))
524 
525  return arrangeFunctionDeclaration(FD);
526 }
527 
528 /// Arrange a thunk that takes 'this' as the first parameter followed by
529 /// varargs. Return a void pointer, regardless of the actual return type.
530 /// The body of the thunk will end in a musttail call to a function of the
531 /// correct type, and the caller will bitcast the function to the correct
532 /// prototype.
533 const CGFunctionInfo &
535  assert(MD->isVirtual() && "only methods have thunks");
537  CanQualType ArgTys[] = {DeriveThisType(MD->getParent(), MD)};
538  return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
539  /*chainCall=*/false, ArgTys,
540  FTP->getExtInfo(), {}, RequiredArgs(1));
541 }
542 
543 const CGFunctionInfo &
545  CXXCtorType CT) {
546  assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
547 
550  const CXXRecordDecl *RD = CD->getParent();
551  ArgTys.push_back(DeriveThisType(RD, CD));
552  if (CT == Ctor_CopyingClosure)
553  ArgTys.push_back(*FTP->param_type_begin());
554  if (RD->getNumVBases() > 0)
555  ArgTys.push_back(Context.IntTy);
557  /*IsVariadic=*/false, /*IsCXXMethod=*/true);
558  return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
559  /*chainCall=*/false, ArgTys,
560  FunctionType::ExtInfo(CC), {},
562 }
563 
564 /// Arrange a call as unto a free function, except possibly with an
565 /// additional number of formal parameters considered required.
566 static const CGFunctionInfo &
568  CodeGenModule &CGM,
569  const CallArgList &args,
570  const FunctionType *fnType,
571  unsigned numExtraRequiredArgs,
572  bool chainCall) {
573  assert(args.size() >= numExtraRequiredArgs);
574 
576 
577  // In most cases, there are no optional arguments.
578  RequiredArgs required = RequiredArgs::All;
579 
580  // If we have a variadic prototype, the required arguments are the
581  // extra prefix plus the arguments in the prototype.
582  if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
583  if (proto->isVariadic())
584  required = RequiredArgs::forPrototypePlus(proto, numExtraRequiredArgs);
585 
586  if (proto->hasExtParameterInfos())
587  addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
588  args.size());
589 
590  // If we don't have a prototype at all, but we're supposed to
591  // explicitly use the variadic convention for unprototyped calls,
592  // treat all of the arguments as required but preserve the nominal
593  // possibility of variadics.
594  } else if (CGM.getTargetCodeGenInfo()
595  .isNoProtoCallVariadic(args,
596  cast<FunctionNoProtoType>(fnType))) {
597  required = RequiredArgs(args.size());
598  }
599 
600  // FIXME: Kill copy.
602  for (const auto &arg : args)
603  argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
605  /*instanceMethod=*/false, chainCall,
606  argTypes, fnType->getExtInfo(), paramInfos,
607  required);
608 }
609 
610 /// Figure out the rules for calling a function with the given formal
611 /// type using the given arguments. The arguments are necessary
612 /// because the function might be unprototyped, in which case it's
613 /// target-dependent in crazy ways.
614 const CGFunctionInfo &
616  const FunctionType *fnType,
617  bool chainCall) {
618  return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
619  chainCall ? 1 : 0, chainCall);
620 }
621 
622 /// A block function is essentially a free function with an
623 /// extra implicit argument.
624 const CGFunctionInfo &
626  const FunctionType *fnType) {
627  return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
628  /*chainCall=*/false);
629 }
630 
631 const CGFunctionInfo &
633  const FunctionArgList &params) {
634  auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
635  auto argTypes = getArgTypesForDeclaration(Context, params);
636 
638  /*instanceMethod*/ false, /*chainCall*/ false,
639  argTypes, proto->getExtInfo(), paramInfos,
641 }
642 
643 const CGFunctionInfo &
645  const CallArgList &args) {
646  // FIXME: Kill copy.
648  for (const auto &Arg : args)
649  argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
651  GetReturnType(resultType), /*instanceMethod=*/false,
652  /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
653  /*paramInfos=*/ {}, RequiredArgs::All);
654 }
655 
656 const CGFunctionInfo &
658  const FunctionArgList &args) {
659  auto argTypes = getArgTypesForDeclaration(Context, args);
660 
662  GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
663  argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
664 }
665 
666 const CGFunctionInfo &
668  ArrayRef<CanQualType> argTypes) {
670  resultType, /*instanceMethod=*/false, /*chainCall=*/false,
671  argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
672 }
673 
674 /// Arrange a call to a C++ method, passing the given arguments.
675 ///
676 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
677 /// does not count `this`.
678 const CGFunctionInfo &
680  const FunctionProtoType *proto,
681  RequiredArgs required,
682  unsigned numPrefixArgs) {
683  assert(numPrefixArgs + 1 <= args.size() &&
684  "Emitting a call with less args than the required prefix?");
685  // Add one to account for `this`. It's a bit awkward here, but we don't count
686  // `this` in similar places elsewhere.
687  auto paramInfos =
688  getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
689 
690  // FIXME: Kill copy.
691  auto argTypes = getArgTypesForCall(Context, args);
692 
693  FunctionType::ExtInfo info = proto->getExtInfo();
695  GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
696  /*chainCall=*/false, argTypes, info, paramInfos, required);
697 }
698 
701  getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
703 }
704 
705 const CGFunctionInfo &
707  const CallArgList &args) {
708  assert(signature.arg_size() <= args.size());
709  if (signature.arg_size() == args.size())
710  return signature;
711 
713  auto sigParamInfos = signature.getExtParameterInfos();
714  if (!sigParamInfos.empty()) {
715  paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
716  paramInfos.resize(args.size());
717  }
718 
719  auto argTypes = getArgTypesForCall(Context, args);
720 
721  assert(signature.getRequiredArgs().allowsOptionalArgs());
722  return arrangeLLVMFunctionInfo(signature.getReturnType(),
723  signature.isInstanceMethod(),
724  signature.isChainCall(),
725  argTypes,
726  signature.getExtInfo(),
727  paramInfos,
728  signature.getRequiredArgs());
729 }
730 
731 namespace clang {
732 namespace CodeGen {
734 }
735 }
736 
737 /// Arrange the argument and result information for an abstract value
738 /// of a given function type. This is the method which all of the
739 /// above functions ultimately defer to.
740 const CGFunctionInfo &
742  bool instanceMethod,
743  bool chainCall,
744  ArrayRef<CanQualType> argTypes,
747  RequiredArgs required) {
748  assert(llvm::all_of(argTypes,
749  [](CanQualType T) { return T.isCanonicalAsParam(); }));
750 
751  // Lookup or create unique function info.
752  llvm::FoldingSetNodeID ID;
753  CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
754  required, resultType, argTypes);
755 
756  void *insertPos = nullptr;
757  CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
758  if (FI)
759  return *FI;
760 
761  unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
762 
763  // Construct the function info. We co-allocate the ArgInfos.
764  FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
765  paramInfos, resultType, argTypes, required);
766  FunctionInfos.InsertNode(FI, insertPos);
767 
768  bool inserted = FunctionsBeingProcessed.insert(FI).second;
769  (void)inserted;
770  assert(inserted && "Recursively being processed?");
771 
772  // Compute ABI information.
773  if (CC == llvm::CallingConv::SPIR_KERNEL) {
774  // Force target independent argument handling for the host visible
775  // kernel functions.
776  computeSPIRKernelABIInfo(CGM, *FI);
777  } else if (info.getCC() == CC_Swift || info.getCC() == CC_SwiftAsync) {
778  swiftcall::computeABIInfo(CGM, *FI);
779  } else {
780  getABIInfo().computeInfo(*FI);
781  }
782 
783  // Loop over all of the computed argument and return value info. If any of
784  // them are direct or extend without a specified coerce type, specify the
785  // default now.
786  ABIArgInfo &retInfo = FI->getReturnInfo();
787  if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
788  retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
789 
790  for (auto &I : FI->arguments())
791  if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
792  I.info.setCoerceToType(ConvertType(I.type));
793 
794  bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
795  assert(erased && "Not in set?");
796 
797  return *FI;
798 }
799 
801  bool instanceMethod,
802  bool chainCall,
803  const FunctionType::ExtInfo &info,
804  ArrayRef<ExtParameterInfo> paramInfos,
805  CanQualType resultType,
806  ArrayRef<CanQualType> argTypes,
807  RequiredArgs required) {
808  assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
809  assert(!required.allowsOptionalArgs() ||
810  required.getNumRequiredArgs() <= argTypes.size());
811 
812  void *buffer =
813  operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
814  argTypes.size() + 1, paramInfos.size()));
815 
816  CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
817  FI->CallingConvention = llvmCC;
818  FI->EffectiveCallingConvention = llvmCC;
819  FI->ASTCallingConvention = info.getCC();
820  FI->InstanceMethod = instanceMethod;
821  FI->ChainCall = chainCall;
822  FI->CmseNSCall = info.getCmseNSCall();
823  FI->NoReturn = info.getNoReturn();
824  FI->ReturnsRetained = info.getProducesResult();
825  FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
826  FI->NoCfCheck = info.getNoCfCheck();
827  FI->Required = required;
828  FI->HasRegParm = info.getHasRegParm();
829  FI->RegParm = info.getRegParm();
830  FI->ArgStruct = nullptr;
831  FI->ArgStructAlign = 0;
832  FI->NumArgs = argTypes.size();
833  FI->HasExtParameterInfos = !paramInfos.empty();
834  FI->getArgsBuffer()[0].type = resultType;
835  for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
836  FI->getArgsBuffer()[i + 1].type = argTypes[i];
837  for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
838  FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
839  return FI;
840 }
841 
842 /***/
843 
844 namespace {
845 // ABIArgInfo::Expand implementation.
846 
847 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
848 struct TypeExpansion {
849  enum TypeExpansionKind {
850  // Elements of constant arrays are expanded recursively.
851  TEK_ConstantArray,
852  // Record fields are expanded recursively (but if record is a union, only
853  // the field with the largest size is expanded).
854  TEK_Record,
855  // For complex types, real and imaginary parts are expanded recursively.
856  TEK_Complex,
857  // All other types are not expandable.
858  TEK_None
859  };
860 
861  const TypeExpansionKind Kind;
862 
863  TypeExpansion(TypeExpansionKind K) : Kind(K) {}
864  virtual ~TypeExpansion() {}
865 };
866 
867 struct ConstantArrayExpansion : TypeExpansion {
868  QualType EltTy;
869  uint64_t NumElts;
870 
871  ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
872  : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
873  static bool classof(const TypeExpansion *TE) {
874  return TE->Kind == TEK_ConstantArray;
875  }
876 };
877 
878 struct RecordExpansion : TypeExpansion {
880 
882 
883  RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
885  : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
886  Fields(std::move(Fields)) {}
887  static bool classof(const TypeExpansion *TE) {
888  return TE->Kind == TEK_Record;
889  }
890 };
891 
892 struct ComplexExpansion : TypeExpansion {
893  QualType EltTy;
894 
895  ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
896  static bool classof(const TypeExpansion *TE) {
897  return TE->Kind == TEK_Complex;
898  }
899 };
900 
901 struct NoExpansion : TypeExpansion {
902  NoExpansion() : TypeExpansion(TEK_None) {}
903  static bool classof(const TypeExpansion *TE) {
904  return TE->Kind == TEK_None;
905  }
906 };
907 } // namespace
908 
909 static std::unique_ptr<TypeExpansion>
910 getTypeExpansion(QualType Ty, const ASTContext &Context) {
911  if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
912  return std::make_unique<ConstantArrayExpansion>(
913  AT->getElementType(), AT->getSize().getZExtValue());
914  }
915  if (const RecordType *RT = Ty->getAs<RecordType>()) {
918  const RecordDecl *RD = RT->getDecl();
919  assert(!RD->hasFlexibleArrayMember() &&
920  "Cannot expand structure with flexible array.");
921  if (RD->isUnion()) {
922  // Unions can be here only in degenerative cases - all the fields are same
923  // after flattening. Thus we have to use the "largest" field.
924  const FieldDecl *LargestFD = nullptr;
925  CharUnits UnionSize = CharUnits::Zero();
926 
927  for (const auto *FD : RD->fields()) {
928  if (FD->isZeroLengthBitField(Context))
929  continue;
930  assert(!FD->isBitField() &&
931  "Cannot expand structure with bit-field members.");
932  CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
933  if (UnionSize < FieldSize) {
934  UnionSize = FieldSize;
935  LargestFD = FD;
936  }
937  }
938  if (LargestFD)
939  Fields.push_back(LargestFD);
940  } else {
941  if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
942  assert(!CXXRD->isDynamicClass() &&
943  "cannot expand vtable pointers in dynamic classes");
944  for (const CXXBaseSpecifier &BS : CXXRD->bases())
945  Bases.push_back(&BS);
946  }
947 
948  for (const auto *FD : RD->fields()) {
949  if (FD->isZeroLengthBitField(Context))
950  continue;
951  assert(!FD->isBitField() &&
952  "Cannot expand structure with bit-field members.");
953  Fields.push_back(FD);
954  }
955  }
956  return std::make_unique<RecordExpansion>(std::move(Bases),
957  std::move(Fields));
958  }
959  if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
960  return std::make_unique<ComplexExpansion>(CT->getElementType());
961  }
962  return std::make_unique<NoExpansion>();
963 }
964 
965 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
966  auto Exp = getTypeExpansion(Ty, Context);
967  if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
968  return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
969  }
970  if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
971  int Res = 0;
972  for (auto BS : RExp->Bases)
973  Res += getExpansionSize(BS->getType(), Context);
974  for (auto FD : RExp->Fields)
975  Res += getExpansionSize(FD->getType(), Context);
976  return Res;
977  }
978  if (isa<ComplexExpansion>(Exp.get()))
979  return 2;
980  assert(isa<NoExpansion>(Exp.get()));
981  return 1;
982 }
983 
984 void
987  auto Exp = getTypeExpansion(Ty, Context);
988  if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
989  for (int i = 0, n = CAExp->NumElts; i < n; i++) {
990  getExpandedTypes(CAExp->EltTy, TI);
991  }
992  } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
993  for (auto BS : RExp->Bases)
994  getExpandedTypes(BS->getType(), TI);
995  for (auto FD : RExp->Fields)
996  getExpandedTypes(FD->getType(), TI);
997  } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
998  llvm::Type *EltTy = ConvertType(CExp->EltTy);
999  *TI++ = EltTy;
1000  *TI++ = EltTy;
1001  } else {
1002  assert(isa<NoExpansion>(Exp.get()));
1003  *TI++ = ConvertType(Ty);
1004  }
1005 }
1006 
1008  ConstantArrayExpansion *CAE,
1009  Address BaseAddr,
1010  llvm::function_ref<void(Address)> Fn) {
1011  CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
1012  CharUnits EltAlign =
1013  BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
1014 
1015  for (int i = 0, n = CAE->NumElts; i < n; i++) {
1016  llvm::Value *EltAddr = CGF.Builder.CreateConstGEP2_32(
1017  BaseAddr.getElementType(), BaseAddr.getPointer(), 0, i);
1018  Fn(Address(EltAddr, EltAlign));
1019  }
1020 }
1021 
1022 void CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV,
1023  llvm::Function::arg_iterator &AI) {
1024  assert(LV.isSimple() &&
1025  "Unexpected non-simple lvalue during struct expansion.");
1026 
1027  auto Exp = getTypeExpansion(Ty, getContext());
1028  if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1030  *this, CAExp, LV.getAddress(*this), [&](Address EltAddr) {
1031  LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
1032  ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1033  });
1034  } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1035  Address This = LV.getAddress(*this);
1036  for (const CXXBaseSpecifier *BS : RExp->Bases) {
1037  // Perform a single step derived-to-base conversion.
1038  Address Base =
1039  GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1040  /*NullCheckValue=*/false, SourceLocation());
1041  LValue SubLV = MakeAddrLValue(Base, BS->getType());
1042 
1043  // Recurse onto bases.
1044  ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1045  }
1046  for (auto FD : RExp->Fields) {
1047  // FIXME: What are the right qualifiers here?
1048  LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1049  ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1050  }
1051  } else if (isa<ComplexExpansion>(Exp.get())) {
1052  auto realValue = &*AI++;
1053  auto imagValue = &*AI++;
1054  EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1055  } else {
1056  // Call EmitStoreOfScalar except when the lvalue is a bitfield to emit a
1057  // primitive store.
1058  assert(isa<NoExpansion>(Exp.get()));
1059  if (LV.isBitField())
1060  EmitStoreThroughLValue(RValue::get(&*AI++), LV);
1061  else
1062  EmitStoreOfScalar(&*AI++, LV);
1063  }
1064 }
1065 
1066 void CodeGenFunction::ExpandTypeToArgs(
1067  QualType Ty, CallArg Arg, llvm::FunctionType *IRFuncTy,
1068  SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1069  auto Exp = getTypeExpansion(Ty, getContext());
1070  if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1071  Address Addr = Arg.hasLValue() ? Arg.getKnownLValue().getAddress(*this)
1074  *this, CAExp, Addr, [&](Address EltAddr) {
1075  CallArg EltArg = CallArg(
1076  convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation()),
1077  CAExp->EltTy);
1078  ExpandTypeToArgs(CAExp->EltTy, EltArg, IRFuncTy, IRCallArgs,
1079  IRCallArgPos);
1080  });
1081  } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1082  Address This = Arg.hasLValue() ? Arg.getKnownLValue().getAddress(*this)
1084  for (const CXXBaseSpecifier *BS : RExp->Bases) {
1085  // Perform a single step derived-to-base conversion.
1086  Address Base =
1087  GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1088  /*NullCheckValue=*/false, SourceLocation());
1089  CallArg BaseArg = CallArg(RValue::getAggregate(Base), BS->getType());
1090 
1091  // Recurse onto bases.
1092  ExpandTypeToArgs(BS->getType(), BaseArg, IRFuncTy, IRCallArgs,
1093  IRCallArgPos);
1094  }
1095 
1096  LValue LV = MakeAddrLValue(This, Ty);
1097  for (auto FD : RExp->Fields) {
1098  CallArg FldArg =
1099  CallArg(EmitRValueForField(LV, FD, SourceLocation()), FD->getType());
1100  ExpandTypeToArgs(FD->getType(), FldArg, IRFuncTy, IRCallArgs,
1101  IRCallArgPos);
1102  }
1103  } else if (isa<ComplexExpansion>(Exp.get())) {
1105  IRCallArgs[IRCallArgPos++] = CV.first;
1106  IRCallArgs[IRCallArgPos++] = CV.second;
1107  } else {
1108  assert(isa<NoExpansion>(Exp.get()));
1109  auto RV = Arg.getKnownRValue();
1110  assert(RV.isScalar() &&
1111  "Unexpected non-scalar rvalue during struct expansion.");
1112 
1113  // Insert a bitcast as needed.
1114  llvm::Value *V = RV.getScalarVal();
1115  if (IRCallArgPos < IRFuncTy->getNumParams() &&
1116  V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1117  V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1118 
1119  IRCallArgs[IRCallArgPos++] = V;
1120  }
1121 }
1122 
1123 /// Create a temporary allocation for the purposes of coercion.
1125  CharUnits MinAlign,
1126  const Twine &Name = "tmp") {
1127  // Don't use an alignment that's worse than what LLVM would prefer.
1128  auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1129  CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1130 
1131  return CGF.CreateTempAlloca(Ty, Align, Name + ".coerce");
1132 }
1133 
1134 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1135 /// accessing some number of bytes out of it, try to gep into the struct to get
1136 /// at its inner goodness. Dive as deep as possible without entering an element
1137 /// with an in-memory size smaller than DstSize.
1138 static Address
1140  llvm::StructType *SrcSTy,
1141  uint64_t DstSize, CodeGenFunction &CGF) {
1142  // We can't dive into a zero-element struct.
1143  if (SrcSTy->getNumElements() == 0) return SrcPtr;
1144 
1145  llvm::Type *FirstElt = SrcSTy->getElementType(0);
1146 
1147  // If the first elt is at least as large as what we're looking for, or if the
1148  // first element is the same size as the whole struct, we can enter it. The
1149  // comparison must be made on the store size and not the alloca size. Using
1150  // the alloca size may overstate the size of the load.
1151  uint64_t FirstEltSize =
1152  CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1153  if (FirstEltSize < DstSize &&
1154  FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1155  return SrcPtr;
1156 
1157  // GEP into the first element.
1158  SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, "coerce.dive");
1159 
1160  // If the first element is a struct, recurse.
1161  llvm::Type *SrcTy = SrcPtr.getElementType();
1162  if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1163  return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1164 
1165  return SrcPtr;
1166 }
1167 
1168 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1169 /// are either integers or pointers. This does a truncation of the value if it
1170 /// is too large or a zero extension if it is too small.
1171 ///
1172 /// This behaves as if the value were coerced through memory, so on big-endian
1173 /// targets the high bits are preserved in a truncation, while little-endian
1174 /// targets preserve the low bits.
1176  llvm::Type *Ty,
1177  CodeGenFunction &CGF) {
1178  if (Val->getType() == Ty)
1179  return Val;
1180 
1181  if (isa<llvm::PointerType>(Val->getType())) {
1182  // If this is Pointer->Pointer avoid conversion to and from int.
1183  if (isa<llvm::PointerType>(Ty))
1184  return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1185 
1186  // Convert the pointer to an integer so we can play with its width.
1187  Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1188  }
1189 
1190  llvm::Type *DestIntTy = Ty;
1191  if (isa<llvm::PointerType>(DestIntTy))
1192  DestIntTy = CGF.IntPtrTy;
1193 
1194  if (Val->getType() != DestIntTy) {
1195  const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1196  if (DL.isBigEndian()) {
1197  // Preserve the high bits on big-endian targets.
1198  // That is what memory coercion does.
1199  uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1200  uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1201 
1202  if (SrcSize > DstSize) {
1203  Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1204  Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1205  } else {
1206  Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1207  Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1208  }
1209  } else {
1210  // Little-endian targets preserve the low bits. No shifts required.
1211  Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1212  }
1213  }
1214 
1215  if (isa<llvm::PointerType>(Ty))
1216  Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1217  return Val;
1218 }
1219 
1220 
1221 
1222 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1223 /// a pointer to an object of type \arg Ty, known to be aligned to
1224 /// \arg SrcAlign bytes.
1225 ///
1226 /// This safely handles the case when the src type is smaller than the
1227 /// destination type; in this situation the values of bits which not
1228 /// present in the src are undefined.
1229 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1230  CodeGenFunction &CGF) {
1231  llvm::Type *SrcTy = Src.getElementType();
1232 
1233  // If SrcTy and Ty are the same, just do a load.
1234  if (SrcTy == Ty)
1235  return CGF.Builder.CreateLoad(Src);
1236 
1237  llvm::TypeSize DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1238 
1239  if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1240  Src = EnterStructPointerForCoercedAccess(Src, SrcSTy,
1241  DstSize.getFixedSize(), CGF);
1242  SrcTy = Src.getElementType();
1243  }
1244 
1245  llvm::TypeSize SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1246 
1247  // If the source and destination are integer or pointer types, just do an
1248  // extension or truncation to the desired type.
1249  if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1250  (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1251  llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1252  return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1253  }
1254 
1255  // If load is legal, just bitcast the src pointer.
1256  if (!SrcSize.isScalable() && !DstSize.isScalable() &&
1257  SrcSize.getFixedSize() >= DstSize.getFixedSize()) {
1258  // Generally SrcSize is never greater than DstSize, since this means we are
1259  // losing bits. However, this can happen in cases where the structure has
1260  // additional padding, for example due to a user specified alignment.
1261  //
1262  // FIXME: Assert that we aren't truncating non-padding bits when have access
1263  // to that information.
1264  Src = CGF.Builder.CreateBitCast(Src,
1265  Ty->getPointerTo(Src.getAddressSpace()));
1266  return CGF.Builder.CreateLoad(Src);
1267  }
1268 
1269  // If coercing a fixed vector to a scalable vector for ABI compatibility, and
1270  // the types match, use the llvm.experimental.vector.insert intrinsic to
1271  // perform the conversion.
1272  if (auto *ScalableDst = dyn_cast<llvm::ScalableVectorType>(Ty)) {
1273  if (auto *FixedSrc = dyn_cast<llvm::FixedVectorType>(SrcTy)) {
1274  // If we are casting a fixed i8 vector to a scalable 16 x i1 predicate
1275  // vector, use a vector insert and bitcast the result.
1276  bool NeedsBitcast = false;
1277  auto PredType =
1278  llvm::ScalableVectorType::get(CGF.Builder.getInt1Ty(), 16);
1279  llvm::Type *OrigType = Ty;
1280  if (ScalableDst == PredType &&
1281  FixedSrc->getElementType() == CGF.Builder.getInt8Ty()) {
1282  ScalableDst = llvm::ScalableVectorType::get(CGF.Builder.getInt8Ty(), 2);
1283  NeedsBitcast = true;
1284  }
1285  if (ScalableDst->getElementType() == FixedSrc->getElementType()) {
1286  auto *Load = CGF.Builder.CreateLoad(Src);
1287  auto *UndefVec = llvm::UndefValue::get(ScalableDst);
1288  auto *Zero = llvm::Constant::getNullValue(CGF.CGM.Int64Ty);
1289  llvm::Value *Result = CGF.Builder.CreateInsertVector(
1290  ScalableDst, UndefVec, Load, Zero, "castScalableSve");
1291  if (NeedsBitcast)
1292  Result = CGF.Builder.CreateBitCast(Result, OrigType);
1293  return Result;
1294  }
1295  }
1296  }
1297 
1298  // Otherwise do coercion through memory. This is stupid, but simple.
1299  Address Tmp =
1300  CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment(), Src.getName());
1301  CGF.Builder.CreateMemCpy(
1302  Tmp.getPointer(), Tmp.getAlignment().getAsAlign(), Src.getPointer(),
1303  Src.getAlignment().getAsAlign(),
1304  llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize.getKnownMinSize()));
1305  return CGF.Builder.CreateLoad(Tmp);
1306 }
1307 
1308 // Function to store a first-class aggregate into memory. We prefer to
1309 // store the elements rather than the aggregate to be more friendly to
1310 // fast-isel.
1311 // FIXME: Do we need to recurse here?
1313  bool DestIsVolatile) {
1314  // Prefer scalar stores to first-class aggregate stores.
1315  if (llvm::StructType *STy = dyn_cast<llvm::StructType>(Val->getType())) {
1316  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1317  Address EltPtr = Builder.CreateStructGEP(Dest, i);
1318  llvm::Value *Elt = Builder.CreateExtractValue(Val, i);
1319  Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1320  }
1321  } else {
1322  Builder.CreateStore(Val, Dest, DestIsVolatile);
1323  }
1324 }
1325 
1326 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1327 /// where the source and destination may have different types. The
1328 /// destination is known to be aligned to \arg DstAlign bytes.
1329 ///
1330 /// This safely handles the case when the src type is larger than the
1331 /// destination type; the upper bits of the src will be lost.
1333  Address Dst,
1334  bool DstIsVolatile,
1335  CodeGenFunction &CGF) {
1336  llvm::Type *SrcTy = Src->getType();
1337  llvm::Type *DstTy = Dst.getElementType();
1338  if (SrcTy == DstTy) {
1339  CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1340  return;
1341  }
1342 
1343  llvm::TypeSize SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1344 
1345  if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1346  Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy,
1347  SrcSize.getFixedSize(), CGF);
1348  DstTy = Dst.getElementType();
1349  }
1350 
1351  llvm::PointerType *SrcPtrTy = llvm::dyn_cast<llvm::PointerType>(SrcTy);
1352  llvm::PointerType *DstPtrTy = llvm::dyn_cast<llvm::PointerType>(DstTy);
1353  if (SrcPtrTy && DstPtrTy &&
1354  SrcPtrTy->getAddressSpace() != DstPtrTy->getAddressSpace()) {
1355  Src = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Src, DstTy);
1356  CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1357  return;
1358  }
1359 
1360  // If the source and destination are integer or pointer types, just do an
1361  // extension or truncation to the desired type.
1362  if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1363  (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1364  Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1365  CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1366  return;
1367  }
1368 
1369  llvm::TypeSize DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1370 
1371  // If store is legal, just bitcast the src pointer.
1372  if (isa<llvm::ScalableVectorType>(SrcTy) ||
1373  isa<llvm::ScalableVectorType>(DstTy) ||
1374  SrcSize.getFixedSize() <= DstSize.getFixedSize()) {
1375  Dst = CGF.Builder.CreateElementBitCast(Dst, SrcTy);
1376  CGF.EmitAggregateStore(Src, Dst, DstIsVolatile);
1377  } else {
1378  // Otherwise do coercion through memory. This is stupid, but
1379  // simple.
1380 
1381  // Generally SrcSize is never greater than DstSize, since this means we are
1382  // losing bits. However, this can happen in cases where the structure has
1383  // additional padding, for example due to a user specified alignment.
1384  //
1385  // FIXME: Assert that we aren't truncating non-padding bits when have access
1386  // to that information.
1387  Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1388  CGF.Builder.CreateStore(Src, Tmp);
1389  CGF.Builder.CreateMemCpy(
1390  Dst.getPointer(), Dst.getAlignment().getAsAlign(), Tmp.getPointer(),
1391  Tmp.getAlignment().getAsAlign(),
1392  llvm::ConstantInt::get(CGF.IntPtrTy, DstSize.getFixedSize()));
1393  }
1394 }
1395 
1397  const ABIArgInfo &info) {
1398  if (unsigned offset = info.getDirectOffset()) {
1399  addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1400  addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1401  CharUnits::fromQuantity(offset));
1402  addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1403  }
1404  return addr;
1405 }
1406 
1407 namespace {
1408 
1409 /// Encapsulates information about the way function arguments from
1410 /// CGFunctionInfo should be passed to actual LLVM IR function.
1411 class ClangToLLVMArgMapping {
1412  static const unsigned InvalidIndex = ~0U;
1413  unsigned InallocaArgNo;
1414  unsigned SRetArgNo;
1415  unsigned TotalIRArgs;
1416 
1417  /// Arguments of LLVM IR function corresponding to single Clang argument.
1418  struct IRArgs {
1419  unsigned PaddingArgIndex;
1420  // Argument is expanded to IR arguments at positions
1421  // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1422  unsigned FirstArgIndex;
1423  unsigned NumberOfArgs;
1424 
1425  IRArgs()
1426  : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1427  NumberOfArgs(0) {}
1428  };
1429 
1430  SmallVector<IRArgs, 8> ArgInfo;
1431 
1432 public:
1433  ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1434  bool OnlyRequiredArgs = false)
1435  : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1436  ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1437  construct(Context, FI, OnlyRequiredArgs);
1438  }
1439 
1440  bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1441  unsigned getInallocaArgNo() const {
1442  assert(hasInallocaArg());
1443  return InallocaArgNo;
1444  }
1445 
1446  bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1447  unsigned getSRetArgNo() const {
1448  assert(hasSRetArg());
1449  return SRetArgNo;
1450  }
1451 
1452  unsigned totalIRArgs() const { return TotalIRArgs; }
1453 
1454  bool hasPaddingArg(unsigned ArgNo) const {
1455  assert(ArgNo < ArgInfo.size());
1456  return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1457  }
1458  unsigned getPaddingArgNo(unsigned ArgNo) const {
1459  assert(hasPaddingArg(ArgNo));
1460  return ArgInfo[ArgNo].PaddingArgIndex;
1461  }
1462 
1463  /// Returns index of first IR argument corresponding to ArgNo, and their
1464  /// quantity.
1465  std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1466  assert(ArgNo < ArgInfo.size());
1467  return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1468  ArgInfo[ArgNo].NumberOfArgs);
1469  }
1470 
1471 private:
1472  void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1473  bool OnlyRequiredArgs);
1474 };
1475 
1476 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1477  const CGFunctionInfo &FI,
1478  bool OnlyRequiredArgs) {
1479  unsigned IRArgNo = 0;
1480  bool SwapThisWithSRet = false;
1481  const ABIArgInfo &RetAI = FI.getReturnInfo();
1482 
1483  if (RetAI.getKind() == ABIArgInfo::Indirect) {
1484  SwapThisWithSRet = RetAI.isSRetAfterThis();
1485  SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1486  }
1487 
1488  unsigned ArgNo = 0;
1489  unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1490  for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1491  ++I, ++ArgNo) {
1492  assert(I != FI.arg_end());
1493  QualType ArgType = I->type;
1494  const ABIArgInfo &AI = I->info;
1495  // Collect data about IR arguments corresponding to Clang argument ArgNo.
1496  auto &IRArgs = ArgInfo[ArgNo];
1497 
1498  if (AI.getPaddingType())
1499  IRArgs.PaddingArgIndex = IRArgNo++;
1500 
1501  switch (AI.getKind()) {
1502  case ABIArgInfo::Extend:
1503  case ABIArgInfo::Direct: {
1504  // FIXME: handle sseregparm someday...
1505  llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1506  if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1507  IRArgs.NumberOfArgs = STy->getNumElements();
1508  } else {
1509  IRArgs.NumberOfArgs = 1;
1510  }
1511  break;
1512  }
1513  case ABIArgInfo::Indirect:
1515  IRArgs.NumberOfArgs = 1;
1516  break;
1517  case ABIArgInfo::Ignore:
1518  case ABIArgInfo::InAlloca:
1519  // ignore and inalloca doesn't have matching LLVM parameters.
1520  IRArgs.NumberOfArgs = 0;
1521  break;
1523  IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1524  break;
1525  case ABIArgInfo::Expand:
1526  IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1527  break;
1528  }
1529 
1530  if (IRArgs.NumberOfArgs > 0) {
1531  IRArgs.FirstArgIndex = IRArgNo;
1532  IRArgNo += IRArgs.NumberOfArgs;
1533  }
1534 
1535  // Skip over the sret parameter when it comes second. We already handled it
1536  // above.
1537  if (IRArgNo == 1 && SwapThisWithSRet)
1538  IRArgNo++;
1539  }
1540  assert(ArgNo == ArgInfo.size());
1541 
1542  if (FI.usesInAlloca())
1543  InallocaArgNo = IRArgNo++;
1544 
1545  TotalIRArgs = IRArgNo;
1546 }
1547 } // namespace
1548 
1549 /***/
1550 
1552  const auto &RI = FI.getReturnInfo();
1553  return RI.isIndirect() || (RI.isInAlloca() && RI.getInAllocaSRet());
1554 }
1555 
1557  return ReturnTypeUsesSRet(FI) &&
1559 }
1560 
1562  if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1563  switch (BT->getKind()) {
1564  default:
1565  return false;
1566  case BuiltinType::Float:
1568  case BuiltinType::Double:
1570  case BuiltinType::LongDouble:
1572  }
1573  }
1574 
1575  return false;
1576 }
1577 
1579  if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1580  if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1581  if (BT->getKind() == BuiltinType::LongDouble)
1583  }
1584  }
1585 
1586  return false;
1587 }
1588 
1590  const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1591  return GetFunctionType(FI);
1592 }
1593 
1594 llvm::FunctionType *
1596 
1597  bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1598  (void)Inserted;
1599  assert(Inserted && "Recursively being processed?");
1600 
1601  llvm::Type *resultType = nullptr;
1602  const ABIArgInfo &retAI = FI.getReturnInfo();
1603  switch (retAI.getKind()) {
1604  case ABIArgInfo::Expand:
1606  llvm_unreachable("Invalid ABI kind for return argument");
1607 
1608  case ABIArgInfo::Extend:
1609  case ABIArgInfo::Direct:
1610  resultType = retAI.getCoerceToType();
1611  break;
1612 
1613  case ABIArgInfo::InAlloca:
1614  if (retAI.getInAllocaSRet()) {
1615  // sret things on win32 aren't void, they return the sret pointer.
1616  QualType ret = FI.getReturnType();
1617  llvm::Type *ty = ConvertType(ret);
1618  unsigned addressSpace = Context.getTargetAddressSpace(ret);
1619  resultType = llvm::PointerType::get(ty, addressSpace);
1620  } else {
1621  resultType = llvm::Type::getVoidTy(getLLVMContext());
1622  }
1623  break;
1624 
1625  case ABIArgInfo::Indirect:
1626  case ABIArgInfo::Ignore:
1627  resultType = llvm::Type::getVoidTy(getLLVMContext());
1628  break;
1629 
1631  resultType = retAI.getUnpaddedCoerceAndExpandType();
1632  break;
1633  }
1634 
1635  ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1636  SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1637 
1638  // Add type for sret argument.
1639  if (IRFunctionArgs.hasSRetArg()) {
1640  QualType Ret = FI.getReturnType();
1641  llvm::Type *Ty = ConvertType(Ret);
1642  unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1643  ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1644  llvm::PointerType::get(Ty, AddressSpace);
1645  }
1646 
1647  // Add type for inalloca argument.
1648  if (IRFunctionArgs.hasInallocaArg()) {
1649  auto ArgStruct = FI.getArgStruct();
1650  assert(ArgStruct);
1651  ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1652  }
1653 
1654  // Add in all of the required arguments.
1655  unsigned ArgNo = 0;
1657  ie = it + FI.getNumRequiredArgs();
1658  for (; it != ie; ++it, ++ArgNo) {
1659  const ABIArgInfo &ArgInfo = it->info;
1660 
1661  // Insert a padding type to ensure proper alignment.
1662  if (IRFunctionArgs.hasPaddingArg(ArgNo))
1663  ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1664  ArgInfo.getPaddingType();
1665 
1666  unsigned FirstIRArg, NumIRArgs;
1667  std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1668 
1669  switch (ArgInfo.getKind()) {
1670  case ABIArgInfo::Ignore:
1671  case ABIArgInfo::InAlloca:
1672  assert(NumIRArgs == 0);
1673  break;
1674 
1675  case ABIArgInfo::Indirect: {
1676  assert(NumIRArgs == 1);
1677  // indirect arguments are always on the stack, which is alloca addr space.
1678  llvm::Type *LTy = ConvertTypeForMem(it->type);
1679  ArgTypes[FirstIRArg] = LTy->getPointerTo(
1680  CGM.getDataLayout().getAllocaAddrSpace());
1681  break;
1682  }
1684  assert(NumIRArgs == 1);
1685  llvm::Type *LTy = ConvertTypeForMem(it->type);
1686  ArgTypes[FirstIRArg] = LTy->getPointerTo(ArgInfo.getIndirectAddrSpace());
1687  break;
1688  }
1689  case ABIArgInfo::Extend:
1690  case ABIArgInfo::Direct: {
1691  // Fast-isel and the optimizer generally like scalar values better than
1692  // FCAs, so we flatten them if this is safe to do for this argument.
1693  llvm::Type *argType = ArgInfo.getCoerceToType();
1694  llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1695  if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1696  assert(NumIRArgs == st->getNumElements());
1697  for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1698  ArgTypes[FirstIRArg + i] = st->getElementType(i);
1699  } else {
1700  assert(NumIRArgs == 1);
1701  ArgTypes[FirstIRArg] = argType;
1702  }
1703  break;
1704  }
1705 
1707  auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1708  for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1709  *ArgTypesIter++ = EltTy;
1710  }
1711  assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1712  break;
1713  }
1714 
1715  case ABIArgInfo::Expand:
1716  auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1717  getExpandedTypes(it->type, ArgTypesIter);
1718  assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1719  break;
1720  }
1721  }
1722 
1723  bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1724  assert(Erased && "Not in set?");
1725 
1726  return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1727 }
1728 
1730  const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1731  const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1732 
1733  if (!isFuncTypeConvertible(FPT))
1734  return llvm::StructType::get(getLLVMContext());
1735 
1736  return GetFunctionType(GD);
1737 }
1738 
1740  llvm::AttrBuilder &FuncAttrs,
1741  const FunctionProtoType *FPT) {
1742  if (!FPT)
1743  return;
1744 
1746  FPT->isNothrow())
1747  FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1748 }
1749 
1750 static void AddAttributesFromAssumes(llvm::AttrBuilder &FuncAttrs,
1751  const Decl *Callee, const Decl *Caller,
1752  bool AssumptionOnCallSite) {
1753  if (!Callee)
1754  return;
1755 
1757 
1758  for (const AssumptionAttr *AA : Callee->specific_attrs<AssumptionAttr>())
1759  AA->getAssumption().split(Attrs, ",");
1760 
1761  if (Caller && Caller->hasAttrs() && AssumptionOnCallSite)
1762  for (const AssumptionAttr *AA : Caller->specific_attrs<AssumptionAttr>())
1763  AA->getAssumption().split(Attrs, ",");
1764 
1765  if (!Attrs.empty())
1766  FuncAttrs.addAttribute(llvm::AssumptionAttrKey,
1767  llvm::join(Attrs.begin(), Attrs.end(), ","));
1768 }
1769 
1771  QualType ReturnType) {
1772  // We can't just discard the return value for a record type with a
1773  // complex destructor or a non-trivially copyable type.
1774  if (const RecordType *RT =
1775  ReturnType.getCanonicalType()->getAs<RecordType>()) {
1776  if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1777  return ClassDecl->hasTrivialDestructor();
1778  }
1779  return ReturnType.isTriviallyCopyableType(Context);
1780 }
1781 
1782 void CodeGenModule::getDefaultFunctionAttributes(StringRef Name,
1783  bool HasOptnone,
1784  bool AttrOnCallSite,
1785  llvm::AttrBuilder &FuncAttrs) {
1786  // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1787  if (!HasOptnone) {
1788  if (CodeGenOpts.OptimizeSize)
1789  FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1790  if (CodeGenOpts.OptimizeSize == 2)
1791  FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1792  }
1793 
1794  if (CodeGenOpts.DisableRedZone)
1795  FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1796  if (CodeGenOpts.IndirectTlsSegRefs)
1797  FuncAttrs.addAttribute("indirect-tls-seg-refs");
1798  if (CodeGenOpts.NoImplicitFloat)
1799  FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1800 
1801  if (AttrOnCallSite) {
1802  // Attributes that should go on the call site only.
1803  if (!CodeGenOpts.SimplifyLibCalls || LangOpts.isNoBuiltinFunc(Name))
1804  FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1805  if (!CodeGenOpts.TrapFuncName.empty())
1806  FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1807  } else {
1808  StringRef FpKind;
1809  switch (CodeGenOpts.getFramePointer()) {
1811  FpKind = "none";
1812  break;
1814  FpKind = "non-leaf";
1815  break;
1817  FpKind = "all";
1818  break;
1819  }
1820  FuncAttrs.addAttribute("frame-pointer", FpKind);
1821 
1822  if (CodeGenOpts.LessPreciseFPMAD)
1823  FuncAttrs.addAttribute("less-precise-fpmad", "true");
1824 
1825  if (CodeGenOpts.NullPointerIsValid)
1826  FuncAttrs.addAttribute(llvm::Attribute::NullPointerIsValid);
1827 
1828  if (CodeGenOpts.FPDenormalMode != llvm::DenormalMode::getIEEE())
1829  FuncAttrs.addAttribute("denormal-fp-math",
1830  CodeGenOpts.FPDenormalMode.str());
1831  if (CodeGenOpts.FP32DenormalMode != CodeGenOpts.FPDenormalMode) {
1832  FuncAttrs.addAttribute(
1833  "denormal-fp-math-f32",
1834  CodeGenOpts.FP32DenormalMode.str());
1835  }
1836 
1837  if (LangOpts.getFPExceptionMode() == LangOptions::FPE_Ignore)
1838  FuncAttrs.addAttribute("no-trapping-math", "true");
1839 
1840  // Strict (compliant) code is the default, so only add this attribute to
1841  // indicate that we are trying to workaround a problem case.
1842  if (!CodeGenOpts.StrictFloatCastOverflow)
1843  FuncAttrs.addAttribute("strict-float-cast-overflow", "false");
1844 
1845  // TODO: Are these all needed?
1846  // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1847  if (LangOpts.NoHonorInfs)
1848  FuncAttrs.addAttribute("no-infs-fp-math", "true");
1849  if (LangOpts.NoHonorNaNs)
1850  FuncAttrs.addAttribute("no-nans-fp-math", "true");
1851  if (LangOpts.ApproxFunc)
1852  FuncAttrs.addAttribute("approx-func-fp-math", "true");
1853  if (LangOpts.UnsafeFPMath)
1854  FuncAttrs.addAttribute("unsafe-fp-math", "true");
1855  if (CodeGenOpts.SoftFloat)
1856  FuncAttrs.addAttribute("use-soft-float", "true");
1857  FuncAttrs.addAttribute("stack-protector-buffer-size",
1858  llvm::utostr(CodeGenOpts.SSPBufferSize));
1859  if (LangOpts.NoSignedZero)
1860  FuncAttrs.addAttribute("no-signed-zeros-fp-math", "true");
1861 
1862  // TODO: Reciprocal estimate codegen options should apply to instructions?
1863  const std::vector<std::string> &Recips = CodeGenOpts.Reciprocals;
1864  if (!Recips.empty())
1865  FuncAttrs.addAttribute("reciprocal-estimates",
1866  llvm::join(Recips, ","));
1867 
1868  if (!CodeGenOpts.PreferVectorWidth.empty() &&
1869  CodeGenOpts.PreferVectorWidth != "none")
1870  FuncAttrs.addAttribute("prefer-vector-width",
1871  CodeGenOpts.PreferVectorWidth);
1872 
1873  if (CodeGenOpts.StackRealignment)
1874  FuncAttrs.addAttribute("stackrealign");
1875  if (CodeGenOpts.Backchain)
1876  FuncAttrs.addAttribute("backchain");
1877  if (CodeGenOpts.EnableSegmentedStacks)
1878  FuncAttrs.addAttribute("split-stack");
1879 
1880  if (CodeGenOpts.SpeculativeLoadHardening)
1881  FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening);
1882  }
1883 
1884  if (getLangOpts().assumeFunctionsAreConvergent()) {
1885  // Conservatively, mark all functions and calls in CUDA and OpenCL as
1886  // convergent (meaning, they may call an intrinsically convergent op, such
1887  // as __syncthreads() / barrier(), and so can't have certain optimizations
1888  // applied around them). LLVM will remove this attribute where it safely
1889  // can.
1890  FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1891  }
1892 
1893  if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1894  // Exceptions aren't supported in CUDA device code.
1895  FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1896  }
1897 
1898  for (StringRef Attr : CodeGenOpts.DefaultFunctionAttrs) {
1899  StringRef Var, Value;
1900  std::tie(Var, Value) = Attr.split('=');
1901  FuncAttrs.addAttribute(Var, Value);
1902  }
1903 }
1904 
1906  llvm::AttrBuilder FuncAttrs;
1907  getDefaultFunctionAttributes(F.getName(), F.hasOptNone(),
1908  /* AttrOnCallSite = */ false, FuncAttrs);
1909  // TODO: call GetCPUAndFeaturesAttributes?
1910  F.addFnAttrs(FuncAttrs);
1911 }
1912 
1914  llvm::AttrBuilder &attrs) {
1915  getDefaultFunctionAttributes(/*function name*/ "", /*optnone*/ false,
1916  /*for call*/ false, attrs);
1917  GetCPUAndFeaturesAttributes(GlobalDecl(), attrs);
1918 }
1919 
1920 static void addNoBuiltinAttributes(llvm::AttrBuilder &FuncAttrs,
1921  const LangOptions &LangOpts,
1922  const NoBuiltinAttr *NBA = nullptr) {
1923  auto AddNoBuiltinAttr = [&FuncAttrs](StringRef BuiltinName) {
1924  SmallString<32> AttributeName;
1925  AttributeName += "no-builtin-";
1926  AttributeName += BuiltinName;
1927  FuncAttrs.addAttribute(AttributeName);
1928  };
1929 
1930  // First, handle the language options passed through -fno-builtin.
1931  if (LangOpts.NoBuiltin) {
1932  // -fno-builtin disables them all.
1933  FuncAttrs.addAttribute("no-builtins");
1934  return;
1935  }
1936 
1937  // Then, add attributes for builtins specified through -fno-builtin-<name>.
1938  llvm::for_each(LangOpts.NoBuiltinFuncs, AddNoBuiltinAttr);
1939 
1940  // Now, let's check the __attribute__((no_builtin("...")) attribute added to
1941  // the source.
1942  if (!NBA)
1943  return;
1944 
1945  // If there is a wildcard in the builtin names specified through the
1946  // attribute, disable them all.
1947  if (llvm::is_contained(NBA->builtinNames(), "*")) {
1948  FuncAttrs.addAttribute("no-builtins");
1949  return;
1950  }
1951 
1952  // And last, add the rest of the builtin names.
1953  llvm::for_each(NBA->builtinNames(), AddNoBuiltinAttr);
1954 }
1955 
1956 static bool DetermineNoUndef(QualType QTy, CodeGenTypes &Types,
1957  const llvm::DataLayout &DL, const ABIArgInfo &AI,
1958  bool CheckCoerce = true) {
1959  llvm::Type *Ty = Types.ConvertTypeForMem(QTy);
1960  if (AI.getKind() == ABIArgInfo::Indirect)
1961  return true;
1962  if (AI.getKind() == ABIArgInfo::Extend)
1963  return true;
1964  if (!DL.typeSizeEqualsStoreSize(Ty))
1965  // TODO: This will result in a modest amount of values not marked noundef
1966  // when they could be. We care about values that *invisibly* contain undef
1967  // bits from the perspective of LLVM IR.
1968  return false;
1969  if (CheckCoerce && AI.canHaveCoerceToType()) {
1970  llvm::Type *CoerceTy = AI.getCoerceToType();
1971  if (llvm::TypeSize::isKnownGT(DL.getTypeSizeInBits(CoerceTy),
1972  DL.getTypeSizeInBits(Ty)))
1973  // If we're coercing to a type with a greater size than the canonical one,
1974  // we're introducing new undef bits.
1975  // Coercing to a type of smaller or equal size is ok, as we know that
1976  // there's no internal padding (typeSizeEqualsStoreSize).
1977  return false;
1978  }
1979  if (QTy->isExtIntType())
1980  return true;
1981  if (QTy->isReferenceType())
1982  return true;
1983  if (QTy->isNullPtrType())
1984  return false;
1985  if (QTy->isMemberPointerType())
1986  // TODO: Some member pointers are `noundef`, but it depends on the ABI. For
1987  // now, never mark them.
1988  return false;
1989  if (QTy->isScalarType()) {
1990  if (const ComplexType *Complex = dyn_cast<ComplexType>(QTy))
1991  return DetermineNoUndef(Complex->getElementType(), Types, DL, AI, false);
1992  return true;
1993  }
1994  if (const VectorType *Vector = dyn_cast<VectorType>(QTy))
1995  return DetermineNoUndef(Vector->getElementType(), Types, DL, AI, false);
1996  if (const MatrixType *Matrix = dyn_cast<MatrixType>(QTy))
1997  return DetermineNoUndef(Matrix->getElementType(), Types, DL, AI, false);
1998  if (const ArrayType *Array = dyn_cast<ArrayType>(QTy))
1999  return DetermineNoUndef(Array->getElementType(), Types, DL, AI, false);
2000 
2001  // TODO: Some structs may be `noundef`, in specific situations.
2002  return false;
2003 }
2004 
2005 /// Construct the IR attribute list of a function or call.
2006 ///
2007 /// When adding an attribute, please consider where it should be handled:
2008 ///
2009 /// - getDefaultFunctionAttributes is for attributes that are essentially
2010 /// part of the global target configuration (but perhaps can be
2011 /// overridden on a per-function basis). Adding attributes there
2012 /// will cause them to also be set in frontends that build on Clang's
2013 /// target-configuration logic, as well as for code defined in library
2014 /// modules such as CUDA's libdevice.
2015 ///
2016 /// - ConstructAttributeList builds on top of getDefaultFunctionAttributes
2017 /// and adds declaration-specific, convention-specific, and
2018 /// frontend-specific logic. The last is of particular importance:
2019 /// attributes that restrict how the frontend generates code must be
2020 /// added here rather than getDefaultFunctionAttributes.
2021 ///
2023  StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
2024  llvm::AttributeList &AttrList, unsigned &CallingConv, bool AttrOnCallSite,
2025  bool IsThunk, const Decl *Caller) {
2026  llvm::AttrBuilder FuncAttrs;
2027  llvm::AttrBuilder RetAttrs;
2028 
2029  // Collect function IR attributes from the CC lowering.
2030  // We'll collect the paramete and result attributes later.
2032  if (FI.isNoReturn())
2033  FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
2034  if (FI.isCmseNSCall())
2035  FuncAttrs.addAttribute("cmse_nonsecure_call");
2036 
2037  // Collect function IR attributes from the callee prototype if we have one.
2039  CalleeInfo.getCalleeFunctionProtoType());
2040 
2041  const Decl *TargetDecl = CalleeInfo.getCalleeDecl().getDecl();
2042 
2043  // Only attach assumptions to call sites in OpenMP mode.
2044  bool AssumptionOnCallSite = getLangOpts().OpenMP && AttrOnCallSite;
2045 
2046  // Attach assumption attributes to the declaration. If this is a call
2047  // site, attach assumptions from the caller to the call as well.
2048  AddAttributesFromAssumes(FuncAttrs, TargetDecl, Caller, AssumptionOnCallSite);
2049 
2050  bool HasOptnone = false;
2051  // The NoBuiltinAttr attached to the target FunctionDecl.
2052  const NoBuiltinAttr *NBA = nullptr;
2053 
2054  // Collect function IR attributes based on declaration-specific
2055  // information.
2056  // FIXME: handle sseregparm someday...
2057  if (TargetDecl) {
2058  if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
2059  FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
2060  if (TargetDecl->hasAttr<NoThrowAttr>())
2061  FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2062  if (TargetDecl->hasAttr<NoReturnAttr>())
2063  FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
2064  if (TargetDecl->hasAttr<ColdAttr>())
2065  FuncAttrs.addAttribute(llvm::Attribute::Cold);
2066  if (TargetDecl->hasAttr<HotAttr>())
2067  FuncAttrs.addAttribute(llvm::Attribute::Hot);
2068  if (TargetDecl->hasAttr<NoDuplicateAttr>())
2069  FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
2070  if (TargetDecl->hasAttr<ConvergentAttr>())
2071  FuncAttrs.addAttribute(llvm::Attribute::Convergent);
2072 
2073  if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
2075  getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
2076  if (AttrOnCallSite && Fn->isReplaceableGlobalAllocationFunction()) {
2077  // A sane operator new returns a non-aliasing pointer.
2078  auto Kind = Fn->getDeclName().getCXXOverloadedOperator();
2079  if (getCodeGenOpts().AssumeSaneOperatorNew &&
2080  (Kind == OO_New || Kind == OO_Array_New))
2081  RetAttrs.addAttribute(llvm::Attribute::NoAlias);
2082  }
2083  const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
2084  const bool IsVirtualCall = MD && MD->isVirtual();
2085  // Don't use [[noreturn]], _Noreturn or [[no_builtin]] for a call to a
2086  // virtual function. These attributes are not inherited by overloads.
2087  if (!(AttrOnCallSite && IsVirtualCall)) {
2088  if (Fn->isNoReturn())
2089  FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
2090  NBA = Fn->getAttr<NoBuiltinAttr>();
2091  }
2092  // Only place nomerge attribute on call sites, never functions. This
2093  // allows it to work on indirect virtual function calls.
2094  if (AttrOnCallSite && TargetDecl->hasAttr<NoMergeAttr>())
2095  FuncAttrs.addAttribute(llvm::Attribute::NoMerge);
2096  }
2097 
2098  // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
2099  if (TargetDecl->hasAttr<ConstAttr>()) {
2100  FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
2101  FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2102  // gcc specifies that 'const' functions have greater restrictions than
2103  // 'pure' functions, so they also cannot have infinite loops.
2104  FuncAttrs.addAttribute(llvm::Attribute::WillReturn);
2105  } else if (TargetDecl->hasAttr<PureAttr>()) {
2106  FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
2107  FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2108  // gcc specifies that 'pure' functions cannot have infinite loops.
2109  FuncAttrs.addAttribute(llvm::Attribute::WillReturn);
2110  } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
2111  FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
2112  FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2113  }
2114  if (TargetDecl->hasAttr<RestrictAttr>())
2115  RetAttrs.addAttribute(llvm::Attribute::NoAlias);
2116  if (TargetDecl->hasAttr<ReturnsNonNullAttr>() &&
2117  !CodeGenOpts.NullPointerIsValid)
2118  RetAttrs.addAttribute(llvm::Attribute::NonNull);
2119  if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
2120  FuncAttrs.addAttribute("no_caller_saved_registers");
2121  if (TargetDecl->hasAttr<AnyX86NoCfCheckAttr>())
2122  FuncAttrs.addAttribute(llvm::Attribute::NoCfCheck);
2123  if (TargetDecl->hasAttr<LeafAttr>())
2124  FuncAttrs.addAttribute(llvm::Attribute::NoCallback);
2125 
2126  HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
2127  if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
2128  Optional<unsigned> NumElemsParam;
2129  if (AllocSize->getNumElemsParam().isValid())
2130  NumElemsParam = AllocSize->getNumElemsParam().getLLVMIndex();
2131  FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam().getLLVMIndex(),
2132  NumElemsParam);
2133  }
2134 
2135  if (TargetDecl->hasAttr<OpenCLKernelAttr>()) {
2136  if (getLangOpts().OpenCLVersion <= 120) {
2137  // OpenCL v1.2 Work groups are always uniform
2138  FuncAttrs.addAttribute("uniform-work-group-size", "true");
2139  } else {
2140  // OpenCL v2.0 Work groups may be whether uniform or not.
2141  // '-cl-uniform-work-group-size' compile option gets a hint
2142  // to the compiler that the global work-size be a multiple of
2143  // the work-group size specified to clEnqueueNDRangeKernel
2144  // (i.e. work groups are uniform).
2145  FuncAttrs.addAttribute("uniform-work-group-size",
2146  llvm::toStringRef(CodeGenOpts.UniformWGSize));
2147  }
2148  }
2149  }
2150 
2151  // Attach "no-builtins" attributes to:
2152  // * call sites: both `nobuiltin` and "no-builtins" or "no-builtin-<name>".
2153  // * definitions: "no-builtins" or "no-builtin-<name>" only.
2154  // The attributes can come from:
2155  // * LangOpts: -ffreestanding, -fno-builtin, -fno-builtin-<name>
2156  // * FunctionDecl attributes: __attribute__((no_builtin(...)))
2157  addNoBuiltinAttributes(FuncAttrs, getLangOpts(), NBA);
2158 
2159  // Collect function IR attributes based on global settiings.
2160  getDefaultFunctionAttributes(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
2161 
2162  // Override some default IR attributes based on declaration-specific
2163  // information.
2164  if (TargetDecl) {
2165  if (TargetDecl->hasAttr<NoSpeculativeLoadHardeningAttr>())
2166  FuncAttrs.removeAttribute(llvm::Attribute::SpeculativeLoadHardening);
2167  if (TargetDecl->hasAttr<SpeculativeLoadHardeningAttr>())
2168  FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening);
2169  if (TargetDecl->hasAttr<NoSplitStackAttr>())
2170  FuncAttrs.removeAttribute("split-stack");
2171 
2172  // Add NonLazyBind attribute to function declarations when -fno-plt
2173  // is used.
2174  // FIXME: what if we just haven't processed the function definition
2175  // yet, or if it's an external definition like C99 inline?
2176  if (CodeGenOpts.NoPLT) {
2177  if (auto *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
2178  if (!Fn->isDefined() && !AttrOnCallSite) {
2179  FuncAttrs.addAttribute(llvm::Attribute::NonLazyBind);
2180  }
2181  }
2182  }
2183  }
2184 
2185  // Add "sample-profile-suffix-elision-policy" attribute for internal linkage
2186  // functions with -funique-internal-linkage-names.
2187  if (TargetDecl && CodeGenOpts.UniqueInternalLinkageNames) {
2188  if (isa<FunctionDecl>(TargetDecl)) {
2189  if (this->getFunctionLinkage(CalleeInfo.getCalleeDecl()) ==
2191  FuncAttrs.addAttribute("sample-profile-suffix-elision-policy",
2192  "selected");
2193  }
2194  }
2195 
2196  // Collect non-call-site function IR attributes from declaration-specific
2197  // information.
2198  if (!AttrOnCallSite) {
2199  if (TargetDecl && TargetDecl->hasAttr<CmseNSEntryAttr>())
2200  FuncAttrs.addAttribute("cmse_nonsecure_entry");
2201 
2202  // Whether tail calls are enabled.
2203  auto shouldDisableTailCalls = [&] {
2204  // Should this be honored in getDefaultFunctionAttributes?
2205  if (CodeGenOpts.DisableTailCalls)
2206  return true;
2207 
2208  if (!TargetDecl)
2209  return false;
2210 
2211  if (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
2212  TargetDecl->hasAttr<AnyX86InterruptAttr>())
2213  return true;
2214 
2215  if (CodeGenOpts.NoEscapingBlockTailCalls) {
2216  if (const auto *BD = dyn_cast<BlockDecl>(TargetDecl))
2217  if (!BD->doesNotEscape())
2218  return true;
2219  }
2220 
2221  return false;
2222  };
2223  if (shouldDisableTailCalls())
2224  FuncAttrs.addAttribute("disable-tail-calls", "true");
2225 
2226  // CPU/feature overrides. addDefaultFunctionDefinitionAttributes
2227  // handles these separately to set them based on the global defaults.
2228  GetCPUAndFeaturesAttributes(CalleeInfo.getCalleeDecl(), FuncAttrs);
2229  }
2230 
2231  // Collect attributes from arguments and return values.
2232  ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
2233 
2234  QualType RetTy = FI.getReturnType();
2235  const ABIArgInfo &RetAI = FI.getReturnInfo();
2236  const llvm::DataLayout &DL = getDataLayout();
2237 
2238  // C++ explicitly makes returning undefined values UB. C's rule only applies
2239  // to used values, so we never mark them noundef for now.
2240  bool HasStrictReturn = getLangOpts().CPlusPlus;
2241  if (TargetDecl && HasStrictReturn) {
2242  if (const FunctionDecl *FDecl = dyn_cast<FunctionDecl>(TargetDecl))
2243  HasStrictReturn &= !FDecl->isExternC();
2244  else if (const VarDecl *VDecl = dyn_cast<VarDecl>(TargetDecl))
2245  // Function pointer
2246  HasStrictReturn &= !VDecl->isExternC();
2247  }
2248 
2249  // We don't want to be too aggressive with the return checking, unless
2250  // it's explicit in the code opts or we're using an appropriate sanitizer.
2251  // Try to respect what the programmer intended.
2252  HasStrictReturn &= getCodeGenOpts().StrictReturn ||
2253  !MayDropFunctionReturn(getContext(), RetTy) ||
2254  getLangOpts().Sanitize.has(SanitizerKind::Memory) ||
2255  getLangOpts().Sanitize.has(SanitizerKind::Return);
2256 
2257  // Determine if the return type could be partially undef
2258  if (CodeGenOpts.EnableNoundefAttrs && HasStrictReturn) {
2259  if (!RetTy->isVoidType() && RetAI.getKind() != ABIArgInfo::Indirect &&
2260  DetermineNoUndef(RetTy, getTypes(), DL, RetAI))
2261  RetAttrs.addAttribute(llvm::Attribute::NoUndef);
2262  }
2263 
2264  switch (RetAI.getKind()) {
2265  case ABIArgInfo::Extend:
2266  if (RetAI.isSignExt())
2267  RetAttrs.addAttribute(llvm::Attribute::SExt);
2268  else
2269  RetAttrs.addAttribute(llvm::Attribute::ZExt);
2270  LLVM_FALLTHROUGH;
2271  case ABIArgInfo::Direct:
2272  if (RetAI.getInReg())
2273  RetAttrs.addAttribute(llvm::Attribute::InReg);
2274  break;
2275  case ABIArgInfo::Ignore:
2276  break;
2277 
2278  case ABIArgInfo::InAlloca:
2279  case ABIArgInfo::Indirect: {
2280  // inalloca and sret disable readnone and readonly
2281  FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2282  .removeAttribute(llvm::Attribute::ReadNone);
2283  break;
2284  }
2285 
2287  break;
2288 
2289  case ABIArgInfo::Expand:
2291  llvm_unreachable("Invalid ABI kind for return argument");
2292  }
2293 
2294  if (!IsThunk) {
2295  // FIXME: fix this properly, https://reviews.llvm.org/D100388
2296  if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
2297  QualType PTy = RefTy->getPointeeType();
2298  if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2299  RetAttrs.addDereferenceableAttr(
2300  getMinimumObjectSize(PTy).getQuantity());
2301  if (getContext().getTargetAddressSpace(PTy) == 0 &&
2302  !CodeGenOpts.NullPointerIsValid)
2303  RetAttrs.addAttribute(llvm::Attribute::NonNull);
2304  if (PTy->isObjectType()) {
2305  llvm::Align Alignment =
2307  RetAttrs.addAlignmentAttr(Alignment);
2308  }
2309  }
2310  }
2311 
2312  bool hasUsedSRet = false;
2313  SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
2314 
2315  // Attach attributes to sret.
2316  if (IRFunctionArgs.hasSRetArg()) {
2317  llvm::AttrBuilder SRETAttrs;
2318  SRETAttrs.addStructRetAttr(getTypes().ConvertTypeForMem(RetTy));
2319  hasUsedSRet = true;
2320  if (RetAI.getInReg())
2321  SRETAttrs.addAttribute(llvm::Attribute::InReg);
2322  SRETAttrs.addAlignmentAttr(RetAI.getIndirectAlign().getQuantity());
2323  ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
2324  llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
2325  }
2326 
2327  // Attach attributes to inalloca argument.
2328  if (IRFunctionArgs.hasInallocaArg()) {
2329  llvm::AttrBuilder Attrs;
2330  Attrs.addInAllocaAttr(FI.getArgStruct());
2331  ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
2332  llvm::AttributeSet::get(getLLVMContext(), Attrs);
2333  }
2334 
2335  // Apply `nonnull`, `dereferencable(N)` and `align N` to the `this` argument,
2336  // unless this is a thunk function.
2337  // FIXME: fix this properly, https://reviews.llvm.org/D100388
2338  if (FI.isInstanceMethod() && !IRFunctionArgs.hasInallocaArg() &&
2339  !FI.arg_begin()->type->isVoidPointerType() && !IsThunk) {
2340  auto IRArgs = IRFunctionArgs.getIRArgs(0);
2341 
2342  assert(IRArgs.second == 1 && "Expected only a single `this` pointer.");
2343 
2344  llvm::AttrBuilder Attrs;
2345 
2346  QualType ThisTy =
2347  FI.arg_begin()->type.castAs<PointerType>()->getPointeeType();
2348 
2349  if (!CodeGenOpts.NullPointerIsValid &&
2350  getContext().getTargetAddressSpace(FI.arg_begin()->type) == 0) {
2351  Attrs.addAttribute(llvm::Attribute::NonNull);
2352  Attrs.addDereferenceableAttr(getMinimumObjectSize(ThisTy).getQuantity());
2353  } else {
2354  // FIXME dereferenceable should be correct here, regardless of
2355  // NullPointerIsValid. However, dereferenceable currently does not always
2356  // respect NullPointerIsValid and may imply nonnull and break the program.
2357  // See https://reviews.llvm.org/D66618 for discussions.
2358  Attrs.addDereferenceableOrNullAttr(
2360  FI.arg_begin()->type.castAs<PointerType>()->getPointeeType())
2361  .getQuantity());
2362  }
2363 
2364  llvm::Align Alignment =
2365  getNaturalTypeAlignment(ThisTy, /*BaseInfo=*/nullptr,
2366  /*TBAAInfo=*/nullptr, /*forPointeeType=*/true)
2367  .getAsAlign();
2368  Attrs.addAlignmentAttr(Alignment);
2369 
2370  ArgAttrs[IRArgs.first] = llvm::AttributeSet::get(getLLVMContext(), Attrs);
2371  }
2372 
2373  unsigned ArgNo = 0;
2375  E = FI.arg_end();
2376  I != E; ++I, ++ArgNo) {
2377  QualType ParamType = I->type;
2378  const ABIArgInfo &AI = I->info;
2379  llvm::AttrBuilder Attrs;
2380 
2381  // Add attribute for padding argument, if necessary.
2382  if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
2383  if (AI.getPaddingInReg()) {
2384  ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
2385  llvm::AttributeSet::get(
2386  getLLVMContext(),
2387  llvm::AttrBuilder().addAttribute(llvm::Attribute::InReg));
2388  }
2389  }
2390 
2391  // Decide whether the argument we're handling could be partially undef
2392  bool ArgNoUndef = DetermineNoUndef(ParamType, getTypes(), DL, AI);
2393  if (CodeGenOpts.EnableNoundefAttrs && ArgNoUndef)
2394  Attrs.addAttribute(llvm::Attribute::NoUndef);
2395 
2396  // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
2397  // have the corresponding parameter variable. It doesn't make
2398  // sense to do it here because parameters are so messed up.
2399  switch (AI.getKind()) {
2400  case ABIArgInfo::Extend:
2401  if (AI.isSignExt())
2402  Attrs.addAttribute(llvm::Attribute::SExt);
2403  else
2404  Attrs.addAttribute(llvm::Attribute::ZExt);
2405  LLVM_FALLTHROUGH;
2406  case ABIArgInfo::Direct:
2407  if (ArgNo == 0 && FI.isChainCall())
2408  Attrs.addAttribute(llvm::Attribute::Nest);
2409  else if (AI.getInReg())
2410  Attrs.addAttribute(llvm::Attribute::InReg);
2411  Attrs.addStackAlignmentAttr(llvm::MaybeAlign(AI.getDirectAlign()));
2412  break;
2413 
2414  case ABIArgInfo::Indirect: {
2415  if (AI.getInReg())
2416  Attrs.addAttribute(llvm::Attribute::InReg);
2417 
2418  if (AI.getIndirectByVal())
2419  Attrs.addByValAttr(getTypes().ConvertTypeForMem(ParamType));
2420 
2421  auto *Decl = ParamType->getAsRecordDecl();
2422  if (CodeGenOpts.PassByValueIsNoAlias && Decl &&
2423  Decl->getArgPassingRestrictions() == RecordDecl::APK_CanPassInRegs)
2424  // When calling the function, the pointer passed in will be the only
2425  // reference to the underlying object. Mark it accordingly.
2426  Attrs.addAttribute(llvm::Attribute::NoAlias);
2427 
2428  // TODO: We could add the byref attribute if not byval, but it would
2429  // require updating many testcases.
2430 
2431  CharUnits Align = AI.getIndirectAlign();
2432 
2433  // In a byval argument, it is important that the required
2434  // alignment of the type is honored, as LLVM might be creating a
2435  // *new* stack object, and needs to know what alignment to give
2436  // it. (Sometimes it can deduce a sensible alignment on its own,
2437  // but not if clang decides it must emit a packed struct, or the
2438  // user specifies increased alignment requirements.)
2439  //
2440  // This is different from indirect *not* byval, where the object
2441  // exists already, and the align attribute is purely
2442  // informative.
2443  assert(!Align.isZero());
2444 
2445  // For now, only add this when we have a byval argument.
2446  // TODO: be less lazy about updating test cases.
2447  if (AI.getIndirectByVal())
2448  Attrs.addAlignmentAttr(Align.getQuantity());
2449 
2450  // byval disables readnone and readonly.
2451  FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2452  .removeAttribute(llvm::Attribute::ReadNone);
2453 
2454  break;
2455  }
2457  CharUnits Align = AI.getIndirectAlign();
2458  Attrs.addByRefAttr(getTypes().ConvertTypeForMem(ParamType));
2459  Attrs.addAlignmentAttr(Align.getQuantity());
2460  break;
2461  }
2462  case ABIArgInfo::Ignore:
2463  case ABIArgInfo::Expand:
2465  break;
2466 
2467  case ABIArgInfo::InAlloca:
2468  // inalloca disables readnone and readonly.
2469  FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2470  .removeAttribute(llvm::Attribute::ReadNone);
2471  continue;
2472  }
2473 
2474  if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2475  QualType PTy = RefTy->getPointeeType();
2476  if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2477  Attrs.addDereferenceableAttr(
2478  getMinimumObjectSize(PTy).getQuantity());
2479  if (getContext().getTargetAddressSpace(PTy) == 0 &&
2480  !CodeGenOpts.NullPointerIsValid)
2481  Attrs.addAttribute(llvm::Attribute::NonNull);
2482  if (PTy->isObjectType()) {
2483  llvm::Align Alignment =
2485  Attrs.addAlignmentAttr(Alignment);
2486  }
2487  }
2488 
2489  switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2491  break;
2492 
2494  // Add 'sret' if we haven't already used it for something, but
2495  // only if the result is void.
2496  if (!hasUsedSRet && RetTy->isVoidType()) {
2497  Attrs.addStructRetAttr(getTypes().ConvertTypeForMem(ParamType));
2498  hasUsedSRet = true;
2499  }
2500 
2501  // Add 'noalias' in either case.
2502  Attrs.addAttribute(llvm::Attribute::NoAlias);
2503 
2504  // Add 'dereferenceable' and 'alignment'.
2505  auto PTy = ParamType->getPointeeType();
2506  if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2507  auto info = getContext().getTypeInfoInChars(PTy);
2508  Attrs.addDereferenceableAttr(info.Width.getQuantity());
2509  Attrs.addAlignmentAttr(info.Align.getAsAlign());
2510  }
2511  break;
2512  }
2513 
2515  Attrs.addAttribute(llvm::Attribute::SwiftError);
2516  break;
2517 
2519  Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2520  break;
2521 
2523  Attrs.addAttribute(llvm::Attribute::SwiftAsync);
2524  break;
2525  }
2526 
2527  if (FI.getExtParameterInfo(ArgNo).isNoEscape())
2528  Attrs.addAttribute(llvm::Attribute::NoCapture);
2529 
2530  if (Attrs.hasAttributes()) {
2531  unsigned FirstIRArg, NumIRArgs;
2532  std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2533  for (unsigned i = 0; i < NumIRArgs; i++)
2534  ArgAttrs[FirstIRArg + i] =
2535  llvm::AttributeSet::get(getLLVMContext(), Attrs);
2536  }
2537  }
2538  assert(ArgNo == FI.arg_size());
2539 
2540  AttrList = llvm::AttributeList::get(
2541  getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2542  llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2543 }
2544 
2545 /// An argument came in as a promoted argument; demote it back to its
2546 /// declared type.
2548  const VarDecl *var,
2549  llvm::Value *value) {
2550  llvm::Type *varType = CGF.ConvertType(var->getType());
2551 
2552  // This can happen with promotions that actually don't change the
2553  // underlying type, like the enum promotions.
2554  if (value->getType() == varType) return value;
2555 
2556  assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2557  && "unexpected promotion type");
2558 
2559  if (isa<llvm::IntegerType>(varType))
2560  return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2561 
2562  return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2563 }
2564 
2565 /// Returns the attribute (either parameter attribute, or function
2566 /// attribute), which declares argument ArgNo to be non-null.
2567 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2568  QualType ArgType, unsigned ArgNo) {
2569  // FIXME: __attribute__((nonnull)) can also be applied to:
2570  // - references to pointers, where the pointee is known to be
2571  // nonnull (apparently a Clang extension)
2572  // - transparent unions containing pointers
2573  // In the former case, LLVM IR cannot represent the constraint. In
2574  // the latter case, we have no guarantee that the transparent union
2575  // is in fact passed as a pointer.
2576  if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2577  return nullptr;
2578  // First, check attribute on parameter itself.
2579  if (PVD) {
2580  if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2581  return ParmNNAttr;
2582  }
2583  // Check function attributes.
2584  if (!FD)
2585  return nullptr;
2586  for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2587  if (NNAttr->isNonNull(ArgNo))
2588  return NNAttr;
2589  }
2590  return nullptr;
2591 }
2592 
2593 namespace {
2594  struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2595  Address Temp;
2596  Address Arg;
2597  CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2598  void Emit(CodeGenFunction &CGF, Flags flags) override {
2599  llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2600  CGF.Builder.CreateStore(errorValue, Arg);
2601  }
2602  };
2603 }
2604 
2606  llvm::Function *Fn,
2607  const FunctionArgList &Args) {
2608  if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2609  // Naked functions don't have prologues.
2610  return;
2611 
2612  // If this is an implicit-return-zero function, go ahead and
2613  // initialize the return value. TODO: it might be nice to have
2614  // a more general mechanism for this that didn't require synthesized
2615  // return statements.
2616  if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2617  if (FD->hasImplicitReturnZero()) {
2618  QualType RetTy = FD->getReturnType().getUnqualifiedType();
2619  llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2620  llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2622  }
2623  }
2624 
2625  // FIXME: We no longer need the types from FunctionArgList; lift up and
2626  // simplify.
2627 
2628  ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2629  assert(Fn->arg_size() == IRFunctionArgs.totalIRArgs());
2630 
2631  // If we're using inalloca, all the memory arguments are GEPs off of the last
2632  // parameter, which is a pointer to the complete memory area.
2633  Address ArgStruct = Address::invalid();
2634  if (IRFunctionArgs.hasInallocaArg()) {
2635  ArgStruct = Address(Fn->getArg(IRFunctionArgs.getInallocaArgNo()),
2636  FI.getArgStructAlignment());
2637 
2638  assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2639  }
2640 
2641  // Name the struct return parameter.
2642  if (IRFunctionArgs.hasSRetArg()) {
2643  auto AI = Fn->getArg(IRFunctionArgs.getSRetArgNo());
2644  AI->setName("agg.result");
2645  AI->addAttr(llvm::Attribute::NoAlias);
2646  }
2647 
2648  // Track if we received the parameter as a pointer (indirect, byval, or
2649  // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2650  // into a local alloca for us.
2652  ArgVals.reserve(Args.size());
2653 
2654  // Create a pointer value for every parameter declaration. This usually
2655  // entails copying one or more LLVM IR arguments into an alloca. Don't push
2656  // any cleanups or do anything that might unwind. We do that separately, so
2657  // we can push the cleanups in the correct order for the ABI.
2658  assert(FI.arg_size() == Args.size() &&
2659  "Mismatch between function signature & arguments.");
2660  unsigned ArgNo = 0;
2662  for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2663  i != e; ++i, ++info_it, ++ArgNo) {
2664  const VarDecl *Arg = *i;
2665  const ABIArgInfo &ArgI = info_it->info;
2666 
2667  bool isPromoted =
2668  isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2669  // We are converting from ABIArgInfo type to VarDecl type directly, unless
2670  // the parameter is promoted. In this case we convert to
2671  // CGFunctionInfo::ArgInfo type with subsequent argument demotion.
2672  QualType Ty = isPromoted ? info_it->type : Arg->getType();
2673  assert(hasScalarEvaluationKind(Ty) ==
2675 
2676  unsigned FirstIRArg, NumIRArgs;
2677  std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2678 
2679  switch (ArgI.getKind()) {
2680  case ABIArgInfo::InAlloca: {
2681  assert(NumIRArgs == 0);
2682  auto FieldIndex = ArgI.getInAllocaFieldIndex();
2683  Address V =
2684  Builder.CreateStructGEP(ArgStruct, FieldIndex, Arg->getName());
2685  if (ArgI.getInAllocaIndirect())
2687  getContext().getTypeAlignInChars(Ty));
2688  ArgVals.push_back(ParamValue::forIndirect(V));
2689  break;
2690  }
2691 
2692  case ABIArgInfo::Indirect:
2694  assert(NumIRArgs == 1);
2695  Address ParamAddr =
2696  Address(Fn->getArg(FirstIRArg), ArgI.getIndirectAlign());
2697 
2698  if (!hasScalarEvaluationKind(Ty)) {
2699  // Aggregates and complex variables are accessed by reference. All we
2700  // need to do is realign the value, if requested. Also, if the address
2701  // may be aliased, copy it to ensure that the parameter variable is
2702  // mutable and has a unique adress, as C requires.
2703  Address V = ParamAddr;
2704  if (ArgI.getIndirectRealign() || ArgI.isIndirectAliased()) {
2705  Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2706 
2707  // Copy from the incoming argument pointer to the temporary with the
2708  // appropriate alignment.
2709  //
2710  // FIXME: We should have a common utility for generating an aggregate
2711  // copy.
2714  AlignedTemp.getPointer(), AlignedTemp.getAlignment().getAsAlign(),
2715  ParamAddr.getPointer(), ParamAddr.getAlignment().getAsAlign(),
2716  llvm::ConstantInt::get(IntPtrTy, Size.getQuantity()));
2717  V = AlignedTemp;
2718  }
2719  ArgVals.push_back(ParamValue::forIndirect(V));
2720  } else {
2721  // Load scalar value from indirect argument.
2722  llvm::Value *V =
2723  EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getBeginLoc());
2724 
2725  if (isPromoted)
2726  V = emitArgumentDemotion(*this, Arg, V);
2727  ArgVals.push_back(ParamValue::forDirect(V));
2728  }
2729  break;
2730  }
2731 
2732  case ABIArgInfo::Extend:
2733  case ABIArgInfo::Direct: {
2734  auto AI = Fn->getArg(FirstIRArg);
2735  llvm::Type *LTy = ConvertType(Arg->getType());
2736 
2737  // Prepare parameter attributes. So far, only attributes for pointer
2738  // parameters are prepared. See
2739  // http://llvm.org/docs/LangRef.html#paramattrs.
2740  if (ArgI.getDirectOffset() == 0 && LTy->isPointerTy() &&
2741  ArgI.getCoerceToType()->isPointerTy()) {
2742  assert(NumIRArgs == 1);
2743 
2744  if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2745  // Set `nonnull` attribute if any.
2746  if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2747  PVD->getFunctionScopeIndex()) &&
2748  !CGM.getCodeGenOpts().NullPointerIsValid)
2749  AI->addAttr(llvm::Attribute::NonNull);
2750 
2751  QualType OTy = PVD->getOriginalType();
2752  if (const auto *ArrTy =
2753  getContext().getAsConstantArrayType(OTy)) {
2754  // A C99 array parameter declaration with the static keyword also
2755  // indicates dereferenceability, and if the size is constant we can
2756  // use the dereferenceable attribute (which requires the size in
2757  // bytes).
2758  if (ArrTy->getSizeModifier() == ArrayType::Static) {
2759  QualType ETy = ArrTy->getElementType();
2760  llvm::Align Alignment =
2762  AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2763  uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2764  if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2765  ArrSize) {
2766  llvm::AttrBuilder Attrs;
2767  Attrs.addDereferenceableAttr(
2768  getContext().getTypeSizeInChars(ETy).getQuantity() *
2769  ArrSize);
2770  AI->addAttrs(Attrs);
2771  } else if (getContext().getTargetInfo().getNullPointerValue(
2772  ETy.getAddressSpace()) == 0 &&
2773  !CGM.getCodeGenOpts().NullPointerIsValid) {
2774  AI->addAttr(llvm::Attribute::NonNull);
2775  }
2776  }
2777  } else if (const auto *ArrTy =
2778  getContext().getAsVariableArrayType(OTy)) {
2779  // For C99 VLAs with the static keyword, we don't know the size so
2780  // we can't use the dereferenceable attribute, but in addrspace(0)
2781  // we know that it must be nonnull.
2782  if (ArrTy->getSizeModifier() == VariableArrayType::Static) {
2783  QualType ETy = ArrTy->getElementType();
2784  llvm::Align Alignment =
2786  AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2787  if (!getContext().getTargetAddressSpace(ETy) &&
2788  !CGM.getCodeGenOpts().NullPointerIsValid)
2789  AI->addAttr(llvm::Attribute::NonNull);
2790  }
2791  }
2792 
2793  // Set `align` attribute if any.
2794  const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2795  if (!AVAttr)
2796  if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2797  AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2798  if (AVAttr && !SanOpts.has(SanitizerKind::Alignment)) {
2799  // If alignment-assumption sanitizer is enabled, we do *not* add
2800  // alignment attribute here, but emit normal alignment assumption,
2801  // so the UBSAN check could function.
2802  llvm::ConstantInt *AlignmentCI =
2803  cast<llvm::ConstantInt>(EmitScalarExpr(AVAttr->getAlignment()));
2804  uint64_t AlignmentInt =
2805  AlignmentCI->getLimitedValue(llvm::Value::MaximumAlignment);
2806  if (AI->getParamAlign().valueOrOne() < AlignmentInt) {
2807  AI->removeAttr(llvm::Attribute::AttrKind::Alignment);
2808  AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(
2809  llvm::Align(AlignmentInt)));
2810  }
2811  }
2812  }
2813 
2814  // Set 'noalias' if an argument type has the `restrict` qualifier.
2815  if (Arg->getType().isRestrictQualified())
2816  AI->addAttr(llvm::Attribute::NoAlias);
2817  }
2818 
2819  // Prepare the argument value. If we have the trivial case, handle it
2820  // with no muss and fuss.
2821  if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2822  ArgI.getCoerceToType() == ConvertType(Ty) &&
2823  ArgI.getDirectOffset() == 0) {
2824  assert(NumIRArgs == 1);
2825 
2826  // LLVM expects swifterror parameters to be used in very restricted
2827  // ways. Copy the value into a less-restricted temporary.
2828  llvm::Value *V = AI;
2829  if (FI.getExtParameterInfo(ArgNo).getABI()
2831  QualType pointeeTy = Ty->getPointeeType();
2832  assert(pointeeTy->isPointerType());
2833  Address temp =
2834  CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2835  Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2836  llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2837  Builder.CreateStore(incomingErrorValue, temp);
2838  V = temp.getPointer();
2839 
2840  // Push a cleanup to copy the value back at the end of the function.
2841  // The convention does not guarantee that the value will be written
2842  // back if the function exits with an unwind exception.
2843  EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2844  }
2845 
2846  // Ensure the argument is the correct type.
2847  if (V->getType() != ArgI.getCoerceToType())
2849 
2850  if (isPromoted)
2851  V = emitArgumentDemotion(*this, Arg, V);
2852 
2853  // Because of merging of function types from multiple decls it is
2854  // possible for the type of an argument to not match the corresponding
2855  // type in the function type. Since we are codegening the callee
2856  // in here, add a cast to the argument type.
2857  llvm::Type *LTy = ConvertType(Arg->getType());
2858  if (V->getType() != LTy)
2859  V = Builder.CreateBitCast(V, LTy);
2860 
2861  ArgVals.push_back(ParamValue::forDirect(V));
2862  break;
2863  }
2864 
2865  // VLST arguments are coerced to VLATs at the function boundary for
2866  // ABI consistency. If this is a VLST that was coerced to
2867  // a VLAT at the function boundary and the types match up, use
2868  // llvm.experimental.vector.extract to convert back to the original
2869  // VLST.
2870  if (auto *VecTyTo = dyn_cast<llvm::FixedVectorType>(ConvertType(Ty))) {
2871  llvm::Value *Coerced = Fn->getArg(FirstIRArg);
2872  if (auto *VecTyFrom =
2873  dyn_cast<llvm::ScalableVectorType>(Coerced->getType())) {
2874  // If we are casting a scalable 16 x i1 predicate vector to a fixed i8
2875  // vector, bitcast the source and use a vector extract.
2876  auto PredType =
2877  llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
2878  if (VecTyFrom == PredType &&
2879  VecTyTo->getElementType() == Builder.getInt8Ty()) {
2880  VecTyFrom = llvm::ScalableVectorType::get(Builder.getInt8Ty(), 2);
2881  Coerced = Builder.CreateBitCast(Coerced, VecTyFrom);
2882  }
2883  if (VecTyFrom->getElementType() == VecTyTo->getElementType()) {
2884  llvm::Value *Zero = llvm::Constant::getNullValue(CGM.Int64Ty);
2885 
2886  assert(NumIRArgs == 1);
2887  Coerced->setName(Arg->getName() + ".coerce");
2888  ArgVals.push_back(ParamValue::forDirect(Builder.CreateExtractVector(
2889  VecTyTo, Coerced, Zero, "castFixedSve")));
2890  break;
2891  }
2892  }
2893  }
2894 
2895  Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2896  Arg->getName());
2897 
2898  // Pointer to store into.
2899  Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2900 
2901  // Fast-isel and the optimizer generally like scalar values better than
2902  // FCAs, so we flatten them if this is safe to do for this argument.
2903  llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2904  if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2905  STy->getNumElements() > 1) {
2906  uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2907  llvm::Type *DstTy = Ptr.getElementType();
2908  uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2909 
2910  Address AddrToStoreInto = Address::invalid();
2911  if (SrcSize <= DstSize) {
2912  AddrToStoreInto = Builder.CreateElementBitCast(Ptr, STy);
2913  } else {
2914  AddrToStoreInto =
2915  CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2916  }
2917 
2918  assert(STy->getNumElements() == NumIRArgs);
2919  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2920  auto AI = Fn->getArg(FirstIRArg + i);
2921  AI->setName(Arg->getName() + ".coerce" + Twine(i));
2922  Address EltPtr = Builder.CreateStructGEP(AddrToStoreInto, i);
2923  Builder.CreateStore(AI, EltPtr);
2924  }
2925 
2926  if (SrcSize > DstSize) {
2927  Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2928  }
2929 
2930  } else {
2931  // Simple case, just do a coerced store of the argument into the alloca.
2932  assert(NumIRArgs == 1);
2933  auto AI = Fn->getArg(FirstIRArg);
2934  AI->setName(Arg->getName() + ".coerce");
2935  CreateCoercedStore(AI, Ptr, /*DstIsVolatile=*/false, *this);
2936  }
2937 
2938  // Match to what EmitParmDecl is expecting for this type.
2940  llvm::Value *V =
2941  EmitLoadOfScalar(Alloca, false, Ty, Arg->getBeginLoc());
2942  if (isPromoted)
2943  V = emitArgumentDemotion(*this, Arg, V);
2944  ArgVals.push_back(ParamValue::forDirect(V));
2945  } else {
2946  ArgVals.push_back(ParamValue::forIndirect(Alloca));
2947  }
2948  break;
2949  }
2950 
2952  // Reconstruct into a temporary.
2953  Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2954  ArgVals.push_back(ParamValue::forIndirect(alloca));
2955 
2956  auto coercionType = ArgI.getCoerceAndExpandType();
2957  alloca = Builder.CreateElementBitCast(alloca, coercionType);
2958 
2959  unsigned argIndex = FirstIRArg;
2960  for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2961  llvm::Type *eltType = coercionType->getElementType(i);
2963  continue;
2964 
2965  auto eltAddr = Builder.CreateStructGEP(alloca, i);
2966  auto elt = Fn->getArg(argIndex++);
2967  Builder.CreateStore(elt, eltAddr);
2968  }
2969  assert(argIndex == FirstIRArg + NumIRArgs);
2970  break;
2971  }
2972 
2973  case ABIArgInfo::Expand: {
2974  // If this structure was expanded into multiple arguments then
2975  // we need to create a temporary and reconstruct it from the
2976  // arguments.
2977  Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2978  LValue LV = MakeAddrLValue(Alloca, Ty);
2979  ArgVals.push_back(ParamValue::forIndirect(Alloca));
2980 
2981  auto FnArgIter = Fn->arg_begin() + FirstIRArg;
2982  ExpandTypeFromArgs(Ty, LV, FnArgIter);
2983  assert(FnArgIter == Fn->arg_begin() + FirstIRArg + NumIRArgs);
2984  for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2985  auto AI = Fn->getArg(FirstIRArg + i);
2986  AI->setName(Arg->getName() + "." + Twine(i));
2987  }
2988  break;
2989  }
2990 
2991  case ABIArgInfo::Ignore:
2992  assert(NumIRArgs == 0);
2993  // Initialize the local variable appropriately.
2994  if (!hasScalarEvaluationKind(Ty)) {
2995  ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2996  } else {
2997  llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2998  ArgVals.push_back(ParamValue::forDirect(U));
2999  }
3000  break;
3001  }
3002  }
3003 
3004  if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3005  for (int I = Args.size() - 1; I >= 0; --I)
3006  EmitParmDecl(*Args[I], ArgVals[I], I + 1);
3007  } else {
3008  for (unsigned I = 0, E = Args.size(); I != E; ++I)
3009  EmitParmDecl(*Args[I], ArgVals[I], I + 1);
3010  }
3011 }
3012 
3013 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
3014  while (insn->use_empty()) {
3015  llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
3016  if (!bitcast) return;
3017 
3018  // This is "safe" because we would have used a ConstantExpr otherwise.
3019  insn = cast<llvm::Instruction>(bitcast->getOperand(0));
3020  bitcast->eraseFromParent();
3021  }
3022 }
3023 
3024 /// Try to emit a fused autorelease of a return result.
3026  llvm::Value *result) {
3027  // We must be immediately followed the cast.
3028  llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
3029  if (BB->empty()) return nullptr;
3030  if (&BB->back() != result) return nullptr;
3031 
3032  llvm::Type *resultType = result->getType();
3033 
3034  // result is in a BasicBlock and is therefore an Instruction.
3035  llvm::Instruction *generator = cast<llvm::Instruction>(result);
3036 
3038 
3039  // Look for:
3040  // %generator = bitcast %type1* %generator2 to %type2*
3041  while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
3042  // We would have emitted this as a constant if the operand weren't
3043  // an Instruction.
3044  generator = cast<llvm::Instruction>(bitcast->getOperand(0));
3045 
3046  // Require the generator to be immediately followed by the cast.
3047  if (generator->getNextNode() != bitcast)
3048  return nullptr;
3049 
3050  InstsToKill.push_back(bitcast);
3051  }
3052 
3053  // Look for:
3054  // %generator = call i8* @objc_retain(i8* %originalResult)
3055  // or
3056  // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
3057  llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
3058  if (!call) return nullptr;
3059 
3060  bool doRetainAutorelease;
3061 
3062  if (call->getCalledOperand() == CGF.CGM.getObjCEntrypoints().objc_retain) {
3063  doRetainAutorelease = true;
3064  } else if (call->getCalledOperand() ==
3066  doRetainAutorelease = false;
3067 
3068  // If we emitted an assembly marker for this call (and the
3069  // ARCEntrypoints field should have been set if so), go looking
3070  // for that call. If we can't find it, we can't do this
3071  // optimization. But it should always be the immediately previous
3072  // instruction, unless we needed bitcasts around the call.
3074  llvm::Instruction *prev = call->getPrevNode();
3075  assert(prev);
3076  if (isa<llvm::BitCastInst>(prev)) {
3077  prev = prev->getPrevNode();
3078  assert(prev);
3079  }
3080  assert(isa<llvm::CallInst>(prev));
3081  assert(cast<llvm::CallInst>(prev)->getCalledOperand() ==
3083  InstsToKill.push_back(prev);
3084  }
3085  } else {
3086  return nullptr;
3087  }
3088 
3089  result = call->getArgOperand(0);
3090  InstsToKill.push_back(call);
3091 
3092  // Keep killing bitcasts, for sanity. Note that we no longer care
3093  // about precise ordering as long as there's exactly one use.
3094  while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
3095  if (!bitcast->hasOneUse()) break;
3096  InstsToKill.push_back(bitcast);
3097  result = bitcast->getOperand(0);
3098  }
3099 
3100  // Delete all the unnecessary instructions, from latest to earliest.
3101  for (auto *I : InstsToKill)
3102  I->eraseFromParent();
3103 
3104  // Do the fused retain/autorelease if we were asked to.
3105  if (doRetainAutorelease)
3106  result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
3107 
3108  // Cast back to the result type.
3109  return CGF.Builder.CreateBitCast(result, resultType);
3110 }
3111 
3112 /// If this is a +1 of the value of an immutable 'self', remove it.
3114  llvm::Value *result) {
3115  // This is only applicable to a method with an immutable 'self'.
3116  const ObjCMethodDecl *method =
3117  dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
3118  if (!method) return nullptr;
3119  const VarDecl *self = method->getSelfDecl();
3120  if (!self->getType().isConstQualified()) return nullptr;
3121 
3122  // Look for a retain call.
3123  llvm::CallInst *retainCall =
3124  dyn_cast<llvm::CallInst>(result->stripPointerCasts());
3125  if (!retainCall || retainCall->getCalledOperand() !=
3127  return nullptr;
3128 
3129  // Look for an ordinary load of 'self'.
3130  llvm::Value *retainedValue = retainCall->getArgOperand(0);
3131  llvm::LoadInst *load =
3132  dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
3133  if (!load || load->isAtomic() || load->isVolatile() ||
3134  load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
3135  return nullptr;
3136 
3137  // Okay! Burn it all down. This relies for correctness on the
3138  // assumption that the retain is emitted as part of the return and
3139  // that thereafter everything is used "linearly".
3140  llvm::Type *resultType = result->getType();
3141  eraseUnusedBitCasts(cast<llvm::Instruction>(result));
3142  assert(retainCall->use_empty());
3143  retainCall->eraseFromParent();
3144  eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
3145 
3146  return CGF.Builder.CreateBitCast(load, resultType);
3147 }
3148 
3149 /// Emit an ARC autorelease of the result of a function.
3150 ///
3151 /// \return the value to actually return from the function
3153  llvm::Value *result) {
3154  // If we're returning 'self', kill the initial retain. This is a
3155  // heuristic attempt to "encourage correctness" in the really unfortunate
3156  // case where we have a return of self during a dealloc and we desperately
3157  // need to avoid the possible autorelease.
3158  if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
3159  return self;
3160 
3161  // At -O0, try to emit a fused retain/autorelease.
3162  if (CGF.shouldUseFusedARCCalls())
3163  if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
3164  return fused;
3165 
3166  return CGF.EmitARCAutoreleaseReturnValue(result);
3167 }
3168 
3169 /// Heuristically search for a dominating store to the return-value slot.
3170 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
3171  // Check if a User is a store which pointerOperand is the ReturnValue.
3172  // We are looking for stores to the ReturnValue, not for stores of the
3173  // ReturnValue to some other location.
3174  auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
3175  auto *SI = dyn_cast<llvm::StoreInst>(U);
3176  if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
3177  return nullptr;
3178  // These aren't actually possible for non-coerced returns, and we
3179  // only care about non-coerced returns on this code path.
3180  assert(!SI->isAtomic() && !SI->isVolatile());
3181  return SI;
3182  };
3183  // If there are multiple uses of the return-value slot, just check
3184  // for something immediately preceding the IP. Sometimes this can
3185  // happen with how we generate implicit-returns; it can also happen
3186  // with noreturn cleanups.
3187  if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
3188  llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
3189  if (IP->empty()) return nullptr;
3190  llvm::Instruction *I = &IP->back();
3191 
3192  // Skip lifetime markers
3193  for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
3194  IE = IP->rend();
3195  II != IE; ++II) {
3196  if (llvm::IntrinsicInst *Intrinsic =
3197  dyn_cast<llvm::IntrinsicInst>(&*II)) {
3198  if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
3199  const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
3200  ++II;
3201  if (II == IE)
3202  break;
3203  if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
3204  continue;
3205  }
3206  }
3207  I = &*II;
3208  break;
3209  }
3210 
3211  return GetStoreIfValid(I);
3212  }
3213 
3214  llvm::StoreInst *store =
3215  GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
3216  if (!store) return nullptr;
3217 
3218  // Now do a first-and-dirty dominance check: just walk up the
3219  // single-predecessors chain from the current insertion point.
3220  llvm::BasicBlock *StoreBB = store->getParent();
3221  llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
3222  while (IP != StoreBB) {
3223  if (!(IP = IP->getSinglePredecessor()))
3224  return nullptr;
3225  }
3226 
3227  // Okay, the store's basic block dominates the insertion point; we
3228  // can do our thing.
3229  return store;
3230 }
3231 
3232 // Helper functions for EmitCMSEClearRecord
3233 
3234 // Set the bits corresponding to a field having width `BitWidth` and located at
3235 // offset `BitOffset` (from the least significant bit) within a storage unit of
3236 // `Bits.size()` bytes. Each element of `Bits` corresponds to one target byte.
3237 // Use little-endian layout, i.e.`Bits[0]` is the LSB.
3238 static void setBitRange(SmallVectorImpl<uint64_t> &Bits, int BitOffset,
3239  int BitWidth, int CharWidth) {
3240  assert(CharWidth <= 64);
3241  assert(static_cast<unsigned>(BitWidth) <= Bits.size() * CharWidth);
3242 
3243  int Pos = 0;
3244  if (BitOffset >= CharWidth) {
3245  Pos += BitOffset / CharWidth;
3246  BitOffset = BitOffset % CharWidth;
3247  }
3248 
3249  const uint64_t Used = (uint64_t(1) << CharWidth) - 1;
3250  if (BitOffset + BitWidth >= CharWidth) {
3251  Bits[Pos++] |= (Used << BitOffset) & Used;
3252  BitWidth -= CharWidth - BitOffset;
3253  BitOffset = 0;
3254  }
3255 
3256  while (BitWidth >= CharWidth) {
3257  Bits[Pos++] = Used;
3258  BitWidth -= CharWidth;
3259  }
3260 
3261  if (BitWidth > 0)
3262  Bits[Pos++] |= (Used >> (CharWidth - BitWidth)) << BitOffset;
3263 }
3264 
3265 // Set the bits corresponding to a field having width `BitWidth` and located at
3266 // offset `BitOffset` (from the least significant bit) within a storage unit of
3267 // `StorageSize` bytes, located at `StorageOffset` in `Bits`. Each element of
3268 // `Bits` corresponds to one target byte. Use target endian layout.
3269 static void setBitRange(SmallVectorImpl<uint64_t> &Bits, int StorageOffset,
3270  int StorageSize, int BitOffset, int BitWidth,
3271  int CharWidth, bool BigEndian) {
3272 
3273  SmallVector<uint64_t, 8> TmpBits(StorageSize);
3274  setBitRange(TmpBits, BitOffset, BitWidth, CharWidth);
3275 
3276  if (BigEndian)
3277  std::reverse(TmpBits.begin(), TmpBits.end());
3278 
3279  for (uint64_t V : TmpBits)
3280  Bits[StorageOffset++] |= V;
3281 }
3282 
3283 static void setUsedBits(CodeGenModule &, QualType, int,
3284  SmallVectorImpl<uint64_t> &);
3285 
3286 // Set the bits in `Bits`, which correspond to the value representations of
3287 // the actual members of the record type `RTy`. Note that this function does
3288 // not handle base classes, virtual tables, etc, since they cannot happen in
3289 // CMSE function arguments or return. The bit mask corresponds to the target
3290 // memory layout, i.e. it's endian dependent.
3291 static void setUsedBits(CodeGenModule &CGM, const RecordType *RTy, int Offset,
3292  SmallVectorImpl<uint64_t> &Bits) {
3293  ASTContext &Context = CGM.getContext();
3294  int CharWidth = Context.getCharWidth();
3295  const RecordDecl *RD = RTy->getDecl()->getDefinition();
3296  const ASTRecordLayout &ASTLayout = Context.getASTRecordLayout(RD);
3297  const CGRecordLayout &Layout = CGM.getTypes().getCGRecordLayout(RD);
3298 
3299  int Idx = 0;
3300  for (auto I = RD->field_begin(), E = RD->field_end(); I != E; ++I, ++Idx) {
3301  const FieldDecl *F = *I;
3302 
3303  if (F->isUnnamedBitfield() || F->isZeroLengthBitField(Context) ||
3305  continue;
3306 
3307  if (F->isBitField()) {
3308  const CGBitFieldInfo &BFI = Layout.getBitFieldInfo(F);
3310  BFI.StorageSize / CharWidth, BFI.Offset,
3311  BFI.Size, CharWidth,
3312  CGM.getDataLayout().isBigEndian());
3313  continue;
3314  }
3315 
3316  setUsedBits(CGM, F->getType(),
3317  Offset + ASTLayout.getFieldOffset(Idx) / CharWidth, Bits);
3318  }
3319 }
3320 
3321 // Set the bits in `Bits`, which correspond to the value representations of
3322 // the elements of an array type `ATy`.
3323 static void setUsedBits(CodeGenModule &CGM, const ConstantArrayType *ATy,
3324  int Offset, SmallVectorImpl<uint64_t> &Bits) {
3325  const ASTContext &Context = CGM.getContext();
3326 
3327  QualType ETy = Context.getBaseElementType(ATy);
3328  int Size = Context.getTypeSizeInChars(ETy).getQuantity();
3329  SmallVector<uint64_t, 4> TmpBits(Size);
3330  setUsedBits(CGM, ETy, 0, TmpBits);
3331 
3332  for (int I = 0, N = Context.getConstantArrayElementCount(ATy); I < N; ++I) {
3333  auto Src = TmpBits.begin();
3334  auto Dst = Bits.begin() + Offset + I * Size;
3335  for (int J = 0; J < Size; ++J)
3336  *Dst++ |= *Src++;
3337  }
3338 }
3339 
3340 // Set the bits in `Bits`, which correspond to the value representations of
3341 // the type `QTy`.
3342 static void setUsedBits(CodeGenModule &CGM, QualType QTy, int Offset,
3343  SmallVectorImpl<uint64_t> &Bits) {
3344  if (const auto *RTy = QTy->getAs<RecordType>())
3345  return setUsedBits(CGM, RTy, Offset, Bits);
3346 
3347  ASTContext &Context = CGM.getContext();
3348  if (const auto *ATy = Context.getAsConstantArrayType(QTy))
3349  return setUsedBits(CGM, ATy, Offset, Bits);
3350 
3351  int Size = Context.getTypeSizeInChars(QTy).getQuantity();
3352  if (Size <= 0)
3353  return;
3354 
3355  std::fill_n(Bits.begin() + Offset, Size,
3356  (uint64_t(1) << Context.getCharWidth()) - 1);
3357 }
3358 
3359 static uint64_t buildMultiCharMask(const SmallVectorImpl<uint64_t> &Bits,
3360  int Pos, int Size, int CharWidth,
3361  bool BigEndian) {
3362  assert(Size > 0);
3363  uint64_t Mask = 0;
3364  if (BigEndian) {
3365  for (auto P = Bits.begin() + Pos, E = Bits.begin() + Pos + Size; P != E;
3366  ++P)
3367  Mask = (Mask << CharWidth) | *P;
3368  } else {
3369  auto P = Bits.begin() + Pos + Size, End = Bits.begin() + Pos;
3370  do
3371  Mask = (Mask << CharWidth) | *--P;
3372  while (P != End);
3373  }
3374  return Mask;
3375 }
3376 
3377 // Emit code to clear the bits in a record, which aren't a part of any user
3378 // declared member, when the record is a function return.
3380  llvm::IntegerType *ITy,
3381  QualType QTy) {
3382  assert(Src->getType() == ITy);
3383  assert(ITy->getScalarSizeInBits() <= 64);
3384 
3385  const llvm::DataLayout &DataLayout = CGM.getDataLayout();
3386  int Size = DataLayout.getTypeStoreSize(ITy);
3387  SmallVector<uint64_t, 4> Bits(Size);
3388  setUsedBits(CGM, QTy->castAs<RecordType>(), 0, Bits);
3389 
3390  int CharWidth = CGM.getContext().getCharWidth();
3391  uint64_t Mask =
3392  buildMultiCharMask(Bits, 0, Size, CharWidth, DataLayout.isBigEndian());
3393 
3394  return Builder.CreateAnd(Src, Mask, "cmse.clear");
3395 }
3396 
3397 // Emit code to clear the bits in a record, which aren't a part of any user
3398 // declared member, when the record is a function argument.
3400  llvm::ArrayType *ATy,
3401  QualType QTy) {
3402  const llvm::DataLayout &DataLayout = CGM.getDataLayout();
3403  int Size = DataLayout.getTypeStoreSize(ATy);
3404  SmallVector<uint64_t, 16> Bits(Size);
3405  setUsedBits(CGM, QTy->castAs<RecordType>(), 0, Bits);
3406 
3407  // Clear each element of the LLVM array.
3408  int CharWidth = CGM.getContext().getCharWidth();
3409  int CharsPerElt =
3410  ATy->getArrayElementType()->getScalarSizeInBits() / CharWidth;
3411  int MaskIndex = 0;
3412  llvm::Value *R = llvm::UndefValue::get(ATy);
3413  for (int I = 0, N = ATy->getArrayNumElements(); I != N; ++I) {
3414  uint64_t Mask = buildMultiCharMask(Bits, MaskIndex, CharsPerElt, CharWidth,
3415  DataLayout.isBigEndian());
3416  MaskIndex += CharsPerElt;
3417  llvm::Value *T0 = Builder.CreateExtractValue(Src, I);
3418  llvm::Value *T1 = Builder.CreateAnd(T0, Mask, "cmse.clear");
3419  R = Builder.CreateInsertValue(R, T1, I);
3420  }
3421 
3422  return R;
3423 }
3424 
3426  bool EmitRetDbgLoc,
3427  SourceLocation EndLoc) {
3428  if (FI.isNoReturn()) {
3429  // Noreturn functions don't return.
3430  EmitUnreachable(EndLoc);
3431  return;
3432  }
3433 
3434  if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
3435  // Naked functions don't have epilogues.
3436  Builder.CreateUnreachable();
3437  return;
3438  }
3439 
3440  // Functions with no result always return void.
3441  if (!ReturnValue.isValid()) {
3442  Builder.CreateRetVoid();
3443  return;
3444  }
3445 
3446  llvm::DebugLoc RetDbgLoc;
3447  llvm::Value *RV = nullptr;
3448  QualType RetTy = FI.getReturnType();
3449  const ABIArgInfo &RetAI = FI.getReturnInfo();
3450 
3451  switch (RetAI.getKind()) {
3452  case ABIArgInfo::InAlloca:
3453  // Aggregrates get evaluated directly into the destination. Sometimes we
3454  // need to return the sret value in a register, though.
3455  assert(hasAggregateEvaluationKind(RetTy));
3456  if (RetAI.getInAllocaSRet()) {
3457  llvm::Function::arg_iterator EI = CurFn->arg_end();
3458  --EI;
3459  llvm::Value *ArgStruct = &*EI;
3461  EI->getType()->getPointerElementType(), ArgStruct,
3462  RetAI.getInAllocaFieldIndex());
3463  llvm::Type *Ty =
3464  cast<llvm::GetElementPtrInst>(SRet)->getResultElementType();
3465  RV = Builder.CreateAlignedLoad(Ty, SRet, getPointerAlign(), "sret");
3466  }
3467  break;
3468 
3469  case ABIArgInfo::Indirect: {
3470  auto AI = CurFn->arg_begin();
3471  if (RetAI.isSRetAfterThis())
3472  ++AI;
3473  switch (getEvaluationKind(RetTy)) {
3474  case TEK_Complex: {
3475  ComplexPairTy RT =
3476  EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
3478  /*isInit*/ true);
3479  break;
3480  }
3481  case TEK_Aggregate:
3482  // Do nothing; aggregrates get evaluated directly into the destination.
3483  break;
3484  case TEK_Scalar:
3486  MakeNaturalAlignAddrLValue(&*AI, RetTy),
3487  /*isInit*/ true);
3488  break;
3489  }
3490  break;
3491  }
3492 
3493  case ABIArgInfo::Extend:
3494  case ABIArgInfo::Direct:
3495  if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
3496  RetAI.getDirectOffset() == 0) {
3497  // The internal return value temp always will have pointer-to-return-type
3498  // type, just do a load.
3499 
3500  // If there is a dominating store to ReturnValue, we can elide
3501  // the load, zap the store, and usually zap the alloca.
3502  if (llvm::StoreInst *SI =
3504  // Reuse the debug location from the store unless there is
3505  // cleanup code to be emitted between the store and return
3506  // instruction.
3507  if (EmitRetDbgLoc && !AutoreleaseResult)
3508  RetDbgLoc = SI->getDebugLoc();
3509  // Get the stored value and nuke the now-dead store.
3510  RV = SI->getValueOperand();
3511  SI->eraseFromParent();
3512 
3513  // Otherwise, we have to do a simple load.
3514  } else {
3516  }
3517  } else {
3518  // If the value is offset in memory, apply the offset now.
3519  Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
3520 
3521  RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
3522  }
3523 
3524  // In ARC, end functions that return a retainable type with a call
3525  // to objc_autoreleaseReturnValue.
3526  if (AutoreleaseResult) {
3527 #ifndef NDEBUG
3528  // Type::isObjCRetainabletype has to be called on a QualType that hasn't
3529  // been stripped of the typedefs, so we cannot use RetTy here. Get the
3530  // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
3531  // CurCodeDecl or BlockInfo.
3532  QualType RT;
3533 
3534  if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
3535  RT = FD->getReturnType();
3536  else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
3537  RT = MD->getReturnType();
3538  else if (isa<BlockDecl>(CurCodeDecl))
3540  else
3541  llvm_unreachable("Unexpected function/method type");
3542 
3543  assert(getLangOpts().ObjCAutoRefCount &&
3544  !FI.isReturnsRetained() &&
3545  RT->isObjCRetainableType());
3546 #endif
3547  RV = emitAutoreleaseOfResult(*this, RV);
3548  }
3549 
3550  break;
3551 
3552  case ABIArgInfo::Ignore:
3553  break;
3554 
3556  auto coercionType = RetAI.getCoerceAndExpandType();
3557 
3558  // Load all of the coerced elements out into results.
3560  Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
3561  for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3562  auto coercedEltType = coercionType->getElementType(i);
3563  if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
3564  continue;
3565 
3566  auto eltAddr = Builder.CreateStructGEP(addr, i);
3567  auto elt = Builder.CreateLoad(eltAddr);
3568  results.push_back(elt);
3569  }
3570 
3571  // If we have one result, it's the single direct result type.
3572  if (results.size() == 1) {
3573  RV = results[0];
3574 
3575  // Otherwise, we need to make a first-class aggregate.
3576  } else {
3577  // Construct a return type that lacks padding elements.
3578  llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
3579 
3580  RV = llvm::UndefValue::get(returnType);
3581  for (unsigned i = 0, e = results.size(); i != e; ++i) {
3582  RV = Builder.CreateInsertValue(RV, results[i], i);
3583  }
3584  }
3585  break;
3586  }
3587  case ABIArgInfo::Expand:
3589  llvm_unreachable("Invalid ABI kind for return argument");
3590  }
3591 
3592  llvm::Instruction *Ret;
3593  if (RV) {
3594  if (CurFuncDecl && CurFuncDecl->hasAttr<CmseNSEntryAttr>()) {
3595  // For certain return types, clear padding bits, as they may reveal
3596  // sensitive information.
3597  // Small struct/union types are passed as integers.
3598  auto *ITy = dyn_cast<llvm::IntegerType>(RV->getType());
3599  if (ITy != nullptr && isa<RecordType>(RetTy.getCanonicalType()))
3600  RV = EmitCMSEClearRecord(RV, ITy, RetTy);
3601  }
3603  Ret = Builder.CreateRet(RV);
3604  } else {
3605  Ret = Builder.CreateRetVoid();
3606  }
3607 
3608  if (RetDbgLoc)
3609  Ret->setDebugLoc(std::move(RetDbgLoc));
3610 }
3611 
3613  // A current decl may not be available when emitting vtable thunks.
3614  if (!CurCodeDecl)
3615  return;
3616 
3617  // If the return block isn't reachable, neither is this check, so don't emit
3618  // it.
3619  if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty())
3620  return;
3621 
3622  ReturnsNonNullAttr *RetNNAttr = nullptr;
3623  if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
3624  RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
3625 
3626  if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
3627  return;
3628 
3629  // Prefer the returns_nonnull attribute if it's present.
3630  SourceLocation AttrLoc;
3631  SanitizerMask CheckKind;
3632  SanitizerHandler Handler;
3633  if (RetNNAttr) {
3634  assert(!requiresReturnValueNullabilityCheck() &&
3635  "Cannot check nullability and the nonnull attribute");
3636  AttrLoc = RetNNAttr->getLocation();
3637  CheckKind = SanitizerKind::ReturnsNonnullAttribute;
3638  Handler = SanitizerHandler::NonnullReturn;
3639  } else {
3640  if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
3641  if (auto *TSI = DD->getTypeSourceInfo())
3642  if (auto FTL = TSI->getTypeLoc().getAsAdjusted<FunctionTypeLoc>())
3643  AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
3644  CheckKind = SanitizerKind::NullabilityReturn;
3645  Handler = SanitizerHandler::NullabilityReturn;
3646  }
3647 
3648  SanitizerScope SanScope(this);
3649 
3650  // Make sure the "return" source location is valid. If we're checking a
3651  // nullability annotation, make sure the preconditions for the check are met.
3652  llvm::BasicBlock *Check = createBasicBlock("nullcheck");
3653  llvm::BasicBlock *NoCheck = createBasicBlock("no.nullcheck");
3654  llvm::Value *SLocPtr = Builder.CreateLoad(ReturnLocation, "return.sloc.load");
3655  llvm::Value *CanNullCheck = Builder.CreateIsNotNull(SLocPtr);
3656  if (requiresReturnValueNullabilityCheck())
3657  CanNullCheck =
3658  Builder.CreateAnd(CanNullCheck, RetValNullabilityPrecondition);
3659  Builder.CreateCondBr(CanNullCheck, Check, NoCheck);
3660  EmitBlock(Check);
3661 
3662  // Now do the null check.
3663  llvm::Value *Cond = Builder.CreateIsNotNull(RV);
3664  llvm::Constant *StaticData[] = {EmitCheckSourceLocation(AttrLoc)};
3665  llvm::Value *DynamicData[] = {SLocPtr};
3666  EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, DynamicData);
3667 
3668  EmitBlock(NoCheck);
3669 
3670 #ifndef NDEBUG
3671  // The return location should not be used after the check has been emitted.
3672  ReturnLocation = Address::invalid();
3673 #endif
3674 }
3675 
3677  const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3678  return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
3679 }
3680 
3682  QualType Ty) {
3683  // FIXME: Generate IR in one pass, rather than going back and fixing up these
3684  // placeholders.
3685  llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
3686  llvm::Type *IRPtrTy = IRTy->getPointerTo();
3687  llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
3688 
3689  // FIXME: When we generate this IR in one pass, we shouldn't need
3690  // this win32-specific alignment hack.
3692  Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
3693 
3694  return AggValueSlot::forAddr(Address(Placeholder, Align),
3695  Ty.getQualifiers(),
3700 }
3701 
3703  const VarDecl *param,
3704  SourceLocation loc) {
3705  // StartFunction converted the ABI-lowered parameter(s) into a
3706  // local alloca. We need to turn that into an r-value suitable
3707  // for EmitCall.
3708  Address local = GetAddrOfLocalVar(param);
3709 
3710  QualType type = param->getType();
3711 
3713  CGM.ErrorUnsupported(param, "forwarded non-trivially copyable parameter");
3714  }
3715 
3716  // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3717  // but the argument needs to be the original pointer.
3718  if (type->isReferenceType()) {
3719  args.add(RValue::get(Builder.CreateLoad(local)), type);
3720 
3721  // In ARC, move out of consumed arguments so that the release cleanup
3722  // entered by StartFunction doesn't cause an over-release. This isn't
3723  // optimal -O0 code generation, but it should get cleaned up when
3724  // optimization is enabled. This also assumes that delegate calls are
3725  // performed exactly once for a set of arguments, but that should be safe.
3726  } else if (getLangOpts().ObjCAutoRefCount &&
3727  param->hasAttr<NSConsumedAttr>() &&
3728  type->isObjCRetainableType()) {
3729  llvm::Value *ptr = Builder.CreateLoad(local);
3730  auto null =
3731  llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3732  Builder.CreateStore(null, local);
3733  args.add(RValue::get(ptr), type);
3734 
3735  // For the most part, we just need to load the alloca, except that
3736  // aggregate r-values are actually pointers to temporaries.
3737  } else {
3738  args.add(convertTempToRValue(local, type, loc), type);
3739  }
3740 
3741  // Deactivate the cleanup for the callee-destructed param that was pushed.
3742  if (type->isRecordType() && !CurFuncIsThunk &&
3743  type->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee() &&
3744  param->needsDestruction(getContext())) {
3746  CalleeDestructedParamCleanups.lookup(cast<ParmVarDecl>(param));
3747  assert(cleanup.isValid() &&
3748  "cleanup for callee-destructed param not recorded");
3749  // This unreachable is a temporary marker which will be removed later.
3750  llvm::Instruction *isActive = Builder.CreateUnreachable();
3751  args.addArgCleanupDeactivation(cleanup, isActive);
3752  }
3753 }
3754 
3755 static bool isProvablyNull(llvm::Value *addr) {
3756  return isa<llvm::ConstantPointerNull>(addr);
3757 }
3758 
3759 /// Emit the actual writing-back of a writeback.
3761  const CallArgList::Writeback &writeback) {
3762  const LValue &srcLV = writeback.Source;
3763  Address srcAddr = srcLV.getAddress(CGF);
3764  assert(!isProvablyNull(srcAddr.getPointer()) &&
3765  "shouldn't have writeback for provably null argument");
3766 
3767  llvm::BasicBlock *contBB = nullptr;
3768 
3769  // If the argument wasn't provably non-null, we need to null check
3770  // before doing the store.
3771  bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3772  CGF.CGM.getDataLayout());
3773  if (!provablyNonNull) {
3774  llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3775  contBB = CGF.createBasicBlock("icr.done");
3776 
3777  llvm::Value *isNull =
3778  CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3779  CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3780  CGF.EmitBlock(writebackBB);
3781  }
3782 
3783  // Load the value to writeback.
3784  llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3785 
3786  // Cast it back, in case we're writing an id to a Foo* or something.
3787  value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3788  "icr.writeback-cast");
3789 
3790  // Perform the writeback.
3791 
3792  // If we have a "to use" value, it's something we need to emit a use
3793  // of. This has to be carefully threaded in: if it's done after the
3794  // release it's potentially undefined behavior (and the optimizer
3795  // will ignore it), and if it happens before the retain then the
3796  // optimizer could move the release there.
3797  if (writeback.ToUse) {
3798  assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3799 
3800  // Retain the new value. No need to block-copy here: the block's
3801  // being passed up the stack.
3802  value = CGF.EmitARCRetainNonBlock(value);
3803 
3804  // Emit the intrinsic use here.
3805  CGF.EmitARCIntrinsicUse(writeback.ToUse);
3806 
3807  // Load the old value (primitively).
3808  llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3809 
3810  // Put the new value in place (primitively).
3811  CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3812 
3813  // Release the old value.
3814  CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3815 
3816  // Otherwise, we can just do a normal lvalue store.
3817  } else {
3818  CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3819  }
3820 
3821  // Jump to the continuation block.
3822  if (!provablyNonNull)
3823  CGF.EmitBlock(contBB);
3824 }
3825 
3827  const CallArgList &args) {
3828  for (const auto &I : args.writebacks())
3829  emitWriteback(CGF, I);
3830 }
3831 
3833  const CallArgList &CallArgs) {
3835  CallArgs.getCleanupsToDeactivate();
3836  // Iterate in reverse to increase the likelihood of popping the cleanup.
3837  for (const auto &I : llvm::reverse(Cleanups)) {
3838  CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3839  I.IsActiveIP->eraseFromParent();
3840  }
3841 }
3842 
3843 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3844  if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3845  if (uop->getOpcode() == UO_AddrOf)
3846  return uop->getSubExpr();
3847  return nullptr;
3848 }
3849 
3850 /// Emit an argument that's being passed call-by-writeback. That is,
3851 /// we are passing the address of an __autoreleased temporary; it
3852 /// might be copy-initialized with the current value of the given
3853 /// address, but it will definitely be copied out of after the call.
3855  const ObjCIndirectCopyRestoreExpr *CRE) {
3856  LValue srcLV;
3857 
3858  // Make an optimistic effort to emit the address as an l-value.
3859  // This can fail if the argument expression is more complicated.
3860  if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3861  srcLV = CGF.EmitLValue(lvExpr);
3862 
3863  // Otherwise, just emit it as a scalar.
3864  } else {
3865  Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3866 
3867  QualType srcAddrType =
3868  CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3869  srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3870  }
3871  Address srcAddr = srcLV.getAddress(CGF);
3872 
3873  // The dest and src types don't necessarily match in LLVM terms
3874  // because of the crazy ObjC compatibility rules.
3875 
3876  llvm::PointerType *destType =
3877  cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3878 
3879  // If the address is a constant null, just pass the appropriate null.
3880  if (isProvablyNull(srcAddr.getPointer())) {
3881  args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3882  CRE->getType());
3883  return;
3884  }
3885 
3886  // Create the temporary.
3887  Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3888  CGF.getPointerAlign(),
3889  "icr.temp");
3890  // Loading an l-value can introduce a cleanup if the l-value is __weak,
3891  // and that cleanup will be conditional if we can't prove that the l-value
3892  // isn't null, so we need to register a dominating point so that the cleanups
3893  // system will make valid IR.
3895 
3896  // Zero-initialize it if we're not doing a copy-initialization.
3897  bool shouldCopy = CRE->shouldCopy();
3898  if (!shouldCopy) {
3899  llvm::Value *null =
3900  llvm::ConstantPointerNull::get(
3901  cast<llvm::PointerType>(destType->getElementType()));
3902  CGF.Builder.CreateStore(null, temp);
3903  }
3904 
3905  llvm::BasicBlock *contBB = nullptr;
3906  llvm::BasicBlock *originBB = nullptr;
3907 
3908  // If the address is *not* known to be non-null, we need to switch.
3909  llvm::Value *finalArgument;
3910 
3911  bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3912  CGF.CGM.getDataLayout());
3913  if (provablyNonNull) {
3914  finalArgument = temp.getPointer();
3915  } else {
3916  llvm::Value *isNull =
3917  CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3918 
3919  finalArgument = CGF.Builder.CreateSelect(isNull,
3920  llvm::ConstantPointerNull::get(destType),
3921  temp.getPointer(), "icr.argument");
3922 
3923  // If we need to copy, then the load has to be conditional, which
3924  // means we need control flow.
3925  if (shouldCopy) {
3926  originBB = CGF.Builder.GetInsertBlock();
3927  contBB = CGF.createBasicBlock("icr.cont");
3928  llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3929  CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3930  CGF.EmitBlock(copyBB);
3931  condEval.begin(CGF);
3932  }
3933  }
3934 
3935  llvm::Value *valueToUse = nullptr;
3936 
3937  // Perform a copy if necessary.
3938  if (shouldCopy) {
3939  RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3940  assert(srcRV.isScalar());
3941 
3942  llvm::Value *src = srcRV.getScalarVal();
3943  src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3944  "icr.cast");
3945 
3946  // Use an ordinary store, not a store-to-lvalue.
3947  CGF.Builder.CreateStore(src, temp);
3948 
3949  // If optimization is enabled, and the value was held in a
3950  // __strong variable, we need to tell the optimizer that this
3951  // value has to stay alive until we're doing the store back.
3952  // This is because the temporary is effectively unretained,
3953  // and so otherwise we can violate the high-level semantics.
3954  if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3956  valueToUse = src;
3957  }
3958  }
3959 
3960  // Finish the control flow if we needed it.
3961  if (shouldCopy && !provablyNonNull) {
3962  llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3963  CGF.EmitBlock(contBB);
3964 
3965  // Make a phi for the value to intrinsically use.
3966  if (valueToUse) {
3967  llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3968  "icr.to-use");
3969  phiToUse->addIncoming(valueToUse, copyBB);
3970  phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3971  originBB);
3972  valueToUse = phiToUse;
3973  }
3974 
3975  condEval.end(CGF);
3976  }
3977 
3978  args.addWriteback(srcLV, temp, valueToUse);
3979  args.add(RValue::get(finalArgument), CRE->getType());
3980 }
3981 
3983  assert(!StackBase);
3984 
3985  // Save the stack.
3986  llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3987  StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3988 }
3989 
3991  if (StackBase) {
3992  // Restore the stack after the call.
3993  llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3994  CGF.Builder.CreateCall(F, StackBase);
3995  }
3996 }
3997 
3999  SourceLocation ArgLoc,
4000  AbstractCallee AC,
4001  unsigned ParmNum) {
4002  if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
4003  SanOpts.has(SanitizerKind::NullabilityArg)))
4004  return;
4005 
4006  // The param decl may be missing in a variadic function.
4007  auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
4008  unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
4009 
4010  // Prefer the nonnull attribute if it's present.
4011  const NonNullAttr *NNAttr = nullptr;
4012  if (SanOpts.has(SanitizerKind::NonnullAttribute))
4013  NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
4014 
4015  bool CanCheckNullability = false;
4016  if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
4017  auto Nullability = PVD->getType()->getNullability(getContext());
4018  CanCheckNullability = Nullability &&
4020  PVD->getTypeSourceInfo();
4021  }
4022 
4023  if (!NNAttr && !CanCheckNullability)
4024  return;
4025 
4026  SourceLocation AttrLoc;
4027  SanitizerMask CheckKind;
4028  SanitizerHandler Handler;
4029  if (NNAttr) {
4030  AttrLoc = NNAttr->getLocation();
4031  CheckKind = SanitizerKind::NonnullAttribute;
4032  Handler = SanitizerHandler::NonnullArg;
4033  } else {
4034  AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
4035  CheckKind = SanitizerKind::NullabilityArg;
4036  Handler = SanitizerHandler::NullabilityArg;
4037  }
4038 
4039  SanitizerScope SanScope(this);
4040  llvm::Value *Cond = EmitNonNullRValueCheck(RV, ArgType);
4041  llvm::Constant *StaticData[] = {
4043  llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
4044  };
4045  EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
4046 }
4047 
4048 // Check if the call is going to use the inalloca convention. This needs to
4049 // agree with CGFunctionInfo::usesInAlloca. The CGFunctionInfo is arranged
4050 // later, so we can't check it directly.
4051 static bool hasInAllocaArgs(CodeGenModule &CGM, CallingConv ExplicitCC,
4052  ArrayRef<QualType> ArgTypes) {
4053  // The Swift calling conventions don't go through the target-specific
4054  // argument classification, they never use inalloca.
4055  // TODO: Consider limiting inalloca use to only calling conventions supported
4056  // by MSVC.
4057  if (ExplicitCC == CC_Swift || ExplicitCC == CC_SwiftAsync)
4058  return false;
4059  if (!CGM.getTarget().getCXXABI().isMicrosoft())
4060  return false;
4061  return llvm::any_of(ArgTypes, [&](QualType Ty) {
4062  return isInAllocaArgument(CGM.getCXXABI(), Ty);
4063  });
4064 }
4065 
4066 #ifndef NDEBUG
4067 // Determine whether the given argument is an Objective-C method
4068 // that may have type parameters in its signature.
4069 static bool isObjCMethodWithTypeParams(const ObjCMethodDecl *method) {
4070  const DeclContext *dc = method->getDeclContext();
4071  if (const ObjCInterfaceDecl *classDecl = dyn_cast<ObjCInterfaceDecl>(dc)) {
4072  return classDecl->getTypeParamListAsWritten();
4073  }
4074 
4075  if (const ObjCCategoryDecl *catDecl = dyn_cast<ObjCCategoryDecl>(dc)) {
4076  return catDecl->getTypeParamList();
4077  }
4078 
4079  return false;
4080 }
4081 #endif
4082 
4083 /// EmitCallArgs - Emit call arguments for a function.
4085  CallArgList &Args, PrototypeWrapper Prototype,
4086  llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
4087  AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
4088  SmallVector<QualType, 16> ArgTypes;
4089 
4090  assert((ParamsToSkip == 0 || Prototype.P) &&
4091  "Can't skip parameters if type info is not provided");
4092 
4093  // This variable only captures *explicitly* written conventions, not those
4094  // applied by default via command line flags or target defaults, such as
4095  // thiscall, aapcs, stdcall via -mrtd, etc. Computing that correctly would
4096  // require knowing if this is a C++ instance method or being able to see
4097  // unprototyped FunctionTypes.
4098  CallingConv ExplicitCC = CC_C;
4099 
4100  // First, if a prototype was provided, use those argument types.
4101  bool IsVariadic = false;
4102  if (Prototype.P) {
4103  const auto *MD = Prototype.P.dyn_cast<const ObjCMethodDecl *>();
4104  if (MD) {
4105  IsVariadic = MD->isVariadic();
4106  ExplicitCC = getCallingConventionForDecl(
4107  MD, CGM.getTarget().getTriple().isOSWindows());
4108  ArgTypes.assign(MD->param_type_begin() + ParamsToSkip,
4109  MD->param_type_end());
4110  } else {
4111  const auto *FPT = Prototype.P.get<const FunctionProtoType *>();
4112  IsVariadic = FPT->isVariadic();
4113  ExplicitCC = FPT->getExtInfo().getCC();
4114  ArgTypes.assign(FPT->param_type_begin() + ParamsToSkip,
4115  FPT->param_type_end());
4116  }
4117 
4118 #ifndef NDEBUG
4119  // Check that the prototyped types match the argument expression types.
4120  bool isGenericMethod = MD && isObjCMethodWithTypeParams(MD);
4121  CallExpr::const_arg_iterator Arg = ArgRange.begin();
4122  for (QualType Ty : ArgTypes) {
4123  assert(Arg != ArgRange.end() && "Running over edge of argument list!");
4124  assert(
4125  (isGenericMethod || Ty->isVariablyModifiedType() ||
4127  getContext()
4128  .getCanonicalType(Ty.getNonReferenceType())
4129  .getTypePtr() ==
4130  getContext().getCanonicalType((*Arg)->getType()).getTypePtr()) &&
4131  "type mismatch in call argument!");
4132  ++Arg;
4133  }
4134 
4135  // Either we've emitted all the call args, or we have a call to variadic
4136  // function.
4137  assert((Arg == ArgRange.end() || IsVariadic) &&
4138  "Extra arguments in non-variadic function!");
4139 #endif
4140  }
4141 
4142  // If we still have any arguments, emit them using the type of the argument.
4143  for (auto *A : llvm::make_range(std::next(ArgRange.begin(), ArgTypes.size()),
4144  ArgRange.end()))
4145  ArgTypes.push_back(IsVariadic ? getVarArgType(A) : A->getType());
4146  assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
4147 
4148  // We must evaluate arguments from right to left in the MS C++ ABI,
4149  // because arguments are destroyed left to right in the callee. As a special
4150  // case, there are certain language constructs that require left-to-right
4151  // evaluation, and in those cases we consider the evaluation order requirement
4152  // to trump the "destruction order is reverse construction order" guarantee.
4153  bool LeftToRight =
4157 
4158  auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
4159  RValue EmittedArg) {
4160  if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
4161  return;
4162  auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
4163  if (PS == nullptr)
4164  return;
4165 
4166  const auto &Context = getContext();
4167  auto SizeTy = Context.getSizeType();
4168  auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
4169  assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
4170  llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
4171  EmittedArg.getScalarVal(),
4172  PS->isDynamic());
4173  Args.add(RValue::get(V), SizeTy);
4174  // If we're emitting args in reverse, be sure to do so with
4175  // pass_object_size, as well.
4176  if (!LeftToRight)
4177  std::swap(Args.back(), *(&Args.back() - 1));
4178  };
4179 
4180  // Insert a stack save if we're going to need any inalloca args.
4181  if (hasInAllocaArgs(CGM, ExplicitCC, ArgTypes)) {
4182  assert(getTarget().getTriple().getArch() == llvm::Triple::x86 &&
4183  "inalloca only supported on x86");
4184  Args.allocateArgumentMemory(*this);
4185  }
4186 
4187  // Evaluate each argument in the appropriate order.
4188  size_t CallArgsStart = Args.size();
4189  for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
4190  unsigned Idx = LeftToRight ? I : E - I - 1;
4191  CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
4192  unsigned InitialArgSize = Args.size();
4193  // If *Arg is an ObjCIndirectCopyRestoreExpr, check that either the types of
4194  // the argument and parameter match or the objc method is parameterized.
4195  assert((!isa<ObjCIndirectCopyRestoreExpr>(*Arg) ||
4196  getContext().hasSameUnqualifiedType((*Arg)->getType(),
4197  ArgTypes[Idx]) ||
4198  (isa<ObjCMethodDecl>(AC.getDecl()) &&
4199  isObjCMethodWithTypeParams(cast<ObjCMethodDecl>(AC.getDecl())))) &&
4200  "Argument and parameter types don't match");
4201  EmitCallArg(Args, *Arg, ArgTypes[Idx]);
4202  // In particular, we depend on it being the last arg in Args, and the
4203  // objectsize bits depend on there only being one arg if !LeftToRight.
4204  assert(InitialArgSize + 1 == Args.size() &&
4205  "The code below depends on only adding one arg per EmitCallArg");
4206  (void)InitialArgSize;
4207  // Since pointer argument are never emitted as LValue, it is safe to emit
4208  // non-null argument check for r-value only.
4209  if (!Args.back().hasLValue()) {
4210  RValue RVArg = Args.back().getKnownRValue();
4211  EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
4212  ParamsToSkip + Idx);
4213  // @llvm.objectsize should never have side-effects and shouldn't need
4214  // destruction/cleanups, so we can safely "emit" it after its arg,
4215  // regardless of right-to-leftness
4216  MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
4217  }
4218  }
4219 
4220  if (!LeftToRight) {
4221  // Un-reverse the arguments we just evaluated so they match up with the LLVM
4222  // IR function.
4223  std::reverse(Args.begin() + CallArgsStart, Args.end());
4224  }
4225 }
4226 
4227 namespace {
4228 
4229 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
4230  DestroyUnpassedArg(Address Addr, QualType Ty)
4231  : Addr(Addr), Ty(Ty) {}
4232 
4233  Address Addr;
4234  QualType Ty;
4235 
4236  void Emit(CodeGenFunction &CGF, Flags flags) override {
4238  if (DtorKind == QualType::DK_cxx_destructor) {
4239  const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
4240  assert(!Dtor->isTrivial());
4241  CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
4242  /*Delegating=*/false, Addr, Ty);
4243  } else {
4244  CGF.callCStructDestructor(CGF.MakeAddrLValue(Addr, Ty));
4245  }
4246  }
4247 };
4248 
4249 struct DisableDebugLocationUpdates {
4250  CodeGenFunction &CGF;
4251  bool disabledDebugInfo;
4252  DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
4253  if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
4254  CGF.disableDebugInfo();
4255  }
4256  ~DisableDebugLocationUpdates() {
4257  if (disabledDebugInfo)
4258  CGF.enableDebugInfo();
4259  }
4260 };
4261 
4262 } // end anonymous namespace
4263 
4265  if (!HasLV)
4266  return RV;
4267  LValue Copy = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty), Ty);
4269  LV.isVolatile());
4270  IsUsed = true;
4271  return RValue::getAggregate(Copy.getAddress(CGF));
4272 }
4273 
4275  LValue Dst = CGF.MakeAddrLValue(Addr, Ty);
4276  if (!HasLV && RV.isScalar())
4277  CGF.EmitStoreOfScalar(RV.getScalarVal(), Dst, /*isInit=*/true);
4278  else if (!HasLV && RV.isComplex())
4279  CGF.EmitStoreOfComplex(RV.getComplexVal(), Dst, /*init=*/true);
4280  else {
4281  auto Addr = HasLV ? LV.getAddress(CGF) : RV.getAggregateAddress();
4282  LValue SrcLV = CGF.MakeAddrLValue(Addr, Ty);
4283  // We assume that call args are never copied into subobjects.
4285  HasLV ? LV.isVolatileQualified()
4286  : RV.isVolatileQualified());
4287  }
4288  IsUsed = true;
4289 }
4290 
4292  QualType type) {
4293  DisableDebugLocationUpdates Dis(*this, E);
4294  if (const ObjCIndirectCopyRestoreExpr *CRE
4295  = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
4296  assert(getLangOpts().ObjCAutoRefCount);
4297  return emitWritebackArg(*this, args, CRE);
4298  }
4299 
4300  assert(type->isReferenceType() == E->isGLValue() &&
4301  "reference binding to unmaterialized r-value!");
4302 
4303  if (E->isGLValue()) {
4304  assert(E->getObjectKind() == OK_Ordinary);
4305  return args.add(EmitReferenceBindingToExpr(E), type);
4306  }
4307 
4308  bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
4309 
4310  // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
4311  // However, we still have to push an EH-only cleanup in case we unwind before
4312  // we make it to the call.
4313  if (type->isRecordType() &&
4314  type->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) {
4315  // If we're using inalloca, use the argument memory. Otherwise, use a
4316  // temporary.
4317  AggValueSlot Slot;
4318  if (args.isUsingInAlloca())
4319  Slot = createPlaceholderSlot(*this, type);
4320  else
4321  Slot = CreateAggTemp(type, "agg.tmp");
4322 
4323  bool DestroyedInCallee = true, NeedsEHCleanup = true;
4324  if (const auto *RD = type->getAsCXXRecordDecl())
4325  DestroyedInCallee = RD->hasNonTrivialDestructor();
4326  else
4327  NeedsEHCleanup = needsEHCleanup(type.isDestructedType());
4328 
4329  if (DestroyedInCallee)
4330  Slot.setExternallyDestructed();
4331 
4332  EmitAggExpr(E, Slot);
4333  RValue RV = Slot.asRValue();
4334  args.add(RV, type);
4335 
4336  if (DestroyedInCallee && NeedsEHCleanup) {
4337  // Create a no-op GEP between the placeholder and the cleanup so we can
4338  // RAUW it successfully. It also serves as a marker of the first
4339  // instruction where the cleanup is active.
4340  pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
4341  type);
4342  // This unreachable is a temporary marker which will be removed later.
4343  llvm::Instruction *IsActive = Builder.CreateUnreachable();
4345  }
4346  return;
4347  }
4348 
4349  if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
4350  cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
4351  LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
4352  assert(L.isSimple());
4353  args.addUncopiedAggregate(L, type);
4354  return;
4355  }
4356 
4357  args.add(EmitAnyExprToTemp(E), type);
4358 }
4359 
4360 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
4361  // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
4362  // implicitly widens null pointer constants that are arguments to varargs
4363  // functions to pointer-sized ints.
4364  if (!getTarget().getTriple().isOSWindows())
4365  return Arg->getType();
4366 
4367  if (Arg->getType()->isIntegerType() &&
4368  getContext().getTypeSize(Arg->getType()) <
4369  getContext().getTargetInfo().getPointerWidth(0) &&
4372  return getContext().getIntPtrType();
4373  }
4374 
4375  return Arg->getType();
4376 }
4377 
4378 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4379 // optimizer it can aggressively ignore unwind edges.
4380 void
4381 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
4382  if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
4383  !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
4384  Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
4386 }
4387 
4388 /// Emits a call to the given no-arguments nounwind runtime function.
4389 llvm::CallInst *
4390 CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee,
4391  const llvm::Twine &name) {
4392  return EmitNounwindRuntimeCall(callee, None, name);
4393 }
4394 
4395 /// Emits a call to the given nounwind runtime function.
4396 llvm::CallInst *
4397 CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee,
4399  const llvm::Twine &name) {
4400  llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
4401  call->setDoesNotThrow();
4402  return call;
4403 }
4404 
4405 /// Emits a simple call (never an invoke) to the given no-arguments
4406 /// runtime function.
4407 llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
4408  const llvm::Twine &name) {
4409  return EmitRuntimeCall(callee, None, name);
4410 }
4411 
4412 // Calls which may throw must have operand bundles indicating which funclet
4413 // they are nested within.
4417  // There is no need for a funclet operand bundle if we aren't inside a
4418  // funclet.
4419  if (!CurrentFuncletPad)
4420  return BundleList;
4421 
4422  // Skip intrinsics which cannot throw.
4423  auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
4424  if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
4425  return BundleList;
4426 
4427  BundleList.emplace_back("funclet", CurrentFuncletPad);
4428  return BundleList;
4429 }
4430 
4431 /// Emits a simple call (never an invoke) to the given runtime function.
4432 llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
4434  const llvm::Twine &name) {
4435  llvm::CallInst *call = Builder.CreateCall(
4436  callee, args, getBundlesForFunclet(callee.getCallee()), name);
4437  call->setCallingConv(getRuntimeCC());
4438  return call;
4439 }
4440 
4441 /// Emits a call or invoke to the given noreturn runtime function.
4443  llvm::FunctionCallee callee, ArrayRef<llvm::Value *> args) {
4445  getBundlesForFunclet(callee.getCallee());
4446 
4447  if (getInvokeDest()) {
4448  llvm::InvokeInst *invoke =
4449  Builder.CreateInvoke(callee,
4451  getInvokeDest(),
4452  args,
4453  BundleList);
4454  invoke->setDoesNotReturn();
4455  invoke->setCallingConv(getRuntimeCC());
4456  } else {
4457  llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
4458  call->setDoesNotReturn();
4459  call->setCallingConv(getRuntimeCC());
4460  Builder.CreateUnreachable();
4461  }
4462 }
4463 
4464 /// Emits a call or invoke instruction to the given nullary runtime function.
4465 llvm::CallBase *
4466 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee,
4467  const Twine &name) {
4468  return EmitRuntimeCallOrInvoke(callee, None, name);
4469 }
4470 
4471 /// Emits a call or invoke instruction to the given runtime function.
4472 llvm::CallBase *
4473 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee,
4475  const Twine &name) {
4476  llvm::CallBase *call = EmitCallOrInvoke(callee, args, name);
4477  call->setCallingConv(getRuntimeCC());
4478  return call;
4479 }
4480 
4481 /// Emits a call or invoke instruction to the given function, depending
4482 /// on the current state of the EH stack.
4483 llvm::CallBase *CodeGenFunction::EmitCallOrInvoke(llvm::FunctionCallee Callee,
4485  const Twine &Name) {
4486  llvm::BasicBlock *InvokeDest = getInvokeDest();
4488  getBundlesForFunclet(Callee.getCallee());
4489 
4490  llvm::CallBase *Inst;
4491  if (!InvokeDest)
4492  Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
4493  else {
4494  llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
4495  Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
4496  Name);
4497  EmitBlock(ContBB);
4498  }
4499 
4500  // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4501  // optimizer it can aggressively ignore unwind edges.
4502  if (CGM.getLangOpts().ObjCAutoRefCount)
4503  AddObjCARCExceptionMetadata(Inst);
4504 
4505  return Inst;
4506 }
4507 
4508 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
4509  llvm::Value *New) {
4510  DeferredReplacements.push_back(
4511  std::make_pair(llvm::WeakTrackingVH(Old), New));
4512 }
4513 
4514 namespace {
4515 
4516 /// Specify given \p NewAlign as the alignment of return value attribute. If
4517 /// such attribute already exists, re-set it to the maximal one of two options.
4518 LLVM_NODISCARD llvm::AttributeList
4519 maybeRaiseRetAlignmentAttribute(llvm::LLVMContext &Ctx,
4520  const llvm::AttributeList &Attrs,
4521  llvm::Align NewAlign) {
4522  llvm::Align CurAlign = Attrs.getRetAlignment().valueOrOne();
4523  if (CurAlign >= NewAlign)
4524  return Attrs;
4525  llvm::Attribute AlignAttr = llvm::Attribute::getWithAlignment(Ctx, NewAlign);
4526  return Attrs.removeRetAttribute(Ctx, llvm::Attribute::AttrKind::Alignment)
4527  .addRetAttribute(Ctx, AlignAttr);
4528 }
4529 
4530 template <typename AlignedAttrTy> class AbstractAssumeAlignedAttrEmitter {
4531 protected:
4532  CodeGenFunction &CGF;
4533 
4534  /// We do nothing if this is, or becomes, nullptr.
4535  const AlignedAttrTy *AA = nullptr;
4536 
4537  llvm::Value *Alignment = nullptr; // May or may not be a constant.
4538  llvm::ConstantInt *OffsetCI = nullptr; // Constant, hopefully zero.
4539 
4540  AbstractAssumeAlignedAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl)
4541  : CGF(CGF_) {
4542  if (!FuncDecl)
4543  return;
4544  AA = FuncDecl->getAttr<AlignedAttrTy>();
4545  }
4546 
4547 public:
4548  /// If we can, materialize the alignment as an attribute on return value.
4549  LLVM_NODISCARD llvm::AttributeList
4550  TryEmitAsCallSiteAttribute(const llvm::AttributeList &Attrs) {
4551  if (!AA || OffsetCI || CGF.SanOpts.has(SanitizerKind::Alignment))
4552  return Attrs;
4553  const auto *AlignmentCI = dyn_cast<llvm::ConstantInt>(Alignment);
4554  if (!AlignmentCI)
4555  return Attrs;
4556  // We may legitimately have non-power-of-2 alignment here.
4557  // If so, this is UB land, emit it via `@llvm.assume` instead.
4558  if (!AlignmentCI->getValue().isPowerOf2())
4559  return Attrs;
4560  llvm::AttributeList NewAttrs = maybeRaiseRetAlignmentAttribute(
4561  CGF.getLLVMContext(), Attrs,
4562  llvm::Align(
4563  AlignmentCI->getLimitedValue(llvm::Value::MaximumAlignment)));
4564  AA = nullptr; // We're done. Disallow doing anything else.
4565  return NewAttrs;
4566  }
4567 
4568  /// Emit alignment assumption.
4569  /// This is a general fallback that we take if either there is an offset,
4570  /// or the alignment is variable or we are sanitizing for alignment.
4571  void EmitAsAnAssumption(SourceLocation Loc, QualType RetTy, RValue &Ret) {
4572  if (!AA)
4573  return;
4574  CGF.emitAlignmentAssumption(Ret.getScalarVal(), RetTy, Loc,
4575  AA->getLocation(), Alignment, OffsetCI);
4576  AA = nullptr; // We're done. Disallow doing anything else.
4577  }
4578 };
4579 
4580 /// Helper data structure to emit `AssumeAlignedAttr`.
4581 class AssumeAlignedAttrEmitter final
4582  : public AbstractAssumeAlignedAttrEmitter<AssumeAlignedAttr> {
4583 public:
4584  AssumeAlignedAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl)
4585  : AbstractAssumeAlignedAttrEmitter(CGF_, FuncDecl) {
4586  if (!AA)
4587  return;
4588  // It is guaranteed that the alignment/offset are constants.
4589  Alignment = cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AA->getAlignment()));
4590  if (Expr *Offset = AA->getOffset()) {
4591  OffsetCI = cast<llvm::ConstantInt>(CGF.EmitScalarExpr(Offset));
4592  if (OffsetCI->isNullValue()) // Canonicalize zero offset to no offset.
4593  OffsetCI = nullptr;
4594  }
4595  }
4596 };
4597 
4598 /// Helper data structure to emit `AllocAlignAttr`.
4599 class AllocAlignAttrEmitter final
4600  : public AbstractAssumeAlignedAttrEmitter<AllocAlignAttr> {
4601 public:
4602  AllocAlignAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl,
4603  const CallArgList &CallArgs)
4604  : AbstractAssumeAlignedAttrEmitter(CGF_, FuncDecl) {
4605  if (!AA)
4606  return;
4607  // Alignment may or may not be a constant, and that is okay.
4608  Alignment = CallArgs[AA->getParamIndex().getLLVMIndex()]
4609  .getRValue(CGF)
4610  .getScalarVal();
4611  }
4612 };
4613 
4614 } // namespace
4615 
4617  const CGCallee &Callee,
4618  ReturnValueSlot ReturnValue,
4619  const CallArgList &CallArgs,
4620  llvm::CallBase **callOrInvoke, bool IsMustTail,
4621  SourceLocation Loc) {
4622  // FIXME: We no longer need the types from CallArgs; lift up and simplify.
4623 
4624  assert(Callee.isOrdinary() || Callee.isVirtual());
4625 
4626  // Handle struct-return functions by passing a pointer to the
4627  // location that we would like to return into.
4628  QualType RetTy = CallInfo.getReturnType();
4629  const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
4630 
4631  llvm::FunctionType *IRFuncTy = getTypes().GetFunctionType(CallInfo);
4632 
4633  const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl().getDecl();
4634  if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl)) {
4635  // We can only guarantee that a function is called from the correct
4636  // context/function based on the appropriate target attributes,
4637  // so only check in the case where we have both always_inline and target
4638  // since otherwise we could be making a conditional call after a check for
4639  // the proper cpu features (and it won't cause code generation issues due to
4640  // function based code generation).
4641  if (TargetDecl->hasAttr<AlwaysInlineAttr>() &&
4642  TargetDecl->hasAttr<TargetAttr>())
4643  checkTargetFeatures(Loc, FD);
4644 
4645  // Some architectures (such as x86-64) have the ABI changed based on
4646  // attribute-target/features. Give them a chance to diagnose.
4648  CGM, Loc, dyn_cast_or_null<FunctionDecl>(CurCodeDecl), FD, CallArgs);
4649  }
4650 
4651 #ifndef NDEBUG
4652  if (!(CallInfo.isVariadic() && CallInfo.getArgStruct())) {
4653  // For an inalloca varargs function, we don't expect CallInfo to match the
4654  // function pointer's type, because the inalloca struct a will have extra
4655  // fields in it for the varargs parameters. Code later in this function
4656  // bitcasts the function pointer to the type derived from CallInfo.
4657  //
4658  // In other cases, we assert that the types match up (until pointers stop
4659  // having pointee types).
4660  llvm::Type *TypeFromVal;
4661  if (Callee.isVirtual())
4662  TypeFromVal = Callee.getVirtualFunctionType();
4663  else
4664  TypeFromVal =
4665  Callee.getFunctionPointer()->getType()->getPointerElementType();
4666  assert(IRFuncTy == TypeFromVal);
4667  }
4668 #endif
4669 
4670  // 1. Set up the arguments.
4671 
4672  // If we're using inalloca, insert the allocation after the stack save.
4673  // FIXME: Do this earlier rather than hacking it in here!
4674  Address ArgMemory = Address::invalid();
4675  if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
4676  const llvm::DataLayout &DL = CGM.getDataLayout();
4677  llvm::Instruction *IP = CallArgs.getStackBase();
4678  llvm::AllocaInst *AI;
4679  if (IP) {
4680  IP = IP->getNextNode();
4681  AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
4682  "argmem", IP);
4683  } else {
4684  AI = CreateTempAlloca(ArgStruct, "argmem");
4685  }
4686  auto Align = CallInfo.getArgStructAlignment();
4687  AI->setAlignment(Align.getAsAlign());
4688  AI->setUsedWithInAlloca(true);
4689  assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
4690  ArgMemory = Address(AI, Align);
4691  }
4692 
4693  ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
4694  SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
4695 
4696  // If the call returns a temporary with struct return, create a temporary
4697  // alloca to hold the result, unless one is given to us.
4698  Address SRetPtr = Address::invalid();
4699  Address SRetAlloca = Address::invalid();
4700  llvm::Value *UnusedReturnSizePtr = nullptr;
4701  if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
4702  if (!ReturnValue.isNull()) {
4703  SRetPtr = ReturnValue.getValue();
4704  } else {
4705  SRetPtr = CreateMemTemp(RetTy, "tmp", &SRetAlloca);
4706  if (HaveInsertPoint() && ReturnValue.isUnused()) {
4707  llvm::TypeSize size =
4708  CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
4709  UnusedReturnSizePtr = EmitLifetimeStart(size, SRetAlloca.getPointer());
4710  }
4711  }
4712  if (IRFunctionArgs.hasSRetArg()) {
4713  IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
4714  } else if (RetAI.isInAlloca()) {
4715  Address Addr =
4716  Builder.CreateStructGEP(ArgMemory, RetAI.getInAllocaFieldIndex());
4717  Builder.CreateStore(SRetPtr.getPointer(), Addr);
4718  }
4719  }
4720 
4721  Address swiftErrorTemp = Address::invalid();
4722  Address swiftErrorArg = Address::invalid();
4723 
4724  // When passing arguments using temporary allocas, we need to add the
4725  // appropriate lifetime markers. This vector keeps track of all the lifetime
4726  // markers that need to be ended right after the call.
4727  SmallVector<CallLifetimeEnd, 2> CallLifetimeEndAfterCall;
4728 
4729  // Translate all of the arguments as necessary to match the IR lowering.
4730  assert(CallInfo.arg_size() == CallArgs.size() &&
4731  "Mismatch between function signature & arguments.");
4732  unsigned ArgNo = 0;
4733  CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
4734  for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
4735  I != E; ++I, ++info_it, ++ArgNo) {
4736  const ABIArgInfo &ArgInfo = info_it->info;
4737 
4738  // Insert a padding argument to ensure proper alignment.
4739  if (IRFunctionArgs.hasPaddingArg(ArgNo))
4740  IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
4741  llvm::UndefValue::get(ArgInfo.getPaddingType());
4742 
4743  unsigned FirstIRArg, NumIRArgs;
4744  std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
4745 
4746  switch (ArgInfo.getKind()) {
4747  case ABIArgInfo::InAlloca: {
4748  assert(NumIRArgs == 0);
4749  assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
4750  if (I->isAggregate()) {
4751  Address Addr = I->hasLValue()
4752  ? I->getKnownLValue().getAddress(*this)
4753  : I->getKnownRValue().getAggregateAddress();
4754  llvm::Instruction *Placeholder =
4755  cast<llvm::Instruction>(Addr.getPointer());
4756 
4757  if (!ArgInfo.getInAllocaIndirect()) {
4758  // Replace the placeholder with the appropriate argument slot GEP.
4759  CGBuilderTy::InsertPoint IP = Builder.saveIP();
4760  Builder.SetInsertPoint(Placeholder);
4761  Addr = Builder.CreateStructGEP(ArgMemory,
4762  ArgInfo.getInAllocaFieldIndex());
4763  Builder.restoreIP(IP);
4764  } else {
4765  // For indirect things such as overaligned structs, replace the
4766  // placeholder with a regular aggregate temporary alloca. Store the
4767  // address of this alloca into the struct.
4768  Addr = CreateMemTemp(info_it->type, "inalloca.indirect.tmp");
4769  Address ArgSlot = Builder.CreateStructGEP(
4770  ArgMemory, ArgInfo.getInAllocaFieldIndex());
4771  Builder.CreateStore(Addr.getPointer(), ArgSlot);
4772  }
4773  deferPlaceholderReplacement(Placeholder, Addr.getPointer());
4774  } else if (ArgInfo.getInAllocaIndirect()) {
4775  // Make a temporary alloca and store the address of it into the argument
4776  // struct.
4778  I->Ty, getContext().getTypeAlignInChars(I->Ty),
4779  "indirect-arg-temp");
4780  I->copyInto(*this, Addr);
4781  Address ArgSlot =
4782  Builder.CreateStructGEP(ArgMemory, ArgInfo.getInAllocaFieldIndex());
4783  Builder.CreateStore(Addr.getPointer(), ArgSlot);
4784  } else {
4785  // Store the RValue into the argument struct.
4786  Address Addr =
4787  Builder.CreateStructGEP(ArgMemory, ArgInfo.getInAllocaFieldIndex());
4788  unsigned AS = Addr.getType()->getPointerAddressSpace();
4789  llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
4790  // There are some cases where a trivial bitcast is not avoidable. The
4791  // definition of a type later in a translation unit may change it's type
4792  // from {}* to (%struct.foo*)*.
4793  if (Addr.getType() != MemType)
4794  Addr = Builder.CreateBitCast(Addr, MemType);
4795  I->copyInto(*this, Addr);
4796  }
4797  break;
4798  }
4799 
4800  case ABIArgInfo::Indirect:
4802  assert(NumIRArgs == 1);
4803  if (!I->isAggregate()) {
4804  // Make a temporary alloca to pass the argument.
4806  I->Ty, ArgInfo.getIndirectAlign(), "indirect-arg-temp");
4807  IRCallArgs[FirstIRArg] = Addr.getPointer();
4808 
4809  I->copyInto(*this, Addr);
4810  } else {
4811  // We want to avoid creating an unnecessary temporary+copy here;
4812  // however, we need one in three cases:
4813  // 1. If the argument is not byval, and we are required to copy the
4814  // source. (This case doesn't occur on any common architecture.)
4815  // 2. If the argument is byval, RV is not sufficiently aligned, and
4816  // we cannot force it to be sufficiently aligned.
4817  // 3. If the argument is byval, but RV is not located in default
4818  // or alloca address space.
4819  Address Addr = I->hasLValue()
4820  ? I->getKnownLValue().getAddress(*this)
4821  : I->getKnownRValue().getAggregateAddress();
4822  llvm::Value *V = Addr.getPointer();
4823  CharUnits Align = ArgInfo.getIndirectAlign();
4824  const llvm::DataLayout *TD = &CGM.getDataLayout();
4825 
4826  assert((FirstIRArg >= IRFuncTy->getNumParams() ||
4827  IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace() ==
4828  TD->getAllocaAddrSpace()) &&
4829  "indirect argument must be in alloca address space");
4830 
4831  bool NeedCopy = false;
4832 
4833  if (Addr.getAlignment() < Align &&
4834  llvm::getOrEnforceKnownAlignment(V, Align.getAsAlign(), *TD) <
4835  Align.getAsAlign()) {
4836  NeedCopy = true;
4837  } else if (I->hasLValue()) {
4838  auto LV = I->getKnownLValue();
4839  auto AS = LV.getAddressSpace();
4840 
4841  if (!ArgInfo.getIndirectByVal() ||
4842  (LV.getAlignment() < getContext().getTypeAlignInChars(I->Ty))) {
4843  NeedCopy = true;
4844  }
4845  if (!getLangOpts().OpenCL) {
4846  if ((ArgInfo.getIndirectByVal() &&
4847  (AS != LangAS::Default &&
4848  AS != CGM.getASTAllocaAddressSpace()))) {
4849  NeedCopy = true;
4850  }
4851  }
4852  // For OpenCL even if RV is located in default or alloca address space
4853  // we don't want to perform address space cast for it.
4854  else if ((ArgInfo.getIndirectByVal() &&
4855  Addr.getType()->getAddressSpace() != IRFuncTy->
4856  getParamType(FirstIRArg)->getPointerAddressSpace())) {
4857  NeedCopy = true;
4858  }
4859  }
4860 
4861  if (NeedCopy) {
4862  // Create an aligned temporary, and copy to it.
4864  I->Ty, ArgInfo.getIndirectAlign(), "byval-temp");
4865  IRCallArgs[FirstIRArg] = AI.getPointer();
4866 
4867  // Emit lifetime markers for the temporary alloca.
4868  llvm::TypeSize ByvalTempElementSize =
4869  CGM.getDataLayout().getTypeAllocSize(AI.getElementType());
4870  llvm::Value *LifetimeSize =
4871  EmitLifetimeStart(ByvalTempElementSize, AI.getPointer());
4872 
4873  // Add cleanup code to emit the end lifetime marker after the call.
4874  if (LifetimeSize) // In case we disabled lifetime markers.
4875  CallLifetimeEndAfterCall.emplace_back(AI, LifetimeSize);
4876 
4877  // Generate the copy.
4878  I->copyInto(*this, AI);
4879  } else {
4880  // Skip the extra memcpy call.
4881  auto *T = V->getType()->getPointerElementType()->getPointerTo(
4882  CGM.getDataLayout().getAllocaAddrSpace());
4883  IRCallArgs[FirstIRArg] = getTargetHooks().performAddrSpaceCast(
4885  true);
4886  }
4887  }
4888  break;
4889  }
4890 
4891  case ABIArgInfo::Ignore:
4892  assert(NumIRArgs == 0);
4893  break;
4894 
4895  case ABIArgInfo::Extend:
4896  case ABIArgInfo::Direct: {
4897  if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
4898  ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
4899  ArgInfo.getDirectOffset() == 0) {
4900  assert(NumIRArgs == 1);
4901  llvm::Value *V;
4902  if (!I->isAggregate())
4903  V = I->getKnownRValue().getScalarVal();
4904  else
4905  V = Builder.CreateLoad(
4906  I->hasLValue() ? I->getKnownLValue().getAddress(*this)
4907  : I->getKnownRValue().getAggregateAddress());
4908 
4909  // Implement swifterror by copying into a new swifterror argument.
4910  // We'll write back in the normal path out of the call.
4911  if (CallInfo.getExtParameterInfo(ArgNo).getABI()
4913  assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
4914 
4915  QualType pointeeTy = I->Ty->getPointeeType();
4916  swiftErrorArg =
4917  Address(V, getContext().getTypeAlignInChars(pointeeTy));
4918 
4919  swiftErrorTemp =
4920  CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
4921  V = swiftErrorTemp.getPointer();
4922  cast<llvm::AllocaInst>(V)->setSwiftError(true);
4923 
4924  llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
4925  Builder.CreateStore(errorValue, swiftErrorTemp);
4926  }
4927 
4928  // We might have to widen integers, but we should never truncate.
4929  if (ArgInfo.getCoerceToType() != V->getType() &&
4930  V->getType()->isIntegerTy())
4931  V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
4932 
4933  // If the argument doesn't match, perform a bitcast to coerce it. This
4934  // can happen due to trivial type mismatches.
4935  if (FirstIRArg < IRFuncTy->getNumParams() &&
4936  V->getType() != IRFuncTy->getParamType(FirstIRArg))
4937  V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
4938 
4939  IRCallArgs[FirstIRArg] = V;
4940  break;
4941  }
4942 
4943  // FIXME: Avoid the conversion through memory if possible.
4944  Address Src = Address::invalid();
4945  if (!I->isAggregate()) {
4946  Src = CreateMemTemp(I->Ty, "coerce");
4947  I->copyInto(*this, Src);
4948  } else {
4949  Src = I->hasLValue() ? I->getKnownLValue().getAddress(*this)
4950  : I->getKnownRValue().getAggregateAddress();
4951  }
4952 
4953  // If the value is offset in memory, apply the offset now.
4954  Src = emitAddressAtOffset(*this, Src, ArgInfo);
4955 
4956  // Fast-isel and the optimizer generally like scalar values better than
4957  // FCAs, so we flatten them if this is safe to do for this argument.
4958  llvm::StructType *STy =
4959  dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
4960  if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
4961  llvm::Type *SrcTy = Src.getElementType();
4962  uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
4963  uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
4964 
4965  // If the source type is smaller than the destination type of the
4966  // coerce-to logic, copy the source value into a temp alloca the size
4967  // of the destination type to allow loading all of it. The bits past
4968  // the source value are left undef.
4969  if (SrcSize < DstSize) {
4970  Address TempAlloca
4971  = CreateTempAlloca(STy, Src.getAlignment(),
4972  Src.getName() + ".coerce");
4973  Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
4974  Src = TempAlloca;
4975  } else {
4976  Src = Builder.CreateBitCast(Src,
4977  STy->getPointerTo(Src.getAddressSpace()));
4978  }
4979 
4980  assert(NumIRArgs == STy->getNumElements());
4981  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
4982  Address EltPtr = Builder.CreateStructGEP(Src, i);
4983  llvm::Value *LI = Builder.CreateLoad(EltPtr);
4984  IRCallArgs[FirstIRArg + i] = LI;
4985  }
4986  } else {
4987  // In the simple case, just pass the coerced loaded value.
4988  assert(NumIRArgs == 1);
4989  llvm::Value *Load =
4990  CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
4991 
4992  if (CallInfo.isCmseNSCall()) {
4993  // For certain parameter types, clear padding bits, as they may reveal
4994  // sensitive information.
4995  // Small struct/union types are passed as integer arrays.
4996  auto *ATy = dyn_cast<llvm::ArrayType>(Load->getType());
4997  if (ATy != nullptr && isa<RecordType>(I->Ty.getCanonicalType()))
4998  Load = EmitCMSEClearRecord(Load, ATy, I->Ty);
4999  }
5000  IRCallArgs[FirstIRArg] = Load;
5001  }
5002 
5003  break;
5004  }
5005 
5007  auto coercionType = ArgInfo.getCoerceAndExpandType();
5008  auto layout = CGM.getDataLayout().getStructLayout(coercionType);
5009 
5010  llvm::Value *tempSize = nullptr;
5011  Address addr = Address::invalid();
5012  Address AllocaAddr = Address::invalid();
5013  if (I->isAggregate()) {
5014  addr = I->hasLValue() ? I->getKnownLValue().getAddress(*this)
5015  : I->getKnownRValue().getAggregateAddress();
5016 
5017  } else {
5018  RValue RV = I->getKnownRValue();
5019  assert(RV.isScalar()); // complex should always just be direct
5020 
5021  llvm::Type *scalarType = RV.getScalarVal()->getType();
5022  auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
5023  auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
5024 
5025  // Materialize to a temporary.
5026  addr =
5027  CreateTempAlloca(RV.getScalarVal()->getType(),
5029  layout->getAlignment().value(), scalarAlign)),
5030  "tmp",
5031  /*ArraySize=*/nullptr, &AllocaAddr);
5032  tempSize = EmitLifetimeStart(scalarSize, AllocaAddr.getPointer());
5033 
5034  Builder.CreateStore(RV.getScalarVal(), addr);
5035  }
5036 
5037  addr = Builder.CreateElementBitCast(addr, coercionType);
5038 
5039  unsigned IRArgPos = FirstIRArg;
5040  for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
5041  llvm::Type *eltType = coercionType->getElementType(i);
5042  if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
5043  Address eltAddr = Builder.CreateStructGEP(addr, i);
5044  llvm::Value *elt = Builder.CreateLoad(eltAddr);
5045  IRCallArgs[IRArgPos++] = elt;
5046  }
5047  assert(IRArgPos == FirstIRArg + NumIRArgs);
5048 
5049  if (tempSize) {
5050  EmitLifetimeEnd(tempSize, AllocaAddr.getPointer());
5051  }
5052 
5053  break;
5054  }
5055 
5056  case ABIArgInfo::Expand: {
5057  unsigned IRArgPos = FirstIRArg;
5058  ExpandTypeToArgs(I->Ty, *I, IRFuncTy, IRCallArgs, IRArgPos);
5059  assert(IRArgPos == FirstIRArg + NumIRArgs);
5060  break;
5061  }
5062  }
5063  }
5064 
5065  const CGCallee &ConcreteCallee = Callee.prepareConcreteCallee(*this);
5066  llvm::Value *CalleePtr = ConcreteCallee.getFunctionPointer();
5067 
5068  // If we're using inalloca, set up that argument.
5069  if (ArgMemory.isValid()) {
5070  llvm::Value *Arg = ArgMemory.getPointer();
5071  if (CallInfo.isVariadic()) {
5072  // When passing non-POD arguments by value to variadic functions, we will
5073  // end up with a variadic prototype and an inalloca call site. In such
5074  // cases, we can't do any parameter mismatch checks. Give up and bitcast
5075  // the callee.
5076  unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
5077  CalleePtr =
5078  Builder.CreateBitCast(CalleePtr, IRFuncTy->getPointerTo(CalleeAS));
5079  } else {
5080  llvm::Type *LastParamTy =
5081  IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
5082  if (Arg->getType() != LastParamTy) {
5083 #ifndef NDEBUG
5084  // Assert that these structs have equivalent element types.
5085  llvm::StructType *FullTy = CallInfo.getArgStruct();
5086  llvm::StructType *DeclaredTy = cast<llvm::StructType>(
5087  cast<llvm::PointerType>(LastParamTy)->getElementType());
5088  assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
5089  for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
5090  DE = DeclaredTy->element_end(),
5091  FI = FullTy->element_begin();
5092  DI != DE; ++DI, ++FI)
5093  assert(*DI == *FI);
5094 #endif
5095  Arg = Builder.CreateBitCast(Arg, LastParamTy);
5096  }
5097  }
5098  assert(IRFunctionArgs.hasInallocaArg());
5099  IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
5100  }
5101 
5102  // 2. Prepare the function pointer.
5103 
5104  // If the callee is a bitcast of a non-variadic function to have a
5105  // variadic function pointer type, check to see if we can remove the
5106  // bitcast. This comes up with unprototyped functions.
5107  //
5108  // This makes the IR nicer, but more importantly it ensures that we
5109  // can inline the function at -O0 if it is marked always_inline.
5110  auto simplifyVariadicCallee = [](llvm::FunctionType *CalleeFT,
5111  llvm::Value *Ptr) -> llvm::Function * {
5112  if (!CalleeFT->isVarArg())
5113  return nullptr;
5114 
5115  // Get underlying value if it's a bitcast
5116  if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr)) {
5117  if (CE->getOpcode() == llvm::Instruction::BitCast)
5118  Ptr = CE->getOperand(0);
5119  }
5120 
5121  llvm::Function *OrigFn = dyn_cast<llvm::Function>(Ptr);
5122  if (!OrigFn)
5123  return nullptr;
5124 
5125  llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
5126 
5127  // If the original type is variadic, or if any of the component types
5128  // disagree, we cannot remove the cast.
5129  if (OrigFT->isVarArg() ||
5130  OrigFT->getNumParams() != CalleeFT->getNumParams() ||
5131  OrigFT->getReturnType() != CalleeFT->getReturnType())
5132  return nullptr;
5133 
5134  for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
5135  if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
5136  return nullptr;
5137 
5138  return OrigFn;
5139  };
5140 
5141  if (llvm::Function *OrigFn = simplifyVariadicCallee(IRFuncTy, CalleePtr)) {
5142  CalleePtr = OrigFn;
5143  IRFuncTy = OrigFn->getFunctionType();
5144  }
5145 
5146  // 3. Perform the actual call.
5147 
5148  // Deactivate any cleanups that we're supposed to do immediately before
5149  // the call.
5150  if (!CallArgs.getCleanupsToDeactivate().empty())
5151  deactivateArgCleanupsBeforeCall(*this, CallArgs);
5152 
5153  // Assert that the arguments we computed match up. The IR verifier
5154  // will catch this, but this is a common enough source of problems
5155  // during IRGen changes that it's way better for debugging to catch
5156  // it ourselves here.
5157 #ifndef NDEBUG
5158  assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
5159  for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
5160  // Inalloca argument can have different type.
5161  if (IRFunctionArgs.hasInallocaArg() &&
5162  i == IRFunctionArgs.getInallocaArgNo())
5163  continue;
5164  if (i < IRFuncTy->getNumParams())
5165  assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
5166  }
5167 #endif
5168 
5169  // Update the largest vector width if any arguments have vector types.
5170  for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
5171  if (auto *VT = dyn_cast<llvm::VectorType>(IRCallArgs[i]->getType()))
5172  LargestVectorWidth =
5173  std::max((uint64_t)LargestVectorWidth,
5174  VT->getPrimitiveSizeInBits().getKnownMinSize());
5175  }
5176 
5177  // Compute the calling convention and attributes.
5178  unsigned CallingConv;
5179  llvm::AttributeList Attrs;
5180  CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
5181  Callee.getAbstractInfo(), Attrs, CallingConv,
5182  /*AttrOnCallSite=*/true,
5183  /*IsThunk=*/false, CurFuncDecl);
5184 
5185  if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl))
5186  if (FD->hasAttr<StrictFPAttr>())
5187  // All calls within a strictfp function are marked strictfp
5188  Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::StrictFP);
5189 
5190  // Add call-site nomerge attribute if exists.
5192  Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoMerge);
5193 
5194  // Apply some call-site-specific attributes.
5195  // TODO: work this into building the attribute set.
5196 
5197  // Apply always_inline to all calls within flatten functions.
5198  // FIXME: should this really take priority over __try, below?
5199  if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
5200  !(TargetDecl && TargetDecl->hasAttr<NoInlineAttr>())) {
5201  Attrs =
5202  Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::AlwaysInline);
5203  }
5204 
5205  // Disable inlining inside SEH __try blocks.
5206  if (isSEHTryScope()) {
5207  Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoInline);
5208  }
5209 
5210  // Decide whether to use a call or an invoke.
5211  bool CannotThrow;
5212  if (currentFunctionUsesSEHTry()) {
5213  // SEH cares about asynchronous exceptions, so everything can "throw."
5214  CannotThrow = false;
5215  } else if (isCleanupPadScope() &&
5217  // The MSVC++ personality will implicitly terminate the program if an
5218  // exception is thrown during a cleanup outside of a try/catch.
5219  // We don't need to model anything in IR to get this behavior.
5220  CannotThrow = true;
5221  } else {
5222  // Otherwise, nounwind call sites will never throw.
5223  CannotThrow = Attrs.hasFnAttr(llvm::Attribute::NoUnwind);
5224 
5225  if (auto *FPtr = dyn_cast<llvm::Function>(CalleePtr))
5226  if (FPtr->hasFnAttribute(llvm::Attribute::NoUnwind))
5227  CannotThrow = true;
5228  }
5229 
5230  // If we made a temporary, be sure to clean up after ourselves. Note that we
5231  // can't depend on being inside of an ExprWithCleanups, so we need to manually
5232  // pop this cleanup later on. Being eager about this is OK, since this
5233  // temporary is 'invisible' outside of the callee.
5234  if (UnusedReturnSizePtr)
5235  pushFullExprCleanup<CallLifetimeEnd>(NormalEHLifetimeMarker, SRetAlloca,
5236  UnusedReturnSizePtr);
5237 
5238  llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
5239 
5241  getBundlesForFunclet(CalleePtr);
5242 
5243  if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl))
5244  if (FD->hasAttr<StrictFPAttr>())
5245  // All calls within a strictfp function are marked strictfp
5246  Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::StrictFP);
5247 
5248  AssumeAlignedAttrEmitter AssumeAlignedAttrEmitter(*this, TargetDecl);
5249  Attrs = AssumeAlignedAttrEmitter.TryEmitAsCallSiteAttribute(Attrs);
5250 
5251  AllocAlignAttrEmitter AllocAlignAttrEmitter(*this, TargetDecl, CallArgs);
5252  Attrs = AllocAlignAttrEmitter.TryEmitAsCallSiteAttribute(Attrs);
5253 
5254  // Emit the actual call/invoke instruction.
5255  llvm::CallBase *CI;
5256  if (!InvokeDest) {
5257  CI = Builder.CreateCall(IRFuncTy, CalleePtr, IRCallArgs, BundleList);
5258  } else {
5259  llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
5260  CI = Builder.CreateInvoke(IRFuncTy, CalleePtr, Cont, InvokeDest, IRCallArgs,
5261  BundleList);
5262  EmitBlock(Cont);
5263  }
5264  if (callOrInvoke)
5265  *callOrInvoke = CI;
5266 
5267  // If this is within a function that has the guard(nocf) attribute and is an
5268  // indirect call, add the "guard_nocf" attribute to this call to indicate that
5269  // Control Flow Guard checks should not be added, even if the call is inlined.
5270  if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) {
5271  if (const auto *A = FD->getAttr<CFGuardAttr>()) {
5272  if (A->getGuard() == CFGuardAttr::GuardArg::nocf && !CI->getCalledFunction())
5273  Attrs = Attrs.addFnAttribute(getLLVMContext(), "guard_nocf");
5274  }
5275  }
5276 
5277  // Apply the attributes and calling convention.
5278  CI->setAttributes(Attrs);
5279  CI->setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
5280 
5281  // Apply various metadata.
5282 
5283  if (!CI->getType()->isVoidTy())
5284  CI->setName("call");
5285 
5286  // Update largest vector width from the return type.
5287  if (auto *VT = dyn_cast<llvm::VectorType>(CI->getType()))
5288  LargestVectorWidth =
5289  std::max((uint64_t)LargestVectorWidth,
5290  VT->getPrimitiveSizeInBits().getKnownMinSize());
5291 
5292  // Insert instrumentation or attach profile metadata at indirect call sites.
5293  // For more details, see the comment before the definition of
5294  // IPVK_IndirectCallTarget in InstrProfData.inc.
5295  if (!CI->getCalledFunction())
5296  PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
5297  CI, CalleePtr);
5298 
5299  // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
5300  // optimizer it can aggressively ignore unwind edges.
5301  if (CGM.getLangOpts().ObjCAutoRefCount)
5302  AddObjCARCExceptionMetadata(CI);
5303 
5304  // Set tail call kind if necessary.
5305  if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
5306  if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
5307  Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
5308  else if (IsMustTail)
5309  Call->setTailCallKind(llvm::CallInst::TCK_MustTail);
5310  }
5311 
5312  // Add metadata for calls to MSAllocator functions
5313  if (getDebugInfo() && TargetDecl &&
5314  TargetDecl->hasAttr<MSAllocatorAttr>())
5316 
5317  // Add metadata if calling an __attribute__((error(""))) or warning fn.
5318  if (TargetDecl && TargetDecl->hasAttr<ErrorAttr>()) {
5319  llvm::ConstantInt *Line =
5320  llvm::ConstantInt::get(Int32Ty, Loc.getRawEncoding());
5321  llvm::ConstantAsMetadata *MD = llvm::ConstantAsMetadata::get(Line);
5322  llvm::MDTuple *MDT = llvm::MDNode::get(getLLVMContext(), {MD});
5323  CI->setMetadata("srcloc", MDT);
5324  }
5325 
5326  // 4. Finish the call.
5327 
5328  // If the call doesn't return, finish the basic block and clear the
5329  // insertion point; this allows the rest of IRGen to discard
5330  // unreachable code.
5331  if (CI->doesNotReturn()) {
5332  if (UnusedReturnSizePtr)
5333  PopCleanupBlock();
5334 
5335  // Strip away the noreturn attribute to better diagnose unreachable UB.
5336  if (SanOpts.has(SanitizerKind::Unreachable)) {
5337  // Also remove from function since CallBase::hasFnAttr additionally checks
5338  // attributes of the called function.
5339  if (auto *F = CI->getCalledFunction())
5340  F->removeFnAttr(llvm::Attribute::NoReturn);
5341  CI->removeFnAttr(llvm::Attribute::NoReturn);
5342 
5343  // Avoid incompatibility with ASan which relies on the `noreturn`
5344  // attribute to insert handler calls.
5345  if (SanOpts.hasOneOf(SanitizerKind::Address |
5346  SanitizerKind::KernelAddress)) {
5347  SanitizerScope SanScope(this);
5348  llvm::IRBuilder<>::InsertPointGuard IPGuard(Builder);
5349  Builder.SetInsertPoint(CI);
5350  auto *FnType = llvm::FunctionType::get(CGM.VoidTy, /*isVarArg=*/false);
5351  llvm::FunctionCallee Fn =
5352  CGM.CreateRuntimeFunction(FnType, "__asan_handle_no_return");
5354  }
5355  }
5356 
5357  EmitUnreachable(Loc);
5358  Builder.ClearInsertionPoint();
5359 
5360  // FIXME: For now, emit a dummy basic block because expr emitters in
5361  // generally are not ready to handle emitting expressions at unreachable
5362  // points.
5364 
5365  // Return a reasonable RValue.
5366  return GetUndefRValue(RetTy);
5367  }
5368 
5369  // If this is a musttail call, return immediately. We do not branch to the
5370  // epilogue in this case.
5371  if (IsMustTail) {
5372  for (auto it = EHStack.find(CurrentCleanupScopeDepth); it !=