CGCall.cpp revision 6857d9d43b082ae825c29cca80f2f6b7c3aa4e5f
1//===----- CGCall.h - Encapsulate calling convention details ----*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// These classes wrap the information about a call or function 11// definition used to handle ABI compliancy. 12// 13//===----------------------------------------------------------------------===// 14 15#include "CGCall.h" 16#include "CodeGenFunction.h" 17#include "CodeGenModule.h" 18#include "clang/Basic/TargetInfo.h" 19#include "clang/AST/ASTContext.h" 20#include "clang/AST/Decl.h" 21#include "clang/AST/DeclCXX.h" 22#include "clang/AST/DeclObjC.h" 23#include "clang/AST/RecordLayout.h" 24#include "llvm/ADT/StringExtras.h" 25#include "llvm/Attributes.h" 26#include "llvm/Support/CallSite.h" 27#include "llvm/Support/MathExtras.h" 28#include "llvm/Target/TargetData.h" 29 30#include "ABIInfo.h" 31 32using namespace clang; 33using namespace CodeGen; 34 35/***/ 36 37// FIXME: Use iterator and sidestep silly type array creation. 38 39const 40CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionNoProtoType *FTNP) { 41 return getFunctionInfo(FTNP->getResultType(), 42 llvm::SmallVector<QualType, 16>()); 43} 44 45const 46CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionProtoType *FTP) { 47 llvm::SmallVector<QualType, 16> ArgTys; 48 // FIXME: Kill copy. 49 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 50 ArgTys.push_back(FTP->getArgType(i)); 51 return getFunctionInfo(FTP->getResultType(), ArgTys); 52} 53 54const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const CXXMethodDecl *MD) { 55 llvm::SmallVector<QualType, 16> ArgTys; 56 // Add the 'this' pointer unless this is a static method. 57 if (MD->isInstance()) 58 ArgTys.push_back(MD->getThisType(Context)); 59 60 const FunctionProtoType *FTP = MD->getType()->getAsFunctionProtoType(); 61 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 62 ArgTys.push_back(FTP->getArgType(i)); 63 return getFunctionInfo(FTP->getResultType(), ArgTys); 64} 65 66const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionDecl *FD) { 67 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD)) 68 if (MD->isInstance()) 69 return getFunctionInfo(MD); 70 71 const FunctionType *FTy = FD->getType()->getAsFunctionType(); 72 if (const FunctionProtoType *FTP = dyn_cast<FunctionProtoType>(FTy)) 73 return getFunctionInfo(FTP); 74 return getFunctionInfo(cast<FunctionNoProtoType>(FTy)); 75} 76 77const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const ObjCMethodDecl *MD) { 78 llvm::SmallVector<QualType, 16> ArgTys; 79 ArgTys.push_back(MD->getSelfDecl()->getType()); 80 ArgTys.push_back(Context.getObjCSelType()); 81 // FIXME: Kill copy? 82 for (ObjCMethodDecl::param_iterator i = MD->param_begin(), 83 e = MD->param_end(); i != e; ++i) 84 ArgTys.push_back((*i)->getType()); 85 return getFunctionInfo(MD->getResultType(), ArgTys); 86} 87 88const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 89 const CallArgList &Args) { 90 // FIXME: Kill copy. 91 llvm::SmallVector<QualType, 16> ArgTys; 92 for (CallArgList::const_iterator i = Args.begin(), e = Args.end(); 93 i != e; ++i) 94 ArgTys.push_back(i->second); 95 return getFunctionInfo(ResTy, ArgTys); 96} 97 98const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 99 const FunctionArgList &Args) { 100 // FIXME: Kill copy. 101 llvm::SmallVector<QualType, 16> ArgTys; 102 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 103 i != e; ++i) 104 ArgTys.push_back(i->second); 105 return getFunctionInfo(ResTy, ArgTys); 106} 107 108const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 109 const llvm::SmallVector<QualType, 16> &ArgTys) { 110 // Lookup or create unique function info. 111 llvm::FoldingSetNodeID ID; 112 CGFunctionInfo::Profile(ID, ResTy, ArgTys.begin(), ArgTys.end()); 113 114 void *InsertPos = 0; 115 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, InsertPos); 116 if (FI) 117 return *FI; 118 119 // Construct the function info. 120 FI = new CGFunctionInfo(ResTy, ArgTys); 121 FunctionInfos.InsertNode(FI, InsertPos); 122 123 // Compute ABI information. 124 getABIInfo().computeInfo(*FI, getContext()); 125 126 return *FI; 127} 128 129/***/ 130 131ABIInfo::~ABIInfo() {} 132 133void ABIArgInfo::dump() const { 134 fprintf(stderr, "(ABIArgInfo Kind="); 135 switch (TheKind) { 136 case Direct: 137 fprintf(stderr, "Direct"); 138 break; 139 case Ignore: 140 fprintf(stderr, "Ignore"); 141 break; 142 case Coerce: 143 fprintf(stderr, "Coerce Type="); 144 getCoerceToType()->print(llvm::errs()); 145 break; 146 case Indirect: 147 fprintf(stderr, "Indirect Align=%d", getIndirectAlign()); 148 break; 149 case Expand: 150 fprintf(stderr, "Expand"); 151 break; 152 } 153 fprintf(stderr, ")\n"); 154} 155 156/***/ 157 158static bool isEmptyRecord(ASTContext &Context, QualType T); 159 160/// isEmptyField - Return true iff a the field is "empty", that is it 161/// is an unnamed bit-field or an (array of) empty record(s). 162static bool isEmptyField(ASTContext &Context, const FieldDecl *FD) { 163 if (FD->isUnnamedBitfield()) 164 return true; 165 166 QualType FT = FD->getType(); 167 // Constant arrays of empty records count as empty, strip them off. 168 while (const ConstantArrayType *AT = Context.getAsConstantArrayType(FT)) 169 FT = AT->getElementType(); 170 171 return isEmptyRecord(Context, FT); 172} 173 174/// isEmptyRecord - Return true iff a structure contains only empty 175/// fields. Note that a structure with a flexible array member is not 176/// considered empty. 177static bool isEmptyRecord(ASTContext &Context, QualType T) { 178 const RecordType *RT = T->getAsRecordType(); 179 if (!RT) 180 return 0; 181 const RecordDecl *RD = RT->getDecl(); 182 if (RD->hasFlexibleArrayMember()) 183 return false; 184 for (RecordDecl::field_iterator i = RD->field_begin(Context), 185 e = RD->field_end(Context); i != e; ++i) 186 if (!isEmptyField(Context, *i)) 187 return false; 188 return true; 189} 190 191/// isSingleElementStruct - Determine if a structure is a "single 192/// element struct", i.e. it has exactly one non-empty field or 193/// exactly one field which is itself a single element 194/// struct. Structures with flexible array members are never 195/// considered single element structs. 196/// 197/// \return The field declaration for the single non-empty field, if 198/// it exists. 199static const Type *isSingleElementStruct(QualType T, ASTContext &Context) { 200 const RecordType *RT = T->getAsStructureType(); 201 if (!RT) 202 return 0; 203 204 const RecordDecl *RD = RT->getDecl(); 205 if (RD->hasFlexibleArrayMember()) 206 return 0; 207 208 const Type *Found = 0; 209 for (RecordDecl::field_iterator i = RD->field_begin(Context), 210 e = RD->field_end(Context); i != e; ++i) { 211 const FieldDecl *FD = *i; 212 QualType FT = FD->getType(); 213 214 // Ignore empty fields. 215 if (isEmptyField(Context, FD)) 216 continue; 217 218 // If we already found an element then this isn't a single-element 219 // struct. 220 if (Found) 221 return 0; 222 223 // Treat single element arrays as the element. 224 while (const ConstantArrayType *AT = Context.getAsConstantArrayType(FT)) { 225 if (AT->getSize().getZExtValue() != 1) 226 break; 227 FT = AT->getElementType(); 228 } 229 230 if (!CodeGenFunction::hasAggregateLLVMType(FT)) { 231 Found = FT.getTypePtr(); 232 } else { 233 Found = isSingleElementStruct(FT, Context); 234 if (!Found) 235 return 0; 236 } 237 } 238 239 return Found; 240} 241 242static bool is32Or64BitBasicType(QualType Ty, ASTContext &Context) { 243 if (!Ty->getAsBuiltinType() && !Ty->isPointerType()) 244 return false; 245 246 uint64_t Size = Context.getTypeSize(Ty); 247 return Size == 32 || Size == 64; 248} 249 250static bool areAllFields32Or64BitBasicType(const RecordDecl *RD, 251 ASTContext &Context) { 252 for (RecordDecl::field_iterator i = RD->field_begin(Context), 253 e = RD->field_end(Context); i != e; ++i) { 254 const FieldDecl *FD = *i; 255 256 if (!is32Or64BitBasicType(FD->getType(), Context)) 257 return false; 258 259 // FIXME: Reject bit-fields wholesale; there are two problems, we don't know 260 // how to expand them yet, and the predicate for telling if a bitfield still 261 // counts as "basic" is more complicated than what we were doing previously. 262 if (FD->isBitField()) 263 return false; 264 } 265 266 return true; 267} 268 269namespace { 270/// DefaultABIInfo - The default implementation for ABI specific 271/// details. This implementation provides information which results in 272/// self-consistent and sensible LLVM IR generation, but does not 273/// conform to any particular ABI. 274class DefaultABIInfo : public ABIInfo { 275 ABIArgInfo classifyReturnType(QualType RetTy, 276 ASTContext &Context) const; 277 278 ABIArgInfo classifyArgumentType(QualType RetTy, 279 ASTContext &Context) const; 280 281 virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const { 282 FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context); 283 for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end(); 284 it != ie; ++it) 285 it->info = classifyArgumentType(it->type, Context); 286 } 287 288 virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 289 CodeGenFunction &CGF) const; 290}; 291 292/// X86_32ABIInfo - The X86-32 ABI information. 293class X86_32ABIInfo : public ABIInfo { 294 ASTContext &Context; 295 bool IsDarwin; 296 297 static bool isRegisterSize(unsigned Size) { 298 return (Size == 8 || Size == 16 || Size == 32 || Size == 64); 299 } 300 301 static bool shouldReturnTypeInRegister(QualType Ty, ASTContext &Context); 302 303public: 304 ABIArgInfo classifyReturnType(QualType RetTy, 305 ASTContext &Context) const; 306 307 ABIArgInfo classifyArgumentType(QualType RetTy, 308 ASTContext &Context) const; 309 310 virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const { 311 FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context); 312 for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end(); 313 it != ie; ++it) 314 it->info = classifyArgumentType(it->type, Context); 315 } 316 317 virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 318 CodeGenFunction &CGF) const; 319 320 X86_32ABIInfo(ASTContext &Context, bool d) 321 : ABIInfo(), Context(Context), IsDarwin(d) {} 322}; 323} 324 325 326/// shouldReturnTypeInRegister - Determine if the given type should be 327/// passed in a register (for the Darwin ABI). 328bool X86_32ABIInfo::shouldReturnTypeInRegister(QualType Ty, 329 ASTContext &Context) { 330 uint64_t Size = Context.getTypeSize(Ty); 331 332 // Type must be register sized. 333 if (!isRegisterSize(Size)) 334 return false; 335 336 if (Ty->isVectorType()) { 337 // 64- and 128- bit vectors inside structures are not returned in 338 // registers. 339 if (Size == 64 || Size == 128) 340 return false; 341 342 return true; 343 } 344 345 // If this is a builtin, pointer, or complex type, it is ok. 346 if (Ty->getAsBuiltinType() || Ty->isPointerType() || Ty->isAnyComplexType()) 347 return true; 348 349 // Arrays are treated like records. 350 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) 351 return shouldReturnTypeInRegister(AT->getElementType(), Context); 352 353 // Otherwise, it must be a record type. 354 const RecordType *RT = Ty->getAsRecordType(); 355 if (!RT) return false; 356 357 // Structure types are passed in register if all fields would be 358 // passed in a register. 359 for (RecordDecl::field_iterator i = RT->getDecl()->field_begin(Context), 360 e = RT->getDecl()->field_end(Context); i != e; ++i) { 361 const FieldDecl *FD = *i; 362 363 // Empty fields are ignored. 364 if (isEmptyField(Context, FD)) 365 continue; 366 367 // Check fields recursively. 368 if (!shouldReturnTypeInRegister(FD->getType(), Context)) 369 return false; 370 } 371 372 return true; 373} 374 375ABIArgInfo X86_32ABIInfo::classifyReturnType(QualType RetTy, 376 ASTContext &Context) const { 377 if (RetTy->isVoidType()) { 378 return ABIArgInfo::getIgnore(); 379 } else if (const VectorType *VT = RetTy->getAsVectorType()) { 380 // On Darwin, some vectors are returned in registers. 381 if (IsDarwin) { 382 uint64_t Size = Context.getTypeSize(RetTy); 383 384 // 128-bit vectors are a special case; they are returned in 385 // registers and we need to make sure to pick a type the LLVM 386 // backend will like. 387 if (Size == 128) 388 return ABIArgInfo::getCoerce(llvm::VectorType::get(llvm::Type::Int64Ty, 389 2)); 390 391 // Always return in register if it fits in a general purpose 392 // register, or if it is 64 bits and has a single element. 393 if ((Size == 8 || Size == 16 || Size == 32) || 394 (Size == 64 && VT->getNumElements() == 1)) 395 return ABIArgInfo::getCoerce(llvm::IntegerType::get(Size)); 396 397 return ABIArgInfo::getIndirect(0); 398 } 399 400 return ABIArgInfo::getDirect(); 401 } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 402 // Structures with flexible arrays are always indirect. 403 if (const RecordType *RT = RetTy->getAsStructureType()) 404 if (RT->getDecl()->hasFlexibleArrayMember()) 405 return ABIArgInfo::getIndirect(0); 406 407 // Outside of Darwin, structs and unions are always indirect. 408 if (!IsDarwin && !RetTy->isAnyComplexType()) 409 return ABIArgInfo::getIndirect(0); 410 411 // Classify "single element" structs as their element type. 412 if (const Type *SeltTy = isSingleElementStruct(RetTy, Context)) { 413 if (const BuiltinType *BT = SeltTy->getAsBuiltinType()) { 414 if (BT->isIntegerType()) { 415 // We need to use the size of the structure, padding 416 // bit-fields can adjust that to be larger than the single 417 // element type. 418 uint64_t Size = Context.getTypeSize(RetTy); 419 return ABIArgInfo::getCoerce(llvm::IntegerType::get((unsigned) Size)); 420 } else if (BT->getKind() == BuiltinType::Float) { 421 assert(Context.getTypeSize(RetTy) == Context.getTypeSize(SeltTy) && 422 "Unexpect single element structure size!"); 423 return ABIArgInfo::getCoerce(llvm::Type::FloatTy); 424 } else if (BT->getKind() == BuiltinType::Double) { 425 assert(Context.getTypeSize(RetTy) == Context.getTypeSize(SeltTy) && 426 "Unexpect single element structure size!"); 427 return ABIArgInfo::getCoerce(llvm::Type::DoubleTy); 428 } 429 } else if (SeltTy->isPointerType()) { 430 // FIXME: It would be really nice if this could come out as the proper 431 // pointer type. 432 llvm::Type *PtrTy = 433 llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 434 return ABIArgInfo::getCoerce(PtrTy); 435 } else if (SeltTy->isVectorType()) { 436 // 64- and 128-bit vectors are never returned in a 437 // register when inside a structure. 438 uint64_t Size = Context.getTypeSize(RetTy); 439 if (Size == 64 || Size == 128) 440 return ABIArgInfo::getIndirect(0); 441 442 return classifyReturnType(QualType(SeltTy, 0), Context); 443 } 444 } 445 446 // Small structures which are register sized are generally returned 447 // in a register. 448 if (X86_32ABIInfo::shouldReturnTypeInRegister(RetTy, Context)) { 449 uint64_t Size = Context.getTypeSize(RetTy); 450 return ABIArgInfo::getCoerce(llvm::IntegerType::get(Size)); 451 } 452 453 return ABIArgInfo::getIndirect(0); 454 } else { 455 return ABIArgInfo::getDirect(); 456 } 457} 458 459ABIArgInfo X86_32ABIInfo::classifyArgumentType(QualType Ty, 460 ASTContext &Context) const { 461 // FIXME: Set alignment on indirect arguments. 462 if (CodeGenFunction::hasAggregateLLVMType(Ty)) { 463 // Structures with flexible arrays are always indirect. 464 if (const RecordType *RT = Ty->getAsStructureType()) 465 if (RT->getDecl()->hasFlexibleArrayMember()) 466 return ABIArgInfo::getIndirect(0); 467 468 // Ignore empty structs. 469 uint64_t Size = Context.getTypeSize(Ty); 470 if (Ty->isStructureType() && Size == 0) 471 return ABIArgInfo::getIgnore(); 472 473 // Expand structs with size <= 128-bits which consist only of 474 // basic types (int, long long, float, double, xxx*). This is 475 // non-recursive and does not ignore empty fields. 476 if (const RecordType *RT = Ty->getAsStructureType()) { 477 if (Context.getTypeSize(Ty) <= 4*32 && 478 areAllFields32Or64BitBasicType(RT->getDecl(), Context)) 479 return ABIArgInfo::getExpand(); 480 } 481 482 return ABIArgInfo::getIndirect(0); 483 } else { 484 return ABIArgInfo::getDirect(); 485 } 486} 487 488llvm::Value *X86_32ABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 489 CodeGenFunction &CGF) const { 490 const llvm::Type *BP = llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 491 const llvm::Type *BPP = llvm::PointerType::getUnqual(BP); 492 493 CGBuilderTy &Builder = CGF.Builder; 494 llvm::Value *VAListAddrAsBPP = Builder.CreateBitCast(VAListAddr, BPP, 495 "ap"); 496 llvm::Value *Addr = Builder.CreateLoad(VAListAddrAsBPP, "ap.cur"); 497 llvm::Type *PTy = 498 llvm::PointerType::getUnqual(CGF.ConvertType(Ty)); 499 llvm::Value *AddrTyped = Builder.CreateBitCast(Addr, PTy); 500 501 uint64_t Offset = 502 llvm::RoundUpToAlignment(CGF.getContext().getTypeSize(Ty) / 8, 4); 503 llvm::Value *NextAddr = 504 Builder.CreateGEP(Addr, 505 llvm::ConstantInt::get(llvm::Type::Int32Ty, Offset), 506 "ap.next"); 507 Builder.CreateStore(NextAddr, VAListAddrAsBPP); 508 509 return AddrTyped; 510} 511 512namespace { 513/// X86_64ABIInfo - The X86_64 ABI information. 514class X86_64ABIInfo : public ABIInfo { 515 enum Class { 516 Integer = 0, 517 SSE, 518 SSEUp, 519 X87, 520 X87Up, 521 ComplexX87, 522 NoClass, 523 Memory 524 }; 525 526 /// merge - Implement the X86_64 ABI merging algorithm. 527 /// 528 /// Merge an accumulating classification \arg Accum with a field 529 /// classification \arg Field. 530 /// 531 /// \param Accum - The accumulating classification. This should 532 /// always be either NoClass or the result of a previous merge 533 /// call. In addition, this should never be Memory (the caller 534 /// should just return Memory for the aggregate). 535 Class merge(Class Accum, Class Field) const; 536 537 /// classify - Determine the x86_64 register classes in which the 538 /// given type T should be passed. 539 /// 540 /// \param Lo - The classification for the parts of the type 541 /// residing in the low word of the containing object. 542 /// 543 /// \param Hi - The classification for the parts of the type 544 /// residing in the high word of the containing object. 545 /// 546 /// \param OffsetBase - The bit offset of this type in the 547 /// containing object. Some parameters are classified different 548 /// depending on whether they straddle an eightbyte boundary. 549 /// 550 /// If a word is unused its result will be NoClass; if a type should 551 /// be passed in Memory then at least the classification of \arg Lo 552 /// will be Memory. 553 /// 554 /// The \arg Lo class will be NoClass iff the argument is ignored. 555 /// 556 /// If the \arg Lo class is ComplexX87, then the \arg Hi class will 557 /// also be ComplexX87. 558 void classify(QualType T, ASTContext &Context, uint64_t OffsetBase, 559 Class &Lo, Class &Hi) const; 560 561 /// getCoerceResult - Given a source type \arg Ty and an LLVM type 562 /// to coerce to, chose the best way to pass Ty in the same place 563 /// that \arg CoerceTo would be passed, but while keeping the 564 /// emitted code as simple as possible. 565 /// 566 /// FIXME: Note, this should be cleaned up to just take an enumeration of all 567 /// the ways we might want to pass things, instead of constructing an LLVM 568 /// type. This makes this code more explicit, and it makes it clearer that we 569 /// are also doing this for correctness in the case of passing scalar types. 570 ABIArgInfo getCoerceResult(QualType Ty, 571 const llvm::Type *CoerceTo, 572 ASTContext &Context) const; 573 574 ABIArgInfo classifyReturnType(QualType RetTy, 575 ASTContext &Context) const; 576 577 ABIArgInfo classifyArgumentType(QualType Ty, 578 ASTContext &Context, 579 unsigned &neededInt, 580 unsigned &neededSSE) const; 581 582public: 583 virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const; 584 585 virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 586 CodeGenFunction &CGF) const; 587}; 588} 589 590X86_64ABIInfo::Class X86_64ABIInfo::merge(Class Accum, 591 Class Field) const { 592 // AMD64-ABI 3.2.3p2: Rule 4. Each field of an object is 593 // classified recursively so that always two fields are 594 // considered. The resulting class is calculated according to 595 // the classes of the fields in the eightbyte: 596 // 597 // (a) If both classes are equal, this is the resulting class. 598 // 599 // (b) If one of the classes is NO_CLASS, the resulting class is 600 // the other class. 601 // 602 // (c) If one of the classes is MEMORY, the result is the MEMORY 603 // class. 604 // 605 // (d) If one of the classes is INTEGER, the result is the 606 // INTEGER. 607 // 608 // (e) If one of the classes is X87, X87UP, COMPLEX_X87 class, 609 // MEMORY is used as class. 610 // 611 // (f) Otherwise class SSE is used. 612 613 // Accum should never be memory (we should have returned) or 614 // ComplexX87 (because this cannot be passed in a structure). 615 assert((Accum != Memory && Accum != ComplexX87) && 616 "Invalid accumulated classification during merge."); 617 if (Accum == Field || Field == NoClass) 618 return Accum; 619 else if (Field == Memory) 620 return Memory; 621 else if (Accum == NoClass) 622 return Field; 623 else if (Accum == Integer || Field == Integer) 624 return Integer; 625 else if (Field == X87 || Field == X87Up || Field == ComplexX87 || 626 Accum == X87 || Accum == X87Up) 627 return Memory; 628 else 629 return SSE; 630} 631 632void X86_64ABIInfo::classify(QualType Ty, 633 ASTContext &Context, 634 uint64_t OffsetBase, 635 Class &Lo, Class &Hi) const { 636 // FIXME: This code can be simplified by introducing a simple value class for 637 // Class pairs with appropriate constructor methods for the various 638 // situations. 639 640 // FIXME: Some of the split computations are wrong; unaligned vectors 641 // shouldn't be passed in registers for example, so there is no chance they 642 // can straddle an eightbyte. Verify & simplify. 643 644 Lo = Hi = NoClass; 645 646 Class &Current = OffsetBase < 64 ? Lo : Hi; 647 Current = Memory; 648 649 if (const BuiltinType *BT = Ty->getAsBuiltinType()) { 650 BuiltinType::Kind k = BT->getKind(); 651 652 if (k == BuiltinType::Void) { 653 Current = NoClass; 654 } else if (k == BuiltinType::Int128 || k == BuiltinType::UInt128) { 655 Lo = Integer; 656 Hi = Integer; 657 } else if (k >= BuiltinType::Bool && k <= BuiltinType::LongLong) { 658 Current = Integer; 659 } else if (k == BuiltinType::Float || k == BuiltinType::Double) { 660 Current = SSE; 661 } else if (k == BuiltinType::LongDouble) { 662 Lo = X87; 663 Hi = X87Up; 664 } 665 // FIXME: _Decimal32 and _Decimal64 are SSE. 666 // FIXME: _float128 and _Decimal128 are (SSE, SSEUp). 667 } else if (const EnumType *ET = Ty->getAsEnumType()) { 668 // Classify the underlying integer type. 669 classify(ET->getDecl()->getIntegerType(), Context, OffsetBase, Lo, Hi); 670 } else if (Ty->hasPointerRepresentation()) { 671 Current = Integer; 672 } else if (const VectorType *VT = Ty->getAsVectorType()) { 673 uint64_t Size = Context.getTypeSize(VT); 674 if (Size == 32) { 675 // gcc passes all <4 x char>, <2 x short>, <1 x int>, <1 x 676 // float> as integer. 677 Current = Integer; 678 679 // If this type crosses an eightbyte boundary, it should be 680 // split. 681 uint64_t EB_Real = (OffsetBase) / 64; 682 uint64_t EB_Imag = (OffsetBase + Size - 1) / 64; 683 if (EB_Real != EB_Imag) 684 Hi = Lo; 685 } else if (Size == 64) { 686 // gcc passes <1 x double> in memory. :( 687 if (VT->getElementType()->isSpecificBuiltinType(BuiltinType::Double)) 688 return; 689 690 // gcc passes <1 x long long> as INTEGER. 691 if (VT->getElementType()->isSpecificBuiltinType(BuiltinType::LongLong)) 692 Current = Integer; 693 else 694 Current = SSE; 695 696 // If this type crosses an eightbyte boundary, it should be 697 // split. 698 if (OffsetBase && OffsetBase != 64) 699 Hi = Lo; 700 } else if (Size == 128) { 701 Lo = SSE; 702 Hi = SSEUp; 703 } 704 } else if (const ComplexType *CT = Ty->getAsComplexType()) { 705 QualType ET = Context.getCanonicalType(CT->getElementType()); 706 707 uint64_t Size = Context.getTypeSize(Ty); 708 if (ET->isIntegralType()) { 709 if (Size <= 64) 710 Current = Integer; 711 else if (Size <= 128) 712 Lo = Hi = Integer; 713 } else if (ET == Context.FloatTy) 714 Current = SSE; 715 else if (ET == Context.DoubleTy) 716 Lo = Hi = SSE; 717 else if (ET == Context.LongDoubleTy) 718 Current = ComplexX87; 719 720 // If this complex type crosses an eightbyte boundary then it 721 // should be split. 722 uint64_t EB_Real = (OffsetBase) / 64; 723 uint64_t EB_Imag = (OffsetBase + Context.getTypeSize(ET)) / 64; 724 if (Hi == NoClass && EB_Real != EB_Imag) 725 Hi = Lo; 726 } else if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) { 727 // Arrays are treated like structures. 728 729 uint64_t Size = Context.getTypeSize(Ty); 730 731 // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger 732 // than two eightbytes, ..., it has class MEMORY. 733 if (Size > 128) 734 return; 735 736 // AMD64-ABI 3.2.3p2: Rule 1. If ..., or it contains unaligned 737 // fields, it has class MEMORY. 738 // 739 // Only need to check alignment of array base. 740 if (OffsetBase % Context.getTypeAlign(AT->getElementType())) 741 return; 742 743 // Otherwise implement simplified merge. We could be smarter about 744 // this, but it isn't worth it and would be harder to verify. 745 Current = NoClass; 746 uint64_t EltSize = Context.getTypeSize(AT->getElementType()); 747 uint64_t ArraySize = AT->getSize().getZExtValue(); 748 for (uint64_t i=0, Offset=OffsetBase; i<ArraySize; ++i, Offset += EltSize) { 749 Class FieldLo, FieldHi; 750 classify(AT->getElementType(), Context, Offset, FieldLo, FieldHi); 751 Lo = merge(Lo, FieldLo); 752 Hi = merge(Hi, FieldHi); 753 if (Lo == Memory || Hi == Memory) 754 break; 755 } 756 757 // Do post merger cleanup (see below). Only case we worry about is Memory. 758 if (Hi == Memory) 759 Lo = Memory; 760 assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp array classification."); 761 } else if (const RecordType *RT = Ty->getAsRecordType()) { 762 uint64_t Size = Context.getTypeSize(Ty); 763 764 // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger 765 // than two eightbytes, ..., it has class MEMORY. 766 if (Size > 128) 767 return; 768 769 const RecordDecl *RD = RT->getDecl(); 770 771 // Assume variable sized types are passed in memory. 772 if (RD->hasFlexibleArrayMember()) 773 return; 774 775 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD); 776 777 // Reset Lo class, this will be recomputed. 778 Current = NoClass; 779 unsigned idx = 0; 780 for (RecordDecl::field_iterator i = RD->field_begin(Context), 781 e = RD->field_end(Context); i != e; ++i, ++idx) { 782 uint64_t Offset = OffsetBase + Layout.getFieldOffset(idx); 783 bool BitField = i->isBitField(); 784 785 // AMD64-ABI 3.2.3p2: Rule 1. If ..., or it contains unaligned 786 // fields, it has class MEMORY. 787 // 788 // Note, skip this test for bit-fields, see below. 789 if (!BitField && Offset % Context.getTypeAlign(i->getType())) { 790 Lo = Memory; 791 return; 792 } 793 794 // Classify this field. 795 // 796 // AMD64-ABI 3.2.3p2: Rule 3. If the size of the aggregate 797 // exceeds a single eightbyte, each is classified 798 // separately. Each eightbyte gets initialized to class 799 // NO_CLASS. 800 Class FieldLo, FieldHi; 801 802 // Bit-fields require special handling, they do not force the 803 // structure to be passed in memory even if unaligned, and 804 // therefore they can straddle an eightbyte. 805 if (BitField) { 806 // Ignore padding bit-fields. 807 if (i->isUnnamedBitfield()) 808 continue; 809 810 uint64_t Offset = OffsetBase + Layout.getFieldOffset(idx); 811 uint64_t Size = i->getBitWidth()->EvaluateAsInt(Context).getZExtValue(); 812 813 uint64_t EB_Lo = Offset / 64; 814 uint64_t EB_Hi = (Offset + Size - 1) / 64; 815 FieldLo = FieldHi = NoClass; 816 if (EB_Lo) { 817 assert(EB_Hi == EB_Lo && "Invalid classification, type > 16 bytes."); 818 FieldLo = NoClass; 819 FieldHi = Integer; 820 } else { 821 FieldLo = Integer; 822 FieldHi = EB_Hi ? Integer : NoClass; 823 } 824 } else 825 classify(i->getType(), Context, Offset, FieldLo, FieldHi); 826 Lo = merge(Lo, FieldLo); 827 Hi = merge(Hi, FieldHi); 828 if (Lo == Memory || Hi == Memory) 829 break; 830 } 831 832 // AMD64-ABI 3.2.3p2: Rule 5. Then a post merger cleanup is done: 833 // 834 // (a) If one of the classes is MEMORY, the whole argument is 835 // passed in memory. 836 // 837 // (b) If SSEUP is not preceeded by SSE, it is converted to SSE. 838 839 // The first of these conditions is guaranteed by how we implement 840 // the merge (just bail). 841 // 842 // The second condition occurs in the case of unions; for example 843 // union { _Complex double; unsigned; }. 844 if (Hi == Memory) 845 Lo = Memory; 846 if (Hi == SSEUp && Lo != SSE) 847 Hi = SSE; 848 } 849} 850 851ABIArgInfo X86_64ABIInfo::getCoerceResult(QualType Ty, 852 const llvm::Type *CoerceTo, 853 ASTContext &Context) const { 854 if (CoerceTo == llvm::Type::Int64Ty) { 855 // Integer and pointer types will end up in a general purpose 856 // register. 857 if (Ty->isIntegralType() || Ty->isPointerType()) 858 return ABIArgInfo::getDirect(); 859 860 } else if (CoerceTo == llvm::Type::DoubleTy) { 861 // FIXME: It would probably be better to make CGFunctionInfo only map using 862 // canonical types than to canonize here. 863 QualType CTy = Context.getCanonicalType(Ty); 864 865 // Float and double end up in a single SSE reg. 866 if (CTy == Context.FloatTy || CTy == Context.DoubleTy) 867 return ABIArgInfo::getDirect(); 868 869 } 870 871 return ABIArgInfo::getCoerce(CoerceTo); 872} 873 874ABIArgInfo X86_64ABIInfo::classifyReturnType(QualType RetTy, 875 ASTContext &Context) const { 876 // AMD64-ABI 3.2.3p4: Rule 1. Classify the return type with the 877 // classification algorithm. 878 X86_64ABIInfo::Class Lo, Hi; 879 classify(RetTy, Context, 0, Lo, Hi); 880 881 // Check some invariants. 882 assert((Hi != Memory || Lo == Memory) && "Invalid memory classification."); 883 assert((Lo != NoClass || Hi == NoClass) && "Invalid null classification."); 884 assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification."); 885 886 const llvm::Type *ResType = 0; 887 switch (Lo) { 888 case NoClass: 889 return ABIArgInfo::getIgnore(); 890 891 case SSEUp: 892 case X87Up: 893 assert(0 && "Invalid classification for lo word."); 894 895 // AMD64-ABI 3.2.3p4: Rule 2. Types of class memory are returned via 896 // hidden argument. 897 case Memory: 898 return ABIArgInfo::getIndirect(0); 899 900 // AMD64-ABI 3.2.3p4: Rule 3. If the class is INTEGER, the next 901 // available register of the sequence %rax, %rdx is used. 902 case Integer: 903 ResType = llvm::Type::Int64Ty; break; 904 905 // AMD64-ABI 3.2.3p4: Rule 4. If the class is SSE, the next 906 // available SSE register of the sequence %xmm0, %xmm1 is used. 907 case SSE: 908 ResType = llvm::Type::DoubleTy; break; 909 910 // AMD64-ABI 3.2.3p4: Rule 6. If the class is X87, the value is 911 // returned on the X87 stack in %st0 as 80-bit x87 number. 912 case X87: 913 ResType = llvm::Type::X86_FP80Ty; break; 914 915 // AMD64-ABI 3.2.3p4: Rule 8. If the class is COMPLEX_X87, the real 916 // part of the value is returned in %st0 and the imaginary part in 917 // %st1. 918 case ComplexX87: 919 assert(Hi == ComplexX87 && "Unexpected ComplexX87 classification."); 920 ResType = llvm::StructType::get(llvm::Type::X86_FP80Ty, 921 llvm::Type::X86_FP80Ty, 922 NULL); 923 break; 924 } 925 926 switch (Hi) { 927 // Memory was handled previously and X87 should 928 // never occur as a hi class. 929 case Memory: 930 case X87: 931 assert(0 && "Invalid classification for hi word."); 932 933 case ComplexX87: // Previously handled. 934 case NoClass: break; 935 936 case Integer: 937 ResType = llvm::StructType::get(ResType, llvm::Type::Int64Ty, NULL); 938 break; 939 case SSE: 940 ResType = llvm::StructType::get(ResType, llvm::Type::DoubleTy, NULL); 941 break; 942 943 // AMD64-ABI 3.2.3p4: Rule 5. If the class is SSEUP, the eightbyte 944 // is passed in the upper half of the last used SSE register. 945 // 946 // SSEUP should always be preceeded by SSE, just widen. 947 case SSEUp: 948 assert(Lo == SSE && "Unexpected SSEUp classification."); 949 ResType = llvm::VectorType::get(llvm::Type::DoubleTy, 2); 950 break; 951 952 // AMD64-ABI 3.2.3p4: Rule 7. If the class is X87UP, the value is 953 // returned together with the previous X87 value in %st0. 954 case X87Up: 955 // If X87Up is preceeded by X87, we don't need to do 956 // anything. However, in some cases with unions it may not be 957 // preceeded by X87. In such situations we follow gcc and pass the 958 // extra bits in an SSE reg. 959 if (Lo != X87) 960 ResType = llvm::StructType::get(ResType, llvm::Type::DoubleTy, NULL); 961 break; 962 } 963 964 return getCoerceResult(RetTy, ResType, Context); 965} 966 967ABIArgInfo X86_64ABIInfo::classifyArgumentType(QualType Ty, ASTContext &Context, 968 unsigned &neededInt, 969 unsigned &neededSSE) const { 970 X86_64ABIInfo::Class Lo, Hi; 971 classify(Ty, Context, 0, Lo, Hi); 972 973 // Check some invariants. 974 // FIXME: Enforce these by construction. 975 assert((Hi != Memory || Lo == Memory) && "Invalid memory classification."); 976 assert((Lo != NoClass || Hi == NoClass) && "Invalid null classification."); 977 assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification."); 978 979 neededInt = 0; 980 neededSSE = 0; 981 const llvm::Type *ResType = 0; 982 switch (Lo) { 983 case NoClass: 984 return ABIArgInfo::getIgnore(); 985 986 // AMD64-ABI 3.2.3p3: Rule 1. If the class is MEMORY, pass the argument 987 // on the stack. 988 case Memory: 989 990 // AMD64-ABI 3.2.3p3: Rule 5. If the class is X87, X87UP or 991 // COMPLEX_X87, it is passed in memory. 992 case X87: 993 case ComplexX87: 994 return ABIArgInfo::getIndirect(0); 995 996 case SSEUp: 997 case X87Up: 998 assert(0 && "Invalid classification for lo word."); 999 1000 // AMD64-ABI 3.2.3p3: Rule 2. If the class is INTEGER, the next 1001 // available register of the sequence %rdi, %rsi, %rdx, %rcx, %r8 1002 // and %r9 is used. 1003 case Integer: 1004 ++neededInt; 1005 ResType = llvm::Type::Int64Ty; 1006 break; 1007 1008 // AMD64-ABI 3.2.3p3: Rule 3. If the class is SSE, the next 1009 // available SSE register is used, the registers are taken in the 1010 // order from %xmm0 to %xmm7. 1011 case SSE: 1012 ++neededSSE; 1013 ResType = llvm::Type::DoubleTy; 1014 break; 1015 } 1016 1017 switch (Hi) { 1018 // Memory was handled previously, ComplexX87 and X87 should 1019 // never occur as hi classes, and X87Up must be preceed by X87, 1020 // which is passed in memory. 1021 case Memory: 1022 case X87: 1023 case ComplexX87: 1024 assert(0 && "Invalid classification for hi word."); 1025 break; 1026 1027 case NoClass: break; 1028 case Integer: 1029 ResType = llvm::StructType::get(ResType, llvm::Type::Int64Ty, NULL); 1030 ++neededInt; 1031 break; 1032 1033 // X87Up generally doesn't occur here (long double is passed in 1034 // memory), except in situations involving unions. 1035 case X87Up: 1036 case SSE: 1037 ResType = llvm::StructType::get(ResType, llvm::Type::DoubleTy, NULL); 1038 ++neededSSE; 1039 break; 1040 1041 // AMD64-ABI 3.2.3p3: Rule 4. If the class is SSEUP, the 1042 // eightbyte is passed in the upper half of the last used SSE 1043 // register. 1044 case SSEUp: 1045 assert(Lo == SSE && "Unexpected SSEUp classification."); 1046 ResType = llvm::VectorType::get(llvm::Type::DoubleTy, 2); 1047 break; 1048 } 1049 1050 return getCoerceResult(Ty, ResType, Context); 1051} 1052 1053void X86_64ABIInfo::computeInfo(CGFunctionInfo &FI, ASTContext &Context) const { 1054 FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context); 1055 1056 // Keep track of the number of assigned registers. 1057 unsigned freeIntRegs = 6, freeSSERegs = 8; 1058 1059 // AMD64-ABI 3.2.3p3: Once arguments are classified, the registers 1060 // get assigned (in left-to-right order) for passing as follows... 1061 for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end(); 1062 it != ie; ++it) { 1063 unsigned neededInt, neededSSE; 1064 it->info = classifyArgumentType(it->type, Context, neededInt, neededSSE); 1065 1066 // AMD64-ABI 3.2.3p3: If there are no registers available for any 1067 // eightbyte of an argument, the whole argument is passed on the 1068 // stack. If registers have already been assigned for some 1069 // eightbytes of such an argument, the assignments get reverted. 1070 if (freeIntRegs >= neededInt && freeSSERegs >= neededSSE) { 1071 freeIntRegs -= neededInt; 1072 freeSSERegs -= neededSSE; 1073 } else { 1074 it->info = ABIArgInfo::getIndirect(0); 1075 } 1076 } 1077} 1078 1079static llvm::Value *EmitVAArgFromMemory(llvm::Value *VAListAddr, 1080 QualType Ty, 1081 CodeGenFunction &CGF) { 1082 llvm::Value *overflow_arg_area_p = 1083 CGF.Builder.CreateStructGEP(VAListAddr, 2, "overflow_arg_area_p"); 1084 llvm::Value *overflow_arg_area = 1085 CGF.Builder.CreateLoad(overflow_arg_area_p, "overflow_arg_area"); 1086 1087 // AMD64-ABI 3.5.7p5: Step 7. Align l->overflow_arg_area upwards to a 16 1088 // byte boundary if alignment needed by type exceeds 8 byte boundary. 1089 uint64_t Align = CGF.getContext().getTypeAlign(Ty) / 8; 1090 if (Align > 8) { 1091 // Note that we follow the ABI & gcc here, even though the type 1092 // could in theory have an alignment greater than 16. This case 1093 // shouldn't ever matter in practice. 1094 1095 // overflow_arg_area = (overflow_arg_area + 15) & ~15; 1096 llvm::Value *Offset = llvm::ConstantInt::get(llvm::Type::Int32Ty, 15); 1097 overflow_arg_area = CGF.Builder.CreateGEP(overflow_arg_area, Offset); 1098 llvm::Value *AsInt = CGF.Builder.CreatePtrToInt(overflow_arg_area, 1099 llvm::Type::Int64Ty); 1100 llvm::Value *Mask = llvm::ConstantInt::get(llvm::Type::Int64Ty, ~15LL); 1101 overflow_arg_area = 1102 CGF.Builder.CreateIntToPtr(CGF.Builder.CreateAnd(AsInt, Mask), 1103 overflow_arg_area->getType(), 1104 "overflow_arg_area.align"); 1105 } 1106 1107 // AMD64-ABI 3.5.7p5: Step 8. Fetch type from l->overflow_arg_area. 1108 const llvm::Type *LTy = CGF.ConvertTypeForMem(Ty); 1109 llvm::Value *Res = 1110 CGF.Builder.CreateBitCast(overflow_arg_area, 1111 llvm::PointerType::getUnqual(LTy)); 1112 1113 // AMD64-ABI 3.5.7p5: Step 9. Set l->overflow_arg_area to: 1114 // l->overflow_arg_area + sizeof(type). 1115 // AMD64-ABI 3.5.7p5: Step 10. Align l->overflow_arg_area upwards to 1116 // an 8 byte boundary. 1117 1118 uint64_t SizeInBytes = (CGF.getContext().getTypeSize(Ty) + 7) / 8; 1119 llvm::Value *Offset = llvm::ConstantInt::get(llvm::Type::Int32Ty, 1120 (SizeInBytes + 7) & ~7); 1121 overflow_arg_area = CGF.Builder.CreateGEP(overflow_arg_area, Offset, 1122 "overflow_arg_area.next"); 1123 CGF.Builder.CreateStore(overflow_arg_area, overflow_arg_area_p); 1124 1125 // AMD64-ABI 3.5.7p5: Step 11. Return the fetched type. 1126 return Res; 1127} 1128 1129llvm::Value *X86_64ABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 1130 CodeGenFunction &CGF) const { 1131 // Assume that va_list type is correct; should be pointer to LLVM type: 1132 // struct { 1133 // i32 gp_offset; 1134 // i32 fp_offset; 1135 // i8* overflow_arg_area; 1136 // i8* reg_save_area; 1137 // }; 1138 unsigned neededInt, neededSSE; 1139 ABIArgInfo AI = classifyArgumentType(Ty, CGF.getContext(), 1140 neededInt, neededSSE); 1141 1142 // AMD64-ABI 3.5.7p5: Step 1. Determine whether type may be passed 1143 // in the registers. If not go to step 7. 1144 if (!neededInt && !neededSSE) 1145 return EmitVAArgFromMemory(VAListAddr, Ty, CGF); 1146 1147 // AMD64-ABI 3.5.7p5: Step 2. Compute num_gp to hold the number of 1148 // general purpose registers needed to pass type and num_fp to hold 1149 // the number of floating point registers needed. 1150 1151 // AMD64-ABI 3.5.7p5: Step 3. Verify whether arguments fit into 1152 // registers. In the case: l->gp_offset > 48 - num_gp * 8 or 1153 // l->fp_offset > 304 - num_fp * 16 go to step 7. 1154 // 1155 // NOTE: 304 is a typo, there are (6 * 8 + 8 * 16) = 176 bytes of 1156 // register save space). 1157 1158 llvm::Value *InRegs = 0; 1159 llvm::Value *gp_offset_p = 0, *gp_offset = 0; 1160 llvm::Value *fp_offset_p = 0, *fp_offset = 0; 1161 if (neededInt) { 1162 gp_offset_p = CGF.Builder.CreateStructGEP(VAListAddr, 0, "gp_offset_p"); 1163 gp_offset = CGF.Builder.CreateLoad(gp_offset_p, "gp_offset"); 1164 InRegs = 1165 CGF.Builder.CreateICmpULE(gp_offset, 1166 llvm::ConstantInt::get(llvm::Type::Int32Ty, 1167 48 - neededInt * 8), 1168 "fits_in_gp"); 1169 } 1170 1171 if (neededSSE) { 1172 fp_offset_p = CGF.Builder.CreateStructGEP(VAListAddr, 1, "fp_offset_p"); 1173 fp_offset = CGF.Builder.CreateLoad(fp_offset_p, "fp_offset"); 1174 llvm::Value *FitsInFP = 1175 CGF.Builder.CreateICmpULE(fp_offset, 1176 llvm::ConstantInt::get(llvm::Type::Int32Ty, 1177 176 - neededSSE * 16), 1178 "fits_in_fp"); 1179 InRegs = InRegs ? CGF.Builder.CreateAnd(InRegs, FitsInFP) : FitsInFP; 1180 } 1181 1182 llvm::BasicBlock *InRegBlock = CGF.createBasicBlock("vaarg.in_reg"); 1183 llvm::BasicBlock *InMemBlock = CGF.createBasicBlock("vaarg.in_mem"); 1184 llvm::BasicBlock *ContBlock = CGF.createBasicBlock("vaarg.end"); 1185 CGF.Builder.CreateCondBr(InRegs, InRegBlock, InMemBlock); 1186 1187 // Emit code to load the value if it was passed in registers. 1188 1189 CGF.EmitBlock(InRegBlock); 1190 1191 // AMD64-ABI 3.5.7p5: Step 4. Fetch type from l->reg_save_area with 1192 // an offset of l->gp_offset and/or l->fp_offset. This may require 1193 // copying to a temporary location in case the parameter is passed 1194 // in different register classes or requires an alignment greater 1195 // than 8 for general purpose registers and 16 for XMM registers. 1196 // 1197 // FIXME: This really results in shameful code when we end up needing to 1198 // collect arguments from different places; often what should result in a 1199 // simple assembling of a structure from scattered addresses has many more 1200 // loads than necessary. Can we clean this up? 1201 const llvm::Type *LTy = CGF.ConvertTypeForMem(Ty); 1202 llvm::Value *RegAddr = 1203 CGF.Builder.CreateLoad(CGF.Builder.CreateStructGEP(VAListAddr, 3), 1204 "reg_save_area"); 1205 if (neededInt && neededSSE) { 1206 // FIXME: Cleanup. 1207 assert(AI.isCoerce() && "Unexpected ABI info for mixed regs"); 1208 const llvm::StructType *ST = cast<llvm::StructType>(AI.getCoerceToType()); 1209 llvm::Value *Tmp = CGF.CreateTempAlloca(ST); 1210 assert(ST->getNumElements() == 2 && "Unexpected ABI info for mixed regs"); 1211 const llvm::Type *TyLo = ST->getElementType(0); 1212 const llvm::Type *TyHi = ST->getElementType(1); 1213 assert((TyLo->isFloatingPoint() ^ TyHi->isFloatingPoint()) && 1214 "Unexpected ABI info for mixed regs"); 1215 const llvm::Type *PTyLo = llvm::PointerType::getUnqual(TyLo); 1216 const llvm::Type *PTyHi = llvm::PointerType::getUnqual(TyHi); 1217 llvm::Value *GPAddr = CGF.Builder.CreateGEP(RegAddr, gp_offset); 1218 llvm::Value *FPAddr = CGF.Builder.CreateGEP(RegAddr, fp_offset); 1219 llvm::Value *RegLoAddr = TyLo->isFloatingPoint() ? FPAddr : GPAddr; 1220 llvm::Value *RegHiAddr = TyLo->isFloatingPoint() ? GPAddr : FPAddr; 1221 llvm::Value *V = 1222 CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegLoAddr, PTyLo)); 1223 CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(Tmp, 0)); 1224 V = CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegHiAddr, PTyHi)); 1225 CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(Tmp, 1)); 1226 1227 RegAddr = CGF.Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(LTy)); 1228 } else if (neededInt) { 1229 RegAddr = CGF.Builder.CreateGEP(RegAddr, gp_offset); 1230 RegAddr = CGF.Builder.CreateBitCast(RegAddr, 1231 llvm::PointerType::getUnqual(LTy)); 1232 } else { 1233 if (neededSSE == 1) { 1234 RegAddr = CGF.Builder.CreateGEP(RegAddr, fp_offset); 1235 RegAddr = CGF.Builder.CreateBitCast(RegAddr, 1236 llvm::PointerType::getUnqual(LTy)); 1237 } else { 1238 assert(neededSSE == 2 && "Invalid number of needed registers!"); 1239 // SSE registers are spaced 16 bytes apart in the register save 1240 // area, we need to collect the two eightbytes together. 1241 llvm::Value *RegAddrLo = CGF.Builder.CreateGEP(RegAddr, fp_offset); 1242 llvm::Value *RegAddrHi = 1243 CGF.Builder.CreateGEP(RegAddrLo, 1244 llvm::ConstantInt::get(llvm::Type::Int32Ty, 16)); 1245 const llvm::Type *DblPtrTy = 1246 llvm::PointerType::getUnqual(llvm::Type::DoubleTy); 1247 const llvm::StructType *ST = llvm::StructType::get(llvm::Type::DoubleTy, 1248 llvm::Type::DoubleTy, 1249 NULL); 1250 llvm::Value *V, *Tmp = CGF.CreateTempAlloca(ST); 1251 V = CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegAddrLo, 1252 DblPtrTy)); 1253 CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(Tmp, 0)); 1254 V = CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegAddrHi, 1255 DblPtrTy)); 1256 CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(Tmp, 1)); 1257 RegAddr = CGF.Builder.CreateBitCast(Tmp, 1258 llvm::PointerType::getUnqual(LTy)); 1259 } 1260 } 1261 1262 // AMD64-ABI 3.5.7p5: Step 5. Set: 1263 // l->gp_offset = l->gp_offset + num_gp * 8 1264 // l->fp_offset = l->fp_offset + num_fp * 16. 1265 if (neededInt) { 1266 llvm::Value *Offset = llvm::ConstantInt::get(llvm::Type::Int32Ty, 1267 neededInt * 8); 1268 CGF.Builder.CreateStore(CGF.Builder.CreateAdd(gp_offset, Offset), 1269 gp_offset_p); 1270 } 1271 if (neededSSE) { 1272 llvm::Value *Offset = llvm::ConstantInt::get(llvm::Type::Int32Ty, 1273 neededSSE * 16); 1274 CGF.Builder.CreateStore(CGF.Builder.CreateAdd(fp_offset, Offset), 1275 fp_offset_p); 1276 } 1277 CGF.EmitBranch(ContBlock); 1278 1279 // Emit code to load the value if it was passed in memory. 1280 1281 CGF.EmitBlock(InMemBlock); 1282 llvm::Value *MemAddr = EmitVAArgFromMemory(VAListAddr, Ty, CGF); 1283 1284 // Return the appropriate result. 1285 1286 CGF.EmitBlock(ContBlock); 1287 llvm::PHINode *ResAddr = CGF.Builder.CreatePHI(RegAddr->getType(), 1288 "vaarg.addr"); 1289 ResAddr->reserveOperandSpace(2); 1290 ResAddr->addIncoming(RegAddr, InRegBlock); 1291 ResAddr->addIncoming(MemAddr, InMemBlock); 1292 1293 return ResAddr; 1294} 1295 1296// ABI Info for PIC16 1297class PIC16ABIInfo : public ABIInfo { 1298 ABIArgInfo classifyReturnType(QualType RetTy, 1299 ASTContext &Context) const; 1300 1301 ABIArgInfo classifyArgumentType(QualType RetTy, 1302 ASTContext &Context) const; 1303 1304 virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const { 1305 FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context); 1306 for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end(); 1307 it != ie; ++it) 1308 it->info = classifyArgumentType(it->type, Context); 1309 } 1310 1311 virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 1312 CodeGenFunction &CGF) const; 1313 1314}; 1315 1316ABIArgInfo PIC16ABIInfo::classifyReturnType(QualType RetTy, 1317 ASTContext &Context) const { 1318 if (RetTy->isVoidType()) { 1319 return ABIArgInfo::getIgnore(); 1320 } else { 1321 return ABIArgInfo::getDirect(); 1322 } 1323} 1324 1325ABIArgInfo PIC16ABIInfo::classifyArgumentType(QualType Ty, 1326 ASTContext &Context) const { 1327 return ABIArgInfo::getDirect(); 1328} 1329 1330llvm::Value *PIC16ABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 1331 CodeGenFunction &CGF) const { 1332 return 0; 1333} 1334 1335class ARMABIInfo : public ABIInfo { 1336 ABIArgInfo classifyReturnType(QualType RetTy, 1337 ASTContext &Context) const; 1338 1339 ABIArgInfo classifyArgumentType(QualType RetTy, 1340 ASTContext &Context) const; 1341 1342 virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const; 1343 1344 virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 1345 CodeGenFunction &CGF) const; 1346}; 1347 1348void ARMABIInfo::computeInfo(CGFunctionInfo &FI, ASTContext &Context) const { 1349 FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context); 1350 for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end(); 1351 it != ie; ++it) { 1352 it->info = classifyArgumentType(it->type, Context); 1353 } 1354} 1355 1356ABIArgInfo ARMABIInfo::classifyArgumentType(QualType Ty, 1357 ASTContext &Context) const { 1358 if (!CodeGenFunction::hasAggregateLLVMType(Ty)) { 1359 return ABIArgInfo::getDirect(); 1360 } 1361 // FIXME: This is kind of nasty... but there isn't much choice because the ARM 1362 // backend doesn't support byval. 1363 // FIXME: This doesn't handle alignment > 64 bits. 1364 const llvm::Type* ElemTy; 1365 unsigned SizeRegs; 1366 if (Context.getTypeAlign(Ty) > 32) { 1367 ElemTy = llvm::Type::Int64Ty; 1368 SizeRegs = (Context.getTypeSize(Ty) + 63) / 64; 1369 } else { 1370 ElemTy = llvm::Type::Int32Ty; 1371 SizeRegs = (Context.getTypeSize(Ty) + 31) / 32; 1372 } 1373 std::vector<const llvm::Type*> LLVMFields; 1374 LLVMFields.push_back(llvm::ArrayType::get(ElemTy, SizeRegs)); 1375 const llvm::Type* STy = llvm::StructType::get(LLVMFields, true); 1376 return ABIArgInfo::getCoerce(STy); 1377} 1378 1379ABIArgInfo ARMABIInfo::classifyReturnType(QualType RetTy, 1380 ASTContext &Context) const { 1381 if (RetTy->isVoidType()) { 1382 return ABIArgInfo::getIgnore(); 1383 } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 1384 // Aggregates <= 4 bytes are returned in r0; other aggregates 1385 // are returned indirectly. 1386 uint64_t Size = Context.getTypeSize(RetTy); 1387 if (Size <= 32) 1388 return ABIArgInfo::getCoerce(llvm::Type::Int32Ty); 1389 return ABIArgInfo::getIndirect(0); 1390 } else { 1391 return ABIArgInfo::getDirect(); 1392 } 1393} 1394 1395llvm::Value *ARMABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 1396 CodeGenFunction &CGF) const { 1397 // FIXME: Need to handle alignment 1398 const llvm::Type *BP = llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 1399 const llvm::Type *BPP = llvm::PointerType::getUnqual(BP); 1400 1401 CGBuilderTy &Builder = CGF.Builder; 1402 llvm::Value *VAListAddrAsBPP = Builder.CreateBitCast(VAListAddr, BPP, 1403 "ap"); 1404 llvm::Value *Addr = Builder.CreateLoad(VAListAddrAsBPP, "ap.cur"); 1405 llvm::Type *PTy = 1406 llvm::PointerType::getUnqual(CGF.ConvertType(Ty)); 1407 llvm::Value *AddrTyped = Builder.CreateBitCast(Addr, PTy); 1408 1409 uint64_t Offset = 1410 llvm::RoundUpToAlignment(CGF.getContext().getTypeSize(Ty) / 8, 4); 1411 llvm::Value *NextAddr = 1412 Builder.CreateGEP(Addr, 1413 llvm::ConstantInt::get(llvm::Type::Int32Ty, Offset), 1414 "ap.next"); 1415 Builder.CreateStore(NextAddr, VAListAddrAsBPP); 1416 1417 return AddrTyped; 1418} 1419 1420ABIArgInfo DefaultABIInfo::classifyReturnType(QualType RetTy, 1421 ASTContext &Context) const { 1422 if (RetTy->isVoidType()) { 1423 return ABIArgInfo::getIgnore(); 1424 } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 1425 return ABIArgInfo::getIndirect(0); 1426 } else { 1427 return ABIArgInfo::getDirect(); 1428 } 1429} 1430 1431ABIArgInfo DefaultABIInfo::classifyArgumentType(QualType Ty, 1432 ASTContext &Context) const { 1433 if (CodeGenFunction::hasAggregateLLVMType(Ty)) { 1434 return ABIArgInfo::getIndirect(0); 1435 } else { 1436 return ABIArgInfo::getDirect(); 1437 } 1438} 1439 1440llvm::Value *DefaultABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty, 1441 CodeGenFunction &CGF) const { 1442 return 0; 1443} 1444 1445const ABIInfo &CodeGenTypes::getABIInfo() const { 1446 if (TheABIInfo) 1447 return *TheABIInfo; 1448 1449 // For now we just cache this in the CodeGenTypes and don't bother 1450 // to free it. 1451 const char *TargetPrefix = getContext().Target.getTargetPrefix(); 1452 if (strcmp(TargetPrefix, "x86") == 0) { 1453 bool IsDarwin = strstr(getContext().Target.getTargetTriple(), "darwin"); 1454 switch (getContext().Target.getPointerWidth(0)) { 1455 case 32: 1456 return *(TheABIInfo = new X86_32ABIInfo(Context, IsDarwin)); 1457 case 64: 1458 return *(TheABIInfo = new X86_64ABIInfo()); 1459 } 1460 } else if (strcmp(TargetPrefix, "arm") == 0) { 1461 // FIXME: Support for OABI? 1462 return *(TheABIInfo = new ARMABIInfo()); 1463 } else if (strcmp(TargetPrefix, "pic16") == 0) { 1464 return *(TheABIInfo = new PIC16ABIInfo()); 1465 } 1466 1467 return *(TheABIInfo = new DefaultABIInfo); 1468} 1469 1470/***/ 1471 1472CGFunctionInfo::CGFunctionInfo(QualType ResTy, 1473 const llvm::SmallVector<QualType, 16> &ArgTys) { 1474 NumArgs = ArgTys.size(); 1475 Args = new ArgInfo[1 + NumArgs]; 1476 Args[0].type = ResTy; 1477 for (unsigned i = 0; i < NumArgs; ++i) 1478 Args[1 + i].type = ArgTys[i]; 1479} 1480 1481/***/ 1482 1483void CodeGenTypes::GetExpandedTypes(QualType Ty, 1484 std::vector<const llvm::Type*> &ArgTys) { 1485 const RecordType *RT = Ty->getAsStructureType(); 1486 assert(RT && "Can only expand structure types."); 1487 const RecordDecl *RD = RT->getDecl(); 1488 assert(!RD->hasFlexibleArrayMember() && 1489 "Cannot expand structure with flexible array."); 1490 1491 for (RecordDecl::field_iterator i = RD->field_begin(Context), 1492 e = RD->field_end(Context); i != e; ++i) { 1493 const FieldDecl *FD = *i; 1494 assert(!FD->isBitField() && 1495 "Cannot expand structure with bit-field members."); 1496 1497 QualType FT = FD->getType(); 1498 if (CodeGenFunction::hasAggregateLLVMType(FT)) { 1499 GetExpandedTypes(FT, ArgTys); 1500 } else { 1501 ArgTys.push_back(ConvertType(FT)); 1502 } 1503 } 1504} 1505 1506llvm::Function::arg_iterator 1507CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV, 1508 llvm::Function::arg_iterator AI) { 1509 const RecordType *RT = Ty->getAsStructureType(); 1510 assert(RT && "Can only expand structure types."); 1511 1512 RecordDecl *RD = RT->getDecl(); 1513 assert(LV.isSimple() && 1514 "Unexpected non-simple lvalue during struct expansion."); 1515 llvm::Value *Addr = LV.getAddress(); 1516 for (RecordDecl::field_iterator i = RD->field_begin(getContext()), 1517 e = RD->field_end(getContext()); i != e; ++i) { 1518 FieldDecl *FD = *i; 1519 QualType FT = FD->getType(); 1520 1521 // FIXME: What are the right qualifiers here? 1522 LValue LV = EmitLValueForField(Addr, FD, false, 0); 1523 if (CodeGenFunction::hasAggregateLLVMType(FT)) { 1524 AI = ExpandTypeFromArgs(FT, LV, AI); 1525 } else { 1526 EmitStoreThroughLValue(RValue::get(AI), LV, FT); 1527 ++AI; 1528 } 1529 } 1530 1531 return AI; 1532} 1533 1534void 1535CodeGenFunction::ExpandTypeToArgs(QualType Ty, RValue RV, 1536 llvm::SmallVector<llvm::Value*, 16> &Args) { 1537 const RecordType *RT = Ty->getAsStructureType(); 1538 assert(RT && "Can only expand structure types."); 1539 1540 RecordDecl *RD = RT->getDecl(); 1541 assert(RV.isAggregate() && "Unexpected rvalue during struct expansion"); 1542 llvm::Value *Addr = RV.getAggregateAddr(); 1543 for (RecordDecl::field_iterator i = RD->field_begin(getContext()), 1544 e = RD->field_end(getContext()); i != e; ++i) { 1545 FieldDecl *FD = *i; 1546 QualType FT = FD->getType(); 1547 1548 // FIXME: What are the right qualifiers here? 1549 LValue LV = EmitLValueForField(Addr, FD, false, 0); 1550 if (CodeGenFunction::hasAggregateLLVMType(FT)) { 1551 ExpandTypeToArgs(FT, RValue::getAggregate(LV.getAddress()), Args); 1552 } else { 1553 RValue RV = EmitLoadOfLValue(LV, FT); 1554 assert(RV.isScalar() && 1555 "Unexpected non-scalar rvalue during struct expansion."); 1556 Args.push_back(RV.getScalarVal()); 1557 } 1558 } 1559} 1560 1561/// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as 1562/// a pointer to an object of type \arg Ty. 1563/// 1564/// This safely handles the case when the src type is smaller than the 1565/// destination type; in this situation the values of bits which not 1566/// present in the src are undefined. 1567static llvm::Value *CreateCoercedLoad(llvm::Value *SrcPtr, 1568 const llvm::Type *Ty, 1569 CodeGenFunction &CGF) { 1570 const llvm::Type *SrcTy = 1571 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 1572 uint64_t SrcSize = CGF.CGM.getTargetData().getTypeAllocSize(SrcTy); 1573 uint64_t DstSize = CGF.CGM.getTargetData().getTypeAllocSize(Ty); 1574 1575 // If load is legal, just bitcast the src pointer. 1576 if (SrcSize >= DstSize) { 1577 // Generally SrcSize is never greater than DstSize, since this means we are 1578 // losing bits. However, this can happen in cases where the structure has 1579 // additional padding, for example due to a user specified alignment. 1580 // 1581 // FIXME: Assert that we aren't truncating non-padding bits when have access 1582 // to that information. 1583 llvm::Value *Casted = 1584 CGF.Builder.CreateBitCast(SrcPtr, llvm::PointerType::getUnqual(Ty)); 1585 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted); 1586 // FIXME: Use better alignment / avoid requiring aligned load. 1587 Load->setAlignment(1); 1588 return Load; 1589 } else { 1590 // Otherwise do coercion through memory. This is stupid, but 1591 // simple. 1592 llvm::Value *Tmp = CGF.CreateTempAlloca(Ty); 1593 llvm::Value *Casted = 1594 CGF.Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(SrcTy)); 1595 llvm::StoreInst *Store = 1596 CGF.Builder.CreateStore(CGF.Builder.CreateLoad(SrcPtr), Casted); 1597 // FIXME: Use better alignment / avoid requiring aligned store. 1598 Store->setAlignment(1); 1599 return CGF.Builder.CreateLoad(Tmp); 1600 } 1601} 1602 1603/// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src, 1604/// where the source and destination may have different types. 1605/// 1606/// This safely handles the case when the src type is larger than the 1607/// destination type; the upper bits of the src will be lost. 1608static void CreateCoercedStore(llvm::Value *Src, 1609 llvm::Value *DstPtr, 1610 CodeGenFunction &CGF) { 1611 const llvm::Type *SrcTy = Src->getType(); 1612 const llvm::Type *DstTy = 1613 cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 1614 1615 uint64_t SrcSize = CGF.CGM.getTargetData().getTypeAllocSize(SrcTy); 1616 uint64_t DstSize = CGF.CGM.getTargetData().getTypeAllocSize(DstTy); 1617 1618 // If store is legal, just bitcast the src pointer. 1619 if (SrcSize >= DstSize) { 1620 // Generally SrcSize is never greater than DstSize, since this means we are 1621 // losing bits. However, this can happen in cases where the structure has 1622 // additional padding, for example due to a user specified alignment. 1623 // 1624 // FIXME: Assert that we aren't truncating non-padding bits when have access 1625 // to that information. 1626 llvm::Value *Casted = 1627 CGF.Builder.CreateBitCast(DstPtr, llvm::PointerType::getUnqual(SrcTy)); 1628 // FIXME: Use better alignment / avoid requiring aligned store. 1629 CGF.Builder.CreateStore(Src, Casted)->setAlignment(1); 1630 } else { 1631 // Otherwise do coercion through memory. This is stupid, but 1632 // simple. 1633 llvm::Value *Tmp = CGF.CreateTempAlloca(SrcTy); 1634 CGF.Builder.CreateStore(Src, Tmp); 1635 llvm::Value *Casted = 1636 CGF.Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(DstTy)); 1637 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted); 1638 // FIXME: Use better alignment / avoid requiring aligned load. 1639 Load->setAlignment(1); 1640 CGF.Builder.CreateStore(Load, DstPtr); 1641 } 1642} 1643 1644/***/ 1645 1646bool CodeGenModule::ReturnTypeUsesSret(const CGFunctionInfo &FI) { 1647 return FI.getReturnInfo().isIndirect(); 1648} 1649 1650const llvm::FunctionType * 1651CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI, bool IsVariadic) { 1652 std::vector<const llvm::Type*> ArgTys; 1653 1654 const llvm::Type *ResultType = 0; 1655 1656 QualType RetTy = FI.getReturnType(); 1657 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1658 switch (RetAI.getKind()) { 1659 case ABIArgInfo::Expand: 1660 assert(0 && "Invalid ABI kind for return argument"); 1661 1662 case ABIArgInfo::Direct: 1663 ResultType = ConvertType(RetTy); 1664 break; 1665 1666 case ABIArgInfo::Indirect: { 1667 assert(!RetAI.getIndirectAlign() && "Align unused on indirect return."); 1668 ResultType = llvm::Type::VoidTy; 1669 const llvm::Type *STy = ConvertType(RetTy); 1670 ArgTys.push_back(llvm::PointerType::get(STy, RetTy.getAddressSpace())); 1671 break; 1672 } 1673 1674 case ABIArgInfo::Ignore: 1675 ResultType = llvm::Type::VoidTy; 1676 break; 1677 1678 case ABIArgInfo::Coerce: 1679 ResultType = RetAI.getCoerceToType(); 1680 break; 1681 } 1682 1683 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 1684 ie = FI.arg_end(); it != ie; ++it) { 1685 const ABIArgInfo &AI = it->info; 1686 1687 switch (AI.getKind()) { 1688 case ABIArgInfo::Ignore: 1689 break; 1690 1691 case ABIArgInfo::Coerce: 1692 ArgTys.push_back(AI.getCoerceToType()); 1693 break; 1694 1695 case ABIArgInfo::Indirect: { 1696 // indirect arguments are always on the stack, which is addr space #0. 1697 const llvm::Type *LTy = ConvertTypeForMem(it->type); 1698 ArgTys.push_back(llvm::PointerType::getUnqual(LTy)); 1699 break; 1700 } 1701 1702 case ABIArgInfo::Direct: 1703 ArgTys.push_back(ConvertType(it->type)); 1704 break; 1705 1706 case ABIArgInfo::Expand: 1707 GetExpandedTypes(it->type, ArgTys); 1708 break; 1709 } 1710 } 1711 1712 return llvm::FunctionType::get(ResultType, ArgTys, IsVariadic); 1713} 1714 1715void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI, 1716 const Decl *TargetDecl, 1717 AttributeListType &PAL) { 1718 unsigned FuncAttrs = 0; 1719 unsigned RetAttrs = 0; 1720 1721 // FIXME: handle sseregparm someday... 1722 if (TargetDecl) { 1723 if (TargetDecl->hasAttr<NoThrowAttr>()) 1724 FuncAttrs |= llvm::Attribute::NoUnwind; 1725 if (TargetDecl->hasAttr<NoReturnAttr>()) 1726 FuncAttrs |= llvm::Attribute::NoReturn; 1727 if (TargetDecl->hasAttr<ConstAttr>()) 1728 FuncAttrs |= llvm::Attribute::ReadNone; 1729 else if (TargetDecl->hasAttr<PureAttr>()) 1730 FuncAttrs |= llvm::Attribute::ReadOnly; 1731 } 1732 1733 QualType RetTy = FI.getReturnType(); 1734 unsigned Index = 1; 1735 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1736 switch (RetAI.getKind()) { 1737 case ABIArgInfo::Direct: 1738 if (RetTy->isPromotableIntegerType()) { 1739 if (RetTy->isSignedIntegerType()) { 1740 RetAttrs |= llvm::Attribute::SExt; 1741 } else if (RetTy->isUnsignedIntegerType()) { 1742 RetAttrs |= llvm::Attribute::ZExt; 1743 } 1744 } 1745 break; 1746 1747 case ABIArgInfo::Indirect: 1748 PAL.push_back(llvm::AttributeWithIndex::get(Index, 1749 llvm::Attribute::StructRet | 1750 llvm::Attribute::NoAlias)); 1751 ++Index; 1752 // sret disables readnone and readonly 1753 FuncAttrs &= ~(llvm::Attribute::ReadOnly | 1754 llvm::Attribute::ReadNone); 1755 break; 1756 1757 case ABIArgInfo::Ignore: 1758 case ABIArgInfo::Coerce: 1759 break; 1760 1761 case ABIArgInfo::Expand: 1762 assert(0 && "Invalid ABI kind for return argument"); 1763 } 1764 1765 if (RetAttrs) 1766 PAL.push_back(llvm::AttributeWithIndex::get(0, RetAttrs)); 1767 1768 // FIXME: we need to honour command line settings also... 1769 // FIXME: RegParm should be reduced in case of nested functions and/or global 1770 // register variable. 1771 signed RegParm = 0; 1772 if (TargetDecl) 1773 if (const RegparmAttr *RegParmAttr = TargetDecl->getAttr<RegparmAttr>()) 1774 RegParm = RegParmAttr->getNumParams(); 1775 1776 unsigned PointerWidth = getContext().Target.getPointerWidth(0); 1777 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 1778 ie = FI.arg_end(); it != ie; ++it) { 1779 QualType ParamType = it->type; 1780 const ABIArgInfo &AI = it->info; 1781 unsigned Attributes = 0; 1782 1783 switch (AI.getKind()) { 1784 case ABIArgInfo::Coerce: 1785 break; 1786 1787 case ABIArgInfo::Indirect: 1788 Attributes |= llvm::Attribute::ByVal; 1789 Attributes |= 1790 llvm::Attribute::constructAlignmentFromInt(AI.getIndirectAlign()); 1791 // byval disables readnone and readonly. 1792 FuncAttrs &= ~(llvm::Attribute::ReadOnly | 1793 llvm::Attribute::ReadNone); 1794 break; 1795 1796 case ABIArgInfo::Direct: 1797 if (ParamType->isPromotableIntegerType()) { 1798 if (ParamType->isSignedIntegerType()) { 1799 Attributes |= llvm::Attribute::SExt; 1800 } else if (ParamType->isUnsignedIntegerType()) { 1801 Attributes |= llvm::Attribute::ZExt; 1802 } 1803 } 1804 if (RegParm > 0 && 1805 (ParamType->isIntegerType() || ParamType->isPointerType())) { 1806 RegParm -= 1807 (Context.getTypeSize(ParamType) + PointerWidth - 1) / PointerWidth; 1808 if (RegParm >= 0) 1809 Attributes |= llvm::Attribute::InReg; 1810 } 1811 // FIXME: handle sseregparm someday... 1812 break; 1813 1814 case ABIArgInfo::Ignore: 1815 // Skip increment, no matching LLVM parameter. 1816 continue; 1817 1818 case ABIArgInfo::Expand: { 1819 std::vector<const llvm::Type*> Tys; 1820 // FIXME: This is rather inefficient. Do we ever actually need to do 1821 // anything here? The result should be just reconstructed on the other 1822 // side, so extension should be a non-issue. 1823 getTypes().GetExpandedTypes(ParamType, Tys); 1824 Index += Tys.size(); 1825 continue; 1826 } 1827 } 1828 1829 if (Attributes) 1830 PAL.push_back(llvm::AttributeWithIndex::get(Index, Attributes)); 1831 ++Index; 1832 } 1833 if (FuncAttrs) 1834 PAL.push_back(llvm::AttributeWithIndex::get(~0, FuncAttrs)); 1835} 1836 1837void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI, 1838 llvm::Function *Fn, 1839 const FunctionArgList &Args) { 1840 // FIXME: We no longer need the types from FunctionArgList; lift up and 1841 // simplify. 1842 1843 // Emit allocs for param decls. Give the LLVM Argument nodes names. 1844 llvm::Function::arg_iterator AI = Fn->arg_begin(); 1845 1846 // Name the struct return argument. 1847 if (CGM.ReturnTypeUsesSret(FI)) { 1848 AI->setName("agg.result"); 1849 ++AI; 1850 } 1851 1852 assert(FI.arg_size() == Args.size() && 1853 "Mismatch between function signature & arguments."); 1854 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin(); 1855 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 1856 i != e; ++i, ++info_it) { 1857 const VarDecl *Arg = i->first; 1858 QualType Ty = info_it->type; 1859 const ABIArgInfo &ArgI = info_it->info; 1860 1861 switch (ArgI.getKind()) { 1862 case ABIArgInfo::Indirect: { 1863 llvm::Value* V = AI; 1864 if (hasAggregateLLVMType(Ty)) { 1865 // Do nothing, aggregates and complex variables are accessed by 1866 // reference. 1867 } else { 1868 // Load scalar value from indirect argument. 1869 V = EmitLoadOfScalar(V, false, Ty); 1870 if (!getContext().typesAreCompatible(Ty, Arg->getType())) { 1871 // This must be a promotion, for something like 1872 // "void a(x) short x; {..." 1873 V = EmitScalarConversion(V, Ty, Arg->getType()); 1874 } 1875 } 1876 EmitParmDecl(*Arg, V); 1877 break; 1878 } 1879 1880 case ABIArgInfo::Direct: { 1881 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1882 llvm::Value* V = AI; 1883 if (hasAggregateLLVMType(Ty)) { 1884 // Create a temporary alloca to hold the argument; the rest of 1885 // codegen expects to access aggregates & complex values by 1886 // reference. 1887 V = CreateTempAlloca(ConvertTypeForMem(Ty)); 1888 Builder.CreateStore(AI, V); 1889 } else { 1890 if (!getContext().typesAreCompatible(Ty, Arg->getType())) { 1891 // This must be a promotion, for something like 1892 // "void a(x) short x; {..." 1893 V = EmitScalarConversion(V, Ty, Arg->getType()); 1894 } 1895 } 1896 EmitParmDecl(*Arg, V); 1897 break; 1898 } 1899 1900 case ABIArgInfo::Expand: { 1901 // If this structure was expanded into multiple arguments then 1902 // we need to create a temporary and reconstruct it from the 1903 // arguments. 1904 std::string Name = Arg->getNameAsString(); 1905 llvm::Value *Temp = CreateTempAlloca(ConvertTypeForMem(Ty), 1906 (Name + ".addr").c_str()); 1907 // FIXME: What are the right qualifiers here? 1908 llvm::Function::arg_iterator End = 1909 ExpandTypeFromArgs(Ty, LValue::MakeAddr(Temp,0), AI); 1910 EmitParmDecl(*Arg, Temp); 1911 1912 // Name the arguments used in expansion and increment AI. 1913 unsigned Index = 0; 1914 for (; AI != End; ++AI, ++Index) 1915 AI->setName(Name + "." + llvm::utostr(Index)); 1916 continue; 1917 } 1918 1919 case ABIArgInfo::Ignore: 1920 // Initialize the local variable appropriately. 1921 if (hasAggregateLLVMType(Ty)) { 1922 EmitParmDecl(*Arg, CreateTempAlloca(ConvertTypeForMem(Ty))); 1923 } else { 1924 EmitParmDecl(*Arg, llvm::UndefValue::get(ConvertType(Arg->getType()))); 1925 } 1926 1927 // Skip increment, no matching LLVM parameter. 1928 continue; 1929 1930 case ABIArgInfo::Coerce: { 1931 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1932 // FIXME: This is very wasteful; EmitParmDecl is just going to drop the 1933 // result in a new alloca anyway, so we could just store into that 1934 // directly if we broke the abstraction down more. 1935 llvm::Value *V = CreateTempAlloca(ConvertTypeForMem(Ty), "coerce"); 1936 CreateCoercedStore(AI, V, *this); 1937 // Match to what EmitParmDecl is expecting for this type. 1938 if (!CodeGenFunction::hasAggregateLLVMType(Ty)) { 1939 V = EmitLoadOfScalar(V, false, Ty); 1940 if (!getContext().typesAreCompatible(Ty, Arg->getType())) { 1941 // This must be a promotion, for something like 1942 // "void a(x) short x; {..." 1943 V = EmitScalarConversion(V, Ty, Arg->getType()); 1944 } 1945 } 1946 EmitParmDecl(*Arg, V); 1947 break; 1948 } 1949 } 1950 1951 ++AI; 1952 } 1953 assert(AI == Fn->arg_end() && "Argument mismatch!"); 1954} 1955 1956void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI, 1957 llvm::Value *ReturnValue) { 1958 llvm::Value *RV = 0; 1959 1960 // Functions with no result always return void. 1961 if (ReturnValue) { 1962 QualType RetTy = FI.getReturnType(); 1963 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1964 1965 switch (RetAI.getKind()) { 1966 case ABIArgInfo::Indirect: 1967 if (RetTy->isAnyComplexType()) { 1968 ComplexPairTy RT = LoadComplexFromAddr(ReturnValue, false); 1969 StoreComplexToAddr(RT, CurFn->arg_begin(), false); 1970 } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 1971 EmitAggregateCopy(CurFn->arg_begin(), ReturnValue, RetTy); 1972 } else { 1973 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue), CurFn->arg_begin(), 1974 false, RetTy); 1975 } 1976 break; 1977 1978 case ABIArgInfo::Direct: 1979 // The internal return value temp always will have 1980 // pointer-to-return-type type. 1981 RV = Builder.CreateLoad(ReturnValue); 1982 break; 1983 1984 case ABIArgInfo::Ignore: 1985 break; 1986 1987 case ABIArgInfo::Coerce: 1988 RV = CreateCoercedLoad(ReturnValue, RetAI.getCoerceToType(), *this); 1989 break; 1990 1991 case ABIArgInfo::Expand: 1992 assert(0 && "Invalid ABI kind for return argument"); 1993 } 1994 } 1995 1996 if (RV) { 1997 Builder.CreateRet(RV); 1998 } else { 1999 Builder.CreateRetVoid(); 2000 } 2001} 2002 2003RValue CodeGenFunction::EmitCallArg(const Expr *E, QualType ArgType) { 2004 if (ArgType->isReferenceType()) 2005 return EmitReferenceBindingToExpr(E, ArgType); 2006 2007 return EmitAnyExprToTemp(E); 2008} 2009 2010RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo, 2011 llvm::Value *Callee, 2012 const CallArgList &CallArgs, 2013 const Decl *TargetDecl) { 2014 // FIXME: We no longer need the types from CallArgs; lift up and simplify. 2015 llvm::SmallVector<llvm::Value*, 16> Args; 2016 2017 // Handle struct-return functions by passing a pointer to the 2018 // location that we would like to return into. 2019 QualType RetTy = CallInfo.getReturnType(); 2020 const ABIArgInfo &RetAI = CallInfo.getReturnInfo(); 2021 if (CGM.ReturnTypeUsesSret(CallInfo)) { 2022 // Create a temporary alloca to hold the result of the call. :( 2023 Args.push_back(CreateTempAlloca(ConvertTypeForMem(RetTy))); 2024 } 2025 2026 assert(CallInfo.arg_size() == CallArgs.size() && 2027 "Mismatch between function signature & arguments."); 2028 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin(); 2029 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end(); 2030 I != E; ++I, ++info_it) { 2031 const ABIArgInfo &ArgInfo = info_it->info; 2032 RValue RV = I->first; 2033 2034 switch (ArgInfo.getKind()) { 2035 case ABIArgInfo::Indirect: 2036 if (RV.isScalar() || RV.isComplex()) { 2037 // Make a temporary alloca to pass the argument. 2038 Args.push_back(CreateTempAlloca(ConvertTypeForMem(I->second))); 2039 if (RV.isScalar()) 2040 EmitStoreOfScalar(RV.getScalarVal(), Args.back(), false, I->second); 2041 else 2042 StoreComplexToAddr(RV.getComplexVal(), Args.back(), false); 2043 } else { 2044 Args.push_back(RV.getAggregateAddr()); 2045 } 2046 break; 2047 2048 case ABIArgInfo::Direct: 2049 if (RV.isScalar()) { 2050 Args.push_back(RV.getScalarVal()); 2051 } else if (RV.isComplex()) { 2052 llvm::Value *Tmp = llvm::UndefValue::get(ConvertType(I->second)); 2053 Tmp = Builder.CreateInsertValue(Tmp, RV.getComplexVal().first, 0); 2054 Tmp = Builder.CreateInsertValue(Tmp, RV.getComplexVal().second, 1); 2055 Args.push_back(Tmp); 2056 } else { 2057 Args.push_back(Builder.CreateLoad(RV.getAggregateAddr())); 2058 } 2059 break; 2060 2061 case ABIArgInfo::Ignore: 2062 break; 2063 2064 case ABIArgInfo::Coerce: { 2065 // FIXME: Avoid the conversion through memory if possible. 2066 llvm::Value *SrcPtr; 2067 if (RV.isScalar()) { 2068 SrcPtr = CreateTempAlloca(ConvertTypeForMem(I->second), "coerce"); 2069 EmitStoreOfScalar(RV.getScalarVal(), SrcPtr, false, I->second); 2070 } else if (RV.isComplex()) { 2071 SrcPtr = CreateTempAlloca(ConvertTypeForMem(I->second), "coerce"); 2072 StoreComplexToAddr(RV.getComplexVal(), SrcPtr, false); 2073 } else 2074 SrcPtr = RV.getAggregateAddr(); 2075 Args.push_back(CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(), 2076 *this)); 2077 break; 2078 } 2079 2080 case ABIArgInfo::Expand: 2081 ExpandTypeToArgs(I->second, RV, Args); 2082 break; 2083 } 2084 } 2085 2086 llvm::BasicBlock *InvokeDest = getInvokeDest(); 2087 CodeGen::AttributeListType AttributeList; 2088 CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList); 2089 llvm::AttrListPtr Attrs = llvm::AttrListPtr::get(AttributeList.begin(), 2090 AttributeList.end()); 2091 2092 llvm::CallSite CS; 2093 if (!InvokeDest || (Attrs.getFnAttributes() & llvm::Attribute::NoUnwind)) { 2094 CS = Builder.CreateCall(Callee, Args.data(), Args.data()+Args.size()); 2095 } else { 2096 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont"); 2097 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, 2098 Args.data(), Args.data()+Args.size()); 2099 EmitBlock(Cont); 2100 } 2101 2102 CS.setAttributes(Attrs); 2103 if (const llvm::Function *F = dyn_cast<llvm::Function>(Callee->stripPointerCasts())) 2104 CS.setCallingConv(F->getCallingConv()); 2105 2106 // If the call doesn't return, finish the basic block and clear the 2107 // insertion point; this allows the rest of IRgen to discard 2108 // unreachable code. 2109 if (CS.doesNotReturn()) { 2110 Builder.CreateUnreachable(); 2111 Builder.ClearInsertionPoint(); 2112 2113 // FIXME: For now, emit a dummy basic block because expr emitters in 2114 // generally are not ready to handle emitting expressions at unreachable 2115 // points. 2116 EnsureInsertPoint(); 2117 2118 // Return a reasonable RValue. 2119 return GetUndefRValue(RetTy); 2120 } 2121 2122 llvm::Instruction *CI = CS.getInstruction(); 2123 if (Builder.isNamePreserving() && CI->getType() != llvm::Type::VoidTy) 2124 CI->setName("call"); 2125 2126 switch (RetAI.getKind()) { 2127 case ABIArgInfo::Indirect: 2128 if (RetTy->isAnyComplexType()) 2129 return RValue::getComplex(LoadComplexFromAddr(Args[0], false)); 2130 if (CodeGenFunction::hasAggregateLLVMType(RetTy)) 2131 return RValue::getAggregate(Args[0]); 2132 return RValue::get(EmitLoadOfScalar(Args[0], false, RetTy)); 2133 2134 case ABIArgInfo::Direct: 2135 if (RetTy->isAnyComplexType()) { 2136 llvm::Value *Real = Builder.CreateExtractValue(CI, 0); 2137 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1); 2138 return RValue::getComplex(std::make_pair(Real, Imag)); 2139 } 2140 if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 2141 llvm::Value *V = CreateTempAlloca(ConvertTypeForMem(RetTy), "agg.tmp"); 2142 Builder.CreateStore(CI, V); 2143 return RValue::getAggregate(V); 2144 } 2145 return RValue::get(CI); 2146 2147 case ABIArgInfo::Ignore: 2148 // If we are ignoring an argument that had a result, make sure to 2149 // construct the appropriate return value for our caller. 2150 return GetUndefRValue(RetTy); 2151 2152 case ABIArgInfo::Coerce: { 2153 // FIXME: Avoid the conversion through memory if possible. 2154 llvm::Value *V = CreateTempAlloca(ConvertTypeForMem(RetTy), "coerce"); 2155 CreateCoercedStore(CI, V, *this); 2156 if (RetTy->isAnyComplexType()) 2157 return RValue::getComplex(LoadComplexFromAddr(V, false)); 2158 if (CodeGenFunction::hasAggregateLLVMType(RetTy)) 2159 return RValue::getAggregate(V); 2160 return RValue::get(EmitLoadOfScalar(V, false, RetTy)); 2161 } 2162 2163 case ABIArgInfo::Expand: 2164 assert(0 && "Invalid ABI kind for return argument"); 2165 } 2166 2167 assert(0 && "Unhandled ABIArgInfo::Kind"); 2168 return RValue::get(0); 2169} 2170 2171/* VarArg handling */ 2172 2173llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) { 2174 return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this); 2175} 2176