CodeGenFunction.cpp revision 88b5396b0897f28d22ae3debf4a0d97b33b6c362
1//===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===// 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// This coordinates the per-function state used while generating code. 11// 12//===----------------------------------------------------------------------===// 13 14#include "CodeGenFunction.h" 15#include "CodeGenModule.h" 16#include "CGDebugInfo.h" 17#include "clang/Basic/TargetInfo.h" 18#include "clang/AST/APValue.h" 19#include "clang/AST/ASTContext.h" 20#include "clang/AST/Decl.h" 21#include "llvm/Support/CFG.h" 22using namespace clang; 23using namespace CodeGen; 24 25CodeGenFunction::CodeGenFunction(CodeGenModule &cgm) 26 : CGM(cgm), Target(CGM.getContext().Target), SwitchInsn(NULL), 27 CaseRangeBlock(NULL) { 28 LLVMIntTy = ConvertType(getContext().IntTy); 29 LLVMPointerWidth = Target.getPointerWidth(0); 30} 31 32ASTContext &CodeGenFunction::getContext() const { 33 return CGM.getContext(); 34} 35 36 37llvm::BasicBlock *CodeGenFunction::getBasicBlockForLabel(const LabelStmt *S) { 38 llvm::BasicBlock *&BB = LabelMap[S]; 39 if (BB) return BB; 40 41 // Create, but don't insert, the new block. 42 return BB = createBasicBlock(S->getName()); 43} 44 45llvm::Constant * 46CodeGenFunction::GetAddrOfStaticLocalVar(const VarDecl *BVD) { 47 return cast<llvm::Constant>(LocalDeclMap[BVD]); 48} 49 50llvm::Value *CodeGenFunction::GetAddrOfLocalVar(const VarDecl *VD) 51{ 52 return LocalDeclMap[VD]; 53} 54 55const llvm::Type *CodeGenFunction::ConvertType(QualType T) { 56 return CGM.getTypes().ConvertType(T); 57} 58 59bool CodeGenFunction::isObjCPointerType(QualType T) { 60 // All Objective-C types are pointers. 61 return T->isObjCInterfaceType() || 62 T->isObjCQualifiedInterfaceType() || T->isObjCQualifiedIdType(); 63} 64 65bool CodeGenFunction::hasAggregateLLVMType(QualType T) { 66 // FIXME: Use positive checks instead of negative ones to be more 67 // robust in the face of extension. 68 return !isObjCPointerType(T) &&!T->isRealType() && !T->isPointerLikeType() && 69 !T->isVoidType() && !T->isVectorType() && !T->isFunctionType() && 70 !T->isBlockPointerType(); 71} 72 73void CodeGenFunction::EmitReturnBlock() { 74 // For cleanliness, we try to avoid emitting the return block for 75 // simple cases. 76 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 77 78 if (CurBB) { 79 assert(!CurBB->getTerminator() && "Unexpected terminated block."); 80 81 // We have a valid insert point, reuse it if there are no explicit 82 // jumps to the return block. 83 if (ReturnBlock->use_empty()) 84 delete ReturnBlock; 85 else 86 EmitBlock(ReturnBlock); 87 return; 88 } 89 90 // Otherwise, if the return block is the target of a single direct 91 // branch then we can just put the code in that block instead. This 92 // cleans up functions which started with a unified return block. 93 if (ReturnBlock->hasOneUse()) { 94 llvm::BranchInst *BI = 95 dyn_cast<llvm::BranchInst>(*ReturnBlock->use_begin()); 96 if (BI && BI->isUnconditional() && BI->getSuccessor(0) == ReturnBlock) { 97 // Reset insertion point and delete the branch. 98 Builder.SetInsertPoint(BI->getParent()); 99 BI->eraseFromParent(); 100 delete ReturnBlock; 101 return; 102 } 103 } 104 105 // FIXME: We are at an unreachable point, there is no reason to emit 106 // the block unless it has uses. However, we still need a place to 107 // put the debug region.end for now. 108 109 EmitBlock(ReturnBlock); 110} 111 112void CodeGenFunction::FinishFunction(SourceLocation EndLoc) { 113 // Finish emission of indirect switches. 114 EmitIndirectSwitches(); 115 116 assert(BreakContinueStack.empty() && 117 "mismatched push/pop in break/continue stack!"); 118 119 // Emit function epilog (to return). 120 EmitReturnBlock(); 121 122 // Emit debug descriptor for function end. 123 if (CGDebugInfo *DI = CGM.getDebugInfo()) { 124 DI->setLocation(EndLoc); 125 DI->EmitRegionEnd(CurFn, Builder); 126 } 127 128 EmitFunctionEpilog(*CurFnInfo, ReturnValue); 129 130 // Remove the AllocaInsertPt instruction, which is just a convenience for us. 131 AllocaInsertPt->eraseFromParent(); 132 AllocaInsertPt = 0; 133} 134 135void CodeGenFunction::StartFunction(const Decl *D, QualType RetTy, 136 llvm::Function *Fn, 137 const FunctionArgList &Args, 138 SourceLocation StartLoc) { 139 CurFuncDecl = D; 140 FnRetTy = RetTy; 141 CurFn = Fn; 142 assert(CurFn->isDeclaration() && "Function already has body?"); 143 144 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn); 145 146 // Create a marker to make it easy to insert allocas into the entryblock 147 // later. Don't create this with the builder, because we don't want it 148 // folded. 149 llvm::Value *Undef = llvm::UndefValue::get(llvm::Type::Int32Ty); 150 AllocaInsertPt = new llvm::BitCastInst(Undef, llvm::Type::Int32Ty, "allocapt", 151 EntryBB); 152 153 ReturnBlock = createBasicBlock("return"); 154 ReturnValue = 0; 155 if (!RetTy->isVoidType()) 156 ReturnValue = CreateTempAlloca(ConvertType(RetTy), "retval"); 157 158 Builder.SetInsertPoint(EntryBB); 159 160 // Emit subprogram debug descriptor. 161 // FIXME: The cast here is a huge hack. 162 if (CGDebugInfo *DI = CGM.getDebugInfo()) { 163 DI->setLocation(StartLoc); 164 if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(D)) { 165 DI->EmitFunctionStart(FD->getIdentifier()->getName(), 166 RetTy, CurFn, Builder); 167 } else { 168 // Just use LLVM function name. 169 DI->EmitFunctionStart(Fn->getName().c_str(), 170 RetTy, CurFn, Builder); 171 } 172 } 173 174 // FIXME: Leaked. 175 CurFnInfo = new CGFunctionInfo(FnRetTy, Args); 176 EmitFunctionProlog(*CurFnInfo, CurFn, Args); 177 178 // If any of the arguments have a variably modified type, make sure to 179 // emit the type size. 180 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 181 i != e; ++i) { 182 QualType Ty = i->second; 183 184 if (Ty->isVariablyModifiedType()) 185 EmitVLASize(Ty); 186 } 187} 188 189void CodeGenFunction::GenerateCode(const FunctionDecl *FD, 190 llvm::Function *Fn) { 191 FunctionArgList Args; 192 if (FD->getNumParams()) { 193 const FunctionTypeProto* FProto = FD->getType()->getAsFunctionTypeProto(); 194 assert(FProto && "Function def must have prototype!"); 195 196 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) 197 Args.push_back(std::make_pair(FD->getParamDecl(i), 198 FProto->getArgType(i))); 199 } 200 201 StartFunction(FD, FD->getResultType(), Fn, Args, 202 cast<CompoundStmt>(FD->getBody())->getLBracLoc()); 203 204 EmitStmt(FD->getBody()); 205 206 const CompoundStmt *S = dyn_cast<CompoundStmt>(FD->getBody()); 207 if (S) { 208 FinishFunction(S->getRBracLoc()); 209 } else { 210 FinishFunction(); 211 } 212} 213 214/// ContainsLabel - Return true if the statement contains a label in it. If 215/// this statement is not executed normally, it not containing a label means 216/// that we can just remove the code. 217bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) { 218 // Null statement, not a label! 219 if (S == 0) return false; 220 221 // If this is a label, we have to emit the code, consider something like: 222 // if (0) { ... foo: bar(); } goto foo; 223 if (isa<LabelStmt>(S)) 224 return true; 225 226 // If this is a case/default statement, and we haven't seen a switch, we have 227 // to emit the code. 228 if (isa<SwitchCase>(S) && !IgnoreCaseStmts) 229 return true; 230 231 // If this is a switch statement, we want to ignore cases below it. 232 if (isa<SwitchStmt>(S)) 233 IgnoreCaseStmts = true; 234 235 // Scan subexpressions for verboten labels. 236 for (Stmt::const_child_iterator I = S->child_begin(), E = S->child_end(); 237 I != E; ++I) 238 if (ContainsLabel(*I, IgnoreCaseStmts)) 239 return true; 240 241 return false; 242} 243 244 245/// ConstantFoldsToSimpleInteger - If the sepcified expression does not fold to 246/// a constant, or if it does but contains a label, return 0. If it constant 247/// folds to 'true' and does not contain a label, return 1, if it constant folds 248/// to 'false' and does not contain a label, return -1. 249int CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond) { 250 // FIXME: Rename and handle conversion of other evaluatable things 251 // to bool. 252 Expr::EvalResult Result; 253 if (!Cond->Evaluate(Result, getContext()) || !Result.Val.isInt() || 254 Result.HasSideEffects) 255 return 0; // Not foldable, not integer or not fully evaluatable. 256 257 if (CodeGenFunction::ContainsLabel(Cond)) 258 return 0; // Contains a label. 259 260 return Result.Val.getInt().getBoolValue() ? 1 : -1; 261} 262 263 264/// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if 265/// statement) to the specified blocks. Based on the condition, this might try 266/// to simplify the codegen of the conditional based on the branch. 267/// 268void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond, 269 llvm::BasicBlock *TrueBlock, 270 llvm::BasicBlock *FalseBlock) { 271 if (const ParenExpr *PE = dyn_cast<ParenExpr>(Cond)) 272 return EmitBranchOnBoolExpr(PE->getSubExpr(), TrueBlock, FalseBlock); 273 274 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) { 275 // Handle X && Y in a condition. 276 if (CondBOp->getOpcode() == BinaryOperator::LAnd) { 277 // If we have "1 && X", simplify the code. "0 && X" would have constant 278 // folded if the case was simple enough. 279 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS()) == 1) { 280 // br(1 && X) -> br(X). 281 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock); 282 } 283 284 // If we have "X && 1", simplify the code to use an uncond branch. 285 // "X && 0" would have been constant folded to 0. 286 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS()) == 1) { 287 // br(X && 1) -> br(X). 288 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock); 289 } 290 291 // Emit the LHS as a conditional. If the LHS conditional is false, we 292 // want to jump to the FalseBlock. 293 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true"); 294 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock); 295 EmitBlock(LHSTrue); 296 297 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock); 298 return; 299 } else if (CondBOp->getOpcode() == BinaryOperator::LOr) { 300 // If we have "0 || X", simplify the code. "1 || X" would have constant 301 // folded if the case was simple enough. 302 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS()) == -1) { 303 // br(0 || X) -> br(X). 304 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock); 305 } 306 307 // If we have "X || 0", simplify the code to use an uncond branch. 308 // "X || 1" would have been constant folded to 1. 309 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS()) == -1) { 310 // br(X || 0) -> br(X). 311 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock); 312 } 313 314 // Emit the LHS as a conditional. If the LHS conditional is true, we 315 // want to jump to the TrueBlock. 316 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false"); 317 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse); 318 EmitBlock(LHSFalse); 319 320 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock); 321 return; 322 } 323 } 324 325 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) { 326 // br(!x, t, f) -> br(x, f, t) 327 if (CondUOp->getOpcode() == UnaryOperator::LNot) 328 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock); 329 } 330 331 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) { 332 // Handle ?: operator. 333 334 // Just ignore GNU ?: extension. 335 if (CondOp->getLHS()) { 336 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f)) 337 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true"); 338 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false"); 339 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock); 340 EmitBlock(LHSBlock); 341 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock); 342 EmitBlock(RHSBlock); 343 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock); 344 return; 345 } 346 } 347 348 // Emit the code with the fully general case. 349 llvm::Value *CondV = EvaluateExprAsBool(Cond); 350 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock); 351} 352 353/// getCGRecordLayout - Return record layout info. 354const CGRecordLayout *CodeGenFunction::getCGRecordLayout(CodeGenTypes &CGT, 355 QualType Ty) { 356 const RecordType *RTy = Ty->getAsRecordType(); 357 assert (RTy && "Unexpected type. RecordType expected here."); 358 359 return CGT.getCGRecordLayout(RTy->getDecl()); 360} 361 362/// ErrorUnsupported - Print out an error that codegen doesn't support the 363/// specified stmt yet. 364void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type, 365 bool OmitOnError) { 366 CGM.ErrorUnsupported(S, Type, OmitOnError); 367} 368 369unsigned CodeGenFunction::GetIDForAddrOfLabel(const LabelStmt *L) { 370 // Use LabelIDs.size() as the new ID if one hasn't been assigned. 371 return LabelIDs.insert(std::make_pair(L, LabelIDs.size())).first->second; 372} 373 374void CodeGenFunction::EmitMemSetToZero(llvm::Value *DestPtr, QualType Ty) 375{ 376 const llvm::Type *BP = llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 377 if (DestPtr->getType() != BP) 378 DestPtr = Builder.CreateBitCast(DestPtr, BP, "tmp"); 379 380 // Get size and alignment info for this aggregate. 381 std::pair<uint64_t, unsigned> TypeInfo = getContext().getTypeInfo(Ty); 382 383 // FIXME: Handle variable sized types. 384 const llvm::Type *IntPtr = llvm::IntegerType::get(LLVMPointerWidth); 385 386 Builder.CreateCall4(CGM.getMemSetFn(), DestPtr, 387 llvm::ConstantInt::getNullValue(llvm::Type::Int8Ty), 388 // TypeInfo.first describes size in bits. 389 llvm::ConstantInt::get(IntPtr, TypeInfo.first/8), 390 llvm::ConstantInt::get(llvm::Type::Int32Ty, 391 TypeInfo.second/8)); 392} 393 394void CodeGenFunction::EmitIndirectSwitches() { 395 llvm::BasicBlock *Default; 396 397 if (IndirectSwitches.empty()) 398 return; 399 400 if (!LabelIDs.empty()) { 401 Default = getBasicBlockForLabel(LabelIDs.begin()->first); 402 } else { 403 // No possible targets for indirect goto, just emit an infinite 404 // loop. 405 Default = createBasicBlock("indirectgoto.loop", CurFn); 406 llvm::BranchInst::Create(Default, Default); 407 } 408 409 for (std::vector<llvm::SwitchInst*>::iterator i = IndirectSwitches.begin(), 410 e = IndirectSwitches.end(); i != e; ++i) { 411 llvm::SwitchInst *I = *i; 412 413 I->setSuccessor(0, Default); 414 for (std::map<const LabelStmt*,unsigned>::iterator LI = LabelIDs.begin(), 415 LE = LabelIDs.end(); LI != LE; ++LI) { 416 I->addCase(llvm::ConstantInt::get(llvm::Type::Int32Ty, 417 LI->second), 418 getBasicBlockForLabel(LI->first)); 419 } 420 } 421} 422 423llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) 424{ 425 // FIXME: This entire method is hardcoded for 32-bit X86. 426 427 const char *TargetPrefix = getContext().Target.getTargetPrefix(); 428 429 if (strcmp(TargetPrefix, "x86") != 0 || 430 getContext().Target.getPointerWidth(0) != 32) 431 return 0; 432 433 const llvm::Type *BP = llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 434 const llvm::Type *BPP = llvm::PointerType::getUnqual(BP); 435 436 llvm::Value *VAListAddrAsBPP = Builder.CreateBitCast(VAListAddr, BPP, 437 "ap"); 438 llvm::Value *Addr = Builder.CreateLoad(VAListAddrAsBPP, "ap.cur"); 439 llvm::Value *AddrTyped = 440 Builder.CreateBitCast(Addr, 441 llvm::PointerType::getUnqual(ConvertType(Ty))); 442 443 uint64_t SizeInBytes = getContext().getTypeSize(Ty) / 8; 444 const unsigned ArgumentSizeInBytes = 4; 445 if (SizeInBytes < ArgumentSizeInBytes) 446 SizeInBytes = ArgumentSizeInBytes; 447 448 llvm::Value *NextAddr = 449 Builder.CreateGEP(Addr, 450 llvm::ConstantInt::get(llvm::Type::Int32Ty, SizeInBytes), 451 "ap.next"); 452 Builder.CreateStore(NextAddr, VAListAddrAsBPP); 453 454 return AddrTyped; 455} 456 457 458llvm::Value *CodeGenFunction::GetVLASize(const VariableArrayType *VAT) 459{ 460 llvm::Value *&SizeEntry = VLASizeMap[VAT]; 461 462 assert(SizeEntry && "Did not emit size for type"); 463 return SizeEntry; 464} 465 466llvm::Value *CodeGenFunction::EmitVLASize(QualType Ty) 467{ 468 assert(Ty->isVariablyModifiedType() && 469 "Must pass variably modified type to EmitVLASizes!"); 470 471 if (const VariableArrayType *VAT = getContext().getAsVariableArrayType(Ty)) { 472 llvm::Value *&SizeEntry = VLASizeMap[VAT]; 473 474 if (!SizeEntry) { 475 // Get the element size; 476 llvm::Value *ElemSize; 477 478 QualType ElemTy = VAT->getElementType(); 479 480 if (ElemTy->isVariableArrayType()) 481 ElemSize = EmitVLASize(ElemTy); 482 else { 483 // FIXME: We use Int32Ty here because the alloca instruction takes a 484 // 32-bit integer. What should we do about overflow? 485 ElemSize = llvm::ConstantInt::get(llvm::Type::Int32Ty, 486 getContext().getTypeSize(ElemTy) / 8); 487 } 488 489 llvm::Value *NumElements = EmitScalarExpr(VAT->getSizeExpr()); 490 491 SizeEntry = Builder.CreateMul(ElemSize, NumElements); 492 } 493 494 return SizeEntry; 495 } else if (const PointerType *PT = Ty->getAsPointerType()) 496 EmitVLASize(PT->getPointeeType()); 497 else { 498 assert(0 && "unknown VM type!"); 499 } 500 501 return 0; 502} 503 504llvm::Value* CodeGenFunction::EmitVAListRef(const Expr* E) { 505 if (CGM.getContext().getBuiltinVaListType()->isArrayType()) { 506 return EmitScalarExpr(E); 507 } 508 return EmitLValue(E).getAddress(); 509} 510