1 //===-- Writer.cpp - Library for converting LLVM code to C ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This library converts LLVM code to C code, compilable by GCC and other C
13 //===----------------------------------------------------------------------===//
15 #include "CTargetMachine.h"
16 #include "llvm/Target/TargetMachineImpls.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Module.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/Pass.h"
22 #include "llvm/PassManager.h"
23 #include "llvm/SymbolTable.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/IntrinsicLowering.h"
26 #include "llvm/Analysis/FindUsedTypes.h"
27 #include "llvm/Analysis/ConstantsScanner.h"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Support/CFG.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/InstVisitor.h"
33 #include "llvm/Support/Mangler.h"
34 #include "Support/StringExtras.h"
35 #include "Config/config.h"
41 /// NameAllUsedStructs - This pass inserts names for any unnamed structure
42 /// types that are used by the program.
44 class CBackendNameAllUsedStructs : public Pass {
45 void getAnalysisUsage(AnalysisUsage &AU) const {
46 AU.addRequired<FindUsedTypes>();
49 virtual const char *getPassName() const {
50 return "C backend type canonicalizer";
53 virtual bool run(Module &M);
56 /// CWriter - This class is the main chunk of code that converts an LLVM
57 /// module to a C translation unit.
58 class CWriter : public FunctionPass, public InstVisitor<CWriter> {
60 IntrinsicLowering &IL;
62 const Module *TheModule;
63 std::map<const Type *, std::string> TypeNames;
65 std::map<const ConstantFP *, unsigned> FPConstantMap;
67 CWriter(std::ostream &o, IntrinsicLowering &il) : Out(o), IL(il) {}
69 virtual const char *getPassName() const { return "C backend"; }
71 virtual bool doInitialization(Module &M);
73 bool runOnFunction(Function &F) {
74 // Output all floating point constants that cannot be printed accurately.
75 printFloatingPointConstants(F);
79 FPConstantMap.clear();
83 virtual bool doFinalization(Module &M) {
90 std::ostream &printType(std::ostream &Out, const Type *Ty,
91 const std::string &VariableName = "",
92 bool IgnoreName = false);
94 void writeOperand(Value *Operand);
95 void writeOperandInternal(Value *Operand);
98 void lowerIntrinsics(Function &F);
100 bool nameAllUsedStructureTypes(Module &M);
101 void printModule(Module *M);
102 void printModuleTypes(const SymbolTable &ST);
103 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
104 void printFloatingPointConstants(Function &F);
105 void printFunctionSignature(const Function *F, bool Prototype);
107 void printFunction(Function &);
109 void printConstant(Constant *CPV);
110 void printConstantArray(ConstantArray *CPA);
112 // isInlinableInst - Attempt to inline instructions into their uses to build
113 // trees as much as possible. To do this, we have to consistently decide
114 // what is acceptable to inline, so that variable declarations don't get
115 // printed and an extra copy of the expr is not emitted.
117 static bool isInlinableInst(const Instruction &I) {
118 // Must be an expression, must be used exactly once. If it is dead, we
119 // emit it inline where it would go.
120 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
121 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
122 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
123 // Don't inline a load across a store or other bad things!
126 // Only inline instruction it it's use is in the same BB as the inst.
127 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
130 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
131 // variables which are accessed with the & operator. This causes GCC to
132 // generate significantly better code than to emit alloca calls directly.
134 static const AllocaInst *isDirectAlloca(const Value *V) {
135 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
136 if (!AI) return false;
137 if (AI->isArrayAllocation())
138 return 0; // FIXME: we can also inline fixed size array allocas!
139 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
144 // Instruction visitation functions
145 friend class InstVisitor<CWriter>;
147 void visitReturnInst(ReturnInst &I);
148 void visitBranchInst(BranchInst &I);
149 void visitSwitchInst(SwitchInst &I);
150 void visitInvokeInst(InvokeInst &I);
151 void visitUnwindInst(UnwindInst &I);
153 void visitPHINode(PHINode &I);
154 void visitBinaryOperator(Instruction &I);
156 void visitCastInst (CastInst &I);
157 void visitSelectInst(SelectInst &I);
158 void visitCallInst (CallInst &I);
159 void visitCallSite (CallSite CS);
160 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
162 void visitMallocInst(MallocInst &I);
163 void visitAllocaInst(AllocaInst &I);
164 void visitFreeInst (FreeInst &I);
165 void visitLoadInst (LoadInst &I);
166 void visitStoreInst (StoreInst &I);
167 void visitGetElementPtrInst(GetElementPtrInst &I);
168 void visitVANextInst(VANextInst &I);
169 void visitVAArgInst (VAArgInst &I);
171 void visitInstruction(Instruction &I) {
172 std::cerr << "C Writer does not know about " << I;
176 void outputLValue(Instruction *I) {
177 Out << " " << Mang->getValueName(I) << " = ";
179 void printPHICopiesForSuccessors(BasicBlock *CurBlock,
181 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
183 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
184 gep_type_iterator E);
188 /// This method inserts names for any unnamed structure types that are used by
189 /// the program, and removes names from structure types that are not used by the
192 bool CBackendNameAllUsedStructs::run(Module &M) {
193 // Get a set of types that are used by the program...
194 std::set<const Type *> UT = getAnalysis<FindUsedTypes>().getTypes();
196 // Loop over the module symbol table, removing types from UT that are
197 // already named, and removing names for structure types that are not used.
199 SymbolTable &MST = M.getSymbolTable();
200 if (MST.find(Type::TypeTy) != MST.end())
201 for (SymbolTable::type_iterator I = MST.type_begin(Type::TypeTy),
202 E = MST.type_end(Type::TypeTy); I != E; ) {
203 SymbolTable::type_iterator It = I++;
204 if (StructType *STy = dyn_cast<StructType>(It->second)) {
205 // If this is not used, remove it from the symbol table.
206 std::set<const Type *>::iterator UTI = UT.find(STy);
208 MST.remove(It->first, It->second);
214 // UT now contains types that are not named. Loop over it, naming
217 bool Changed = false;
218 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
220 if (const StructType *ST = dyn_cast<StructType>(*I)) {
221 ((Value*)ST)->setName("unnamed", &MST);
228 // Pass the Type* and the variable name and this prints out the variable
231 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
232 const std::string &NameSoFar,
234 if (Ty->isPrimitiveType())
235 switch (Ty->getPrimitiveID()) {
236 case Type::VoidTyID: return Out << "void " << NameSoFar;
237 case Type::BoolTyID: return Out << "bool " << NameSoFar;
238 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
239 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
240 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
241 case Type::ShortTyID: return Out << "short " << NameSoFar;
242 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
243 case Type::IntTyID: return Out << "int " << NameSoFar;
244 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
245 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
246 case Type::FloatTyID: return Out << "float " << NameSoFar;
247 case Type::DoubleTyID: return Out << "double " << NameSoFar;
249 std::cerr << "Unknown primitive type: " << Ty << "\n";
253 // Check to see if the type is named.
254 if (!IgnoreName || isa<OpaqueType>(Ty)) {
255 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
256 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
259 switch (Ty->getPrimitiveID()) {
260 case Type::FunctionTyID: {
261 const FunctionType *MTy = cast<FunctionType>(Ty);
262 std::stringstream FunctionInnards;
263 FunctionInnards << " (" << NameSoFar << ") (";
264 for (FunctionType::param_iterator I = MTy->param_begin(),
265 E = MTy->param_end(); I != E; ++I) {
266 if (I != MTy->param_begin())
267 FunctionInnards << ", ";
268 printType(FunctionInnards, *I, "");
270 if (MTy->isVarArg()) {
271 if (MTy->getNumParams())
272 FunctionInnards << ", ...";
273 } else if (!MTy->getNumParams()) {
274 FunctionInnards << "void";
276 FunctionInnards << ")";
277 std::string tstr = FunctionInnards.str();
278 printType(Out, MTy->getReturnType(), tstr);
281 case Type::StructTyID: {
282 const StructType *STy = cast<StructType>(Ty);
283 Out << NameSoFar + " {\n";
285 for (StructType::element_iterator I = STy->element_begin(),
286 E = STy->element_end(); I != E; ++I) {
288 printType(Out, *I, "field" + utostr(Idx++));
294 case Type::PointerTyID: {
295 const PointerType *PTy = cast<PointerType>(Ty);
296 std::string ptrName = "*" + NameSoFar;
298 if (isa<ArrayType>(PTy->getElementType()))
299 ptrName = "(" + ptrName + ")";
301 return printType(Out, PTy->getElementType(), ptrName);
304 case Type::ArrayTyID: {
305 const ArrayType *ATy = cast<ArrayType>(Ty);
306 unsigned NumElements = ATy->getNumElements();
307 return printType(Out, ATy->getElementType(),
308 NameSoFar + "[" + utostr(NumElements) + "]");
311 case Type::OpaqueTyID: {
312 static int Count = 0;
313 std::string TyName = "struct opaque_" + itostr(Count++);
314 assert(TypeNames.find(Ty) == TypeNames.end());
315 TypeNames[Ty] = TyName;
316 return Out << TyName << " " << NameSoFar;
319 assert(0 && "Unhandled case in getTypeProps!");
326 void CWriter::printConstantArray(ConstantArray *CPA) {
328 // As a special case, print the array as a string if it is an array of
329 // ubytes or an array of sbytes with positive values.
331 const Type *ETy = CPA->getType()->getElementType();
332 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
334 // Make sure the last character is a null char, as automatically added by C
335 if (isString && (CPA->getNumOperands() == 0 ||
336 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
341 // Keep track of whether the last number was a hexadecimal escape
342 bool LastWasHex = false;
344 // Do not include the last character, which we know is null
345 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
346 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
348 // Print it out literally if it is a printable character. The only thing
349 // to be careful about is when the last letter output was a hex escape
350 // code, in which case we have to be careful not to print out hex digits
351 // explicitly (the C compiler thinks it is a continuation of the previous
352 // character, sheesh...)
354 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
356 if (C == '"' || C == '\\')
363 case '\n': Out << "\\n"; break;
364 case '\t': Out << "\\t"; break;
365 case '\r': Out << "\\r"; break;
366 case '\v': Out << "\\v"; break;
367 case '\a': Out << "\\a"; break;
368 case '\"': Out << "\\\""; break;
369 case '\'': Out << "\\\'"; break;
372 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
373 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
382 if (CPA->getNumOperands()) {
384 printConstant(cast<Constant>(CPA->getOperand(0)));
385 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
387 printConstant(cast<Constant>(CPA->getOperand(i)));
394 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
395 // textually as a double (rather than as a reference to a stack-allocated
396 // variable). We decide this by converting CFP to a string and back into a
397 // double, and then checking whether the conversion results in a bit-equal
398 // double to the original value of CFP. This depends on us and the target C
399 // compiler agreeing on the conversion process (which is pretty likely since we
400 // only deal in IEEE FP).
402 bool isFPCSafeToPrint(const ConstantFP *CFP) {
405 sprintf(Buffer, "%a", CFP->getValue());
407 if (!strncmp(Buffer, "0x", 2) ||
408 !strncmp(Buffer, "-0x", 3) ||
409 !strncmp(Buffer, "+0x", 3))
410 return atof(Buffer) == CFP->getValue();
413 std::string StrVal = ftostr(CFP->getValue());
415 while (StrVal[0] == ' ')
416 StrVal.erase(StrVal.begin());
418 // Check to make sure that the stringized number is not some string like "Inf"
419 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
420 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
421 ((StrVal[0] == '-' || StrVal[0] == '+') &&
422 (StrVal[1] >= '0' && StrVal[1] <= '9')))
423 // Reparse stringized version!
424 return atof(StrVal.c_str()) == CFP->getValue();
429 // printConstant - The LLVM Constant to C Constant converter.
430 void CWriter::printConstant(Constant *CPV) {
431 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
432 switch (CE->getOpcode()) {
433 case Instruction::Cast:
435 printType(Out, CPV->getType());
437 printConstant(CE->getOperand(0));
441 case Instruction::GetElementPtr:
443 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
447 case Instruction::Select:
449 printConstant(CE->getOperand(0));
451 printConstant(CE->getOperand(1));
453 printConstant(CE->getOperand(2));
456 case Instruction::Add:
457 case Instruction::Sub:
458 case Instruction::Mul:
459 case Instruction::Div:
460 case Instruction::Rem:
461 case Instruction::SetEQ:
462 case Instruction::SetNE:
463 case Instruction::SetLT:
464 case Instruction::SetLE:
465 case Instruction::SetGT:
466 case Instruction::SetGE:
467 case Instruction::Shl:
468 case Instruction::Shr:
470 printConstant(CE->getOperand(0));
471 switch (CE->getOpcode()) {
472 case Instruction::Add: Out << " + "; break;
473 case Instruction::Sub: Out << " - "; break;
474 case Instruction::Mul: Out << " * "; break;
475 case Instruction::Div: Out << " / "; break;
476 case Instruction::Rem: Out << " % "; break;
477 case Instruction::SetEQ: Out << " == "; break;
478 case Instruction::SetNE: Out << " != "; break;
479 case Instruction::SetLT: Out << " < "; break;
480 case Instruction::SetLE: Out << " <= "; break;
481 case Instruction::SetGT: Out << " > "; break;
482 case Instruction::SetGE: Out << " >= "; break;
483 case Instruction::Shl: Out << " << "; break;
484 case Instruction::Shr: Out << " >> "; break;
485 default: assert(0 && "Illegal opcode here!");
487 printConstant(CE->getOperand(1));
492 std::cerr << "CWriter Error: Unhandled constant expression: "
498 switch (CPV->getType()->getPrimitiveID()) {
500 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
501 case Type::SByteTyID:
502 case Type::ShortTyID:
503 Out << cast<ConstantSInt>(CPV)->getValue(); break;
505 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
506 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
508 Out << cast<ConstantSInt>(CPV)->getValue();
512 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
514 case Type::UByteTyID:
515 case Type::UShortTyID:
516 Out << cast<ConstantUInt>(CPV)->getValue(); break;
518 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
519 case Type::ULongTyID:
520 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
522 case Type::FloatTyID:
523 case Type::DoubleTyID: {
524 ConstantFP *FPC = cast<ConstantFP>(CPV);
525 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
526 if (I != FPConstantMap.end()) {
527 // Because of FP precision problems we must load from a stack allocated
528 // value that holds the value in hex.
529 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
530 << "*)&FPConstant" << I->second << ")";
533 // Print out the constant as a floating point number.
535 sprintf(Buffer, "%a", FPC->getValue());
536 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
538 Out << ftostr(FPC->getValue());
544 case Type::ArrayTyID:
545 if (isa<ConstantAggregateZero>(CPV)) {
546 const ArrayType *AT = cast<ArrayType>(CPV->getType());
548 if (AT->getNumElements()) {
550 Constant *CZ = Constant::getNullValue(AT->getElementType());
552 for (unsigned i = 1, e = AT->getNumElements(); i != e; ++i) {
559 printConstantArray(cast<ConstantArray>(CPV));
563 case Type::StructTyID:
564 if (isa<ConstantAggregateZero>(CPV)) {
565 const StructType *ST = cast<StructType>(CPV->getType());
567 if (ST->getNumElements()) {
569 printConstant(Constant::getNullValue(ST->getElementType(0)));
570 for (unsigned i = 1, e = ST->getNumElements(); i != e; ++i) {
572 printConstant(Constant::getNullValue(ST->getElementType(i)));
578 if (CPV->getNumOperands()) {
580 printConstant(cast<Constant>(CPV->getOperand(0)));
581 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
583 printConstant(cast<Constant>(CPV->getOperand(i)));
590 case Type::PointerTyID:
591 if (isa<ConstantPointerNull>(CPV)) {
593 printType(Out, CPV->getType());
594 Out << ")/*NULL*/0)";
596 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
597 writeOperand(CPR->getValue());
602 std::cerr << "Unknown constant type: " << CPV << "\n";
607 void CWriter::writeOperandInternal(Value *Operand) {
608 if (Instruction *I = dyn_cast<Instruction>(Operand))
609 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
610 // Should we inline this instruction to build a tree?
617 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
620 Out << Mang->getValueName(Operand);
624 void CWriter::writeOperand(Value *Operand) {
625 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
626 Out << "(&"; // Global variables are references as their addresses by llvm
628 writeOperandInternal(Operand);
630 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
634 // generateCompilerSpecificCode - This is where we add conditional compilation
635 // directives to cater to specific compilers as need be.
637 static void generateCompilerSpecificCode(std::ostream& Out) {
638 // Alloca is hard to get, and we don't want to include stdlib.h here...
639 Out << "/* get a declaration for alloca */\n"
641 << "extern void *__builtin_alloca(unsigned long);\n"
642 << "#define alloca(x) __builtin_alloca(x)\n"
644 << "#ifndef __FreeBSD__\n"
645 << "#include <alloca.h>\n"
649 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
650 // If we aren't being compiled with GCC, just drop these attributes.
651 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
652 << "#define __attribute__(X)\n"
656 // At some point, we should support "external weak" vs. "weak" linkages.
657 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
658 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
659 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
660 << "#elif defined(__GNUC__)\n"
661 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
663 << "#define __EXTERNAL_WEAK__\n"
667 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
668 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
669 << "#define __ATTRIBUTE_WEAK__\n"
670 << "#elif defined(__GNUC__)\n"
671 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
673 << "#define __ATTRIBUTE_WEAK__\n"
677 bool CWriter::doInitialization(Module &M) {
683 // Ensure that all structure types have names...
684 Mang = new Mangler(M);
686 // get declaration for alloca
687 Out << "/* Provide Declarations */\n";
688 Out << "#include <stdarg.h>\n"; // Varargs support
689 Out << "#include <setjmp.h>\n"; // Unwind support
690 generateCompilerSpecificCode(Out);
692 // Provide a definition for `bool' if not compiling with a C++ compiler.
694 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
696 << "\n\n/* Support for floating point constants */\n"
697 << "typedef unsigned long long ConstantDoubleTy;\n"
698 << "typedef unsigned int ConstantFloatTy;\n"
700 << "\n\n/* Global Declarations */\n";
702 // First output all the declarations for the program, because C requires
703 // Functions & globals to be declared before they are used.
706 // Loop over the symbol table, emitting all named constants...
707 printModuleTypes(M.getSymbolTable());
709 // Global variable declarations...
711 Out << "\n/* External Global Variable Declarations */\n";
712 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I) {
713 if (I->hasExternalLinkage()) {
715 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
721 // Function declarations
723 Out << "\n/* Function Declarations */\n";
724 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
725 // Don't print declarations for intrinsic functions.
726 if (!I->getIntrinsicID() &&
727 I->getName() != "setjmp" && I->getName() != "longjmp") {
728 printFunctionSignature(I, true);
729 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
730 if (I->hasLinkOnceLinkage()) Out << " __ATTRIBUTE_WEAK__";
736 // Output the global variable declarations
738 Out << "\n\n/* Global Variable Declarations */\n";
739 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
740 if (!I->isExternal()) {
742 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
744 if (I->hasLinkOnceLinkage())
745 Out << " __attribute__((common))";
746 else if (I->hasWeakLinkage())
747 Out << " __ATTRIBUTE_WEAK__";
752 // Output the global variable definitions and contents...
754 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
755 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
756 if (!I->isExternal()) {
757 if (I->hasInternalLinkage())
759 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
760 if (I->hasLinkOnceLinkage())
761 Out << " __attribute__((common))";
762 else if (I->hasWeakLinkage())
763 Out << " __ATTRIBUTE_WEAK__";
765 // If the initializer is not null, emit the initializer. If it is null,
766 // we try to avoid emitting large amounts of zeros. The problem with
767 // this, however, occurs when the variable has weak linkage. In this
768 // case, the assembler will complain about the variable being both weak
769 // and common, so we disable this optimization.
770 if (!I->getInitializer()->isNullValue()) {
772 writeOperand(I->getInitializer());
773 } else if (I->hasWeakLinkage()) {
774 // We have to specify an initializer, but it doesn't have to be
775 // complete. If the value is an aggregate, print out { 0 }, and let
776 // the compiler figure out the rest of the zeros.
778 if (isa<StructType>(I->getInitializer()->getType()) ||
779 isa<ArrayType>(I->getInitializer()->getType())) {
782 // Just print it out normally.
783 writeOperand(I->getInitializer());
791 Out << "\n\n/* Function Bodies */\n";
796 /// Output all floating point constants that cannot be printed accurately...
797 void CWriter::printFloatingPointConstants(Function &F) {
808 // Scan the module for floating point constants. If any FP constant is used
809 // in the function, we want to redirect it here so that we do not depend on
810 // the precision of the printed form, unless the printed form preserves
813 static unsigned FPCounter = 0;
814 for (constant_iterator I = constant_begin(&F), E = constant_end(&F);
816 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
817 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
818 !FPConstantMap.count(FPC)) {
819 double Val = FPC->getValue();
821 FPConstantMap[FPC] = FPCounter; // Number the FP constants
823 if (FPC->getType() == Type::DoubleTy) {
825 Out << "static const ConstantDoubleTy FPConstant" << FPCounter++
826 << " = 0x" << std::hex << DBLUnion.U << std::dec
827 << "ULL; /* " << Val << " */\n";
828 } else if (FPC->getType() == Type::FloatTy) {
830 Out << "static const ConstantFloatTy FPConstant" << FPCounter++
831 << " = 0x" << std::hex << FLTUnion.U << std::dec
832 << "U; /* " << Val << " */\n";
834 assert(0 && "Unknown float type!");
841 /// printSymbolTable - Run through symbol table looking for type names. If a
842 /// type name is found, emit it's declaration...
844 void CWriter::printModuleTypes(const SymbolTable &ST) {
845 // If there are no type names, exit early.
846 if (ST.find(Type::TypeTy) == ST.end())
849 // We are only interested in the type plane of the symbol table...
850 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
851 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
853 // Print out forward declarations for structure types before anything else!
854 Out << "/* Structure forward decls */\n";
855 for (; I != End; ++I)
856 if (const Type *STy = dyn_cast<StructType>(I->second)) {
857 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
858 Out << Name << ";\n";
859 TypeNames.insert(std::make_pair(STy, Name));
864 // Now we can print out typedefs...
865 Out << "/* Typedefs */\n";
866 for (I = ST.type_begin(Type::TypeTy); I != End; ++I) {
867 const Type *Ty = cast<Type>(I->second);
868 std::string Name = "l_" + Mangler::makeNameProper(I->first);
870 printType(Out, Ty, Name);
876 // Keep track of which structures have been printed so far...
877 std::set<const StructType *> StructPrinted;
879 // Loop over all structures then push them into the stack so they are
880 // printed in the correct order.
882 Out << "/* Structure contents */\n";
883 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
884 if (const StructType *STy = dyn_cast<StructType>(I->second))
885 // Only print out used types!
886 printContainedStructs(STy, StructPrinted);
889 // Push the struct onto the stack and recursively push all structs
890 // this one depends on.
891 void CWriter::printContainedStructs(const Type *Ty,
892 std::set<const StructType*> &StructPrinted){
893 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
894 //Check to see if we have already printed this struct
895 if (StructPrinted.count(STy) == 0) {
896 // Print all contained types first...
897 for (StructType::element_iterator I = STy->element_begin(),
898 E = STy->element_end(); I != E; ++I) {
899 const Type *Ty1 = I->get();
900 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
901 printContainedStructs(*I, StructPrinted);
904 //Print structure type out..
905 StructPrinted.insert(STy);
906 std::string Name = TypeNames[STy];
907 printType(Out, STy, Name, true);
911 // If it is an array, check contained types and continue
912 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
913 const Type *Ty1 = ATy->getElementType();
914 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
915 printContainedStructs(Ty1, StructPrinted);
920 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
921 if (F->hasInternalLinkage()) Out << "static ";
923 // Loop over the arguments, printing them...
924 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
926 std::stringstream FunctionInnards;
928 // Print out the name...
929 FunctionInnards << Mang->getValueName(F) << "(";
931 if (!F->isExternal()) {
934 if (F->abegin()->hasName() || !Prototype)
935 ArgName = Mang->getValueName(F->abegin());
936 printType(FunctionInnards, F->afront().getType(), ArgName);
937 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
939 FunctionInnards << ", ";
940 if (I->hasName() || !Prototype)
941 ArgName = Mang->getValueName(I);
944 printType(FunctionInnards, I->getType(), ArgName);
948 // Loop over the arguments, printing them...
949 for (FunctionType::param_iterator I = FT->param_begin(),
950 E = FT->param_end(); I != E; ++I) {
951 if (I != FT->param_begin()) FunctionInnards << ", ";
952 printType(FunctionInnards, *I);
956 // Finish printing arguments... if this is a vararg function, print the ...,
957 // unless there are no known types, in which case, we just emit ().
959 if (FT->isVarArg() && FT->getNumParams()) {
960 if (FT->getNumParams()) FunctionInnards << ", ";
961 FunctionInnards << "..."; // Output varargs portion of signature!
962 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
963 FunctionInnards << "void"; // ret() -> ret(void) in C.
965 FunctionInnards << ")";
966 // Print out the return type and the entire signature for that matter
967 printType(Out, F->getReturnType(), FunctionInnards.str());
970 void CWriter::printFunction(Function &F) {
971 printFunctionSignature(&F, false);
974 // print local variable information for the function
975 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I)
976 if (const AllocaInst *AI = isDirectAlloca(&*I)) {
978 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
979 Out << "; /* Address exposed local */\n";
980 } else if (I->getType() != Type::VoidTy && !isInlinableInst(*I)) {
982 printType(Out, I->getType(), Mang->getValueName(&*I));
985 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
987 printType(Out, I->getType(),
988 Mang->getValueName(&*I)+"__PHI_TEMPORARY");
995 // print the basic blocks
996 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
997 BasicBlock *Prev = BB->getPrev();
999 // Don't print the label for the basic block if there are no uses, or if the
1000 // only terminator use is the predecessor basic block's terminator. We have
1001 // to scan the use list because PHI nodes use basic blocks too but do not
1002 // require a label to be generated.
1004 bool NeedsLabel = false;
1005 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
1007 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
1008 if (TI != Prev->getTerminator() ||
1009 isa<SwitchInst>(Prev->getTerminator()) ||
1010 isa<InvokeInst>(Prev->getTerminator())) {
1015 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1017 // Output all of the instructions in the basic block...
1018 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
1019 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1020 if (II->getType() != Type::VoidTy)
1029 // Don't emit prefix or suffix for the terminator...
1030 visit(*BB->getTerminator());
1036 // Specific Instruction type classes... note that all of the casts are
1037 // necessary because we use the instruction classes as opaque types...
1039 void CWriter::visitReturnInst(ReturnInst &I) {
1040 // Don't output a void return if this is the last basic block in the function
1041 if (I.getNumOperands() == 0 &&
1042 &*--I.getParent()->getParent()->end() == I.getParent() &&
1043 !I.getParent()->size() == 1) {
1048 if (I.getNumOperands()) {
1050 writeOperand(I.getOperand(0));
1055 void CWriter::visitSwitchInst(SwitchInst &SI) {
1056 printPHICopiesForSuccessors(SI.getParent(), 0);
1059 writeOperand(SI.getOperand(0));
1060 Out << ") {\n default:\n";
1061 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1063 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1065 writeOperand(SI.getOperand(i));
1067 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1068 printBranchToBlock(SI.getParent(), Succ, 2);
1069 if (Succ == SI.getParent()->getNext())
1075 void CWriter::visitInvokeInst(InvokeInst &II) {
1076 assert(0 && "Lowerinvoke pass didn't work!");
1080 void CWriter::visitUnwindInst(UnwindInst &I) {
1081 assert(0 && "Lowerinvoke pass didn't work!");
1084 static bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1085 // If PHI nodes need copies, we need the copy code...
1086 if (isa<PHINode>(To->front()) ||
1087 From->getNext() != To) // Not directly successor, need goto
1090 // Otherwise we don't need the code.
1094 void CWriter::printPHICopiesForSuccessors(BasicBlock *CurBlock,
1096 for (succ_iterator SI = succ_begin(CurBlock), E = succ_end(CurBlock);
1098 for (BasicBlock::iterator I = SI->begin();
1099 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1100 // now we have to do the printing
1101 Out << std::string(Indent, ' ');
1102 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1103 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBlock)));
1104 Out << "; /* for PHI node */\n";
1109 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1111 if (CurBB->getNext() != Succ ||
1112 isa<InvokeInst>(CurBB->getTerminator()) ||
1113 isa<SwitchInst>(CurBB->getTerminator())) {
1114 Out << std::string(Indent, ' ') << " goto ";
1120 // Branch instruction printing - Avoid printing out a branch to a basic block
1121 // that immediately succeeds the current one.
1123 void CWriter::visitBranchInst(BranchInst &I) {
1124 printPHICopiesForSuccessors(I.getParent(), 0);
1126 if (I.isConditional()) {
1127 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1129 writeOperand(I.getCondition());
1132 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1134 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1135 Out << " } else {\n";
1136 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1139 // First goto not necessary, assume second one is...
1141 writeOperand(I.getCondition());
1144 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1149 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1154 // PHI nodes get copied into temporary values at the end of predecessor basic
1155 // blocks. We now need to copy these temporary values into the REAL value for
1157 void CWriter::visitPHINode(PHINode &I) {
1159 Out << "__PHI_TEMPORARY";
1163 void CWriter::visitBinaryOperator(Instruction &I) {
1164 // binary instructions, shift instructions, setCond instructions.
1165 assert(!isa<PointerType>(I.getType()));
1167 // We must cast the results of binary operations which might be promoted.
1168 bool needsCast = false;
1169 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1170 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1171 || (I.getType() == Type::FloatTy)) {
1174 printType(Out, I.getType());
1178 writeOperand(I.getOperand(0));
1180 switch (I.getOpcode()) {
1181 case Instruction::Add: Out << " + "; break;
1182 case Instruction::Sub: Out << " - "; break;
1183 case Instruction::Mul: Out << "*"; break;
1184 case Instruction::Div: Out << "/"; break;
1185 case Instruction::Rem: Out << "%"; break;
1186 case Instruction::And: Out << " & "; break;
1187 case Instruction::Or: Out << " | "; break;
1188 case Instruction::Xor: Out << " ^ "; break;
1189 case Instruction::SetEQ: Out << " == "; break;
1190 case Instruction::SetNE: Out << " != "; break;
1191 case Instruction::SetLE: Out << " <= "; break;
1192 case Instruction::SetGE: Out << " >= "; break;
1193 case Instruction::SetLT: Out << " < "; break;
1194 case Instruction::SetGT: Out << " > "; break;
1195 case Instruction::Shl : Out << " << "; break;
1196 case Instruction::Shr : Out << " >> "; break;
1197 default: std::cerr << "Invalid operator type!" << I; abort();
1200 writeOperand(I.getOperand(1));
1207 void CWriter::visitCastInst(CastInst &I) {
1208 if (I.getType() == Type::BoolTy) {
1210 writeOperand(I.getOperand(0));
1215 printType(Out, I.getType());
1217 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1218 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1219 // Avoid "cast to pointer from integer of different size" warnings
1223 writeOperand(I.getOperand(0));
1226 void CWriter::visitSelectInst(SelectInst &I) {
1228 writeOperand(I.getCondition());
1230 writeOperand(I.getTrueValue());
1232 writeOperand(I.getFalseValue());
1237 void CWriter::lowerIntrinsics(Function &F) {
1238 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1239 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1240 if (CallInst *CI = dyn_cast<CallInst>(I++))
1241 if (Function *F = CI->getCalledFunction())
1242 switch (F->getIntrinsicID()) {
1243 case Intrinsic::not_intrinsic:
1244 case Intrinsic::vastart:
1245 case Intrinsic::vacopy:
1246 case Intrinsic::vaend:
1247 case Intrinsic::returnaddress:
1248 case Intrinsic::frameaddress:
1249 case Intrinsic::setjmp:
1250 case Intrinsic::longjmp:
1251 // We directly implement these intrinsics
1254 // All other intrinsic calls we must lower.
1255 Instruction *Before = CI->getPrev();
1256 IL.LowerIntrinsicCall(CI);
1257 if (Before) { // Move iterator to instruction after call
1267 void CWriter::visitCallInst(CallInst &I) {
1268 // Handle intrinsic function calls first...
1269 if (Function *F = I.getCalledFunction())
1270 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1272 default: assert(0 && "Unknown LLVM intrinsic!");
1273 case Intrinsic::vastart:
1276 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1277 // Output the last argument to the enclosing function...
1278 if (I.getParent()->getParent()->aempty()) {
1279 std::cerr << "The C backend does not currently support zero "
1280 << "argument varargs functions, such as '"
1281 << I.getParent()->getParent()->getName() << "'!\n";
1284 writeOperand(&I.getParent()->getParent()->aback());
1287 case Intrinsic::vaend:
1288 Out << "va_end(*(va_list*)&";
1289 writeOperand(I.getOperand(1));
1292 case Intrinsic::vacopy:
1294 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1295 Out << "*(va_list*)&";
1296 writeOperand(I.getOperand(1));
1299 case Intrinsic::returnaddress:
1300 Out << "__builtin_return_address(";
1301 writeOperand(I.getOperand(1));
1304 case Intrinsic::frameaddress:
1305 Out << "__builtin_frame_address(";
1306 writeOperand(I.getOperand(1));
1309 case Intrinsic::setjmp:
1310 Out << "setjmp(*(jmp_buf*)";
1311 writeOperand(I.getOperand(1));
1314 case Intrinsic::longjmp:
1315 Out << "longjmp(*(jmp_buf*)";
1316 writeOperand(I.getOperand(1));
1318 writeOperand(I.getOperand(2));
1326 void CWriter::visitCallSite(CallSite CS) {
1327 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1328 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1329 const Type *RetTy = FTy->getReturnType();
1331 writeOperand(CS.getCalledValue());
1334 if (CS.arg_begin() != CS.arg_end()) {
1335 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1338 for (++AI; AI != AE; ++AI) {
1346 void CWriter::visitMallocInst(MallocInst &I) {
1347 assert(0 && "lowerallocations pass didn't work!");
1350 void CWriter::visitAllocaInst(AllocaInst &I) {
1352 printType(Out, I.getType());
1353 Out << ") alloca(sizeof(";
1354 printType(Out, I.getType()->getElementType());
1356 if (I.isArrayAllocation()) {
1358 writeOperand(I.getOperand(0));
1363 void CWriter::visitFreeInst(FreeInst &I) {
1364 assert(0 && "lowerallocations pass didn't work!");
1367 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1368 gep_type_iterator E) {
1369 bool HasImplicitAddress = false;
1370 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1371 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1372 HasImplicitAddress = true;
1373 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1374 HasImplicitAddress = true;
1375 Ptr = CPR->getValue(); // Get to the global...
1376 } else if (isDirectAlloca(Ptr)) {
1377 HasImplicitAddress = true;
1381 if (!HasImplicitAddress)
1382 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1384 writeOperandInternal(Ptr);
1388 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1389 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1392 writeOperandInternal(Ptr);
1394 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1396 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1399 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1400 "Can only have implicit address with direct accessing");
1402 if (HasImplicitAddress) {
1404 } else if (CI && CI->isNullValue()) {
1405 gep_type_iterator TmpI = I; ++TmpI;
1407 // Print out the -> operator if possible...
1408 if (TmpI != E && isa<StructType>(*TmpI)) {
1409 Out << (HasImplicitAddress ? "." : "->");
1410 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1416 if (isa<StructType>(*I)) {
1417 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1420 writeOperand(I.getOperand());
1425 void CWriter::visitLoadInst(LoadInst &I) {
1427 writeOperand(I.getOperand(0));
1430 void CWriter::visitStoreInst(StoreInst &I) {
1432 writeOperand(I.getPointerOperand());
1434 writeOperand(I.getOperand(0));
1437 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1439 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1443 void CWriter::visitVANextInst(VANextInst &I) {
1444 Out << Mang->getValueName(I.getOperand(0));
1445 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1446 printType(Out, I.getArgType());
1450 void CWriter::visitVAArgInst(VAArgInst &I) {
1452 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1453 writeOperand(I.getOperand(0));
1454 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1455 printType(Out, I.getType());
1456 Out << ");\n va_end(Tmp); }";
1459 //===----------------------------------------------------------------------===//
1460 // External Interface declaration
1461 //===----------------------------------------------------------------------===//
1463 bool CTargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &o) {
1464 PM.add(createLowerAllocationsPass());
1465 PM.add(createLowerInvokePass());
1466 PM.add(new CBackendNameAllUsedStructs());
1467 PM.add(new CWriter(o, getIntrinsicLowering()));
1471 TargetMachine *llvm::allocateCTargetMachine(const Module &M,
1472 IntrinsicLowering *IL) {
1473 return new CTargetMachine(M, IL);