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.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Assembly/CWriter.h"
15 #include "llvm/Constants.h"
16 #include "llvm/DerivedTypes.h"
17 #include "llvm/Module.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/Pass.h"
20 #include "llvm/SymbolTable.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/Analysis/FindUsedTypes.h"
23 #include "llvm/Analysis/ConstantsScanner.h"
24 #include "llvm/Support/InstVisitor.h"
25 #include "llvm/Support/InstIterator.h"
26 #include "llvm/Support/CallSite.h"
27 #include "llvm/Support/Mangler.h"
28 #include "Support/StringExtras.h"
29 #include "Support/STLExtras.h"
30 #include "Config/config.h"
35 class CWriter : public Pass, public InstVisitor<CWriter> {
38 const Module *TheModule;
39 std::map<const Type *, std::string> TypeNames;
40 std::set<const Value*> MangledGlobals;
41 bool needsMalloc, emittedInvoke;
43 std::map<const ConstantFP *, unsigned> FPConstantMap;
45 CWriter(std::ostream &o) : Out(o) {}
47 void getAnalysisUsage(AnalysisUsage &AU) const {
49 AU.addRequired<FindUsedTypes>();
52 virtual bool run(Module &M) {
56 // Ensure that all structure types have names...
57 bool Changed = nameAllUsedStructureTypes(M);
58 Mang = new Mangler(M);
66 MangledGlobals.clear();
70 std::ostream &printType(std::ostream &Out, const Type *Ty,
71 const std::string &VariableName = "",
72 bool IgnoreName = false, bool namedContext = true);
74 void writeOperand(Value *Operand);
75 void writeOperandInternal(Value *Operand);
78 bool nameAllUsedStructureTypes(Module &M);
79 void printModule(Module *M);
80 void printFloatingPointConstants(Module &M);
81 void printSymbolTable(const SymbolTable &ST);
82 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
83 void printFunctionSignature(const Function *F, bool Prototype);
85 void printFunction(Function *);
87 void printConstant(Constant *CPV);
88 void printConstantArray(ConstantArray *CPA);
90 // isInlinableInst - Attempt to inline instructions into their uses to build
91 // trees as much as possible. To do this, we have to consistently decide
92 // what is acceptable to inline, so that variable declarations don't get
93 // printed and an extra copy of the expr is not emitted.
95 static bool isInlinableInst(const Instruction &I) {
96 // Must be an expression, must be used exactly once. If it is dead, we
97 // emit it inline where it would go.
98 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
99 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
100 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
101 // Don't inline a load across a store or other bad things!
104 // Only inline instruction it it's use is in the same BB as the inst.
105 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
108 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
109 // variables which are accessed with the & operator. This causes GCC to
110 // generate significantly better code than to emit alloca calls directly.
112 static const AllocaInst *isDirectAlloca(const Value *V) {
113 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
114 if (!AI) return false;
115 if (AI->isArrayAllocation())
116 return 0; // FIXME: we can also inline fixed size array allocas!
117 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
122 // Instruction visitation functions
123 friend class InstVisitor<CWriter>;
125 void visitReturnInst(ReturnInst &I);
126 void visitBranchInst(BranchInst &I);
127 void visitSwitchInst(SwitchInst &I);
128 void visitInvokeInst(InvokeInst &I);
129 void visitUnwindInst(UnwindInst &I);
131 void visitPHINode(PHINode &I);
132 void visitBinaryOperator(Instruction &I);
134 void visitCastInst (CastInst &I);
135 void visitCallInst (CallInst &I);
136 void visitCallSite (CallSite CS);
137 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
139 void visitMallocInst(MallocInst &I);
140 void visitAllocaInst(AllocaInst &I);
141 void visitFreeInst (FreeInst &I);
142 void visitLoadInst (LoadInst &I);
143 void visitStoreInst (StoreInst &I);
144 void visitGetElementPtrInst(GetElementPtrInst &I);
145 void visitVANextInst(VANextInst &I);
146 void visitVAArgInst (VAArgInst &I);
148 void visitInstruction(Instruction &I) {
149 std::cerr << "C Writer does not know about " << I;
153 void outputLValue(Instruction *I) {
154 Out << " " << Mang->getValueName(I) << " = ";
156 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
158 void printIndexingExpression(Value *Ptr, User::op_iterator I,
159 User::op_iterator E);
163 // A pointer type should not use parens around *'s alone, e.g., (**)
164 inline bool ptrTypeNameNeedsParens(const std::string &NameSoFar) {
165 return NameSoFar.find_last_not_of('*') != std::string::npos;
168 // Pass the Type* and the variable name and this prints out the variable
171 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
172 const std::string &NameSoFar,
173 bool IgnoreName, bool namedContext) {
174 if (Ty->isPrimitiveType())
175 switch (Ty->getPrimitiveID()) {
176 case Type::VoidTyID: return Out << "void " << NameSoFar;
177 case Type::BoolTyID: return Out << "bool " << NameSoFar;
178 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
179 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
180 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
181 case Type::ShortTyID: return Out << "short " << NameSoFar;
182 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
183 case Type::IntTyID: return Out << "int " << NameSoFar;
184 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
185 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
186 case Type::FloatTyID: return Out << "float " << NameSoFar;
187 case Type::DoubleTyID: return Out << "double " << NameSoFar;
189 std::cerr << "Unknown primitive type: " << Ty << "\n";
193 // Check to see if the type is named.
194 if (!IgnoreName || isa<OpaqueType>(Ty)) {
195 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
196 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
199 switch (Ty->getPrimitiveID()) {
200 case Type::FunctionTyID: {
201 const FunctionType *MTy = cast<FunctionType>(Ty);
202 std::stringstream FunctionInnards;
203 FunctionInnards << " (" << NameSoFar << ") (";
204 for (FunctionType::ParamTypes::const_iterator
205 I = MTy->getParamTypes().begin(),
206 E = MTy->getParamTypes().end(); I != E; ++I) {
207 if (I != MTy->getParamTypes().begin())
208 FunctionInnards << ", ";
209 printType(FunctionInnards, *I, "");
211 if (MTy->isVarArg()) {
212 if (!MTy->getParamTypes().empty())
213 FunctionInnards << ", ...";
214 } else if (MTy->getParamTypes().empty()) {
215 FunctionInnards << "void";
217 FunctionInnards << ")";
218 std::string tstr = FunctionInnards.str();
219 printType(Out, MTy->getReturnType(), tstr);
222 case Type::StructTyID: {
223 const StructType *STy = cast<StructType>(Ty);
224 Out << NameSoFar + " {\n";
226 for (StructType::ElementTypes::const_iterator
227 I = STy->getElementTypes().begin(),
228 E = STy->getElementTypes().end(); I != E; ++I) {
230 printType(Out, *I, "field" + utostr(Idx++));
236 case Type::PointerTyID: {
237 const PointerType *PTy = cast<PointerType>(Ty);
238 std::string ptrName = "*" + NameSoFar;
240 // Do not need parens around "* NameSoFar" if NameSoFar consists only
241 // of zero or more '*' chars *and* this is not an unnamed pointer type
242 // such as the result type in a cast statement. Otherwise, enclose in ( ).
243 if (ptrTypeNameNeedsParens(NameSoFar) || !namedContext ||
244 PTy->getElementType()->getPrimitiveID() == Type::ArrayTyID)
245 ptrName = "(" + ptrName + ")"; //
247 return printType(Out, PTy->getElementType(), ptrName);
250 case Type::ArrayTyID: {
251 const ArrayType *ATy = cast<ArrayType>(Ty);
252 unsigned NumElements = ATy->getNumElements();
253 return printType(Out, ATy->getElementType(),
254 NameSoFar + "[" + utostr(NumElements) + "]");
257 case Type::OpaqueTyID: {
258 static int Count = 0;
259 std::string TyName = "struct opaque_" + itostr(Count++);
260 assert(TypeNames.find(Ty) == TypeNames.end());
261 TypeNames[Ty] = TyName;
262 return Out << TyName << " " << NameSoFar;
265 assert(0 && "Unhandled case in getTypeProps!");
272 void CWriter::printConstantArray(ConstantArray *CPA) {
274 // As a special case, print the array as a string if it is an array of
275 // ubytes or an array of sbytes with positive values.
277 const Type *ETy = CPA->getType()->getElementType();
278 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
280 // Make sure the last character is a null char, as automatically added by C
281 if (isString && (CPA->getNumOperands() == 0 ||
282 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
287 // Keep track of whether the last number was a hexadecimal escape
288 bool LastWasHex = false;
290 // Do not include the last character, which we know is null
291 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
292 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
294 // Print it out literally if it is a printable character. The only thing
295 // to be careful about is when the last letter output was a hex escape
296 // code, in which case we have to be careful not to print out hex digits
297 // explicitly (the C compiler thinks it is a continuation of the previous
298 // character, sheesh...)
300 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
302 if (C == '"' || C == '\\')
309 case '\n': Out << "\\n"; break;
310 case '\t': Out << "\\t"; break;
311 case '\r': Out << "\\r"; break;
312 case '\v': Out << "\\v"; break;
313 case '\a': Out << "\\a"; break;
314 case '\"': Out << "\\\""; break;
315 case '\'': Out << "\\\'"; break;
318 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
319 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
328 if (CPA->getNumOperands()) {
330 printConstant(cast<Constant>(CPA->getOperand(0)));
331 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
333 printConstant(cast<Constant>(CPA->getOperand(i)));
340 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
341 // textually as a double (rather than as a reference to a stack-allocated
342 // variable). We decide this by converting CFP to a string and back into a
343 // double, and then checking whether the conversion results in a bit-equal
344 // double to the original value of CFP. This depends on us and the target C
345 // compiler agreeing on the conversion process (which is pretty likely since we
346 // only deal in IEEE FP).
348 static bool isFPCSafeToPrint(const ConstantFP *CFP) {
351 sprintf(Buffer, "%a", CFP->getValue());
353 if (!strncmp(Buffer, "0x", 2) ||
354 !strncmp(Buffer, "-0x", 3) ||
355 !strncmp(Buffer, "+0x", 3))
356 return atof(Buffer) == CFP->getValue();
359 std::string StrVal = ftostr(CFP->getValue());
361 while (StrVal[0] == ' ')
362 StrVal.erase(StrVal.begin());
364 // Check to make sure that the stringized number is not some string like "Inf"
365 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
366 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
367 ((StrVal[0] == '-' || StrVal[0] == '+') &&
368 (StrVal[1] >= '0' && StrVal[1] <= '9')))
369 // Reparse stringized version!
370 return atof(StrVal.c_str()) == CFP->getValue();
375 // printConstant - The LLVM Constant to C Constant converter.
376 void CWriter::printConstant(Constant *CPV) {
377 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
378 switch (CE->getOpcode()) {
379 case Instruction::Cast:
381 printType(Out, CPV->getType());
383 printConstant(CE->getOperand(0));
387 case Instruction::GetElementPtr:
389 printIndexingExpression(CE->getOperand(0),
390 CPV->op_begin()+1, CPV->op_end());
393 case Instruction::Add:
394 case Instruction::Sub:
395 case Instruction::Mul:
396 case Instruction::Div:
397 case Instruction::Rem:
398 case Instruction::SetEQ:
399 case Instruction::SetNE:
400 case Instruction::SetLT:
401 case Instruction::SetLE:
402 case Instruction::SetGT:
403 case Instruction::SetGE:
405 printConstant(CE->getOperand(0));
406 switch (CE->getOpcode()) {
407 case Instruction::Add: Out << " + "; break;
408 case Instruction::Sub: Out << " - "; break;
409 case Instruction::Mul: Out << " * "; break;
410 case Instruction::Div: Out << " / "; break;
411 case Instruction::Rem: Out << " % "; break;
412 case Instruction::SetEQ: Out << " == "; break;
413 case Instruction::SetNE: Out << " != "; break;
414 case Instruction::SetLT: Out << " < "; break;
415 case Instruction::SetLE: Out << " <= "; break;
416 case Instruction::SetGT: Out << " > "; break;
417 case Instruction::SetGE: Out << " >= "; break;
418 default: assert(0 && "Illegal opcode here!");
420 printConstant(CE->getOperand(1));
425 std::cerr << "CWriter Error: Unhandled constant expression: "
431 switch (CPV->getType()->getPrimitiveID()) {
433 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
434 case Type::SByteTyID:
435 case Type::ShortTyID:
436 Out << cast<ConstantSInt>(CPV)->getValue(); break;
438 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
439 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
441 Out << cast<ConstantSInt>(CPV)->getValue();
445 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
447 case Type::UByteTyID:
448 case Type::UShortTyID:
449 Out << cast<ConstantUInt>(CPV)->getValue(); break;
451 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
452 case Type::ULongTyID:
453 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
455 case Type::FloatTyID:
456 case Type::DoubleTyID: {
457 ConstantFP *FPC = cast<ConstantFP>(CPV);
458 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
459 if (I != FPConstantMap.end()) {
460 // Because of FP precision problems we must load from a stack allocated
461 // value that holds the value in hex.
462 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
463 << "*)&FPConstant" << I->second << ")";
466 // Print out the constant as a floating point number.
468 sprintf(Buffer, "%a", FPC->getValue());
469 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
471 Out << ftostr(FPC->getValue());
477 case Type::ArrayTyID:
478 printConstantArray(cast<ConstantArray>(CPV));
481 case Type::StructTyID: {
483 if (CPV->getNumOperands()) {
485 printConstant(cast<Constant>(CPV->getOperand(0)));
486 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
488 printConstant(cast<Constant>(CPV->getOperand(i)));
495 case Type::PointerTyID:
496 if (isa<ConstantPointerNull>(CPV)) {
498 printType(Out, CPV->getType());
499 Out << ")/*NULL*/0)";
501 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
502 writeOperand(CPR->getValue());
507 std::cerr << "Unknown constant type: " << CPV << "\n";
512 void CWriter::writeOperandInternal(Value *Operand) {
513 if (Instruction *I = dyn_cast<Instruction>(Operand))
514 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
515 // Should we inline this instruction to build a tree?
522 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
525 Out << Mang->getValueName(Operand);
529 void CWriter::writeOperand(Value *Operand) {
530 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
531 Out << "(&"; // Global variables are references as their addresses by llvm
533 writeOperandInternal(Operand);
535 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
539 // nameAllUsedStructureTypes - If there are structure types in the module that
540 // are used but do not have names assigned to them in the symbol table yet then
541 // we assign them names now.
543 bool CWriter::nameAllUsedStructureTypes(Module &M) {
544 // Get a set of types that are used by the program...
545 std::set<const Type *> UT = getAnalysis<FindUsedTypes>().getTypes();
547 // Loop over the module symbol table, removing types from UT that are already
550 SymbolTable &MST = M.getSymbolTable();
551 if (MST.find(Type::TypeTy) != MST.end())
552 for (SymbolTable::type_iterator I = MST.type_begin(Type::TypeTy),
553 E = MST.type_end(Type::TypeTy); I != E; ++I)
554 UT.erase(cast<Type>(I->second));
556 // UT now contains types that are not named. Loop over it, naming structure
559 bool Changed = false;
560 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
562 if (const StructType *ST = dyn_cast<StructType>(*I)) {
563 ((Value*)ST)->setName("unnamed", &MST);
569 // generateCompilerSpecificCode - This is where we add conditional compilation
570 // directives to cater to specific compilers as need be.
572 static void generateCompilerSpecificCode(std::ostream& Out) {
573 // Alloca is hard to get, and we don't want to include stdlib.h here...
574 Out << "/* get a declaration for alloca */\n"
576 << "extern void *__builtin_alloca(unsigned long);\n"
577 << "#define alloca(x) __builtin_alloca(x)\n"
579 << "#ifndef __FreeBSD__\n"
580 << "#include <alloca.h>\n"
584 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
585 // If we aren't being compiled with GCC, just drop these attributes.
586 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
587 << "#define __attribute__(X)\n"
591 void CWriter::printModule(Module *M) {
592 // Calculate which global values have names that will collide when we throw
593 // away type information.
594 { // Scope to delete the FoundNames set when we are done with it...
595 std::set<std::string> FoundNames;
596 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
597 if (I->hasName()) // If the global has a name...
598 if (FoundNames.count(I->getName())) // And the name is already used
599 MangledGlobals.insert(I); // Mangle the name
601 FoundNames.insert(I->getName()); // Otherwise, keep track of name
603 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
604 if (I->hasName()) // If the global has a name...
605 if (FoundNames.count(I->getName())) // And the name is already used
606 MangledGlobals.insert(I); // Mangle the name
608 FoundNames.insert(I->getName()); // Otherwise, keep track of name
611 // get declaration for alloca
612 Out << "/* Provide Declarations */\n";
613 Out << "#include <stdarg.h>\n";
614 Out << "#include <setjmp.h>\n";
615 generateCompilerSpecificCode(Out);
617 // Provide a definition for `bool' if not compiling with a C++ compiler.
619 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
621 << "\n\n/* Support for floating point constants */\n"
622 << "typedef unsigned long long ConstantDoubleTy;\n"
623 << "typedef unsigned int ConstantFloatTy;\n"
625 << "\n\n/* Support for the invoke instruction */\n"
626 << "extern struct __llvm_jmpbuf_list_t {\n"
627 << " jmp_buf buf; struct __llvm_jmpbuf_list_t *next;\n"
628 << "} *__llvm_jmpbuf_list;\n"
630 << "\n\n/* Global Declarations */\n";
632 // First output all the declarations for the program, because C requires
633 // Functions & globals to be declared before they are used.
636 // Loop over the symbol table, emitting all named constants...
637 printSymbolTable(M->getSymbolTable());
639 // Global variable declarations...
641 Out << "\n/* External Global Variable Declarations */\n";
642 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) {
643 if (I->hasExternalLinkage()) {
645 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
651 // Function declarations
653 Out << "\n/* Function Declarations */\n";
655 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
656 // If the function is external and the name collides don't print it.
657 // Sometimes the bytecode likes to have multiple "declarations" for
658 // external functions
659 if ((I->hasInternalLinkage() || !MangledGlobals.count(I)) &&
660 !I->getIntrinsicID()) {
661 printFunctionSignature(I, true);
667 // Print Malloc prototype if needed
669 Out << "\n/* Malloc to make sun happy */\n";
670 Out << "extern void * malloc();\n\n";
673 // Output the global variable declarations
675 Out << "\n\n/* Global Variable Declarations */\n";
676 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
677 if (!I->isExternal()) {
679 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
685 // Output the global variable definitions and contents...
687 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
688 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
689 if (!I->isExternal()) {
690 if (I->hasInternalLinkage())
692 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
693 if (I->hasLinkOnceLinkage())
694 Out << " __attribute__((common))";
695 else if (I->hasWeakLinkage())
696 Out << " __attribute__((weak))";
697 if (!I->getInitializer()->isNullValue()) {
699 writeOperand(I->getInitializer());
705 // Output all floating point constants that cannot be printed accurately...
706 printFloatingPointConstants(*M);
708 // Output all of the functions...
709 emittedInvoke = false;
711 Out << "\n\n/* Function Bodies */\n";
712 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
716 // If the program included an invoke instruction, we need to output the
717 // support code for it here!
719 Out << "\n/* More support for the invoke instruction */\n"
720 << "struct __llvm_jmpbuf_list_t *__llvm_jmpbuf_list "
721 << "__attribute__((common)) = 0;\n";
724 // Done with global FP constants
725 FPConstantMap.clear();
728 /// Output all floating point constants that cannot be printed accurately...
729 void CWriter::printFloatingPointConstants(Module &M) {
732 unsigned long long U;
740 // Scan the module for floating point constants. If any FP constant is used
741 // in the function, we want to redirect it here so that we do not depend on
742 // the precision of the printed form, unless the printed form preserves
745 unsigned FPCounter = 0;
746 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
747 for (constant_iterator I = constant_begin(F), E = constant_end(F);
749 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
750 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
751 !FPConstantMap.count(FPC)) {
752 double Val = FPC->getValue();
754 FPConstantMap[FPC] = FPCounter; // Number the FP constants
756 if (FPC->getType() == Type::DoubleTy) {
758 Out << "const ConstantDoubleTy FPConstant" << FPCounter++
759 << " = 0x" << std::hex << DBLUnion.U << std::dec
760 << "ULL; /* " << Val << " */\n";
761 } else if (FPC->getType() == Type::FloatTy) {
763 Out << "const ConstantFloatTy FPConstant" << FPCounter++
764 << " = 0x" << std::hex << FLTUnion.U << std::dec
765 << "U; /* " << Val << " */\n";
767 assert(0 && "Unknown float type!");
774 /// printSymbolTable - Run through symbol table looking for type names. If a
775 /// type name is found, emit it's declaration...
777 void CWriter::printSymbolTable(const SymbolTable &ST) {
778 // If there are no type names, exit early.
779 if (ST.find(Type::TypeTy) == ST.end())
782 // We are only interested in the type plane of the symbol table...
783 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
784 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
786 // Print out forward declarations for structure types before anything else!
787 Out << "/* Structure forward decls */\n";
788 for (; I != End; ++I)
789 if (const Type *STy = dyn_cast<StructType>(I->second)) {
790 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
791 Out << Name << ";\n";
792 TypeNames.insert(std::make_pair(STy, Name));
797 // Now we can print out typedefs...
798 Out << "/* Typedefs */\n";
799 for (I = ST.type_begin(Type::TypeTy); I != End; ++I) {
800 const Type *Ty = cast<Type>(I->second);
801 std::string Name = "l_" + Mangler::makeNameProper(I->first);
803 printType(Out, Ty, Name);
809 // Keep track of which structures have been printed so far...
810 std::set<const StructType *> StructPrinted;
812 // Loop over all structures then push them into the stack so they are
813 // printed in the correct order.
815 Out << "/* Structure contents */\n";
816 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
817 if (const StructType *STy = dyn_cast<StructType>(I->second))
818 printContainedStructs(STy, StructPrinted);
821 // Push the struct onto the stack and recursively push all structs
822 // this one depends on.
823 void CWriter::printContainedStructs(const Type *Ty,
824 std::set<const StructType*> &StructPrinted){
825 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
826 //Check to see if we have already printed this struct
827 if (StructPrinted.count(STy) == 0) {
828 // Print all contained types first...
829 for (StructType::ElementTypes::const_iterator
830 I = STy->getElementTypes().begin(),
831 E = STy->getElementTypes().end(); I != E; ++I) {
832 const Type *Ty1 = I->get();
833 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
834 printContainedStructs(*I, StructPrinted);
837 //Print structure type out..
838 StructPrinted.insert(STy);
839 std::string Name = TypeNames[STy];
840 printType(Out, STy, Name, true);
844 // If it is an array, check contained types and continue
845 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
846 const Type *Ty1 = ATy->getElementType();
847 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
848 printContainedStructs(Ty1, StructPrinted);
853 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
854 // If the program provides its own malloc prototype we don't need
855 // to include the general one.
856 if (Mang->getValueName(F) == "malloc")
859 if (F->hasInternalLinkage()) Out << "static ";
860 if (F->hasLinkOnceLinkage()) Out << "inline ";
862 // Loop over the arguments, printing them...
863 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
865 std::stringstream FunctionInnards;
867 // Print out the name...
868 FunctionInnards << Mang->getValueName(F) << "(";
870 if (!F->isExternal()) {
873 if (F->abegin()->hasName() || !Prototype)
874 ArgName = Mang->getValueName(F->abegin());
875 printType(FunctionInnards, F->afront().getType(), ArgName);
876 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
878 FunctionInnards << ", ";
879 if (I->hasName() || !Prototype)
880 ArgName = Mang->getValueName(I);
883 printType(FunctionInnards, I->getType(), ArgName);
887 // Loop over the arguments, printing them...
888 for (FunctionType::ParamTypes::const_iterator I =
889 FT->getParamTypes().begin(),
890 E = FT->getParamTypes().end(); I != E; ++I) {
891 if (I != FT->getParamTypes().begin()) FunctionInnards << ", ";
892 printType(FunctionInnards, *I);
896 // Finish printing arguments... if this is a vararg function, print the ...,
897 // unless there are no known types, in which case, we just emit ().
899 if (FT->isVarArg() && !FT->getParamTypes().empty()) {
900 if (FT->getParamTypes().size()) FunctionInnards << ", ";
901 FunctionInnards << "..."; // Output varargs portion of signature!
903 FunctionInnards << ")";
904 // Print out the return type and the entire signature for that matter
905 printType(Out, F->getReturnType(), FunctionInnards.str());
907 if (F->hasWeakLinkage()) Out << " __attribute((weak))";
910 void CWriter::printFunction(Function *F) {
911 if (F->isExternal()) return;
913 printFunctionSignature(F, false);
916 // print local variable information for the function
917 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
918 if (const AllocaInst *AI = isDirectAlloca(*I)) {
920 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
921 Out << "; /* Address exposed local */\n";
922 } else if ((*I)->getType() != Type::VoidTy && !isInlinableInst(**I)) {
924 printType(Out, (*I)->getType(), Mang->getValueName(*I));
927 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
929 printType(Out, (*I)->getType(),
930 Mang->getValueName(*I)+"__PHI_TEMPORARY");
937 // print the basic blocks
938 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
939 BasicBlock *Prev = BB->getPrev();
941 // Don't print the label for the basic block if there are no uses, or if the
942 // only terminator use is the predecessor basic block's terminator. We have
943 // to scan the use list because PHI nodes use basic blocks too but do not
944 // require a label to be generated.
946 bool NeedsLabel = false;
947 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
949 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
950 if (TI != Prev->getTerminator() ||
951 isa<SwitchInst>(Prev->getTerminator()) ||
952 isa<InvokeInst>(Prev->getTerminator())) {
957 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
959 // Output all of the instructions in the basic block...
960 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
961 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
962 if (II->getType() != Type::VoidTy)
971 // Don't emit prefix or suffix for the terminator...
972 visit(*BB->getTerminator());
978 // Specific Instruction type classes... note that all of the casts are
979 // necessary because we use the instruction classes as opaque types...
981 void CWriter::visitReturnInst(ReturnInst &I) {
982 // Don't output a void return if this is the last basic block in the function
983 if (I.getNumOperands() == 0 &&
984 &*--I.getParent()->getParent()->end() == I.getParent() &&
985 !I.getParent()->size() == 1) {
990 if (I.getNumOperands()) {
992 writeOperand(I.getOperand(0));
997 void CWriter::visitSwitchInst(SwitchInst &SI) {
999 writeOperand(SI.getOperand(0));
1000 Out << ") {\n default:\n";
1001 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1003 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1005 writeOperand(SI.getOperand(i));
1007 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1008 printBranchToBlock(SI.getParent(), Succ, 2);
1009 if (Succ == SI.getParent()->getNext())
1015 void CWriter::visitInvokeInst(InvokeInst &II) {
1017 << " struct __llvm_jmpbuf_list_t Entry;\n"
1018 << " Entry.next = __llvm_jmpbuf_list;\n"
1019 << " if (setjmp(Entry.buf)) {\n"
1020 << " __llvm_jmpbuf_list = Entry.next;\n";
1021 printBranchToBlock(II.getParent(), II.getExceptionalDest(), 4);
1023 << " __llvm_jmpbuf_list = &Entry;\n"
1026 if (II.getType() != Type::VoidTy) outputLValue(&II);
1029 << " __llvm_jmpbuf_list = Entry.next;\n"
1031 printBranchToBlock(II.getParent(), II.getNormalDest(), 0);
1032 emittedInvoke = true;
1036 void CWriter::visitUnwindInst(UnwindInst &I) {
1037 // The unwind instructions causes a control flow transfer out of the current
1038 // function, unwinding the stack until a caller who used the invoke
1039 // instruction is found. In this context, we code generated the invoke
1040 // instruction to add an entry to the top of the jmpbuf_list. Thus, here we
1041 // just have to longjmp to the specified handler.
1042 Out << " if (__llvm_jmpbuf_list == 0) { /* unwind */\n"
1043 << " extern write();\n"
1044 << " ((void (*)(int, void*, unsigned))write)(2,\n"
1045 << " \"throw found with no handler!\\n\", 31); abort();\n"
1047 << " longjmp(__llvm_jmpbuf_list->buf, 1);\n";
1048 emittedInvoke = true;
1051 static bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1052 // If PHI nodes need copies, we need the copy code...
1053 if (isa<PHINode>(To->front()) ||
1054 From->getNext() != To) // Not directly successor, need goto
1057 // Otherwise we don't need the code.
1061 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1063 for (BasicBlock::iterator I = Succ->begin();
1064 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1065 // now we have to do the printing
1066 Out << std::string(Indent, ' ');
1067 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1068 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBB)));
1069 Out << "; /* for PHI node */\n";
1072 if (CurBB->getNext() != Succ ||
1073 isa<InvokeInst>(CurBB->getTerminator()) ||
1074 isa<SwitchInst>(CurBB->getTerminator())) {
1075 Out << std::string(Indent, ' ') << " goto ";
1081 // Branch instruction printing - Avoid printing out a branch to a basic block
1082 // that immediately succeeds the current one.
1084 void CWriter::visitBranchInst(BranchInst &I) {
1085 if (I.isConditional()) {
1086 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1088 writeOperand(I.getCondition());
1091 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1093 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1094 Out << " } else {\n";
1095 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1098 // First goto not necessary, assume second one is...
1100 writeOperand(I.getCondition());
1103 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1108 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1113 // PHI nodes get copied into temporary values at the end of predecessor basic
1114 // blocks. We now need to copy these temporary values into the REAL value for
1116 void CWriter::visitPHINode(PHINode &I) {
1118 Out << "__PHI_TEMPORARY";
1122 void CWriter::visitBinaryOperator(Instruction &I) {
1123 // binary instructions, shift instructions, setCond instructions.
1124 assert(!isa<PointerType>(I.getType()));
1126 // We must cast the results of binary operations which might be promoted.
1127 bool needsCast = false;
1128 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1129 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1130 || (I.getType() == Type::FloatTy)) {
1133 printType(Out, I.getType(), "", false, false);
1137 writeOperand(I.getOperand(0));
1139 switch (I.getOpcode()) {
1140 case Instruction::Add: Out << " + "; break;
1141 case Instruction::Sub: Out << " - "; break;
1142 case Instruction::Mul: Out << "*"; break;
1143 case Instruction::Div: Out << "/"; break;
1144 case Instruction::Rem: Out << "%"; break;
1145 case Instruction::And: Out << " & "; break;
1146 case Instruction::Or: Out << " | "; break;
1147 case Instruction::Xor: Out << " ^ "; break;
1148 case Instruction::SetEQ: Out << " == "; break;
1149 case Instruction::SetNE: Out << " != "; break;
1150 case Instruction::SetLE: Out << " <= "; break;
1151 case Instruction::SetGE: Out << " >= "; break;
1152 case Instruction::SetLT: Out << " < "; break;
1153 case Instruction::SetGT: Out << " > "; break;
1154 case Instruction::Shl : Out << " << "; break;
1155 case Instruction::Shr : Out << " >> "; break;
1156 default: std::cerr << "Invalid operator type!" << I; abort();
1159 writeOperand(I.getOperand(1));
1166 void CWriter::visitCastInst(CastInst &I) {
1167 if (I.getType() == Type::BoolTy) {
1169 writeOperand(I.getOperand(0));
1174 printType(Out, I.getType(), "", /*ignoreName*/false, /*namedContext*/false);
1176 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1177 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1178 // Avoid "cast to pointer from integer of different size" warnings
1182 writeOperand(I.getOperand(0));
1185 void CWriter::visitCallInst(CallInst &I) {
1186 // Handle intrinsic function calls first...
1187 if (Function *F = I.getCalledFunction())
1188 if (LLVMIntrinsic::ID ID = (LLVMIntrinsic::ID)F->getIntrinsicID()) {
1190 default: assert(0 && "Unknown LLVM intrinsic!");
1191 case LLVMIntrinsic::va_start:
1194 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1195 // Output the last argument to the enclosing function...
1196 if (I.getParent()->getParent()->aempty()) {
1197 std::cerr << "The C backend does not currently support zero "
1198 << "argument varargs functions, such as '"
1199 << I.getParent()->getParent()->getName() << "'!\n";
1202 writeOperand(&I.getParent()->getParent()->aback());
1205 case LLVMIntrinsic::va_end:
1206 Out << "va_end(*(va_list*)&";
1207 writeOperand(I.getOperand(1));
1210 case LLVMIntrinsic::va_copy:
1212 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1213 Out << "*(va_list*)&";
1214 writeOperand(I.getOperand(1));
1217 case LLVMIntrinsic::setjmp:
1218 case LLVMIntrinsic::sigsetjmp:
1219 // This intrinsic should never exist in the program, but until we get
1220 // setjmp/longjmp transformations going on, we should codegen it to
1221 // something reasonable. This will allow code that never calls longjmp
1225 case LLVMIntrinsic::longjmp:
1226 case LLVMIntrinsic::siglongjmp:
1227 // Longjmp is not implemented, and never will be. It would cause an
1236 void CWriter::visitCallSite(CallSite CS) {
1237 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1238 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1239 const Type *RetTy = FTy->getReturnType();
1241 writeOperand(CS.getCalledValue());
1244 if (CS.arg_begin() != CS.arg_end()) {
1245 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1248 for (++AI; AI != AE; ++AI) {
1256 void CWriter::visitMallocInst(MallocInst &I) {
1258 printType(Out, I.getType());
1259 Out << ")malloc(sizeof(";
1260 printType(Out, I.getType()->getElementType());
1263 if (I.isArrayAllocation()) {
1265 writeOperand(I.getOperand(0));
1270 void CWriter::visitAllocaInst(AllocaInst &I) {
1272 printType(Out, I.getType());
1273 Out << ") alloca(sizeof(";
1274 printType(Out, I.getType()->getElementType());
1276 if (I.isArrayAllocation()) {
1278 writeOperand(I.getOperand(0));
1283 void CWriter::visitFreeInst(FreeInst &I) {
1284 Out << "free((char*)";
1285 writeOperand(I.getOperand(0));
1289 void CWriter::printIndexingExpression(Value *Ptr, User::op_iterator I,
1290 User::op_iterator E) {
1291 bool HasImplicitAddress = false;
1292 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1293 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1294 HasImplicitAddress = true;
1295 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1296 HasImplicitAddress = true;
1297 Ptr = CPR->getValue(); // Get to the global...
1298 } else if (isDirectAlloca(Ptr)) {
1299 HasImplicitAddress = true;
1303 if (!HasImplicitAddress)
1304 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1306 writeOperandInternal(Ptr);
1310 const Constant *CI = dyn_cast<Constant>(I);
1311 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1314 writeOperandInternal(Ptr);
1316 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1318 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1321 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1322 "Can only have implicit address with direct accessing");
1324 if (HasImplicitAddress) {
1326 } else if (CI && CI->isNullValue() && I+1 != E) {
1327 // Print out the -> operator if possible...
1328 if ((*(I+1))->getType() == Type::UByteTy) {
1329 Out << (HasImplicitAddress ? "." : "->");
1330 Out << "field" << cast<ConstantUInt>(*(I+1))->getValue();
1336 if ((*I)->getType() == Type::LongTy) {
1341 Out << ".field" << cast<ConstantUInt>(*I)->getValue();
1345 void CWriter::visitLoadInst(LoadInst &I) {
1347 writeOperand(I.getOperand(0));
1350 void CWriter::visitStoreInst(StoreInst &I) {
1352 writeOperand(I.getPointerOperand());
1354 writeOperand(I.getOperand(0));
1357 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1359 printIndexingExpression(I.getPointerOperand(), I.idx_begin(), I.idx_end());
1362 void CWriter::visitVANextInst(VANextInst &I) {
1363 Out << Mang->getValueName(I.getOperand(0));
1364 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1365 printType(Out, I.getArgType(), "", /*ignoreName*/false,
1366 /*namedContext*/false);
1370 void CWriter::visitVAArgInst(VAArgInst &I) {
1372 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1373 writeOperand(I.getOperand(0));
1374 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1375 printType(Out, I.getType(), "", /*ignoreName*/false, /*namedContext*/false);
1376 Out << ");\n va_end(Tmp); }";
1380 //===----------------------------------------------------------------------===//
1381 // External Interface declaration
1382 //===----------------------------------------------------------------------===//
1384 Pass *createWriteToCPass(std::ostream &o) { return new CWriter(o); }