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/CallSite.h"
25 #include "llvm/Support/GetElementPtrTypeIterator.h"
26 #include "llvm/Support/InstVisitor.h"
27 #include "llvm/Support/InstIterator.h"
28 #include "llvm/Support/Mangler.h"
29 #include "Support/StringExtras.h"
30 #include "Support/STLExtras.h"
31 #include "Config/config.h"
38 class CWriter : public Pass, public InstVisitor<CWriter> {
41 const Module *TheModule;
44 std::map<const Type *, std::string> TypeNames;
45 std::set<const Value*> MangledGlobals;
46 bool needsMalloc, emittedInvoke;
48 std::map<const ConstantFP *, unsigned> FPConstantMap;
50 CWriter(std::ostream &o) : Out(o) {}
52 void getAnalysisUsage(AnalysisUsage &AU) const {
54 AU.addRequired<FindUsedTypes>();
57 virtual bool run(Module &M) {
60 FUT = &getAnalysis<FindUsedTypes>();
62 // Ensure that all structure types have names...
63 bool Changed = nameAllUsedStructureTypes(M);
64 Mang = new Mangler(M);
72 MangledGlobals.clear();
76 std::ostream &printType(std::ostream &Out, const Type *Ty,
77 const std::string &VariableName = "",
78 bool IgnoreName = false);
80 void writeOperand(Value *Operand);
81 void writeOperandInternal(Value *Operand);
84 bool nameAllUsedStructureTypes(Module &M);
85 void printModule(Module *M);
86 void printFloatingPointConstants(Module &M);
87 void printSymbolTable(const SymbolTable &ST);
88 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
89 void printFunctionSignature(const Function *F, bool Prototype);
91 void printFunction(Function *);
93 void printConstant(Constant *CPV);
94 void printConstantArray(ConstantArray *CPA);
96 // isInlinableInst - Attempt to inline instructions into their uses to build
97 // trees as much as possible. To do this, we have to consistently decide
98 // what is acceptable to inline, so that variable declarations don't get
99 // printed and an extra copy of the expr is not emitted.
101 static bool isInlinableInst(const Instruction &I) {
102 // Must be an expression, must be used exactly once. If it is dead, we
103 // emit it inline where it would go.
104 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
105 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
106 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
107 // Don't inline a load across a store or other bad things!
110 // Only inline instruction it it's use is in the same BB as the inst.
111 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
114 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
115 // variables which are accessed with the & operator. This causes GCC to
116 // generate significantly better code than to emit alloca calls directly.
118 static const AllocaInst *isDirectAlloca(const Value *V) {
119 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
120 if (!AI) return false;
121 if (AI->isArrayAllocation())
122 return 0; // FIXME: we can also inline fixed size array allocas!
123 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
128 // Instruction visitation functions
129 friend class InstVisitor<CWriter>;
131 void visitReturnInst(ReturnInst &I);
132 void visitBranchInst(BranchInst &I);
133 void visitSwitchInst(SwitchInst &I);
134 void visitInvokeInst(InvokeInst &I);
135 void visitUnwindInst(UnwindInst &I);
137 void visitPHINode(PHINode &I);
138 void visitBinaryOperator(Instruction &I);
140 void visitCastInst (CastInst &I);
141 void visitCallInst (CallInst &I);
142 void visitCallSite (CallSite CS);
143 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
145 void visitMallocInst(MallocInst &I);
146 void visitAllocaInst(AllocaInst &I);
147 void visitFreeInst (FreeInst &I);
148 void visitLoadInst (LoadInst &I);
149 void visitStoreInst (StoreInst &I);
150 void visitGetElementPtrInst(GetElementPtrInst &I);
151 void visitVANextInst(VANextInst &I);
152 void visitVAArgInst (VAArgInst &I);
154 void visitInstruction(Instruction &I) {
155 std::cerr << "C Writer does not know about " << I;
159 void outputLValue(Instruction *I) {
160 Out << " " << Mang->getValueName(I) << " = ";
162 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
164 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
165 gep_type_iterator E);
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,
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::param_iterator I = MTy->param_begin(),
205 E = MTy->param_end(); I != E; ++I) {
206 if (I != MTy->param_begin())
207 FunctionInnards << ", ";
208 printType(FunctionInnards, *I, "");
210 if (MTy->isVarArg()) {
211 if (MTy->getNumParams())
212 FunctionInnards << ", ...";
213 } else if (!MTy->getNumParams()) {
214 FunctionInnards << "void";
216 FunctionInnards << ")";
217 std::string tstr = FunctionInnards.str();
218 printType(Out, MTy->getReturnType(), tstr);
221 case Type::StructTyID: {
222 const StructType *STy = cast<StructType>(Ty);
223 Out << NameSoFar + " {\n";
225 for (StructType::element_iterator I = STy->element_begin(),
226 E = STy->element_end(); I != E; ++I) {
228 printType(Out, *I, "field" + utostr(Idx++));
234 case Type::PointerTyID: {
235 const PointerType *PTy = cast<PointerType>(Ty);
236 std::string ptrName = "*" + NameSoFar;
238 if (isa<ArrayType>(PTy->getElementType()))
239 ptrName = "(" + ptrName + ")";
241 return printType(Out, PTy->getElementType(), ptrName);
244 case Type::ArrayTyID: {
245 const ArrayType *ATy = cast<ArrayType>(Ty);
246 unsigned NumElements = ATy->getNumElements();
247 return printType(Out, ATy->getElementType(),
248 NameSoFar + "[" + utostr(NumElements) + "]");
251 case Type::OpaqueTyID: {
252 static int Count = 0;
253 std::string TyName = "struct opaque_" + itostr(Count++);
254 assert(TypeNames.find(Ty) == TypeNames.end());
255 TypeNames[Ty] = TyName;
256 return Out << TyName << " " << NameSoFar;
259 assert(0 && "Unhandled case in getTypeProps!");
266 void CWriter::printConstantArray(ConstantArray *CPA) {
268 // As a special case, print the array as a string if it is an array of
269 // ubytes or an array of sbytes with positive values.
271 const Type *ETy = CPA->getType()->getElementType();
272 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
274 // Make sure the last character is a null char, as automatically added by C
275 if (isString && (CPA->getNumOperands() == 0 ||
276 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
281 // Keep track of whether the last number was a hexadecimal escape
282 bool LastWasHex = false;
284 // Do not include the last character, which we know is null
285 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
286 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
288 // Print it out literally if it is a printable character. The only thing
289 // to be careful about is when the last letter output was a hex escape
290 // code, in which case we have to be careful not to print out hex digits
291 // explicitly (the C compiler thinks it is a continuation of the previous
292 // character, sheesh...)
294 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
296 if (C == '"' || C == '\\')
303 case '\n': Out << "\\n"; break;
304 case '\t': Out << "\\t"; break;
305 case '\r': Out << "\\r"; break;
306 case '\v': Out << "\\v"; break;
307 case '\a': Out << "\\a"; break;
308 case '\"': Out << "\\\""; break;
309 case '\'': Out << "\\\'"; break;
312 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
313 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
322 if (CPA->getNumOperands()) {
324 printConstant(cast<Constant>(CPA->getOperand(0)));
325 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
327 printConstant(cast<Constant>(CPA->getOperand(i)));
334 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
335 // textually as a double (rather than as a reference to a stack-allocated
336 // variable). We decide this by converting CFP to a string and back into a
337 // double, and then checking whether the conversion results in a bit-equal
338 // double to the original value of CFP. This depends on us and the target C
339 // compiler agreeing on the conversion process (which is pretty likely since we
340 // only deal in IEEE FP).
342 bool isFPCSafeToPrint(const ConstantFP *CFP) {
345 sprintf(Buffer, "%a", CFP->getValue());
347 if (!strncmp(Buffer, "0x", 2) ||
348 !strncmp(Buffer, "-0x", 3) ||
349 !strncmp(Buffer, "+0x", 3))
350 return atof(Buffer) == CFP->getValue();
353 std::string StrVal = ftostr(CFP->getValue());
355 while (StrVal[0] == ' ')
356 StrVal.erase(StrVal.begin());
358 // Check to make sure that the stringized number is not some string like "Inf"
359 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
360 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
361 ((StrVal[0] == '-' || StrVal[0] == '+') &&
362 (StrVal[1] >= '0' && StrVal[1] <= '9')))
363 // Reparse stringized version!
364 return atof(StrVal.c_str()) == CFP->getValue();
369 // printConstant - The LLVM Constant to C Constant converter.
370 void CWriter::printConstant(Constant *CPV) {
371 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
372 switch (CE->getOpcode()) {
373 case Instruction::Cast:
375 printType(Out, CPV->getType());
377 printConstant(CE->getOperand(0));
381 case Instruction::GetElementPtr:
383 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
387 case Instruction::Add:
388 case Instruction::Sub:
389 case Instruction::Mul:
390 case Instruction::Div:
391 case Instruction::Rem:
392 case Instruction::SetEQ:
393 case Instruction::SetNE:
394 case Instruction::SetLT:
395 case Instruction::SetLE:
396 case Instruction::SetGT:
397 case Instruction::SetGE:
398 case Instruction::Shl:
399 case Instruction::Shr:
401 printConstant(CE->getOperand(0));
402 switch (CE->getOpcode()) {
403 case Instruction::Add: Out << " + "; break;
404 case Instruction::Sub: Out << " - "; break;
405 case Instruction::Mul: Out << " * "; break;
406 case Instruction::Div: Out << " / "; break;
407 case Instruction::Rem: Out << " % "; break;
408 case Instruction::SetEQ: Out << " == "; break;
409 case Instruction::SetNE: Out << " != "; break;
410 case Instruction::SetLT: Out << " < "; break;
411 case Instruction::SetLE: Out << " <= "; break;
412 case Instruction::SetGT: Out << " > "; break;
413 case Instruction::SetGE: Out << " >= "; break;
414 case Instruction::Shl: Out << " << "; break;
415 case Instruction::Shr: Out << " >> "; break;
416 default: assert(0 && "Illegal opcode here!");
418 printConstant(CE->getOperand(1));
423 std::cerr << "CWriter Error: Unhandled constant expression: "
429 switch (CPV->getType()->getPrimitiveID()) {
431 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
432 case Type::SByteTyID:
433 case Type::ShortTyID:
434 Out << cast<ConstantSInt>(CPV)->getValue(); break;
436 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
437 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
439 Out << cast<ConstantSInt>(CPV)->getValue();
443 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
445 case Type::UByteTyID:
446 case Type::UShortTyID:
447 Out << cast<ConstantUInt>(CPV)->getValue(); break;
449 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
450 case Type::ULongTyID:
451 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
453 case Type::FloatTyID:
454 case Type::DoubleTyID: {
455 ConstantFP *FPC = cast<ConstantFP>(CPV);
456 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
457 if (I != FPConstantMap.end()) {
458 // Because of FP precision problems we must load from a stack allocated
459 // value that holds the value in hex.
460 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
461 << "*)&FPConstant" << I->second << ")";
464 // Print out the constant as a floating point number.
466 sprintf(Buffer, "%a", FPC->getValue());
467 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
469 Out << ftostr(FPC->getValue());
475 case Type::ArrayTyID:
476 printConstantArray(cast<ConstantArray>(CPV));
479 case Type::StructTyID: {
481 if (CPV->getNumOperands()) {
483 printConstant(cast<Constant>(CPV->getOperand(0)));
484 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
486 printConstant(cast<Constant>(CPV->getOperand(i)));
493 case Type::PointerTyID:
494 if (isa<ConstantPointerNull>(CPV)) {
496 printType(Out, CPV->getType());
497 Out << ")/*NULL*/0)";
499 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
500 writeOperand(CPR->getValue());
505 std::cerr << "Unknown constant type: " << CPV << "\n";
510 void CWriter::writeOperandInternal(Value *Operand) {
511 if (Instruction *I = dyn_cast<Instruction>(Operand))
512 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
513 // Should we inline this instruction to build a tree?
520 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
523 Out << Mang->getValueName(Operand);
527 void CWriter::writeOperand(Value *Operand) {
528 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
529 Out << "(&"; // Global variables are references as their addresses by llvm
531 writeOperandInternal(Operand);
533 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
537 // nameAllUsedStructureTypes - If there are structure types in the module that
538 // are used but do not have names assigned to them in the symbol table yet then
539 // we assign them names now.
541 bool CWriter::nameAllUsedStructureTypes(Module &M) {
542 // Get a set of types that are used by the program...
543 std::set<const Type *> UT = FUT->getTypes();
545 // Loop over the module symbol table, removing types from UT that are already
548 SymbolTable &MST = M.getSymbolTable();
549 if (MST.find(Type::TypeTy) != MST.end())
550 for (SymbolTable::type_iterator I = MST.type_begin(Type::TypeTy),
551 E = MST.type_end(Type::TypeTy); I != E; ++I)
552 UT.erase(cast<Type>(I->second));
554 // UT now contains types that are not named. Loop over it, naming structure
557 bool Changed = false;
558 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
560 if (const StructType *ST = dyn_cast<StructType>(*I)) {
561 ((Value*)ST)->setName("unnamed", &MST);
567 // generateCompilerSpecificCode - This is where we add conditional compilation
568 // directives to cater to specific compilers as need be.
570 static void generateCompilerSpecificCode(std::ostream& Out) {
571 // Alloca is hard to get, and we don't want to include stdlib.h here...
572 Out << "/* get a declaration for alloca */\n"
574 << "extern void *__builtin_alloca(unsigned long);\n"
575 << "#define alloca(x) __builtin_alloca(x)\n"
577 << "#ifndef __FreeBSD__\n"
578 << "#include <alloca.h>\n"
582 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
583 // If we aren't being compiled with GCC, just drop these attributes.
584 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
585 << "#define __attribute__(X)\n"
589 // At some point, we should support "external weak" vs. "weak" linkages.
590 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
591 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
592 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
593 << "#elif defined(__GNUC__)\n"
594 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
596 << "#define __EXTERNAL_WEAK__\n"
600 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
601 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
602 << "#define __ATTRIBUTE_WEAK__\n"
603 << "#elif defined(__GNUC__)\n"
604 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
606 << "#define __ATTRIBUTE_WEAK__\n"
610 // generateProcessorSpecificCode - This is where we add conditional compilation
611 // directives to cater to specific processors as need be.
613 static void generateProcessorSpecificCode(std::ostream& Out) {
614 // According to ANSI C, longjmp'ing to a setjmp could invalidate any
615 // non-volatile variable in the scope of the setjmp. For now, we are not
616 // doing analysis to determine which variables need to be marked volatile, so
617 // we just mark them all.
619 // HOWEVER, many targets implement setjmp by saving and restoring the register
620 // file, so they DON'T need variables to be marked volatile, and this is a
621 // HUGE pessimization for them. For this reason, on known-good processors, we
622 // do not emit volatile qualifiers.
623 Out << "#if defined(__386__) || defined(__i386__) || \\\n"
624 << " defined(i386) || defined(WIN32)\n"
625 << "/* setjmp does not require variables to be marked volatile */"
626 << "#define VOLATILE_FOR_SETJMP\n"
628 << "#define VOLATILE_FOR_SETJMP volatile\n"
633 void CWriter::printModule(Module *M) {
634 // Calculate which global values have names that will collide when we throw
635 // away type information.
636 { // Scope to delete the FoundNames set when we are done with it...
637 std::set<std::string> FoundNames;
638 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
639 if (I->hasName()) // If the global has a name...
640 if (FoundNames.count(I->getName())) // And the name is already used
641 MangledGlobals.insert(I); // Mangle the name
643 FoundNames.insert(I->getName()); // Otherwise, keep track of name
645 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
646 if (I->hasName()) // If the global has a name...
647 if (FoundNames.count(I->getName())) // And the name is already used
648 MangledGlobals.insert(I); // Mangle the name
650 FoundNames.insert(I->getName()); // Otherwise, keep track of name
653 // get declaration for alloca
654 Out << "/* Provide Declarations */\n";
655 Out << "#include <stdarg.h>\n"; // Varargs support
656 Out << "#include <setjmp.h>\n"; // Unwind support
657 generateCompilerSpecificCode(Out);
658 generateProcessorSpecificCode(Out);
660 // Provide a definition for `bool' if not compiling with a C++ compiler.
662 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
664 << "\n\n/* Support for floating point constants */\n"
665 << "typedef unsigned long long ConstantDoubleTy;\n"
666 << "typedef unsigned int ConstantFloatTy;\n"
668 << "\n\n/* Support for the invoke instruction */\n"
669 << "extern struct __llvm_jmpbuf_list_t {\n"
670 << " jmp_buf buf; struct __llvm_jmpbuf_list_t *next;\n"
671 << "} *__llvm_jmpbuf_list;\n"
673 << "\n\n/* Global Declarations */\n";
675 // First output all the declarations for the program, because C requires
676 // Functions & globals to be declared before they are used.
679 // Loop over the symbol table, emitting all named constants...
680 printSymbolTable(M->getSymbolTable());
682 // Global variable declarations...
684 Out << "\n/* External Global Variable Declarations */\n";
685 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) {
686 if (I->hasExternalLinkage()) {
688 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
694 // Function declarations
696 Out << "\n/* Function Declarations */\n";
698 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
699 // If the function is external and the name collides don't print it.
700 // Sometimes the bytecode likes to have multiple "declarations" for
701 // external functions
702 if ((I->hasInternalLinkage() || !MangledGlobals.count(I)) &&
703 !I->getIntrinsicID()) {
704 printFunctionSignature(I, true);
705 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
711 // Print Malloc prototype if needed
713 Out << "\n/* Malloc to make sun happy */\n";
714 Out << "extern void * malloc();\n\n";
717 // Output the global variable declarations
719 Out << "\n\n/* Global Variable Declarations */\n";
720 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
721 if (!I->isExternal()) {
723 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
725 if (I->hasLinkOnceLinkage())
726 Out << " __attribute__((common))";
727 else if (I->hasWeakLinkage())
728 Out << " __ATTRIBUTE_WEAK__";
733 // Output the global variable definitions and contents...
735 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
736 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
737 if (!I->isExternal()) {
738 if (I->hasInternalLinkage())
740 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
741 if (I->hasLinkOnceLinkage())
742 Out << " __attribute__((common))";
743 else if (I->hasWeakLinkage())
744 Out << " __ATTRIBUTE_WEAK__";
746 // If the initializer is not null, emit the initializer. If it is null,
747 // we try to avoid emitting large amounts of zeros. The problem with
748 // this, however, occurs when the variable has weak linkage. In this
749 // case, the assembler will complain about the variable being both weak
750 // and common, so we disable this optimization.
751 if (!I->getInitializer()->isNullValue() ||
752 I->hasWeakLinkage()) {
754 writeOperand(I->getInitializer());
760 // Output all floating point constants that cannot be printed accurately...
761 printFloatingPointConstants(*M);
763 // Output all of the functions...
764 emittedInvoke = false;
766 Out << "\n\n/* Function Bodies */\n";
767 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
771 // If the program included an invoke instruction, we need to output the
772 // support code for it here!
774 Out << "\n/* More support for the invoke instruction */\n"
775 << "struct __llvm_jmpbuf_list_t *__llvm_jmpbuf_list "
776 << "__attribute__((common)) = 0;\n";
779 // Done with global FP constants
780 FPConstantMap.clear();
783 /// Output all floating point constants that cannot be printed accurately...
784 void CWriter::printFloatingPointConstants(Module &M) {
787 unsigned long long U;
795 // Scan the module for floating point constants. If any FP constant is used
796 // in the function, we want to redirect it here so that we do not depend on
797 // the precision of the printed form, unless the printed form preserves
800 unsigned FPCounter = 0;
801 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
802 for (constant_iterator I = constant_begin(F), E = constant_end(F);
804 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
805 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
806 !FPConstantMap.count(FPC)) {
807 double Val = FPC->getValue();
809 FPConstantMap[FPC] = FPCounter; // Number the FP constants
811 if (FPC->getType() == Type::DoubleTy) {
813 Out << "const ConstantDoubleTy FPConstant" << FPCounter++
814 << " = 0x" << std::hex << DBLUnion.U << std::dec
815 << "ULL; /* " << Val << " */\n";
816 } else if (FPC->getType() == Type::FloatTy) {
818 Out << "const ConstantFloatTy FPConstant" << FPCounter++
819 << " = 0x" << std::hex << FLTUnion.U << std::dec
820 << "U; /* " << Val << " */\n";
822 assert(0 && "Unknown float type!");
829 /// printSymbolTable - Run through symbol table looking for type names. If a
830 /// type name is found, emit it's declaration...
832 void CWriter::printSymbolTable(const SymbolTable &ST) {
833 // If there are no type names, exit early.
834 if (ST.find(Type::TypeTy) == ST.end())
837 // We are only interested in the type plane of the symbol table...
838 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
839 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
841 // Print out forward declarations for structure types before anything else!
842 Out << "/* Structure forward decls */\n";
843 for (; I != End; ++I)
844 if (const Type *STy = dyn_cast<StructType>(I->second))
845 // Only print out used types!
846 if (FUT->getTypes().count(STy)) {
847 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
848 Out << Name << ";\n";
849 TypeNames.insert(std::make_pair(STy, Name));
854 // Now we can print out typedefs...
855 Out << "/* Typedefs */\n";
856 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
857 // Only print out used types!
858 if (FUT->getTypes().count(cast<Type>(I->second))) {
859 const Type *Ty = cast<Type>(I->second);
860 std::string Name = "l_" + Mangler::makeNameProper(I->first);
862 printType(Out, Ty, Name);
868 // Keep track of which structures have been printed so far...
869 std::set<const StructType *> StructPrinted;
871 // Loop over all structures then push them into the stack so they are
872 // printed in the correct order.
874 Out << "/* Structure contents */\n";
875 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
876 if (const StructType *STy = dyn_cast<StructType>(I->second))
877 // Only print out used types!
878 if (FUT->getTypes().count(STy))
879 printContainedStructs(STy, StructPrinted);
882 // Push the struct onto the stack and recursively push all structs
883 // this one depends on.
884 void CWriter::printContainedStructs(const Type *Ty,
885 std::set<const StructType*> &StructPrinted){
886 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
887 //Check to see if we have already printed this struct
888 if (StructPrinted.count(STy) == 0) {
889 // Print all contained types first...
890 for (StructType::element_iterator I = STy->element_begin(),
891 E = STy->element_end(); I != E; ++I) {
892 const Type *Ty1 = I->get();
893 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
894 printContainedStructs(*I, StructPrinted);
897 //Print structure type out..
898 StructPrinted.insert(STy);
899 std::string Name = TypeNames[STy];
900 printType(Out, STy, Name, true);
904 // If it is an array, check contained types and continue
905 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
906 const Type *Ty1 = ATy->getElementType();
907 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
908 printContainedStructs(Ty1, StructPrinted);
913 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
914 // If the program provides its own malloc prototype we don't need
915 // to include the general one.
916 if (Mang->getValueName(F) == "malloc")
919 if (F->hasInternalLinkage()) Out << "static ";
920 if (F->hasLinkOnceLinkage()) Out << "inline ";
922 // Loop over the arguments, printing them...
923 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
925 std::stringstream FunctionInnards;
927 // Print out the name...
928 FunctionInnards << Mang->getValueName(F) << "(";
930 if (!F->isExternal()) {
933 if (F->abegin()->hasName() || !Prototype)
934 ArgName = Mang->getValueName(F->abegin());
935 printType(FunctionInnards, F->afront().getType(), ArgName);
936 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
938 FunctionInnards << ", ";
939 if (I->hasName() || !Prototype)
940 ArgName = Mang->getValueName(I);
943 printType(FunctionInnards, I->getType(), ArgName);
947 // Loop over the arguments, printing them...
948 for (FunctionType::param_iterator I = FT->param_begin(),
949 E = FT->param_end(); I != E; ++I) {
950 if (I != FT->param_begin()) FunctionInnards << ", ";
951 printType(FunctionInnards, *I);
955 // Finish printing arguments... if this is a vararg function, print the ...,
956 // unless there are no known types, in which case, we just emit ().
958 if (FT->isVarArg() && FT->getNumParams()) {
959 if (FT->getNumParams()) FunctionInnards << ", ";
960 FunctionInnards << "..."; // Output varargs portion of signature!
961 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
962 FunctionInnards << "void"; // ret() -> ret(void) in C.
964 FunctionInnards << ")";
965 // Print out the return type and the entire signature for that matter
966 printType(Out, F->getReturnType(), FunctionInnards.str());
969 void CWriter::printFunction(Function *F) {
970 if (F->isExternal()) return;
972 printFunctionSignature(F, false);
975 // Determine whether or not the function contains any invoke instructions.
976 bool HasInvoke = false;
977 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
978 if (isa<InvokeInst>(I->getTerminator())) {
983 // print local variable information for the function
984 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
985 if (const AllocaInst *AI = isDirectAlloca(*I)) {
987 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
988 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
989 Out << "; /* Address exposed local */\n";
990 } else if ((*I)->getType() != Type::VoidTy && !isInlinableInst(**I)) {
992 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
993 printType(Out, (*I)->getType(), Mang->getValueName(*I));
996 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
998 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
999 printType(Out, (*I)->getType(),
1000 Mang->getValueName(*I)+"__PHI_TEMPORARY");
1007 // print the basic blocks
1008 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1009 BasicBlock *Prev = BB->getPrev();
1011 // Don't print the label for the basic block if there are no uses, or if the
1012 // only terminator use is the predecessor basic block's terminator. We have
1013 // to scan the use list because PHI nodes use basic blocks too but do not
1014 // require a label to be generated.
1016 bool NeedsLabel = false;
1017 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
1019 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
1020 if (TI != Prev->getTerminator() ||
1021 isa<SwitchInst>(Prev->getTerminator()) ||
1022 isa<InvokeInst>(Prev->getTerminator())) {
1027 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1029 // Output all of the instructions in the basic block...
1030 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
1031 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1032 if (II->getType() != Type::VoidTy)
1041 // Don't emit prefix or suffix for the terminator...
1042 visit(*BB->getTerminator());
1048 // Specific Instruction type classes... note that all of the casts are
1049 // necessary because we use the instruction classes as opaque types...
1051 void CWriter::visitReturnInst(ReturnInst &I) {
1052 // Don't output a void return if this is the last basic block in the function
1053 if (I.getNumOperands() == 0 &&
1054 &*--I.getParent()->getParent()->end() == I.getParent() &&
1055 !I.getParent()->size() == 1) {
1060 if (I.getNumOperands()) {
1062 writeOperand(I.getOperand(0));
1067 void CWriter::visitSwitchInst(SwitchInst &SI) {
1069 writeOperand(SI.getOperand(0));
1070 Out << ") {\n default:\n";
1071 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1073 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1075 writeOperand(SI.getOperand(i));
1077 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1078 printBranchToBlock(SI.getParent(), Succ, 2);
1079 if (Succ == SI.getParent()->getNext())
1085 void CWriter::visitInvokeInst(InvokeInst &II) {
1087 << " struct __llvm_jmpbuf_list_t Entry;\n"
1088 << " Entry.next = __llvm_jmpbuf_list;\n"
1089 << " if (setjmp(Entry.buf)) {\n"
1090 << " __llvm_jmpbuf_list = Entry.next;\n";
1091 printBranchToBlock(II.getParent(), II.getUnwindDest(), 4);
1093 << " __llvm_jmpbuf_list = &Entry;\n"
1096 if (II.getType() != Type::VoidTy) outputLValue(&II);
1099 << " __llvm_jmpbuf_list = Entry.next;\n"
1101 printBranchToBlock(II.getParent(), II.getNormalDest(), 0);
1102 emittedInvoke = true;
1106 void CWriter::visitUnwindInst(UnwindInst &I) {
1107 // The unwind instructions causes a control flow transfer out of the current
1108 // function, unwinding the stack until a caller who used the invoke
1109 // instruction is found. In this context, we code generated the invoke
1110 // instruction to add an entry to the top of the jmpbuf_list. Thus, here we
1111 // just have to longjmp to the specified handler.
1112 Out << " if (__llvm_jmpbuf_list == 0) { /* unwind */\n"
1114 << " extern signed long long write();\n"
1116 << " extern write();\n"
1118 << " ((void (*)(int, void*, unsigned))write)(2,\n"
1119 << " \"throw found with no handler!\\n\", 31); abort();\n"
1121 << " longjmp(__llvm_jmpbuf_list->buf, 1);\n";
1122 emittedInvoke = true;
1125 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1126 // If PHI nodes need copies, we need the copy code...
1127 if (isa<PHINode>(To->front()) ||
1128 From->getNext() != To) // Not directly successor, need goto
1131 // Otherwise we don't need the code.
1135 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1137 for (BasicBlock::iterator I = Succ->begin();
1138 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1139 // now we have to do the printing
1140 Out << std::string(Indent, ' ');
1141 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1142 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBB)));
1143 Out << "; /* for PHI node */\n";
1146 if (CurBB->getNext() != Succ ||
1147 isa<InvokeInst>(CurBB->getTerminator()) ||
1148 isa<SwitchInst>(CurBB->getTerminator())) {
1149 Out << std::string(Indent, ' ') << " goto ";
1155 // Branch instruction printing - Avoid printing out a branch to a basic block
1156 // that immediately succeeds the current one.
1158 void CWriter::visitBranchInst(BranchInst &I) {
1159 if (I.isConditional()) {
1160 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1162 writeOperand(I.getCondition());
1165 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1167 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1168 Out << " } else {\n";
1169 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1172 // First goto not necessary, assume second one is...
1174 writeOperand(I.getCondition());
1177 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1182 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1187 // PHI nodes get copied into temporary values at the end of predecessor basic
1188 // blocks. We now need to copy these temporary values into the REAL value for
1190 void CWriter::visitPHINode(PHINode &I) {
1192 Out << "__PHI_TEMPORARY";
1196 void CWriter::visitBinaryOperator(Instruction &I) {
1197 // binary instructions, shift instructions, setCond instructions.
1198 assert(!isa<PointerType>(I.getType()));
1200 // We must cast the results of binary operations which might be promoted.
1201 bool needsCast = false;
1202 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1203 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1204 || (I.getType() == Type::FloatTy)) {
1207 printType(Out, I.getType());
1211 writeOperand(I.getOperand(0));
1213 switch (I.getOpcode()) {
1214 case Instruction::Add: Out << " + "; break;
1215 case Instruction::Sub: Out << " - "; break;
1216 case Instruction::Mul: Out << "*"; break;
1217 case Instruction::Div: Out << "/"; break;
1218 case Instruction::Rem: Out << "%"; break;
1219 case Instruction::And: Out << " & "; break;
1220 case Instruction::Or: Out << " | "; break;
1221 case Instruction::Xor: Out << " ^ "; break;
1222 case Instruction::SetEQ: Out << " == "; break;
1223 case Instruction::SetNE: Out << " != "; break;
1224 case Instruction::SetLE: Out << " <= "; break;
1225 case Instruction::SetGE: Out << " >= "; break;
1226 case Instruction::SetLT: Out << " < "; break;
1227 case Instruction::SetGT: Out << " > "; break;
1228 case Instruction::Shl : Out << " << "; break;
1229 case Instruction::Shr : Out << " >> "; break;
1230 default: std::cerr << "Invalid operator type!" << I; abort();
1233 writeOperand(I.getOperand(1));
1240 void CWriter::visitCastInst(CastInst &I) {
1241 if (I.getType() == Type::BoolTy) {
1243 writeOperand(I.getOperand(0));
1248 printType(Out, I.getType());
1250 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1251 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1252 // Avoid "cast to pointer from integer of different size" warnings
1256 writeOperand(I.getOperand(0));
1259 void CWriter::visitCallInst(CallInst &I) {
1260 // Handle intrinsic function calls first...
1261 if (Function *F = I.getCalledFunction())
1262 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1264 default: assert(0 && "Unknown LLVM intrinsic!");
1265 case Intrinsic::va_start:
1268 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1269 // Output the last argument to the enclosing function...
1270 if (I.getParent()->getParent()->aempty()) {
1271 std::cerr << "The C backend does not currently support zero "
1272 << "argument varargs functions, such as '"
1273 << I.getParent()->getParent()->getName() << "'!\n";
1276 writeOperand(&I.getParent()->getParent()->aback());
1279 case Intrinsic::va_end:
1280 Out << "va_end(*(va_list*)&";
1281 writeOperand(I.getOperand(1));
1284 case Intrinsic::va_copy:
1286 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1287 Out << "*(va_list*)&";
1288 writeOperand(I.getOperand(1));
1291 case Intrinsic::setjmp:
1292 case Intrinsic::sigsetjmp:
1293 // This intrinsic should never exist in the program, but until we get
1294 // setjmp/longjmp transformations going on, we should codegen it to
1295 // something reasonable. This will allow code that never calls longjmp
1299 case Intrinsic::longjmp:
1300 case Intrinsic::siglongjmp:
1301 // Longjmp is not implemented, and never will be. It would cause an
1310 void CWriter::visitCallSite(CallSite CS) {
1311 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1312 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1313 const Type *RetTy = FTy->getReturnType();
1315 writeOperand(CS.getCalledValue());
1318 if (CS.arg_begin() != CS.arg_end()) {
1319 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1322 for (++AI; AI != AE; ++AI) {
1330 void CWriter::visitMallocInst(MallocInst &I) {
1332 printType(Out, I.getType());
1333 Out << ")malloc(sizeof(";
1334 printType(Out, I.getType()->getElementType());
1337 if (I.isArrayAllocation()) {
1339 writeOperand(I.getOperand(0));
1344 void CWriter::visitAllocaInst(AllocaInst &I) {
1346 printType(Out, I.getType());
1347 Out << ") alloca(sizeof(";
1348 printType(Out, I.getType()->getElementType());
1350 if (I.isArrayAllocation()) {
1352 writeOperand(I.getOperand(0));
1357 void CWriter::visitFreeInst(FreeInst &I) {
1358 Out << "free((char*)";
1359 writeOperand(I.getOperand(0));
1363 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1364 gep_type_iterator E) {
1365 bool HasImplicitAddress = false;
1366 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1367 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1368 HasImplicitAddress = true;
1369 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1370 HasImplicitAddress = true;
1371 Ptr = CPR->getValue(); // Get to the global...
1372 } else if (isDirectAlloca(Ptr)) {
1373 HasImplicitAddress = true;
1377 if (!HasImplicitAddress)
1378 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1380 writeOperandInternal(Ptr);
1384 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1385 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1388 writeOperandInternal(Ptr);
1390 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1392 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1395 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1396 "Can only have implicit address with direct accessing");
1398 if (HasImplicitAddress) {
1400 } else if (CI && CI->isNullValue()) {
1401 gep_type_iterator TmpI = I; ++TmpI;
1403 // Print out the -> operator if possible...
1404 if (TmpI != E && isa<StructType>(*TmpI)) {
1405 Out << (HasImplicitAddress ? "." : "->");
1406 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1412 if (isa<StructType>(*I)) {
1413 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1416 writeOperand(I.getOperand());
1421 void CWriter::visitLoadInst(LoadInst &I) {
1423 writeOperand(I.getOperand(0));
1426 void CWriter::visitStoreInst(StoreInst &I) {
1428 writeOperand(I.getPointerOperand());
1430 writeOperand(I.getOperand(0));
1433 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1435 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1439 void CWriter::visitVANextInst(VANextInst &I) {
1440 Out << Mang->getValueName(I.getOperand(0));
1441 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1442 printType(Out, I.getArgType());
1446 void CWriter::visitVAArgInst(VAArgInst &I) {
1448 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1449 writeOperand(I.getOperand(0));
1450 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1451 printType(Out, I.getType());
1452 Out << ");\n va_end(Tmp); }";
1457 //===----------------------------------------------------------------------===//
1458 // External Interface declaration
1459 //===----------------------------------------------------------------------===//
1461 Pass *createWriteToCPass(std::ostream &o) { return new CWriter(o); }
1463 } // End llvm namespace