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::ElementTypes::const_iterator
226 I = STy->getElementTypes().begin(),
227 E = STy->getElementTypes().end(); I != E; ++I) {
229 printType(Out, *I, "field" + utostr(Idx++));
235 case Type::PointerTyID: {
236 const PointerType *PTy = cast<PointerType>(Ty);
237 std::string ptrName = "*" + NameSoFar;
239 if (isa<ArrayType>(PTy->getElementType()))
240 ptrName = "(" + ptrName + ")";
242 return printType(Out, PTy->getElementType(), ptrName);
245 case Type::ArrayTyID: {
246 const ArrayType *ATy = cast<ArrayType>(Ty);
247 unsigned NumElements = ATy->getNumElements();
248 return printType(Out, ATy->getElementType(),
249 NameSoFar + "[" + utostr(NumElements) + "]");
252 case Type::OpaqueTyID: {
253 static int Count = 0;
254 std::string TyName = "struct opaque_" + itostr(Count++);
255 assert(TypeNames.find(Ty) == TypeNames.end());
256 TypeNames[Ty] = TyName;
257 return Out << TyName << " " << NameSoFar;
260 assert(0 && "Unhandled case in getTypeProps!");
267 void CWriter::printConstantArray(ConstantArray *CPA) {
269 // As a special case, print the array as a string if it is an array of
270 // ubytes or an array of sbytes with positive values.
272 const Type *ETy = CPA->getType()->getElementType();
273 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
275 // Make sure the last character is a null char, as automatically added by C
276 if (isString && (CPA->getNumOperands() == 0 ||
277 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
282 // Keep track of whether the last number was a hexadecimal escape
283 bool LastWasHex = false;
285 // Do not include the last character, which we know is null
286 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
287 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
289 // Print it out literally if it is a printable character. The only thing
290 // to be careful about is when the last letter output was a hex escape
291 // code, in which case we have to be careful not to print out hex digits
292 // explicitly (the C compiler thinks it is a continuation of the previous
293 // character, sheesh...)
295 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
297 if (C == '"' || C == '\\')
304 case '\n': Out << "\\n"; break;
305 case '\t': Out << "\\t"; break;
306 case '\r': Out << "\\r"; break;
307 case '\v': Out << "\\v"; break;
308 case '\a': Out << "\\a"; break;
309 case '\"': Out << "\\\""; break;
310 case '\'': Out << "\\\'"; break;
313 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
314 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
323 if (CPA->getNumOperands()) {
325 printConstant(cast<Constant>(CPA->getOperand(0)));
326 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
328 printConstant(cast<Constant>(CPA->getOperand(i)));
335 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
336 // textually as a double (rather than as a reference to a stack-allocated
337 // variable). We decide this by converting CFP to a string and back into a
338 // double, and then checking whether the conversion results in a bit-equal
339 // double to the original value of CFP. This depends on us and the target C
340 // compiler agreeing on the conversion process (which is pretty likely since we
341 // only deal in IEEE FP).
343 bool isFPCSafeToPrint(const ConstantFP *CFP) {
346 sprintf(Buffer, "%a", CFP->getValue());
348 if (!strncmp(Buffer, "0x", 2) ||
349 !strncmp(Buffer, "-0x", 3) ||
350 !strncmp(Buffer, "+0x", 3))
351 return atof(Buffer) == CFP->getValue();
354 std::string StrVal = ftostr(CFP->getValue());
356 while (StrVal[0] == ' ')
357 StrVal.erase(StrVal.begin());
359 // Check to make sure that the stringized number is not some string like "Inf"
360 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
361 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
362 ((StrVal[0] == '-' || StrVal[0] == '+') &&
363 (StrVal[1] >= '0' && StrVal[1] <= '9')))
364 // Reparse stringized version!
365 return atof(StrVal.c_str()) == CFP->getValue();
370 // printConstant - The LLVM Constant to C Constant converter.
371 void CWriter::printConstant(Constant *CPV) {
372 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
373 switch (CE->getOpcode()) {
374 case Instruction::Cast:
376 printType(Out, CPV->getType());
378 printConstant(CE->getOperand(0));
382 case Instruction::GetElementPtr:
384 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
388 case Instruction::Add:
389 case Instruction::Sub:
390 case Instruction::Mul:
391 case Instruction::Div:
392 case Instruction::Rem:
393 case Instruction::SetEQ:
394 case Instruction::SetNE:
395 case Instruction::SetLT:
396 case Instruction::SetLE:
397 case Instruction::SetGT:
398 case Instruction::SetGE:
399 case Instruction::Shl:
400 case Instruction::Shr:
402 printConstant(CE->getOperand(0));
403 switch (CE->getOpcode()) {
404 case Instruction::Add: Out << " + "; break;
405 case Instruction::Sub: Out << " - "; break;
406 case Instruction::Mul: Out << " * "; break;
407 case Instruction::Div: Out << " / "; break;
408 case Instruction::Rem: Out << " % "; break;
409 case Instruction::SetEQ: Out << " == "; break;
410 case Instruction::SetNE: Out << " != "; break;
411 case Instruction::SetLT: Out << " < "; break;
412 case Instruction::SetLE: Out << " <= "; break;
413 case Instruction::SetGT: Out << " > "; break;
414 case Instruction::SetGE: Out << " >= "; break;
415 case Instruction::Shl: Out << " << "; break;
416 case Instruction::Shr: Out << " >> "; break;
417 default: assert(0 && "Illegal opcode here!");
419 printConstant(CE->getOperand(1));
424 std::cerr << "CWriter Error: Unhandled constant expression: "
430 switch (CPV->getType()->getPrimitiveID()) {
432 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
433 case Type::SByteTyID:
434 case Type::ShortTyID:
435 Out << cast<ConstantSInt>(CPV)->getValue(); break;
437 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
438 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
440 Out << cast<ConstantSInt>(CPV)->getValue();
444 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
446 case Type::UByteTyID:
447 case Type::UShortTyID:
448 Out << cast<ConstantUInt>(CPV)->getValue(); break;
450 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
451 case Type::ULongTyID:
452 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
454 case Type::FloatTyID:
455 case Type::DoubleTyID: {
456 ConstantFP *FPC = cast<ConstantFP>(CPV);
457 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
458 if (I != FPConstantMap.end()) {
459 // Because of FP precision problems we must load from a stack allocated
460 // value that holds the value in hex.
461 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
462 << "*)&FPConstant" << I->second << ")";
465 // Print out the constant as a floating point number.
467 sprintf(Buffer, "%a", FPC->getValue());
468 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
470 Out << ftostr(FPC->getValue());
476 case Type::ArrayTyID:
477 printConstantArray(cast<ConstantArray>(CPV));
480 case Type::StructTyID: {
482 if (CPV->getNumOperands()) {
484 printConstant(cast<Constant>(CPV->getOperand(0)));
485 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
487 printConstant(cast<Constant>(CPV->getOperand(i)));
494 case Type::PointerTyID:
495 if (isa<ConstantPointerNull>(CPV)) {
497 printType(Out, CPV->getType());
498 Out << ")/*NULL*/0)";
500 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
501 writeOperand(CPR->getValue());
506 std::cerr << "Unknown constant type: " << CPV << "\n";
511 void CWriter::writeOperandInternal(Value *Operand) {
512 if (Instruction *I = dyn_cast<Instruction>(Operand))
513 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
514 // Should we inline this instruction to build a tree?
521 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
524 Out << Mang->getValueName(Operand);
528 void CWriter::writeOperand(Value *Operand) {
529 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
530 Out << "(&"; // Global variables are references as their addresses by llvm
532 writeOperandInternal(Operand);
534 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
538 // nameAllUsedStructureTypes - If there are structure types in the module that
539 // are used but do not have names assigned to them in the symbol table yet then
540 // we assign them names now.
542 bool CWriter::nameAllUsedStructureTypes(Module &M) {
543 // Get a set of types that are used by the program...
544 std::set<const Type *> UT = FUT->getTypes();
546 // Loop over the module symbol table, removing types from UT that are already
549 SymbolTable &MST = M.getSymbolTable();
550 if (MST.find(Type::TypeTy) != MST.end())
551 for (SymbolTable::type_iterator I = MST.type_begin(Type::TypeTy),
552 E = MST.type_end(Type::TypeTy); I != E; ++I)
553 UT.erase(cast<Type>(I->second));
555 // UT now contains types that are not named. Loop over it, naming structure
558 bool Changed = false;
559 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
561 if (const StructType *ST = dyn_cast<StructType>(*I)) {
562 ((Value*)ST)->setName("unnamed", &MST);
568 // generateCompilerSpecificCode - This is where we add conditional compilation
569 // directives to cater to specific compilers as need be.
571 static void generateCompilerSpecificCode(std::ostream& Out) {
572 // Alloca is hard to get, and we don't want to include stdlib.h here...
573 Out << "/* get a declaration for alloca */\n"
575 << "extern void *__builtin_alloca(unsigned long);\n"
576 << "#define alloca(x) __builtin_alloca(x)\n"
578 << "#ifndef __FreeBSD__\n"
579 << "#include <alloca.h>\n"
583 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
584 // If we aren't being compiled with GCC, just drop these attributes.
585 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
586 << "#define __attribute__(X)\n"
590 // At some point, we should support "external weak" vs. "weak" linkages.
591 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
592 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
593 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
594 << "#elif defined(__GNUC__)\n"
595 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
597 << "#define __EXTERNAL_WEAK__\n"
601 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
602 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
603 << "#define __ATTRIBUTE_WEAK__\n"
604 << "#elif defined(__GNUC__)\n"
605 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
607 << "#define __ATTRIBUTE_WEAK__\n"
611 // generateProcessorSpecificCode - This is where we add conditional compilation
612 // directives to cater to specific processors as need be.
614 static void generateProcessorSpecificCode(std::ostream& Out) {
615 // According to ANSI C, longjmp'ing to a setjmp could invalidate any
616 // non-volatile variable in the scope of the setjmp. For now, we are not
617 // doing analysis to determine which variables need to be marked volatile, so
618 // we just mark them all.
620 // HOWEVER, many targets implement setjmp by saving and restoring the register
621 // file, so they DON'T need variables to be marked volatile, and this is a
622 // HUGE pessimization for them. For this reason, on known-good processors, we
623 // do not emit volatile qualifiers.
624 Out << "#if defined(__386__) || defined(__i386__) || \\\n"
625 << " defined(i386) || defined(WIN32)\n"
626 << "/* setjmp does not require variables to be marked volatile */"
627 << "#define VOLATILE_FOR_SETJMP\n"
629 << "#define VOLATILE_FOR_SETJMP volatile\n"
634 void CWriter::printModule(Module *M) {
635 // Calculate which global values have names that will collide when we throw
636 // away type information.
637 { // Scope to delete the FoundNames set when we are done with it...
638 std::set<std::string> FoundNames;
639 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
640 if (I->hasName()) // If the global has a name...
641 if (FoundNames.count(I->getName())) // And the name is already used
642 MangledGlobals.insert(I); // Mangle the name
644 FoundNames.insert(I->getName()); // Otherwise, keep track of name
646 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
647 if (I->hasName()) // If the global has a name...
648 if (FoundNames.count(I->getName())) // And the name is already used
649 MangledGlobals.insert(I); // Mangle the name
651 FoundNames.insert(I->getName()); // Otherwise, keep track of name
654 // get declaration for alloca
655 Out << "/* Provide Declarations */\n";
656 Out << "#include <stdarg.h>\n"; // Varargs support
657 Out << "#include <setjmp.h>\n"; // Unwind support
658 generateCompilerSpecificCode(Out);
659 generateProcessorSpecificCode(Out);
661 // Provide a definition for `bool' if not compiling with a C++ compiler.
663 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
665 << "\n\n/* Support for floating point constants */\n"
666 << "typedef unsigned long long ConstantDoubleTy;\n"
667 << "typedef unsigned int ConstantFloatTy;\n"
669 << "\n\n/* Support for the invoke instruction */\n"
670 << "extern struct __llvm_jmpbuf_list_t {\n"
671 << " jmp_buf buf; struct __llvm_jmpbuf_list_t *next;\n"
672 << "} *__llvm_jmpbuf_list;\n"
674 << "\n\n/* Global Declarations */\n";
676 // First output all the declarations for the program, because C requires
677 // Functions & globals to be declared before they are used.
680 // Loop over the symbol table, emitting all named constants...
681 printSymbolTable(M->getSymbolTable());
683 // Global variable declarations...
685 Out << "\n/* External Global Variable Declarations */\n";
686 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) {
687 if (I->hasExternalLinkage()) {
689 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
695 // Function declarations
697 Out << "\n/* Function Declarations */\n";
699 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
700 // If the function is external and the name collides don't print it.
701 // Sometimes the bytecode likes to have multiple "declarations" for
702 // external functions
703 if ((I->hasInternalLinkage() || !MangledGlobals.count(I)) &&
704 !I->getIntrinsicID()) {
705 printFunctionSignature(I, true);
706 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
712 // Print Malloc prototype if needed
714 Out << "\n/* Malloc to make sun happy */\n";
715 Out << "extern void * malloc();\n\n";
718 // Output the global variable declarations
720 Out << "\n\n/* Global Variable Declarations */\n";
721 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
722 if (!I->isExternal()) {
724 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
726 if (I->hasLinkOnceLinkage())
727 Out << " __attribute__((common))";
728 else if (I->hasWeakLinkage())
729 Out << " __ATTRIBUTE_WEAK__";
734 // Output the global variable definitions and contents...
736 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
737 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
738 if (!I->isExternal()) {
739 if (I->hasInternalLinkage())
741 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
742 if (I->hasLinkOnceLinkage())
743 Out << " __attribute__((common))";
744 else if (I->hasWeakLinkage())
745 Out << " __ATTRIBUTE_WEAK__";
747 // If the initializer is not null, emit the initializer. If it is null,
748 // we try to avoid emitting large amounts of zeros. The problem with
749 // this, however, occurs when the variable has weak linkage. In this
750 // case, the assembler will complain about the variable being both weak
751 // and common, so we disable this optimization.
752 if (!I->getInitializer()->isNullValue() ||
753 I->hasWeakLinkage()) {
755 writeOperand(I->getInitializer());
761 // Output all floating point constants that cannot be printed accurately...
762 printFloatingPointConstants(*M);
764 // Output all of the functions...
765 emittedInvoke = false;
767 Out << "\n\n/* Function Bodies */\n";
768 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
772 // If the program included an invoke instruction, we need to output the
773 // support code for it here!
775 Out << "\n/* More support for the invoke instruction */\n"
776 << "struct __llvm_jmpbuf_list_t *__llvm_jmpbuf_list "
777 << "__attribute__((common)) = 0;\n";
780 // Done with global FP constants
781 FPConstantMap.clear();
784 /// Output all floating point constants that cannot be printed accurately...
785 void CWriter::printFloatingPointConstants(Module &M) {
788 unsigned long long U;
796 // Scan the module for floating point constants. If any FP constant is used
797 // in the function, we want to redirect it here so that we do not depend on
798 // the precision of the printed form, unless the printed form preserves
801 unsigned FPCounter = 0;
802 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
803 for (constant_iterator I = constant_begin(F), E = constant_end(F);
805 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
806 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
807 !FPConstantMap.count(FPC)) {
808 double Val = FPC->getValue();
810 FPConstantMap[FPC] = FPCounter; // Number the FP constants
812 if (FPC->getType() == Type::DoubleTy) {
814 Out << "const ConstantDoubleTy FPConstant" << FPCounter++
815 << " = 0x" << std::hex << DBLUnion.U << std::dec
816 << "ULL; /* " << Val << " */\n";
817 } else if (FPC->getType() == Type::FloatTy) {
819 Out << "const ConstantFloatTy FPConstant" << FPCounter++
820 << " = 0x" << std::hex << FLTUnion.U << std::dec
821 << "U; /* " << Val << " */\n";
823 assert(0 && "Unknown float type!");
830 /// printSymbolTable - Run through symbol table looking for type names. If a
831 /// type name is found, emit it's declaration...
833 void CWriter::printSymbolTable(const SymbolTable &ST) {
834 // If there are no type names, exit early.
835 if (ST.find(Type::TypeTy) == ST.end())
838 // We are only interested in the type plane of the symbol table...
839 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
840 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
842 // Print out forward declarations for structure types before anything else!
843 Out << "/* Structure forward decls */\n";
844 for (; I != End; ++I)
845 if (const Type *STy = dyn_cast<StructType>(I->second))
846 // Only print out used types!
847 if (FUT->getTypes().count(STy)) {
848 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
849 Out << Name << ";\n";
850 TypeNames.insert(std::make_pair(STy, Name));
855 // Now we can print out typedefs...
856 Out << "/* Typedefs */\n";
857 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
858 // Only print out used types!
859 if (FUT->getTypes().count(cast<Type>(I->second))) {
860 const Type *Ty = cast<Type>(I->second);
861 std::string Name = "l_" + Mangler::makeNameProper(I->first);
863 printType(Out, Ty, Name);
869 // Keep track of which structures have been printed so far...
870 std::set<const StructType *> StructPrinted;
872 // Loop over all structures then push them into the stack so they are
873 // printed in the correct order.
875 Out << "/* Structure contents */\n";
876 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
877 if (const StructType *STy = dyn_cast<StructType>(I->second))
878 // Only print out used types!
879 if (FUT->getTypes().count(STy))
880 printContainedStructs(STy, StructPrinted);
883 // Push the struct onto the stack and recursively push all structs
884 // this one depends on.
885 void CWriter::printContainedStructs(const Type *Ty,
886 std::set<const StructType*> &StructPrinted){
887 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
888 //Check to see if we have already printed this struct
889 if (StructPrinted.count(STy) == 0) {
890 // Print all contained types first...
891 for (StructType::ElementTypes::const_iterator
892 I = STy->getElementTypes().begin(),
893 E = STy->getElementTypes().end(); I != E; ++I) {
894 const Type *Ty1 = I->get();
895 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
896 printContainedStructs(*I, StructPrinted);
899 //Print structure type out..
900 StructPrinted.insert(STy);
901 std::string Name = TypeNames[STy];
902 printType(Out, STy, Name, true);
906 // If it is an array, check contained types and continue
907 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
908 const Type *Ty1 = ATy->getElementType();
909 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
910 printContainedStructs(Ty1, StructPrinted);
915 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
916 // If the program provides its own malloc prototype we don't need
917 // to include the general one.
918 if (Mang->getValueName(F) == "malloc")
921 if (F->hasInternalLinkage()) Out << "static ";
922 if (F->hasLinkOnceLinkage()) Out << "inline ";
924 // Loop over the arguments, printing them...
925 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
927 std::stringstream FunctionInnards;
929 // Print out the name...
930 FunctionInnards << Mang->getValueName(F) << "(";
932 if (!F->isExternal()) {
935 if (F->abegin()->hasName() || !Prototype)
936 ArgName = Mang->getValueName(F->abegin());
937 printType(FunctionInnards, F->afront().getType(), ArgName);
938 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
940 FunctionInnards << ", ";
941 if (I->hasName() || !Prototype)
942 ArgName = Mang->getValueName(I);
945 printType(FunctionInnards, I->getType(), ArgName);
949 // Loop over the arguments, printing them...
950 for (FunctionType::param_iterator I = FT->param_begin(),
951 E = FT->param_end(); I != E; ++I) {
952 if (I != FT->param_begin()) FunctionInnards << ", ";
953 printType(FunctionInnards, *I);
957 // Finish printing arguments... if this is a vararg function, print the ...,
958 // unless there are no known types, in which case, we just emit ().
960 if (FT->isVarArg() && FT->getNumParams()) {
961 if (FT->getNumParams()) FunctionInnards << ", ";
962 FunctionInnards << "..."; // Output varargs portion of signature!
963 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
964 FunctionInnards << "void"; // ret() -> ret(void) in C.
966 FunctionInnards << ")";
967 // Print out the return type and the entire signature for that matter
968 printType(Out, F->getReturnType(), FunctionInnards.str());
971 void CWriter::printFunction(Function *F) {
972 if (F->isExternal()) return;
974 printFunctionSignature(F, false);
977 // Determine whether or not the function contains any invoke instructions.
978 bool HasInvoke = false;
979 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
980 if (isa<InvokeInst>(I->getTerminator())) {
985 // print local variable information for the function
986 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
987 if (const AllocaInst *AI = isDirectAlloca(*I)) {
989 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
990 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
991 Out << "; /* Address exposed local */\n";
992 } else if ((*I)->getType() != Type::VoidTy && !isInlinableInst(**I)) {
994 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
995 printType(Out, (*I)->getType(), Mang->getValueName(*I));
998 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
1000 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
1001 printType(Out, (*I)->getType(),
1002 Mang->getValueName(*I)+"__PHI_TEMPORARY");
1009 // print the basic blocks
1010 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1011 BasicBlock *Prev = BB->getPrev();
1013 // Don't print the label for the basic block if there are no uses, or if the
1014 // only terminator use is the predecessor basic block's terminator. We have
1015 // to scan the use list because PHI nodes use basic blocks too but do not
1016 // require a label to be generated.
1018 bool NeedsLabel = false;
1019 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
1021 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
1022 if (TI != Prev->getTerminator() ||
1023 isa<SwitchInst>(Prev->getTerminator()) ||
1024 isa<InvokeInst>(Prev->getTerminator())) {
1029 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1031 // Output all of the instructions in the basic block...
1032 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
1033 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1034 if (II->getType() != Type::VoidTy)
1043 // Don't emit prefix or suffix for the terminator...
1044 visit(*BB->getTerminator());
1050 // Specific Instruction type classes... note that all of the casts are
1051 // necessary because we use the instruction classes as opaque types...
1053 void CWriter::visitReturnInst(ReturnInst &I) {
1054 // Don't output a void return if this is the last basic block in the function
1055 if (I.getNumOperands() == 0 &&
1056 &*--I.getParent()->getParent()->end() == I.getParent() &&
1057 !I.getParent()->size() == 1) {
1062 if (I.getNumOperands()) {
1064 writeOperand(I.getOperand(0));
1069 void CWriter::visitSwitchInst(SwitchInst &SI) {
1071 writeOperand(SI.getOperand(0));
1072 Out << ") {\n default:\n";
1073 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1075 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1077 writeOperand(SI.getOperand(i));
1079 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1080 printBranchToBlock(SI.getParent(), Succ, 2);
1081 if (Succ == SI.getParent()->getNext())
1087 void CWriter::visitInvokeInst(InvokeInst &II) {
1089 << " struct __llvm_jmpbuf_list_t Entry;\n"
1090 << " Entry.next = __llvm_jmpbuf_list;\n"
1091 << " if (setjmp(Entry.buf)) {\n"
1092 << " __llvm_jmpbuf_list = Entry.next;\n";
1093 printBranchToBlock(II.getParent(), II.getUnwindDest(), 4);
1095 << " __llvm_jmpbuf_list = &Entry;\n"
1098 if (II.getType() != Type::VoidTy) outputLValue(&II);
1101 << " __llvm_jmpbuf_list = Entry.next;\n"
1103 printBranchToBlock(II.getParent(), II.getNormalDest(), 0);
1104 emittedInvoke = true;
1108 void CWriter::visitUnwindInst(UnwindInst &I) {
1109 // The unwind instructions causes a control flow transfer out of the current
1110 // function, unwinding the stack until a caller who used the invoke
1111 // instruction is found. In this context, we code generated the invoke
1112 // instruction to add an entry to the top of the jmpbuf_list. Thus, here we
1113 // just have to longjmp to the specified handler.
1114 Out << " if (__llvm_jmpbuf_list == 0) { /* unwind */\n"
1116 << " extern signed long long write();\n"
1118 << " extern write();\n"
1120 << " ((void (*)(int, void*, unsigned))write)(2,\n"
1121 << " \"throw found with no handler!\\n\", 31); abort();\n"
1123 << " longjmp(__llvm_jmpbuf_list->buf, 1);\n";
1124 emittedInvoke = true;
1127 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1128 // If PHI nodes need copies, we need the copy code...
1129 if (isa<PHINode>(To->front()) ||
1130 From->getNext() != To) // Not directly successor, need goto
1133 // Otherwise we don't need the code.
1137 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1139 for (BasicBlock::iterator I = Succ->begin();
1140 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1141 // now we have to do the printing
1142 Out << std::string(Indent, ' ');
1143 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1144 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBB)));
1145 Out << "; /* for PHI node */\n";
1148 if (CurBB->getNext() != Succ ||
1149 isa<InvokeInst>(CurBB->getTerminator()) ||
1150 isa<SwitchInst>(CurBB->getTerminator())) {
1151 Out << std::string(Indent, ' ') << " goto ";
1157 // Branch instruction printing - Avoid printing out a branch to a basic block
1158 // that immediately succeeds the current one.
1160 void CWriter::visitBranchInst(BranchInst &I) {
1161 if (I.isConditional()) {
1162 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1164 writeOperand(I.getCondition());
1167 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1169 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1170 Out << " } else {\n";
1171 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1174 // First goto not necessary, assume second one is...
1176 writeOperand(I.getCondition());
1179 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1184 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1189 // PHI nodes get copied into temporary values at the end of predecessor basic
1190 // blocks. We now need to copy these temporary values into the REAL value for
1192 void CWriter::visitPHINode(PHINode &I) {
1194 Out << "__PHI_TEMPORARY";
1198 void CWriter::visitBinaryOperator(Instruction &I) {
1199 // binary instructions, shift instructions, setCond instructions.
1200 assert(!isa<PointerType>(I.getType()));
1202 // We must cast the results of binary operations which might be promoted.
1203 bool needsCast = false;
1204 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1205 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1206 || (I.getType() == Type::FloatTy)) {
1209 printType(Out, I.getType());
1213 writeOperand(I.getOperand(0));
1215 switch (I.getOpcode()) {
1216 case Instruction::Add: Out << " + "; break;
1217 case Instruction::Sub: Out << " - "; break;
1218 case Instruction::Mul: Out << "*"; break;
1219 case Instruction::Div: Out << "/"; break;
1220 case Instruction::Rem: Out << "%"; break;
1221 case Instruction::And: Out << " & "; break;
1222 case Instruction::Or: Out << " | "; break;
1223 case Instruction::Xor: Out << " ^ "; break;
1224 case Instruction::SetEQ: Out << " == "; break;
1225 case Instruction::SetNE: Out << " != "; break;
1226 case Instruction::SetLE: Out << " <= "; break;
1227 case Instruction::SetGE: Out << " >= "; break;
1228 case Instruction::SetLT: Out << " < "; break;
1229 case Instruction::SetGT: Out << " > "; break;
1230 case Instruction::Shl : Out << " << "; break;
1231 case Instruction::Shr : Out << " >> "; break;
1232 default: std::cerr << "Invalid operator type!" << I; abort();
1235 writeOperand(I.getOperand(1));
1242 void CWriter::visitCastInst(CastInst &I) {
1243 if (I.getType() == Type::BoolTy) {
1245 writeOperand(I.getOperand(0));
1250 printType(Out, I.getType());
1252 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1253 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1254 // Avoid "cast to pointer from integer of different size" warnings
1258 writeOperand(I.getOperand(0));
1261 void CWriter::visitCallInst(CallInst &I) {
1262 // Handle intrinsic function calls first...
1263 if (Function *F = I.getCalledFunction())
1264 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1266 default: assert(0 && "Unknown LLVM intrinsic!");
1267 case Intrinsic::va_start:
1270 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1271 // Output the last argument to the enclosing function...
1272 if (I.getParent()->getParent()->aempty()) {
1273 std::cerr << "The C backend does not currently support zero "
1274 << "argument varargs functions, such as '"
1275 << I.getParent()->getParent()->getName() << "'!\n";
1278 writeOperand(&I.getParent()->getParent()->aback());
1281 case Intrinsic::va_end:
1282 Out << "va_end(*(va_list*)&";
1283 writeOperand(I.getOperand(1));
1286 case Intrinsic::va_copy:
1288 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1289 Out << "*(va_list*)&";
1290 writeOperand(I.getOperand(1));
1293 case Intrinsic::setjmp:
1294 case Intrinsic::sigsetjmp:
1295 // This intrinsic should never exist in the program, but until we get
1296 // setjmp/longjmp transformations going on, we should codegen it to
1297 // something reasonable. This will allow code that never calls longjmp
1301 case Intrinsic::longjmp:
1302 case Intrinsic::siglongjmp:
1303 // Longjmp is not implemented, and never will be. It would cause an
1312 void CWriter::visitCallSite(CallSite CS) {
1313 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1314 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1315 const Type *RetTy = FTy->getReturnType();
1317 writeOperand(CS.getCalledValue());
1320 if (CS.arg_begin() != CS.arg_end()) {
1321 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1324 for (++AI; AI != AE; ++AI) {
1332 void CWriter::visitMallocInst(MallocInst &I) {
1334 printType(Out, I.getType());
1335 Out << ")malloc(sizeof(";
1336 printType(Out, I.getType()->getElementType());
1339 if (I.isArrayAllocation()) {
1341 writeOperand(I.getOperand(0));
1346 void CWriter::visitAllocaInst(AllocaInst &I) {
1348 printType(Out, I.getType());
1349 Out << ") alloca(sizeof(";
1350 printType(Out, I.getType()->getElementType());
1352 if (I.isArrayAllocation()) {
1354 writeOperand(I.getOperand(0));
1359 void CWriter::visitFreeInst(FreeInst &I) {
1360 Out << "free((char*)";
1361 writeOperand(I.getOperand(0));
1365 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1366 gep_type_iterator E) {
1367 bool HasImplicitAddress = false;
1368 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1369 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1370 HasImplicitAddress = true;
1371 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1372 HasImplicitAddress = true;
1373 Ptr = CPR->getValue(); // Get to the global...
1374 } else if (isDirectAlloca(Ptr)) {
1375 HasImplicitAddress = true;
1379 if (!HasImplicitAddress)
1380 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1382 writeOperandInternal(Ptr);
1386 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1387 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1390 writeOperandInternal(Ptr);
1392 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1394 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1397 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1398 "Can only have implicit address with direct accessing");
1400 if (HasImplicitAddress) {
1402 } else if (CI && CI->isNullValue()) {
1403 gep_type_iterator TmpI = I; ++TmpI;
1405 // Print out the -> operator if possible...
1406 if (TmpI != E && isa<StructType>(*TmpI)) {
1407 Out << (HasImplicitAddress ? "." : "->");
1408 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1414 if (isa<StructType>(*I)) {
1415 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1418 writeOperand(I.getOperand());
1423 void CWriter::visitLoadInst(LoadInst &I) {
1425 writeOperand(I.getOperand(0));
1428 void CWriter::visitStoreInst(StoreInst &I) {
1430 writeOperand(I.getPointerOperand());
1432 writeOperand(I.getOperand(0));
1435 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1437 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1441 void CWriter::visitVANextInst(VANextInst &I) {
1442 Out << Mang->getValueName(I.getOperand(0));
1443 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1444 printType(Out, I.getArgType());
1448 void CWriter::visitVAArgInst(VAArgInst &I) {
1450 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1451 writeOperand(I.getOperand(0));
1452 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1453 printType(Out, I.getType());
1454 Out << ");\n va_end(Tmp); }";
1459 //===----------------------------------------------------------------------===//
1460 // External Interface declaration
1461 //===----------------------------------------------------------------------===//
1463 Pass *createWriteToCPass(std::ostream &o) { return new CWriter(o); }
1465 } // End llvm namespace