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::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 if (isa<ArrayType>(PTy->getElementType()))
241 ptrName = "(" + ptrName + ")";
243 return printType(Out, PTy->getElementType(), ptrName);
246 case Type::ArrayTyID: {
247 const ArrayType *ATy = cast<ArrayType>(Ty);
248 unsigned NumElements = ATy->getNumElements();
249 return printType(Out, ATy->getElementType(),
250 NameSoFar + "[" + utostr(NumElements) + "]");
253 case Type::OpaqueTyID: {
254 static int Count = 0;
255 std::string TyName = "struct opaque_" + itostr(Count++);
256 assert(TypeNames.find(Ty) == TypeNames.end());
257 TypeNames[Ty] = TyName;
258 return Out << TyName << " " << NameSoFar;
261 assert(0 && "Unhandled case in getTypeProps!");
268 void CWriter::printConstantArray(ConstantArray *CPA) {
270 // As a special case, print the array as a string if it is an array of
271 // ubytes or an array of sbytes with positive values.
273 const Type *ETy = CPA->getType()->getElementType();
274 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
276 // Make sure the last character is a null char, as automatically added by C
277 if (isString && (CPA->getNumOperands() == 0 ||
278 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
283 // Keep track of whether the last number was a hexadecimal escape
284 bool LastWasHex = false;
286 // Do not include the last character, which we know is null
287 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
288 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
290 // Print it out literally if it is a printable character. The only thing
291 // to be careful about is when the last letter output was a hex escape
292 // code, in which case we have to be careful not to print out hex digits
293 // explicitly (the C compiler thinks it is a continuation of the previous
294 // character, sheesh...)
296 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
298 if (C == '"' || C == '\\')
305 case '\n': Out << "\\n"; break;
306 case '\t': Out << "\\t"; break;
307 case '\r': Out << "\\r"; break;
308 case '\v': Out << "\\v"; break;
309 case '\a': Out << "\\a"; break;
310 case '\"': Out << "\\\""; break;
311 case '\'': Out << "\\\'"; break;
314 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
315 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
324 if (CPA->getNumOperands()) {
326 printConstant(cast<Constant>(CPA->getOperand(0)));
327 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
329 printConstant(cast<Constant>(CPA->getOperand(i)));
336 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
337 // textually as a double (rather than as a reference to a stack-allocated
338 // variable). We decide this by converting CFP to a string and back into a
339 // double, and then checking whether the conversion results in a bit-equal
340 // double to the original value of CFP. This depends on us and the target C
341 // compiler agreeing on the conversion process (which is pretty likely since we
342 // only deal in IEEE FP).
344 bool isFPCSafeToPrint(const ConstantFP *CFP) {
347 sprintf(Buffer, "%a", CFP->getValue());
349 if (!strncmp(Buffer, "0x", 2) ||
350 !strncmp(Buffer, "-0x", 3) ||
351 !strncmp(Buffer, "+0x", 3))
352 return atof(Buffer) == CFP->getValue();
355 std::string StrVal = ftostr(CFP->getValue());
357 while (StrVal[0] == ' ')
358 StrVal.erase(StrVal.begin());
360 // Check to make sure that the stringized number is not some string like "Inf"
361 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
362 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
363 ((StrVal[0] == '-' || StrVal[0] == '+') &&
364 (StrVal[1] >= '0' && StrVal[1] <= '9')))
365 // Reparse stringized version!
366 return atof(StrVal.c_str()) == CFP->getValue();
371 // printConstant - The LLVM Constant to C Constant converter.
372 void CWriter::printConstant(Constant *CPV) {
373 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
374 switch (CE->getOpcode()) {
375 case Instruction::Cast:
377 printType(Out, CPV->getType());
379 printConstant(CE->getOperand(0));
383 case Instruction::GetElementPtr:
385 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
389 case Instruction::Add:
390 case Instruction::Sub:
391 case Instruction::Mul:
392 case Instruction::Div:
393 case Instruction::Rem:
394 case Instruction::SetEQ:
395 case Instruction::SetNE:
396 case Instruction::SetLT:
397 case Instruction::SetLE:
398 case Instruction::SetGT:
399 case Instruction::SetGE:
400 case Instruction::Shl:
401 case Instruction::Shr:
403 printConstant(CE->getOperand(0));
404 switch (CE->getOpcode()) {
405 case Instruction::Add: Out << " + "; break;
406 case Instruction::Sub: Out << " - "; break;
407 case Instruction::Mul: Out << " * "; break;
408 case Instruction::Div: Out << " / "; break;
409 case Instruction::Rem: Out << " % "; break;
410 case Instruction::SetEQ: Out << " == "; break;
411 case Instruction::SetNE: Out << " != "; break;
412 case Instruction::SetLT: Out << " < "; break;
413 case Instruction::SetLE: Out << " <= "; break;
414 case Instruction::SetGT: Out << " > "; break;
415 case Instruction::SetGE: Out << " >= "; break;
416 case Instruction::Shl: Out << " << "; break;
417 case Instruction::Shr: 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 = FUT->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 // generateProcessorSpecificCode - This is where we add conditional compilation
592 // directives to cater to specific processors as need be.
594 static void generateProcessorSpecificCode(std::ostream& Out) {
595 // According to ANSI C, longjmp'ing to a setjmp could invalidate any
596 // non-volatile variable in the scope of the setjmp. For now, we are not
597 // doing analysis to determine which variables need to be marked volatile, so
598 // we just mark them all.
600 // HOWEVER, many targets implement setjmp by saving and restoring the register
601 // file, so they DON'T need variables to be marked volatile, and this is a
602 // HUGE pessimization for them. For this reason, on known-good processors, we
603 // do not emit volatile qualifiers.
604 Out << "#if defined(__386__) || defined(__i386__) || \\\n"
605 << " defined(i386) || defined(WIN32)\n"
606 << "/* setjmp does not require variables to be marked volatile */"
607 << "#define VOLATILE_FOR_SETJMP\n"
609 << "#define VOLATILE_FOR_SETJMP volatile\n"
614 void CWriter::printModule(Module *M) {
615 // Calculate which global values have names that will collide when we throw
616 // away type information.
617 { // Scope to delete the FoundNames set when we are done with it...
618 std::set<std::string> FoundNames;
619 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
620 if (I->hasName()) // If the global has a name...
621 if (FoundNames.count(I->getName())) // And the name is already used
622 MangledGlobals.insert(I); // Mangle the name
624 FoundNames.insert(I->getName()); // Otherwise, keep track of name
626 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
627 if (I->hasName()) // If the global has a name...
628 if (FoundNames.count(I->getName())) // And the name is already used
629 MangledGlobals.insert(I); // Mangle the name
631 FoundNames.insert(I->getName()); // Otherwise, keep track of name
634 // get declaration for alloca
635 Out << "/* Provide Declarations */\n";
636 Out << "#include <stdarg.h>\n"; // Varargs support
637 Out << "#include <setjmp.h>\n"; // Unwind support
638 generateCompilerSpecificCode(Out);
639 generateProcessorSpecificCode(Out);
641 // Provide a definition for `bool' if not compiling with a C++ compiler.
643 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
645 << "\n\n/* Support for floating point constants */\n"
646 << "typedef unsigned long long ConstantDoubleTy;\n"
647 << "typedef unsigned int ConstantFloatTy;\n"
649 << "\n\n/* Support for the invoke instruction */\n"
650 << "extern struct __llvm_jmpbuf_list_t {\n"
651 << " jmp_buf buf; struct __llvm_jmpbuf_list_t *next;\n"
652 << "} *__llvm_jmpbuf_list;\n"
654 << "\n\n/* Global Declarations */\n";
656 // First output all the declarations for the program, because C requires
657 // Functions & globals to be declared before they are used.
660 // Loop over the symbol table, emitting all named constants...
661 printSymbolTable(M->getSymbolTable());
663 // Global variable declarations...
665 Out << "\n/* External Global Variable Declarations */\n";
666 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) {
667 if (I->hasExternalLinkage()) {
669 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
675 // Function declarations
677 Out << "\n/* Function Declarations */\n";
679 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
680 // If the function is external and the name collides don't print it.
681 // Sometimes the bytecode likes to have multiple "declarations" for
682 // external functions
683 if ((I->hasInternalLinkage() || !MangledGlobals.count(I)) &&
684 !I->getIntrinsicID()) {
685 printFunctionSignature(I, true);
686 if (I->hasWeakLinkage()) Out << " __attribute__((weak))";
692 // Print Malloc prototype if needed
694 Out << "\n/* Malloc to make sun happy */\n";
695 Out << "extern void * malloc();\n\n";
698 // Output the global variable declarations
700 Out << "\n\n/* Global Variable Declarations */\n";
701 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
702 if (!I->isExternal()) {
704 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
706 if (I->hasLinkOnceLinkage())
707 Out << " __attribute__((common))";
708 else if (I->hasWeakLinkage())
709 Out << " __attribute__((weak))";
714 // Output the global variable definitions and contents...
716 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
717 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
718 if (!I->isExternal()) {
719 if (I->hasInternalLinkage())
721 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
722 if (I->hasLinkOnceLinkage())
723 Out << " __attribute__((common))";
724 else if (I->hasWeakLinkage())
725 Out << " __attribute__((weak))";
727 // If the initializer is not null, emit the initializer. If it is null,
728 // we try to avoid emitting large amounts of zeros. The problem with
729 // this, however, occurs when the variable has weak linkage. In this
730 // case, the assembler will complain about the variable being both weak
731 // and common, so we disable this optimization.
732 if (!I->getInitializer()->isNullValue() ||
733 I->hasWeakLinkage()) {
735 writeOperand(I->getInitializer());
741 // Output all floating point constants that cannot be printed accurately...
742 printFloatingPointConstants(*M);
744 // Output all of the functions...
745 emittedInvoke = false;
747 Out << "\n\n/* Function Bodies */\n";
748 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
752 // If the program included an invoke instruction, we need to output the
753 // support code for it here!
755 Out << "\n/* More support for the invoke instruction */\n"
756 << "struct __llvm_jmpbuf_list_t *__llvm_jmpbuf_list "
757 << "__attribute__((common)) = 0;\n";
760 // Done with global FP constants
761 FPConstantMap.clear();
764 /// Output all floating point constants that cannot be printed accurately...
765 void CWriter::printFloatingPointConstants(Module &M) {
768 unsigned long long U;
776 // Scan the module for floating point constants. If any FP constant is used
777 // in the function, we want to redirect it here so that we do not depend on
778 // the precision of the printed form, unless the printed form preserves
781 unsigned FPCounter = 0;
782 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
783 for (constant_iterator I = constant_begin(F), E = constant_end(F);
785 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
786 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
787 !FPConstantMap.count(FPC)) {
788 double Val = FPC->getValue();
790 FPConstantMap[FPC] = FPCounter; // Number the FP constants
792 if (FPC->getType() == Type::DoubleTy) {
794 Out << "const ConstantDoubleTy FPConstant" << FPCounter++
795 << " = 0x" << std::hex << DBLUnion.U << std::dec
796 << "ULL; /* " << Val << " */\n";
797 } else if (FPC->getType() == Type::FloatTy) {
799 Out << "const ConstantFloatTy FPConstant" << FPCounter++
800 << " = 0x" << std::hex << FLTUnion.U << std::dec
801 << "U; /* " << Val << " */\n";
803 assert(0 && "Unknown float type!");
810 /// printSymbolTable - Run through symbol table looking for type names. If a
811 /// type name is found, emit it's declaration...
813 void CWriter::printSymbolTable(const SymbolTable &ST) {
814 // If there are no type names, exit early.
815 if (ST.find(Type::TypeTy) == ST.end())
818 // We are only interested in the type plane of the symbol table...
819 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
820 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
822 // Print out forward declarations for structure types before anything else!
823 Out << "/* Structure forward decls */\n";
824 for (; I != End; ++I)
825 if (const Type *STy = dyn_cast<StructType>(I->second))
826 // Only print out used types!
827 if (FUT->getTypes().count(STy)) {
828 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
829 Out << Name << ";\n";
830 TypeNames.insert(std::make_pair(STy, Name));
835 // Now we can print out typedefs...
836 Out << "/* Typedefs */\n";
837 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
838 // Only print out used types!
839 if (FUT->getTypes().count(cast<Type>(I->second))) {
840 const Type *Ty = cast<Type>(I->second);
841 std::string Name = "l_" + Mangler::makeNameProper(I->first);
843 printType(Out, Ty, Name);
849 // Keep track of which structures have been printed so far...
850 std::set<const StructType *> StructPrinted;
852 // Loop over all structures then push them into the stack so they are
853 // printed in the correct order.
855 Out << "/* Structure contents */\n";
856 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
857 if (const StructType *STy = dyn_cast<StructType>(I->second))
858 // Only print out used types!
859 if (FUT->getTypes().count(STy))
860 printContainedStructs(STy, StructPrinted);
863 // Push the struct onto the stack and recursively push all structs
864 // this one depends on.
865 void CWriter::printContainedStructs(const Type *Ty,
866 std::set<const StructType*> &StructPrinted){
867 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
868 //Check to see if we have already printed this struct
869 if (StructPrinted.count(STy) == 0) {
870 // Print all contained types first...
871 for (StructType::ElementTypes::const_iterator
872 I = STy->getElementTypes().begin(),
873 E = STy->getElementTypes().end(); I != E; ++I) {
874 const Type *Ty1 = I->get();
875 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
876 printContainedStructs(*I, StructPrinted);
879 //Print structure type out..
880 StructPrinted.insert(STy);
881 std::string Name = TypeNames[STy];
882 printType(Out, STy, Name, true);
886 // If it is an array, check contained types and continue
887 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
888 const Type *Ty1 = ATy->getElementType();
889 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
890 printContainedStructs(Ty1, StructPrinted);
895 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
896 // If the program provides its own malloc prototype we don't need
897 // to include the general one.
898 if (Mang->getValueName(F) == "malloc")
901 if (F->hasInternalLinkage()) Out << "static ";
902 if (F->hasLinkOnceLinkage()) Out << "inline ";
904 // Loop over the arguments, printing them...
905 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
907 std::stringstream FunctionInnards;
909 // Print out the name...
910 FunctionInnards << Mang->getValueName(F) << "(";
912 if (!F->isExternal()) {
915 if (F->abegin()->hasName() || !Prototype)
916 ArgName = Mang->getValueName(F->abegin());
917 printType(FunctionInnards, F->afront().getType(), ArgName);
918 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
920 FunctionInnards << ", ";
921 if (I->hasName() || !Prototype)
922 ArgName = Mang->getValueName(I);
925 printType(FunctionInnards, I->getType(), ArgName);
929 // Loop over the arguments, printing them...
930 for (FunctionType::ParamTypes::const_iterator I =
931 FT->getParamTypes().begin(),
932 E = FT->getParamTypes().end(); I != E; ++I) {
933 if (I != FT->getParamTypes().begin()) FunctionInnards << ", ";
934 printType(FunctionInnards, *I);
938 // Finish printing arguments... if this is a vararg function, print the ...,
939 // unless there are no known types, in which case, we just emit ().
941 if (FT->isVarArg() && !FT->getParamTypes().empty()) {
942 if (FT->getParamTypes().size()) FunctionInnards << ", ";
943 FunctionInnards << "..."; // Output varargs portion of signature!
945 FunctionInnards << ")";
946 // Print out the return type and the entire signature for that matter
947 printType(Out, F->getReturnType(), FunctionInnards.str());
950 void CWriter::printFunction(Function *F) {
951 if (F->isExternal()) return;
953 printFunctionSignature(F, false);
956 // Determine whether or not the function contains any invoke instructions.
957 bool HasInvoke = false;
958 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
959 if (isa<InvokeInst>(I->getTerminator())) {
964 // print local variable information for the function
965 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
966 if (const AllocaInst *AI = isDirectAlloca(*I)) {
968 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
969 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
970 Out << "; /* Address exposed local */\n";
971 } else if ((*I)->getType() != Type::VoidTy && !isInlinableInst(**I)) {
973 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
974 printType(Out, (*I)->getType(), Mang->getValueName(*I));
977 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
979 if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
980 printType(Out, (*I)->getType(),
981 Mang->getValueName(*I)+"__PHI_TEMPORARY");
988 // print the basic blocks
989 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
990 BasicBlock *Prev = BB->getPrev();
992 // Don't print the label for the basic block if there are no uses, or if the
993 // only terminator use is the predecessor basic block's terminator. We have
994 // to scan the use list because PHI nodes use basic blocks too but do not
995 // require a label to be generated.
997 bool NeedsLabel = false;
998 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
1000 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
1001 if (TI != Prev->getTerminator() ||
1002 isa<SwitchInst>(Prev->getTerminator()) ||
1003 isa<InvokeInst>(Prev->getTerminator())) {
1008 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1010 // Output all of the instructions in the basic block...
1011 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
1012 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1013 if (II->getType() != Type::VoidTy)
1022 // Don't emit prefix or suffix for the terminator...
1023 visit(*BB->getTerminator());
1029 // Specific Instruction type classes... note that all of the casts are
1030 // necessary because we use the instruction classes as opaque types...
1032 void CWriter::visitReturnInst(ReturnInst &I) {
1033 // Don't output a void return if this is the last basic block in the function
1034 if (I.getNumOperands() == 0 &&
1035 &*--I.getParent()->getParent()->end() == I.getParent() &&
1036 !I.getParent()->size() == 1) {
1041 if (I.getNumOperands()) {
1043 writeOperand(I.getOperand(0));
1048 void CWriter::visitSwitchInst(SwitchInst &SI) {
1050 writeOperand(SI.getOperand(0));
1051 Out << ") {\n default:\n";
1052 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1054 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1056 writeOperand(SI.getOperand(i));
1058 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1059 printBranchToBlock(SI.getParent(), Succ, 2);
1060 if (Succ == SI.getParent()->getNext())
1066 void CWriter::visitInvokeInst(InvokeInst &II) {
1068 << " struct __llvm_jmpbuf_list_t Entry;\n"
1069 << " Entry.next = __llvm_jmpbuf_list;\n"
1070 << " if (setjmp(Entry.buf)) {\n"
1071 << " __llvm_jmpbuf_list = Entry.next;\n";
1072 printBranchToBlock(II.getParent(), II.getExceptionalDest(), 4);
1074 << " __llvm_jmpbuf_list = &Entry;\n"
1077 if (II.getType() != Type::VoidTy) outputLValue(&II);
1080 << " __llvm_jmpbuf_list = Entry.next;\n"
1082 printBranchToBlock(II.getParent(), II.getNormalDest(), 0);
1083 emittedInvoke = true;
1087 void CWriter::visitUnwindInst(UnwindInst &I) {
1088 // The unwind instructions causes a control flow transfer out of the current
1089 // function, unwinding the stack until a caller who used the invoke
1090 // instruction is found. In this context, we code generated the invoke
1091 // instruction to add an entry to the top of the jmpbuf_list. Thus, here we
1092 // just have to longjmp to the specified handler.
1093 Out << " if (__llvm_jmpbuf_list == 0) { /* unwind */\n"
1094 << " extern write();\n"
1095 << " ((void (*)(int, void*, unsigned))write)(2,\n"
1096 << " \"throw found with no handler!\\n\", 31); abort();\n"
1098 << " longjmp(__llvm_jmpbuf_list->buf, 1);\n";
1099 emittedInvoke = true;
1102 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1103 // If PHI nodes need copies, we need the copy code...
1104 if (isa<PHINode>(To->front()) ||
1105 From->getNext() != To) // Not directly successor, need goto
1108 // Otherwise we don't need the code.
1112 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1114 for (BasicBlock::iterator I = Succ->begin();
1115 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1116 // now we have to do the printing
1117 Out << std::string(Indent, ' ');
1118 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1119 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBB)));
1120 Out << "; /* for PHI node */\n";
1123 if (CurBB->getNext() != Succ ||
1124 isa<InvokeInst>(CurBB->getTerminator()) ||
1125 isa<SwitchInst>(CurBB->getTerminator())) {
1126 Out << std::string(Indent, ' ') << " goto ";
1132 // Branch instruction printing - Avoid printing out a branch to a basic block
1133 // that immediately succeeds the current one.
1135 void CWriter::visitBranchInst(BranchInst &I) {
1136 if (I.isConditional()) {
1137 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1139 writeOperand(I.getCondition());
1142 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1144 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1145 Out << " } else {\n";
1146 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1149 // First goto not necessary, assume second one is...
1151 writeOperand(I.getCondition());
1154 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1159 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1164 // PHI nodes get copied into temporary values at the end of predecessor basic
1165 // blocks. We now need to copy these temporary values into the REAL value for
1167 void CWriter::visitPHINode(PHINode &I) {
1169 Out << "__PHI_TEMPORARY";
1173 void CWriter::visitBinaryOperator(Instruction &I) {
1174 // binary instructions, shift instructions, setCond instructions.
1175 assert(!isa<PointerType>(I.getType()));
1177 // We must cast the results of binary operations which might be promoted.
1178 bool needsCast = false;
1179 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1180 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1181 || (I.getType() == Type::FloatTy)) {
1184 printType(Out, I.getType());
1188 writeOperand(I.getOperand(0));
1190 switch (I.getOpcode()) {
1191 case Instruction::Add: Out << " + "; break;
1192 case Instruction::Sub: Out << " - "; break;
1193 case Instruction::Mul: Out << "*"; break;
1194 case Instruction::Div: Out << "/"; break;
1195 case Instruction::Rem: Out << "%"; break;
1196 case Instruction::And: Out << " & "; break;
1197 case Instruction::Or: Out << " | "; break;
1198 case Instruction::Xor: Out << " ^ "; break;
1199 case Instruction::SetEQ: Out << " == "; break;
1200 case Instruction::SetNE: Out << " != "; break;
1201 case Instruction::SetLE: Out << " <= "; break;
1202 case Instruction::SetGE: Out << " >= "; break;
1203 case Instruction::SetLT: Out << " < "; break;
1204 case Instruction::SetGT: Out << " > "; break;
1205 case Instruction::Shl : Out << " << "; break;
1206 case Instruction::Shr : Out << " >> "; break;
1207 default: std::cerr << "Invalid operator type!" << I; abort();
1210 writeOperand(I.getOperand(1));
1217 void CWriter::visitCastInst(CastInst &I) {
1218 if (I.getType() == Type::BoolTy) {
1220 writeOperand(I.getOperand(0));
1225 printType(Out, I.getType());
1227 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1228 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1229 // Avoid "cast to pointer from integer of different size" warnings
1233 writeOperand(I.getOperand(0));
1236 void CWriter::visitCallInst(CallInst &I) {
1237 // Handle intrinsic function calls first...
1238 if (Function *F = I.getCalledFunction())
1239 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1241 default: assert(0 && "Unknown LLVM intrinsic!");
1242 case Intrinsic::va_start:
1245 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1246 // Output the last argument to the enclosing function...
1247 if (I.getParent()->getParent()->aempty()) {
1248 std::cerr << "The C backend does not currently support zero "
1249 << "argument varargs functions, such as '"
1250 << I.getParent()->getParent()->getName() << "'!\n";
1253 writeOperand(&I.getParent()->getParent()->aback());
1256 case Intrinsic::va_end:
1257 Out << "va_end(*(va_list*)&";
1258 writeOperand(I.getOperand(1));
1261 case Intrinsic::va_copy:
1263 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1264 Out << "*(va_list*)&";
1265 writeOperand(I.getOperand(1));
1268 case Intrinsic::setjmp:
1269 case Intrinsic::sigsetjmp:
1270 // This intrinsic should never exist in the program, but until we get
1271 // setjmp/longjmp transformations going on, we should codegen it to
1272 // something reasonable. This will allow code that never calls longjmp
1276 case Intrinsic::longjmp:
1277 case Intrinsic::siglongjmp:
1278 // Longjmp is not implemented, and never will be. It would cause an
1287 void CWriter::visitCallSite(CallSite CS) {
1288 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1289 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1290 const Type *RetTy = FTy->getReturnType();
1292 writeOperand(CS.getCalledValue());
1295 if (CS.arg_begin() != CS.arg_end()) {
1296 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1299 for (++AI; AI != AE; ++AI) {
1307 void CWriter::visitMallocInst(MallocInst &I) {
1309 printType(Out, I.getType());
1310 Out << ")malloc(sizeof(";
1311 printType(Out, I.getType()->getElementType());
1314 if (I.isArrayAllocation()) {
1316 writeOperand(I.getOperand(0));
1321 void CWriter::visitAllocaInst(AllocaInst &I) {
1323 printType(Out, I.getType());
1324 Out << ") alloca(sizeof(";
1325 printType(Out, I.getType()->getElementType());
1327 if (I.isArrayAllocation()) {
1329 writeOperand(I.getOperand(0));
1334 void CWriter::visitFreeInst(FreeInst &I) {
1335 Out << "free((char*)";
1336 writeOperand(I.getOperand(0));
1340 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1341 gep_type_iterator E) {
1342 bool HasImplicitAddress = false;
1343 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1344 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1345 HasImplicitAddress = true;
1346 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1347 HasImplicitAddress = true;
1348 Ptr = CPR->getValue(); // Get to the global...
1349 } else if (isDirectAlloca(Ptr)) {
1350 HasImplicitAddress = true;
1354 if (!HasImplicitAddress)
1355 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1357 writeOperandInternal(Ptr);
1361 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1362 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1365 writeOperandInternal(Ptr);
1367 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1369 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1372 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1373 "Can only have implicit address with direct accessing");
1375 if (HasImplicitAddress) {
1377 } else if (CI && CI->isNullValue()) {
1378 gep_type_iterator TmpI = I; ++TmpI;
1380 // Print out the -> operator if possible...
1381 if (TmpI != E && isa<StructType>(*TmpI)) {
1382 Out << (HasImplicitAddress ? "." : "->");
1383 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1389 if (isa<StructType>(*I)) {
1390 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1393 writeOperand(I.getOperand());
1398 void CWriter::visitLoadInst(LoadInst &I) {
1400 writeOperand(I.getOperand(0));
1403 void CWriter::visitStoreInst(StoreInst &I) {
1405 writeOperand(I.getPointerOperand());
1407 writeOperand(I.getOperand(0));
1410 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1412 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1416 void CWriter::visitVANextInst(VANextInst &I) {
1417 Out << Mang->getValueName(I.getOperand(0));
1418 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1419 printType(Out, I.getArgType());
1423 void CWriter::visitVAArgInst(VAArgInst &I) {
1425 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1426 writeOperand(I.getOperand(0));
1427 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1428 printType(Out, I.getType());
1429 Out << ");\n va_end(Tmp); }";
1434 //===----------------------------------------------------------------------===//
1435 // External Interface declaration
1436 //===----------------------------------------------------------------------===//
1438 Pass *createWriteToCPass(std::ostream &o) { return new CWriter(o); }
1440 } // End llvm namespace