1 //===-- Writer.cpp - Library for converting LLVM code to C ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This library converts LLVM code to C code, compilable by GCC and other C
13 //===----------------------------------------------------------------------===//
15 #include "CTargetMachine.h"
16 #include "llvm/Target/TargetMachineImpls.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Module.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/Pass.h"
22 #include "llvm/PassManager.h"
23 #include "llvm/SymbolTable.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/IntrinsicLowering.h"
26 #include "llvm/Analysis/FindUsedTypes.h"
27 #include "llvm/Analysis/ConstantsScanner.h"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/InstVisitor.h"
32 #include "llvm/Support/Mangler.h"
33 #include "Support/StringExtras.h"
39 class CWriter : public Pass, public InstVisitor<CWriter> {
41 IntrinsicLowering &IL;
43 const Module *TheModule;
46 std::map<const Type *, std::string> TypeNames;
48 std::map<const ConstantFP *, unsigned> FPConstantMap;
50 CWriter(std::ostream &o, IntrinsicLowering &il) : Out(o), IL(il) {}
52 void getAnalysisUsage(AnalysisUsage &AU) const {
53 AU.addRequired<FindUsedTypes>();
56 virtual const char *getPassName() const { return "C backend"; }
58 bool doInitialization(Module &M);
60 // First pass, lower all unhandled intrinsics.
65 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
75 std::ostream &printType(std::ostream &Out, const Type *Ty,
76 const std::string &VariableName = "",
77 bool IgnoreName = false);
79 void writeOperand(Value *Operand);
80 void writeOperandInternal(Value *Operand);
83 void lowerIntrinsics(Module &M);
85 bool nameAllUsedStructureTypes(Module &M);
86 void printModule(Module *M);
87 void printFloatingPointConstants(Module &M);
88 void printSymbolTable(const SymbolTable &ST);
89 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
90 void printFunctionSignature(const Function *F, bool Prototype);
92 void printFunction(Function &);
94 void printConstant(Constant *CPV);
95 void printConstantArray(ConstantArray *CPA);
97 // isInlinableInst - Attempt to inline instructions into their uses to build
98 // trees as much as possible. To do this, we have to consistently decide
99 // what is acceptable to inline, so that variable declarations don't get
100 // printed and an extra copy of the expr is not emitted.
102 static bool isInlinableInst(const Instruction &I) {
103 // Must be an expression, must be used exactly once. If it is dead, we
104 // emit it inline where it would go.
105 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
106 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
107 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
108 // Don't inline a load across a store or other bad things!
111 // Only inline instruction it it's use is in the same BB as the inst.
112 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
115 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
116 // variables which are accessed with the & operator. This causes GCC to
117 // generate significantly better code than to emit alloca calls directly.
119 static const AllocaInst *isDirectAlloca(const Value *V) {
120 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
121 if (!AI) return false;
122 if (AI->isArrayAllocation())
123 return 0; // FIXME: we can also inline fixed size array allocas!
124 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
129 // Instruction visitation functions
130 friend class InstVisitor<CWriter>;
132 void visitReturnInst(ReturnInst &I);
133 void visitBranchInst(BranchInst &I);
134 void visitSwitchInst(SwitchInst &I);
135 void visitInvokeInst(InvokeInst &I);
136 void visitUnwindInst(UnwindInst &I);
138 void visitPHINode(PHINode &I);
139 void visitBinaryOperator(Instruction &I);
141 void visitCastInst (CastInst &I);
142 void visitCallInst (CallInst &I);
143 void visitCallSite (CallSite CS);
144 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
146 void visitMallocInst(MallocInst &I);
147 void visitAllocaInst(AllocaInst &I);
148 void visitFreeInst (FreeInst &I);
149 void visitLoadInst (LoadInst &I);
150 void visitStoreInst (StoreInst &I);
151 void visitGetElementPtrInst(GetElementPtrInst &I);
152 void visitVANextInst(VANextInst &I);
153 void visitVAArgInst (VAArgInst &I);
155 void visitInstruction(Instruction &I) {
156 std::cerr << "C Writer does not know about " << I;
160 void outputLValue(Instruction *I) {
161 Out << " " << Mang->getValueName(I) << " = ";
163 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
165 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
166 gep_type_iterator E);
170 // Pass the Type* and the variable name and this prints out the variable
173 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
174 const std::string &NameSoFar,
176 if (Ty->isPrimitiveType())
177 switch (Ty->getPrimitiveID()) {
178 case Type::VoidTyID: return Out << "void " << NameSoFar;
179 case Type::BoolTyID: return Out << "bool " << NameSoFar;
180 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
181 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
182 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
183 case Type::ShortTyID: return Out << "short " << NameSoFar;
184 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
185 case Type::IntTyID: return Out << "int " << NameSoFar;
186 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
187 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
188 case Type::FloatTyID: return Out << "float " << NameSoFar;
189 case Type::DoubleTyID: return Out << "double " << NameSoFar;
191 std::cerr << "Unknown primitive type: " << Ty << "\n";
195 // Check to see if the type is named.
196 if (!IgnoreName || isa<OpaqueType>(Ty)) {
197 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
198 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
201 switch (Ty->getPrimitiveID()) {
202 case Type::FunctionTyID: {
203 const FunctionType *MTy = cast<FunctionType>(Ty);
204 std::stringstream FunctionInnards;
205 FunctionInnards << " (" << NameSoFar << ") (";
206 for (FunctionType::param_iterator I = MTy->param_begin(),
207 E = MTy->param_end(); I != E; ++I) {
208 if (I != MTy->param_begin())
209 FunctionInnards << ", ";
210 printType(FunctionInnards, *I, "");
212 if (MTy->isVarArg()) {
213 if (MTy->getNumParams())
214 FunctionInnards << ", ...";
215 } else if (!MTy->getNumParams()) {
216 FunctionInnards << "void";
218 FunctionInnards << ")";
219 std::string tstr = FunctionInnards.str();
220 printType(Out, MTy->getReturnType(), tstr);
223 case Type::StructTyID: {
224 const StructType *STy = cast<StructType>(Ty);
225 Out << NameSoFar + " {\n";
227 for (StructType::element_iterator I = STy->element_begin(),
228 E = STy->element_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 if (isa<ConstantAggregateZero>(CPV)) {
479 const ArrayType *AT = cast<ArrayType>(CPV->getType());
481 if (AT->getNumElements()) {
483 Constant *CZ = Constant::getNullValue(AT->getElementType());
485 for (unsigned i = 1, e = AT->getNumElements(); i != e; ++i) {
492 printConstantArray(cast<ConstantArray>(CPV));
496 case Type::StructTyID:
497 if (isa<ConstantAggregateZero>(CPV)) {
498 const StructType *ST = cast<StructType>(CPV->getType());
500 if (ST->getNumElements()) {
502 printConstant(Constant::getNullValue(ST->getElementType(0)));
503 for (unsigned i = 1, e = ST->getNumElements(); i != e; ++i) {
505 printConstant(Constant::getNullValue(ST->getElementType(i)));
511 if (CPV->getNumOperands()) {
513 printConstant(cast<Constant>(CPV->getOperand(0)));
514 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
516 printConstant(cast<Constant>(CPV->getOperand(i)));
523 case Type::PointerTyID:
524 if (isa<ConstantPointerNull>(CPV)) {
526 printType(Out, CPV->getType());
527 Out << ")/*NULL*/0)";
529 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
530 writeOperand(CPR->getValue());
535 std::cerr << "Unknown constant type: " << CPV << "\n";
540 void CWriter::writeOperandInternal(Value *Operand) {
541 if (Instruction *I = dyn_cast<Instruction>(Operand))
542 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
543 // Should we inline this instruction to build a tree?
550 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
553 Out << Mang->getValueName(Operand);
557 void CWriter::writeOperand(Value *Operand) {
558 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
559 Out << "(&"; // Global variables are references as their addresses by llvm
561 writeOperandInternal(Operand);
563 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
567 // nameAllUsedStructureTypes - If there are structure types in the module that
568 // are used but do not have names assigned to them in the symbol table yet then
569 // we assign them names now.
571 bool CWriter::nameAllUsedStructureTypes(Module &M) {
572 // Get a set of types that are used by the program...
573 std::set<const Type *> UT = FUT->getTypes();
575 // Loop over the module symbol table, removing types from UT that are already
578 SymbolTable &MST = M.getSymbolTable();
579 if (MST.find(Type::TypeTy) != MST.end())
580 for (SymbolTable::type_iterator I = MST.type_begin(Type::TypeTy),
581 E = MST.type_end(Type::TypeTy); I != E; ++I)
582 UT.erase(cast<Type>(I->second));
584 // UT now contains types that are not named. Loop over it, naming structure
587 bool Changed = false;
588 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
590 if (const StructType *ST = dyn_cast<StructType>(*I)) {
591 ((Value*)ST)->setName("unnamed", &MST);
597 // generateCompilerSpecificCode - This is where we add conditional compilation
598 // directives to cater to specific compilers as need be.
600 static void generateCompilerSpecificCode(std::ostream& Out) {
601 // Alloca is hard to get, and we don't want to include stdlib.h here...
602 Out << "/* get a declaration for alloca */\n"
604 << "extern void *__builtin_alloca(unsigned long);\n"
605 << "#define alloca(x) __builtin_alloca(x)\n"
607 << "#ifndef __FreeBSD__\n"
608 << "#include <alloca.h>\n"
612 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
613 // If we aren't being compiled with GCC, just drop these attributes.
614 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
615 << "#define __attribute__(X)\n"
619 // At some point, we should support "external weak" vs. "weak" linkages.
620 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
621 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
622 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
623 << "#elif defined(__GNUC__)\n"
624 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
626 << "#define __EXTERNAL_WEAK__\n"
630 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
631 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
632 << "#define __ATTRIBUTE_WEAK__\n"
633 << "#elif defined(__GNUC__)\n"
634 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
636 << "#define __ATTRIBUTE_WEAK__\n"
640 bool CWriter::doInitialization(Module &M) {
643 FUT = &getAnalysis<FindUsedTypes>();
645 // Ensure that all structure types have names...
646 bool Changed = nameAllUsedStructureTypes(M);
647 Mang = new Mangler(M);
649 // get declaration for alloca
650 Out << "/* Provide Declarations */\n";
651 Out << "#include <stdarg.h>\n"; // Varargs support
652 Out << "#include <setjmp.h>\n"; // Unwind support
653 generateCompilerSpecificCode(Out);
655 // Provide a definition for `bool' if not compiling with a C++ compiler.
657 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
659 << "\n\n/* Support for floating point constants */\n"
660 << "typedef unsigned long long ConstantDoubleTy;\n"
661 << "typedef unsigned int ConstantFloatTy;\n"
663 << "\n\n/* Global Declarations */\n";
665 // First output all the declarations for the program, because C requires
666 // Functions & globals to be declared before they are used.
669 // Loop over the symbol table, emitting all named constants...
670 printSymbolTable(M.getSymbolTable());
672 // Global variable declarations...
674 Out << "\n/* External Global Variable Declarations */\n";
675 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I) {
676 if (I->hasExternalLinkage()) {
678 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
684 // Function declarations
686 Out << "\n/* Function Declarations */\n";
687 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
688 // Don't print declarations for intrinsic functions.
689 if (!I->getIntrinsicID()) {
690 printFunctionSignature(I, true);
691 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
697 // Output the global variable declarations
699 Out << "\n\n/* Global Variable Declarations */\n";
700 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
701 if (!I->isExternal()) {
703 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
705 if (I->hasLinkOnceLinkage())
706 Out << " __attribute__((common))";
707 else if (I->hasWeakLinkage())
708 Out << " __ATTRIBUTE_WEAK__";
713 // Output the global variable definitions and contents...
715 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
716 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
717 if (!I->isExternal()) {
718 if (I->hasInternalLinkage())
720 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
721 if (I->hasLinkOnceLinkage())
722 Out << " __attribute__((common))";
723 else if (I->hasWeakLinkage())
724 Out << " __ATTRIBUTE_WEAK__";
726 // If the initializer is not null, emit the initializer. If it is null,
727 // we try to avoid emitting large amounts of zeros. The problem with
728 // this, however, occurs when the variable has weak linkage. In this
729 // case, the assembler will complain about the variable being both weak
730 // and common, so we disable this optimization.
731 if (!I->getInitializer()->isNullValue()) {
733 writeOperand(I->getInitializer());
734 } else if (I->hasWeakLinkage()) {
735 // We have to specify an initializer, but it doesn't have to be
736 // complete. If the value is an aggregate, print out { 0 }, and let
737 // the compiler figure out the rest of the zeros.
739 if (isa<StructType>(I->getInitializer()->getType()) ||
740 isa<ArrayType>(I->getInitializer()->getType())) {
743 // Just print it out normally.
744 writeOperand(I->getInitializer());
751 // Output all floating point constants that cannot be printed accurately...
752 printFloatingPointConstants(M);
755 Out << "\n\n/* Function Bodies */\n";
760 /// Output all floating point constants that cannot be printed accurately...
761 void CWriter::printFloatingPointConstants(Module &M) {
764 unsigned long long U;
772 // Scan the module for floating point constants. If any FP constant is used
773 // in the function, we want to redirect it here so that we do not depend on
774 // the precision of the printed form, unless the printed form preserves
777 unsigned FPCounter = 0;
778 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
779 for (constant_iterator I = constant_begin(F), E = constant_end(F);
781 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
782 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
783 !FPConstantMap.count(FPC)) {
784 double Val = FPC->getValue();
786 FPConstantMap[FPC] = FPCounter; // Number the FP constants
788 if (FPC->getType() == Type::DoubleTy) {
790 Out << "const ConstantDoubleTy FPConstant" << FPCounter++
791 << " = 0x" << std::hex << DBLUnion.U << std::dec
792 << "ULL; /* " << Val << " */\n";
793 } else if (FPC->getType() == Type::FloatTy) {
795 Out << "const ConstantFloatTy FPConstant" << FPCounter++
796 << " = 0x" << std::hex << FLTUnion.U << std::dec
797 << "U; /* " << Val << " */\n";
799 assert(0 && "Unknown float type!");
806 /// printSymbolTable - Run through symbol table looking for type names. If a
807 /// type name is found, emit it's declaration...
809 void CWriter::printSymbolTable(const SymbolTable &ST) {
810 // If there are no type names, exit early.
811 if (ST.find(Type::TypeTy) == ST.end())
814 // We are only interested in the type plane of the symbol table...
815 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
816 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
818 // Print out forward declarations for structure types before anything else!
819 Out << "/* Structure forward decls */\n";
820 for (; I != End; ++I)
821 if (const Type *STy = dyn_cast<StructType>(I->second))
822 // Only print out used types!
823 if (FUT->getTypes().count(STy)) {
824 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
825 Out << Name << ";\n";
826 TypeNames.insert(std::make_pair(STy, Name));
831 // Now we can print out typedefs...
832 Out << "/* Typedefs */\n";
833 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
834 // Only print out used types!
835 if (FUT->getTypes().count(cast<Type>(I->second))) {
836 const Type *Ty = cast<Type>(I->second);
837 std::string Name = "l_" + Mangler::makeNameProper(I->first);
839 printType(Out, Ty, Name);
845 // Keep track of which structures have been printed so far...
846 std::set<const StructType *> StructPrinted;
848 // Loop over all structures then push them into the stack so they are
849 // printed in the correct order.
851 Out << "/* Structure contents */\n";
852 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
853 if (const StructType *STy = dyn_cast<StructType>(I->second))
854 // Only print out used types!
855 if (FUT->getTypes().count(STy))
856 printContainedStructs(STy, StructPrinted);
859 // Push the struct onto the stack and recursively push all structs
860 // this one depends on.
861 void CWriter::printContainedStructs(const Type *Ty,
862 std::set<const StructType*> &StructPrinted){
863 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
864 //Check to see if we have already printed this struct
865 if (StructPrinted.count(STy) == 0) {
866 // Print all contained types first...
867 for (StructType::element_iterator I = STy->element_begin(),
868 E = STy->element_end(); I != E; ++I) {
869 const Type *Ty1 = I->get();
870 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
871 printContainedStructs(*I, StructPrinted);
874 //Print structure type out..
875 StructPrinted.insert(STy);
876 std::string Name = TypeNames[STy];
877 printType(Out, STy, Name, true);
881 // If it is an array, check contained types and continue
882 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
883 const Type *Ty1 = ATy->getElementType();
884 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
885 printContainedStructs(Ty1, StructPrinted);
890 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
891 if (F->hasInternalLinkage()) Out << "static ";
892 if (F->hasLinkOnceLinkage()) Out << "inline ";
894 // Loop over the arguments, printing them...
895 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
897 std::stringstream FunctionInnards;
899 // Print out the name...
900 FunctionInnards << Mang->getValueName(F) << "(";
902 if (!F->isExternal()) {
905 if (F->abegin()->hasName() || !Prototype)
906 ArgName = Mang->getValueName(F->abegin());
907 printType(FunctionInnards, F->afront().getType(), ArgName);
908 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
910 FunctionInnards << ", ";
911 if (I->hasName() || !Prototype)
912 ArgName = Mang->getValueName(I);
915 printType(FunctionInnards, I->getType(), ArgName);
919 // Loop over the arguments, printing them...
920 for (FunctionType::param_iterator I = FT->param_begin(),
921 E = FT->param_end(); I != E; ++I) {
922 if (I != FT->param_begin()) FunctionInnards << ", ";
923 printType(FunctionInnards, *I);
927 // Finish printing arguments... if this is a vararg function, print the ...,
928 // unless there are no known types, in which case, we just emit ().
930 if (FT->isVarArg() && FT->getNumParams()) {
931 if (FT->getNumParams()) FunctionInnards << ", ";
932 FunctionInnards << "..."; // Output varargs portion of signature!
933 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
934 FunctionInnards << "void"; // ret() -> ret(void) in C.
936 FunctionInnards << ")";
937 // Print out the return type and the entire signature for that matter
938 printType(Out, F->getReturnType(), FunctionInnards.str());
941 void CWriter::printFunction(Function &F) {
942 printFunctionSignature(&F, false);
945 // print local variable information for the function
946 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I)
947 if (const AllocaInst *AI = isDirectAlloca(*I)) {
949 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
950 Out << "; /* Address exposed local */\n";
951 } else if ((*I)->getType() != Type::VoidTy && !isInlinableInst(**I)) {
953 printType(Out, (*I)->getType(), Mang->getValueName(*I));
956 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
958 printType(Out, (*I)->getType(),
959 Mang->getValueName(*I)+"__PHI_TEMPORARY");
966 // print the basic blocks
967 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
968 BasicBlock *Prev = BB->getPrev();
970 // Don't print the label for the basic block if there are no uses, or if the
971 // only terminator use is the predecessor basic block's terminator. We have
972 // to scan the use list because PHI nodes use basic blocks too but do not
973 // require a label to be generated.
975 bool NeedsLabel = false;
976 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
978 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
979 if (TI != Prev->getTerminator() ||
980 isa<SwitchInst>(Prev->getTerminator()) ||
981 isa<InvokeInst>(Prev->getTerminator())) {
986 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
988 // Output all of the instructions in the basic block...
989 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
990 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
991 if (II->getType() != Type::VoidTy)
1000 // Don't emit prefix or suffix for the terminator...
1001 visit(*BB->getTerminator());
1007 // Specific Instruction type classes... note that all of the casts are
1008 // necessary because we use the instruction classes as opaque types...
1010 void CWriter::visitReturnInst(ReturnInst &I) {
1011 // Don't output a void return if this is the last basic block in the function
1012 if (I.getNumOperands() == 0 &&
1013 &*--I.getParent()->getParent()->end() == I.getParent() &&
1014 !I.getParent()->size() == 1) {
1019 if (I.getNumOperands()) {
1021 writeOperand(I.getOperand(0));
1026 void CWriter::visitSwitchInst(SwitchInst &SI) {
1028 writeOperand(SI.getOperand(0));
1029 Out << ") {\n default:\n";
1030 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1032 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1034 writeOperand(SI.getOperand(i));
1036 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1037 printBranchToBlock(SI.getParent(), Succ, 2);
1038 if (Succ == SI.getParent()->getNext())
1044 void CWriter::visitInvokeInst(InvokeInst &II) {
1045 assert(0 && "Lowerinvoke pass didn't work!");
1049 void CWriter::visitUnwindInst(UnwindInst &I) {
1050 assert(0 && "Lowerinvoke pass didn't work!");
1053 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1054 // If PHI nodes need copies, we need the copy code...
1055 if (isa<PHINode>(To->front()) ||
1056 From->getNext() != To) // Not directly successor, need goto
1059 // Otherwise we don't need the code.
1063 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1065 for (BasicBlock::iterator I = Succ->begin();
1066 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1067 // now we have to do the printing
1068 Out << std::string(Indent, ' ');
1069 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1070 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBB)));
1071 Out << "; /* for PHI node */\n";
1074 if (CurBB->getNext() != Succ ||
1075 isa<InvokeInst>(CurBB->getTerminator()) ||
1076 isa<SwitchInst>(CurBB->getTerminator())) {
1077 Out << std::string(Indent, ' ') << " goto ";
1083 // Branch instruction printing - Avoid printing out a branch to a basic block
1084 // that immediately succeeds the current one.
1086 void CWriter::visitBranchInst(BranchInst &I) {
1087 if (I.isConditional()) {
1088 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1090 writeOperand(I.getCondition());
1093 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1095 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1096 Out << " } else {\n";
1097 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1100 // First goto not necessary, assume second one is...
1102 writeOperand(I.getCondition());
1105 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1110 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1115 // PHI nodes get copied into temporary values at the end of predecessor basic
1116 // blocks. We now need to copy these temporary values into the REAL value for
1118 void CWriter::visitPHINode(PHINode &I) {
1120 Out << "__PHI_TEMPORARY";
1124 void CWriter::visitBinaryOperator(Instruction &I) {
1125 // binary instructions, shift instructions, setCond instructions.
1126 assert(!isa<PointerType>(I.getType()));
1128 // We must cast the results of binary operations which might be promoted.
1129 bool needsCast = false;
1130 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1131 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1132 || (I.getType() == Type::FloatTy)) {
1135 printType(Out, I.getType());
1139 writeOperand(I.getOperand(0));
1141 switch (I.getOpcode()) {
1142 case Instruction::Add: Out << " + "; break;
1143 case Instruction::Sub: Out << " - "; break;
1144 case Instruction::Mul: Out << "*"; break;
1145 case Instruction::Div: Out << "/"; break;
1146 case Instruction::Rem: Out << "%"; break;
1147 case Instruction::And: Out << " & "; break;
1148 case Instruction::Or: Out << " | "; break;
1149 case Instruction::Xor: Out << " ^ "; break;
1150 case Instruction::SetEQ: Out << " == "; break;
1151 case Instruction::SetNE: Out << " != "; break;
1152 case Instruction::SetLE: Out << " <= "; break;
1153 case Instruction::SetGE: Out << " >= "; break;
1154 case Instruction::SetLT: Out << " < "; break;
1155 case Instruction::SetGT: Out << " > "; break;
1156 case Instruction::Shl : Out << " << "; break;
1157 case Instruction::Shr : Out << " >> "; break;
1158 default: std::cerr << "Invalid operator type!" << I; abort();
1161 writeOperand(I.getOperand(1));
1168 void CWriter::visitCastInst(CastInst &I) {
1169 if (I.getType() == Type::BoolTy) {
1171 writeOperand(I.getOperand(0));
1176 printType(Out, I.getType());
1178 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1179 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1180 // Avoid "cast to pointer from integer of different size" warnings
1184 writeOperand(I.getOperand(0));
1187 void CWriter::lowerIntrinsics(Module &M) {
1188 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1189 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1190 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1191 if (CallInst *CI = dyn_cast<CallInst>(I++))
1192 if (Function *F = CI->getCalledFunction())
1193 switch (F->getIntrinsicID()) {
1194 case Intrinsic::not_intrinsic:
1195 case Intrinsic::va_start:
1196 case Intrinsic::va_copy:
1197 case Intrinsic::va_end:
1198 case Intrinsic::returnaddress:
1199 case Intrinsic::frameaddress:
1200 case Intrinsic::setjmp:
1201 case Intrinsic::longjmp:
1202 // We directly implement these intrinsics
1205 // All other intrinsic calls we must lower.
1206 Instruction *Before = CI->getPrev();
1207 IL.LowerIntrinsicCall(CI);
1208 if (Before) { // Move iterator to instruction after call
1218 void CWriter::visitCallInst(CallInst &I) {
1219 // Handle intrinsic function calls first...
1220 if (Function *F = I.getCalledFunction())
1221 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1223 default: assert(0 && "Unknown LLVM intrinsic!");
1224 case Intrinsic::va_start:
1227 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1228 // Output the last argument to the enclosing function...
1229 if (I.getParent()->getParent()->aempty()) {
1230 std::cerr << "The C backend does not currently support zero "
1231 << "argument varargs functions, such as '"
1232 << I.getParent()->getParent()->getName() << "'!\n";
1235 writeOperand(&I.getParent()->getParent()->aback());
1238 case Intrinsic::va_end:
1239 Out << "va_end(*(va_list*)&";
1240 writeOperand(I.getOperand(1));
1243 case Intrinsic::va_copy:
1245 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1246 Out << "*(va_list*)&";
1247 writeOperand(I.getOperand(1));
1250 case Intrinsic::returnaddress:
1251 Out << "__builtin_return_address(";
1252 writeOperand(I.getOperand(1));
1255 case Intrinsic::frameaddress:
1256 Out << "__builtin_frame_address(";
1257 writeOperand(I.getOperand(1));
1260 case Intrinsic::setjmp:
1261 Out << "setjmp(*(jmp_buf*)";
1262 writeOperand(I.getOperand(1));
1265 case Intrinsic::longjmp:
1266 Out << "longjmp(*(jmp_buf*)";
1267 writeOperand(I.getOperand(1));
1269 writeOperand(I.getOperand(2));
1277 void CWriter::visitCallSite(CallSite CS) {
1278 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1279 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1280 const Type *RetTy = FTy->getReturnType();
1282 writeOperand(CS.getCalledValue());
1285 if (CS.arg_begin() != CS.arg_end()) {
1286 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1289 for (++AI; AI != AE; ++AI) {
1297 void CWriter::visitMallocInst(MallocInst &I) {
1298 assert(0 && "lowerallocations pass didn't work!");
1301 void CWriter::visitAllocaInst(AllocaInst &I) {
1303 printType(Out, I.getType());
1304 Out << ") alloca(sizeof(";
1305 printType(Out, I.getType()->getElementType());
1307 if (I.isArrayAllocation()) {
1309 writeOperand(I.getOperand(0));
1314 void CWriter::visitFreeInst(FreeInst &I) {
1315 assert(0 && "lowerallocations pass didn't work!");
1318 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1319 gep_type_iterator E) {
1320 bool HasImplicitAddress = false;
1321 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1322 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1323 HasImplicitAddress = true;
1324 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1325 HasImplicitAddress = true;
1326 Ptr = CPR->getValue(); // Get to the global...
1327 } else if (isDirectAlloca(Ptr)) {
1328 HasImplicitAddress = true;
1332 if (!HasImplicitAddress)
1333 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1335 writeOperandInternal(Ptr);
1339 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1340 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1343 writeOperandInternal(Ptr);
1345 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1347 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1350 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1351 "Can only have implicit address with direct accessing");
1353 if (HasImplicitAddress) {
1355 } else if (CI && CI->isNullValue()) {
1356 gep_type_iterator TmpI = I; ++TmpI;
1358 // Print out the -> operator if possible...
1359 if (TmpI != E && isa<StructType>(*TmpI)) {
1360 Out << (HasImplicitAddress ? "." : "->");
1361 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1367 if (isa<StructType>(*I)) {
1368 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1371 writeOperand(I.getOperand());
1376 void CWriter::visitLoadInst(LoadInst &I) {
1378 writeOperand(I.getOperand(0));
1381 void CWriter::visitStoreInst(StoreInst &I) {
1383 writeOperand(I.getPointerOperand());
1385 writeOperand(I.getOperand(0));
1388 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1390 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1394 void CWriter::visitVANextInst(VANextInst &I) {
1395 Out << Mang->getValueName(I.getOperand(0));
1396 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1397 printType(Out, I.getArgType());
1401 void CWriter::visitVAArgInst(VAArgInst &I) {
1403 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1404 writeOperand(I.getOperand(0));
1405 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1406 printType(Out, I.getType());
1407 Out << ");\n va_end(Tmp); }";
1410 //===----------------------------------------------------------------------===//
1411 // External Interface declaration
1412 //===----------------------------------------------------------------------===//
1414 bool CTargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &o) {
1415 PM.add(createLowerAllocationsPass());
1416 PM.add(createLowerInvokePass());
1417 PM.add(new CWriter(o, getIntrinsicLowering()));
1421 TargetMachine *llvm::allocateCTargetMachine(const Module &M,
1422 IntrinsicLowering *IL) {
1423 return new CTargetMachine(M, IL);