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/Analysis/ConstantsScanner.h"
26 #include "llvm/Analysis/FindUsedTypes.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/Transforms/Scalar.h"
30 #include "llvm/Support/CallSite.h"
31 #include "llvm/Support/CFG.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/InstVisitor.h"
34 #include "llvm/Support/Mangler.h"
35 #include "Support/StringExtras.h"
36 #include "Config/config.h"
42 /// NameAllUsedStructs - This pass inserts names for any unnamed structure
43 /// types that are used by the program.
45 class CBackendNameAllUsedStructs : public Pass {
46 void getAnalysisUsage(AnalysisUsage &AU) const {
47 AU.addRequired<FindUsedTypes>();
50 virtual const char *getPassName() const {
51 return "C backend type canonicalizer";
54 virtual bool run(Module &M);
57 /// CWriter - This class is the main chunk of code that converts an LLVM
58 /// module to a C translation unit.
59 class CWriter : public FunctionPass, public InstVisitor<CWriter> {
61 IntrinsicLowering &IL;
64 const Module *TheModule;
65 std::map<const Type *, std::string> TypeNames;
67 std::map<const ConstantFP *, unsigned> FPConstantMap;
69 CWriter(std::ostream &o, IntrinsicLowering &il) : Out(o), IL(il) {}
71 virtual const char *getPassName() const { return "C backend"; }
73 void getAnalysisUsage(AnalysisUsage &AU) const {
74 AU.addRequired<LoopInfo>();
78 virtual bool doInitialization(Module &M);
80 bool runOnFunction(Function &F) {
81 LI = &getAnalysis<LoopInfo>();
83 // Output all floating point constants that cannot be printed accurately.
84 printFloatingPointConstants(F);
88 FPConstantMap.clear();
92 virtual bool doFinalization(Module &M) {
99 std::ostream &printType(std::ostream &Out, const Type *Ty,
100 const std::string &VariableName = "",
101 bool IgnoreName = false);
103 void writeOperand(Value *Operand);
104 void writeOperandInternal(Value *Operand);
107 void lowerIntrinsics(Function &F);
109 bool nameAllUsedStructureTypes(Module &M);
110 void printModule(Module *M);
111 void printModuleTypes(const SymbolTable &ST);
112 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
113 void printFloatingPointConstants(Function &F);
114 void printFunctionSignature(const Function *F, bool Prototype);
116 void printFunction(Function &);
117 void printBasicBlock(BasicBlock *BB);
118 void printLoop(Loop *L);
120 void printConstant(Constant *CPV);
121 void printConstantArray(ConstantArray *CPA);
123 // isInlinableInst - Attempt to inline instructions into their uses to build
124 // trees as much as possible. To do this, we have to consistently decide
125 // what is acceptable to inline, so that variable declarations don't get
126 // printed and an extra copy of the expr is not emitted.
128 static bool isInlinableInst(const Instruction &I) {
129 // Always inline setcc instructions, even if they are shared by multiple
130 // expressions. GCC generates horrible code if we don't.
131 if (isa<SetCondInst>(I)) return true;
133 // Must be an expression, must be used exactly once. If it is dead, we
134 // emit it inline where it would go.
135 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
136 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
137 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
138 // Don't inline a load across a store or other bad things!
141 // Only inline instruction it it's use is in the same BB as the inst.
142 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
145 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
146 // variables which are accessed with the & operator. This causes GCC to
147 // generate significantly better code than to emit alloca calls directly.
149 static const AllocaInst *isDirectAlloca(const Value *V) {
150 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
151 if (!AI) return false;
152 if (AI->isArrayAllocation())
153 return 0; // FIXME: we can also inline fixed size array allocas!
154 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
159 // Instruction visitation functions
160 friend class InstVisitor<CWriter>;
162 void visitReturnInst(ReturnInst &I);
163 void visitBranchInst(BranchInst &I);
164 void visitSwitchInst(SwitchInst &I);
165 void visitInvokeInst(InvokeInst &I) {
166 assert(0 && "Lowerinvoke pass didn't work!");
169 void visitUnwindInst(UnwindInst &I) {
170 assert(0 && "Lowerinvoke pass didn't work!");
173 void visitPHINode(PHINode &I);
174 void visitBinaryOperator(Instruction &I);
176 void visitCastInst (CastInst &I);
177 void visitSelectInst(SelectInst &I);
178 void visitCallInst (CallInst &I);
179 void visitCallSite (CallSite CS);
180 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
182 void visitMallocInst(MallocInst &I);
183 void visitAllocaInst(AllocaInst &I);
184 void visitFreeInst (FreeInst &I);
185 void visitLoadInst (LoadInst &I);
186 void visitStoreInst (StoreInst &I);
187 void visitGetElementPtrInst(GetElementPtrInst &I);
188 void visitVANextInst(VANextInst &I);
189 void visitVAArgInst (VAArgInst &I);
191 void visitInstruction(Instruction &I) {
192 std::cerr << "C Writer does not know about " << I;
196 void outputLValue(Instruction *I) {
197 Out << " " << Mang->getValueName(I) << " = ";
200 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To);
201 void printPHICopiesForSuccessors(BasicBlock *CurBlock,
203 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
205 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
206 gep_type_iterator E);
210 /// This method inserts names for any unnamed structure types that are used by
211 /// the program, and removes names from structure types that are not used by the
214 bool CBackendNameAllUsedStructs::run(Module &M) {
215 // Get a set of types that are used by the program...
216 std::set<const Type *> UT = getAnalysis<FindUsedTypes>().getTypes();
218 // Loop over the module symbol table, removing types from UT that are
219 // already named, and removing names for structure types that are not used.
221 SymbolTable &MST = M.getSymbolTable();
222 for (SymbolTable::type_iterator TI = MST.type_begin(), TE = MST.type_end();
224 SymbolTable::type_iterator I = TI++;
225 if (StructType *STy = dyn_cast<StructType>(I->second)) {
226 // If this is not used, remove it from the symbol table.
227 std::set<const Type *>::iterator UTI = UT.find(STy);
235 // UT now contains types that are not named. Loop over it, naming
238 bool Changed = false;
239 unsigned RenameCounter = 0;
240 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
242 if (const StructType *ST = dyn_cast<StructType>(*I)) {
243 while (M.addTypeName("unnamed"+utostr(RenameCounter), ST))
251 // Pass the Type* and the variable name and this prints out the variable
254 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
255 const std::string &NameSoFar,
257 if (Ty->isPrimitiveType())
258 switch (Ty->getTypeID()) {
259 case Type::VoidTyID: return Out << "void " << NameSoFar;
260 case Type::BoolTyID: return Out << "bool " << NameSoFar;
261 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
262 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
263 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
264 case Type::ShortTyID: return Out << "short " << NameSoFar;
265 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
266 case Type::IntTyID: return Out << "int " << NameSoFar;
267 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
268 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
269 case Type::FloatTyID: return Out << "float " << NameSoFar;
270 case Type::DoubleTyID: return Out << "double " << NameSoFar;
272 std::cerr << "Unknown primitive type: " << Ty << "\n";
276 // Check to see if the type is named.
277 if (!IgnoreName || isa<OpaqueType>(Ty)) {
278 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
279 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
282 switch (Ty->getTypeID()) {
283 case Type::FunctionTyID: {
284 const FunctionType *MTy = cast<FunctionType>(Ty);
285 std::stringstream FunctionInnards;
286 FunctionInnards << " (" << NameSoFar << ") (";
287 for (FunctionType::param_iterator I = MTy->param_begin(),
288 E = MTy->param_end(); I != E; ++I) {
289 if (I != MTy->param_begin())
290 FunctionInnards << ", ";
291 printType(FunctionInnards, *I, "");
293 if (MTy->isVarArg()) {
294 if (MTy->getNumParams())
295 FunctionInnards << ", ...";
296 } else if (!MTy->getNumParams()) {
297 FunctionInnards << "void";
299 FunctionInnards << ")";
300 std::string tstr = FunctionInnards.str();
301 printType(Out, MTy->getReturnType(), tstr);
304 case Type::StructTyID: {
305 const StructType *STy = cast<StructType>(Ty);
306 Out << NameSoFar + " {\n";
308 for (StructType::element_iterator I = STy->element_begin(),
309 E = STy->element_end(); I != E; ++I) {
311 printType(Out, *I, "field" + utostr(Idx++));
317 case Type::PointerTyID: {
318 const PointerType *PTy = cast<PointerType>(Ty);
319 std::string ptrName = "*" + NameSoFar;
321 if (isa<ArrayType>(PTy->getElementType()))
322 ptrName = "(" + ptrName + ")";
324 return printType(Out, PTy->getElementType(), ptrName);
327 case Type::ArrayTyID: {
328 const ArrayType *ATy = cast<ArrayType>(Ty);
329 unsigned NumElements = ATy->getNumElements();
330 return printType(Out, ATy->getElementType(),
331 NameSoFar + "[" + utostr(NumElements) + "]");
334 case Type::OpaqueTyID: {
335 static int Count = 0;
336 std::string TyName = "struct opaque_" + itostr(Count++);
337 assert(TypeNames.find(Ty) == TypeNames.end());
338 TypeNames[Ty] = TyName;
339 return Out << TyName << " " << NameSoFar;
342 assert(0 && "Unhandled case in getTypeProps!");
349 void CWriter::printConstantArray(ConstantArray *CPA) {
351 // As a special case, print the array as a string if it is an array of
352 // ubytes or an array of sbytes with positive values.
354 const Type *ETy = CPA->getType()->getElementType();
355 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
357 // Make sure the last character is a null char, as automatically added by C
358 if (isString && (CPA->getNumOperands() == 0 ||
359 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
364 // Keep track of whether the last number was a hexadecimal escape
365 bool LastWasHex = false;
367 // Do not include the last character, which we know is null
368 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
369 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
371 // Print it out literally if it is a printable character. The only thing
372 // to be careful about is when the last letter output was a hex escape
373 // code, in which case we have to be careful not to print out hex digits
374 // explicitly (the C compiler thinks it is a continuation of the previous
375 // character, sheesh...)
377 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
379 if (C == '"' || C == '\\')
386 case '\n': Out << "\\n"; break;
387 case '\t': Out << "\\t"; break;
388 case '\r': Out << "\\r"; break;
389 case '\v': Out << "\\v"; break;
390 case '\a': Out << "\\a"; break;
391 case '\"': Out << "\\\""; break;
392 case '\'': Out << "\\\'"; break;
395 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
396 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
405 if (CPA->getNumOperands()) {
407 printConstant(cast<Constant>(CPA->getOperand(0)));
408 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
410 printConstant(cast<Constant>(CPA->getOperand(i)));
417 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
418 // textually as a double (rather than as a reference to a stack-allocated
419 // variable). We decide this by converting CFP to a string and back into a
420 // double, and then checking whether the conversion results in a bit-equal
421 // double to the original value of CFP. This depends on us and the target C
422 // compiler agreeing on the conversion process (which is pretty likely since we
423 // only deal in IEEE FP).
425 static bool isFPCSafeToPrint(const ConstantFP *CFP) {
428 sprintf(Buffer, "%a", CFP->getValue());
430 if (!strncmp(Buffer, "0x", 2) ||
431 !strncmp(Buffer, "-0x", 3) ||
432 !strncmp(Buffer, "+0x", 3))
433 return atof(Buffer) == CFP->getValue();
436 std::string StrVal = ftostr(CFP->getValue());
438 while (StrVal[0] == ' ')
439 StrVal.erase(StrVal.begin());
441 // Check to make sure that the stringized number is not some string like "Inf"
442 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
443 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
444 ((StrVal[0] == '-' || StrVal[0] == '+') &&
445 (StrVal[1] >= '0' && StrVal[1] <= '9')))
446 // Reparse stringized version!
447 return atof(StrVal.c_str()) == CFP->getValue();
452 // printConstant - The LLVM Constant to C Constant converter.
453 void CWriter::printConstant(Constant *CPV) {
454 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
455 switch (CE->getOpcode()) {
456 case Instruction::Cast:
458 printType(Out, CPV->getType());
460 printConstant(CE->getOperand(0));
464 case Instruction::GetElementPtr:
466 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
470 case Instruction::Select:
472 printConstant(CE->getOperand(0));
474 printConstant(CE->getOperand(1));
476 printConstant(CE->getOperand(2));
479 case Instruction::Add:
480 case Instruction::Sub:
481 case Instruction::Mul:
482 case Instruction::Div:
483 case Instruction::Rem:
484 case Instruction::SetEQ:
485 case Instruction::SetNE:
486 case Instruction::SetLT:
487 case Instruction::SetLE:
488 case Instruction::SetGT:
489 case Instruction::SetGE:
490 case Instruction::Shl:
491 case Instruction::Shr:
493 printConstant(CE->getOperand(0));
494 switch (CE->getOpcode()) {
495 case Instruction::Add: Out << " + "; break;
496 case Instruction::Sub: Out << " - "; break;
497 case Instruction::Mul: Out << " * "; break;
498 case Instruction::Div: Out << " / "; break;
499 case Instruction::Rem: Out << " % "; break;
500 case Instruction::SetEQ: Out << " == "; break;
501 case Instruction::SetNE: Out << " != "; break;
502 case Instruction::SetLT: Out << " < "; break;
503 case Instruction::SetLE: Out << " <= "; break;
504 case Instruction::SetGT: Out << " > "; break;
505 case Instruction::SetGE: Out << " >= "; break;
506 case Instruction::Shl: Out << " << "; break;
507 case Instruction::Shr: Out << " >> "; break;
508 default: assert(0 && "Illegal opcode here!");
510 printConstant(CE->getOperand(1));
515 std::cerr << "CWriter Error: Unhandled constant expression: "
521 switch (CPV->getType()->getTypeID()) {
523 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
524 case Type::SByteTyID:
525 case Type::ShortTyID:
526 Out << cast<ConstantSInt>(CPV)->getValue(); break;
528 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
529 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
531 Out << cast<ConstantSInt>(CPV)->getValue();
535 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
537 case Type::UByteTyID:
538 case Type::UShortTyID:
539 Out << cast<ConstantUInt>(CPV)->getValue(); break;
541 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
542 case Type::ULongTyID:
543 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
545 case Type::FloatTyID:
546 case Type::DoubleTyID: {
547 ConstantFP *FPC = cast<ConstantFP>(CPV);
548 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
549 if (I != FPConstantMap.end()) {
550 // Because of FP precision problems we must load from a stack allocated
551 // value that holds the value in hex.
552 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
553 << "*)&FPConstant" << I->second << ")";
556 // Print out the constant as a floating point number.
558 sprintf(Buffer, "%a", FPC->getValue());
559 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
561 Out << ftostr(FPC->getValue());
567 case Type::ArrayTyID:
568 if (isa<ConstantAggregateZero>(CPV)) {
569 const ArrayType *AT = cast<ArrayType>(CPV->getType());
571 if (AT->getNumElements()) {
573 Constant *CZ = Constant::getNullValue(AT->getElementType());
575 for (unsigned i = 1, e = AT->getNumElements(); i != e; ++i) {
582 printConstantArray(cast<ConstantArray>(CPV));
586 case Type::StructTyID:
587 if (isa<ConstantAggregateZero>(CPV)) {
588 const StructType *ST = cast<StructType>(CPV->getType());
590 if (ST->getNumElements()) {
592 printConstant(Constant::getNullValue(ST->getElementType(0)));
593 for (unsigned i = 1, e = ST->getNumElements(); i != e; ++i) {
595 printConstant(Constant::getNullValue(ST->getElementType(i)));
601 if (CPV->getNumOperands()) {
603 printConstant(cast<Constant>(CPV->getOperand(0)));
604 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
606 printConstant(cast<Constant>(CPV->getOperand(i)));
613 case Type::PointerTyID:
614 if (isa<ConstantPointerNull>(CPV)) {
616 printType(Out, CPV->getType());
617 Out << ")/*NULL*/0)";
619 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
620 writeOperand(CPR->getValue());
625 std::cerr << "Unknown constant type: " << CPV << "\n";
630 void CWriter::writeOperandInternal(Value *Operand) {
631 if (Instruction *I = dyn_cast<Instruction>(Operand))
632 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
633 // Should we inline this instruction to build a tree?
640 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
643 Out << Mang->getValueName(Operand);
647 void CWriter::writeOperand(Value *Operand) {
648 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
649 Out << "(&"; // Global variables are references as their addresses by llvm
651 writeOperandInternal(Operand);
653 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
657 // generateCompilerSpecificCode - This is where we add conditional compilation
658 // directives to cater to specific compilers as need be.
660 static void generateCompilerSpecificCode(std::ostream& Out) {
661 // Alloca is hard to get, and we don't want to include stdlib.h here...
662 Out << "/* get a declaration for alloca */\n"
663 << "#if defined(sun) || defined(__CYGWIN__)\n"
664 << "extern void *__builtin_alloca(unsigned long);\n"
665 << "#define alloca(x) __builtin_alloca(x)\n"
667 << "#ifndef __FreeBSD__\n"
668 << "#include <alloca.h>\n"
672 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
673 // If we aren't being compiled with GCC, just drop these attributes.
674 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
675 << "#define __attribute__(X)\n"
679 // At some point, we should support "external weak" vs. "weak" linkages.
680 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
681 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
682 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
683 << "#elif defined(__GNUC__)\n"
684 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
686 << "#define __EXTERNAL_WEAK__\n"
690 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
691 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
692 << "#define __ATTRIBUTE_WEAK__\n"
693 << "#elif defined(__GNUC__)\n"
694 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
696 << "#define __ATTRIBUTE_WEAK__\n"
700 bool CWriter::doInitialization(Module &M) {
706 // Ensure that all structure types have names...
707 Mang = new Mangler(M);
709 // get declaration for alloca
710 Out << "/* Provide Declarations */\n";
711 Out << "#include <stdarg.h>\n"; // Varargs support
712 Out << "#include <setjmp.h>\n"; // Unwind support
713 generateCompilerSpecificCode(Out);
715 // Provide a definition for `bool' if not compiling with a C++ compiler.
717 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
719 << "\n\n/* Support for floating point constants */\n"
720 << "typedef unsigned long long ConstantDoubleTy;\n"
721 << "typedef unsigned int ConstantFloatTy;\n"
723 << "\n\n/* Global Declarations */\n";
725 // First output all the declarations for the program, because C requires
726 // Functions & globals to be declared before they are used.
729 // Loop over the symbol table, emitting all named constants...
730 printModuleTypes(M.getSymbolTable());
732 // Global variable declarations...
734 Out << "\n/* External Global Variable Declarations */\n";
735 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I) {
736 if (I->hasExternalLinkage()) {
738 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
744 // Function declarations
746 Out << "\n/* Function Declarations */\n";
747 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
748 // Don't print declarations for intrinsic functions.
749 if (!I->getIntrinsicID() &&
750 I->getName() != "setjmp" && I->getName() != "longjmp") {
751 printFunctionSignature(I, true);
752 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
753 if (I->hasLinkOnceLinkage()) Out << " __ATTRIBUTE_WEAK__";
759 // Output the global variable declarations
761 Out << "\n\n/* Global Variable Declarations */\n";
762 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
763 if (!I->isExternal()) {
765 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
767 if (I->hasLinkOnceLinkage())
768 Out << " __attribute__((common))";
769 else if (I->hasWeakLinkage())
770 Out << " __ATTRIBUTE_WEAK__";
775 // Output the global variable definitions and contents...
777 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
778 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
779 if (!I->isExternal()) {
780 if (I->hasInternalLinkage())
782 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
783 if (I->hasLinkOnceLinkage())
784 Out << " __attribute__((common))";
785 else if (I->hasWeakLinkage())
786 Out << " __ATTRIBUTE_WEAK__";
788 // If the initializer is not null, emit the initializer. If it is null,
789 // we try to avoid emitting large amounts of zeros. The problem with
790 // this, however, occurs when the variable has weak linkage. In this
791 // case, the assembler will complain about the variable being both weak
792 // and common, so we disable this optimization.
793 if (!I->getInitializer()->isNullValue()) {
795 writeOperand(I->getInitializer());
796 } else if (I->hasWeakLinkage()) {
797 // We have to specify an initializer, but it doesn't have to be
798 // complete. If the value is an aggregate, print out { 0 }, and let
799 // the compiler figure out the rest of the zeros.
801 if (isa<StructType>(I->getInitializer()->getType()) ||
802 isa<ArrayType>(I->getInitializer()->getType())) {
805 // Just print it out normally.
806 writeOperand(I->getInitializer());
814 Out << "\n\n/* Function Bodies */\n";
819 /// Output all floating point constants that cannot be printed accurately...
820 void CWriter::printFloatingPointConstants(Function &F) {
831 // Scan the module for floating point constants. If any FP constant is used
832 // in the function, we want to redirect it here so that we do not depend on
833 // the precision of the printed form, unless the printed form preserves
836 static unsigned FPCounter = 0;
837 for (constant_iterator I = constant_begin(&F), E = constant_end(&F);
839 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
840 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
841 !FPConstantMap.count(FPC)) {
842 double Val = FPC->getValue();
844 FPConstantMap[FPC] = FPCounter; // Number the FP constants
846 if (FPC->getType() == Type::DoubleTy) {
848 Out << "static const ConstantDoubleTy FPConstant" << FPCounter++
849 << " = 0x" << std::hex << DBLUnion.U << std::dec
850 << "ULL; /* " << Val << " */\n";
851 } else if (FPC->getType() == Type::FloatTy) {
853 Out << "static const ConstantFloatTy FPConstant" << FPCounter++
854 << " = 0x" << std::hex << FLTUnion.U << std::dec
855 << "U; /* " << Val << " */\n";
857 assert(0 && "Unknown float type!");
864 /// printSymbolTable - Run through symbol table looking for type names. If a
865 /// type name is found, emit it's declaration...
867 void CWriter::printModuleTypes(const SymbolTable &ST) {
868 // If there are no type names, exit early.
869 if ( ! ST.hasTypes() )
872 // We are only interested in the type plane of the symbol table...
873 SymbolTable::type_const_iterator I = ST.type_begin();
874 SymbolTable::type_const_iterator End = ST.type_end();
876 // Print out forward declarations for structure types before anything else!
877 Out << "/* Structure forward decls */\n";
878 for (; I != End; ++I)
879 if (const Type *STy = dyn_cast<StructType>(I->second)) {
880 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
881 Out << Name << ";\n";
882 TypeNames.insert(std::make_pair(STy, Name));
887 // Now we can print out typedefs...
888 Out << "/* Typedefs */\n";
889 for (I = ST.type_begin(); I != End; ++I) {
890 const Type *Ty = cast<Type>(I->second);
891 std::string Name = "l_" + Mangler::makeNameProper(I->first);
893 printType(Out, Ty, Name);
899 // Keep track of which structures have been printed so far...
900 std::set<const StructType *> StructPrinted;
902 // Loop over all structures then push them into the stack so they are
903 // printed in the correct order.
905 Out << "/* Structure contents */\n";
906 for (I = ST.type_begin(); I != End; ++I)
907 if (const StructType *STy = dyn_cast<StructType>(I->second))
908 // Only print out used types!
909 printContainedStructs(STy, StructPrinted);
912 // Push the struct onto the stack and recursively push all structs
913 // this one depends on.
914 void CWriter::printContainedStructs(const Type *Ty,
915 std::set<const StructType*> &StructPrinted){
916 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
917 //Check to see if we have already printed this struct
918 if (StructPrinted.count(STy) == 0) {
919 // Print all contained types first...
920 for (StructType::element_iterator I = STy->element_begin(),
921 E = STy->element_end(); I != E; ++I) {
922 const Type *Ty1 = I->get();
923 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
924 printContainedStructs(*I, StructPrinted);
927 //Print structure type out..
928 StructPrinted.insert(STy);
929 std::string Name = TypeNames[STy];
930 printType(Out, STy, Name, true);
934 // If it is an array, check contained types and continue
935 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
936 const Type *Ty1 = ATy->getElementType();
937 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
938 printContainedStructs(Ty1, StructPrinted);
943 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
944 if (F->hasInternalLinkage()) Out << "static ";
946 // Loop over the arguments, printing them...
947 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
949 std::stringstream FunctionInnards;
951 // Print out the name...
952 FunctionInnards << Mang->getValueName(F) << "(";
954 if (!F->isExternal()) {
957 if (F->abegin()->hasName() || !Prototype)
958 ArgName = Mang->getValueName(F->abegin());
959 printType(FunctionInnards, F->afront().getType(), ArgName);
960 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
962 FunctionInnards << ", ";
963 if (I->hasName() || !Prototype)
964 ArgName = Mang->getValueName(I);
967 printType(FunctionInnards, I->getType(), ArgName);
971 // Loop over the arguments, printing them...
972 for (FunctionType::param_iterator I = FT->param_begin(),
973 E = FT->param_end(); I != E; ++I) {
974 if (I != FT->param_begin()) FunctionInnards << ", ";
975 printType(FunctionInnards, *I);
979 // Finish printing arguments... if this is a vararg function, print the ...,
980 // unless there are no known types, in which case, we just emit ().
982 if (FT->isVarArg() && FT->getNumParams()) {
983 if (FT->getNumParams()) FunctionInnards << ", ";
984 FunctionInnards << "..."; // Output varargs portion of signature!
985 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
986 FunctionInnards << "void"; // ret() -> ret(void) in C.
988 FunctionInnards << ")";
989 // Print out the return type and the entire signature for that matter
990 printType(Out, F->getReturnType(), FunctionInnards.str());
993 void CWriter::printFunction(Function &F) {
994 printFunctionSignature(&F, false);
997 // print local variable information for the function
998 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I)
999 if (const AllocaInst *AI = isDirectAlloca(&*I)) {
1001 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
1002 Out << "; /* Address exposed local */\n";
1003 } else if (I->getType() != Type::VoidTy && !isInlinableInst(*I)) {
1005 printType(Out, I->getType(), Mang->getValueName(&*I));
1008 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
1010 printType(Out, I->getType(),
1011 Mang->getValueName(&*I)+"__PHI_TEMPORARY");
1018 // print the basic blocks
1019 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1020 if (Loop *L = LI->getLoopFor(BB)) {
1021 if (L->getHeader() == BB && L->getParentLoop() == 0)
1024 printBasicBlock(BB);
1031 void CWriter::printLoop(Loop *L) {
1032 Out << " do { /* Syntactic loop '" << L->getHeader()->getName()
1033 << "' to make GCC happy */\n";
1034 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
1035 BasicBlock *BB = L->getBlocks()[i];
1036 Loop *BBLoop = LI->getLoopFor(BB);
1038 printBasicBlock(BB);
1039 else if (BB == BBLoop->getHeader() && BBLoop->getParentLoop() == L)
1042 Out << " } while (1); /* end of syntactic loop '"
1043 << L->getHeader()->getName() << "' */\n";
1046 void CWriter::printBasicBlock(BasicBlock *BB) {
1048 // Don't print the label for the basic block if there are no uses, or if
1049 // the only terminator use is the predecessor basic block's terminator.
1050 // We have to scan the use list because PHI nodes use basic blocks too but
1051 // do not require a label to be generated.
1053 bool NeedsLabel = false;
1054 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1055 if (isGotoCodeNecessary(*PI, BB)) {
1060 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1062 // Output all of the instructions in the basic block...
1063 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E;
1065 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1066 if (II->getType() != Type::VoidTy)
1075 // Don't emit prefix or suffix for the terminator...
1076 visit(*BB->getTerminator());
1080 // Specific Instruction type classes... note that all of the casts are
1081 // necessary because we use the instruction classes as opaque types...
1083 void CWriter::visitReturnInst(ReturnInst &I) {
1084 // Don't output a void return if this is the last basic block in the function
1085 if (I.getNumOperands() == 0 &&
1086 &*--I.getParent()->getParent()->end() == I.getParent() &&
1087 !I.getParent()->size() == 1) {
1092 if (I.getNumOperands()) {
1094 writeOperand(I.getOperand(0));
1099 void CWriter::visitSwitchInst(SwitchInst &SI) {
1100 printPHICopiesForSuccessors(SI.getParent(), 0);
1103 writeOperand(SI.getOperand(0));
1104 Out << ") {\n default:\n";
1105 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1107 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1109 writeOperand(SI.getOperand(i));
1111 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1112 printBranchToBlock(SI.getParent(), Succ, 2);
1113 if (Succ == SI.getParent()->getNext())
1119 bool CWriter::isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1120 /// FIXME: This should be reenabled, but loop reordering safe!!
1123 if (From->getNext() != To) // Not the direct successor, we need a goto
1126 //isa<SwitchInst>(From->getTerminator())
1129 if (LI->getLoopFor(From) != LI->getLoopFor(To))
1134 void CWriter::printPHICopiesForSuccessors(BasicBlock *CurBlock,
1136 for (succ_iterator SI = succ_begin(CurBlock), E = succ_end(CurBlock);
1138 for (BasicBlock::iterator I = SI->begin();
1139 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1140 // now we have to do the printing
1141 Out << std::string(Indent, ' ');
1142 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1143 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBlock)));
1144 Out << "; /* for PHI node */\n";
1149 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1151 if (isGotoCodeNecessary(CurBB, Succ)) {
1152 Out << std::string(Indent, ' ') << " goto ";
1158 // Branch instruction printing - Avoid printing out a branch to a basic block
1159 // that immediately succeeds the current one.
1161 void CWriter::visitBranchInst(BranchInst &I) {
1162 printPHICopiesForSuccessors(I.getParent(), 0);
1164 if (I.isConditional()) {
1165 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1167 writeOperand(I.getCondition());
1170 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1172 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1173 Out << " } else {\n";
1174 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1177 // First goto not necessary, assume second one is...
1179 writeOperand(I.getCondition());
1182 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1187 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1192 // PHI nodes get copied into temporary values at the end of predecessor basic
1193 // blocks. We now need to copy these temporary values into the REAL value for
1195 void CWriter::visitPHINode(PHINode &I) {
1197 Out << "__PHI_TEMPORARY";
1201 void CWriter::visitBinaryOperator(Instruction &I) {
1202 // binary instructions, shift instructions, setCond instructions.
1203 assert(!isa<PointerType>(I.getType()));
1205 // We must cast the results of binary operations which might be promoted.
1206 bool needsCast = false;
1207 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1208 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1209 || (I.getType() == Type::FloatTy)) {
1212 printType(Out, I.getType());
1216 writeOperand(I.getOperand(0));
1218 switch (I.getOpcode()) {
1219 case Instruction::Add: Out << " + "; break;
1220 case Instruction::Sub: Out << " - "; break;
1221 case Instruction::Mul: Out << "*"; break;
1222 case Instruction::Div: Out << "/"; break;
1223 case Instruction::Rem: Out << "%"; break;
1224 case Instruction::And: Out << " & "; break;
1225 case Instruction::Or: Out << " | "; break;
1226 case Instruction::Xor: Out << " ^ "; break;
1227 case Instruction::SetEQ: Out << " == "; break;
1228 case Instruction::SetNE: Out << " != "; break;
1229 case Instruction::SetLE: Out << " <= "; break;
1230 case Instruction::SetGE: Out << " >= "; break;
1231 case Instruction::SetLT: Out << " < "; break;
1232 case Instruction::SetGT: Out << " > "; break;
1233 case Instruction::Shl : Out << " << "; break;
1234 case Instruction::Shr : Out << " >> "; break;
1235 default: std::cerr << "Invalid operator type!" << I; abort();
1238 writeOperand(I.getOperand(1));
1245 void CWriter::visitCastInst(CastInst &I) {
1246 if (I.getType() == Type::BoolTy) {
1248 writeOperand(I.getOperand(0));
1253 printType(Out, I.getType());
1255 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1256 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1257 // Avoid "cast to pointer from integer of different size" warnings
1261 writeOperand(I.getOperand(0));
1264 void CWriter::visitSelectInst(SelectInst &I) {
1266 writeOperand(I.getCondition());
1268 writeOperand(I.getTrueValue());
1270 writeOperand(I.getFalseValue());
1275 void CWriter::lowerIntrinsics(Function &F) {
1276 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1277 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1278 if (CallInst *CI = dyn_cast<CallInst>(I++))
1279 if (Function *F = CI->getCalledFunction())
1280 switch (F->getIntrinsicID()) {
1281 case Intrinsic::not_intrinsic:
1282 case Intrinsic::vastart:
1283 case Intrinsic::vacopy:
1284 case Intrinsic::vaend:
1285 case Intrinsic::returnaddress:
1286 case Intrinsic::frameaddress:
1287 case Intrinsic::setjmp:
1288 case Intrinsic::longjmp:
1289 // We directly implement these intrinsics
1292 // All other intrinsic calls we must lower.
1293 Instruction *Before = CI->getPrev();
1294 IL.LowerIntrinsicCall(CI);
1295 if (Before) { // Move iterator to instruction after call
1305 void CWriter::visitCallInst(CallInst &I) {
1306 // Handle intrinsic function calls first...
1307 if (Function *F = I.getCalledFunction())
1308 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1310 default: assert(0 && "Unknown LLVM intrinsic!");
1311 case Intrinsic::vastart:
1314 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1315 // Output the last argument to the enclosing function...
1316 if (I.getParent()->getParent()->aempty()) {
1317 std::cerr << "The C backend does not currently support zero "
1318 << "argument varargs functions, such as '"
1319 << I.getParent()->getParent()->getName() << "'!\n";
1322 writeOperand(&I.getParent()->getParent()->aback());
1325 case Intrinsic::vaend:
1326 Out << "va_end(*(va_list*)&";
1327 writeOperand(I.getOperand(1));
1330 case Intrinsic::vacopy:
1332 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1333 Out << "*(va_list*)&";
1334 writeOperand(I.getOperand(1));
1337 case Intrinsic::returnaddress:
1338 Out << "__builtin_return_address(";
1339 writeOperand(I.getOperand(1));
1342 case Intrinsic::frameaddress:
1343 Out << "__builtin_frame_address(";
1344 writeOperand(I.getOperand(1));
1347 case Intrinsic::setjmp:
1348 Out << "setjmp(*(jmp_buf*)";
1349 writeOperand(I.getOperand(1));
1352 case Intrinsic::longjmp:
1353 Out << "longjmp(*(jmp_buf*)";
1354 writeOperand(I.getOperand(1));
1356 writeOperand(I.getOperand(2));
1364 void CWriter::visitCallSite(CallSite CS) {
1365 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1366 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1367 const Type *RetTy = FTy->getReturnType();
1369 writeOperand(CS.getCalledValue());
1372 if (CS.arg_begin() != CS.arg_end()) {
1373 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1376 for (++AI; AI != AE; ++AI) {
1384 void CWriter::visitMallocInst(MallocInst &I) {
1385 assert(0 && "lowerallocations pass didn't work!");
1388 void CWriter::visitAllocaInst(AllocaInst &I) {
1390 printType(Out, I.getType());
1391 Out << ") alloca(sizeof(";
1392 printType(Out, I.getType()->getElementType());
1394 if (I.isArrayAllocation()) {
1396 writeOperand(I.getOperand(0));
1401 void CWriter::visitFreeInst(FreeInst &I) {
1402 assert(0 && "lowerallocations pass didn't work!");
1405 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1406 gep_type_iterator E) {
1407 bool HasImplicitAddress = false;
1408 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1409 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1410 HasImplicitAddress = true;
1411 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1412 HasImplicitAddress = true;
1413 Ptr = CPR->getValue(); // Get to the global...
1414 } else if (isDirectAlloca(Ptr)) {
1415 HasImplicitAddress = true;
1419 if (!HasImplicitAddress)
1420 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1422 writeOperandInternal(Ptr);
1426 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1427 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1430 writeOperandInternal(Ptr);
1432 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1434 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1437 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1438 "Can only have implicit address with direct accessing");
1440 if (HasImplicitAddress) {
1442 } else if (CI && CI->isNullValue()) {
1443 gep_type_iterator TmpI = I; ++TmpI;
1445 // Print out the -> operator if possible...
1446 if (TmpI != E && isa<StructType>(*TmpI)) {
1447 Out << (HasImplicitAddress ? "." : "->");
1448 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1454 if (isa<StructType>(*I)) {
1455 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1458 writeOperand(I.getOperand());
1463 void CWriter::visitLoadInst(LoadInst &I) {
1465 writeOperand(I.getOperand(0));
1468 void CWriter::visitStoreInst(StoreInst &I) {
1470 writeOperand(I.getPointerOperand());
1472 writeOperand(I.getOperand(0));
1475 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1477 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1481 void CWriter::visitVANextInst(VANextInst &I) {
1482 Out << Mang->getValueName(I.getOperand(0));
1483 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1484 printType(Out, I.getArgType());
1488 void CWriter::visitVAArgInst(VAArgInst &I) {
1490 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1491 writeOperand(I.getOperand(0));
1492 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1493 printType(Out, I.getType());
1494 Out << ");\n va_end(Tmp); }";
1497 //===----------------------------------------------------------------------===//
1498 // External Interface declaration
1499 //===----------------------------------------------------------------------===//
1501 bool CTargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &o) {
1502 PM.add(createLowerGCPass());
1503 PM.add(createLowerAllocationsPass());
1504 PM.add(createLowerInvokePass());
1505 PM.add(new CBackendNameAllUsedStructs());
1506 PM.add(new CWriter(o, getIntrinsicLowering()));
1510 TargetMachine *llvm::allocateCTargetMachine(const Module &M,
1511 IntrinsicLowering *IL) {
1512 return new CTargetMachine(M, IL);