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"
43 /// NameAllUsedStructs - This pass inserts names for any unnamed structure
44 /// types that are used by the program.
46 class CBackendNameAllUsedStructs : public Pass {
47 void getAnalysisUsage(AnalysisUsage &AU) const {
48 AU.addRequired<FindUsedTypes>();
51 virtual const char *getPassName() const {
52 return "C backend type canonicalizer";
55 virtual bool run(Module &M);
58 /// CWriter - This class is the main chunk of code that converts an LLVM
59 /// module to a C translation unit.
60 class CWriter : public FunctionPass, public InstVisitor<CWriter> {
62 IntrinsicLowering &IL;
65 const Module *TheModule;
66 std::map<const Type *, std::string> TypeNames;
68 std::map<const ConstantFP *, unsigned> FPConstantMap;
70 CWriter(std::ostream &o, IntrinsicLowering &il) : Out(o), IL(il) {}
72 virtual const char *getPassName() const { return "C backend"; }
74 void getAnalysisUsage(AnalysisUsage &AU) const {
75 AU.addRequired<LoopInfo>();
79 virtual bool doInitialization(Module &M);
81 bool runOnFunction(Function &F) {
82 LI = &getAnalysis<LoopInfo>();
84 // Output all floating point constants that cannot be printed accurately.
85 printFloatingPointConstants(F);
89 FPConstantMap.clear();
93 virtual bool doFinalization(Module &M) {
100 std::ostream &printType(std::ostream &Out, const Type *Ty,
101 const std::string &VariableName = "",
102 bool IgnoreName = false);
104 void writeOperand(Value *Operand);
105 void writeOperandInternal(Value *Operand);
108 void lowerIntrinsics(Function &F);
110 bool nameAllUsedStructureTypes(Module &M);
111 void printModule(Module *M);
112 void printModuleTypes(const SymbolTable &ST);
113 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
114 void printFloatingPointConstants(Function &F);
115 void printFunctionSignature(const Function *F, bool Prototype);
117 void printFunction(Function &);
118 void printBasicBlock(BasicBlock *BB);
119 void printLoop(Loop *L);
121 void printConstant(Constant *CPV);
122 void printConstantArray(ConstantArray *CPA);
124 // isInlinableInst - Attempt to inline instructions into their uses to build
125 // trees as much as possible. To do this, we have to consistently decide
126 // what is acceptable to inline, so that variable declarations don't get
127 // printed and an extra copy of the expr is not emitted.
129 static bool isInlinableInst(const Instruction &I) {
130 // Always inline setcc instructions, even if they are shared by multiple
131 // expressions. GCC generates horrible code if we don't.
132 if (isa<SetCondInst>(I)) return true;
134 // Must be an expression, must be used exactly once. If it is dead, we
135 // emit it inline where it would go.
136 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
137 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
138 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
139 // Don't inline a load across a store or other bad things!
142 // Only inline instruction it it's use is in the same BB as the inst.
143 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
146 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
147 // variables which are accessed with the & operator. This causes GCC to
148 // generate significantly better code than to emit alloca calls directly.
150 static const AllocaInst *isDirectAlloca(const Value *V) {
151 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
152 if (!AI) return false;
153 if (AI->isArrayAllocation())
154 return 0; // FIXME: we can also inline fixed size array allocas!
155 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
160 // Instruction visitation functions
161 friend class InstVisitor<CWriter>;
163 void visitReturnInst(ReturnInst &I);
164 void visitBranchInst(BranchInst &I);
165 void visitSwitchInst(SwitchInst &I);
166 void visitInvokeInst(InvokeInst &I) {
167 assert(0 && "Lowerinvoke pass didn't work!");
170 void visitUnwindInst(UnwindInst &I) {
171 assert(0 && "Lowerinvoke pass didn't work!");
174 void visitPHINode(PHINode &I);
175 void visitBinaryOperator(Instruction &I);
177 void visitCastInst (CastInst &I);
178 void visitSelectInst(SelectInst &I);
179 void visitCallInst (CallInst &I);
180 void visitCallSite (CallSite CS);
181 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
183 void visitMallocInst(MallocInst &I);
184 void visitAllocaInst(AllocaInst &I);
185 void visitFreeInst (FreeInst &I);
186 void visitLoadInst (LoadInst &I);
187 void visitStoreInst (StoreInst &I);
188 void visitGetElementPtrInst(GetElementPtrInst &I);
189 void visitVANextInst(VANextInst &I);
190 void visitVAArgInst (VAArgInst &I);
192 void visitInstruction(Instruction &I) {
193 std::cerr << "C Writer does not know about " << I;
197 void outputLValue(Instruction *I) {
198 Out << " " << Mang->getValueName(I) << " = ";
201 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To);
202 void printPHICopiesForSuccessors(BasicBlock *CurBlock,
204 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
206 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
207 gep_type_iterator E);
211 /// This method inserts names for any unnamed structure types that are used by
212 /// the program, and removes names from structure types that are not used by the
215 bool CBackendNameAllUsedStructs::run(Module &M) {
216 // Get a set of types that are used by the program...
217 std::set<const Type *> UT = getAnalysis<FindUsedTypes>().getTypes();
219 // Loop over the module symbol table, removing types from UT that are
220 // already named, and removing names for structure types that are not used.
222 SymbolTable &MST = M.getSymbolTable();
223 for (SymbolTable::type_iterator TI = MST.type_begin(), TE = MST.type_end();
225 SymbolTable::type_iterator I = TI++;
226 if (const StructType *STy = dyn_cast<StructType>(I->second)) {
227 // If this is not used, remove it from the symbol table.
228 std::set<const Type *>::iterator UTI = UT.find(STy);
236 // UT now contains types that are not named. Loop over it, naming
239 bool Changed = false;
240 unsigned RenameCounter = 0;
241 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
243 if (const StructType *ST = dyn_cast<StructType>(*I)) {
244 while (M.addTypeName("unnamed"+utostr(RenameCounter), ST))
252 // Pass the Type* and the variable name and this prints out the variable
255 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
256 const std::string &NameSoFar,
258 if (Ty->isPrimitiveType())
259 switch (Ty->getTypeID()) {
260 case Type::VoidTyID: return Out << "void " << NameSoFar;
261 case Type::BoolTyID: return Out << "bool " << NameSoFar;
262 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
263 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
264 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
265 case Type::ShortTyID: return Out << "short " << NameSoFar;
266 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
267 case Type::IntTyID: return Out << "int " << NameSoFar;
268 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
269 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
270 case Type::FloatTyID: return Out << "float " << NameSoFar;
271 case Type::DoubleTyID: return Out << "double " << NameSoFar;
273 std::cerr << "Unknown primitive type: " << Ty << "\n";
277 // Check to see if the type is named.
278 if (!IgnoreName || isa<OpaqueType>(Ty)) {
279 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
280 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
283 switch (Ty->getTypeID()) {
284 case Type::FunctionTyID: {
285 const FunctionType *MTy = cast<FunctionType>(Ty);
286 std::stringstream FunctionInnards;
287 FunctionInnards << " (" << NameSoFar << ") (";
288 for (FunctionType::param_iterator I = MTy->param_begin(),
289 E = MTy->param_end(); I != E; ++I) {
290 if (I != MTy->param_begin())
291 FunctionInnards << ", ";
292 printType(FunctionInnards, *I, "");
294 if (MTy->isVarArg()) {
295 if (MTy->getNumParams())
296 FunctionInnards << ", ...";
297 } else if (!MTy->getNumParams()) {
298 FunctionInnards << "void";
300 FunctionInnards << ")";
301 std::string tstr = FunctionInnards.str();
302 printType(Out, MTy->getReturnType(), tstr);
305 case Type::StructTyID: {
306 const StructType *STy = cast<StructType>(Ty);
307 Out << NameSoFar + " {\n";
309 for (StructType::element_iterator I = STy->element_begin(),
310 E = STy->element_end(); I != E; ++I) {
312 printType(Out, *I, "field" + utostr(Idx++));
318 case Type::PointerTyID: {
319 const PointerType *PTy = cast<PointerType>(Ty);
320 std::string ptrName = "*" + NameSoFar;
322 if (isa<ArrayType>(PTy->getElementType()))
323 ptrName = "(" + ptrName + ")";
325 return printType(Out, PTy->getElementType(), ptrName);
328 case Type::ArrayTyID: {
329 const ArrayType *ATy = cast<ArrayType>(Ty);
330 unsigned NumElements = ATy->getNumElements();
331 return printType(Out, ATy->getElementType(),
332 NameSoFar + "[" + utostr(NumElements) + "]");
335 case Type::OpaqueTyID: {
336 static int Count = 0;
337 std::string TyName = "struct opaque_" + itostr(Count++);
338 assert(TypeNames.find(Ty) == TypeNames.end());
339 TypeNames[Ty] = TyName;
340 return Out << TyName << " " << NameSoFar;
343 assert(0 && "Unhandled case in getTypeProps!");
350 void CWriter::printConstantArray(ConstantArray *CPA) {
352 // As a special case, print the array as a string if it is an array of
353 // ubytes or an array of sbytes with positive values.
355 const Type *ETy = CPA->getType()->getElementType();
356 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
358 // Make sure the last character is a null char, as automatically added by C
359 if (isString && (CPA->getNumOperands() == 0 ||
360 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
365 // Keep track of whether the last number was a hexadecimal escape
366 bool LastWasHex = false;
368 // Do not include the last character, which we know is null
369 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
370 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
372 // Print it out literally if it is a printable character. The only thing
373 // to be careful about is when the last letter output was a hex escape
374 // code, in which case we have to be careful not to print out hex digits
375 // explicitly (the C compiler thinks it is a continuation of the previous
376 // character, sheesh...)
378 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
380 if (C == '"' || C == '\\')
387 case '\n': Out << "\\n"; break;
388 case '\t': Out << "\\t"; break;
389 case '\r': Out << "\\r"; break;
390 case '\v': Out << "\\v"; break;
391 case '\a': Out << "\\a"; break;
392 case '\"': Out << "\\\""; break;
393 case '\'': Out << "\\\'"; break;
396 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
397 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
406 if (CPA->getNumOperands()) {
408 printConstant(cast<Constant>(CPA->getOperand(0)));
409 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
411 printConstant(cast<Constant>(CPA->getOperand(i)));
418 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
419 // textually as a double (rather than as a reference to a stack-allocated
420 // variable). We decide this by converting CFP to a string and back into a
421 // double, and then checking whether the conversion results in a bit-equal
422 // double to the original value of CFP. This depends on us and the target C
423 // compiler agreeing on the conversion process (which is pretty likely since we
424 // only deal in IEEE FP).
426 static bool isFPCSafeToPrint(const ConstantFP *CFP) {
429 sprintf(Buffer, "%a", CFP->getValue());
431 if (!strncmp(Buffer, "0x", 2) ||
432 !strncmp(Buffer, "-0x", 3) ||
433 !strncmp(Buffer, "+0x", 3))
434 return atof(Buffer) == CFP->getValue();
437 std::string StrVal = ftostr(CFP->getValue());
439 while (StrVal[0] == ' ')
440 StrVal.erase(StrVal.begin());
442 // Check to make sure that the stringized number is not some string like "Inf"
443 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
444 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
445 ((StrVal[0] == '-' || StrVal[0] == '+') &&
446 (StrVal[1] >= '0' && StrVal[1] <= '9')))
447 // Reparse stringized version!
448 return atof(StrVal.c_str()) == CFP->getValue();
453 // printConstant - The LLVM Constant to C Constant converter.
454 void CWriter::printConstant(Constant *CPV) {
455 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
456 switch (CE->getOpcode()) {
457 case Instruction::Cast:
459 printType(Out, CPV->getType());
461 printConstant(CE->getOperand(0));
465 case Instruction::GetElementPtr:
467 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
471 case Instruction::Select:
473 printConstant(CE->getOperand(0));
475 printConstant(CE->getOperand(1));
477 printConstant(CE->getOperand(2));
480 case Instruction::Add:
481 case Instruction::Sub:
482 case Instruction::Mul:
483 case Instruction::Div:
484 case Instruction::Rem:
485 case Instruction::SetEQ:
486 case Instruction::SetNE:
487 case Instruction::SetLT:
488 case Instruction::SetLE:
489 case Instruction::SetGT:
490 case Instruction::SetGE:
491 case Instruction::Shl:
492 case Instruction::Shr:
494 printConstant(CE->getOperand(0));
495 switch (CE->getOpcode()) {
496 case Instruction::Add: Out << " + "; break;
497 case Instruction::Sub: Out << " - "; break;
498 case Instruction::Mul: Out << " * "; break;
499 case Instruction::Div: Out << " / "; break;
500 case Instruction::Rem: Out << " % "; break;
501 case Instruction::SetEQ: Out << " == "; break;
502 case Instruction::SetNE: Out << " != "; break;
503 case Instruction::SetLT: Out << " < "; break;
504 case Instruction::SetLE: Out << " <= "; break;
505 case Instruction::SetGT: Out << " > "; break;
506 case Instruction::SetGE: Out << " >= "; break;
507 case Instruction::Shl: Out << " << "; break;
508 case Instruction::Shr: Out << " >> "; break;
509 default: assert(0 && "Illegal opcode here!");
511 printConstant(CE->getOperand(1));
516 std::cerr << "CWriter Error: Unhandled constant expression: "
522 switch (CPV->getType()->getTypeID()) {
524 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
525 case Type::SByteTyID:
526 case Type::ShortTyID:
527 Out << cast<ConstantSInt>(CPV)->getValue(); break;
529 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
530 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
532 Out << cast<ConstantSInt>(CPV)->getValue();
536 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
538 case Type::UByteTyID:
539 case Type::UShortTyID:
540 Out << cast<ConstantUInt>(CPV)->getValue(); break;
542 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
543 case Type::ULongTyID:
544 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
546 case Type::FloatTyID:
547 case Type::DoubleTyID: {
548 ConstantFP *FPC = cast<ConstantFP>(CPV);
549 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
550 if (I != FPConstantMap.end()) {
551 // Because of FP precision problems we must load from a stack allocated
552 // value that holds the value in hex.
553 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
554 << "*)&FPConstant" << I->second << ")";
557 // Print out the constant as a floating point number.
559 sprintf(Buffer, "%a", FPC->getValue());
560 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
562 Out << ftostr(FPC->getValue());
568 case Type::ArrayTyID:
569 if (isa<ConstantAggregateZero>(CPV)) {
570 const ArrayType *AT = cast<ArrayType>(CPV->getType());
572 if (AT->getNumElements()) {
574 Constant *CZ = Constant::getNullValue(AT->getElementType());
576 for (unsigned i = 1, e = AT->getNumElements(); i != e; ++i) {
583 printConstantArray(cast<ConstantArray>(CPV));
587 case Type::StructTyID:
588 if (isa<ConstantAggregateZero>(CPV)) {
589 const StructType *ST = cast<StructType>(CPV->getType());
591 if (ST->getNumElements()) {
593 printConstant(Constant::getNullValue(ST->getElementType(0)));
594 for (unsigned i = 1, e = ST->getNumElements(); i != e; ++i) {
596 printConstant(Constant::getNullValue(ST->getElementType(i)));
602 if (CPV->getNumOperands()) {
604 printConstant(cast<Constant>(CPV->getOperand(0)));
605 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
607 printConstant(cast<Constant>(CPV->getOperand(i)));
614 case Type::PointerTyID:
615 if (isa<ConstantPointerNull>(CPV)) {
617 printType(Out, CPV->getType());
618 Out << ")/*NULL*/0)";
620 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
621 writeOperand(CPR->getValue());
626 std::cerr << "Unknown constant type: " << CPV << "\n";
631 void CWriter::writeOperandInternal(Value *Operand) {
632 if (Instruction *I = dyn_cast<Instruction>(Operand))
633 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
634 // Should we inline this instruction to build a tree?
641 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
644 Out << Mang->getValueName(Operand);
648 void CWriter::writeOperand(Value *Operand) {
649 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
650 Out << "(&"; // Global variables are references as their addresses by llvm
652 writeOperandInternal(Operand);
654 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
658 // generateCompilerSpecificCode - This is where we add conditional compilation
659 // directives to cater to specific compilers as need be.
661 static void generateCompilerSpecificCode(std::ostream& Out) {
662 // Alloca is hard to get, and we don't want to include stdlib.h here...
663 Out << "/* get a declaration for alloca */\n"
664 << "#if defined(sun) || defined(__CYGWIN__)\n"
665 << "extern void *__builtin_alloca(unsigned long);\n"
666 << "#define alloca(x) __builtin_alloca(x)\n"
668 << "#ifndef __FreeBSD__\n"
669 << "#include <alloca.h>\n"
673 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
674 // If we aren't being compiled with GCC, just drop these attributes.
675 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
676 << "#define __attribute__(X)\n"
680 // At some point, we should support "external weak" vs. "weak" linkages.
681 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
682 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
683 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
684 << "#elif defined(__GNUC__)\n"
685 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
687 << "#define __EXTERNAL_WEAK__\n"
691 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
692 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
693 << "#define __ATTRIBUTE_WEAK__\n"
694 << "#elif defined(__GNUC__)\n"
695 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
697 << "#define __ATTRIBUTE_WEAK__\n"
701 bool CWriter::doInitialization(Module &M) {
707 // Ensure that all structure types have names...
708 Mang = new Mangler(M);
710 // get declaration for alloca
711 Out << "/* Provide Declarations */\n";
712 Out << "#include <stdarg.h>\n"; // Varargs support
713 Out << "#include <setjmp.h>\n"; // Unwind support
714 generateCompilerSpecificCode(Out);
716 // Provide a definition for `bool' if not compiling with a C++ compiler.
718 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
720 << "\n\n/* Support for floating point constants */\n"
721 << "typedef unsigned long long ConstantDoubleTy;\n"
722 << "typedef unsigned int ConstantFloatTy;\n"
724 << "\n\n/* Global Declarations */\n";
726 // First output all the declarations for the program, because C requires
727 // Functions & globals to be declared before they are used.
730 // Loop over the symbol table, emitting all named constants...
731 printModuleTypes(M.getSymbolTable());
733 // Global variable declarations...
735 Out << "\n/* External Global Variable Declarations */\n";
736 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I) {
737 if (I->hasExternalLinkage()) {
739 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
745 // Function declarations
747 Out << "\n/* Function Declarations */\n";
748 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
749 // Don't print declarations for intrinsic functions.
750 if (!I->getIntrinsicID() &&
751 I->getName() != "setjmp" && I->getName() != "longjmp") {
752 printFunctionSignature(I, true);
753 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
754 if (I->hasLinkOnceLinkage()) Out << " __ATTRIBUTE_WEAK__";
760 // Output the global variable declarations
762 Out << "\n\n/* Global Variable Declarations */\n";
763 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
764 if (!I->isExternal()) {
766 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
768 if (I->hasLinkOnceLinkage())
769 Out << " __attribute__((common))";
770 else if (I->hasWeakLinkage())
771 Out << " __ATTRIBUTE_WEAK__";
776 // Output the global variable definitions and contents...
778 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
779 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
780 if (!I->isExternal()) {
781 if (I->hasInternalLinkage())
783 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
784 if (I->hasLinkOnceLinkage())
785 Out << " __attribute__((common))";
786 else if (I->hasWeakLinkage())
787 Out << " __ATTRIBUTE_WEAK__";
789 // If the initializer is not null, emit the initializer. If it is null,
790 // we try to avoid emitting large amounts of zeros. The problem with
791 // this, however, occurs when the variable has weak linkage. In this
792 // case, the assembler will complain about the variable being both weak
793 // and common, so we disable this optimization.
794 if (!I->getInitializer()->isNullValue()) {
796 writeOperand(I->getInitializer());
797 } else if (I->hasWeakLinkage()) {
798 // We have to specify an initializer, but it doesn't have to be
799 // complete. If the value is an aggregate, print out { 0 }, and let
800 // the compiler figure out the rest of the zeros.
802 if (isa<StructType>(I->getInitializer()->getType()) ||
803 isa<ArrayType>(I->getInitializer()->getType())) {
806 // Just print it out normally.
807 writeOperand(I->getInitializer());
815 Out << "\n\n/* Function Bodies */\n";
820 /// Output all floating point constants that cannot be printed accurately...
821 void CWriter::printFloatingPointConstants(Function &F) {
832 // Scan the module for floating point constants. If any FP constant is used
833 // in the function, we want to redirect it here so that we do not depend on
834 // the precision of the printed form, unless the printed form preserves
837 static unsigned FPCounter = 0;
838 for (constant_iterator I = constant_begin(&F), E = constant_end(&F);
840 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
841 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
842 !FPConstantMap.count(FPC)) {
843 double Val = FPC->getValue();
845 FPConstantMap[FPC] = FPCounter; // Number the FP constants
847 if (FPC->getType() == Type::DoubleTy) {
849 Out << "static const ConstantDoubleTy FPConstant" << FPCounter++
850 << " = 0x" << std::hex << DBLUnion.U << std::dec
851 << "ULL; /* " << Val << " */\n";
852 } else if (FPC->getType() == Type::FloatTy) {
854 Out << "static const ConstantFloatTy FPConstant" << FPCounter++
855 << " = 0x" << std::hex << FLTUnion.U << std::dec
856 << "U; /* " << Val << " */\n";
858 assert(0 && "Unknown float type!");
865 /// printSymbolTable - Run through symbol table looking for type names. If a
866 /// type name is found, emit it's declaration...
868 void CWriter::printModuleTypes(const SymbolTable &ST) {
869 // If there are no type names, exit early.
870 if ( ! ST.hasTypes() )
873 // We are only interested in the type plane of the symbol table...
874 SymbolTable::type_const_iterator I = ST.type_begin();
875 SymbolTable::type_const_iterator End = ST.type_end();
877 // Print out forward declarations for structure types before anything else!
878 Out << "/* Structure forward decls */\n";
879 for (; I != End; ++I)
880 if (const Type *STy = dyn_cast<StructType>(I->second)) {
881 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
882 Out << Name << ";\n";
883 TypeNames.insert(std::make_pair(STy, Name));
888 // Now we can print out typedefs...
889 Out << "/* Typedefs */\n";
890 for (I = ST.type_begin(); I != End; ++I) {
891 const Type *Ty = cast<Type>(I->second);
892 std::string Name = "l_" + Mangler::makeNameProper(I->first);
894 printType(Out, Ty, Name);
900 // Keep track of which structures have been printed so far...
901 std::set<const StructType *> StructPrinted;
903 // Loop over all structures then push them into the stack so they are
904 // printed in the correct order.
906 Out << "/* Structure contents */\n";
907 for (I = ST.type_begin(); I != End; ++I)
908 if (const StructType *STy = dyn_cast<StructType>(I->second))
909 // Only print out used types!
910 printContainedStructs(STy, StructPrinted);
913 // Push the struct onto the stack and recursively push all structs
914 // this one depends on.
915 void CWriter::printContainedStructs(const Type *Ty,
916 std::set<const StructType*> &StructPrinted){
917 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
918 //Check to see if we have already printed this struct
919 if (StructPrinted.count(STy) == 0) {
920 // Print all contained types first...
921 for (StructType::element_iterator I = STy->element_begin(),
922 E = STy->element_end(); I != E; ++I) {
923 const Type *Ty1 = I->get();
924 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
925 printContainedStructs(*I, StructPrinted);
928 //Print structure type out..
929 StructPrinted.insert(STy);
930 std::string Name = TypeNames[STy];
931 printType(Out, STy, Name, true);
935 // If it is an array, check contained types and continue
936 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
937 const Type *Ty1 = ATy->getElementType();
938 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
939 printContainedStructs(Ty1, StructPrinted);
944 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
945 if (F->hasInternalLinkage()) Out << "static ";
947 // Loop over the arguments, printing them...
948 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
950 std::stringstream FunctionInnards;
952 // Print out the name...
953 FunctionInnards << Mang->getValueName(F) << "(";
955 if (!F->isExternal()) {
958 if (F->abegin()->hasName() || !Prototype)
959 ArgName = Mang->getValueName(F->abegin());
960 printType(FunctionInnards, F->afront().getType(), ArgName);
961 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
963 FunctionInnards << ", ";
964 if (I->hasName() || !Prototype)
965 ArgName = Mang->getValueName(I);
968 printType(FunctionInnards, I->getType(), ArgName);
972 // Loop over the arguments, printing them...
973 for (FunctionType::param_iterator I = FT->param_begin(),
974 E = FT->param_end(); I != E; ++I) {
975 if (I != FT->param_begin()) FunctionInnards << ", ";
976 printType(FunctionInnards, *I);
980 // Finish printing arguments... if this is a vararg function, print the ...,
981 // unless there are no known types, in which case, we just emit ().
983 if (FT->isVarArg() && FT->getNumParams()) {
984 if (FT->getNumParams()) FunctionInnards << ", ";
985 FunctionInnards << "..."; // Output varargs portion of signature!
986 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
987 FunctionInnards << "void"; // ret() -> ret(void) in C.
989 FunctionInnards << ")";
990 // Print out the return type and the entire signature for that matter
991 printType(Out, F->getReturnType(), FunctionInnards.str());
994 void CWriter::printFunction(Function &F) {
995 printFunctionSignature(&F, false);
998 // print local variable information for the function
999 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I)
1000 if (const AllocaInst *AI = isDirectAlloca(&*I)) {
1002 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
1003 Out << "; /* Address exposed local */\n";
1004 } else if (I->getType() != Type::VoidTy && !isInlinableInst(*I)) {
1006 printType(Out, I->getType(), Mang->getValueName(&*I));
1009 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
1011 printType(Out, I->getType(),
1012 Mang->getValueName(&*I)+"__PHI_TEMPORARY");
1019 // print the basic blocks
1020 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1021 if (Loop *L = LI->getLoopFor(BB)) {
1022 if (L->getHeader() == BB && L->getParentLoop() == 0)
1025 printBasicBlock(BB);
1032 void CWriter::printLoop(Loop *L) {
1033 Out << " do { /* Syntactic loop '" << L->getHeader()->getName()
1034 << "' to make GCC happy */\n";
1035 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
1036 BasicBlock *BB = L->getBlocks()[i];
1037 Loop *BBLoop = LI->getLoopFor(BB);
1039 printBasicBlock(BB);
1040 else if (BB == BBLoop->getHeader() && BBLoop->getParentLoop() == L)
1043 Out << " } while (1); /* end of syntactic loop '"
1044 << L->getHeader()->getName() << "' */\n";
1047 void CWriter::printBasicBlock(BasicBlock *BB) {
1049 // Don't print the label for the basic block if there are no uses, or if
1050 // the only terminator use is the predecessor basic block's terminator.
1051 // We have to scan the use list because PHI nodes use basic blocks too but
1052 // do not require a label to be generated.
1054 bool NeedsLabel = false;
1055 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1056 if (isGotoCodeNecessary(*PI, BB)) {
1061 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1063 // Output all of the instructions in the basic block...
1064 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E;
1066 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1067 if (II->getType() != Type::VoidTy)
1076 // Don't emit prefix or suffix for the terminator...
1077 visit(*BB->getTerminator());
1081 // Specific Instruction type classes... note that all of the casts are
1082 // necessary because we use the instruction classes as opaque types...
1084 void CWriter::visitReturnInst(ReturnInst &I) {
1085 // Don't output a void return if this is the last basic block in the function
1086 if (I.getNumOperands() == 0 &&
1087 &*--I.getParent()->getParent()->end() == I.getParent() &&
1088 !I.getParent()->size() == 1) {
1093 if (I.getNumOperands()) {
1095 writeOperand(I.getOperand(0));
1100 void CWriter::visitSwitchInst(SwitchInst &SI) {
1101 printPHICopiesForSuccessors(SI.getParent(), 0);
1104 writeOperand(SI.getOperand(0));
1105 Out << ") {\n default:\n";
1106 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1108 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1110 writeOperand(SI.getOperand(i));
1112 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1113 printBranchToBlock(SI.getParent(), Succ, 2);
1114 if (Succ == SI.getParent()->getNext())
1120 bool CWriter::isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1121 /// FIXME: This should be reenabled, but loop reordering safe!!
1124 if (From->getNext() != To) // Not the direct successor, we need a goto
1127 //isa<SwitchInst>(From->getTerminator())
1130 if (LI->getLoopFor(From) != LI->getLoopFor(To))
1135 void CWriter::printPHICopiesForSuccessors(BasicBlock *CurBlock,
1137 for (succ_iterator SI = succ_begin(CurBlock), E = succ_end(CurBlock);
1139 for (BasicBlock::iterator I = SI->begin();
1140 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1141 // now we have to do the printing
1142 Out << std::string(Indent, ' ');
1143 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1144 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBlock)));
1145 Out << "; /* for PHI node */\n";
1150 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1152 if (isGotoCodeNecessary(CurBB, Succ)) {
1153 Out << std::string(Indent, ' ') << " goto ";
1159 // Branch instruction printing - Avoid printing out a branch to a basic block
1160 // that immediately succeeds the current one.
1162 void CWriter::visitBranchInst(BranchInst &I) {
1163 printPHICopiesForSuccessors(I.getParent(), 0);
1165 if (I.isConditional()) {
1166 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1168 writeOperand(I.getCondition());
1171 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1173 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1174 Out << " } else {\n";
1175 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1178 // First goto not necessary, assume second one is...
1180 writeOperand(I.getCondition());
1183 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1188 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1193 // PHI nodes get copied into temporary values at the end of predecessor basic
1194 // blocks. We now need to copy these temporary values into the REAL value for
1196 void CWriter::visitPHINode(PHINode &I) {
1198 Out << "__PHI_TEMPORARY";
1202 void CWriter::visitBinaryOperator(Instruction &I) {
1203 // binary instructions, shift instructions, setCond instructions.
1204 assert(!isa<PointerType>(I.getType()));
1206 // We must cast the results of binary operations which might be promoted.
1207 bool needsCast = false;
1208 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1209 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1210 || (I.getType() == Type::FloatTy)) {
1213 printType(Out, I.getType());
1217 writeOperand(I.getOperand(0));
1219 switch (I.getOpcode()) {
1220 case Instruction::Add: Out << " + "; break;
1221 case Instruction::Sub: Out << " - "; break;
1222 case Instruction::Mul: Out << "*"; break;
1223 case Instruction::Div: Out << "/"; break;
1224 case Instruction::Rem: Out << "%"; break;
1225 case Instruction::And: Out << " & "; break;
1226 case Instruction::Or: Out << " | "; break;
1227 case Instruction::Xor: Out << " ^ "; break;
1228 case Instruction::SetEQ: Out << " == "; break;
1229 case Instruction::SetNE: Out << " != "; break;
1230 case Instruction::SetLE: Out << " <= "; break;
1231 case Instruction::SetGE: Out << " >= "; break;
1232 case Instruction::SetLT: Out << " < "; break;
1233 case Instruction::SetGT: Out << " > "; break;
1234 case Instruction::Shl : Out << " << "; break;
1235 case Instruction::Shr : Out << " >> "; break;
1236 default: std::cerr << "Invalid operator type!" << I; abort();
1239 writeOperand(I.getOperand(1));
1246 void CWriter::visitCastInst(CastInst &I) {
1247 if (I.getType() == Type::BoolTy) {
1249 writeOperand(I.getOperand(0));
1254 printType(Out, I.getType());
1256 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1257 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1258 // Avoid "cast to pointer from integer of different size" warnings
1262 writeOperand(I.getOperand(0));
1265 void CWriter::visitSelectInst(SelectInst &I) {
1267 writeOperand(I.getCondition());
1269 writeOperand(I.getTrueValue());
1271 writeOperand(I.getFalseValue());
1276 void CWriter::lowerIntrinsics(Function &F) {
1277 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1278 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1279 if (CallInst *CI = dyn_cast<CallInst>(I++))
1280 if (Function *F = CI->getCalledFunction())
1281 switch (F->getIntrinsicID()) {
1282 case Intrinsic::not_intrinsic:
1283 case Intrinsic::vastart:
1284 case Intrinsic::vacopy:
1285 case Intrinsic::vaend:
1286 case Intrinsic::returnaddress:
1287 case Intrinsic::frameaddress:
1288 case Intrinsic::setjmp:
1289 case Intrinsic::longjmp:
1290 // We directly implement these intrinsics
1293 // All other intrinsic calls we must lower.
1294 Instruction *Before = CI->getPrev();
1295 IL.LowerIntrinsicCall(CI);
1296 if (Before) { // Move iterator to instruction after call
1306 void CWriter::visitCallInst(CallInst &I) {
1307 // Handle intrinsic function calls first...
1308 if (Function *F = I.getCalledFunction())
1309 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1311 default: assert(0 && "Unknown LLVM intrinsic!");
1312 case Intrinsic::vastart:
1315 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1316 // Output the last argument to the enclosing function...
1317 if (I.getParent()->getParent()->aempty()) {
1318 std::cerr << "The C backend does not currently support zero "
1319 << "argument varargs functions, such as '"
1320 << I.getParent()->getParent()->getName() << "'!\n";
1323 writeOperand(&I.getParent()->getParent()->aback());
1326 case Intrinsic::vaend:
1327 Out << "va_end(*(va_list*)&";
1328 writeOperand(I.getOperand(1));
1331 case Intrinsic::vacopy:
1333 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1334 Out << "*(va_list*)&";
1335 writeOperand(I.getOperand(1));
1338 case Intrinsic::returnaddress:
1339 Out << "__builtin_return_address(";
1340 writeOperand(I.getOperand(1));
1343 case Intrinsic::frameaddress:
1344 Out << "__builtin_frame_address(";
1345 writeOperand(I.getOperand(1));
1348 case Intrinsic::setjmp:
1349 Out << "setjmp(*(jmp_buf*)";
1350 writeOperand(I.getOperand(1));
1353 case Intrinsic::longjmp:
1354 Out << "longjmp(*(jmp_buf*)";
1355 writeOperand(I.getOperand(1));
1357 writeOperand(I.getOperand(2));
1365 void CWriter::visitCallSite(CallSite CS) {
1366 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1367 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1368 const Type *RetTy = FTy->getReturnType();
1370 writeOperand(CS.getCalledValue());
1373 if (CS.arg_begin() != CS.arg_end()) {
1374 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1377 for (++AI; AI != AE; ++AI) {
1385 void CWriter::visitMallocInst(MallocInst &I) {
1386 assert(0 && "lowerallocations pass didn't work!");
1389 void CWriter::visitAllocaInst(AllocaInst &I) {
1391 printType(Out, I.getType());
1392 Out << ") alloca(sizeof(";
1393 printType(Out, I.getType()->getElementType());
1395 if (I.isArrayAllocation()) {
1397 writeOperand(I.getOperand(0));
1402 void CWriter::visitFreeInst(FreeInst &I) {
1403 assert(0 && "lowerallocations pass didn't work!");
1406 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1407 gep_type_iterator E) {
1408 bool HasImplicitAddress = false;
1409 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1410 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1411 HasImplicitAddress = true;
1412 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1413 HasImplicitAddress = true;
1414 Ptr = CPR->getValue(); // Get to the global...
1415 } else if (isDirectAlloca(Ptr)) {
1416 HasImplicitAddress = true;
1420 if (!HasImplicitAddress)
1421 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1423 writeOperandInternal(Ptr);
1427 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1428 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1431 writeOperandInternal(Ptr);
1433 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1435 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1438 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1439 "Can only have implicit address with direct accessing");
1441 if (HasImplicitAddress) {
1443 } else if (CI && CI->isNullValue()) {
1444 gep_type_iterator TmpI = I; ++TmpI;
1446 // Print out the -> operator if possible...
1447 if (TmpI != E && isa<StructType>(*TmpI)) {
1448 Out << (HasImplicitAddress ? "." : "->");
1449 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1455 if (isa<StructType>(*I)) {
1456 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1459 writeOperand(I.getOperand());
1464 void CWriter::visitLoadInst(LoadInst &I) {
1466 writeOperand(I.getOperand(0));
1469 void CWriter::visitStoreInst(StoreInst &I) {
1471 writeOperand(I.getPointerOperand());
1473 writeOperand(I.getOperand(0));
1476 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1478 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1482 void CWriter::visitVANextInst(VANextInst &I) {
1483 Out << Mang->getValueName(I.getOperand(0));
1484 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1485 printType(Out, I.getArgType());
1489 void CWriter::visitVAArgInst(VAArgInst &I) {
1491 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1492 writeOperand(I.getOperand(0));
1493 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1494 printType(Out, I.getType());
1495 Out << ");\n va_end(Tmp); }";
1498 //===----------------------------------------------------------------------===//
1499 // External Interface declaration
1500 //===----------------------------------------------------------------------===//
1502 bool CTargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &o) {
1503 PM.add(createLowerGCPass());
1504 PM.add(createLowerAllocationsPass());
1505 PM.add(createLowerInvokePass());
1506 PM.add(new CBackendNameAllUsedStructs());
1507 PM.add(new CWriter(o, getIntrinsicLowering()));
1511 TargetMachine *llvm::allocateCTargetMachine(const Module &M,
1512 IntrinsicLowering *IL) {
1513 return new CTargetMachine(M, IL);