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/Target/TargetMachineRegistry.h"
31 #include "llvm/Support/CallSite.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/InstVisitor.h"
35 #include "llvm/Support/Mangler.h"
36 #include "Support/StringExtras.h"
37 #include "Config/config.h"
44 // Register the target.
45 RegisterTarget<CTargetMachine> X("c", "C backend");
47 /// NameAllUsedStructs - This pass inserts names for any unnamed structure
48 /// types that are used by the program.
50 class CBackendNameAllUsedStructs : public Pass {
51 void getAnalysisUsage(AnalysisUsage &AU) const {
52 AU.addRequired<FindUsedTypes>();
55 virtual const char *getPassName() const {
56 return "C backend type canonicalizer";
59 virtual bool run(Module &M);
62 /// CWriter - This class is the main chunk of code that converts an LLVM
63 /// module to a C translation unit.
64 class CWriter : public FunctionPass, public InstVisitor<CWriter> {
66 IntrinsicLowering &IL;
69 const Module *TheModule;
70 std::map<const Type *, std::string> TypeNames;
72 std::map<const ConstantFP *, unsigned> FPConstantMap;
74 CWriter(std::ostream &o, IntrinsicLowering &il) : Out(o), IL(il) {}
76 virtual const char *getPassName() const { return "C backend"; }
78 void getAnalysisUsage(AnalysisUsage &AU) const {
79 AU.addRequired<LoopInfo>();
83 virtual bool doInitialization(Module &M);
85 bool runOnFunction(Function &F) {
86 LI = &getAnalysis<LoopInfo>();
88 // Output all floating point constants that cannot be printed accurately.
89 printFloatingPointConstants(F);
93 FPConstantMap.clear();
97 virtual bool doFinalization(Module &M) {
104 std::ostream &printType(std::ostream &Out, const Type *Ty,
105 const std::string &VariableName = "",
106 bool IgnoreName = false);
108 void writeOperand(Value *Operand);
109 void writeOperandInternal(Value *Operand);
112 void lowerIntrinsics(Function &F);
114 bool nameAllUsedStructureTypes(Module &M);
115 void printModule(Module *M);
116 void printModuleTypes(const SymbolTable &ST);
117 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
118 void printFloatingPointConstants(Function &F);
119 void printFunctionSignature(const Function *F, bool Prototype);
121 void printFunction(Function &);
122 void printBasicBlock(BasicBlock *BB);
123 void printLoop(Loop *L);
125 void printConstant(Constant *CPV);
126 void printConstantArray(ConstantArray *CPA);
128 // isInlinableInst - Attempt to inline instructions into their uses to build
129 // trees as much as possible. To do this, we have to consistently decide
130 // what is acceptable to inline, so that variable declarations don't get
131 // printed and an extra copy of the expr is not emitted.
133 static bool isInlinableInst(const Instruction &I) {
134 // Always inline setcc instructions, even if they are shared by multiple
135 // expressions. GCC generates horrible code if we don't.
136 if (isa<SetCondInst>(I)) return true;
138 // Must be an expression, must be used exactly once. If it is dead, we
139 // emit it inline where it would go.
140 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
141 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
142 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
143 // Don't inline a load across a store or other bad things!
146 // Only inline instruction it it's use is in the same BB as the inst.
147 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
150 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
151 // variables which are accessed with the & operator. This causes GCC to
152 // generate significantly better code than to emit alloca calls directly.
154 static const AllocaInst *isDirectAlloca(const Value *V) {
155 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
156 if (!AI) return false;
157 if (AI->isArrayAllocation())
158 return 0; // FIXME: we can also inline fixed size array allocas!
159 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
164 // Instruction visitation functions
165 friend class InstVisitor<CWriter>;
167 void visitReturnInst(ReturnInst &I);
168 void visitBranchInst(BranchInst &I);
169 void visitSwitchInst(SwitchInst &I);
170 void visitInvokeInst(InvokeInst &I) {
171 assert(0 && "Lowerinvoke pass didn't work!");
174 void visitUnwindInst(UnwindInst &I) {
175 assert(0 && "Lowerinvoke pass didn't work!");
178 void visitPHINode(PHINode &I);
179 void visitBinaryOperator(Instruction &I);
181 void visitCastInst (CastInst &I);
182 void visitSelectInst(SelectInst &I);
183 void visitCallInst (CallInst &I);
184 void visitCallSite (CallSite CS);
185 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
187 void visitMallocInst(MallocInst &I);
188 void visitAllocaInst(AllocaInst &I);
189 void visitFreeInst (FreeInst &I);
190 void visitLoadInst (LoadInst &I);
191 void visitStoreInst (StoreInst &I);
192 void visitGetElementPtrInst(GetElementPtrInst &I);
193 void visitVANextInst(VANextInst &I);
194 void visitVAArgInst (VAArgInst &I);
196 void visitInstruction(Instruction &I) {
197 std::cerr << "C Writer does not know about " << I;
201 void outputLValue(Instruction *I) {
202 Out << " " << Mang->getValueName(I) << " = ";
205 bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To);
206 void printPHICopiesForSuccessors(BasicBlock *CurBlock,
208 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
210 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
211 gep_type_iterator E);
215 /// This method inserts names for any unnamed structure types that are used by
216 /// the program, and removes names from structure types that are not used by the
219 bool CBackendNameAllUsedStructs::run(Module &M) {
220 // Get a set of types that are used by the program...
221 std::set<const Type *> UT = getAnalysis<FindUsedTypes>().getTypes();
223 // Loop over the module symbol table, removing types from UT that are
224 // already named, and removing names for structure types that are not used.
226 SymbolTable &MST = M.getSymbolTable();
227 for (SymbolTable::type_iterator TI = MST.type_begin(), TE = MST.type_end();
229 SymbolTable::type_iterator I = TI++;
230 if (const StructType *STy = dyn_cast<StructType>(I->second)) {
231 // If this is not used, remove it from the symbol table.
232 std::set<const Type *>::iterator UTI = UT.find(STy);
240 // UT now contains types that are not named. Loop over it, naming
243 bool Changed = false;
244 unsigned RenameCounter = 0;
245 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
247 if (const StructType *ST = dyn_cast<StructType>(*I)) {
248 while (M.addTypeName("unnamed"+utostr(RenameCounter), ST))
256 // Pass the Type* and the variable name and this prints out the variable
259 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
260 const std::string &NameSoFar,
262 if (Ty->isPrimitiveType())
263 switch (Ty->getTypeID()) {
264 case Type::VoidTyID: return Out << "void " << NameSoFar;
265 case Type::BoolTyID: return Out << "bool " << NameSoFar;
266 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
267 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
268 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
269 case Type::ShortTyID: return Out << "short " << NameSoFar;
270 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
271 case Type::IntTyID: return Out << "int " << NameSoFar;
272 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
273 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
274 case Type::FloatTyID: return Out << "float " << NameSoFar;
275 case Type::DoubleTyID: return Out << "double " << NameSoFar;
277 std::cerr << "Unknown primitive type: " << Ty << "\n";
281 // Check to see if the type is named.
282 if (!IgnoreName || isa<OpaqueType>(Ty)) {
283 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
284 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
287 switch (Ty->getTypeID()) {
288 case Type::FunctionTyID: {
289 const FunctionType *MTy = cast<FunctionType>(Ty);
290 std::stringstream FunctionInnards;
291 FunctionInnards << " (" << NameSoFar << ") (";
292 for (FunctionType::param_iterator I = MTy->param_begin(),
293 E = MTy->param_end(); I != E; ++I) {
294 if (I != MTy->param_begin())
295 FunctionInnards << ", ";
296 printType(FunctionInnards, *I, "");
298 if (MTy->isVarArg()) {
299 if (MTy->getNumParams())
300 FunctionInnards << ", ...";
301 } else if (!MTy->getNumParams()) {
302 FunctionInnards << "void";
304 FunctionInnards << ")";
305 std::string tstr = FunctionInnards.str();
306 printType(Out, MTy->getReturnType(), tstr);
309 case Type::StructTyID: {
310 const StructType *STy = cast<StructType>(Ty);
311 Out << NameSoFar + " {\n";
313 for (StructType::element_iterator I = STy->element_begin(),
314 E = STy->element_end(); I != E; ++I) {
316 printType(Out, *I, "field" + utostr(Idx++));
322 case Type::PointerTyID: {
323 const PointerType *PTy = cast<PointerType>(Ty);
324 std::string ptrName = "*" + NameSoFar;
326 if (isa<ArrayType>(PTy->getElementType()))
327 ptrName = "(" + ptrName + ")";
329 return printType(Out, PTy->getElementType(), ptrName);
332 case Type::ArrayTyID: {
333 const ArrayType *ATy = cast<ArrayType>(Ty);
334 unsigned NumElements = ATy->getNumElements();
335 return printType(Out, ATy->getElementType(),
336 NameSoFar + "[" + utostr(NumElements) + "]");
339 case Type::OpaqueTyID: {
340 static int Count = 0;
341 std::string TyName = "struct opaque_" + itostr(Count++);
342 assert(TypeNames.find(Ty) == TypeNames.end());
343 TypeNames[Ty] = TyName;
344 return Out << TyName << " " << NameSoFar;
347 assert(0 && "Unhandled case in getTypeProps!");
354 void CWriter::printConstantArray(ConstantArray *CPA) {
356 // As a special case, print the array as a string if it is an array of
357 // ubytes or an array of sbytes with positive values.
359 const Type *ETy = CPA->getType()->getElementType();
360 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
362 // Make sure the last character is a null char, as automatically added by C
363 if (isString && (CPA->getNumOperands() == 0 ||
364 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
369 // Keep track of whether the last number was a hexadecimal escape
370 bool LastWasHex = false;
372 // Do not include the last character, which we know is null
373 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
374 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
376 // Print it out literally if it is a printable character. The only thing
377 // to be careful about is when the last letter output was a hex escape
378 // code, in which case we have to be careful not to print out hex digits
379 // explicitly (the C compiler thinks it is a continuation of the previous
380 // character, sheesh...)
382 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
384 if (C == '"' || C == '\\')
391 case '\n': Out << "\\n"; break;
392 case '\t': Out << "\\t"; break;
393 case '\r': Out << "\\r"; break;
394 case '\v': Out << "\\v"; break;
395 case '\a': Out << "\\a"; break;
396 case '\"': Out << "\\\""; break;
397 case '\'': Out << "\\\'"; break;
400 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
401 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
410 if (CPA->getNumOperands()) {
412 printConstant(cast<Constant>(CPA->getOperand(0)));
413 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
415 printConstant(cast<Constant>(CPA->getOperand(i)));
422 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
423 // textually as a double (rather than as a reference to a stack-allocated
424 // variable). We decide this by converting CFP to a string and back into a
425 // double, and then checking whether the conversion results in a bit-equal
426 // double to the original value of CFP. This depends on us and the target C
427 // compiler agreeing on the conversion process (which is pretty likely since we
428 // only deal in IEEE FP).
430 static bool isFPCSafeToPrint(const ConstantFP *CFP) {
433 sprintf(Buffer, "%a", CFP->getValue());
435 if (!strncmp(Buffer, "0x", 2) ||
436 !strncmp(Buffer, "-0x", 3) ||
437 !strncmp(Buffer, "+0x", 3))
438 return atof(Buffer) == CFP->getValue();
441 std::string StrVal = ftostr(CFP->getValue());
443 while (StrVal[0] == ' ')
444 StrVal.erase(StrVal.begin());
446 // Check to make sure that the stringized number is not some string like "Inf"
447 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
448 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
449 ((StrVal[0] == '-' || StrVal[0] == '+') &&
450 (StrVal[1] >= '0' && StrVal[1] <= '9')))
451 // Reparse stringized version!
452 return atof(StrVal.c_str()) == CFP->getValue();
457 // printConstant - The LLVM Constant to C Constant converter.
458 void CWriter::printConstant(Constant *CPV) {
459 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
460 switch (CE->getOpcode()) {
461 case Instruction::Cast:
463 printType(Out, CPV->getType());
465 printConstant(CE->getOperand(0));
469 case Instruction::GetElementPtr:
471 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
475 case Instruction::Select:
477 printConstant(CE->getOperand(0));
479 printConstant(CE->getOperand(1));
481 printConstant(CE->getOperand(2));
484 case Instruction::Add:
485 case Instruction::Sub:
486 case Instruction::Mul:
487 case Instruction::Div:
488 case Instruction::Rem:
489 case Instruction::SetEQ:
490 case Instruction::SetNE:
491 case Instruction::SetLT:
492 case Instruction::SetLE:
493 case Instruction::SetGT:
494 case Instruction::SetGE:
495 case Instruction::Shl:
496 case Instruction::Shr:
498 printConstant(CE->getOperand(0));
499 switch (CE->getOpcode()) {
500 case Instruction::Add: Out << " + "; break;
501 case Instruction::Sub: Out << " - "; break;
502 case Instruction::Mul: Out << " * "; break;
503 case Instruction::Div: Out << " / "; break;
504 case Instruction::Rem: Out << " % "; break;
505 case Instruction::SetEQ: Out << " == "; break;
506 case Instruction::SetNE: Out << " != "; break;
507 case Instruction::SetLT: Out << " < "; break;
508 case Instruction::SetLE: Out << " <= "; break;
509 case Instruction::SetGT: Out << " > "; break;
510 case Instruction::SetGE: Out << " >= "; break;
511 case Instruction::Shl: Out << " << "; break;
512 case Instruction::Shr: Out << " >> "; break;
513 default: assert(0 && "Illegal opcode here!");
515 printConstant(CE->getOperand(1));
520 std::cerr << "CWriter Error: Unhandled constant expression: "
526 switch (CPV->getType()->getTypeID()) {
528 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
529 case Type::SByteTyID:
530 case Type::ShortTyID:
531 Out << cast<ConstantSInt>(CPV)->getValue(); break;
533 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
534 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
536 Out << cast<ConstantSInt>(CPV)->getValue();
540 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
542 case Type::UByteTyID:
543 case Type::UShortTyID:
544 Out << cast<ConstantUInt>(CPV)->getValue(); break;
546 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
547 case Type::ULongTyID:
548 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
550 case Type::FloatTyID:
551 case Type::DoubleTyID: {
552 ConstantFP *FPC = cast<ConstantFP>(CPV);
553 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
554 if (I != FPConstantMap.end()) {
555 // Because of FP precision problems we must load from a stack allocated
556 // value that holds the value in hex.
557 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
558 << "*)&FPConstant" << I->second << ")";
561 // Print out the constant as a floating point number.
563 sprintf(Buffer, "%a", FPC->getValue());
564 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
566 Out << ftostr(FPC->getValue());
572 case Type::ArrayTyID:
573 if (isa<ConstantAggregateZero>(CPV)) {
574 const ArrayType *AT = cast<ArrayType>(CPV->getType());
576 if (AT->getNumElements()) {
578 Constant *CZ = Constant::getNullValue(AT->getElementType());
580 for (unsigned i = 1, e = AT->getNumElements(); i != e; ++i) {
587 printConstantArray(cast<ConstantArray>(CPV));
591 case Type::StructTyID:
592 if (isa<ConstantAggregateZero>(CPV)) {
593 const StructType *ST = cast<StructType>(CPV->getType());
595 if (ST->getNumElements()) {
597 printConstant(Constant::getNullValue(ST->getElementType(0)));
598 for (unsigned i = 1, e = ST->getNumElements(); i != e; ++i) {
600 printConstant(Constant::getNullValue(ST->getElementType(i)));
606 if (CPV->getNumOperands()) {
608 printConstant(cast<Constant>(CPV->getOperand(0)));
609 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
611 printConstant(cast<Constant>(CPV->getOperand(i)));
618 case Type::PointerTyID:
619 if (isa<ConstantPointerNull>(CPV)) {
621 printType(Out, CPV->getType());
622 Out << ")/*NULL*/0)";
624 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
625 writeOperand(CPR->getValue());
630 std::cerr << "Unknown constant type: " << CPV << "\n";
635 void CWriter::writeOperandInternal(Value *Operand) {
636 if (Instruction *I = dyn_cast<Instruction>(Operand))
637 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
638 // Should we inline this instruction to build a tree?
645 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
648 Out << Mang->getValueName(Operand);
652 void CWriter::writeOperand(Value *Operand) {
653 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
654 Out << "(&"; // Global variables are references as their addresses by llvm
656 writeOperandInternal(Operand);
658 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
662 // generateCompilerSpecificCode - This is where we add conditional compilation
663 // directives to cater to specific compilers as need be.
665 static void generateCompilerSpecificCode(std::ostream& Out) {
666 // Alloca is hard to get, and we don't want to include stdlib.h here...
667 Out << "/* get a declaration for alloca */\n"
668 << "#if defined(sun) || defined(__CYGWIN__)\n"
669 << "extern void *__builtin_alloca(unsigned long);\n"
670 << "#define alloca(x) __builtin_alloca(x)\n"
672 << "#ifndef __FreeBSD__\n"
673 << "#include <alloca.h>\n"
677 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
678 // If we aren't being compiled with GCC, just drop these attributes.
679 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
680 << "#define __attribute__(X)\n"
684 // At some point, we should support "external weak" vs. "weak" linkages.
685 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
686 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
687 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
688 << "#elif defined(__GNUC__)\n"
689 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
691 << "#define __EXTERNAL_WEAK__\n"
695 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
696 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
697 << "#define __ATTRIBUTE_WEAK__\n"
698 << "#elif defined(__GNUC__)\n"
699 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
701 << "#define __ATTRIBUTE_WEAK__\n"
705 bool CWriter::doInitialization(Module &M) {
711 // Ensure that all structure types have names...
712 Mang = new Mangler(M);
714 // get declaration for alloca
715 Out << "/* Provide Declarations */\n";
716 Out << "#include <stdarg.h>\n"; // Varargs support
717 Out << "#include <setjmp.h>\n"; // Unwind support
718 generateCompilerSpecificCode(Out);
720 // Provide a definition for `bool' if not compiling with a C++ compiler.
722 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
724 << "\n\n/* Support for floating point constants */\n"
725 << "typedef unsigned long long ConstantDoubleTy;\n"
726 << "typedef unsigned int ConstantFloatTy;\n"
728 << "\n\n/* Global Declarations */\n";
730 // First output all the declarations for the program, because C requires
731 // Functions & globals to be declared before they are used.
734 // Loop over the symbol table, emitting all named constants...
735 printModuleTypes(M.getSymbolTable());
737 // Global variable declarations...
739 Out << "\n/* External Global Variable Declarations */\n";
740 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I) {
741 if (I->hasExternalLinkage()) {
743 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
749 // Function declarations
751 Out << "\n/* Function Declarations */\n";
752 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
753 // Don't print declarations for intrinsic functions.
754 if (!I->getIntrinsicID() &&
755 I->getName() != "setjmp" && I->getName() != "longjmp") {
756 printFunctionSignature(I, true);
757 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
758 if (I->hasLinkOnceLinkage()) Out << " __ATTRIBUTE_WEAK__";
764 // Output the global variable declarations
766 Out << "\n\n/* Global Variable Declarations */\n";
767 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
768 if (!I->isExternal()) {
770 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
772 if (I->hasLinkOnceLinkage())
773 Out << " __attribute__((common))";
774 else if (I->hasWeakLinkage())
775 Out << " __ATTRIBUTE_WEAK__";
780 // Output the global variable definitions and contents...
782 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
783 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
784 if (!I->isExternal()) {
785 if (I->hasInternalLinkage())
787 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
788 if (I->hasLinkOnceLinkage())
789 Out << " __attribute__((common))";
790 else if (I->hasWeakLinkage())
791 Out << " __ATTRIBUTE_WEAK__";
793 // If the initializer is not null, emit the initializer. If it is null,
794 // we try to avoid emitting large amounts of zeros. The problem with
795 // this, however, occurs when the variable has weak linkage. In this
796 // case, the assembler will complain about the variable being both weak
797 // and common, so we disable this optimization.
798 if (!I->getInitializer()->isNullValue()) {
800 writeOperand(I->getInitializer());
801 } else if (I->hasWeakLinkage()) {
802 // We have to specify an initializer, but it doesn't have to be
803 // complete. If the value is an aggregate, print out { 0 }, and let
804 // the compiler figure out the rest of the zeros.
806 if (isa<StructType>(I->getInitializer()->getType()) ||
807 isa<ArrayType>(I->getInitializer()->getType())) {
810 // Just print it out normally.
811 writeOperand(I->getInitializer());
819 Out << "\n\n/* Function Bodies */\n";
824 /// Output all floating point constants that cannot be printed accurately...
825 void CWriter::printFloatingPointConstants(Function &F) {
836 // Scan the module for floating point constants. If any FP constant is used
837 // in the function, we want to redirect it here so that we do not depend on
838 // the precision of the printed form, unless the printed form preserves
841 static unsigned FPCounter = 0;
842 for (constant_iterator I = constant_begin(&F), E = constant_end(&F);
844 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
845 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
846 !FPConstantMap.count(FPC)) {
847 double Val = FPC->getValue();
849 FPConstantMap[FPC] = FPCounter; // Number the FP constants
851 if (FPC->getType() == Type::DoubleTy) {
853 Out << "static const ConstantDoubleTy FPConstant" << FPCounter++
854 << " = 0x" << std::hex << DBLUnion.U << std::dec
855 << "ULL; /* " << Val << " */\n";
856 } else if (FPC->getType() == Type::FloatTy) {
858 Out << "static const ConstantFloatTy FPConstant" << FPCounter++
859 << " = 0x" << std::hex << FLTUnion.U << std::dec
860 << "U; /* " << Val << " */\n";
862 assert(0 && "Unknown float type!");
869 /// printSymbolTable - Run through symbol table looking for type names. If a
870 /// type name is found, emit it's declaration...
872 void CWriter::printModuleTypes(const SymbolTable &ST) {
873 // If there are no type names, exit early.
874 if ( ! ST.hasTypes() )
877 // We are only interested in the type plane of the symbol table...
878 SymbolTable::type_const_iterator I = ST.type_begin();
879 SymbolTable::type_const_iterator End = ST.type_end();
881 // Print out forward declarations for structure types before anything else!
882 Out << "/* Structure forward decls */\n";
883 for (; I != End; ++I)
884 if (const Type *STy = dyn_cast<StructType>(I->second)) {
885 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
886 Out << Name << ";\n";
887 TypeNames.insert(std::make_pair(STy, Name));
892 // Now we can print out typedefs...
893 Out << "/* Typedefs */\n";
894 for (I = ST.type_begin(); I != End; ++I) {
895 const Type *Ty = cast<Type>(I->second);
896 std::string Name = "l_" + Mangler::makeNameProper(I->first);
898 printType(Out, Ty, Name);
904 // Keep track of which structures have been printed so far...
905 std::set<const StructType *> StructPrinted;
907 // Loop over all structures then push them into the stack so they are
908 // printed in the correct order.
910 Out << "/* Structure contents */\n";
911 for (I = ST.type_begin(); I != End; ++I)
912 if (const StructType *STy = dyn_cast<StructType>(I->second))
913 // Only print out used types!
914 printContainedStructs(STy, StructPrinted);
917 // Push the struct onto the stack and recursively push all structs
918 // this one depends on.
919 void CWriter::printContainedStructs(const Type *Ty,
920 std::set<const StructType*> &StructPrinted){
921 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
922 //Check to see if we have already printed this struct
923 if (StructPrinted.count(STy) == 0) {
924 // Print all contained types first...
925 for (StructType::element_iterator I = STy->element_begin(),
926 E = STy->element_end(); I != E; ++I) {
927 const Type *Ty1 = I->get();
928 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
929 printContainedStructs(*I, StructPrinted);
932 //Print structure type out..
933 StructPrinted.insert(STy);
934 std::string Name = TypeNames[STy];
935 printType(Out, STy, Name, true);
939 // If it is an array, check contained types and continue
940 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
941 const Type *Ty1 = ATy->getElementType();
942 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
943 printContainedStructs(Ty1, StructPrinted);
948 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
949 if (F->hasInternalLinkage()) Out << "static ";
951 // Loop over the arguments, printing them...
952 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
954 std::stringstream FunctionInnards;
956 // Print out the name...
957 FunctionInnards << Mang->getValueName(F) << "(";
959 if (!F->isExternal()) {
962 if (F->abegin()->hasName() || !Prototype)
963 ArgName = Mang->getValueName(F->abegin());
964 printType(FunctionInnards, F->afront().getType(), ArgName);
965 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
967 FunctionInnards << ", ";
968 if (I->hasName() || !Prototype)
969 ArgName = Mang->getValueName(I);
972 printType(FunctionInnards, I->getType(), ArgName);
976 // Loop over the arguments, printing them...
977 for (FunctionType::param_iterator I = FT->param_begin(),
978 E = FT->param_end(); I != E; ++I) {
979 if (I != FT->param_begin()) FunctionInnards << ", ";
980 printType(FunctionInnards, *I);
984 // Finish printing arguments... if this is a vararg function, print the ...,
985 // unless there are no known types, in which case, we just emit ().
987 if (FT->isVarArg() && FT->getNumParams()) {
988 if (FT->getNumParams()) FunctionInnards << ", ";
989 FunctionInnards << "..."; // Output varargs portion of signature!
990 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
991 FunctionInnards << "void"; // ret() -> ret(void) in C.
993 FunctionInnards << ")";
994 // Print out the return type and the entire signature for that matter
995 printType(Out, F->getReturnType(), FunctionInnards.str());
998 void CWriter::printFunction(Function &F) {
999 printFunctionSignature(&F, false);
1002 // print local variable information for the function
1003 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I)
1004 if (const AllocaInst *AI = isDirectAlloca(&*I)) {
1006 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
1007 Out << "; /* Address exposed local */\n";
1008 } else if (I->getType() != Type::VoidTy && !isInlinableInst(*I)) {
1010 printType(Out, I->getType(), Mang->getValueName(&*I));
1013 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
1015 printType(Out, I->getType(),
1016 Mang->getValueName(&*I)+"__PHI_TEMPORARY");
1023 // print the basic blocks
1024 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1025 if (Loop *L = LI->getLoopFor(BB)) {
1026 if (L->getHeader() == BB && L->getParentLoop() == 0)
1029 printBasicBlock(BB);
1036 void CWriter::printLoop(Loop *L) {
1037 Out << " do { /* Syntactic loop '" << L->getHeader()->getName()
1038 << "' to make GCC happy */\n";
1039 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
1040 BasicBlock *BB = L->getBlocks()[i];
1041 Loop *BBLoop = LI->getLoopFor(BB);
1043 printBasicBlock(BB);
1044 else if (BB == BBLoop->getHeader() && BBLoop->getParentLoop() == L)
1047 Out << " } while (1); /* end of syntactic loop '"
1048 << L->getHeader()->getName() << "' */\n";
1051 void CWriter::printBasicBlock(BasicBlock *BB) {
1053 // Don't print the label for the basic block if there are no uses, or if
1054 // the only terminator use is the predecessor basic block's terminator.
1055 // We have to scan the use list because PHI nodes use basic blocks too but
1056 // do not require a label to be generated.
1058 bool NeedsLabel = false;
1059 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1060 if (isGotoCodeNecessary(*PI, BB)) {
1065 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1067 // Output all of the instructions in the basic block...
1068 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E;
1070 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1071 if (II->getType() != Type::VoidTy)
1080 // Don't emit prefix or suffix for the terminator...
1081 visit(*BB->getTerminator());
1085 // Specific Instruction type classes... note that all of the casts are
1086 // necessary because we use the instruction classes as opaque types...
1088 void CWriter::visitReturnInst(ReturnInst &I) {
1089 // Don't output a void return if this is the last basic block in the function
1090 if (I.getNumOperands() == 0 &&
1091 &*--I.getParent()->getParent()->end() == I.getParent() &&
1092 !I.getParent()->size() == 1) {
1097 if (I.getNumOperands()) {
1099 writeOperand(I.getOperand(0));
1104 void CWriter::visitSwitchInst(SwitchInst &SI) {
1105 printPHICopiesForSuccessors(SI.getParent(), 0);
1108 writeOperand(SI.getOperand(0));
1109 Out << ") {\n default:\n";
1110 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1112 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1114 writeOperand(SI.getOperand(i));
1116 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1117 printBranchToBlock(SI.getParent(), Succ, 2);
1118 if (Succ == SI.getParent()->getNext())
1124 bool CWriter::isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1125 /// FIXME: This should be reenabled, but loop reordering safe!!
1128 if (From->getNext() != To) // Not the direct successor, we need a goto
1131 //isa<SwitchInst>(From->getTerminator())
1134 if (LI->getLoopFor(From) != LI->getLoopFor(To))
1139 void CWriter::printPHICopiesForSuccessors(BasicBlock *CurBlock,
1141 for (succ_iterator SI = succ_begin(CurBlock), E = succ_end(CurBlock);
1143 for (BasicBlock::iterator I = SI->begin();
1144 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1145 // now we have to do the printing
1146 Out << std::string(Indent, ' ');
1147 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1148 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBlock)));
1149 Out << "; /* for PHI node */\n";
1154 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1156 if (isGotoCodeNecessary(CurBB, Succ)) {
1157 Out << std::string(Indent, ' ') << " goto ";
1163 // Branch instruction printing - Avoid printing out a branch to a basic block
1164 // that immediately succeeds the current one.
1166 void CWriter::visitBranchInst(BranchInst &I) {
1167 printPHICopiesForSuccessors(I.getParent(), 0);
1169 if (I.isConditional()) {
1170 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1172 writeOperand(I.getCondition());
1175 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1177 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1178 Out << " } else {\n";
1179 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1182 // First goto not necessary, assume second one is...
1184 writeOperand(I.getCondition());
1187 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1192 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1197 // PHI nodes get copied into temporary values at the end of predecessor basic
1198 // blocks. We now need to copy these temporary values into the REAL value for
1200 void CWriter::visitPHINode(PHINode &I) {
1202 Out << "__PHI_TEMPORARY";
1206 void CWriter::visitBinaryOperator(Instruction &I) {
1207 // binary instructions, shift instructions, setCond instructions.
1208 assert(!isa<PointerType>(I.getType()));
1210 // We must cast the results of binary operations which might be promoted.
1211 bool needsCast = false;
1212 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1213 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1214 || (I.getType() == Type::FloatTy)) {
1217 printType(Out, I.getType());
1221 writeOperand(I.getOperand(0));
1223 switch (I.getOpcode()) {
1224 case Instruction::Add: Out << " + "; break;
1225 case Instruction::Sub: Out << " - "; break;
1226 case Instruction::Mul: Out << "*"; break;
1227 case Instruction::Div: Out << "/"; break;
1228 case Instruction::Rem: Out << "%"; break;
1229 case Instruction::And: Out << " & "; break;
1230 case Instruction::Or: Out << " | "; break;
1231 case Instruction::Xor: Out << " ^ "; break;
1232 case Instruction::SetEQ: Out << " == "; break;
1233 case Instruction::SetNE: Out << " != "; break;
1234 case Instruction::SetLE: Out << " <= "; break;
1235 case Instruction::SetGE: Out << " >= "; break;
1236 case Instruction::SetLT: Out << " < "; break;
1237 case Instruction::SetGT: Out << " > "; break;
1238 case Instruction::Shl : Out << " << "; break;
1239 case Instruction::Shr : Out << " >> "; break;
1240 default: std::cerr << "Invalid operator type!" << I; abort();
1243 writeOperand(I.getOperand(1));
1250 void CWriter::visitCastInst(CastInst &I) {
1251 if (I.getType() == Type::BoolTy) {
1253 writeOperand(I.getOperand(0));
1258 printType(Out, I.getType());
1260 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1261 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1262 // Avoid "cast to pointer from integer of different size" warnings
1266 writeOperand(I.getOperand(0));
1269 void CWriter::visitSelectInst(SelectInst &I) {
1271 writeOperand(I.getCondition());
1273 writeOperand(I.getTrueValue());
1275 writeOperand(I.getFalseValue());
1280 void CWriter::lowerIntrinsics(Function &F) {
1281 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1282 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1283 if (CallInst *CI = dyn_cast<CallInst>(I++))
1284 if (Function *F = CI->getCalledFunction())
1285 switch (F->getIntrinsicID()) {
1286 case Intrinsic::not_intrinsic:
1287 case Intrinsic::vastart:
1288 case Intrinsic::vacopy:
1289 case Intrinsic::vaend:
1290 case Intrinsic::returnaddress:
1291 case Intrinsic::frameaddress:
1292 case Intrinsic::setjmp:
1293 case Intrinsic::longjmp:
1294 // We directly implement these intrinsics
1297 // All other intrinsic calls we must lower.
1298 Instruction *Before = CI->getPrev();
1299 IL.LowerIntrinsicCall(CI);
1300 if (Before) { // Move iterator to instruction after call
1310 void CWriter::visitCallInst(CallInst &I) {
1311 // Handle intrinsic function calls first...
1312 if (Function *F = I.getCalledFunction())
1313 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1315 default: assert(0 && "Unknown LLVM intrinsic!");
1316 case Intrinsic::vastart:
1319 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1320 // Output the last argument to the enclosing function...
1321 if (I.getParent()->getParent()->aempty()) {
1322 std::cerr << "The C backend does not currently support zero "
1323 << "argument varargs functions, such as '"
1324 << I.getParent()->getParent()->getName() << "'!\n";
1327 writeOperand(&I.getParent()->getParent()->aback());
1330 case Intrinsic::vaend:
1331 Out << "va_end(*(va_list*)&";
1332 writeOperand(I.getOperand(1));
1335 case Intrinsic::vacopy:
1337 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1338 Out << "*(va_list*)&";
1339 writeOperand(I.getOperand(1));
1342 case Intrinsic::returnaddress:
1343 Out << "__builtin_return_address(";
1344 writeOperand(I.getOperand(1));
1347 case Intrinsic::frameaddress:
1348 Out << "__builtin_frame_address(";
1349 writeOperand(I.getOperand(1));
1352 case Intrinsic::setjmp:
1353 Out << "setjmp(*(jmp_buf*)";
1354 writeOperand(I.getOperand(1));
1357 case Intrinsic::longjmp:
1358 Out << "longjmp(*(jmp_buf*)";
1359 writeOperand(I.getOperand(1));
1361 writeOperand(I.getOperand(2));
1369 void CWriter::visitCallSite(CallSite CS) {
1370 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1371 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1372 const Type *RetTy = FTy->getReturnType();
1374 writeOperand(CS.getCalledValue());
1377 if (CS.arg_begin() != CS.arg_end()) {
1378 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1381 for (++AI; AI != AE; ++AI) {
1389 void CWriter::visitMallocInst(MallocInst &I) {
1390 assert(0 && "lowerallocations pass didn't work!");
1393 void CWriter::visitAllocaInst(AllocaInst &I) {
1395 printType(Out, I.getType());
1396 Out << ") alloca(sizeof(";
1397 printType(Out, I.getType()->getElementType());
1399 if (I.isArrayAllocation()) {
1401 writeOperand(I.getOperand(0));
1406 void CWriter::visitFreeInst(FreeInst &I) {
1407 assert(0 && "lowerallocations pass didn't work!");
1410 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1411 gep_type_iterator E) {
1412 bool HasImplicitAddress = false;
1413 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1414 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1415 HasImplicitAddress = true;
1416 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1417 HasImplicitAddress = true;
1418 Ptr = CPR->getValue(); // Get to the global...
1419 } else if (isDirectAlloca(Ptr)) {
1420 HasImplicitAddress = true;
1424 if (!HasImplicitAddress)
1425 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1427 writeOperandInternal(Ptr);
1431 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1432 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1435 writeOperandInternal(Ptr);
1437 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1439 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1442 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1443 "Can only have implicit address with direct accessing");
1445 if (HasImplicitAddress) {
1447 } else if (CI && CI->isNullValue()) {
1448 gep_type_iterator TmpI = I; ++TmpI;
1450 // Print out the -> operator if possible...
1451 if (TmpI != E && isa<StructType>(*TmpI)) {
1452 Out << (HasImplicitAddress ? "." : "->");
1453 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1459 if (isa<StructType>(*I)) {
1460 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1463 writeOperand(I.getOperand());
1468 void CWriter::visitLoadInst(LoadInst &I) {
1470 writeOperand(I.getOperand(0));
1473 void CWriter::visitStoreInst(StoreInst &I) {
1475 writeOperand(I.getPointerOperand());
1477 writeOperand(I.getOperand(0));
1480 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1482 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1486 void CWriter::visitVANextInst(VANextInst &I) {
1487 Out << Mang->getValueName(I.getOperand(0));
1488 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1489 printType(Out, I.getArgType());
1493 void CWriter::visitVAArgInst(VAArgInst &I) {
1495 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1496 writeOperand(I.getOperand(0));
1497 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1498 printType(Out, I.getType());
1499 Out << ");\n va_end(Tmp); }";
1502 //===----------------------------------------------------------------------===//
1503 // External Interface declaration
1504 //===----------------------------------------------------------------------===//
1506 bool CTargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &o) {
1507 PM.add(createLowerGCPass());
1508 PM.add(createLowerAllocationsPass());
1509 PM.add(createLowerInvokePass());
1510 PM.add(new CBackendNameAllUsedStructs());
1511 PM.add(new CWriter(o, getIntrinsicLowering()));
1515 TargetMachine *llvm::allocateCTargetMachine(const Module &M,
1516 IntrinsicLowering *IL) {
1517 return new CTargetMachine(M, IL);