1 //===-- Writer.cpp - Library for converting LLVM code to C ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This library converts LLVM code to C code, compilable by GCC and other C
13 //===----------------------------------------------------------------------===//
15 #include "CTargetMachine.h"
16 #include "llvm/Target/TargetMachineImpls.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Module.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/Pass.h"
22 #include "llvm/PassManager.h"
23 #include "llvm/SymbolTable.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/IntrinsicLowering.h"
26 #include "llvm/Analysis/FindUsedTypes.h"
27 #include "llvm/Analysis/ConstantsScanner.h"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Support/CFG.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/InstVisitor.h"
33 #include "llvm/Support/Mangler.h"
34 #include "Support/StringExtras.h"
35 #include "Config/config.h"
41 class CWriter : public Pass, public InstVisitor<CWriter> {
43 IntrinsicLowering &IL;
45 const Module *TheModule;
48 std::map<const Type *, std::string> TypeNames;
50 std::map<const ConstantFP *, unsigned> FPConstantMap;
52 CWriter(std::ostream &o, IntrinsicLowering &il) : Out(o), IL(il) {}
54 void getAnalysisUsage(AnalysisUsage &AU) const {
55 AU.addRequired<FindUsedTypes>();
58 virtual const char *getPassName() const { return "C backend"; }
60 bool doInitialization(Module &M);
62 // First pass, lower all unhandled intrinsics.
67 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
77 std::ostream &printType(std::ostream &Out, const Type *Ty,
78 const std::string &VariableName = "",
79 bool IgnoreName = false);
81 void writeOperand(Value *Operand);
82 void writeOperandInternal(Value *Operand);
85 void lowerIntrinsics(Module &M);
87 bool nameAllUsedStructureTypes(Module &M);
88 void printModule(Module *M);
89 void printFloatingPointConstants(Module &M);
90 void printSymbolTable(const SymbolTable &ST);
91 void printContainedStructs(const Type *Ty, std::set<const StructType *> &);
92 void printFunctionSignature(const Function *F, bool Prototype);
94 void printFunction(Function &);
96 void printConstant(Constant *CPV);
97 void printConstantArray(ConstantArray *CPA);
99 // isInlinableInst - Attempt to inline instructions into their uses to build
100 // trees as much as possible. To do this, we have to consistently decide
101 // what is acceptable to inline, so that variable declarations don't get
102 // printed and an extra copy of the expr is not emitted.
104 static bool isInlinableInst(const Instruction &I) {
105 // Must be an expression, must be used exactly once. If it is dead, we
106 // emit it inline where it would go.
107 if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
108 isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
109 isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
110 // Don't inline a load across a store or other bad things!
113 // Only inline instruction it it's use is in the same BB as the inst.
114 return I.getParent() == cast<Instruction>(I.use_back())->getParent();
117 // isDirectAlloca - Define fixed sized allocas in the entry block as direct
118 // variables which are accessed with the & operator. This causes GCC to
119 // generate significantly better code than to emit alloca calls directly.
121 static const AllocaInst *isDirectAlloca(const Value *V) {
122 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
123 if (!AI) return false;
124 if (AI->isArrayAllocation())
125 return 0; // FIXME: we can also inline fixed size array allocas!
126 if (AI->getParent() != &AI->getParent()->getParent()->getEntryBlock())
131 // Instruction visitation functions
132 friend class InstVisitor<CWriter>;
134 void visitReturnInst(ReturnInst &I);
135 void visitBranchInst(BranchInst &I);
136 void visitSwitchInst(SwitchInst &I);
137 void visitInvokeInst(InvokeInst &I);
138 void visitUnwindInst(UnwindInst &I);
140 void visitPHINode(PHINode &I);
141 void visitBinaryOperator(Instruction &I);
143 void visitCastInst (CastInst &I);
144 void visitSelectInst(SelectInst &I);
145 void visitCallInst (CallInst &I);
146 void visitCallSite (CallSite CS);
147 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); }
149 void visitMallocInst(MallocInst &I);
150 void visitAllocaInst(AllocaInst &I);
151 void visitFreeInst (FreeInst &I);
152 void visitLoadInst (LoadInst &I);
153 void visitStoreInst (StoreInst &I);
154 void visitGetElementPtrInst(GetElementPtrInst &I);
155 void visitVANextInst(VANextInst &I);
156 void visitVAArgInst (VAArgInst &I);
158 void visitInstruction(Instruction &I) {
159 std::cerr << "C Writer does not know about " << I;
163 void outputLValue(Instruction *I) {
164 Out << " " << Mang->getValueName(I) << " = ";
166 void printPHICopiesForSuccessors(BasicBlock *CurBlock,
168 void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
170 void printIndexingExpression(Value *Ptr, gep_type_iterator I,
171 gep_type_iterator E);
175 // Pass the Type* and the variable name and this prints out the variable
178 std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
179 const std::string &NameSoFar,
181 if (Ty->isPrimitiveType())
182 switch (Ty->getPrimitiveID()) {
183 case Type::VoidTyID: return Out << "void " << NameSoFar;
184 case Type::BoolTyID: return Out << "bool " << NameSoFar;
185 case Type::UByteTyID: return Out << "unsigned char " << NameSoFar;
186 case Type::SByteTyID: return Out << "signed char " << NameSoFar;
187 case Type::UShortTyID: return Out << "unsigned short " << NameSoFar;
188 case Type::ShortTyID: return Out << "short " << NameSoFar;
189 case Type::UIntTyID: return Out << "unsigned " << NameSoFar;
190 case Type::IntTyID: return Out << "int " << NameSoFar;
191 case Type::ULongTyID: return Out << "unsigned long long " << NameSoFar;
192 case Type::LongTyID: return Out << "signed long long " << NameSoFar;
193 case Type::FloatTyID: return Out << "float " << NameSoFar;
194 case Type::DoubleTyID: return Out << "double " << NameSoFar;
196 std::cerr << "Unknown primitive type: " << Ty << "\n";
200 // Check to see if the type is named.
201 if (!IgnoreName || isa<OpaqueType>(Ty)) {
202 std::map<const Type *, std::string>::iterator I = TypeNames.find(Ty);
203 if (I != TypeNames.end()) return Out << I->second << " " << NameSoFar;
206 switch (Ty->getPrimitiveID()) {
207 case Type::FunctionTyID: {
208 const FunctionType *MTy = cast<FunctionType>(Ty);
209 std::stringstream FunctionInnards;
210 FunctionInnards << " (" << NameSoFar << ") (";
211 for (FunctionType::param_iterator I = MTy->param_begin(),
212 E = MTy->param_end(); I != E; ++I) {
213 if (I != MTy->param_begin())
214 FunctionInnards << ", ";
215 printType(FunctionInnards, *I, "");
217 if (MTy->isVarArg()) {
218 if (MTy->getNumParams())
219 FunctionInnards << ", ...";
220 } else if (!MTy->getNumParams()) {
221 FunctionInnards << "void";
223 FunctionInnards << ")";
224 std::string tstr = FunctionInnards.str();
225 printType(Out, MTy->getReturnType(), tstr);
228 case Type::StructTyID: {
229 const StructType *STy = cast<StructType>(Ty);
230 Out << NameSoFar + " {\n";
232 for (StructType::element_iterator I = STy->element_begin(),
233 E = STy->element_end(); I != E; ++I) {
235 printType(Out, *I, "field" + utostr(Idx++));
241 case Type::PointerTyID: {
242 const PointerType *PTy = cast<PointerType>(Ty);
243 std::string ptrName = "*" + NameSoFar;
245 if (isa<ArrayType>(PTy->getElementType()))
246 ptrName = "(" + ptrName + ")";
248 return printType(Out, PTy->getElementType(), ptrName);
251 case Type::ArrayTyID: {
252 const ArrayType *ATy = cast<ArrayType>(Ty);
253 unsigned NumElements = ATy->getNumElements();
254 return printType(Out, ATy->getElementType(),
255 NameSoFar + "[" + utostr(NumElements) + "]");
258 case Type::OpaqueTyID: {
259 static int Count = 0;
260 std::string TyName = "struct opaque_" + itostr(Count++);
261 assert(TypeNames.find(Ty) == TypeNames.end());
262 TypeNames[Ty] = TyName;
263 return Out << TyName << " " << NameSoFar;
266 assert(0 && "Unhandled case in getTypeProps!");
273 void CWriter::printConstantArray(ConstantArray *CPA) {
275 // As a special case, print the array as a string if it is an array of
276 // ubytes or an array of sbytes with positive values.
278 const Type *ETy = CPA->getType()->getElementType();
279 bool isString = (ETy == Type::SByteTy || ETy == Type::UByteTy);
281 // Make sure the last character is a null char, as automatically added by C
282 if (isString && (CPA->getNumOperands() == 0 ||
283 !cast<Constant>(*(CPA->op_end()-1))->isNullValue()))
288 // Keep track of whether the last number was a hexadecimal escape
289 bool LastWasHex = false;
291 // Do not include the last character, which we know is null
292 for (unsigned i = 0, e = CPA->getNumOperands()-1; i != e; ++i) {
293 unsigned char C = cast<ConstantInt>(CPA->getOperand(i))->getRawValue();
295 // Print it out literally if it is a printable character. The only thing
296 // to be careful about is when the last letter output was a hex escape
297 // code, in which case we have to be careful not to print out hex digits
298 // explicitly (the C compiler thinks it is a continuation of the previous
299 // character, sheesh...)
301 if (isprint(C) && (!LastWasHex || !isxdigit(C))) {
303 if (C == '"' || C == '\\')
310 case '\n': Out << "\\n"; break;
311 case '\t': Out << "\\t"; break;
312 case '\r': Out << "\\r"; break;
313 case '\v': Out << "\\v"; break;
314 case '\a': Out << "\\a"; break;
315 case '\"': Out << "\\\""; break;
316 case '\'': Out << "\\\'"; break;
319 Out << (char)(( C/16 < 10) ? ( C/16 +'0') : ( C/16 -10+'A'));
320 Out << (char)(((C&15) < 10) ? ((C&15)+'0') : ((C&15)-10+'A'));
329 if (CPA->getNumOperands()) {
331 printConstant(cast<Constant>(CPA->getOperand(0)));
332 for (unsigned i = 1, e = CPA->getNumOperands(); i != e; ++i) {
334 printConstant(cast<Constant>(CPA->getOperand(i)));
341 // isFPCSafeToPrint - Returns true if we may assume that CFP may be written out
342 // textually as a double (rather than as a reference to a stack-allocated
343 // variable). We decide this by converting CFP to a string and back into a
344 // double, and then checking whether the conversion results in a bit-equal
345 // double to the original value of CFP. This depends on us and the target C
346 // compiler agreeing on the conversion process (which is pretty likely since we
347 // only deal in IEEE FP).
349 bool isFPCSafeToPrint(const ConstantFP *CFP) {
352 sprintf(Buffer, "%a", CFP->getValue());
354 if (!strncmp(Buffer, "0x", 2) ||
355 !strncmp(Buffer, "-0x", 3) ||
356 !strncmp(Buffer, "+0x", 3))
357 return atof(Buffer) == CFP->getValue();
360 std::string StrVal = ftostr(CFP->getValue());
362 while (StrVal[0] == ' ')
363 StrVal.erase(StrVal.begin());
365 // Check to make sure that the stringized number is not some string like "Inf"
366 // or NaN. Check that the string matches the "[-+]?[0-9]" regex.
367 if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
368 ((StrVal[0] == '-' || StrVal[0] == '+') &&
369 (StrVal[1] >= '0' && StrVal[1] <= '9')))
370 // Reparse stringized version!
371 return atof(StrVal.c_str()) == CFP->getValue();
376 // printConstant - The LLVM Constant to C Constant converter.
377 void CWriter::printConstant(Constant *CPV) {
378 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
379 switch (CE->getOpcode()) {
380 case Instruction::Cast:
382 printType(Out, CPV->getType());
384 printConstant(CE->getOperand(0));
388 case Instruction::GetElementPtr:
390 printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
394 case Instruction::Select:
396 printConstant(CE->getOperand(0));
398 printConstant(CE->getOperand(1));
400 printConstant(CE->getOperand(2));
403 case Instruction::Add:
404 case Instruction::Sub:
405 case Instruction::Mul:
406 case Instruction::Div:
407 case Instruction::Rem:
408 case Instruction::SetEQ:
409 case Instruction::SetNE:
410 case Instruction::SetLT:
411 case Instruction::SetLE:
412 case Instruction::SetGT:
413 case Instruction::SetGE:
414 case Instruction::Shl:
415 case Instruction::Shr:
417 printConstant(CE->getOperand(0));
418 switch (CE->getOpcode()) {
419 case Instruction::Add: Out << " + "; break;
420 case Instruction::Sub: Out << " - "; break;
421 case Instruction::Mul: Out << " * "; break;
422 case Instruction::Div: Out << " / "; break;
423 case Instruction::Rem: Out << " % "; break;
424 case Instruction::SetEQ: Out << " == "; break;
425 case Instruction::SetNE: Out << " != "; break;
426 case Instruction::SetLT: Out << " < "; break;
427 case Instruction::SetLE: Out << " <= "; break;
428 case Instruction::SetGT: Out << " > "; break;
429 case Instruction::SetGE: Out << " >= "; break;
430 case Instruction::Shl: Out << " << "; break;
431 case Instruction::Shr: Out << " >> "; break;
432 default: assert(0 && "Illegal opcode here!");
434 printConstant(CE->getOperand(1));
439 std::cerr << "CWriter Error: Unhandled constant expression: "
445 switch (CPV->getType()->getPrimitiveID()) {
447 Out << (CPV == ConstantBool::False ? "0" : "1"); break;
448 case Type::SByteTyID:
449 case Type::ShortTyID:
450 Out << cast<ConstantSInt>(CPV)->getValue(); break;
452 if ((int)cast<ConstantSInt>(CPV)->getValue() == (int)0x80000000)
453 Out << "((int)0x80000000)"; // Handle MININT specially to avoid warning
455 Out << cast<ConstantSInt>(CPV)->getValue();
459 Out << cast<ConstantSInt>(CPV)->getValue() << "ll"; break;
461 case Type::UByteTyID:
462 case Type::UShortTyID:
463 Out << cast<ConstantUInt>(CPV)->getValue(); break;
465 Out << cast<ConstantUInt>(CPV)->getValue() << "u"; break;
466 case Type::ULongTyID:
467 Out << cast<ConstantUInt>(CPV)->getValue() << "ull"; break;
469 case Type::FloatTyID:
470 case Type::DoubleTyID: {
471 ConstantFP *FPC = cast<ConstantFP>(CPV);
472 std::map<const ConstantFP*, unsigned>::iterator I = FPConstantMap.find(FPC);
473 if (I != FPConstantMap.end()) {
474 // Because of FP precision problems we must load from a stack allocated
475 // value that holds the value in hex.
476 Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
477 << "*)&FPConstant" << I->second << ")";
480 // Print out the constant as a floating point number.
482 sprintf(Buffer, "%a", FPC->getValue());
483 Out << Buffer << " /*" << FPC->getValue() << "*/ ";
485 Out << ftostr(FPC->getValue());
491 case Type::ArrayTyID:
492 if (isa<ConstantAggregateZero>(CPV)) {
493 const ArrayType *AT = cast<ArrayType>(CPV->getType());
495 if (AT->getNumElements()) {
497 Constant *CZ = Constant::getNullValue(AT->getElementType());
499 for (unsigned i = 1, e = AT->getNumElements(); i != e; ++i) {
506 printConstantArray(cast<ConstantArray>(CPV));
510 case Type::StructTyID:
511 if (isa<ConstantAggregateZero>(CPV)) {
512 const StructType *ST = cast<StructType>(CPV->getType());
514 if (ST->getNumElements()) {
516 printConstant(Constant::getNullValue(ST->getElementType(0)));
517 for (unsigned i = 1, e = ST->getNumElements(); i != e; ++i) {
519 printConstant(Constant::getNullValue(ST->getElementType(i)));
525 if (CPV->getNumOperands()) {
527 printConstant(cast<Constant>(CPV->getOperand(0)));
528 for (unsigned i = 1, e = CPV->getNumOperands(); i != e; ++i) {
530 printConstant(cast<Constant>(CPV->getOperand(i)));
537 case Type::PointerTyID:
538 if (isa<ConstantPointerNull>(CPV)) {
540 printType(Out, CPV->getType());
541 Out << ")/*NULL*/0)";
543 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(CPV)) {
544 writeOperand(CPR->getValue());
549 std::cerr << "Unknown constant type: " << CPV << "\n";
554 void CWriter::writeOperandInternal(Value *Operand) {
555 if (Instruction *I = dyn_cast<Instruction>(Operand))
556 if (isInlinableInst(*I) && !isDirectAlloca(I)) {
557 // Should we inline this instruction to build a tree?
564 if (Constant *CPV = dyn_cast<Constant>(Operand)) {
567 Out << Mang->getValueName(Operand);
571 void CWriter::writeOperand(Value *Operand) {
572 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
573 Out << "(&"; // Global variables are references as their addresses by llvm
575 writeOperandInternal(Operand);
577 if (isa<GlobalVariable>(Operand) || isDirectAlloca(Operand))
581 // nameAllUsedStructureTypes - If there are structure types in the module that
582 // are used but do not have names assigned to them in the symbol table yet then
583 // we assign them names now.
585 bool CWriter::nameAllUsedStructureTypes(Module &M) {
586 // Get a set of types that are used by the program...
587 std::set<const Type *> UT = FUT->getTypes();
589 // Loop over the module symbol table, removing types from UT that are already
592 SymbolTable &MST = M.getSymbolTable();
593 if (MST.find(Type::TypeTy) != MST.end())
594 for (SymbolTable::type_iterator I = MST.type_begin(Type::TypeTy),
595 E = MST.type_end(Type::TypeTy); I != E; ++I)
596 UT.erase(cast<Type>(I->second));
598 // UT now contains types that are not named. Loop over it, naming structure
601 bool Changed = false;
602 for (std::set<const Type *>::const_iterator I = UT.begin(), E = UT.end();
604 if (const StructType *ST = dyn_cast<StructType>(*I)) {
605 ((Value*)ST)->setName("unnamed", &MST);
611 // generateCompilerSpecificCode - This is where we add conditional compilation
612 // directives to cater to specific compilers as need be.
614 static void generateCompilerSpecificCode(std::ostream& Out) {
615 // Alloca is hard to get, and we don't want to include stdlib.h here...
616 Out << "/* get a declaration for alloca */\n"
618 << "extern void *__builtin_alloca(unsigned long);\n"
619 << "#define alloca(x) __builtin_alloca(x)\n"
621 << "#ifndef __FreeBSD__\n"
622 << "#include <alloca.h>\n"
626 // We output GCC specific attributes to preserve 'linkonce'ness on globals.
627 // If we aren't being compiled with GCC, just drop these attributes.
628 Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
629 << "#define __attribute__(X)\n"
633 // At some point, we should support "external weak" vs. "weak" linkages.
634 // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
635 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
636 << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
637 << "#elif defined(__GNUC__)\n"
638 << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
640 << "#define __EXTERNAL_WEAK__\n"
644 // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
645 Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
646 << "#define __ATTRIBUTE_WEAK__\n"
647 << "#elif defined(__GNUC__)\n"
648 << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
650 << "#define __ATTRIBUTE_WEAK__\n"
654 bool CWriter::doInitialization(Module &M) {
657 FUT = &getAnalysis<FindUsedTypes>();
659 // Ensure that all structure types have names...
660 bool Changed = nameAllUsedStructureTypes(M);
661 Mang = new Mangler(M);
663 // get declaration for alloca
664 Out << "/* Provide Declarations */\n";
665 Out << "#include <stdarg.h>\n"; // Varargs support
666 Out << "#include <setjmp.h>\n"; // Unwind support
667 generateCompilerSpecificCode(Out);
669 // Provide a definition for `bool' if not compiling with a C++ compiler.
671 << "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
673 << "\n\n/* Support for floating point constants */\n"
674 << "typedef unsigned long long ConstantDoubleTy;\n"
675 << "typedef unsigned int ConstantFloatTy;\n"
677 << "\n\n/* Global Declarations */\n";
679 // First output all the declarations for the program, because C requires
680 // Functions & globals to be declared before they are used.
683 // Loop over the symbol table, emitting all named constants...
684 printSymbolTable(M.getSymbolTable());
686 // Global variable declarations...
688 Out << "\n/* External Global Variable Declarations */\n";
689 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I) {
690 if (I->hasExternalLinkage()) {
692 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
698 // Function declarations
700 Out << "\n/* Function Declarations */\n";
701 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
702 // Don't print declarations for intrinsic functions.
703 if (!I->getIntrinsicID()) {
704 printFunctionSignature(I, true);
705 if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
706 if (I->hasLinkOnceLinkage()) Out << " __ATTRIBUTE_WEAK__";
712 // Output the global variable declarations
714 Out << "\n\n/* Global Variable Declarations */\n";
715 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
716 if (!I->isExternal()) {
718 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
720 if (I->hasLinkOnceLinkage())
721 Out << " __attribute__((common))";
722 else if (I->hasWeakLinkage())
723 Out << " __ATTRIBUTE_WEAK__";
728 // Output the global variable definitions and contents...
730 Out << "\n\n/* Global Variable Definitions and Initialization */\n";
731 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
732 if (!I->isExternal()) {
733 if (I->hasInternalLinkage())
735 printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
736 if (I->hasLinkOnceLinkage())
737 Out << " __attribute__((common))";
738 else if (I->hasWeakLinkage())
739 Out << " __ATTRIBUTE_WEAK__";
741 // If the initializer is not null, emit the initializer. If it is null,
742 // we try to avoid emitting large amounts of zeros. The problem with
743 // this, however, occurs when the variable has weak linkage. In this
744 // case, the assembler will complain about the variable being both weak
745 // and common, so we disable this optimization.
746 if (!I->getInitializer()->isNullValue()) {
748 writeOperand(I->getInitializer());
749 } else if (I->hasWeakLinkage()) {
750 // We have to specify an initializer, but it doesn't have to be
751 // complete. If the value is an aggregate, print out { 0 }, and let
752 // the compiler figure out the rest of the zeros.
754 if (isa<StructType>(I->getInitializer()->getType()) ||
755 isa<ArrayType>(I->getInitializer()->getType())) {
758 // Just print it out normally.
759 writeOperand(I->getInitializer());
766 // Output all floating point constants that cannot be printed accurately...
767 printFloatingPointConstants(M);
770 Out << "\n\n/* Function Bodies */\n";
775 /// Output all floating point constants that cannot be printed accurately...
776 void CWriter::printFloatingPointConstants(Module &M) {
779 unsigned long long U;
787 // Scan the module for floating point constants. If any FP constant is used
788 // in the function, we want to redirect it here so that we do not depend on
789 // the precision of the printed form, unless the printed form preserves
792 unsigned FPCounter = 0;
793 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
794 for (constant_iterator I = constant_begin(F), E = constant_end(F);
796 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(*I))
797 if (!isFPCSafeToPrint(FPC) && // Do not put in FPConstantMap if safe.
798 !FPConstantMap.count(FPC)) {
799 double Val = FPC->getValue();
801 FPConstantMap[FPC] = FPCounter; // Number the FP constants
803 if (FPC->getType() == Type::DoubleTy) {
805 Out << "static const ConstantDoubleTy FPConstant" << FPCounter++
806 << " = 0x" << std::hex << DBLUnion.U << std::dec
807 << "ULL; /* " << Val << " */\n";
808 } else if (FPC->getType() == Type::FloatTy) {
810 Out << "static const ConstantFloatTy FPConstant" << FPCounter++
811 << " = 0x" << std::hex << FLTUnion.U << std::dec
812 << "U; /* " << Val << " */\n";
814 assert(0 && "Unknown float type!");
821 /// printSymbolTable - Run through symbol table looking for type names. If a
822 /// type name is found, emit it's declaration...
824 void CWriter::printSymbolTable(const SymbolTable &ST) {
825 // If there are no type names, exit early.
826 if (ST.find(Type::TypeTy) == ST.end())
829 // We are only interested in the type plane of the symbol table...
830 SymbolTable::type_const_iterator I = ST.type_begin(Type::TypeTy);
831 SymbolTable::type_const_iterator End = ST.type_end(Type::TypeTy);
833 // Print out forward declarations for structure types before anything else!
834 Out << "/* Structure forward decls */\n";
835 for (; I != End; ++I)
836 if (const Type *STy = dyn_cast<StructType>(I->second))
837 // Only print out used types!
838 if (FUT->getTypes().count(STy)) {
839 std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
840 Out << Name << ";\n";
841 TypeNames.insert(std::make_pair(STy, Name));
846 // Now we can print out typedefs...
847 Out << "/* Typedefs */\n";
848 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
849 // Only print out used types!
850 if (FUT->getTypes().count(cast<Type>(I->second))) {
851 const Type *Ty = cast<Type>(I->second);
852 std::string Name = "l_" + Mangler::makeNameProper(I->first);
854 printType(Out, Ty, Name);
860 // Keep track of which structures have been printed so far...
861 std::set<const StructType *> StructPrinted;
863 // Loop over all structures then push them into the stack so they are
864 // printed in the correct order.
866 Out << "/* Structure contents */\n";
867 for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
868 if (const StructType *STy = dyn_cast<StructType>(I->second))
869 // Only print out used types!
870 if (FUT->getTypes().count(STy))
871 printContainedStructs(STy, StructPrinted);
874 // Push the struct onto the stack and recursively push all structs
875 // this one depends on.
876 void CWriter::printContainedStructs(const Type *Ty,
877 std::set<const StructType*> &StructPrinted){
878 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
879 //Check to see if we have already printed this struct
880 if (StructPrinted.count(STy) == 0) {
881 // Print all contained types first...
882 for (StructType::element_iterator I = STy->element_begin(),
883 E = STy->element_end(); I != E; ++I) {
884 const Type *Ty1 = I->get();
885 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
886 printContainedStructs(*I, StructPrinted);
889 //Print structure type out..
890 StructPrinted.insert(STy);
891 std::string Name = TypeNames[STy];
892 printType(Out, STy, Name, true);
896 // If it is an array, check contained types and continue
897 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)){
898 const Type *Ty1 = ATy->getElementType();
899 if (isa<StructType>(Ty1) || isa<ArrayType>(Ty1))
900 printContainedStructs(Ty1, StructPrinted);
905 void CWriter::printFunctionSignature(const Function *F, bool Prototype) {
906 if (F->hasInternalLinkage()) Out << "static ";
908 // Loop over the arguments, printing them...
909 const FunctionType *FT = cast<FunctionType>(F->getFunctionType());
911 std::stringstream FunctionInnards;
913 // Print out the name...
914 FunctionInnards << Mang->getValueName(F) << "(";
916 if (!F->isExternal()) {
919 if (F->abegin()->hasName() || !Prototype)
920 ArgName = Mang->getValueName(F->abegin());
921 printType(FunctionInnards, F->afront().getType(), ArgName);
922 for (Function::const_aiterator I = ++F->abegin(), E = F->aend();
924 FunctionInnards << ", ";
925 if (I->hasName() || !Prototype)
926 ArgName = Mang->getValueName(I);
929 printType(FunctionInnards, I->getType(), ArgName);
933 // Loop over the arguments, printing them...
934 for (FunctionType::param_iterator I = FT->param_begin(),
935 E = FT->param_end(); I != E; ++I) {
936 if (I != FT->param_begin()) FunctionInnards << ", ";
937 printType(FunctionInnards, *I);
941 // Finish printing arguments... if this is a vararg function, print the ...,
942 // unless there are no known types, in which case, we just emit ().
944 if (FT->isVarArg() && FT->getNumParams()) {
945 if (FT->getNumParams()) FunctionInnards << ", ";
946 FunctionInnards << "..."; // Output varargs portion of signature!
947 } else if (!FT->isVarArg() && FT->getNumParams() == 0) {
948 FunctionInnards << "void"; // ret() -> ret(void) in C.
950 FunctionInnards << ")";
951 // Print out the return type and the entire signature for that matter
952 printType(Out, F->getReturnType(), FunctionInnards.str());
955 void CWriter::printFunction(Function &F) {
956 printFunctionSignature(&F, false);
959 // print local variable information for the function
960 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I)
961 if (const AllocaInst *AI = isDirectAlloca(&*I)) {
963 printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
964 Out << "; /* Address exposed local */\n";
965 } else if (I->getType() != Type::VoidTy && !isInlinableInst(*I)) {
967 printType(Out, I->getType(), Mang->getValueName(&*I));
970 if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
972 printType(Out, I->getType(),
973 Mang->getValueName(&*I)+"__PHI_TEMPORARY");
980 // print the basic blocks
981 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
982 BasicBlock *Prev = BB->getPrev();
984 // Don't print the label for the basic block if there are no uses, or if the
985 // only terminator use is the predecessor basic block's terminator. We have
986 // to scan the use list because PHI nodes use basic blocks too but do not
987 // require a label to be generated.
989 bool NeedsLabel = false;
990 for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
992 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(*UI))
993 if (TI != Prev->getTerminator() ||
994 isa<SwitchInst>(Prev->getTerminator()) ||
995 isa<InvokeInst>(Prev->getTerminator())) {
1000 if (NeedsLabel) Out << Mang->getValueName(BB) << ":\n";
1002 // Output all of the instructions in the basic block...
1003 for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ++II){
1004 if (!isInlinableInst(*II) && !isDirectAlloca(II)) {
1005 if (II->getType() != Type::VoidTy)
1014 // Don't emit prefix or suffix for the terminator...
1015 visit(*BB->getTerminator());
1021 // Specific Instruction type classes... note that all of the casts are
1022 // necessary because we use the instruction classes as opaque types...
1024 void CWriter::visitReturnInst(ReturnInst &I) {
1025 // Don't output a void return if this is the last basic block in the function
1026 if (I.getNumOperands() == 0 &&
1027 &*--I.getParent()->getParent()->end() == I.getParent() &&
1028 !I.getParent()->size() == 1) {
1033 if (I.getNumOperands()) {
1035 writeOperand(I.getOperand(0));
1040 void CWriter::visitSwitchInst(SwitchInst &SI) {
1041 printPHICopiesForSuccessors(SI.getParent(), 0);
1044 writeOperand(SI.getOperand(0));
1045 Out << ") {\n default:\n";
1046 printBranchToBlock(SI.getParent(), SI.getDefaultDest(), 2);
1048 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) {
1050 writeOperand(SI.getOperand(i));
1052 BasicBlock *Succ = cast<BasicBlock>(SI.getOperand(i+1));
1053 printBranchToBlock(SI.getParent(), Succ, 2);
1054 if (Succ == SI.getParent()->getNext())
1060 void CWriter::visitInvokeInst(InvokeInst &II) {
1061 assert(0 && "Lowerinvoke pass didn't work!");
1065 void CWriter::visitUnwindInst(UnwindInst &I) {
1066 assert(0 && "Lowerinvoke pass didn't work!");
1069 static bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
1070 // If PHI nodes need copies, we need the copy code...
1071 if (isa<PHINode>(To->front()) ||
1072 From->getNext() != To) // Not directly successor, need goto
1075 // Otherwise we don't need the code.
1079 void CWriter::printPHICopiesForSuccessors(BasicBlock *CurBlock,
1081 for (succ_iterator SI = succ_begin(CurBlock), E = succ_end(CurBlock);
1083 for (BasicBlock::iterator I = SI->begin();
1084 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1085 // now we have to do the printing
1086 Out << std::string(Indent, ' ');
1087 Out << " " << Mang->getValueName(I) << "__PHI_TEMPORARY = ";
1088 writeOperand(PN->getIncomingValue(PN->getBasicBlockIndex(CurBlock)));
1089 Out << "; /* for PHI node */\n";
1094 void CWriter::printBranchToBlock(BasicBlock *CurBB, BasicBlock *Succ,
1096 if (CurBB->getNext() != Succ ||
1097 isa<InvokeInst>(CurBB->getTerminator()) ||
1098 isa<SwitchInst>(CurBB->getTerminator())) {
1099 Out << std::string(Indent, ' ') << " goto ";
1105 // Branch instruction printing - Avoid printing out a branch to a basic block
1106 // that immediately succeeds the current one.
1108 void CWriter::visitBranchInst(BranchInst &I) {
1109 printPHICopiesForSuccessors(I.getParent(), 0);
1111 if (I.isConditional()) {
1112 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(0))) {
1114 writeOperand(I.getCondition());
1117 printBranchToBlock(I.getParent(), I.getSuccessor(0), 2);
1119 if (isGotoCodeNecessary(I.getParent(), I.getSuccessor(1))) {
1120 Out << " } else {\n";
1121 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1124 // First goto not necessary, assume second one is...
1126 writeOperand(I.getCondition());
1129 printBranchToBlock(I.getParent(), I.getSuccessor(1), 2);
1134 printBranchToBlock(I.getParent(), I.getSuccessor(0), 0);
1139 // PHI nodes get copied into temporary values at the end of predecessor basic
1140 // blocks. We now need to copy these temporary values into the REAL value for
1142 void CWriter::visitPHINode(PHINode &I) {
1144 Out << "__PHI_TEMPORARY";
1148 void CWriter::visitBinaryOperator(Instruction &I) {
1149 // binary instructions, shift instructions, setCond instructions.
1150 assert(!isa<PointerType>(I.getType()));
1152 // We must cast the results of binary operations which might be promoted.
1153 bool needsCast = false;
1154 if ((I.getType() == Type::UByteTy) || (I.getType() == Type::SByteTy)
1155 || (I.getType() == Type::UShortTy) || (I.getType() == Type::ShortTy)
1156 || (I.getType() == Type::FloatTy)) {
1159 printType(Out, I.getType());
1163 writeOperand(I.getOperand(0));
1165 switch (I.getOpcode()) {
1166 case Instruction::Add: Out << " + "; break;
1167 case Instruction::Sub: Out << " - "; break;
1168 case Instruction::Mul: Out << "*"; break;
1169 case Instruction::Div: Out << "/"; break;
1170 case Instruction::Rem: Out << "%"; break;
1171 case Instruction::And: Out << " & "; break;
1172 case Instruction::Or: Out << " | "; break;
1173 case Instruction::Xor: Out << " ^ "; break;
1174 case Instruction::SetEQ: Out << " == "; break;
1175 case Instruction::SetNE: Out << " != "; break;
1176 case Instruction::SetLE: Out << " <= "; break;
1177 case Instruction::SetGE: Out << " >= "; break;
1178 case Instruction::SetLT: Out << " < "; break;
1179 case Instruction::SetGT: Out << " > "; break;
1180 case Instruction::Shl : Out << " << "; break;
1181 case Instruction::Shr : Out << " >> "; break;
1182 default: std::cerr << "Invalid operator type!" << I; abort();
1185 writeOperand(I.getOperand(1));
1192 void CWriter::visitCastInst(CastInst &I) {
1193 if (I.getType() == Type::BoolTy) {
1195 writeOperand(I.getOperand(0));
1200 printType(Out, I.getType());
1202 if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
1203 isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
1204 // Avoid "cast to pointer from integer of different size" warnings
1208 writeOperand(I.getOperand(0));
1211 void CWriter::visitSelectInst(SelectInst &I) {
1213 writeOperand(I.getCondition());
1215 writeOperand(I.getTrueValue());
1217 writeOperand(I.getFalseValue());
1222 void CWriter::lowerIntrinsics(Module &M) {
1223 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1224 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1225 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1226 if (CallInst *CI = dyn_cast<CallInst>(I++))
1227 if (Function *F = CI->getCalledFunction())
1228 switch (F->getIntrinsicID()) {
1229 case Intrinsic::not_intrinsic:
1230 case Intrinsic::vastart:
1231 case Intrinsic::vacopy:
1232 case Intrinsic::vaend:
1233 case Intrinsic::returnaddress:
1234 case Intrinsic::frameaddress:
1235 case Intrinsic::setjmp:
1236 case Intrinsic::longjmp:
1237 // We directly implement these intrinsics
1240 // All other intrinsic calls we must lower.
1241 Instruction *Before = CI->getPrev();
1242 IL.LowerIntrinsicCall(CI);
1243 if (Before) { // Move iterator to instruction after call
1253 void CWriter::visitCallInst(CallInst &I) {
1254 // Handle intrinsic function calls first...
1255 if (Function *F = I.getCalledFunction())
1256 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1258 default: assert(0 && "Unknown LLVM intrinsic!");
1259 case Intrinsic::vastart:
1262 Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1263 // Output the last argument to the enclosing function...
1264 if (I.getParent()->getParent()->aempty()) {
1265 std::cerr << "The C backend does not currently support zero "
1266 << "argument varargs functions, such as '"
1267 << I.getParent()->getParent()->getName() << "'!\n";
1270 writeOperand(&I.getParent()->getParent()->aback());
1273 case Intrinsic::vaend:
1274 Out << "va_end(*(va_list*)&";
1275 writeOperand(I.getOperand(1));
1278 case Intrinsic::vacopy:
1280 Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1281 Out << "*(va_list*)&";
1282 writeOperand(I.getOperand(1));
1285 case Intrinsic::returnaddress:
1286 Out << "__builtin_return_address(";
1287 writeOperand(I.getOperand(1));
1290 case Intrinsic::frameaddress:
1291 Out << "__builtin_frame_address(";
1292 writeOperand(I.getOperand(1));
1295 case Intrinsic::setjmp:
1296 Out << "setjmp(*(jmp_buf*)";
1297 writeOperand(I.getOperand(1));
1300 case Intrinsic::longjmp:
1301 Out << "longjmp(*(jmp_buf*)";
1302 writeOperand(I.getOperand(1));
1304 writeOperand(I.getOperand(2));
1312 void CWriter::visitCallSite(CallSite CS) {
1313 const PointerType *PTy = cast<PointerType>(CS.getCalledValue()->getType());
1314 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1315 const Type *RetTy = FTy->getReturnType();
1317 writeOperand(CS.getCalledValue());
1320 if (CS.arg_begin() != CS.arg_end()) {
1321 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
1324 for (++AI; AI != AE; ++AI) {
1332 void CWriter::visitMallocInst(MallocInst &I) {
1333 assert(0 && "lowerallocations pass didn't work!");
1336 void CWriter::visitAllocaInst(AllocaInst &I) {
1338 printType(Out, I.getType());
1339 Out << ") alloca(sizeof(";
1340 printType(Out, I.getType()->getElementType());
1342 if (I.isArrayAllocation()) {
1344 writeOperand(I.getOperand(0));
1349 void CWriter::visitFreeInst(FreeInst &I) {
1350 assert(0 && "lowerallocations pass didn't work!");
1353 void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
1354 gep_type_iterator E) {
1355 bool HasImplicitAddress = false;
1356 // If accessing a global value with no indexing, avoid *(&GV) syndrome
1357 if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
1358 HasImplicitAddress = true;
1359 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Ptr)) {
1360 HasImplicitAddress = true;
1361 Ptr = CPR->getValue(); // Get to the global...
1362 } else if (isDirectAlloca(Ptr)) {
1363 HasImplicitAddress = true;
1367 if (!HasImplicitAddress)
1368 Out << "*"; // Implicit zero first argument: '*x' is equivalent to 'x[0]'
1370 writeOperandInternal(Ptr);
1374 const Constant *CI = dyn_cast<Constant>(I.getOperand());
1375 if (HasImplicitAddress && (!CI || !CI->isNullValue()))
1378 writeOperandInternal(Ptr);
1380 if (HasImplicitAddress && (!CI || !CI->isNullValue())) {
1382 HasImplicitAddress = false; // HIA is only true if we haven't addressed yet
1385 assert(!HasImplicitAddress || (CI && CI->isNullValue()) &&
1386 "Can only have implicit address with direct accessing");
1388 if (HasImplicitAddress) {
1390 } else if (CI && CI->isNullValue()) {
1391 gep_type_iterator TmpI = I; ++TmpI;
1393 // Print out the -> operator if possible...
1394 if (TmpI != E && isa<StructType>(*TmpI)) {
1395 Out << (HasImplicitAddress ? "." : "->");
1396 Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
1402 if (isa<StructType>(*I)) {
1403 Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
1406 writeOperand(I.getOperand());
1411 void CWriter::visitLoadInst(LoadInst &I) {
1413 writeOperand(I.getOperand(0));
1416 void CWriter::visitStoreInst(StoreInst &I) {
1418 writeOperand(I.getPointerOperand());
1420 writeOperand(I.getOperand(0));
1423 void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
1425 printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
1429 void CWriter::visitVANextInst(VANextInst &I) {
1430 Out << Mang->getValueName(I.getOperand(0));
1431 Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
1432 printType(Out, I.getArgType());
1436 void CWriter::visitVAArgInst(VAArgInst &I) {
1438 Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
1439 writeOperand(I.getOperand(0));
1440 Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
1441 printType(Out, I.getType());
1442 Out << ");\n va_end(Tmp); }";
1445 //===----------------------------------------------------------------------===//
1446 // External Interface declaration
1447 //===----------------------------------------------------------------------===//
1449 bool CTargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &o) {
1450 PM.add(createLowerAllocationsPass());
1451 PM.add(createLowerInvokePass());
1452 PM.add(new CWriter(o, getIntrinsicLowering()));
1456 TargetMachine *llvm::allocateCTargetMachine(const Module &M,
1457 IntrinsicLowering *IL) {
1458 return new CTargetMachine(M, IL);