//===-- Writer.cpp - Library for converting LLVM code to C ----------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
//
// This library converts LLVM code to C code, compilable by GCC.
//
#include "llvm/Intrinsics.h"
#include "llvm/Analysis/FindUsedTypes.h"
#include "llvm/Analysis/ConstantsScanner.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/InstIterator.h"
-#include "llvm/Support/CallSite.h"
#include "llvm/Support/Mangler.h"
#include "Support/StringExtras.h"
#include "Support/STLExtras.h"
+#include "Config/config.h"
#include <algorithm>
#include <sstream>
-
-/* FIXME: This should be autoconf'd! */
-#define HAS_C99_HEXADECIMAL_CONSTANTS 1
+namespace llvm {
namespace {
class CWriter : public Pass, public InstVisitor<CWriter> {
std::ostream &Out;
Mangler *Mang;
const Module *TheModule;
+ FindUsedTypes *FUT;
+
std::map<const Type *, std::string> TypeNames;
std::set<const Value*> MangledGlobals;
bool needsMalloc, emittedInvoke;
virtual bool run(Module &M) {
// Initialize
TheModule = &M;
+ FUT = &getAnalysis<FindUsedTypes>();
// Ensure that all structure types have names...
bool Changed = nameAllUsedStructureTypes(M);
std::ostream &printType(std::ostream &Out, const Type *Ty,
const std::string &VariableName = "",
- bool IgnoreName = false, bool namedContext = true);
+ bool IgnoreName = false);
void writeOperand(Value *Operand);
void writeOperandInternal(Value *Operand);
static bool isInlinableInst(const Instruction &I) {
// Must be an expression, must be used exactly once. If it is dead, we
// emit it inline where it would go.
- if (I.getType() == Type::VoidTy || I.use_size() != 1 ||
+ if (I.getType() == Type::VoidTy || !I.hasOneUse() ||
isa<TerminatorInst>(I) || isa<CallInst>(I) || isa<PHINode>(I) ||
- isa<LoadInst>(I) || isa<VarArgInst>(I))
+ isa<LoadInst>(I) || isa<VAArgInst>(I) || isa<VANextInst>(I))
// Don't inline a load across a store or other bad things!
return false;
void visitLoadInst (LoadInst &I);
void visitStoreInst (StoreInst &I);
void visitGetElementPtrInst(GetElementPtrInst &I);
- void visitVarArgInst(VarArgInst &I);
+ void visitVANextInst(VANextInst &I);
+ void visitVAArgInst (VAArgInst &I);
void visitInstruction(Instruction &I) {
std::cerr << "C Writer does not know about " << I;
}
void printBranchToBlock(BasicBlock *CurBlock, BasicBlock *SuccBlock,
unsigned Indent);
- void printIndexingExpression(Value *Ptr, User::op_iterator I,
- User::op_iterator E);
+ void printIndexingExpression(Value *Ptr, gep_type_iterator I,
+ gep_type_iterator E);
};
-}
-
-// A pointer type should not use parens around *'s alone, e.g., (**)
-inline bool ptrTypeNameNeedsParens(const std::string &NameSoFar) {
- return NameSoFar.find_last_not_of('*') != std::string::npos;
-}
// Pass the Type* and the variable name and this prints out the variable
// declaration.
//
std::ostream &CWriter::printType(std::ostream &Out, const Type *Ty,
const std::string &NameSoFar,
- bool IgnoreName, bool namedContext) {
+ bool IgnoreName) {
if (Ty->isPrimitiveType())
switch (Ty->getPrimitiveID()) {
case Type::VoidTyID: return Out << "void " << NameSoFar;
const PointerType *PTy = cast<PointerType>(Ty);
std::string ptrName = "*" + NameSoFar;
- // Do not need parens around "* NameSoFar" if NameSoFar consists only
- // of zero or more '*' chars *and* this is not an unnamed pointer type
- // such as the result type in a cast statement. Otherwise, enclose in ( ).
- if (ptrTypeNameNeedsParens(NameSoFar) || !namedContext ||
- PTy->getElementType()->getPrimitiveID() == Type::ArrayTyID)
- ptrName = "(" + ptrName + ")"; //
+ if (isa<ArrayType>(PTy->getElementType()))
+ ptrName = "(" + ptrName + ")";
return printType(Out, PTy->getElementType(), ptrName);
}
// compiler agreeing on the conversion process (which is pretty likely since we
// only deal in IEEE FP).
//
-static bool isFPCSafeToPrint(const ConstantFP *CFP) {
-#if HAS_C99_HEXADECIMAL_CONSTANTS
+bool isFPCSafeToPrint(const ConstantFP *CFP) {
+#if HAVE_PRINTF_A
char Buffer[100];
sprintf(Buffer, "%a", CFP->getValue());
case Instruction::GetElementPtr:
Out << "(&(";
- printIndexingExpression(CE->getOperand(0),
- CPV->op_begin()+1, CPV->op_end());
+ printIndexingExpression(CE->getOperand(0), gep_type_begin(CPV),
+ gep_type_end(CPV));
Out << "))";
return;
case Instruction::Add:
case Instruction::SetLE:
case Instruction::SetGT:
case Instruction::SetGE:
+ case Instruction::Shl:
+ case Instruction::Shr:
Out << "(";
printConstant(CE->getOperand(0));
switch (CE->getOpcode()) {
case Instruction::SetLE: Out << " <= "; break;
case Instruction::SetGT: Out << " > "; break;
case Instruction::SetGE: Out << " >= "; break;
+ case Instruction::Shl: Out << " << "; break;
+ case Instruction::Shr: Out << " >> "; break;
default: assert(0 && "Illegal opcode here!");
}
printConstant(CE->getOperand(1));
Out << "(*(" << (FPC->getType() == Type::FloatTy ? "float" : "double")
<< "*)&FPConstant" << I->second << ")";
} else {
-#if HAS_C99_HEXADECIMAL_CONSTANTS
+#if HAVE_PRINTF_A
// Print out the constant as a floating point number.
char Buffer[100];
sprintf(Buffer, "%a", FPC->getValue());
//
bool CWriter::nameAllUsedStructureTypes(Module &M) {
// Get a set of types that are used by the program...
- std::set<const Type *> UT = getAnalysis<FindUsedTypes>().getTypes();
+ std::set<const Type *> UT = FUT->getTypes();
// Loop over the module symbol table, removing types from UT that are already
// named.
// If we aren't being compiled with GCC, just drop these attributes.
Out << "#ifndef __GNUC__ /* Can only support \"linkonce\" vars with GCC */\n"
<< "#define __attribute__(X)\n"
- << "#endif\n";
+ << "#endif\n\n";
+
+#if 0
+ // At some point, we should support "external weak" vs. "weak" linkages.
+ // On Mac OS X, "external weak" is spelled "__attribute__((weak_import))".
+ Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
+ << "#define __EXTERNAL_WEAK__ __attribute__((weak_import))\n"
+ << "#elif defined(__GNUC__)\n"
+ << "#define __EXTERNAL_WEAK__ __attribute__((weak))\n"
+ << "#else\n"
+ << "#define __EXTERNAL_WEAK__\n"
+ << "#endif\n\n";
+#endif
+
+ // For now, turn off the weak linkage attribute on Mac OS X. (See above.)
+ Out << "#if defined(__GNUC__) && defined(__APPLE_CC__)\n"
+ << "#define __ATTRIBUTE_WEAK__\n"
+ << "#elif defined(__GNUC__)\n"
+ << "#define __ATTRIBUTE_WEAK__ __attribute__((weak))\n"
+ << "#else\n"
+ << "#define __ATTRIBUTE_WEAK__\n"
+ << "#endif\n\n";
+}
+
+// generateProcessorSpecificCode - This is where we add conditional compilation
+// directives to cater to specific processors as need be.
+//
+static void generateProcessorSpecificCode(std::ostream& Out) {
+ // According to ANSI C, longjmp'ing to a setjmp could invalidate any
+ // non-volatile variable in the scope of the setjmp. For now, we are not
+ // doing analysis to determine which variables need to be marked volatile, so
+ // we just mark them all.
+ //
+ // HOWEVER, many targets implement setjmp by saving and restoring the register
+ // file, so they DON'T need variables to be marked volatile, and this is a
+ // HUGE pessimization for them. For this reason, on known-good processors, we
+ // do not emit volatile qualifiers.
+ Out << "#if defined(__386__) || defined(__i386__) || \\\n"
+ << " defined(i386) || defined(WIN32)\n"
+ << "/* setjmp does not require variables to be marked volatile */"
+ << "#define VOLATILE_FOR_SETJMP\n"
+ << "#else\n"
+ << "#define VOLATILE_FOR_SETJMP volatile\n"
+ << "#endif\n\n";
}
+
void CWriter::printModule(Module *M) {
// Calculate which global values have names that will collide when we throw
// away type information.
// get declaration for alloca
Out << "/* Provide Declarations */\n";
- Out << "#include <stdarg.h>\n";
- Out << "#include <setjmp.h>\n";
+ Out << "#include <stdarg.h>\n"; // Varargs support
+ Out << "#include <setjmp.h>\n"; // Unwind support
generateCompilerSpecificCode(Out);
-
+ generateProcessorSpecificCode(Out);
+
// Provide a definition for `bool' if not compiling with a C++ compiler.
Out << "\n"
<< "#ifndef __cplusplus\ntypedef unsigned char bool;\n#endif\n"
if ((I->hasInternalLinkage() || !MangledGlobals.count(I)) &&
!I->getIntrinsicID()) {
printFunctionSignature(I, true);
+ if (I->hasWeakLinkage()) Out << " __ATTRIBUTE_WEAK__";
Out << ";\n";
}
}
if (!I->isExternal()) {
Out << "extern ";
printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
-
+
+ if (I->hasLinkOnceLinkage())
+ Out << " __attribute__((common))";
+ else if (I->hasWeakLinkage())
+ Out << " __ATTRIBUTE_WEAK__";
Out << ";\n";
}
}
printType(Out, I->getType()->getElementType(), Mang->getValueName(I));
if (I->hasLinkOnceLinkage())
Out << " __attribute__((common))";
- if (!I->getInitializer()->isNullValue()) {
+ else if (I->hasWeakLinkage())
+ Out << " __ATTRIBUTE_WEAK__";
+
+ // If the initializer is not null, emit the initializer. If it is null,
+ // we try to avoid emitting large amounts of zeros. The problem with
+ // this, however, occurs when the variable has weak linkage. In this
+ // case, the assembler will complain about the variable being both weak
+ // and common, so we disable this optimization.
+ if (!I->getInitializer()->isNullValue() ||
+ I->hasWeakLinkage()) {
Out << " = " ;
writeOperand(I->getInitializer());
}
// Print out forward declarations for structure types before anything else!
Out << "/* Structure forward decls */\n";
for (; I != End; ++I)
- if (const Type *STy = dyn_cast<StructType>(I->second)) {
- std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
- Out << Name << ";\n";
- TypeNames.insert(std::make_pair(STy, Name));
- }
+ if (const Type *STy = dyn_cast<StructType>(I->second))
+ // Only print out used types!
+ if (FUT->getTypes().count(STy)) {
+ std::string Name = "struct l_" + Mangler::makeNameProper(I->first);
+ Out << Name << ";\n";
+ TypeNames.insert(std::make_pair(STy, Name));
+ }
Out << "\n";
// Now we can print out typedefs...
Out << "/* Typedefs */\n";
- for (I = ST.type_begin(Type::TypeTy); I != End; ++I) {
- const Type *Ty = cast<Type>(I->second);
- std::string Name = "l_" + Mangler::makeNameProper(I->first);
- Out << "typedef ";
- printType(Out, Ty, Name);
- Out << ";\n";
- }
-
+ for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
+ // Only print out used types!
+ if (FUT->getTypes().count(cast<Type>(I->second))) {
+ const Type *Ty = cast<Type>(I->second);
+ std::string Name = "l_" + Mangler::makeNameProper(I->first);
+ Out << "typedef ";
+ printType(Out, Ty, Name);
+ Out << ";\n";
+ }
+
Out << "\n";
// Keep track of which structures have been printed so far...
Out << "/* Structure contents */\n";
for (I = ST.type_begin(Type::TypeTy); I != End; ++I)
if (const StructType *STy = dyn_cast<StructType>(I->second))
- printContainedStructs(STy, StructPrinted);
+ // Only print out used types!
+ if (FUT->getTypes().count(STy))
+ printContainedStructs(STy, StructPrinted);
}
// Push the struct onto the stack and recursively push all structs
if (FT->isVarArg() && !FT->getParamTypes().empty()) {
if (FT->getParamTypes().size()) FunctionInnards << ", ";
FunctionInnards << "..."; // Output varargs portion of signature!
+ } else if (!FT->isVarArg() && FT->getParamTypes().empty()) {
+ FunctionInnards << "void"; // ret() -> ret(void) in C.
}
FunctionInnards << ")";
// Print out the return type and the entire signature for that matter
printFunctionSignature(F, false);
Out << " {\n";
+ // Determine whether or not the function contains any invoke instructions.
+ bool HasInvoke = false;
+ for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
+ if (isa<InvokeInst>(I->getTerminator())) {
+ HasInvoke = true;
+ break;
+ }
+
// print local variable information for the function
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
if (const AllocaInst *AI = isDirectAlloca(*I)) {
Out << " ";
+ if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
printType(Out, AI->getAllocatedType(), Mang->getValueName(AI));
Out << "; /* Address exposed local */\n";
} else if ((*I)->getType() != Type::VoidTy && !isInlinableInst(**I)) {
Out << " ";
+ if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
printType(Out, (*I)->getType(), Mang->getValueName(*I));
Out << ";\n";
if (isa<PHINode>(*I)) { // Print out PHI node temporaries as well...
Out << " ";
+ if (HasInvoke) Out << "VOLATILE_FOR_SETJMP ";
printType(Out, (*I)->getType(),
Mang->getValueName(*I)+"__PHI_TEMPORARY");
Out << ";\n";
// instruction to add an entry to the top of the jmpbuf_list. Thus, here we
// just have to longjmp to the specified handler.
Out << " if (__llvm_jmpbuf_list == 0) { /* unwind */\n"
+ << "#ifdef _LP64\n"
+ << " extern signed long long write();\n"
+ << "#else\n"
<< " extern write();\n"
+ << "#endif\n"
<< " ((void (*)(int, void*, unsigned))write)(2,\n"
<< " \"throw found with no handler!\\n\", 31); abort();\n"
<< " }\n"
emittedInvoke = true;
}
-static bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
+bool isGotoCodeNecessary(BasicBlock *From, BasicBlock *To) {
// If PHI nodes need copies, we need the copy code...
if (isa<PHINode>(To->front()) ||
From->getNext() != To) // Not directly successor, need goto
Out << "; /* for PHI node */\n";
}
- if (CurBB->getNext() != Succ || isa<InvokeInst>(CurBB->getTerminator())) {
+ if (CurBB->getNext() != Succ ||
+ isa<InvokeInst>(CurBB->getTerminator()) ||
+ isa<SwitchInst>(CurBB->getTerminator())) {
Out << std::string(Indent, ' ') << " goto ";
writeOperand(Succ);
Out << ";\n";
|| (I.getType() == Type::FloatTy)) {
needsCast = true;
Out << "((";
- printType(Out, I.getType(), "", false, false);
+ printType(Out, I.getType());
Out << ")(";
}
return;
}
Out << "(";
- printType(Out, I.getType(), "", /*ignoreName*/false, /*namedContext*/false);
+ printType(Out, I.getType());
Out << ")";
if (isa<PointerType>(I.getType())&&I.getOperand(0)->getType()->isIntegral() ||
isa<PointerType>(I.getOperand(0)->getType())&&I.getType()->isIntegral()) {
void CWriter::visitCallInst(CallInst &I) {
// Handle intrinsic function calls first...
if (Function *F = I.getCalledFunction())
- if (LLVMIntrinsic::ID ID = (LLVMIntrinsic::ID)F->getIntrinsicID()) {
+ if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
switch (ID) {
default: assert(0 && "Unknown LLVM intrinsic!");
- case LLVMIntrinsic::va_start:
- Out << "va_start(*(va_list*)";
- writeOperand(I.getOperand(1));
- Out << ", ";
+ case Intrinsic::va_start:
+ Out << "0; ";
+
+ Out << "va_start(*(va_list*)&" << Mang->getValueName(&I) << ", ";
// Output the last argument to the enclosing function...
+ if (I.getParent()->getParent()->aempty()) {
+ std::cerr << "The C backend does not currently support zero "
+ << "argument varargs functions, such as '"
+ << I.getParent()->getParent()->getName() << "'!\n";
+ abort();
+ }
writeOperand(&I.getParent()->getParent()->aback());
Out << ")";
return;
- case LLVMIntrinsic::va_end:
- Out << "va_end(*(va_list*)";
+ case Intrinsic::va_end:
+ Out << "va_end(*(va_list*)&";
writeOperand(I.getOperand(1));
Out << ")";
return;
- case LLVMIntrinsic::va_copy:
- Out << "va_copy(*(va_list*)";
+ case Intrinsic::va_copy:
+ Out << "0;";
+ Out << "va_copy(*(va_list*)&" << Mang->getValueName(&I) << ", ";
+ Out << "*(va_list*)&";
writeOperand(I.getOperand(1));
- Out << ", (va_list)";
- writeOperand(I.getOperand(2));
Out << ")";
return;
-
- case LLVMIntrinsic::setjmp:
- case LLVMIntrinsic::sigsetjmp:
+ case Intrinsic::setjmp:
+ case Intrinsic::sigsetjmp:
// This intrinsic should never exist in the program, but until we get
// setjmp/longjmp transformations going on, we should codegen it to
// something reasonable. This will allow code that never calls longjmp
// to work.
Out << "0";
return;
- case LLVMIntrinsic::longjmp:
- case LLVMIntrinsic::siglongjmp:
+ case Intrinsic::longjmp:
+ case Intrinsic::siglongjmp:
// Longjmp is not implemented, and never will be. It would cause an
// exception throw.
Out << "abort()";
Out << ")";
}
-void CWriter::printIndexingExpression(Value *Ptr, User::op_iterator I,
- User::op_iterator E) {
+void CWriter::printIndexingExpression(Value *Ptr, gep_type_iterator I,
+ gep_type_iterator E) {
bool HasImplicitAddress = false;
// If accessing a global value with no indexing, avoid *(&GV) syndrome
if (GlobalValue *V = dyn_cast<GlobalValue>(Ptr)) {
return;
}
- const Constant *CI = dyn_cast<Constant>(I);
+ const Constant *CI = dyn_cast<Constant>(I.getOperand());
if (HasImplicitAddress && (!CI || !CI->isNullValue()))
Out << "(&";
if (HasImplicitAddress) {
++I;
- } else if (CI && CI->isNullValue() && I+1 != E) {
+ } else if (CI && CI->isNullValue()) {
+ gep_type_iterator TmpI = I; ++TmpI;
+
// Print out the -> operator if possible...
- if ((*(I+1))->getType() == Type::UByteTy) {
+ if (TmpI != E && isa<StructType>(*TmpI)) {
Out << (HasImplicitAddress ? "." : "->");
- Out << "field" << cast<ConstantUInt>(*(I+1))->getValue();
- I += 2;
- }
+ Out << "field" << cast<ConstantUInt>(TmpI.getOperand())->getValue();
+ I = ++TmpI;
+ }
}
for (; I != E; ++I)
- if ((*I)->getType() == Type::LongTy) {
+ if (isa<StructType>(*I)) {
+ Out << ".field" << cast<ConstantUInt>(I.getOperand())->getValue();
+ } else {
Out << "[";
- writeOperand(*I);
+ writeOperand(I.getOperand());
Out << "]";
- } else {
- Out << ".field" << cast<ConstantUInt>(*I)->getValue();
}
}
void CWriter::visitGetElementPtrInst(GetElementPtrInst &I) {
Out << "&";
- printIndexingExpression(I.getPointerOperand(), I.idx_begin(), I.idx_end());
+ printIndexingExpression(I.getPointerOperand(), gep_type_begin(I),
+ gep_type_end(I));
}
-void CWriter::visitVarArgInst(VarArgInst &I) {
- Out << "va_arg((va_list)*";
- writeOperand(I.getOperand(0));
- Out << ", ";
- printType(Out, I.getType(), "", /*ignoreName*/false, /*namedContext*/false);
+void CWriter::visitVANextInst(VANextInst &I) {
+ Out << Mang->getValueName(I.getOperand(0));
+ Out << "; va_arg(*(va_list*)&" << Mang->getValueName(&I) << ", ";
+ printType(Out, I.getArgType());
Out << ")";
}
+void CWriter::visitVAArgInst(VAArgInst &I) {
+ Out << "0;\n";
+ Out << "{ va_list Tmp; va_copy(Tmp, *(va_list*)&";
+ writeOperand(I.getOperand(0));
+ Out << ");\n " << Mang->getValueName(&I) << " = va_arg(Tmp, ";
+ printType(Out, I.getType());
+ Out << ");\n va_end(Tmp); }";
+}
+
+}
//===----------------------------------------------------------------------===//
// External Interface declaration
//===----------------------------------------------------------------------===//
Pass *createWriteToCPass(std::ostream &o) { return new CWriter(o); }
+
+} // End llvm namespace