1 //===-- Writer.cpp - Library for writing LLVM bytecode files --------------===//
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 implements the functionality defined in llvm/Bytecode/Writer.h
12 // Note that this file uses an unusual technique of outputting all the bytecode
13 // to a vector of unsigned char, then copies the vector to an ostream. The
14 // reason for this is that we must do "seeking" in the stream to do back-
15 // patching, and some very important ostreams that we want to support (like
16 // pipes) do not support seeking. :( :( :(
18 //===----------------------------------------------------------------------===//
20 #include "WriterInternals.h"
21 #include "llvm/Bytecode/WriteBytecodePass.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/InlineAsm.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Module.h"
28 #include "llvm/SymbolTable.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Support/Compressor.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/System/Program.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/Statistic.h"
39 /// This value needs to be incremented every time the bytecode format changes
40 /// so that the reader can distinguish which format of the bytecode file has
42 /// @brief The bytecode version number
43 const unsigned BCVersionNum = 6;
45 static RegisterPass<WriteBytecodePass> X("emitbytecode", "Bytecode Writer");
48 BytesWritten("bytecodewriter", "Number of bytecode bytes written");
50 //===----------------------------------------------------------------------===//
51 //=== Output Primitives ===//
52 //===----------------------------------------------------------------------===//
54 // output - If a position is specified, it must be in the valid portion of the
55 // string... note that this should be inlined always so only the relevant IF
56 // body should be included.
57 inline void BytecodeWriter::output(unsigned i, int pos) {
58 if (pos == -1) { // Be endian clean, little endian is our friend
59 Out.push_back((unsigned char)i);
60 Out.push_back((unsigned char)(i >> 8));
61 Out.push_back((unsigned char)(i >> 16));
62 Out.push_back((unsigned char)(i >> 24));
64 Out[pos ] = (unsigned char)i;
65 Out[pos+1] = (unsigned char)(i >> 8);
66 Out[pos+2] = (unsigned char)(i >> 16);
67 Out[pos+3] = (unsigned char)(i >> 24);
71 inline void BytecodeWriter::output(int i) {
75 /// output_vbr - Output an unsigned value, by using the least number of bytes
76 /// possible. This is useful because many of our "infinite" values are really
77 /// very small most of the time; but can be large a few times.
78 /// Data format used: If you read a byte with the high bit set, use the low
79 /// seven bits as data and then read another byte.
80 inline void BytecodeWriter::output_vbr(uint64_t i) {
82 if (i < 0x80) { // done?
83 Out.push_back((unsigned char)i); // We know the high bit is clear...
87 // Nope, we are bigger than a character, output the next 7 bits and set the
88 // high bit to say that there is more coming...
89 Out.push_back(0x80 | ((unsigned char)i & 0x7F));
90 i >>= 7; // Shift out 7 bits now...
94 inline void BytecodeWriter::output_vbr(unsigned i) {
96 if (i < 0x80) { // done?
97 Out.push_back((unsigned char)i); // We know the high bit is clear...
101 // Nope, we are bigger than a character, output the next 7 bits and set the
102 // high bit to say that there is more coming...
103 Out.push_back(0x80 | ((unsigned char)i & 0x7F));
104 i >>= 7; // Shift out 7 bits now...
108 inline void BytecodeWriter::output_typeid(unsigned i) {
112 this->output_vbr(0x00FFFFFF);
117 inline void BytecodeWriter::output_vbr(int64_t i) {
119 output_vbr(((uint64_t)(-i) << 1) | 1); // Set low order sign bit...
121 output_vbr((uint64_t)i << 1); // Low order bit is clear.
125 inline void BytecodeWriter::output_vbr(int i) {
127 output_vbr(((unsigned)(-i) << 1) | 1); // Set low order sign bit...
129 output_vbr((unsigned)i << 1); // Low order bit is clear.
132 inline void BytecodeWriter::output(const std::string &s) {
133 unsigned Len = s.length();
134 output_vbr(Len); // Strings may have an arbitrary length.
135 Out.insert(Out.end(), s.begin(), s.end());
138 inline void BytecodeWriter::output_data(const void *Ptr, const void *End) {
139 Out.insert(Out.end(), (const unsigned char*)Ptr, (const unsigned char*)End);
142 inline void BytecodeWriter::output_float(float& FloatVal) {
143 /// FIXME: This isn't optimal, it has size problems on some platforms
144 /// where FP is not IEEE.
145 uint32_t i = FloatToBits(FloatVal);
146 Out.push_back( static_cast<unsigned char>( (i ) & 0xFF));
147 Out.push_back( static_cast<unsigned char>( (i >> 8 ) & 0xFF));
148 Out.push_back( static_cast<unsigned char>( (i >> 16) & 0xFF));
149 Out.push_back( static_cast<unsigned char>( (i >> 24) & 0xFF));
152 inline void BytecodeWriter::output_double(double& DoubleVal) {
153 /// FIXME: This isn't optimal, it has size problems on some platforms
154 /// where FP is not IEEE.
155 uint64_t i = DoubleToBits(DoubleVal);
156 Out.push_back( static_cast<unsigned char>( (i ) & 0xFF));
157 Out.push_back( static_cast<unsigned char>( (i >> 8 ) & 0xFF));
158 Out.push_back( static_cast<unsigned char>( (i >> 16) & 0xFF));
159 Out.push_back( static_cast<unsigned char>( (i >> 24) & 0xFF));
160 Out.push_back( static_cast<unsigned char>( (i >> 32) & 0xFF));
161 Out.push_back( static_cast<unsigned char>( (i >> 40) & 0xFF));
162 Out.push_back( static_cast<unsigned char>( (i >> 48) & 0xFF));
163 Out.push_back( static_cast<unsigned char>( (i >> 56) & 0xFF));
166 inline BytecodeBlock::BytecodeBlock(unsigned ID, BytecodeWriter &w,
167 bool elideIfEmpty, bool hasLongFormat)
168 : Id(ID), Writer(w), ElideIfEmpty(elideIfEmpty), HasLongFormat(hasLongFormat){
172 w.output(0U); // For length in long format
174 w.output(0U); /// Place holder for ID and length for this block
179 inline BytecodeBlock::~BytecodeBlock() { // Do backpatch when block goes out
181 if (Loc == Writer.size() && ElideIfEmpty) {
182 // If the block is empty, and we are allowed to, do not emit the block at
184 Writer.resize(Writer.size()-(HasLongFormat?8:4));
189 Writer.output(unsigned(Writer.size()-Loc), int(Loc-4));
191 Writer.output(unsigned(Writer.size()-Loc) << 5 | (Id & 0x1F), int(Loc-4));
194 //===----------------------------------------------------------------------===//
195 //=== Constant Output ===//
196 //===----------------------------------------------------------------------===//
198 void BytecodeWriter::outputType(const Type *T) {
199 output_vbr((unsigned)T->getTypeID());
201 // That's all there is to handling primitive types...
202 if (T->isPrimitiveType()) {
203 return; // We might do this if we alias a prim type: %x = type int
206 switch (T->getTypeID()) { // Handle derived types now.
207 case Type::FunctionTyID: {
208 const FunctionType *MT = cast<FunctionType>(T);
209 int Slot = Table.getSlot(MT->getReturnType());
210 assert(Slot != -1 && "Type used but not available!!");
211 output_typeid((unsigned)Slot);
213 // Output the number of arguments to function (+1 if varargs):
214 output_vbr((unsigned)MT->getNumParams()+MT->isVarArg());
216 // Output all of the arguments...
217 FunctionType::param_iterator I = MT->param_begin();
218 for (; I != MT->param_end(); ++I) {
219 Slot = Table.getSlot(*I);
220 assert(Slot != -1 && "Type used but not available!!");
221 output_typeid((unsigned)Slot);
224 // Terminate list with VoidTy if we are a varargs function...
226 output_typeid((unsigned)Type::VoidTyID);
230 case Type::ArrayTyID: {
231 const ArrayType *AT = cast<ArrayType>(T);
232 int Slot = Table.getSlot(AT->getElementType());
233 assert(Slot != -1 && "Type used but not available!!");
234 output_typeid((unsigned)Slot);
235 output_vbr(AT->getNumElements());
239 case Type::PackedTyID: {
240 const PackedType *PT = cast<PackedType>(T);
241 int Slot = Table.getSlot(PT->getElementType());
242 assert(Slot != -1 && "Type used but not available!!");
243 output_typeid((unsigned)Slot);
244 output_vbr(PT->getNumElements());
249 case Type::StructTyID: {
250 const StructType *ST = cast<StructType>(T);
252 // Output all of the element types...
253 for (StructType::element_iterator I = ST->element_begin(),
254 E = ST->element_end(); I != E; ++I) {
255 int Slot = Table.getSlot(*I);
256 assert(Slot != -1 && "Type used but not available!!");
257 output_typeid((unsigned)Slot);
260 // Terminate list with VoidTy
261 output_typeid((unsigned)Type::VoidTyID);
265 case Type::PointerTyID: {
266 const PointerType *PT = cast<PointerType>(T);
267 int Slot = Table.getSlot(PT->getElementType());
268 assert(Slot != -1 && "Type used but not available!!");
269 output_typeid((unsigned)Slot);
273 case Type::OpaqueTyID:
274 // No need to emit anything, just the count of opaque types is enough.
278 std::cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
279 << " Type '" << T->getDescription() << "'\n";
284 void BytecodeWriter::outputConstant(const Constant *CPV) {
285 assert((CPV->getType()->isPrimitiveType() || !CPV->isNullValue()) &&
286 "Shouldn't output null constants!");
288 // We must check for a ConstantExpr before switching by type because
289 // a ConstantExpr can be of any type, and has no explicit value.
291 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
292 // FIXME: Encoding of constant exprs could be much more compact!
293 assert(CE->getNumOperands() > 0 && "ConstantExpr with 0 operands");
294 assert(CE->getNumOperands() != 1 || CE->getOpcode() == Instruction::Cast);
295 output_vbr(1+CE->getNumOperands()); // flags as an expr
296 output_vbr(CE->getOpcode()); // Put out the CE op code
298 for (User::const_op_iterator OI = CE->op_begin(); OI != CE->op_end(); ++OI){
299 int Slot = Table.getSlot(*OI);
300 assert(Slot != -1 && "Unknown constant used in ConstantExpr!!");
301 output_vbr((unsigned)Slot);
302 Slot = Table.getSlot((*OI)->getType());
303 output_typeid((unsigned)Slot);
306 } else if (isa<UndefValue>(CPV)) {
307 output_vbr(1U); // 1 -> UndefValue constant.
310 output_vbr(0U); // flag as not a ConstantExpr (i.e. 0 operands)
313 switch (CPV->getType()->getTypeID()) {
314 case Type::BoolTyID: // Boolean Types
315 if (cast<ConstantBool>(CPV)->getValue())
321 case Type::UByteTyID: // Unsigned integer types...
322 case Type::UShortTyID:
324 case Type::ULongTyID:
325 output_vbr(cast<ConstantInt>(CPV)->getZExtValue());
328 case Type::SByteTyID: // Signed integer types...
329 case Type::ShortTyID:
332 output_vbr(cast<ConstantInt>(CPV)->getSExtValue());
335 case Type::ArrayTyID: {
336 const ConstantArray *CPA = cast<ConstantArray>(CPV);
337 assert(!CPA->isString() && "Constant strings should be handled specially!");
339 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i) {
340 int Slot = Table.getSlot(CPA->getOperand(i));
341 assert(Slot != -1 && "Constant used but not available!!");
342 output_vbr((unsigned)Slot);
347 case Type::PackedTyID: {
348 const ConstantPacked *CP = cast<ConstantPacked>(CPV);
350 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i) {
351 int Slot = Table.getSlot(CP->getOperand(i));
352 assert(Slot != -1 && "Constant used but not available!!");
353 output_vbr((unsigned)Slot);
358 case Type::StructTyID: {
359 const ConstantStruct *CPS = cast<ConstantStruct>(CPV);
361 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i) {
362 int Slot = Table.getSlot(CPS->getOperand(i));
363 assert(Slot != -1 && "Constant used but not available!!");
364 output_vbr((unsigned)Slot);
369 case Type::PointerTyID:
370 assert(0 && "No non-null, non-constant-expr constants allowed!");
373 case Type::FloatTyID: { // Floating point types...
374 float Tmp = (float)cast<ConstantFP>(CPV)->getValue();
378 case Type::DoubleTyID: {
379 double Tmp = cast<ConstantFP>(CPV)->getValue();
385 case Type::LabelTyID:
387 std::cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
388 << " type '" << *CPV->getType() << "'\n";
394 /// outputInlineAsm - InlineAsm's get emitted to the constant pool, so they can
395 /// be shared by multiple uses.
396 void BytecodeWriter::outputInlineAsm(const InlineAsm *IA) {
397 // Output a marker, so we know when we have one one parsing the constant pool.
398 // Note that this encoding is 5 bytes: not very efficient for a marker. Since
399 // unique inline asms are rare, this should hardly matter.
402 output(IA->getAsmString());
403 output(IA->getConstraintString());
404 output_vbr(unsigned(IA->hasSideEffects()));
407 void BytecodeWriter::outputConstantStrings() {
408 SlotCalculator::string_iterator I = Table.string_begin();
409 SlotCalculator::string_iterator E = Table.string_end();
410 if (I == E) return; // No strings to emit
412 // If we have != 0 strings to emit, output them now. Strings are emitted into
413 // the 'void' type plane.
414 output_vbr(unsigned(E-I));
415 output_typeid(Type::VoidTyID);
417 // Emit all of the strings.
418 for (I = Table.string_begin(); I != E; ++I) {
419 const ConstantArray *Str = *I;
420 int Slot = Table.getSlot(Str->getType());
421 assert(Slot != -1 && "Constant string of unknown type?");
422 output_typeid((unsigned)Slot);
424 // Now that we emitted the type (which indicates the size of the string),
425 // emit all of the characters.
426 std::string Val = Str->getAsString();
427 output_data(Val.c_str(), Val.c_str()+Val.size());
431 //===----------------------------------------------------------------------===//
432 //=== Instruction Output ===//
433 //===----------------------------------------------------------------------===//
435 // outputInstructionFormat0 - Output those weird instructions that have a large
436 // number of operands or have large operands themselves.
438 // Format: [opcode] [type] [numargs] [arg0] [arg1] ... [arg<numargs-1>]
440 void BytecodeWriter::outputInstructionFormat0(const Instruction *I,
442 const SlotCalculator &Table,
444 // Opcode must have top two bits clear...
445 output_vbr(Opcode << 2); // Instruction Opcode ID
446 output_typeid(Type); // Result type
448 unsigned NumArgs = I->getNumOperands();
449 output_vbr(NumArgs + (isa<CastInst>(I) ||
450 isa<VAArgInst>(I) || Opcode == 56 || Opcode == 58));
452 if (!isa<GetElementPtrInst>(&I)) {
453 for (unsigned i = 0; i < NumArgs; ++i) {
454 int Slot = Table.getSlot(I->getOperand(i));
455 assert(Slot >= 0 && "No slot number for value!?!?");
456 output_vbr((unsigned)Slot);
459 if (isa<CastInst>(I) || isa<VAArgInst>(I)) {
460 int Slot = Table.getSlot(I->getType());
461 assert(Slot != -1 && "Cast return type unknown?");
462 output_typeid((unsigned)Slot);
463 } else if (Opcode == 56) { // Invoke escape sequence
464 output_vbr(cast<InvokeInst>(I)->getCallingConv());
465 } else if (Opcode == 58) { // Call escape sequence
466 output_vbr((cast<CallInst>(I)->getCallingConv() << 1) |
467 unsigned(cast<CallInst>(I)->isTailCall()));
470 int Slot = Table.getSlot(I->getOperand(0));
471 assert(Slot >= 0 && "No slot number for value!?!?");
472 output_vbr(unsigned(Slot));
474 // We need to encode the type of sequential type indices into their slot #
476 for (gep_type_iterator TI = gep_type_begin(I), E = gep_type_end(I);
477 Idx != NumArgs; ++TI, ++Idx) {
478 Slot = Table.getSlot(I->getOperand(Idx));
479 assert(Slot >= 0 && "No slot number for value!?!?");
481 if (isa<SequentialType>(*TI)) {
483 switch (I->getOperand(Idx)->getType()->getTypeID()) {
484 default: assert(0 && "Unknown index type!");
485 case Type::UIntTyID: IdxId = 0; break;
486 case Type::IntTyID: IdxId = 1; break;
487 case Type::ULongTyID: IdxId = 2; break;
488 case Type::LongTyID: IdxId = 3; break;
490 Slot = (Slot << 2) | IdxId;
492 output_vbr(unsigned(Slot));
498 // outputInstrVarArgsCall - Output the absurdly annoying varargs function calls.
499 // This are more annoying than most because the signature of the call does not
500 // tell us anything about the types of the arguments in the varargs portion.
501 // Because of this, we encode (as type 0) all of the argument types explicitly
502 // before the argument value. This really sucks, but you shouldn't be using
503 // varargs functions in your code! *death to printf*!
505 // Format: [opcode] [type] [numargs] [arg0] [arg1] ... [arg<numargs-1>]
507 void BytecodeWriter::outputInstrVarArgsCall(const Instruction *I,
509 const SlotCalculator &Table,
511 assert(isa<CallInst>(I) || isa<InvokeInst>(I));
512 // Opcode must have top two bits clear...
513 output_vbr(Opcode << 2); // Instruction Opcode ID
514 output_typeid(Type); // Result type (varargs type)
516 const PointerType *PTy = cast<PointerType>(I->getOperand(0)->getType());
517 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
518 unsigned NumParams = FTy->getNumParams();
520 unsigned NumFixedOperands;
521 if (isa<CallInst>(I)) {
522 // Output an operand for the callee and each fixed argument, then two for
523 // each variable argument.
524 NumFixedOperands = 1+NumParams;
526 assert(isa<InvokeInst>(I) && "Not call or invoke??");
527 // Output an operand for the callee and destinations, then two for each
528 // variable argument.
529 NumFixedOperands = 3+NumParams;
531 output_vbr(2 * I->getNumOperands()-NumFixedOperands +
532 unsigned(Opcode == 56 || Opcode == 58));
534 // The type for the function has already been emitted in the type field of the
535 // instruction. Just emit the slot # now.
536 for (unsigned i = 0; i != NumFixedOperands; ++i) {
537 int Slot = Table.getSlot(I->getOperand(i));
538 assert(Slot >= 0 && "No slot number for value!?!?");
539 output_vbr((unsigned)Slot);
542 for (unsigned i = NumFixedOperands, e = I->getNumOperands(); i != e; ++i) {
543 // Output Arg Type ID
544 int Slot = Table.getSlot(I->getOperand(i)->getType());
545 assert(Slot >= 0 && "No slot number for value!?!?");
546 output_typeid((unsigned)Slot);
548 // Output arg ID itself
549 Slot = Table.getSlot(I->getOperand(i));
550 assert(Slot >= 0 && "No slot number for value!?!?");
551 output_vbr((unsigned)Slot);
554 // If this is the escape sequence for call, emit the tailcall/cc info.
556 const CallInst *CI = cast<CallInst>(I);
557 output_vbr((CI->getCallingConv() << 1) | unsigned(CI->isTailCall()));
558 } else if (Opcode == 56) { // Invoke escape sequence.
559 output_vbr(cast<InvokeInst>(I)->getCallingConv());
564 // outputInstructionFormat1 - Output one operand instructions, knowing that no
565 // operand index is >= 2^12.
567 inline void BytecodeWriter::outputInstructionFormat1(const Instruction *I,
571 // bits Instruction format:
572 // --------------------------
573 // 01-00: Opcode type, fixed to 1.
575 // 19-08: Resulting type plane
576 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
578 output(1 | (Opcode << 2) | (Type << 8) | (Slots[0] << 20));
582 // outputInstructionFormat2 - Output two operand instructions, knowing that no
583 // operand index is >= 2^8.
585 inline void BytecodeWriter::outputInstructionFormat2(const Instruction *I,
589 // bits Instruction format:
590 // --------------------------
591 // 01-00: Opcode type, fixed to 2.
593 // 15-08: Resulting type plane
597 output(2 | (Opcode << 2) | (Type << 8) | (Slots[0] << 16) | (Slots[1] << 24));
601 // outputInstructionFormat3 - Output three operand instructions, knowing that no
602 // operand index is >= 2^6.
604 inline void BytecodeWriter::outputInstructionFormat3(const Instruction *I,
608 // bits Instruction format:
609 // --------------------------
610 // 01-00: Opcode type, fixed to 3.
612 // 13-08: Resulting type plane
617 output(3 | (Opcode << 2) | (Type << 8) |
618 (Slots[0] << 14) | (Slots[1] << 20) | (Slots[2] << 26));
621 void BytecodeWriter::outputInstruction(const Instruction &I) {
622 assert(I.getOpcode() < 56 && "Opcode too big???");
623 unsigned Opcode = I.getOpcode();
624 unsigned NumOperands = I.getNumOperands();
626 // Encode 'tail call' as 61, 'volatile load' as 62, and 'volatile store' as
628 if (const CallInst *CI = dyn_cast<CallInst>(&I)) {
629 if (CI->getCallingConv() == CallingConv::C) {
630 if (CI->isTailCall())
631 Opcode = 61; // CCC + Tail Call
633 ; // Opcode = Instruction::Call
634 } else if (CI->getCallingConv() == CallingConv::Fast) {
635 if (CI->isTailCall())
636 Opcode = 59; // FastCC + TailCall
638 Opcode = 60; // FastCC + Not Tail Call
640 Opcode = 58; // Call escape sequence.
642 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(&I)) {
643 if (II->getCallingConv() == CallingConv::Fast)
644 Opcode = 57; // FastCC invoke.
645 else if (II->getCallingConv() != CallingConv::C)
646 Opcode = 56; // Invoke escape sequence.
648 } else if (isa<LoadInst>(I) && cast<LoadInst>(I).isVolatile()) {
650 } else if (isa<StoreInst>(I) && cast<StoreInst>(I).isVolatile()) {
654 // Figure out which type to encode with the instruction. Typically we want
655 // the type of the first parameter, as opposed to the type of the instruction
656 // (for example, with setcc, we always know it returns bool, but the type of
657 // the first param is actually interesting). But if we have no arguments
658 // we take the type of the instruction itself.
661 switch (I.getOpcode()) {
662 case Instruction::Select:
663 case Instruction::Malloc:
664 case Instruction::Alloca:
665 Ty = I.getType(); // These ALWAYS want to encode the return type
667 case Instruction::Store:
668 Ty = I.getOperand(1)->getType(); // Encode the pointer type...
669 assert(isa<PointerType>(Ty) && "Store to nonpointer type!?!?");
671 default: // Otherwise use the default behavior...
672 Ty = NumOperands ? I.getOperand(0)->getType() : I.getType();
677 int Slot = Table.getSlot(Ty);
678 assert(Slot != -1 && "Type not available!!?!");
679 Type = (unsigned)Slot;
681 // Varargs calls and invokes are encoded entirely different from any other
683 if (const CallInst *CI = dyn_cast<CallInst>(&I)){
684 const PointerType *Ty =cast<PointerType>(CI->getCalledValue()->getType());
685 if (cast<FunctionType>(Ty->getElementType())->isVarArg()) {
686 outputInstrVarArgsCall(CI, Opcode, Table, Type);
689 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(&I)) {
690 const PointerType *Ty =cast<PointerType>(II->getCalledValue()->getType());
691 if (cast<FunctionType>(Ty->getElementType())->isVarArg()) {
692 outputInstrVarArgsCall(II, Opcode, Table, Type);
697 if (NumOperands <= 3) {
698 // Make sure that we take the type number into consideration. We don't want
699 // to overflow the field size for the instruction format we select.
701 unsigned MaxOpSlot = Type;
702 unsigned Slots[3]; Slots[0] = (1 << 12)-1; // Marker to signify 0 operands
704 for (unsigned i = 0; i != NumOperands; ++i) {
705 int slot = Table.getSlot(I.getOperand(i));
706 assert(slot != -1 && "Broken bytecode!");
707 if (unsigned(slot) > MaxOpSlot) MaxOpSlot = unsigned(slot);
708 Slots[i] = unsigned(slot);
711 // Handle the special cases for various instructions...
712 if (isa<CastInst>(I) || isa<VAArgInst>(I)) {
713 // Cast has to encode the destination type as the second argument in the
714 // packet, or else we won't know what type to cast to!
715 Slots[1] = Table.getSlot(I.getType());
716 assert(Slots[1] != ~0U && "Cast return type unknown?");
717 if (Slots[1] > MaxOpSlot) MaxOpSlot = Slots[1];
719 } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(&I)) {
720 assert(NumOperands == 1 && "Bogus allocation!");
721 if (AI->getAlignment()) {
722 Slots[1] = Log2_32(AI->getAlignment())+1;
723 if (Slots[1] > MaxOpSlot) MaxOpSlot = Slots[1];
726 } else if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
727 // We need to encode the type of sequential type indices into their slot #
729 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
731 if (isa<SequentialType>(*I)) {
733 switch (GEP->getOperand(Idx)->getType()->getTypeID()) {
734 default: assert(0 && "Unknown index type!");
735 case Type::UIntTyID: IdxId = 0; break;
736 case Type::IntTyID: IdxId = 1; break;
737 case Type::ULongTyID: IdxId = 2; break;
738 case Type::LongTyID: IdxId = 3; break;
740 Slots[Idx] = (Slots[Idx] << 2) | IdxId;
741 if (Slots[Idx] > MaxOpSlot) MaxOpSlot = Slots[Idx];
743 } else if (Opcode == 58) {
744 // If this is the escape sequence for call, emit the tailcall/cc info.
745 const CallInst &CI = cast<CallInst>(I);
747 if (NumOperands <= 3) {
748 Slots[NumOperands-1] =
749 (CI.getCallingConv() << 1)|unsigned(CI.isTailCall());
750 if (Slots[NumOperands-1] > MaxOpSlot)
751 MaxOpSlot = Slots[NumOperands-1];
753 } else if (Opcode == 56) {
754 // Invoke escape seq has at least 4 operands to encode.
758 // Decide which instruction encoding to use. This is determined primarily
759 // by the number of operands, and secondarily by whether or not the max
760 // operand will fit into the instruction encoding. More operands == fewer
763 switch (NumOperands) {
766 if (MaxOpSlot < (1 << 12)-1) { // -1 because we use 4095 to indicate 0 ops
767 outputInstructionFormat1(&I, Opcode, Slots, Type);
773 if (MaxOpSlot < (1 << 8)) {
774 outputInstructionFormat2(&I, Opcode, Slots, Type);
780 if (MaxOpSlot < (1 << 6)) {
781 outputInstructionFormat3(&I, Opcode, Slots, Type);
790 // If we weren't handled before here, we either have a large number of
791 // operands or a large operand index that we are referring to.
792 outputInstructionFormat0(&I, Opcode, Table, Type);
795 //===----------------------------------------------------------------------===//
796 //=== Block Output ===//
797 //===----------------------------------------------------------------------===//
799 BytecodeWriter::BytecodeWriter(std::vector<unsigned char> &o, const Module *M)
802 // Emit the signature...
803 static const unsigned char *Sig = (const unsigned char*)"llvm";
804 output_data(Sig, Sig+4);
806 // Emit the top level CLASS block.
807 BytecodeBlock ModuleBlock(BytecodeFormat::ModuleBlockID, *this, false, true);
809 bool isBigEndian = M->getEndianness() == Module::BigEndian;
810 bool hasLongPointers = M->getPointerSize() == Module::Pointer64;
811 bool hasNoEndianness = M->getEndianness() == Module::AnyEndianness;
812 bool hasNoPointerSize = M->getPointerSize() == Module::AnyPointerSize;
814 // Output the version identifier and other information.
815 unsigned Version = (BCVersionNum << 4) |
816 (unsigned)isBigEndian | (hasLongPointers << 1) |
817 (hasNoEndianness << 2) |
818 (hasNoPointerSize << 3);
821 // The Global type plane comes first
823 BytecodeBlock CPool(BytecodeFormat::GlobalTypePlaneBlockID, *this);
824 outputTypes(Type::FirstDerivedTyID);
827 // The ModuleInfoBlock follows directly after the type information
828 outputModuleInfoBlock(M);
830 // Output module level constants, used for global variable initializers
831 outputConstants(false);
833 // Do the whole module now! Process each function at a time...
834 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I)
837 // If needed, output the symbol table for the module...
838 outputSymbolTable(M->getSymbolTable());
841 void BytecodeWriter::outputTypes(unsigned TypeNum) {
842 // Write the type plane for types first because earlier planes (e.g. for a
843 // primitive type like float) may have constants constructed using types
844 // coming later (e.g., via getelementptr from a pointer type). The type
845 // plane is needed before types can be fwd or bkwd referenced.
846 const std::vector<const Type*>& Types = Table.getTypes();
847 assert(!Types.empty() && "No types at all?");
848 assert(TypeNum <= Types.size() && "Invalid TypeNo index");
850 unsigned NumEntries = Types.size() - TypeNum;
852 // Output type header: [num entries]
853 output_vbr(NumEntries);
855 for (unsigned i = TypeNum; i < TypeNum+NumEntries; ++i)
856 outputType(Types[i]);
859 // Helper function for outputConstants().
860 // Writes out all the constants in the plane Plane starting at entry StartNo.
862 void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
863 &Plane, unsigned StartNo) {
864 unsigned ValNo = StartNo;
866 // Scan through and ignore function arguments, global values, and constant
868 for (; ValNo < Plane.size() &&
869 (isa<Argument>(Plane[ValNo]) || isa<GlobalValue>(Plane[ValNo]) ||
870 (isa<ConstantArray>(Plane[ValNo]) &&
871 cast<ConstantArray>(Plane[ValNo])->isString())); ValNo++)
874 unsigned NC = ValNo; // Number of constants
875 for (; NC < Plane.size() && (isa<Constant>(Plane[NC]) ||
876 isa<InlineAsm>(Plane[NC])); NC++)
878 NC -= ValNo; // Convert from index into count
879 if (NC == 0) return; // Skip empty type planes...
881 // FIXME: Most slabs only have 1 or 2 entries! We should encode this much
884 // Put out type header: [num entries][type id number]
888 // Put out the Type ID Number...
889 int Slot = Table.getSlot(Plane.front()->getType());
890 assert (Slot != -1 && "Type in constant pool but not in function!!");
891 output_typeid((unsigned)Slot);
893 for (unsigned i = ValNo; i < ValNo+NC; ++i) {
894 const Value *V = Plane[i];
895 if (const Constant *C = dyn_cast<Constant>(V))
898 outputInlineAsm(cast<InlineAsm>(V));
902 static inline bool hasNullValue(const Type *Ty) {
903 return Ty != Type::LabelTy && Ty != Type::VoidTy && !isa<OpaqueType>(Ty);
906 void BytecodeWriter::outputConstants(bool isFunction) {
907 BytecodeBlock CPool(BytecodeFormat::ConstantPoolBlockID, *this,
908 true /* Elide block if empty */);
910 unsigned NumPlanes = Table.getNumPlanes();
913 // Output the type plane before any constants!
914 outputTypes(Table.getModuleTypeLevel());
916 // Output module-level string constants before any other constants.
917 outputConstantStrings();
919 for (unsigned pno = 0; pno != NumPlanes; pno++) {
920 const std::vector<const Value*> &Plane = Table.getPlane(pno);
921 if (!Plane.empty()) { // Skip empty type planes...
923 if (isFunction) // Don't re-emit module constants
924 ValNo += Table.getModuleLevel(pno);
926 if (hasNullValue(Plane[0]->getType())) {
927 // Skip zero initializer
932 // Write out constants in the plane
933 outputConstantsInPlane(Plane, ValNo);
938 static unsigned getEncodedLinkage(const GlobalValue *GV) {
939 switch (GV->getLinkage()) {
940 default: assert(0 && "Invalid linkage!");
941 case GlobalValue::ExternalLinkage: return 0;
942 case GlobalValue::WeakLinkage: return 1;
943 case GlobalValue::AppendingLinkage: return 2;
944 case GlobalValue::InternalLinkage: return 3;
945 case GlobalValue::LinkOnceLinkage: return 4;
946 case GlobalValue::DLLImportLinkage: return 5;
947 case GlobalValue::DLLExportLinkage: return 6;
948 case GlobalValue::ExternalWeakLinkage: return 7;
952 void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
953 BytecodeBlock ModuleInfoBlock(BytecodeFormat::ModuleGlobalInfoBlockID, *this);
955 // Give numbers to sections as we encounter them.
956 unsigned SectionIDCounter = 0;
957 std::vector<std::string> SectionNames;
958 std::map<std::string, unsigned> SectionID;
960 // Output the types for the global variables in the module...
961 for (Module::const_global_iterator I = M->global_begin(),
962 End = M->global_end(); I != End; ++I) {
963 int Slot = Table.getSlot(I->getType());
964 assert(Slot != -1 && "Module global vars is broken!");
966 assert((I->hasInitializer() || !I->hasInternalLinkage()) &&
967 "Global must have an initializer or have external linkage!");
969 // Fields: bit0 = isConstant, bit1 = hasInitializer, bit2-4=Linkage,
970 // bit5+ = Slot # for type.
971 bool HasExtensionWord = (I->getAlignment() != 0) || I->hasSection();
973 // If we need to use the extension byte, set linkage=3(internal) and
974 // initializer = 0 (impossible!).
975 if (!HasExtensionWord) {
976 unsigned oSlot = ((unsigned)Slot << 5) | (getEncodedLinkage(I) << 2) |
977 (I->hasInitializer() << 1) | (unsigned)I->isConstant();
980 unsigned oSlot = ((unsigned)Slot << 5) | (3 << 2) |
981 (0 << 1) | (unsigned)I->isConstant();
984 // The extension word has this format: bit 0 = has initializer, bit 1-3 =
985 // linkage, bit 4-8 = alignment (log2), bit 9 = has SectionID,
986 // bits 10+ = future use.
987 unsigned ExtWord = (unsigned)I->hasInitializer() |
988 (getEncodedLinkage(I) << 1) |
989 ((Log2_32(I->getAlignment())+1) << 4) |
990 ((unsigned)I->hasSection() << 9);
992 if (I->hasSection()) {
993 // Give section names unique ID's.
994 unsigned &Entry = SectionID[I->getSection()];
996 Entry = ++SectionIDCounter;
997 SectionNames.push_back(I->getSection());
1003 // If we have an initializer, output it now.
1004 if (I->hasInitializer()) {
1005 Slot = Table.getSlot((Value*)I->getInitializer());
1006 assert(Slot != -1 && "No slot for global var initializer!");
1007 output_vbr((unsigned)Slot);
1010 output_typeid((unsigned)Table.getSlot(Type::VoidTy));
1012 // Output the types of the functions in this module.
1013 for (Module::const_iterator I = M->begin(), End = M->end(); I != End; ++I) {
1014 int Slot = Table.getSlot(I->getType());
1015 assert(Slot != -1 && "Module slot calculator is broken!");
1016 assert(Slot >= Type::FirstDerivedTyID && "Derived type not in range!");
1017 assert(((Slot << 6) >> 6) == Slot && "Slot # too big!");
1018 unsigned CC = I->getCallingConv()+1;
1019 unsigned ID = (Slot << 5) | (CC & 15);
1021 if (I->isExternal()) // If external, we don't have an FunctionInfo block.
1024 if (I->getAlignment() || I->hasSection() || (CC & ~15) != 0 ||
1025 (I->isExternal() && I->hasDLLImportLinkage()) ||
1026 (I->isExternal() && I->hasExternalWeakLinkage())
1028 ID |= 1 << 31; // Do we need an extension word?
1032 if (ID & (1 << 31)) {
1033 // Extension byte: bits 0-4 = alignment, bits 5-9 = top nibble of calling
1034 // convention, bit 10 = hasSectionID., bits 11-12 = external linkage type
1035 unsigned extLinkage = 0;
1037 if (I->isExternal()) {
1038 if (I->hasDLLImportLinkage()) {
1040 } else if (I->hasExternalWeakLinkage()) {
1045 ID = (Log2_32(I->getAlignment())+1) | ((CC >> 4) << 5) |
1046 (I->hasSection() << 10) |
1047 ((extLinkage & 3) << 11);
1050 // Give section names unique ID's.
1051 if (I->hasSection()) {
1052 unsigned &Entry = SectionID[I->getSection()];
1054 Entry = ++SectionIDCounter;
1055 SectionNames.push_back(I->getSection());
1061 output_vbr((unsigned)Table.getSlot(Type::VoidTy) << 5);
1063 // Emit the list of dependent libraries for the Module.
1064 Module::lib_iterator LI = M->lib_begin();
1065 Module::lib_iterator LE = M->lib_end();
1066 output_vbr(unsigned(LE - LI)); // Emit the number of dependent libraries.
1067 for (; LI != LE; ++LI)
1070 // Output the target triple from the module
1071 output(M->getTargetTriple());
1073 // Emit the table of section names.
1074 output_vbr((unsigned)SectionNames.size());
1075 for (unsigned i = 0, e = SectionNames.size(); i != e; ++i)
1076 output(SectionNames[i]);
1078 // Output the inline asm string.
1079 output(M->getModuleInlineAsm());
1082 void BytecodeWriter::outputInstructions(const Function *F) {
1083 BytecodeBlock ILBlock(BytecodeFormat::InstructionListBlockID, *this);
1084 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1085 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
1086 outputInstruction(*I);
1089 void BytecodeWriter::outputFunction(const Function *F) {
1090 // If this is an external function, there is nothing else to emit!
1091 if (F->isExternal()) return;
1093 BytecodeBlock FunctionBlock(BytecodeFormat::FunctionBlockID, *this);
1094 output_vbr(getEncodedLinkage(F));
1096 // Get slot information about the function...
1097 Table.incorporateFunction(F);
1099 if (Table.getCompactionTable().empty()) {
1100 // Output information about the constants in the function if the compaction
1101 // table is not being used.
1102 outputConstants(true);
1104 // Otherwise, emit the compaction table.
1105 outputCompactionTable();
1108 // Output all of the instructions in the body of the function
1109 outputInstructions(F);
1111 // If needed, output the symbol table for the function...
1112 outputSymbolTable(F->getSymbolTable());
1114 Table.purgeFunction();
1117 void BytecodeWriter::outputCompactionTablePlane(unsigned PlaneNo,
1118 const std::vector<const Value*> &Plane,
1120 unsigned End = Table.getModuleLevel(PlaneNo);
1121 if (Plane.empty() || StartNo == End || End == 0) return; // Nothing to emit
1122 assert(StartNo < End && "Cannot emit negative range!");
1123 assert(StartNo < Plane.size() && End <= Plane.size());
1125 // Do not emit the null initializer!
1128 // Figure out which encoding to use. By far the most common case we have is
1129 // to emit 0-2 entries in a compaction table plane.
1130 switch (End-StartNo) {
1131 case 0: // Avoid emitting two vbr's if possible.
1134 output_vbr((PlaneNo << 2) | End-StartNo);
1137 // Output the number of things.
1138 output_vbr((unsigned(End-StartNo) << 2) | 3);
1139 output_typeid(PlaneNo); // Emit the type plane this is
1143 for (unsigned i = StartNo; i != End; ++i)
1144 output_vbr(Table.getGlobalSlot(Plane[i]));
1147 void BytecodeWriter::outputCompactionTypes(unsigned StartNo) {
1148 // Get the compaction type table from the slot calculator
1149 const std::vector<const Type*> &CTypes = Table.getCompactionTypes();
1151 // The compaction types may have been uncompactified back to the
1152 // global types. If so, we just write an empty table
1153 if (CTypes.size() == 0) {
1158 assert(CTypes.size() >= StartNo && "Invalid compaction types start index");
1160 // Determine how many types to write
1161 unsigned NumTypes = CTypes.size() - StartNo;
1163 // Output the number of types.
1164 output_vbr(NumTypes);
1166 for (unsigned i = StartNo; i < StartNo+NumTypes; ++i)
1167 output_typeid(Table.getGlobalSlot(CTypes[i]));
1170 void BytecodeWriter::outputCompactionTable() {
1171 // Avoid writing the compaction table at all if there is no content.
1172 if (Table.getCompactionTypes().size() >= Type::FirstDerivedTyID ||
1173 (!Table.CompactionTableIsEmpty())) {
1174 BytecodeBlock CTB(BytecodeFormat::CompactionTableBlockID, *this,
1175 true/*ElideIfEmpty*/);
1176 const std::vector<std::vector<const Value*> > &CT =
1177 Table.getCompactionTable();
1179 // First things first, emit the type compaction table if there is one.
1180 outputCompactionTypes(Type::FirstDerivedTyID);
1182 for (unsigned i = 0, e = CT.size(); i != e; ++i)
1183 outputCompactionTablePlane(i, CT[i], 0);
1187 void BytecodeWriter::outputSymbolTable(const SymbolTable &MST) {
1188 // Do not output the Bytecode block for an empty symbol table, it just wastes
1190 if (MST.isEmpty()) return;
1192 BytecodeBlock SymTabBlock(BytecodeFormat::SymbolTableBlockID, *this,
1193 true/*ElideIfEmpty*/);
1195 // Write the number of types
1196 output_vbr(MST.num_types());
1198 // Write each of the types
1199 for (SymbolTable::type_const_iterator TI = MST.type_begin(),
1200 TE = MST.type_end(); TI != TE; ++TI) {
1201 // Symtab entry:[def slot #][name]
1202 output_typeid((unsigned)Table.getSlot(TI->second));
1206 // Now do each of the type planes in order.
1207 for (SymbolTable::plane_const_iterator PI = MST.plane_begin(),
1208 PE = MST.plane_end(); PI != PE; ++PI) {
1209 SymbolTable::value_const_iterator I = MST.value_begin(PI->first);
1210 SymbolTable::value_const_iterator End = MST.value_end(PI->first);
1213 if (I == End) continue; // Don't mess with an absent type...
1215 // Write the number of values in this plane
1216 output_vbr((unsigned)PI->second.size());
1218 // Write the slot number of the type for this plane
1219 Slot = Table.getSlot(PI->first);
1220 assert(Slot != -1 && "Type in symtab, but not in table!");
1221 output_typeid((unsigned)Slot);
1223 // Write each of the values in this plane
1224 for (; I != End; ++I) {
1225 // Symtab entry: [def slot #][name]
1226 Slot = Table.getSlot(I->second);
1227 assert(Slot != -1 && "Value in symtab but has no slot number!!");
1228 output_vbr((unsigned)Slot);
1234 void llvm::WriteBytecodeToFile(const Module *M, std::ostream &Out,
1236 assert(M && "You can't write a null module!!");
1238 // Make sure that std::cout is put into binary mode for systems
1240 if (&Out == std::cout)
1241 sys::Program::ChangeStdoutToBinary();
1243 // Create a vector of unsigned char for the bytecode output. We
1244 // reserve 256KBytes of space in the vector so that we avoid doing
1245 // lots of little allocations. 256KBytes is sufficient for a large
1246 // proportion of the bytecode files we will encounter. Larger files
1247 // will be automatically doubled in size as needed (std::vector
1249 std::vector<unsigned char> Buffer;
1250 Buffer.reserve(256 * 1024);
1252 // The BytecodeWriter populates Buffer for us.
1253 BytecodeWriter BCW(Buffer, M);
1255 // Keep track of how much we've written
1256 BytesWritten += Buffer.size();
1258 // Determine start and end points of the Buffer
1259 const unsigned char *FirstByte = &Buffer.front();
1261 // If we're supposed to compress this mess ...
1264 // We signal compression by using an alternate magic number for the
1265 // file. The compressed bytecode file's magic number is "llvc" instead
1267 char compressed_magic[4];
1268 compressed_magic[0] = 'l';
1269 compressed_magic[1] = 'l';
1270 compressed_magic[2] = 'v';
1271 compressed_magic[3] = 'c';
1273 Out.write(compressed_magic,4);
1275 // Compress everything after the magic number (which we altered)
1276 uint64_t zipSize = Compressor::compressToStream(
1277 (char*)(FirstByte+4), // Skip the magic number
1278 Buffer.size()-4, // Skip the magic number
1279 Out // Where to write compressed data
1284 // We're not compressing, so just write the entire block.
1285 Out.write((char*)FirstByte, Buffer.size());
1288 // make sure it hits disk now