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 #define DEBUG_TYPE "bcwriter"
21 #include "WriterInternals.h"
22 #include "llvm/Bytecode/WriteBytecodePass.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/ParameterAttributes.h"
27 #include "llvm/InlineAsm.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Module.h"
30 #include "llvm/TypeSymbolTable.h"
31 #include "llvm/ValueSymbolTable.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/Compressor.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/Streams.h"
36 #include "llvm/System/Program.h"
37 #include "llvm/ADT/SmallVector.h"
38 #include "llvm/ADT/STLExtras.h"
39 #include "llvm/ADT/Statistic.h"
44 /// This value needs to be incremented every time the bytecode format changes
45 /// so that the reader can distinguish which format of the bytecode file has
47 /// @brief The bytecode version number
48 const unsigned BCVersionNum = 7;
50 static RegisterPass<WriteBytecodePass> X("emitbytecode", "Bytecode Writer");
52 STATISTIC(BytesWritten, "Number of bytecode bytes written");
54 //===----------------------------------------------------------------------===//
55 //=== Output Primitives ===//
56 //===----------------------------------------------------------------------===//
58 // output - If a position is specified, it must be in the valid portion of the
59 // string... note that this should be inlined always so only the relevant IF
60 // body should be included.
61 inline void BytecodeWriter::output(unsigned i, int pos) {
62 if (pos == -1) { // Be endian clean, little endian is our friend
63 Out.push_back((unsigned char)i);
64 Out.push_back((unsigned char)(i >> 8));
65 Out.push_back((unsigned char)(i >> 16));
66 Out.push_back((unsigned char)(i >> 24));
68 Out[pos ] = (unsigned char)i;
69 Out[pos+1] = (unsigned char)(i >> 8);
70 Out[pos+2] = (unsigned char)(i >> 16);
71 Out[pos+3] = (unsigned char)(i >> 24);
75 inline void BytecodeWriter::output(int32_t i) {
79 /// output_vbr - Output an unsigned value, by using the least number of bytes
80 /// possible. This is useful because many of our "infinite" values are really
81 /// very small most of the time; but can be large a few times.
82 /// Data format used: If you read a byte with the high bit set, use the low
83 /// seven bits as data and then read another byte.
84 inline void BytecodeWriter::output_vbr(uint64_t i) {
86 if (i < 0x80) { // done?
87 Out.push_back((unsigned char)i); // We know the high bit is clear...
91 // Nope, we are bigger than a character, output the next 7 bits and set the
92 // high bit to say that there is more coming...
93 Out.push_back(0x80 | ((unsigned char)i & 0x7F));
94 i >>= 7; // Shift out 7 bits now...
98 inline void BytecodeWriter::output_vbr(uint32_t i) {
100 if (i < 0x80) { // done?
101 Out.push_back((unsigned char)i); // We know the high bit is clear...
105 // Nope, we are bigger than a character, output the next 7 bits and set the
106 // high bit to say that there is more coming...
107 Out.push_back(0x80 | ((unsigned char)i & 0x7F));
108 i >>= 7; // Shift out 7 bits now...
112 inline void BytecodeWriter::output_typeid(unsigned i) {
116 this->output_vbr(0x00FFFFFF);
121 inline void BytecodeWriter::output_vbr(int64_t i) {
123 output_vbr(((uint64_t)(-i) << 1) | 1); // Set low order sign bit...
125 output_vbr((uint64_t)i << 1); // Low order bit is clear.
129 inline void BytecodeWriter::output_vbr(int i) {
131 output_vbr(((unsigned)(-i) << 1) | 1); // Set low order sign bit...
133 output_vbr((unsigned)i << 1); // Low order bit is clear.
136 inline void BytecodeWriter::output_str(const char *Str, unsigned Len) {
137 output_vbr(Len); // Strings may have an arbitrary length.
138 Out.insert(Out.end(), Str, Str+Len);
141 inline void BytecodeWriter::output_data(const void *Ptr, const void *End) {
142 Out.insert(Out.end(), (const unsigned char*)Ptr, (const unsigned char*)End);
145 inline void BytecodeWriter::output_float(float& FloatVal) {
146 /// FIXME: This isn't optimal, it has size problems on some platforms
147 /// where FP is not IEEE.
148 uint32_t i = FloatToBits(FloatVal);
149 Out.push_back( static_cast<unsigned char>( (i ) & 0xFF));
150 Out.push_back( static_cast<unsigned char>( (i >> 8 ) & 0xFF));
151 Out.push_back( static_cast<unsigned char>( (i >> 16) & 0xFF));
152 Out.push_back( static_cast<unsigned char>( (i >> 24) & 0xFF));
155 inline void BytecodeWriter::output_double(double& DoubleVal) {
156 /// FIXME: This isn't optimal, it has size problems on some platforms
157 /// where FP is not IEEE.
158 uint64_t i = DoubleToBits(DoubleVal);
159 Out.push_back( static_cast<unsigned char>( (i ) & 0xFF));
160 Out.push_back( static_cast<unsigned char>( (i >> 8 ) & 0xFF));
161 Out.push_back( static_cast<unsigned char>( (i >> 16) & 0xFF));
162 Out.push_back( static_cast<unsigned char>( (i >> 24) & 0xFF));
163 Out.push_back( static_cast<unsigned char>( (i >> 32) & 0xFF));
164 Out.push_back( static_cast<unsigned char>( (i >> 40) & 0xFF));
165 Out.push_back( static_cast<unsigned char>( (i >> 48) & 0xFF));
166 Out.push_back( static_cast<unsigned char>( (i >> 56) & 0xFF));
169 inline BytecodeBlock::BytecodeBlock(unsigned ID, BytecodeWriter &w,
170 bool elideIfEmpty, bool hasLongFormat)
171 : Id(ID), Writer(w), ElideIfEmpty(elideIfEmpty), HasLongFormat(hasLongFormat){
175 w.output(0U); // For length in long format
177 w.output(0U); /// Place holder for ID and length for this block
182 inline BytecodeBlock::~BytecodeBlock() { // Do backpatch when block goes out
184 if (Loc == Writer.size() && ElideIfEmpty) {
185 // If the block is empty, and we are allowed to, do not emit the block at
187 Writer.resize(Writer.size()-(HasLongFormat?8:4));
192 Writer.output(unsigned(Writer.size()-Loc), int(Loc-4));
194 Writer.output(unsigned(Writer.size()-Loc) << 5 | (Id & 0x1F), int(Loc-4));
197 //===----------------------------------------------------------------------===//
198 //=== Constant Output ===//
199 //===----------------------------------------------------------------------===//
201 void BytecodeWriter::outputParamAttrsList(const ParamAttrsList *Attrs) {
203 output_vbr(unsigned(0));
206 unsigned numAttrs = Attrs->size();
207 output_vbr(numAttrs);
208 for (unsigned i = 0; i < numAttrs; ++i) {
209 uint16_t index = Attrs->getParamIndex(i);
210 uint16_t attrs = Attrs->getParamAttrs(index);
211 output_vbr(uint32_t(index));
212 output_vbr(uint32_t(attrs));
216 void BytecodeWriter::outputType(const Type *T) {
217 const StructType* STy = dyn_cast<StructType>(T);
218 if(STy && STy->isPacked())
219 output_vbr((unsigned)Type::PackedStructTyID);
221 output_vbr((unsigned)T->getTypeID());
223 // That's all there is to handling primitive types...
224 if (T->isPrimitiveType())
225 return; // We might do this if we alias a prim type: %x = type int
227 switch (T->getTypeID()) { // Handle derived types now.
228 case Type::IntegerTyID:
229 output_vbr(cast<IntegerType>(T)->getBitWidth());
231 case Type::FunctionTyID: {
232 const FunctionType *FT = cast<FunctionType>(T);
233 output_typeid(Table.getTypeSlot(FT->getReturnType()));
235 // Output the number of arguments to function (+1 if varargs):
236 output_vbr((unsigned)FT->getNumParams()+FT->isVarArg());
238 // Output all of the arguments...
239 FunctionType::param_iterator I = FT->param_begin();
240 for (; I != FT->param_end(); ++I)
241 output_typeid(Table.getTypeSlot(*I));
243 // Terminate list with VoidTy if we are a varargs function...
245 output_typeid((unsigned)Type::VoidTyID);
247 // Put out all the parameter attributes
248 outputParamAttrsList(FT->getParamAttrs());
252 case Type::ArrayTyID: {
253 const ArrayType *AT = cast<ArrayType>(T);
254 output_typeid(Table.getTypeSlot(AT->getElementType()));
255 output_vbr(AT->getNumElements());
259 case Type::VectorTyID: {
260 const VectorType *PT = cast<VectorType>(T);
261 output_typeid(Table.getTypeSlot(PT->getElementType()));
262 output_vbr(PT->getNumElements());
266 case Type::StructTyID: {
267 const StructType *ST = cast<StructType>(T);
268 // Output all of the element types...
269 for (StructType::element_iterator I = ST->element_begin(),
270 E = ST->element_end(); I != E; ++I) {
271 output_typeid(Table.getTypeSlot(*I));
274 // Terminate list with VoidTy
275 output_typeid((unsigned)Type::VoidTyID);
279 case Type::PointerTyID:
280 output_typeid(Table.getTypeSlot(cast<PointerType>(T)->getElementType()));
283 case Type::OpaqueTyID:
284 // No need to emit anything, just the count of opaque types is enough.
288 cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
289 << " Type '" << T->getDescription() << "'\n";
294 void BytecodeWriter::outputConstant(const Constant *CPV) {
295 assert(((CPV->getType()->isPrimitiveType() || CPV->getType()->isInteger()) ||
296 !CPV->isNullValue()) && "Shouldn't output null constants!");
298 // We must check for a ConstantExpr before switching by type because
299 // a ConstantExpr can be of any type, and has no explicit value.
301 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
302 // FIXME: Encoding of constant exprs could be much more compact!
303 assert(CE->getNumOperands() > 0 && "ConstantExpr with 0 operands");
304 assert(CE->getNumOperands() != 1 || CE->isCast());
305 output_vbr(1+CE->getNumOperands()); // flags as an expr
306 output_vbr(CE->getOpcode()); // Put out the CE op code
308 for (User::const_op_iterator OI = CE->op_begin(); OI != CE->op_end(); ++OI){
309 output_vbr(Table.getSlot(*OI));
310 output_typeid(Table.getTypeSlot((*OI)->getType()));
313 output_vbr((unsigned)CE->getPredicate());
315 } else if (isa<UndefValue>(CPV)) {
316 output_vbr(1U); // 1 -> UndefValue constant.
319 output_vbr(0U); // flag as not a ConstantExpr (i.e. 0 operands)
322 switch (CPV->getType()->getTypeID()) {
323 case Type::IntegerTyID: { // Integer types...
324 const ConstantInt *CI = cast<ConstantInt>(CPV);
325 unsigned NumBits = cast<IntegerType>(CPV->getType())->getBitWidth();
327 output_vbr(uint32_t(CI->getZExtValue()));
328 else if (NumBits <= 64)
329 output_vbr(uint64_t(CI->getZExtValue()));
331 // We have an arbitrary precision integer value to write whose
332 // bit width is > 64. However, in canonical unsigned integer
333 // format it is likely that the high bits are going to be zero.
334 // So, we only write the number of active words.
335 uint32_t activeWords = CI->getValue().getActiveWords();
336 const uint64_t *rawData = CI->getValue().getRawData();
337 output_vbr(activeWords);
338 for (uint32_t i = 0; i < activeWords; ++i)
339 output_vbr(rawData[i]);
344 case Type::ArrayTyID: {
345 const ConstantArray *CPA = cast<ConstantArray>(CPV);
346 assert(!CPA->isString() && "Constant strings should be handled specially!");
348 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i)
349 output_vbr(Table.getSlot(CPA->getOperand(i)));
353 case Type::VectorTyID: {
354 const ConstantVector *CP = cast<ConstantVector>(CPV);
355 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
356 output_vbr(Table.getSlot(CP->getOperand(i)));
360 case Type::StructTyID: {
361 const ConstantStruct *CPS = cast<ConstantStruct>(CPV);
363 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i)
364 output_vbr(Table.getSlot(CPS->getOperand(i)));
368 case Type::PointerTyID:
369 assert(0 && "No non-null, non-constant-expr constants allowed!");
372 case Type::FloatTyID: { // Floating point types...
373 float Tmp = (float)cast<ConstantFP>(CPV)->getValue();
377 case Type::DoubleTyID: {
378 double Tmp = cast<ConstantFP>(CPV)->getValue();
384 case Type::LabelTyID:
386 cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
387 << " type '" << *CPV->getType() << "'\n";
393 /// outputInlineAsm - InlineAsm's get emitted to the constant pool, so they can
394 /// be shared by multiple uses.
395 void BytecodeWriter::outputInlineAsm(const InlineAsm *IA) {
396 // Output a marker, so we know when we have one one parsing the constant pool.
397 // Note that this encoding is 5 bytes: not very efficient for a marker. Since
398 // unique inline asms are rare, this should hardly matter.
401 output(IA->getAsmString());
402 output(IA->getConstraintString());
403 output_vbr(unsigned(IA->hasSideEffects()));
406 void BytecodeWriter::outputConstantStrings() {
407 SlotCalculator::string_iterator I = Table.string_begin();
408 SlotCalculator::string_iterator E = Table.string_end();
409 if (I == E) return; // No strings to emit
411 // If we have != 0 strings to emit, output them now. Strings are emitted into
412 // the 'void' type plane.
413 output_vbr(unsigned(E-I));
414 output_typeid(Type::VoidTyID);
416 // Emit all of the strings.
417 for (I = Table.string_begin(); I != E; ++I) {
418 const ConstantArray *Str = *I;
419 output_typeid(Table.getTypeSlot(Str->getType()));
421 // Now that we emitted the type (which indicates the size of the string),
422 // emit all of the characters.
423 std::string Val = Str->getAsString();
424 output_data(Val.c_str(), Val.c_str()+Val.size());
428 //===----------------------------------------------------------------------===//
429 //=== Instruction Output ===//
430 //===----------------------------------------------------------------------===//
432 // outputInstructionFormat0 - Output those weird instructions that have a large
433 // number of operands or have large operands themselves.
435 // Format: [opcode] [type] [numargs] [arg0] [arg1] ... [arg<numargs-1>]
437 void BytecodeWriter::outputInstructionFormat0(const Instruction *I,
439 const SlotCalculator &Table,
441 // Opcode must have top two bits clear...
442 output_vbr(Opcode << 2); // Instruction Opcode ID
443 output_typeid(Type); // Result type
445 unsigned NumArgs = I->getNumOperands();
446 bool HasExtraArg = false;
447 if (isa<CastInst>(I) || isa<InvokeInst>(I) ||
448 isa<CmpInst>(I) || isa<VAArgInst>(I) || Opcode == 58)
450 if (const AllocationInst *AI = dyn_cast<AllocationInst>(I))
451 HasExtraArg = AI->getAlignment() != 0;
453 output_vbr(NumArgs + HasExtraArg);
455 if (!isa<GetElementPtrInst>(&I)) {
456 for (unsigned i = 0; i < NumArgs; ++i)
457 output_vbr(Table.getSlot(I->getOperand(i)));
459 if (isa<CastInst>(I) || isa<VAArgInst>(I)) {
460 output_typeid(Table.getTypeSlot(I->getType()));
461 } else if (isa<CmpInst>(I)) {
462 output_vbr(unsigned(cast<CmpInst>(I)->getPredicate()));
463 } else if (isa<InvokeInst>(I)) {
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()));
468 } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(I)) {
469 if (AI->getAlignment())
470 output_vbr((unsigned)Log2_32(AI->getAlignment())+1);
473 output_vbr(Table.getSlot(I->getOperand(0)));
475 // We need to encode the type of sequential type indices into their slot #
477 for (gep_type_iterator TI = gep_type_begin(I), E = gep_type_end(I);
478 Idx != NumArgs; ++TI, ++Idx) {
479 unsigned Slot = Table.getSlot(I->getOperand(Idx));
481 if (isa<SequentialType>(*TI)) {
482 // These should be either 32-bits or 64-bits, however, with bit
483 // accurate types we just distinguish between less than or equal to
484 // 32-bits or greater than 32-bits.
486 cast<IntegerType>(I->getOperand(Idx)->getType())->getBitWidth();
487 assert(BitWidth == 32 || BitWidth == 64 &&
488 "Invalid bitwidth for GEP index");
489 unsigned IdxId = BitWidth == 32 ? 0 : 1;
490 Slot = (Slot << 1) | IdxId;
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 == 58 || isa<InvokeInst>(I)));
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 output_vbr(Table.getSlot(I->getOperand(i)));
539 for (unsigned i = NumFixedOperands, e = I->getNumOperands(); i != e; ++i) {
540 // Output Arg Type ID
541 output_typeid(Table.getTypeSlot(I->getOperand(i)->getType()));
543 // Output arg ID itself
544 output_vbr(Table.getSlot(I->getOperand(i)));
547 if (isa<InvokeInst>(I)) {
548 // Emit the tail call/calling conv for invoke instructions
549 output_vbr(cast<InvokeInst>(I)->getCallingConv());
550 } else if (Opcode == 58) {
551 const CallInst *CI = cast<CallInst>(I);
552 output_vbr((CI->getCallingConv() << 1) | unsigned(CI->isTailCall()));
557 // outputInstructionFormat1 - Output one operand instructions, knowing that no
558 // operand index is >= 2^12.
560 inline void BytecodeWriter::outputInstructionFormat1(const Instruction *I,
564 // bits Instruction format:
565 // --------------------------
566 // 01-00: Opcode type, fixed to 1.
568 // 19-08: Resulting type plane
569 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
571 output(1 | (Opcode << 2) | (Type << 8) | (Slots[0] << 20));
575 // outputInstructionFormat2 - Output two operand instructions, knowing that no
576 // operand index is >= 2^8.
578 inline void BytecodeWriter::outputInstructionFormat2(const Instruction *I,
582 // bits Instruction format:
583 // --------------------------
584 // 01-00: Opcode type, fixed to 2.
586 // 15-08: Resulting type plane
590 output(2 | (Opcode << 2) | (Type << 8) | (Slots[0] << 16) | (Slots[1] << 24));
594 // outputInstructionFormat3 - Output three operand instructions, knowing that no
595 // operand index is >= 2^6.
597 inline void BytecodeWriter::outputInstructionFormat3(const Instruction *I,
601 // bits Instruction format:
602 // --------------------------
603 // 01-00: Opcode type, fixed to 3.
605 // 13-08: Resulting type plane
610 output(3 | (Opcode << 2) | (Type << 8) |
611 (Slots[0] << 14) | (Slots[1] << 20) | (Slots[2] << 26));
614 void BytecodeWriter::outputInstruction(const Instruction &I) {
615 assert(I.getOpcode() < 57 && "Opcode too big???");
616 unsigned Opcode = I.getOpcode();
617 unsigned NumOperands = I.getNumOperands();
619 // Encode 'tail call' as 61
621 if (const CallInst *CI = dyn_cast<CallInst>(&I)) {
622 if (CI->getCallingConv() == CallingConv::C) {
623 if (CI->isTailCall())
624 Opcode = 61; // CCC + Tail Call
626 ; // Opcode = Instruction::Call
627 } else if (CI->getCallingConv() == CallingConv::Fast) {
628 if (CI->isTailCall())
629 Opcode = 59; // FastCC + TailCall
631 Opcode = 60; // FastCC + Not Tail Call
633 Opcode = 58; // Call escape sequence.
637 // Figure out which type to encode with the instruction. Typically we want
638 // the type of the first parameter, as opposed to the type of the instruction
639 // (for example, with setcc, we always know it returns bool, but the type of
640 // the first param is actually interesting). But if we have no arguments
641 // we take the type of the instruction itself.
644 switch (I.getOpcode()) {
645 case Instruction::Select:
646 case Instruction::Malloc:
647 case Instruction::Alloca:
648 Ty = I.getType(); // These ALWAYS want to encode the return type
650 case Instruction::Store:
651 Ty = I.getOperand(1)->getType(); // Encode the pointer type...
652 assert(isa<PointerType>(Ty) && "Store to nonpointer type!?!?");
654 default: // Otherwise use the default behavior...
655 Ty = NumOperands ? I.getOperand(0)->getType() : I.getType();
659 unsigned Type = Table.getTypeSlot(Ty);
661 // Varargs calls and invokes are encoded entirely different from any other
663 if (const CallInst *CI = dyn_cast<CallInst>(&I)){
664 const PointerType *Ty =cast<PointerType>(CI->getCalledValue()->getType());
665 if (cast<FunctionType>(Ty->getElementType())->isVarArg()) {
666 outputInstrVarArgsCall(CI, Opcode, Table, Type);
669 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(&I)) {
670 const PointerType *Ty =cast<PointerType>(II->getCalledValue()->getType());
671 if (cast<FunctionType>(Ty->getElementType())->isVarArg()) {
672 outputInstrVarArgsCall(II, Opcode, Table, Type);
677 if (NumOperands <= 3) {
678 // Make sure that we take the type number into consideration. We don't want
679 // to overflow the field size for the instruction format we select.
681 unsigned MaxOpSlot = Type;
682 unsigned Slots[3]; Slots[0] = (1 << 12)-1; // Marker to signify 0 operands
684 for (unsigned i = 0; i != NumOperands; ++i) {
685 unsigned Slot = Table.getSlot(I.getOperand(i));
686 if (Slot > MaxOpSlot) MaxOpSlot = Slot;
690 // Handle the special cases for various instructions...
691 if (isa<CastInst>(I) || isa<VAArgInst>(I)) {
692 // Cast has to encode the destination type as the second argument in the
693 // packet, or else we won't know what type to cast to!
694 Slots[1] = Table.getTypeSlot(I.getType());
695 if (Slots[1] > MaxOpSlot) MaxOpSlot = Slots[1];
697 } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(&I)) {
698 assert(NumOperands == 1 && "Bogus allocation!");
699 if (AI->getAlignment()) {
700 Slots[1] = Log2_32(AI->getAlignment())+1;
701 if (Slots[1] > MaxOpSlot) MaxOpSlot = Slots[1];
704 } else if (isa<ICmpInst>(I) || isa<FCmpInst>(I)) {
705 // We need to encode the compare instruction's predicate as the third
706 // operand. Its not really a slot, but we don't want to break the
707 // instruction format for these instructions.
709 assert(NumOperands == 3 && "CmpInst with wrong number of operands?");
710 Slots[2] = unsigned(cast<CmpInst>(&I)->getPredicate());
711 if (Slots[2] > MaxOpSlot)
712 MaxOpSlot = Slots[2];
713 } else if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
714 // We need to encode the type of sequential type indices into their slot #
716 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
718 if (isa<SequentialType>(*I)) {
719 // These should be either 32-bits or 64-bits, however, with bit
720 // accurate types we just distinguish between less than or equal to
721 // 32-bits or greater than 32-bits.
723 cast<IntegerType>(GEP->getOperand(Idx)->getType())->getBitWidth();
724 assert(BitWidth == 32 || BitWidth == 64 &&
725 "Invalid bitwidth for GEP index");
726 unsigned IdxId = BitWidth == 32 ? 0 : 1;
727 Slots[Idx] = (Slots[Idx] << 1) | IdxId;
728 if (Slots[Idx] > MaxOpSlot) MaxOpSlot = Slots[Idx];
730 } else if (Opcode == 58) {
731 // If this is the escape sequence for call, emit the tailcall/cc info.
732 const CallInst &CI = cast<CallInst>(I);
734 if (NumOperands <= 3) {
735 Slots[NumOperands-1] =
736 (CI.getCallingConv() << 1)|unsigned(CI.isTailCall());
737 if (Slots[NumOperands-1] > MaxOpSlot)
738 MaxOpSlot = Slots[NumOperands-1];
740 } else if (isa<InvokeInst>(I)) {
741 // Invoke escape seq has at least 4 operands to encode.
743 } else if (const LoadInst *LI = dyn_cast<LoadInst>(&I)) {
744 // Encode attributed load as opcode 62
745 // We need to encode the attributes of the load instruction as the second
746 // operand. Its not really a slot, but we don't want to break the
747 // instruction format for these instructions.
748 if (LI->getAlignment() || LI->isVolatile()) {
750 Slots[1] = ((Log2_32(LI->getAlignment())+1)<<1) +
751 (LI->isVolatile() ? 1 : 0);
752 if (Slots[1] > MaxOpSlot)
753 MaxOpSlot = Slots[1];
756 } else if (const StoreInst *SI = dyn_cast<StoreInst>(&I)) {
757 // Encode attributed store as opcode 63
758 // We need to encode the attributes of the store instruction as the third
759 // operand. Its not really a slot, but we don't want to break the
760 // instruction format for these instructions.
761 if (SI->getAlignment() || SI->isVolatile()) {
763 Slots[2] = ((Log2_32(SI->getAlignment())+1)<<1) +
764 (SI->isVolatile() ? 1 : 0);
765 if (Slots[2] > MaxOpSlot)
766 MaxOpSlot = Slots[2];
771 // Decide which instruction encoding to use. This is determined primarily
772 // by the number of operands, and secondarily by whether or not the max
773 // operand will fit into the instruction encoding. More operands == fewer
776 switch (NumOperands) {
779 if (MaxOpSlot < (1 << 12)-1) { // -1 because we use 4095 to indicate 0 ops
780 outputInstructionFormat1(&I, Opcode, Slots, Type);
786 if (MaxOpSlot < (1 << 8)) {
787 outputInstructionFormat2(&I, Opcode, Slots, Type);
793 if (MaxOpSlot < (1 << 6)) {
794 outputInstructionFormat3(&I, Opcode, Slots, Type);
803 // If we weren't handled before here, we either have a large number of
804 // operands or a large operand index that we are referring to.
805 outputInstructionFormat0(&I, Opcode, Table, Type);
808 //===----------------------------------------------------------------------===//
809 //=== Block Output ===//
810 //===----------------------------------------------------------------------===//
812 BytecodeWriter::BytecodeWriter(std::vector<unsigned char> &o, const Module *M)
815 // Emit the signature...
816 static const unsigned char *Sig = (const unsigned char*)"llvm";
817 output_data(Sig, Sig+4);
819 // Emit the top level CLASS block.
820 BytecodeBlock ModuleBlock(BytecodeFormat::ModuleBlockID, *this, false, true);
822 // Output the version identifier
823 output_vbr(BCVersionNum);
825 // The Global type plane comes first
827 BytecodeBlock CPool(BytecodeFormat::GlobalTypePlaneBlockID, *this);
828 outputTypes(Type::FirstDerivedTyID);
831 // The ModuleInfoBlock follows directly after the type information
832 outputModuleInfoBlock(M);
834 // Output module level constants, used for global variable initializers
837 // Do the whole module now! Process each function at a time...
838 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I)
841 // Output the symbole table for types
842 outputTypeSymbolTable(M->getTypeSymbolTable());
844 // Output the symbol table for values
845 outputValueSymbolTable(M->getValueSymbolTable());
848 void BytecodeWriter::outputTypes(unsigned TypeNum) {
849 // Write the type plane for types first because earlier planes (e.g. for a
850 // primitive type like float) may have constants constructed using types
851 // coming later (e.g., via getelementptr from a pointer type). The type
852 // plane is needed before types can be fwd or bkwd referenced.
853 const std::vector<const Type*>& Types = Table.getTypes();
854 assert(!Types.empty() && "No types at all?");
855 assert(TypeNum <= Types.size() && "Invalid TypeNo index");
857 unsigned NumEntries = Types.size() - TypeNum;
859 // Output type header: [num entries]
860 output_vbr(NumEntries);
862 for (unsigned i = TypeNum; i < TypeNum+NumEntries; ++i)
863 outputType(Types[i]);
866 // Helper function for outputConstants().
867 // Writes out all the constants in the plane Plane starting at entry StartNo.
869 void BytecodeWriter::outputConstantsInPlane(const Value *const *Plane,
872 unsigned ValNo = StartNo;
874 // Scan through and ignore function arguments, global values, and constant
876 for (; ValNo < PlaneSize &&
877 (isa<Argument>(Plane[ValNo]) || isa<GlobalValue>(Plane[ValNo]) ||
878 (isa<ConstantArray>(Plane[ValNo]) &&
879 cast<ConstantArray>(Plane[ValNo])->isString())); ValNo++)
882 unsigned NC = ValNo; // Number of constants
883 for (; NC < PlaneSize && (isa<Constant>(Plane[NC]) ||
884 isa<InlineAsm>(Plane[NC])); NC++)
886 NC -= ValNo; // Convert from index into count
887 if (NC == 0) return; // Skip empty type planes...
889 // FIXME: Most slabs only have 1 or 2 entries! We should encode this much
892 // Put out type header: [num entries][type id number]
896 // Put out the Type ID Number.
897 output_typeid(Table.getTypeSlot(Plane[0]->getType()));
899 for (unsigned i = ValNo; i < ValNo+NC; ++i) {
900 const Value *V = Plane[i];
901 if (const Constant *C = dyn_cast<Constant>(V))
904 outputInlineAsm(cast<InlineAsm>(V));
908 static inline bool hasNullValue(const Type *Ty) {
909 return Ty != Type::LabelTy && Ty != Type::VoidTy && !isa<OpaqueType>(Ty);
912 void BytecodeWriter::outputConstants() {
913 BytecodeBlock CPool(BytecodeFormat::ConstantPoolBlockID, *this,
914 true /* Elide block if empty */);
916 unsigned NumPlanes = Table.getNumPlanes();
918 // Output module-level string constants before any other constants.
919 outputConstantStrings();
921 for (unsigned pno = 0; pno != NumPlanes; pno++) {
922 const SlotCalculator::TypePlane &Plane = Table.getPlane(pno);
923 if (!Plane.empty()) { // Skip empty type planes...
925 if (hasNullValue(Plane[0]->getType())) {
926 // Skip zero initializer
930 // Write out constants in the plane
931 outputConstantsInPlane(&Plane[0], Plane.size(), ValNo);
936 static unsigned getEncodedLinkage(const GlobalValue *GV) {
937 switch (GV->getLinkage()) {
938 default: assert(0 && "Invalid linkage!");
939 case GlobalValue::ExternalLinkage: return 0;
940 case GlobalValue::WeakLinkage: return 1;
941 case GlobalValue::AppendingLinkage: return 2;
942 case GlobalValue::InternalLinkage: return 3;
943 case GlobalValue::LinkOnceLinkage: return 4;
944 case GlobalValue::DLLImportLinkage: return 5;
945 case GlobalValue::DLLExportLinkage: return 6;
946 case GlobalValue::ExternalWeakLinkage: return 7;
950 static unsigned getEncodedVisibility(const GlobalValue *GV) {
951 switch (GV->getVisibility()) {
952 default: assert(0 && "Invalid visibility!");
953 case GlobalValue::DefaultVisibility: return 0;
954 case GlobalValue::HiddenVisibility: return 1;
958 void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
959 BytecodeBlock ModuleInfoBlock(BytecodeFormat::ModuleGlobalInfoBlockID, *this);
961 // Give numbers to sections as we encounter them.
962 unsigned SectionIDCounter = 0;
963 std::vector<std::string> SectionNames;
964 std::map<std::string, unsigned> SectionID;
966 // Output the types for the global variables in the module...
967 for (Module::const_global_iterator I = M->global_begin(),
968 End = M->global_end(); I != End; ++I) {
969 unsigned Slot = Table.getTypeSlot(I->getType());
971 assert((I->hasInitializer() || !I->hasInternalLinkage()) &&
972 "Global must have an initializer or have external linkage!");
974 // Fields: bit0 = isConstant, bit1 = hasInitializer, bit2-4=Linkage,
975 // bit5 = isThreadLocal, bit6+ = Slot # for type.
976 bool HasExtensionWord = (I->getAlignment() != 0) ||
978 (I->getVisibility() != GlobalValue::DefaultVisibility);
980 // If we need to use the extension byte, set linkage=3(internal) and
981 // initializer = 0 (impossible!).
982 if (!HasExtensionWord) {
983 unsigned oSlot = (Slot << 6)| (((unsigned)I->isThreadLocal()) << 5) |
984 (getEncodedLinkage(I) << 2) | (I->hasInitializer() << 1)
985 | (unsigned)I->isConstant();
988 unsigned oSlot = (Slot << 6) | (((unsigned)I->isThreadLocal()) << 5) |
989 (3 << 2) | (0 << 1) | (unsigned)I->isConstant();
992 // The extension word has this format: bit 0 = has initializer, bit 1-3 =
993 // linkage, bit 4-8 = alignment (log2), bit 9 = has SectionID,
994 // bits 10-12 = visibility, bits 13+ = future use.
995 unsigned ExtWord = (unsigned)I->hasInitializer() |
996 (getEncodedLinkage(I) << 1) |
997 ((Log2_32(I->getAlignment())+1) << 4) |
998 ((unsigned)I->hasSection() << 9) |
999 (getEncodedVisibility(I) << 10);
1000 output_vbr(ExtWord);
1001 if (I->hasSection()) {
1002 // Give section names unique ID's.
1003 unsigned &Entry = SectionID[I->getSection()];
1005 Entry = ++SectionIDCounter;
1006 SectionNames.push_back(I->getSection());
1012 // If we have an initializer, output it now.
1013 if (I->hasInitializer())
1014 output_vbr(Table.getSlot((Value*)I->getInitializer()));
1016 output_typeid(Table.getTypeSlot(Type::VoidTy));
1018 // Output the types of the functions in this module.
1019 for (Module::const_iterator I = M->begin(), End = M->end(); I != End; ++I) {
1020 unsigned Slot = Table.getTypeSlot(I->getType());
1021 assert(((Slot << 6) >> 6) == Slot && "Slot # too big!");
1022 unsigned CC = I->getCallingConv()+1;
1023 unsigned ID = (Slot << 5) | (CC & 15);
1025 if (I->isDeclaration()) // If external, we don't have an FunctionInfo block.
1028 if (I->getAlignment() || I->hasSection() || (CC & ~15) != 0 ||
1029 (I->isDeclaration() && I->hasDLLImportLinkage()) ||
1030 (I->isDeclaration() && I->hasExternalWeakLinkage())
1032 ID |= 1 << 31; // Do we need an extension word?
1036 if (ID & (1 << 31)) {
1037 // Extension byte: bits 0-4 = alignment, bits 5-9 = top nibble of calling
1038 // convention, bit 10 = hasSectionID., bits 11-12 = external linkage type
1039 unsigned extLinkage = 0;
1041 if (I->isDeclaration()) {
1042 if (I->hasDLLImportLinkage()) {
1044 } else if (I->hasExternalWeakLinkage()) {
1049 ID = (Log2_32(I->getAlignment())+1) | ((CC >> 4) << 5) |
1050 (I->hasSection() << 10) |
1051 ((extLinkage & 3) << 11);
1054 // Give section names unique ID's.
1055 if (I->hasSection()) {
1056 unsigned &Entry = SectionID[I->getSection()];
1058 Entry = ++SectionIDCounter;
1059 SectionNames.push_back(I->getSection());
1065 output_vbr(Table.getTypeSlot(Type::VoidTy) << 5);
1067 // Emit the list of dependent libraries for the Module.
1068 Module::lib_iterator LI = M->lib_begin();
1069 Module::lib_iterator LE = M->lib_end();
1070 output_vbr(unsigned(LE - LI)); // Emit the number of dependent libraries.
1071 for (; LI != LE; ++LI)
1074 // Output the target triple from the module
1075 output(M->getTargetTriple());
1077 // Output the data layout from the module
1078 output(M->getDataLayout());
1080 // Emit the table of section names.
1081 output_vbr((unsigned)SectionNames.size());
1082 for (unsigned i = 0, e = SectionNames.size(); i != e; ++i)
1083 output(SectionNames[i]);
1085 // Output the inline asm string.
1086 output(M->getModuleInlineAsm());
1089 void BytecodeWriter::outputInstructions(const Function *F) {
1090 BytecodeBlock ILBlock(BytecodeFormat::InstructionListBlockID, *this);
1091 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1092 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
1093 outputInstruction(*I);
1096 void BytecodeWriter::outputFunction(const Function *F) {
1097 // If this is an external function, there is nothing else to emit!
1098 if (F->isDeclaration()) return;
1100 BytecodeBlock FunctionBlock(BytecodeFormat::FunctionBlockID, *this);
1101 unsigned rWord = (getEncodedVisibility(F) << 16) | getEncodedLinkage(F);
1104 // Get slot information about the function...
1105 Table.incorporateFunction(F);
1107 // Output all of the instructions in the body of the function
1108 outputInstructions(F);
1110 // If needed, output the symbol table for the function...
1111 outputValueSymbolTable(F->getValueSymbolTable());
1113 Table.purgeFunction();
1117 void BytecodeWriter::outputTypeSymbolTable(const TypeSymbolTable &TST) {
1118 // Do not output the block for an empty symbol table, it just wastes
1120 if (TST.empty()) return;
1122 // Create a header for the symbol table
1123 BytecodeBlock SymTabBlock(BytecodeFormat::TypeSymbolTableBlockID, *this,
1124 true/*ElideIfEmpty*/);
1125 // Write the number of types
1126 output_vbr(TST.size());
1128 // Write each of the types
1129 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end();
1131 // Symtab entry:[def slot #][name]
1132 output_typeid(Table.getTypeSlot(TI->second));
1137 void BytecodeWriter::outputValueSymbolTable(const ValueSymbolTable &VST) {
1138 // Do not output the Bytecode block for an empty symbol table, it just wastes
1140 if (VST.empty()) return;
1142 BytecodeBlock SymTabBlock(BytecodeFormat::ValueSymbolTableBlockID, *this,
1143 true/*ElideIfEmpty*/);
1145 // Organize the symbol table by type
1146 typedef SmallVector<const ValueName*, 8> PlaneMapVector;
1147 typedef DenseMap<const Type*, PlaneMapVector> PlaneMap;
1149 for (ValueSymbolTable::const_iterator SI = VST.begin(), SE = VST.end();
1151 Planes[SI->getValue()->getType()].push_back(&*SI);
1153 for (PlaneMap::iterator PI = Planes.begin(), PE = Planes.end();
1155 PlaneMapVector::const_iterator I = PI->second.begin();
1156 PlaneMapVector::const_iterator End = PI->second.end();
1158 if (I == End) continue; // Don't mess with an absent type...
1160 // Write the number of values in this plane
1161 output_vbr((unsigned)PI->second.size());
1163 // Write the slot number of the type for this plane
1164 output_typeid(Table.getTypeSlot(PI->first));
1166 // Write each of the values in this plane
1167 for (; I != End; ++I) {
1168 // Symtab entry: [def slot #][name]
1169 output_vbr(Table.getSlot((*I)->getValue()));
1170 output_str((*I)->getKeyData(), (*I)->getKeyLength());
1175 void llvm::WriteBytecodeToFile(const Module *M, OStream &Out,
1177 assert(M && "You can't write a null module!!");
1179 // Make sure that std::cout is put into binary mode for systems
1182 sys::Program::ChangeStdoutToBinary();
1184 // Create a vector of unsigned char for the bytecode output. We
1185 // reserve 256KBytes of space in the vector so that we avoid doing
1186 // lots of little allocations. 256KBytes is sufficient for a large
1187 // proportion of the bytecode files we will encounter. Larger files
1188 // will be automatically doubled in size as needed (std::vector
1190 std::vector<unsigned char> Buffer;
1191 Buffer.reserve(256 * 1024);
1193 // The BytecodeWriter populates Buffer for us.
1194 BytecodeWriter BCW(Buffer, M);
1196 // Keep track of how much we've written
1197 BytesWritten += Buffer.size();
1199 // Determine start and end points of the Buffer
1200 const unsigned char *FirstByte = &Buffer.front();
1202 // If we're supposed to compress this mess ...
1205 // We signal compression by using an alternate magic number for the
1206 // file. The compressed bytecode file's magic number is "llvc" instead
1208 char compressed_magic[4];
1209 compressed_magic[0] = 'l';
1210 compressed_magic[1] = 'l';
1211 compressed_magic[2] = 'v';
1212 compressed_magic[3] = 'c';
1214 Out.stream()->write(compressed_magic,4);
1216 // Compress everything after the magic number (which we altered)
1217 Compressor::compressToStream(
1218 (char*)(FirstByte+4), // Skip the magic number
1219 Buffer.size()-4, // Skip the magic number
1220 *Out.stream() // Where to write compressed data
1225 // We're not compressing, so just write the entire block.
1226 Out.stream()->write((char*)FirstByte, Buffer.size());
1229 // make sure it hits disk now
1230 Out.stream()->flush();