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 "bytecodewriter"
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/InlineAsm.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Module.h"
29 #include "llvm/SymbolTable.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/Compressor.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/Streams.h"
34 #include "llvm/System/Program.h"
35 #include "llvm/ADT/STLExtras.h"
36 #include "llvm/ADT/Statistic.h"
41 /// This value needs to be incremented every time the bytecode format changes
42 /// so that the reader can distinguish which format of the bytecode file has
44 /// @brief The bytecode version number
45 const unsigned BCVersionNum = 7;
47 static RegisterPass<WriteBytecodePass> X("emitbytecode", "Bytecode Writer");
49 STATISTIC(BytesWritten, "Number of bytecode bytes written");
51 //===----------------------------------------------------------------------===//
52 //=== Output Primitives ===//
53 //===----------------------------------------------------------------------===//
55 // output - If a position is specified, it must be in the valid portion of the
56 // string... note that this should be inlined always so only the relevant IF
57 // body should be included.
58 inline void BytecodeWriter::output(unsigned i, int pos) {
59 if (pos == -1) { // Be endian clean, little endian is our friend
60 Out.push_back((unsigned char)i);
61 Out.push_back((unsigned char)(i >> 8));
62 Out.push_back((unsigned char)(i >> 16));
63 Out.push_back((unsigned char)(i >> 24));
65 Out[pos ] = (unsigned char)i;
66 Out[pos+1] = (unsigned char)(i >> 8);
67 Out[pos+2] = (unsigned char)(i >> 16);
68 Out[pos+3] = (unsigned char)(i >> 24);
72 inline void BytecodeWriter::output(int i) {
76 /// output_vbr - Output an unsigned value, by using the least number of bytes
77 /// possible. This is useful because many of our "infinite" values are really
78 /// very small most of the time; but can be large a few times.
79 /// Data format used: If you read a byte with the high bit set, use the low
80 /// seven bits as data and then read another byte.
81 inline void BytecodeWriter::output_vbr(uint64_t i) {
83 if (i < 0x80) { // done?
84 Out.push_back((unsigned char)i); // We know the high bit is clear...
88 // Nope, we are bigger than a character, output the next 7 bits and set the
89 // high bit to say that there is more coming...
90 Out.push_back(0x80 | ((unsigned char)i & 0x7F));
91 i >>= 7; // Shift out 7 bits now...
95 inline void BytecodeWriter::output_vbr(unsigned i) {
97 if (i < 0x80) { // done?
98 Out.push_back((unsigned char)i); // We know the high bit is clear...
102 // Nope, we are bigger than a character, output the next 7 bits and set the
103 // high bit to say that there is more coming...
104 Out.push_back(0x80 | ((unsigned char)i & 0x7F));
105 i >>= 7; // Shift out 7 bits now...
109 inline void BytecodeWriter::output_typeid(unsigned i) {
113 this->output_vbr(0x00FFFFFF);
118 inline void BytecodeWriter::output_vbr(int64_t i) {
120 output_vbr(((uint64_t)(-i) << 1) | 1); // Set low order sign bit...
122 output_vbr((uint64_t)i << 1); // Low order bit is clear.
126 inline void BytecodeWriter::output_vbr(int i) {
128 output_vbr(((unsigned)(-i) << 1) | 1); // Set low order sign bit...
130 output_vbr((unsigned)i << 1); // Low order bit is clear.
133 inline void BytecodeWriter::output(const std::string &s) {
134 unsigned Len = s.length();
135 output_vbr(Len); // Strings may have an arbitrary length.
136 Out.insert(Out.end(), s.begin(), s.end());
139 inline void BytecodeWriter::output_data(const void *Ptr, const void *End) {
140 Out.insert(Out.end(), (const unsigned char*)Ptr, (const unsigned char*)End);
143 inline void BytecodeWriter::output_float(float& FloatVal) {
144 /// FIXME: This isn't optimal, it has size problems on some platforms
145 /// where FP is not IEEE.
146 uint32_t i = FloatToBits(FloatVal);
147 Out.push_back( static_cast<unsigned char>( (i ) & 0xFF));
148 Out.push_back( static_cast<unsigned char>( (i >> 8 ) & 0xFF));
149 Out.push_back( static_cast<unsigned char>( (i >> 16) & 0xFF));
150 Out.push_back( static_cast<unsigned char>( (i >> 24) & 0xFF));
153 inline void BytecodeWriter::output_double(double& DoubleVal) {
154 /// FIXME: This isn't optimal, it has size problems on some platforms
155 /// where FP is not IEEE.
156 uint64_t i = DoubleToBits(DoubleVal);
157 Out.push_back( static_cast<unsigned char>( (i ) & 0xFF));
158 Out.push_back( static_cast<unsigned char>( (i >> 8 ) & 0xFF));
159 Out.push_back( static_cast<unsigned char>( (i >> 16) & 0xFF));
160 Out.push_back( static_cast<unsigned char>( (i >> 24) & 0xFF));
161 Out.push_back( static_cast<unsigned char>( (i >> 32) & 0xFF));
162 Out.push_back( static_cast<unsigned char>( (i >> 40) & 0xFF));
163 Out.push_back( static_cast<unsigned char>( (i >> 48) & 0xFF));
164 Out.push_back( static_cast<unsigned char>( (i >> 56) & 0xFF));
167 inline BytecodeBlock::BytecodeBlock(unsigned ID, BytecodeWriter &w,
168 bool elideIfEmpty, bool hasLongFormat)
169 : Id(ID), Writer(w), ElideIfEmpty(elideIfEmpty), HasLongFormat(hasLongFormat){
173 w.output(0U); // For length in long format
175 w.output(0U); /// Place holder for ID and length for this block
180 inline BytecodeBlock::~BytecodeBlock() { // Do backpatch when block goes out
182 if (Loc == Writer.size() && ElideIfEmpty) {
183 // If the block is empty, and we are allowed to, do not emit the block at
185 Writer.resize(Writer.size()-(HasLongFormat?8:4));
190 Writer.output(unsigned(Writer.size()-Loc), int(Loc-4));
192 Writer.output(unsigned(Writer.size()-Loc) << 5 | (Id & 0x1F), int(Loc-4));
195 //===----------------------------------------------------------------------===//
196 //=== Constant Output ===//
197 //===----------------------------------------------------------------------===//
199 void BytecodeWriter::outputType(const Type *T) {
200 const StructType* STy = dyn_cast<StructType>(T);
201 if(STy && STy->isPacked())
202 output_vbr((unsigned)Type::BC_ONLY_PackedStructTyID);
204 output_vbr((unsigned)T->getTypeID());
206 // That's all there is to handling primitive types...
207 if (T->isPrimitiveType()) {
208 return; // We might do this if we alias a prim type: %x = type int
211 switch (T->getTypeID()) { // Handle derived types now.
212 case Type::FunctionTyID: {
213 const FunctionType *MT = cast<FunctionType>(T);
214 int Slot = Table.getSlot(MT->getReturnType());
215 assert(Slot != -1 && "Type used but not available!!");
216 output_typeid((unsigned)Slot);
218 // Output the number of arguments to function (+1 if varargs):
219 output_vbr((unsigned)MT->getNumParams()+MT->isVarArg());
221 // Output all of the arguments...
222 FunctionType::param_iterator I = MT->param_begin();
223 for (; I != MT->param_end(); ++I) {
224 Slot = Table.getSlot(*I);
225 assert(Slot != -1 && "Type used but not available!!");
226 output_typeid((unsigned)Slot);
229 // Terminate list with VoidTy if we are a varargs function...
231 output_typeid((unsigned)Type::VoidTyID);
235 case Type::ArrayTyID: {
236 const ArrayType *AT = cast<ArrayType>(T);
237 int Slot = Table.getSlot(AT->getElementType());
238 assert(Slot != -1 && "Type used but not available!!");
239 output_typeid((unsigned)Slot);
240 output_vbr(AT->getNumElements());
244 case Type::PackedTyID: {
245 const PackedType *PT = cast<PackedType>(T);
246 int Slot = Table.getSlot(PT->getElementType());
247 assert(Slot != -1 && "Type used but not available!!");
248 output_typeid((unsigned)Slot);
249 output_vbr(PT->getNumElements());
253 case Type::StructTyID: {
254 const StructType *ST = cast<StructType>(T);
255 // Output all of the element types...
256 for (StructType::element_iterator I = ST->element_begin(),
257 E = ST->element_end(); I != E; ++I) {
258 int Slot = Table.getSlot(*I);
259 assert(Slot != -1 && "Type used but not available!!");
260 output_typeid((unsigned)Slot);
263 // Terminate list with VoidTy
264 output_typeid((unsigned)Type::VoidTyID);
268 case Type::PointerTyID: {
269 const PointerType *PT = cast<PointerType>(T);
270 int Slot = Table.getSlot(PT->getElementType());
271 assert(Slot != -1 && "Type used but not available!!");
272 output_typeid((unsigned)Slot);
276 case Type::OpaqueTyID:
277 // No need to emit anything, just the count of opaque types is enough.
281 cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
282 << " Type '" << T->getDescription() << "'\n";
287 void BytecodeWriter::outputConstant(const Constant *CPV) {
288 assert((CPV->getType()->isPrimitiveType() || !CPV->isNullValue()) &&
289 "Shouldn't output null constants!");
291 // We must check for a ConstantExpr before switching by type because
292 // a ConstantExpr can be of any type, and has no explicit value.
294 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
295 // FIXME: Encoding of constant exprs could be much more compact!
296 assert(CE->getNumOperands() > 0 && "ConstantExpr with 0 operands");
297 assert(CE->getNumOperands() != 1 || CE->isCast());
298 output_vbr(1+CE->getNumOperands()); // flags as an expr
299 output_vbr(CE->getOpcode()); // Put out the CE op code
301 for (User::const_op_iterator OI = CE->op_begin(); OI != CE->op_end(); ++OI){
302 int Slot = Table.getSlot(*OI);
303 assert(Slot != -1 && "Unknown constant used in ConstantExpr!!");
304 output_vbr((unsigned)Slot);
305 Slot = Table.getSlot((*OI)->getType());
306 output_typeid((unsigned)Slot);
309 output_vbr((unsigned)CE->getPredicate());
311 } else if (isa<UndefValue>(CPV)) {
312 output_vbr(1U); // 1 -> UndefValue constant.
315 output_vbr(0U); // flag as not a ConstantExpr (i.e. 0 operands)
318 switch (CPV->getType()->getTypeID()) {
319 case Type::BoolTyID: // Boolean Types
320 if (cast<ConstantBool>(CPV)->getValue())
326 case Type::UByteTyID: // Unsigned integer types...
327 case Type::UShortTyID:
329 case Type::ULongTyID:
330 output_vbr(cast<ConstantInt>(CPV)->getZExtValue());
333 case Type::SByteTyID: // Signed integer types...
334 case Type::ShortTyID:
337 output_vbr(cast<ConstantInt>(CPV)->getSExtValue());
340 case Type::ArrayTyID: {
341 const ConstantArray *CPA = cast<ConstantArray>(CPV);
342 assert(!CPA->isString() && "Constant strings should be handled specially!");
344 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i) {
345 int Slot = Table.getSlot(CPA->getOperand(i));
346 assert(Slot != -1 && "Constant used but not available!!");
347 output_vbr((unsigned)Slot);
352 case Type::PackedTyID: {
353 const ConstantPacked *CP = cast<ConstantPacked>(CPV);
355 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i) {
356 int Slot = Table.getSlot(CP->getOperand(i));
357 assert(Slot != -1 && "Constant used but not available!!");
358 output_vbr((unsigned)Slot);
363 case Type::StructTyID: {
364 const ConstantStruct *CPS = cast<ConstantStruct>(CPV);
366 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i) {
367 int Slot = Table.getSlot(CPS->getOperand(i));
368 assert(Slot != -1 && "Constant used but not available!!");
369 output_vbr((unsigned)Slot);
374 case Type::PointerTyID:
375 assert(0 && "No non-null, non-constant-expr constants allowed!");
378 case Type::FloatTyID: { // Floating point types...
379 float Tmp = (float)cast<ConstantFP>(CPV)->getValue();
383 case Type::DoubleTyID: {
384 double Tmp = cast<ConstantFP>(CPV)->getValue();
390 case Type::LabelTyID:
392 cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
393 << " type '" << *CPV->getType() << "'\n";
399 /// outputInlineAsm - InlineAsm's get emitted to the constant pool, so they can
400 /// be shared by multiple uses.
401 void BytecodeWriter::outputInlineAsm(const InlineAsm *IA) {
402 // Output a marker, so we know when we have one one parsing the constant pool.
403 // Note that this encoding is 5 bytes: not very efficient for a marker. Since
404 // unique inline asms are rare, this should hardly matter.
407 output(IA->getAsmString());
408 output(IA->getConstraintString());
409 output_vbr(unsigned(IA->hasSideEffects()));
412 void BytecodeWriter::outputConstantStrings() {
413 SlotCalculator::string_iterator I = Table.string_begin();
414 SlotCalculator::string_iterator E = Table.string_end();
415 if (I == E) return; // No strings to emit
417 // If we have != 0 strings to emit, output them now. Strings are emitted into
418 // the 'void' type plane.
419 output_vbr(unsigned(E-I));
420 output_typeid(Type::VoidTyID);
422 // Emit all of the strings.
423 for (I = Table.string_begin(); I != E; ++I) {
424 const ConstantArray *Str = *I;
425 int Slot = Table.getSlot(Str->getType());
426 assert(Slot != -1 && "Constant string of unknown type?");
427 output_typeid((unsigned)Slot);
429 // Now that we emitted the type (which indicates the size of the string),
430 // emit all of the characters.
431 std::string Val = Str->getAsString();
432 output_data(Val.c_str(), Val.c_str()+Val.size());
436 //===----------------------------------------------------------------------===//
437 //=== Instruction Output ===//
438 //===----------------------------------------------------------------------===//
440 // outputInstructionFormat0 - Output those weird instructions that have a large
441 // number of operands or have large operands themselves.
443 // Format: [opcode] [type] [numargs] [arg0] [arg1] ... [arg<numargs-1>]
445 void BytecodeWriter::outputInstructionFormat0(const Instruction *I,
447 const SlotCalculator &Table,
449 // Opcode must have top two bits clear...
450 output_vbr(Opcode << 2); // Instruction Opcode ID
451 output_typeid(Type); // Result type
453 unsigned NumArgs = I->getNumOperands();
454 output_vbr(NumArgs + (isa<CastInst>(I) || isa<InvokeInst>(I) ||
455 isa<CmpInst>(I) || isa<VAArgInst>(I) || Opcode == 58));
457 if (!isa<GetElementPtrInst>(&I)) {
458 for (unsigned i = 0; i < NumArgs; ++i) {
459 int Slot = Table.getSlot(I->getOperand(i));
460 assert(Slot >= 0 && "No slot number for value!?!?");
461 output_vbr((unsigned)Slot);
464 if (isa<CastInst>(I) || isa<VAArgInst>(I)) {
465 int Slot = Table.getSlot(I->getType());
466 assert(Slot != -1 && "Cast return type unknown?");
467 output_typeid((unsigned)Slot);
468 } else if (isa<CmpInst>(I)) {
469 output_vbr(unsigned(cast<CmpInst>(I)->getPredicate()));
470 } else if (isa<InvokeInst>(I)) {
471 output_vbr(cast<InvokeInst>(I)->getCallingConv());
472 } else if (Opcode == 58) { // Call escape sequence
473 output_vbr((cast<CallInst>(I)->getCallingConv() << 1) |
474 unsigned(cast<CallInst>(I)->isTailCall()));
477 int Slot = Table.getSlot(I->getOperand(0));
478 assert(Slot >= 0 && "No slot number for value!?!?");
479 output_vbr(unsigned(Slot));
481 // We need to encode the type of sequential type indices into their slot #
483 for (gep_type_iterator TI = gep_type_begin(I), E = gep_type_end(I);
484 Idx != NumArgs; ++TI, ++Idx) {
485 Slot = Table.getSlot(I->getOperand(Idx));
486 assert(Slot >= 0 && "No slot number for value!?!?");
488 if (isa<SequentialType>(*TI)) {
490 switch (I->getOperand(Idx)->getType()->getTypeID()) {
491 default: assert(0 && "Unknown index type!");
492 case Type::UIntTyID: IdxId = 0; break;
493 case Type::IntTyID: IdxId = 1; break;
494 case Type::ULongTyID: IdxId = 2; break;
495 case Type::LongTyID: IdxId = 3; break;
497 Slot = (Slot << 2) | IdxId;
499 output_vbr(unsigned(Slot));
505 // outputInstrVarArgsCall - Output the absurdly annoying varargs function calls.
506 // This are more annoying than most because the signature of the call does not
507 // tell us anything about the types of the arguments in the varargs portion.
508 // Because of this, we encode (as type 0) all of the argument types explicitly
509 // before the argument value. This really sucks, but you shouldn't be using
510 // varargs functions in your code! *death to printf*!
512 // Format: [opcode] [type] [numargs] [arg0] [arg1] ... [arg<numargs-1>]
514 void BytecodeWriter::outputInstrVarArgsCall(const Instruction *I,
516 const SlotCalculator &Table,
518 assert(isa<CallInst>(I) || isa<InvokeInst>(I));
519 // Opcode must have top two bits clear...
520 output_vbr(Opcode << 2); // Instruction Opcode ID
521 output_typeid(Type); // Result type (varargs type)
523 const PointerType *PTy = cast<PointerType>(I->getOperand(0)->getType());
524 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
525 unsigned NumParams = FTy->getNumParams();
527 unsigned NumFixedOperands;
528 if (isa<CallInst>(I)) {
529 // Output an operand for the callee and each fixed argument, then two for
530 // each variable argument.
531 NumFixedOperands = 1+NumParams;
533 assert(isa<InvokeInst>(I) && "Not call or invoke??");
534 // Output an operand for the callee and destinations, then two for each
535 // variable argument.
536 NumFixedOperands = 3+NumParams;
538 output_vbr(2 * I->getNumOperands()-NumFixedOperands +
539 unsigned(Opcode == 58 || isa<InvokeInst>(I)));
541 // The type for the function has already been emitted in the type field of the
542 // instruction. Just emit the slot # now.
543 for (unsigned i = 0; i != NumFixedOperands; ++i) {
544 int Slot = Table.getSlot(I->getOperand(i));
545 assert(Slot >= 0 && "No slot number for value!?!?");
546 output_vbr((unsigned)Slot);
549 for (unsigned i = NumFixedOperands, e = I->getNumOperands(); i != e; ++i) {
550 // Output Arg Type ID
551 int Slot = Table.getSlot(I->getOperand(i)->getType());
552 assert(Slot >= 0 && "No slot number for value!?!?");
553 output_typeid((unsigned)Slot);
555 // Output arg ID itself
556 Slot = Table.getSlot(I->getOperand(i));
557 assert(Slot >= 0 && "No slot number for value!?!?");
558 output_vbr((unsigned)Slot);
561 if (isa<InvokeInst>(I)) {
562 // Emit the tail call/calling conv for invoke instructions
563 output_vbr(cast<InvokeInst>(I)->getCallingConv());
564 } else if (Opcode == 58) {
565 const CallInst *CI = cast<CallInst>(I);
566 output_vbr((CI->getCallingConv() << 1) | unsigned(CI->isTailCall()));
571 // outputInstructionFormat1 - Output one operand instructions, knowing that no
572 // operand index is >= 2^12.
574 inline void BytecodeWriter::outputInstructionFormat1(const Instruction *I,
578 // bits Instruction format:
579 // --------------------------
580 // 01-00: Opcode type, fixed to 1.
582 // 19-08: Resulting type plane
583 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
585 output(1 | (Opcode << 2) | (Type << 8) | (Slots[0] << 20));
589 // outputInstructionFormat2 - Output two operand instructions, knowing that no
590 // operand index is >= 2^8.
592 inline void BytecodeWriter::outputInstructionFormat2(const Instruction *I,
596 // bits Instruction format:
597 // --------------------------
598 // 01-00: Opcode type, fixed to 2.
600 // 15-08: Resulting type plane
604 output(2 | (Opcode << 2) | (Type << 8) | (Slots[0] << 16) | (Slots[1] << 24));
608 // outputInstructionFormat3 - Output three operand instructions, knowing that no
609 // operand index is >= 2^6.
611 inline void BytecodeWriter::outputInstructionFormat3(const Instruction *I,
615 // bits Instruction format:
616 // --------------------------
617 // 01-00: Opcode type, fixed to 3.
619 // 13-08: Resulting type plane
624 output(3 | (Opcode << 2) | (Type << 8) |
625 (Slots[0] << 14) | (Slots[1] << 20) | (Slots[2] << 26));
628 void BytecodeWriter::outputInstruction(const Instruction &I) {
629 assert(I.getOpcode() < 57 && "Opcode too big???");
630 unsigned Opcode = I.getOpcode();
631 unsigned NumOperands = I.getNumOperands();
633 // Encode 'tail call' as 61, 'volatile load' as 62, and 'volatile store' as
635 if (const CallInst *CI = dyn_cast<CallInst>(&I)) {
636 if (CI->getCallingConv() == CallingConv::C) {
637 if (CI->isTailCall())
638 Opcode = 61; // CCC + Tail Call
640 ; // Opcode = Instruction::Call
641 } else if (CI->getCallingConv() == CallingConv::Fast) {
642 if (CI->isTailCall())
643 Opcode = 59; // FastCC + TailCall
645 Opcode = 60; // FastCC + Not Tail Call
647 Opcode = 58; // Call escape sequence.
649 } else if (isa<LoadInst>(I) && cast<LoadInst>(I).isVolatile()) {
651 } else if (isa<StoreInst>(I) && cast<StoreInst>(I).isVolatile()) {
655 // Figure out which type to encode with the instruction. Typically we want
656 // the type of the first parameter, as opposed to the type of the instruction
657 // (for example, with setcc, we always know it returns bool, but the type of
658 // the first param is actually interesting). But if we have no arguments
659 // we take the type of the instruction itself.
662 switch (I.getOpcode()) {
663 case Instruction::Select:
664 case Instruction::Malloc:
665 case Instruction::Alloca:
666 Ty = I.getType(); // These ALWAYS want to encode the return type
668 case Instruction::Store:
669 Ty = I.getOperand(1)->getType(); // Encode the pointer type...
670 assert(isa<PointerType>(Ty) && "Store to nonpointer type!?!?");
672 default: // Otherwise use the default behavior...
673 Ty = NumOperands ? I.getOperand(0)->getType() : I.getType();
678 int Slot = Table.getSlot(Ty);
679 assert(Slot != -1 && "Type not available!!?!");
680 Type = (unsigned)Slot;
682 // Varargs calls and invokes are encoded entirely different from any other
684 if (const CallInst *CI = dyn_cast<CallInst>(&I)){
685 const PointerType *Ty =cast<PointerType>(CI->getCalledValue()->getType());
686 if (cast<FunctionType>(Ty->getElementType())->isVarArg()) {
687 outputInstrVarArgsCall(CI, Opcode, Table, Type);
690 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(&I)) {
691 const PointerType *Ty =cast<PointerType>(II->getCalledValue()->getType());
692 if (cast<FunctionType>(Ty->getElementType())->isVarArg()) {
693 outputInstrVarArgsCall(II, Opcode, Table, Type);
698 if (NumOperands <= 3) {
699 // Make sure that we take the type number into consideration. We don't want
700 // to overflow the field size for the instruction format we select.
702 unsigned MaxOpSlot = Type;
703 unsigned Slots[3]; Slots[0] = (1 << 12)-1; // Marker to signify 0 operands
705 for (unsigned i = 0; i != NumOperands; ++i) {
706 int slot = Table.getSlot(I.getOperand(i));
707 assert(slot != -1 && "Broken bytecode!");
708 if (unsigned(slot) > MaxOpSlot) MaxOpSlot = unsigned(slot);
709 Slots[i] = unsigned(slot);
712 // Handle the special cases for various instructions...
713 if (isa<CastInst>(I) || isa<VAArgInst>(I)) {
714 // Cast has to encode the destination type as the second argument in the
715 // packet, or else we won't know what type to cast to!
716 Slots[1] = Table.getSlot(I.getType());
717 assert(Slots[1] != ~0U && "Cast return type unknown?");
718 if (Slots[1] > MaxOpSlot) MaxOpSlot = Slots[1];
720 } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(&I)) {
721 assert(NumOperands == 1 && "Bogus allocation!");
722 if (AI->getAlignment()) {
723 Slots[1] = Log2_32(AI->getAlignment())+1;
724 if (Slots[1] > MaxOpSlot) MaxOpSlot = Slots[1];
727 } else if (isa<ICmpInst>(I) || isa<FCmpInst>(I)) {
728 // We need to encode the compare instruction's predicate as the third
729 // operand. Its not really a slot, but we don't want to break the
730 // instruction format for these instructions.
732 assert(NumOperands == 3 && "CmpInst with wrong number of operands?");
733 Slots[2] = unsigned(cast<CmpInst>(&I)->getPredicate());
734 if (Slots[2] > MaxOpSlot)
735 MaxOpSlot = Slots[2];
736 } else if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
737 // We need to encode the type of sequential type indices into their slot #
739 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
741 if (isa<SequentialType>(*I)) {
743 switch (GEP->getOperand(Idx)->getType()->getTypeID()) {
744 default: assert(0 && "Unknown index type!");
745 case Type::UIntTyID: IdxId = 0; break;
746 case Type::IntTyID: IdxId = 1; break;
747 case Type::ULongTyID: IdxId = 2; break;
748 case Type::LongTyID: IdxId = 3; break;
750 Slots[Idx] = (Slots[Idx] << 2) | IdxId;
751 if (Slots[Idx] > MaxOpSlot) MaxOpSlot = Slots[Idx];
753 } else if (Opcode == 58) {
754 // If this is the escape sequence for call, emit the tailcall/cc info.
755 const CallInst &CI = cast<CallInst>(I);
757 if (NumOperands <= 3) {
758 Slots[NumOperands-1] =
759 (CI.getCallingConv() << 1)|unsigned(CI.isTailCall());
760 if (Slots[NumOperands-1] > MaxOpSlot)
761 MaxOpSlot = Slots[NumOperands-1];
763 } else if (isa<InvokeInst>(I)) {
764 // Invoke escape seq has at least 4 operands to encode.
768 // Decide which instruction encoding to use. This is determined primarily
769 // by the number of operands, and secondarily by whether or not the max
770 // operand will fit into the instruction encoding. More operands == fewer
773 switch (NumOperands) {
776 if (MaxOpSlot < (1 << 12)-1) { // -1 because we use 4095 to indicate 0 ops
777 outputInstructionFormat1(&I, Opcode, Slots, Type);
783 if (MaxOpSlot < (1 << 8)) {
784 outputInstructionFormat2(&I, Opcode, Slots, Type);
790 if (MaxOpSlot < (1 << 6)) {
791 outputInstructionFormat3(&I, Opcode, Slots, Type);
800 // If we weren't handled before here, we either have a large number of
801 // operands or a large operand index that we are referring to.
802 outputInstructionFormat0(&I, Opcode, Table, Type);
805 //===----------------------------------------------------------------------===//
806 //=== Block Output ===//
807 //===----------------------------------------------------------------------===//
809 BytecodeWriter::BytecodeWriter(std::vector<unsigned char> &o, const Module *M)
812 // Emit the signature...
813 static const unsigned char *Sig = (const unsigned char*)"llvm";
814 output_data(Sig, Sig+4);
816 // Emit the top level CLASS block.
817 BytecodeBlock ModuleBlock(BytecodeFormat::ModuleBlockID, *this, false, true);
819 bool isBigEndian = M->getEndianness() == Module::BigEndian;
820 bool hasLongPointers = M->getPointerSize() == Module::Pointer64;
821 bool hasNoEndianness = M->getEndianness() == Module::AnyEndianness;
822 bool hasNoPointerSize = M->getPointerSize() == Module::AnyPointerSize;
824 // Output the version identifier and other information.
825 unsigned Version = (BCVersionNum << 4) |
826 (unsigned)isBigEndian | (hasLongPointers << 1) |
827 (hasNoEndianness << 2) |
828 (hasNoPointerSize << 3);
831 // The Global type plane comes first
833 BytecodeBlock CPool(BytecodeFormat::GlobalTypePlaneBlockID, *this);
834 outputTypes(Type::FirstDerivedTyID);
837 // The ModuleInfoBlock follows directly after the type information
838 outputModuleInfoBlock(M);
840 // Output module level constants, used for global variable initializers
841 outputConstants(false);
843 // Do the whole module now! Process each function at a time...
844 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I)
847 // If needed, output the symbol table for the module...
848 outputSymbolTable(M->getSymbolTable());
851 void BytecodeWriter::outputTypes(unsigned TypeNum) {
852 // Write the type plane for types first because earlier planes (e.g. for a
853 // primitive type like float) may have constants constructed using types
854 // coming later (e.g., via getelementptr from a pointer type). The type
855 // plane is needed before types can be fwd or bkwd referenced.
856 const std::vector<const Type*>& Types = Table.getTypes();
857 assert(!Types.empty() && "No types at all?");
858 assert(TypeNum <= Types.size() && "Invalid TypeNo index");
860 unsigned NumEntries = Types.size() - TypeNum;
862 // Output type header: [num entries]
863 output_vbr(NumEntries);
865 for (unsigned i = TypeNum; i < TypeNum+NumEntries; ++i)
866 outputType(Types[i]);
869 // Helper function for outputConstants().
870 // Writes out all the constants in the plane Plane starting at entry StartNo.
872 void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
873 &Plane, unsigned StartNo) {
874 unsigned ValNo = StartNo;
876 // Scan through and ignore function arguments, global values, and constant
878 for (; ValNo < Plane.size() &&
879 (isa<Argument>(Plane[ValNo]) || isa<GlobalValue>(Plane[ValNo]) ||
880 (isa<ConstantArray>(Plane[ValNo]) &&
881 cast<ConstantArray>(Plane[ValNo])->isString())); ValNo++)
884 unsigned NC = ValNo; // Number of constants
885 for (; NC < Plane.size() && (isa<Constant>(Plane[NC]) ||
886 isa<InlineAsm>(Plane[NC])); NC++)
888 NC -= ValNo; // Convert from index into count
889 if (NC == 0) return; // Skip empty type planes...
891 // FIXME: Most slabs only have 1 or 2 entries! We should encode this much
894 // Put out type header: [num entries][type id number]
898 // Put out the Type ID Number...
899 int Slot = Table.getSlot(Plane.front()->getType());
900 assert (Slot != -1 && "Type in constant pool but not in function!!");
901 output_typeid((unsigned)Slot);
903 for (unsigned i = ValNo; i < ValNo+NC; ++i) {
904 const Value *V = Plane[i];
905 if (const Constant *C = dyn_cast<Constant>(V))
908 outputInlineAsm(cast<InlineAsm>(V));
912 static inline bool hasNullValue(const Type *Ty) {
913 return Ty != Type::LabelTy && Ty != Type::VoidTy && !isa<OpaqueType>(Ty);
916 void BytecodeWriter::outputConstants(bool isFunction) {
917 BytecodeBlock CPool(BytecodeFormat::ConstantPoolBlockID, *this,
918 true /* Elide block if empty */);
920 unsigned NumPlanes = Table.getNumPlanes();
923 // Output the type plane before any constants!
924 outputTypes(Table.getModuleTypeLevel());
926 // Output module-level string constants before any other constants.
927 outputConstantStrings();
929 for (unsigned pno = 0; pno != NumPlanes; pno++) {
930 const std::vector<const Value*> &Plane = Table.getPlane(pno);
931 if (!Plane.empty()) { // Skip empty type planes...
933 if (isFunction) // Don't re-emit module constants
934 ValNo += Table.getModuleLevel(pno);
936 if (hasNullValue(Plane[0]->getType())) {
937 // Skip zero initializer
942 // Write out constants in the plane
943 outputConstantsInPlane(Plane, ValNo);
948 static unsigned getEncodedLinkage(const GlobalValue *GV) {
949 switch (GV->getLinkage()) {
950 default: assert(0 && "Invalid linkage!");
951 case GlobalValue::ExternalLinkage: return 0;
952 case GlobalValue::WeakLinkage: return 1;
953 case GlobalValue::AppendingLinkage: return 2;
954 case GlobalValue::InternalLinkage: return 3;
955 case GlobalValue::LinkOnceLinkage: return 4;
956 case GlobalValue::DLLImportLinkage: return 5;
957 case GlobalValue::DLLExportLinkage: return 6;
958 case GlobalValue::ExternalWeakLinkage: return 7;
962 void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
963 BytecodeBlock ModuleInfoBlock(BytecodeFormat::ModuleGlobalInfoBlockID, *this);
965 // Give numbers to sections as we encounter them.
966 unsigned SectionIDCounter = 0;
967 std::vector<std::string> SectionNames;
968 std::map<std::string, unsigned> SectionID;
970 // Output the types for the global variables in the module...
971 for (Module::const_global_iterator I = M->global_begin(),
972 End = M->global_end(); I != End; ++I) {
973 int Slot = Table.getSlot(I->getType());
974 assert(Slot != -1 && "Module global vars is broken!");
976 assert((I->hasInitializer() || !I->hasInternalLinkage()) &&
977 "Global must have an initializer or have external linkage!");
979 // Fields: bit0 = isConstant, bit1 = hasInitializer, bit2-4=Linkage,
980 // bit5+ = Slot # for type.
981 bool HasExtensionWord = (I->getAlignment() != 0) || I->hasSection();
983 // If we need to use the extension byte, set linkage=3(internal) and
984 // initializer = 0 (impossible!).
985 if (!HasExtensionWord) {
986 unsigned oSlot = ((unsigned)Slot << 5) | (getEncodedLinkage(I) << 2) |
987 (I->hasInitializer() << 1) | (unsigned)I->isConstant();
990 unsigned oSlot = ((unsigned)Slot << 5) | (3 << 2) |
991 (0 << 1) | (unsigned)I->isConstant();
994 // The extension word has this format: bit 0 = has initializer, bit 1-3 =
995 // linkage, bit 4-8 = alignment (log2), bit 9 = has SectionID,
996 // bits 10+ = future use.
997 unsigned ExtWord = (unsigned)I->hasInitializer() |
998 (getEncodedLinkage(I) << 1) |
999 ((Log2_32(I->getAlignment())+1) << 4) |
1000 ((unsigned)I->hasSection() << 9);
1001 output_vbr(ExtWord);
1002 if (I->hasSection()) {
1003 // Give section names unique ID's.
1004 unsigned &Entry = SectionID[I->getSection()];
1006 Entry = ++SectionIDCounter;
1007 SectionNames.push_back(I->getSection());
1013 // If we have an initializer, output it now.
1014 if (I->hasInitializer()) {
1015 Slot = Table.getSlot((Value*)I->getInitializer());
1016 assert(Slot != -1 && "No slot for global var initializer!");
1017 output_vbr((unsigned)Slot);
1020 output_typeid((unsigned)Table.getSlot(Type::VoidTy));
1022 // Output the types of the functions in this module.
1023 for (Module::const_iterator I = M->begin(), End = M->end(); I != End; ++I) {
1024 int Slot = Table.getSlot(I->getType());
1025 assert(Slot != -1 && "Module slot calculator is broken!");
1026 assert(Slot >= Type::FirstDerivedTyID && "Derived type not in range!");
1027 assert(((Slot << 6) >> 6) == Slot && "Slot # too big!");
1028 unsigned CC = I->getCallingConv()+1;
1029 unsigned ID = (Slot << 5) | (CC & 15);
1031 if (I->isExternal()) // If external, we don't have an FunctionInfo block.
1034 if (I->getAlignment() || I->hasSection() || (CC & ~15) != 0 ||
1035 (I->isExternal() && I->hasDLLImportLinkage()) ||
1036 (I->isExternal() && I->hasExternalWeakLinkage())
1038 ID |= 1 << 31; // Do we need an extension word?
1042 if (ID & (1 << 31)) {
1043 // Extension byte: bits 0-4 = alignment, bits 5-9 = top nibble of calling
1044 // convention, bit 10 = hasSectionID., bits 11-12 = external linkage type
1045 unsigned extLinkage = 0;
1047 if (I->isExternal()) {
1048 if (I->hasDLLImportLinkage()) {
1050 } else if (I->hasExternalWeakLinkage()) {
1055 ID = (Log2_32(I->getAlignment())+1) | ((CC >> 4) << 5) |
1056 (I->hasSection() << 10) |
1057 ((extLinkage & 3) << 11);
1060 // Give section names unique ID's.
1061 if (I->hasSection()) {
1062 unsigned &Entry = SectionID[I->getSection()];
1064 Entry = ++SectionIDCounter;
1065 SectionNames.push_back(I->getSection());
1071 output_vbr((unsigned)Table.getSlot(Type::VoidTy) << 5);
1073 // Emit the list of dependent libraries for the Module.
1074 Module::lib_iterator LI = M->lib_begin();
1075 Module::lib_iterator LE = M->lib_end();
1076 output_vbr(unsigned(LE - LI)); // Emit the number of dependent libraries.
1077 for (; LI != LE; ++LI)
1080 // Output the target triple from the module
1081 output(M->getTargetTriple());
1083 // Emit the table of section names.
1084 output_vbr((unsigned)SectionNames.size());
1085 for (unsigned i = 0, e = SectionNames.size(); i != e; ++i)
1086 output(SectionNames[i]);
1088 // Output the inline asm string.
1089 output(M->getModuleInlineAsm());
1092 void BytecodeWriter::outputInstructions(const Function *F) {
1093 BytecodeBlock ILBlock(BytecodeFormat::InstructionListBlockID, *this);
1094 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1095 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
1096 outputInstruction(*I);
1099 void BytecodeWriter::outputFunction(const Function *F) {
1100 // If this is an external function, there is nothing else to emit!
1101 if (F->isExternal()) return;
1103 BytecodeBlock FunctionBlock(BytecodeFormat::FunctionBlockID, *this);
1104 output_vbr(getEncodedLinkage(F));
1106 // Get slot information about the function...
1107 Table.incorporateFunction(F);
1109 if (Table.getCompactionTable().empty()) {
1110 // Output information about the constants in the function if the compaction
1111 // table is not being used.
1112 outputConstants(true);
1114 // Otherwise, emit the compaction table.
1115 outputCompactionTable();
1118 // Output all of the instructions in the body of the function
1119 outputInstructions(F);
1121 // If needed, output the symbol table for the function...
1122 outputSymbolTable(F->getSymbolTable());
1124 Table.purgeFunction();
1127 void BytecodeWriter::outputCompactionTablePlane(unsigned PlaneNo,
1128 const std::vector<const Value*> &Plane,
1130 unsigned End = Table.getModuleLevel(PlaneNo);
1131 if (Plane.empty() || StartNo == End || End == 0) return; // Nothing to emit
1132 assert(StartNo < End && "Cannot emit negative range!");
1133 assert(StartNo < Plane.size() && End <= Plane.size());
1135 // Do not emit the null initializer!
1138 // Figure out which encoding to use. By far the most common case we have is
1139 // to emit 0-2 entries in a compaction table plane.
1140 switch (End-StartNo) {
1141 case 0: // Avoid emitting two vbr's if possible.
1144 output_vbr((PlaneNo << 2) | End-StartNo);
1147 // Output the number of things.
1148 output_vbr((unsigned(End-StartNo) << 2) | 3);
1149 output_typeid(PlaneNo); // Emit the type plane this is
1153 for (unsigned i = StartNo; i != End; ++i)
1154 output_vbr(Table.getGlobalSlot(Plane[i]));
1157 void BytecodeWriter::outputCompactionTypes(unsigned StartNo) {
1158 // Get the compaction type table from the slot calculator
1159 const std::vector<const Type*> &CTypes = Table.getCompactionTypes();
1161 // The compaction types may have been uncompactified back to the
1162 // global types. If so, we just write an empty table
1163 if (CTypes.size() == 0) {
1168 assert(CTypes.size() >= StartNo && "Invalid compaction types start index");
1170 // Determine how many types to write
1171 unsigned NumTypes = CTypes.size() - StartNo;
1173 // Output the number of types.
1174 output_vbr(NumTypes);
1176 for (unsigned i = StartNo; i < StartNo+NumTypes; ++i)
1177 output_typeid(Table.getGlobalSlot(CTypes[i]));
1180 void BytecodeWriter::outputCompactionTable() {
1181 // Avoid writing the compaction table at all if there is no content.
1182 if (Table.getCompactionTypes().size() >= Type::FirstDerivedTyID ||
1183 (!Table.CompactionTableIsEmpty())) {
1184 BytecodeBlock CTB(BytecodeFormat::CompactionTableBlockID, *this,
1185 true/*ElideIfEmpty*/);
1186 const std::vector<std::vector<const Value*> > &CT =
1187 Table.getCompactionTable();
1189 // First things first, emit the type compaction table if there is one.
1190 outputCompactionTypes(Type::FirstDerivedTyID);
1192 for (unsigned i = 0, e = CT.size(); i != e; ++i)
1193 outputCompactionTablePlane(i, CT[i], 0);
1197 void BytecodeWriter::outputSymbolTable(const SymbolTable &MST) {
1198 // Do not output the Bytecode block for an empty symbol table, it just wastes
1200 if (MST.isEmpty()) return;
1202 BytecodeBlock SymTabBlock(BytecodeFormat::SymbolTableBlockID, *this,
1203 true/*ElideIfEmpty*/);
1205 // Write the number of types
1206 output_vbr(MST.num_types());
1208 // Write each of the types
1209 for (SymbolTable::type_const_iterator TI = MST.type_begin(),
1210 TE = MST.type_end(); TI != TE; ++TI) {
1211 // Symtab entry:[def slot #][name]
1212 output_typeid((unsigned)Table.getSlot(TI->second));
1216 // Now do each of the type planes in order.
1217 for (SymbolTable::plane_const_iterator PI = MST.plane_begin(),
1218 PE = MST.plane_end(); PI != PE; ++PI) {
1219 SymbolTable::value_const_iterator I = MST.value_begin(PI->first);
1220 SymbolTable::value_const_iterator End = MST.value_end(PI->first);
1223 if (I == End) continue; // Don't mess with an absent type...
1225 // Write the number of values in this plane
1226 output_vbr((unsigned)PI->second.size());
1228 // Write the slot number of the type for this plane
1229 Slot = Table.getSlot(PI->first);
1230 assert(Slot != -1 && "Type in symtab, but not in table!");
1231 output_typeid((unsigned)Slot);
1233 // Write each of the values in this plane
1234 for (; I != End; ++I) {
1235 // Symtab entry: [def slot #][name]
1236 Slot = Table.getSlot(I->second);
1237 assert(Slot != -1 && "Value in symtab but has no slot number!!");
1238 output_vbr((unsigned)Slot);
1244 void llvm::WriteBytecodeToFile(const Module *M, OStream &Out,
1246 assert(M && "You can't write a null module!!");
1248 // Make sure that std::cout is put into binary mode for systems
1251 sys::Program::ChangeStdoutToBinary();
1253 // Create a vector of unsigned char for the bytecode output. We
1254 // reserve 256KBytes of space in the vector so that we avoid doing
1255 // lots of little allocations. 256KBytes is sufficient for a large
1256 // proportion of the bytecode files we will encounter. Larger files
1257 // will be automatically doubled in size as needed (std::vector
1259 std::vector<unsigned char> Buffer;
1260 Buffer.reserve(256 * 1024);
1262 // The BytecodeWriter populates Buffer for us.
1263 BytecodeWriter BCW(Buffer, M);
1265 // Keep track of how much we've written
1266 BytesWritten += Buffer.size();
1268 // Determine start and end points of the Buffer
1269 const unsigned char *FirstByte = &Buffer.front();
1271 // If we're supposed to compress this mess ...
1274 // We signal compression by using an alternate magic number for the
1275 // file. The compressed bytecode file's magic number is "llvc" instead
1277 char compressed_magic[4];
1278 compressed_magic[0] = 'l';
1279 compressed_magic[1] = 'l';
1280 compressed_magic[2] = 'v';
1281 compressed_magic[3] = 'c';
1283 Out.stream()->write(compressed_magic,4);
1285 // Compress everything after the magic number (which we altered)
1286 Compressor::compressToStream(
1287 (char*)(FirstByte+4), // Skip the magic number
1288 Buffer.size()-4, // Skip the magic number
1289 *Out.stream() // Where to write compressed data
1294 // We're not compressing, so just write the entire block.
1295 Out.stream()->write((char*)FirstByte, Buffer.size());
1298 // make sure it hits disk now
1299 Out.stream()->flush();