1 //===- BitstreamWriter.h - Low-level bitstream writer interface -*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This header defines the BitstreamWriter class.  This class can be used to
11 // write an arbitrary bitstream, regardless of its contents.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_BITCODE_BITSTREAMWRITER_H
16 #define LLVM_BITCODE_BITSTREAMWRITER_H
17 
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/Bitcode/BitCodes.h"
21 #include <vector>
22 
23 namespace llvm {
24 
25 class BitstreamWriter {
26   SmallVectorImpl<char> &Out;
27 
28   /// CurBit - Always between 0 and 31 inclusive, specifies the next bit to use.
29   unsigned CurBit;
30 
31   /// CurValue - The current value.  Only bits < CurBit are valid.
32   uint32_t CurValue;
33 
34   /// CurCodeSize - This is the declared size of code values used for the
35   /// current block, in bits.
36   unsigned CurCodeSize;
37 
38   /// BlockInfoCurBID - When emitting a BLOCKINFO_BLOCK, this is the currently
39   /// selected BLOCK ID.
40   unsigned BlockInfoCurBID;
41 
42   /// CurAbbrevs - Abbrevs installed at in this block.
43   std::vector<IntrusiveRefCntPtr<BitCodeAbbrev>> CurAbbrevs;
44 
45   struct Block {
46     unsigned PrevCodeSize;
47     unsigned StartSizeWord;
48     std::vector<IntrusiveRefCntPtr<BitCodeAbbrev>> PrevAbbrevs;
BlockBlock49     Block(unsigned PCS, unsigned SSW) : PrevCodeSize(PCS), StartSizeWord(SSW) {}
50   };
51 
52   /// BlockScope - This tracks the current blocks that we have entered.
53   std::vector<Block> BlockScope;
54 
55   /// BlockInfo - This contains information emitted to BLOCKINFO_BLOCK blocks.
56   /// These describe abbreviations that all blocks of the specified ID inherit.
57   struct BlockInfo {
58     unsigned BlockID;
59     std::vector<IntrusiveRefCntPtr<BitCodeAbbrev>> Abbrevs;
60   };
61   std::vector<BlockInfo> BlockInfoRecords;
62 
63   // BackpatchWord - Backpatch a 32-bit word in the output with the specified
64   // value.
BackpatchWord(unsigned ByteNo,unsigned NewWord)65   void BackpatchWord(unsigned ByteNo, unsigned NewWord) {
66     Out[ByteNo++] = (unsigned char)(NewWord >>  0);
67     Out[ByteNo++] = (unsigned char)(NewWord >>  8);
68     Out[ByteNo++] = (unsigned char)(NewWord >> 16);
69     Out[ByteNo  ] = (unsigned char)(NewWord >> 24);
70   }
71 
WriteByte(unsigned char Value)72   void WriteByte(unsigned char Value) {
73     Out.push_back(Value);
74   }
75 
WriteWord(unsigned Value)76   void WriteWord(unsigned Value) {
77     unsigned char Bytes[4] = {
78       (unsigned char)(Value >>  0),
79       (unsigned char)(Value >>  8),
80       (unsigned char)(Value >> 16),
81       (unsigned char)(Value >> 24) };
82     Out.append(&Bytes[0], &Bytes[4]);
83   }
84 
GetBufferOffset()85   unsigned GetBufferOffset() const {
86     return Out.size();
87   }
88 
GetWordIndex()89   unsigned GetWordIndex() const {
90     unsigned Offset = GetBufferOffset();
91     assert((Offset & 3) == 0 && "Not 32-bit aligned");
92     return Offset / 4;
93   }
94 
95 public:
BitstreamWriter(SmallVectorImpl<char> & O)96   explicit BitstreamWriter(SmallVectorImpl<char> &O)
97     : Out(O), CurBit(0), CurValue(0), CurCodeSize(2) {}
98 
~BitstreamWriter()99   ~BitstreamWriter() {
100     assert(CurBit == 0 && "Unflushed data remaining");
101     assert(BlockScope.empty() && CurAbbrevs.empty() && "Block imbalance");
102   }
103 
104   /// \brief Retrieve the current position in the stream, in bits.
GetCurrentBitNo()105   uint64_t GetCurrentBitNo() const { return GetBufferOffset() * 8 + CurBit; }
106 
107   //===--------------------------------------------------------------------===//
108   // Basic Primitives for emitting bits to the stream.
109   //===--------------------------------------------------------------------===//
110 
Emit(uint32_t Val,unsigned NumBits)111   void Emit(uint32_t Val, unsigned NumBits) {
112     assert(NumBits && NumBits <= 32 && "Invalid value size!");
113     assert((Val & ~(~0U >> (32-NumBits))) == 0 && "High bits set!");
114     CurValue |= Val << CurBit;
115     if (CurBit + NumBits < 32) {
116       CurBit += NumBits;
117       return;
118     }
119 
120     // Add the current word.
121     WriteWord(CurValue);
122 
123     if (CurBit)
124       CurValue = Val >> (32-CurBit);
125     else
126       CurValue = 0;
127     CurBit = (CurBit+NumBits) & 31;
128   }
129 
Emit64(uint64_t Val,unsigned NumBits)130   void Emit64(uint64_t Val, unsigned NumBits) {
131     if (NumBits <= 32)
132       Emit((uint32_t)Val, NumBits);
133     else {
134       Emit((uint32_t)Val, 32);
135       Emit((uint32_t)(Val >> 32), NumBits-32);
136     }
137   }
138 
FlushToWord()139   void FlushToWord() {
140     if (CurBit) {
141       WriteWord(CurValue);
142       CurBit = 0;
143       CurValue = 0;
144     }
145   }
146 
EmitVBR(uint32_t Val,unsigned NumBits)147   void EmitVBR(uint32_t Val, unsigned NumBits) {
148     assert(NumBits <= 32 && "Too many bits to emit!");
149     uint32_t Threshold = 1U << (NumBits-1);
150 
151     // Emit the bits with VBR encoding, NumBits-1 bits at a time.
152     while (Val >= Threshold) {
153       Emit((Val & ((1 << (NumBits-1))-1)) | (1 << (NumBits-1)), NumBits);
154       Val >>= NumBits-1;
155     }
156 
157     Emit(Val, NumBits);
158   }
159 
EmitVBR64(uint64_t Val,unsigned NumBits)160   void EmitVBR64(uint64_t Val, unsigned NumBits) {
161     assert(NumBits <= 32 && "Too many bits to emit!");
162     if ((uint32_t)Val == Val)
163       return EmitVBR((uint32_t)Val, NumBits);
164 
165     uint32_t Threshold = 1U << (NumBits-1);
166 
167     // Emit the bits with VBR encoding, NumBits-1 bits at a time.
168     while (Val >= Threshold) {
169       Emit(((uint32_t)Val & ((1 << (NumBits-1))-1)) |
170            (1 << (NumBits-1)), NumBits);
171       Val >>= NumBits-1;
172     }
173 
174     Emit((uint32_t)Val, NumBits);
175   }
176 
177   /// EmitCode - Emit the specified code.
EmitCode(unsigned Val)178   void EmitCode(unsigned Val) {
179     Emit(Val, CurCodeSize);
180   }
181 
182   //===--------------------------------------------------------------------===//
183   // Block Manipulation
184   //===--------------------------------------------------------------------===//
185 
186   /// getBlockInfo - If there is block info for the specified ID, return it,
187   /// otherwise return null.
getBlockInfo(unsigned BlockID)188   BlockInfo *getBlockInfo(unsigned BlockID) {
189     // Common case, the most recent entry matches BlockID.
190     if (!BlockInfoRecords.empty() && BlockInfoRecords.back().BlockID == BlockID)
191       return &BlockInfoRecords.back();
192 
193     for (unsigned i = 0, e = static_cast<unsigned>(BlockInfoRecords.size());
194          i != e; ++i)
195       if (BlockInfoRecords[i].BlockID == BlockID)
196         return &BlockInfoRecords[i];
197     return nullptr;
198   }
199 
EnterSubblock(unsigned BlockID,unsigned CodeLen)200   void EnterSubblock(unsigned BlockID, unsigned CodeLen) {
201     // Block header:
202     //    [ENTER_SUBBLOCK, blockid, newcodelen, <align4bytes>, blocklen]
203     EmitCode(bitc::ENTER_SUBBLOCK);
204     EmitVBR(BlockID, bitc::BlockIDWidth);
205     EmitVBR(CodeLen, bitc::CodeLenWidth);
206     FlushToWord();
207 
208     unsigned BlockSizeWordIndex = GetWordIndex();
209     unsigned OldCodeSize = CurCodeSize;
210 
211     // Emit a placeholder, which will be replaced when the block is popped.
212     Emit(0, bitc::BlockSizeWidth);
213 
214     CurCodeSize = CodeLen;
215 
216     // Push the outer block's abbrev set onto the stack, start out with an
217     // empty abbrev set.
218     BlockScope.push_back(Block(OldCodeSize, BlockSizeWordIndex));
219     BlockScope.back().PrevAbbrevs.swap(CurAbbrevs);
220 
221     // If there is a blockinfo for this BlockID, add all the predefined abbrevs
222     // to the abbrev list.
223     if (BlockInfo *Info = getBlockInfo(BlockID)) {
224       CurAbbrevs.insert(CurAbbrevs.end(), Info->Abbrevs.begin(),
225                         Info->Abbrevs.end());
226     }
227   }
228 
ExitBlock()229   void ExitBlock() {
230     assert(!BlockScope.empty() && "Block scope imbalance!");
231     const Block &B = BlockScope.back();
232 
233     // Block tail:
234     //    [END_BLOCK, <align4bytes>]
235     EmitCode(bitc::END_BLOCK);
236     FlushToWord();
237 
238     // Compute the size of the block, in words, not counting the size field.
239     unsigned SizeInWords = GetWordIndex() - B.StartSizeWord - 1;
240     unsigned ByteNo = B.StartSizeWord*4;
241 
242     // Update the block size field in the header of this sub-block.
243     BackpatchWord(ByteNo, SizeInWords);
244 
245     // Restore the inner block's code size and abbrev table.
246     CurCodeSize = B.PrevCodeSize;
247     CurAbbrevs = std::move(B.PrevAbbrevs);
248     BlockScope.pop_back();
249   }
250 
251   //===--------------------------------------------------------------------===//
252   // Record Emission
253   //===--------------------------------------------------------------------===//
254 
255 private:
256   /// EmitAbbreviatedLiteral - Emit a literal value according to its abbrev
257   /// record.  This is a no-op, since the abbrev specifies the literal to use.
258   template<typename uintty>
EmitAbbreviatedLiteral(const BitCodeAbbrevOp & Op,uintty V)259   void EmitAbbreviatedLiteral(const BitCodeAbbrevOp &Op, uintty V) {
260     assert(Op.isLiteral() && "Not a literal");
261     // If the abbrev specifies the literal value to use, don't emit
262     // anything.
263     assert(V == Op.getLiteralValue() &&
264            "Invalid abbrev for record!");
265   }
266 
267   /// EmitAbbreviatedField - Emit a single scalar field value with the specified
268   /// encoding.
269   template<typename uintty>
EmitAbbreviatedField(const BitCodeAbbrevOp & Op,uintty V)270   void EmitAbbreviatedField(const BitCodeAbbrevOp &Op, uintty V) {
271     assert(!Op.isLiteral() && "Literals should use EmitAbbreviatedLiteral!");
272 
273     // Encode the value as we are commanded.
274     switch (Op.getEncoding()) {
275     default: llvm_unreachable("Unknown encoding!");
276     case BitCodeAbbrevOp::Fixed:
277       if (Op.getEncodingData())
278         Emit((unsigned)V, (unsigned)Op.getEncodingData());
279       break;
280     case BitCodeAbbrevOp::VBR:
281       if (Op.getEncodingData())
282         EmitVBR64(V, (unsigned)Op.getEncodingData());
283       break;
284     case BitCodeAbbrevOp::Char6:
285       Emit(BitCodeAbbrevOp::EncodeChar6((char)V), 6);
286       break;
287     }
288   }
289 
290   /// EmitRecordWithAbbrevImpl - This is the core implementation of the record
291   /// emission code.  If BlobData is non-null, then it specifies an array of
292   /// data that should be emitted as part of the Blob or Array operand that is
293   /// known to exist at the end of the record.
294   template<typename uintty>
EmitRecordWithAbbrevImpl(unsigned Abbrev,SmallVectorImpl<uintty> & Vals,StringRef Blob)295   void EmitRecordWithAbbrevImpl(unsigned Abbrev, SmallVectorImpl<uintty> &Vals,
296                                 StringRef Blob) {
297     const char *BlobData = Blob.data();
298     unsigned BlobLen = (unsigned) Blob.size();
299     unsigned AbbrevNo = Abbrev-bitc::FIRST_APPLICATION_ABBREV;
300     assert(AbbrevNo < CurAbbrevs.size() && "Invalid abbrev #!");
301     const BitCodeAbbrev *Abbv = CurAbbrevs[AbbrevNo].get();
302 
303     EmitCode(Abbrev);
304 
305     unsigned RecordIdx = 0;
306     for (unsigned i = 0, e = static_cast<unsigned>(Abbv->getNumOperandInfos());
307          i != e; ++i) {
308       const BitCodeAbbrevOp &Op = Abbv->getOperandInfo(i);
309       if (Op.isLiteral()) {
310         assert(RecordIdx < Vals.size() && "Invalid abbrev/record");
311         EmitAbbreviatedLiteral(Op, Vals[RecordIdx]);
312         ++RecordIdx;
313       } else if (Op.getEncoding() == BitCodeAbbrevOp::Array) {
314         // Array case.
315         assert(i+2 == e && "array op not second to last?");
316         const BitCodeAbbrevOp &EltEnc = Abbv->getOperandInfo(++i);
317 
318         // If this record has blob data, emit it, otherwise we must have record
319         // entries to encode this way.
320         if (BlobData) {
321           assert(RecordIdx == Vals.size() &&
322                  "Blob data and record entries specified for array!");
323           // Emit a vbr6 to indicate the number of elements present.
324           EmitVBR(static_cast<uint32_t>(BlobLen), 6);
325 
326           // Emit each field.
327           for (unsigned i = 0; i != BlobLen; ++i)
328             EmitAbbreviatedField(EltEnc, (unsigned char)BlobData[i]);
329 
330           // Know that blob data is consumed for assertion below.
331           BlobData = nullptr;
332         } else {
333           // Emit a vbr6 to indicate the number of elements present.
334           EmitVBR(static_cast<uint32_t>(Vals.size()-RecordIdx), 6);
335 
336           // Emit each field.
337           for (unsigned e = Vals.size(); RecordIdx != e; ++RecordIdx)
338             EmitAbbreviatedField(EltEnc, Vals[RecordIdx]);
339         }
340       } else if (Op.getEncoding() == BitCodeAbbrevOp::Blob) {
341         // If this record has blob data, emit it, otherwise we must have record
342         // entries to encode this way.
343 
344         // Emit a vbr6 to indicate the number of elements present.
345         if (BlobData) {
346           EmitVBR(static_cast<uint32_t>(BlobLen), 6);
347           assert(RecordIdx == Vals.size() &&
348                  "Blob data and record entries specified for blob operand!");
349         } else {
350           EmitVBR(static_cast<uint32_t>(Vals.size()-RecordIdx), 6);
351         }
352 
353         // Flush to a 32-bit alignment boundary.
354         FlushToWord();
355 
356         // Emit each field as a literal byte.
357         if (BlobData) {
358           for (unsigned i = 0; i != BlobLen; ++i)
359             WriteByte((unsigned char)BlobData[i]);
360 
361           // Know that blob data is consumed for assertion below.
362           BlobData = nullptr;
363         } else {
364           for (unsigned e = Vals.size(); RecordIdx != e; ++RecordIdx) {
365             assert(isUInt<8>(Vals[RecordIdx]) &&
366                    "Value too large to emit as blob");
367             WriteByte((unsigned char)Vals[RecordIdx]);
368           }
369         }
370 
371         // Align end to 32-bits.
372         while (GetBufferOffset() & 3)
373           WriteByte(0);
374       } else {  // Single scalar field.
375         assert(RecordIdx < Vals.size() && "Invalid abbrev/record");
376         EmitAbbreviatedField(Op, Vals[RecordIdx]);
377         ++RecordIdx;
378       }
379     }
380     assert(RecordIdx == Vals.size() && "Not all record operands emitted!");
381     assert(BlobData == nullptr &&
382            "Blob data specified for record that doesn't use it!");
383   }
384 
385 public:
386 
387   /// EmitRecord - Emit the specified record to the stream, using an abbrev if
388   /// we have one to compress the output.
389   template<typename uintty>
390   void EmitRecord(unsigned Code, SmallVectorImpl<uintty> &Vals,
391                   unsigned Abbrev = 0) {
392     if (!Abbrev) {
393       // If we don't have an abbrev to use, emit this in its fully unabbreviated
394       // form.
395       EmitCode(bitc::UNABBREV_RECORD);
396       EmitVBR(Code, 6);
397       EmitVBR(static_cast<uint32_t>(Vals.size()), 6);
398       for (unsigned i = 0, e = static_cast<unsigned>(Vals.size()); i != e; ++i)
399         EmitVBR64(Vals[i], 6);
400       return;
401     }
402 
403     // Insert the code into Vals to treat it uniformly.
404     Vals.insert(Vals.begin(), Code);
405 
406     EmitRecordWithAbbrev(Abbrev, Vals);
407   }
408 
409   /// EmitRecordWithAbbrev - Emit a record with the specified abbreviation.
410   /// Unlike EmitRecord, the code for the record should be included in Vals as
411   /// the first entry.
412   template<typename uintty>
EmitRecordWithAbbrev(unsigned Abbrev,SmallVectorImpl<uintty> & Vals)413   void EmitRecordWithAbbrev(unsigned Abbrev, SmallVectorImpl<uintty> &Vals) {
414     EmitRecordWithAbbrevImpl(Abbrev, Vals, StringRef());
415   }
416 
417   /// EmitRecordWithBlob - Emit the specified record to the stream, using an
418   /// abbrev that includes a blob at the end.  The blob data to emit is
419   /// specified by the pointer and length specified at the end.  In contrast to
420   /// EmitRecord, this routine expects that the first entry in Vals is the code
421   /// of the record.
422   template<typename uintty>
EmitRecordWithBlob(unsigned Abbrev,SmallVectorImpl<uintty> & Vals,StringRef Blob)423   void EmitRecordWithBlob(unsigned Abbrev, SmallVectorImpl<uintty> &Vals,
424                           StringRef Blob) {
425     EmitRecordWithAbbrevImpl(Abbrev, Vals, Blob);
426   }
427   template<typename uintty>
EmitRecordWithBlob(unsigned Abbrev,SmallVectorImpl<uintty> & Vals,const char * BlobData,unsigned BlobLen)428   void EmitRecordWithBlob(unsigned Abbrev, SmallVectorImpl<uintty> &Vals,
429                           const char *BlobData, unsigned BlobLen) {
430     return EmitRecordWithAbbrevImpl(Abbrev, Vals, StringRef(BlobData, BlobLen));
431   }
432 
433   /// EmitRecordWithArray - Just like EmitRecordWithBlob, works with records
434   /// that end with an array.
435   template<typename uintty>
EmitRecordWithArray(unsigned Abbrev,SmallVectorImpl<uintty> & Vals,StringRef Array)436   void EmitRecordWithArray(unsigned Abbrev, SmallVectorImpl<uintty> &Vals,
437                           StringRef Array) {
438     EmitRecordWithAbbrevImpl(Abbrev, Vals, Array);
439   }
440   template<typename uintty>
EmitRecordWithArray(unsigned Abbrev,SmallVectorImpl<uintty> & Vals,const char * ArrayData,unsigned ArrayLen)441   void EmitRecordWithArray(unsigned Abbrev, SmallVectorImpl<uintty> &Vals,
442                           const char *ArrayData, unsigned ArrayLen) {
443     return EmitRecordWithAbbrevImpl(Abbrev, Vals, StringRef(ArrayData,
444                                                             ArrayLen));
445   }
446 
447   //===--------------------------------------------------------------------===//
448   // Abbrev Emission
449   //===--------------------------------------------------------------------===//
450 
451 private:
452   // Emit the abbreviation as a DEFINE_ABBREV record.
EncodeAbbrev(BitCodeAbbrev * Abbv)453   void EncodeAbbrev(BitCodeAbbrev *Abbv) {
454     EmitCode(bitc::DEFINE_ABBREV);
455     EmitVBR(Abbv->getNumOperandInfos(), 5);
456     for (unsigned i = 0, e = static_cast<unsigned>(Abbv->getNumOperandInfos());
457          i != e; ++i) {
458       const BitCodeAbbrevOp &Op = Abbv->getOperandInfo(i);
459       Emit(Op.isLiteral(), 1);
460       if (Op.isLiteral()) {
461         EmitVBR64(Op.getLiteralValue(), 8);
462       } else {
463         Emit(Op.getEncoding(), 3);
464         if (Op.hasEncodingData())
465           EmitVBR64(Op.getEncodingData(), 5);
466       }
467     }
468   }
469 public:
470 
471   /// EmitAbbrev - This emits an abbreviation to the stream.  Note that this
472   /// method takes ownership of the specified abbrev.
EmitAbbrev(BitCodeAbbrev * Abbv)473   unsigned EmitAbbrev(BitCodeAbbrev *Abbv) {
474     // Emit the abbreviation as a record.
475     EncodeAbbrev(Abbv);
476     CurAbbrevs.push_back(Abbv);
477     return static_cast<unsigned>(CurAbbrevs.size())-1 +
478       bitc::FIRST_APPLICATION_ABBREV;
479   }
480 
481   //===--------------------------------------------------------------------===//
482   // BlockInfo Block Emission
483   //===--------------------------------------------------------------------===//
484 
485   /// EnterBlockInfoBlock - Start emitting the BLOCKINFO_BLOCK.
EnterBlockInfoBlock(unsigned CodeWidth)486   void EnterBlockInfoBlock(unsigned CodeWidth) {
487     EnterSubblock(bitc::BLOCKINFO_BLOCK_ID, CodeWidth);
488     BlockInfoCurBID = ~0U;
489   }
490 private:
491   /// SwitchToBlockID - If we aren't already talking about the specified block
492   /// ID, emit a BLOCKINFO_CODE_SETBID record.
SwitchToBlockID(unsigned BlockID)493   void SwitchToBlockID(unsigned BlockID) {
494     if (BlockInfoCurBID == BlockID) return;
495     SmallVector<unsigned, 2> V;
496     V.push_back(BlockID);
497     EmitRecord(bitc::BLOCKINFO_CODE_SETBID, V);
498     BlockInfoCurBID = BlockID;
499   }
500 
getOrCreateBlockInfo(unsigned BlockID)501   BlockInfo &getOrCreateBlockInfo(unsigned BlockID) {
502     if (BlockInfo *BI = getBlockInfo(BlockID))
503       return *BI;
504 
505     // Otherwise, add a new record.
506     BlockInfoRecords.push_back(BlockInfo());
507     BlockInfoRecords.back().BlockID = BlockID;
508     return BlockInfoRecords.back();
509   }
510 
511 public:
512 
513   /// EmitBlockInfoAbbrev - Emit a DEFINE_ABBREV record for the specified
514   /// BlockID.
EmitBlockInfoAbbrev(unsigned BlockID,BitCodeAbbrev * Abbv)515   unsigned EmitBlockInfoAbbrev(unsigned BlockID, BitCodeAbbrev *Abbv) {
516     SwitchToBlockID(BlockID);
517     EncodeAbbrev(Abbv);
518 
519     // Add the abbrev to the specified block record.
520     BlockInfo &Info = getOrCreateBlockInfo(BlockID);
521     Info.Abbrevs.push_back(Abbv);
522 
523     return Info.Abbrevs.size()-1+bitc::FIRST_APPLICATION_ABBREV;
524   }
525 };
526 
527 
528 } // End llvm namespace
529 
530 #endif
531