1//===- DbiStreamBuilder.cpp - PDB Dbi Stream Creation -----------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// The data structures defined in this file are based on the reference
10// implementation which is available at
11// https://github.com/Microsoft/microsoft-pdb/blob/master/PDB/dbi/gsi.cpp
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/DebugInfo/PDB/Native/GSIStreamBuilder.h"
16#include "llvm/DebugInfo/CodeView/RecordName.h"
17#include "llvm/DebugInfo/CodeView/RecordSerialization.h"
18#include "llvm/DebugInfo/CodeView/SymbolRecord.h"
19#include "llvm/DebugInfo/CodeView/SymbolSerializer.h"
20#include "llvm/DebugInfo/MSF/MSFBuilder.h"
21#include "llvm/DebugInfo/MSF/MSFCommon.h"
22#include "llvm/DebugInfo/MSF/MappedBlockStream.h"
23#include "llvm/DebugInfo/PDB/Native/GlobalsStream.h"
24#include "llvm/DebugInfo/PDB/Native/Hash.h"
25#include "llvm/DebugInfo/PDB/Native/RawTypes.h"
26#include "llvm/Support/BinaryItemStream.h"
27#include "llvm/Support/BinaryStreamWriter.h"
28#include "llvm/Support/FormatVariadic.h"
29#include "llvm/Support/Parallel.h"
30#include "llvm/Support/TimeProfiler.h"
31#include <algorithm>
32#include <vector>
33
34using namespace llvm;
35using namespace llvm::msf;
36using namespace llvm::pdb;
37using namespace llvm::codeview;
38
39// Helper class for building the public and global PDB hash table buckets.
40struct llvm::pdb::GSIHashStreamBuilder {
41 // Sum of the size of all public or global records.
42 uint64_t RecordByteSize = 0;
43
44 std::vector<PSHashRecord> HashRecords;
45
46 // The hash bitmap has `ceil((IPHR_HASH + 1) / 32)` words in it. The
47 // reference implementation builds a hash table with IPHR_HASH buckets in it.
48 // The last bucket is used to link together free hash table cells in a linked
49 // list, but it is always empty in the compressed, on-disk format. However,
50 // the bitmap must have a bit for it.
51 std::array<support::ulittle32_t, (IPHR_HASH + 32) / 32> HashBitmap;
52
53 std::vector<support::ulittle32_t> HashBuckets;
54
55 uint32_t calculateSerializedLength() const;
56 Error commit(BinaryStreamWriter &Writer);
57
58 void finalizePublicBuckets();
59 void finalizeGlobalBuckets(uint32_t RecordZeroOffset);
60
61 // Assign public and global symbol records into hash table buckets.
62 // Modifies the list of records to store the bucket index, but does not
63 // change the order.
64 void finalizeBuckets(uint32_t RecordZeroOffset,
65 MutableArrayRef<BulkPublic> Globals);
66};
67
68// DenseMapInfo implementation for deduplicating symbol records.
69struct llvm::pdb::SymbolDenseMapInfo {
70 static unsigned getHashValue(const CVSymbol &Val) {
71 return xxh3_64bits(data: Val.RecordData);
72 }
73 static bool isEqual(const CVSymbol &LHS, const CVSymbol &RHS) {
74 return LHS.RecordData == RHS.RecordData;
75 }
76};
77
78namespace {
79LLVM_PACKED_START
80struct PublicSym32Layout {
81 RecordPrefix Prefix;
82 PublicSym32Header Pub;
83 // char Name[];
84};
85LLVM_PACKED_END
86} // namespace
87
88// Calculate how much memory this public needs when serialized.
89static uint32_t sizeOfPublic(const BulkPublic &Pub) {
90 uint32_t NameLen = Pub.NameLen;
91 NameLen = std::min(a: NameLen,
92 b: uint32_t(MaxRecordLength - sizeof(PublicSym32Layout) - 1));
93 return alignTo(Value: sizeof(PublicSym32Layout) + NameLen + 1, Align: 4);
94}
95
96static CVSymbol serializePublic(uint8_t *Mem, const BulkPublic &Pub) {
97 // Assume the caller has allocated sizeOfPublic bytes.
98 uint32_t NameLen = std::min(
99 a: Pub.NameLen, b: uint32_t(MaxRecordLength - sizeof(PublicSym32Layout) - 1));
100 size_t Size = alignTo(Value: sizeof(PublicSym32Layout) + NameLen + 1, Align: 4);
101 assert(Size == sizeOfPublic(Pub));
102 auto *FixedMem = reinterpret_cast<PublicSym32Layout *>(Mem);
103 FixedMem->Prefix.RecordKind = static_cast<uint16_t>(codeview::S_PUB32);
104 FixedMem->Prefix.RecordLen = static_cast<uint16_t>(Size - 2);
105 FixedMem->Pub.Flags = Pub.Flags;
106 FixedMem->Pub.Offset = Pub.Offset;
107 FixedMem->Pub.Segment = Pub.Segment;
108 char *NameMem = reinterpret_cast<char *>(FixedMem + 1);
109 memcpy(dest: NameMem, src: Pub.Name, n: NameLen);
110 // Zero the null terminator and remaining bytes.
111 memset(s: &NameMem[NameLen], c: 0, n: Size - sizeof(PublicSym32Layout) - NameLen);
112 return CVSymbol(ArrayRef(Mem, Size));
113}
114
115uint32_t GSIHashStreamBuilder::calculateSerializedLength() const {
116 uint32_t Size = sizeof(GSIHashHeader);
117 Size += HashRecords.size() * sizeof(PSHashRecord);
118 Size += HashBitmap.size() * sizeof(uint32_t);
119 Size += HashBuckets.size() * sizeof(uint32_t);
120 return Size;
121}
122
123Error GSIHashStreamBuilder::commit(BinaryStreamWriter &Writer) {
124 GSIHashHeader Header;
125 Header.VerSignature = GSIHashHeader::HdrSignature;
126 Header.VerHdr = GSIHashHeader::HdrVersion;
127 Header.HrSize = HashRecords.size() * sizeof(PSHashRecord);
128 Header.NumBuckets = HashBitmap.size() * 4 + HashBuckets.size() * 4;
129
130 if (auto EC = Writer.writeObject(Obj: Header))
131 return EC;
132
133 if (auto EC = Writer.writeArray(Array: ArrayRef(HashRecords)))
134 return EC;
135 if (auto EC = Writer.writeArray(Array: ArrayRef(HashBitmap)))
136 return EC;
137 if (auto EC = Writer.writeArray(Array: ArrayRef(HashBuckets)))
138 return EC;
139 return Error::success();
140}
141
142static bool isAsciiString(StringRef S) {
143 return llvm::all_of(Range&: S, P: [](char C) { return unsigned(C) < 0x80; });
144}
145
146// See `caseInsensitiveComparePchPchCchCch` in gsi.cpp
147static int gsiRecordCmp(StringRef S1, StringRef S2) {
148 size_t LS = S1.size();
149 size_t RS = S2.size();
150 // Shorter strings always compare less than longer strings.
151 if (LS != RS)
152 return (LS > RS) - (LS < RS);
153
154 // If either string contains non ascii characters, memcmp them.
155 if (LLVM_UNLIKELY(!isAsciiString(S1) || !isAsciiString(S2)))
156 return memcmp(s1: S1.data(), s2: S2.data(), n: LS);
157
158 // Both strings are ascii, perform a case-insensitive comparison.
159 return S1.compare_insensitive(RHS: S2);
160}
161
162void GSIStreamBuilder::finalizePublicBuckets() {
163 PSH->finalizeBuckets(RecordZeroOffset: 0, Globals: Publics);
164}
165
166void GSIStreamBuilder::finalizeGlobalBuckets(uint32_t RecordZeroOffset) {
167 // Build up a list of globals to be bucketed. Use the BulkPublic data
168 // structure for this purpose, even though these are global records, not
169 // public records. Most of the same fields are required:
170 // - Name
171 // - NameLen
172 // - SymOffset
173 // - BucketIdx
174 // The dead fields are Offset, Segment, and Flags.
175 std::vector<BulkPublic> Records;
176 Records.resize(new_size: Globals.size());
177 uint32_t SymOffset = RecordZeroOffset;
178 for (size_t I = 0, E = Globals.size(); I < E; ++I) {
179 StringRef Name = getSymbolName(Sym: Globals[I]);
180 Records[I].Name = Name.data();
181 Records[I].NameLen = Name.size();
182 Records[I].SymOffset = SymOffset;
183 SymOffset += Globals[I].length();
184 }
185
186 GSH->finalizeBuckets(RecordZeroOffset, Globals: Records);
187}
188
189void GSIHashStreamBuilder::finalizeBuckets(
190 uint32_t RecordZeroOffset, MutableArrayRef<BulkPublic> Records) {
191 // Hash every name in parallel.
192 parallelFor(Begin: 0, End: Records.size(), Fn: [&](size_t I) {
193 Records[I].setBucketIdx(hashStringV1(Str: Records[I].getName()) % IPHR_HASH);
194 });
195
196 // Count up the size of each bucket. Then, use an exclusive prefix sum to
197 // calculate the bucket start offsets. This is C++17 std::exclusive_scan, but
198 // we can't use it yet.
199 uint32_t BucketStarts[IPHR_HASH] = {0};
200 for (const BulkPublic &P : Records)
201 ++BucketStarts[P.BucketIdx];
202 uint32_t Sum = 0;
203 for (uint32_t &B : BucketStarts) {
204 uint32_t Size = B;
205 B = Sum;
206 Sum += Size;
207 }
208
209 // Place globals into the hash table in bucket order. When placing a global,
210 // update the bucket start. Every hash table slot should be filled. Always use
211 // a refcount of one for now.
212 HashRecords.resize(new_size: Records.size());
213 uint32_t BucketCursors[IPHR_HASH];
214 memcpy(dest: BucketCursors, src: BucketStarts, n: sizeof(BucketCursors));
215 for (int I = 0, E = Records.size(); I < E; ++I) {
216 uint32_t HashIdx = BucketCursors[Records[I].BucketIdx]++;
217 HashRecords[HashIdx].Off = I;
218 HashRecords[HashIdx].CRef = 1;
219 }
220
221 // Within the buckets, sort each bucket by memcmp of the symbol's name. It's
222 // important that we use the same sorting algorithm as is used by the
223 // reference implementation to ensure that the search for a record within a
224 // bucket can properly early-out when it detects the record won't be found.
225 // The algorithm used here corresponds to the function
226 // caseInsensitiveComparePchPchCchCch in the reference implementation.
227 parallelFor(Begin: 0, End: IPHR_HASH, Fn: [&](size_t I) {
228 auto B = HashRecords.begin() + BucketStarts[I];
229 auto E = HashRecords.begin() + BucketCursors[I];
230 if (B == E)
231 return;
232 auto BucketCmp = [Records](const PSHashRecord &LHash,
233 const PSHashRecord &RHash) {
234 const BulkPublic &L = Records[uint32_t(LHash.Off)];
235 const BulkPublic &R = Records[uint32_t(RHash.Off)];
236 assert(L.BucketIdx == R.BucketIdx);
237 int Cmp = gsiRecordCmp(S1: L.getName(), S2: R.getName());
238 if (Cmp != 0)
239 return Cmp < 0;
240 // This comparison is necessary to make the sorting stable in the presence
241 // of two static globals with the same name. The easiest way to observe
242 // this is with S_LDATA32 records.
243 return L.SymOffset < R.SymOffset;
244 };
245 llvm::sort(Start: B, End: E, Comp: BucketCmp);
246
247 // After we are done sorting, replace the global indices with the stream
248 // offsets of each global. Add one when writing symbol offsets to disk.
249 // See GSI1::fixSymRecs.
250 for (PSHashRecord &HRec : make_range(x: B, y: E))
251 HRec.Off = Records[uint32_t(HRec.Off)].SymOffset + 1;
252 });
253
254 // For each non-empty bucket, push the bucket start offset into HashBuckets
255 // and set a bit in the hash bitmap.
256 for (uint32_t I = 0; I < HashBitmap.size(); ++I) {
257 uint32_t Word = 0;
258 for (uint32_t J = 0; J < 32; ++J) {
259 // Skip empty buckets.
260 uint32_t BucketIdx = I * 32 + J;
261 if (BucketIdx >= IPHR_HASH ||
262 BucketStarts[BucketIdx] == BucketCursors[BucketIdx])
263 continue;
264 Word |= (1U << J);
265
266 // Calculate what the offset of the first hash record in the chain would
267 // be if it were inflated to contain 32-bit pointers. On a 32-bit system,
268 // each record would be 12 bytes. See HROffsetCalc in gsi.h.
269 const int SizeOfHROffsetCalc = 12;
270 ulittle32_t ChainStartOff =
271 ulittle32_t(BucketStarts[BucketIdx] * SizeOfHROffsetCalc);
272 HashBuckets.push_back(x: ChainStartOff);
273 }
274 HashBitmap[I] = Word;
275 }
276}
277
278GSIStreamBuilder::GSIStreamBuilder(msf::MSFBuilder &Msf)
279 : Msf(Msf), PSH(std::make_unique<GSIHashStreamBuilder>()),
280 GSH(std::make_unique<GSIHashStreamBuilder>()) {}
281
282GSIStreamBuilder::~GSIStreamBuilder() = default;
283
284uint32_t GSIStreamBuilder::calculatePublicsHashStreamSize() const {
285 uint32_t Size = 0;
286 Size += sizeof(PublicsStreamHeader);
287 Size += PSH->calculateSerializedLength();
288 Size += Publics.size() * sizeof(uint32_t); // AddrMap
289 // FIXME: Add thunk map and section offsets for incremental linking.
290
291 return Size;
292}
293
294uint32_t GSIStreamBuilder::calculateGlobalsHashStreamSize() const {
295 return GSH->calculateSerializedLength();
296}
297
298Error GSIStreamBuilder::finalizeMsfLayout() {
299 // First we write public symbol records, then we write global symbol records.
300 finalizePublicBuckets();
301 finalizeGlobalBuckets(RecordZeroOffset: PSH->RecordByteSize);
302
303 Expected<uint32_t> Idx = Msf.addStream(Size: calculateGlobalsHashStreamSize());
304 if (!Idx)
305 return Idx.takeError();
306 GlobalsStreamIndex = *Idx;
307
308 Idx = Msf.addStream(Size: calculatePublicsHashStreamSize());
309 if (!Idx)
310 return Idx.takeError();
311 PublicsStreamIndex = *Idx;
312
313 uint64_t RecordBytes = PSH->RecordByteSize + GSH->RecordByteSize;
314 if (RecordBytes > UINT32_MAX)
315 return make_error<StringError>(
316 Args: formatv(Fmt: "the public symbols ({0} bytes) and global symbols ({1} bytes) "
317 "are too large to fit in a PDB file; "
318 "the maximum total is {2} bytes.",
319 Vals&: PSH->RecordByteSize, Vals&: GSH->RecordByteSize, UINT32_MAX),
320 Args: inconvertibleErrorCode());
321
322 Idx = Msf.addStream(Size: RecordBytes);
323 if (!Idx)
324 return Idx.takeError();
325 RecordStreamIndex = *Idx;
326 return Error::success();
327}
328
329void GSIStreamBuilder::addPublicSymbols(std::vector<BulkPublic> &&PublicsIn) {
330 assert(Publics.empty() && PSH->RecordByteSize == 0 &&
331 "publics can only be added once");
332 Publics = std::move(PublicsIn);
333
334 // Sort the symbols by name. PDBs contain lots of symbols, so use parallelism.
335 parallelSort(R&: Publics, Comp: [](const BulkPublic &L, const BulkPublic &R) {
336 return L.getName() < R.getName();
337 });
338
339 // Assign offsets and calculate the length of the public symbol records.
340 uint32_t SymOffset = 0;
341 for (BulkPublic &Pub : Publics) {
342 Pub.SymOffset = SymOffset;
343 SymOffset += sizeOfPublic(Pub);
344 }
345
346 // Remember the length of the public stream records.
347 PSH->RecordByteSize = SymOffset;
348}
349
350void GSIStreamBuilder::addGlobalSymbol(const ProcRefSym &Sym) {
351 serializeAndAddGlobal(Symbol: Sym);
352}
353
354void GSIStreamBuilder::addGlobalSymbol(const DataSym &Sym) {
355 serializeAndAddGlobal(Symbol: Sym);
356}
357
358void GSIStreamBuilder::addGlobalSymbol(const ConstantSym &Sym) {
359 serializeAndAddGlobal(Symbol: Sym);
360}
361
362template <typename T>
363void GSIStreamBuilder::serializeAndAddGlobal(const T &Symbol) {
364 T Copy(Symbol);
365 addGlobalSymbol(SymbolSerializer::writeOneSymbol(Copy, Msf.getAllocator(),
366 CodeViewContainer::Pdb));
367}
368
369void GSIStreamBuilder::addGlobalSymbol(const codeview::CVSymbol &Symbol) {
370 // Ignore duplicate typedefs and constants.
371 if (Symbol.kind() == S_UDT || Symbol.kind() == S_CONSTANT) {
372 auto Iter = GlobalsSeen.insert(V: Symbol);
373 if (!Iter.second)
374 return;
375 }
376 GSH->RecordByteSize += Symbol.length();
377 Globals.push_back(x: Symbol);
378}
379
380// Serialize each public and write it.
381static Error writePublics(BinaryStreamWriter &Writer,
382 ArrayRef<BulkPublic> Publics) {
383 std::vector<uint8_t> Storage;
384 for (const BulkPublic &Pub : Publics) {
385 Storage.resize(new_size: sizeOfPublic(Pub));
386 serializePublic(Mem: Storage.data(), Pub);
387 if (Error E = Writer.writeBytes(Buffer: Storage))
388 return E;
389 }
390 return Error::success();
391}
392
393static Error writeRecords(BinaryStreamWriter &Writer,
394 ArrayRef<CVSymbol> Records) {
395 BinaryItemStream<CVSymbol> ItemStream(llvm::endianness::little);
396 ItemStream.setItems(Records);
397 BinaryStreamRef RecordsRef(ItemStream);
398 return Writer.writeStreamRef(Ref: RecordsRef);
399}
400
401Error GSIStreamBuilder::commitSymbolRecordStream(
402 WritableBinaryStreamRef Stream) {
403 BinaryStreamWriter Writer(Stream);
404
405 // Write public symbol records first, followed by global symbol records. This
406 // must match the order that we assume in finalizeMsfLayout when computing
407 // PSHZero and GSHZero.
408 if (auto EC = writePublics(Writer, Publics))
409 return EC;
410 if (auto EC = writeRecords(Writer, Records: Globals))
411 return EC;
412
413 return Error::success();
414}
415
416static std::vector<support::ulittle32_t>
417computeAddrMap(ArrayRef<BulkPublic> Publics) {
418 // Build a parallel vector of indices into the Publics vector, and sort it by
419 // address.
420 std::vector<ulittle32_t> PubAddrMap;
421 PubAddrMap.reserve(n: Publics.size());
422 for (int I = 0, E = Publics.size(); I < E; ++I)
423 PubAddrMap.push_back(x: ulittle32_t(I));
424
425 auto AddrCmp = [Publics](const ulittle32_t &LIdx, const ulittle32_t &RIdx) {
426 const BulkPublic &L = Publics[LIdx];
427 const BulkPublic &R = Publics[RIdx];
428 if (L.Segment != R.Segment)
429 return L.Segment < R.Segment;
430 if (L.Offset != R.Offset)
431 return L.Offset < R.Offset;
432 // parallelSort is unstable, so we have to do name comparison to ensure
433 // that two names for the same location come out in a deterministic order.
434 return L.getName() < R.getName();
435 };
436 parallelSort(R&: PubAddrMap, Comp: AddrCmp);
437
438 // Rewrite the public symbol indices into symbol offsets.
439 for (ulittle32_t &Entry : PubAddrMap)
440 Entry = Publics[Entry].SymOffset;
441 return PubAddrMap;
442}
443
444Error GSIStreamBuilder::commitPublicsHashStream(
445 WritableBinaryStreamRef Stream) {
446 BinaryStreamWriter Writer(Stream);
447 PublicsStreamHeader Header;
448
449 // FIXME: Fill these in. They are for incremental linking.
450 Header.SymHash = PSH->calculateSerializedLength();
451 Header.AddrMap = Publics.size() * 4;
452 Header.NumThunks = 0;
453 Header.SizeOfThunk = 0;
454 Header.ISectThunkTable = 0;
455 memset(s: Header.Padding, c: 0, n: sizeof(Header.Padding));
456 Header.OffThunkTable = 0;
457 Header.NumSections = 0;
458 if (auto EC = Writer.writeObject(Obj: Header))
459 return EC;
460
461 if (auto EC = PSH->commit(Writer))
462 return EC;
463
464 std::vector<support::ulittle32_t> PubAddrMap = computeAddrMap(Publics);
465 assert(PubAddrMap.size() == Publics.size());
466 if (auto EC = Writer.writeArray(Array: ArrayRef(PubAddrMap)))
467 return EC;
468
469 return Error::success();
470}
471
472Error GSIStreamBuilder::commitGlobalsHashStream(
473 WritableBinaryStreamRef Stream) {
474 BinaryStreamWriter Writer(Stream);
475 return GSH->commit(Writer);
476}
477
478Error GSIStreamBuilder::commit(const msf::MSFLayout &Layout,
479 WritableBinaryStreamRef Buffer) {
480 llvm::TimeTraceScope timeScope("Commit GSI stream");
481 auto GS = WritableMappedBlockStream::createIndexedStream(
482 Layout, MsfData: Buffer, StreamIndex: getGlobalsStreamIndex(), Allocator&: Msf.getAllocator());
483 auto PS = WritableMappedBlockStream::createIndexedStream(
484 Layout, MsfData: Buffer, StreamIndex: getPublicsStreamIndex(), Allocator&: Msf.getAllocator());
485 auto PRS = WritableMappedBlockStream::createIndexedStream(
486 Layout, MsfData: Buffer, StreamIndex: getRecordStreamIndex(), Allocator&: Msf.getAllocator());
487
488 if (auto EC = commitSymbolRecordStream(Stream: *PRS))
489 return EC;
490 if (auto EC = commitGlobalsHashStream(Stream: *GS))
491 return EC;
492 if (auto EC = commitPublicsHashStream(Stream: *PS))
493 return EC;
494 return Error::success();
495}
496