1 //===- llvm/CodeGen/DwarfAccelTable.cpp - Dwarf Accelerator Tables --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains support for writing dwarf accelerator tables.
12 //===----------------------------------------------------------------------===//
14 #include "DwarfAccelTable.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/StringMap.h"
17 #include "llvm/ADT/Twine.h"
18 #include "llvm/BinaryFormat/Dwarf.h"
19 #include "llvm/CodeGen/AsmPrinter.h"
20 #include "llvm/CodeGen/DIE.h"
21 #include "llvm/MC/MCExpr.h"
22 #include "llvm/MC/MCStreamer.h"
23 #include "llvm/Support/raw_ostream.h"
34 // The length of the header data is always going to be 4 + 4 + 4*NumAtoms.
35 DwarfAccelTable::DwarfAccelTable(ArrayRef<DwarfAccelTable::Atom> atomList)
36 : Header(8 + (atomList.size() * 4)), HeaderData(atomList),
39 void DwarfAccelTable::AddName(DwarfStringPoolEntryRef Name, const DIE *die,
41 assert(Data.empty() && "Already finalized!");
42 // If the string is in the list already then add this die to the list
43 // otherwise add a new one.
44 DataArray &DIEs = Entries[Name.getString()];
45 assert(!DIEs.Name || DIEs.Name == Name);
47 DIEs.Values.push_back(new (Allocator) HashDataContents(die, Flags));
50 void DwarfAccelTable::ComputeBucketCount() {
51 // First get the number of unique hashes.
52 std::vector<uint32_t> uniques(Data.size());
53 for (size_t i = 0, e = Data.size(); i < e; ++i)
54 uniques[i] = Data[i]->HashValue;
55 array_pod_sort(uniques.begin(), uniques.end());
56 std::vector<uint32_t>::iterator p =
57 std::unique(uniques.begin(), uniques.end());
58 uint32_t num = std::distance(uniques.begin(), p);
60 // Then compute the bucket size, minimum of 1 bucket.
62 Header.bucket_count = num / 4;
64 Header.bucket_count = num / 2;
66 Header.bucket_count = num > 0 ? num : 1;
68 Header.hashes_count = num;
71 // compareDIEs - comparison predicate that sorts DIEs by their offset.
72 static bool compareDIEs(const DwarfAccelTable::HashDataContents *A,
73 const DwarfAccelTable::HashDataContents *B) {
74 return A->Die->getOffset() < B->Die->getOffset();
77 void DwarfAccelTable::FinalizeTable(AsmPrinter *Asm, StringRef Prefix) {
78 // Create the individual hash data outputs.
79 Data.reserve(Entries.size());
80 for (StringMap<DataArray>::iterator EI = Entries.begin(), EE = Entries.end();
83 // Unique the entries.
84 std::stable_sort(EI->second.Values.begin(), EI->second.Values.end(), compareDIEs);
85 EI->second.Values.erase(
86 std::unique(EI->second.Values.begin(), EI->second.Values.end()),
87 EI->second.Values.end());
89 HashData *Entry = new (Allocator) HashData(EI->getKey(), EI->second);
90 Data.push_back(Entry);
93 // Figure out how many buckets we need, then compute the bucket
94 // contents and the final ordering. We'll emit the hashes and offsets
95 // by doing a walk during the emission phase. We add temporary
96 // symbols to the data so that we can reference them during the offset
97 // later, we'll emit them when we emit the data.
100 // Compute bucket contents and final ordering.
101 Buckets.resize(Header.bucket_count);
102 for (size_t i = 0, e = Data.size(); i < e; ++i) {
103 uint32_t bucket = Data[i]->HashValue % Header.bucket_count;
104 Buckets[bucket].push_back(Data[i]);
105 Data[i]->Sym = Asm->createTempSymbol(Prefix);
108 // Sort the contents of the buckets by hash value so that hash
109 // collisions end up together. Stable sort makes testing easier and
110 // doesn't cost much more.
111 for (size_t i = 0; i < Buckets.size(); ++i)
112 std::stable_sort(Buckets[i].begin(), Buckets[i].end(),
113 [] (HashData *LHS, HashData *RHS) {
114 return LHS->HashValue < RHS->HashValue;
118 // Emits the header for the table via the AsmPrinter.
119 void DwarfAccelTable::EmitHeader(AsmPrinter *Asm) {
120 Asm->OutStreamer->AddComment("Header Magic");
121 Asm->EmitInt32(Header.magic);
122 Asm->OutStreamer->AddComment("Header Version");
123 Asm->EmitInt16(Header.version);
124 Asm->OutStreamer->AddComment("Header Hash Function");
125 Asm->EmitInt16(Header.hash_function);
126 Asm->OutStreamer->AddComment("Header Bucket Count");
127 Asm->EmitInt32(Header.bucket_count);
128 Asm->OutStreamer->AddComment("Header Hash Count");
129 Asm->EmitInt32(Header.hashes_count);
130 Asm->OutStreamer->AddComment("Header Data Length");
131 Asm->EmitInt32(Header.header_data_len);
132 Asm->OutStreamer->AddComment("HeaderData Die Offset Base");
133 Asm->EmitInt32(HeaderData.die_offset_base);
134 Asm->OutStreamer->AddComment("HeaderData Atom Count");
135 Asm->EmitInt32(HeaderData.Atoms.size());
136 for (size_t i = 0; i < HeaderData.Atoms.size(); i++) {
137 Atom A = HeaderData.Atoms[i];
138 Asm->OutStreamer->AddComment(dwarf::AtomTypeString(A.type));
139 Asm->EmitInt16(A.type);
140 Asm->OutStreamer->AddComment(dwarf::FormEncodingString(A.form));
141 Asm->EmitInt16(A.form);
145 // Walk through and emit the buckets for the table. Each index is
146 // an offset into the list of hashes.
147 void DwarfAccelTable::EmitBuckets(AsmPrinter *Asm) {
149 for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
150 Asm->OutStreamer->AddComment("Bucket " + Twine(i));
151 if (!Buckets[i].empty())
152 Asm->EmitInt32(index);
154 Asm->EmitInt32(std::numeric_limits<uint32_t>::max());
155 // Buckets point in the list of hashes, not to the data. Do not
156 // increment the index multiple times in case of hash collisions.
157 uint64_t PrevHash = std::numeric_limits<uint64_t>::max();
158 for (auto *HD : Buckets[i]) {
159 uint32_t HashValue = HD->HashValue;
160 if (PrevHash != HashValue)
162 PrevHash = HashValue;
167 // Walk through the buckets and emit the individual hashes for each
169 void DwarfAccelTable::EmitHashes(AsmPrinter *Asm) {
170 uint64_t PrevHash = std::numeric_limits<uint64_t>::max();
171 for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
172 for (HashList::const_iterator HI = Buckets[i].begin(),
173 HE = Buckets[i].end();
175 uint32_t HashValue = (*HI)->HashValue;
176 if (PrevHash == HashValue)
178 Asm->OutStreamer->AddComment("Hash in Bucket " + Twine(i));
179 Asm->EmitInt32(HashValue);
180 PrevHash = HashValue;
185 // Walk through the buckets and emit the individual offsets for each
186 // element in each bucket. This is done via a symbol subtraction from the
187 // beginning of the section. The non-section symbol will be output later
188 // when we emit the actual data.
189 void DwarfAccelTable::emitOffsets(AsmPrinter *Asm, const MCSymbol *SecBegin) {
190 uint64_t PrevHash = std::numeric_limits<uint64_t>::max();
191 for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
192 for (HashList::const_iterator HI = Buckets[i].begin(),
193 HE = Buckets[i].end();
195 uint32_t HashValue = (*HI)->HashValue;
196 if (PrevHash == HashValue)
198 PrevHash = HashValue;
199 Asm->OutStreamer->AddComment("Offset in Bucket " + Twine(i));
200 MCContext &Context = Asm->OutStreamer->getContext();
201 const MCExpr *Sub = MCBinaryExpr::createSub(
202 MCSymbolRefExpr::create((*HI)->Sym, Context),
203 MCSymbolRefExpr::create(SecBegin, Context), Context);
204 Asm->OutStreamer->EmitValue(Sub, sizeof(uint32_t));
209 // Walk through the buckets and emit the full data for each element in
210 // the bucket. For the string case emit the dies and the various offsets.
211 // Terminate each HashData bucket with 0.
212 void DwarfAccelTable::EmitData(AsmPrinter *Asm, DwarfDebug *D) {
213 for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
214 uint64_t PrevHash = std::numeric_limits<uint64_t>::max();
215 for (HashList::const_iterator HI = Buckets[i].begin(),
216 HE = Buckets[i].end();
218 // Terminate the previous entry if there is no hash collision
219 // with the current one.
220 if (PrevHash != std::numeric_limits<uint64_t>::max() &&
221 PrevHash != (*HI)->HashValue)
223 // Remember to emit the label for our offset.
224 Asm->OutStreamer->EmitLabel((*HI)->Sym);
225 Asm->OutStreamer->AddComment((*HI)->Str);
226 Asm->emitDwarfStringOffset((*HI)->Data.Name);
227 Asm->OutStreamer->AddComment("Num DIEs");
228 Asm->EmitInt32((*HI)->Data.Values.size());
229 for (HashDataContents *HD : (*HI)->Data.Values) {
230 // Emit the DIE offset
231 Asm->EmitInt32(HD->Die->getDebugSectionOffset());
232 // If we have multiple Atoms emit that info too.
233 // FIXME: A bit of a hack, we either emit only one atom or all info.
234 if (HeaderData.Atoms.size() > 1) {
235 Asm->EmitInt16(HD->Die->getTag());
236 Asm->EmitInt8(HD->Flags);
239 PrevHash = (*HI)->HashValue;
241 // Emit the final end marker for the bucket.
242 if (!Buckets[i].empty())
247 // Emit the entire data structure to the output file.
248 void DwarfAccelTable::emit(AsmPrinter *Asm, const MCSymbol *SecBegin,
260 emitOffsets(Asm, SecBegin);
262 // Emit the hash data.
267 void DwarfAccelTable::print(raw_ostream &OS) {
269 HeaderData.print(OS);
272 for (StringMap<DataArray>::const_iterator EI = Entries.begin(),
275 OS << "Name: " << EI->getKeyData() << "\n";
276 for (HashDataContents *HD : EI->second.Values)
280 OS << "Buckets and Hashes: \n";
281 for (size_t i = 0, e = Buckets.size(); i < e; ++i)
282 for (HashList::const_iterator HI = Buckets[i].begin(),
283 HE = Buckets[i].end();
288 for (std::vector<HashData *>::const_iterator DI = Data.begin(),