1 //=-- llvm/CodeGen/DwarfAccelTable.cpp - Dwarf Accelerator Tables -*- 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 file contains support for writing dwarf accelerator tables.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "DwarfAccelTable.h"
15 #include "DwarfCompileUnit.h"
16 #include "DwarfDebug.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/Twine.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/MC/MCSymbol.h"
24 #include "llvm/Support/Debug.h"
25 
26 using namespace llvm;
27 
28 // The length of the header data is always going to be 4 + 4 + 4*NumAtoms.
DwarfAccelTable(ArrayRef<DwarfAccelTable::Atom> atomList)29 DwarfAccelTable::DwarfAccelTable(ArrayRef<DwarfAccelTable::Atom> atomList)
30     : Header(8 + (atomList.size() * 4)), HeaderData(atomList),
31       Entries(Allocator) {}
32 
AddName(StringRef Name,MCSymbol * StrSym,const DIE * die,char Flags)33 void DwarfAccelTable::AddName(StringRef Name, MCSymbol *StrSym, const DIE *die,
34                               char Flags) {
35   assert(Data.empty() && "Already finalized!");
36   // If the string is in the list already then add this die to the list
37   // otherwise add a new one.
38   DataArray &DIEs = Entries[Name];
39   assert(!DIEs.StrSym || DIEs.StrSym == StrSym);
40   DIEs.StrSym = StrSym;
41   DIEs.Values.push_back(new (Allocator) HashDataContents(die, Flags));
42 }
43 
ComputeBucketCount(void)44 void DwarfAccelTable::ComputeBucketCount(void) {
45   // First get the number of unique hashes.
46   std::vector<uint32_t> uniques(Data.size());
47   for (size_t i = 0, e = Data.size(); i < e; ++i)
48     uniques[i] = Data[i]->HashValue;
49   array_pod_sort(uniques.begin(), uniques.end());
50   std::vector<uint32_t>::iterator p =
51       std::unique(uniques.begin(), uniques.end());
52   uint32_t num = std::distance(uniques.begin(), p);
53 
54   // Then compute the bucket size, minimum of 1 bucket.
55   if (num > 1024)
56     Header.bucket_count = num / 4;
57   else if (num > 16)
58     Header.bucket_count = num / 2;
59   else
60     Header.bucket_count = num > 0 ? num : 1;
61 
62   Header.hashes_count = num;
63 }
64 
65 // compareDIEs - comparison predicate that sorts DIEs by their offset.
compareDIEs(const DwarfAccelTable::HashDataContents * A,const DwarfAccelTable::HashDataContents * B)66 static bool compareDIEs(const DwarfAccelTable::HashDataContents *A,
67                         const DwarfAccelTable::HashDataContents *B) {
68   return A->Die->getOffset() < B->Die->getOffset();
69 }
70 
FinalizeTable(AsmPrinter * Asm,StringRef Prefix)71 void DwarfAccelTable::FinalizeTable(AsmPrinter *Asm, StringRef Prefix) {
72   // Create the individual hash data outputs.
73   Data.reserve(Entries.size());
74   for (StringMap<DataArray>::iterator EI = Entries.begin(), EE = Entries.end();
75        EI != EE; ++EI) {
76 
77     // Unique the entries.
78     std::stable_sort(EI->second.Values.begin(), EI->second.Values.end(), compareDIEs);
79     EI->second.Values.erase(
80         std::unique(EI->second.Values.begin(), EI->second.Values.end()),
81         EI->second.Values.end());
82 
83     HashData *Entry = new (Allocator) HashData(EI->getKey(), EI->second);
84     Data.push_back(Entry);
85   }
86 
87   // Figure out how many buckets we need, then compute the bucket
88   // contents and the final ordering. We'll emit the hashes and offsets
89   // by doing a walk during the emission phase. We add temporary
90   // symbols to the data so that we can reference them during the offset
91   // later, we'll emit them when we emit the data.
92   ComputeBucketCount();
93 
94   // Compute bucket contents and final ordering.
95   Buckets.resize(Header.bucket_count);
96   for (size_t i = 0, e = Data.size(); i < e; ++i) {
97     uint32_t bucket = Data[i]->HashValue % Header.bucket_count;
98     Buckets[bucket].push_back(Data[i]);
99     Data[i]->Sym = Asm->createTempSymbol(Prefix);
100   }
101 
102   // Sort the contents of the buckets by hash value so that hash
103   // collisions end up together. Stable sort makes testing easier and
104   // doesn't cost much more.
105   for (size_t i = 0; i < Buckets.size(); ++i)
106     std::stable_sort(Buckets[i].begin(), Buckets[i].end(),
107                      [] (HashData *LHS, HashData *RHS) {
108                        return LHS->HashValue < RHS->HashValue;
109                      });
110 }
111 
112 // Emits the header for the table via the AsmPrinter.
EmitHeader(AsmPrinter * Asm)113 void DwarfAccelTable::EmitHeader(AsmPrinter *Asm) {
114   Asm->OutStreamer.AddComment("Header Magic");
115   Asm->EmitInt32(Header.magic);
116   Asm->OutStreamer.AddComment("Header Version");
117   Asm->EmitInt16(Header.version);
118   Asm->OutStreamer.AddComment("Header Hash Function");
119   Asm->EmitInt16(Header.hash_function);
120   Asm->OutStreamer.AddComment("Header Bucket Count");
121   Asm->EmitInt32(Header.bucket_count);
122   Asm->OutStreamer.AddComment("Header Hash Count");
123   Asm->EmitInt32(Header.hashes_count);
124   Asm->OutStreamer.AddComment("Header Data Length");
125   Asm->EmitInt32(Header.header_data_len);
126   Asm->OutStreamer.AddComment("HeaderData Die Offset Base");
127   Asm->EmitInt32(HeaderData.die_offset_base);
128   Asm->OutStreamer.AddComment("HeaderData Atom Count");
129   Asm->EmitInt32(HeaderData.Atoms.size());
130   for (size_t i = 0; i < HeaderData.Atoms.size(); i++) {
131     Atom A = HeaderData.Atoms[i];
132     Asm->OutStreamer.AddComment(dwarf::AtomTypeString(A.type));
133     Asm->EmitInt16(A.type);
134     Asm->OutStreamer.AddComment(dwarf::FormEncodingString(A.form));
135     Asm->EmitInt16(A.form);
136   }
137 }
138 
139 // Walk through and emit the buckets for the table. Each index is
140 // an offset into the list of hashes.
EmitBuckets(AsmPrinter * Asm)141 void DwarfAccelTable::EmitBuckets(AsmPrinter *Asm) {
142   unsigned index = 0;
143   for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
144     Asm->OutStreamer.AddComment("Bucket " + Twine(i));
145     if (Buckets[i].size() != 0)
146       Asm->EmitInt32(index);
147     else
148       Asm->EmitInt32(UINT32_MAX);
149     // Buckets point in the list of hashes, not to the data. Do not
150     // increment the index multiple times in case of hash collisions.
151     uint64_t PrevHash = UINT64_MAX;
152     for (auto *HD : Buckets[i]) {
153       uint32_t HashValue = HD->HashValue;
154       if (PrevHash != HashValue)
155         ++index;
156       PrevHash = HashValue;
157     }
158   }
159 }
160 
161 // Walk through the buckets and emit the individual hashes for each
162 // bucket.
EmitHashes(AsmPrinter * Asm)163 void DwarfAccelTable::EmitHashes(AsmPrinter *Asm) {
164   uint64_t PrevHash = UINT64_MAX;
165   for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
166     for (HashList::const_iterator HI = Buckets[i].begin(),
167                                   HE = Buckets[i].end();
168          HI != HE; ++HI) {
169       uint32_t HashValue = (*HI)->HashValue;
170       if (PrevHash == HashValue)
171         continue;
172       Asm->OutStreamer.AddComment("Hash in Bucket " + Twine(i));
173       Asm->EmitInt32(HashValue);
174       PrevHash = HashValue;
175     }
176   }
177 }
178 
179 // Walk through the buckets and emit the individual offsets for each
180 // element in each bucket. This is done via a symbol subtraction from the
181 // beginning of the section. The non-section symbol will be output later
182 // when we emit the actual data.
emitOffsets(AsmPrinter * Asm,const MCSymbol * SecBegin)183 void DwarfAccelTable::emitOffsets(AsmPrinter *Asm, const MCSymbol *SecBegin) {
184   uint64_t PrevHash = UINT64_MAX;
185   for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
186     for (HashList::const_iterator HI = Buckets[i].begin(),
187                                   HE = Buckets[i].end();
188          HI != HE; ++HI) {
189       uint32_t HashValue = (*HI)->HashValue;
190       if (PrevHash == HashValue)
191         continue;
192       PrevHash = HashValue;
193       Asm->OutStreamer.AddComment("Offset in Bucket " + Twine(i));
194       MCContext &Context = Asm->OutStreamer.getContext();
195       const MCExpr *Sub = MCBinaryExpr::CreateSub(
196           MCSymbolRefExpr::Create((*HI)->Sym, Context),
197           MCSymbolRefExpr::Create(SecBegin, Context), Context);
198       Asm->OutStreamer.EmitValue(Sub, sizeof(uint32_t));
199     }
200   }
201 }
202 
203 // Walk through the buckets and emit the full data for each element in
204 // the bucket. For the string case emit the dies and the various offsets.
205 // Terminate each HashData bucket with 0.
EmitData(AsmPrinter * Asm,DwarfDebug * D)206 void DwarfAccelTable::EmitData(AsmPrinter *Asm, DwarfDebug *D) {
207   for (size_t i = 0, e = Buckets.size(); i < e; ++i) {
208     uint64_t PrevHash = UINT64_MAX;
209     for (HashList::const_iterator HI = Buckets[i].begin(),
210                                   HE = Buckets[i].end();
211          HI != HE; ++HI) {
212       // Terminate the previous entry if there is no hash collision
213       // with the current one.
214       if (PrevHash != UINT64_MAX && PrevHash != (*HI)->HashValue)
215         Asm->EmitInt32(0);
216       // Remember to emit the label for our offset.
217       Asm->OutStreamer.EmitLabel((*HI)->Sym);
218       Asm->OutStreamer.AddComment((*HI)->Str);
219       Asm->emitSectionOffset((*HI)->Data.StrSym);
220       Asm->OutStreamer.AddComment("Num DIEs");
221       Asm->EmitInt32((*HI)->Data.Values.size());
222       for (HashDataContents *HD : (*HI)->Data.Values) {
223         // Emit the DIE offset
224         DwarfCompileUnit *CU = D->lookupUnit(HD->Die->getUnit());
225         assert(CU && "Accelerated DIE should belong to a CU.");
226         Asm->EmitInt32(HD->Die->getOffset() + CU->getDebugInfoOffset());
227         // If we have multiple Atoms emit that info too.
228         // FIXME: A bit of a hack, we either emit only one atom or all info.
229         if (HeaderData.Atoms.size() > 1) {
230           Asm->EmitInt16(HD->Die->getTag());
231           Asm->EmitInt8(HD->Flags);
232         }
233       }
234       PrevHash = (*HI)->HashValue;
235     }
236     // Emit the final end marker for the bucket.
237     if (!Buckets[i].empty())
238       Asm->EmitInt32(0);
239   }
240 }
241 
242 // Emit the entire data structure to the output file.
emit(AsmPrinter * Asm,const MCSymbol * SecBegin,DwarfDebug * D)243 void DwarfAccelTable::emit(AsmPrinter *Asm, const MCSymbol *SecBegin,
244                            DwarfDebug *D) {
245   // Emit the header.
246   EmitHeader(Asm);
247 
248   // Emit the buckets.
249   EmitBuckets(Asm);
250 
251   // Emit the hashes.
252   EmitHashes(Asm);
253 
254   // Emit the offsets.
255   emitOffsets(Asm, SecBegin);
256 
257   // Emit the hash data.
258   EmitData(Asm, D);
259 }
260 
261 #ifndef NDEBUG
print(raw_ostream & O)262 void DwarfAccelTable::print(raw_ostream &O) {
263 
264   Header.print(O);
265   HeaderData.print(O);
266 
267   O << "Entries: \n";
268   for (StringMap<DataArray>::const_iterator EI = Entries.begin(),
269                                             EE = Entries.end();
270        EI != EE; ++EI) {
271     O << "Name: " << EI->getKeyData() << "\n";
272     for (HashDataContents *HD : EI->second.Values)
273       HD->print(O);
274   }
275 
276   O << "Buckets and Hashes: \n";
277   for (size_t i = 0, e = Buckets.size(); i < e; ++i)
278     for (HashList::const_iterator HI = Buckets[i].begin(),
279                                   HE = Buckets[i].end();
280          HI != HE; ++HI)
281       (*HI)->print(O);
282 
283   O << "Data: \n";
284   for (std::vector<HashData *>::const_iterator DI = Data.begin(),
285                                                DE = Data.end();
286        DI != DE; ++DI)
287     (*DI)->print(O);
288 }
289 #endif
290