RuntimeDyld cleanup:
[oota-llvm.git] / lib / ExecutionEngine / RuntimeDyld / RuntimeDyldMachO.cpp
1 //===-- RuntimeDyldMachO.cpp - Run-time dynamic linker for MC-JIT -*- 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 // Implementation of the MC-JIT runtime dynamic linker.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #define DEBUG_TYPE "dyld"
15 #include "llvm/ADT/OwningPtr.h"
16 #include "llvm/ADT/StringRef.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "RuntimeDyldMachO.h"
19 using namespace llvm;
20 using namespace llvm::object;
21
22 namespace llvm {
23
24 void RuntimeDyldMachO::resolveRelocation(uint8_t *LocalAddress,
25                                          uint64_t FinalAddress,
26                                          uint64_t Value,
27                                          uint32_t Type,
28                                          int64_t Addend) {
29   bool isPCRel = (Type >> 24) & 1;
30   unsigned MachoType = (Type >> 28) & 0xf;
31   unsigned Size = 1 << ((Type >> 25) & 3);
32
33   DEBUG(dbgs() << "resolveRelocation LocalAddress: " 
34         << format("%p", LocalAddress)
35         << " FinalAddress: " << format("%p", FinalAddress)
36         << " Value: " << format("%p", Value)
37         << " Addend: " << Addend
38         << " isPCRel: " << isPCRel
39         << " MachoType: " << MachoType
40         << " Size: " << Size
41         << "\n");
42
43   // This just dispatches to the proper target specific routine.
44   switch (Arch) {
45   default: llvm_unreachable("Unsupported CPU type!");
46   case Triple::x86_64:
47     resolveX86_64Relocation(LocalAddress,
48                             FinalAddress,
49                             (uintptr_t)Value,
50                             isPCRel,
51                             MachoType,
52                             Size,
53                             Addend);
54     break;
55   case Triple::x86:
56     resolveI386Relocation(LocalAddress,
57                           FinalAddress,
58                           (uintptr_t)Value,
59                           isPCRel,
60                           Type,
61                           Size,
62                           Addend);
63     break;
64   case Triple::arm:    // Fall through.
65   case Triple::thumb:
66     resolveARMRelocation(LocalAddress,
67                          FinalAddress,
68                          (uintptr_t)Value,
69                          isPCRel,
70                          MachoType,
71                          Size,
72                          Addend);
73     break;
74   }
75 }
76
77 bool RuntimeDyldMachO::resolveI386Relocation(uint8_t *LocalAddress,
78                                              uint64_t FinalAddress,
79                                              uint64_t Value,
80                                              bool isPCRel,
81                                              unsigned Type,
82                                              unsigned Size,
83                                              int64_t Addend) {
84   if (isPCRel)
85     Value -= FinalAddress + 4; // see resolveX86_64Relocation
86
87   switch (Type) {
88   default:
89     llvm_unreachable("Invalid relocation type!");
90   case macho::RIT_Vanilla: {
91     uint8_t *p = LocalAddress;
92     uint64_t ValueToWrite = Value + Addend;
93     for (unsigned i = 0; i < Size; ++i) {
94       *p++ = (uint8_t)(ValueToWrite & 0xff);
95       ValueToWrite >>= 8;
96     }
97   }
98   case macho::RIT_Difference:
99   case macho::RIT_Generic_LocalDifference:
100   case macho::RIT_Generic_PreboundLazyPointer:
101     return Error("Relocation type not implemented yet!");
102   }
103 }
104
105 bool RuntimeDyldMachO::resolveX86_64Relocation(uint8_t *LocalAddress,
106                                                uint64_t FinalAddress,
107                                                uint64_t Value,
108                                                bool isPCRel,
109                                                unsigned Type,
110                                                unsigned Size,
111                                                int64_t Addend) {
112   // If the relocation is PC-relative, the value to be encoded is the
113   // pointer difference.
114   if (isPCRel)
115     // FIXME: It seems this value needs to be adjusted by 4 for an effective PC
116     // address. Is that expected? Only for branches, perhaps?
117     Value -= FinalAddress + 4;
118
119   switch(Type) {
120   default:
121     llvm_unreachable("Invalid relocation type!");
122   case macho::RIT_X86_64_Signed1:
123   case macho::RIT_X86_64_Signed2:
124   case macho::RIT_X86_64_Signed4:
125   case macho::RIT_X86_64_Signed:
126   case macho::RIT_X86_64_Unsigned:
127   case macho::RIT_X86_64_Branch: {
128     Value += Addend;
129     // Mask in the target value a byte at a time (we don't have an alignment
130     // guarantee for the target address, so this is safest).
131     uint8_t *p = (uint8_t*)LocalAddress;
132     for (unsigned i = 0; i < Size; ++i) {
133       *p++ = (uint8_t)Value;
134       Value >>= 8;
135     }
136     return false;
137   }
138   case macho::RIT_X86_64_GOTLoad:
139   case macho::RIT_X86_64_GOT:
140   case macho::RIT_X86_64_Subtractor:
141   case macho::RIT_X86_64_TLV:
142     return Error("Relocation type not implemented yet!");
143   }
144 }
145
146 bool RuntimeDyldMachO::resolveARMRelocation(uint8_t *LocalAddress,
147                                             uint64_t FinalAddress,
148                                             uint64_t Value,
149                                             bool isPCRel,
150                                             unsigned Type,
151                                             unsigned Size,
152                                             int64_t Addend) {
153   // If the relocation is PC-relative, the value to be encoded is the
154   // pointer difference.
155   if (isPCRel) {
156     Value -= FinalAddress;
157     // ARM PCRel relocations have an effective-PC offset of two instructions
158     // (four bytes in Thumb mode, 8 bytes in ARM mode).
159     // FIXME: For now, assume ARM mode.
160     Value -= 8;
161   }
162
163   switch(Type) {
164   default:
165     llvm_unreachable("Invalid relocation type!");
166   case macho::RIT_Vanilla: {
167     // Mask in the target value a byte at a time (we don't have an alignment
168     // guarantee for the target address, so this is safest).
169     uint8_t *p = (uint8_t*)LocalAddress;
170     for (unsigned i = 0; i < Size; ++i) {
171       *p++ = (uint8_t)Value;
172       Value >>= 8;
173     }
174     break;
175   }
176   case macho::RIT_ARM_Branch24Bit: {
177     // Mask the value into the target address. We know instructions are
178     // 32-bit aligned, so we can do it all at once.
179     uint32_t *p = (uint32_t*)LocalAddress;
180     // The low two bits of the value are not encoded.
181     Value >>= 2;
182     // Mask the value to 24 bits.
183     Value &= 0xffffff;
184     // FIXME: If the destination is a Thumb function (and the instruction
185     // is a non-predicated BL instruction), we need to change it to a BLX
186     // instruction instead.
187
188     // Insert the value into the instruction.
189     *p = (*p & ~0xffffff) | Value;
190     break;
191   }
192   case macho::RIT_ARM_ThumbBranch22Bit:
193   case macho::RIT_ARM_ThumbBranch32Bit:
194   case macho::RIT_ARM_Half:
195   case macho::RIT_ARM_HalfDifference:
196   case macho::RIT_Pair:
197   case macho::RIT_Difference:
198   case macho::RIT_ARM_LocalDifference:
199   case macho::RIT_ARM_PreboundLazyPointer:
200     return Error("Relocation type not implemented yet!");
201   }
202   return false;
203 }
204
205 void RuntimeDyldMachO::processRelocationRef(const ObjRelocationInfo &Rel,
206                                             ObjectImage &Obj,
207                                             ObjSectionToIDMap &ObjSectionToID,
208                                             const SymbolTableMap &Symbols,
209                                             StubMap &Stubs) {
210
211   uint32_t RelType = (uint32_t) (Rel.Type & 0xffffffffL);
212   RelocationValueRef Value;
213   SectionEntry &Section = Sections[Rel.SectionID];
214   uint8_t *Target = Section.Address + Rel.Offset;
215
216   bool isExtern = (RelType >> 27) & 1;
217   if (isExtern) {
218     // Obtain the symbol name which is referenced in the relocation
219     StringRef TargetName;
220     const SymbolRef &Symbol = Rel.Symbol;
221     Symbol.getName(TargetName);
222     // First search for the symbol in the local symbol table
223     SymbolTableMap::const_iterator lsi = Symbols.find(TargetName.data());
224     if (lsi != Symbols.end()) {
225       Value.SectionID = lsi->second.first;
226       Value.Addend = lsi->second.second;
227     } else {
228       // Search for the symbol in the global symbol table
229       SymbolTableMap::const_iterator gsi = GlobalSymbolTable.find(TargetName.data());
230       if (gsi != GlobalSymbolTable.end()) {
231         Value.SectionID = gsi->second.first;
232         Value.Addend = gsi->second.second;
233       } else
234         Value.SymbolName = TargetName.data();
235     }
236   } else {
237     error_code err;
238     uint8_t sectionIndex = static_cast<uint8_t>(RelType & 0xFF);
239     section_iterator si = Obj.begin_sections(),
240                      se = Obj.end_sections();
241     for (uint8_t i = 1; i < sectionIndex; i++) {
242       error_code err;
243       si.increment(err);
244       if (si == se)
245         break;
246     }
247     assert(si != se && "No section containing relocation!");
248     Value.SectionID = findOrEmitSection(Obj, *si, true, ObjSectionToID);
249     Value.Addend = *(const intptr_t *)Target;
250     if (Value.Addend) {
251       // The MachO addend is an offset from the current section.  We need it
252       // to be an offset from the destination section
253       Value.Addend += Section.ObjAddress - Sections[Value.SectionID].ObjAddress;
254     }
255   }
256
257   if (Arch == Triple::arm && RelType == macho::RIT_ARM_Branch24Bit) {
258     // This is an ARM branch relocation, need to use a stub function.
259
260     //  Look up for existing stub.
261     StubMap::const_iterator i = Stubs.find(Value);
262     if (i != Stubs.end())
263       resolveRelocation(Target, (uint64_t)Target,
264                         (uint64_t)Section.Address + i->second,
265                         RelType, 0);
266     else {
267       // Create a new stub function.
268       Stubs[Value] = Section.StubOffset;
269       uint8_t *StubTargetAddr = createStubFunction(Section.Address +
270                                                    Section.StubOffset);
271       RelocationEntry RE(Rel.SectionID, StubTargetAddr - Section.Address,
272                          macho::RIT_Vanilla, Value.Addend);
273       if (Value.SymbolName)
274         addRelocationForSymbol(RE, Value.SymbolName);
275       else
276         addRelocationForSection(RE, Value.SectionID);
277       resolveRelocation(Target, (uint64_t)Target,
278                         (uint64_t)Section.Address + Section.StubOffset,
279                         RelType, 0);
280       Section.StubOffset += getMaxStubSize();
281     }
282   } else {
283     RelocationEntry RE(Rel.SectionID, Rel.Offset, RelType, Value.Addend);
284     if (Value.SymbolName)
285       addRelocationForSymbol(RE, Value.SymbolName);
286     else
287       addRelocationForSection(RE, Value.SectionID);
288   }
289 }
290
291
292 bool RuntimeDyldMachO::isCompatibleFormat(
293         const MemoryBuffer *InputBuffer) const {
294   StringRef Magic = InputBuffer->getBuffer().slice(0, 4);
295   if (Magic == "\xFE\xED\xFA\xCE") return true;
296   if (Magic == "\xCE\xFA\xED\xFE") return true;
297   if (Magic == "\xFE\xED\xFA\xCF") return true;
298   if (Magic == "\xCF\xFA\xED\xFE") return true;
299   return false;
300 }
301
302 } // end namespace llvm