1 //===-- RuntimeDyldImpl.h - 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 // Interface for the implementations of runtime dynamic linker facilities.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #ifndef LLVM_LIB_EXECUTIONENGINE_RUNTIMEDYLD_RUNTIMEDYLDIMPL_H
15 #define LLVM_LIB_EXECUTIONENGINE_RUNTIMEDYLD_RUNTIMEDYLDIMPL_H
16
CreateTestColumn(int8_t mask,int8_t index,int64_t timestamp)17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/StringMap.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/ExecutionEngine/RTDyldMemoryManager.h"
21 #include "llvm/ExecutionEngine/RuntimeDyld.h"
22 #include "llvm/ExecutionEngine/RuntimeDyldChecker.h"
23 #include "llvm/Object/ObjectFile.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/ErrorHandling.h"
26 #include "llvm/Support/Format.h"
27 #include "llvm/Support/Host.h"
28 #include "llvm/Support/Mutex.h"
29 #include "llvm/Support/SwapByteOrder.h"
30 #include <map>
31 #include <system_error>
32 #include <unordered_map>
33
34 using namespace llvm;
35 using namespace llvm::object;
36
37 namespace llvm {
CreateTestRowValue(std::vector<std::tuple<int8_t,int8_t,int64_t>> column_specs)38
39 class Twine;
40
41 #define UNIMPLEMENTED_RELOC(RelType) \
42 case RelType: \
43 return make_error<RuntimeDyldError>("Unimplemented relocation: " #RelType)
44
45 /// SectionEntry - represents a section emitted into memory by the dynamic
46 /// linker.
47 class SectionEntry {
48 /// Name - section name.
49 std::string Name;
50
CreateRowTombstone(int64_t timestamp)51 /// Address - address in the linker's memory where the section resides.
52 uint8_t *Address;
53
54 /// Size - section size. Doesn't include the stubs.
VerifyRowValueColumns(std::vector<std::shared_ptr<ColumnBase>> & columns,std::size_t index_of_vector,int8_t expected_mask,int8_t expected_index,int64_t expected_timestamp)55 size_t Size;
56
57 /// LoadAddress - the address of the section in the target process's memory.
58 /// Used for situations in which JIT-ed code is being executed in the address
59 /// space of a separate process. If the code executes in the same address
60 /// space where it was JIT-ed, this just equals Address.
61 uint64_t LoadAddress;
62
63 /// StubOffset - used for architectures with stub functions for far
64 /// relocations (like ARM).
65 uintptr_t StubOffset;
66
ToMicroSeconds(int64_t seconds)67 /// The total amount of space allocated for this section. This includes the
68 /// section size and the maximum amount of space that the stubs can occupy.
69 size_t AllocationSize;
70
ToSeconds(int64_t microseconds)71 /// ObjAddress - address of the section in the in-memory object file. Used
72 /// for calculating relocations in some object formats (like MachO).
73 uintptr_t ObjAddress;
74
75 public:
76 SectionEntry(StringRef name, uint8_t *address, size_t size,
77 size_t allocationSize, uintptr_t objAddress)
78 : Name(name), Address(address), Size(size),
79 LoadAddress(reinterpret_cast<uintptr_t>(address)), StubOffset(size),
80 AllocationSize(allocationSize), ObjAddress(objAddress) {
81 // AllocationSize is used only in asserts, prevent an "unused private field"
82 // warning:
83 (void)AllocationSize;
84 }
85
86 StringRef getName() const { return Name; }
87
88 uint8_t *getAddress() const { return Address; }
89
90 /// Return the address of this section with an offset.
91 uint8_t *getAddressWithOffset(unsigned OffsetBytes) const {
92 assert(OffsetBytes <= AllocationSize && "Offset out of bounds!");
93 return Address + OffsetBytes;
94 }
95
96 size_t getSize() const { return Size; }
97
98 uint64_t getLoadAddress() const { return LoadAddress; }
99 void setLoadAddress(uint64_t LA) { LoadAddress = LA; }
100
101 /// Return the load address of this section with an offset.
102 uint64_t getLoadAddressWithOffset(unsigned OffsetBytes) const {
103 assert(OffsetBytes <= AllocationSize && "Offset out of bounds!");
104 return LoadAddress + OffsetBytes;
105 }
106
107 uintptr_t getStubOffset() const { return StubOffset; }
108
109 void advanceStubOffset(unsigned StubSize) {
110 StubOffset += StubSize;
111 assert(StubOffset <= AllocationSize && "Not enough space allocated!");
112 }
113
114 uintptr_t getObjAddress() const { return ObjAddress; }
115 };
116
117 /// RelocationEntry - used to represent relocations internally in the dynamic
118 /// linker.
119 class RelocationEntry {
120 public:
121 /// SectionID - the section this relocation points to.
122 unsigned SectionID;
123
124 /// Offset - offset into the section.
125 uint64_t Offset;
126
127 /// RelType - relocation type.
128 uint32_t RelType;
129
130 /// Addend - the relocation addend encoded in the instruction itself. Also
131 /// used to make a relocation section relative instead of symbol relative.
132 int64_t Addend;
133
134 struct SectionPair {
135 uint32_t SectionA;
136 uint32_t SectionB;
137 };
138
139 /// SymOffset - Section offset of the relocation entry's symbol (used for GOT
140 /// lookup).
141 union {
142 uint64_t SymOffset;
143 SectionPair Sections;
144 };
145
146 /// True if this is a PCRel relocation (MachO specific).
147 bool IsPCRel;
148
149 /// The size of this relocation (MachO specific).
150 unsigned Size;
151
152 // ARM (MachO and COFF) specific.
153 bool IsTargetThumbFunc = false;
154
155 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend)
156 : SectionID(id), Offset(offset), RelType(type), Addend(addend),
157 SymOffset(0), IsPCRel(false), Size(0), IsTargetThumbFunc(false) {}
158
159 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
160 uint64_t symoffset)
161 : SectionID(id), Offset(offset), RelType(type), Addend(addend),
162 SymOffset(symoffset), IsPCRel(false), Size(0),
163 IsTargetThumbFunc(false) {}
164
165 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
166 bool IsPCRel, unsigned Size)
167 : SectionID(id), Offset(offset), RelType(type), Addend(addend),
168 SymOffset(0), IsPCRel(IsPCRel), Size(Size), IsTargetThumbFunc(false) {}
169
170 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
171 unsigned SectionA, uint64_t SectionAOffset, unsigned SectionB,
172 uint64_t SectionBOffset, bool IsPCRel, unsigned Size)
173 : SectionID(id), Offset(offset), RelType(type),
174 Addend(SectionAOffset - SectionBOffset + addend), IsPCRel(IsPCRel),
175 Size(Size), IsTargetThumbFunc(false) {
176 Sections.SectionA = SectionA;
177 Sections.SectionB = SectionB;
178 }
179
180 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
181 unsigned SectionA, uint64_t SectionAOffset, unsigned SectionB,
182 uint64_t SectionBOffset, bool IsPCRel, unsigned Size,
183 bool IsTargetThumbFunc)
184 : SectionID(id), Offset(offset), RelType(type),
185 Addend(SectionAOffset - SectionBOffset + addend), IsPCRel(IsPCRel),
186 Size(Size), IsTargetThumbFunc(IsTargetThumbFunc) {
187 Sections.SectionA = SectionA;
188 Sections.SectionB = SectionB;
189 }
190 };
191
192 class RelocationValueRef {
193 public:
194 unsigned SectionID;
195 uint64_t Offset;
196 int64_t Addend;
197 const char *SymbolName;
198 bool IsStubThumb = false;
199 RelocationValueRef() : SectionID(0), Offset(0), Addend(0),
200 SymbolName(nullptr) {}
201
202 inline bool operator==(const RelocationValueRef &Other) const {
203 return SectionID == Other.SectionID && Offset == Other.Offset &&
204 Addend == Other.Addend && SymbolName == Other.SymbolName &&
205 IsStubThumb == Other.IsStubThumb;
206 }
207 inline bool operator<(const RelocationValueRef &Other) const {
208 if (SectionID != Other.SectionID)
209 return SectionID < Other.SectionID;
210 if (Offset != Other.Offset)
211 return Offset < Other.Offset;
212 if (Addend != Other.Addend)
213 return Addend < Other.Addend;
214 if (IsStubThumb != Other.IsStubThumb)
215 return IsStubThumb < Other.IsStubThumb;
216 return SymbolName < Other.SymbolName;
217 }
218 };
219
220 /// Symbol info for RuntimeDyld.
221 class SymbolTableEntry {
222 public:
223 SymbolTableEntry() = default;
224
225 SymbolTableEntry(unsigned SectionID, uint64_t Offset, JITSymbolFlags Flags)
226 : Offset(Offset), SectionID(SectionID), Flags(Flags) {}
227
228 unsigned getSectionID() const { return SectionID; }
229 uint64_t getOffset() const { return Offset; }
230 void setOffset(uint64_t NewOffset) { Offset = NewOffset; }
231
232 JITSymbolFlags getFlags() const { return Flags; }
233
234 private:
235 uint64_t Offset = 0;
236 unsigned SectionID = 0;
237 JITSymbolFlags Flags = JITSymbolFlags::None;
238 };
239
240 typedef StringMap<SymbolTableEntry> RTDyldSymbolTable;
241
242 class RuntimeDyldImpl {
243 friend class RuntimeDyld::LoadedObjectInfo;
244 friend class RuntimeDyldCheckerImpl;
245 protected:
246 static const unsigned AbsoluteSymbolSection = ~0U;
247
248 // The MemoryManager to load objects into.
249 RuntimeDyld::MemoryManager &MemMgr;
250
251 // The symbol resolver to use for external symbols.
252 JITSymbolResolver &Resolver;
253
254 // Attached RuntimeDyldChecker instance. Null if no instance attached.
255 RuntimeDyldCheckerImpl *Checker;
256
257 // A list of all sections emitted by the dynamic linker. These sections are
258 // referenced in the code by means of their index in this list - SectionID.
259 typedef SmallVector<SectionEntry, 64> SectionList;
260 SectionList Sections;
261
262 typedef unsigned SID; // Type for SectionIDs
263 #define RTDYLD_INVALID_SECTION_ID ((RuntimeDyldImpl::SID)(-1))
264
265 // Keep a map of sections from object file to the SectionID which
266 // references it.
267 typedef std::map<SectionRef, unsigned> ObjSectionToIDMap;
268
269 // A global symbol table for symbols from all loaded modules.
270 RTDyldSymbolTable GlobalSymbolTable;
271
272 // Keep a map of common symbols to their info pairs
273 typedef std::vector<SymbolRef> CommonSymbolList;
274
275 // For each symbol, keep a list of relocations based on it. Anytime
276 // its address is reassigned (the JIT re-compiled the function, e.g.),
277 // the relocations get re-resolved.
278 // The symbol (or section) the relocation is sourced from is the Key
279 // in the relocation list where it's stored.
280 typedef SmallVector<RelocationEntry, 64> RelocationList;
281 // Relocations to sections already loaded. Indexed by SectionID which is the
282 // source of the address. The target where the address will be written is
283 // SectionID/Offset in the relocation itself.
284 std::unordered_map<unsigned, RelocationList> Relocations;
285
286 // Relocations to external symbols that are not yet resolved. Symbols are
287 // external when they aren't found in the global symbol table of all loaded
288 // modules. This map is indexed by symbol name.
289 StringMap<RelocationList> ExternalSymbolRelocations;
290
291
292 typedef std::map<RelocationValueRef, uintptr_t> StubMap;
293
294 Triple::ArchType Arch;
295 bool IsTargetLittleEndian;
296 bool IsMipsO32ABI;
297 bool IsMipsN32ABI;
298 bool IsMipsN64ABI;
299
300 // True if all sections should be passed to the memory manager, false if only
301 // sections containing relocations should be. Defaults to 'false'.
302 bool ProcessAllSections;
303
304 // This mutex prevents simultaneously loading objects from two different
305 // threads. This keeps us from having to protect individual data structures
306 // and guarantees that section allocation requests to the memory manager
307 // won't be interleaved between modules. It is also used in mapSectionAddress
308 // and resolveRelocations to protect write access to internal data structures.
309 //
310 // loadObject may be called on the same thread during the handling of of
311 // processRelocations, and that's OK. The handling of the relocation lists
312 // is written in such a way as to work correctly if new elements are added to
313 // the end of the list while the list is being processed.
314 sys::Mutex lock;
315
316 virtual unsigned getMaxStubSize() = 0;
317 virtual unsigned getStubAlignment() = 0;
318
319 bool HasError;
320 std::string ErrorStr;
321
322 uint64_t getSectionLoadAddress(unsigned SectionID) const {
323 return Sections[SectionID].getLoadAddress();
324 }
325
326 uint8_t *getSectionAddress(unsigned SectionID) const {
327 return Sections[SectionID].getAddress();
328 }
329
330 void writeInt16BE(uint8_t *Addr, uint16_t Value) {
331 if (IsTargetLittleEndian)
332 sys::swapByteOrder(Value);
333 *Addr = (Value >> 8) & 0xFF;
334 *(Addr + 1) = Value & 0xFF;
335 }
336
337 void writeInt32BE(uint8_t *Addr, uint32_t Value) {
338 if (IsTargetLittleEndian)
339 sys::swapByteOrder(Value);
340 *Addr = (Value >> 24) & 0xFF;
341 *(Addr + 1) = (Value >> 16) & 0xFF;
342 *(Addr + 2) = (Value >> 8) & 0xFF;
343 *(Addr + 3) = Value & 0xFF;
344 }
345
346 void writeInt64BE(uint8_t *Addr, uint64_t Value) {
347 if (IsTargetLittleEndian)
348 sys::swapByteOrder(Value);
349 *Addr = (Value >> 56) & 0xFF;
350 *(Addr + 1) = (Value >> 48) & 0xFF;
351 *(Addr + 2) = (Value >> 40) & 0xFF;
352 *(Addr + 3) = (Value >> 32) & 0xFF;
353 *(Addr + 4) = (Value >> 24) & 0xFF;
354 *(Addr + 5) = (Value >> 16) & 0xFF;
355 *(Addr + 6) = (Value >> 8) & 0xFF;
356 *(Addr + 7) = Value & 0xFF;
357 }
358
359 virtual void setMipsABI(const ObjectFile &Obj) {
360 IsMipsO32ABI = false;
361 IsMipsN32ABI = false;
362 IsMipsN64ABI = false;
363 }
364
365 /// Endian-aware read Read the least significant Size bytes from Src.
366 uint64_t readBytesUnaligned(uint8_t *Src, unsigned Size) const;
367
368 /// Endian-aware write. Write the least significant Size bytes from Value to
369 /// Dst.
370 void writeBytesUnaligned(uint64_t Value, uint8_t *Dst, unsigned Size) const;
371
372 /// Generate JITSymbolFlags from a libObject symbol.
373 virtual JITSymbolFlags getJITSymbolFlags(const BasicSymbolRef &Sym);
374
375 /// Modify the given target address based on the given symbol flags.
376 /// This can be used by subclasses to tweak addresses based on symbol flags,
377 /// For example: the MachO/ARM target uses it to set the low bit if the target
378 /// is a thumb symbol.
379 virtual uint64_t modifyAddressBasedOnFlags(uint64_t Addr,
380 JITSymbolFlags Flags) const {
381 return Addr;
382 }
383
384 /// Given the common symbols discovered in the object file, emit a
385 /// new section for them and update the symbol mappings in the object and
386 /// symbol table.
387 Error emitCommonSymbols(const ObjectFile &Obj,
388 CommonSymbolList &CommonSymbols, uint64_t CommonSize,
389 uint32_t CommonAlign);
390
391 /// Emits section data from the object file to the MemoryManager.
392 /// \param IsCode if it's true then allocateCodeSection() will be
393 /// used for emits, else allocateDataSection() will be used.
394 /// \return SectionID.
395 Expected<unsigned> emitSection(const ObjectFile &Obj,
396 const SectionRef &Section,
397 bool IsCode);
398
399 /// Find Section in LocalSections. If the secton is not found - emit
400 /// it and store in LocalSections.
401 /// \param IsCode if it's true then allocateCodeSection() will be
402 /// used for emmits, else allocateDataSection() will be used.
403 /// \return SectionID.
404 Expected<unsigned> findOrEmitSection(const ObjectFile &Obj,
405 const SectionRef &Section, bool IsCode,
406 ObjSectionToIDMap &LocalSections);
407
408 // Add a relocation entry that uses the given section.
409 void addRelocationForSection(const RelocationEntry &RE, unsigned SectionID);
410
411 // Add a relocation entry that uses the given symbol. This symbol may
412 // be found in the global symbol table, or it may be external.
413 void addRelocationForSymbol(const RelocationEntry &RE, StringRef SymbolName);
414
415 /// Emits long jump instruction to Addr.
416 /// \return Pointer to the memory area for emitting target address.
417 uint8_t *createStubFunction(uint8_t *Addr, unsigned AbiVariant = 0);
418
419 /// Resolves relocations from Relocs list with address from Value.
420 void resolveRelocationList(const RelocationList &Relocs, uint64_t Value);
421
422 /// A object file specific relocation resolver
423 /// \param RE The relocation to be resolved
424 /// \param Value Target symbol address to apply the relocation action
425 virtual void resolveRelocation(const RelocationEntry &RE, uint64_t Value) = 0;
426
427 /// Parses one or more object file relocations (some object files use
428 /// relocation pairs) and stores it to Relocations or SymbolRelocations
429 /// (this depends on the object file type).
430 /// \return Iterator to the next relocation that needs to be parsed.
431 virtual Expected<relocation_iterator>
432 processRelocationRef(unsigned SectionID, relocation_iterator RelI,
433 const ObjectFile &Obj, ObjSectionToIDMap &ObjSectionToID,
434 StubMap &Stubs) = 0;
435
436 /// Resolve relocations to external symbols.
437 Error resolveExternalSymbols();
438
439 // Compute an upper bound of the memory that is required to load all
440 // sections
441 Error computeTotalAllocSize(const ObjectFile &Obj,
442 uint64_t &CodeSize, uint32_t &CodeAlign,
443 uint64_t &RODataSize, uint32_t &RODataAlign,
444 uint64_t &RWDataSize, uint32_t &RWDataAlign);
445
446 // Compute GOT size
447 unsigned computeGOTSize(const ObjectFile &Obj);
448
449 // Compute the stub buffer size required for a section
450 unsigned computeSectionStubBufSize(const ObjectFile &Obj,
451 const SectionRef &Section);
452
453 // Implementation of the generic part of the loadObject algorithm.
454 Expected<ObjSectionToIDMap> loadObjectImpl(const object::ObjectFile &Obj);
455
456 // Return size of Global Offset Table (GOT) entry
457 virtual size_t getGOTEntrySize() { return 0; }
458
459 // Return true if the relocation R may require allocating a GOT entry.
460 virtual bool relocationNeedsGot(const RelocationRef &R) const {
461 return false;
462 }
463
464 // Return true if the relocation R may require allocating a stub.
465 virtual bool relocationNeedsStub(const RelocationRef &R) const {
466 return true; // Conservative answer
467 }
468
469 public:
470 RuntimeDyldImpl(RuntimeDyld::MemoryManager &MemMgr,
471 JITSymbolResolver &Resolver)
472 : MemMgr(MemMgr), Resolver(Resolver), Checker(nullptr),
473 ProcessAllSections(false), HasError(false) {
474 }
475
476 virtual ~RuntimeDyldImpl();
477
478 void setProcessAllSections(bool ProcessAllSections) {
479 this->ProcessAllSections = ProcessAllSections;
480 }
481
482 void setRuntimeDyldChecker(RuntimeDyldCheckerImpl *Checker) {
483 this->Checker = Checker;
484 }
485
486 virtual std::unique_ptr<RuntimeDyld::LoadedObjectInfo>
487 loadObject(const object::ObjectFile &Obj) = 0;
488
489 uint8_t* getSymbolLocalAddress(StringRef Name) const {
490 // FIXME: Just look up as a function for now. Overly simple of course.
491 // Work in progress.
492 RTDyldSymbolTable::const_iterator pos = GlobalSymbolTable.find(Name);
493 if (pos == GlobalSymbolTable.end())
494 return nullptr;
495 const auto &SymInfo = pos->second;
496 // Absolute symbols do not have a local address.
497 if (SymInfo.getSectionID() == AbsoluteSymbolSection)
498 return nullptr;
499 return getSectionAddress(SymInfo.getSectionID()) + SymInfo.getOffset();
500 }
501
502 JITEvaluatedSymbol getSymbol(StringRef Name) const {
503 // FIXME: Just look up as a function for now. Overly simple of course.
504 // Work in progress.
505 RTDyldSymbolTable::const_iterator pos = GlobalSymbolTable.find(Name);
506 if (pos == GlobalSymbolTable.end())
507 return nullptr;
508 const auto &SymEntry = pos->second;
509 uint64_t SectionAddr = 0;
510 if (SymEntry.getSectionID() != AbsoluteSymbolSection)
511 SectionAddr = getSectionLoadAddress(SymEntry.getSectionID());
512 uint64_t TargetAddr = SectionAddr + SymEntry.getOffset();
513
514 // FIXME: Have getSymbol should return the actual address and the client
515 // modify it based on the flags. This will require clients to be
516 // aware of the target architecture, which we should build
517 // infrastructure for.
518 TargetAddr = modifyAddressBasedOnFlags(TargetAddr, SymEntry.getFlags());
519 return JITEvaluatedSymbol(TargetAddr, SymEntry.getFlags());
520 }
521
522 std::map<StringRef, JITEvaluatedSymbol> getSymbolTable() const {
523 std::map<StringRef, JITEvaluatedSymbol> Result;
524
525 for (auto &KV : GlobalSymbolTable) {
526 auto SectionID = KV.second.getSectionID();
527 uint64_t SectionAddr = 0;
528 if (SectionID != AbsoluteSymbolSection)
529 SectionAddr = getSectionLoadAddress(SectionID);
530 Result[KV.first()] =
531 JITEvaluatedSymbol(SectionAddr + KV.second.getOffset(), KV.second.getFlags());
532 }
533
534 return Result;
535 }
536
537 void resolveRelocations();
538
539 void reassignSectionAddress(unsigned SectionID, uint64_t Addr);
540
541 void mapSectionAddress(const void *LocalAddress, uint64_t TargetAddress);
542
543 // Is the linker in an error state?
544 bool hasError() { return HasError; }
545
546 // Mark the error condition as handled and continue.
547 void clearError() { HasError = false; }
548
549 // Get the error message.
550 StringRef getErrorString() { return ErrorStr; }
551
552 virtual bool isCompatibleFile(const ObjectFile &Obj) const = 0;
553
554 virtual void registerEHFrames();
555
556 void deregisterEHFrames();
557
558 virtual Error finalizeLoad(const ObjectFile &ObjImg,
559 ObjSectionToIDMap &SectionMap) {
560 return Error::success();
561 }
562 };
563
564 } // end namespace llvm
565
566 #endif
567