1//===-- RuntimeDyldImpl.h - Run-time dynamic linker for MC-JIT --*- 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// Interface for the implementations of runtime dynamic linker facilities.
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef LLVM_LIB_EXECUTIONENGINE_RUNTIMEDYLD_RUNTIMEDYLDIMPL_H
14#define LLVM_LIB_EXECUTIONENGINE_RUNTIMEDYLD_RUNTIMEDYLDIMPL_H
15
16#include "llvm/ADT/SmallVector.h"
17#include "llvm/ADT/StringMap.h"
18#include "llvm/ExecutionEngine/Orc/SymbolStringPool.h"
19#include "llvm/ExecutionEngine/RTDyldMemoryManager.h"
20#include "llvm/ExecutionEngine/RuntimeDyld.h"
21#include "llvm/ExecutionEngine/RuntimeDyldChecker.h"
22#include "llvm/Object/ObjectFile.h"
23#include "llvm/Support/Debug.h"
24#include "llvm/Support/ErrorHandling.h"
25#include "llvm/Support/Format.h"
26#include "llvm/Support/Mutex.h"
27#include "llvm/Support/SwapByteOrder.h"
28#include "llvm/TargetParser/Host.h"
29#include "llvm/TargetParser/Triple.h"
30#include <deque>
31#include <map>
32#include <system_error>
33#include <unordered_map>
34
35using namespace llvm;
36using namespace llvm::object;
37
38namespace llvm {
39
40#define UNIMPLEMENTED_RELOC(RelType) \
41 case RelType: \
42 return make_error<RuntimeDyldError>("Unimplemented relocation: " #RelType)
43
44/// SectionEntry - represents a section emitted into memory by the dynamic
45/// linker.
46class SectionEntry {
47 /// Name - section name.
48 std::string Name;
49
50 /// Address - address in the linker's memory where the section resides.
51 uint8_t *Address;
52
53 /// Size - section size. Doesn't include the stubs.
54 size_t Size;
55
56 /// LoadAddress - the address of the section in the target process's memory.
57 /// Used for situations in which JIT-ed code is being executed in the address
58 /// space of a separate process. If the code executes in the same address
59 /// space where it was JIT-ed, this just equals Address.
60 uint64_t LoadAddress;
61
62 /// StubOffset - used for architectures with stub functions for far
63 /// relocations (like ARM).
64 uintptr_t StubOffset;
65
66 /// The total amount of space allocated for this section. This includes the
67 /// section size and the maximum amount of space that the stubs can occupy.
68 size_t AllocationSize;
69
70 /// ObjAddress - address of the section in the in-memory object file. Used
71 /// for calculating relocations in some object formats (like MachO).
72 uintptr_t ObjAddress;
73
74public:
75 SectionEntry(StringRef name, uint8_t *address, size_t size,
76 size_t allocationSize, uintptr_t objAddress)
77 : Name(std::string(name)), Address(address), Size(size),
78 LoadAddress(reinterpret_cast<uintptr_t>(address)), StubOffset(size),
79 AllocationSize(allocationSize), ObjAddress(objAddress) {
80 // AllocationSize is used only in asserts, prevent an "unused private field"
81 // warning:
82 (void)AllocationSize;
83 }
84
85 StringRef getName() const { return Name; }
86
87 uint8_t *getAddress() const { return Address; }
88
89 /// Return the address of this section with an offset.
90 uint8_t *getAddressWithOffset(unsigned OffsetBytes) const {
91 assert(OffsetBytes <= AllocationSize && "Offset out of bounds!");
92 return Address + OffsetBytes;
93 }
94
95 size_t getSize() const { return Size; }
96
97 uint64_t getLoadAddress() const { return LoadAddress; }
98 void setLoadAddress(uint64_t LA) { LoadAddress = LA; }
99
100 /// Return the load address of this section with an offset.
101 uint64_t getLoadAddressWithOffset(unsigned OffsetBytes) const {
102 assert(OffsetBytes <= AllocationSize && "Offset out of bounds!");
103 return LoadAddress + OffsetBytes;
104 }
105
106 uintptr_t getStubOffset() const { return StubOffset; }
107
108 void advanceStubOffset(unsigned StubSize) {
109 StubOffset += StubSize;
110 assert(StubOffset <= AllocationSize && "Not enough space allocated!");
111 }
112
113 uintptr_t getObjAddress() const { return ObjAddress; }
114};
115
116/// RelocationEntry - used to represent relocations internally in the dynamic
117/// linker.
118class RelocationEntry {
119public:
120 /// Offset - offset into the section.
121 uint64_t Offset;
122
123 /// Addend - the relocation addend encoded in the instruction itself. Also
124 /// used to make a relocation section relative instead of symbol relative.
125 int64_t Addend;
126
127 /// SectionID - the section this relocation points to.
128 unsigned SectionID;
129
130 /// RelType - relocation type.
131 uint32_t RelType;
132
133 struct SectionPair {
134 uint32_t SectionA;
135 uint32_t SectionB;
136 };
137
138 /// SymOffset - Section offset of the relocation entry's symbol (used for GOT
139 /// lookup).
140 union {
141 uint64_t SymOffset;
142 SectionPair Sections;
143 };
144
145 /// The size of this relocation (MachO specific).
146 unsigned Size;
147
148 /// True if this is a PCRel relocation (MachO specific).
149 bool IsPCRel : 1;
150
151 // ARM (MachO and COFF) specific.
152 bool IsTargetThumbFunc : 1;
153
154 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend)
155 : Offset(offset), Addend(addend), SectionID(id), RelType(type),
156 SymOffset(0), Size(0), IsPCRel(false), IsTargetThumbFunc(false) {}
157
158 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
159 uint64_t symoffset)
160 : Offset(offset), Addend(addend), SectionID(id), RelType(type),
161 SymOffset(symoffset), Size(0), IsPCRel(false),
162 IsTargetThumbFunc(false) {}
163
164 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
165 bool IsPCRel, unsigned Size)
166 : Offset(offset), Addend(addend), SectionID(id), RelType(type),
167 SymOffset(0), Size(Size), IsPCRel(IsPCRel), IsTargetThumbFunc(false) {}
168
169 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
170 unsigned SectionA, uint64_t SectionAOffset, unsigned SectionB,
171 uint64_t SectionBOffset, bool IsPCRel, unsigned Size)
172 : Offset(offset), Addend(SectionAOffset - SectionBOffset + addend),
173 SectionID(id), RelType(type), Size(Size), IsPCRel(IsPCRel),
174 IsTargetThumbFunc(false) {
175 Sections.SectionA = SectionA;
176 Sections.SectionB = SectionB;
177 }
178
179 RelocationEntry(unsigned id, uint64_t offset, uint32_t type, int64_t addend,
180 unsigned SectionA, uint64_t SectionAOffset, unsigned SectionB,
181 uint64_t SectionBOffset, bool IsPCRel, unsigned Size,
182 bool IsTargetThumbFunc)
183 : Offset(offset), Addend(SectionAOffset - SectionBOffset + addend),
184 SectionID(id), RelType(type), Size(Size), IsPCRel(IsPCRel),
185 IsTargetThumbFunc(IsTargetThumbFunc) {
186 Sections.SectionA = SectionA;
187 Sections.SectionB = SectionB;
188 }
189};
190
191class RelocationValueRef {
192public:
193 unsigned SectionID = 0;
194 uint64_t Offset = 0;
195 int64_t Addend = 0;
196 const char *SymbolName = nullptr;
197 bool IsStubThumb = false;
198
199 inline bool operator==(const RelocationValueRef &Other) const {
200 return SectionID == Other.SectionID && Offset == Other.Offset &&
201 Addend == Other.Addend && SymbolName == Other.SymbolName &&
202 IsStubThumb == Other.IsStubThumb;
203 }
204 inline bool operator<(const RelocationValueRef &Other) const {
205 return std::tie(args: SectionID, args: Offset, args: Addend, args: IsStubThumb, args: SymbolName) <
206 std::tie(args: Other.SectionID, args: Other.Offset, args: Other.Addend,
207 args: Other.IsStubThumb, args: Other.SymbolName);
208 }
209};
210
211/// Symbol info for RuntimeDyld.
212class SymbolTableEntry {
213public:
214 SymbolTableEntry() = default;
215
216 SymbolTableEntry(unsigned SectionID, uint64_t Offset, JITSymbolFlags Flags)
217 : Offset(Offset), SectionID(SectionID), Flags(Flags) {}
218
219 unsigned getSectionID() const { return SectionID; }
220 uint64_t getOffset() const { return Offset; }
221 void setOffset(uint64_t NewOffset) { Offset = NewOffset; }
222
223 JITSymbolFlags getFlags() const { return Flags; }
224
225private:
226 uint64_t Offset = 0;
227 unsigned SectionID = 0;
228 JITSymbolFlags Flags = JITSymbolFlags::None;
229};
230
231typedef StringMap<SymbolTableEntry> RTDyldSymbolTable;
232
233class RuntimeDyldImpl {
234 friend class RuntimeDyld::LoadedObjectInfo;
235protected:
236 static const unsigned AbsoluteSymbolSection = ~0U;
237
238 // The MemoryManager to load objects into.
239 RuntimeDyld::MemoryManager &MemMgr;
240
241 // The symbol resolver to use for external symbols.
242 JITSymbolResolver &Resolver;
243
244 // A list of all sections emitted by the dynamic linker. These sections are
245 // referenced in the code by means of their index in this list - SectionID.
246 // Because references may be kept while the list grows, use a container that
247 // guarantees reference stability.
248 typedef std::deque<SectionEntry> SectionList;
249 SectionList Sections;
250
251 typedef unsigned SID; // Type for SectionIDs
252#define RTDYLD_INVALID_SECTION_ID ((RuntimeDyldImpl::SID)(-1))
253
254 // Keep a map of sections from object file to the SectionID which
255 // references it.
256 typedef std::map<SectionRef, unsigned> ObjSectionToIDMap;
257
258 // A global symbol table for symbols from all loaded modules.
259 RTDyldSymbolTable GlobalSymbolTable;
260
261 // Keep a map of common symbols to their info pairs
262 typedef std::vector<SymbolRef> CommonSymbolList;
263
264 // For each symbol, keep a list of relocations based on it. Anytime
265 // its address is reassigned (the JIT re-compiled the function, e.g.),
266 // the relocations get re-resolved.
267 // The symbol (or section) the relocation is sourced from is the Key
268 // in the relocation list where it's stored.
269 typedef SmallVector<RelocationEntry, 64> RelocationList;
270 // Relocations to sections already loaded. Indexed by SectionID which is the
271 // source of the address. The target where the address will be written is
272 // SectionID/Offset in the relocation itself.
273 std::unordered_map<unsigned, RelocationList> Relocations;
274
275 // Relocations to external symbols that are not yet resolved. Symbols are
276 // external when they aren't found in the global symbol table of all loaded
277 // modules. This map is indexed by symbol name.
278 StringMap<RelocationList> ExternalSymbolRelocations;
279
280
281 typedef std::map<RelocationValueRef, uintptr_t> StubMap;
282
283 Triple::ArchType Arch;
284 bool IsTargetLittleEndian;
285 bool IsMipsO32ABI;
286 bool IsMipsN32ABI;
287 bool IsMipsN64ABI;
288
289 // True if all sections should be passed to the memory manager, false if only
290 // sections containing relocations should be. Defaults to 'false'.
291 bool ProcessAllSections;
292
293 // This mutex prevents simultaneously loading objects from two different
294 // threads. This keeps us from having to protect individual data structures
295 // and guarantees that section allocation requests to the memory manager
296 // won't be interleaved between modules. It is also used in mapSectionAddress
297 // and resolveRelocations to protect write access to internal data structures.
298 //
299 // loadObject may be called on the same thread during the handling of
300 // processRelocations, and that's OK. The handling of the relocation lists
301 // is written in such a way as to work correctly if new elements are added to
302 // the end of the list while the list is being processed.
303 sys::Mutex lock;
304
305 using NotifyStubEmittedFunction =
306 RuntimeDyld::NotifyStubEmittedFunction;
307 NotifyStubEmittedFunction NotifyStubEmitted;
308
309 virtual unsigned getMaxStubSize() const = 0;
310 virtual Align getStubAlignment() = 0;
311
312 bool HasError;
313 std::string ErrorStr;
314
315 void writeInt16BE(uint8_t *Addr, uint16_t Value) {
316 llvm::support::endian::write<uint16_t>(memory: Addr, value: Value,
317 endian: IsTargetLittleEndian
318 ? llvm::endianness::little
319 : llvm::endianness::big);
320 }
321
322 void writeInt32BE(uint8_t *Addr, uint32_t Value) {
323 llvm::support::endian::write<uint32_t>(memory: Addr, value: Value,
324 endian: IsTargetLittleEndian
325 ? llvm::endianness::little
326 : llvm::endianness::big);
327 }
328
329 void writeInt64BE(uint8_t *Addr, uint64_t Value) {
330 llvm::support::endian::write<uint64_t>(memory: Addr, value: Value,
331 endian: IsTargetLittleEndian
332 ? llvm::endianness::little
333 : llvm::endianness::big);
334 }
335
336 virtual void setMipsABI(const ObjectFile &Obj) {
337 IsMipsO32ABI = false;
338 IsMipsN32ABI = false;
339 IsMipsN64ABI = false;
340 }
341
342 /// Endian-aware read Read the least significant Size bytes from Src.
343 uint64_t readBytesUnaligned(uint8_t *Src, unsigned Size) const;
344
345 /// Endian-aware write. Write the least significant Size bytes from Value to
346 /// Dst.
347 void writeBytesUnaligned(uint64_t Value, uint8_t *Dst, unsigned Size) const;
348
349 /// Generate JITSymbolFlags from a libObject symbol.
350 virtual Expected<JITSymbolFlags> getJITSymbolFlags(const SymbolRef &Sym);
351
352 /// Modify the given target address based on the given symbol flags.
353 /// This can be used by subclasses to tweak addresses based on symbol flags,
354 /// For example: the MachO/ARM target uses it to set the low bit if the target
355 /// is a thumb symbol.
356 virtual uint64_t modifyAddressBasedOnFlags(uint64_t Addr,
357 JITSymbolFlags Flags) const {
358 return Addr;
359 }
360
361 /// Given the common symbols discovered in the object file, emit a
362 /// new section for them and update the symbol mappings in the object and
363 /// symbol table.
364 Error emitCommonSymbols(const ObjectFile &Obj,
365 CommonSymbolList &CommonSymbols, uint64_t CommonSize,
366 uint32_t CommonAlign);
367
368 /// Emits section data from the object file to the MemoryManager.
369 /// \param IsCode if it's true then allocateCodeSection() will be
370 /// used for emits, else allocateDataSection() will be used.
371 /// \return SectionID.
372 Expected<unsigned> emitSection(const ObjectFile &Obj,
373 const SectionRef &Section,
374 bool IsCode);
375
376 /// Find Section in LocalSections. If the secton is not found - emit
377 /// it and store in LocalSections.
378 /// \param IsCode if it's true then allocateCodeSection() will be
379 /// used for emmits, else allocateDataSection() will be used.
380 /// \return SectionID.
381 Expected<unsigned> findOrEmitSection(const ObjectFile &Obj,
382 const SectionRef &Section, bool IsCode,
383 ObjSectionToIDMap &LocalSections);
384
385 // Add a relocation entry that uses the given section.
386 void addRelocationForSection(const RelocationEntry &RE, unsigned SectionID);
387
388 // Add a relocation entry that uses the given symbol. This symbol may
389 // be found in the global symbol table, or it may be external.
390 void addRelocationForSymbol(const RelocationEntry &RE, StringRef SymbolName);
391
392 /// Emits long jump instruction to Addr.
393 /// \return Pointer to the memory area for emitting target address.
394 uint8_t *createStubFunction(uint8_t *Addr, unsigned AbiVariant = 0);
395
396 /// Resolves relocations from Relocs list with address from Value.
397 void resolveRelocationList(const RelocationList &Relocs, uint64_t Value);
398
399 /// A object file specific relocation resolver
400 /// \param RE The relocation to be resolved
401 /// \param Value Target symbol address to apply the relocation action
402 virtual void resolveRelocation(const RelocationEntry &RE, uint64_t Value) = 0;
403
404 /// Parses one or more object file relocations (some object files use
405 /// relocation pairs) and stores it to Relocations or SymbolRelocations
406 /// (this depends on the object file type).
407 /// \return Iterator to the next relocation that needs to be parsed.
408 virtual Expected<relocation_iterator>
409 processRelocationRef(unsigned SectionID, relocation_iterator RelI,
410 const ObjectFile &Obj, ObjSectionToIDMap &ObjSectionToID,
411 StubMap &Stubs) = 0;
412
413 void applyExternalSymbolRelocations(
414 const StringMap<JITEvaluatedSymbol> ExternalSymbolMap);
415
416 /// Resolve relocations to external symbols.
417 Error resolveExternalSymbols();
418
419 // Compute an upper bound of the memory that is required to load all
420 // sections
421 Error computeTotalAllocSize(const ObjectFile &Obj, uint64_t &CodeSize,
422 Align &CodeAlign, uint64_t &RODataSize,
423 Align &RODataAlign, uint64_t &RWDataSize,
424 Align &RWDataAlign);
425
426 // Compute GOT size
427 unsigned computeGOTSize(const ObjectFile &Obj);
428
429 // Compute the stub buffer size required for a section
430 unsigned computeSectionStubBufSize(const ObjectFile &Obj,
431 const SectionRef &Section);
432
433 // Implementation of the generic part of the loadObject algorithm.
434 Expected<ObjSectionToIDMap> loadObjectImpl(const object::ObjectFile &Obj);
435
436 // Return size of Global Offset Table (GOT) entry
437 virtual size_t getGOTEntrySize() { return 0; }
438
439 // Hook for the subclasses to do further processing when a symbol is added to
440 // the global symbol table. This function may modify the symbol table entry.
441 virtual void processNewSymbol(const SymbolRef &ObjSymbol, SymbolTableEntry& Entry) {}
442
443 // Return true if the relocation R may require allocating a GOT entry.
444 virtual bool relocationNeedsGot(const RelocationRef &R) const {
445 return false;
446 }
447
448 // Return true if the relocation R may require allocating a stub.
449 virtual bool relocationNeedsStub(const RelocationRef &R) const {
450 return true; // Conservative answer
451 }
452
453 // Return true if the relocation R may require allocating a DLL import stub.
454 virtual bool relocationNeedsDLLImportStub(const RelocationRef &R) const {
455 return false;
456 }
457
458 // Add the size of a DLL import stub to the buffer size
459 virtual unsigned sizeAfterAddingDLLImportStub(unsigned Size) const {
460 return Size;
461 }
462
463public:
464 RuntimeDyldImpl(RuntimeDyld::MemoryManager &MemMgr,
465 JITSymbolResolver &Resolver)
466 : MemMgr(MemMgr), Resolver(Resolver),
467 ProcessAllSections(false), HasError(false) {
468 }
469
470 virtual ~RuntimeDyldImpl();
471
472 void setProcessAllSections(bool ProcessAllSections) {
473 this->ProcessAllSections = ProcessAllSections;
474 }
475
476 virtual std::unique_ptr<RuntimeDyld::LoadedObjectInfo>
477 loadObject(const object::ObjectFile &Obj) = 0;
478
479 uint64_t getSectionLoadAddress(unsigned SectionID) const {
480 if (SectionID == AbsoluteSymbolSection)
481 return 0;
482 else
483 return Sections[SectionID].getLoadAddress();
484 }
485
486 uint8_t *getSectionAddress(unsigned SectionID) const {
487 if (SectionID == AbsoluteSymbolSection)
488 return nullptr;
489 else
490 return Sections[SectionID].getAddress();
491 }
492
493 StringRef getSectionContent(unsigned SectionID) const {
494 if (SectionID == AbsoluteSymbolSection)
495 return {};
496 else
497 return StringRef(
498 reinterpret_cast<char *>(Sections[SectionID].getAddress()),
499 Sections[SectionID].getStubOffset() + getMaxStubSize());
500 }
501
502 uint8_t* getSymbolLocalAddress(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(Key: Name);
506 if (pos == GlobalSymbolTable.end())
507 return nullptr;
508 const auto &SymInfo = pos->second;
509 // Absolute symbols do not have a local address.
510 if (SymInfo.getSectionID() == AbsoluteSymbolSection)
511 return nullptr;
512 return getSectionAddress(SectionID: SymInfo.getSectionID()) + SymInfo.getOffset();
513 }
514
515 unsigned getSymbolSectionID(StringRef Name) const {
516 auto GSTItr = GlobalSymbolTable.find(Key: Name);
517 if (GSTItr == GlobalSymbolTable.end())
518 return ~0U;
519 return GSTItr->second.getSectionID();
520 }
521
522 JITEvaluatedSymbol getSymbol(StringRef Name) const {
523 // FIXME: Just look up as a function for now. Overly simple of course.
524 // Work in progress.
525 RTDyldSymbolTable::const_iterator pos = GlobalSymbolTable.find(Key: Name);
526 if (pos == GlobalSymbolTable.end())
527 return nullptr;
528 const auto &SymEntry = pos->second;
529 uint64_t SectionAddr = 0;
530 if (SymEntry.getSectionID() != AbsoluteSymbolSection)
531 SectionAddr = getSectionLoadAddress(SectionID: SymEntry.getSectionID());
532 uint64_t TargetAddr = SectionAddr + SymEntry.getOffset();
533
534 // FIXME: Have getSymbol should return the actual address and the client
535 // modify it based on the flags. This will require clients to be
536 // aware of the target architecture, which we should build
537 // infrastructure for.
538 TargetAddr = modifyAddressBasedOnFlags(Addr: TargetAddr, Flags: SymEntry.getFlags());
539 return JITEvaluatedSymbol(TargetAddr, SymEntry.getFlags());
540 }
541
542 std::map<StringRef, JITEvaluatedSymbol> getSymbolTable() const {
543 std::map<StringRef, JITEvaluatedSymbol> Result;
544
545 for (const auto &KV : GlobalSymbolTable) {
546 auto SectionID = KV.second.getSectionID();
547 uint64_t SectionAddr = getSectionLoadAddress(SectionID);
548 Result[KV.first()] =
549 JITEvaluatedSymbol(SectionAddr + KV.second.getOffset(), KV.second.getFlags());
550 }
551
552 return Result;
553 }
554
555 void resolveRelocations();
556
557 void resolveLocalRelocations();
558
559 static void finalizeAsync(
560 std::unique_ptr<RuntimeDyldImpl> This,
561 unique_function<void(object::OwningBinary<object::ObjectFile>,
562 std::unique_ptr<RuntimeDyld::LoadedObjectInfo>,
563 Error)>
564 OnEmitted,
565 object::OwningBinary<object::ObjectFile> O,
566 std::unique_ptr<RuntimeDyld::LoadedObjectInfo> Info);
567
568 void reassignSectionAddress(unsigned SectionID, uint64_t Addr);
569
570 void mapSectionAddress(const void *LocalAddress, uint64_t TargetAddress);
571
572 // Is the linker in an error state?
573 bool hasError() { return HasError; }
574
575 // Mark the error condition as handled and continue.
576 void clearError() { HasError = false; }
577
578 // Get the error message.
579 StringRef getErrorString() { return ErrorStr; }
580
581 virtual bool isCompatibleFile(const ObjectFile &Obj) const = 0;
582
583 void setNotifyStubEmitted(NotifyStubEmittedFunction NotifyStubEmitted) {
584 this->NotifyStubEmitted = std::move(NotifyStubEmitted);
585 }
586
587 virtual void registerEHFrames();
588
589 void deregisterEHFrames();
590
591 virtual Error finalizeLoad(const ObjectFile &ObjImg,
592 ObjSectionToIDMap &SectionMap) {
593 return Error::success();
594 }
595};
596
597} // end namespace llvm
598
599#endif
600