1 | //===-- RuntimeDyldELF.cpp - 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 | // Implementation of ELF support for the MC-JIT runtime dynamic linker. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "RuntimeDyldELF.h" |
14 | #include "RuntimeDyldCheckerImpl.h" |
15 | #include "Targets/RuntimeDyldELFMips.h" |
16 | #include "llvm/ADT/STLExtras.h" |
17 | #include "llvm/ADT/StringRef.h" |
18 | #include "llvm/BinaryFormat/ELF.h" |
19 | #include "llvm/Object/ELFObjectFile.h" |
20 | #include "llvm/Object/ObjectFile.h" |
21 | #include "llvm/Support/Endian.h" |
22 | #include "llvm/Support/MemoryBuffer.h" |
23 | #include "llvm/TargetParser/Triple.h" |
24 | |
25 | using namespace llvm; |
26 | using namespace llvm::object; |
27 | using namespace llvm::support::endian; |
28 | |
29 | #define DEBUG_TYPE "dyld" |
30 | |
31 | static void or32le(void *P, int32_t V) { write32le(P, V: read32le(P) | V); } |
32 | |
33 | static void or32AArch64Imm(void *L, uint64_t Imm) { |
34 | or32le(P: L, V: (Imm & 0xFFF) << 10); |
35 | } |
36 | |
37 | template <class T> static void write(bool isBE, void *P, T V) { |
38 | isBE ? write<T, llvm::endianness::big>(P, V) |
39 | : write<T, llvm::endianness::little>(P, V); |
40 | } |
41 | |
42 | static void write32AArch64Addr(void *L, uint64_t Imm) { |
43 | uint32_t ImmLo = (Imm & 0x3) << 29; |
44 | uint32_t ImmHi = (Imm & 0x1FFFFC) << 3; |
45 | uint64_t Mask = (0x3 << 29) | (0x1FFFFC << 3); |
46 | write32le(P: L, V: (read32le(P: L) & ~Mask) | ImmLo | ImmHi); |
47 | } |
48 | |
49 | // Return the bits [Start, End] from Val shifted Start bits. |
50 | // For instance, getBits(0xF0, 4, 8) returns 0xF. |
51 | static uint64_t getBits(uint64_t Val, int Start, int End) { |
52 | uint64_t Mask = ((uint64_t)1 << (End + 1 - Start)) - 1; |
53 | return (Val >> Start) & Mask; |
54 | } |
55 | |
56 | namespace { |
57 | |
58 | template <class ELFT> class DyldELFObject : public ELFObjectFile<ELFT> { |
59 | LLVM_ELF_IMPORT_TYPES_ELFT(ELFT) |
60 | |
61 | typedef typename ELFT::uint addr_type; |
62 | |
63 | DyldELFObject(ELFObjectFile<ELFT> &&Obj); |
64 | |
65 | public: |
66 | static Expected<std::unique_ptr<DyldELFObject>> |
67 | create(MemoryBufferRef Wrapper); |
68 | |
69 | void updateSectionAddress(const SectionRef &Sec, uint64_t Addr); |
70 | |
71 | void updateSymbolAddress(const SymbolRef &SymRef, uint64_t Addr); |
72 | |
73 | // Methods for type inquiry through isa, cast and dyn_cast |
74 | static bool classof(const Binary *v) { |
75 | return (isa<ELFObjectFile<ELFT>>(v) && |
76 | classof(cast<ELFObjectFile<ELFT>>(v))); |
77 | } |
78 | static bool classof(const ELFObjectFile<ELFT> *v) { |
79 | return v->isDyldType(); |
80 | } |
81 | }; |
82 | |
83 | |
84 | |
85 | // The MemoryBuffer passed into this constructor is just a wrapper around the |
86 | // actual memory. Ultimately, the Binary parent class will take ownership of |
87 | // this MemoryBuffer object but not the underlying memory. |
88 | template <class ELFT> |
89 | DyldELFObject<ELFT>::DyldELFObject(ELFObjectFile<ELFT> &&Obj) |
90 | : ELFObjectFile<ELFT>(std::move(Obj)) { |
91 | this->isDyldELFObject = true; |
92 | } |
93 | |
94 | template <class ELFT> |
95 | Expected<std::unique_ptr<DyldELFObject<ELFT>>> |
96 | DyldELFObject<ELFT>::create(MemoryBufferRef Wrapper) { |
97 | auto Obj = ELFObjectFile<ELFT>::create(Wrapper); |
98 | if (auto E = Obj.takeError()) |
99 | return std::move(E); |
100 | std::unique_ptr<DyldELFObject<ELFT>> Ret( |
101 | new DyldELFObject<ELFT>(std::move(*Obj))); |
102 | return std::move(Ret); |
103 | } |
104 | |
105 | template <class ELFT> |
106 | void DyldELFObject<ELFT>::updateSectionAddress(const SectionRef &Sec, |
107 | uint64_t Addr) { |
108 | DataRefImpl ShdrRef = Sec.getRawDataRefImpl(); |
109 | Elf_Shdr *shdr = |
110 | const_cast<Elf_Shdr *>(reinterpret_cast<const Elf_Shdr *>(ShdrRef.p)); |
111 | |
112 | // This assumes the address passed in matches the target address bitness |
113 | // The template-based type cast handles everything else. |
114 | shdr->sh_addr = static_cast<addr_type>(Addr); |
115 | } |
116 | |
117 | template <class ELFT> |
118 | void DyldELFObject<ELFT>::updateSymbolAddress(const SymbolRef &SymRef, |
119 | uint64_t Addr) { |
120 | |
121 | Elf_Sym *sym = const_cast<Elf_Sym *>( |
122 | ELFObjectFile<ELFT>::getSymbol(SymRef.getRawDataRefImpl())); |
123 | |
124 | // This assumes the address passed in matches the target address bitness |
125 | // The template-based type cast handles everything else. |
126 | sym->st_value = static_cast<addr_type>(Addr); |
127 | } |
128 | |
129 | class LoadedELFObjectInfo final |
130 | : public LoadedObjectInfoHelper<LoadedELFObjectInfo, |
131 | RuntimeDyld::LoadedObjectInfo> { |
132 | public: |
133 | LoadedELFObjectInfo(RuntimeDyldImpl &RTDyld, ObjSectionToIDMap ObjSecToIDMap) |
134 | : LoadedObjectInfoHelper(RTDyld, std::move(ObjSecToIDMap)) {} |
135 | |
136 | OwningBinary<ObjectFile> |
137 | getObjectForDebug(const ObjectFile &Obj) const override; |
138 | }; |
139 | |
140 | template <typename ELFT> |
141 | static Expected<std::unique_ptr<DyldELFObject<ELFT>>> |
142 | createRTDyldELFObject(MemoryBufferRef Buffer, const ObjectFile &SourceObject, |
143 | const LoadedELFObjectInfo &L) { |
144 | typedef typename ELFT::Shdr Elf_Shdr; |
145 | typedef typename ELFT::uint addr_type; |
146 | |
147 | Expected<std::unique_ptr<DyldELFObject<ELFT>>> ObjOrErr = |
148 | DyldELFObject<ELFT>::create(Buffer); |
149 | if (Error E = ObjOrErr.takeError()) |
150 | return std::move(E); |
151 | |
152 | std::unique_ptr<DyldELFObject<ELFT>> Obj = std::move(*ObjOrErr); |
153 | |
154 | // Iterate over all sections in the object. |
155 | auto SI = SourceObject.section_begin(); |
156 | for (const auto &Sec : Obj->sections()) { |
157 | Expected<StringRef> NameOrErr = Sec.getName(); |
158 | if (!NameOrErr) { |
159 | consumeError(Err: NameOrErr.takeError()); |
160 | continue; |
161 | } |
162 | |
163 | if (*NameOrErr != "" ) { |
164 | DataRefImpl ShdrRef = Sec.getRawDataRefImpl(); |
165 | Elf_Shdr *shdr = const_cast<Elf_Shdr *>( |
166 | reinterpret_cast<const Elf_Shdr *>(ShdrRef.p)); |
167 | |
168 | if (uint64_t SecLoadAddr = L.getSectionLoadAddress(Sec: *SI)) { |
169 | // This assumes that the address passed in matches the target address |
170 | // bitness. The template-based type cast handles everything else. |
171 | shdr->sh_addr = static_cast<addr_type>(SecLoadAddr); |
172 | } |
173 | } |
174 | ++SI; |
175 | } |
176 | |
177 | return std::move(Obj); |
178 | } |
179 | |
180 | static OwningBinary<ObjectFile> |
181 | createELFDebugObject(const ObjectFile &Obj, const LoadedELFObjectInfo &L) { |
182 | assert(Obj.isELF() && "Not an ELF object file." ); |
183 | |
184 | std::unique_ptr<MemoryBuffer> Buffer = |
185 | MemoryBuffer::getMemBufferCopy(InputData: Obj.getData(), BufferName: Obj.getFileName()); |
186 | |
187 | Expected<std::unique_ptr<ObjectFile>> DebugObj(nullptr); |
188 | handleAllErrors(E: DebugObj.takeError()); |
189 | if (Obj.getBytesInAddress() == 4 && Obj.isLittleEndian()) |
190 | DebugObj = |
191 | createRTDyldELFObject<ELF32LE>(Buffer: Buffer->getMemBufferRef(), SourceObject: Obj, L); |
192 | else if (Obj.getBytesInAddress() == 4 && !Obj.isLittleEndian()) |
193 | DebugObj = |
194 | createRTDyldELFObject<ELF32BE>(Buffer: Buffer->getMemBufferRef(), SourceObject: Obj, L); |
195 | else if (Obj.getBytesInAddress() == 8 && !Obj.isLittleEndian()) |
196 | DebugObj = |
197 | createRTDyldELFObject<ELF64BE>(Buffer: Buffer->getMemBufferRef(), SourceObject: Obj, L); |
198 | else if (Obj.getBytesInAddress() == 8 && Obj.isLittleEndian()) |
199 | DebugObj = |
200 | createRTDyldELFObject<ELF64LE>(Buffer: Buffer->getMemBufferRef(), SourceObject: Obj, L); |
201 | else |
202 | llvm_unreachable("Unexpected ELF format" ); |
203 | |
204 | handleAllErrors(E: DebugObj.takeError()); |
205 | return OwningBinary<ObjectFile>(std::move(*DebugObj), std::move(Buffer)); |
206 | } |
207 | |
208 | OwningBinary<ObjectFile> |
209 | LoadedELFObjectInfo::getObjectForDebug(const ObjectFile &Obj) const { |
210 | return createELFDebugObject(Obj, L: *this); |
211 | } |
212 | |
213 | } // anonymous namespace |
214 | |
215 | namespace llvm { |
216 | |
217 | RuntimeDyldELF::RuntimeDyldELF(RuntimeDyld::MemoryManager &MemMgr, |
218 | JITSymbolResolver &Resolver) |
219 | : RuntimeDyldImpl(MemMgr, Resolver), GOTSectionID(0), CurrentGOTIndex(0) {} |
220 | RuntimeDyldELF::~RuntimeDyldELF() = default; |
221 | |
222 | void RuntimeDyldELF::registerEHFrames() { |
223 | for (SID EHFrameSID : UnregisteredEHFrameSections) { |
224 | uint8_t *EHFrameAddr = Sections[EHFrameSID].getAddress(); |
225 | uint64_t EHFrameLoadAddr = Sections[EHFrameSID].getLoadAddress(); |
226 | size_t EHFrameSize = Sections[EHFrameSID].getSize(); |
227 | MemMgr.registerEHFrames(Addr: EHFrameAddr, LoadAddr: EHFrameLoadAddr, Size: EHFrameSize); |
228 | } |
229 | UnregisteredEHFrameSections.clear(); |
230 | } |
231 | |
232 | std::unique_ptr<RuntimeDyldELF> |
233 | llvm::RuntimeDyldELF::create(Triple::ArchType Arch, |
234 | RuntimeDyld::MemoryManager &MemMgr, |
235 | JITSymbolResolver &Resolver) { |
236 | switch (Arch) { |
237 | default: |
238 | return std::make_unique<RuntimeDyldELF>(args&: MemMgr, args&: Resolver); |
239 | case Triple::mips: |
240 | case Triple::mipsel: |
241 | case Triple::mips64: |
242 | case Triple::mips64el: |
243 | return std::make_unique<RuntimeDyldELFMips>(args&: MemMgr, args&: Resolver); |
244 | } |
245 | } |
246 | |
247 | std::unique_ptr<RuntimeDyld::LoadedObjectInfo> |
248 | RuntimeDyldELF::loadObject(const object::ObjectFile &O) { |
249 | if (auto ObjSectionToIDOrErr = loadObjectImpl(Obj: O)) |
250 | return std::make_unique<LoadedELFObjectInfo>(args&: *this, args&: *ObjSectionToIDOrErr); |
251 | else { |
252 | HasError = true; |
253 | raw_string_ostream ErrStream(ErrorStr); |
254 | logAllUnhandledErrors(E: ObjSectionToIDOrErr.takeError(), OS&: ErrStream); |
255 | return nullptr; |
256 | } |
257 | } |
258 | |
259 | void RuntimeDyldELF::resolveX86_64Relocation(const SectionEntry &Section, |
260 | uint64_t Offset, uint64_t Value, |
261 | uint32_t Type, int64_t Addend, |
262 | uint64_t SymOffset) { |
263 | switch (Type) { |
264 | default: |
265 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
266 | break; |
267 | case ELF::R_X86_64_NONE: |
268 | break; |
269 | case ELF::R_X86_64_8: { |
270 | Value += Addend; |
271 | assert((int64_t)Value <= INT8_MAX && (int64_t)Value >= INT8_MIN); |
272 | uint8_t TruncatedAddr = (Value & 0xFF); |
273 | *Section.getAddressWithOffset(OffsetBytes: Offset) = TruncatedAddr; |
274 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , TruncatedAddr) << " at " |
275 | << format("%p\n" , Section.getAddressWithOffset(Offset))); |
276 | break; |
277 | } |
278 | case ELF::R_X86_64_16: { |
279 | Value += Addend; |
280 | assert((int64_t)Value <= INT16_MAX && (int64_t)Value >= INT16_MIN); |
281 | uint16_t TruncatedAddr = (Value & 0xFFFF); |
282 | support::ulittle16_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
283 | TruncatedAddr; |
284 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , TruncatedAddr) << " at " |
285 | << format("%p\n" , Section.getAddressWithOffset(Offset))); |
286 | break; |
287 | } |
288 | case ELF::R_X86_64_64: { |
289 | support::ulittle64_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
290 | Value + Addend; |
291 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , (Value + Addend)) << " at " |
292 | << format("%p\n" , Section.getAddressWithOffset(Offset))); |
293 | break; |
294 | } |
295 | case ELF::R_X86_64_32: |
296 | case ELF::R_X86_64_32S: { |
297 | Value += Addend; |
298 | assert((Type == ELF::R_X86_64_32 && (Value <= UINT32_MAX)) || |
299 | (Type == ELF::R_X86_64_32S && |
300 | ((int64_t)Value <= INT32_MAX && (int64_t)Value >= INT32_MIN))); |
301 | uint32_t TruncatedAddr = (Value & 0xFFFFFFFF); |
302 | support::ulittle32_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
303 | TruncatedAddr; |
304 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , TruncatedAddr) << " at " |
305 | << format("%p\n" , Section.getAddressWithOffset(Offset))); |
306 | break; |
307 | } |
308 | case ELF::R_X86_64_PC8: { |
309 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
310 | int64_t RealOffset = Value + Addend - FinalAddress; |
311 | assert(isInt<8>(RealOffset)); |
312 | int8_t TruncOffset = (RealOffset & 0xFF); |
313 | Section.getAddress()[Offset] = TruncOffset; |
314 | break; |
315 | } |
316 | case ELF::R_X86_64_PC32: { |
317 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
318 | int64_t RealOffset = Value + Addend - FinalAddress; |
319 | assert(isInt<32>(RealOffset)); |
320 | int32_t TruncOffset = (RealOffset & 0xFFFFFFFF); |
321 | support::ulittle32_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
322 | TruncOffset; |
323 | break; |
324 | } |
325 | case ELF::R_X86_64_PC64: { |
326 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
327 | int64_t RealOffset = Value + Addend - FinalAddress; |
328 | support::ulittle64_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
329 | RealOffset; |
330 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , RealOffset) << " at " |
331 | << format("%p\n" , FinalAddress)); |
332 | break; |
333 | } |
334 | case ELF::R_X86_64_GOTOFF64: { |
335 | // Compute Value - GOTBase. |
336 | uint64_t GOTBase = 0; |
337 | for (const auto &Section : Sections) { |
338 | if (Section.getName() == ".got" ) { |
339 | GOTBase = Section.getLoadAddressWithOffset(OffsetBytes: 0); |
340 | break; |
341 | } |
342 | } |
343 | assert(GOTBase != 0 && "missing GOT" ); |
344 | int64_t GOTOffset = Value - GOTBase + Addend; |
345 | support::ulittle64_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = GOTOffset; |
346 | break; |
347 | } |
348 | case ELF::R_X86_64_DTPMOD64: { |
349 | // We only have one DSO, so the module id is always 1. |
350 | support::ulittle64_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = 1; |
351 | break; |
352 | } |
353 | case ELF::R_X86_64_DTPOFF64: |
354 | case ELF::R_X86_64_TPOFF64: { |
355 | // DTPOFF64 should resolve to the offset in the TLS block, TPOFF64 to the |
356 | // offset in the *initial* TLS block. Since we are statically linking, all |
357 | // TLS blocks already exist in the initial block, so resolve both |
358 | // relocations equally. |
359 | support::ulittle64_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
360 | Value + Addend; |
361 | break; |
362 | } |
363 | case ELF::R_X86_64_DTPOFF32: |
364 | case ELF::R_X86_64_TPOFF32: { |
365 | // As for the (D)TPOFF64 relocations above, both DTPOFF32 and TPOFF32 can |
366 | // be resolved equally. |
367 | int64_t RealValue = Value + Addend; |
368 | assert(RealValue >= INT32_MIN && RealValue <= INT32_MAX); |
369 | int32_t TruncValue = RealValue; |
370 | support::ulittle32_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
371 | TruncValue; |
372 | break; |
373 | } |
374 | } |
375 | } |
376 | |
377 | void RuntimeDyldELF::resolveX86Relocation(const SectionEntry &Section, |
378 | uint64_t Offset, uint32_t Value, |
379 | uint32_t Type, int32_t Addend) { |
380 | switch (Type) { |
381 | case ELF::R_386_32: { |
382 | support::ulittle32_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
383 | Value + Addend; |
384 | break; |
385 | } |
386 | // Handle R_386_PLT32 like R_386_PC32 since it should be able to |
387 | // reach any 32 bit address. |
388 | case ELF::R_386_PLT32: |
389 | case ELF::R_386_PC32: { |
390 | uint32_t FinalAddress = |
391 | Section.getLoadAddressWithOffset(OffsetBytes: Offset) & 0xFFFFFFFF; |
392 | uint32_t RealOffset = Value + Addend - FinalAddress; |
393 | support::ulittle32_t::ref(Section.getAddressWithOffset(OffsetBytes: Offset)) = |
394 | RealOffset; |
395 | break; |
396 | } |
397 | default: |
398 | // There are other relocation types, but it appears these are the |
399 | // only ones currently used by the LLVM ELF object writer |
400 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
401 | break; |
402 | } |
403 | } |
404 | |
405 | void RuntimeDyldELF::resolveAArch64Relocation(const SectionEntry &Section, |
406 | uint64_t Offset, uint64_t Value, |
407 | uint32_t Type, int64_t Addend) { |
408 | uint32_t *TargetPtr = |
409 | reinterpret_cast<uint32_t *>(Section.getAddressWithOffset(OffsetBytes: Offset)); |
410 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
411 | // Data should use target endian. Code should always use little endian. |
412 | bool isBE = Arch == Triple::aarch64_be; |
413 | |
414 | LLVM_DEBUG(dbgs() << "resolveAArch64Relocation, LocalAddress: 0x" |
415 | << format("%llx" , Section.getAddressWithOffset(Offset)) |
416 | << " FinalAddress: 0x" << format("%llx" , FinalAddress) |
417 | << " Value: 0x" << format("%llx" , Value) << " Type: 0x" |
418 | << format("%x" , Type) << " Addend: 0x" |
419 | << format("%llx" , Addend) << "\n" ); |
420 | |
421 | switch (Type) { |
422 | default: |
423 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
424 | break; |
425 | case ELF::R_AARCH64_NONE: |
426 | break; |
427 | case ELF::R_AARCH64_ABS16: { |
428 | uint64_t Result = Value + Addend; |
429 | assert(Result == static_cast<uint64_t>(llvm::SignExtend64(Result, 16)) || |
430 | (Result >> 16) == 0); |
431 | write(isBE, P: TargetPtr, V: static_cast<uint16_t>(Result & 0xffffU)); |
432 | break; |
433 | } |
434 | case ELF::R_AARCH64_ABS32: { |
435 | uint64_t Result = Value + Addend; |
436 | assert(Result == static_cast<uint64_t>(llvm::SignExtend64(Result, 32)) || |
437 | (Result >> 32) == 0); |
438 | write(isBE, P: TargetPtr, V: static_cast<uint32_t>(Result & 0xffffffffU)); |
439 | break; |
440 | } |
441 | case ELF::R_AARCH64_ABS64: |
442 | write(isBE, P: TargetPtr, V: Value + Addend); |
443 | break; |
444 | case ELF::R_AARCH64_PLT32: { |
445 | uint64_t Result = Value + Addend - FinalAddress; |
446 | assert(static_cast<int64_t>(Result) >= INT32_MIN && |
447 | static_cast<int64_t>(Result) <= INT32_MAX); |
448 | write(isBE, P: TargetPtr, V: static_cast<uint32_t>(Result)); |
449 | break; |
450 | } |
451 | case ELF::R_AARCH64_PREL16: { |
452 | uint64_t Result = Value + Addend - FinalAddress; |
453 | assert(static_cast<int64_t>(Result) >= INT16_MIN && |
454 | static_cast<int64_t>(Result) <= UINT16_MAX); |
455 | write(isBE, P: TargetPtr, V: static_cast<uint16_t>(Result & 0xffffU)); |
456 | break; |
457 | } |
458 | case ELF::R_AARCH64_PREL32: { |
459 | uint64_t Result = Value + Addend - FinalAddress; |
460 | assert(static_cast<int64_t>(Result) >= INT32_MIN && |
461 | static_cast<int64_t>(Result) <= UINT32_MAX); |
462 | write(isBE, P: TargetPtr, V: static_cast<uint32_t>(Result & 0xffffffffU)); |
463 | break; |
464 | } |
465 | case ELF::R_AARCH64_PREL64: |
466 | write(isBE, P: TargetPtr, V: Value + Addend - FinalAddress); |
467 | break; |
468 | case ELF::R_AARCH64_CONDBR19: { |
469 | uint64_t BranchImm = Value + Addend - FinalAddress; |
470 | |
471 | assert(isInt<21>(BranchImm)); |
472 | *TargetPtr &= 0xff00001fU; |
473 | // Immediate:20:2 goes in bits 23:5 of Bcc, CBZ, CBNZ |
474 | or32le(P: TargetPtr, V: (BranchImm & 0x001FFFFC) << 3); |
475 | break; |
476 | } |
477 | case ELF::R_AARCH64_TSTBR14: { |
478 | uint64_t BranchImm = Value + Addend - FinalAddress; |
479 | |
480 | assert(isInt<16>(BranchImm)); |
481 | |
482 | uint32_t RawInstr = *(support::little32_t *)TargetPtr; |
483 | *(support::little32_t *)TargetPtr = RawInstr & 0xfff8001fU; |
484 | |
485 | // Immediate:15:2 goes in bits 18:5 of TBZ, TBNZ |
486 | or32le(P: TargetPtr, V: (BranchImm & 0x0000FFFC) << 3); |
487 | break; |
488 | } |
489 | case ELF::R_AARCH64_CALL26: // fallthrough |
490 | case ELF::R_AARCH64_JUMP26: { |
491 | // Operation: S+A-P. Set Call or B immediate value to bits fff_fffc of the |
492 | // calculation. |
493 | uint64_t BranchImm = Value + Addend - FinalAddress; |
494 | |
495 | // "Check that -2^27 <= result < 2^27". |
496 | assert(isInt<28>(BranchImm)); |
497 | or32le(P: TargetPtr, V: (BranchImm & 0x0FFFFFFC) >> 2); |
498 | break; |
499 | } |
500 | case ELF::R_AARCH64_MOVW_UABS_G3: |
501 | or32le(P: TargetPtr, V: ((Value + Addend) & 0xFFFF000000000000) >> 43); |
502 | break; |
503 | case ELF::R_AARCH64_MOVW_UABS_G2_NC: |
504 | or32le(P: TargetPtr, V: ((Value + Addend) & 0xFFFF00000000) >> 27); |
505 | break; |
506 | case ELF::R_AARCH64_MOVW_UABS_G1_NC: |
507 | or32le(P: TargetPtr, V: ((Value + Addend) & 0xFFFF0000) >> 11); |
508 | break; |
509 | case ELF::R_AARCH64_MOVW_UABS_G0_NC: |
510 | or32le(P: TargetPtr, V: ((Value + Addend) & 0xFFFF) << 5); |
511 | break; |
512 | case ELF::R_AARCH64_ADR_PREL_PG_HI21: { |
513 | // Operation: Page(S+A) - Page(P) |
514 | uint64_t Result = |
515 | ((Value + Addend) & ~0xfffULL) - (FinalAddress & ~0xfffULL); |
516 | |
517 | // Check that -2^32 <= X < 2^32 |
518 | assert(isInt<33>(Result) && "overflow check failed for relocation" ); |
519 | |
520 | // Immediate goes in bits 30:29 + 5:23 of ADRP instruction, taken |
521 | // from bits 32:12 of X. |
522 | write32AArch64Addr(L: TargetPtr, Imm: Result >> 12); |
523 | break; |
524 | } |
525 | case ELF::R_AARCH64_ADD_ABS_LO12_NC: |
526 | // Operation: S + A |
527 | // Immediate goes in bits 21:10 of LD/ST instruction, taken |
528 | // from bits 11:0 of X |
529 | or32AArch64Imm(L: TargetPtr, Imm: Value + Addend); |
530 | break; |
531 | case ELF::R_AARCH64_LDST8_ABS_LO12_NC: |
532 | // Operation: S + A |
533 | // Immediate goes in bits 21:10 of LD/ST instruction, taken |
534 | // from bits 11:0 of X |
535 | or32AArch64Imm(L: TargetPtr, Imm: getBits(Val: Value + Addend, Start: 0, End: 11)); |
536 | break; |
537 | case ELF::R_AARCH64_LDST16_ABS_LO12_NC: |
538 | // Operation: S + A |
539 | // Immediate goes in bits 21:10 of LD/ST instruction, taken |
540 | // from bits 11:1 of X |
541 | or32AArch64Imm(L: TargetPtr, Imm: getBits(Val: Value + Addend, Start: 1, End: 11)); |
542 | break; |
543 | case ELF::R_AARCH64_LDST32_ABS_LO12_NC: |
544 | // Operation: S + A |
545 | // Immediate goes in bits 21:10 of LD/ST instruction, taken |
546 | // from bits 11:2 of X |
547 | or32AArch64Imm(L: TargetPtr, Imm: getBits(Val: Value + Addend, Start: 2, End: 11)); |
548 | break; |
549 | case ELF::R_AARCH64_LDST64_ABS_LO12_NC: |
550 | // Operation: S + A |
551 | // Immediate goes in bits 21:10 of LD/ST instruction, taken |
552 | // from bits 11:3 of X |
553 | or32AArch64Imm(L: TargetPtr, Imm: getBits(Val: Value + Addend, Start: 3, End: 11)); |
554 | break; |
555 | case ELF::R_AARCH64_LDST128_ABS_LO12_NC: |
556 | // Operation: S + A |
557 | // Immediate goes in bits 21:10 of LD/ST instruction, taken |
558 | // from bits 11:4 of X |
559 | or32AArch64Imm(L: TargetPtr, Imm: getBits(Val: Value + Addend, Start: 4, End: 11)); |
560 | break; |
561 | case ELF::R_AARCH64_LD_PREL_LO19: { |
562 | // Operation: S + A - P |
563 | uint64_t Result = Value + Addend - FinalAddress; |
564 | |
565 | // "Check that -2^20 <= result < 2^20". |
566 | assert(isInt<21>(Result)); |
567 | |
568 | *TargetPtr &= 0xff00001fU; |
569 | // Immediate goes in bits 23:5 of LD imm instruction, taken |
570 | // from bits 20:2 of X |
571 | *TargetPtr |= ((Result & 0xffc) << (5 - 2)); |
572 | break; |
573 | } |
574 | case ELF::R_AARCH64_ADR_PREL_LO21: { |
575 | // Operation: S + A - P |
576 | uint64_t Result = Value + Addend - FinalAddress; |
577 | |
578 | // "Check that -2^20 <= result < 2^20". |
579 | assert(isInt<21>(Result)); |
580 | |
581 | *TargetPtr &= 0x9f00001fU; |
582 | // Immediate goes in bits 23:5, 30:29 of ADR imm instruction, taken |
583 | // from bits 20:0 of X |
584 | *TargetPtr |= ((Result & 0xffc) << (5 - 2)); |
585 | *TargetPtr |= (Result & 0x3) << 29; |
586 | break; |
587 | } |
588 | } |
589 | } |
590 | |
591 | void RuntimeDyldELF::resolveARMRelocation(const SectionEntry &Section, |
592 | uint64_t Offset, uint32_t Value, |
593 | uint32_t Type, int32_t Addend) { |
594 | // TODO: Add Thumb relocations. |
595 | uint32_t *TargetPtr = |
596 | reinterpret_cast<uint32_t *>(Section.getAddressWithOffset(OffsetBytes: Offset)); |
597 | uint32_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset) & 0xFFFFFFFF; |
598 | Value += Addend; |
599 | |
600 | LLVM_DEBUG(dbgs() << "resolveARMRelocation, LocalAddress: " |
601 | << Section.getAddressWithOffset(Offset) |
602 | << " FinalAddress: " << format("%p" , FinalAddress) |
603 | << " Value: " << format("%x" , Value) |
604 | << " Type: " << format("%x" , Type) |
605 | << " Addend: " << format("%x" , Addend) << "\n" ); |
606 | |
607 | switch (Type) { |
608 | default: |
609 | llvm_unreachable("Not implemented relocation type!" ); |
610 | |
611 | case ELF::R_ARM_NONE: |
612 | break; |
613 | // Write a 31bit signed offset |
614 | case ELF::R_ARM_PREL31: |
615 | support::ulittle32_t::ref{TargetPtr} = |
616 | (support::ulittle32_t::ref{TargetPtr} & 0x80000000) | |
617 | ((Value - FinalAddress) & ~0x80000000); |
618 | break; |
619 | case ELF::R_ARM_TARGET1: |
620 | case ELF::R_ARM_ABS32: |
621 | support::ulittle32_t::ref{TargetPtr} = Value; |
622 | break; |
623 | // Write first 16 bit of 32 bit value to the mov instruction. |
624 | // Last 4 bit should be shifted. |
625 | case ELF::R_ARM_MOVW_ABS_NC: |
626 | case ELF::R_ARM_MOVT_ABS: |
627 | if (Type == ELF::R_ARM_MOVW_ABS_NC) |
628 | Value = Value & 0xFFFF; |
629 | else if (Type == ELF::R_ARM_MOVT_ABS) |
630 | Value = (Value >> 16) & 0xFFFF; |
631 | support::ulittle32_t::ref{TargetPtr} = |
632 | (support::ulittle32_t::ref{TargetPtr} & ~0x000F0FFF) | (Value & 0xFFF) | |
633 | (((Value >> 12) & 0xF) << 16); |
634 | break; |
635 | // Write 24 bit relative value to the branch instruction. |
636 | case ELF::R_ARM_PC24: // Fall through. |
637 | case ELF::R_ARM_CALL: // Fall through. |
638 | case ELF::R_ARM_JUMP24: |
639 | int32_t RelValue = static_cast<int32_t>(Value - FinalAddress - 8); |
640 | RelValue = (RelValue & 0x03FFFFFC) >> 2; |
641 | assert((support::ulittle32_t::ref{TargetPtr} & 0xFFFFFF) == 0xFFFFFE); |
642 | support::ulittle32_t::ref{TargetPtr} = |
643 | (support::ulittle32_t::ref{TargetPtr} & 0xFF000000) | RelValue; |
644 | break; |
645 | } |
646 | } |
647 | |
648 | void RuntimeDyldELF::setMipsABI(const ObjectFile &Obj) { |
649 | if (Arch == Triple::UnknownArch || |
650 | Triple::getArchTypePrefix(Kind: Arch) != "mips" ) { |
651 | IsMipsO32ABI = false; |
652 | IsMipsN32ABI = false; |
653 | IsMipsN64ABI = false; |
654 | return; |
655 | } |
656 | if (auto *E = dyn_cast<ELFObjectFileBase>(Val: &Obj)) { |
657 | unsigned AbiVariant = E->getPlatformFlags(); |
658 | IsMipsO32ABI = AbiVariant & ELF::EF_MIPS_ABI_O32; |
659 | IsMipsN32ABI = AbiVariant & ELF::EF_MIPS_ABI2; |
660 | } |
661 | IsMipsN64ABI = Obj.getFileFormatName() == "elf64-mips" ; |
662 | } |
663 | |
664 | // Return the .TOC. section and offset. |
665 | Error RuntimeDyldELF::findPPC64TOCSection(const ELFObjectFileBase &Obj, |
666 | ObjSectionToIDMap &LocalSections, |
667 | RelocationValueRef &Rel) { |
668 | // Set a default SectionID in case we do not find a TOC section below. |
669 | // This may happen for references to TOC base base (sym@toc, .odp |
670 | // relocation) without a .toc directive. In this case just use the |
671 | // first section (which is usually the .odp) since the code won't |
672 | // reference the .toc base directly. |
673 | Rel.SymbolName = nullptr; |
674 | Rel.SectionID = 0; |
675 | |
676 | // The TOC consists of sections .got, .toc, .tocbss, .plt in that |
677 | // order. The TOC starts where the first of these sections starts. |
678 | for (auto &Section : Obj.sections()) { |
679 | Expected<StringRef> NameOrErr = Section.getName(); |
680 | if (!NameOrErr) |
681 | return NameOrErr.takeError(); |
682 | StringRef SectionName = *NameOrErr; |
683 | |
684 | if (SectionName == ".got" |
685 | || SectionName == ".toc" |
686 | || SectionName == ".tocbss" |
687 | || SectionName == ".plt" ) { |
688 | if (auto SectionIDOrErr = |
689 | findOrEmitSection(Obj, Section, IsCode: false, LocalSections)) |
690 | Rel.SectionID = *SectionIDOrErr; |
691 | else |
692 | return SectionIDOrErr.takeError(); |
693 | break; |
694 | } |
695 | } |
696 | |
697 | // Per the ppc64-elf-linux ABI, The TOC base is TOC value plus 0x8000 |
698 | // thus permitting a full 64 Kbytes segment. |
699 | Rel.Addend = 0x8000; |
700 | |
701 | return Error::success(); |
702 | } |
703 | |
704 | // Returns the sections and offset associated with the ODP entry referenced |
705 | // by Symbol. |
706 | Error RuntimeDyldELF::findOPDEntrySection(const ELFObjectFileBase &Obj, |
707 | ObjSectionToIDMap &LocalSections, |
708 | RelocationValueRef &Rel) { |
709 | // Get the ELF symbol value (st_value) to compare with Relocation offset in |
710 | // .opd entries |
711 | for (section_iterator si = Obj.section_begin(), se = Obj.section_end(); |
712 | si != se; ++si) { |
713 | |
714 | Expected<section_iterator> RelSecOrErr = si->getRelocatedSection(); |
715 | if (!RelSecOrErr) |
716 | report_fatal_error(reason: Twine(toString(E: RelSecOrErr.takeError()))); |
717 | |
718 | section_iterator RelSecI = *RelSecOrErr; |
719 | if (RelSecI == Obj.section_end()) |
720 | continue; |
721 | |
722 | Expected<StringRef> NameOrErr = RelSecI->getName(); |
723 | if (!NameOrErr) |
724 | return NameOrErr.takeError(); |
725 | StringRef RelSectionName = *NameOrErr; |
726 | |
727 | if (RelSectionName != ".opd" ) |
728 | continue; |
729 | |
730 | for (elf_relocation_iterator i = si->relocation_begin(), |
731 | e = si->relocation_end(); |
732 | i != e;) { |
733 | // The R_PPC64_ADDR64 relocation indicates the first field |
734 | // of a .opd entry |
735 | uint64_t TypeFunc = i->getType(); |
736 | if (TypeFunc != ELF::R_PPC64_ADDR64) { |
737 | ++i; |
738 | continue; |
739 | } |
740 | |
741 | uint64_t TargetSymbolOffset = i->getOffset(); |
742 | symbol_iterator TargetSymbol = i->getSymbol(); |
743 | int64_t Addend; |
744 | if (auto AddendOrErr = i->getAddend()) |
745 | Addend = *AddendOrErr; |
746 | else |
747 | return AddendOrErr.takeError(); |
748 | |
749 | ++i; |
750 | if (i == e) |
751 | break; |
752 | |
753 | // Just check if following relocation is a R_PPC64_TOC |
754 | uint64_t TypeTOC = i->getType(); |
755 | if (TypeTOC != ELF::R_PPC64_TOC) |
756 | continue; |
757 | |
758 | // Finally compares the Symbol value and the target symbol offset |
759 | // to check if this .opd entry refers to the symbol the relocation |
760 | // points to. |
761 | if (Rel.Addend != (int64_t)TargetSymbolOffset) |
762 | continue; |
763 | |
764 | section_iterator TSI = Obj.section_end(); |
765 | if (auto TSIOrErr = TargetSymbol->getSection()) |
766 | TSI = *TSIOrErr; |
767 | else |
768 | return TSIOrErr.takeError(); |
769 | assert(TSI != Obj.section_end() && "TSI should refer to a valid section" ); |
770 | |
771 | bool IsCode = TSI->isText(); |
772 | if (auto SectionIDOrErr = findOrEmitSection(Obj, Section: *TSI, IsCode, |
773 | LocalSections)) |
774 | Rel.SectionID = *SectionIDOrErr; |
775 | else |
776 | return SectionIDOrErr.takeError(); |
777 | Rel.Addend = (intptr_t)Addend; |
778 | return Error::success(); |
779 | } |
780 | } |
781 | llvm_unreachable("Attempting to get address of ODP entry!" ); |
782 | } |
783 | |
784 | // Relocation masks following the #lo(value), #hi(value), #ha(value), |
785 | // #higher(value), #highera(value), #highest(value), and #highesta(value) |
786 | // macros defined in section 4.5.1. Relocation Types of the PPC-elf64abi |
787 | // document. |
788 | |
789 | static inline uint16_t applyPPClo(uint64_t value) { return value & 0xffff; } |
790 | |
791 | static inline uint16_t applyPPChi(uint64_t value) { |
792 | return (value >> 16) & 0xffff; |
793 | } |
794 | |
795 | static inline uint16_t applyPPCha (uint64_t value) { |
796 | return ((value + 0x8000) >> 16) & 0xffff; |
797 | } |
798 | |
799 | static inline uint16_t applyPPChigher(uint64_t value) { |
800 | return (value >> 32) & 0xffff; |
801 | } |
802 | |
803 | static inline uint16_t applyPPChighera (uint64_t value) { |
804 | return ((value + 0x8000) >> 32) & 0xffff; |
805 | } |
806 | |
807 | static inline uint16_t applyPPChighest(uint64_t value) { |
808 | return (value >> 48) & 0xffff; |
809 | } |
810 | |
811 | static inline uint16_t applyPPChighesta (uint64_t value) { |
812 | return ((value + 0x8000) >> 48) & 0xffff; |
813 | } |
814 | |
815 | void RuntimeDyldELF::resolvePPC32Relocation(const SectionEntry &Section, |
816 | uint64_t Offset, uint64_t Value, |
817 | uint32_t Type, int64_t Addend) { |
818 | uint8_t *LocalAddress = Section.getAddressWithOffset(OffsetBytes: Offset); |
819 | switch (Type) { |
820 | default: |
821 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
822 | break; |
823 | case ELF::R_PPC_ADDR16_LO: |
824 | writeInt16BE(Addr: LocalAddress, Value: applyPPClo(value: Value + Addend)); |
825 | break; |
826 | case ELF::R_PPC_ADDR16_HI: |
827 | writeInt16BE(Addr: LocalAddress, Value: applyPPChi(value: Value + Addend)); |
828 | break; |
829 | case ELF::R_PPC_ADDR16_HA: |
830 | writeInt16BE(Addr: LocalAddress, Value: applyPPCha(value: Value + Addend)); |
831 | break; |
832 | } |
833 | } |
834 | |
835 | void RuntimeDyldELF::resolvePPC64Relocation(const SectionEntry &Section, |
836 | uint64_t Offset, uint64_t Value, |
837 | uint32_t Type, int64_t Addend) { |
838 | uint8_t *LocalAddress = Section.getAddressWithOffset(OffsetBytes: Offset); |
839 | switch (Type) { |
840 | default: |
841 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
842 | break; |
843 | case ELF::R_PPC64_ADDR16: |
844 | writeInt16BE(Addr: LocalAddress, Value: applyPPClo(value: Value + Addend)); |
845 | break; |
846 | case ELF::R_PPC64_ADDR16_DS: |
847 | writeInt16BE(Addr: LocalAddress, Value: applyPPClo(value: Value + Addend) & ~3); |
848 | break; |
849 | case ELF::R_PPC64_ADDR16_LO: |
850 | writeInt16BE(Addr: LocalAddress, Value: applyPPClo(value: Value + Addend)); |
851 | break; |
852 | case ELF::R_PPC64_ADDR16_LO_DS: |
853 | writeInt16BE(Addr: LocalAddress, Value: applyPPClo(value: Value + Addend) & ~3); |
854 | break; |
855 | case ELF::R_PPC64_ADDR16_HI: |
856 | case ELF::R_PPC64_ADDR16_HIGH: |
857 | writeInt16BE(Addr: LocalAddress, Value: applyPPChi(value: Value + Addend)); |
858 | break; |
859 | case ELF::R_PPC64_ADDR16_HA: |
860 | case ELF::R_PPC64_ADDR16_HIGHA: |
861 | writeInt16BE(Addr: LocalAddress, Value: applyPPCha(value: Value + Addend)); |
862 | break; |
863 | case ELF::R_PPC64_ADDR16_HIGHER: |
864 | writeInt16BE(Addr: LocalAddress, Value: applyPPChigher(value: Value + Addend)); |
865 | break; |
866 | case ELF::R_PPC64_ADDR16_HIGHERA: |
867 | writeInt16BE(Addr: LocalAddress, Value: applyPPChighera(value: Value + Addend)); |
868 | break; |
869 | case ELF::R_PPC64_ADDR16_HIGHEST: |
870 | writeInt16BE(Addr: LocalAddress, Value: applyPPChighest(value: Value + Addend)); |
871 | break; |
872 | case ELF::R_PPC64_ADDR16_HIGHESTA: |
873 | writeInt16BE(Addr: LocalAddress, Value: applyPPChighesta(value: Value + Addend)); |
874 | break; |
875 | case ELF::R_PPC64_ADDR14: { |
876 | assert(((Value + Addend) & 3) == 0); |
877 | // Preserve the AA/LK bits in the branch instruction |
878 | uint8_t aalk = *(LocalAddress + 3); |
879 | writeInt16BE(Addr: LocalAddress + 2, Value: (aalk & 3) | ((Value + Addend) & 0xfffc)); |
880 | } break; |
881 | case ELF::R_PPC64_REL16_LO: { |
882 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
883 | uint64_t Delta = Value - FinalAddress + Addend; |
884 | writeInt16BE(Addr: LocalAddress, Value: applyPPClo(value: Delta)); |
885 | } break; |
886 | case ELF::R_PPC64_REL16_HI: { |
887 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
888 | uint64_t Delta = Value - FinalAddress + Addend; |
889 | writeInt16BE(Addr: LocalAddress, Value: applyPPChi(value: Delta)); |
890 | } break; |
891 | case ELF::R_PPC64_REL16_HA: { |
892 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
893 | uint64_t Delta = Value - FinalAddress + Addend; |
894 | writeInt16BE(Addr: LocalAddress, Value: applyPPCha(value: Delta)); |
895 | } break; |
896 | case ELF::R_PPC64_ADDR32: { |
897 | int64_t Result = static_cast<int64_t>(Value + Addend); |
898 | if (SignExtend64<32>(x: Result) != Result) |
899 | llvm_unreachable("Relocation R_PPC64_ADDR32 overflow" ); |
900 | writeInt32BE(Addr: LocalAddress, Value: Result); |
901 | } break; |
902 | case ELF::R_PPC64_REL24: { |
903 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
904 | int64_t delta = static_cast<int64_t>(Value - FinalAddress + Addend); |
905 | if (SignExtend64<26>(x: delta) != delta) |
906 | llvm_unreachable("Relocation R_PPC64_REL24 overflow" ); |
907 | // We preserve bits other than LI field, i.e. PO and AA/LK fields. |
908 | uint32_t Inst = readBytesUnaligned(Src: LocalAddress, Size: 4); |
909 | writeInt32BE(Addr: LocalAddress, Value: (Inst & 0xFC000003) | (delta & 0x03FFFFFC)); |
910 | } break; |
911 | case ELF::R_PPC64_REL32: { |
912 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
913 | int64_t delta = static_cast<int64_t>(Value - FinalAddress + Addend); |
914 | if (SignExtend64<32>(x: delta) != delta) |
915 | llvm_unreachable("Relocation R_PPC64_REL32 overflow" ); |
916 | writeInt32BE(Addr: LocalAddress, Value: delta); |
917 | } break; |
918 | case ELF::R_PPC64_REL64: { |
919 | uint64_t FinalAddress = Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
920 | uint64_t Delta = Value - FinalAddress + Addend; |
921 | writeInt64BE(Addr: LocalAddress, Value: Delta); |
922 | } break; |
923 | case ELF::R_PPC64_ADDR64: |
924 | writeInt64BE(Addr: LocalAddress, Value: Value + Addend); |
925 | break; |
926 | } |
927 | } |
928 | |
929 | void RuntimeDyldELF::resolveSystemZRelocation(const SectionEntry &Section, |
930 | uint64_t Offset, uint64_t Value, |
931 | uint32_t Type, int64_t Addend) { |
932 | uint8_t *LocalAddress = Section.getAddressWithOffset(OffsetBytes: Offset); |
933 | switch (Type) { |
934 | default: |
935 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
936 | break; |
937 | case ELF::R_390_PC16DBL: |
938 | case ELF::R_390_PLT16DBL: { |
939 | int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
940 | assert(int16_t(Delta / 2) * 2 == Delta && "R_390_PC16DBL overflow" ); |
941 | writeInt16BE(Addr: LocalAddress, Value: Delta / 2); |
942 | break; |
943 | } |
944 | case ELF::R_390_PC32DBL: |
945 | case ELF::R_390_PLT32DBL: { |
946 | int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
947 | assert(int32_t(Delta / 2) * 2 == Delta && "R_390_PC32DBL overflow" ); |
948 | writeInt32BE(Addr: LocalAddress, Value: Delta / 2); |
949 | break; |
950 | } |
951 | case ELF::R_390_PC16: { |
952 | int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
953 | assert(int16_t(Delta) == Delta && "R_390_PC16 overflow" ); |
954 | writeInt16BE(Addr: LocalAddress, Value: Delta); |
955 | break; |
956 | } |
957 | case ELF::R_390_PC32: { |
958 | int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
959 | assert(int32_t(Delta) == Delta && "R_390_PC32 overflow" ); |
960 | writeInt32BE(Addr: LocalAddress, Value: Delta); |
961 | break; |
962 | } |
963 | case ELF::R_390_PC64: { |
964 | int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(OffsetBytes: Offset); |
965 | writeInt64BE(Addr: LocalAddress, Value: Delta); |
966 | break; |
967 | } |
968 | case ELF::R_390_8: |
969 | *LocalAddress = (uint8_t)(Value + Addend); |
970 | break; |
971 | case ELF::R_390_16: |
972 | writeInt16BE(Addr: LocalAddress, Value: Value + Addend); |
973 | break; |
974 | case ELF::R_390_32: |
975 | writeInt32BE(Addr: LocalAddress, Value: Value + Addend); |
976 | break; |
977 | case ELF::R_390_64: |
978 | writeInt64BE(Addr: LocalAddress, Value: Value + Addend); |
979 | break; |
980 | } |
981 | } |
982 | |
983 | void RuntimeDyldELF::resolveBPFRelocation(const SectionEntry &Section, |
984 | uint64_t Offset, uint64_t Value, |
985 | uint32_t Type, int64_t Addend) { |
986 | bool isBE = Arch == Triple::bpfeb; |
987 | |
988 | switch (Type) { |
989 | default: |
990 | report_fatal_error(reason: "Relocation type not implemented yet!" ); |
991 | break; |
992 | case ELF::R_BPF_NONE: |
993 | case ELF::R_BPF_64_64: |
994 | case ELF::R_BPF_64_32: |
995 | case ELF::R_BPF_64_NODYLD32: |
996 | break; |
997 | case ELF::R_BPF_64_ABS64: { |
998 | write(isBE, P: Section.getAddressWithOffset(OffsetBytes: Offset), V: Value + Addend); |
999 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , (Value + Addend)) << " at " |
1000 | << format("%p\n" , Section.getAddressWithOffset(Offset))); |
1001 | break; |
1002 | } |
1003 | case ELF::R_BPF_64_ABS32: { |
1004 | Value += Addend; |
1005 | assert(Value <= UINT32_MAX); |
1006 | write(isBE, P: Section.getAddressWithOffset(OffsetBytes: Offset), V: static_cast<uint32_t>(Value)); |
1007 | LLVM_DEBUG(dbgs() << "Writing " << format("%p" , Value) << " at " |
1008 | << format("%p\n" , Section.getAddressWithOffset(Offset))); |
1009 | break; |
1010 | } |
1011 | } |
1012 | } |
1013 | |
1014 | // The target location for the relocation is described by RE.SectionID and |
1015 | // RE.Offset. RE.SectionID can be used to find the SectionEntry. Each |
1016 | // SectionEntry has three members describing its location. |
1017 | // SectionEntry::Address is the address at which the section has been loaded |
1018 | // into memory in the current (host) process. SectionEntry::LoadAddress is the |
1019 | // address that the section will have in the target process. |
1020 | // SectionEntry::ObjAddress is the address of the bits for this section in the |
1021 | // original emitted object image (also in the current address space). |
1022 | // |
1023 | // Relocations will be applied as if the section were loaded at |
1024 | // SectionEntry::LoadAddress, but they will be applied at an address based |
1025 | // on SectionEntry::Address. SectionEntry::ObjAddress will be used to refer to |
1026 | // Target memory contents if they are required for value calculations. |
1027 | // |
1028 | // The Value parameter here is the load address of the symbol for the |
1029 | // relocation to be applied. For relocations which refer to symbols in the |
1030 | // current object Value will be the LoadAddress of the section in which |
1031 | // the symbol resides (RE.Addend provides additional information about the |
1032 | // symbol location). For external symbols, Value will be the address of the |
1033 | // symbol in the target address space. |
1034 | void RuntimeDyldELF::resolveRelocation(const RelocationEntry &RE, |
1035 | uint64_t Value) { |
1036 | const SectionEntry &Section = Sections[RE.SectionID]; |
1037 | return resolveRelocation(Section, Offset: RE.Offset, Value, Type: RE.RelType, Addend: RE.Addend, |
1038 | SymOffset: RE.SymOffset, SectionID: RE.SectionID); |
1039 | } |
1040 | |
1041 | void RuntimeDyldELF::resolveRelocation(const SectionEntry &Section, |
1042 | uint64_t Offset, uint64_t Value, |
1043 | uint32_t Type, int64_t Addend, |
1044 | uint64_t SymOffset, SID SectionID) { |
1045 | switch (Arch) { |
1046 | case Triple::x86_64: |
1047 | resolveX86_64Relocation(Section, Offset, Value, Type, Addend, SymOffset); |
1048 | break; |
1049 | case Triple::x86: |
1050 | resolveX86Relocation(Section, Offset, Value: (uint32_t)(Value & 0xffffffffL), Type, |
1051 | Addend: (uint32_t)(Addend & 0xffffffffL)); |
1052 | break; |
1053 | case Triple::aarch64: |
1054 | case Triple::aarch64_be: |
1055 | resolveAArch64Relocation(Section, Offset, Value, Type, Addend); |
1056 | break; |
1057 | case Triple::arm: // Fall through. |
1058 | case Triple::armeb: |
1059 | case Triple::thumb: |
1060 | case Triple::thumbeb: |
1061 | resolveARMRelocation(Section, Offset, Value: (uint32_t)(Value & 0xffffffffL), Type, |
1062 | Addend: (uint32_t)(Addend & 0xffffffffL)); |
1063 | break; |
1064 | case Triple::ppc: // Fall through. |
1065 | case Triple::ppcle: |
1066 | resolvePPC32Relocation(Section, Offset, Value, Type, Addend); |
1067 | break; |
1068 | case Triple::ppc64: // Fall through. |
1069 | case Triple::ppc64le: |
1070 | resolvePPC64Relocation(Section, Offset, Value, Type, Addend); |
1071 | break; |
1072 | case Triple::systemz: |
1073 | resolveSystemZRelocation(Section, Offset, Value, Type, Addend); |
1074 | break; |
1075 | case Triple::bpfel: |
1076 | case Triple::bpfeb: |
1077 | resolveBPFRelocation(Section, Offset, Value, Type, Addend); |
1078 | break; |
1079 | default: |
1080 | llvm_unreachable("Unsupported CPU type!" ); |
1081 | } |
1082 | } |
1083 | |
1084 | void *RuntimeDyldELF::computePlaceholderAddress(unsigned SectionID, uint64_t Offset) const { |
1085 | return (void *)(Sections[SectionID].getObjAddress() + Offset); |
1086 | } |
1087 | |
1088 | void RuntimeDyldELF::processSimpleRelocation(unsigned SectionID, uint64_t Offset, unsigned RelType, RelocationValueRef Value) { |
1089 | RelocationEntry RE(SectionID, Offset, RelType, Value.Addend, Value.Offset); |
1090 | if (Value.SymbolName) |
1091 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1092 | else |
1093 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1094 | } |
1095 | |
1096 | uint32_t RuntimeDyldELF::getMatchingLoRelocation(uint32_t RelType, |
1097 | bool IsLocal) const { |
1098 | switch (RelType) { |
1099 | case ELF::R_MICROMIPS_GOT16: |
1100 | if (IsLocal) |
1101 | return ELF::R_MICROMIPS_LO16; |
1102 | break; |
1103 | case ELF::R_MICROMIPS_HI16: |
1104 | return ELF::R_MICROMIPS_LO16; |
1105 | case ELF::R_MIPS_GOT16: |
1106 | if (IsLocal) |
1107 | return ELF::R_MIPS_LO16; |
1108 | break; |
1109 | case ELF::R_MIPS_HI16: |
1110 | return ELF::R_MIPS_LO16; |
1111 | case ELF::R_MIPS_PCHI16: |
1112 | return ELF::R_MIPS_PCLO16; |
1113 | default: |
1114 | break; |
1115 | } |
1116 | return ELF::R_MIPS_NONE; |
1117 | } |
1118 | |
1119 | // Sometimes we don't need to create thunk for a branch. |
1120 | // This typically happens when branch target is located |
1121 | // in the same object file. In such case target is either |
1122 | // a weak symbol or symbol in a different executable section. |
1123 | // This function checks if branch target is located in the |
1124 | // same object file and if distance between source and target |
1125 | // fits R_AARCH64_CALL26 relocation. If both conditions are |
1126 | // met, it emits direct jump to the target and returns true. |
1127 | // Otherwise false is returned and thunk is created. |
1128 | bool RuntimeDyldELF::resolveAArch64ShortBranch( |
1129 | unsigned SectionID, relocation_iterator RelI, |
1130 | const RelocationValueRef &Value) { |
1131 | uint64_t TargetOffset; |
1132 | unsigned TargetSectionID; |
1133 | if (Value.SymbolName) { |
1134 | auto Loc = GlobalSymbolTable.find(Key: Value.SymbolName); |
1135 | |
1136 | // Don't create direct branch for external symbols. |
1137 | if (Loc == GlobalSymbolTable.end()) |
1138 | return false; |
1139 | |
1140 | const auto &SymInfo = Loc->second; |
1141 | |
1142 | TargetSectionID = SymInfo.getSectionID(); |
1143 | TargetOffset = SymInfo.getOffset(); |
1144 | } else { |
1145 | TargetSectionID = Value.SectionID; |
1146 | TargetOffset = 0; |
1147 | } |
1148 | |
1149 | // We don't actually know the load addresses at this point, so if the |
1150 | // branch is cross-section, we don't know exactly how far away it is. |
1151 | if (TargetSectionID != SectionID) |
1152 | return false; |
1153 | |
1154 | uint64_t SourceOffset = RelI->getOffset(); |
1155 | |
1156 | // R_AARCH64_CALL26 requires immediate to be in range -2^27 <= imm < 2^27 |
1157 | // If distance between source and target is out of range then we should |
1158 | // create thunk. |
1159 | if (!isInt<28>(x: TargetOffset + Value.Addend - SourceOffset)) |
1160 | return false; |
1161 | |
1162 | RelocationEntry RE(SectionID, SourceOffset, RelI->getType(), Value.Addend); |
1163 | if (Value.SymbolName) |
1164 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1165 | else |
1166 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1167 | |
1168 | return true; |
1169 | } |
1170 | |
1171 | void RuntimeDyldELF::resolveAArch64Branch(unsigned SectionID, |
1172 | const RelocationValueRef &Value, |
1173 | relocation_iterator RelI, |
1174 | StubMap &Stubs) { |
1175 | |
1176 | LLVM_DEBUG(dbgs() << "\t\tThis is an AArch64 branch relocation." ); |
1177 | SectionEntry &Section = Sections[SectionID]; |
1178 | |
1179 | uint64_t Offset = RelI->getOffset(); |
1180 | unsigned RelType = RelI->getType(); |
1181 | // Look for an existing stub. |
1182 | StubMap::const_iterator i = Stubs.find(x: Value); |
1183 | if (i != Stubs.end()) { |
1184 | resolveRelocation(Section, Offset, |
1185 | Value: Section.getLoadAddressWithOffset(OffsetBytes: i->second), Type: RelType, Addend: 0); |
1186 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1187 | } else if (!resolveAArch64ShortBranch(SectionID, RelI, Value)) { |
1188 | // Create a new stub function. |
1189 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1190 | Stubs[Value] = Section.getStubOffset(); |
1191 | uint8_t *StubTargetAddr = createStubFunction( |
1192 | Addr: Section.getAddressWithOffset(OffsetBytes: Section.getStubOffset())); |
1193 | |
1194 | RelocationEntry REmovz_g3(SectionID, StubTargetAddr - Section.getAddress(), |
1195 | ELF::R_AARCH64_MOVW_UABS_G3, Value.Addend); |
1196 | RelocationEntry REmovk_g2(SectionID, |
1197 | StubTargetAddr - Section.getAddress() + 4, |
1198 | ELF::R_AARCH64_MOVW_UABS_G2_NC, Value.Addend); |
1199 | RelocationEntry REmovk_g1(SectionID, |
1200 | StubTargetAddr - Section.getAddress() + 8, |
1201 | ELF::R_AARCH64_MOVW_UABS_G1_NC, Value.Addend); |
1202 | RelocationEntry REmovk_g0(SectionID, |
1203 | StubTargetAddr - Section.getAddress() + 12, |
1204 | ELF::R_AARCH64_MOVW_UABS_G0_NC, Value.Addend); |
1205 | |
1206 | if (Value.SymbolName) { |
1207 | addRelocationForSymbol(RE: REmovz_g3, SymbolName: Value.SymbolName); |
1208 | addRelocationForSymbol(RE: REmovk_g2, SymbolName: Value.SymbolName); |
1209 | addRelocationForSymbol(RE: REmovk_g1, SymbolName: Value.SymbolName); |
1210 | addRelocationForSymbol(RE: REmovk_g0, SymbolName: Value.SymbolName); |
1211 | } else { |
1212 | addRelocationForSection(RE: REmovz_g3, SectionID: Value.SectionID); |
1213 | addRelocationForSection(RE: REmovk_g2, SectionID: Value.SectionID); |
1214 | addRelocationForSection(RE: REmovk_g1, SectionID: Value.SectionID); |
1215 | addRelocationForSection(RE: REmovk_g0, SectionID: Value.SectionID); |
1216 | } |
1217 | resolveRelocation(Section, Offset, |
1218 | Value: Section.getLoadAddressWithOffset(OffsetBytes: Section.getStubOffset()), |
1219 | Type: RelType, Addend: 0); |
1220 | Section.advanceStubOffset(StubSize: getMaxStubSize()); |
1221 | } |
1222 | } |
1223 | |
1224 | Expected<relocation_iterator> |
1225 | RuntimeDyldELF::processRelocationRef( |
1226 | unsigned SectionID, relocation_iterator RelI, const ObjectFile &O, |
1227 | ObjSectionToIDMap &ObjSectionToID, StubMap &Stubs) { |
1228 | const auto &Obj = cast<ELFObjectFileBase>(Val: O); |
1229 | uint64_t RelType = RelI->getType(); |
1230 | int64_t Addend = 0; |
1231 | if (Expected<int64_t> AddendOrErr = ELFRelocationRef(*RelI).getAddend()) |
1232 | Addend = *AddendOrErr; |
1233 | else |
1234 | consumeError(Err: AddendOrErr.takeError()); |
1235 | elf_symbol_iterator Symbol = RelI->getSymbol(); |
1236 | |
1237 | // Obtain the symbol name which is referenced in the relocation |
1238 | StringRef TargetName; |
1239 | if (Symbol != Obj.symbol_end()) { |
1240 | if (auto TargetNameOrErr = Symbol->getName()) |
1241 | TargetName = *TargetNameOrErr; |
1242 | else |
1243 | return TargetNameOrErr.takeError(); |
1244 | } |
1245 | LLVM_DEBUG(dbgs() << "\t\tRelType: " << RelType << " Addend: " << Addend |
1246 | << " TargetName: " << TargetName << "\n" ); |
1247 | RelocationValueRef Value; |
1248 | // First search for the symbol in the local symbol table |
1249 | SymbolRef::Type SymType = SymbolRef::ST_Unknown; |
1250 | |
1251 | // Search for the symbol in the global symbol table |
1252 | RTDyldSymbolTable::const_iterator gsi = GlobalSymbolTable.end(); |
1253 | if (Symbol != Obj.symbol_end()) { |
1254 | gsi = GlobalSymbolTable.find(Key: TargetName.data()); |
1255 | Expected<SymbolRef::Type> SymTypeOrErr = Symbol->getType(); |
1256 | if (!SymTypeOrErr) { |
1257 | std::string Buf; |
1258 | raw_string_ostream OS(Buf); |
1259 | logAllUnhandledErrors(E: SymTypeOrErr.takeError(), OS); |
1260 | report_fatal_error(reason: Twine(OS.str())); |
1261 | } |
1262 | SymType = *SymTypeOrErr; |
1263 | } |
1264 | if (gsi != GlobalSymbolTable.end()) { |
1265 | const auto &SymInfo = gsi->second; |
1266 | Value.SectionID = SymInfo.getSectionID(); |
1267 | Value.Offset = SymInfo.getOffset(); |
1268 | Value.Addend = SymInfo.getOffset() + Addend; |
1269 | } else { |
1270 | switch (SymType) { |
1271 | case SymbolRef::ST_Debug: { |
1272 | // TODO: Now ELF SymbolRef::ST_Debug = STT_SECTION, it's not obviously |
1273 | // and can be changed by another developers. Maybe best way is add |
1274 | // a new symbol type ST_Section to SymbolRef and use it. |
1275 | auto SectionOrErr = Symbol->getSection(); |
1276 | if (!SectionOrErr) { |
1277 | std::string Buf; |
1278 | raw_string_ostream OS(Buf); |
1279 | logAllUnhandledErrors(E: SectionOrErr.takeError(), OS); |
1280 | report_fatal_error(reason: Twine(OS.str())); |
1281 | } |
1282 | section_iterator si = *SectionOrErr; |
1283 | if (si == Obj.section_end()) |
1284 | llvm_unreachable("Symbol section not found, bad object file format!" ); |
1285 | LLVM_DEBUG(dbgs() << "\t\tThis is section symbol\n" ); |
1286 | bool isCode = si->isText(); |
1287 | if (auto SectionIDOrErr = findOrEmitSection(Obj, Section: (*si), IsCode: isCode, |
1288 | LocalSections&: ObjSectionToID)) |
1289 | Value.SectionID = *SectionIDOrErr; |
1290 | else |
1291 | return SectionIDOrErr.takeError(); |
1292 | Value.Addend = Addend; |
1293 | break; |
1294 | } |
1295 | case SymbolRef::ST_Data: |
1296 | case SymbolRef::ST_Function: |
1297 | case SymbolRef::ST_Other: |
1298 | case SymbolRef::ST_Unknown: { |
1299 | Value.SymbolName = TargetName.data(); |
1300 | Value.Addend = Addend; |
1301 | |
1302 | // Absolute relocations will have a zero symbol ID (STN_UNDEF), which |
1303 | // will manifest here as a NULL symbol name. |
1304 | // We can set this as a valid (but empty) symbol name, and rely |
1305 | // on addRelocationForSymbol to handle this. |
1306 | if (!Value.SymbolName) |
1307 | Value.SymbolName = "" ; |
1308 | break; |
1309 | } |
1310 | default: |
1311 | llvm_unreachable("Unresolved symbol type!" ); |
1312 | break; |
1313 | } |
1314 | } |
1315 | |
1316 | uint64_t Offset = RelI->getOffset(); |
1317 | |
1318 | LLVM_DEBUG(dbgs() << "\t\tSectionID: " << SectionID << " Offset: " << Offset |
1319 | << "\n" ); |
1320 | if ((Arch == Triple::aarch64 || Arch == Triple::aarch64_be)) { |
1321 | if ((RelType == ELF::R_AARCH64_CALL26 || |
1322 | RelType == ELF::R_AARCH64_JUMP26) && |
1323 | MemMgr.allowStubAllocation()) { |
1324 | resolveAArch64Branch(SectionID, Value, RelI, Stubs); |
1325 | } else if (RelType == ELF::R_AARCH64_ADR_GOT_PAGE) { |
1326 | // Create new GOT entry or find existing one. If GOT entry is |
1327 | // to be created, then we also emit ABS64 relocation for it. |
1328 | uint64_t GOTOffset = findOrAllocGOTEntry(Value, GOTRelType: ELF::R_AARCH64_ABS64); |
1329 | resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset: GOTOffset + Addend, |
1330 | Type: ELF::R_AARCH64_ADR_PREL_PG_HI21); |
1331 | |
1332 | } else if (RelType == ELF::R_AARCH64_LD64_GOT_LO12_NC) { |
1333 | uint64_t GOTOffset = findOrAllocGOTEntry(Value, GOTRelType: ELF::R_AARCH64_ABS64); |
1334 | resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset: GOTOffset + Addend, |
1335 | Type: ELF::R_AARCH64_LDST64_ABS_LO12_NC); |
1336 | } else { |
1337 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1338 | } |
1339 | } else if (Arch == Triple::arm) { |
1340 | if (RelType == ELF::R_ARM_PC24 || RelType == ELF::R_ARM_CALL || |
1341 | RelType == ELF::R_ARM_JUMP24) { |
1342 | // This is an ARM branch relocation, need to use a stub function. |
1343 | LLVM_DEBUG(dbgs() << "\t\tThis is an ARM branch relocation.\n" ); |
1344 | SectionEntry &Section = Sections[SectionID]; |
1345 | |
1346 | // Look for an existing stub. |
1347 | StubMap::const_iterator i = Stubs.find(x: Value); |
1348 | if (i != Stubs.end()) { |
1349 | resolveRelocation(Section, Offset, |
1350 | Value: Section.getLoadAddressWithOffset(OffsetBytes: i->second), Type: RelType, |
1351 | Addend: 0); |
1352 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1353 | } else { |
1354 | // Create a new stub function. |
1355 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1356 | Stubs[Value] = Section.getStubOffset(); |
1357 | uint8_t *StubTargetAddr = createStubFunction( |
1358 | Addr: Section.getAddressWithOffset(OffsetBytes: Section.getStubOffset())); |
1359 | RelocationEntry RE(SectionID, StubTargetAddr - Section.getAddress(), |
1360 | ELF::R_ARM_ABS32, Value.Addend); |
1361 | if (Value.SymbolName) |
1362 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1363 | else |
1364 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1365 | |
1366 | resolveRelocation( |
1367 | Section, Offset, |
1368 | Value: Section.getLoadAddressWithOffset(OffsetBytes: Section.getStubOffset()), Type: RelType, |
1369 | Addend: 0); |
1370 | Section.advanceStubOffset(StubSize: getMaxStubSize()); |
1371 | } |
1372 | } else { |
1373 | uint32_t *Placeholder = |
1374 | reinterpret_cast<uint32_t*>(computePlaceholderAddress(SectionID, Offset)); |
1375 | if (RelType == ELF::R_ARM_PREL31 || RelType == ELF::R_ARM_TARGET1 || |
1376 | RelType == ELF::R_ARM_ABS32) { |
1377 | Value.Addend += *Placeholder; |
1378 | } else if (RelType == ELF::R_ARM_MOVW_ABS_NC || RelType == ELF::R_ARM_MOVT_ABS) { |
1379 | // See ELF for ARM documentation |
1380 | Value.Addend += (int16_t)((*Placeholder & 0xFFF) | (((*Placeholder >> 16) & 0xF) << 12)); |
1381 | } |
1382 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1383 | } |
1384 | } else if (IsMipsO32ABI) { |
1385 | uint8_t *Placeholder = reinterpret_cast<uint8_t *>( |
1386 | computePlaceholderAddress(SectionID, Offset)); |
1387 | uint32_t Opcode = readBytesUnaligned(Src: Placeholder, Size: 4); |
1388 | if (RelType == ELF::R_MIPS_26) { |
1389 | // This is an Mips branch relocation, need to use a stub function. |
1390 | LLVM_DEBUG(dbgs() << "\t\tThis is a Mips branch relocation." ); |
1391 | SectionEntry &Section = Sections[SectionID]; |
1392 | |
1393 | // Extract the addend from the instruction. |
1394 | // We shift up by two since the Value will be down shifted again |
1395 | // when applying the relocation. |
1396 | uint32_t Addend = (Opcode & 0x03ffffff) << 2; |
1397 | |
1398 | Value.Addend += Addend; |
1399 | |
1400 | // Look up for existing stub. |
1401 | StubMap::const_iterator i = Stubs.find(x: Value); |
1402 | if (i != Stubs.end()) { |
1403 | RelocationEntry RE(SectionID, Offset, RelType, i->second); |
1404 | addRelocationForSection(RE, SectionID); |
1405 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1406 | } else { |
1407 | // Create a new stub function. |
1408 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1409 | Stubs[Value] = Section.getStubOffset(); |
1410 | |
1411 | unsigned AbiVariant = Obj.getPlatformFlags(); |
1412 | |
1413 | uint8_t *StubTargetAddr = createStubFunction( |
1414 | Addr: Section.getAddressWithOffset(OffsetBytes: Section.getStubOffset()), AbiVariant); |
1415 | |
1416 | // Creating Hi and Lo relocations for the filled stub instructions. |
1417 | RelocationEntry REHi(SectionID, StubTargetAddr - Section.getAddress(), |
1418 | ELF::R_MIPS_HI16, Value.Addend); |
1419 | RelocationEntry RELo(SectionID, |
1420 | StubTargetAddr - Section.getAddress() + 4, |
1421 | ELF::R_MIPS_LO16, Value.Addend); |
1422 | |
1423 | if (Value.SymbolName) { |
1424 | addRelocationForSymbol(RE: REHi, SymbolName: Value.SymbolName); |
1425 | addRelocationForSymbol(RE: RELo, SymbolName: Value.SymbolName); |
1426 | } else { |
1427 | addRelocationForSection(RE: REHi, SectionID: Value.SectionID); |
1428 | addRelocationForSection(RE: RELo, SectionID: Value.SectionID); |
1429 | } |
1430 | |
1431 | RelocationEntry RE(SectionID, Offset, RelType, Section.getStubOffset()); |
1432 | addRelocationForSection(RE, SectionID); |
1433 | Section.advanceStubOffset(StubSize: getMaxStubSize()); |
1434 | } |
1435 | } else if (RelType == ELF::R_MIPS_HI16 || RelType == ELF::R_MIPS_PCHI16) { |
1436 | int64_t Addend = (Opcode & 0x0000ffff) << 16; |
1437 | RelocationEntry RE(SectionID, Offset, RelType, Addend); |
1438 | PendingRelocs.push_back(Elt: std::make_pair(x&: Value, y&: RE)); |
1439 | } else if (RelType == ELF::R_MIPS_LO16 || RelType == ELF::R_MIPS_PCLO16) { |
1440 | int64_t Addend = Value.Addend + SignExtend32<16>(X: Opcode & 0x0000ffff); |
1441 | for (auto I = PendingRelocs.begin(); I != PendingRelocs.end();) { |
1442 | const RelocationValueRef &MatchingValue = I->first; |
1443 | RelocationEntry &Reloc = I->second; |
1444 | if (MatchingValue == Value && |
1445 | RelType == getMatchingLoRelocation(RelType: Reloc.RelType) && |
1446 | SectionID == Reloc.SectionID) { |
1447 | Reloc.Addend += Addend; |
1448 | if (Value.SymbolName) |
1449 | addRelocationForSymbol(RE: Reloc, SymbolName: Value.SymbolName); |
1450 | else |
1451 | addRelocationForSection(RE: Reloc, SectionID: Value.SectionID); |
1452 | I = PendingRelocs.erase(CI: I); |
1453 | } else |
1454 | ++I; |
1455 | } |
1456 | RelocationEntry RE(SectionID, Offset, RelType, Addend); |
1457 | if (Value.SymbolName) |
1458 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1459 | else |
1460 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1461 | } else { |
1462 | if (RelType == ELF::R_MIPS_32) |
1463 | Value.Addend += Opcode; |
1464 | else if (RelType == ELF::R_MIPS_PC16) |
1465 | Value.Addend += SignExtend32<18>(X: (Opcode & 0x0000ffff) << 2); |
1466 | else if (RelType == ELF::R_MIPS_PC19_S2) |
1467 | Value.Addend += SignExtend32<21>(X: (Opcode & 0x0007ffff) << 2); |
1468 | else if (RelType == ELF::R_MIPS_PC21_S2) |
1469 | Value.Addend += SignExtend32<23>(X: (Opcode & 0x001fffff) << 2); |
1470 | else if (RelType == ELF::R_MIPS_PC26_S2) |
1471 | Value.Addend += SignExtend32<28>(X: (Opcode & 0x03ffffff) << 2); |
1472 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1473 | } |
1474 | } else if (IsMipsN32ABI || IsMipsN64ABI) { |
1475 | uint32_t r_type = RelType & 0xff; |
1476 | RelocationEntry RE(SectionID, Offset, RelType, Value.Addend); |
1477 | if (r_type == ELF::R_MIPS_CALL16 || r_type == ELF::R_MIPS_GOT_PAGE |
1478 | || r_type == ELF::R_MIPS_GOT_DISP) { |
1479 | StringMap<uint64_t>::iterator i = GOTSymbolOffsets.find(Key: TargetName); |
1480 | if (i != GOTSymbolOffsets.end()) |
1481 | RE.SymOffset = i->second; |
1482 | else { |
1483 | RE.SymOffset = allocateGOTEntries(no: 1); |
1484 | GOTSymbolOffsets[TargetName] = RE.SymOffset; |
1485 | } |
1486 | if (Value.SymbolName) |
1487 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1488 | else |
1489 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1490 | } else if (RelType == ELF::R_MIPS_26) { |
1491 | // This is an Mips branch relocation, need to use a stub function. |
1492 | LLVM_DEBUG(dbgs() << "\t\tThis is a Mips branch relocation." ); |
1493 | SectionEntry &Section = Sections[SectionID]; |
1494 | |
1495 | // Look up for existing stub. |
1496 | StubMap::const_iterator i = Stubs.find(x: Value); |
1497 | if (i != Stubs.end()) { |
1498 | RelocationEntry RE(SectionID, Offset, RelType, i->second); |
1499 | addRelocationForSection(RE, SectionID); |
1500 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1501 | } else { |
1502 | // Create a new stub function. |
1503 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1504 | Stubs[Value] = Section.getStubOffset(); |
1505 | |
1506 | unsigned AbiVariant = Obj.getPlatformFlags(); |
1507 | |
1508 | uint8_t *StubTargetAddr = createStubFunction( |
1509 | Addr: Section.getAddressWithOffset(OffsetBytes: Section.getStubOffset()), AbiVariant); |
1510 | |
1511 | if (IsMipsN32ABI) { |
1512 | // Creating Hi and Lo relocations for the filled stub instructions. |
1513 | RelocationEntry REHi(SectionID, StubTargetAddr - Section.getAddress(), |
1514 | ELF::R_MIPS_HI16, Value.Addend); |
1515 | RelocationEntry RELo(SectionID, |
1516 | StubTargetAddr - Section.getAddress() + 4, |
1517 | ELF::R_MIPS_LO16, Value.Addend); |
1518 | if (Value.SymbolName) { |
1519 | addRelocationForSymbol(RE: REHi, SymbolName: Value.SymbolName); |
1520 | addRelocationForSymbol(RE: RELo, SymbolName: Value.SymbolName); |
1521 | } else { |
1522 | addRelocationForSection(RE: REHi, SectionID: Value.SectionID); |
1523 | addRelocationForSection(RE: RELo, SectionID: Value.SectionID); |
1524 | } |
1525 | } else { |
1526 | // Creating Highest, Higher, Hi and Lo relocations for the filled stub |
1527 | // instructions. |
1528 | RelocationEntry REHighest(SectionID, |
1529 | StubTargetAddr - Section.getAddress(), |
1530 | ELF::R_MIPS_HIGHEST, Value.Addend); |
1531 | RelocationEntry REHigher(SectionID, |
1532 | StubTargetAddr - Section.getAddress() + 4, |
1533 | ELF::R_MIPS_HIGHER, Value.Addend); |
1534 | RelocationEntry REHi(SectionID, |
1535 | StubTargetAddr - Section.getAddress() + 12, |
1536 | ELF::R_MIPS_HI16, Value.Addend); |
1537 | RelocationEntry RELo(SectionID, |
1538 | StubTargetAddr - Section.getAddress() + 20, |
1539 | ELF::R_MIPS_LO16, Value.Addend); |
1540 | if (Value.SymbolName) { |
1541 | addRelocationForSymbol(RE: REHighest, SymbolName: Value.SymbolName); |
1542 | addRelocationForSymbol(RE: REHigher, SymbolName: Value.SymbolName); |
1543 | addRelocationForSymbol(RE: REHi, SymbolName: Value.SymbolName); |
1544 | addRelocationForSymbol(RE: RELo, SymbolName: Value.SymbolName); |
1545 | } else { |
1546 | addRelocationForSection(RE: REHighest, SectionID: Value.SectionID); |
1547 | addRelocationForSection(RE: REHigher, SectionID: Value.SectionID); |
1548 | addRelocationForSection(RE: REHi, SectionID: Value.SectionID); |
1549 | addRelocationForSection(RE: RELo, SectionID: Value.SectionID); |
1550 | } |
1551 | } |
1552 | RelocationEntry RE(SectionID, Offset, RelType, Section.getStubOffset()); |
1553 | addRelocationForSection(RE, SectionID); |
1554 | Section.advanceStubOffset(StubSize: getMaxStubSize()); |
1555 | } |
1556 | } else { |
1557 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1558 | } |
1559 | |
1560 | } else if (Arch == Triple::ppc64 || Arch == Triple::ppc64le) { |
1561 | if (RelType == ELF::R_PPC64_REL24) { |
1562 | // Determine ABI variant in use for this object. |
1563 | unsigned AbiVariant = Obj.getPlatformFlags(); |
1564 | AbiVariant &= ELF::EF_PPC64_ABI; |
1565 | // A PPC branch relocation will need a stub function if the target is |
1566 | // an external symbol (either Value.SymbolName is set, or SymType is |
1567 | // Symbol::ST_Unknown) or if the target address is not within the |
1568 | // signed 24-bits branch address. |
1569 | SectionEntry &Section = Sections[SectionID]; |
1570 | uint8_t *Target = Section.getAddressWithOffset(OffsetBytes: Offset); |
1571 | bool RangeOverflow = false; |
1572 | bool IsExtern = Value.SymbolName || SymType == SymbolRef::ST_Unknown; |
1573 | if (!IsExtern) { |
1574 | if (AbiVariant != 2) { |
1575 | // In the ELFv1 ABI, a function call may point to the .opd entry, |
1576 | // so the final symbol value is calculated based on the relocation |
1577 | // values in the .opd section. |
1578 | if (auto Err = findOPDEntrySection(Obj, LocalSections&: ObjSectionToID, Rel&: Value)) |
1579 | return std::move(Err); |
1580 | } else { |
1581 | // In the ELFv2 ABI, a function symbol may provide a local entry |
1582 | // point, which must be used for direct calls. |
1583 | if (Value.SectionID == SectionID){ |
1584 | uint8_t SymOther = Symbol->getOther(); |
1585 | Value.Addend += ELF::decodePPC64LocalEntryOffset(Other: SymOther); |
1586 | } |
1587 | } |
1588 | uint8_t *RelocTarget = |
1589 | Sections[Value.SectionID].getAddressWithOffset(OffsetBytes: Value.Addend); |
1590 | int64_t delta = static_cast<int64_t>(Target - RelocTarget); |
1591 | // If it is within 26-bits branch range, just set the branch target |
1592 | if (SignExtend64<26>(x: delta) != delta) { |
1593 | RangeOverflow = true; |
1594 | } else if ((AbiVariant != 2) || |
1595 | (AbiVariant == 2 && Value.SectionID == SectionID)) { |
1596 | RelocationEntry RE(SectionID, Offset, RelType, Value.Addend); |
1597 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1598 | } |
1599 | } |
1600 | if (IsExtern || (AbiVariant == 2 && Value.SectionID != SectionID) || |
1601 | RangeOverflow) { |
1602 | // It is an external symbol (either Value.SymbolName is set, or |
1603 | // SymType is SymbolRef::ST_Unknown) or out of range. |
1604 | StubMap::const_iterator i = Stubs.find(x: Value); |
1605 | if (i != Stubs.end()) { |
1606 | // Symbol function stub already created, just relocate to it |
1607 | resolveRelocation(Section, Offset, |
1608 | Value: Section.getLoadAddressWithOffset(OffsetBytes: i->second), |
1609 | Type: RelType, Addend: 0); |
1610 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1611 | } else { |
1612 | // Create a new stub function. |
1613 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1614 | Stubs[Value] = Section.getStubOffset(); |
1615 | uint8_t *StubTargetAddr = createStubFunction( |
1616 | Addr: Section.getAddressWithOffset(OffsetBytes: Section.getStubOffset()), |
1617 | AbiVariant); |
1618 | RelocationEntry RE(SectionID, StubTargetAddr - Section.getAddress(), |
1619 | ELF::R_PPC64_ADDR64, Value.Addend); |
1620 | |
1621 | // Generates the 64-bits address loads as exemplified in section |
1622 | // 4.5.1 in PPC64 ELF ABI. Note that the relocations need to |
1623 | // apply to the low part of the instructions, so we have to update |
1624 | // the offset according to the target endianness. |
1625 | uint64_t StubRelocOffset = StubTargetAddr - Section.getAddress(); |
1626 | if (!IsTargetLittleEndian) |
1627 | StubRelocOffset += 2; |
1628 | |
1629 | RelocationEntry REhst(SectionID, StubRelocOffset + 0, |
1630 | ELF::R_PPC64_ADDR16_HIGHEST, Value.Addend); |
1631 | RelocationEntry REhr(SectionID, StubRelocOffset + 4, |
1632 | ELF::R_PPC64_ADDR16_HIGHER, Value.Addend); |
1633 | RelocationEntry REh(SectionID, StubRelocOffset + 12, |
1634 | ELF::R_PPC64_ADDR16_HI, Value.Addend); |
1635 | RelocationEntry REl(SectionID, StubRelocOffset + 16, |
1636 | ELF::R_PPC64_ADDR16_LO, Value.Addend); |
1637 | |
1638 | if (Value.SymbolName) { |
1639 | addRelocationForSymbol(RE: REhst, SymbolName: Value.SymbolName); |
1640 | addRelocationForSymbol(RE: REhr, SymbolName: Value.SymbolName); |
1641 | addRelocationForSymbol(RE: REh, SymbolName: Value.SymbolName); |
1642 | addRelocationForSymbol(RE: REl, SymbolName: Value.SymbolName); |
1643 | } else { |
1644 | addRelocationForSection(RE: REhst, SectionID: Value.SectionID); |
1645 | addRelocationForSection(RE: REhr, SectionID: Value.SectionID); |
1646 | addRelocationForSection(RE: REh, SectionID: Value.SectionID); |
1647 | addRelocationForSection(RE: REl, SectionID: Value.SectionID); |
1648 | } |
1649 | |
1650 | resolveRelocation( |
1651 | Section, Offset, |
1652 | Value: Section.getLoadAddressWithOffset(OffsetBytes: Section.getStubOffset()), |
1653 | Type: RelType, Addend: 0); |
1654 | Section.advanceStubOffset(StubSize: getMaxStubSize()); |
1655 | } |
1656 | if (IsExtern || (AbiVariant == 2 && Value.SectionID != SectionID)) { |
1657 | // Restore the TOC for external calls |
1658 | if (AbiVariant == 2) |
1659 | writeInt32BE(Addr: Target + 4, Value: 0xE8410018); // ld r2,24(r1) |
1660 | else |
1661 | writeInt32BE(Addr: Target + 4, Value: 0xE8410028); // ld r2,40(r1) |
1662 | } |
1663 | } |
1664 | } else if (RelType == ELF::R_PPC64_TOC16 || |
1665 | RelType == ELF::R_PPC64_TOC16_DS || |
1666 | RelType == ELF::R_PPC64_TOC16_LO || |
1667 | RelType == ELF::R_PPC64_TOC16_LO_DS || |
1668 | RelType == ELF::R_PPC64_TOC16_HI || |
1669 | RelType == ELF::R_PPC64_TOC16_HA) { |
1670 | // These relocations are supposed to subtract the TOC address from |
1671 | // the final value. This does not fit cleanly into the RuntimeDyld |
1672 | // scheme, since there may be *two* sections involved in determining |
1673 | // the relocation value (the section of the symbol referred to by the |
1674 | // relocation, and the TOC section associated with the current module). |
1675 | // |
1676 | // Fortunately, these relocations are currently only ever generated |
1677 | // referring to symbols that themselves reside in the TOC, which means |
1678 | // that the two sections are actually the same. Thus they cancel out |
1679 | // and we can immediately resolve the relocation right now. |
1680 | switch (RelType) { |
1681 | case ELF::R_PPC64_TOC16: RelType = ELF::R_PPC64_ADDR16; break; |
1682 | case ELF::R_PPC64_TOC16_DS: RelType = ELF::R_PPC64_ADDR16_DS; break; |
1683 | case ELF::R_PPC64_TOC16_LO: RelType = ELF::R_PPC64_ADDR16_LO; break; |
1684 | case ELF::R_PPC64_TOC16_LO_DS: RelType = ELF::R_PPC64_ADDR16_LO_DS; break; |
1685 | case ELF::R_PPC64_TOC16_HI: RelType = ELF::R_PPC64_ADDR16_HI; break; |
1686 | case ELF::R_PPC64_TOC16_HA: RelType = ELF::R_PPC64_ADDR16_HA; break; |
1687 | default: llvm_unreachable("Wrong relocation type." ); |
1688 | } |
1689 | |
1690 | RelocationValueRef TOCValue; |
1691 | if (auto Err = findPPC64TOCSection(Obj, LocalSections&: ObjSectionToID, Rel&: TOCValue)) |
1692 | return std::move(Err); |
1693 | if (Value.SymbolName || Value.SectionID != TOCValue.SectionID) |
1694 | llvm_unreachable("Unsupported TOC relocation." ); |
1695 | Value.Addend -= TOCValue.Addend; |
1696 | resolveRelocation(Section: Sections[SectionID], Offset, Value: Value.Addend, Type: RelType, Addend: 0); |
1697 | } else { |
1698 | // There are two ways to refer to the TOC address directly: either |
1699 | // via a ELF::R_PPC64_TOC relocation (where both symbol and addend are |
1700 | // ignored), or via any relocation that refers to the magic ".TOC." |
1701 | // symbols (in which case the addend is respected). |
1702 | if (RelType == ELF::R_PPC64_TOC) { |
1703 | RelType = ELF::R_PPC64_ADDR64; |
1704 | if (auto Err = findPPC64TOCSection(Obj, LocalSections&: ObjSectionToID, Rel&: Value)) |
1705 | return std::move(Err); |
1706 | } else if (TargetName == ".TOC." ) { |
1707 | if (auto Err = findPPC64TOCSection(Obj, LocalSections&: ObjSectionToID, Rel&: Value)) |
1708 | return std::move(Err); |
1709 | Value.Addend += Addend; |
1710 | } |
1711 | |
1712 | RelocationEntry RE(SectionID, Offset, RelType, Value.Addend); |
1713 | |
1714 | if (Value.SymbolName) |
1715 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1716 | else |
1717 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1718 | } |
1719 | } else if (Arch == Triple::systemz && |
1720 | (RelType == ELF::R_390_PLT32DBL || RelType == ELF::R_390_GOTENT)) { |
1721 | // Create function stubs for both PLT and GOT references, regardless of |
1722 | // whether the GOT reference is to data or code. The stub contains the |
1723 | // full address of the symbol, as needed by GOT references, and the |
1724 | // executable part only adds an overhead of 8 bytes. |
1725 | // |
1726 | // We could try to conserve space by allocating the code and data |
1727 | // parts of the stub separately. However, as things stand, we allocate |
1728 | // a stub for every relocation, so using a GOT in JIT code should be |
1729 | // no less space efficient than using an explicit constant pool. |
1730 | LLVM_DEBUG(dbgs() << "\t\tThis is a SystemZ indirect relocation." ); |
1731 | SectionEntry &Section = Sections[SectionID]; |
1732 | |
1733 | // Look for an existing stub. |
1734 | StubMap::const_iterator i = Stubs.find(x: Value); |
1735 | uintptr_t StubAddress; |
1736 | if (i != Stubs.end()) { |
1737 | StubAddress = uintptr_t(Section.getAddressWithOffset(OffsetBytes: i->second)); |
1738 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1739 | } else { |
1740 | // Create a new stub function. |
1741 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1742 | |
1743 | uintptr_t BaseAddress = uintptr_t(Section.getAddress()); |
1744 | StubAddress = |
1745 | alignTo(Size: BaseAddress + Section.getStubOffset(), A: getStubAlignment()); |
1746 | unsigned StubOffset = StubAddress - BaseAddress; |
1747 | |
1748 | Stubs[Value] = StubOffset; |
1749 | createStubFunction(Addr: (uint8_t *)StubAddress); |
1750 | RelocationEntry RE(SectionID, StubOffset + 8, ELF::R_390_64, |
1751 | Value.Offset); |
1752 | if (Value.SymbolName) |
1753 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1754 | else |
1755 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1756 | Section.advanceStubOffset(StubSize: getMaxStubSize()); |
1757 | } |
1758 | |
1759 | if (RelType == ELF::R_390_GOTENT) |
1760 | resolveRelocation(Section, Offset, Value: StubAddress + 8, Type: ELF::R_390_PC32DBL, |
1761 | Addend); |
1762 | else |
1763 | resolveRelocation(Section, Offset, Value: StubAddress, Type: RelType, Addend); |
1764 | } else if (Arch == Triple::x86_64) { |
1765 | if (RelType == ELF::R_X86_64_PLT32) { |
1766 | // The way the PLT relocations normally work is that the linker allocates |
1767 | // the |
1768 | // PLT and this relocation makes a PC-relative call into the PLT. The PLT |
1769 | // entry will then jump to an address provided by the GOT. On first call, |
1770 | // the |
1771 | // GOT address will point back into PLT code that resolves the symbol. After |
1772 | // the first call, the GOT entry points to the actual function. |
1773 | // |
1774 | // For local functions we're ignoring all of that here and just replacing |
1775 | // the PLT32 relocation type with PC32, which will translate the relocation |
1776 | // into a PC-relative call directly to the function. For external symbols we |
1777 | // can't be sure the function will be within 2^32 bytes of the call site, so |
1778 | // we need to create a stub, which calls into the GOT. This case is |
1779 | // equivalent to the usual PLT implementation except that we use the stub |
1780 | // mechanism in RuntimeDyld (which puts stubs at the end of the section) |
1781 | // rather than allocating a PLT section. |
1782 | if (Value.SymbolName && MemMgr.allowStubAllocation()) { |
1783 | // This is a call to an external function. |
1784 | // Look for an existing stub. |
1785 | SectionEntry *Section = &Sections[SectionID]; |
1786 | StubMap::const_iterator i = Stubs.find(x: Value); |
1787 | uintptr_t StubAddress; |
1788 | if (i != Stubs.end()) { |
1789 | StubAddress = uintptr_t(Section->getAddress()) + i->second; |
1790 | LLVM_DEBUG(dbgs() << " Stub function found\n" ); |
1791 | } else { |
1792 | // Create a new stub function (equivalent to a PLT entry). |
1793 | LLVM_DEBUG(dbgs() << " Create a new stub function\n" ); |
1794 | |
1795 | uintptr_t BaseAddress = uintptr_t(Section->getAddress()); |
1796 | StubAddress = alignTo(Size: BaseAddress + Section->getStubOffset(), |
1797 | A: getStubAlignment()); |
1798 | unsigned StubOffset = StubAddress - BaseAddress; |
1799 | Stubs[Value] = StubOffset; |
1800 | createStubFunction(Addr: (uint8_t *)StubAddress); |
1801 | |
1802 | // Bump our stub offset counter |
1803 | Section->advanceStubOffset(StubSize: getMaxStubSize()); |
1804 | |
1805 | // Allocate a GOT Entry |
1806 | uint64_t GOTOffset = allocateGOTEntries(no: 1); |
1807 | // This potentially creates a new Section which potentially |
1808 | // invalidates the Section pointer, so reload it. |
1809 | Section = &Sections[SectionID]; |
1810 | |
1811 | // The load of the GOT address has an addend of -4 |
1812 | resolveGOTOffsetRelocation(SectionID, Offset: StubOffset + 2, GOTOffset: GOTOffset - 4, |
1813 | Type: ELF::R_X86_64_PC32); |
1814 | |
1815 | // Fill in the value of the symbol we're targeting into the GOT |
1816 | addRelocationForSymbol( |
1817 | RE: computeGOTOffsetRE(GOTOffset, SymbolOffset: 0, Type: ELF::R_X86_64_64), |
1818 | SymbolName: Value.SymbolName); |
1819 | } |
1820 | |
1821 | // Make the target call a call into the stub table. |
1822 | resolveRelocation(Section: *Section, Offset, Value: StubAddress, Type: ELF::R_X86_64_PC32, |
1823 | Addend); |
1824 | } else { |
1825 | Value.Addend += support::ulittle32_t::ref( |
1826 | computePlaceholderAddress(SectionID, Offset)); |
1827 | processSimpleRelocation(SectionID, Offset, RelType: ELF::R_X86_64_PC32, Value); |
1828 | } |
1829 | } else if (RelType == ELF::R_X86_64_GOTPCREL || |
1830 | RelType == ELF::R_X86_64_GOTPCRELX || |
1831 | RelType == ELF::R_X86_64_REX_GOTPCRELX) { |
1832 | uint64_t GOTOffset = allocateGOTEntries(no: 1); |
1833 | resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset: GOTOffset + Addend, |
1834 | Type: ELF::R_X86_64_PC32); |
1835 | |
1836 | // Fill in the value of the symbol we're targeting into the GOT |
1837 | RelocationEntry RE = |
1838 | computeGOTOffsetRE(GOTOffset, SymbolOffset: Value.Offset, Type: ELF::R_X86_64_64); |
1839 | if (Value.SymbolName) |
1840 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1841 | else |
1842 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1843 | } else if (RelType == ELF::R_X86_64_GOT64) { |
1844 | // Fill in a 64-bit GOT offset. |
1845 | uint64_t GOTOffset = allocateGOTEntries(no: 1); |
1846 | resolveRelocation(Section: Sections[SectionID], Offset, Value: GOTOffset, |
1847 | Type: ELF::R_X86_64_64, Addend: 0); |
1848 | |
1849 | // Fill in the value of the symbol we're targeting into the GOT |
1850 | RelocationEntry RE = |
1851 | computeGOTOffsetRE(GOTOffset, SymbolOffset: Value.Offset, Type: ELF::R_X86_64_64); |
1852 | if (Value.SymbolName) |
1853 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1854 | else |
1855 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1856 | } else if (RelType == ELF::R_X86_64_GOTPC32) { |
1857 | // Materialize the address of the base of the GOT relative to the PC. |
1858 | // This doesn't create a GOT entry, but it does mean we need a GOT |
1859 | // section. |
1860 | (void)allocateGOTEntries(no: 0); |
1861 | resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset: Addend, Type: ELF::R_X86_64_PC32); |
1862 | } else if (RelType == ELF::R_X86_64_GOTPC64) { |
1863 | (void)allocateGOTEntries(no: 0); |
1864 | resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset: Addend, Type: ELF::R_X86_64_PC64); |
1865 | } else if (RelType == ELF::R_X86_64_GOTOFF64) { |
1866 | // GOTOFF relocations ultimately require a section difference relocation. |
1867 | (void)allocateGOTEntries(no: 0); |
1868 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1869 | } else if (RelType == ELF::R_X86_64_PC32) { |
1870 | Value.Addend += support::ulittle32_t::ref(computePlaceholderAddress(SectionID, Offset)); |
1871 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1872 | } else if (RelType == ELF::R_X86_64_PC64) { |
1873 | Value.Addend += support::ulittle64_t::ref(computePlaceholderAddress(SectionID, Offset)); |
1874 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1875 | } else if (RelType == ELF::R_X86_64_GOTTPOFF) { |
1876 | processX86_64GOTTPOFFRelocation(SectionID, Offset, Value, Addend); |
1877 | } else if (RelType == ELF::R_X86_64_TLSGD || |
1878 | RelType == ELF::R_X86_64_TLSLD) { |
1879 | // The next relocation must be the relocation for __tls_get_addr. |
1880 | ++RelI; |
1881 | auto &GetAddrRelocation = *RelI; |
1882 | processX86_64TLSRelocation(SectionID, Offset, RelType, Value, Addend, |
1883 | GetAddrRelocation); |
1884 | } else { |
1885 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1886 | } |
1887 | } else { |
1888 | if (Arch == Triple::x86) { |
1889 | Value.Addend += support::ulittle32_t::ref(computePlaceholderAddress(SectionID, Offset)); |
1890 | } |
1891 | processSimpleRelocation(SectionID, Offset, RelType, Value); |
1892 | } |
1893 | return ++RelI; |
1894 | } |
1895 | |
1896 | void RuntimeDyldELF::processX86_64GOTTPOFFRelocation(unsigned SectionID, |
1897 | uint64_t Offset, |
1898 | RelocationValueRef Value, |
1899 | int64_t Addend) { |
1900 | // Use the approach from "x86-64 Linker Optimizations" from the TLS spec |
1901 | // to replace the GOTTPOFF relocation with a TPOFF relocation. The spec |
1902 | // only mentions one optimization even though there are two different |
1903 | // code sequences for the Initial Exec TLS Model. We match the code to |
1904 | // find out which one was used. |
1905 | |
1906 | // A possible TLS code sequence and its replacement |
1907 | struct CodeSequence { |
1908 | // The expected code sequence |
1909 | ArrayRef<uint8_t> ExpectedCodeSequence; |
1910 | // The negative offset of the GOTTPOFF relocation to the beginning of |
1911 | // the sequence |
1912 | uint64_t TLSSequenceOffset; |
1913 | // The new code sequence |
1914 | ArrayRef<uint8_t> NewCodeSequence; |
1915 | // The offset of the new TPOFF relocation |
1916 | uint64_t TpoffRelocationOffset; |
1917 | }; |
1918 | |
1919 | std::array<CodeSequence, 2> CodeSequences; |
1920 | |
1921 | // Initial Exec Code Model Sequence |
1922 | { |
1923 | static const std::initializer_list<uint8_t> ExpectedCodeSequenceList = { |
1924 | 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, |
1925 | 0x00, // mov %fs:0, %rax |
1926 | 0x48, 0x03, 0x05, 0x00, 0x00, 0x00, 0x00 // add x@gotpoff(%rip), |
1927 | // %rax |
1928 | }; |
1929 | CodeSequences[0].ExpectedCodeSequence = |
1930 | ArrayRef<uint8_t>(ExpectedCodeSequenceList); |
1931 | CodeSequences[0].TLSSequenceOffset = 12; |
1932 | |
1933 | static const std::initializer_list<uint8_t> NewCodeSequenceList = { |
1934 | 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 0x00, // mov %fs:0, %rax |
1935 | 0x48, 0x8d, 0x80, 0x00, 0x00, 0x00, 0x00 // lea x@tpoff(%rax), %rax |
1936 | }; |
1937 | CodeSequences[0].NewCodeSequence = ArrayRef<uint8_t>(NewCodeSequenceList); |
1938 | CodeSequences[0].TpoffRelocationOffset = 12; |
1939 | } |
1940 | |
1941 | // Initial Exec Code Model Sequence, II |
1942 | { |
1943 | static const std::initializer_list<uint8_t> ExpectedCodeSequenceList = { |
1944 | 0x48, 0x8b, 0x05, 0x00, 0x00, 0x00, 0x00, // mov x@gotpoff(%rip), %rax |
1945 | 0x64, 0x48, 0x8b, 0x00, 0x00, 0x00, 0x00 // mov %fs:(%rax), %rax |
1946 | }; |
1947 | CodeSequences[1].ExpectedCodeSequence = |
1948 | ArrayRef<uint8_t>(ExpectedCodeSequenceList); |
1949 | CodeSequences[1].TLSSequenceOffset = 3; |
1950 | |
1951 | static const std::initializer_list<uint8_t> NewCodeSequenceList = { |
1952 | 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00, // 6 byte nop |
1953 | 0x64, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 0x00, // mov %fs:x@tpoff, %rax |
1954 | }; |
1955 | CodeSequences[1].NewCodeSequence = ArrayRef<uint8_t>(NewCodeSequenceList); |
1956 | CodeSequences[1].TpoffRelocationOffset = 10; |
1957 | } |
1958 | |
1959 | bool Resolved = false; |
1960 | auto &Section = Sections[SectionID]; |
1961 | for (const auto &C : CodeSequences) { |
1962 | assert(C.ExpectedCodeSequence.size() == C.NewCodeSequence.size() && |
1963 | "Old and new code sequences must have the same size" ); |
1964 | |
1965 | if (Offset < C.TLSSequenceOffset || |
1966 | (Offset - C.TLSSequenceOffset + C.NewCodeSequence.size()) > |
1967 | Section.getSize()) { |
1968 | // This can't be a matching sequence as it doesn't fit in the current |
1969 | // section |
1970 | continue; |
1971 | } |
1972 | |
1973 | auto TLSSequenceStartOffset = Offset - C.TLSSequenceOffset; |
1974 | auto *TLSSequence = Section.getAddressWithOffset(OffsetBytes: TLSSequenceStartOffset); |
1975 | if (ArrayRef<uint8_t>(TLSSequence, C.ExpectedCodeSequence.size()) != |
1976 | C.ExpectedCodeSequence) { |
1977 | continue; |
1978 | } |
1979 | |
1980 | memcpy(dest: TLSSequence, src: C.NewCodeSequence.data(), n: C.NewCodeSequence.size()); |
1981 | |
1982 | // The original GOTTPOFF relocation has an addend as it is PC relative, |
1983 | // so it needs to be corrected. The TPOFF32 relocation is used as an |
1984 | // absolute value (which is an offset from %fs:0), so remove the addend |
1985 | // again. |
1986 | RelocationEntry RE(SectionID, |
1987 | TLSSequenceStartOffset + C.TpoffRelocationOffset, |
1988 | ELF::R_X86_64_TPOFF32, Value.Addend - Addend); |
1989 | |
1990 | if (Value.SymbolName) |
1991 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
1992 | else |
1993 | addRelocationForSection(RE, SectionID: Value.SectionID); |
1994 | |
1995 | Resolved = true; |
1996 | break; |
1997 | } |
1998 | |
1999 | if (!Resolved) { |
2000 | // The GOTTPOFF relocation was not used in one of the sequences |
2001 | // described in the spec, so we can't optimize it to a TPOFF |
2002 | // relocation. |
2003 | uint64_t GOTOffset = allocateGOTEntries(no: 1); |
2004 | resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset: GOTOffset + Addend, |
2005 | Type: ELF::R_X86_64_PC32); |
2006 | RelocationEntry RE = |
2007 | computeGOTOffsetRE(GOTOffset, SymbolOffset: Value.Offset, Type: ELF::R_X86_64_TPOFF64); |
2008 | if (Value.SymbolName) |
2009 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
2010 | else |
2011 | addRelocationForSection(RE, SectionID: Value.SectionID); |
2012 | } |
2013 | } |
2014 | |
2015 | void RuntimeDyldELF::processX86_64TLSRelocation( |
2016 | unsigned SectionID, uint64_t Offset, uint64_t RelType, |
2017 | RelocationValueRef Value, int64_t Addend, |
2018 | const RelocationRef &GetAddrRelocation) { |
2019 | // Since we are statically linking and have no additional DSOs, we can resolve |
2020 | // the relocation directly without using __tls_get_addr. |
2021 | // Use the approach from "x86-64 Linker Optimizations" from the TLS spec |
2022 | // to replace it with the Local Exec relocation variant. |
2023 | |
2024 | // Find out whether the code was compiled with the large or small memory |
2025 | // model. For this we look at the next relocation which is the relocation |
2026 | // for the __tls_get_addr function. If it's a 32 bit relocation, it's the |
2027 | // small code model, with a 64 bit relocation it's the large code model. |
2028 | bool IsSmallCodeModel; |
2029 | // Is the relocation for the __tls_get_addr a PC-relative GOT relocation? |
2030 | bool IsGOTPCRel = false; |
2031 | |
2032 | switch (GetAddrRelocation.getType()) { |
2033 | case ELF::R_X86_64_GOTPCREL: |
2034 | case ELF::R_X86_64_REX_GOTPCRELX: |
2035 | case ELF::R_X86_64_GOTPCRELX: |
2036 | IsGOTPCRel = true; |
2037 | [[fallthrough]]; |
2038 | case ELF::R_X86_64_PLT32: |
2039 | IsSmallCodeModel = true; |
2040 | break; |
2041 | case ELF::R_X86_64_PLTOFF64: |
2042 | IsSmallCodeModel = false; |
2043 | break; |
2044 | default: |
2045 | report_fatal_error( |
2046 | reason: "invalid TLS relocations for General/Local Dynamic TLS Model: " |
2047 | "expected PLT or GOT relocation for __tls_get_addr function" ); |
2048 | } |
2049 | |
2050 | // The negative offset to the start of the TLS code sequence relative to |
2051 | // the offset of the TLSGD/TLSLD relocation |
2052 | uint64_t TLSSequenceOffset; |
2053 | // The expected start of the code sequence |
2054 | ArrayRef<uint8_t> ExpectedCodeSequence; |
2055 | // The new TLS code sequence that will replace the existing code |
2056 | ArrayRef<uint8_t> NewCodeSequence; |
2057 | |
2058 | if (RelType == ELF::R_X86_64_TLSGD) { |
2059 | // The offset of the new TPOFF32 relocation (offset starting from the |
2060 | // beginning of the whole TLS sequence) |
2061 | uint64_t TpoffRelocOffset; |
2062 | |
2063 | if (IsSmallCodeModel) { |
2064 | if (!IsGOTPCRel) { |
2065 | static const std::initializer_list<uint8_t> CodeSequence = { |
2066 | 0x66, // data16 (no-op prefix) |
2067 | 0x48, 0x8d, 0x3d, 0x00, 0x00, |
2068 | 0x00, 0x00, // lea <disp32>(%rip), %rdi |
2069 | 0x66, 0x66, // two data16 prefixes |
2070 | 0x48, // rex64 (no-op prefix) |
2071 | 0xe8, 0x00, 0x00, 0x00, 0x00 // call __tls_get_addr@plt |
2072 | }; |
2073 | ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); |
2074 | TLSSequenceOffset = 4; |
2075 | } else { |
2076 | // This code sequence is not described in the TLS spec but gcc |
2077 | // generates it sometimes. |
2078 | static const std::initializer_list<uint8_t> CodeSequence = { |
2079 | 0x66, // data16 (no-op prefix) |
2080 | 0x48, 0x8d, 0x3d, 0x00, 0x00, |
2081 | 0x00, 0x00, // lea <disp32>(%rip), %rdi |
2082 | 0x66, // data16 prefix (no-op prefix) |
2083 | 0x48, // rex64 (no-op prefix) |
2084 | 0xff, 0x15, 0x00, 0x00, 0x00, |
2085 | 0x00 // call *__tls_get_addr@gotpcrel(%rip) |
2086 | }; |
2087 | ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); |
2088 | TLSSequenceOffset = 4; |
2089 | } |
2090 | |
2091 | // The replacement code for the small code model. It's the same for |
2092 | // both sequences. |
2093 | static const std::initializer_list<uint8_t> SmallSequence = { |
2094 | 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, |
2095 | 0x00, // mov %fs:0, %rax |
2096 | 0x48, 0x8d, 0x80, 0x00, 0x00, 0x00, 0x00 // lea x@tpoff(%rax), |
2097 | // %rax |
2098 | }; |
2099 | NewCodeSequence = ArrayRef<uint8_t>(SmallSequence); |
2100 | TpoffRelocOffset = 12; |
2101 | } else { |
2102 | static const std::initializer_list<uint8_t> CodeSequence = { |
2103 | 0x48, 0x8d, 0x3d, 0x00, 0x00, 0x00, 0x00, // lea <disp32>(%rip), |
2104 | // %rdi |
2105 | 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
2106 | 0x00, // movabs $__tls_get_addr@pltoff, %rax |
2107 | 0x48, 0x01, 0xd8, // add %rbx, %rax |
2108 | 0xff, 0xd0 // call *%rax |
2109 | }; |
2110 | ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); |
2111 | TLSSequenceOffset = 3; |
2112 | |
2113 | // The replacement code for the large code model |
2114 | static const std::initializer_list<uint8_t> LargeSequence = { |
2115 | 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, |
2116 | 0x00, // mov %fs:0, %rax |
2117 | 0x48, 0x8d, 0x80, 0x00, 0x00, 0x00, 0x00, // lea x@tpoff(%rax), |
2118 | // %rax |
2119 | 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00 // nopw 0x0(%rax,%rax,1) |
2120 | }; |
2121 | NewCodeSequence = ArrayRef<uint8_t>(LargeSequence); |
2122 | TpoffRelocOffset = 12; |
2123 | } |
2124 | |
2125 | // The TLSGD/TLSLD relocations are PC-relative, so they have an addend. |
2126 | // The new TPOFF32 relocations is used as an absolute offset from |
2127 | // %fs:0, so remove the TLSGD/TLSLD addend again. |
2128 | RelocationEntry RE(SectionID, Offset - TLSSequenceOffset + TpoffRelocOffset, |
2129 | ELF::R_X86_64_TPOFF32, Value.Addend - Addend); |
2130 | if (Value.SymbolName) |
2131 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
2132 | else |
2133 | addRelocationForSection(RE, SectionID: Value.SectionID); |
2134 | } else if (RelType == ELF::R_X86_64_TLSLD) { |
2135 | if (IsSmallCodeModel) { |
2136 | if (!IsGOTPCRel) { |
2137 | static const std::initializer_list<uint8_t> CodeSequence = { |
2138 | 0x48, 0x8d, 0x3d, 0x00, 0x00, 0x00, // leaq <disp32>(%rip), %rdi |
2139 | 0x00, 0xe8, 0x00, 0x00, 0x00, 0x00 // call __tls_get_addr@plt |
2140 | }; |
2141 | ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); |
2142 | TLSSequenceOffset = 3; |
2143 | |
2144 | // The replacement code for the small code model |
2145 | static const std::initializer_list<uint8_t> SmallSequence = { |
2146 | 0x66, 0x66, 0x66, // three data16 prefixes (no-op) |
2147 | 0x64, 0x48, 0x8b, 0x04, 0x25, |
2148 | 0x00, 0x00, 0x00, 0x00 // mov %fs:0, %rax |
2149 | }; |
2150 | NewCodeSequence = ArrayRef<uint8_t>(SmallSequence); |
2151 | } else { |
2152 | // This code sequence is not described in the TLS spec but gcc |
2153 | // generates it sometimes. |
2154 | static const std::initializer_list<uint8_t> CodeSequence = { |
2155 | 0x48, 0x8d, 0x3d, 0x00, |
2156 | 0x00, 0x00, 0x00, // leaq <disp32>(%rip), %rdi |
2157 | 0xff, 0x15, 0x00, 0x00, |
2158 | 0x00, 0x00 // call |
2159 | // *__tls_get_addr@gotpcrel(%rip) |
2160 | }; |
2161 | ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); |
2162 | TLSSequenceOffset = 3; |
2163 | |
2164 | // The replacement is code is just like above but it needs to be |
2165 | // one byte longer. |
2166 | static const std::initializer_list<uint8_t> SmallSequence = { |
2167 | 0x0f, 0x1f, 0x40, 0x00, // 4 byte nop |
2168 | 0x64, 0x48, 0x8b, 0x04, 0x25, |
2169 | 0x00, 0x00, 0x00, 0x00 // mov %fs:0, %rax |
2170 | }; |
2171 | NewCodeSequence = ArrayRef<uint8_t>(SmallSequence); |
2172 | } |
2173 | } else { |
2174 | // This is the same sequence as for the TLSGD sequence with the large |
2175 | // memory model above |
2176 | static const std::initializer_list<uint8_t> CodeSequence = { |
2177 | 0x48, 0x8d, 0x3d, 0x00, 0x00, 0x00, 0x00, // lea <disp32>(%rip), |
2178 | // %rdi |
2179 | 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
2180 | 0x48, // movabs $__tls_get_addr@pltoff, %rax |
2181 | 0x01, 0xd8, // add %rbx, %rax |
2182 | 0xff, 0xd0 // call *%rax |
2183 | }; |
2184 | ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); |
2185 | TLSSequenceOffset = 3; |
2186 | |
2187 | // The replacement code for the large code model |
2188 | static const std::initializer_list<uint8_t> LargeSequence = { |
2189 | 0x66, 0x66, 0x66, // three data16 prefixes (no-op) |
2190 | 0x66, 0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, |
2191 | 0x00, // 10 byte nop |
2192 | 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 0x00 // mov %fs:0,%rax |
2193 | }; |
2194 | NewCodeSequence = ArrayRef<uint8_t>(LargeSequence); |
2195 | } |
2196 | } else { |
2197 | llvm_unreachable("both TLS relocations handled above" ); |
2198 | } |
2199 | |
2200 | assert(ExpectedCodeSequence.size() == NewCodeSequence.size() && |
2201 | "Old and new code sequences must have the same size" ); |
2202 | |
2203 | auto &Section = Sections[SectionID]; |
2204 | if (Offset < TLSSequenceOffset || |
2205 | (Offset - TLSSequenceOffset + NewCodeSequence.size()) > |
2206 | Section.getSize()) { |
2207 | report_fatal_error(reason: "unexpected end of section in TLS sequence" ); |
2208 | } |
2209 | |
2210 | auto *TLSSequence = Section.getAddressWithOffset(OffsetBytes: Offset - TLSSequenceOffset); |
2211 | if (ArrayRef<uint8_t>(TLSSequence, ExpectedCodeSequence.size()) != |
2212 | ExpectedCodeSequence) { |
2213 | report_fatal_error( |
2214 | reason: "invalid TLS sequence for Global/Local Dynamic TLS Model" ); |
2215 | } |
2216 | |
2217 | memcpy(dest: TLSSequence, src: NewCodeSequence.data(), n: NewCodeSequence.size()); |
2218 | } |
2219 | |
2220 | size_t RuntimeDyldELF::getGOTEntrySize() { |
2221 | // We don't use the GOT in all of these cases, but it's essentially free |
2222 | // to put them all here. |
2223 | size_t Result = 0; |
2224 | switch (Arch) { |
2225 | case Triple::x86_64: |
2226 | case Triple::aarch64: |
2227 | case Triple::aarch64_be: |
2228 | case Triple::ppc64: |
2229 | case Triple::ppc64le: |
2230 | case Triple::systemz: |
2231 | Result = sizeof(uint64_t); |
2232 | break; |
2233 | case Triple::x86: |
2234 | case Triple::arm: |
2235 | case Triple::thumb: |
2236 | Result = sizeof(uint32_t); |
2237 | break; |
2238 | case Triple::mips: |
2239 | case Triple::mipsel: |
2240 | case Triple::mips64: |
2241 | case Triple::mips64el: |
2242 | if (IsMipsO32ABI || IsMipsN32ABI) |
2243 | Result = sizeof(uint32_t); |
2244 | else if (IsMipsN64ABI) |
2245 | Result = sizeof(uint64_t); |
2246 | else |
2247 | llvm_unreachable("Mips ABI not handled" ); |
2248 | break; |
2249 | default: |
2250 | llvm_unreachable("Unsupported CPU type!" ); |
2251 | } |
2252 | return Result; |
2253 | } |
2254 | |
2255 | uint64_t RuntimeDyldELF::allocateGOTEntries(unsigned no) { |
2256 | if (GOTSectionID == 0) { |
2257 | GOTSectionID = Sections.size(); |
2258 | // Reserve a section id. We'll allocate the section later |
2259 | // once we know the total size |
2260 | Sections.push_back(x: SectionEntry(".got" , nullptr, 0, 0, 0)); |
2261 | } |
2262 | uint64_t StartOffset = CurrentGOTIndex * getGOTEntrySize(); |
2263 | CurrentGOTIndex += no; |
2264 | return StartOffset; |
2265 | } |
2266 | |
2267 | uint64_t RuntimeDyldELF::findOrAllocGOTEntry(const RelocationValueRef &Value, |
2268 | unsigned GOTRelType) { |
2269 | auto E = GOTOffsetMap.insert(x: {Value, 0}); |
2270 | if (E.second) { |
2271 | uint64_t GOTOffset = allocateGOTEntries(no: 1); |
2272 | |
2273 | // Create relocation for newly created GOT entry |
2274 | RelocationEntry RE = |
2275 | computeGOTOffsetRE(GOTOffset, SymbolOffset: Value.Offset, Type: GOTRelType); |
2276 | if (Value.SymbolName) |
2277 | addRelocationForSymbol(RE, SymbolName: Value.SymbolName); |
2278 | else |
2279 | addRelocationForSection(RE, SectionID: Value.SectionID); |
2280 | |
2281 | E.first->second = GOTOffset; |
2282 | } |
2283 | |
2284 | return E.first->second; |
2285 | } |
2286 | |
2287 | void RuntimeDyldELF::resolveGOTOffsetRelocation(unsigned SectionID, |
2288 | uint64_t Offset, |
2289 | uint64_t GOTOffset, |
2290 | uint32_t Type) { |
2291 | // Fill in the relative address of the GOT Entry into the stub |
2292 | RelocationEntry GOTRE(SectionID, Offset, Type, GOTOffset); |
2293 | addRelocationForSection(RE: GOTRE, SectionID: GOTSectionID); |
2294 | } |
2295 | |
2296 | RelocationEntry RuntimeDyldELF::computeGOTOffsetRE(uint64_t GOTOffset, |
2297 | uint64_t SymbolOffset, |
2298 | uint32_t Type) { |
2299 | return RelocationEntry(GOTSectionID, GOTOffset, Type, SymbolOffset); |
2300 | } |
2301 | |
2302 | void RuntimeDyldELF::processNewSymbol(const SymbolRef &ObjSymbol, SymbolTableEntry& Symbol) { |
2303 | // This should never return an error as `processNewSymbol` wouldn't have been |
2304 | // called if getFlags() returned an error before. |
2305 | auto ObjSymbolFlags = cantFail(ValOrErr: ObjSymbol.getFlags()); |
2306 | |
2307 | if (ObjSymbolFlags & SymbolRef::SF_Indirect) { |
2308 | if (IFuncStubSectionID == 0) { |
2309 | // Create a dummy section for the ifunc stubs. It will be actually |
2310 | // allocated in finalizeLoad() below. |
2311 | IFuncStubSectionID = Sections.size(); |
2312 | Sections.push_back( |
2313 | x: SectionEntry(".text.__llvm_IFuncStubs" , nullptr, 0, 0, 0)); |
2314 | // First 64B are reserverd for the IFunc resolver |
2315 | IFuncStubOffset = 64; |
2316 | } |
2317 | |
2318 | IFuncStubs.push_back(Elt: IFuncStub{.StubOffset: IFuncStubOffset, .OriginalSymbol: Symbol}); |
2319 | // Modify the symbol so that it points to the ifunc stub instead of to the |
2320 | // resolver function. |
2321 | Symbol = SymbolTableEntry(IFuncStubSectionID, IFuncStubOffset, |
2322 | Symbol.getFlags()); |
2323 | IFuncStubOffset += getMaxIFuncStubSize(); |
2324 | } |
2325 | } |
2326 | |
2327 | Error RuntimeDyldELF::finalizeLoad(const ObjectFile &Obj, |
2328 | ObjSectionToIDMap &SectionMap) { |
2329 | if (IsMipsO32ABI) |
2330 | if (!PendingRelocs.empty()) |
2331 | return make_error<RuntimeDyldError>(Args: "Can't find matching LO16 reloc" ); |
2332 | |
2333 | // Create the IFunc stubs if necessary. This must be done before processing |
2334 | // the GOT entries, as the IFunc stubs may create some. |
2335 | if (IFuncStubSectionID != 0) { |
2336 | uint8_t *IFuncStubsAddr = MemMgr.allocateCodeSection( |
2337 | Size: IFuncStubOffset, Alignment: 1, SectionID: IFuncStubSectionID, SectionName: ".text.__llvm_IFuncStubs" ); |
2338 | if (!IFuncStubsAddr) |
2339 | return make_error<RuntimeDyldError>( |
2340 | Args: "Unable to allocate memory for IFunc stubs!" ); |
2341 | Sections[IFuncStubSectionID] = |
2342 | SectionEntry(".text.__llvm_IFuncStubs" , IFuncStubsAddr, IFuncStubOffset, |
2343 | IFuncStubOffset, 0); |
2344 | |
2345 | createIFuncResolver(Addr: IFuncStubsAddr); |
2346 | |
2347 | LLVM_DEBUG(dbgs() << "Creating IFunc stubs SectionID: " |
2348 | << IFuncStubSectionID << " Addr: " |
2349 | << Sections[IFuncStubSectionID].getAddress() << '\n'); |
2350 | for (auto &IFuncStub : IFuncStubs) { |
2351 | auto &Symbol = IFuncStub.OriginalSymbol; |
2352 | LLVM_DEBUG(dbgs() << "\tSectionID: " << Symbol.getSectionID() |
2353 | << " Offset: " << format("%p" , Symbol.getOffset()) |
2354 | << " IFuncStubOffset: " |
2355 | << format("%p\n" , IFuncStub.StubOffset)); |
2356 | createIFuncStub(IFuncStubSectionID, IFuncResolverOffset: 0, IFuncStubOffset: IFuncStub.StubOffset, |
2357 | IFuncSectionID: Symbol.getSectionID(), IFuncOffset: Symbol.getOffset()); |
2358 | } |
2359 | |
2360 | IFuncStubSectionID = 0; |
2361 | IFuncStubOffset = 0; |
2362 | IFuncStubs.clear(); |
2363 | } |
2364 | |
2365 | // If necessary, allocate the global offset table |
2366 | if (GOTSectionID != 0) { |
2367 | // Allocate memory for the section |
2368 | size_t TotalSize = CurrentGOTIndex * getGOTEntrySize(); |
2369 | uint8_t *Addr = MemMgr.allocateDataSection(Size: TotalSize, Alignment: getGOTEntrySize(), |
2370 | SectionID: GOTSectionID, SectionName: ".got" , IsReadOnly: false); |
2371 | if (!Addr) |
2372 | return make_error<RuntimeDyldError>(Args: "Unable to allocate memory for GOT!" ); |
2373 | |
2374 | Sections[GOTSectionID] = |
2375 | SectionEntry(".got" , Addr, TotalSize, TotalSize, 0); |
2376 | |
2377 | // For now, initialize all GOT entries to zero. We'll fill them in as |
2378 | // needed when GOT-based relocations are applied. |
2379 | memset(s: Addr, c: 0, n: TotalSize); |
2380 | if (IsMipsN32ABI || IsMipsN64ABI) { |
2381 | // To correctly resolve Mips GOT relocations, we need a mapping from |
2382 | // object's sections to GOTs. |
2383 | for (section_iterator SI = Obj.section_begin(), SE = Obj.section_end(); |
2384 | SI != SE; ++SI) { |
2385 | if (SI->relocation_begin() != SI->relocation_end()) { |
2386 | Expected<section_iterator> RelSecOrErr = SI->getRelocatedSection(); |
2387 | if (!RelSecOrErr) |
2388 | return make_error<RuntimeDyldError>( |
2389 | Args: toString(E: RelSecOrErr.takeError())); |
2390 | |
2391 | section_iterator RelocatedSection = *RelSecOrErr; |
2392 | ObjSectionToIDMap::iterator i = SectionMap.find(x: *RelocatedSection); |
2393 | assert(i != SectionMap.end()); |
2394 | SectionToGOTMap[i->second] = GOTSectionID; |
2395 | } |
2396 | } |
2397 | GOTSymbolOffsets.clear(); |
2398 | } |
2399 | } |
2400 | |
2401 | // Look for and record the EH frame section. |
2402 | ObjSectionToIDMap::iterator i, e; |
2403 | for (i = SectionMap.begin(), e = SectionMap.end(); i != e; ++i) { |
2404 | const SectionRef &Section = i->first; |
2405 | |
2406 | StringRef Name; |
2407 | Expected<StringRef> NameOrErr = Section.getName(); |
2408 | if (NameOrErr) |
2409 | Name = *NameOrErr; |
2410 | else |
2411 | consumeError(Err: NameOrErr.takeError()); |
2412 | |
2413 | if (Name == ".eh_frame" ) { |
2414 | UnregisteredEHFrameSections.push_back(Elt: i->second); |
2415 | break; |
2416 | } |
2417 | } |
2418 | |
2419 | GOTOffsetMap.clear(); |
2420 | GOTSectionID = 0; |
2421 | CurrentGOTIndex = 0; |
2422 | |
2423 | return Error::success(); |
2424 | } |
2425 | |
2426 | bool RuntimeDyldELF::isCompatibleFile(const object::ObjectFile &Obj) const { |
2427 | return Obj.isELF(); |
2428 | } |
2429 | |
2430 | void RuntimeDyldELF::createIFuncResolver(uint8_t *Addr) const { |
2431 | if (Arch == Triple::x86_64) { |
2432 | // The adddres of the GOT1 entry is in %r11, the GOT2 entry is in %r11+8 |
2433 | // (see createIFuncStub() for details) |
2434 | // The following code first saves all registers that contain the original |
2435 | // function arguments as those registers are not saved by the resolver |
2436 | // function. %r11 is saved as well so that the GOT2 entry can be updated |
2437 | // afterwards. Then it calls the actual IFunc resolver function whose |
2438 | // address is stored in GOT2. After the resolver function returns, all |
2439 | // saved registers are restored and the return value is written to GOT1. |
2440 | // Finally, jump to the now resolved function. |
2441 | // clang-format off |
2442 | const uint8_t StubCode[] = { |
2443 | 0x57, // push %rdi |
2444 | 0x56, // push %rsi |
2445 | 0x52, // push %rdx |
2446 | 0x51, // push %rcx |
2447 | 0x41, 0x50, // push %r8 |
2448 | 0x41, 0x51, // push %r9 |
2449 | 0x41, 0x53, // push %r11 |
2450 | 0x41, 0xff, 0x53, 0x08, // call *0x8(%r11) |
2451 | 0x41, 0x5b, // pop %r11 |
2452 | 0x41, 0x59, // pop %r9 |
2453 | 0x41, 0x58, // pop %r8 |
2454 | 0x59, // pop %rcx |
2455 | 0x5a, // pop %rdx |
2456 | 0x5e, // pop %rsi |
2457 | 0x5f, // pop %rdi |
2458 | 0x49, 0x89, 0x03, // mov %rax,(%r11) |
2459 | 0xff, 0xe0 // jmp *%rax |
2460 | }; |
2461 | // clang-format on |
2462 | static_assert(sizeof(StubCode) <= 64, |
2463 | "maximum size of the IFunc resolver is 64B" ); |
2464 | memcpy(dest: Addr, src: StubCode, n: sizeof(StubCode)); |
2465 | } else { |
2466 | report_fatal_error( |
2467 | reason: "IFunc resolver is not supported for target architecture" ); |
2468 | } |
2469 | } |
2470 | |
2471 | void RuntimeDyldELF::createIFuncStub(unsigned IFuncStubSectionID, |
2472 | uint64_t IFuncResolverOffset, |
2473 | uint64_t IFuncStubOffset, |
2474 | unsigned IFuncSectionID, |
2475 | uint64_t IFuncOffset) { |
2476 | auto &IFuncStubSection = Sections[IFuncStubSectionID]; |
2477 | auto *Addr = IFuncStubSection.getAddressWithOffset(OffsetBytes: IFuncStubOffset); |
2478 | |
2479 | if (Arch == Triple::x86_64) { |
2480 | // The first instruction loads a PC-relative address into %r11 which is a |
2481 | // GOT entry for this stub. This initially contains the address to the |
2482 | // IFunc resolver. We can use %r11 here as it's caller saved but not used |
2483 | // to pass any arguments. In fact, x86_64 ABI even suggests using %r11 for |
2484 | // code in the PLT. The IFunc resolver will use %r11 to update the GOT |
2485 | // entry. |
2486 | // |
2487 | // The next instruction just jumps to the address contained in the GOT |
2488 | // entry. As mentioned above, we do this two-step jump by first setting |
2489 | // %r11 so that the IFunc resolver has access to it. |
2490 | // |
2491 | // The IFunc resolver of course also needs to know the actual address of |
2492 | // the actual IFunc resolver function. This will be stored in a GOT entry |
2493 | // right next to the first one for this stub. So, the IFunc resolver will |
2494 | // be able to call it with %r11+8. |
2495 | // |
2496 | // In total, two adjacent GOT entries (+relocation) and one additional |
2497 | // relocation are required: |
2498 | // GOT1: Address of the IFunc resolver. |
2499 | // GOT2: Address of the IFunc resolver function. |
2500 | // IFuncStubOffset+3: 32-bit PC-relative address of GOT1. |
2501 | uint64_t GOT1 = allocateGOTEntries(no: 2); |
2502 | uint64_t GOT2 = GOT1 + getGOTEntrySize(); |
2503 | |
2504 | RelocationEntry RE1(GOTSectionID, GOT1, ELF::R_X86_64_64, |
2505 | IFuncResolverOffset, {}); |
2506 | addRelocationForSection(RE: RE1, SectionID: IFuncStubSectionID); |
2507 | RelocationEntry RE2(GOTSectionID, GOT2, ELF::R_X86_64_64, IFuncOffset, {}); |
2508 | addRelocationForSection(RE: RE2, SectionID: IFuncSectionID); |
2509 | |
2510 | const uint8_t StubCode[] = { |
2511 | 0x4c, 0x8d, 0x1d, 0x00, 0x00, 0x00, 0x00, // leaq 0x0(%rip),%r11 |
2512 | 0x41, 0xff, 0x23 // jmpq *(%r11) |
2513 | }; |
2514 | assert(sizeof(StubCode) <= getMaxIFuncStubSize() && |
2515 | "IFunc stub size must not exceed getMaxIFuncStubSize()" ); |
2516 | memcpy(dest: Addr, src: StubCode, n: sizeof(StubCode)); |
2517 | |
2518 | // The PC-relative value starts 4 bytes from the end of the leaq |
2519 | // instruction, so the addend is -4. |
2520 | resolveGOTOffsetRelocation(SectionID: IFuncStubSectionID, Offset: IFuncStubOffset + 3, |
2521 | GOTOffset: GOT1 - 4, Type: ELF::R_X86_64_PC32); |
2522 | } else { |
2523 | report_fatal_error(reason: "IFunc stub is not supported for target architecture" ); |
2524 | } |
2525 | } |
2526 | |
2527 | unsigned RuntimeDyldELF::getMaxIFuncStubSize() const { |
2528 | if (Arch == Triple::x86_64) { |
2529 | return 10; |
2530 | } |
2531 | return 0; |
2532 | } |
2533 | |
2534 | bool RuntimeDyldELF::relocationNeedsGot(const RelocationRef &R) const { |
2535 | unsigned RelTy = R.getType(); |
2536 | if (Arch == Triple::aarch64 || Arch == Triple::aarch64_be) |
2537 | return RelTy == ELF::R_AARCH64_ADR_GOT_PAGE || |
2538 | RelTy == ELF::R_AARCH64_LD64_GOT_LO12_NC; |
2539 | |
2540 | if (Arch == Triple::x86_64) |
2541 | return RelTy == ELF::R_X86_64_GOTPCREL || |
2542 | RelTy == ELF::R_X86_64_GOTPCRELX || |
2543 | RelTy == ELF::R_X86_64_GOT64 || |
2544 | RelTy == ELF::R_X86_64_REX_GOTPCRELX; |
2545 | return false; |
2546 | } |
2547 | |
2548 | bool RuntimeDyldELF::relocationNeedsStub(const RelocationRef &R) const { |
2549 | if (Arch != Triple::x86_64) |
2550 | return true; // Conservative answer |
2551 | |
2552 | switch (R.getType()) { |
2553 | default: |
2554 | return true; // Conservative answer |
2555 | |
2556 | |
2557 | case ELF::R_X86_64_GOTPCREL: |
2558 | case ELF::R_X86_64_GOTPCRELX: |
2559 | case ELF::R_X86_64_REX_GOTPCRELX: |
2560 | case ELF::R_X86_64_GOTPC64: |
2561 | case ELF::R_X86_64_GOT64: |
2562 | case ELF::R_X86_64_GOTOFF64: |
2563 | case ELF::R_X86_64_PC32: |
2564 | case ELF::R_X86_64_PC64: |
2565 | case ELF::R_X86_64_64: |
2566 | // We know that these reloation types won't need a stub function. This list |
2567 | // can be extended as needed. |
2568 | return false; |
2569 | } |
2570 | } |
2571 | |
2572 | } // namespace llvm |
2573 | |