Convert another use of sys::identifyFileType.
[oota-llvm.git] / lib / ExecutionEngine / RuntimeDyld / RuntimeDyld.cpp
1 //===-- RuntimeDyld.cpp - Run-time dynamic linker for MC-JIT ----*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Implementation of the MC-JIT runtime dynamic linker.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #define DEBUG_TYPE "dyld"
15 #include "llvm/ExecutionEngine/RuntimeDyld.h"
16 #include "ObjectImageCommon.h"
17 #include "RuntimeDyldELF.h"
18 #include "RuntimeDyldImpl.h"
19 #include "RuntimeDyldMachO.h"
20 #include "llvm/Support/FileSystem.h"
21 #include "llvm/Support/MathExtras.h"
22 #include "llvm/Object/ELF.h"
23
24 using namespace llvm;
25 using namespace llvm::object;
26
27 // Empty out-of-line virtual destructor as the key function.
28 RuntimeDyldImpl::~RuntimeDyldImpl() {}
29
30 namespace llvm {
31
32 StringRef RuntimeDyldImpl::getEHFrameSection() {
33   return StringRef();
34 }
35
36 // Resolve the relocations for all symbols we currently know about.
37 void RuntimeDyldImpl::resolveRelocations() {
38   // First, resolve relocations associated with external symbols.
39   resolveExternalSymbols();
40
41   // Just iterate over the sections we have and resolve all the relocations
42   // in them. Gross overkill, but it gets the job done.
43   for (int i = 0, e = Sections.size(); i != e; ++i) {
44     uint64_t Addr = Sections[i].LoadAddress;
45     DEBUG(dbgs() << "Resolving relocations Section #" << i
46             << "\t" << format("%p", (uint8_t *)Addr)
47             << "\n");
48     resolveRelocationList(Relocations[i], Addr);
49   }
50 }
51
52 void RuntimeDyldImpl::mapSectionAddress(const void *LocalAddress,
53                                         uint64_t TargetAddress) {
54   for (unsigned i = 0, e = Sections.size(); i != e; ++i) {
55     if (Sections[i].Address == LocalAddress) {
56       reassignSectionAddress(i, TargetAddress);
57       return;
58     }
59   }
60   llvm_unreachable("Attempting to remap address of unknown section!");
61 }
62
63 // Subclasses can implement this method to create specialized image instances.
64 // The caller owns the pointer that is returned.
65 ObjectImage *RuntimeDyldImpl::createObjectImage(ObjectBuffer *InputBuffer) {
66   return new ObjectImageCommon(InputBuffer);
67 }
68
69 ObjectImage *RuntimeDyldImpl::loadObject(ObjectBuffer *InputBuffer) {
70   OwningPtr<ObjectImage> obj(createObjectImage(InputBuffer));
71   if (!obj)
72     report_fatal_error("Unable to create object image from memory buffer!");
73
74   Arch = (Triple::ArchType)obj->getArch();
75
76   // Symbols found in this object
77   StringMap<SymbolLoc> LocalSymbols;
78   // Used sections from the object file
79   ObjSectionToIDMap LocalSections;
80
81   // Common symbols requiring allocation, with their sizes and alignments
82   CommonSymbolMap CommonSymbols;
83   // Maximum required total memory to allocate all common symbols
84   uint64_t CommonSize = 0;
85
86   error_code err;
87   // Parse symbols
88   DEBUG(dbgs() << "Parse symbols:\n");
89   for (symbol_iterator i = obj->begin_symbols(), e = obj->end_symbols();
90        i != e; i.increment(err)) {
91     Check(err);
92     object::SymbolRef::Type SymType;
93     StringRef Name;
94     Check(i->getType(SymType));
95     Check(i->getName(Name));
96
97     uint32_t flags;
98     Check(i->getFlags(flags));
99
100     bool isCommon = flags & SymbolRef::SF_Common;
101     if (isCommon) {
102       // Add the common symbols to a list.  We'll allocate them all below.
103       uint32_t Align;
104       Check(i->getAlignment(Align));
105       uint64_t Size = 0;
106       Check(i->getSize(Size));
107       CommonSize += Size + Align;
108       CommonSymbols[*i] = CommonSymbolInfo(Size, Align);
109     } else {
110       if (SymType == object::SymbolRef::ST_Function ||
111           SymType == object::SymbolRef::ST_Data ||
112           SymType == object::SymbolRef::ST_Unknown) {
113         uint64_t FileOffset;
114         StringRef SectionData;
115         bool IsCode;
116         section_iterator si = obj->end_sections();
117         Check(i->getFileOffset(FileOffset));
118         Check(i->getSection(si));
119         if (si == obj->end_sections()) continue;
120         Check(si->getContents(SectionData));
121         Check(si->isText(IsCode));
122         const uint8_t* SymPtr = (const uint8_t*)InputBuffer->getBufferStart() +
123                                 (uintptr_t)FileOffset;
124         uintptr_t SectOffset = (uintptr_t)(SymPtr -
125                                            (const uint8_t*)SectionData.begin());
126         unsigned SectionID = findOrEmitSection(*obj, *si, IsCode, LocalSections);
127         LocalSymbols[Name.data()] = SymbolLoc(SectionID, SectOffset);
128         DEBUG(dbgs() << "\tFileOffset: " << format("%p", (uintptr_t)FileOffset)
129                      << " flags: " << flags
130                      << " SID: " << SectionID
131                      << " Offset: " << format("%p", SectOffset));
132         GlobalSymbolTable[Name] = SymbolLoc(SectionID, SectOffset);
133       }
134     }
135     DEBUG(dbgs() << "\tType: " << SymType << " Name: " << Name << "\n");
136   }
137
138   // Allocate common symbols
139   if (CommonSize != 0)
140     emitCommonSymbols(*obj, CommonSymbols, CommonSize, LocalSymbols);
141
142   // Parse and process relocations
143   DEBUG(dbgs() << "Parse relocations:\n");
144   for (section_iterator si = obj->begin_sections(),
145        se = obj->end_sections(); si != se; si.increment(err)) {
146     Check(err);
147     bool isFirstRelocation = true;
148     unsigned SectionID = 0;
149     StubMap Stubs;
150     section_iterator RelocatedSection = si->getRelocatedSection();
151
152     for (relocation_iterator i = si->begin_relocations(),
153          e = si->end_relocations(); i != e; i.increment(err)) {
154       Check(err);
155
156       // If it's the first relocation in this section, find its SectionID
157       if (isFirstRelocation) {
158         SectionID =
159             findOrEmitSection(*obj, *RelocatedSection, true, LocalSections);
160         DEBUG(dbgs() << "\tSectionID: " << SectionID << "\n");
161         isFirstRelocation = false;
162       }
163
164       processRelocationRef(SectionID, *i, *obj, LocalSections, LocalSymbols,
165                            Stubs);
166     }
167   }
168
169   return obj.take();
170 }
171
172 void RuntimeDyldImpl::emitCommonSymbols(ObjectImage &Obj,
173                                         const CommonSymbolMap &CommonSymbols,
174                                         uint64_t TotalSize,
175                                         SymbolTableMap &SymbolTable) {
176   // Allocate memory for the section
177   unsigned SectionID = Sections.size();
178   uint8_t *Addr = MemMgr->allocateDataSection(TotalSize, sizeof(void*),
179                                               SectionID, false);
180   if (!Addr)
181     report_fatal_error("Unable to allocate memory for common symbols!");
182   uint64_t Offset = 0;
183   Sections.push_back(SectionEntry(StringRef(), Addr, TotalSize, 0));
184   memset(Addr, 0, TotalSize);
185
186   DEBUG(dbgs() << "emitCommonSection SectionID: " << SectionID
187                << " new addr: " << format("%p", Addr)
188                << " DataSize: " << TotalSize
189                << "\n");
190
191   // Assign the address of each symbol
192   for (CommonSymbolMap::const_iterator it = CommonSymbols.begin(),
193        itEnd = CommonSymbols.end(); it != itEnd; it++) {
194     uint64_t Size = it->second.first;
195     uint64_t Align = it->second.second;
196     StringRef Name;
197     it->first.getName(Name);
198     if (Align) {
199       // This symbol has an alignment requirement.
200       uint64_t AlignOffset = OffsetToAlignment((uint64_t)Addr, Align);
201       Addr += AlignOffset;
202       Offset += AlignOffset;
203       DEBUG(dbgs() << "Allocating common symbol " << Name << " address " <<
204                       format("%p\n", Addr));
205     }
206     Obj.updateSymbolAddress(it->first, (uint64_t)Addr);
207     SymbolTable[Name.data()] = SymbolLoc(SectionID, Offset);
208     Offset += Size;
209     Addr += Size;
210   }
211 }
212
213 unsigned RuntimeDyldImpl::emitSection(ObjectImage &Obj,
214                                       const SectionRef &Section,
215                                       bool IsCode) {
216
217   unsigned StubBufSize = 0,
218            StubSize = getMaxStubSize();
219   error_code err;
220   const ObjectFile *ObjFile = Obj.getObjectFile();
221   // FIXME: this is an inefficient way to handle this. We should computed the
222   // necessary section allocation size in loadObject by walking all the sections
223   // once.
224   if (StubSize > 0) {
225     for (section_iterator SI = ObjFile->begin_sections(),
226            SE = ObjFile->end_sections();
227          SI != SE; SI.increment(err), Check(err)) {
228       section_iterator RelSecI = SI->getRelocatedSection();
229       if (!(RelSecI == Section))
230         continue;
231
232       for (relocation_iterator I = SI->begin_relocations(),
233              E = SI->end_relocations(); I != E; I.increment(err), Check(err)) {
234         StubBufSize += StubSize;
235       }
236     }
237   }
238
239   StringRef data;
240   uint64_t Alignment64;
241   Check(Section.getContents(data));
242   Check(Section.getAlignment(Alignment64));
243
244   unsigned Alignment = (unsigned)Alignment64 & 0xffffffffL;
245   bool IsRequired;
246   bool IsVirtual;
247   bool IsZeroInit;
248   bool IsReadOnly;
249   uint64_t DataSize;
250   StringRef Name;
251   Check(Section.isRequiredForExecution(IsRequired));
252   Check(Section.isVirtual(IsVirtual));
253   Check(Section.isZeroInit(IsZeroInit));
254   Check(Section.isReadOnlyData(IsReadOnly));
255   Check(Section.getSize(DataSize));
256   Check(Section.getName(Name));
257   if (StubSize > 0) {
258     unsigned StubAlignment = getStubAlignment();
259     unsigned EndAlignment = (DataSize | Alignment) & -(DataSize | Alignment);
260     if (StubAlignment > EndAlignment)
261       StubBufSize += StubAlignment - EndAlignment;
262   }
263
264   unsigned Allocate;
265   unsigned SectionID = Sections.size();
266   uint8_t *Addr;
267   const char *pData = 0;
268
269   // Some sections, such as debug info, don't need to be loaded for execution.
270   // Leave those where they are.
271   if (IsRequired) {
272     Allocate = DataSize + StubBufSize;
273     Addr = IsCode
274       ? MemMgr->allocateCodeSection(Allocate, Alignment, SectionID)
275       : MemMgr->allocateDataSection(Allocate, Alignment, SectionID, IsReadOnly);
276     if (!Addr)
277       report_fatal_error("Unable to allocate section memory!");
278
279     // Virtual sections have no data in the object image, so leave pData = 0
280     if (!IsVirtual)
281       pData = data.data();
282
283     // Zero-initialize or copy the data from the image
284     if (IsZeroInit || IsVirtual)
285       memset(Addr, 0, DataSize);
286     else
287       memcpy(Addr, pData, DataSize);
288
289     DEBUG(dbgs() << "emitSection SectionID: " << SectionID
290                  << " Name: " << Name
291                  << " obj addr: " << format("%p", pData)
292                  << " new addr: " << format("%p", Addr)
293                  << " DataSize: " << DataSize
294                  << " StubBufSize: " << StubBufSize
295                  << " Allocate: " << Allocate
296                  << "\n");
297     Obj.updateSectionAddress(Section, (uint64_t)Addr);
298   }
299   else {
300     // Even if we didn't load the section, we need to record an entry for it
301     // to handle later processing (and by 'handle' I mean don't do anything
302     // with these sections).
303     Allocate = 0;
304     Addr = 0;
305     DEBUG(dbgs() << "emitSection SectionID: " << SectionID
306                  << " Name: " << Name
307                  << " obj addr: " << format("%p", data.data())
308                  << " new addr: 0"
309                  << " DataSize: " << DataSize
310                  << " StubBufSize: " << StubBufSize
311                  << " Allocate: " << Allocate
312                  << "\n");
313   }
314
315   Sections.push_back(SectionEntry(Name, Addr, DataSize, (uintptr_t)pData));
316   return SectionID;
317 }
318
319 unsigned RuntimeDyldImpl::findOrEmitSection(ObjectImage &Obj,
320                                             const SectionRef &Section,
321                                             bool IsCode,
322                                             ObjSectionToIDMap &LocalSections) {
323
324   unsigned SectionID = 0;
325   ObjSectionToIDMap::iterator i = LocalSections.find(Section);
326   if (i != LocalSections.end())
327     SectionID = i->second;
328   else {
329     SectionID = emitSection(Obj, Section, IsCode);
330     LocalSections[Section] = SectionID;
331   }
332   return SectionID;
333 }
334
335 void RuntimeDyldImpl::addRelocationForSection(const RelocationEntry &RE,
336                                               unsigned SectionID) {
337   Relocations[SectionID].push_back(RE);
338 }
339
340 void RuntimeDyldImpl::addRelocationForSymbol(const RelocationEntry &RE,
341                                              StringRef SymbolName) {
342   // Relocation by symbol.  If the symbol is found in the global symbol table,
343   // create an appropriate section relocation.  Otherwise, add it to
344   // ExternalSymbolRelocations.
345   SymbolTableMap::const_iterator Loc =
346       GlobalSymbolTable.find(SymbolName);
347   if (Loc == GlobalSymbolTable.end()) {
348     ExternalSymbolRelocations[SymbolName].push_back(RE);
349   } else {
350     // Copy the RE since we want to modify its addend.
351     RelocationEntry RECopy = RE;
352     RECopy.Addend += Loc->second.second;
353     Relocations[Loc->second.first].push_back(RECopy);
354   }
355 }
356
357 uint8_t *RuntimeDyldImpl::createStubFunction(uint8_t *Addr) {
358   if (Arch == Triple::aarch64) {
359     // This stub has to be able to access the full address space,
360     // since symbol lookup won't necessarily find a handy, in-range,
361     // PLT stub for functions which could be anywhere.
362     uint32_t *StubAddr = (uint32_t*)Addr;
363
364     // Stub can use ip0 (== x16) to calculate address
365     *StubAddr = 0xd2e00010; // movz ip0, #:abs_g3:<addr>
366     StubAddr++;
367     *StubAddr = 0xf2c00010; // movk ip0, #:abs_g2_nc:<addr>
368     StubAddr++;
369     *StubAddr = 0xf2a00010; // movk ip0, #:abs_g1_nc:<addr>
370     StubAddr++;
371     *StubAddr = 0xf2800010; // movk ip0, #:abs_g0_nc:<addr>
372     StubAddr++;
373     *StubAddr = 0xd61f0200; // br ip0
374
375     return Addr;
376   } else if (Arch == Triple::arm) {
377     // TODO: There is only ARM far stub now. We should add the Thumb stub,
378     // and stubs for branches Thumb - ARM and ARM - Thumb.
379     uint32_t *StubAddr = (uint32_t*)Addr;
380     *StubAddr = 0xe51ff004; // ldr pc,<label>
381     return (uint8_t*)++StubAddr;
382   } else if (Arch == Triple::mipsel || Arch == Triple::mips) {
383     uint32_t *StubAddr = (uint32_t*)Addr;
384     // 0:   3c190000        lui     t9,%hi(addr).
385     // 4:   27390000        addiu   t9,t9,%lo(addr).
386     // 8:   03200008        jr      t9.
387     // c:   00000000        nop.
388     const unsigned LuiT9Instr = 0x3c190000, AdduiT9Instr = 0x27390000;
389     const unsigned JrT9Instr = 0x03200008, NopInstr = 0x0;
390
391     *StubAddr = LuiT9Instr;
392     StubAddr++;
393     *StubAddr = AdduiT9Instr;
394     StubAddr++;
395     *StubAddr = JrT9Instr;
396     StubAddr++;
397     *StubAddr = NopInstr;
398     return Addr;
399   } else if (Arch == Triple::ppc64) {
400     // PowerPC64 stub: the address points to a function descriptor
401     // instead of the function itself. Load the function address
402     // on r11 and sets it to control register. Also loads the function
403     // TOC in r2 and environment pointer to r11.
404     writeInt32BE(Addr,    0x3D800000); // lis   r12, highest(addr)
405     writeInt32BE(Addr+4,  0x618C0000); // ori   r12, higher(addr)
406     writeInt32BE(Addr+8,  0x798C07C6); // sldi  r12, r12, 32
407     writeInt32BE(Addr+12, 0x658C0000); // oris  r12, r12, h(addr)
408     writeInt32BE(Addr+16, 0x618C0000); // ori   r12, r12, l(addr)
409     writeInt32BE(Addr+20, 0xF8410028); // std   r2,  40(r1)
410     writeInt32BE(Addr+24, 0xE96C0000); // ld    r11, 0(r12)
411     writeInt32BE(Addr+28, 0xE84C0008); // ld    r2,  0(r12)
412     writeInt32BE(Addr+32, 0x7D6903A6); // mtctr r11
413     writeInt32BE(Addr+36, 0xE96C0010); // ld    r11, 16(r2)
414     writeInt32BE(Addr+40, 0x4E800420); // bctr
415
416     return Addr;
417   } else if (Arch == Triple::systemz) {
418     writeInt16BE(Addr,    0xC418);     // lgrl %r1,.+8
419     writeInt16BE(Addr+2,  0x0000);
420     writeInt16BE(Addr+4,  0x0004);
421     writeInt16BE(Addr+6,  0x07F1);     // brc 15,%r1
422     // 8-byte address stored at Addr + 8
423     return Addr;
424   }
425   return Addr;
426 }
427
428 // Assign an address to a symbol name and resolve all the relocations
429 // associated with it.
430 void RuntimeDyldImpl::reassignSectionAddress(unsigned SectionID,
431                                              uint64_t Addr) {
432   // The address to use for relocation resolution is not
433   // the address of the local section buffer. We must be doing
434   // a remote execution environment of some sort. Relocations can't
435   // be applied until all the sections have been moved.  The client must
436   // trigger this with a call to MCJIT::finalize() or
437   // RuntimeDyld::resolveRelocations().
438   //
439   // Addr is a uint64_t because we can't assume the pointer width
440   // of the target is the same as that of the host. Just use a generic
441   // "big enough" type.
442   Sections[SectionID].LoadAddress = Addr;
443 }
444
445 void RuntimeDyldImpl::resolveRelocationList(const RelocationList &Relocs,
446                                             uint64_t Value) {
447   for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
448     const RelocationEntry &RE = Relocs[i];
449     // Ignore relocations for sections that were not loaded
450     if (Sections[RE.SectionID].Address == 0)
451       continue;
452     resolveRelocation(RE, Value);
453   }
454 }
455
456 void RuntimeDyldImpl::resolveExternalSymbols() {
457   StringMap<RelocationList>::iterator i = ExternalSymbolRelocations.begin(),
458                                       e = ExternalSymbolRelocations.end();
459   for (; i != e; i++) {
460     StringRef Name = i->first();
461     RelocationList &Relocs = i->second;
462     SymbolTableMap::const_iterator Loc = GlobalSymbolTable.find(Name);
463     if (Loc == GlobalSymbolTable.end()) {
464       if (Name.size() == 0) {
465         // This is an absolute symbol, use an address of zero.
466         DEBUG(dbgs() << "Resolving absolute relocations." << "\n");
467         resolveRelocationList(Relocs, 0);
468       } else {
469         // This is an external symbol, try to get its address from
470         // MemoryManager.
471         uint8_t *Addr = (uint8_t*) MemMgr->getPointerToNamedFunction(Name.data(),
472                                                                    true);
473         DEBUG(dbgs() << "Resolving relocations Name: " << Name
474                 << "\t" << format("%p", Addr)
475                 << "\n");
476         resolveRelocationList(Relocs, (uintptr_t)Addr);
477       }
478     } else {
479       report_fatal_error("Expected external symbol");
480     }
481   }
482 }
483
484
485 //===----------------------------------------------------------------------===//
486 // RuntimeDyld class implementation
487 RuntimeDyld::RuntimeDyld(RTDyldMemoryManager *mm) {
488   // FIXME: There's a potential issue lurking here if a single instance of
489   // RuntimeDyld is used to load multiple objects.  The current implementation
490   // associates a single memory manager with a RuntimeDyld instance.  Even
491   // though the public class spawns a new 'impl' instance for each load,
492   // they share a single memory manager.  This can become a problem when page
493   // permissions are applied.
494   Dyld = 0;
495   MM = mm;
496 }
497
498 RuntimeDyld::~RuntimeDyld() {
499   delete Dyld;
500 }
501
502 ObjectImage *RuntimeDyld::loadObject(ObjectBuffer *InputBuffer) {
503   if (!Dyld) {
504     sys::fs::file_magic Type =
505         sys::fs::identify_magic(InputBuffer->getBuffer());
506     switch (Type) {
507     case sys::fs::file_magic::elf_relocatable:
508     case sys::fs::file_magic::elf_executable:
509     case sys::fs::file_magic::elf_shared_object:
510     case sys::fs::file_magic::elf_core:
511       Dyld = new RuntimeDyldELF(MM);
512       break;
513     case sys::fs::file_magic::macho_object:
514     case sys::fs::file_magic::macho_executable:
515     case sys::fs::file_magic::macho_fixed_virtual_memory_shared_lib:
516     case sys::fs::file_magic::macho_core:
517     case sys::fs::file_magic::macho_preload_executable:
518     case sys::fs::file_magic::macho_dynamically_linked_shared_lib:
519     case sys::fs::file_magic::macho_dynamic_linker:
520     case sys::fs::file_magic::macho_bundle:
521     case sys::fs::file_magic::macho_dynamically_linked_shared_lib_stub:
522     case sys::fs::file_magic::macho_dsym_companion:
523       Dyld = new RuntimeDyldMachO(MM);
524       break;
525     case sys::fs::file_magic::unknown:
526     case sys::fs::file_magic::bitcode:
527     case sys::fs::file_magic::archive:
528     case sys::fs::file_magic::coff_object:
529     case sys::fs::file_magic::pecoff_executable:
530       report_fatal_error("Incompatible object format!");
531     }
532   } else {
533     if (!Dyld->isCompatibleFormat(InputBuffer))
534       report_fatal_error("Incompatible object format!");
535   }
536
537   return Dyld->loadObject(InputBuffer);
538 }
539
540 void *RuntimeDyld::getSymbolAddress(StringRef Name) {
541   return Dyld->getSymbolAddress(Name);
542 }
543
544 uint64_t RuntimeDyld::getSymbolLoadAddress(StringRef Name) {
545   return Dyld->getSymbolLoadAddress(Name);
546 }
547
548 void RuntimeDyld::resolveRelocations() {
549   Dyld->resolveRelocations();
550 }
551
552 void RuntimeDyld::reassignSectionAddress(unsigned SectionID,
553                                          uint64_t Addr) {
554   Dyld->reassignSectionAddress(SectionID, Addr);
555 }
556
557 void RuntimeDyld::mapSectionAddress(const void *LocalAddress,
558                                     uint64_t TargetAddress) {
559   Dyld->mapSectionAddress(LocalAddress, TargetAddress);
560 }
561
562 StringRef RuntimeDyld::getErrorString() {
563   return Dyld->getErrorString();
564 }
565
566 StringRef RuntimeDyld::getEHFrameSection() {
567   return Dyld->getEHFrameSection();
568 }
569
570 } // end namespace llvm