1 //===-- ELFWriter.cpp - Target-independent ELF Writer code ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the target-independent ELF writer. This file writes out
11 // the ELF file in the following order:
14 // #2. '.text' section
15 // #3. '.data' section
16 // #4. '.bss' section (conceptual position in file)
18 // #X. '.shstrtab' section
21 // The entries in the section table are laid out as:
22 // #0. Null entry [required]
23 // #1. ".text" entry - the program code
24 // #2. ".data" entry - global variables with initializers. [ if needed ]
25 // #3. ".bss" entry - global variables without initializers. [ if needed ]
27 // #N. ".shstrtab" entry - String table for the section names.
29 // NOTE: This code should eventually be extended to support 64-bit ELF (this
30 // won't be hard), but we haven't done so yet!
32 //===----------------------------------------------------------------------===//
34 #include "llvm/CodeGen/ELFWriter.h"
35 #include "llvm/Module.h"
36 #include "llvm/CodeGen/MachineCodeEmitter.h"
37 #include "llvm/CodeGen/MachineConstantPool.h"
38 #include "llvm/Target/TargetMachine.h"
39 #include "llvm/Support/Mangler.h"
43 //===----------------------------------------------------------------------===//
44 // ELFCodeEmitter Implementation
45 //===----------------------------------------------------------------------===//
48 /// ELFCodeEmitter - This class is used by the ELFWriter to emit the code for
49 /// functions to the ELF file.
50 class ELFCodeEmitter : public MachineCodeEmitter {
52 ELFWriter::ELFSection *ES; // Section to write to.
53 std::vector<unsigned char> *OutBuffer;
56 ELFCodeEmitter(ELFWriter &ew) : EW(ew), OutBuffer(0) {}
58 void startFunction(MachineFunction &F);
59 void finishFunction(MachineFunction &F);
61 void emitConstantPool(MachineConstantPool *MCP) {
62 if (MCP->isEmpty()) return;
65 virtual void emitByte(unsigned char B) {
66 OutBuffer->push_back(B);
68 virtual void emitWordAt(unsigned W, unsigned *Ptr) {
71 virtual void emitWord(unsigned W) {
74 virtual uint64_t getCurrentPCValue() {
75 return OutBuffer->size();
77 virtual uint64_t getCurrentPCOffset() {
78 return OutBuffer->size()-FnStart;
80 void addRelocation(const MachineRelocation &MR) {
81 assert(0 && "relo not handled yet!");
83 virtual uint64_t getConstantPoolEntryAddress(unsigned Index) {
84 assert(0 && "CP not implementated yet!");
88 virtual unsigned char* allocateGlobal(unsigned size, unsigned alignment) {
89 assert(0 && "Globals not implemented yet!");
93 /// JIT SPECIFIC FUNCTIONS - DO NOT IMPLEMENT THESE HERE!
94 void startFunctionStub(unsigned StubSize) {
95 assert(0 && "JIT specific function called!");
98 void *finishFunctionStub(const Function *F) {
99 assert(0 && "JIT specific function called!");
106 /// startFunction - This callback is invoked when a new machine function is
107 /// about to be emitted.
108 void ELFCodeEmitter::startFunction(MachineFunction &F) {
109 // Align the output buffer to the appropriate alignment.
110 unsigned Align = 16; // FIXME: GENERICIZE!!
111 // Get the ELF Section that this function belongs in.
112 ES = &EW.getSection(".text", ELFWriter::ELFSection::SHT_PROGBITS,
113 ELFWriter::ELFSection::SHF_EXECINSTR |
114 ELFWriter::ELFSection::SHF_ALLOC);
115 OutBuffer = &ES->SectionData;
117 // Upgrade the section alignment if required.
118 if (ES->Align < Align) ES->Align = Align;
120 // Add padding zeros to the end of the buffer to make sure that the
121 // function will start on the correct byte alignment within the section.
122 size_t SectionOff = OutBuffer->size();
123 ELFWriter::align(*OutBuffer, Align);
125 FnStart = OutBuffer->size();
128 /// finishFunction - This callback is invoked after the function is completely
130 void ELFCodeEmitter::finishFunction(MachineFunction &F) {
131 // We now know the size of the function, add a symbol to represent it.
132 ELFWriter::ELFSym FnSym(F.getFunction());
134 // Figure out the binding (linkage) of the symbol.
135 switch (F.getFunction()->getLinkage()) {
137 // appending linkage is illegal for functions.
138 assert(0 && "Unknown linkage type!");
139 case GlobalValue::ExternalLinkage:
140 FnSym.SetBind(ELFWriter::ELFSym::STB_GLOBAL);
142 case GlobalValue::LinkOnceLinkage:
143 case GlobalValue::WeakLinkage:
144 FnSym.SetBind(ELFWriter::ELFSym::STB_WEAK);
146 case GlobalValue::InternalLinkage:
147 FnSym.SetBind(ELFWriter::ELFSym::STB_LOCAL);
151 ES->Size = OutBuffer->size();
153 FnSym.SetType(ELFWriter::ELFSym::STT_FUNC);
154 FnSym.SectionIdx = ES->SectionIdx;
155 FnSym.Value = FnStart; // Value = Offset from start of Section.
156 FnSym.Size = OutBuffer->size()-FnStart;
158 // Finally, add it to the symtab.
159 EW.SymbolTable.push_back(FnSym);
162 //===----------------------------------------------------------------------===//
163 // ELFWriter Implementation
164 //===----------------------------------------------------------------------===//
166 ELFWriter::ELFWriter(std::ostream &o, TargetMachine &tm) : O(o), TM(tm) {
167 e_machine = 0; // e_machine defaults to 'No Machine'
168 e_flags = 0; // e_flags defaults to 0, no flags.
170 is64Bit = TM.getTargetData().getPointerSizeInBits() == 64;
171 isLittleEndian = TM.getTargetData().isLittleEndian();
173 // Create the machine code emitter object for this target.
174 MCE = new ELFCodeEmitter(*this);
178 ELFWriter::~ELFWriter() {
182 // doInitialization - Emit the file header and all of the global variables for
183 // the module to the ELF file.
184 bool ELFWriter::doInitialization(Module &M) {
185 Mang = new Mangler(M);
187 // Local alias to shortenify coming code.
188 std::vector<unsigned char> &FH = FileHeader;
190 outbyte(FH, 0x7F); // EI_MAG0
191 outbyte(FH, 'E'); // EI_MAG1
192 outbyte(FH, 'L'); // EI_MAG2
193 outbyte(FH, 'F'); // EI_MAG3
194 outbyte(FH, is64Bit ? 2 : 1); // EI_CLASS
195 outbyte(FH, isLittleEndian ? 1 : 2); // EI_DATA
196 outbyte(FH, 1); // EI_VERSION
197 FH.resize(16); // EI_PAD up to 16 bytes.
199 // This should change for shared objects.
200 outhalf(FH, 1); // e_type = ET_REL
201 outhalf(FH, e_machine); // e_machine = whatever the target wants
202 outword(FH, 1); // e_version = 1
203 outaddr(FH, 0); // e_entry = 0 -> no entry point in .o file
204 outaddr(FH, 0); // e_phoff = 0 -> no program header for .o
206 ELFHeader_e_shoff_Offset = FH.size();
207 outaddr(FH, 0); // e_shoff
208 outword(FH, e_flags); // e_flags = whatever the target wants
210 outhalf(FH, is64Bit ? 64 : 52); // e_ehsize = ELF header size
211 outhalf(FH, 0); // e_phentsize = prog header entry size
212 outhalf(FH, 0); // e_phnum = # prog header entries = 0
213 outhalf(FH, is64Bit ? 64 : 40); // e_shentsize = sect hdr entry size
216 ELFHeader_e_shnum_Offset = FH.size();
217 outhalf(FH, 0); // e_shnum = # of section header ents
218 ELFHeader_e_shstrndx_Offset = FH.size();
219 outhalf(FH, 0); // e_shstrndx = Section # of '.shstrtab'
221 // Add the null section, which is required to be first in the file.
222 getSection("", 0, 0);
224 // Start up the symbol table. The first entry in the symtab is the null
226 SymbolTable.push_back(ELFSym(0));
231 void ELFWriter::EmitGlobal(GlobalVariable *GV) {
232 // If this is an external global, emit it now. TODO: Note that it would be
233 // better to ignore the symbol here and only add it to the symbol table if
235 if (!GV->hasInitializer()) {
236 ELFSym ExternalSym(GV);
237 ExternalSym.SetBind(ELFSym::STB_GLOBAL);
238 ExternalSym.SetType(ELFSym::STT_NOTYPE);
239 ExternalSym.SectionIdx = ELFSection::SHN_UNDEF;
240 SymbolTable.push_back(ExternalSym);
244 const Type *GVType = (const Type*)GV->getType();
245 unsigned Align = TM.getTargetData().getTypeAlignment(GVType);
246 unsigned Size = TM.getTargetData().getTypeSize(GVType);
248 // If this global has a zero initializer, it is part of the .bss or common
250 if (GV->getInitializer()->isNullValue()) {
251 // If this global is part of the common block, add it now. Variables are
252 // part of the common block if they are zero initialized and allowed to be
253 // merged with other symbols.
254 if (GV->hasLinkOnceLinkage() || GV->hasWeakLinkage()) {
255 ELFSym CommonSym(GV);
256 // Value for common symbols is the alignment required.
257 CommonSym.Value = Align;
258 CommonSym.Size = Size;
259 CommonSym.SetBind(ELFSym::STB_GLOBAL);
260 CommonSym.SetType(ELFSym::STT_OBJECT);
261 // TODO SOMEDAY: add ELF visibility.
262 CommonSym.SectionIdx = ELFSection::SHN_COMMON;
263 SymbolTable.push_back(CommonSym);
267 // Otherwise, this symbol is part of the .bss section. Emit it now.
269 // Handle alignment. Ensure section is aligned at least as much as required
271 ELFSection &BSSSection = getBSSSection();
272 BSSSection.Align = std::max(BSSSection.Align, Align);
274 // Within the section, emit enough virtual padding to get us to an alignment
277 BSSSection.Size = (BSSSection.Size + Align - 1) & ~(Align-1);
280 BSSSym.Value = BSSSection.Size;
282 BSSSym.SetType(ELFSym::STT_OBJECT);
284 switch (GV->getLinkage()) {
285 default: // weak/linkonce handled above
286 assert(0 && "Unexpected linkage type!");
287 case GlobalValue::AppendingLinkage: // FIXME: This should be improved!
288 case GlobalValue::ExternalLinkage:
289 BSSSym.SetBind(ELFSym::STB_GLOBAL);
291 case GlobalValue::InternalLinkage:
292 BSSSym.SetBind(ELFSym::STB_LOCAL);
296 // Set the idx of the .bss section
297 BSSSym.SectionIdx = BSSSection.SectionIdx;
298 SymbolTable.push_back(BSSSym);
300 // Reserve space in the .bss section for this symbol.
301 BSSSection.Size += Size;
305 // FIXME: handle .rodata
306 //assert(!GV->isConstant() && "unimp");
308 // FIXME: handle .data
309 //assert(0 && "unimp");
313 bool ELFWriter::runOnMachineFunction(MachineFunction &MF) {
314 // Nothing to do here, this is all done through the MCE object above.
318 /// doFinalization - Now that the module has been completely processed, emit
319 /// the ELF file to 'O'.
320 bool ELFWriter::doFinalization(Module &M) {
321 // Okay, the ELF header and .text sections have been completed, build the
322 // .data, .bss, and "common" sections next.
323 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
327 // Emit the symbol table now, if non-empty.
330 // FIXME: Emit the relocations now.
332 // Emit the string table for the sections in the ELF file we have.
333 EmitSectionTableStringTable();
335 // Emit the sections to the .o file, and emit the section table for the file.
336 OutputSectionsAndSectionTable();
338 // We are done with the abstract symbols.
342 // Release the name mangler object.
343 delete Mang; Mang = 0;
347 /// EmitSymbolTable - If the current symbol table is non-empty, emit the string
348 /// table for it and then the symbol table itself.
349 void ELFWriter::EmitSymbolTable() {
350 if (SymbolTable.size() == 1) return; // Only the null entry.
352 // FIXME: compact all local symbols to the start of the symtab.
353 unsigned FirstNonLocalSymbol = 1;
355 ELFSection &StrTab = getSection(".strtab", ELFSection::SHT_STRTAB, 0);
358 DataBuffer &StrTabBuf = StrTab.SectionData;
360 // Set the zero'th symbol to a null byte, as required.
361 outbyte(StrTabBuf, 0);
362 SymbolTable[0].NameIdx = 0;
364 for (unsigned i = 1, e = SymbolTable.size(); i != e; ++i) {
365 // Use the name mangler to uniquify the LLVM symbol.
366 std::string Name = Mang->getValueName(SymbolTable[i].GV);
369 SymbolTable[i].NameIdx = 0;
371 SymbolTable[i].NameIdx = Index;
373 // Add the name to the output buffer, including the null terminator.
374 StrTabBuf.insert(StrTabBuf.end(), Name.begin(), Name.end());
376 // Add a null terminator.
377 StrTabBuf.push_back(0);
379 // Keep track of the number of bytes emitted to this section.
380 Index += Name.size()+1;
383 assert(Index == StrTabBuf.size());
386 // Now that we have emitted the string table and know the offset into the
387 // string table of each symbol, emit the symbol table itself.
388 ELFSection &SymTab = getSection(".symtab", ELFSection::SHT_SYMTAB, 0);
389 SymTab.Align = is64Bit ? 8 : 4;
390 SymTab.Link = SymTab.SectionIdx; // Section Index of .strtab.
391 SymTab.Info = FirstNonLocalSymbol; // First non-STB_LOCAL symbol.
392 SymTab.EntSize = 16; // Size of each symtab entry. FIXME: wrong for ELF64
393 DataBuffer &SymTabBuf = SymTab.SectionData;
395 if (!is64Bit) { // 32-bit and 64-bit formats are shuffled a bit.
396 for (unsigned i = 0, e = SymbolTable.size(); i != e; ++i) {
397 ELFSym &Sym = SymbolTable[i];
398 outword(SymTabBuf, Sym.NameIdx);
399 outaddr32(SymTabBuf, Sym.Value);
400 outword(SymTabBuf, Sym.Size);
401 outbyte(SymTabBuf, Sym.Info);
402 outbyte(SymTabBuf, Sym.Other);
403 outhalf(SymTabBuf, Sym.SectionIdx);
406 for (unsigned i = 0, e = SymbolTable.size(); i != e; ++i) {
407 ELFSym &Sym = SymbolTable[i];
408 outword(SymTabBuf, Sym.NameIdx);
409 outbyte(SymTabBuf, Sym.Info);
410 outbyte(SymTabBuf, Sym.Other);
411 outhalf(SymTabBuf, Sym.SectionIdx);
412 outaddr64(SymTabBuf, Sym.Value);
413 outxword(SymTabBuf, Sym.Size);
417 SymTab.Size = SymTabBuf.size();
420 /// EmitSectionTableStringTable - This method adds and emits a section for the
421 /// ELF Section Table string table: the string table that holds all of the
423 void ELFWriter::EmitSectionTableStringTable() {
424 // First step: add the section for the string table to the list of sections:
425 ELFSection &SHStrTab = getSection(".shstrtab", ELFSection::SHT_STRTAB, 0);
427 // Now that we know which section number is the .shstrtab section, update the
428 // e_shstrndx entry in the ELF header.
429 fixhalf(FileHeader, SHStrTab.SectionIdx, ELFHeader_e_shstrndx_Offset);
431 // Set the NameIdx of each section in the string table and emit the bytes for
434 DataBuffer &Buf = SHStrTab.SectionData;
436 for (std::list<ELFSection>::iterator I = SectionList.begin(),
437 E = SectionList.end(); I != E; ++I) {
438 // Set the index into the table. Note if we have lots of entries with
439 // common suffixes, we could memoize them here if we cared.
442 // Add the name to the output buffer, including the null terminator.
443 Buf.insert(Buf.end(), I->Name.begin(), I->Name.end());
445 // Add a null terminator.
448 // Keep track of the number of bytes emitted to this section.
449 Index += I->Name.size()+1;
452 // Set the size of .shstrtab now that we know what it is.
453 assert(Index == Buf.size());
454 SHStrTab.Size = Index;
457 /// OutputSectionsAndSectionTable - Now that we have constructed the file header
458 /// and all of the sections, emit these to the ostream destination and emit the
460 void ELFWriter::OutputSectionsAndSectionTable() {
461 // Pass #1: Compute the file offset for each section.
462 size_t FileOff = FileHeader.size(); // File header first.
464 // Emit all of the section data in order.
465 for (std::list<ELFSection>::iterator I = SectionList.begin(),
466 E = SectionList.end(); I != E; ++I) {
467 // Align FileOff to whatever the alignment restrictions of the section are.
469 FileOff = (FileOff+I->Align-1) & ~(I->Align-1);
471 FileOff += I->SectionData.size();
474 // Align Section Header.
475 unsigned TableAlign = is64Bit ? 8 : 4;
476 FileOff = (FileOff+TableAlign-1) & ~(TableAlign-1);
478 // Now that we know where all of the sections will be emitted, set the e_shnum
479 // entry in the ELF header.
480 fixhalf(FileHeader, NumSections, ELFHeader_e_shnum_Offset);
482 // Now that we know the offset in the file of the section table, update the
483 // e_shoff address in the ELF header.
484 fixaddr(FileHeader, FileOff, ELFHeader_e_shoff_Offset);
486 // Now that we know all of the data in the file header, emit it and all of the
488 O.write((char*)&FileHeader[0], FileHeader.size());
489 FileOff = FileHeader.size();
490 DataBuffer().swap(FileHeader);
494 // Emit all of the section data and build the section table itself.
495 while (!SectionList.empty()) {
496 const ELFSection &S = *SectionList.begin();
498 // Align FileOff to whatever the alignment restrictions of the section are.
500 for (size_t NewFileOff = (FileOff+S.Align-1) & ~(S.Align-1);
501 FileOff != NewFileOff; ++FileOff)
503 O.write((char*)&S.SectionData[0], S.SectionData.size());
504 FileOff += S.SectionData.size();
506 outword(Table, S.NameIdx); // sh_name - Symbol table name idx
507 outword(Table, S.Type); // sh_type - Section contents & semantics
508 outword(Table, S.Flags); // sh_flags - Section flags.
509 outaddr(Table, S.Addr); // sh_addr - The mem addr this section is in.
510 outaddr(Table, S.Offset); // sh_offset - Offset from the file start.
511 outword(Table, S.Size); // sh_size - The section size.
512 outword(Table, S.Link); // sh_link - Section header table index link.
513 outword(Table, S.Info); // sh_info - Auxillary information.
514 outword(Table, S.Align); // sh_addralign - Alignment of section.
515 outword(Table, S.EntSize); // sh_entsize - Size of entries in the section.
517 SectionList.pop_front();
520 // Align output for the section table.
521 for (size_t NewFileOff = (FileOff+TableAlign-1) & ~(TableAlign-1);
522 FileOff != NewFileOff; ++FileOff)
525 // Emit the section table itself.
526 O.write((char*)&Table[0], Table.size());