1 //===-- MachOWriter.cpp - Target-independent Mach-O Writer code -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the target-independent Mach-O writer. This file writes
11 // out the Mach-O file in the following order:
13 // #1 FatHeader (universal-only)
14 // #2 FatArch (universal-only, 1 per universal arch)
23 //===----------------------------------------------------------------------===//
25 #include "MachOWriter.h"
26 #include "llvm/Constants.h"
27 #include "llvm/DerivedTypes.h"
28 #include "llvm/Module.h"
29 #include "llvm/PassManager.h"
30 #include "llvm/CodeGen/FileWriters.h"
31 #include "llvm/CodeGen/MachineCodeEmitter.h"
32 #include "llvm/CodeGen/MachineConstantPool.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/Target/TargetAsmInfo.h"
35 #include "llvm/Target/TargetJITInfo.h"
36 #include "llvm/Support/Mangler.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/OutputBuffer.h"
39 #include "llvm/Support/Streams.h"
44 /// AddMachOWriter - Concrete function to add the Mach-O writer to the function
46 MachineCodeEmitter *llvm::AddMachOWriter(FunctionPassManager &FPM,
49 MachOWriter *MOW = new MachOWriter(O, TM);
51 return &MOW->getMachineCodeEmitter();
54 //===----------------------------------------------------------------------===//
55 // MachOCodeEmitter Implementation
56 //===----------------------------------------------------------------------===//
59 /// MachOCodeEmitter - This class is used by the MachOWriter to emit the code
60 /// for functions to the Mach-O file.
61 class MachOCodeEmitter : public MachineCodeEmitter {
64 /// Target machine description.
67 /// is64Bit/isLittleEndian - This information is inferred from the target
68 /// machine directly, indicating what header values and flags to set.
69 bool is64Bit, isLittleEndian;
71 /// Relocations - These are the relocations that the function needs, as
73 std::vector<MachineRelocation> Relocations;
75 /// CPLocations - This is a map of constant pool indices to offsets from the
76 /// start of the section for that constant pool index.
77 std::vector<intptr_t> CPLocations;
79 /// CPSections - This is a map of constant pool indices to the MachOSection
80 /// containing the constant pool entry for that index.
81 std::vector<unsigned> CPSections;
83 /// JTLocations - This is a map of jump table indices to offsets from the
84 /// start of the section for that jump table index.
85 std::vector<intptr_t> JTLocations;
87 /// MBBLocations - This vector is a mapping from MBB ID's to their address.
88 /// It is filled in by the StartMachineBasicBlock callback and queried by
89 /// the getMachineBasicBlockAddress callback.
90 std::vector<intptr_t> MBBLocations;
93 MachOCodeEmitter(MachOWriter &mow) : MOW(mow), TM(MOW.TM) {
94 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
95 isLittleEndian = TM.getTargetData()->isLittleEndian();
98 virtual void startFunction(MachineFunction &MF);
99 virtual bool finishFunction(MachineFunction &MF);
101 virtual void addRelocation(const MachineRelocation &MR) {
102 Relocations.push_back(MR);
105 void emitConstantPool(MachineConstantPool *MCP);
106 void emitJumpTables(MachineJumpTableInfo *MJTI);
108 virtual intptr_t getConstantPoolEntryAddress(unsigned Index) const {
109 assert(CPLocations.size() > Index && "CP not emitted!");
110 return CPLocations[Index];
112 virtual intptr_t getJumpTableEntryAddress(unsigned Index) const {
113 assert(JTLocations.size() > Index && "JT not emitted!");
114 return JTLocations[Index];
117 virtual void StartMachineBasicBlock(MachineBasicBlock *MBB) {
118 if (MBBLocations.size() <= (unsigned)MBB->getNumber())
119 MBBLocations.resize((MBB->getNumber()+1)*2);
120 MBBLocations[MBB->getNumber()] = getCurrentPCOffset();
123 virtual intptr_t getMachineBasicBlockAddress(MachineBasicBlock *MBB) const {
124 assert(MBBLocations.size() > (unsigned)MBB->getNumber() &&
125 MBBLocations[MBB->getNumber()] && "MBB not emitted!");
126 return MBBLocations[MBB->getNumber()];
129 virtual intptr_t getLabelAddress(uint64_t Label) const {
130 assert(0 && "get Label not implemented");
135 virtual void emitLabel(uint64_t LabelID) {
136 assert(0 && "emit Label not implemented");
141 virtual void setModuleInfo(llvm::MachineModuleInfo* MMI) { }
143 /// JIT SPECIFIC FUNCTIONS - DO NOT IMPLEMENT THESE HERE!
144 virtual void startFunctionStub(unsigned StubSize, unsigned Alignment = 1) {
145 assert(0 && "JIT specific function called!");
148 virtual void *finishFunctionStub(const Function *F) {
149 assert(0 && "JIT specific function called!");
156 /// startFunction - This callback is invoked when a new machine function is
157 /// about to be emitted.
158 void MachOCodeEmitter::startFunction(MachineFunction &MF) {
159 const TargetData *TD = TM.getTargetData();
160 const Function *F = MF.getFunction();
162 // Align the output buffer to the appropriate alignment, power of 2.
163 unsigned FnAlign = F->getAlignment();
164 unsigned TDAlign = TD->getPrefTypeAlignment(F->getType());
165 unsigned Align = Log2_32(std::max(FnAlign, TDAlign));
166 assert(!(Align & (Align-1)) && "Alignment is not a power of two!");
168 // Get the Mach-O Section that this function belongs in.
169 MachOWriter::MachOSection *MOS = MOW.getTextSection();
171 // FIXME: better memory management
172 MOS->SectionData.reserve(4096);
173 BufferBegin = &MOS->SectionData[0];
174 BufferEnd = BufferBegin + MOS->SectionData.capacity();
176 // Upgrade the section alignment if required.
177 if (MOS->align < Align) MOS->align = Align;
179 // Round the size up to the correct alignment for starting the new function.
180 if ((MOS->size & ((1 << Align) - 1)) != 0) {
181 MOS->size += (1 << Align);
182 MOS->size &= ~((1 << Align) - 1);
185 // FIXME: Using MOS->size directly here instead of calculating it from the
186 // output buffer size (impossible because the code emitter deals only in raw
187 // bytes) forces us to manually synchronize size and write padding zero bytes
188 // to the output buffer for all non-text sections. For text sections, we do
189 // not synchonize the output buffer, and we just blow up if anyone tries to
190 // write non-code to it. An assert should probably be added to
191 // AddSymbolToSection to prevent calling it on the text section.
192 CurBufferPtr = BufferBegin + MOS->size;
194 // Clear per-function data structures.
198 MBBLocations.clear();
201 /// finishFunction - This callback is invoked after the function is completely
203 bool MachOCodeEmitter::finishFunction(MachineFunction &MF) {
204 // Get the Mach-O Section that this function belongs in.
205 MachOWriter::MachOSection *MOS = MOW.getTextSection();
207 // Get a symbol for the function to add to the symbol table
208 // FIXME: it seems like we should call something like AddSymbolToSection
209 // in startFunction rather than changing the section size and symbol n_value
211 const GlobalValue *FuncV = MF.getFunction();
212 MachOSym FnSym(FuncV, MOW.Mang->getValueName(FuncV), MOS->Index, TM);
213 FnSym.n_value = MOS->size;
214 MOS->size = CurBufferPtr - BufferBegin;
216 // Emit constant pool to appropriate section(s)
217 emitConstantPool(MF.getConstantPool());
219 // Emit jump tables to appropriate section
220 emitJumpTables(MF.getJumpTableInfo());
222 // If we have emitted any relocations to function-specific objects such as
223 // basic blocks, constant pools entries, or jump tables, record their
224 // addresses now so that we can rewrite them with the correct addresses
226 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
227 MachineRelocation &MR = Relocations[i];
230 if (MR.isBasicBlock()) {
231 Addr = getMachineBasicBlockAddress(MR.getBasicBlock());
232 MR.setConstantVal(MOS->Index);
233 MR.setResultPointer((void*)Addr);
234 } else if (MR.isJumpTableIndex()) {
235 Addr = getJumpTableEntryAddress(MR.getJumpTableIndex());
236 MR.setConstantVal(MOW.getJumpTableSection()->Index);
237 MR.setResultPointer((void*)Addr);
238 } else if (MR.isConstantPoolIndex()) {
239 Addr = getConstantPoolEntryAddress(MR.getConstantPoolIndex());
240 MR.setConstantVal(CPSections[MR.getConstantPoolIndex()]);
241 MR.setResultPointer((void*)Addr);
242 } else if (MR.isGlobalValue()) {
243 // FIXME: This should be a set or something that uniques
244 MOW.PendingGlobals.push_back(MR.getGlobalValue());
246 assert(0 && "Unhandled relocation type");
248 MOS->Relocations.push_back(MR);
252 // Finally, add it to the symtab.
253 MOW.SymbolTable.push_back(FnSym);
257 /// emitConstantPool - For each constant pool entry, figure out which section
258 /// the constant should live in, allocate space for it, and emit it to the
259 /// Section data buffer.
260 void MachOCodeEmitter::emitConstantPool(MachineConstantPool *MCP) {
261 const std::vector<MachineConstantPoolEntry> &CP = MCP->getConstants();
262 if (CP.empty()) return;
264 // FIXME: handle PIC codegen
265 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
266 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
268 // Although there is no strict necessity that I am aware of, we will do what
269 // gcc for OS X does and put each constant pool entry in a section of constant
270 // objects of a certain size. That means that float constants go in the
271 // literal4 section, and double objects go in literal8, etc.
273 // FIXME: revisit this decision if we ever do the "stick everything into one
274 // "giant object for PIC" optimization.
275 for (unsigned i = 0, e = CP.size(); i != e; ++i) {
276 const Type *Ty = CP[i].getType();
277 unsigned Size = TM.getTargetData()->getABITypeSize(Ty);
279 MachOWriter::MachOSection *Sec = MOW.getConstSection(CP[i].Val.ConstVal);
280 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
282 CPLocations.push_back(Sec->SectionData.size());
283 CPSections.push_back(Sec->Index);
285 // FIXME: remove when we have unified size + output buffer
288 // Allocate space in the section for the global.
289 // FIXME: need alignment?
290 // FIXME: share between here and AddSymbolToSection?
291 for (unsigned j = 0; j < Size; ++j)
292 SecDataOut.outbyte(0);
294 MOW.InitMem(CP[i].Val.ConstVal, &Sec->SectionData[0], CPLocations[i],
295 TM.getTargetData(), Sec->Relocations);
299 /// emitJumpTables - Emit all the jump tables for a given jump table info
300 /// record to the appropriate section.
301 void MachOCodeEmitter::emitJumpTables(MachineJumpTableInfo *MJTI) {
302 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
303 if (JT.empty()) return;
305 // FIXME: handle PIC codegen
306 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
307 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
309 MachOWriter::MachOSection *Sec = MOW.getJumpTableSection();
310 unsigned TextSecIndex = MOW.getTextSection()->Index;
311 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
313 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
314 // For each jump table, record its offset from the start of the section,
315 // reserve space for the relocations to the MBBs, and add the relocations.
316 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
317 JTLocations.push_back(Sec->SectionData.size());
318 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi) {
319 MachineRelocation MR(MOW.GetJTRelocation(Sec->SectionData.size(),
321 MR.setResultPointer((void *)JTLocations[i]);
322 MR.setConstantVal(TextSecIndex);
323 Sec->Relocations.push_back(MR);
324 SecDataOut.outaddr(0);
327 // FIXME: remove when we have unified size + output buffer
328 Sec->size = Sec->SectionData.size();
331 //===----------------------------------------------------------------------===//
332 // MachOWriter Implementation
333 //===----------------------------------------------------------------------===//
335 char MachOWriter::ID = 0;
336 MachOWriter::MachOWriter(std::ostream &o, TargetMachine &tm)
337 : MachineFunctionPass((intptr_t)&ID), O(o), TM(tm) {
338 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
339 isLittleEndian = TM.getTargetData()->isLittleEndian();
341 // Create the machine code emitter object for this target.
342 MCE = new MachOCodeEmitter(*this);
345 MachOWriter::~MachOWriter() {
349 void MachOWriter::AddSymbolToSection(MachOSection *Sec, GlobalVariable *GV) {
350 const Type *Ty = GV->getType()->getElementType();
351 unsigned Size = TM.getTargetData()->getABITypeSize(Ty);
352 unsigned Align = TM.getTargetData()->getPreferredAlignment(GV);
354 // Reserve space in the .bss section for this symbol while maintaining the
355 // desired section alignment, which must be at least as much as required by
357 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
360 uint64_t OrigSize = Sec->size;
361 Align = Log2_32(Align);
362 Sec->align = std::max(unsigned(Sec->align), Align);
363 Sec->size = (Sec->size + Align - 1) & ~(Align-1);
365 // Add alignment padding to buffer as well.
366 // FIXME: remove when we have unified size + output buffer
367 unsigned AlignedSize = Sec->size - OrigSize;
368 for (unsigned i = 0; i < AlignedSize; ++i)
369 SecDataOut.outbyte(0);
371 // Globals without external linkage apparently do not go in the symbol table.
372 if (GV->getLinkage() != GlobalValue::InternalLinkage) {
373 MachOSym Sym(GV, Mang->getValueName(GV), Sec->Index, TM);
374 Sym.n_value = Sec->size;
375 SymbolTable.push_back(Sym);
378 // Record the offset of the symbol, and then allocate space for it.
379 // FIXME: remove when we have unified size + output buffer
382 // Now that we know what section the GlovalVariable is going to be emitted
383 // into, update our mappings.
384 // FIXME: We may also need to update this when outputting non-GlobalVariable
385 // GlobalValues such as functions.
387 GVOffset[GV] = Sec->SectionData.size();
389 // Allocate space in the section for the global.
390 for (unsigned i = 0; i < Size; ++i)
391 SecDataOut.outbyte(0);
394 void MachOWriter::EmitGlobal(GlobalVariable *GV) {
395 const Type *Ty = GV->getType()->getElementType();
396 unsigned Size = TM.getTargetData()->getABITypeSize(Ty);
397 bool NoInit = !GV->hasInitializer();
399 // If this global has a zero initializer, it is part of the .bss or common
401 if (NoInit || GV->getInitializer()->isNullValue()) {
402 // If this global is part of the common block, add it now. Variables are
403 // part of the common block if they are zero initialized and allowed to be
404 // merged with other symbols.
405 if (NoInit || GV->hasLinkOnceLinkage() || GV->hasWeakLinkage()) {
406 MachOSym ExtOrCommonSym(GV, Mang->getValueName(GV), MachOSym::NO_SECT,TM);
407 // For undefined (N_UNDF) external (N_EXT) types, n_value is the size in
408 // bytes of the symbol.
409 ExtOrCommonSym.n_value = Size;
410 SymbolTable.push_back(ExtOrCommonSym);
411 // Remember that we've seen this symbol
415 // Otherwise, this symbol is part of the .bss section.
416 MachOSection *BSS = getBSSSection();
417 AddSymbolToSection(BSS, GV);
421 // Scalar read-only data goes in a literal section if the scalar is 4, 8, or
422 // 16 bytes, or a cstring. Other read only data goes into a regular const
423 // section. Read-write data goes in the data section.
424 MachOSection *Sec = GV->isConstant() ? getConstSection(GV->getInitializer()) :
426 AddSymbolToSection(Sec, GV);
427 InitMem(GV->getInitializer(), &Sec->SectionData[0], GVOffset[GV],
428 TM.getTargetData(), Sec->Relocations);
432 bool MachOWriter::runOnMachineFunction(MachineFunction &MF) {
433 // Nothing to do here, this is all done through the MCE object.
437 bool MachOWriter::doInitialization(Module &M) {
438 // Set the magic value, now that we know the pointer size and endianness
439 Header.setMagic(isLittleEndian, is64Bit);
442 // FIXME: this only works for object files, we do not support the creation
443 // of dynamic libraries or executables at this time.
444 Header.filetype = MachOHeader::MH_OBJECT;
446 Mang = new Mangler(M);
450 /// doFinalization - Now that the module has been completely processed, emit
451 /// the Mach-O file to 'O'.
452 bool MachOWriter::doFinalization(Module &M) {
453 // FIXME: we don't handle debug info yet, we should probably do that.
455 // Okay, the.text section has been completed, build the .data, .bss, and
456 // "common" sections next.
457 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
461 // Emit the header and load commands.
462 EmitHeaderAndLoadCommands();
464 // Emit the various sections and their relocation info.
467 // Write the symbol table and the string table to the end of the file.
468 O.write((char*)&SymT[0], SymT.size());
469 O.write((char*)&StrT[0], StrT.size());
471 // We are done with the abstract symbols.
474 DynamicSymbolTable.clear();
476 // Release the name mangler object.
477 delete Mang; Mang = 0;
481 void MachOWriter::EmitHeaderAndLoadCommands() {
482 // Step #0: Fill in the segment load command size, since we need it to figure
483 // out the rest of the header fields
484 MachOSegment SEG("", is64Bit);
485 SEG.nsects = SectionList.size();
486 SEG.cmdsize = SEG.cmdSize(is64Bit) +
487 SEG.nsects * SectionList[0]->cmdSize(is64Bit);
489 // Step #1: calculate the number of load commands. We always have at least
490 // one, for the LC_SEGMENT load command, plus two for the normal
491 // and dynamic symbol tables, if there are any symbols.
492 Header.ncmds = SymbolTable.empty() ? 1 : 3;
494 // Step #2: calculate the size of the load commands
495 Header.sizeofcmds = SEG.cmdsize;
496 if (!SymbolTable.empty())
497 Header.sizeofcmds += SymTab.cmdsize + DySymTab.cmdsize;
499 // Step #3: write the header to the file
500 // Local alias to shortenify coming code.
501 DataBuffer &FH = Header.HeaderData;
502 OutputBuffer FHOut(FH, is64Bit, isLittleEndian);
504 FHOut.outword(Header.magic);
505 FHOut.outword(TM.getMachOWriterInfo()->getCPUType());
506 FHOut.outword(TM.getMachOWriterInfo()->getCPUSubType());
507 FHOut.outword(Header.filetype);
508 FHOut.outword(Header.ncmds);
509 FHOut.outword(Header.sizeofcmds);
510 FHOut.outword(Header.flags);
512 FHOut.outword(Header.reserved);
514 // Step #4: Finish filling in the segment load command and write it out
515 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
516 E = SectionList.end(); I != E; ++I)
517 SEG.filesize += (*I)->size;
519 SEG.vmsize = SEG.filesize;
520 SEG.fileoff = Header.cmdSize(is64Bit) + Header.sizeofcmds;
522 FHOut.outword(SEG.cmd);
523 FHOut.outword(SEG.cmdsize);
524 FHOut.outstring(SEG.segname, 16);
525 FHOut.outaddr(SEG.vmaddr);
526 FHOut.outaddr(SEG.vmsize);
527 FHOut.outaddr(SEG.fileoff);
528 FHOut.outaddr(SEG.filesize);
529 FHOut.outword(SEG.maxprot);
530 FHOut.outword(SEG.initprot);
531 FHOut.outword(SEG.nsects);
532 FHOut.outword(SEG.flags);
534 // Step #5: Finish filling in the fields of the MachOSections
535 uint64_t currentAddr = 0;
536 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
537 E = SectionList.end(); I != E; ++I) {
538 MachOSection *MOS = *I;
539 MOS->addr = currentAddr;
540 MOS->offset = currentAddr + SEG.fileoff;
542 // FIXME: do we need to do something with alignment here?
543 currentAddr += MOS->size;
546 // Step #6: Emit the symbol table to temporary buffers, so that we know the
547 // size of the string table when we write the next load command. This also
548 // sorts and assigns indices to each of the symbols, which is necessary for
549 // emitting relocations to externally-defined objects.
550 BufferSymbolAndStringTable();
552 // Step #7: Calculate the number of relocations for each section and write out
553 // the section commands for each section
554 currentAddr += SEG.fileoff;
555 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
556 E = SectionList.end(); I != E; ++I) {
557 MachOSection *MOS = *I;
558 // Convert the relocations to target-specific relocations, and fill in the
559 // relocation offset for this section.
560 CalculateRelocations(*MOS);
561 MOS->reloff = MOS->nreloc ? currentAddr : 0;
562 currentAddr += MOS->nreloc * 8;
564 // write the finalized section command to the output buffer
565 FHOut.outstring(MOS->sectname, 16);
566 FHOut.outstring(MOS->segname, 16);
567 FHOut.outaddr(MOS->addr);
568 FHOut.outaddr(MOS->size);
569 FHOut.outword(MOS->offset);
570 FHOut.outword(MOS->align);
571 FHOut.outword(MOS->reloff);
572 FHOut.outword(MOS->nreloc);
573 FHOut.outword(MOS->flags);
574 FHOut.outword(MOS->reserved1);
575 FHOut.outword(MOS->reserved2);
577 FHOut.outword(MOS->reserved3);
580 // Step #8: Emit LC_SYMTAB/LC_DYSYMTAB load commands
581 SymTab.symoff = currentAddr;
582 SymTab.nsyms = SymbolTable.size();
583 SymTab.stroff = SymTab.symoff + SymT.size();
584 SymTab.strsize = StrT.size();
585 FHOut.outword(SymTab.cmd);
586 FHOut.outword(SymTab.cmdsize);
587 FHOut.outword(SymTab.symoff);
588 FHOut.outword(SymTab.nsyms);
589 FHOut.outword(SymTab.stroff);
590 FHOut.outword(SymTab.strsize);
592 // FIXME: set DySymTab fields appropriately
593 // We should probably just update these in BufferSymbolAndStringTable since
594 // thats where we're partitioning up the different kinds of symbols.
595 FHOut.outword(DySymTab.cmd);
596 FHOut.outword(DySymTab.cmdsize);
597 FHOut.outword(DySymTab.ilocalsym);
598 FHOut.outword(DySymTab.nlocalsym);
599 FHOut.outword(DySymTab.iextdefsym);
600 FHOut.outword(DySymTab.nextdefsym);
601 FHOut.outword(DySymTab.iundefsym);
602 FHOut.outword(DySymTab.nundefsym);
603 FHOut.outword(DySymTab.tocoff);
604 FHOut.outword(DySymTab.ntoc);
605 FHOut.outword(DySymTab.modtaboff);
606 FHOut.outword(DySymTab.nmodtab);
607 FHOut.outword(DySymTab.extrefsymoff);
608 FHOut.outword(DySymTab.nextrefsyms);
609 FHOut.outword(DySymTab.indirectsymoff);
610 FHOut.outword(DySymTab.nindirectsyms);
611 FHOut.outword(DySymTab.extreloff);
612 FHOut.outword(DySymTab.nextrel);
613 FHOut.outword(DySymTab.locreloff);
614 FHOut.outword(DySymTab.nlocrel);
616 O.write((char*)&FH[0], FH.size());
619 /// EmitSections - Now that we have constructed the file header and load
620 /// commands, emit the data for each section to the file.
621 void MachOWriter::EmitSections() {
622 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
623 E = SectionList.end(); I != E; ++I)
624 // Emit the contents of each section
625 O.write((char*)&(*I)->SectionData[0], (*I)->size);
626 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
627 E = SectionList.end(); I != E; ++I)
628 // Emit the relocation entry data for each section.
629 O.write((char*)&(*I)->RelocBuffer[0], (*I)->RelocBuffer.size());
632 /// PartitionByLocal - Simple boolean predicate that returns true if Sym is
633 /// a local symbol rather than an external symbol.
634 bool MachOWriter::PartitionByLocal(const MachOSym &Sym) {
635 return (Sym.n_type & (MachOSym::N_EXT | MachOSym::N_PEXT)) == 0;
638 /// PartitionByDefined - Simple boolean predicate that returns true if Sym is
639 /// defined in this module.
640 bool MachOWriter::PartitionByDefined(const MachOSym &Sym) {
641 // FIXME: Do N_ABS or N_INDR count as defined?
642 return (Sym.n_type & MachOSym::N_SECT) == MachOSym::N_SECT;
645 /// BufferSymbolAndStringTable - Sort the symbols we encountered and assign them
646 /// each a string table index so that they appear in the correct order in the
648 void MachOWriter::BufferSymbolAndStringTable() {
649 // The order of the symbol table is:
651 // 2. defined external symbols (sorted by name)
652 // 3. undefined external symbols (sorted by name)
654 // Before sorting the symbols, check the PendingGlobals for any undefined
655 // globals that need to be put in the symbol table.
656 for (std::vector<GlobalValue*>::iterator I = PendingGlobals.begin(),
657 E = PendingGlobals.end(); I != E; ++I) {
658 if (GVOffset[*I] == 0 && GVSection[*I] == 0) {
659 MachOSym UndfSym(*I, Mang->getValueName(*I), MachOSym::NO_SECT, TM);
660 SymbolTable.push_back(UndfSym);
665 // Sort the symbols by name, so that when we partition the symbols by scope
666 // of definition, we won't have to sort by name within each partition.
667 std::sort(SymbolTable.begin(), SymbolTable.end(), MachOSymCmp());
669 // Parition the symbol table entries so that all local symbols come before
670 // all symbols with external linkage. { 1 | 2 3 }
671 std::partition(SymbolTable.begin(), SymbolTable.end(), PartitionByLocal);
673 // Advance iterator to beginning of external symbols and partition so that
674 // all external symbols defined in this module come before all external
675 // symbols defined elsewhere. { 1 | 2 | 3 }
676 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
677 E = SymbolTable.end(); I != E; ++I) {
678 if (!PartitionByLocal(*I)) {
679 std::partition(I, E, PartitionByDefined);
684 // Calculate the starting index for each of the local, extern defined, and
685 // undefined symbols, as well as the number of each to put in the LC_DYSYMTAB
687 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
688 E = SymbolTable.end(); I != E; ++I) {
689 if (PartitionByLocal(*I)) {
690 ++DySymTab.nlocalsym;
691 ++DySymTab.iextdefsym;
692 ++DySymTab.iundefsym;
693 } else if (PartitionByDefined(*I)) {
694 ++DySymTab.nextdefsym;
695 ++DySymTab.iundefsym;
697 ++DySymTab.nundefsym;
701 // Write out a leading zero byte when emitting string table, for n_strx == 0
702 // which means an empty string.
703 OutputBuffer StrTOut(StrT, is64Bit, isLittleEndian);
706 // The order of the string table is:
707 // 1. strings for external symbols
708 // 2. strings for local symbols
709 // Since this is the opposite order from the symbol table, which we have just
710 // sorted, we can walk the symbol table backwards to output the string table.
711 for (std::vector<MachOSym>::reverse_iterator I = SymbolTable.rbegin(),
712 E = SymbolTable.rend(); I != E; ++I) {
713 if (I->GVName == "") {
716 I->n_strx = StrT.size();
717 StrTOut.outstring(I->GVName, I->GVName.length()+1);
721 OutputBuffer SymTOut(SymT, is64Bit, isLittleEndian);
724 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
725 E = SymbolTable.end(); I != E; ++I, ++index) {
726 // Add the section base address to the section offset in the n_value field
727 // to calculate the full address.
728 // FIXME: handle symbols where the n_value field is not the address
729 GlobalValue *GV = const_cast<GlobalValue*>(I->GV);
730 if (GV && GVSection[GV])
731 I->n_value += GVSection[GV]->addr;
732 if (GV && (GVOffset[GV] == -1))
733 GVOffset[GV] = index;
735 // Emit nlist to buffer
736 SymTOut.outword(I->n_strx);
737 SymTOut.outbyte(I->n_type);
738 SymTOut.outbyte(I->n_sect);
739 SymTOut.outhalf(I->n_desc);
740 SymTOut.outaddr(I->n_value);
744 /// CalculateRelocations - For each MachineRelocation in the current section,
745 /// calculate the index of the section containing the object to be relocated,
746 /// and the offset into that section. From this information, create the
747 /// appropriate target-specific MachORelocation type and add buffer it to be
748 /// written out after we are finished writing out sections.
749 void MachOWriter::CalculateRelocations(MachOSection &MOS) {
750 for (unsigned i = 0, e = MOS.Relocations.size(); i != e; ++i) {
751 MachineRelocation &MR = MOS.Relocations[i];
752 unsigned TargetSection = MR.getConstantVal();
753 unsigned TargetAddr = 0;
754 unsigned TargetIndex = 0;
756 // This is a scattered relocation entry if it points to a global value with
757 // a non-zero offset.
758 bool Scattered = false;
761 // Since we may not have seen the GlobalValue we were interested in yet at
762 // the time we emitted the relocation for it, fix it up now so that it
763 // points to the offset into the correct section.
764 if (MR.isGlobalValue()) {
765 GlobalValue *GV = MR.getGlobalValue();
766 MachOSection *MOSPtr = GVSection[GV];
767 intptr_t Offset = GVOffset[GV];
769 // If we have never seen the global before, it must be to a symbol
770 // defined in another module (N_UNDF).
772 // FIXME: need to append stub suffix
775 TargetIndex = GVOffset[GV];
777 Scattered = TargetSection != 0;
778 TargetSection = MOSPtr->Index;
780 MR.setResultPointer((void*)Offset);
783 // If the symbol is locally defined, pass in the address of the section and
784 // the section index to the code which will generate the target relocation.
786 MachOSection &To = *SectionList[TargetSection - 1];
787 TargetAddr = To.addr;
788 TargetIndex = To.Index;
791 OutputBuffer RelocOut(MOS.RelocBuffer, is64Bit, isLittleEndian);
792 OutputBuffer SecOut(MOS.SectionData, is64Bit, isLittleEndian);
794 MOS.nreloc += GetTargetRelocation(MR, MOS.Index, TargetAddr, TargetIndex,
795 RelocOut, SecOut, Scattered, Extern);
799 // InitMem - Write the value of a Constant to the specified memory location,
800 // converting it into bytes and relocations.
801 void MachOWriter::InitMem(const Constant *C, void *Addr, intptr_t Offset,
802 const TargetData *TD,
803 std::vector<MachineRelocation> &MRs) {
804 typedef std::pair<const Constant*, intptr_t> CPair;
805 std::vector<CPair> WorkList;
807 WorkList.push_back(CPair(C,(intptr_t)Addr + Offset));
809 intptr_t ScatteredOffset = 0;
811 while (!WorkList.empty()) {
812 const Constant *PC = WorkList.back().first;
813 intptr_t PA = WorkList.back().second;
816 if (isa<UndefValue>(PC)) {
818 } else if (const ConstantVector *CP = dyn_cast<ConstantVector>(PC)) {
819 unsigned ElementSize =
820 TD->getABITypeSize(CP->getType()->getElementType());
821 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
822 WorkList.push_back(CPair(CP->getOperand(i), PA+i*ElementSize));
823 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(PC)) {
825 // FIXME: Handle ConstantExpression. See EE::getConstantValue()
827 switch (CE->getOpcode()) {
828 case Instruction::GetElementPtr: {
829 SmallVector<Value*, 8> Indices(CE->op_begin()+1, CE->op_end());
830 ScatteredOffset = TD->getIndexedOffset(CE->getOperand(0)->getType(),
831 &Indices[0], Indices.size());
832 WorkList.push_back(CPair(CE->getOperand(0), PA));
835 case Instruction::Add:
837 cerr << "ConstantExpr not handled as global var init: " << *CE << "\n";
841 } else if (PC->getType()->isFirstClassType()) {
842 unsigned char *ptr = (unsigned char *)PA;
843 switch (PC->getType()->getTypeID()) {
844 case Type::IntegerTyID: {
845 unsigned NumBits = cast<IntegerType>(PC->getType())->getBitWidth();
846 uint64_t val = cast<ConstantInt>(PC)->getZExtValue();
849 else if (NumBits <= 16) {
850 if (TD->isBigEndian())
851 val = ByteSwap_16(val);
854 } else if (NumBits <= 32) {
855 if (TD->isBigEndian())
856 val = ByteSwap_32(val);
861 } else if (NumBits <= 64) {
862 if (TD->isBigEndian())
863 val = ByteSwap_64(val);
873 assert(0 && "Not implemented: bit widths > 64");
877 case Type::FloatTyID: {
878 uint32_t val = cast<ConstantFP>(PC)->getValueAPF().convertToAPInt().
880 if (TD->isBigEndian())
881 val = ByteSwap_32(val);
888 case Type::DoubleTyID: {
889 uint64_t val = cast<ConstantFP>(PC)->getValueAPF().convertToAPInt().
891 if (TD->isBigEndian())
892 val = ByteSwap_64(val);
903 case Type::PointerTyID:
904 if (isa<ConstantPointerNull>(PC))
905 memset(ptr, 0, TD->getPointerSize());
906 else if (const GlobalValue* GV = dyn_cast<GlobalValue>(PC)) {
907 // FIXME: what about function stubs?
908 MRs.push_back(MachineRelocation::getGV(PA-(intptr_t)Addr,
909 MachineRelocation::VANILLA,
910 const_cast<GlobalValue*>(GV),
914 assert(0 && "Unknown constant pointer type!");
917 cerr << "ERROR: Constant unimp for type: " << *PC->getType() << "\n";
920 } else if (isa<ConstantAggregateZero>(PC)) {
921 memset((void*)PA, 0, (size_t)TD->getABITypeSize(PC->getType()));
922 } else if (const ConstantArray *CPA = dyn_cast<ConstantArray>(PC)) {
923 unsigned ElementSize =
924 TD->getABITypeSize(CPA->getType()->getElementType());
925 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i)
926 WorkList.push_back(CPair(CPA->getOperand(i), PA+i*ElementSize));
927 } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(PC)) {
928 const StructLayout *SL =
929 TD->getStructLayout(cast<StructType>(CPS->getType()));
930 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i)
931 WorkList.push_back(CPair(CPS->getOperand(i),
932 PA+SL->getElementOffset(i)));
934 cerr << "Bad Type: " << *PC->getType() << "\n";
935 assert(0 && "Unknown constant type to initialize memory with!");
940 MachOSym::MachOSym(const GlobalValue *gv, std::string name, uint8_t sect,
942 GV(gv), n_strx(0), n_type(sect == NO_SECT ? N_UNDF : N_SECT), n_sect(sect),
943 n_desc(0), n_value(0) {
945 const TargetAsmInfo *TAI = TM.getTargetAsmInfo();
947 switch (GV->getLinkage()) {
949 assert(0 && "Unexpected linkage type!");
951 case GlobalValue::WeakLinkage:
952 case GlobalValue::LinkOnceLinkage:
953 assert(!isa<Function>(gv) && "Unexpected linkage type for Function!");
954 case GlobalValue::ExternalLinkage:
955 GVName = TAI->getGlobalPrefix() + name;
956 n_type |= GV->hasHiddenVisibility() ? N_PEXT : N_EXT;
958 case GlobalValue::InternalLinkage:
959 GVName = TAI->getGlobalPrefix() + name;