1 //===-- MachOWriter.cpp - Target-independent Mach-O Writer code -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman 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 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 "llvm/Constants.h"
26 #include "llvm/DerivedTypes.h"
27 #include "llvm/Module.h"
28 #include "llvm/CodeGen/MachineCodeEmitter.h"
29 #include "llvm/CodeGen/MachineConstantPool.h"
30 #include "llvm/CodeGen/MachineJumpTableInfo.h"
31 #include "llvm/CodeGen/MachOWriter.h"
32 #include "llvm/ExecutionEngine/ExecutionEngine.h"
33 #include "llvm/Target/TargetAsmInfo.h"
34 #include "llvm/Target/TargetJITInfo.h"
35 #include "llvm/Support/Mangler.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/OutputBuffer.h"
38 #include "llvm/Support/Streams.h"
43 //===----------------------------------------------------------------------===//
44 // MachOCodeEmitter Implementation
45 //===----------------------------------------------------------------------===//
48 /// MachOCodeEmitter - This class is used by the MachOWriter to emit the code
49 /// for functions to the Mach-O file.
50 class MachOCodeEmitter : public MachineCodeEmitter {
53 /// Target machine description.
56 /// is64Bit/isLittleEndian - This information is inferred from the target
57 /// machine directly, indicating what header values and flags to set.
58 bool is64Bit, isLittleEndian;
60 /// Relocations - These are the relocations that the function needs, as
62 std::vector<MachineRelocation> Relocations;
64 /// CPLocations - This is a map of constant pool indices to offsets from the
65 /// start of the section for that constant pool index.
66 std::vector<intptr_t> CPLocations;
68 /// CPSections - This is a map of constant pool indices to the MachOSection
69 /// containing the constant pool entry for that index.
70 std::vector<unsigned> CPSections;
72 /// JTLocations - This is a map of jump table indices to offsets from the
73 /// start of the section for that jump table index.
74 std::vector<intptr_t> JTLocations;
76 /// MBBLocations - This vector is a mapping from MBB ID's to their address.
77 /// It is filled in by the StartMachineBasicBlock callback and queried by
78 /// the getMachineBasicBlockAddress callback.
79 std::vector<intptr_t> MBBLocations;
82 MachOCodeEmitter(MachOWriter &mow) : MOW(mow), TM(MOW.TM) {
83 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
84 isLittleEndian = TM.getTargetData()->isLittleEndian();
87 virtual void startFunction(MachineFunction &F);
88 virtual bool finishFunction(MachineFunction &F);
90 virtual void addRelocation(const MachineRelocation &MR) {
91 Relocations.push_back(MR);
94 void emitConstantPool(MachineConstantPool *MCP);
95 void emitJumpTables(MachineJumpTableInfo *MJTI);
97 virtual intptr_t getConstantPoolEntryAddress(unsigned Index) const {
98 assert(CPLocations.size() > Index && "CP not emitted!");
99 return CPLocations[Index];
101 virtual intptr_t getJumpTableEntryAddress(unsigned Index) const {
102 assert(JTLocations.size() > Index && "JT not emitted!");
103 return JTLocations[Index];
106 virtual void StartMachineBasicBlock(MachineBasicBlock *MBB) {
107 if (MBBLocations.size() <= (unsigned)MBB->getNumber())
108 MBBLocations.resize((MBB->getNumber()+1)*2);
109 MBBLocations[MBB->getNumber()] = getCurrentPCOffset();
112 virtual intptr_t getMachineBasicBlockAddress(MachineBasicBlock *MBB) const {
113 assert(MBBLocations.size() > (unsigned)MBB->getNumber() &&
114 MBBLocations[MBB->getNumber()] && "MBB not emitted!");
115 return MBBLocations[MBB->getNumber()];
118 /// JIT SPECIFIC FUNCTIONS - DO NOT IMPLEMENT THESE HERE!
119 virtual void startFunctionStub(unsigned StubSize, unsigned Alignment = 1) {
120 assert(0 && "JIT specific function called!");
123 virtual void *finishFunctionStub(const Function *F) {
124 assert(0 && "JIT specific function called!");
131 /// startFunction - This callback is invoked when a new machine function is
132 /// about to be emitted.
133 void MachOCodeEmitter::startFunction(MachineFunction &F) {
134 // Align the output buffer to the appropriate alignment, power of 2.
135 // FIXME: MachineFunction or TargetData should probably carry an alignment
136 // field for functions that we can query here instead of hard coding 4 in both
137 // the object writer and asm printer.
140 // Get the Mach-O Section that this function belongs in.
141 MachOWriter::MachOSection *MOS = MOW.getTextSection();
143 // FIXME: better memory management
144 MOS->SectionData.reserve(4096);
145 BufferBegin = &MOS->SectionData[0];
146 BufferEnd = BufferBegin + MOS->SectionData.capacity();
148 // FIXME: Using MOS->size directly here instead of calculating it from the
149 // output buffer size (impossible because the code emitter deals only in raw
150 // bytes) forces us to manually synchronize size and write padding zero bytes
151 // to the output buffer for all non-text sections. For text sections, we do
152 // not synchonize the output buffer, and we just blow up if anyone tries to
153 // write non-code to it. An assert should probably be added to
154 // AddSymbolToSection to prevent calling it on the text section.
155 CurBufferPtr = BufferBegin + MOS->size;
157 // Upgrade the section alignment if required.
158 if (MOS->align < Align) MOS->align = Align;
160 // Clear per-function data structures.
164 MBBLocations.clear();
167 /// finishFunction - This callback is invoked after the function is completely
169 bool MachOCodeEmitter::finishFunction(MachineFunction &F) {
170 // Get the Mach-O Section that this function belongs in.
171 MachOWriter::MachOSection *MOS = MOW.getTextSection();
173 MOS->size += CurBufferPtr - BufferBegin;
175 // Get a symbol for the function to add to the symbol table
176 const GlobalValue *FuncV = F.getFunction();
177 MachOSym FnSym(FuncV, MOW.Mang->getValueName(FuncV), MOS->Index, TM);
179 // Emit constant pool to appropriate section(s)
180 emitConstantPool(F.getConstantPool());
182 // Emit jump tables to appropriate section
183 emitJumpTables(F.getJumpTableInfo());
185 // If we have emitted any relocations to function-specific objects such as
186 // basic blocks, constant pools entries, or jump tables, record their
187 // addresses now so that we can rewrite them with the correct addresses
189 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
190 MachineRelocation &MR = Relocations[i];
193 if (MR.isBasicBlock()) {
194 Addr = getMachineBasicBlockAddress(MR.getBasicBlock());
195 MR.setConstantVal(MOS->Index);
196 MR.setResultPointer((void*)Addr);
197 } else if (MR.isJumpTableIndex()) {
198 Addr = getJumpTableEntryAddress(MR.getJumpTableIndex());
199 MR.setConstantVal(MOW.getJumpTableSection()->Index);
200 MR.setResultPointer((void*)Addr);
201 } else if (MR.isConstantPoolIndex()) {
202 Addr = getConstantPoolEntryAddress(MR.getConstantPoolIndex());
203 MR.setConstantVal(CPSections[MR.getConstantPoolIndex()]);
204 MR.setResultPointer((void*)Addr);
205 } else if (!MR.isGlobalValue()) {
206 assert(0 && "Unhandled relocation type");
208 MOS->Relocations.push_back(MR);
212 // Finally, add it to the symtab.
213 MOW.SymbolTable.push_back(FnSym);
217 /// emitConstantPool - For each constant pool entry, figure out which section
218 /// the constant should live in, allocate space for it, and emit it to the
219 /// Section data buffer.
220 void MachOCodeEmitter::emitConstantPool(MachineConstantPool *MCP) {
221 const std::vector<MachineConstantPoolEntry> &CP = MCP->getConstants();
222 if (CP.empty()) return;
224 // FIXME: handle PIC codegen
225 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
226 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
228 // Although there is no strict necessity that I am aware of, we will do what
229 // gcc for OS X does and put each constant pool entry in a section of constant
230 // objects of a certain size. That means that float constants go in the
231 // literal4 section, and double objects go in literal8, etc.
233 // FIXME: revisit this decision if we ever do the "stick everything into one
234 // "giant object for PIC" optimization.
235 for (unsigned i = 0, e = CP.size(); i != e; ++i) {
236 const Type *Ty = CP[i].getType();
237 unsigned Size = TM.getTargetData()->getTypeSize(Ty);
239 MachOWriter::MachOSection *Sec = MOW.getConstSection(Ty);
240 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
242 CPLocations.push_back(Sec->SectionData.size());
243 CPSections.push_back(Sec->Index);
245 // FIXME: remove when we have unified size + output buffer
248 // Allocate space in the section for the global.
249 // FIXME: need alignment?
250 // FIXME: share between here and AddSymbolToSection?
251 for (unsigned j = 0; j < Size; ++j)
252 SecDataOut.outbyte(0);
254 MOW.InitMem(CP[i].Val.ConstVal, &Sec->SectionData[0], CPLocations[i],
255 TM.getTargetData(), Sec->Relocations);
259 /// emitJumpTables - Emit all the jump tables for a given jump table info
260 /// record to the appropriate section.
261 void MachOCodeEmitter::emitJumpTables(MachineJumpTableInfo *MJTI) {
262 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
263 if (JT.empty()) return;
265 // FIXME: handle PIC codegen
266 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
267 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
269 MachOWriter::MachOSection *Sec = MOW.getJumpTableSection();
270 unsigned TextSecIndex = MOW.getTextSection()->Index;
271 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
273 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
274 // For each jump table, record its offset from the start of the section,
275 // reserve space for the relocations to the MBBs, and add the relocations.
276 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
277 JTLocations.push_back(Sec->SectionData.size());
278 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi) {
279 MachineRelocation MR(MOW.GetJTRelocation(Sec->SectionData.size(),
281 MR.setResultPointer((void *)JTLocations[i]);
282 MR.setConstantVal(TextSecIndex);
283 Sec->Relocations.push_back(MR);
284 SecDataOut.outaddr(0);
287 // FIXME: remove when we have unified size + output buffer
288 Sec->size = Sec->SectionData.size();
291 //===----------------------------------------------------------------------===//
292 // MachOWriter Implementation
293 //===----------------------------------------------------------------------===//
295 MachOWriter::MachOWriter(std::ostream &o, TargetMachine &tm) : O(o), TM(tm) {
296 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
297 isLittleEndian = TM.getTargetData()->isLittleEndian();
299 // Create the machine code emitter object for this target.
300 MCE = new MachOCodeEmitter(*this);
303 MachOWriter::~MachOWriter() {
307 void MachOWriter::AddSymbolToSection(MachOSection *Sec, GlobalVariable *GV) {
308 const Type *Ty = GV->getType()->getElementType();
309 unsigned Size = TM.getTargetData()->getTypeSize(Ty);
310 unsigned Align = GV->getAlignment();
312 Align = TM.getTargetData()->getTypeAlignmentPref(Ty);
314 MachOSym Sym(GV, Mang->getValueName(GV), Sec->Index, TM);
316 // Reserve space in the .bss section for this symbol while maintaining the
317 // desired section alignment, which must be at least as much as required by
319 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
322 uint64_t OrigSize = Sec->size;
323 Align = Log2_32(Align);
324 Sec->align = std::max(unsigned(Sec->align), Align);
325 Sec->size = (Sec->size + Align - 1) & ~(Align-1);
327 // Add alignment padding to buffer as well.
328 // FIXME: remove when we have unified size + output buffer
329 unsigned AlignedSize = Sec->size - OrigSize;
330 for (unsigned i = 0; i < AlignedSize; ++i)
331 SecDataOut.outbyte(0);
333 // Record the offset of the symbol, and then allocate space for it.
334 // FIXME: remove when we have unified size + output buffer
335 Sym.n_value = Sec->size;
337 SymbolTable.push_back(Sym);
339 // Now that we know what section the GlovalVariable is going to be emitted
340 // into, update our mappings.
341 // FIXME: We may also need to update this when outputting non-GlobalVariable
342 // GlobalValues such as functions.
344 GVOffset[GV] = Sec->SectionData.size();
346 // Allocate space in the section for the global.
347 for (unsigned i = 0; i < Size; ++i)
348 SecDataOut.outbyte(0);
351 void MachOWriter::EmitGlobal(GlobalVariable *GV) {
352 const Type *Ty = GV->getType()->getElementType();
353 unsigned Size = TM.getTargetData()->getTypeSize(Ty);
354 bool NoInit = !GV->hasInitializer();
356 // If this global has a zero initializer, it is part of the .bss or common
358 if (NoInit || GV->getInitializer()->isNullValue()) {
359 // If this global is part of the common block, add it now. Variables are
360 // part of the common block if they are zero initialized and allowed to be
361 // merged with other symbols.
362 if (NoInit || GV->hasLinkOnceLinkage() || GV->hasWeakLinkage()) {
363 MachOSym ExtOrCommonSym(GV, Mang->getValueName(GV), MachOSym::NO_SECT,TM);
364 // For undefined (N_UNDF) external (N_EXT) types, n_value is the size in
365 // bytes of the symbol.
366 ExtOrCommonSym.n_value = Size;
367 // If the symbol is external, we'll put it on a list of symbols whose
368 // addition to the symbol table is being pended until we find a reference
370 PendingSyms.push_back(ExtOrCommonSym);
372 SymbolTable.push_back(ExtOrCommonSym);
375 // Otherwise, this symbol is part of the .bss section.
376 MachOSection *BSS = getBSSSection();
377 AddSymbolToSection(BSS, GV);
381 // Scalar read-only data goes in a literal section if the scalar is 4, 8, or
382 // 16 bytes, or a cstring. Other read only data goes into a regular const
383 // section. Read-write data goes in the data section.
384 MachOSection *Sec = GV->isConstant() ? getConstSection(Ty) : getDataSection();
385 AddSymbolToSection(Sec, GV);
386 InitMem(GV->getInitializer(), &Sec->SectionData[0], GVOffset[GV],
387 TM.getTargetData(), Sec->Relocations);
391 bool MachOWriter::runOnMachineFunction(MachineFunction &MF) {
392 // Nothing to do here, this is all done through the MCE object.
396 bool MachOWriter::doInitialization(Module &M) {
397 // Set the magic value, now that we know the pointer size and endianness
398 Header.setMagic(isLittleEndian, is64Bit);
401 // FIXME: this only works for object files, we do not support the creation
402 // of dynamic libraries or executables at this time.
403 Header.filetype = MachOHeader::MH_OBJECT;
405 Mang = new Mangler(M);
409 /// doFinalization - Now that the module has been completely processed, emit
410 /// the Mach-O file to 'O'.
411 bool MachOWriter::doFinalization(Module &M) {
412 // FIXME: we don't handle debug info yet, we should probably do that.
414 // Okay, the.text section has been completed, build the .data, .bss, and
415 // "common" sections next.
416 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
420 // Emit the header and load commands.
421 EmitHeaderAndLoadCommands();
423 // Emit the various sections and their relocation info.
426 // Write the symbol table and the string table to the end of the file.
427 O.write((char*)&SymT[0], SymT.size());
428 O.write((char*)&StrT[0], StrT.size());
430 // We are done with the abstract symbols.
433 DynamicSymbolTable.clear();
435 // Release the name mangler object.
436 delete Mang; Mang = 0;
440 void MachOWriter::EmitHeaderAndLoadCommands() {
441 // Step #0: Fill in the segment load command size, since we need it to figure
442 // out the rest of the header fields
443 MachOSegment SEG("", is64Bit);
444 SEG.nsects = SectionList.size();
445 SEG.cmdsize = SEG.cmdSize(is64Bit) +
446 SEG.nsects * SectionList[0]->cmdSize(is64Bit);
448 // Step #1: calculate the number of load commands. We always have at least
449 // one, for the LC_SEGMENT load command, plus two for the normal
450 // and dynamic symbol tables, if there are any symbols.
451 Header.ncmds = SymbolTable.empty() ? 1 : 3;
453 // Step #2: calculate the size of the load commands
454 Header.sizeofcmds = SEG.cmdsize;
455 if (!SymbolTable.empty())
456 Header.sizeofcmds += SymTab.cmdsize + DySymTab.cmdsize;
458 // Step #3: write the header to the file
459 // Local alias to shortenify coming code.
460 DataBuffer &FH = Header.HeaderData;
461 OutputBuffer FHOut(FH, is64Bit, isLittleEndian);
463 FHOut.outword(Header.magic);
464 FHOut.outword(TM.getMachOWriterInfo()->getCPUType());
465 FHOut.outword(TM.getMachOWriterInfo()->getCPUSubType());
466 FHOut.outword(Header.filetype);
467 FHOut.outword(Header.ncmds);
468 FHOut.outword(Header.sizeofcmds);
469 FHOut.outword(Header.flags);
471 FHOut.outword(Header.reserved);
473 // Step #4: Finish filling in the segment load command and write it out
474 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
475 E = SectionList.end(); I != E; ++I)
476 SEG.filesize += (*I)->size;
478 SEG.vmsize = SEG.filesize;
479 SEG.fileoff = Header.cmdSize(is64Bit) + Header.sizeofcmds;
481 FHOut.outword(SEG.cmd);
482 FHOut.outword(SEG.cmdsize);
483 FHOut.outstring(SEG.segname, 16);
484 FHOut.outaddr(SEG.vmaddr);
485 FHOut.outaddr(SEG.vmsize);
486 FHOut.outaddr(SEG.fileoff);
487 FHOut.outaddr(SEG.filesize);
488 FHOut.outword(SEG.maxprot);
489 FHOut.outword(SEG.initprot);
490 FHOut.outword(SEG.nsects);
491 FHOut.outword(SEG.flags);
493 // Step #5: Finish filling in the fields of the MachOSections
494 uint64_t currentAddr = 0;
495 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
496 E = SectionList.end(); I != E; ++I) {
497 MachOSection *MOS = *I;
498 MOS->addr = currentAddr;
499 MOS->offset = currentAddr + SEG.fileoff;
501 // FIXME: do we need to do something with alignment here?
502 currentAddr += MOS->size;
505 // Step #6: Calculate the number of relocations for each section and write out
506 // the section commands for each section
507 currentAddr += SEG.fileoff;
508 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
509 E = SectionList.end(); I != E; ++I) {
510 MachOSection *MOS = *I;
511 // Convert the relocations to target-specific relocations, and fill in the
512 // relocation offset for this section.
513 CalculateRelocations(*MOS);
514 MOS->reloff = MOS->nreloc ? currentAddr : 0;
515 currentAddr += MOS->nreloc * 8;
517 // write the finalized section command to the output buffer
518 FHOut.outstring(MOS->sectname, 16);
519 FHOut.outstring(MOS->segname, 16);
520 FHOut.outaddr(MOS->addr);
521 FHOut.outaddr(MOS->size);
522 FHOut.outword(MOS->offset);
523 FHOut.outword(MOS->align);
524 FHOut.outword(MOS->reloff);
525 FHOut.outword(MOS->nreloc);
526 FHOut.outword(MOS->flags);
527 FHOut.outword(MOS->reserved1);
528 FHOut.outword(MOS->reserved2);
530 FHOut.outword(MOS->reserved3);
533 // Step #7: Emit the symbol table to temporary buffers, so that we know the
534 // size of the string table when we write the next load command.
535 BufferSymbolAndStringTable();
537 // Step #8: Emit LC_SYMTAB/LC_DYSYMTAB load commands
538 SymTab.symoff = currentAddr;
539 SymTab.nsyms = SymbolTable.size();
540 SymTab.stroff = SymTab.symoff + SymT.size();
541 SymTab.strsize = StrT.size();
542 FHOut.outword(SymTab.cmd);
543 FHOut.outword(SymTab.cmdsize);
544 FHOut.outword(SymTab.symoff);
545 FHOut.outword(SymTab.nsyms);
546 FHOut.outword(SymTab.stroff);
547 FHOut.outword(SymTab.strsize);
549 // FIXME: set DySymTab fields appropriately
550 // We should probably just update these in BufferSymbolAndStringTable since
551 // thats where we're partitioning up the different kinds of symbols.
552 FHOut.outword(DySymTab.cmd);
553 FHOut.outword(DySymTab.cmdsize);
554 FHOut.outword(DySymTab.ilocalsym);
555 FHOut.outword(DySymTab.nlocalsym);
556 FHOut.outword(DySymTab.iextdefsym);
557 FHOut.outword(DySymTab.nextdefsym);
558 FHOut.outword(DySymTab.iundefsym);
559 FHOut.outword(DySymTab.nundefsym);
560 FHOut.outword(DySymTab.tocoff);
561 FHOut.outword(DySymTab.ntoc);
562 FHOut.outword(DySymTab.modtaboff);
563 FHOut.outword(DySymTab.nmodtab);
564 FHOut.outword(DySymTab.extrefsymoff);
565 FHOut.outword(DySymTab.nextrefsyms);
566 FHOut.outword(DySymTab.indirectsymoff);
567 FHOut.outword(DySymTab.nindirectsyms);
568 FHOut.outword(DySymTab.extreloff);
569 FHOut.outword(DySymTab.nextrel);
570 FHOut.outword(DySymTab.locreloff);
571 FHOut.outword(DySymTab.nlocrel);
573 O.write((char*)&FH[0], FH.size());
576 /// EmitSections - Now that we have constructed the file header and load
577 /// commands, emit the data for each section to the file.
578 void MachOWriter::EmitSections() {
579 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
580 E = SectionList.end(); I != E; ++I)
581 // Emit the contents of each section
582 O.write((char*)&(*I)->SectionData[0], (*I)->size);
583 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
584 E = SectionList.end(); I != E; ++I)
585 // Emit the relocation entry data for each section.
586 O.write((char*)&(*I)->RelocBuffer[0], (*I)->RelocBuffer.size());
589 /// PartitionByLocal - Simple boolean predicate that returns true if Sym is
590 /// a local symbol rather than an external symbol.
591 bool MachOWriter::PartitionByLocal(const MachOSym &Sym) {
592 return (Sym.n_type & (MachOSym::N_EXT | MachOSym::N_PEXT)) == 0;
595 /// PartitionByDefined - Simple boolean predicate that returns true if Sym is
596 /// defined in this module.
597 bool MachOWriter::PartitionByDefined(const MachOSym &Sym) {
598 // FIXME: Do N_ABS or N_INDR count as defined?
599 return (Sym.n_type & MachOSym::N_SECT) == MachOSym::N_SECT;
602 /// BufferSymbolAndStringTable - Sort the symbols we encountered and assign them
603 /// each a string table index so that they appear in the correct order in the
605 void MachOWriter::BufferSymbolAndStringTable() {
606 // The order of the symbol table is:
608 // 2. defined external symbols (sorted by name)
609 // 3. undefined external symbols (sorted by name)
611 // Sort the symbols by name, so that when we partition the symbols by scope
612 // of definition, we won't have to sort by name within each partition.
613 std::sort(SymbolTable.begin(), SymbolTable.end(), MachOSymCmp());
615 // Parition the symbol table entries so that all local symbols come before
616 // all symbols with external linkage. { 1 | 2 3 }
617 std::partition(SymbolTable.begin(), SymbolTable.end(), PartitionByLocal);
619 // Advance iterator to beginning of external symbols and partition so that
620 // all external symbols defined in this module come before all external
621 // symbols defined elsewhere. { 1 | 2 | 3 }
622 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
623 E = SymbolTable.end(); I != E; ++I) {
624 if (!PartitionByLocal(*I)) {
625 std::partition(I, E, PartitionByDefined);
630 // Calculate the starting index for each of the local, extern defined, and
631 // undefined symbols, as well as the number of each to put in the LC_DYSYMTAB
633 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
634 E = SymbolTable.end(); I != E; ++I) {
635 if (PartitionByLocal(*I)) {
636 ++DySymTab.nlocalsym;
637 ++DySymTab.iextdefsym;
638 } else if (PartitionByDefined(*I)) {
639 ++DySymTab.nextdefsym;
640 ++DySymTab.iundefsym;
642 ++DySymTab.nundefsym;
646 // Write out a leading zero byte when emitting string table, for n_strx == 0
647 // which means an empty string.
648 OutputBuffer StrTOut(StrT, is64Bit, isLittleEndian);
651 // The order of the string table is:
652 // 1. strings for external symbols
653 // 2. strings for local symbols
654 // Since this is the opposite order from the symbol table, which we have just
655 // sorted, we can walk the symbol table backwards to output the string table.
656 for (std::vector<MachOSym>::reverse_iterator I = SymbolTable.rbegin(),
657 E = SymbolTable.rend(); I != E; ++I) {
658 if (I->GVName == "") {
661 I->n_strx = StrT.size();
662 StrTOut.outstring(I->GVName, I->GVName.length()+1);
666 OutputBuffer SymTOut(SymT, is64Bit, isLittleEndian);
668 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
669 E = SymbolTable.end(); I != E; ++I) {
670 // Add the section base address to the section offset in the n_value field
671 // to calculate the full address.
672 // FIXME: handle symbols where the n_value field is not the address
673 GlobalValue *GV = const_cast<GlobalValue*>(I->GV);
674 if (GV && GVSection[GV])
675 I->n_value += GVSection[GV]->addr;
677 // Emit nlist to buffer
678 SymTOut.outword(I->n_strx);
679 SymTOut.outbyte(I->n_type);
680 SymTOut.outbyte(I->n_sect);
681 SymTOut.outhalf(I->n_desc);
682 SymTOut.outaddr(I->n_value);
686 /// CalculateRelocations - For each MachineRelocation in the current section,
687 /// calculate the index of the section containing the object to be relocated,
688 /// and the offset into that section. From this information, create the
689 /// appropriate target-specific MachORelocation type and add buffer it to be
690 /// written out after we are finished writing out sections.
691 void MachOWriter::CalculateRelocations(MachOSection &MOS) {
692 for (unsigned i = 0, e = MOS.Relocations.size(); i != e; ++i) {
693 MachineRelocation &MR = MOS.Relocations[i];
694 unsigned TargetSection = MR.getConstantVal();
696 // Since we may not have seen the GlobalValue we were interested in yet at
697 // the time we emitted the relocation for it, fix it up now so that it
698 // points to the offset into the correct section.
699 if (MR.isGlobalValue()) {
700 GlobalValue *GV = MR.getGlobalValue();
701 MachOSection *MOSPtr = GVSection[GV];
702 intptr_t offset = GVOffset[GV];
704 assert(MOSPtr && "Trying to relocate unknown global!");
706 TargetSection = MOSPtr->Index;
707 MR.setResultPointer((void*)offset);
710 GetTargetRelocation(MR, MOS, *SectionList[TargetSection-1]);
714 // InitMem - Write the value of a Constant to the specified memory location,
715 // converting it into bytes and relocations.
716 void MachOWriter::InitMem(const Constant *C, void *Addr, intptr_t Offset,
717 const TargetData *TD,
718 std::vector<MachineRelocation> &MRs) {
719 typedef std::pair<const Constant*, intptr_t> CPair;
720 std::vector<CPair> WorkList;
722 WorkList.push_back(CPair(C,(intptr_t)Addr + Offset));
724 while (!WorkList.empty()) {
725 const Constant *PC = WorkList.back().first;
726 intptr_t PA = WorkList.back().second;
729 if (isa<UndefValue>(PC)) {
731 } else if (const ConstantPacked *CP = dyn_cast<ConstantPacked>(PC)) {
732 unsigned ElementSize = TD->getTypeSize(CP->getType()->getElementType());
733 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
734 WorkList.push_back(CPair(CP->getOperand(i), PA+i*ElementSize));
735 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(PC)) {
737 // FIXME: Handle ConstantExpression. See EE::getConstantValue()
739 switch (CE->getOpcode()) {
740 case Instruction::GetElementPtr:
741 case Instruction::Add:
743 cerr << "ConstantExpr not handled as global var init: " << *CE << "\n";
747 } else if (PC->getType()->isFirstClassType()) {
748 unsigned char *ptr = (unsigned char *)PA;
749 switch (PC->getType()->getTypeID()) {
750 case Type::IntegerTyID: {
751 unsigned NumBits = cast<IntegerType>(PC->getType())->getBitWidth();
752 uint64_t val = cast<ConstantInt>(PC)->getZExtValue();
755 else if (NumBits <= 16) {
756 if (TD->isBigEndian())
757 val = ByteSwap_16(val);
760 } else if (NumBits <= 32) {
761 if (TD->isBigEndian())
762 val = ByteSwap_32(val);
767 } else if (NumBits <= 64) {
768 if (TD->isBigEndian())
769 val = ByteSwap_64(val);
779 assert(0 && "Not implemented: bit widths > 64");
783 case Type::FloatTyID: {
784 uint64_t val = FloatToBits(cast<ConstantFP>(PC)->getValue());
785 if (TD->isBigEndian())
786 val = ByteSwap_32(val);
793 case Type::DoubleTyID: {
794 uint64_t val = DoubleToBits(cast<ConstantFP>(PC)->getValue());
795 if (TD->isBigEndian())
796 val = ByteSwap_64(val);
807 case Type::PointerTyID:
808 if (isa<ConstantPointerNull>(C))
809 memset(ptr, 0, TD->getPointerSize());
810 else if (const GlobalValue* GV = dyn_cast<GlobalValue>(C))
811 // FIXME: what about function stubs?
812 MRs.push_back(MachineRelocation::getGV(PA-(intptr_t)Addr,
813 MachineRelocation::VANILLA,
814 const_cast<GlobalValue*>(GV)));
816 assert(0 && "Unknown constant pointer type!");
819 cerr << "ERROR: Constant unimp for type: " << *PC->getType() << "\n";
822 } else if (isa<ConstantAggregateZero>(PC)) {
823 memset((void*)PA, 0, (size_t)TD->getTypeSize(PC->getType()));
824 } else if (const ConstantArray *CPA = dyn_cast<ConstantArray>(PC)) {
825 unsigned ElementSize = TD->getTypeSize(CPA->getType()->getElementType());
826 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i)
827 WorkList.push_back(CPair(CPA->getOperand(i), PA+i*ElementSize));
828 } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(PC)) {
829 const StructLayout *SL =
830 TD->getStructLayout(cast<StructType>(CPS->getType()));
831 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i)
832 WorkList.push_back(CPair(CPS->getOperand(i), PA+SL->MemberOffsets[i]));
834 cerr << "Bad Type: " << *PC->getType() << "\n";
835 assert(0 && "Unknown constant type to initialize memory with!");
840 MachOSym::MachOSym(const GlobalValue *gv, std::string name, uint8_t sect,
842 GV(gv), n_strx(0), n_type(sect == NO_SECT ? N_UNDF : N_SECT), n_sect(sect),
843 n_desc(0), n_value(0) {
845 const TargetAsmInfo *TAI = TM.getTargetAsmInfo();
847 switch (GV->getLinkage()) {
849 assert(0 && "Unexpected linkage type!");
851 case GlobalValue::WeakLinkage:
852 case GlobalValue::LinkOnceLinkage:
853 assert(!isa<Function>(gv) && "Unexpected linkage type for Function!");
854 case GlobalValue::ExternalLinkage:
855 GVName = TAI->getGlobalPrefix() + name;
858 case GlobalValue::InternalLinkage:
859 GVName = TAI->getPrivateGlobalPrefix() + name;