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"
43 /// AddMachOWriter - Concrete function to add the Mach-O writer to the function
45 MachineCodeEmitter *llvm::AddMachOWriter(FunctionPassManager &FPM,
48 MachOWriter *MOW = new MachOWriter(O, TM);
50 return &MOW->getMachineCodeEmitter();
53 //===----------------------------------------------------------------------===//
54 // MachOCodeEmitter Implementation
55 //===----------------------------------------------------------------------===//
58 /// MachOCodeEmitter - This class is used by the MachOWriter to emit the code
59 /// for functions to the Mach-O file.
60 class MachOCodeEmitter : public MachineCodeEmitter {
63 /// Target machine description.
66 /// is64Bit/isLittleEndian - This information is inferred from the target
67 /// machine directly, indicating what header values and flags to set.
68 bool is64Bit, isLittleEndian;
70 /// Relocations - These are the relocations that the function needs, as
72 std::vector<MachineRelocation> Relocations;
74 /// CPLocations - This is a map of constant pool indices to offsets from the
75 /// start of the section for that constant pool index.
76 std::vector<intptr_t> CPLocations;
78 /// CPSections - This is a map of constant pool indices to the MachOSection
79 /// containing the constant pool entry for that index.
80 std::vector<unsigned> CPSections;
82 /// JTLocations - This is a map of jump table indices to offsets from the
83 /// start of the section for that jump table index.
84 std::vector<intptr_t> JTLocations;
86 /// MBBLocations - This vector is a mapping from MBB ID's to their address.
87 /// It is filled in by the StartMachineBasicBlock callback and queried by
88 /// the getMachineBasicBlockAddress callback.
89 std::vector<intptr_t> MBBLocations;
92 MachOCodeEmitter(MachOWriter &mow) : MOW(mow), TM(MOW.TM) {
93 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
94 isLittleEndian = TM.getTargetData()->isLittleEndian();
97 virtual void startFunction(MachineFunction &MF);
98 virtual bool finishFunction(MachineFunction &MF);
100 virtual void addRelocation(const MachineRelocation &MR) {
101 Relocations.push_back(MR);
104 void emitConstantPool(MachineConstantPool *MCP);
105 void emitJumpTables(MachineJumpTableInfo *MJTI);
107 virtual intptr_t getConstantPoolEntryAddress(unsigned Index) const {
108 assert(CPLocations.size() > Index && "CP not emitted!");
109 return CPLocations[Index];
111 virtual intptr_t getJumpTableEntryAddress(unsigned Index) const {
112 assert(JTLocations.size() > Index && "JT not emitted!");
113 return JTLocations[Index];
116 virtual void StartMachineBasicBlock(MachineBasicBlock *MBB) {
117 if (MBBLocations.size() <= (unsigned)MBB->getNumber())
118 MBBLocations.resize((MBB->getNumber()+1)*2);
119 MBBLocations[MBB->getNumber()] = getCurrentPCOffset();
122 virtual intptr_t getMachineBasicBlockAddress(MachineBasicBlock *MBB) const {
123 assert(MBBLocations.size() > (unsigned)MBB->getNumber() &&
124 MBBLocations[MBB->getNumber()] && "MBB not emitted!");
125 return MBBLocations[MBB->getNumber()];
128 virtual intptr_t getLabelAddress(uint64_t Label) const {
129 assert(0 && "get Label not implemented");
134 virtual void emitLabel(uint64_t LabelID) {
135 assert(0 && "emit Label not implemented");
140 virtual void setModuleInfo(llvm::MachineModuleInfo* MMI) { }
142 /// JIT SPECIFIC FUNCTIONS - DO NOT IMPLEMENT THESE HERE!
143 virtual void startFunctionStub(unsigned StubSize, unsigned Alignment = 1) {
144 assert(0 && "JIT specific function called!");
147 virtual void *finishFunctionStub(const Function *F) {
148 assert(0 && "JIT specific function called!");
155 /// startFunction - This callback is invoked when a new machine function is
156 /// about to be emitted.
157 void MachOCodeEmitter::startFunction(MachineFunction &MF) {
158 const TargetData *TD = TM.getTargetData();
159 const Function *F = MF.getFunction();
161 // Align the output buffer to the appropriate alignment, power of 2.
162 unsigned FnAlign = F->getAlignment();
163 unsigned TDAlign = TD->getPrefTypeAlignment(F->getType());
164 unsigned Align = Log2_32(std::max(FnAlign, TDAlign));
165 assert(!(Align & (Align-1)) && "Alignment is not a power of two!");
167 // Get the Mach-O Section that this function belongs in.
168 MachOWriter::MachOSection *MOS = MOW.getTextSection();
170 // FIXME: better memory management
171 MOS->SectionData.reserve(4096);
172 BufferBegin = &MOS->SectionData[0];
173 BufferEnd = BufferBegin + MOS->SectionData.capacity();
175 // Upgrade the section alignment if required.
176 if (MOS->align < Align) MOS->align = Align;
178 // Round the size up to the correct alignment for starting the new function.
179 if ((MOS->size & ((1 << Align) - 1)) != 0) {
180 MOS->size += (1 << Align);
181 MOS->size &= ~((1 << Align) - 1);
184 // FIXME: Using MOS->size directly here instead of calculating it from the
185 // output buffer size (impossible because the code emitter deals only in raw
186 // bytes) forces us to manually synchronize size and write padding zero bytes
187 // to the output buffer for all non-text sections. For text sections, we do
188 // not synchonize the output buffer, and we just blow up if anyone tries to
189 // write non-code to it. An assert should probably be added to
190 // AddSymbolToSection to prevent calling it on the text section.
191 CurBufferPtr = BufferBegin + MOS->size;
193 // Clear per-function data structures.
197 MBBLocations.clear();
200 /// finishFunction - This callback is invoked after the function is completely
202 bool MachOCodeEmitter::finishFunction(MachineFunction &MF) {
203 // Get the Mach-O Section that this function belongs in.
204 MachOWriter::MachOSection *MOS = MOW.getTextSection();
206 // Get a symbol for the function to add to the symbol table
207 // FIXME: it seems like we should call something like AddSymbolToSection
208 // in startFunction rather than changing the section size and symbol n_value
210 const GlobalValue *FuncV = MF.getFunction();
211 MachOSym FnSym(FuncV, MOW.Mang->getValueName(FuncV), MOS->Index, TM);
212 FnSym.n_value = MOS->size;
213 MOS->size = CurBufferPtr - BufferBegin;
215 // Emit constant pool to appropriate section(s)
216 emitConstantPool(MF.getConstantPool());
218 // Emit jump tables to appropriate section
219 emitJumpTables(MF.getJumpTableInfo());
221 // If we have emitted any relocations to function-specific objects such as
222 // basic blocks, constant pools entries, or jump tables, record their
223 // addresses now so that we can rewrite them with the correct addresses
225 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
226 MachineRelocation &MR = Relocations[i];
229 if (MR.isBasicBlock()) {
230 Addr = getMachineBasicBlockAddress(MR.getBasicBlock());
231 MR.setConstantVal(MOS->Index);
232 MR.setResultPointer((void*)Addr);
233 } else if (MR.isJumpTableIndex()) {
234 Addr = getJumpTableEntryAddress(MR.getJumpTableIndex());
235 MR.setConstantVal(MOW.getJumpTableSection()->Index);
236 MR.setResultPointer((void*)Addr);
237 } else if (MR.isConstantPoolIndex()) {
238 Addr = getConstantPoolEntryAddress(MR.getConstantPoolIndex());
239 MR.setConstantVal(CPSections[MR.getConstantPoolIndex()]);
240 MR.setResultPointer((void*)Addr);
241 } else if (MR.isGlobalValue()) {
242 // FIXME: This should be a set or something that uniques
243 MOW.PendingGlobals.push_back(MR.getGlobalValue());
245 assert(0 && "Unhandled relocation type");
247 MOS->Relocations.push_back(MR);
251 // Finally, add it to the symtab.
252 MOW.SymbolTable.push_back(FnSym);
256 /// emitConstantPool - For each constant pool entry, figure out which section
257 /// the constant should live in, allocate space for it, and emit it to the
258 /// Section data buffer.
259 void MachOCodeEmitter::emitConstantPool(MachineConstantPool *MCP) {
260 const std::vector<MachineConstantPoolEntry> &CP = MCP->getConstants();
261 if (CP.empty()) return;
263 // FIXME: handle PIC codegen
264 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
265 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
267 // Although there is no strict necessity that I am aware of, we will do what
268 // gcc for OS X does and put each constant pool entry in a section of constant
269 // objects of a certain size. That means that float constants go in the
270 // literal4 section, and double objects go in literal8, etc.
272 // FIXME: revisit this decision if we ever do the "stick everything into one
273 // "giant object for PIC" optimization.
274 for (unsigned i = 0, e = CP.size(); i != e; ++i) {
275 const Type *Ty = CP[i].getType();
276 unsigned Size = TM.getTargetData()->getABITypeSize(Ty);
278 MachOWriter::MachOSection *Sec = MOW.getConstSection(CP[i].Val.ConstVal);
279 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
281 CPLocations.push_back(Sec->SectionData.size());
282 CPSections.push_back(Sec->Index);
284 // FIXME: remove when we have unified size + output buffer
287 // Allocate space in the section for the global.
288 // FIXME: need alignment?
289 // FIXME: share between here and AddSymbolToSection?
290 for (unsigned j = 0; j < Size; ++j)
291 SecDataOut.outbyte(0);
293 MOW.InitMem(CP[i].Val.ConstVal, &Sec->SectionData[0], CPLocations[i],
294 TM.getTargetData(), Sec->Relocations);
298 /// emitJumpTables - Emit all the jump tables for a given jump table info
299 /// record to the appropriate section.
300 void MachOCodeEmitter::emitJumpTables(MachineJumpTableInfo *MJTI) {
301 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
302 if (JT.empty()) return;
304 // FIXME: handle PIC codegen
305 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
306 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
308 MachOWriter::MachOSection *Sec = MOW.getJumpTableSection();
309 unsigned TextSecIndex = MOW.getTextSection()->Index;
310 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
312 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
313 // For each jump table, record its offset from the start of the section,
314 // reserve space for the relocations to the MBBs, and add the relocations.
315 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
316 JTLocations.push_back(Sec->SectionData.size());
317 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi) {
318 MachineRelocation MR(MOW.GetJTRelocation(Sec->SectionData.size(),
320 MR.setResultPointer((void *)JTLocations[i]);
321 MR.setConstantVal(TextSecIndex);
322 Sec->Relocations.push_back(MR);
323 SecDataOut.outaddr(0);
326 // FIXME: remove when we have unified size + output buffer
327 Sec->size = Sec->SectionData.size();
330 //===----------------------------------------------------------------------===//
331 // MachOWriter Implementation
332 //===----------------------------------------------------------------------===//
334 char MachOWriter::ID = 0;
335 MachOWriter::MachOWriter(std::ostream &o, TargetMachine &tm)
336 : MachineFunctionPass((intptr_t)&ID), O(o), TM(tm) {
337 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
338 isLittleEndian = TM.getTargetData()->isLittleEndian();
340 // Create the machine code emitter object for this target.
341 MCE = new MachOCodeEmitter(*this);
344 MachOWriter::~MachOWriter() {
348 void MachOWriter::AddSymbolToSection(MachOSection *Sec, GlobalVariable *GV) {
349 const Type *Ty = GV->getType()->getElementType();
350 unsigned Size = TM.getTargetData()->getABITypeSize(Ty);
351 unsigned Align = TM.getTargetData()->getPreferredAlignment(GV);
353 // Reserve space in the .bss section for this symbol while maintaining the
354 // desired section alignment, which must be at least as much as required by
356 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
359 uint64_t OrigSize = Sec->size;
360 Align = Log2_32(Align);
361 Sec->align = std::max(unsigned(Sec->align), Align);
362 Sec->size = (Sec->size + Align - 1) & ~(Align-1);
364 // Add alignment padding to buffer as well.
365 // FIXME: remove when we have unified size + output buffer
366 unsigned AlignedSize = Sec->size - OrigSize;
367 for (unsigned i = 0; i < AlignedSize; ++i)
368 SecDataOut.outbyte(0);
370 // Globals without external linkage apparently do not go in the symbol table.
371 if (GV->getLinkage() != GlobalValue::InternalLinkage) {
372 MachOSym Sym(GV, Mang->getValueName(GV), Sec->Index, TM);
373 Sym.n_value = Sec->size;
374 SymbolTable.push_back(Sym);
377 // Record the offset of the symbol, and then allocate space for it.
378 // FIXME: remove when we have unified size + output buffer
381 // Now that we know what section the GlovalVariable is going to be emitted
382 // into, update our mappings.
383 // FIXME: We may also need to update this when outputting non-GlobalVariable
384 // GlobalValues such as functions.
386 GVOffset[GV] = Sec->SectionData.size();
388 // Allocate space in the section for the global.
389 for (unsigned i = 0; i < Size; ++i)
390 SecDataOut.outbyte(0);
393 void MachOWriter::EmitGlobal(GlobalVariable *GV) {
394 const Type *Ty = GV->getType()->getElementType();
395 unsigned Size = TM.getTargetData()->getABITypeSize(Ty);
396 bool NoInit = !GV->hasInitializer();
398 // If this global has a zero initializer, it is part of the .bss or common
400 if (NoInit || GV->getInitializer()->isNullValue()) {
401 // If this global is part of the common block, add it now. Variables are
402 // part of the common block if they are zero initialized and allowed to be
403 // merged with other symbols.
404 if (NoInit || GV->hasLinkOnceLinkage() || GV->hasWeakLinkage()) {
405 MachOSym ExtOrCommonSym(GV, Mang->getValueName(GV), MachOSym::NO_SECT,TM);
406 // For undefined (N_UNDF) external (N_EXT) types, n_value is the size in
407 // bytes of the symbol.
408 ExtOrCommonSym.n_value = Size;
409 SymbolTable.push_back(ExtOrCommonSym);
410 // Remember that we've seen this symbol
414 // Otherwise, this symbol is part of the .bss section.
415 MachOSection *BSS = getBSSSection();
416 AddSymbolToSection(BSS, GV);
420 // Scalar read-only data goes in a literal section if the scalar is 4, 8, or
421 // 16 bytes, or a cstring. Other read only data goes into a regular const
422 // section. Read-write data goes in the data section.
423 MachOSection *Sec = GV->isConstant() ? getConstSection(GV->getInitializer()) :
425 AddSymbolToSection(Sec, GV);
426 InitMem(GV->getInitializer(), &Sec->SectionData[0], GVOffset[GV],
427 TM.getTargetData(), Sec->Relocations);
431 bool MachOWriter::runOnMachineFunction(MachineFunction &MF) {
432 // Nothing to do here, this is all done through the MCE object.
436 bool MachOWriter::doInitialization(Module &M) {
437 // Set the magic value, now that we know the pointer size and endianness
438 Header.setMagic(isLittleEndian, is64Bit);
441 // FIXME: this only works for object files, we do not support the creation
442 // of dynamic libraries or executables at this time.
443 Header.filetype = MachOHeader::MH_OBJECT;
445 Mang = new Mangler(M);
449 /// doFinalization - Now that the module has been completely processed, emit
450 /// the Mach-O file to 'O'.
451 bool MachOWriter::doFinalization(Module &M) {
452 // FIXME: we don't handle debug info yet, we should probably do that.
454 // Okay, the.text section has been completed, build the .data, .bss, and
455 // "common" sections next.
456 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
460 // Emit the header and load commands.
461 EmitHeaderAndLoadCommands();
463 // Emit the various sections and their relocation info.
466 // Write the symbol table and the string table to the end of the file.
467 O.write((char*)&SymT[0], SymT.size());
468 O.write((char*)&StrT[0], StrT.size());
470 // We are done with the abstract symbols.
473 DynamicSymbolTable.clear();
475 // Release the name mangler object.
476 delete Mang; Mang = 0;
480 void MachOWriter::EmitHeaderAndLoadCommands() {
481 // Step #0: Fill in the segment load command size, since we need it to figure
482 // out the rest of the header fields
483 MachOSegment SEG("", is64Bit);
484 SEG.nsects = SectionList.size();
485 SEG.cmdsize = SEG.cmdSize(is64Bit) +
486 SEG.nsects * SectionList[0]->cmdSize(is64Bit);
488 // Step #1: calculate the number of load commands. We always have at least
489 // one, for the LC_SEGMENT load command, plus two for the normal
490 // and dynamic symbol tables, if there are any symbols.
491 Header.ncmds = SymbolTable.empty() ? 1 : 3;
493 // Step #2: calculate the size of the load commands
494 Header.sizeofcmds = SEG.cmdsize;
495 if (!SymbolTable.empty())
496 Header.sizeofcmds += SymTab.cmdsize + DySymTab.cmdsize;
498 // Step #3: write the header to the file
499 // Local alias to shortenify coming code.
500 DataBuffer &FH = Header.HeaderData;
501 OutputBuffer FHOut(FH, is64Bit, isLittleEndian);
503 FHOut.outword(Header.magic);
504 FHOut.outword(TM.getMachOWriterInfo()->getCPUType());
505 FHOut.outword(TM.getMachOWriterInfo()->getCPUSubType());
506 FHOut.outword(Header.filetype);
507 FHOut.outword(Header.ncmds);
508 FHOut.outword(Header.sizeofcmds);
509 FHOut.outword(Header.flags);
511 FHOut.outword(Header.reserved);
513 // Step #4: Finish filling in the segment load command and write it out
514 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
515 E = SectionList.end(); I != E; ++I)
516 SEG.filesize += (*I)->size;
518 SEG.vmsize = SEG.filesize;
519 SEG.fileoff = Header.cmdSize(is64Bit) + Header.sizeofcmds;
521 FHOut.outword(SEG.cmd);
522 FHOut.outword(SEG.cmdsize);
523 FHOut.outstring(SEG.segname, 16);
524 FHOut.outaddr(SEG.vmaddr);
525 FHOut.outaddr(SEG.vmsize);
526 FHOut.outaddr(SEG.fileoff);
527 FHOut.outaddr(SEG.filesize);
528 FHOut.outword(SEG.maxprot);
529 FHOut.outword(SEG.initprot);
530 FHOut.outword(SEG.nsects);
531 FHOut.outword(SEG.flags);
533 // Step #5: Finish filling in the fields of the MachOSections
534 uint64_t currentAddr = 0;
535 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
536 E = SectionList.end(); I != E; ++I) {
537 MachOSection *MOS = *I;
538 MOS->addr = currentAddr;
539 MOS->offset = currentAddr + SEG.fileoff;
541 // FIXME: do we need to do something with alignment here?
542 currentAddr += MOS->size;
545 // Step #6: Emit the symbol table to temporary buffers, so that we know the
546 // size of the string table when we write the next load command. This also
547 // sorts and assigns indices to each of the symbols, which is necessary for
548 // emitting relocations to externally-defined objects.
549 BufferSymbolAndStringTable();
551 // Step #7: Calculate the number of relocations for each section and write out
552 // the section commands for each section
553 currentAddr += SEG.fileoff;
554 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
555 E = SectionList.end(); I != E; ++I) {
556 MachOSection *MOS = *I;
557 // Convert the relocations to target-specific relocations, and fill in the
558 // relocation offset for this section.
559 CalculateRelocations(*MOS);
560 MOS->reloff = MOS->nreloc ? currentAddr : 0;
561 currentAddr += MOS->nreloc * 8;
563 // write the finalized section command to the output buffer
564 FHOut.outstring(MOS->sectname, 16);
565 FHOut.outstring(MOS->segname, 16);
566 FHOut.outaddr(MOS->addr);
567 FHOut.outaddr(MOS->size);
568 FHOut.outword(MOS->offset);
569 FHOut.outword(MOS->align);
570 FHOut.outword(MOS->reloff);
571 FHOut.outword(MOS->nreloc);
572 FHOut.outword(MOS->flags);
573 FHOut.outword(MOS->reserved1);
574 FHOut.outword(MOS->reserved2);
576 FHOut.outword(MOS->reserved3);
579 // Step #8: Emit LC_SYMTAB/LC_DYSYMTAB load commands
580 SymTab.symoff = currentAddr;
581 SymTab.nsyms = SymbolTable.size();
582 SymTab.stroff = SymTab.symoff + SymT.size();
583 SymTab.strsize = StrT.size();
584 FHOut.outword(SymTab.cmd);
585 FHOut.outword(SymTab.cmdsize);
586 FHOut.outword(SymTab.symoff);
587 FHOut.outword(SymTab.nsyms);
588 FHOut.outword(SymTab.stroff);
589 FHOut.outword(SymTab.strsize);
591 // FIXME: set DySymTab fields appropriately
592 // We should probably just update these in BufferSymbolAndStringTable since
593 // thats where we're partitioning up the different kinds of symbols.
594 FHOut.outword(DySymTab.cmd);
595 FHOut.outword(DySymTab.cmdsize);
596 FHOut.outword(DySymTab.ilocalsym);
597 FHOut.outword(DySymTab.nlocalsym);
598 FHOut.outword(DySymTab.iextdefsym);
599 FHOut.outword(DySymTab.nextdefsym);
600 FHOut.outword(DySymTab.iundefsym);
601 FHOut.outword(DySymTab.nundefsym);
602 FHOut.outword(DySymTab.tocoff);
603 FHOut.outword(DySymTab.ntoc);
604 FHOut.outword(DySymTab.modtaboff);
605 FHOut.outword(DySymTab.nmodtab);
606 FHOut.outword(DySymTab.extrefsymoff);
607 FHOut.outword(DySymTab.nextrefsyms);
608 FHOut.outword(DySymTab.indirectsymoff);
609 FHOut.outword(DySymTab.nindirectsyms);
610 FHOut.outword(DySymTab.extreloff);
611 FHOut.outword(DySymTab.nextrel);
612 FHOut.outword(DySymTab.locreloff);
613 FHOut.outword(DySymTab.nlocrel);
615 O.write((char*)&FH[0], FH.size());
618 /// EmitSections - Now that we have constructed the file header and load
619 /// commands, emit the data for each section to the file.
620 void MachOWriter::EmitSections() {
621 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
622 E = SectionList.end(); I != E; ++I)
623 // Emit the contents of each section
624 O.write((char*)&(*I)->SectionData[0], (*I)->size);
625 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
626 E = SectionList.end(); I != E; ++I)
627 // Emit the relocation entry data for each section.
628 O.write((char*)&(*I)->RelocBuffer[0], (*I)->RelocBuffer.size());
631 /// PartitionByLocal - Simple boolean predicate that returns true if Sym is
632 /// a local symbol rather than an external symbol.
633 bool MachOWriter::PartitionByLocal(const MachOSym &Sym) {
634 return (Sym.n_type & (MachOSym::N_EXT | MachOSym::N_PEXT)) == 0;
637 /// PartitionByDefined - Simple boolean predicate that returns true if Sym is
638 /// defined in this module.
639 bool MachOWriter::PartitionByDefined(const MachOSym &Sym) {
640 // FIXME: Do N_ABS or N_INDR count as defined?
641 return (Sym.n_type & MachOSym::N_SECT) == MachOSym::N_SECT;
644 /// BufferSymbolAndStringTable - Sort the symbols we encountered and assign them
645 /// each a string table index so that they appear in the correct order in the
647 void MachOWriter::BufferSymbolAndStringTable() {
648 // The order of the symbol table is:
650 // 2. defined external symbols (sorted by name)
651 // 3. undefined external symbols (sorted by name)
653 // Before sorting the symbols, check the PendingGlobals for any undefined
654 // globals that need to be put in the symbol table.
655 for (std::vector<GlobalValue*>::iterator I = PendingGlobals.begin(),
656 E = PendingGlobals.end(); I != E; ++I) {
657 if (GVOffset[*I] == 0 && GVSection[*I] == 0) {
658 MachOSym UndfSym(*I, Mang->getValueName(*I), MachOSym::NO_SECT, TM);
659 SymbolTable.push_back(UndfSym);
664 // Sort the symbols by name, so that when we partition the symbols by scope
665 // of definition, we won't have to sort by name within each partition.
666 std::sort(SymbolTable.begin(), SymbolTable.end(), MachOSymCmp());
668 // Parition the symbol table entries so that all local symbols come before
669 // all symbols with external linkage. { 1 | 2 3 }
670 std::partition(SymbolTable.begin(), SymbolTable.end(), PartitionByLocal);
672 // Advance iterator to beginning of external symbols and partition so that
673 // all external symbols defined in this module come before all external
674 // symbols defined elsewhere. { 1 | 2 | 3 }
675 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
676 E = SymbolTable.end(); I != E; ++I) {
677 if (!PartitionByLocal(*I)) {
678 std::partition(I, E, PartitionByDefined);
683 // Calculate the starting index for each of the local, extern defined, and
684 // undefined symbols, as well as the number of each to put in the LC_DYSYMTAB
686 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
687 E = SymbolTable.end(); I != E; ++I) {
688 if (PartitionByLocal(*I)) {
689 ++DySymTab.nlocalsym;
690 ++DySymTab.iextdefsym;
691 ++DySymTab.iundefsym;
692 } else if (PartitionByDefined(*I)) {
693 ++DySymTab.nextdefsym;
694 ++DySymTab.iundefsym;
696 ++DySymTab.nundefsym;
700 // Write out a leading zero byte when emitting string table, for n_strx == 0
701 // which means an empty string.
702 OutputBuffer StrTOut(StrT, is64Bit, isLittleEndian);
705 // The order of the string table is:
706 // 1. strings for external symbols
707 // 2. strings for local symbols
708 // Since this is the opposite order from the symbol table, which we have just
709 // sorted, we can walk the symbol table backwards to output the string table.
710 for (std::vector<MachOSym>::reverse_iterator I = SymbolTable.rbegin(),
711 E = SymbolTable.rend(); I != E; ++I) {
712 if (I->GVName == "") {
715 I->n_strx = StrT.size();
716 StrTOut.outstring(I->GVName, I->GVName.length()+1);
720 OutputBuffer SymTOut(SymT, is64Bit, isLittleEndian);
723 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
724 E = SymbolTable.end(); I != E; ++I, ++index) {
725 // Add the section base address to the section offset in the n_value field
726 // to calculate the full address.
727 // FIXME: handle symbols where the n_value field is not the address
728 GlobalValue *GV = const_cast<GlobalValue*>(I->GV);
729 if (GV && GVSection[GV])
730 I->n_value += GVSection[GV]->addr;
731 if (GV && (GVOffset[GV] == -1))
732 GVOffset[GV] = index;
734 // Emit nlist to buffer
735 SymTOut.outword(I->n_strx);
736 SymTOut.outbyte(I->n_type);
737 SymTOut.outbyte(I->n_sect);
738 SymTOut.outhalf(I->n_desc);
739 SymTOut.outaddr(I->n_value);
743 /// CalculateRelocations - For each MachineRelocation in the current section,
744 /// calculate the index of the section containing the object to be relocated,
745 /// and the offset into that section. From this information, create the
746 /// appropriate target-specific MachORelocation type and add buffer it to be
747 /// written out after we are finished writing out sections.
748 void MachOWriter::CalculateRelocations(MachOSection &MOS) {
749 for (unsigned i = 0, e = MOS.Relocations.size(); i != e; ++i) {
750 MachineRelocation &MR = MOS.Relocations[i];
751 unsigned TargetSection = MR.getConstantVal();
752 unsigned TargetAddr = 0;
753 unsigned TargetIndex = 0;
755 // This is a scattered relocation entry if it points to a global value with
756 // a non-zero offset.
757 bool Scattered = false;
760 // Since we may not have seen the GlobalValue we were interested in yet at
761 // the time we emitted the relocation for it, fix it up now so that it
762 // points to the offset into the correct section.
763 if (MR.isGlobalValue()) {
764 GlobalValue *GV = MR.getGlobalValue();
765 MachOSection *MOSPtr = GVSection[GV];
766 intptr_t Offset = GVOffset[GV];
768 // If we have never seen the global before, it must be to a symbol
769 // defined in another module (N_UNDF).
771 // FIXME: need to append stub suffix
774 TargetIndex = GVOffset[GV];
776 Scattered = TargetSection != 0;
777 TargetSection = MOSPtr->Index;
779 MR.setResultPointer((void*)Offset);
782 // If the symbol is locally defined, pass in the address of the section and
783 // the section index to the code which will generate the target relocation.
785 MachOSection &To = *SectionList[TargetSection - 1];
786 TargetAddr = To.addr;
787 TargetIndex = To.Index;
790 OutputBuffer RelocOut(MOS.RelocBuffer, is64Bit, isLittleEndian);
791 OutputBuffer SecOut(MOS.SectionData, is64Bit, isLittleEndian);
793 MOS.nreloc += GetTargetRelocation(MR, MOS.Index, TargetAddr, TargetIndex,
794 RelocOut, SecOut, Scattered, Extern);
798 // InitMem - Write the value of a Constant to the specified memory location,
799 // converting it into bytes and relocations.
800 void MachOWriter::InitMem(const Constant *C, void *Addr, intptr_t Offset,
801 const TargetData *TD,
802 std::vector<MachineRelocation> &MRs) {
803 typedef std::pair<const Constant*, intptr_t> CPair;
804 std::vector<CPair> WorkList;
806 WorkList.push_back(CPair(C,(intptr_t)Addr + Offset));
808 intptr_t ScatteredOffset = 0;
810 while (!WorkList.empty()) {
811 const Constant *PC = WorkList.back().first;
812 intptr_t PA = WorkList.back().second;
815 if (isa<UndefValue>(PC)) {
817 } else if (const ConstantVector *CP = dyn_cast<ConstantVector>(PC)) {
818 unsigned ElementSize =
819 TD->getABITypeSize(CP->getType()->getElementType());
820 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
821 WorkList.push_back(CPair(CP->getOperand(i), PA+i*ElementSize));
822 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(PC)) {
824 // FIXME: Handle ConstantExpression. See EE::getConstantValue()
826 switch (CE->getOpcode()) {
827 case Instruction::GetElementPtr: {
828 SmallVector<Value*, 8> Indices(CE->op_begin()+1, CE->op_end());
829 ScatteredOffset = TD->getIndexedOffset(CE->getOperand(0)->getType(),
830 &Indices[0], Indices.size());
831 WorkList.push_back(CPair(CE->getOperand(0), PA));
834 case Instruction::Add:
836 cerr << "ConstantExpr not handled as global var init: " << *CE << "\n";
840 } else if (PC->getType()->isFirstClassType()) {
841 unsigned char *ptr = (unsigned char *)PA;
842 switch (PC->getType()->getTypeID()) {
843 case Type::IntegerTyID: {
844 unsigned NumBits = cast<IntegerType>(PC->getType())->getBitWidth();
845 uint64_t val = cast<ConstantInt>(PC)->getZExtValue();
848 else if (NumBits <= 16) {
849 if (TD->isBigEndian())
850 val = ByteSwap_16(val);
853 } else if (NumBits <= 32) {
854 if (TD->isBigEndian())
855 val = ByteSwap_32(val);
860 } else if (NumBits <= 64) {
861 if (TD->isBigEndian())
862 val = ByteSwap_64(val);
872 assert(0 && "Not implemented: bit widths > 64");
876 case Type::FloatTyID: {
877 uint32_t val = cast<ConstantFP>(PC)->getValueAPF().convertToAPInt().
879 if (TD->isBigEndian())
880 val = ByteSwap_32(val);
887 case Type::DoubleTyID: {
888 uint64_t val = cast<ConstantFP>(PC)->getValueAPF().convertToAPInt().
890 if (TD->isBigEndian())
891 val = ByteSwap_64(val);
902 case Type::PointerTyID:
903 if (isa<ConstantPointerNull>(PC))
904 memset(ptr, 0, TD->getPointerSize());
905 else if (const GlobalValue* GV = dyn_cast<GlobalValue>(PC)) {
906 // FIXME: what about function stubs?
907 MRs.push_back(MachineRelocation::getGV(PA-(intptr_t)Addr,
908 MachineRelocation::VANILLA,
909 const_cast<GlobalValue*>(GV),
913 assert(0 && "Unknown constant pointer type!");
916 cerr << "ERROR: Constant unimp for type: " << *PC->getType() << "\n";
919 } else if (isa<ConstantAggregateZero>(PC)) {
920 memset((void*)PA, 0, (size_t)TD->getABITypeSize(PC->getType()));
921 } else if (const ConstantArray *CPA = dyn_cast<ConstantArray>(PC)) {
922 unsigned ElementSize =
923 TD->getABITypeSize(CPA->getType()->getElementType());
924 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i)
925 WorkList.push_back(CPair(CPA->getOperand(i), PA+i*ElementSize));
926 } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(PC)) {
927 const StructLayout *SL =
928 TD->getStructLayout(cast<StructType>(CPS->getType()));
929 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i)
930 WorkList.push_back(CPair(CPS->getOperand(i),
931 PA+SL->getElementOffset(i)));
933 cerr << "Bad Type: " << *PC->getType() << "\n";
934 assert(0 && "Unknown constant type to initialize memory with!");
939 MachOSym::MachOSym(const GlobalValue *gv, std::string name, uint8_t sect,
941 GV(gv), n_strx(0), n_type(sect == NO_SECT ? N_UNDF : N_SECT), n_sect(sect),
942 n_desc(0), n_value(0) {
944 const TargetAsmInfo *TAI = TM.getTargetAsmInfo();
946 switch (GV->getLinkage()) {
948 assert(0 && "Unexpected linkage type!");
950 case GlobalValue::WeakLinkage:
951 case GlobalValue::LinkOnceLinkage:
952 assert(!isa<Function>(gv) && "Unexpected linkage type for Function!");
953 case GlobalValue::ExternalLinkage:
954 GVName = TAI->getGlobalPrefix() + name;
955 n_type |= GV->hasHiddenVisibility() ? N_PEXT : N_EXT;
957 case GlobalValue::InternalLinkage:
958 GVName = TAI->getGlobalPrefix() + name;