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 // Upgrade the section alignment if required.
149 if (MOS->align < Align) MOS->align = Align;
151 // Round the size up to the correct alignment for starting the new function.
152 if ((MOS->size & ((1 << Align) - 1)) != 0) {
153 MOS->size += (1 << Align);
154 MOS->size &= ~((1 << Align) - 1);
157 // FIXME: Using MOS->size directly here instead of calculating it from the
158 // output buffer size (impossible because the code emitter deals only in raw
159 // bytes) forces us to manually synchronize size and write padding zero bytes
160 // to the output buffer for all non-text sections. For text sections, we do
161 // not synchonize the output buffer, and we just blow up if anyone tries to
162 // write non-code to it. An assert should probably be added to
163 // AddSymbolToSection to prevent calling it on the text section.
164 CurBufferPtr = BufferBegin + MOS->size;
166 // Clear per-function data structures.
170 MBBLocations.clear();
173 /// finishFunction - This callback is invoked after the function is completely
175 bool MachOCodeEmitter::finishFunction(MachineFunction &F) {
176 // Get the Mach-O Section that this function belongs in.
177 MachOWriter::MachOSection *MOS = MOW.getTextSection();
179 // Get a symbol for the function to add to the symbol table
180 // FIXME: it seems like we should call something like AddSymbolToSection
181 // in startFunction rather than changing the section size and symbol n_value
183 const GlobalValue *FuncV = F.getFunction();
184 MachOSym FnSym(FuncV, MOW.Mang->getValueName(FuncV), MOS->Index, TM);
185 FnSym.n_value = MOS->size;
186 MOS->size = CurBufferPtr - BufferBegin;
188 // Emit constant pool to appropriate section(s)
189 emitConstantPool(F.getConstantPool());
191 // Emit jump tables to appropriate section
192 emitJumpTables(F.getJumpTableInfo());
194 // If we have emitted any relocations to function-specific objects such as
195 // basic blocks, constant pools entries, or jump tables, record their
196 // addresses now so that we can rewrite them with the correct addresses
198 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
199 MachineRelocation &MR = Relocations[i];
202 if (MR.isBasicBlock()) {
203 Addr = getMachineBasicBlockAddress(MR.getBasicBlock());
204 MR.setConstantVal(MOS->Index);
205 MR.setResultPointer((void*)Addr);
206 } else if (MR.isJumpTableIndex()) {
207 Addr = getJumpTableEntryAddress(MR.getJumpTableIndex());
208 MR.setConstantVal(MOW.getJumpTableSection()->Index);
209 MR.setResultPointer((void*)Addr);
210 } else if (MR.isConstantPoolIndex()) {
211 Addr = getConstantPoolEntryAddress(MR.getConstantPoolIndex());
212 MR.setConstantVal(CPSections[MR.getConstantPoolIndex()]);
213 MR.setResultPointer((void*)Addr);
214 } else if (!MR.isGlobalValue()) {
215 assert(0 && "Unhandled relocation type");
217 MOS->Relocations.push_back(MR);
221 // Finally, add it to the symtab.
222 MOW.SymbolTable.push_back(FnSym);
226 /// emitConstantPool - For each constant pool entry, figure out which section
227 /// the constant should live in, allocate space for it, and emit it to the
228 /// Section data buffer.
229 void MachOCodeEmitter::emitConstantPool(MachineConstantPool *MCP) {
230 const std::vector<MachineConstantPoolEntry> &CP = MCP->getConstants();
231 if (CP.empty()) return;
233 // FIXME: handle PIC codegen
234 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
235 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
237 // Although there is no strict necessity that I am aware of, we will do what
238 // gcc for OS X does and put each constant pool entry in a section of constant
239 // objects of a certain size. That means that float constants go in the
240 // literal4 section, and double objects go in literal8, etc.
242 // FIXME: revisit this decision if we ever do the "stick everything into one
243 // "giant object for PIC" optimization.
244 for (unsigned i = 0, e = CP.size(); i != e; ++i) {
245 const Type *Ty = CP[i].getType();
246 unsigned Size = TM.getTargetData()->getTypeSize(Ty);
248 MachOWriter::MachOSection *Sec = MOW.getConstSection(CP[i].Val.ConstVal);
249 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
251 CPLocations.push_back(Sec->SectionData.size());
252 CPSections.push_back(Sec->Index);
254 // FIXME: remove when we have unified size + output buffer
257 // Allocate space in the section for the global.
258 // FIXME: need alignment?
259 // FIXME: share between here and AddSymbolToSection?
260 for (unsigned j = 0; j < Size; ++j)
261 SecDataOut.outbyte(0);
263 MOW.InitMem(CP[i].Val.ConstVal, &Sec->SectionData[0], CPLocations[i],
264 TM.getTargetData(), Sec->Relocations);
268 /// emitJumpTables - Emit all the jump tables for a given jump table info
269 /// record to the appropriate section.
270 void MachOCodeEmitter::emitJumpTables(MachineJumpTableInfo *MJTI) {
271 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
272 if (JT.empty()) return;
274 // FIXME: handle PIC codegen
275 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
276 assert(!isPIC && "PIC codegen not yet handled for mach-o jump tables!");
278 MachOWriter::MachOSection *Sec = MOW.getJumpTableSection();
279 unsigned TextSecIndex = MOW.getTextSection()->Index;
280 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
282 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
283 // For each jump table, record its offset from the start of the section,
284 // reserve space for the relocations to the MBBs, and add the relocations.
285 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
286 JTLocations.push_back(Sec->SectionData.size());
287 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi) {
288 MachineRelocation MR(MOW.GetJTRelocation(Sec->SectionData.size(),
290 MR.setResultPointer((void *)JTLocations[i]);
291 MR.setConstantVal(TextSecIndex);
292 Sec->Relocations.push_back(MR);
293 SecDataOut.outaddr(0);
296 // FIXME: remove when we have unified size + output buffer
297 Sec->size = Sec->SectionData.size();
300 //===----------------------------------------------------------------------===//
301 // MachOWriter Implementation
302 //===----------------------------------------------------------------------===//
304 MachOWriter::MachOWriter(std::ostream &o, TargetMachine &tm) : O(o), TM(tm) {
305 is64Bit = TM.getTargetData()->getPointerSizeInBits() == 64;
306 isLittleEndian = TM.getTargetData()->isLittleEndian();
308 // Create the machine code emitter object for this target.
309 MCE = new MachOCodeEmitter(*this);
312 MachOWriter::~MachOWriter() {
316 void MachOWriter::AddSymbolToSection(MachOSection *Sec, GlobalVariable *GV) {
317 const Type *Ty = GV->getType()->getElementType();
318 unsigned Size = TM.getTargetData()->getTypeSize(Ty);
319 unsigned Align = GV->getAlignment();
321 Align = TM.getTargetData()->getTypeAlignmentPref(Ty);
323 MachOSym Sym(GV, Mang->getValueName(GV), Sec->Index, TM);
325 // Reserve space in the .bss section for this symbol while maintaining the
326 // desired section alignment, which must be at least as much as required by
328 OutputBuffer SecDataOut(Sec->SectionData, is64Bit, isLittleEndian);
331 uint64_t OrigSize = Sec->size;
332 Align = Log2_32(Align);
333 Sec->align = std::max(unsigned(Sec->align), Align);
334 Sec->size = (Sec->size + Align - 1) & ~(Align-1);
336 // Add alignment padding to buffer as well.
337 // FIXME: remove when we have unified size + output buffer
338 unsigned AlignedSize = Sec->size - OrigSize;
339 for (unsigned i = 0; i < AlignedSize; ++i)
340 SecDataOut.outbyte(0);
342 // Record the offset of the symbol, and then allocate space for it.
343 // FIXME: remove when we have unified size + output buffer
344 Sym.n_value = Sec->size;
346 SymbolTable.push_back(Sym);
348 // Now that we know what section the GlovalVariable is going to be emitted
349 // into, update our mappings.
350 // FIXME: We may also need to update this when outputting non-GlobalVariable
351 // GlobalValues such as functions.
353 GVOffset[GV] = Sec->SectionData.size();
355 // Allocate space in the section for the global.
356 for (unsigned i = 0; i < Size; ++i)
357 SecDataOut.outbyte(0);
360 void MachOWriter::EmitGlobal(GlobalVariable *GV) {
361 const Type *Ty = GV->getType()->getElementType();
362 unsigned Size = TM.getTargetData()->getTypeSize(Ty);
363 bool NoInit = !GV->hasInitializer();
365 // If this global has a zero initializer, it is part of the .bss or common
367 if (NoInit || GV->getInitializer()->isNullValue()) {
368 // If this global is part of the common block, add it now. Variables are
369 // part of the common block if they are zero initialized and allowed to be
370 // merged with other symbols.
371 if (NoInit || GV->hasLinkOnceLinkage() || GV->hasWeakLinkage()) {
372 MachOSym ExtOrCommonSym(GV, Mang->getValueName(GV), MachOSym::NO_SECT,TM);
373 // For undefined (N_UNDF) external (N_EXT) types, n_value is the size in
374 // bytes of the symbol.
375 ExtOrCommonSym.n_value = Size;
376 // If the symbol is external, we'll put it on a list of symbols whose
377 // addition to the symbol table is being pended until we find a reference
379 PendingSyms.push_back(ExtOrCommonSym);
381 SymbolTable.push_back(ExtOrCommonSym);
384 // Otherwise, this symbol is part of the .bss section.
385 MachOSection *BSS = getBSSSection();
386 AddSymbolToSection(BSS, GV);
390 // Scalar read-only data goes in a literal section if the scalar is 4, 8, or
391 // 16 bytes, or a cstring. Other read only data goes into a regular const
392 // section. Read-write data goes in the data section.
393 MachOSection *Sec = GV->isConstant() ? getConstSection(GV->getInitializer()) :
395 AddSymbolToSection(Sec, GV);
396 InitMem(GV->getInitializer(), &Sec->SectionData[0], GVOffset[GV],
397 TM.getTargetData(), Sec->Relocations);
401 bool MachOWriter::runOnMachineFunction(MachineFunction &MF) {
402 // Nothing to do here, this is all done through the MCE object.
406 bool MachOWriter::doInitialization(Module &M) {
407 // Set the magic value, now that we know the pointer size and endianness
408 Header.setMagic(isLittleEndian, is64Bit);
411 // FIXME: this only works for object files, we do not support the creation
412 // of dynamic libraries or executables at this time.
413 Header.filetype = MachOHeader::MH_OBJECT;
415 Mang = new Mangler(M);
419 /// doFinalization - Now that the module has been completely processed, emit
420 /// the Mach-O file to 'O'.
421 bool MachOWriter::doFinalization(Module &M) {
422 // FIXME: we don't handle debug info yet, we should probably do that.
424 // Okay, the.text section has been completed, build the .data, .bss, and
425 // "common" sections next.
426 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
430 // Emit the header and load commands.
431 EmitHeaderAndLoadCommands();
433 // Emit the various sections and their relocation info.
436 // Write the symbol table and the string table to the end of the file.
437 O.write((char*)&SymT[0], SymT.size());
438 O.write((char*)&StrT[0], StrT.size());
440 // We are done with the abstract symbols.
443 DynamicSymbolTable.clear();
445 // Release the name mangler object.
446 delete Mang; Mang = 0;
450 void MachOWriter::EmitHeaderAndLoadCommands() {
451 // Step #0: Fill in the segment load command size, since we need it to figure
452 // out the rest of the header fields
453 MachOSegment SEG("", is64Bit);
454 SEG.nsects = SectionList.size();
455 SEG.cmdsize = SEG.cmdSize(is64Bit) +
456 SEG.nsects * SectionList[0]->cmdSize(is64Bit);
458 // Step #1: calculate the number of load commands. We always have at least
459 // one, for the LC_SEGMENT load command, plus two for the normal
460 // and dynamic symbol tables, if there are any symbols.
461 Header.ncmds = SymbolTable.empty() ? 1 : 3;
463 // Step #2: calculate the size of the load commands
464 Header.sizeofcmds = SEG.cmdsize;
465 if (!SymbolTable.empty())
466 Header.sizeofcmds += SymTab.cmdsize + DySymTab.cmdsize;
468 // Step #3: write the header to the file
469 // Local alias to shortenify coming code.
470 DataBuffer &FH = Header.HeaderData;
471 OutputBuffer FHOut(FH, is64Bit, isLittleEndian);
473 FHOut.outword(Header.magic);
474 FHOut.outword(TM.getMachOWriterInfo()->getCPUType());
475 FHOut.outword(TM.getMachOWriterInfo()->getCPUSubType());
476 FHOut.outword(Header.filetype);
477 FHOut.outword(Header.ncmds);
478 FHOut.outword(Header.sizeofcmds);
479 FHOut.outword(Header.flags);
481 FHOut.outword(Header.reserved);
483 // Step #4: Finish filling in the segment load command and write it out
484 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
485 E = SectionList.end(); I != E; ++I)
486 SEG.filesize += (*I)->size;
488 SEG.vmsize = SEG.filesize;
489 SEG.fileoff = Header.cmdSize(is64Bit) + Header.sizeofcmds;
491 FHOut.outword(SEG.cmd);
492 FHOut.outword(SEG.cmdsize);
493 FHOut.outstring(SEG.segname, 16);
494 FHOut.outaddr(SEG.vmaddr);
495 FHOut.outaddr(SEG.vmsize);
496 FHOut.outaddr(SEG.fileoff);
497 FHOut.outaddr(SEG.filesize);
498 FHOut.outword(SEG.maxprot);
499 FHOut.outword(SEG.initprot);
500 FHOut.outword(SEG.nsects);
501 FHOut.outword(SEG.flags);
503 // Step #5: Finish filling in the fields of the MachOSections
504 uint64_t currentAddr = 0;
505 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
506 E = SectionList.end(); I != E; ++I) {
507 MachOSection *MOS = *I;
508 MOS->addr = currentAddr;
509 MOS->offset = currentAddr + SEG.fileoff;
511 // FIXME: do we need to do something with alignment here?
512 currentAddr += MOS->size;
515 // Step #6: Calculate the number of relocations for each section and write out
516 // the section commands for each section
517 currentAddr += SEG.fileoff;
518 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
519 E = SectionList.end(); I != E; ++I) {
520 MachOSection *MOS = *I;
521 // Convert the relocations to target-specific relocations, and fill in the
522 // relocation offset for this section.
523 CalculateRelocations(*MOS);
524 MOS->reloff = MOS->nreloc ? currentAddr : 0;
525 currentAddr += MOS->nreloc * 8;
527 // write the finalized section command to the output buffer
528 FHOut.outstring(MOS->sectname, 16);
529 FHOut.outstring(MOS->segname, 16);
530 FHOut.outaddr(MOS->addr);
531 FHOut.outaddr(MOS->size);
532 FHOut.outword(MOS->offset);
533 FHOut.outword(MOS->align);
534 FHOut.outword(MOS->reloff);
535 FHOut.outword(MOS->nreloc);
536 FHOut.outword(MOS->flags);
537 FHOut.outword(MOS->reserved1);
538 FHOut.outword(MOS->reserved2);
540 FHOut.outword(MOS->reserved3);
543 // Step #7: Emit the symbol table to temporary buffers, so that we know the
544 // size of the string table when we write the next load command.
545 BufferSymbolAndStringTable();
547 // Step #8: Emit LC_SYMTAB/LC_DYSYMTAB load commands
548 SymTab.symoff = currentAddr;
549 SymTab.nsyms = SymbolTable.size();
550 SymTab.stroff = SymTab.symoff + SymT.size();
551 SymTab.strsize = StrT.size();
552 FHOut.outword(SymTab.cmd);
553 FHOut.outword(SymTab.cmdsize);
554 FHOut.outword(SymTab.symoff);
555 FHOut.outword(SymTab.nsyms);
556 FHOut.outword(SymTab.stroff);
557 FHOut.outword(SymTab.strsize);
559 // FIXME: set DySymTab fields appropriately
560 // We should probably just update these in BufferSymbolAndStringTable since
561 // thats where we're partitioning up the different kinds of symbols.
562 FHOut.outword(DySymTab.cmd);
563 FHOut.outword(DySymTab.cmdsize);
564 FHOut.outword(DySymTab.ilocalsym);
565 FHOut.outword(DySymTab.nlocalsym);
566 FHOut.outword(DySymTab.iextdefsym);
567 FHOut.outword(DySymTab.nextdefsym);
568 FHOut.outword(DySymTab.iundefsym);
569 FHOut.outword(DySymTab.nundefsym);
570 FHOut.outword(DySymTab.tocoff);
571 FHOut.outword(DySymTab.ntoc);
572 FHOut.outword(DySymTab.modtaboff);
573 FHOut.outword(DySymTab.nmodtab);
574 FHOut.outword(DySymTab.extrefsymoff);
575 FHOut.outword(DySymTab.nextrefsyms);
576 FHOut.outword(DySymTab.indirectsymoff);
577 FHOut.outword(DySymTab.nindirectsyms);
578 FHOut.outword(DySymTab.extreloff);
579 FHOut.outword(DySymTab.nextrel);
580 FHOut.outword(DySymTab.locreloff);
581 FHOut.outword(DySymTab.nlocrel);
583 O.write((char*)&FH[0], FH.size());
586 /// EmitSections - Now that we have constructed the file header and load
587 /// commands, emit the data for each section to the file.
588 void MachOWriter::EmitSections() {
589 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
590 E = SectionList.end(); I != E; ++I)
591 // Emit the contents of each section
592 O.write((char*)&(*I)->SectionData[0], (*I)->size);
593 for (std::vector<MachOSection*>::iterator I = SectionList.begin(),
594 E = SectionList.end(); I != E; ++I)
595 // Emit the relocation entry data for each section.
596 O.write((char*)&(*I)->RelocBuffer[0], (*I)->RelocBuffer.size());
599 /// PartitionByLocal - Simple boolean predicate that returns true if Sym is
600 /// a local symbol rather than an external symbol.
601 bool MachOWriter::PartitionByLocal(const MachOSym &Sym) {
602 return (Sym.n_type & (MachOSym::N_EXT | MachOSym::N_PEXT)) == 0;
605 /// PartitionByDefined - Simple boolean predicate that returns true if Sym is
606 /// defined in this module.
607 bool MachOWriter::PartitionByDefined(const MachOSym &Sym) {
608 // FIXME: Do N_ABS or N_INDR count as defined?
609 return (Sym.n_type & MachOSym::N_SECT) == MachOSym::N_SECT;
612 /// BufferSymbolAndStringTable - Sort the symbols we encountered and assign them
613 /// each a string table index so that they appear in the correct order in the
615 void MachOWriter::BufferSymbolAndStringTable() {
616 // The order of the symbol table is:
618 // 2. defined external symbols (sorted by name)
619 // 3. undefined external symbols (sorted by name)
621 // Sort the symbols by name, so that when we partition the symbols by scope
622 // of definition, we won't have to sort by name within each partition.
623 std::sort(SymbolTable.begin(), SymbolTable.end(), MachOSymCmp());
625 // Parition the symbol table entries so that all local symbols come before
626 // all symbols with external linkage. { 1 | 2 3 }
627 std::partition(SymbolTable.begin(), SymbolTable.end(), PartitionByLocal);
629 // Advance iterator to beginning of external symbols and partition so that
630 // all external symbols defined in this module come before all external
631 // symbols defined elsewhere. { 1 | 2 | 3 }
632 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
633 E = SymbolTable.end(); I != E; ++I) {
634 if (!PartitionByLocal(*I)) {
635 std::partition(I, E, PartitionByDefined);
640 // Calculate the starting index for each of the local, extern defined, and
641 // undefined symbols, as well as the number of each to put in the LC_DYSYMTAB
643 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
644 E = SymbolTable.end(); I != E; ++I) {
645 if (PartitionByLocal(*I)) {
646 ++DySymTab.nlocalsym;
647 ++DySymTab.iextdefsym;
648 ++DySymTab.iundefsym;
649 } else if (PartitionByDefined(*I)) {
650 ++DySymTab.nextdefsym;
651 ++DySymTab.iundefsym;
653 ++DySymTab.nundefsym;
657 // Write out a leading zero byte when emitting string table, for n_strx == 0
658 // which means an empty string.
659 OutputBuffer StrTOut(StrT, is64Bit, isLittleEndian);
662 // The order of the string table is:
663 // 1. strings for external symbols
664 // 2. strings for local symbols
665 // Since this is the opposite order from the symbol table, which we have just
666 // sorted, we can walk the symbol table backwards to output the string table.
667 for (std::vector<MachOSym>::reverse_iterator I = SymbolTable.rbegin(),
668 E = SymbolTable.rend(); I != E; ++I) {
669 if (I->GVName == "") {
672 I->n_strx = StrT.size();
673 StrTOut.outstring(I->GVName, I->GVName.length()+1);
677 OutputBuffer SymTOut(SymT, is64Bit, isLittleEndian);
679 for (std::vector<MachOSym>::iterator I = SymbolTable.begin(),
680 E = SymbolTable.end(); I != E; ++I) {
681 // Add the section base address to the section offset in the n_value field
682 // to calculate the full address.
683 // FIXME: handle symbols where the n_value field is not the address
684 GlobalValue *GV = const_cast<GlobalValue*>(I->GV);
685 if (GV && GVSection[GV])
686 I->n_value += GVSection[GV]->addr;
688 // Emit nlist to buffer
689 SymTOut.outword(I->n_strx);
690 SymTOut.outbyte(I->n_type);
691 SymTOut.outbyte(I->n_sect);
692 SymTOut.outhalf(I->n_desc);
693 SymTOut.outaddr(I->n_value);
697 /// CalculateRelocations - For each MachineRelocation in the current section,
698 /// calculate the index of the section containing the object to be relocated,
699 /// and the offset into that section. From this information, create the
700 /// appropriate target-specific MachORelocation type and add buffer it to be
701 /// written out after we are finished writing out sections.
702 void MachOWriter::CalculateRelocations(MachOSection &MOS) {
703 for (unsigned i = 0, e = MOS.Relocations.size(); i != e; ++i) {
704 MachineRelocation &MR = MOS.Relocations[i];
705 unsigned TargetSection = MR.getConstantVal();
707 // This is a scattered relocation entry if it points to a global value with
708 // a non-zero offset.
709 bool Scattered = false;
711 // Since we may not have seen the GlobalValue we were interested in yet at
712 // the time we emitted the relocation for it, fix it up now so that it
713 // points to the offset into the correct section.
714 if (MR.isGlobalValue()) {
715 GlobalValue *GV = MR.getGlobalValue();
716 MachOSection *MOSPtr = GVSection[GV];
717 intptr_t Offset = GVOffset[GV];
718 Scattered = TargetSection != 0;
721 cerr << "Trying to relocate unknown global " << *GV << '\n';
726 TargetSection = MOSPtr->Index;
727 MR.setResultPointer((void*)Offset);
730 OutputBuffer RelocOut(MOS.RelocBuffer, is64Bit, isLittleEndian);
731 OutputBuffer SecOut(MOS.SectionData, is64Bit, isLittleEndian);
732 MachOSection &To = *SectionList[TargetSection - 1];
734 MOS.nreloc += GetTargetRelocation(MR, MOS.Index, To.addr, To.Index,
735 RelocOut, SecOut, Scattered);
739 // InitMem - Write the value of a Constant to the specified memory location,
740 // converting it into bytes and relocations.
741 void MachOWriter::InitMem(const Constant *C, void *Addr, intptr_t Offset,
742 const TargetData *TD,
743 std::vector<MachineRelocation> &MRs) {
744 typedef std::pair<const Constant*, intptr_t> CPair;
745 std::vector<CPair> WorkList;
747 WorkList.push_back(CPair(C,(intptr_t)Addr + Offset));
749 intptr_t ScatteredOffset = 0;
751 while (!WorkList.empty()) {
752 const Constant *PC = WorkList.back().first;
753 intptr_t PA = WorkList.back().second;
756 if (isa<UndefValue>(PC)) {
758 } else if (const ConstantPacked *CP = dyn_cast<ConstantPacked>(PC)) {
759 unsigned ElementSize = TD->getTypeSize(CP->getType()->getElementType());
760 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
761 WorkList.push_back(CPair(CP->getOperand(i), PA+i*ElementSize));
762 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(PC)) {
764 // FIXME: Handle ConstantExpression. See EE::getConstantValue()
766 switch (CE->getOpcode()) {
767 case Instruction::GetElementPtr: {
768 std::vector<Value*> Indexes(CE->op_begin()+1, CE->op_end());
769 ScatteredOffset = TD->getIndexedOffset(CE->getOperand(0)->getType(),
771 WorkList.push_back(CPair(CE->getOperand(0), PA));
774 case Instruction::Add:
776 cerr << "ConstantExpr not handled as global var init: " << *CE << "\n";
780 } else if (PC->getType()->isFirstClassType()) {
781 unsigned char *ptr = (unsigned char *)PA;
782 switch (PC->getType()->getTypeID()) {
783 case Type::IntegerTyID: {
784 unsigned NumBits = cast<IntegerType>(PC->getType())->getBitWidth();
785 uint64_t val = cast<ConstantInt>(PC)->getZExtValue();
788 else if (NumBits <= 16) {
789 if (TD->isBigEndian())
790 val = ByteSwap_16(val);
793 } else if (NumBits <= 32) {
794 if (TD->isBigEndian())
795 val = ByteSwap_32(val);
800 } else if (NumBits <= 64) {
801 if (TD->isBigEndian())
802 val = ByteSwap_64(val);
812 assert(0 && "Not implemented: bit widths > 64");
816 case Type::FloatTyID: {
817 uint64_t val = FloatToBits(cast<ConstantFP>(PC)->getValue());
818 if (TD->isBigEndian())
819 val = ByteSwap_32(val);
826 case Type::DoubleTyID: {
827 uint64_t val = DoubleToBits(cast<ConstantFP>(PC)->getValue());
828 if (TD->isBigEndian())
829 val = ByteSwap_64(val);
840 case Type::PointerTyID:
841 if (isa<ConstantPointerNull>(PC))
842 memset(ptr, 0, TD->getPointerSize());
843 else if (const GlobalValue* GV = dyn_cast<GlobalValue>(PC)) {
844 // FIXME: what about function stubs?
845 MRs.push_back(MachineRelocation::getGV(PA-(intptr_t)Addr,
846 MachineRelocation::VANILLA,
847 const_cast<GlobalValue*>(GV),
851 assert(0 && "Unknown constant pointer type!");
854 cerr << "ERROR: Constant unimp for type: " << *PC->getType() << "\n";
857 } else if (isa<ConstantAggregateZero>(PC)) {
858 memset((void*)PA, 0, (size_t)TD->getTypeSize(PC->getType()));
859 } else if (const ConstantArray *CPA = dyn_cast<ConstantArray>(PC)) {
860 unsigned ElementSize = TD->getTypeSize(CPA->getType()->getElementType());
861 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i)
862 WorkList.push_back(CPair(CPA->getOperand(i), PA+i*ElementSize));
863 } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(PC)) {
864 const StructLayout *SL =
865 TD->getStructLayout(cast<StructType>(CPS->getType()));
866 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i)
867 WorkList.push_back(CPair(CPS->getOperand(i), PA+SL->MemberOffsets[i]));
869 cerr << "Bad Type: " << *PC->getType() << "\n";
870 assert(0 && "Unknown constant type to initialize memory with!");
875 MachOSym::MachOSym(const GlobalValue *gv, std::string name, uint8_t sect,
877 GV(gv), n_strx(0), n_type(sect == NO_SECT ? N_UNDF : N_SECT), n_sect(sect),
878 n_desc(0), n_value(0) {
880 const TargetAsmInfo *TAI = TM.getTargetAsmInfo();
882 switch (GV->getLinkage()) {
884 assert(0 && "Unexpected linkage type!");
886 case GlobalValue::WeakLinkage:
887 case GlobalValue::LinkOnceLinkage:
888 assert(!isa<Function>(gv) && "Unexpected linkage type for Function!");
889 case GlobalValue::ExternalLinkage:
890 GVName = TAI->getGlobalPrefix() + name;
891 n_type |= GV->hasHiddenVisibility() ? N_PEXT : N_EXT;
893 case GlobalValue::InternalLinkage:
894 GVName = TAI->getGlobalPrefix() + name;