lib/MC/MCAssembler.cpp

   1 //===- lib/MC/MCAssembler.cpp - Assembler Backend Implementation ----------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9
  10 #define DEBUG_TYPE "assembler"
  11 #include "llvm/MC/MCAssembler.h"
  12 #include "llvm/MC/MCAsmLayout.h"
  13 #include "llvm/MC/MCCodeEmitter.h"
  14 #include "llvm/MC/MCExpr.h"
  15 #include "llvm/MC/MCObjectWriter.h"
  16 #include "llvm/MC/MCSymbol.h"
  17 #include "llvm/MC/MCValue.h"
  18 #include "llvm/ADT/OwningPtr.h"
  19 #include "llvm/ADT/Statistic.h"
  20 #include "llvm/ADT/StringExtras.h"
  21 #include "llvm/ADT/Twine.h"
  22 #include "llvm/Support/Debug.h"
  23 #include "llvm/Support/ErrorHandling.h"
  24 #include "llvm/Support/raw_ostream.h"
  25 #include "llvm/Target/TargetRegistry.h"
  26 #include "llvm/Target/TargetAsmBackend.h"
  27
  28 #include <vector>
  29 using namespace llvm;
  30
  31 namespace {
  32 namespace stats {
  33 STATISTIC(EmittedFragments, "Number of emitted assembler fragments");
  34 STATISTIC(EvaluateFixup, "Number of evaluated fixups");
  35 STATISTIC(FragmentLayouts, "Number of fragment layouts");
  36 STATISTIC(ObjectBytes, "Number of emitted object file bytes");
  37 STATISTIC(RelaxationSteps, "Number of assembler layout and relaxation steps");
  38 STATISTIC(RelaxedInstructions, "Number of relaxed instructions");
  39 STATISTIC(SectionLayouts, "Number of section layouts");
  40 }
  41 }
  42
  43 // FIXME FIXME FIXME: There are number of places in this file where we convert
  44 // what is a 64-bit assembler value used for computation into a value in the
  45 // object file, which may truncate it. We should detect that truncation where
  46 // invalid and report errors back.
  47
  48 /* *** */
  49
  50 MCAsmLayout::MCAsmLayout(MCAssembler &Asm) : Assembler(Asm) {
  51   // Compute the section layout order. Virtual sections must go last.
  52   for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
  53     if (!Asm.getBackend().isVirtualSection(it->getSection()))
  54       SectionOrder.push_back(&*it);
  55   for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
  56     if (Asm.getBackend().isVirtualSection(it->getSection()))
  57       SectionOrder.push_back(&*it);
  58 }
  59
  60 void MCAsmLayout::UpdateForSlide(MCFragment *F, int SlideAmount) {
  61   // We shouldn't have to do anything special to support negative slides, and it
  62   // is a perfectly valid thing to do as long as other parts of the system can
  63   // guarantee convergence.
  64   assert(SlideAmount >= 0 && "Negative slides not yet supported");
  65
  66   // Update the layout by simply recomputing the layout for the entire
  67   // file. This is trivially correct, but very slow.
  68   //
  69   // FIXME-PERF: This is O(N^2), but will be eliminated once we get smarter.
  70
  71   // Layout the sections in order.
  72   LayoutFile();
  73 }
  74
  75 void MCAsmLayout::FragmentReplaced(MCFragment *Src, MCFragment *Dst) {
  76   Dst->Offset = Src->Offset;
  77   Dst->EffectiveSize = Src->EffectiveSize;
  78 }
  79
  80 uint64_t MCAsmLayout::getFragmentAddress(const MCFragment *F) const {
  81   assert(F->getParent() && "Missing section()!");
  82   return getSectionAddress(F->getParent()) + getFragmentOffset(F);
  83 }
  84
  85 uint64_t MCAsmLayout::getFragmentEffectiveSize(const MCFragment *F) const {
  86   assert(F->EffectiveSize != ~UINT64_C(0) && "Address not set!");
  87   return F->EffectiveSize;
  88 }
  89
  90 uint64_t MCAsmLayout::getFragmentOffset(const MCFragment *F) const {
  91   assert(F->Offset != ~UINT64_C(0) && "Address not set!");
  92   return F->Offset;
  93 }
  94
  95 uint64_t MCAsmLayout::getSymbolAddress(const MCSymbolData *SD) const {
  96   assert(SD->getFragment() && "Invalid getAddress() on undefined symbol!");
  97   return getFragmentAddress(SD->getFragment()) + SD->getOffset();
  98 }
  99
 100 uint64_t MCAsmLayout::getSectionAddress(const MCSectionData *SD) const {
 101   assert(SD->Address != ~UINT64_C(0) && "Address not set!");
 102   return SD->Address;
 103 }
 104
 105 uint64_t MCAsmLayout::getSectionAddressSize(const MCSectionData *SD) const {
 106   // The size is the last fragment's end offset.
 107   const MCFragment &F = SD->getFragmentList().back();
 108   return getFragmentOffset(&F) + getFragmentEffectiveSize(&F);
 109 }
 110
 111 uint64_t MCAsmLayout::getSectionFileSize(const MCSectionData *SD) const {
 112   // Virtual sections have no file size.
 113   if (getAssembler().getBackend().isVirtualSection(SD->getSection()))
 114     return 0;
 115
 116   // Otherwise, the file size is the same as the address space size.
 117   return getSectionAddressSize(SD);
 118 }
 119
 120 uint64_t MCAsmLayout::getSectionSize(const MCSectionData *SD) const {
 121   // The logical size is the address space size minus any tail padding.
 122   uint64_t Size = getSectionAddressSize(SD);
 123   const MCAlignFragment *AF =
 124     dyn_cast<MCAlignFragment>(&(SD->getFragmentList().back()));
 125   if (AF && AF->hasOnlyAlignAddress())
 126     Size -= getFragmentEffectiveSize(AF);
 127
 128   return Size;
 129 }
 130
 131 /* *** */
 132
 133 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
 134 }
 135
 136 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent)
 137   : Kind(_Kind), Parent(_Parent), Atom(0), EffectiveSize(~UINT64_C(0))
 138 {
 139   if (Parent)
 140     Parent->getFragmentList().push_back(this);
 141 }
 142
 143 MCFragment::~MCFragment() {
 144 }
 145
 146 /* *** */
 147
 148 MCSectionData::MCSectionData() : Section(0) {}
 149
 150 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
 151   : Section(&_Section),
 152     Alignment(1),
 153     Address(~UINT64_C(0)),
 154     HasInstructions(false)
 155 {
 156   if (A)
 157     A->getSectionList().push_back(this);
 158 }
 159
 160 /* *** */
 161
 162 MCSymbolData::MCSymbolData() : Symbol(0) {}
 163
 164 MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment,
 165                            uint64_t _Offset, MCAssembler *A)
 166   : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset),
 167     IsExternal(false), IsPrivateExtern(false),
 168     CommonSize(0), CommonAlign(0), Flags(0), Index(0)
 169 {
 170   if (A)
 171     A->getSymbolList().push_back(this);
 172 }
 173
 174 /* *** */
 175
 176 MCAssembler::MCAssembler(MCContext &_Context, TargetAsmBackend &_Backend,
 177                          MCCodeEmitter &_Emitter, raw_ostream &_OS)
 178   : Context(_Context), Backend(_Backend), Emitter(_Emitter),
 179     OS(_OS), RelaxAll(false), SubsectionsViaSymbols(false)
 180 {
 181 }
 182
 183 MCAssembler::~MCAssembler() {
 184 }
 185
 186 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
 187                                                 const MCAsmFixup &Fixup,
 188                                                 const MCValue Target,
 189                                                 const MCSection *BaseSection) {
 190   // The effective fixup address is
 191   //     addr(atom(A)) + offset(A)
 192   //   - addr(atom(B)) - offset(B)
 193   //   - addr(<base symbol>) + <fixup offset from base symbol>
 194   // and the offsets are not relocatable, so the fixup is fully resolved when
 195   //  addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
 196   //
 197   // The simple (Darwin, except on x86_64) way of dealing with this was to
 198   // assume that any reference to a temporary symbol *must* be a temporary
 199   // symbol in the same atom, unless the sections differ. Therefore, any PCrel
 200   // relocation to a temporary symbol (in the same section) is fully
 201   // resolved. This also works in conjunction with absolutized .set, which
 202   // requires the compiler to use .set to absolutize the differences between
 203   // symbols which the compiler knows to be assembly time constants, so we don't
 204   // need to worry about considering symbol differences fully resolved.
 205
 206   // Non-relative fixups are only resolved if constant.
 207   if (!BaseSection)
 208     return Target.isAbsolute();
 209
 210   // Otherwise, relative fixups are only resolved if not a difference and the
 211   // target is a temporary in the same section.
 212   if (Target.isAbsolute() || Target.getSymB())
 213     return false;
 214
 215   const MCSymbol *A = &Target.getSymA()->getSymbol();
 216   if (!A->isTemporary() || !A->isInSection() ||
 217       &A->getSection() != BaseSection)
 218     return false;
 219
 220   return true;
 221 }
 222
 223 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
 224                                           const MCAsmLayout &Layout,
 225                                           const MCAsmFixup &Fixup,
 226                                           const MCValue Target,
 227                                           const MCSymbolData *BaseSymbol) {
 228   // The effective fixup address is
 229   //     addr(atom(A)) + offset(A)
 230   //   - addr(atom(B)) - offset(B)
 231   //   - addr(BaseSymbol) + <fixup offset from base symbol>
 232   // and the offsets are not relocatable, so the fixup is fully resolved when
 233   //  addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
 234   //
 235   // Note that "false" is almost always conservatively correct (it means we emit
 236   // a relocation which is unnecessary), except when it would force us to emit a
 237   // relocation which the target cannot encode.
 238
 239   const MCSymbolData *A_Base = 0, *B_Base = 0;
 240   if (const MCSymbolRefExpr *A = Target.getSymA()) {
 241     // Modified symbol references cannot be resolved.
 242     if (A->getKind() != MCSymbolRefExpr::VK_None)
 243       return false;
 244
 245     A_Base = Asm.getAtom(Layout, &Asm.getSymbolData(A->getSymbol()));
 246     if (!A_Base)
 247       return false;
 248   }
 249
 250   if (const MCSymbolRefExpr *B = Target.getSymB()) {
 251     // Modified symbol references cannot be resolved.
 252     if (B->getKind() != MCSymbolRefExpr::VK_None)
 253       return false;
 254
 255     B_Base = Asm.getAtom(Layout, &Asm.getSymbolData(B->getSymbol()));
 256     if (!B_Base)
 257       return false;
 258   }
 259
 260   // If there is no base, A and B have to be the same atom for this fixup to be
 261   // fully resolved.
 262   if (!BaseSymbol)
 263     return A_Base == B_Base;
 264
 265   // Otherwise, B must be missing and A must be the base.
 266   return !B_Base && BaseSymbol == A_Base;
 267 }
 268
 269 bool MCAssembler::isSymbolLinkerVisible(const MCSymbolData *SD) const {
 270   // Non-temporary labels should always be visible to the linker.
 271   if (!SD->getSymbol().isTemporary())
 272     return true;
 273
 274   // Absolute temporary labels are never visible.
 275   if (!SD->getFragment())
 276     return false;
 277
 278   // Otherwise, check if the section requires symbols even for temporary labels.
 279   return getBackend().doesSectionRequireSymbols(
 280     SD->getFragment()->getParent()->getSection());
 281 }
 282
 283 const MCSymbolData *MCAssembler::getAtom(const MCAsmLayout &Layout,
 284                                          const MCSymbolData *SD) const {
 285   // Linker visible symbols define atoms.
 286   if (isSymbolLinkerVisible(SD))
 287     return SD;
 288
 289   // Absolute and undefined symbols have no defining atom.
 290   if (!SD->getFragment())
 291     return 0;
 292
 293   // Non-linker visible symbols in sections which can't be atomized have no
 294   // defining atom.
 295   if (!getBackend().isSectionAtomizable(
 296         SD->getFragment()->getParent()->getSection()))
 297     return 0;
 298
 299   // Otherwise, return the atom for the containing fragment.
 300   return SD->getFragment()->getAtom();
 301 }
 302
 303 bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout,
 304                                 const MCAsmFixup &Fixup, const MCFragment *DF,
 305                                 MCValue &Target, uint64_t &Value) const {
 306   ++stats::EvaluateFixup;
 307
 308   if (!Fixup.Value->EvaluateAsRelocatable(Target, &Layout))
 309     report_fatal_error("expected relocatable expression");
 310
 311   // FIXME: How do non-scattered symbols work in ELF? I presume the linker
 312   // doesn't support small relocations, but then under what criteria does the
 313   // assembler allow symbol differences?
 314
 315   Value = Target.getConstant();
 316
 317   bool IsPCRel =
 318     Emitter.getFixupKindInfo(Fixup.Kind).Flags & MCFixupKindInfo::FKF_IsPCRel;
 319   bool IsResolved = true;
 320   if (const MCSymbolRefExpr *A = Target.getSymA()) {
 321     if (A->getSymbol().isDefined())
 322       Value += Layout.getSymbolAddress(&getSymbolData(A->getSymbol()));
 323     else
 324       IsResolved = false;
 325   }
 326   if (const MCSymbolRefExpr *B = Target.getSymB()) {
 327     if (B->getSymbol().isDefined())
 328       Value -= Layout.getSymbolAddress(&getSymbolData(B->getSymbol()));
 329     else
 330       IsResolved = false;
 331   }
 332
 333   // If we are using scattered symbols, determine whether this value is actually
 334   // resolved; scattering may cause atoms to move.
 335   if (IsResolved && getBackend().hasScatteredSymbols()) {
 336     if (getBackend().hasReliableSymbolDifference()) {
 337       // If this is a PCrel relocation, find the base atom (identified by its
 338       // symbol) that the fixup value is relative to.
 339       const MCSymbolData *BaseSymbol = 0;
 340       if (IsPCRel) {
 341         BaseSymbol = DF->getAtom();
 342         if (!BaseSymbol)
 343           IsResolved = false;
 344       }
 345
 346       if (IsResolved)
 347         IsResolved = isScatteredFixupFullyResolved(*this, Layout, Fixup, Target,
 348                                                    BaseSymbol);
 349     } else {
 350       const MCSection *BaseSection = 0;
 351       if (IsPCRel)
 352         BaseSection = &DF->getParent()->getSection();
 353
 354       IsResolved = isScatteredFixupFullyResolvedSimple(*this, Fixup, Target,
 355                                                        BaseSection);
 356     }
 357   }
 358
 359   if (IsPCRel)
 360     Value -= Layout.getFragmentAddress(DF) + Fixup.Offset;
 361
 362   return IsResolved;
 363 }
 364
 365 uint64_t MCAssembler::ComputeFragmentSize(MCAsmLayout &Layout,
 366                                           const MCFragment &F,
 367                                           uint64_t SectionAddress,
 368                                           uint64_t FragmentOffset) const {
 369   switch (F.getKind()) {
 370   case MCFragment::FT_Data:
 371     return cast<MCDataFragment>(F).getContents().size();
 372   case MCFragment::FT_Fill:
 373     return cast<MCFillFragment>(F).getSize();
 374   case MCFragment::FT_Inst:
 375     return cast<MCInstFragment>(F).getInstSize();
 376
 377   case MCFragment::FT_Align: {
 378     const MCAlignFragment &AF = cast<MCAlignFragment>(F);
 379
 380     assert((!AF.hasOnlyAlignAddress() || !AF.getNextNode()) &&
 381            "Invalid OnlyAlignAddress bit, not the last fragment!");
 382
 383     uint64_t Size = OffsetToAlignment(SectionAddress + FragmentOffset,
 384                                       AF.getAlignment());
 385
 386     // Honor MaxBytesToEmit.
 387     if (Size > AF.getMaxBytesToEmit())
 388       return 0;
 389
 390     return Size;
 391   }
 392
 393   case MCFragment::FT_Org: {
 394     const MCOrgFragment &OF = cast<MCOrgFragment>(F);
 395
 396     // FIXME: We should compute this sooner, we don't want to recurse here, and
 397     // we would like to be more functional.
 398     int64_t TargetLocation;
 399     if (!OF.getOffset().EvaluateAsAbsolute(TargetLocation, &Layout))
 400       report_fatal_error("expected assembly-time absolute expression");
 401
 402     // FIXME: We need a way to communicate this error.
 403     int64_t Offset = TargetLocation - FragmentOffset;
 404     if (Offset < 0)
 405       report_fatal_error("invalid .org offset '" + Twine(TargetLocation) +
 406                          "' (at offset '" + Twine(FragmentOffset) + "'");
 407
 408     return Offset;
 409   }
 410   }
 411
 412   assert(0 && "invalid fragment kind");
 413   return 0;
 414 }
 415
 416 void MCAsmLayout::LayoutFile() {
 417   for (unsigned i = 0, e = getSectionOrder().size(); i != e; ++i) {
 418     MCSectionData *SD = getSectionOrder()[i];
 419
 420     LayoutSection(SD);
 421     for (MCSectionData::iterator it = SD->begin(),
 422            ie = SD->end(); it != ie; ++it)
 423       LayoutFragment(it);
 424   }
 425 }
 426
 427 void MCAsmLayout::LayoutFragment(MCFragment *F) {
 428   uint64_t StartAddress = getSectionAddress(F->getParent());
 429
 430   // Get the fragment start address.
 431   uint64_t Address = StartAddress;
 432   MCSectionData::iterator it = F;
 433   if (MCFragment *Prev = F->getPrevNode())
 434     Address = (StartAddress + getFragmentOffset(Prev) +
 435                getFragmentEffectiveSize(Prev));
 436
 437   ++stats::FragmentLayouts;
 438
 439   // Compute fragment offset and size.
 440   F->Offset = Address - StartAddress;
 441   F->EffectiveSize = getAssembler().ComputeFragmentSize(*this, *F, StartAddress,
 442                                                         F->Offset);
 443 }
 444
 445 void MCAsmLayout::LayoutSection(MCSectionData *SD) {
 446   unsigned SectionOrderIndex = SD->getLayoutOrder();
 447
 448   ++stats::SectionLayouts;
 449
 450   // Compute the section start address.
 451   uint64_t StartAddress = 0;
 452   if (SectionOrderIndex) {
 453     MCSectionData *Prev = getSectionOrder()[SectionOrderIndex - 1];
 454     StartAddress = getSectionAddress(Prev) + getSectionAddressSize(Prev);
 455   }
 456
 457   // Honor the section alignment requirements.
 458   StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
 459
 460   // Set the section address.
 461   SD->Address = StartAddress;
 462 }
 463
 464 /// WriteFragmentData - Write the \arg F data to the output file.
 465 static void WriteFragmentData(const MCAssembler &Asm, const MCAsmLayout &Layout,
 466                               const MCFragment &F, MCObjectWriter *OW) {
 467   uint64_t Start = OW->getStream().tell();
 468   (void) Start;
 469
 470   ++stats::EmittedFragments;
 471
 472   // FIXME: Embed in fragments instead?
 473   uint64_t FragmentSize = Layout.getFragmentEffectiveSize(&F);
 474   switch (F.getKind()) {
 475   case MCFragment::FT_Align: {
 476     MCAlignFragment &AF = cast<MCAlignFragment>(F);
 477     uint64_t Count = FragmentSize / AF.getValueSize();
 478
 479     assert(AF.getValueSize() && "Invalid virtual align in concrete fragment!");
 480
 481     // FIXME: This error shouldn't actually occur (the front end should emit
 482     // multiple .align directives to enforce the semantics it wants), but is
 483     // severe enough that we want to report it. How to handle this?
 484     if (Count * AF.getValueSize() != FragmentSize)
 485       report_fatal_error("undefined .align directive, value size '" +
 486                         Twine(AF.getValueSize()) +
 487                         "' is not a divisor of padding size '" +
 488                         Twine(FragmentSize) + "'");
 489
 490     // See if we are aligning with nops, and if so do that first to try to fill
 491     // the Count bytes.  Then if that did not fill any bytes or there are any
 492     // bytes left to fill use the the Value and ValueSize to fill the rest.
 493     // If we are aligning with nops, ask that target to emit the right data.
 494     if (AF.hasEmitNops()) {
 495       if (!Asm.getBackend().WriteNopData(Count, OW))
 496         report_fatal_error("unable to write nop sequence of " +
 497                           Twine(Count) + " bytes");
 498       break;
 499     }
 500
 501     // Otherwise, write out in multiples of the value size.
 502     for (uint64_t i = 0; i != Count; ++i) {
 503       switch (AF.getValueSize()) {
 504       default:
 505         assert(0 && "Invalid size!");
 506       case 1: OW->Write8 (uint8_t (AF.getValue())); break;
 507       case 2: OW->Write16(uint16_t(AF.getValue())); break;
 508       case 4: OW->Write32(uint32_t(AF.getValue())); break;
 509       case 8: OW->Write64(uint64_t(AF.getValue())); break;
 510       }
 511     }
 512     break;
 513   }
 514
 515   case MCFragment::FT_Data: {
 516     MCDataFragment &DF = cast<MCDataFragment>(F);
 517     assert(FragmentSize == DF.getContents().size() && "Invalid size!");
 518     OW->WriteBytes(DF.getContents().str());
 519     break;
 520   }
 521
 522   case MCFragment::FT_Fill: {
 523     MCFillFragment &FF = cast<MCFillFragment>(F);
 524
 525     assert(FF.getValueSize() && "Invalid virtual align in concrete fragment!");
 526
 527     for (uint64_t i = 0, e = FF.getSize() / FF.getValueSize(); i != e; ++i) {
 528       switch (FF.getValueSize()) {
 529       default:
 530         assert(0 && "Invalid size!");
 531       case 1: OW->Write8 (uint8_t (FF.getValue())); break;
 532       case 2: OW->Write16(uint16_t(FF.getValue())); break;
 533       case 4: OW->Write32(uint32_t(FF.getValue())); break;
 534       case 8: OW->Write64(uint64_t(FF.getValue())); break;
 535       }
 536     }
 537     break;
 538   }
 539
 540   case MCFragment::FT_Inst:
 541     llvm_unreachable("unexpected inst fragment after lowering");
 542     break;
 543
 544   case MCFragment::FT_Org: {
 545     MCOrgFragment &OF = cast<MCOrgFragment>(F);
 546
 547     for (uint64_t i = 0, e = FragmentSize; i != e; ++i)
 548       OW->Write8(uint8_t(OF.getValue()));
 549
 550     break;
 551   }
 552   }
 553
 554   assert(OW->getStream().tell() - Start == FragmentSize);
 555 }
 556
 557 void MCAssembler::WriteSectionData(const MCSectionData *SD,
 558                                    const MCAsmLayout &Layout,
 559                                    MCObjectWriter *OW) const {
 560   // Ignore virtual sections.
 561   if (getBackend().isVirtualSection(SD->getSection())) {
 562     assert(Layout.getSectionFileSize(SD) == 0 && "Invalid size for section!");
 563
 564     // Check that contents are only things legal inside a virtual section.
 565     for (MCSectionData::const_iterator it = SD->begin(),
 566            ie = SD->end(); it != ie; ++it) {
 567       switch (it->getKind()) {
 568       default:
 569         assert(0 && "Invalid fragment in virtual section!");
 570       case MCFragment::FT_Align:
 571         assert(!cast<MCAlignFragment>(it)->getValueSize() &&
 572                "Invalid align in virtual section!");
 573         break;
 574       case MCFragment::FT_Fill:
 575         assert(!cast<MCFillFragment>(it)->getValueSize() &&
 576                "Invalid fill in virtual section!");
 577         break;
 578       }
 579     }
 580
 581     return;
 582   }
 583
 584   uint64_t Start = OW->getStream().tell();
 585   (void) Start;
 586
 587   for (MCSectionData::const_iterator it = SD->begin(),
 588          ie = SD->end(); it != ie; ++it)
 589     WriteFragmentData(*this, Layout, *it, OW);
 590
 591   assert(OW->getStream().tell() - Start == Layout.getSectionFileSize(SD));
 592 }
 593
 594 void MCAssembler::Finish() {
 595   DEBUG_WITH_TYPE("mc-dump", {
 596       llvm::errs() << "assembler backend - pre-layout\n--\n";
 597       dump(); });
 598
 599   // Create the layout object.
 600   MCAsmLayout Layout(*this);
 601
 602   // Insert additional align fragments for concrete sections to explicitly pad
 603   // the previous section to match their alignment requirements. This is for
 604   // 'gas' compatibility, it shouldn't strictly be necessary.
 605   //
 606   // FIXME: This may be Mach-O specific.
 607   for (unsigned i = 1, e = Layout.getSectionOrder().size(); i < e; ++i) {
 608     MCSectionData *SD = Layout.getSectionOrder()[i];
 609
 610     // Ignore sections without alignment requirements.
 611     unsigned Align = SD->getAlignment();
 612     if (Align <= 1)
 613       continue;
 614
 615     // Ignore virtual sections, they don't cause file size modifications.
 616     if (getBackend().isVirtualSection(SD->getSection()))
 617       continue;
 618
 619     // Otherwise, create a new align fragment at the end of the previous
 620     // section.
 621     MCAlignFragment *AF = new MCAlignFragment(Align, 0, 1, Align,
 622                                               Layout.getSectionOrder()[i - 1]);
 623     AF->setOnlyAlignAddress(true);
 624   }
 625
 626   // Create dummy fragments and assign section ordinals.
 627   unsigned SectionIndex = 0;
 628   for (MCAssembler::iterator it = begin(), ie = end(); it != ie; ++it) {
 629     // Create dummy fragments to eliminate any empty sections, this simplifies
 630     // layout.
 631     if (it->getFragmentList().empty()) {
 632       unsigned ValueSize = 1;
 633       if (getBackend().isVirtualSection(it->getSection()))
 634         ValueSize = 1;
 635       new MCFillFragment(0, 1, 0, it);
 636     }
 637
 638     it->setOrdinal(SectionIndex++);
 639   }
 640
 641   // Assign layout order indices to sections and fragments.
 642   unsigned FragmentIndex = 0;
 643   for (unsigned i = 0, e = Layout.getSectionOrder().size(); i != e; ++i) {
 644     MCSectionData *SD = Layout.getSectionOrder()[i];
 645     SD->setLayoutOrder(i);
 646
 647     for (MCSectionData::iterator it2 = SD->begin(),
 648            ie2 = SD->end(); it2 != ie2; ++it2)
 649       it2->setLayoutOrder(FragmentIndex++);
 650   }
 651
 652   // Layout until everything fits.
 653   while (LayoutOnce(Layout))
 654     continue;
 655
 656   DEBUG_WITH_TYPE("mc-dump", {
 657       llvm::errs() << "assembler backend - post-relaxation\n--\n";
 658       dump(); });
 659
 660   // Finalize the layout, including fragment lowering.
 661   FinishLayout(Layout);
 662
 663   DEBUG_WITH_TYPE("mc-dump", {
 664       llvm::errs() << "assembler backend - final-layout\n--\n";
 665       dump(); });
 666
 667   uint64_t StartOffset = OS.tell();
 668   llvm::OwningPtr<MCObjectWriter> Writer(getBackend().createObjectWriter(OS));
 669   if (!Writer)
 670     report_fatal_error("unable to create object writer!");
 671
 672   // Allow the object writer a chance to perform post-layout binding (for
 673   // example, to set the index fields in the symbol data).
 674   Writer->ExecutePostLayoutBinding(*this);
 675
 676   // Evaluate and apply the fixups, generating relocation entries as necessary.
 677   for (MCAssembler::iterator it = begin(), ie = end(); it != ie; ++it) {
 678     for (MCSectionData::iterator it2 = it->begin(),
 679            ie2 = it->end(); it2 != ie2; ++it2) {
 680       MCDataFragment *DF = dyn_cast<MCDataFragment>(it2);
 681       if (!DF)
 682         continue;
 683
 684       for (MCDataFragment::fixup_iterator it3 = DF->fixup_begin(),
 685              ie3 = DF->fixup_end(); it3 != ie3; ++it3) {
 686         MCAsmFixup &Fixup = *it3;
 687
 688         // Evaluate the fixup.
 689         MCValue Target;
 690         uint64_t FixedValue;
 691         if (!EvaluateFixup(Layout, Fixup, DF, Target, FixedValue)) {
 692           // The fixup was unresolved, we need a relocation. Inform the object
 693           // writer of the relocation, and give it an opportunity to adjust the
 694           // fixup value if need be.
 695           Writer->RecordRelocation(*this, Layout, DF, Fixup, Target,FixedValue);
 696         }
 697
 698         getBackend().ApplyFixup(Fixup, *DF, FixedValue);
 699       }
 700     }
 701   }
 702
 703   // Write the object file.
 704   Writer->WriteObject(*this, Layout);
 705   OS.flush();
 706
 707   stats::ObjectBytes += OS.tell() - StartOffset;
 708 }
 709
 710 bool MCAssembler::FixupNeedsRelaxation(const MCAsmFixup &Fixup,
 711                                        const MCFragment *DF,
 712                                        const MCAsmLayout &Layout) const {
 713   if (getRelaxAll())
 714     return true;
 715
 716   // If we cannot resolve the fixup value, it requires relaxation.
 717   MCValue Target;
 718   uint64_t Value;
 719   if (!EvaluateFixup(Layout, Fixup, DF, Target, Value))
 720     return true;
 721
 722   // Otherwise, relax if the value is too big for a (signed) i8.
 723   //
 724   // FIXME: This is target dependent!
 725   return int64_t(Value) != int64_t(int8_t(Value));
 726 }
 727
 728 bool MCAssembler::FragmentNeedsRelaxation(const MCInstFragment *IF,
 729                                           const MCAsmLayout &Layout) const {
 730   // If this inst doesn't ever need relaxation, ignore it. This occurs when we
 731   // are intentionally pushing out inst fragments, or because we relaxed a
 732   // previous instruction to one that doesn't need relaxation.
 733   if (!getBackend().MayNeedRelaxation(IF->getInst(), IF->getFixups()))
 734     return false;
 735
 736   for (MCInstFragment::const_fixup_iterator it = IF->fixup_begin(),
 737          ie = IF->fixup_end(); it != ie; ++it)
 738     if (FixupNeedsRelaxation(*it, IF, Layout))
 739       return true;
 740
 741   return false;
 742 }
 743
 744 bool MCAssembler::LayoutOnce(MCAsmLayout &Layout) {
 745   ++stats::RelaxationSteps;
 746
 747   // Layout the sections in order.
 748   Layout.LayoutFile();
 749
 750   // Scan for fragments that need relaxation.
 751   bool WasRelaxed = false;
 752   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 753     MCSectionData &SD = *it;
 754
 755     for (MCSectionData::iterator it2 = SD.begin(),
 756            ie2 = SD.end(); it2 != ie2; ++it2) {
 757       // Check if this is an instruction fragment that needs relaxation.
 758       MCInstFragment *IF = dyn_cast<MCInstFragment>(it2);
 759       if (!IF || !FragmentNeedsRelaxation(IF, Layout))
 760         continue;
 761
 762       ++stats::RelaxedInstructions;
 763
 764       // FIXME-PERF: We could immediately lower out instructions if we can tell
 765       // they are fully resolved, to avoid retesting on later passes.
 766
 767       // Relax the fragment.
 768
 769       MCInst Relaxed;
 770       getBackend().RelaxInstruction(IF, Relaxed);
 771
 772       // Encode the new instruction.
 773       //
 774       // FIXME-PERF: If it matters, we could let the target do this. It can
 775       // probably do so more efficiently in many cases.
 776       SmallVector<MCFixup, 4> Fixups;
 777       SmallString<256> Code;
 778       raw_svector_ostream VecOS(Code);
 779       getEmitter().EncodeInstruction(Relaxed, VecOS, Fixups);
 780       VecOS.flush();
 781
 782       // Update the instruction fragment.
 783       int SlideAmount = Code.size() - IF->getInstSize();
 784       IF->setInst(Relaxed);
 785       IF->getCode() = Code;
 786       IF->getFixups().clear();
 787       for (unsigned i = 0, e = Fixups.size(); i != e; ++i) {
 788         MCFixup &F = Fixups[i];
 789         IF->getFixups().push_back(MCAsmFixup(F.getOffset(), *F.getValue(),
 790                                              F.getKind()));
 791       }
 792
 793       // Update the layout, and remember that we relaxed. If we are relaxing
 794       // everything, we can skip this step since nothing will depend on updating
 795       // the values.
 796       if (!getRelaxAll())
 797         Layout.UpdateForSlide(IF, SlideAmount);
 798       WasRelaxed = true;
 799     }
 800   }
 801
 802   return WasRelaxed;
 803 }
 804
 805 void MCAssembler::FinishLayout(MCAsmLayout &Layout) {
 806   // Lower out any instruction fragments, to simplify the fixup application and
 807   // output.
 808   //
 809   // FIXME-PERF: We don't have to do this, but the assumption is that it is
 810   // cheap (we will mostly end up eliminating fragments and appending on to data
 811   // fragments), so the extra complexity downstream isn't worth it. Evaluate
 812   // this assumption.
 813   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 814     MCSectionData &SD = *it;
 815
 816     for (MCSectionData::iterator it2 = SD.begin(),
 817            ie2 = SD.end(); it2 != ie2; ++it2) {
 818       MCInstFragment *IF = dyn_cast<MCInstFragment>(it2);
 819       if (!IF)
 820         continue;
 821
 822       // Create a new data fragment for the instruction.
 823       //
 824       // FIXME-PERF: Reuse previous data fragment if possible.
 825       MCDataFragment *DF = new MCDataFragment();
 826       SD.getFragmentList().insert(it2, DF);
 827
 828       // Update the data fragments layout data.
 829       DF->setParent(IF->getParent());
 830       DF->setAtom(IF->getAtom());
 831       DF->setLayoutOrder(IF->getLayoutOrder());
 832       Layout.FragmentReplaced(IF, DF);
 833
 834       // Copy in the data and the fixups.
 835       DF->getContents().append(IF->getCode().begin(), IF->getCode().end());
 836       for (unsigned i = 0, e = IF->getFixups().size(); i != e; ++i)
 837         DF->getFixups().push_back(IF->getFixups()[i]);
 838
 839       // Delete the instruction fragment and update the iterator.
 840       SD.getFragmentList().erase(IF);
 841       it2 = DF;
 842     }
 843   }
 844 }
 845
 846 // Debugging methods
 847
 848 namespace llvm {
 849
 850 raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) {
 851   OS << "<MCAsmFixup" << " Offset:" << AF.Offset << " Value:" << *AF.Value
 852      << " Kind:" << AF.Kind << ">";
 853   return OS;
 854 }
 855
 856 }
 857
 858 void MCFragment::dump() {
 859   raw_ostream &OS = llvm::errs();
 860
 861   OS << "<MCFragment " << (void*) this << " LayoutOrder:" << LayoutOrder
 862      << " Offset:" << Offset << " EffectiveSize:" << EffectiveSize << ">";
 863 }
 864
 865 void MCAlignFragment::dump() {
 866   raw_ostream &OS = llvm::errs();
 867
 868   OS << "<MCAlignFragment ";
 869   this->MCFragment::dump();
 870   if (hasEmitNops())
 871     OS << " (emit nops)";
 872   if (hasOnlyAlignAddress())
 873     OS << " (only align section)";
 874   OS << "\n       ";
 875   OS << " Alignment:" << getAlignment()
 876      << " Value:" << getValue() << " ValueSize:" << getValueSize()
 877      << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">";
 878 }
 879
 880 void MCDataFragment::dump() {
 881   raw_ostream &OS = llvm::errs();
 882
 883   OS << "<MCDataFragment ";
 884   this->MCFragment::dump();
 885   OS << "\n       ";
 886   OS << " Contents:[";
 887   for (unsigned i = 0, e = getContents().size(); i != e; ++i) {
 888     if (i) OS << ",";
 889     OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF);
 890   }
 891   OS << "] (" << getContents().size() << " bytes)";
 892
 893   if (!getFixups().empty()) {
 894     OS << ",\n       ";
 895     OS << " Fixups:[";
 896     for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) {
 897       if (it != fixup_begin()) OS << ",\n                ";
 898       OS << *it;
 899     }
 900     OS << "]";
 901   }
 902
 903   OS << ">";
 904 }
 905
 906 void MCFillFragment::dump() {
 907   raw_ostream &OS = llvm::errs();
 908
 909   OS << "<MCFillFragment ";
 910   this->MCFragment::dump();
 911   OS << "\n       ";
 912   OS << " Value:" << getValue() << " ValueSize:" << getValueSize()
 913      << " Size:" << getSize() << ">";
 914 }
 915
 916 void MCInstFragment::dump() {
 917   raw_ostream &OS = llvm::errs();
 918
 919   OS << "<MCInstFragment ";
 920   this->MCFragment::dump();
 921   OS << "\n       ";
 922   OS << " Inst:";
 923   getInst().dump_pretty(OS);
 924   OS << ">";
 925 }
 926
 927 void MCOrgFragment::dump() {
 928   raw_ostream &OS = llvm::errs();
 929
 930   OS << "<MCOrgFragment ";
 931   this->MCFragment::dump();
 932   OS << "\n       ";
 933   OS << " Offset:" << getOffset() << " Value:" << getValue() << ">";
 934 }
 935
 936 void MCSectionData::dump() {
 937   raw_ostream &OS = llvm::errs();
 938
 939   OS << "<MCSectionData";
 940   OS << " Alignment:" << getAlignment() << " Address:" << Address
 941      << " Fragments:[\n      ";
 942   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 943     if (it != begin()) OS << ",\n      ";
 944     it->dump();
 945   }
 946   OS << "]>";
 947 }
 948
 949 void MCSymbolData::dump() {
 950   raw_ostream &OS = llvm::errs();
 951
 952   OS << "<MCSymbolData Symbol:" << getSymbol()
 953      << " Fragment:" << getFragment() << " Offset:" << getOffset()
 954      << " Flags:" << getFlags() << " Index:" << getIndex();
 955   if (isCommon())
 956     OS << " (common, size:" << getCommonSize()
 957        << " align: " << getCommonAlignment() << ")";
 958   if (isExternal())
 959     OS << " (external)";
 960   if (isPrivateExtern())
 961     OS << " (private extern)";
 962   OS << ">";
 963 }
 964
 965 void MCAssembler::dump() {
 966   raw_ostream &OS = llvm::errs();
 967
 968   OS << "<MCAssembler\n";
 969   OS << "  Sections:[\n    ";
 970   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 971     if (it != begin()) OS << ",\n    ";
 972     it->dump();
 973   }
 974   OS << "],\n";
 975   OS << "  Symbols:[";
 976
 977   for (symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
 978     if (it != symbol_begin()) OS << ",\n           ";
 979     it->dump();
 980   }
 981   OS << "]>\n";
 982 }