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   for (unsigned i = 0, e = getSectionOrder().size(); i != e; ++i)
  73     getAssembler().LayoutSection(*this, i);
  74 }
  75
  76 void MCAsmLayout::FragmentReplaced(MCFragment *Src, MCFragment *Dst) {
  77   Dst->Offset = Src->Offset;
  78   Dst->EffectiveSize = Src->EffectiveSize;
  79 }
  80
  81 uint64_t MCAsmLayout::getFragmentAddress(const MCFragment *F) const {
  82   assert(F->getParent() && "Missing section()!");
  83   return getSectionAddress(F->getParent()) + getFragmentOffset(F);
  84 }
  85
  86 uint64_t MCAsmLayout::getFragmentEffectiveSize(const MCFragment *F) const {
  87   assert(F->EffectiveSize != ~UINT64_C(0) && "Address not set!");
  88   return F->EffectiveSize;
  89 }
  90
  91 void MCAsmLayout::setFragmentEffectiveSize(MCFragment *F, uint64_t Value) {
  92   F->EffectiveSize = Value;
  93 }
  94
  95 uint64_t MCAsmLayout::getFragmentOffset(const MCFragment *F) const {
  96   assert(F->Offset != ~UINT64_C(0) && "Address not set!");
  97   return F->Offset;
  98 }
  99
 100 void MCAsmLayout::setFragmentOffset(MCFragment *F, uint64_t Value) {
 101   F->Offset = Value;
 102 }
 103
 104 uint64_t MCAsmLayout::getSymbolAddress(const MCSymbolData *SD) const {
 105   assert(SD->getFragment() && "Invalid getAddress() on undefined symbol!");
 106   return getFragmentAddress(SD->getFragment()) + SD->getOffset();
 107 }
 108
 109 uint64_t MCAsmLayout::getSectionAddress(const MCSectionData *SD) const {
 110   assert(SD->Address != ~UINT64_C(0) && "Address not set!");
 111   return SD->Address;
 112 }
 113
 114 void MCAsmLayout::setSectionAddress(MCSectionData *SD, uint64_t Value) {
 115   SD->Address = Value;
 116 }
 117
 118 uint64_t MCAsmLayout::getSectionAddressSize(const MCSectionData *SD) const {
 119   // Otherwise, the size is the last fragment's end offset.
 120   const MCFragment &F = SD->getFragmentList().back();
 121   return getFragmentOffset(&F) + getFragmentEffectiveSize(&F);
 122 }
 123
 124 uint64_t MCAsmLayout::getSectionFileSize(const MCSectionData *SD) const {
 125   // Virtual sections have no file size.
 126   if (getAssembler().getBackend().isVirtualSection(SD->getSection()))
 127     return 0;
 128
 129   // Otherwise, the file size is the same as the address space size.
 130   return getSectionAddressSize(SD);
 131 }
 132
 133 uint64_t MCAsmLayout::getSectionSize(const MCSectionData *SD) const {
 134   // The logical size is the address space size minus any tail padding.
 135   uint64_t Size = getSectionAddressSize(SD);
 136   const MCAlignFragment *AF =
 137     dyn_cast<MCAlignFragment>(&(SD->getFragmentList().back()));
 138   if (AF && AF->hasOnlyAlignAddress())
 139     Size -= getFragmentEffectiveSize(AF);
 140
 141   return Size;
 142 }
 143
 144 /* *** */
 145
 146 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
 147 }
 148
 149 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent)
 150   : Kind(_Kind), Parent(_Parent), Atom(0), EffectiveSize(~UINT64_C(0))
 151 {
 152   if (Parent)
 153     Parent->getFragmentList().push_back(this);
 154 }
 155
 156 MCFragment::~MCFragment() {
 157 }
 158
 159 /* *** */
 160
 161 MCSectionData::MCSectionData() : Section(0) {}
 162
 163 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
 164   : Section(&_Section),
 165     Alignment(1),
 166     Address(~UINT64_C(0)),
 167     HasInstructions(false)
 168 {
 169   if (A)
 170     A->getSectionList().push_back(this);
 171 }
 172
 173 /* *** */
 174
 175 MCSymbolData::MCSymbolData() : Symbol(0) {}
 176
 177 MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment,
 178                            uint64_t _Offset, MCAssembler *A)
 179   : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset),
 180     IsExternal(false), IsPrivateExtern(false),
 181     CommonSize(0), CommonAlign(0), Flags(0), Index(0)
 182 {
 183   if (A)
 184     A->getSymbolList().push_back(this);
 185 }
 186
 187 /* *** */
 188
 189 MCAssembler::MCAssembler(MCContext &_Context, TargetAsmBackend &_Backend,
 190                          MCCodeEmitter &_Emitter, raw_ostream &_OS)
 191   : Context(_Context), Backend(_Backend), Emitter(_Emitter),
 192     OS(_OS), RelaxAll(false), SubsectionsViaSymbols(false)
 193 {
 194 }
 195
 196 MCAssembler::~MCAssembler() {
 197 }
 198
 199 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
 200                                                 const MCAsmFixup &Fixup,
 201                                                 const MCValue Target,
 202                                                 const MCSection *BaseSection) {
 203   // The effective fixup address is
 204   //     addr(atom(A)) + offset(A)
 205   //   - addr(atom(B)) - offset(B)
 206   //   - addr(<base symbol>) + <fixup offset from base symbol>
 207   // and the offsets are not relocatable, so the fixup is fully resolved when
 208   //  addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
 209   //
 210   // The simple (Darwin, except on x86_64) way of dealing with this was to
 211   // assume that any reference to a temporary symbol *must* be a temporary
 212   // symbol in the same atom, unless the sections differ. Therefore, any PCrel
 213   // relocation to a temporary symbol (in the same section) is fully
 214   // resolved. This also works in conjunction with absolutized .set, which
 215   // requires the compiler to use .set to absolutize the differences between
 216   // symbols which the compiler knows to be assembly time constants, so we don't
 217   // need to worry about considering symbol differences fully resolved.
 218
 219   // Non-relative fixups are only resolved if constant.
 220   if (!BaseSection)
 221     return Target.isAbsolute();
 222
 223   // Otherwise, relative fixups are only resolved if not a difference and the
 224   // target is a temporary in the same section.
 225   if (Target.isAbsolute() || Target.getSymB())
 226     return false;
 227
 228   const MCSymbol *A = &Target.getSymA()->getSymbol();
 229   if (!A->isTemporary() || !A->isInSection() ||
 230       &A->getSection() != BaseSection)
 231     return false;
 232
 233   return true;
 234 }
 235
 236 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
 237                                           const MCAsmLayout &Layout,
 238                                           const MCAsmFixup &Fixup,
 239                                           const MCValue Target,
 240                                           const MCSymbolData *BaseSymbol) {
 241   // The effective fixup address is
 242   //     addr(atom(A)) + offset(A)
 243   //   - addr(atom(B)) - offset(B)
 244   //   - addr(BaseSymbol) + <fixup offset from base symbol>
 245   // and the offsets are not relocatable, so the fixup is fully resolved when
 246   //  addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
 247   //
 248   // Note that "false" is almost always conservatively correct (it means we emit
 249   // a relocation which is unnecessary), except when it would force us to emit a
 250   // relocation which the target cannot encode.
 251
 252   const MCSymbolData *A_Base = 0, *B_Base = 0;
 253   if (const MCSymbolRefExpr *A = Target.getSymA()) {
 254     // Modified symbol references cannot be resolved.
 255     if (A->getKind() != MCSymbolRefExpr::VK_None)
 256       return false;
 257
 258     A_Base = Asm.getAtom(Layout, &Asm.getSymbolData(A->getSymbol()));
 259     if (!A_Base)
 260       return false;
 261   }
 262
 263   if (const MCSymbolRefExpr *B = Target.getSymB()) {
 264     // Modified symbol references cannot be resolved.
 265     if (B->getKind() != MCSymbolRefExpr::VK_None)
 266       return false;
 267
 268     B_Base = Asm.getAtom(Layout, &Asm.getSymbolData(B->getSymbol()));
 269     if (!B_Base)
 270       return false;
 271   }
 272
 273   // If there is no base, A and B have to be the same atom for this fixup to be
 274   // fully resolved.
 275   if (!BaseSymbol)
 276     return A_Base == B_Base;
 277
 278   // Otherwise, B must be missing and A must be the base.
 279   return !B_Base && BaseSymbol == A_Base;
 280 }
 281
 282 bool MCAssembler::isSymbolLinkerVisible(const MCSymbolData *SD) const {
 283   // Non-temporary labels should always be visible to the linker.
 284   if (!SD->getSymbol().isTemporary())
 285     return true;
 286
 287   // Absolute temporary labels are never visible.
 288   if (!SD->getFragment())
 289     return false;
 290
 291   // Otherwise, check if the section requires symbols even for temporary labels.
 292   return getBackend().doesSectionRequireSymbols(
 293     SD->getFragment()->getParent()->getSection());
 294 }
 295
 296 const MCSymbolData *MCAssembler::getAtom(const MCAsmLayout &Layout,
 297                                          const MCSymbolData *SD) const {
 298   // Linker visible symbols define atoms.
 299   if (isSymbolLinkerVisible(SD))
 300     return SD;
 301
 302   // Absolute and undefined symbols have no defining atom.
 303   if (!SD->getFragment())
 304     return 0;
 305
 306   // Non-linker visible symbols in sections which can't be atomized have no
 307   // defining atom.
 308   if (!getBackend().isSectionAtomizable(
 309         SD->getFragment()->getParent()->getSection()))
 310     return 0;
 311
 312   // Otherwise, return the atom for the containing fragment.
 313   return SD->getFragment()->getAtom();
 314 }
 315
 316 bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout,
 317                                 const MCAsmFixup &Fixup, const MCFragment *DF,
 318                                 MCValue &Target, uint64_t &Value) const {
 319   ++stats::EvaluateFixup;
 320
 321   if (!Fixup.Value->EvaluateAsRelocatable(Target, &Layout))
 322     report_fatal_error("expected relocatable expression");
 323
 324   // FIXME: How do non-scattered symbols work in ELF? I presume the linker
 325   // doesn't support small relocations, but then under what criteria does the
 326   // assembler allow symbol differences?
 327
 328   Value = Target.getConstant();
 329
 330   bool IsPCRel =
 331     Emitter.getFixupKindInfo(Fixup.Kind).Flags & MCFixupKindInfo::FKF_IsPCRel;
 332   bool IsResolved = true;
 333   if (const MCSymbolRefExpr *A = Target.getSymA()) {
 334     if (A->getSymbol().isDefined())
 335       Value += Layout.getSymbolAddress(&getSymbolData(A->getSymbol()));
 336     else
 337       IsResolved = false;
 338   }
 339   if (const MCSymbolRefExpr *B = Target.getSymB()) {
 340     if (B->getSymbol().isDefined())
 341       Value -= Layout.getSymbolAddress(&getSymbolData(B->getSymbol()));
 342     else
 343       IsResolved = false;
 344   }
 345
 346   // If we are using scattered symbols, determine whether this value is actually
 347   // resolved; scattering may cause atoms to move.
 348   if (IsResolved && getBackend().hasScatteredSymbols()) {
 349     if (getBackend().hasReliableSymbolDifference()) {
 350       // If this is a PCrel relocation, find the base atom (identified by its
 351       // symbol) that the fixup value is relative to.
 352       const MCSymbolData *BaseSymbol = 0;
 353       if (IsPCRel) {
 354         BaseSymbol = DF->getAtom();
 355         if (!BaseSymbol)
 356           IsResolved = false;
 357       }
 358
 359       if (IsResolved)
 360         IsResolved = isScatteredFixupFullyResolved(*this, Layout, Fixup, Target,
 361                                                    BaseSymbol);
 362     } else {
 363       const MCSection *BaseSection = 0;
 364       if (IsPCRel)
 365         BaseSection = &DF->getParent()->getSection();
 366
 367       IsResolved = isScatteredFixupFullyResolvedSimple(*this, Fixup, Target,
 368                                                        BaseSection);
 369     }
 370   }
 371
 372   if (IsPCRel)
 373     Value -= Layout.getFragmentAddress(DF) + Fixup.Offset;
 374
 375   return IsResolved;
 376 }
 377
 378 void MCAssembler::LayoutFragment(MCAsmLayout &Layout, MCFragment &F) {
 379   uint64_t StartAddress = Layout.getSectionAddress(F.getParent());
 380
 381   // Get the fragment start address.
 382   uint64_t Address = StartAddress;
 383   MCSectionData::iterator it = &F;
 384   if (MCFragment *Prev = F.getPrevNode())
 385     Address = (StartAddress + Layout.getFragmentOffset(Prev) +
 386                Layout.getFragmentEffectiveSize(Prev));
 387
 388   ++stats::FragmentLayouts;
 389
 390   uint64_t FragmentOffset = Address - StartAddress;
 391   Layout.setFragmentOffset(&F, FragmentOffset);
 392
 393   // Evaluate fragment size.
 394   uint64_t EffectiveSize = 0;
 395   switch (F.getKind()) {
 396   case MCFragment::FT_Align: {
 397     MCAlignFragment &AF = cast<MCAlignFragment>(F);
 398
 399     assert((!AF.hasOnlyAlignAddress() || !AF.getNextNode()) &&
 400            "Invalid OnlyAlignAddress bit, not the last fragment!");
 401
 402     EffectiveSize = OffsetToAlignment(Address, AF.getAlignment());
 403     if (EffectiveSize > AF.getMaxBytesToEmit())
 404       EffectiveSize = 0;
 405     break;
 406   }
 407
 408   case MCFragment::FT_Data:
 409     EffectiveSize = cast<MCDataFragment>(F).getContents().size();
 410     break;
 411
 412   case MCFragment::FT_Fill: {
 413     EffectiveSize = cast<MCFillFragment>(F).getSize();
 414     break;
 415   }
 416
 417   case MCFragment::FT_Inst:
 418     EffectiveSize = cast<MCInstFragment>(F).getInstSize();
 419     break;
 420
 421   case MCFragment::FT_Org: {
 422     MCOrgFragment &OF = cast<MCOrgFragment>(F);
 423
 424     int64_t TargetLocation;
 425     if (!OF.getOffset().EvaluateAsAbsolute(TargetLocation, &Layout))
 426       report_fatal_error("expected assembly-time absolute expression");
 427
 428     // FIXME: We need a way to communicate this error.
 429     int64_t Offset = TargetLocation - FragmentOffset;
 430     if (Offset < 0)
 431       report_fatal_error("invalid .org offset '" + Twine(TargetLocation) +
 432                          "' (at offset '" + Twine(FragmentOffset) + "'");
 433
 434     EffectiveSize = Offset;
 435     break;
 436   }
 437   }
 438
 439   Layout.setFragmentEffectiveSize(&F, EffectiveSize);
 440 }
 441
 442 void MCAssembler::LayoutSection(MCAsmLayout &Layout,
 443                                 unsigned SectionOrderIndex) {
 444   MCSectionData &SD = *Layout.getSectionOrder()[SectionOrderIndex];
 445
 446   ++stats::SectionLayouts;
 447
 448   // Compute the section start address.
 449   uint64_t StartAddress = 0;
 450   if (SectionOrderIndex) {
 451     MCSectionData *Prev = Layout.getSectionOrder()[SectionOrderIndex - 1];
 452     StartAddress = (Layout.getSectionAddress(Prev) +
 453                     Layout.getSectionAddressSize(Prev));
 454   }
 455
 456   // Honor the section alignment requirements.
 457   StartAddress = RoundUpToAlignment(StartAddress, SD.getAlignment());
 458
 459   // Set the section address.
 460   Layout.setSectionAddress(&SD, StartAddress);
 461
 462   for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it)
 463     LayoutFragment(Layout, *it);
 464 }
 465
 466 /// WriteFragmentData - Write the \arg F data to the output file.
 467 static void WriteFragmentData(const MCAssembler &Asm, const MCAsmLayout &Layout,
 468                               const MCFragment &F, MCObjectWriter *OW) {
 469   uint64_t Start = OW->getStream().tell();
 470   (void) Start;
 471
 472   ++stats::EmittedFragments;
 473
 474   // FIXME: Embed in fragments instead?
 475   uint64_t FragmentSize = Layout.getFragmentEffectiveSize(&F);
 476   switch (F.getKind()) {
 477   case MCFragment::FT_Align: {
 478     MCAlignFragment &AF = cast<MCAlignFragment>(F);
 479     uint64_t Count = FragmentSize / AF.getValueSize();
 480
 481     assert(AF.getValueSize() && "Invalid virtual align in concrete fragment!");
 482
 483     // FIXME: This error shouldn't actually occur (the front end should emit
 484     // multiple .align directives to enforce the semantics it wants), but is
 485     // severe enough that we want to report it. How to handle this?
 486     if (Count * AF.getValueSize() != FragmentSize)
 487       report_fatal_error("undefined .align directive, value size '" +
 488                         Twine(AF.getValueSize()) +
 489                         "' is not a divisor of padding size '" +
 490                         Twine(FragmentSize) + "'");
 491
 492     // See if we are aligning with nops, and if so do that first to try to fill
 493     // the Count bytes.  Then if that did not fill any bytes or there are any
 494     // bytes left to fill use the the Value and ValueSize to fill the rest.
 495     // If we are aligning with nops, ask that target to emit the right data.
 496     if (AF.hasEmitNops()) {
 497       if (!Asm.getBackend().WriteNopData(Count, OW))
 498         report_fatal_error("unable to write nop sequence of " +
 499                           Twine(Count) + " bytes");
 500       break;
 501     }
 502
 503     // Otherwise, write out in multiples of the value size.
 504     for (uint64_t i = 0; i != Count; ++i) {
 505       switch (AF.getValueSize()) {
 506       default:
 507         assert(0 && "Invalid size!");
 508       case 1: OW->Write8 (uint8_t (AF.getValue())); break;
 509       case 2: OW->Write16(uint16_t(AF.getValue())); break;
 510       case 4: OW->Write32(uint32_t(AF.getValue())); break;
 511       case 8: OW->Write64(uint64_t(AF.getValue())); break;
 512       }
 513     }
 514     break;
 515   }
 516
 517   case MCFragment::FT_Data: {
 518     MCDataFragment &DF = cast<MCDataFragment>(F);
 519     assert(FragmentSize == DF.getContents().size() && "Invalid size!");
 520     OW->WriteBytes(DF.getContents().str());
 521     break;
 522   }
 523
 524   case MCFragment::FT_Fill: {
 525     MCFillFragment &FF = cast<MCFillFragment>(F);
 526
 527     assert(FF.getValueSize() && "Invalid virtual align in concrete fragment!");
 528
 529     for (uint64_t i = 0, e = FF.getSize() / FF.getValueSize(); i != e; ++i) {
 530       switch (FF.getValueSize()) {
 531       default:
 532         assert(0 && "Invalid size!");
 533       case 1: OW->Write8 (uint8_t (FF.getValue())); break;
 534       case 2: OW->Write16(uint16_t(FF.getValue())); break;
 535       case 4: OW->Write32(uint32_t(FF.getValue())); break;
 536       case 8: OW->Write64(uint64_t(FF.getValue())); break;
 537       }
 538     }
 539     break;
 540   }
 541
 542   case MCFragment::FT_Inst:
 543     llvm_unreachable("unexpected inst fragment after lowering");
 544     break;
 545
 546   case MCFragment::FT_Org: {
 547     MCOrgFragment &OF = cast<MCOrgFragment>(F);
 548
 549     for (uint64_t i = 0, e = FragmentSize; i != e; ++i)
 550       OW->Write8(uint8_t(OF.getValue()));
 551
 552     break;
 553   }
 554   }
 555
 556   assert(OW->getStream().tell() - Start == FragmentSize);
 557 }
 558
 559 void MCAssembler::WriteSectionData(const MCSectionData *SD,
 560                                    const MCAsmLayout &Layout,
 561                                    MCObjectWriter *OW) const {
 562   // Ignore virtual sections.
 563   if (getBackend().isVirtualSection(SD->getSection())) {
 564     assert(Layout.getSectionFileSize(SD) == 0 && "Invalid size for section!");
 565
 566     // Check that contents are only things legal inside a virtual section.
 567     for (MCSectionData::const_iterator it = SD->begin(),
 568            ie = SD->end(); it != ie; ++it) {
 569       switch (it->getKind()) {
 570       default:
 571         assert(0 && "Invalid fragment in virtual section!");
 572       case MCFragment::FT_Align:
 573         assert(!cast<MCAlignFragment>(it)->getValueSize() &&
 574                "Invalid align in virtual section!");
 575         break;
 576       case MCFragment::FT_Fill:
 577         assert(!cast<MCFillFragment>(it)->getValueSize() &&
 578                "Invalid fill in virtual section!");
 579         break;
 580       }
 581     }
 582
 583     return;
 584   }
 585
 586   uint64_t Start = OW->getStream().tell();
 587   (void) Start;
 588
 589   for (MCSectionData::const_iterator it = SD->begin(),
 590          ie = SD->end(); it != ie; ++it)
 591     WriteFragmentData(*this, Layout, *it, OW);
 592
 593   assert(OW->getStream().tell() - Start == Layout.getSectionFileSize(SD));
 594 }
 595
 596 void MCAssembler::Finish() {
 597   DEBUG_WITH_TYPE("mc-dump", {
 598       llvm::errs() << "assembler backend - pre-layout\n--\n";
 599       dump(); });
 600
 601   // Create the layout object.
 602   MCAsmLayout Layout(*this);
 603
 604   // Assign layout order indices.
 605   for (unsigned i = 0, e = Layout.getSectionOrder().size(); i != e; ++i)
 606     Layout.getSectionOrder()[i]->setLayoutOrder(i);
 607
 608   // Insert additional align fragments for concrete sections to explicitly pad
 609   // the previous section to match their alignment requirements. This is for
 610   // 'gas' compatibility, it shouldn't strictly be necessary.
 611   //
 612   // FIXME: This may be Mach-O specific.
 613   for (unsigned i = 1, e = Layout.getSectionOrder().size(); i < e; ++i) {
 614     MCSectionData *SD = Layout.getSectionOrder()[i];
 615
 616     // Ignore sections without alignment requirements.
 617     unsigned Align = SD->getAlignment();
 618     if (Align <= 1)
 619       continue;
 620
 621     // Ignore virtual sections, they don't cause file size modifications.
 622     if (getBackend().isVirtualSection(SD->getSection()))
 623       continue;
 624
 625     // Otherwise, create a new align fragment at the end of the previous
 626     // section.
 627     MCAlignFragment *AF = new MCAlignFragment(Align, 0, 1, Align,
 628                                               Layout.getSectionOrder()[i - 1]);
 629     AF->setOnlyAlignAddress(true);
 630   }
 631
 632   // Assign section and fragment ordinals, all subsequent backend code is
 633   // responsible for updating these in place.
 634   unsigned SectionIndex = 0;
 635   unsigned FragmentIndex = 0;
 636   for (MCAssembler::iterator it = begin(), ie = end(); it != ie; ++it) {
 637     // Create dummy fragments to eliminate any empty sections, this simplifies
 638     // layout.
 639     if (it->getFragmentList().empty()) {
 640       unsigned ValueSize = 1;
 641       if (getBackend().isVirtualSection(it->getSection()))
 642         ValueSize = 1;
 643       new MCFillFragment(0, 1, 0, it);
 644     }
 645
 646     it->setOrdinal(SectionIndex++);
 647
 648     for (MCSectionData::iterator it2 = it->begin(),
 649            ie2 = it->end(); it2 != ie2; ++it2)
 650       it2->setOrdinal(FragmentIndex++);
 651   }
 652
 653   // Layout until everything fits.
 654   while (LayoutOnce(Layout))
 655     continue;
 656
 657   DEBUG_WITH_TYPE("mc-dump", {
 658       llvm::errs() << "assembler backend - post-relaxation\n--\n";
 659       dump(); });
 660
 661   // Finalize the layout, including fragment lowering.
 662   FinishLayout(Layout);
 663
 664   DEBUG_WITH_TYPE("mc-dump", {
 665       llvm::errs() << "assembler backend - final-layout\n--\n";
 666       dump(); });
 667
 668   uint64_t StartOffset = OS.tell();
 669   llvm::OwningPtr<MCObjectWriter> Writer(getBackend().createObjectWriter(OS));
 670   if (!Writer)
 671     report_fatal_error("unable to create object writer!");
 672
 673   // Allow the object writer a chance to perform post-layout binding (for
 674   // example, to set the index fields in the symbol data).
 675   Writer->ExecutePostLayoutBinding(*this);
 676
 677   // Evaluate and apply the fixups, generating relocation entries as necessary.
 678   for (MCAssembler::iterator it = begin(), ie = end(); it != ie; ++it) {
 679     for (MCSectionData::iterator it2 = it->begin(),
 680            ie2 = it->end(); it2 != ie2; ++it2) {
 681       MCDataFragment *DF = dyn_cast<MCDataFragment>(it2);
 682       if (!DF)
 683         continue;
 684
 685       for (MCDataFragment::fixup_iterator it3 = DF->fixup_begin(),
 686              ie3 = DF->fixup_end(); it3 != ie3; ++it3) {
 687         MCAsmFixup &Fixup = *it3;
 688
 689         // Evaluate the fixup.
 690         MCValue Target;
 691         uint64_t FixedValue;
 692         if (!EvaluateFixup(Layout, Fixup, DF, Target, FixedValue)) {
 693           // The fixup was unresolved, we need a relocation. Inform the object
 694           // writer of the relocation, and give it an opportunity to adjust the
 695           // fixup value if need be.
 696           Writer->RecordRelocation(*this, Layout, DF, Fixup, Target,FixedValue);
 697         }
 698
 699         getBackend().ApplyFixup(Fixup, *DF, FixedValue);
 700       }
 701     }
 702   }
 703
 704   // Write the object file.
 705   Writer->WriteObject(*this, Layout);
 706   OS.flush();
 707
 708   stats::ObjectBytes += OS.tell() - StartOffset;
 709 }
 710
 711 bool MCAssembler::FixupNeedsRelaxation(const MCAsmFixup &Fixup,
 712                                        const MCFragment *DF,
 713                                        const MCAsmLayout &Layout) const {
 714   if (getRelaxAll())
 715     return true;
 716
 717   // If we cannot resolve the fixup value, it requires relaxation.
 718   MCValue Target;
 719   uint64_t Value;
 720   if (!EvaluateFixup(Layout, Fixup, DF, Target, Value))
 721     return true;
 722
 723   // Otherwise, relax if the value is too big for a (signed) i8.
 724   //
 725   // FIXME: This is target dependent!
 726   return int64_t(Value) != int64_t(int8_t(Value));
 727 }
 728
 729 bool MCAssembler::FragmentNeedsRelaxation(const MCInstFragment *IF,
 730                                           const MCAsmLayout &Layout) const {
 731   // If this inst doesn't ever need relaxation, ignore it. This occurs when we
 732   // are intentionally pushing out inst fragments, or because we relaxed a
 733   // previous instruction to one that doesn't need relaxation.
 734   if (!getBackend().MayNeedRelaxation(IF->getInst(), IF->getFixups()))
 735     return false;
 736
 737   for (MCInstFragment::const_fixup_iterator it = IF->fixup_begin(),
 738          ie = IF->fixup_end(); it != ie; ++it)
 739     if (FixupNeedsRelaxation(*it, IF, Layout))
 740       return true;
 741
 742   return false;
 743 }
 744
 745 bool MCAssembler::LayoutOnce(MCAsmLayout &Layout) {
 746   ++stats::RelaxationSteps;
 747
 748   // Layout the sections in order.
 749   for (unsigned i = 0, e = Layout.getSectionOrder().size(); i != e; ++i)
 750     LayoutSection(Layout, i);
 751
 752   // Scan for fragments that need relaxation.
 753   bool WasRelaxed = false;
 754   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 755     MCSectionData &SD = *it;
 756
 757     for (MCSectionData::iterator it2 = SD.begin(),
 758            ie2 = SD.end(); it2 != ie2; ++it2) {
 759       // Check if this is an instruction fragment that needs relaxation.
 760       MCInstFragment *IF = dyn_cast<MCInstFragment>(it2);
 761       if (!IF || !FragmentNeedsRelaxation(IF, Layout))
 762         continue;
 763
 764       ++stats::RelaxedInstructions;
 765
 766       // FIXME-PERF: We could immediately lower out instructions if we can tell
 767       // they are fully resolved, to avoid retesting on later passes.
 768
 769       // Relax the fragment.
 770
 771       MCInst Relaxed;
 772       getBackend().RelaxInstruction(IF, Relaxed);
 773
 774       // Encode the new instruction.
 775       //
 776       // FIXME-PERF: If it matters, we could let the target do this. It can
 777       // probably do so more efficiently in many cases.
 778       SmallVector<MCFixup, 4> Fixups;
 779       SmallString<256> Code;
 780       raw_svector_ostream VecOS(Code);
 781       getEmitter().EncodeInstruction(Relaxed, VecOS, Fixups);
 782       VecOS.flush();
 783
 784       // Update the instruction fragment.
 785       int SlideAmount = Code.size() - IF->getInstSize();
 786       IF->setInst(Relaxed);
 787       IF->getCode() = Code;
 788       IF->getFixups().clear();
 789       for (unsigned i = 0, e = Fixups.size(); i != e; ++i) {
 790         MCFixup &F = Fixups[i];
 791         IF->getFixups().push_back(MCAsmFixup(F.getOffset(), *F.getValue(),
 792                                              F.getKind()));
 793       }
 794
 795       // Update the layout, and remember that we relaxed. If we are relaxing
 796       // everything, we can skip this step since nothing will depend on updating
 797       // the values.
 798       if (!getRelaxAll())
 799         Layout.UpdateForSlide(IF, SlideAmount);
 800       WasRelaxed = true;
 801     }
 802   }
 803
 804   return WasRelaxed;
 805 }
 806
 807 void MCAssembler::FinishLayout(MCAsmLayout &Layout) {
 808   // Lower out any instruction fragments, to simplify the fixup application and
 809   // output.
 810   //
 811   // FIXME-PERF: We don't have to do this, but the assumption is that it is
 812   // cheap (we will mostly end up eliminating fragments and appending on to data
 813   // fragments), so the extra complexity downstream isn't worth it. Evaluate
 814   // this assumption.
 815   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 816     MCSectionData &SD = *it;
 817
 818     for (MCSectionData::iterator it2 = SD.begin(),
 819            ie2 = SD.end(); it2 != ie2; ++it2) {
 820       MCInstFragment *IF = dyn_cast<MCInstFragment>(it2);
 821       if (!IF)
 822         continue;
 823
 824       // Create a new data fragment for the instruction.
 825       //
 826       // FIXME-PERF: Reuse previous data fragment if possible.
 827       MCDataFragment *DF = new MCDataFragment();
 828       SD.getFragmentList().insert(it2, DF);
 829
 830       // Update the data fragments layout data.
 831       DF->setParent(IF->getParent());
 832       DF->setAtom(IF->getAtom());
 833       DF->setOrdinal(IF->getOrdinal());
 834       Layout.FragmentReplaced(IF, DF);
 835
 836       // Copy in the data and the fixups.
 837       DF->getContents().append(IF->getCode().begin(), IF->getCode().end());
 838       for (unsigned i = 0, e = IF->getFixups().size(); i != e; ++i)
 839         DF->getFixups().push_back(IF->getFixups()[i]);
 840
 841       // Delete the instruction fragment and update the iterator.
 842       SD.getFragmentList().erase(IF);
 843       it2 = DF;
 844     }
 845   }
 846 }
 847
 848 // Debugging methods
 849
 850 namespace llvm {
 851
 852 raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) {
 853   OS << "<MCAsmFixup" << " Offset:" << AF.Offset << " Value:" << *AF.Value
 854      << " Kind:" << AF.Kind << ">";
 855   return OS;
 856 }
 857
 858 }
 859
 860 void MCFragment::dump() {
 861   raw_ostream &OS = llvm::errs();
 862
 863   OS << "<MCFragment " << (void*) this << " Offset:" << Offset
 864      << " EffectiveSize:" << EffectiveSize << ">";
 865 }
 866
 867 void MCAlignFragment::dump() {
 868   raw_ostream &OS = llvm::errs();
 869
 870   OS << "<MCAlignFragment ";
 871   this->MCFragment::dump();
 872   if (hasEmitNops())
 873     OS << " (emit nops)";
 874   if (hasOnlyAlignAddress())
 875     OS << " (only align section)";
 876   OS << "\n       ";
 877   OS << " Alignment:" << getAlignment()
 878      << " Value:" << getValue() << " ValueSize:" << getValueSize()
 879      << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">";
 880 }
 881
 882 void MCDataFragment::dump() {
 883   raw_ostream &OS = llvm::errs();
 884
 885   OS << "<MCDataFragment ";
 886   this->MCFragment::dump();
 887   OS << "\n       ";
 888   OS << " Contents:[";
 889   for (unsigned i = 0, e = getContents().size(); i != e; ++i) {
 890     if (i) OS << ",";
 891     OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF);
 892   }
 893   OS << "] (" << getContents().size() << " bytes)";
 894
 895   if (!getFixups().empty()) {
 896     OS << ",\n       ";
 897     OS << " Fixups:[";
 898     for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) {
 899       if (it != fixup_begin()) OS << ",\n                ";
 900       OS << *it;
 901     }
 902     OS << "]";
 903   }
 904
 905   OS << ">";
 906 }
 907
 908 void MCFillFragment::dump() {
 909   raw_ostream &OS = llvm::errs();
 910
 911   OS << "<MCFillFragment ";
 912   this->MCFragment::dump();
 913   OS << "\n       ";
 914   OS << " Value:" << getValue() << " ValueSize:" << getValueSize()
 915      << " Size:" << getSize() << ">";
 916 }
 917
 918 void MCInstFragment::dump() {
 919   raw_ostream &OS = llvm::errs();
 920
 921   OS << "<MCInstFragment ";
 922   this->MCFragment::dump();
 923   OS << "\n       ";
 924   OS << " Inst:";
 925   getInst().dump_pretty(OS);
 926   OS << ">";
 927 }
 928
 929 void MCOrgFragment::dump() {
 930   raw_ostream &OS = llvm::errs();
 931
 932   OS << "<MCOrgFragment ";
 933   this->MCFragment::dump();
 934   OS << "\n       ";
 935   OS << " Offset:" << getOffset() << " Value:" << getValue() << ">";
 936 }
 937
 938 void MCSectionData::dump() {
 939   raw_ostream &OS = llvm::errs();
 940
 941   OS << "<MCSectionData";
 942   OS << " Alignment:" << getAlignment() << " Address:" << Address
 943      << " Fragments:[\n      ";
 944   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 945     if (it != begin()) OS << ",\n      ";
 946     it->dump();
 947   }
 948   OS << "]>";
 949 }
 950
 951 void MCSymbolData::dump() {
 952   raw_ostream &OS = llvm::errs();
 953
 954   OS << "<MCSymbolData Symbol:" << getSymbol()
 955      << " Fragment:" << getFragment() << " Offset:" << getOffset()
 956      << " Flags:" << getFlags() << " Index:" << getIndex();
 957   if (isCommon())
 958     OS << " (common, size:" << getCommonSize()
 959        << " align: " << getCommonAlignment() << ")";
 960   if (isExternal())
 961     OS << " (external)";
 962   if (isPrivateExtern())
 963     OS << " (private extern)";
 964   OS << ">";
 965 }
 966
 967 void MCAssembler::dump() {
 968   raw_ostream &OS = llvm::errs();
 969
 970   OS << "<MCAssembler\n";
 971   OS << "  Sections:[\n    ";
 972   for (iterator it = begin(), ie = end(); it != ie; ++it) {
 973     if (it != begin()) OS << ",\n    ";
 974     it->dump();
 975   }
 976   OS << "],\n";
 977   OS << "  Symbols:[";
 978
 979   for (symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
 980     if (it != symbol_begin()) OS << ",\n           ";
 981     it->dump();
 982   }
 983   OS << "]>\n";
 984 }