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