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 #include "llvm/MC/MCAssembler.h"
11 #include "llvm/ADT/ArrayRef.h"
12 #include "llvm/ADT/SmallString.h"
13 #include "llvm/ADT/SmallVector.h"
14 #include "llvm/ADT/Statistic.h"
15 #include "llvm/ADT/StringRef.h"
16 #include "llvm/ADT/Twine.h"
17 #include "llvm/MC/MCAsmBackend.h"
18 #include "llvm/MC/MCAsmInfo.h"
19 #include "llvm/MC/MCAsmLayout.h"
20 #include "llvm/MC/MCCodeEmitter.h"
21 #include "llvm/MC/MCCodeView.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCDwarf.h"
24 #include "llvm/MC/MCExpr.h"
25 #include "llvm/MC/MCFixup.h"
26 #include "llvm/MC/MCFixupKindInfo.h"
27 #include "llvm/MC/MCFragment.h"
28 #include "llvm/MC/MCInst.h"
29 #include "llvm/MC/MCObjectWriter.h"
30 #include "llvm/MC/MCSection.h"
31 #include "llvm/MC/MCSectionELF.h"
32 #include "llvm/MC/MCSymbol.h"
33 #include "llvm/MC/MCValue.h"
34 #include "llvm/Support/Casting.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/LEB128.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include <cassert>
41 #include <cstdint>
42 #include <cstring>
43 #include <tuple>
44 #include <utility>
45 
46 using namespace llvm;
47 
48 #define DEBUG_TYPE "assembler"
49 
50 namespace {
51 namespace stats {
52 
53 STATISTIC(EmittedFragments, "Number of emitted assembler fragments - total");
54 STATISTIC(EmittedRelaxableFragments,
55           "Number of emitted assembler fragments - relaxable");
56 STATISTIC(EmittedDataFragments,
57           "Number of emitted assembler fragments - data");
58 STATISTIC(EmittedCompactEncodedInstFragments,
59           "Number of emitted assembler fragments - compact encoded inst");
60 STATISTIC(EmittedAlignFragments,
61           "Number of emitted assembler fragments - align");
62 STATISTIC(EmittedFillFragments,
63           "Number of emitted assembler fragments - fill");
64 STATISTIC(EmittedOrgFragments,
65           "Number of emitted assembler fragments - org");
66 STATISTIC(evaluateFixup, "Number of evaluated fixups");
67 STATISTIC(FragmentLayouts, "Number of fragment layouts");
68 STATISTIC(ObjectBytes, "Number of emitted object file bytes");
69 STATISTIC(RelaxationSteps, "Number of assembler layout and relaxation steps");
70 STATISTIC(RelaxedInstructions, "Number of relaxed instructions");
71 STATISTIC(PaddingFragmentsRelaxations,
72           "Number of Padding Fragments relaxations");
73 STATISTIC(PaddingFragmentsBytes,
74           "Total size of all padding from adding Fragments");
75 
76 } // end namespace stats
77 } // end anonymous namespace
78 
79 // FIXME FIXME FIXME: There are number of places in this file where we convert
80 // what is a 64-bit assembler value used for computation into a value in the
81 // object file, which may truncate it. We should detect that truncation where
82 // invalid and report errors back.
83 
84 /* *** */
85 
MCAssembler(MCContext & Context,std::unique_ptr<MCAsmBackend> Backend,std::unique_ptr<MCCodeEmitter> Emitter,std::unique_ptr<MCObjectWriter> Writer)86 MCAssembler::MCAssembler(MCContext &Context,
87                          std::unique_ptr<MCAsmBackend> Backend,
88                          std::unique_ptr<MCCodeEmitter> Emitter,
89                          std::unique_ptr<MCObjectWriter> Writer)
90     : Context(Context), Backend(std::move(Backend)),
91       Emitter(std::move(Emitter)), Writer(std::move(Writer)),
92       BundleAlignSize(0), RelaxAll(false), SubsectionsViaSymbols(false),
93       IncrementalLinkerCompatible(false), ELFHeaderEFlags(0) {
94   VersionInfo.Major = 0; // Major version == 0 for "none specified"
95 }
96 
97 MCAssembler::~MCAssembler() = default;
98 
reset()99 void MCAssembler::reset() {
100   Sections.clear();
101   Symbols.clear();
102   IndirectSymbols.clear();
103   DataRegions.clear();
104   LinkerOptions.clear();
105   FileNames.clear();
106   ThumbFuncs.clear();
107   BundleAlignSize = 0;
108   RelaxAll = false;
109   SubsectionsViaSymbols = false;
110   IncrementalLinkerCompatible = false;
111   ELFHeaderEFlags = 0;
112   LOHContainer.reset();
113   VersionInfo.Major = 0;
114   VersionInfo.SDKVersion = VersionTuple();
115 
116   // reset objects owned by us
117   if (getBackendPtr())
118     getBackendPtr()->reset();
119   if (getEmitterPtr())
120     getEmitterPtr()->reset();
121   if (getWriterPtr())
122     getWriterPtr()->reset();
123   getLOHContainer().reset();
124 }
125 
registerSection(MCSection & Section)126 bool MCAssembler::registerSection(MCSection &Section) {
127   if (Section.isRegistered())
128     return false;
129   Sections.push_back(&Section);
130   Section.setIsRegistered(true);
131   return true;
132 }
133 
isThumbFunc(const MCSymbol * Symbol) const134 bool MCAssembler::isThumbFunc(const MCSymbol *Symbol) const {
135   if (ThumbFuncs.count(Symbol))
136     return true;
137 
138   if (!Symbol->isVariable())
139     return false;
140 
141   const MCExpr *Expr = Symbol->getVariableValue();
142 
143   MCValue V;
144   if (!Expr->evaluateAsRelocatable(V, nullptr, nullptr))
145     return false;
146 
147   if (V.getSymB() || V.getRefKind() != MCSymbolRefExpr::VK_None)
148     return false;
149 
150   const MCSymbolRefExpr *Ref = V.getSymA();
151   if (!Ref)
152     return false;
153 
154   if (Ref->getKind() != MCSymbolRefExpr::VK_None)
155     return false;
156 
157   const MCSymbol &Sym = Ref->getSymbol();
158   if (!isThumbFunc(&Sym))
159     return false;
160 
161   ThumbFuncs.insert(Symbol); // Cache it.
162   return true;
163 }
164 
isSymbolLinkerVisible(const MCSymbol & Symbol) const165 bool MCAssembler::isSymbolLinkerVisible(const MCSymbol &Symbol) const {
166   // Non-temporary labels should always be visible to the linker.
167   if (!Symbol.isTemporary())
168     return true;
169 
170   // Absolute temporary labels are never visible.
171   if (!Symbol.isInSection())
172     return false;
173 
174   if (Symbol.isUsedInReloc())
175     return true;
176 
177   return false;
178 }
179 
getAtom(const MCSymbol & S) const180 const MCSymbol *MCAssembler::getAtom(const MCSymbol &S) const {
181   // Linker visible symbols define atoms.
182   if (isSymbolLinkerVisible(S))
183     return &S;
184 
185   // Absolute and undefined symbols have no defining atom.
186   if (!S.isInSection())
187     return nullptr;
188 
189   // Non-linker visible symbols in sections which can't be atomized have no
190   // defining atom.
191   if (!getContext().getAsmInfo()->isSectionAtomizableBySymbols(
192           *S.getFragment()->getParent()))
193     return nullptr;
194 
195   // Otherwise, return the atom for the containing fragment.
196   return S.getFragment()->getAtom();
197 }
198 
evaluateFixup(const MCAsmLayout & Layout,const MCFixup & Fixup,const MCFragment * DF,MCValue & Target,uint64_t & Value,bool & WasForced) const199 bool MCAssembler::evaluateFixup(const MCAsmLayout &Layout,
200                                 const MCFixup &Fixup, const MCFragment *DF,
201                                 MCValue &Target, uint64_t &Value,
202                                 bool &WasForced) const {
203   ++stats::evaluateFixup;
204 
205   // FIXME: This code has some duplication with recordRelocation. We should
206   // probably merge the two into a single callback that tries to evaluate a
207   // fixup and records a relocation if one is needed.
208 
209   // On error claim to have completely evaluated the fixup, to prevent any
210   // further processing from being done.
211   const MCExpr *Expr = Fixup.getValue();
212   MCContext &Ctx = getContext();
213   Value = 0;
214   WasForced = false;
215   if (!Expr->evaluateAsRelocatable(Target, &Layout, &Fixup)) {
216     Ctx.reportError(Fixup.getLoc(), "expected relocatable expression");
217     return true;
218   }
219   if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
220     if (RefB->getKind() != MCSymbolRefExpr::VK_None) {
221       Ctx.reportError(Fixup.getLoc(),
222                       "unsupported subtraction of qualified symbol");
223       return true;
224     }
225   }
226 
227   assert(getBackendPtr() && "Expected assembler backend");
228   bool IsPCRel = getBackendPtr()->getFixupKindInfo(Fixup.getKind()).Flags &
229                  MCFixupKindInfo::FKF_IsPCRel;
230 
231   bool IsResolved = false;
232   if (IsPCRel) {
233     if (Target.getSymB()) {
234       IsResolved = false;
235     } else if (!Target.getSymA()) {
236       IsResolved = false;
237     } else {
238       const MCSymbolRefExpr *A = Target.getSymA();
239       const MCSymbol &SA = A->getSymbol();
240       if (A->getKind() != MCSymbolRefExpr::VK_None || SA.isUndefined()) {
241         IsResolved = false;
242       } else if (auto *Writer = getWriterPtr()) {
243         IsResolved = Writer->isSymbolRefDifferenceFullyResolvedImpl(
244             *this, SA, *DF, false, true);
245       }
246     }
247   } else {
248     IsResolved = Target.isAbsolute();
249   }
250 
251   Value = Target.getConstant();
252 
253   if (const MCSymbolRefExpr *A = Target.getSymA()) {
254     const MCSymbol &Sym = A->getSymbol();
255     if (Sym.isDefined())
256       Value += Layout.getSymbolOffset(Sym);
257   }
258   if (const MCSymbolRefExpr *B = Target.getSymB()) {
259     const MCSymbol &Sym = B->getSymbol();
260     if (Sym.isDefined())
261       Value -= Layout.getSymbolOffset(Sym);
262   }
263 
264   bool ShouldAlignPC = getBackend().getFixupKindInfo(Fixup.getKind()).Flags &
265                        MCFixupKindInfo::FKF_IsAlignedDownTo32Bits;
266   assert((ShouldAlignPC ? IsPCRel : true) &&
267     "FKF_IsAlignedDownTo32Bits is only allowed on PC-relative fixups!");
268 
269   if (IsPCRel) {
270     uint32_t Offset = Layout.getFragmentOffset(DF) + Fixup.getOffset();
271 
272     // A number of ARM fixups in Thumb mode require that the effective PC
273     // address be determined as the 32-bit aligned version of the actual offset.
274     if (ShouldAlignPC) Offset &= ~0x3;
275     Value -= Offset;
276   }
277 
278   // Let the backend force a relocation if needed.
279   if (IsResolved && getBackend().shouldForceRelocation(*this, Fixup, Target)) {
280     IsResolved = false;
281     WasForced = true;
282   }
283 
284   return IsResolved;
285 }
286 
computeFragmentSize(const MCAsmLayout & Layout,const MCFragment & F) const287 uint64_t MCAssembler::computeFragmentSize(const MCAsmLayout &Layout,
288                                           const MCFragment &F) const {
289   assert(getBackendPtr() && "Requires assembler backend");
290   switch (F.getKind()) {
291   case MCFragment::FT_Data:
292     return cast<MCDataFragment>(F).getContents().size();
293   case MCFragment::FT_Relaxable:
294     return cast<MCRelaxableFragment>(F).getContents().size();
295   case MCFragment::FT_CompactEncodedInst:
296     return cast<MCCompactEncodedInstFragment>(F).getContents().size();
297   case MCFragment::FT_Fill: {
298     auto &FF = cast<MCFillFragment>(F);
299     int64_t NumValues = 0;
300     if (!FF.getNumValues().evaluateAsAbsolute(NumValues, Layout)) {
301       getContext().reportError(FF.getLoc(),
302                                "expected assembly-time absolute expression");
303       return 0;
304     }
305     int64_t Size = NumValues * FF.getValueSize();
306     if (Size < 0) {
307       getContext().reportError(FF.getLoc(), "invalid number of bytes");
308       return 0;
309     }
310     return Size;
311   }
312 
313   case MCFragment::FT_LEB:
314     return cast<MCLEBFragment>(F).getContents().size();
315 
316   case MCFragment::FT_Padding:
317     return cast<MCPaddingFragment>(F).getSize();
318 
319   case MCFragment::FT_SymbolId:
320     return 4;
321 
322   case MCFragment::FT_Align: {
323     const MCAlignFragment &AF = cast<MCAlignFragment>(F);
324     unsigned Offset = Layout.getFragmentOffset(&AF);
325     unsigned Size = OffsetToAlignment(Offset, AF.getAlignment());
326     // If we are padding with nops, force the padding to be larger than the
327     // minimum nop size.
328     if (Size > 0 && AF.hasEmitNops()) {
329       while (Size % getBackend().getMinimumNopSize())
330         Size += AF.getAlignment();
331     }
332     if (Size > AF.getMaxBytesToEmit())
333       return 0;
334     return Size;
335   }
336 
337   case MCFragment::FT_Org: {
338     const MCOrgFragment &OF = cast<MCOrgFragment>(F);
339     MCValue Value;
340     if (!OF.getOffset().evaluateAsValue(Value, Layout)) {
341       getContext().reportError(OF.getLoc(),
342                                "expected assembly-time absolute expression");
343         return 0;
344     }
345 
346     uint64_t FragmentOffset = Layout.getFragmentOffset(&OF);
347     int64_t TargetLocation = Value.getConstant();
348     if (const MCSymbolRefExpr *A = Value.getSymA()) {
349       uint64_t Val;
350       if (!Layout.getSymbolOffset(A->getSymbol(), Val)) {
351         getContext().reportError(OF.getLoc(), "expected absolute expression");
352         return 0;
353       }
354       TargetLocation += Val;
355     }
356     int64_t Size = TargetLocation - FragmentOffset;
357     if (Size < 0 || Size >= 0x40000000) {
358       getContext().reportError(
359           OF.getLoc(), "invalid .org offset '" + Twine(TargetLocation) +
360                            "' (at offset '" + Twine(FragmentOffset) + "')");
361       return 0;
362     }
363     return Size;
364   }
365 
366   case MCFragment::FT_Dwarf:
367     return cast<MCDwarfLineAddrFragment>(F).getContents().size();
368   case MCFragment::FT_DwarfFrame:
369     return cast<MCDwarfCallFrameFragment>(F).getContents().size();
370   case MCFragment::FT_CVInlineLines:
371     return cast<MCCVInlineLineTableFragment>(F).getContents().size();
372   case MCFragment::FT_CVDefRange:
373     return cast<MCCVDefRangeFragment>(F).getContents().size();
374   case MCFragment::FT_Dummy:
375     llvm_unreachable("Should not have been added");
376   }
377 
378   llvm_unreachable("invalid fragment kind");
379 }
380 
layoutFragment(MCFragment * F)381 void MCAsmLayout::layoutFragment(MCFragment *F) {
382   MCFragment *Prev = F->getPrevNode();
383 
384   // We should never try to recompute something which is valid.
385   assert(!isFragmentValid(F) && "Attempt to recompute a valid fragment!");
386   // We should never try to compute the fragment layout if its predecessor
387   // isn't valid.
388   assert((!Prev || isFragmentValid(Prev)) &&
389          "Attempt to compute fragment before its predecessor!");
390 
391   ++stats::FragmentLayouts;
392 
393   // Compute fragment offset and size.
394   if (Prev)
395     F->Offset = Prev->Offset + getAssembler().computeFragmentSize(*this, *Prev);
396   else
397     F->Offset = 0;
398   LastValidFragment[F->getParent()] = F;
399 
400   // If bundling is enabled and this fragment has instructions in it, it has to
401   // obey the bundling restrictions. With padding, we'll have:
402   //
403   //
404   //        BundlePadding
405   //             |||
406   // -------------------------------------
407   //   Prev  |##########|       F        |
408   // -------------------------------------
409   //                    ^
410   //                    |
411   //                    F->Offset
412   //
413   // The fragment's offset will point to after the padding, and its computed
414   // size won't include the padding.
415   //
416   // When the -mc-relax-all flag is used, we optimize bundling by writting the
417   // padding directly into fragments when the instructions are emitted inside
418   // the streamer. When the fragment is larger than the bundle size, we need to
419   // ensure that it's bundle aligned. This means that if we end up with
420   // multiple fragments, we must emit bundle padding between fragments.
421   //
422   // ".align N" is an example of a directive that introduces multiple
423   // fragments. We could add a special case to handle ".align N" by emitting
424   // within-fragment padding (which would produce less padding when N is less
425   // than the bundle size), but for now we don't.
426   //
427   if (Assembler.isBundlingEnabled() && F->hasInstructions()) {
428     assert(isa<MCEncodedFragment>(F) &&
429            "Only MCEncodedFragment implementations have instructions");
430     MCEncodedFragment *EF = cast<MCEncodedFragment>(F);
431     uint64_t FSize = Assembler.computeFragmentSize(*this, *EF);
432 
433     if (!Assembler.getRelaxAll() && FSize > Assembler.getBundleAlignSize())
434       report_fatal_error("Fragment can't be larger than a bundle size");
435 
436     uint64_t RequiredBundlePadding =
437         computeBundlePadding(Assembler, EF, EF->Offset, FSize);
438     if (RequiredBundlePadding > UINT8_MAX)
439       report_fatal_error("Padding cannot exceed 255 bytes");
440     EF->setBundlePadding(static_cast<uint8_t>(RequiredBundlePadding));
441     EF->Offset += RequiredBundlePadding;
442   }
443 }
444 
registerSymbol(const MCSymbol & Symbol,bool * Created)445 void MCAssembler::registerSymbol(const MCSymbol &Symbol, bool *Created) {
446   bool New = !Symbol.isRegistered();
447   if (Created)
448     *Created = New;
449   if (New) {
450     Symbol.setIsRegistered(true);
451     Symbols.push_back(&Symbol);
452   }
453 }
454 
writeFragmentPadding(raw_ostream & OS,const MCEncodedFragment & EF,uint64_t FSize) const455 void MCAssembler::writeFragmentPadding(raw_ostream &OS,
456                                        const MCEncodedFragment &EF,
457                                        uint64_t FSize) const {
458   assert(getBackendPtr() && "Expected assembler backend");
459   // Should NOP padding be written out before this fragment?
460   unsigned BundlePadding = EF.getBundlePadding();
461   if (BundlePadding > 0) {
462     assert(isBundlingEnabled() &&
463            "Writing bundle padding with disabled bundling");
464     assert(EF.hasInstructions() &&
465            "Writing bundle padding for a fragment without instructions");
466 
467     unsigned TotalLength = BundlePadding + static_cast<unsigned>(FSize);
468     if (EF.alignToBundleEnd() && TotalLength > getBundleAlignSize()) {
469       // If the padding itself crosses a bundle boundary, it must be emitted
470       // in 2 pieces, since even nop instructions must not cross boundaries.
471       //             v--------------v   <- BundleAlignSize
472       //        v---------v             <- BundlePadding
473       // ----------------------------
474       // | Prev |####|####|    F    |
475       // ----------------------------
476       //        ^-------------------^   <- TotalLength
477       unsigned DistanceToBoundary = TotalLength - getBundleAlignSize();
478       if (!getBackend().writeNopData(OS, DistanceToBoundary))
479         report_fatal_error("unable to write NOP sequence of " +
480                            Twine(DistanceToBoundary) + " bytes");
481       BundlePadding -= DistanceToBoundary;
482     }
483     if (!getBackend().writeNopData(OS, BundlePadding))
484       report_fatal_error("unable to write NOP sequence of " +
485                          Twine(BundlePadding) + " bytes");
486   }
487 }
488 
489 /// Write the fragment \p F to the output file.
writeFragment(raw_ostream & OS,const MCAssembler & Asm,const MCAsmLayout & Layout,const MCFragment & F)490 static void writeFragment(raw_ostream &OS, const MCAssembler &Asm,
491                           const MCAsmLayout &Layout, const MCFragment &F) {
492   // FIXME: Embed in fragments instead?
493   uint64_t FragmentSize = Asm.computeFragmentSize(Layout, F);
494 
495   support::endianness Endian = Asm.getBackend().Endian;
496 
497   if (const MCEncodedFragment *EF = dyn_cast<MCEncodedFragment>(&F))
498     Asm.writeFragmentPadding(OS, *EF, FragmentSize);
499 
500   // This variable (and its dummy usage) is to participate in the assert at
501   // the end of the function.
502   uint64_t Start = OS.tell();
503   (void) Start;
504 
505   ++stats::EmittedFragments;
506 
507   switch (F.getKind()) {
508   case MCFragment::FT_Align: {
509     ++stats::EmittedAlignFragments;
510     const MCAlignFragment &AF = cast<MCAlignFragment>(F);
511     assert(AF.getValueSize() && "Invalid virtual align in concrete fragment!");
512 
513     uint64_t Count = FragmentSize / AF.getValueSize();
514 
515     // FIXME: This error shouldn't actually occur (the front end should emit
516     // multiple .align directives to enforce the semantics it wants), but is
517     // severe enough that we want to report it. How to handle this?
518     if (Count * AF.getValueSize() != FragmentSize)
519       report_fatal_error("undefined .align directive, value size '" +
520                         Twine(AF.getValueSize()) +
521                         "' is not a divisor of padding size '" +
522                         Twine(FragmentSize) + "'");
523 
524     // See if we are aligning with nops, and if so do that first to try to fill
525     // the Count bytes.  Then if that did not fill any bytes or there are any
526     // bytes left to fill use the Value and ValueSize to fill the rest.
527     // If we are aligning with nops, ask that target to emit the right data.
528     if (AF.hasEmitNops()) {
529       if (!Asm.getBackend().writeNopData(OS, Count))
530         report_fatal_error("unable to write nop sequence of " +
531                           Twine(Count) + " bytes");
532       break;
533     }
534 
535     // Otherwise, write out in multiples of the value size.
536     for (uint64_t i = 0; i != Count; ++i) {
537       switch (AF.getValueSize()) {
538       default: llvm_unreachable("Invalid size!");
539       case 1: OS << char(AF.getValue()); break;
540       case 2:
541         support::endian::write<uint16_t>(OS, AF.getValue(), Endian);
542         break;
543       case 4:
544         support::endian::write<uint32_t>(OS, AF.getValue(), Endian);
545         break;
546       case 8:
547         support::endian::write<uint64_t>(OS, AF.getValue(), Endian);
548         break;
549       }
550     }
551     break;
552   }
553 
554   case MCFragment::FT_Data:
555     ++stats::EmittedDataFragments;
556     OS << cast<MCDataFragment>(F).getContents();
557     break;
558 
559   case MCFragment::FT_Relaxable:
560     ++stats::EmittedRelaxableFragments;
561     OS << cast<MCRelaxableFragment>(F).getContents();
562     break;
563 
564   case MCFragment::FT_CompactEncodedInst:
565     ++stats::EmittedCompactEncodedInstFragments;
566     OS << cast<MCCompactEncodedInstFragment>(F).getContents();
567     break;
568 
569   case MCFragment::FT_Fill: {
570     ++stats::EmittedFillFragments;
571     const MCFillFragment &FF = cast<MCFillFragment>(F);
572     uint64_t V = FF.getValue();
573     unsigned VSize = FF.getValueSize();
574     const unsigned MaxChunkSize = 16;
575     char Data[MaxChunkSize];
576     // Duplicate V into Data as byte vector to reduce number of
577     // writes done. As such, do endian conversion here.
578     for (unsigned I = 0; I != VSize; ++I) {
579       unsigned index = Endian == support::little ? I : (VSize - I - 1);
580       Data[I] = uint8_t(V >> (index * 8));
581     }
582     for (unsigned I = VSize; I < MaxChunkSize; ++I)
583       Data[I] = Data[I - VSize];
584 
585     // Set to largest multiple of VSize in Data.
586     const unsigned NumPerChunk = MaxChunkSize / VSize;
587     // Set ChunkSize to largest multiple of VSize in Data
588     const unsigned ChunkSize = VSize * NumPerChunk;
589 
590     // Do copies by chunk.
591     StringRef Ref(Data, ChunkSize);
592     for (uint64_t I = 0, E = FragmentSize / ChunkSize; I != E; ++I)
593       OS << Ref;
594 
595     // do remainder if needed.
596     unsigned TrailingCount = FragmentSize % ChunkSize;
597     if (TrailingCount)
598       OS.write(Data, TrailingCount);
599     break;
600   }
601 
602   case MCFragment::FT_LEB: {
603     const MCLEBFragment &LF = cast<MCLEBFragment>(F);
604     OS << LF.getContents();
605     break;
606   }
607 
608   case MCFragment::FT_Padding: {
609     if (!Asm.getBackend().writeNopData(OS, FragmentSize))
610       report_fatal_error("unable to write nop sequence of " +
611                          Twine(FragmentSize) + " bytes");
612     break;
613   }
614 
615   case MCFragment::FT_SymbolId: {
616     const MCSymbolIdFragment &SF = cast<MCSymbolIdFragment>(F);
617     support::endian::write<uint32_t>(OS, SF.getSymbol()->getIndex(), Endian);
618     break;
619   }
620 
621   case MCFragment::FT_Org: {
622     ++stats::EmittedOrgFragments;
623     const MCOrgFragment &OF = cast<MCOrgFragment>(F);
624 
625     for (uint64_t i = 0, e = FragmentSize; i != e; ++i)
626       OS << char(OF.getValue());
627 
628     break;
629   }
630 
631   case MCFragment::FT_Dwarf: {
632     const MCDwarfLineAddrFragment &OF = cast<MCDwarfLineAddrFragment>(F);
633     OS << OF.getContents();
634     break;
635   }
636   case MCFragment::FT_DwarfFrame: {
637     const MCDwarfCallFrameFragment &CF = cast<MCDwarfCallFrameFragment>(F);
638     OS << CF.getContents();
639     break;
640   }
641   case MCFragment::FT_CVInlineLines: {
642     const auto &OF = cast<MCCVInlineLineTableFragment>(F);
643     OS << OF.getContents();
644     break;
645   }
646   case MCFragment::FT_CVDefRange: {
647     const auto &DRF = cast<MCCVDefRangeFragment>(F);
648     OS << DRF.getContents();
649     break;
650   }
651   case MCFragment::FT_Dummy:
652     llvm_unreachable("Should not have been added");
653   }
654 
655   assert(OS.tell() - Start == FragmentSize &&
656          "The stream should advance by fragment size");
657 }
658 
writeSectionData(raw_ostream & OS,const MCSection * Sec,const MCAsmLayout & Layout) const659 void MCAssembler::writeSectionData(raw_ostream &OS, const MCSection *Sec,
660                                    const MCAsmLayout &Layout) const {
661   assert(getBackendPtr() && "Expected assembler backend");
662 
663   // Ignore virtual sections.
664   if (Sec->isVirtualSection()) {
665     assert(Layout.getSectionFileSize(Sec) == 0 && "Invalid size for section!");
666 
667     // Check that contents are only things legal inside a virtual section.
668     for (const MCFragment &F : *Sec) {
669       switch (F.getKind()) {
670       default: llvm_unreachable("Invalid fragment in virtual section!");
671       case MCFragment::FT_Data: {
672         // Check that we aren't trying to write a non-zero contents (or fixups)
673         // into a virtual section. This is to support clients which use standard
674         // directives to fill the contents of virtual sections.
675         const MCDataFragment &DF = cast<MCDataFragment>(F);
676         if (DF.fixup_begin() != DF.fixup_end())
677           report_fatal_error("cannot have fixups in virtual section!");
678         for (unsigned i = 0, e = DF.getContents().size(); i != e; ++i)
679           if (DF.getContents()[i]) {
680             if (auto *ELFSec = dyn_cast<const MCSectionELF>(Sec))
681               report_fatal_error("non-zero initializer found in section '" +
682                   ELFSec->getSectionName() + "'");
683             else
684               report_fatal_error("non-zero initializer found in virtual section");
685           }
686         break;
687       }
688       case MCFragment::FT_Align:
689         // Check that we aren't trying to write a non-zero value into a virtual
690         // section.
691         assert((cast<MCAlignFragment>(F).getValueSize() == 0 ||
692                 cast<MCAlignFragment>(F).getValue() == 0) &&
693                "Invalid align in virtual section!");
694         break;
695       case MCFragment::FT_Fill:
696         assert((cast<MCFillFragment>(F).getValue() == 0) &&
697                "Invalid fill in virtual section!");
698         break;
699       }
700     }
701 
702     return;
703   }
704 
705   uint64_t Start = OS.tell();
706   (void)Start;
707 
708   for (const MCFragment &F : *Sec)
709     writeFragment(OS, *this, Layout, F);
710 
711   assert(OS.tell() - Start == Layout.getSectionAddressSize(Sec));
712 }
713 
714 std::tuple<MCValue, uint64_t, bool>
handleFixup(const MCAsmLayout & Layout,MCFragment & F,const MCFixup & Fixup)715 MCAssembler::handleFixup(const MCAsmLayout &Layout, MCFragment &F,
716                          const MCFixup &Fixup) {
717   // Evaluate the fixup.
718   MCValue Target;
719   uint64_t FixedValue;
720   bool WasForced;
721   bool IsResolved = evaluateFixup(Layout, Fixup, &F, Target, FixedValue,
722                                   WasForced);
723   if (!IsResolved) {
724     // The fixup was unresolved, we need a relocation. Inform the object
725     // writer of the relocation, and give it an opportunity to adjust the
726     // fixup value if need be.
727     if (Target.getSymA() && Target.getSymB() &&
728         getBackend().requiresDiffExpressionRelocations()) {
729       // The fixup represents the difference between two symbols, which the
730       // backend has indicated must be resolved at link time. Split up the fixup
731       // into two relocations, one for the add, and one for the sub, and emit
732       // both of these. The constant will be associated with the add half of the
733       // expression.
734       MCFixup FixupAdd = MCFixup::createAddFor(Fixup);
735       MCValue TargetAdd =
736           MCValue::get(Target.getSymA(), nullptr, Target.getConstant());
737       getWriter().recordRelocation(*this, Layout, &F, FixupAdd, TargetAdd,
738                                    FixedValue);
739       MCFixup FixupSub = MCFixup::createSubFor(Fixup);
740       MCValue TargetSub = MCValue::get(Target.getSymB());
741       getWriter().recordRelocation(*this, Layout, &F, FixupSub, TargetSub,
742                                    FixedValue);
743     } else {
744       getWriter().recordRelocation(*this, Layout, &F, Fixup, Target,
745                                    FixedValue);
746     }
747   }
748   return std::make_tuple(Target, FixedValue, IsResolved);
749 }
750 
layout(MCAsmLayout & Layout)751 void MCAssembler::layout(MCAsmLayout &Layout) {
752   assert(getBackendPtr() && "Expected assembler backend");
753   DEBUG_WITH_TYPE("mc-dump", {
754       errs() << "assembler backend - pre-layout\n--\n";
755       dump(); });
756 
757   // Create dummy fragments and assign section ordinals.
758   unsigned SectionIndex = 0;
759   for (MCSection &Sec : *this) {
760     // Create dummy fragments to eliminate any empty sections, this simplifies
761     // layout.
762     if (Sec.getFragmentList().empty())
763       new MCDataFragment(&Sec);
764 
765     Sec.setOrdinal(SectionIndex++);
766   }
767 
768   // Assign layout order indices to sections and fragments.
769   for (unsigned i = 0, e = Layout.getSectionOrder().size(); i != e; ++i) {
770     MCSection *Sec = Layout.getSectionOrder()[i];
771     Sec->setLayoutOrder(i);
772 
773     unsigned FragmentIndex = 0;
774     for (MCFragment &Frag : *Sec)
775       Frag.setLayoutOrder(FragmentIndex++);
776   }
777 
778   // Layout until everything fits.
779   while (layoutOnce(Layout))
780     if (getContext().hadError())
781       return;
782 
783   DEBUG_WITH_TYPE("mc-dump", {
784       errs() << "assembler backend - post-relaxation\n--\n";
785       dump(); });
786 
787   // Finalize the layout, including fragment lowering.
788   finishLayout(Layout);
789 
790   DEBUG_WITH_TYPE("mc-dump", {
791       errs() << "assembler backend - final-layout\n--\n";
792       dump(); });
793 
794   // Allow the object writer a chance to perform post-layout binding (for
795   // example, to set the index fields in the symbol data).
796   getWriter().executePostLayoutBinding(*this, Layout);
797 
798   // Evaluate and apply the fixups, generating relocation entries as necessary.
799   for (MCSection &Sec : *this) {
800     for (MCFragment &Frag : Sec) {
801       // Data and relaxable fragments both have fixups.  So only process
802       // those here.
803       // FIXME: Is there a better way to do this?  MCEncodedFragmentWithFixups
804       // being templated makes this tricky.
805       if (isa<MCEncodedFragment>(&Frag) &&
806           isa<MCCompactEncodedInstFragment>(&Frag))
807         continue;
808       if (!isa<MCEncodedFragment>(&Frag) && !isa<MCCVDefRangeFragment>(&Frag))
809         continue;
810       ArrayRef<MCFixup> Fixups;
811       MutableArrayRef<char> Contents;
812       const MCSubtargetInfo *STI = nullptr;
813       if (auto *FragWithFixups = dyn_cast<MCDataFragment>(&Frag)) {
814         Fixups = FragWithFixups->getFixups();
815         Contents = FragWithFixups->getContents();
816         STI = FragWithFixups->getSubtargetInfo();
817         assert(!FragWithFixups->hasInstructions() || STI != nullptr);
818       } else if (auto *FragWithFixups = dyn_cast<MCRelaxableFragment>(&Frag)) {
819         Fixups = FragWithFixups->getFixups();
820         Contents = FragWithFixups->getContents();
821         STI = FragWithFixups->getSubtargetInfo();
822         assert(!FragWithFixups->hasInstructions() || STI != nullptr);
823       } else if (auto *FragWithFixups = dyn_cast<MCCVDefRangeFragment>(&Frag)) {
824         Fixups = FragWithFixups->getFixups();
825         Contents = FragWithFixups->getContents();
826       } else if (auto *FragWithFixups = dyn_cast<MCDwarfLineAddrFragment>(&Frag)) {
827         Fixups = FragWithFixups->getFixups();
828         Contents = FragWithFixups->getContents();
829       } else
830         llvm_unreachable("Unknown fragment with fixups!");
831       for (const MCFixup &Fixup : Fixups) {
832         uint64_t FixedValue;
833         bool IsResolved;
834         MCValue Target;
835         std::tie(Target, FixedValue, IsResolved) =
836             handleFixup(Layout, Frag, Fixup);
837         getBackend().applyFixup(*this, Fixup, Target, Contents, FixedValue,
838                                 IsResolved, STI);
839       }
840     }
841   }
842 }
843 
Finish()844 void MCAssembler::Finish() {
845   // Create the layout object.
846   MCAsmLayout Layout(*this);
847   layout(Layout);
848 
849   // Write the object file.
850   stats::ObjectBytes += getWriter().writeObject(*this, Layout);
851 }
852 
fixupNeedsRelaxation(const MCFixup & Fixup,const MCRelaxableFragment * DF,const MCAsmLayout & Layout) const853 bool MCAssembler::fixupNeedsRelaxation(const MCFixup &Fixup,
854                                        const MCRelaxableFragment *DF,
855                                        const MCAsmLayout &Layout) const {
856   assert(getBackendPtr() && "Expected assembler backend");
857   MCValue Target;
858   uint64_t Value;
859   bool WasForced;
860   bool Resolved = evaluateFixup(Layout, Fixup, DF, Target, Value, WasForced);
861   if (Target.getSymA() &&
862       Target.getSymA()->getKind() == MCSymbolRefExpr::VK_X86_ABS8 &&
863       Fixup.getKind() == FK_Data_1)
864     return false;
865   return getBackend().fixupNeedsRelaxationAdvanced(Fixup, Resolved, Value, DF,
866                                                    Layout, WasForced);
867 }
868 
fragmentNeedsRelaxation(const MCRelaxableFragment * F,const MCAsmLayout & Layout) const869 bool MCAssembler::fragmentNeedsRelaxation(const MCRelaxableFragment *F,
870                                           const MCAsmLayout &Layout) const {
871   assert(getBackendPtr() && "Expected assembler backend");
872   // If this inst doesn't ever need relaxation, ignore it. This occurs when we
873   // are intentionally pushing out inst fragments, or because we relaxed a
874   // previous instruction to one that doesn't need relaxation.
875   if (!getBackend().mayNeedRelaxation(F->getInst(), *F->getSubtargetInfo()))
876     return false;
877 
878   for (const MCFixup &Fixup : F->getFixups())
879     if (fixupNeedsRelaxation(Fixup, F, Layout))
880       return true;
881 
882   return false;
883 }
884 
relaxInstruction(MCAsmLayout & Layout,MCRelaxableFragment & F)885 bool MCAssembler::relaxInstruction(MCAsmLayout &Layout,
886                                    MCRelaxableFragment &F) {
887   assert(getEmitterPtr() &&
888          "Expected CodeEmitter defined for relaxInstruction");
889   if (!fragmentNeedsRelaxation(&F, Layout))
890     return false;
891 
892   ++stats::RelaxedInstructions;
893 
894   // FIXME-PERF: We could immediately lower out instructions if we can tell
895   // they are fully resolved, to avoid retesting on later passes.
896 
897   // Relax the fragment.
898 
899   MCInst Relaxed;
900   getBackend().relaxInstruction(F.getInst(), *F.getSubtargetInfo(), Relaxed);
901 
902   // Encode the new instruction.
903   //
904   // FIXME-PERF: If it matters, we could let the target do this. It can
905   // probably do so more efficiently in many cases.
906   SmallVector<MCFixup, 4> Fixups;
907   SmallString<256> Code;
908   raw_svector_ostream VecOS(Code);
909   getEmitter().encodeInstruction(Relaxed, VecOS, Fixups, *F.getSubtargetInfo());
910 
911   // Update the fragment.
912   F.setInst(Relaxed);
913   F.getContents() = Code;
914   F.getFixups() = Fixups;
915 
916   return true;
917 }
918 
relaxPaddingFragment(MCAsmLayout & Layout,MCPaddingFragment & PF)919 bool MCAssembler::relaxPaddingFragment(MCAsmLayout &Layout,
920                                        MCPaddingFragment &PF) {
921   assert(getBackendPtr() && "Expected assembler backend");
922   uint64_t OldSize = PF.getSize();
923   if (!getBackend().relaxFragment(&PF, Layout))
924     return false;
925   uint64_t NewSize = PF.getSize();
926 
927   ++stats::PaddingFragmentsRelaxations;
928   stats::PaddingFragmentsBytes += NewSize;
929   stats::PaddingFragmentsBytes -= OldSize;
930   return true;
931 }
932 
relaxLEB(MCAsmLayout & Layout,MCLEBFragment & LF)933 bool MCAssembler::relaxLEB(MCAsmLayout &Layout, MCLEBFragment &LF) {
934   uint64_t OldSize = LF.getContents().size();
935   int64_t Value;
936   bool Abs = LF.getValue().evaluateKnownAbsolute(Value, Layout);
937   if (!Abs)
938     report_fatal_error("sleb128 and uleb128 expressions must be absolute");
939   SmallString<8> &Data = LF.getContents();
940   Data.clear();
941   raw_svector_ostream OSE(Data);
942   // The compiler can generate EH table assembly that is impossible to assemble
943   // without either adding padding to an LEB fragment or adding extra padding
944   // to a later alignment fragment. To accommodate such tables, relaxation can
945   // only increase an LEB fragment size here, not decrease it. See PR35809.
946   if (LF.isSigned())
947     encodeSLEB128(Value, OSE, OldSize);
948   else
949     encodeULEB128(Value, OSE, OldSize);
950   return OldSize != LF.getContents().size();
951 }
952 
relaxDwarfLineAddr(MCAsmLayout & Layout,MCDwarfLineAddrFragment & DF)953 bool MCAssembler::relaxDwarfLineAddr(MCAsmLayout &Layout,
954                                      MCDwarfLineAddrFragment &DF) {
955   MCContext &Context = Layout.getAssembler().getContext();
956   uint64_t OldSize = DF.getContents().size();
957   int64_t AddrDelta;
958   bool Abs;
959   if (getBackend().requiresDiffExpressionRelocations())
960     Abs = DF.getAddrDelta().evaluateAsAbsolute(AddrDelta, Layout);
961   else {
962     Abs = DF.getAddrDelta().evaluateKnownAbsolute(AddrDelta, Layout);
963     assert(Abs && "We created a line delta with an invalid expression");
964   }
965   int64_t LineDelta;
966   LineDelta = DF.getLineDelta();
967   SmallVectorImpl<char> &Data = DF.getContents();
968   Data.clear();
969   raw_svector_ostream OSE(Data);
970   DF.getFixups().clear();
971 
972   if (Abs) {
973     MCDwarfLineAddr::Encode(Context, getDWARFLinetableParams(), LineDelta,
974                             AddrDelta, OSE);
975   } else {
976     uint32_t Offset;
977     uint32_t Size;
978     bool SetDelta = MCDwarfLineAddr::FixedEncode(Context,
979                                                  getDWARFLinetableParams(),
980                                                  LineDelta, AddrDelta,
981                                                  OSE, &Offset, &Size);
982     // Add Fixups for address delta or new address.
983     const MCExpr *FixupExpr;
984     if (SetDelta) {
985       FixupExpr = &DF.getAddrDelta();
986     } else {
987       const MCBinaryExpr *ABE = cast<MCBinaryExpr>(&DF.getAddrDelta());
988       FixupExpr = ABE->getLHS();
989     }
990     DF.getFixups().push_back(
991         MCFixup::create(Offset, FixupExpr,
992                         MCFixup::getKindForSize(Size, false /*isPCRel*/)));
993   }
994 
995   return OldSize != Data.size();
996 }
997 
relaxDwarfCallFrameFragment(MCAsmLayout & Layout,MCDwarfCallFrameFragment & DF)998 bool MCAssembler::relaxDwarfCallFrameFragment(MCAsmLayout &Layout,
999                                               MCDwarfCallFrameFragment &DF) {
1000   MCContext &Context = Layout.getAssembler().getContext();
1001   uint64_t OldSize = DF.getContents().size();
1002   int64_t AddrDelta;
1003   bool Abs = DF.getAddrDelta().evaluateKnownAbsolute(AddrDelta, Layout);
1004   assert(Abs && "We created call frame with an invalid expression");
1005   (void) Abs;
1006   SmallString<8> &Data = DF.getContents();
1007   Data.clear();
1008   raw_svector_ostream OSE(Data);
1009   MCDwarfFrameEmitter::EncodeAdvanceLoc(Context, AddrDelta, OSE);
1010   return OldSize != Data.size();
1011 }
1012 
relaxCVInlineLineTable(MCAsmLayout & Layout,MCCVInlineLineTableFragment & F)1013 bool MCAssembler::relaxCVInlineLineTable(MCAsmLayout &Layout,
1014                                          MCCVInlineLineTableFragment &F) {
1015   unsigned OldSize = F.getContents().size();
1016   getContext().getCVContext().encodeInlineLineTable(Layout, F);
1017   return OldSize != F.getContents().size();
1018 }
1019 
relaxCVDefRange(MCAsmLayout & Layout,MCCVDefRangeFragment & F)1020 bool MCAssembler::relaxCVDefRange(MCAsmLayout &Layout,
1021                                   MCCVDefRangeFragment &F) {
1022   unsigned OldSize = F.getContents().size();
1023   getContext().getCVContext().encodeDefRange(Layout, F);
1024   return OldSize != F.getContents().size();
1025 }
1026 
layoutSectionOnce(MCAsmLayout & Layout,MCSection & Sec)1027 bool MCAssembler::layoutSectionOnce(MCAsmLayout &Layout, MCSection &Sec) {
1028   // Holds the first fragment which needed relaxing during this layout. It will
1029   // remain NULL if none were relaxed.
1030   // When a fragment is relaxed, all the fragments following it should get
1031   // invalidated because their offset is going to change.
1032   MCFragment *FirstRelaxedFragment = nullptr;
1033 
1034   // Attempt to relax all the fragments in the section.
1035   for (MCSection::iterator I = Sec.begin(), IE = Sec.end(); I != IE; ++I) {
1036     // Check if this is a fragment that needs relaxation.
1037     bool RelaxedFrag = false;
1038     switch(I->getKind()) {
1039     default:
1040       break;
1041     case MCFragment::FT_Relaxable:
1042       assert(!getRelaxAll() &&
1043              "Did not expect a MCRelaxableFragment in RelaxAll mode");
1044       RelaxedFrag = relaxInstruction(Layout, *cast<MCRelaxableFragment>(I));
1045       break;
1046     case MCFragment::FT_Dwarf:
1047       RelaxedFrag = relaxDwarfLineAddr(Layout,
1048                                        *cast<MCDwarfLineAddrFragment>(I));
1049       break;
1050     case MCFragment::FT_DwarfFrame:
1051       RelaxedFrag =
1052         relaxDwarfCallFrameFragment(Layout,
1053                                     *cast<MCDwarfCallFrameFragment>(I));
1054       break;
1055     case MCFragment::FT_LEB:
1056       RelaxedFrag = relaxLEB(Layout, *cast<MCLEBFragment>(I));
1057       break;
1058     case MCFragment::FT_Padding:
1059       RelaxedFrag = relaxPaddingFragment(Layout, *cast<MCPaddingFragment>(I));
1060       break;
1061     case MCFragment::FT_CVInlineLines:
1062       RelaxedFrag =
1063           relaxCVInlineLineTable(Layout, *cast<MCCVInlineLineTableFragment>(I));
1064       break;
1065     case MCFragment::FT_CVDefRange:
1066       RelaxedFrag = relaxCVDefRange(Layout, *cast<MCCVDefRangeFragment>(I));
1067       break;
1068     }
1069     if (RelaxedFrag && !FirstRelaxedFragment)
1070       FirstRelaxedFragment = &*I;
1071   }
1072   if (FirstRelaxedFragment) {
1073     Layout.invalidateFragmentsFrom(FirstRelaxedFragment);
1074     return true;
1075   }
1076   return false;
1077 }
1078 
layoutOnce(MCAsmLayout & Layout)1079 bool MCAssembler::layoutOnce(MCAsmLayout &Layout) {
1080   ++stats::RelaxationSteps;
1081 
1082   bool WasRelaxed = false;
1083   for (iterator it = begin(), ie = end(); it != ie; ++it) {
1084     MCSection &Sec = *it;
1085     while (layoutSectionOnce(Layout, Sec))
1086       WasRelaxed = true;
1087   }
1088 
1089   return WasRelaxed;
1090 }
1091 
finishLayout(MCAsmLayout & Layout)1092 void MCAssembler::finishLayout(MCAsmLayout &Layout) {
1093   assert(getBackendPtr() && "Expected assembler backend");
1094   // The layout is done. Mark every fragment as valid.
1095   for (unsigned int i = 0, n = Layout.getSectionOrder().size(); i != n; ++i) {
1096     MCSection &Section = *Layout.getSectionOrder()[i];
1097     Layout.getFragmentOffset(&*Section.rbegin());
1098     computeFragmentSize(Layout, *Section.rbegin());
1099   }
1100   getBackend().finishLayout(*this, Layout);
1101 }
1102 
1103 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump() const1104 LLVM_DUMP_METHOD void MCAssembler::dump() const{
1105   raw_ostream &OS = errs();
1106 
1107   OS << "<MCAssembler\n";
1108   OS << "  Sections:[\n    ";
1109   for (const_iterator it = begin(), ie = end(); it != ie; ++it) {
1110     if (it != begin()) OS << ",\n    ";
1111     it->dump();
1112   }
1113   OS << "],\n";
1114   OS << "  Symbols:[";
1115 
1116   for (const_symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
1117     if (it != symbol_begin()) OS << ",\n           ";
1118     OS << "(";
1119     it->dump();
1120     OS << ", Index:" << it->getIndex() << ", ";
1121     OS << ")";
1122   }
1123   OS << "]>\n";
1124 }
1125 #endif
1126