1 //===- lib/MC/MCAssembler.cpp - Assembler Backend Implementation ----------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 #include "llvm/MC/MCAssembler.h"
10 #include "llvm/ADT/ArrayRef.h"
11 #include "llvm/ADT/SmallString.h"
12 #include "llvm/ADT/SmallVector.h"
13 #include "llvm/ADT/Statistic.h"
14 #include "llvm/ADT/StringRef.h"
15 #include "llvm/ADT/Twine.h"
16 #include "llvm/MC/MCAsmBackend.h"
17 #include "llvm/MC/MCAsmInfo.h"
18 #include "llvm/MC/MCAsmLayout.h"
19 #include "llvm/MC/MCCodeEmitter.h"
20 #include "llvm/MC/MCCodeView.h"
21 #include "llvm/MC/MCContext.h"
22 #include "llvm/MC/MCDwarf.h"
23 #include "llvm/MC/MCExpr.h"
24 #include "llvm/MC/MCFixup.h"
25 #include "llvm/MC/MCFixupKindInfo.h"
26 #include "llvm/MC/MCFragment.h"
27 #include "llvm/MC/MCInst.h"
28 #include "llvm/MC/MCObjectWriter.h"
29 #include "llvm/MC/MCSection.h"
30 #include "llvm/MC/MCSectionELF.h"
31 #include "llvm/MC/MCSymbol.h"
32 #include "llvm/MC/MCValue.h"
33 #include "llvm/Support/Alignment.h"
34 #include "llvm/Support/Casting.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/EndianStream.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/LEB128.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/raw_ostream.h"
49 #define DEBUG_TYPE "assembler"
54 STATISTIC(EmittedFragments
, "Number of emitted assembler fragments - total");
55 STATISTIC(EmittedRelaxableFragments
,
56 "Number of emitted assembler fragments - relaxable");
57 STATISTIC(EmittedDataFragments
,
58 "Number of emitted assembler fragments - data");
59 STATISTIC(EmittedCompactEncodedInstFragments
,
60 "Number of emitted assembler fragments - compact encoded inst");
61 STATISTIC(EmittedAlignFragments
,
62 "Number of emitted assembler fragments - align");
63 STATISTIC(EmittedFillFragments
,
64 "Number of emitted assembler fragments - fill");
65 STATISTIC(EmittedNopsFragments
, "Number of emitted assembler fragments - nops");
66 STATISTIC(EmittedOrgFragments
, "Number of emitted assembler fragments - org");
67 STATISTIC(evaluateFixup
, "Number of evaluated fixups");
68 STATISTIC(FragmentLayouts
, "Number of fragment layouts");
69 STATISTIC(ObjectBytes
, "Number of emitted object file bytes");
70 STATISTIC(RelaxationSteps
, "Number of assembler layout and relaxation steps");
71 STATISTIC(RelaxedInstructions
, "Number of relaxed instructions");
73 } // end namespace stats
74 } // end anonymous namespace
76 // FIXME FIXME FIXME: There are number of places in this file where we convert
77 // what is a 64-bit assembler value used for computation into a value in the
78 // object file, which may truncate it. We should detect that truncation where
79 // invalid and report errors back.
83 MCAssembler::MCAssembler(MCContext
&Context
,
84 std::unique_ptr
<MCAsmBackend
> Backend
,
85 std::unique_ptr
<MCCodeEmitter
> Emitter
,
86 std::unique_ptr
<MCObjectWriter
> Writer
)
87 : Context(Context
), Backend(std::move(Backend
)),
88 Emitter(std::move(Emitter
)), Writer(std::move(Writer
)),
89 BundleAlignSize(0), RelaxAll(false), SubsectionsViaSymbols(false),
90 IncrementalLinkerCompatible(false), ELFHeaderEFlags(0) {
91 VersionInfo
.Major
= 0; // Major version == 0 for "none specified"
94 MCAssembler::~MCAssembler() = default;
96 void MCAssembler::reset() {
99 IndirectSymbols
.clear();
101 LinkerOptions
.clear();
106 SubsectionsViaSymbols
= false;
107 IncrementalLinkerCompatible
= false;
109 LOHContainer
.reset();
110 VersionInfo
.Major
= 0;
111 VersionInfo
.SDKVersion
= VersionTuple();
113 // reset objects owned by us
115 getBackendPtr()->reset();
117 getEmitterPtr()->reset();
119 getWriterPtr()->reset();
120 getLOHContainer().reset();
123 bool MCAssembler::registerSection(MCSection
&Section
) {
124 if (Section
.isRegistered())
126 Sections
.push_back(&Section
);
127 Section
.setIsRegistered(true);
131 bool MCAssembler::isThumbFunc(const MCSymbol
*Symbol
) const {
132 if (ThumbFuncs
.count(Symbol
))
135 if (!Symbol
->isVariable())
138 const MCExpr
*Expr
= Symbol
->getVariableValue();
141 if (!Expr
->evaluateAsRelocatable(V
, nullptr, nullptr))
144 if (V
.getSymB() || V
.getRefKind() != MCSymbolRefExpr::VK_None
)
147 const MCSymbolRefExpr
*Ref
= V
.getSymA();
151 if (Ref
->getKind() != MCSymbolRefExpr::VK_None
)
154 const MCSymbol
&Sym
= Ref
->getSymbol();
155 if (!isThumbFunc(&Sym
))
158 ThumbFuncs
.insert(Symbol
); // Cache it.
162 bool MCAssembler::isSymbolLinkerVisible(const MCSymbol
&Symbol
) const {
163 // Non-temporary labels should always be visible to the linker.
164 if (!Symbol
.isTemporary())
167 if (Symbol
.isUsedInReloc())
173 const MCSymbol
*MCAssembler::getAtom(const MCSymbol
&S
) const {
174 // Linker visible symbols define atoms.
175 if (isSymbolLinkerVisible(S
))
178 // Absolute and undefined symbols have no defining atom.
179 if (!S
.isInSection())
182 // Non-linker visible symbols in sections which can't be atomized have no
184 if (!getContext().getAsmInfo()->isSectionAtomizableBySymbols(
185 *S
.getFragment()->getParent()))
188 // Otherwise, return the atom for the containing fragment.
189 return S
.getFragment()->getAtom();
192 bool MCAssembler::evaluateFixup(const MCAsmLayout
&Layout
,
193 const MCFixup
&Fixup
, const MCFragment
*DF
,
194 MCValue
&Target
, uint64_t &Value
,
195 bool &WasForced
) const {
196 ++stats::evaluateFixup
;
198 // FIXME: This code has some duplication with recordRelocation. We should
199 // probably merge the two into a single callback that tries to evaluate a
200 // fixup and records a relocation if one is needed.
202 // On error claim to have completely evaluated the fixup, to prevent any
203 // further processing from being done.
204 const MCExpr
*Expr
= Fixup
.getValue();
205 MCContext
&Ctx
= getContext();
208 if (!Expr
->evaluateAsRelocatable(Target
, &Layout
, &Fixup
)) {
209 Ctx
.reportError(Fixup
.getLoc(), "expected relocatable expression");
212 if (const MCSymbolRefExpr
*RefB
= Target
.getSymB()) {
213 if (RefB
->getKind() != MCSymbolRefExpr::VK_None
) {
214 Ctx
.reportError(Fixup
.getLoc(),
215 "unsupported subtraction of qualified symbol");
220 assert(getBackendPtr() && "Expected assembler backend");
221 bool IsTarget
= getBackendPtr()->getFixupKindInfo(Fixup
.getKind()).Flags
&
222 MCFixupKindInfo::FKF_IsTarget
;
225 return getBackend().evaluateTargetFixup(*this, Layout
, Fixup
, DF
, Target
,
228 unsigned FixupFlags
= getBackendPtr()->getFixupKindInfo(Fixup
.getKind()).Flags
;
229 bool IsPCRel
= getBackendPtr()->getFixupKindInfo(Fixup
.getKind()).Flags
&
230 MCFixupKindInfo::FKF_IsPCRel
;
232 bool IsResolved
= false;
234 if (Target
.getSymB()) {
236 } else if (!Target
.getSymA()) {
239 const MCSymbolRefExpr
*A
= Target
.getSymA();
240 const MCSymbol
&SA
= A
->getSymbol();
241 if (A
->getKind() != MCSymbolRefExpr::VK_None
|| SA
.isUndefined()) {
243 } else if (auto *Writer
= getWriterPtr()) {
244 IsResolved
= (FixupFlags
& MCFixupKindInfo::FKF_Constant
) ||
245 Writer
->isSymbolRefDifferenceFullyResolvedImpl(
246 *this, SA
, *DF
, false, true);
250 IsResolved
= Target
.isAbsolute();
253 Value
= Target
.getConstant();
255 if (const MCSymbolRefExpr
*A
= Target
.getSymA()) {
256 const MCSymbol
&Sym
= A
->getSymbol();
258 Value
+= Layout
.getSymbolOffset(Sym
);
260 if (const MCSymbolRefExpr
*B
= Target
.getSymB()) {
261 const MCSymbol
&Sym
= B
->getSymbol();
263 Value
-= Layout
.getSymbolOffset(Sym
);
266 bool ShouldAlignPC
= getBackend().getFixupKindInfo(Fixup
.getKind()).Flags
&
267 MCFixupKindInfo::FKF_IsAlignedDownTo32Bits
;
268 assert((ShouldAlignPC
? IsPCRel
: true) &&
269 "FKF_IsAlignedDownTo32Bits is only allowed on PC-relative fixups!");
272 uint32_t Offset
= Layout
.getFragmentOffset(DF
) + Fixup
.getOffset();
274 // A number of ARM fixups in Thumb mode require that the effective PC
275 // address be determined as the 32-bit aligned version of the actual offset.
276 if (ShouldAlignPC
) Offset
&= ~0x3;
280 // Let the backend force a relocation if needed.
281 if (IsResolved
&& getBackend().shouldForceRelocation(*this, Fixup
, Target
)) {
289 uint64_t MCAssembler::computeFragmentSize(const MCAsmLayout
&Layout
,
290 const MCFragment
&F
) const {
291 assert(getBackendPtr() && "Requires assembler backend");
292 switch (F
.getKind()) {
293 case MCFragment::FT_Data
:
294 return cast
<MCDataFragment
>(F
).getContents().size();
295 case MCFragment::FT_Relaxable
:
296 return cast
<MCRelaxableFragment
>(F
).getContents().size();
297 case MCFragment::FT_CompactEncodedInst
:
298 return cast
<MCCompactEncodedInstFragment
>(F
).getContents().size();
299 case MCFragment::FT_Fill
: {
300 auto &FF
= cast
<MCFillFragment
>(F
);
301 int64_t NumValues
= 0;
302 if (!FF
.getNumValues().evaluateAsAbsolute(NumValues
, Layout
)) {
303 getContext().reportError(FF
.getLoc(),
304 "expected assembly-time absolute expression");
307 int64_t Size
= NumValues
* FF
.getValueSize();
309 getContext().reportError(FF
.getLoc(), "invalid number of bytes");
315 case MCFragment::FT_Nops
:
316 return cast
<MCNopsFragment
>(F
).getNumBytes();
318 case MCFragment::FT_LEB
:
319 return cast
<MCLEBFragment
>(F
).getContents().size();
321 case MCFragment::FT_BoundaryAlign
:
322 return cast
<MCBoundaryAlignFragment
>(F
).getSize();
324 case MCFragment::FT_SymbolId
:
327 case MCFragment::FT_Align
: {
328 const MCAlignFragment
&AF
= cast
<MCAlignFragment
>(F
);
329 unsigned Offset
= Layout
.getFragmentOffset(&AF
);
330 unsigned Size
= offsetToAlignment(Offset
, Align(AF
.getAlignment()));
332 // Insert extra Nops for code alignment if the target define
333 // shouldInsertExtraNopBytesForCodeAlign target hook.
334 if (AF
.getParent()->UseCodeAlign() && AF
.hasEmitNops() &&
335 getBackend().shouldInsertExtraNopBytesForCodeAlign(AF
, Size
))
338 // If we are padding with nops, force the padding to be larger than the
340 if (Size
> 0 && AF
.hasEmitNops()) {
341 while (Size
% getBackend().getMinimumNopSize())
342 Size
+= AF
.getAlignment();
344 if (Size
> AF
.getMaxBytesToEmit())
349 case MCFragment::FT_Org
: {
350 const MCOrgFragment
&OF
= cast
<MCOrgFragment
>(F
);
352 if (!OF
.getOffset().evaluateAsValue(Value
, Layout
)) {
353 getContext().reportError(OF
.getLoc(),
354 "expected assembly-time absolute expression");
358 uint64_t FragmentOffset
= Layout
.getFragmentOffset(&OF
);
359 int64_t TargetLocation
= Value
.getConstant();
360 if (const MCSymbolRefExpr
*A
= Value
.getSymA()) {
362 if (!Layout
.getSymbolOffset(A
->getSymbol(), Val
)) {
363 getContext().reportError(OF
.getLoc(), "expected absolute expression");
366 TargetLocation
+= Val
;
368 int64_t Size
= TargetLocation
- FragmentOffset
;
369 if (Size
< 0 || Size
>= 0x40000000) {
370 getContext().reportError(
371 OF
.getLoc(), "invalid .org offset '" + Twine(TargetLocation
) +
372 "' (at offset '" + Twine(FragmentOffset
) + "')");
378 case MCFragment::FT_Dwarf
:
379 return cast
<MCDwarfLineAddrFragment
>(F
).getContents().size();
380 case MCFragment::FT_DwarfFrame
:
381 return cast
<MCDwarfCallFrameFragment
>(F
).getContents().size();
382 case MCFragment::FT_CVInlineLines
:
383 return cast
<MCCVInlineLineTableFragment
>(F
).getContents().size();
384 case MCFragment::FT_CVDefRange
:
385 return cast
<MCCVDefRangeFragment
>(F
).getContents().size();
386 case MCFragment::FT_PseudoProbe
:
387 return cast
<MCPseudoProbeAddrFragment
>(F
).getContents().size();
388 case MCFragment::FT_Dummy
:
389 llvm_unreachable("Should not have been added");
392 llvm_unreachable("invalid fragment kind");
395 void MCAsmLayout::layoutFragment(MCFragment
*F
) {
396 MCFragment
*Prev
= F
->getPrevNode();
398 // We should never try to recompute something which is valid.
399 assert(!isFragmentValid(F
) && "Attempt to recompute a valid fragment!");
400 // We should never try to compute the fragment layout if its predecessor
402 assert((!Prev
|| isFragmentValid(Prev
)) &&
403 "Attempt to compute fragment before its predecessor!");
405 assert(!F
->IsBeingLaidOut
&& "Already being laid out!");
406 F
->IsBeingLaidOut
= true;
408 ++stats::FragmentLayouts
;
410 // Compute fragment offset and size.
412 F
->Offset
= Prev
->Offset
+ getAssembler().computeFragmentSize(*this, *Prev
);
415 F
->IsBeingLaidOut
= false;
416 LastValidFragment
[F
->getParent()] = F
;
418 // If bundling is enabled and this fragment has instructions in it, it has to
419 // obey the bundling restrictions. With padding, we'll have:
424 // -------------------------------------
425 // Prev |##########| F |
426 // -------------------------------------
431 // The fragment's offset will point to after the padding, and its computed
432 // size won't include the padding.
434 // When the -mc-relax-all flag is used, we optimize bundling by writting the
435 // padding directly into fragments when the instructions are emitted inside
436 // the streamer. When the fragment is larger than the bundle size, we need to
437 // ensure that it's bundle aligned. This means that if we end up with
438 // multiple fragments, we must emit bundle padding between fragments.
440 // ".align N" is an example of a directive that introduces multiple
441 // fragments. We could add a special case to handle ".align N" by emitting
442 // within-fragment padding (which would produce less padding when N is less
443 // than the bundle size), but for now we don't.
445 if (Assembler
.isBundlingEnabled() && F
->hasInstructions()) {
446 assert(isa
<MCEncodedFragment
>(F
) &&
447 "Only MCEncodedFragment implementations have instructions");
448 MCEncodedFragment
*EF
= cast
<MCEncodedFragment
>(F
);
449 uint64_t FSize
= Assembler
.computeFragmentSize(*this, *EF
);
451 if (!Assembler
.getRelaxAll() && FSize
> Assembler
.getBundleAlignSize())
452 report_fatal_error("Fragment can't be larger than a bundle size");
454 uint64_t RequiredBundlePadding
=
455 computeBundlePadding(Assembler
, EF
, EF
->Offset
, FSize
);
456 if (RequiredBundlePadding
> UINT8_MAX
)
457 report_fatal_error("Padding cannot exceed 255 bytes");
458 EF
->setBundlePadding(static_cast<uint8_t>(RequiredBundlePadding
));
459 EF
->Offset
+= RequiredBundlePadding
;
463 void MCAssembler::registerSymbol(const MCSymbol
&Symbol
, bool *Created
) {
464 bool New
= !Symbol
.isRegistered();
468 Symbol
.setIsRegistered(true);
469 Symbols
.push_back(&Symbol
);
473 void MCAssembler::writeFragmentPadding(raw_ostream
&OS
,
474 const MCEncodedFragment
&EF
,
475 uint64_t FSize
) const {
476 assert(getBackendPtr() && "Expected assembler backend");
477 // Should NOP padding be written out before this fragment?
478 unsigned BundlePadding
= EF
.getBundlePadding();
479 if (BundlePadding
> 0) {
480 assert(isBundlingEnabled() &&
481 "Writing bundle padding with disabled bundling");
482 assert(EF
.hasInstructions() &&
483 "Writing bundle padding for a fragment without instructions");
485 unsigned TotalLength
= BundlePadding
+ static_cast<unsigned>(FSize
);
486 if (EF
.alignToBundleEnd() && TotalLength
> getBundleAlignSize()) {
487 // If the padding itself crosses a bundle boundary, it must be emitted
488 // in 2 pieces, since even nop instructions must not cross boundaries.
489 // v--------------v <- BundleAlignSize
490 // v---------v <- BundlePadding
491 // ----------------------------
492 // | Prev |####|####| F |
493 // ----------------------------
494 // ^-------------------^ <- TotalLength
495 unsigned DistanceToBoundary
= TotalLength
- getBundleAlignSize();
496 if (!getBackend().writeNopData(OS
, DistanceToBoundary
))
497 report_fatal_error("unable to write NOP sequence of " +
498 Twine(DistanceToBoundary
) + " bytes");
499 BundlePadding
-= DistanceToBoundary
;
501 if (!getBackend().writeNopData(OS
, BundlePadding
))
502 report_fatal_error("unable to write NOP sequence of " +
503 Twine(BundlePadding
) + " bytes");
507 /// Write the fragment \p F to the output file.
508 static void writeFragment(raw_ostream
&OS
, const MCAssembler
&Asm
,
509 const MCAsmLayout
&Layout
, const MCFragment
&F
) {
510 // FIXME: Embed in fragments instead?
511 uint64_t FragmentSize
= Asm
.computeFragmentSize(Layout
, F
);
513 support::endianness Endian
= Asm
.getBackend().Endian
;
515 if (const MCEncodedFragment
*EF
= dyn_cast
<MCEncodedFragment
>(&F
))
516 Asm
.writeFragmentPadding(OS
, *EF
, FragmentSize
);
518 // This variable (and its dummy usage) is to participate in the assert at
519 // the end of the function.
520 uint64_t Start
= OS
.tell();
523 ++stats::EmittedFragments
;
525 switch (F
.getKind()) {
526 case MCFragment::FT_Align
: {
527 ++stats::EmittedAlignFragments
;
528 const MCAlignFragment
&AF
= cast
<MCAlignFragment
>(F
);
529 assert(AF
.getValueSize() && "Invalid virtual align in concrete fragment!");
531 uint64_t Count
= FragmentSize
/ AF
.getValueSize();
533 // FIXME: This error shouldn't actually occur (the front end should emit
534 // multiple .align directives to enforce the semantics it wants), but is
535 // severe enough that we want to report it. How to handle this?
536 if (Count
* AF
.getValueSize() != FragmentSize
)
537 report_fatal_error("undefined .align directive, value size '" +
538 Twine(AF
.getValueSize()) +
539 "' is not a divisor of padding size '" +
540 Twine(FragmentSize
) + "'");
542 // See if we are aligning with nops, and if so do that first to try to fill
543 // the Count bytes. Then if that did not fill any bytes or there are any
544 // bytes left to fill use the Value and ValueSize to fill the rest.
545 // If we are aligning with nops, ask that target to emit the right data.
546 if (AF
.hasEmitNops()) {
547 if (!Asm
.getBackend().writeNopData(OS
, Count
))
548 report_fatal_error("unable to write nop sequence of " +
549 Twine(Count
) + " bytes");
553 // Otherwise, write out in multiples of the value size.
554 for (uint64_t i
= 0; i
!= Count
; ++i
) {
555 switch (AF
.getValueSize()) {
556 default: llvm_unreachable("Invalid size!");
557 case 1: OS
<< char(AF
.getValue()); break;
559 support::endian::write
<uint16_t>(OS
, AF
.getValue(), Endian
);
562 support::endian::write
<uint32_t>(OS
, AF
.getValue(), Endian
);
565 support::endian::write
<uint64_t>(OS
, AF
.getValue(), Endian
);
572 case MCFragment::FT_Data
:
573 ++stats::EmittedDataFragments
;
574 OS
<< cast
<MCDataFragment
>(F
).getContents();
577 case MCFragment::FT_Relaxable
:
578 ++stats::EmittedRelaxableFragments
;
579 OS
<< cast
<MCRelaxableFragment
>(F
).getContents();
582 case MCFragment::FT_CompactEncodedInst
:
583 ++stats::EmittedCompactEncodedInstFragments
;
584 OS
<< cast
<MCCompactEncodedInstFragment
>(F
).getContents();
587 case MCFragment::FT_Fill
: {
588 ++stats::EmittedFillFragments
;
589 const MCFillFragment
&FF
= cast
<MCFillFragment
>(F
);
590 uint64_t V
= FF
.getValue();
591 unsigned VSize
= FF
.getValueSize();
592 const unsigned MaxChunkSize
= 16;
593 char Data
[MaxChunkSize
];
594 assert(0 < VSize
&& VSize
<= MaxChunkSize
&& "Illegal fragment fill size");
595 // Duplicate V into Data as byte vector to reduce number of
596 // writes done. As such, do endian conversion here.
597 for (unsigned I
= 0; I
!= VSize
; ++I
) {
598 unsigned index
= Endian
== support::little
? I
: (VSize
- I
- 1);
599 Data
[I
] = uint8_t(V
>> (index
* 8));
601 for (unsigned I
= VSize
; I
< MaxChunkSize
; ++I
)
602 Data
[I
] = Data
[I
- VSize
];
604 // Set to largest multiple of VSize in Data.
605 const unsigned NumPerChunk
= MaxChunkSize
/ VSize
;
606 // Set ChunkSize to largest multiple of VSize in Data
607 const unsigned ChunkSize
= VSize
* NumPerChunk
;
609 // Do copies by chunk.
610 StringRef
Ref(Data
, ChunkSize
);
611 for (uint64_t I
= 0, E
= FragmentSize
/ ChunkSize
; I
!= E
; ++I
)
614 // do remainder if needed.
615 unsigned TrailingCount
= FragmentSize
% ChunkSize
;
617 OS
.write(Data
, TrailingCount
);
621 case MCFragment::FT_Nops
: {
622 ++stats::EmittedNopsFragments
;
623 const MCNopsFragment
&NF
= cast
<MCNopsFragment
>(F
);
624 int64_t NumBytes
= NF
.getNumBytes();
625 int64_t ControlledNopLength
= NF
.getControlledNopLength();
626 int64_t MaximumNopLength
= Asm
.getBackend().getMaximumNopSize();
628 assert(NumBytes
> 0 && "Expected positive NOPs fragment size");
629 assert(ControlledNopLength
>= 0 && "Expected non-negative NOP size");
631 if (ControlledNopLength
> MaximumNopLength
) {
632 Asm
.getContext().reportError(NF
.getLoc(),
633 "illegal NOP size " +
634 std::to_string(ControlledNopLength
) +
635 ". (expected within [0, " +
636 std::to_string(MaximumNopLength
) + "])");
637 // Clamp the NOP length as reportError does not stop the execution
639 ControlledNopLength
= MaximumNopLength
;
642 // Use maximum value if the size of each NOP is not specified
643 if (!ControlledNopLength
)
644 ControlledNopLength
= MaximumNopLength
;
647 uint64_t NumBytesToEmit
=
648 (uint64_t)std::min(NumBytes
, ControlledNopLength
);
649 assert(NumBytesToEmit
&& "try to emit empty NOP instruction");
650 if (!Asm
.getBackend().writeNopData(OS
, NumBytesToEmit
)) {
651 report_fatal_error("unable to write nop sequence of the remaining " +
652 Twine(NumBytesToEmit
) + " bytes");
655 NumBytes
-= NumBytesToEmit
;
660 case MCFragment::FT_LEB
: {
661 const MCLEBFragment
&LF
= cast
<MCLEBFragment
>(F
);
662 OS
<< LF
.getContents();
666 case MCFragment::FT_BoundaryAlign
: {
667 if (!Asm
.getBackend().writeNopData(OS
, FragmentSize
))
668 report_fatal_error("unable to write nop sequence of " +
669 Twine(FragmentSize
) + " bytes");
673 case MCFragment::FT_SymbolId
: {
674 const MCSymbolIdFragment
&SF
= cast
<MCSymbolIdFragment
>(F
);
675 support::endian::write
<uint32_t>(OS
, SF
.getSymbol()->getIndex(), Endian
);
679 case MCFragment::FT_Org
: {
680 ++stats::EmittedOrgFragments
;
681 const MCOrgFragment
&OF
= cast
<MCOrgFragment
>(F
);
683 for (uint64_t i
= 0, e
= FragmentSize
; i
!= e
; ++i
)
684 OS
<< char(OF
.getValue());
689 case MCFragment::FT_Dwarf
: {
690 const MCDwarfLineAddrFragment
&OF
= cast
<MCDwarfLineAddrFragment
>(F
);
691 OS
<< OF
.getContents();
694 case MCFragment::FT_DwarfFrame
: {
695 const MCDwarfCallFrameFragment
&CF
= cast
<MCDwarfCallFrameFragment
>(F
);
696 OS
<< CF
.getContents();
699 case MCFragment::FT_CVInlineLines
: {
700 const auto &OF
= cast
<MCCVInlineLineTableFragment
>(F
);
701 OS
<< OF
.getContents();
704 case MCFragment::FT_CVDefRange
: {
705 const auto &DRF
= cast
<MCCVDefRangeFragment
>(F
);
706 OS
<< DRF
.getContents();
709 case MCFragment::FT_PseudoProbe
: {
710 const MCPseudoProbeAddrFragment
&PF
= cast
<MCPseudoProbeAddrFragment
>(F
);
711 OS
<< PF
.getContents();
714 case MCFragment::FT_Dummy
:
715 llvm_unreachable("Should not have been added");
718 assert(OS
.tell() - Start
== FragmentSize
&&
719 "The stream should advance by fragment size");
722 void MCAssembler::writeSectionData(raw_ostream
&OS
, const MCSection
*Sec
,
723 const MCAsmLayout
&Layout
) const {
724 assert(getBackendPtr() && "Expected assembler backend");
726 // Ignore virtual sections.
727 if (Sec
->isVirtualSection()) {
728 assert(Layout
.getSectionFileSize(Sec
) == 0 && "Invalid size for section!");
730 // Check that contents are only things legal inside a virtual section.
731 for (const MCFragment
&F
: *Sec
) {
732 switch (F
.getKind()) {
733 default: llvm_unreachable("Invalid fragment in virtual section!");
734 case MCFragment::FT_Data
: {
735 // Check that we aren't trying to write a non-zero contents (or fixups)
736 // into a virtual section. This is to support clients which use standard
737 // directives to fill the contents of virtual sections.
738 const MCDataFragment
&DF
= cast
<MCDataFragment
>(F
);
739 if (DF
.fixup_begin() != DF
.fixup_end())
740 getContext().reportError(SMLoc(), Sec
->getVirtualSectionKind() +
741 " section '" + Sec
->getName() +
742 "' cannot have fixups");
743 for (unsigned i
= 0, e
= DF
.getContents().size(); i
!= e
; ++i
)
744 if (DF
.getContents()[i
]) {
745 getContext().reportError(SMLoc(),
746 Sec
->getVirtualSectionKind() +
747 " section '" + Sec
->getName() +
748 "' cannot have non-zero initializers");
753 case MCFragment::FT_Align
:
754 // Check that we aren't trying to write a non-zero value into a virtual
756 assert((cast
<MCAlignFragment
>(F
).getValueSize() == 0 ||
757 cast
<MCAlignFragment
>(F
).getValue() == 0) &&
758 "Invalid align in virtual section!");
760 case MCFragment::FT_Fill
:
761 assert((cast
<MCFillFragment
>(F
).getValue() == 0) &&
762 "Invalid fill in virtual section!");
764 case MCFragment::FT_Org
:
772 uint64_t Start
= OS
.tell();
775 for (const MCFragment
&F
: *Sec
)
776 writeFragment(OS
, *this, Layout
, F
);
778 assert(getContext().hadError() ||
779 OS
.tell() - Start
== Layout
.getSectionAddressSize(Sec
));
782 std::tuple
<MCValue
, uint64_t, bool>
783 MCAssembler::handleFixup(const MCAsmLayout
&Layout
, MCFragment
&F
,
784 const MCFixup
&Fixup
) {
785 // Evaluate the fixup.
789 bool IsResolved
= evaluateFixup(Layout
, Fixup
, &F
, Target
, FixedValue
,
792 // The fixup was unresolved, we need a relocation. Inform the object
793 // writer of the relocation, and give it an opportunity to adjust the
794 // fixup value if need be.
795 getWriter().recordRelocation(*this, Layout
, &F
, Fixup
, Target
, FixedValue
);
797 return std::make_tuple(Target
, FixedValue
, IsResolved
);
800 void MCAssembler::layout(MCAsmLayout
&Layout
) {
801 assert(getBackendPtr() && "Expected assembler backend");
802 DEBUG_WITH_TYPE("mc-dump", {
803 errs() << "assembler backend - pre-layout\n--\n";
806 // Create dummy fragments and assign section ordinals.
807 unsigned SectionIndex
= 0;
808 for (MCSection
&Sec
: *this) {
809 // Create dummy fragments to eliminate any empty sections, this simplifies
811 if (Sec
.getFragmentList().empty())
812 new MCDataFragment(&Sec
);
814 Sec
.setOrdinal(SectionIndex
++);
817 // Assign layout order indices to sections and fragments.
818 for (unsigned i
= 0, e
= Layout
.getSectionOrder().size(); i
!= e
; ++i
) {
819 MCSection
*Sec
= Layout
.getSectionOrder()[i
];
820 Sec
->setLayoutOrder(i
);
822 unsigned FragmentIndex
= 0;
823 for (MCFragment
&Frag
: *Sec
)
824 Frag
.setLayoutOrder(FragmentIndex
++);
827 // Layout until everything fits.
828 while (layoutOnce(Layout
)) {
829 if (getContext().hadError())
831 // Size of fragments in one section can depend on the size of fragments in
832 // another. If any fragment has changed size, we have to re-layout (and
833 // as a result possibly further relax) all.
834 for (MCSection
&Sec
: *this)
835 Layout
.invalidateFragmentsFrom(&*Sec
.begin());
838 DEBUG_WITH_TYPE("mc-dump", {
839 errs() << "assembler backend - post-relaxation\n--\n";
842 // Finalize the layout, including fragment lowering.
843 finishLayout(Layout
);
845 DEBUG_WITH_TYPE("mc-dump", {
846 errs() << "assembler backend - final-layout\n--\n";
849 // Allow the object writer a chance to perform post-layout binding (for
850 // example, to set the index fields in the symbol data).
851 getWriter().executePostLayoutBinding(*this, Layout
);
853 // Evaluate and apply the fixups, generating relocation entries as necessary.
854 for (MCSection
&Sec
: *this) {
855 for (MCFragment
&Frag
: Sec
) {
856 ArrayRef
<MCFixup
> Fixups
;
857 MutableArrayRef
<char> Contents
;
858 const MCSubtargetInfo
*STI
= nullptr;
860 // Process MCAlignFragment and MCEncodedFragmentWithFixups here.
861 switch (Frag
.getKind()) {
864 case MCFragment::FT_Align
: {
865 MCAlignFragment
&AF
= cast
<MCAlignFragment
>(Frag
);
866 // Insert fixup type for code alignment if the target define
867 // shouldInsertFixupForCodeAlign target hook.
868 if (Sec
.UseCodeAlign() && AF
.hasEmitNops())
869 getBackend().shouldInsertFixupForCodeAlign(*this, Layout
, AF
);
872 case MCFragment::FT_Data
: {
873 MCDataFragment
&DF
= cast
<MCDataFragment
>(Frag
);
874 Fixups
= DF
.getFixups();
875 Contents
= DF
.getContents();
876 STI
= DF
.getSubtargetInfo();
877 assert(!DF
.hasInstructions() || STI
!= nullptr);
880 case MCFragment::FT_Relaxable
: {
881 MCRelaxableFragment
&RF
= cast
<MCRelaxableFragment
>(Frag
);
882 Fixups
= RF
.getFixups();
883 Contents
= RF
.getContents();
884 STI
= RF
.getSubtargetInfo();
885 assert(!RF
.hasInstructions() || STI
!= nullptr);
888 case MCFragment::FT_CVDefRange
: {
889 MCCVDefRangeFragment
&CF
= cast
<MCCVDefRangeFragment
>(Frag
);
890 Fixups
= CF
.getFixups();
891 Contents
= CF
.getContents();
894 case MCFragment::FT_Dwarf
: {
895 MCDwarfLineAddrFragment
&DF
= cast
<MCDwarfLineAddrFragment
>(Frag
);
896 Fixups
= DF
.getFixups();
897 Contents
= DF
.getContents();
900 case MCFragment::FT_DwarfFrame
: {
901 MCDwarfCallFrameFragment
&DF
= cast
<MCDwarfCallFrameFragment
>(Frag
);
902 Fixups
= DF
.getFixups();
903 Contents
= DF
.getContents();
906 case MCFragment::FT_PseudoProbe
: {
907 MCPseudoProbeAddrFragment
&PF
= cast
<MCPseudoProbeAddrFragment
>(Frag
);
908 Fixups
= PF
.getFixups();
909 Contents
= PF
.getContents();
913 for (const MCFixup
&Fixup
: Fixups
) {
917 std::tie(Target
, FixedValue
, IsResolved
) =
918 handleFixup(Layout
, Frag
, Fixup
);
919 getBackend().applyFixup(*this, Fixup
, Target
, Contents
, FixedValue
,
926 void MCAssembler::Finish() {
927 // Create the layout object.
928 MCAsmLayout
Layout(*this);
931 // Write the object file.
932 stats::ObjectBytes
+= getWriter().writeObject(*this, Layout
);
935 bool MCAssembler::fixupNeedsRelaxation(const MCFixup
&Fixup
,
936 const MCRelaxableFragment
*DF
,
937 const MCAsmLayout
&Layout
) const {
938 assert(getBackendPtr() && "Expected assembler backend");
942 bool Resolved
= evaluateFixup(Layout
, Fixup
, DF
, Target
, Value
, WasForced
);
943 if (Target
.getSymA() &&
944 Target
.getSymA()->getKind() == MCSymbolRefExpr::VK_X86_ABS8
&&
945 Fixup
.getKind() == FK_Data_1
)
947 return getBackend().fixupNeedsRelaxationAdvanced(Fixup
, Resolved
, Value
, DF
,
951 bool MCAssembler::fragmentNeedsRelaxation(const MCRelaxableFragment
*F
,
952 const MCAsmLayout
&Layout
) const {
953 assert(getBackendPtr() && "Expected assembler backend");
954 // If this inst doesn't ever need relaxation, ignore it. This occurs when we
955 // are intentionally pushing out inst fragments, or because we relaxed a
956 // previous instruction to one that doesn't need relaxation.
957 if (!getBackend().mayNeedRelaxation(F
->getInst(), *F
->getSubtargetInfo()))
960 for (const MCFixup
&Fixup
: F
->getFixups())
961 if (fixupNeedsRelaxation(Fixup
, F
, Layout
))
967 bool MCAssembler::relaxInstruction(MCAsmLayout
&Layout
,
968 MCRelaxableFragment
&F
) {
969 assert(getEmitterPtr() &&
970 "Expected CodeEmitter defined for relaxInstruction");
971 if (!fragmentNeedsRelaxation(&F
, Layout
))
974 ++stats::RelaxedInstructions
;
976 // FIXME-PERF: We could immediately lower out instructions if we can tell
977 // they are fully resolved, to avoid retesting on later passes.
979 // Relax the fragment.
981 MCInst Relaxed
= F
.getInst();
982 getBackend().relaxInstruction(Relaxed
, *F
.getSubtargetInfo());
984 // Encode the new instruction.
986 // FIXME-PERF: If it matters, we could let the target do this. It can
987 // probably do so more efficiently in many cases.
988 SmallVector
<MCFixup
, 4> Fixups
;
989 SmallString
<256> Code
;
990 raw_svector_ostream
VecOS(Code
);
991 getEmitter().encodeInstruction(Relaxed
, VecOS
, Fixups
, *F
.getSubtargetInfo());
993 // Update the fragment.
995 F
.getContents() = Code
;
996 F
.getFixups() = Fixups
;
1001 bool MCAssembler::relaxLEB(MCAsmLayout
&Layout
, MCLEBFragment
&LF
) {
1002 uint64_t OldSize
= LF
.getContents().size();
1004 bool Abs
= LF
.getValue().evaluateKnownAbsolute(Value
, Layout
);
1006 report_fatal_error("sleb128 and uleb128 expressions must be absolute");
1007 SmallString
<8> &Data
= LF
.getContents();
1009 raw_svector_ostream
OSE(Data
);
1010 // The compiler can generate EH table assembly that is impossible to assemble
1011 // without either adding padding to an LEB fragment or adding extra padding
1012 // to a later alignment fragment. To accommodate such tables, relaxation can
1013 // only increase an LEB fragment size here, not decrease it. See PR35809.
1015 encodeSLEB128(Value
, OSE
, OldSize
);
1017 encodeULEB128(Value
, OSE
, OldSize
);
1018 return OldSize
!= LF
.getContents().size();
1021 /// Check if the branch crosses the boundary.
1023 /// \param StartAddr start address of the fused/unfused branch.
1024 /// \param Size size of the fused/unfused branch.
1025 /// \param BoundaryAlignment alignment requirement of the branch.
1026 /// \returns true if the branch cross the boundary.
1027 static bool mayCrossBoundary(uint64_t StartAddr
, uint64_t Size
,
1028 Align BoundaryAlignment
) {
1029 uint64_t EndAddr
= StartAddr
+ Size
;
1030 return (StartAddr
>> Log2(BoundaryAlignment
)) !=
1031 ((EndAddr
- 1) >> Log2(BoundaryAlignment
));
1034 /// Check if the branch is against the boundary.
1036 /// \param StartAddr start address of the fused/unfused branch.
1037 /// \param Size size of the fused/unfused branch.
1038 /// \param BoundaryAlignment alignment requirement of the branch.
1039 /// \returns true if the branch is against the boundary.
1040 static bool isAgainstBoundary(uint64_t StartAddr
, uint64_t Size
,
1041 Align BoundaryAlignment
) {
1042 uint64_t EndAddr
= StartAddr
+ Size
;
1043 return (EndAddr
& (BoundaryAlignment
.value() - 1)) == 0;
1046 /// Check if the branch needs padding.
1048 /// \param StartAddr start address of the fused/unfused branch.
1049 /// \param Size size of the fused/unfused branch.
1050 /// \param BoundaryAlignment alignment requirement of the branch.
1051 /// \returns true if the branch needs padding.
1052 static bool needPadding(uint64_t StartAddr
, uint64_t Size
,
1053 Align BoundaryAlignment
) {
1054 return mayCrossBoundary(StartAddr
, Size
, BoundaryAlignment
) ||
1055 isAgainstBoundary(StartAddr
, Size
, BoundaryAlignment
);
1058 bool MCAssembler::relaxBoundaryAlign(MCAsmLayout
&Layout
,
1059 MCBoundaryAlignFragment
&BF
) {
1060 // BoundaryAlignFragment that doesn't need to align any fragment should not be
1062 if (!BF
.getLastFragment())
1065 uint64_t AlignedOffset
= Layout
.getFragmentOffset(&BF
);
1066 uint64_t AlignedSize
= 0;
1067 for (const MCFragment
*F
= BF
.getLastFragment(); F
!= &BF
;
1068 F
= F
->getPrevNode())
1069 AlignedSize
+= computeFragmentSize(Layout
, *F
);
1071 Align BoundaryAlignment
= BF
.getAlignment();
1072 uint64_t NewSize
= needPadding(AlignedOffset
, AlignedSize
, BoundaryAlignment
)
1073 ? offsetToAlignment(AlignedOffset
, BoundaryAlignment
)
1075 if (NewSize
== BF
.getSize())
1077 BF
.setSize(NewSize
);
1078 Layout
.invalidateFragmentsFrom(&BF
);
1082 bool MCAssembler::relaxDwarfLineAddr(MCAsmLayout
&Layout
,
1083 MCDwarfLineAddrFragment
&DF
) {
1086 if (getBackend().relaxDwarfLineAddr(DF
, Layout
, WasRelaxed
))
1089 MCContext
&Context
= Layout
.getAssembler().getContext();
1090 uint64_t OldSize
= DF
.getContents().size();
1092 bool Abs
= DF
.getAddrDelta().evaluateKnownAbsolute(AddrDelta
, Layout
);
1093 assert(Abs
&& "We created a line delta with an invalid expression");
1096 LineDelta
= DF
.getLineDelta();
1097 SmallVectorImpl
<char> &Data
= DF
.getContents();
1099 raw_svector_ostream
OSE(Data
);
1100 DF
.getFixups().clear();
1102 MCDwarfLineAddr::Encode(Context
, getDWARFLinetableParams(), LineDelta
,
1104 return OldSize
!= Data
.size();
1107 bool MCAssembler::relaxDwarfCallFrameFragment(MCAsmLayout
&Layout
,
1108 MCDwarfCallFrameFragment
&DF
) {
1110 if (getBackend().relaxDwarfCFA(DF
, Layout
, WasRelaxed
))
1113 MCContext
&Context
= Layout
.getAssembler().getContext();
1114 uint64_t OldSize
= DF
.getContents().size();
1116 bool Abs
= DF
.getAddrDelta().evaluateKnownAbsolute(AddrDelta
, Layout
);
1117 assert(Abs
&& "We created call frame with an invalid expression");
1119 SmallVectorImpl
<char> &Data
= DF
.getContents();
1121 raw_svector_ostream
OSE(Data
);
1122 DF
.getFixups().clear();
1124 MCDwarfFrameEmitter::EncodeAdvanceLoc(Context
, AddrDelta
, OSE
);
1125 return OldSize
!= Data
.size();
1128 bool MCAssembler::relaxCVInlineLineTable(MCAsmLayout
&Layout
,
1129 MCCVInlineLineTableFragment
&F
) {
1130 unsigned OldSize
= F
.getContents().size();
1131 getContext().getCVContext().encodeInlineLineTable(Layout
, F
);
1132 return OldSize
!= F
.getContents().size();
1135 bool MCAssembler::relaxCVDefRange(MCAsmLayout
&Layout
,
1136 MCCVDefRangeFragment
&F
) {
1137 unsigned OldSize
= F
.getContents().size();
1138 getContext().getCVContext().encodeDefRange(Layout
, F
);
1139 return OldSize
!= F
.getContents().size();
1142 bool MCAssembler::relaxPseudoProbeAddr(MCAsmLayout
&Layout
,
1143 MCPseudoProbeAddrFragment
&PF
) {
1144 uint64_t OldSize
= PF
.getContents().size();
1146 bool Abs
= PF
.getAddrDelta().evaluateKnownAbsolute(AddrDelta
, Layout
);
1147 assert(Abs
&& "We created a pseudo probe with an invalid expression");
1149 SmallVectorImpl
<char> &Data
= PF
.getContents();
1151 raw_svector_ostream
OSE(Data
);
1152 PF
.getFixups().clear();
1154 // AddrDelta is a signed integer
1155 encodeSLEB128(AddrDelta
, OSE
, OldSize
);
1156 return OldSize
!= Data
.size();
1159 bool MCAssembler::relaxFragment(MCAsmLayout
&Layout
, MCFragment
&F
) {
1160 switch(F
.getKind()) {
1163 case MCFragment::FT_Relaxable
:
1164 assert(!getRelaxAll() &&
1165 "Did not expect a MCRelaxableFragment in RelaxAll mode");
1166 return relaxInstruction(Layout
, cast
<MCRelaxableFragment
>(F
));
1167 case MCFragment::FT_Dwarf
:
1168 return relaxDwarfLineAddr(Layout
, cast
<MCDwarfLineAddrFragment
>(F
));
1169 case MCFragment::FT_DwarfFrame
:
1170 return relaxDwarfCallFrameFragment(Layout
,
1171 cast
<MCDwarfCallFrameFragment
>(F
));
1172 case MCFragment::FT_LEB
:
1173 return relaxLEB(Layout
, cast
<MCLEBFragment
>(F
));
1174 case MCFragment::FT_BoundaryAlign
:
1175 return relaxBoundaryAlign(Layout
, cast
<MCBoundaryAlignFragment
>(F
));
1176 case MCFragment::FT_CVInlineLines
:
1177 return relaxCVInlineLineTable(Layout
, cast
<MCCVInlineLineTableFragment
>(F
));
1178 case MCFragment::FT_CVDefRange
:
1179 return relaxCVDefRange(Layout
, cast
<MCCVDefRangeFragment
>(F
));
1180 case MCFragment::FT_PseudoProbe
:
1181 return relaxPseudoProbeAddr(Layout
, cast
<MCPseudoProbeAddrFragment
>(F
));
1185 bool MCAssembler::layoutSectionOnce(MCAsmLayout
&Layout
, MCSection
&Sec
) {
1186 // Holds the first fragment which needed relaxing during this layout. It will
1187 // remain NULL if none were relaxed.
1188 // When a fragment is relaxed, all the fragments following it should get
1189 // invalidated because their offset is going to change.
1190 MCFragment
*FirstRelaxedFragment
= nullptr;
1192 // Attempt to relax all the fragments in the section.
1193 for (MCFragment
&Frag
: Sec
) {
1194 // Check if this is a fragment that needs relaxation.
1195 bool RelaxedFrag
= relaxFragment(Layout
, Frag
);
1196 if (RelaxedFrag
&& !FirstRelaxedFragment
)
1197 FirstRelaxedFragment
= &Frag
;
1199 if (FirstRelaxedFragment
) {
1200 Layout
.invalidateFragmentsFrom(FirstRelaxedFragment
);
1206 bool MCAssembler::layoutOnce(MCAsmLayout
&Layout
) {
1207 ++stats::RelaxationSteps
;
1209 bool WasRelaxed
= false;
1210 for (MCSection
&Sec
: *this) {
1211 while (layoutSectionOnce(Layout
, Sec
))
1218 void MCAssembler::finishLayout(MCAsmLayout
&Layout
) {
1219 assert(getBackendPtr() && "Expected assembler backend");
1220 // The layout is done. Mark every fragment as valid.
1221 for (unsigned int i
= 0, n
= Layout
.getSectionOrder().size(); i
!= n
; ++i
) {
1222 MCSection
&Section
= *Layout
.getSectionOrder()[i
];
1223 Layout
.getFragmentOffset(&*Section
.getFragmentList().rbegin());
1224 computeFragmentSize(Layout
, *Section
.getFragmentList().rbegin());
1226 getBackend().finishLayout(*this, Layout
);
1229 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1230 LLVM_DUMP_METHOD
void MCAssembler::dump() const{
1231 raw_ostream
&OS
= errs();
1233 OS
<< "<MCAssembler\n";
1234 OS
<< " Sections:[\n ";
1235 for (const_iterator it
= begin(), ie
= end(); it
!= ie
; ++it
) {
1236 if (it
!= begin()) OS
<< ",\n ";
1242 for (const_symbol_iterator it
= symbol_begin(), ie
= symbol_end(); it
!= ie
; ++it
) {
1243 if (it
!= symbol_begin()) OS
<< ",\n ";
1246 OS
<< ", Index:" << it
->getIndex() << ", ";