1 //===-- X86MachObjectWriter.cpp - X86 Mach-O Writer -----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "MCTargetDesc/X86MCTargetDesc.h"
11 #include "MCTargetDesc/X86FixupKinds.h"
12 #include "llvm/ADT/Twine.h"
13 #include "llvm/MC/MCAsmInfo.h"
14 #include "llvm/MC/MCAsmLayout.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCContext.h"
17 #include "llvm/MC/MCMachObjectWriter.h"
18 #include "llvm/MC/MCSectionMachO.h"
19 #include "llvm/MC/MCValue.h"
20 #include "llvm/Support/ErrorHandling.h"
21 #include "llvm/Support/Format.h"
22 #include "llvm/Support/MachO.h"
27 class X86MachObjectWriter : public MCMachObjectTargetWriter {
28 bool RecordScatteredRelocation(MachObjectWriter *Writer,
29 const MCAssembler &Asm,
30 const MCAsmLayout &Layout,
31 const MCFragment *Fragment,
35 uint64_t &FixedValue);
36 void RecordTLVPRelocation(MachObjectWriter *Writer,
37 const MCAssembler &Asm,
38 const MCAsmLayout &Layout,
39 const MCFragment *Fragment,
42 uint64_t &FixedValue);
44 void RecordX86Relocation(MachObjectWriter *Writer,
45 const MCAssembler &Asm,
46 const MCAsmLayout &Layout,
47 const MCFragment *Fragment,
50 uint64_t &FixedValue);
51 void RecordX86_64Relocation(MachObjectWriter *Writer, MCAssembler &Asm,
52 const MCAsmLayout &Layout,
53 const MCFragment *Fragment, const MCFixup &Fixup,
54 MCValue Target, uint64_t &FixedValue);
57 X86MachObjectWriter(bool Is64Bit, uint32_t CPUType,
59 : MCMachObjectTargetWriter(Is64Bit, CPUType, CPUSubtype,
60 /*UseAggressiveSymbolFolding=*/Is64Bit) {}
62 void RecordRelocation(MachObjectWriter *Writer, MCAssembler &Asm,
63 const MCAsmLayout &Layout, const MCFragment *Fragment,
64 const MCFixup &Fixup, MCValue Target,
65 uint64_t &FixedValue) override {
66 if (Writer->is64Bit())
67 RecordX86_64Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
70 RecordX86Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
76 static bool isFixupKindRIPRel(unsigned Kind) {
77 return Kind == X86::reloc_riprel_4byte ||
78 Kind == X86::reloc_riprel_4byte_movq_load;
81 static unsigned getFixupKindLog2Size(unsigned Kind) {
84 llvm_unreachable("invalid fixup kind!");
86 case FK_Data_1: return 0;
88 case FK_Data_2: return 1;
90 // FIXME: Remove these!!!
91 case X86::reloc_riprel_4byte:
92 case X86::reloc_riprel_4byte_movq_load:
93 case X86::reloc_signed_4byte:
94 case FK_Data_4: return 2;
95 case FK_Data_8: return 3;
99 void X86MachObjectWriter::RecordX86_64Relocation(
100 MachObjectWriter *Writer, MCAssembler &Asm, const MCAsmLayout &Layout,
101 const MCFragment *Fragment, const MCFixup &Fixup, MCValue Target,
102 uint64_t &FixedValue) {
103 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
104 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
105 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
108 uint32_t FixupOffset =
109 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
110 uint32_t FixupAddress =
111 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
114 unsigned IsExtern = 0;
116 const MCSymbol *RelSymbol = nullptr;
118 Value = Target.getConstant();
121 // Compensate for the relocation offset, Darwin x86_64 relocations only have
122 // the addend and appear to have attempted to define it to be the actual
123 // expression addend without the PCrel bias. However, instructions with data
124 // following the relocation are not accommodated for (see comment below
125 // regarding SIGNED{1,2,4}), so it isn't exactly that either.
126 Value += 1LL << Log2Size;
129 if (Target.isAbsolute()) { // constant
130 // SymbolNum of 0 indicates the absolute section.
131 Type = MachO::X86_64_RELOC_UNSIGNED;
133 // FIXME: I believe this is broken, I don't think the linker can understand
134 // it. I think it would require a local relocation, but I'm not sure if that
135 // would work either. The official way to get an absolute PCrel relocation
136 // is to use an absolute symbol (which we don't support yet).
139 Type = MachO::X86_64_RELOC_BRANCH;
141 } else if (Target.getSymB()) { // A - B + constant
142 const MCSymbol *A = &Target.getSymA()->getSymbol();
143 if (A->isTemporary())
144 A = &Writer->findAliasedSymbol(*A);
145 const MCSymbolData &A_SD = Asm.getSymbolData(*A);
146 const MCSymbol *A_Base = Asm.getAtom(&A_SD);
148 const MCSymbol *B = &Target.getSymB()->getSymbol();
149 if (B->isTemporary())
150 B = &Writer->findAliasedSymbol(*B);
151 const MCSymbolData &B_SD = Asm.getSymbolData(*B);
152 const MCSymbol *B_Base = Asm.getAtom(&B_SD);
154 // Neither symbol can be modified.
155 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
156 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
157 report_fatal_error("unsupported relocation of modified symbol", false);
159 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
160 // implement most of these correctly.
162 report_fatal_error("unsupported pc-relative relocation of difference",
165 // The support for the situation where one or both of the symbols would
166 // require a local relocation is handled just like if the symbols were
167 // external. This is certainly used in the case of debug sections where the
168 // section has only temporary symbols and thus the symbols don't have base
169 // symbols. This is encoded using the section ordinal and non-extern
170 // relocation entries.
172 // Darwin 'as' doesn't emit correct relocations for this (it ends up with a
173 // single SIGNED relocation); reject it for now. Except the case where both
174 // symbols don't have a base, equal but both NULL.
175 if (A_Base == B_Base && A_Base)
176 report_fatal_error("unsupported relocation with identical base", false);
178 // A subtraction expression where either symbol is undefined is a
179 // non-relocatable expression.
180 if (A->isUndefined() || B->isUndefined()) {
181 StringRef Name = A->isUndefined() ? A->getName() : B->getName();
182 Asm.getContext().FatalError(Fixup.getLoc(),
183 "unsupported relocation with subtraction expression, symbol '" +
184 Name + "' can not be undefined in a subtraction expression");
188 Writer->getSymbolAddress(&A_SD, Layout) -
189 (!A_Base ? 0 : Writer->getSymbolAddress(&A_Base->getData(), Layout));
191 Writer->getSymbolAddress(&B_SD, Layout) -
192 (!B_Base ? 0 : Writer->getSymbolAddress(&B_Base->getData(), Layout));
195 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
196 Type = MachO::X86_64_RELOC_UNSIGNED;
198 MachO::any_relocation_info MRE;
199 MRE.r_word0 = FixupOffset;
201 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
202 Writer->addRelocation(A_Base ? &A_Base->getData() : nullptr,
203 Fragment->getParent(), MRE);
208 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
209 Type = MachO::X86_64_RELOC_SUBTRACTOR;
211 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
212 if (Symbol->isTemporary() && Value) {
213 const MCSection &Sec = Symbol->getSection();
214 if (!Asm.getContext().getAsmInfo()->isSectionAtomizableBySymbols(Sec))
215 Asm.addLocalUsedInReloc(*Symbol);
217 const MCSymbolData &SD = Asm.getSymbolData(*Symbol);
218 RelSymbol = Asm.getAtom(&SD);
220 // Relocations inside debug sections always use local relocations when
221 // possible. This seems to be done because the debugger doesn't fully
222 // understand x86_64 relocation entries, and expects to find values that
223 // have already been fixed up.
224 if (Symbol->isInSection()) {
225 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
226 Fragment->getParent()->getSection());
227 if (Section.hasAttribute(MachO::S_ATTR_DEBUG))
231 // x86_64 almost always uses external relocations, except when there is no
232 // symbol to use as a base address (a local symbol with no preceding
233 // non-local symbol).
235 // Add the local offset, if needed.
236 if (&RelSymbol->getData() != &SD)
237 Value += Layout.getSymbolOffset(&SD) -
238 Layout.getSymbolOffset(&RelSymbol->getData());
239 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
240 // The index is the section ordinal (1-based).
241 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
242 Value += Writer->getSymbolAddress(&SD, Layout);
245 Value -= FixupAddress + (1 << Log2Size);
246 } else if (Symbol->isVariable()) {
247 const MCExpr *Value = Symbol->getVariableValue();
249 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout,
250 Writer->getSectionAddressMap());
255 report_fatal_error("unsupported relocation of variable '" +
256 Symbol->getName() + "'", false);
259 report_fatal_error("unsupported relocation of undefined symbol '" +
260 Symbol->getName() + "'", false);
263 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
266 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
267 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
268 // rewrite the movq to an leaq at link time if the symbol ends up in
269 // the same linkage unit.
270 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
271 Type = MachO::X86_64_RELOC_GOT_LOAD;
273 Type = MachO::X86_64_RELOC_GOT;
274 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
275 Type = MachO::X86_64_RELOC_TLV;
276 } else if (Modifier != MCSymbolRefExpr::VK_None) {
277 report_fatal_error("unsupported symbol modifier in relocation",
280 Type = MachO::X86_64_RELOC_SIGNED;
282 // The Darwin x86_64 relocation format has a problem where it cannot
283 // encode an address (L<foo> + <constant>) which is outside the atom
284 // containing L<foo>. Generally, this shouldn't occur but it does
285 // happen when we have a RIPrel instruction with data following the
286 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
287 // adjustment Darwin x86_64 uses, the offset is still negative and the
288 // linker has no way to recognize this.
290 // To work around this, Darwin uses several special relocation types
291 // to indicate the offsets. However, the specification or
292 // implementation of these seems to also be incomplete; they should
293 // adjust the addend as well based on the actual encoded instruction
294 // (the additional bias), but instead appear to just look at the final
296 switch (-(Target.getConstant() + (1LL << Log2Size))) {
297 case 1: Type = MachO::X86_64_RELOC_SIGNED_1; break;
298 case 2: Type = MachO::X86_64_RELOC_SIGNED_2; break;
299 case 4: Type = MachO::X86_64_RELOC_SIGNED_4; break;
303 if (Modifier != MCSymbolRefExpr::VK_None)
304 report_fatal_error("unsupported symbol modifier in branch "
305 "relocation", false);
307 Type = MachO::X86_64_RELOC_BRANCH;
310 if (Modifier == MCSymbolRefExpr::VK_GOT) {
311 Type = MachO::X86_64_RELOC_GOT;
312 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
313 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
314 // case all we do is set the PCrel bit in the relocation entry; this is
315 // used with exception handling, for example. The source is required to
316 // include any necessary offset directly.
317 Type = MachO::X86_64_RELOC_GOT;
319 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
320 report_fatal_error("TLVP symbol modifier should have been rip-rel",
322 } else if (Modifier != MCSymbolRefExpr::VK_None)
323 report_fatal_error("unsupported symbol modifier in relocation", false);
325 Type = MachO::X86_64_RELOC_UNSIGNED;
326 unsigned Kind = Fixup.getKind();
327 if (Kind == X86::reloc_signed_4byte)
328 report_fatal_error("32-bit absolute addressing is not supported in "
329 "64-bit mode", false);
334 // x86_64 always writes custom values into the fixups.
337 // struct relocation_info (8 bytes)
338 MachO::any_relocation_info MRE;
339 MRE.r_word0 = FixupOffset;
340 MRE.r_word1 = (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) |
341 (IsExtern << 27) | (Type << 28);
342 Writer->addRelocation(RelSymbol ? &RelSymbol->getData() : nullptr,
343 Fragment->getParent(), MRE);
346 bool X86MachObjectWriter::RecordScatteredRelocation(MachObjectWriter *Writer,
347 const MCAssembler &Asm,
348 const MCAsmLayout &Layout,
349 const MCFragment *Fragment,
350 const MCFixup &Fixup,
353 uint64_t &FixedValue) {
354 uint64_t OriginalFixedValue = FixedValue;
355 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
356 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
357 unsigned Type = MachO::GENERIC_RELOC_VANILLA;
360 const MCSymbol *A = &Target.getSymA()->getSymbol();
361 const MCSymbolData *A_SD = &Asm.getSymbolData(*A);
363 if (!A_SD->getFragment())
364 report_fatal_error("symbol '" + A->getName() +
365 "' can not be undefined in a subtraction expression",
368 uint32_t Value = Writer->getSymbolAddress(A_SD, Layout);
369 uint64_t SecAddr = Writer->getSectionAddress(A_SD->getFragment()->getParent());
370 FixedValue += SecAddr;
373 if (const MCSymbolRefExpr *B = Target.getSymB()) {
374 const MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
376 if (!B_SD->getFragment())
377 report_fatal_error("symbol '" + B->getSymbol().getName() +
378 "' can not be undefined in a subtraction expression",
381 // Select the appropriate difference relocation type.
383 // Note that there is no longer any semantic difference between these two
384 // relocation types from the linkers point of view, this is done solely for
385 // pedantic compatibility with 'as'.
386 Type = A_SD->isExternal() ? (unsigned)MachO::GENERIC_RELOC_SECTDIFF :
387 (unsigned)MachO::GENERIC_RELOC_LOCAL_SECTDIFF;
388 Value2 = Writer->getSymbolAddress(B_SD, Layout);
389 FixedValue -= Writer->getSectionAddress(B_SD->getFragment()->getParent());
392 // Relocations are written out in reverse order, so the PAIR comes first.
393 if (Type == MachO::GENERIC_RELOC_SECTDIFF ||
394 Type == MachO::GENERIC_RELOC_LOCAL_SECTDIFF) {
395 // If the offset is too large to fit in a scattered relocation,
396 // we're hosed. It's an unfortunate limitation of the MachO format.
397 if (FixupOffset > 0xffffff) {
399 format("0x%x", FixupOffset).print(Buffer, sizeof(Buffer));
400 Asm.getContext().FatalError(Fixup.getLoc(),
401 Twine("Section too large, can't encode "
402 "r_address (") + Buffer +
403 ") into 24 bits of scattered "
404 "relocation entry.");
405 llvm_unreachable("fatal error returned?!");
408 MachO::any_relocation_info MRE;
409 MRE.r_word0 = ((0 << 0) | // r_address
410 (MachO::GENERIC_RELOC_PAIR << 24) | // r_type
414 MRE.r_word1 = Value2;
415 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
417 // If the offset is more than 24-bits, it won't fit in a scattered
418 // relocation offset field, so we fall back to using a non-scattered
419 // relocation. This is a bit risky, as if the offset reaches out of
420 // the block and the linker is doing scattered loading on this
421 // symbol, things can go badly.
423 // Required for 'as' compatibility.
424 if (FixupOffset > 0xffffff) {
425 FixedValue = OriginalFixedValue;
430 MachO::any_relocation_info MRE;
431 MRE.r_word0 = ((FixupOffset << 0) |
437 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
441 void X86MachObjectWriter::RecordTLVPRelocation(MachObjectWriter *Writer,
442 const MCAssembler &Asm,
443 const MCAsmLayout &Layout,
444 const MCFragment *Fragment,
445 const MCFixup &Fixup,
447 uint64_t &FixedValue) {
448 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
450 "Should only be called with a 32-bit TLVP relocation!");
452 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
453 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
454 unsigned IsPCRel = 0;
456 // Get the symbol data.
457 const MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
459 // We're only going to have a second symbol in pic mode and it'll be a
460 // subtraction from the picbase. For 32-bit pic the addend is the difference
461 // between the picbase and the next address. For 32-bit static the addend is
463 if (Target.getSymB()) {
464 // If this is a subtraction then we're pcrel.
465 uint32_t FixupAddress =
466 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
467 const MCSymbolData *SD_B =
468 &Asm.getSymbolData(Target.getSymB()->getSymbol());
470 FixedValue = (FixupAddress - Writer->getSymbolAddress(SD_B, Layout) +
471 Target.getConstant());
472 FixedValue += 1ULL << Log2Size;
477 // struct relocation_info (8 bytes)
478 MachO::any_relocation_info MRE;
481 (IsPCRel << 24) | (Log2Size << 25) | (MachO::GENERIC_RELOC_TLV << 28);
482 Writer->addRelocation(SD_A, Fragment->getParent(), MRE);
485 void X86MachObjectWriter::RecordX86Relocation(MachObjectWriter *Writer,
486 const MCAssembler &Asm,
487 const MCAsmLayout &Layout,
488 const MCFragment *Fragment,
489 const MCFixup &Fixup,
491 uint64_t &FixedValue) {
492 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
493 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
495 // If this is a 32-bit TLVP reloc it's handled a bit differently.
496 if (Target.getSymA() &&
497 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
498 RecordTLVPRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
503 // If this is a difference or a defined symbol plus an offset, then we need a
504 // scattered relocation entry. Differences always require scattered
506 if (Target.getSymB()) {
507 RecordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
508 Target, Log2Size, FixedValue);
512 // Get the symbol data, if any.
513 const MCSymbolData *SD = nullptr;
514 if (Target.getSymA())
515 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
517 // If this is an internal relocation with an offset, it also needs a scattered
519 uint32_t Offset = Target.getConstant();
521 Offset += 1 << Log2Size;
522 // Try to record the scattered relocation if needed. Fall back to non
523 // scattered if necessary (see comments in RecordScatteredRelocation()
525 if (Offset && SD && !Writer->doesSymbolRequireExternRelocation(SD) &&
526 RecordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
527 Target, Log2Size, FixedValue))
531 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
534 const MCSymbolData *RelSymbol = nullptr;
536 if (Target.isAbsolute()) { // constant
537 // SymbolNum of 0 indicates the absolute section.
539 // FIXME: Currently, these are never generated (see code below). I cannot
540 // find a case where they are actually emitted.
541 Type = MachO::GENERIC_RELOC_VANILLA;
543 // Resolve constant variables.
544 if (SD->getSymbol().isVariable()) {
546 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
547 Res, Layout, Writer->getSectionAddressMap())) {
553 // Check whether we need an external or internal relocation.
554 if (Writer->doesSymbolRequireExternRelocation(SD)) {
556 // For external relocations, make sure to offset the fixup value to
557 // compensate for the addend of the symbol address, if it was
558 // undefined. This occurs with weak definitions, for example.
559 if (!SD->getSymbol().isUndefined())
560 FixedValue -= Layout.getSymbolOffset(SD);
562 // The index is the section ordinal (1-based).
563 const MCSectionData &SymSD = Asm.getSectionData(
564 SD->getSymbol().getSection());
565 Index = SymSD.getOrdinal() + 1;
566 FixedValue += Writer->getSectionAddress(&SymSD);
569 FixedValue -= Writer->getSectionAddress(Fragment->getParent());
571 Type = MachO::GENERIC_RELOC_VANILLA;
574 // struct relocation_info (8 bytes)
575 MachO::any_relocation_info MRE;
576 MRE.r_word0 = FixupOffset;
578 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
579 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
582 MCObjectWriter *llvm::createX86MachObjectWriter(raw_pwrite_stream &OS,
583 bool Is64Bit, uint32_t CPUType,
584 uint32_t CPUSubtype) {
585 return createMachObjectWriter(new X86MachObjectWriter(Is64Bit,
588 OS, /*IsLittleEndian=*/true);