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, uint32_t CPUSubtype)
58 : MCMachObjectTargetWriter(Is64Bit, CPUType, CPUSubtype) {}
60 void recordRelocation(MachObjectWriter *Writer, MCAssembler &Asm,
61 const MCAsmLayout &Layout, const MCFragment *Fragment,
62 const MCFixup &Fixup, MCValue Target,
63 uint64_t &FixedValue) override {
64 if (Writer->is64Bit())
65 RecordX86_64Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
68 RecordX86Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
74 static bool isFixupKindRIPRel(unsigned Kind) {
75 return Kind == X86::reloc_riprel_4byte ||
76 Kind == X86::reloc_riprel_4byte_movq_load;
79 static unsigned getFixupKindLog2Size(unsigned Kind) {
82 llvm_unreachable("invalid fixup kind!");
84 case FK_Data_1: return 0;
86 case FK_Data_2: return 1;
88 // FIXME: Remove these!!!
89 case X86::reloc_riprel_4byte:
90 case X86::reloc_riprel_4byte_movq_load:
91 case X86::reloc_signed_4byte:
92 case FK_Data_4: return 2;
93 case FK_Data_8: return 3;
97 void X86MachObjectWriter::RecordX86_64Relocation(
98 MachObjectWriter *Writer, MCAssembler &Asm, const MCAsmLayout &Layout,
99 const MCFragment *Fragment, const MCFixup &Fixup, MCValue Target,
100 uint64_t &FixedValue) {
101 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
102 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
103 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
106 uint32_t FixupOffset =
107 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
108 uint32_t FixupAddress =
109 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
112 unsigned IsExtern = 0;
114 const MCSymbol *RelSymbol = nullptr;
116 Value = Target.getConstant();
119 // Compensate for the relocation offset, Darwin x86_64 relocations only have
120 // the addend and appear to have attempted to define it to be the actual
121 // expression addend without the PCrel bias. However, instructions with data
122 // following the relocation are not accommodated for (see comment below
123 // regarding SIGNED{1,2,4}), so it isn't exactly that either.
124 Value += 1LL << Log2Size;
127 if (Target.isAbsolute()) { // constant
128 // SymbolNum of 0 indicates the absolute section.
129 Type = MachO::X86_64_RELOC_UNSIGNED;
131 // FIXME: I believe this is broken, I don't think the linker can understand
132 // it. I think it would require a local relocation, but I'm not sure if that
133 // would work either. The official way to get an absolute PCrel relocation
134 // is to use an absolute symbol (which we don't support yet).
137 Type = MachO::X86_64_RELOC_BRANCH;
139 } else if (Target.getSymB()) { // A - B + constant
140 const MCSymbol *A = &Target.getSymA()->getSymbol();
141 if (A->isTemporary())
142 A = &Writer->findAliasedSymbol(*A);
143 const MCSymbol *A_Base = Asm.getAtom(*A);
145 const MCSymbol *B = &Target.getSymB()->getSymbol();
146 if (B->isTemporary())
147 B = &Writer->findAliasedSymbol(*B);
148 const MCSymbol *B_Base = Asm.getAtom(*B);
150 // Neither symbol can be modified.
151 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
152 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
153 report_fatal_error("unsupported relocation of modified symbol", false);
155 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
156 // implement most of these correctly.
158 report_fatal_error("unsupported pc-relative relocation of difference",
161 // The support for the situation where one or both of the symbols would
162 // require a local relocation is handled just like if the symbols were
163 // external. This is certainly used in the case of debug sections where the
164 // section has only temporary symbols and thus the symbols don't have base
165 // symbols. This is encoded using the section ordinal and non-extern
166 // relocation entries.
168 // Darwin 'as' doesn't emit correct relocations for this (it ends up with a
169 // single SIGNED relocation); reject it for now. Except the case where both
170 // symbols don't have a base, equal but both NULL.
171 if (A_Base == B_Base && A_Base)
172 report_fatal_error("unsupported relocation with identical base", false);
174 // A subtraction expression where either symbol is undefined is a
175 // non-relocatable expression.
176 if (A->isUndefined() || B->isUndefined()) {
177 StringRef Name = A->isUndefined() ? A->getName() : B->getName();
178 Asm.getContext().reportFatalError(Fixup.getLoc(),
179 "unsupported relocation with subtraction expression, symbol '" +
180 Name + "' can not be undefined in a subtraction expression");
183 Value += Writer->getSymbolAddress(*A, Layout) -
184 (!A_Base ? 0 : Writer->getSymbolAddress(*A_Base, Layout));
185 Value -= Writer->getSymbolAddress(*B, Layout) -
186 (!B_Base ? 0 : Writer->getSymbolAddress(*B_Base, Layout));
189 Index = A->getFragment()->getParent()->getOrdinal() + 1;
190 Type = MachO::X86_64_RELOC_UNSIGNED;
192 MachO::any_relocation_info MRE;
193 MRE.r_word0 = FixupOffset;
195 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
196 Writer->addRelocation(A_Base, Fragment->getParent(), MRE);
201 Index = B->getFragment()->getParent()->getOrdinal() + 1;
202 Type = MachO::X86_64_RELOC_SUBTRACTOR;
204 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
205 if (Symbol->isTemporary() && Value) {
206 const MCSection &Sec = Symbol->getSection();
207 if (!Asm.getContext().getAsmInfo()->isSectionAtomizableBySymbols(Sec))
208 Symbol->setUsedInReloc();
210 RelSymbol = Asm.getAtom(*Symbol);
212 // Relocations inside debug sections always use local relocations when
213 // possible. This seems to be done because the debugger doesn't fully
214 // understand x86_64 relocation entries, and expects to find values that
215 // have already been fixed up.
216 if (Symbol->isInSection()) {
217 const MCSectionMachO &Section =
218 static_cast<const MCSectionMachO &>(*Fragment->getParent());
219 if (Section.hasAttribute(MachO::S_ATTR_DEBUG))
223 // x86_64 almost always uses external relocations, except when there is no
224 // symbol to use as a base address (a local symbol with no preceding
225 // non-local symbol).
227 // Add the local offset, if needed.
228 if (RelSymbol != Symbol)
229 Value += Layout.getSymbolOffset(*Symbol) -
230 Layout.getSymbolOffset(*RelSymbol);
231 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
232 // The index is the section ordinal (1-based).
233 Index = Symbol->getFragment()->getParent()->getOrdinal() + 1;
234 Value += Writer->getSymbolAddress(*Symbol, Layout);
237 Value -= FixupAddress + (1 << Log2Size);
238 } else if (Symbol->isVariable()) {
239 const MCExpr *Value = Symbol->getVariableValue();
241 bool isAbs = Value->evaluateAsAbsolute(Res, Layout,
242 Writer->getSectionAddressMap());
247 report_fatal_error("unsupported relocation of variable '" +
248 Symbol->getName() + "'", false);
251 report_fatal_error("unsupported relocation of undefined symbol '" +
252 Symbol->getName() + "'", false);
255 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
258 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
259 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
260 // rewrite the movq to an leaq at link time if the symbol ends up in
261 // the same linkage unit.
262 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
263 Type = MachO::X86_64_RELOC_GOT_LOAD;
265 Type = MachO::X86_64_RELOC_GOT;
266 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
267 Type = MachO::X86_64_RELOC_TLV;
268 } else if (Modifier != MCSymbolRefExpr::VK_None) {
269 report_fatal_error("unsupported symbol modifier in relocation",
272 Type = MachO::X86_64_RELOC_SIGNED;
274 // The Darwin x86_64 relocation format has a problem where it cannot
275 // encode an address (L<foo> + <constant>) which is outside the atom
276 // containing L<foo>. Generally, this shouldn't occur but it does
277 // happen when we have a RIPrel instruction with data following the
278 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
279 // adjustment Darwin x86_64 uses, the offset is still negative and the
280 // linker has no way to recognize this.
282 // To work around this, Darwin uses several special relocation types
283 // to indicate the offsets. However, the specification or
284 // implementation of these seems to also be incomplete; they should
285 // adjust the addend as well based on the actual encoded instruction
286 // (the additional bias), but instead appear to just look at the final
288 switch (-(Target.getConstant() + (1LL << Log2Size))) {
289 case 1: Type = MachO::X86_64_RELOC_SIGNED_1; break;
290 case 2: Type = MachO::X86_64_RELOC_SIGNED_2; break;
291 case 4: Type = MachO::X86_64_RELOC_SIGNED_4; break;
295 if (Modifier != MCSymbolRefExpr::VK_None)
296 report_fatal_error("unsupported symbol modifier in branch "
297 "relocation", false);
299 Type = MachO::X86_64_RELOC_BRANCH;
302 if (Modifier == MCSymbolRefExpr::VK_GOT) {
303 Type = MachO::X86_64_RELOC_GOT;
304 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
305 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
306 // case all we do is set the PCrel bit in the relocation entry; this is
307 // used with exception handling, for example. The source is required to
308 // include any necessary offset directly.
309 Type = MachO::X86_64_RELOC_GOT;
311 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
312 report_fatal_error("TLVP symbol modifier should have been rip-rel",
314 } else if (Modifier != MCSymbolRefExpr::VK_None)
315 report_fatal_error("unsupported symbol modifier in relocation", false);
317 Type = MachO::X86_64_RELOC_UNSIGNED;
318 unsigned Kind = Fixup.getKind();
319 if (Kind == X86::reloc_signed_4byte)
320 report_fatal_error("32-bit absolute addressing is not supported in "
321 "64-bit mode", false);
326 // x86_64 always writes custom values into the fixups.
329 // struct relocation_info (8 bytes)
330 MachO::any_relocation_info MRE;
331 MRE.r_word0 = FixupOffset;
332 MRE.r_word1 = (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) |
333 (IsExtern << 27) | (Type << 28);
334 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
337 bool X86MachObjectWriter::recordScatteredRelocation(MachObjectWriter *Writer,
338 const MCAssembler &Asm,
339 const MCAsmLayout &Layout,
340 const MCFragment *Fragment,
341 const MCFixup &Fixup,
344 uint64_t &FixedValue) {
345 uint64_t OriginalFixedValue = FixedValue;
346 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
347 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
348 unsigned Type = MachO::GENERIC_RELOC_VANILLA;
351 const MCSymbol *A = &Target.getSymA()->getSymbol();
353 if (!A->getFragment())
354 report_fatal_error("symbol '" + A->getName() +
355 "' can not be undefined in a subtraction expression",
358 uint32_t Value = Writer->getSymbolAddress(*A, Layout);
359 uint64_t SecAddr = Writer->getSectionAddress(A->getFragment()->getParent());
360 FixedValue += SecAddr;
363 if (const MCSymbolRefExpr *B = Target.getSymB()) {
364 const MCSymbol *SB = &B->getSymbol();
366 if (!SB->getFragment())
367 report_fatal_error("symbol '" + B->getSymbol().getName() +
368 "' can not be undefined in a subtraction expression",
371 // Select the appropriate difference relocation type.
373 // Note that there is no longer any semantic difference between these two
374 // relocation types from the linkers point of view, this is done solely for
375 // pedantic compatibility with 'as'.
376 Type = A->isExternal() ? (unsigned)MachO::GENERIC_RELOC_SECTDIFF
377 : (unsigned)MachO::GENERIC_RELOC_LOCAL_SECTDIFF;
378 Value2 = Writer->getSymbolAddress(B->getSymbol(), Layout);
379 FixedValue -= Writer->getSectionAddress(SB->getFragment()->getParent());
382 // Relocations are written out in reverse order, so the PAIR comes first.
383 if (Type == MachO::GENERIC_RELOC_SECTDIFF ||
384 Type == MachO::GENERIC_RELOC_LOCAL_SECTDIFF) {
385 // If the offset is too large to fit in a scattered relocation,
386 // we're hosed. It's an unfortunate limitation of the MachO format.
387 if (FixupOffset > 0xffffff) {
389 format("0x%x", FixupOffset).print(Buffer, sizeof(Buffer));
390 Asm.getContext().reportFatalError(Fixup.getLoc(),
391 Twine("Section too large, can't encode "
392 "r_address (") + Buffer +
393 ") into 24 bits of scattered "
394 "relocation entry.");
395 llvm_unreachable("fatal error returned?!");
398 MachO::any_relocation_info MRE;
399 MRE.r_word0 = ((0 << 0) | // r_address
400 (MachO::GENERIC_RELOC_PAIR << 24) | // r_type
404 MRE.r_word1 = Value2;
405 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
407 // If the offset is more than 24-bits, it won't fit in a scattered
408 // relocation offset field, so we fall back to using a non-scattered
409 // relocation. This is a bit risky, as if the offset reaches out of
410 // the block and the linker is doing scattered loading on this
411 // symbol, things can go badly.
413 // Required for 'as' compatibility.
414 if (FixupOffset > 0xffffff) {
415 FixedValue = OriginalFixedValue;
420 MachO::any_relocation_info MRE;
421 MRE.r_word0 = ((FixupOffset << 0) |
427 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
431 void X86MachObjectWriter::recordTLVPRelocation(MachObjectWriter *Writer,
432 const MCAssembler &Asm,
433 const MCAsmLayout &Layout,
434 const MCFragment *Fragment,
435 const MCFixup &Fixup,
437 uint64_t &FixedValue) {
438 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
440 "Should only be called with a 32-bit TLVP relocation!");
442 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
443 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
444 unsigned IsPCRel = 0;
446 // We're only going to have a second symbol in pic mode and it'll be a
447 // subtraction from the picbase. For 32-bit pic the addend is the difference
448 // between the picbase and the next address. For 32-bit static the addend is
450 if (Target.getSymB()) {
451 // If this is a subtraction then we're pcrel.
452 uint32_t FixupAddress =
453 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
457 Writer->getSymbolAddress(Target.getSymB()->getSymbol(), Layout) +
458 Target.getConstant();
459 FixedValue += 1ULL << Log2Size;
464 // struct relocation_info (8 bytes)
465 MachO::any_relocation_info MRE;
468 (IsPCRel << 24) | (Log2Size << 25) | (MachO::GENERIC_RELOC_TLV << 28);
469 Writer->addRelocation(&Target.getSymA()->getSymbol(), Fragment->getParent(),
473 void X86MachObjectWriter::RecordX86Relocation(MachObjectWriter *Writer,
474 const MCAssembler &Asm,
475 const MCAsmLayout &Layout,
476 const MCFragment *Fragment,
477 const MCFixup &Fixup,
479 uint64_t &FixedValue) {
480 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
481 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
483 // If this is a 32-bit TLVP reloc it's handled a bit differently.
484 if (Target.getSymA() &&
485 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
486 recordTLVPRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
491 // If this is a difference or a defined symbol plus an offset, then we need a
492 // scattered relocation entry. Differences always require scattered
494 if (Target.getSymB()) {
495 recordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
496 Target, Log2Size, FixedValue);
500 // Get the symbol data, if any.
501 const MCSymbol *A = nullptr;
502 if (Target.getSymA())
503 A = &Target.getSymA()->getSymbol();
505 // If this is an internal relocation with an offset, it also needs a scattered
507 uint32_t Offset = Target.getConstant();
509 Offset += 1 << Log2Size;
510 // Try to record the scattered relocation if needed. Fall back to non
511 // scattered if necessary (see comments in recordScatteredRelocation()
513 if (Offset && A && !Writer->doesSymbolRequireExternRelocation(*A) &&
514 recordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
515 Log2Size, FixedValue))
519 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
522 const MCSymbol *RelSymbol = nullptr;
524 if (Target.isAbsolute()) { // constant
525 // SymbolNum of 0 indicates the absolute section.
527 // FIXME: Currently, these are never generated (see code below). I cannot
528 // find a case where they are actually emitted.
529 Type = MachO::GENERIC_RELOC_VANILLA;
531 // Resolve constant variables.
532 if (A->isVariable()) {
534 if (A->getVariableValue()->evaluateAsAbsolute(
535 Res, Layout, Writer->getSectionAddressMap())) {
541 // Check whether we need an external or internal relocation.
542 if (Writer->doesSymbolRequireExternRelocation(*A)) {
544 // For external relocations, make sure to offset the fixup value to
545 // compensate for the addend of the symbol address, if it was
546 // undefined. This occurs with weak definitions, for example.
547 if (!A->isUndefined())
548 FixedValue -= Layout.getSymbolOffset(*A);
550 // The index is the section ordinal (1-based).
551 const MCSection &Sec = A->getSection();
552 Index = Sec.getOrdinal() + 1;
553 FixedValue += Writer->getSectionAddress(&Sec);
556 FixedValue -= Writer->getSectionAddress(Fragment->getParent());
558 Type = MachO::GENERIC_RELOC_VANILLA;
561 // struct relocation_info (8 bytes)
562 MachO::any_relocation_info MRE;
563 MRE.r_word0 = FixupOffset;
565 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
566 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
569 MCObjectWriter *llvm::createX86MachObjectWriter(raw_pwrite_stream &OS,
570 bool Is64Bit, uint32_t CPUType,
571 uint32_t CPUSubtype) {
572 return createMachObjectWriter(new X86MachObjectWriter(Is64Bit,
575 OS, /*IsLittleEndian=*/true);