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().reportError(Fixup.getLoc(),
179 "unsupported relocation with subtraction expression, symbol '" +
180 Name + "' can not be undefined in a subtraction expression");
184 Value += Writer->getSymbolAddress(*A, Layout) -
185 (!A_Base ? 0 : Writer->getSymbolAddress(*A_Base, Layout));
186 Value -= Writer->getSymbolAddress(*B, Layout) -
187 (!B_Base ? 0 : Writer->getSymbolAddress(*B_Base, Layout));
190 Index = A->getFragment()->getParent()->getOrdinal() + 1;
191 Type = MachO::X86_64_RELOC_UNSIGNED;
193 MachO::any_relocation_info MRE;
194 MRE.r_word0 = FixupOffset;
196 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
197 Writer->addRelocation(A_Base, Fragment->getParent(), MRE);
202 Index = B->getFragment()->getParent()->getOrdinal() + 1;
203 Type = MachO::X86_64_RELOC_SUBTRACTOR;
205 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
206 if (Symbol->isTemporary() && Value) {
207 const MCSection &Sec = Symbol->getSection();
208 if (!Asm.getContext().getAsmInfo()->isSectionAtomizableBySymbols(Sec))
209 Symbol->setUsedInReloc();
211 RelSymbol = Asm.getAtom(*Symbol);
213 // Relocations inside debug sections always use local relocations when
214 // possible. This seems to be done because the debugger doesn't fully
215 // understand x86_64 relocation entries, and expects to find values that
216 // have already been fixed up.
217 if (Symbol->isInSection()) {
218 const MCSectionMachO &Section =
219 static_cast<const MCSectionMachO &>(*Fragment->getParent());
220 if (Section.hasAttribute(MachO::S_ATTR_DEBUG))
224 // x86_64 almost always uses external relocations, except when there is no
225 // symbol to use as a base address (a local symbol with no preceding
226 // non-local symbol).
228 // Add the local offset, if needed.
229 if (RelSymbol != Symbol)
230 Value += Layout.getSymbolOffset(*Symbol) -
231 Layout.getSymbolOffset(*RelSymbol);
232 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
233 // The index is the section ordinal (1-based).
234 Index = Symbol->getFragment()->getParent()->getOrdinal() + 1;
235 Value += Writer->getSymbolAddress(*Symbol, Layout);
238 Value -= FixupAddress + (1 << Log2Size);
239 } else if (Symbol->isVariable()) {
240 const MCExpr *Value = Symbol->getVariableValue();
242 bool isAbs = Value->evaluateAsAbsolute(Res, Layout,
243 Writer->getSectionAddressMap());
248 report_fatal_error("unsupported relocation of variable '" +
249 Symbol->getName() + "'", false);
252 report_fatal_error("unsupported relocation of undefined symbol '" +
253 Symbol->getName() + "'", false);
256 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
259 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
260 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
261 // rewrite the movq to an leaq at link time if the symbol ends up in
262 // the same linkage unit.
263 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
264 Type = MachO::X86_64_RELOC_GOT_LOAD;
266 Type = MachO::X86_64_RELOC_GOT;
267 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
268 Type = MachO::X86_64_RELOC_TLV;
269 } else if (Modifier != MCSymbolRefExpr::VK_None) {
270 report_fatal_error("unsupported symbol modifier in relocation",
273 Type = MachO::X86_64_RELOC_SIGNED;
275 // The Darwin x86_64 relocation format has a problem where it cannot
276 // encode an address (L<foo> + <constant>) which is outside the atom
277 // containing L<foo>. Generally, this shouldn't occur but it does
278 // happen when we have a RIPrel instruction with data following the
279 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
280 // adjustment Darwin x86_64 uses, the offset is still negative and the
281 // linker has no way to recognize this.
283 // To work around this, Darwin uses several special relocation types
284 // to indicate the offsets. However, the specification or
285 // implementation of these seems to also be incomplete; they should
286 // adjust the addend as well based on the actual encoded instruction
287 // (the additional bias), but instead appear to just look at the final
289 switch (-(Target.getConstant() + (1LL << Log2Size))) {
290 case 1: Type = MachO::X86_64_RELOC_SIGNED_1; break;
291 case 2: Type = MachO::X86_64_RELOC_SIGNED_2; break;
292 case 4: Type = MachO::X86_64_RELOC_SIGNED_4; break;
296 if (Modifier != MCSymbolRefExpr::VK_None)
297 report_fatal_error("unsupported symbol modifier in branch "
298 "relocation", false);
300 Type = MachO::X86_64_RELOC_BRANCH;
303 if (Modifier == MCSymbolRefExpr::VK_GOT) {
304 Type = MachO::X86_64_RELOC_GOT;
305 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
306 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
307 // case all we do is set the PCrel bit in the relocation entry; this is
308 // used with exception handling, for example. The source is required to
309 // include any necessary offset directly.
310 Type = MachO::X86_64_RELOC_GOT;
312 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
313 report_fatal_error("TLVP symbol modifier should have been rip-rel",
315 } else if (Modifier != MCSymbolRefExpr::VK_None)
316 report_fatal_error("unsupported symbol modifier in relocation", false);
318 Type = MachO::X86_64_RELOC_UNSIGNED;
319 unsigned Kind = Fixup.getKind();
320 if (Kind == X86::reloc_signed_4byte)
321 report_fatal_error("32-bit absolute addressing is not supported in "
322 "64-bit mode", false);
327 // x86_64 always writes custom values into the fixups.
330 // struct relocation_info (8 bytes)
331 MachO::any_relocation_info MRE;
332 MRE.r_word0 = FixupOffset;
333 MRE.r_word1 = (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) |
334 (IsExtern << 27) | (Type << 28);
335 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
338 bool X86MachObjectWriter::recordScatteredRelocation(MachObjectWriter *Writer,
339 const MCAssembler &Asm,
340 const MCAsmLayout &Layout,
341 const MCFragment *Fragment,
342 const MCFixup &Fixup,
345 uint64_t &FixedValue) {
346 uint64_t OriginalFixedValue = FixedValue;
347 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
348 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
349 unsigned Type = MachO::GENERIC_RELOC_VANILLA;
352 const MCSymbol *A = &Target.getSymA()->getSymbol();
354 if (!A->getFragment())
355 report_fatal_error("symbol '" + A->getName() +
356 "' can not be undefined in a subtraction expression",
359 uint32_t Value = Writer->getSymbolAddress(*A, Layout);
360 uint64_t SecAddr = Writer->getSectionAddress(A->getFragment()->getParent());
361 FixedValue += SecAddr;
364 if (const MCSymbolRefExpr *B = Target.getSymB()) {
365 const MCSymbol *SB = &B->getSymbol();
367 if (!SB->getFragment())
368 report_fatal_error("symbol '" + B->getSymbol().getName() +
369 "' can not be undefined in a subtraction expression",
372 // Select the appropriate difference relocation type.
374 // Note that there is no longer any semantic difference between these two
375 // relocation types from the linkers point of view, this is done solely for
376 // pedantic compatibility with 'as'.
377 Type = A->isExternal() ? (unsigned)MachO::GENERIC_RELOC_SECTDIFF
378 : (unsigned)MachO::GENERIC_RELOC_LOCAL_SECTDIFF;
379 Value2 = Writer->getSymbolAddress(B->getSymbol(), Layout);
380 FixedValue -= Writer->getSectionAddress(SB->getFragment()->getParent());
383 // Relocations are written out in reverse order, so the PAIR comes first.
384 if (Type == MachO::GENERIC_RELOC_SECTDIFF ||
385 Type == MachO::GENERIC_RELOC_LOCAL_SECTDIFF) {
386 // If the offset is too large to fit in a scattered relocation,
387 // we're hosed. It's an unfortunate limitation of the MachO format.
388 if (FixupOffset > 0xffffff) {
390 format("0x%x", FixupOffset).print(Buffer, sizeof(Buffer));
391 Asm.getContext().reportError(Fixup.getLoc(),
392 Twine("Section too large, can't encode "
393 "r_address (") + Buffer +
394 ") into 24 bits of scattered "
395 "relocation entry.");
399 MachO::any_relocation_info MRE;
400 MRE.r_word0 = ((0 << 0) | // r_address
401 (MachO::GENERIC_RELOC_PAIR << 24) | // r_type
405 MRE.r_word1 = Value2;
406 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
408 // If the offset is more than 24-bits, it won't fit in a scattered
409 // relocation offset field, so we fall back to using a non-scattered
410 // relocation. This is a bit risky, as if the offset reaches out of
411 // the block and the linker is doing scattered loading on this
412 // symbol, things can go badly.
414 // Required for 'as' compatibility.
415 if (FixupOffset > 0xffffff) {
416 FixedValue = OriginalFixedValue;
421 MachO::any_relocation_info MRE;
422 MRE.r_word0 = ((FixupOffset << 0) |
428 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
432 void X86MachObjectWriter::recordTLVPRelocation(MachObjectWriter *Writer,
433 const MCAssembler &Asm,
434 const MCAsmLayout &Layout,
435 const MCFragment *Fragment,
436 const MCFixup &Fixup,
438 uint64_t &FixedValue) {
439 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
441 "Should only be called with a 32-bit TLVP relocation!");
443 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
444 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
445 unsigned IsPCRel = 0;
447 // We're only going to have a second symbol in pic mode and it'll be a
448 // subtraction from the picbase. For 32-bit pic the addend is the difference
449 // between the picbase and the next address. For 32-bit static the addend is
451 if (Target.getSymB()) {
452 // If this is a subtraction then we're pcrel.
453 uint32_t FixupAddress =
454 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
458 Writer->getSymbolAddress(Target.getSymB()->getSymbol(), Layout) +
459 Target.getConstant();
460 FixedValue += 1ULL << Log2Size;
465 // struct relocation_info (8 bytes)
466 MachO::any_relocation_info MRE;
469 (IsPCRel << 24) | (Log2Size << 25) | (MachO::GENERIC_RELOC_TLV << 28);
470 Writer->addRelocation(&Target.getSymA()->getSymbol(), Fragment->getParent(),
474 void X86MachObjectWriter::RecordX86Relocation(MachObjectWriter *Writer,
475 const MCAssembler &Asm,
476 const MCAsmLayout &Layout,
477 const MCFragment *Fragment,
478 const MCFixup &Fixup,
480 uint64_t &FixedValue) {
481 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
482 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
484 // If this is a 32-bit TLVP reloc it's handled a bit differently.
485 if (Target.getSymA() &&
486 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
487 recordTLVPRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
492 // If this is a difference or a defined symbol plus an offset, then we need a
493 // scattered relocation entry. Differences always require scattered
495 if (Target.getSymB()) {
496 recordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
497 Target, Log2Size, FixedValue);
501 // Get the symbol data, if any.
502 const MCSymbol *A = nullptr;
503 if (Target.getSymA())
504 A = &Target.getSymA()->getSymbol();
506 // If this is an internal relocation with an offset, it also needs a scattered
508 uint32_t Offset = Target.getConstant();
510 Offset += 1 << Log2Size;
511 // Try to record the scattered relocation if needed. Fall back to non
512 // scattered if necessary (see comments in recordScatteredRelocation()
514 if (Offset && A && !Writer->doesSymbolRequireExternRelocation(*A) &&
515 recordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
516 Log2Size, FixedValue))
520 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
523 const MCSymbol *RelSymbol = nullptr;
525 if (Target.isAbsolute()) { // constant
526 // SymbolNum of 0 indicates the absolute section.
528 // FIXME: Currently, these are never generated (see code below). I cannot
529 // find a case where they are actually emitted.
530 Type = MachO::GENERIC_RELOC_VANILLA;
532 // Resolve constant variables.
533 if (A->isVariable()) {
535 if (A->getVariableValue()->evaluateAsAbsolute(
536 Res, Layout, Writer->getSectionAddressMap())) {
542 // Check whether we need an external or internal relocation.
543 if (Writer->doesSymbolRequireExternRelocation(*A)) {
545 // For external relocations, make sure to offset the fixup value to
546 // compensate for the addend of the symbol address, if it was
547 // undefined. This occurs with weak definitions, for example.
548 if (!A->isUndefined())
549 FixedValue -= Layout.getSymbolOffset(*A);
551 // The index is the section ordinal (1-based).
552 const MCSection &Sec = A->getSection();
553 Index = Sec.getOrdinal() + 1;
554 FixedValue += Writer->getSectionAddress(&Sec);
557 FixedValue -= Writer->getSectionAddress(Fragment->getParent());
559 Type = MachO::GENERIC_RELOC_VANILLA;
562 // struct relocation_info (8 bytes)
563 MachO::any_relocation_info MRE;
564 MRE.r_word0 = FixupOffset;
566 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
567 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
570 MCObjectWriter *llvm::createX86MachObjectWriter(raw_pwrite_stream &OS,
571 bool Is64Bit, uint32_t CPUType,
572 uint32_t CPUSubtype) {
573 return createMachObjectWriter(new X86MachObjectWriter(Is64Bit,
576 OS, /*IsLittleEndian=*/true);