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/MCAsmLayout.h"
14 #include "llvm/MC/MCAssembler.h"
15 #include "llvm/MC/MCContext.h"
16 #include "llvm/MC/MCMachObjectWriter.h"
17 #include "llvm/MC/MCSectionMachO.h"
18 #include "llvm/MC/MCValue.h"
19 #include "llvm/Support/ErrorHandling.h"
20 #include "llvm/Support/Format.h"
21 #include "llvm/Support/MachO.h"
26 class X86MachObjectWriter : public MCMachObjectTargetWriter {
27 bool RecordScatteredRelocation(MachObjectWriter *Writer,
28 const MCAssembler &Asm,
29 const MCAsmLayout &Layout,
30 const MCFragment *Fragment,
34 uint64_t &FixedValue);
35 void RecordTLVPRelocation(MachObjectWriter *Writer,
36 const MCAssembler &Asm,
37 const MCAsmLayout &Layout,
38 const MCFragment *Fragment,
41 uint64_t &FixedValue);
43 void RecordX86Relocation(MachObjectWriter *Writer,
44 const MCAssembler &Asm,
45 const MCAsmLayout &Layout,
46 const MCFragment *Fragment,
49 uint64_t &FixedValue);
50 void RecordX86_64Relocation(MachObjectWriter *Writer,
51 const MCAssembler &Asm,
52 const MCAsmLayout &Layout,
53 const MCFragment *Fragment,
56 uint64_t &FixedValue);
58 X86MachObjectWriter(bool Is64Bit, uint32_t CPUType,
60 : MCMachObjectTargetWriter(Is64Bit, CPUType, CPUSubtype,
61 /*UseAggressiveSymbolFolding=*/Is64Bit) {}
63 void RecordRelocation(MachObjectWriter *Writer,
64 const MCAssembler &Asm, const MCAsmLayout &Layout,
65 const MCFragment *Fragment, const MCFixup &Fixup,
66 MCValue Target, uint64_t &FixedValue) {
67 if (Writer->is64Bit())
68 RecordX86_64Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
71 RecordX86Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
77 static bool isFixupKindRIPRel(unsigned Kind) {
78 return Kind == X86::reloc_riprel_4byte ||
79 Kind == X86::reloc_riprel_4byte_movq_load;
82 static unsigned getFixupKindLog2Size(unsigned Kind) {
85 llvm_unreachable("invalid fixup kind!");
87 case FK_Data_1: return 0;
89 case FK_Data_2: return 1;
91 // FIXME: Remove these!!!
92 case X86::reloc_riprel_4byte:
93 case X86::reloc_riprel_4byte_movq_load:
94 case X86::reloc_signed_4byte:
95 case FK_Data_4: return 2;
96 case FK_Data_8: return 3;
100 void X86MachObjectWriter::RecordX86_64Relocation(MachObjectWriter *Writer,
101 const MCAssembler &Asm,
102 const MCAsmLayout &Layout,
103 const MCFragment *Fragment,
104 const MCFixup &Fixup,
106 uint64_t &FixedValue) {
107 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
108 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
109 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
112 uint32_t FixupOffset =
113 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
114 uint32_t FixupAddress =
115 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
118 unsigned IsExtern = 0;
121 Value = Target.getConstant();
124 // Compensate for the relocation offset, Darwin x86_64 relocations only have
125 // the addend and appear to have attempted to define it to be the actual
126 // expression addend without the PCrel bias. However, instructions with data
127 // following the relocation are not accommodated for (see comment below
128 // regarding SIGNED{1,2,4}), so it isn't exactly that either.
129 Value += 1LL << Log2Size;
132 if (Target.isAbsolute()) { // constant
133 // SymbolNum of 0 indicates the absolute section.
134 Type = MachO::X86_64_RELOC_UNSIGNED;
137 // FIXME: I believe this is broken, I don't think the linker can understand
138 // it. I think it would require a local relocation, but I'm not sure if that
139 // would work either. The official way to get an absolute PCrel relocation
140 // is to use an absolute symbol (which we don't support yet).
143 Type = MachO::X86_64_RELOC_BRANCH;
145 } else if (Target.getSymB()) { // A - B + constant
146 const MCSymbol *A = &Target.getSymA()->getSymbol();
147 if (A->isTemporary())
148 A = &A->AliasedSymbol();
149 MCSymbolData &A_SD = Asm.getSymbolData(*A);
150 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
152 const MCSymbol *B = &Target.getSymB()->getSymbol();
153 if (B->isTemporary())
154 B = &B->AliasedSymbol();
155 MCSymbolData &B_SD = Asm.getSymbolData(*B);
156 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
158 // Neither symbol can be modified.
159 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
160 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
161 report_fatal_error("unsupported relocation of modified symbol", false);
163 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
164 // implement most of these correctly.
166 report_fatal_error("unsupported pc-relative relocation of difference",
169 // The support for the situation where one or both of the symbols would
170 // require a local relocation is handled just like if the symbols were
171 // external. This is certainly used in the case of debug sections where the
172 // section has only temporary symbols and thus the symbols don't have base
173 // symbols. This is encoded using the section ordinal and non-extern
174 // relocation entries.
176 // Darwin 'as' doesn't emit correct relocations for this (it ends up with a
177 // single SIGNED relocation); reject it for now. Except the case where both
178 // symbols don't have a base, equal but both NULL.
179 if (A_Base == B_Base && A_Base)
180 report_fatal_error("unsupported relocation with identical base", false);
182 // A subtraction expression where both symbols are undefined is a
183 // non-relocatable expression.
184 if (A->isUndefined() && B->isUndefined())
185 report_fatal_error("unsupported relocation with subtraction expression",
188 Value += Writer->getSymbolAddress(&A_SD, Layout) -
189 (A_Base == NULL ? 0 : Writer->getSymbolAddress(A_Base, Layout));
190 Value -= Writer->getSymbolAddress(&B_SD, Layout) -
191 (B_Base == NULL ? 0 : Writer->getSymbolAddress(B_Base, Layout));
194 Index = A_Base->getIndex();
198 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
201 Type = MachO::X86_64_RELOC_UNSIGNED;
203 MachO::any_relocation_info MRE;
204 MRE.r_word0 = FixupOffset;
205 MRE.r_word1 = ((Index << 0) |
210 Writer->addRelocation(Fragment->getParent(), MRE);
213 Index = B_Base->getIndex();
217 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
220 Type = MachO::X86_64_RELOC_SUBTRACTOR;
222 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
223 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
224 const MCSymbolData *Base = Asm.getAtom(&SD);
226 // Relocations inside debug sections always use local relocations when
227 // possible. This seems to be done because the debugger doesn't fully
228 // understand x86_64 relocation entries, and expects to find values that
229 // have already been fixed up.
230 if (Symbol->isInSection()) {
231 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
232 Fragment->getParent()->getSection());
233 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
237 // x86_64 almost always uses external relocations, except when there is no
238 // symbol to use as a base address (a local symbol with no preceding
239 // non-local symbol).
241 Index = Base->getIndex();
244 // Add the local offset, if needed.
246 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
247 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
248 // The index is the section ordinal (1-based).
249 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
251 Value += Writer->getSymbolAddress(&SD, Layout);
254 Value -= FixupAddress + (1 << Log2Size);
255 } else if (Symbol->isVariable()) {
256 const MCExpr *Value = Symbol->getVariableValue();
258 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout,
259 Writer->getSectionAddressMap());
264 report_fatal_error("unsupported relocation of variable '" +
265 Symbol->getName() + "'", false);
268 report_fatal_error("unsupported relocation of undefined symbol '" +
269 Symbol->getName() + "'", false);
272 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
275 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
276 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
277 // rewrite the movq to an leaq at link time if the symbol ends up in
278 // the same linkage unit.
279 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
280 Type = MachO::X86_64_RELOC_GOT_LOAD;
282 Type = MachO::X86_64_RELOC_GOT;
283 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
284 Type = MachO::X86_64_RELOC_TLV;
285 } else if (Modifier != MCSymbolRefExpr::VK_None) {
286 report_fatal_error("unsupported symbol modifier in relocation",
289 Type = MachO::X86_64_RELOC_SIGNED;
291 // The Darwin x86_64 relocation format has a problem where it cannot
292 // encode an address (L<foo> + <constant>) which is outside the atom
293 // containing L<foo>. Generally, this shouldn't occur but it does
294 // happen when we have a RIPrel instruction with data following the
295 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
296 // adjustment Darwin x86_64 uses, the offset is still negative and the
297 // linker has no way to recognize this.
299 // To work around this, Darwin uses several special relocation types
300 // to indicate the offsets. However, the specification or
301 // implementation of these seems to also be incomplete; they should
302 // adjust the addend as well based on the actual encoded instruction
303 // (the additional bias), but instead appear to just look at the final
305 switch (-(Target.getConstant() + (1LL << Log2Size))) {
306 case 1: Type = MachO::X86_64_RELOC_SIGNED_1; break;
307 case 2: Type = MachO::X86_64_RELOC_SIGNED_2; break;
308 case 4: Type = MachO::X86_64_RELOC_SIGNED_4; break;
312 if (Modifier != MCSymbolRefExpr::VK_None)
313 report_fatal_error("unsupported symbol modifier in branch "
314 "relocation", false);
316 Type = MachO::X86_64_RELOC_BRANCH;
319 if (Modifier == MCSymbolRefExpr::VK_GOT) {
320 Type = MachO::X86_64_RELOC_GOT;
321 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
322 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
323 // case all we do is set the PCrel bit in the relocation entry; this is
324 // used with exception handling, for example. The source is required to
325 // include any necessary offset directly.
326 Type = MachO::X86_64_RELOC_GOT;
328 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
329 report_fatal_error("TLVP symbol modifier should have been rip-rel",
331 } else if (Modifier != MCSymbolRefExpr::VK_None)
332 report_fatal_error("unsupported symbol modifier in relocation", false);
334 Type = MachO::X86_64_RELOC_UNSIGNED;
335 unsigned Kind = Fixup.getKind();
336 if (Kind == X86::reloc_signed_4byte)
337 report_fatal_error("32-bit absolute addressing is not supported in "
338 "64-bit mode", false);
343 // x86_64 always writes custom values into the fixups.
346 // struct relocation_info (8 bytes)
347 MachO::any_relocation_info MRE;
348 MRE.r_word0 = FixupOffset;
349 MRE.r_word1 = ((Index << 0) |
354 Writer->addRelocation(Fragment->getParent(), MRE);
357 bool X86MachObjectWriter::RecordScatteredRelocation(MachObjectWriter *Writer,
358 const MCAssembler &Asm,
359 const MCAsmLayout &Layout,
360 const MCFragment *Fragment,
361 const MCFixup &Fixup,
364 uint64_t &FixedValue) {
365 uint64_t OriginalFixedValue = FixedValue;
366 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
367 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
368 unsigned Type = MachO::GENERIC_RELOC_VANILLA;
371 const MCSymbol *A = &Target.getSymA()->getSymbol();
372 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
374 if (!A_SD->getFragment())
375 report_fatal_error("symbol '" + A->getName() +
376 "' can not be undefined in a subtraction expression",
379 uint32_t Value = Writer->getSymbolAddress(A_SD, Layout);
380 uint64_t SecAddr = Writer->getSectionAddress(A_SD->getFragment()->getParent());
381 FixedValue += SecAddr;
384 if (const MCSymbolRefExpr *B = Target.getSymB()) {
385 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
387 if (!B_SD->getFragment())
388 report_fatal_error("symbol '" + B->getSymbol().getName() +
389 "' can not be undefined in a subtraction expression",
392 // Select the appropriate difference relocation type.
394 // Note that there is no longer any semantic difference between these two
395 // relocation types from the linkers point of view, this is done solely for
396 // pedantic compatibility with 'as'.
397 Type = A_SD->isExternal() ? (unsigned)MachO::GENERIC_RELOC_SECTDIFF :
398 (unsigned)MachO::GENERIC_RELOC_LOCAL_SECTDIFF;
399 Value2 = Writer->getSymbolAddress(B_SD, Layout);
400 FixedValue -= Writer->getSectionAddress(B_SD->getFragment()->getParent());
403 // Relocations are written out in reverse order, so the PAIR comes first.
404 if (Type == MachO::GENERIC_RELOC_SECTDIFF ||
405 Type == MachO::GENERIC_RELOC_LOCAL_SECTDIFF) {
406 // If the offset is too large to fit in a scattered relocation,
407 // we're hosed. It's an unfortunate limitation of the MachO format.
408 if (FixupOffset > 0xffffff) {
410 format("0x%x", FixupOffset).print(Buffer, sizeof(Buffer));
411 Asm.getContext().FatalError(Fixup.getLoc(),
412 Twine("Section too large, can't encode "
413 "r_address (") + Buffer +
414 ") into 24 bits of scattered "
415 "relocation entry.");
416 llvm_unreachable("fatal error returned?!");
419 MachO::any_relocation_info MRE;
420 MRE.r_word0 = ((0 << 0) | // r_address
421 (MachO::GENERIC_RELOC_PAIR << 24) | // r_type
425 MRE.r_word1 = Value2;
426 Writer->addRelocation(Fragment->getParent(), MRE);
428 // If the offset is more than 24-bits, it won't fit in a scattered
429 // relocation offset field, so we fall back to using a non-scattered
430 // relocation. This is a bit risky, as if the offset reaches out of
431 // the block and the linker is doing scattered loading on this
432 // symbol, things can go badly.
434 // Required for 'as' compatibility.
435 if (FixupOffset > 0xffffff) {
436 FixedValue = OriginalFixedValue;
441 MachO::any_relocation_info MRE;
442 MRE.r_word0 = ((FixupOffset << 0) |
448 Writer->addRelocation(Fragment->getParent(), MRE);
452 void X86MachObjectWriter::RecordTLVPRelocation(MachObjectWriter *Writer,
453 const MCAssembler &Asm,
454 const MCAsmLayout &Layout,
455 const MCFragment *Fragment,
456 const MCFixup &Fixup,
458 uint64_t &FixedValue) {
459 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
461 "Should only be called with a 32-bit TLVP relocation!");
463 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
464 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
465 unsigned IsPCRel = 0;
467 // Get the symbol data.
468 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
469 unsigned Index = SD_A->getIndex();
471 // We're only going to have a second symbol in pic mode and it'll be a
472 // subtraction from the picbase. For 32-bit pic the addend is the difference
473 // between the picbase and the next address. For 32-bit static the addend is
475 if (Target.getSymB()) {
476 // If this is a subtraction then we're pcrel.
477 uint32_t FixupAddress =
478 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
479 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
481 FixedValue = (FixupAddress - Writer->getSymbolAddress(SD_B, Layout) +
482 Target.getConstant());
483 FixedValue += 1ULL << Log2Size;
488 // struct relocation_info (8 bytes)
489 MachO::any_relocation_info MRE;
491 MRE.r_word1 = ((Index << 0) |
494 (1 << 27) | // r_extern
495 (MachO::GENERIC_RELOC_TLV << 28)); // r_type
496 Writer->addRelocation(Fragment->getParent(), MRE);
499 void X86MachObjectWriter::RecordX86Relocation(MachObjectWriter *Writer,
500 const MCAssembler &Asm,
501 const MCAsmLayout &Layout,
502 const MCFragment *Fragment,
503 const MCFixup &Fixup,
505 uint64_t &FixedValue) {
506 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
507 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
509 // If this is a 32-bit TLVP reloc it's handled a bit differently.
510 if (Target.getSymA() &&
511 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
512 RecordTLVPRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
517 // If this is a difference or a defined symbol plus an offset, then we need a
518 // scattered relocation entry. Differences always require scattered
520 if (Target.getSymB()) {
521 RecordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
522 Target, Log2Size, FixedValue);
526 // Get the symbol data, if any.
527 MCSymbolData *SD = 0;
528 if (Target.getSymA())
529 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
531 // If this is an internal relocation with an offset, it also needs a scattered
533 uint32_t Offset = Target.getConstant();
535 Offset += 1 << Log2Size;
536 // Try to record the scattered relocation if needed. Fall back to non
537 // scattered if necessary (see comments in RecordScatteredRelocation()
539 if (Offset && SD && !Writer->doesSymbolRequireExternRelocation(SD) &&
540 RecordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
541 Target, Log2Size, FixedValue))
545 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
547 unsigned IsExtern = 0;
550 if (Target.isAbsolute()) { // constant
551 // SymbolNum of 0 indicates the absolute section.
553 // FIXME: Currently, these are never generated (see code below). I cannot
554 // find a case where they are actually emitted.
555 Type = MachO::GENERIC_RELOC_VANILLA;
557 // Resolve constant variables.
558 if (SD->getSymbol().isVariable()) {
560 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
561 Res, Layout, Writer->getSectionAddressMap())) {
567 // Check whether we need an external or internal relocation.
568 if (Writer->doesSymbolRequireExternRelocation(SD)) {
570 Index = SD->getIndex();
571 // For external relocations, make sure to offset the fixup value to
572 // compensate for the addend of the symbol address, if it was
573 // undefined. This occurs with weak definitions, for example.
574 if (!SD->Symbol->isUndefined())
575 FixedValue -= Layout.getSymbolOffset(SD);
577 // The index is the section ordinal (1-based).
578 const MCSectionData &SymSD = Asm.getSectionData(
579 SD->getSymbol().getSection());
580 Index = SymSD.getOrdinal() + 1;
581 FixedValue += Writer->getSectionAddress(&SymSD);
584 FixedValue -= Writer->getSectionAddress(Fragment->getParent());
586 Type = MachO::GENERIC_RELOC_VANILLA;
589 // struct relocation_info (8 bytes)
590 MachO::any_relocation_info MRE;
591 MRE.r_word0 = FixupOffset;
592 MRE.r_word1 = ((Index << 0) |
597 Writer->addRelocation(Fragment->getParent(), MRE);
600 MCObjectWriter *llvm::createX86MachObjectWriter(raw_ostream &OS,
603 uint32_t CPUSubtype) {
604 return createMachObjectWriter(new X86MachObjectWriter(Is64Bit,
607 OS, /*IsLittleEndian=*/true);