1 //===-- AArch64MachObjectWriter.cpp - ARM Mach Object Writer --------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8
9 #include "MCTargetDesc/AArch64FixupKinds.h"
10 #include "MCTargetDesc/AArch64MCTargetDesc.h"
11 #include "llvm/ADT/Twine.h"
12 #include "llvm/BinaryFormat/MachO.h"
13 #include "llvm/MC/MCAsmInfo.h"
14 #include "llvm/MC/MCAsmInfoDarwin.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCContext.h"
17 #include "llvm/MC/MCExpr.h"
18 #include "llvm/MC/MCFixup.h"
19 #include "llvm/MC/MCFragment.h"
20 #include "llvm/MC/MCMachObjectWriter.h"
21 #include "llvm/MC/MCSection.h"
22 #include "llvm/MC/MCSectionMachO.h"
23 #include "llvm/MC/MCSymbol.h"
24 #include "llvm/MC/MCValue.h"
25 #include "llvm/Support/Casting.h"
26 #include "llvm/Support/MathExtras.h"
27 #include <cassert>
28 #include <cstdint>
29
30 using namespace llvm;
31
32 namespace {
33
34 class AArch64MachObjectWriter : public MCMachObjectTargetWriter {
35 bool getAArch64FixupKindMachOInfo(const MCFixup &Fixup, unsigned &RelocType,
36 const MCSymbolRefExpr *Sym,
37 unsigned &Log2Size, const MCAssembler &Asm);
38
39 public:
AArch64MachObjectWriter(uint32_t CPUType,uint32_t CPUSubtype,bool IsILP32)40 AArch64MachObjectWriter(uint32_t CPUType, uint32_t CPUSubtype, bool IsILP32)
41 : MCMachObjectTargetWriter(!IsILP32 /* is64Bit */, CPUType, CPUSubtype) {}
42
43 void recordRelocation(MachObjectWriter *Writer, MCAssembler &Asm,
44 const MCFragment *Fragment, const MCFixup &Fixup,
45 MCValue Target, uint64_t &FixedValue) override;
46 };
47
48 } // end anonymous namespace
49
getAArch64FixupKindMachOInfo(const MCFixup & Fixup,unsigned & RelocType,const MCSymbolRefExpr * Sym,unsigned & Log2Size,const MCAssembler & Asm)50 bool AArch64MachObjectWriter::getAArch64FixupKindMachOInfo(
51 const MCFixup &Fixup, unsigned &RelocType, const MCSymbolRefExpr *Sym,
52 unsigned &Log2Size, const MCAssembler &Asm) {
53 RelocType = unsigned(MachO::ARM64_RELOC_UNSIGNED);
54 Log2Size = ~0U;
55
56 switch (Fixup.getTargetKind()) {
57 default:
58 return false;
59
60 case FK_Data_1:
61 Log2Size = Log2_32(1);
62 return true;
63 case FK_Data_2:
64 Log2Size = Log2_32(2);
65 return true;
66 case FK_Data_4:
67 Log2Size = Log2_32(4);
68 if (Sym->getKind() == MCSymbolRefExpr::VK_GOT)
69 RelocType = unsigned(MachO::ARM64_RELOC_POINTER_TO_GOT);
70 return true;
71 case FK_Data_8:
72 Log2Size = Log2_32(8);
73 if (Sym->getKind() == MCSymbolRefExpr::VK_GOT)
74 RelocType = unsigned(MachO::ARM64_RELOC_POINTER_TO_GOT);
75 return true;
76 case AArch64::fixup_aarch64_add_imm12:
77 case AArch64::fixup_aarch64_ldst_imm12_scale1:
78 case AArch64::fixup_aarch64_ldst_imm12_scale2:
79 case AArch64::fixup_aarch64_ldst_imm12_scale4:
80 case AArch64::fixup_aarch64_ldst_imm12_scale8:
81 case AArch64::fixup_aarch64_ldst_imm12_scale16:
82 Log2Size = Log2_32(4);
83 switch (Sym->getKind()) {
84 default:
85 return false;
86 case MCSymbolRefExpr::VK_PAGEOFF:
87 RelocType = unsigned(MachO::ARM64_RELOC_PAGEOFF12);
88 return true;
89 case MCSymbolRefExpr::VK_GOTPAGEOFF:
90 RelocType = unsigned(MachO::ARM64_RELOC_GOT_LOAD_PAGEOFF12);
91 return true;
92 case MCSymbolRefExpr::VK_TLVPPAGEOFF:
93 RelocType = unsigned(MachO::ARM64_RELOC_TLVP_LOAD_PAGEOFF12);
94 return true;
95 }
96 case AArch64::fixup_aarch64_pcrel_adrp_imm21:
97 Log2Size = Log2_32(4);
98 // This encompasses the relocation for the whole 21-bit value.
99 switch (Sym->getKind()) {
100 default:
101 Asm.getContext().reportError(Fixup.getLoc(),
102 "ADR/ADRP relocations must be GOT relative");
103 return false;
104 case MCSymbolRefExpr::VK_PAGE:
105 RelocType = unsigned(MachO::ARM64_RELOC_PAGE21);
106 return true;
107 case MCSymbolRefExpr::VK_GOTPAGE:
108 RelocType = unsigned(MachO::ARM64_RELOC_GOT_LOAD_PAGE21);
109 return true;
110 case MCSymbolRefExpr::VK_TLVPPAGE:
111 RelocType = unsigned(MachO::ARM64_RELOC_TLVP_LOAD_PAGE21);
112 return true;
113 }
114 return true;
115 case AArch64::fixup_aarch64_pcrel_branch26:
116 case AArch64::fixup_aarch64_pcrel_call26:
117 Log2Size = Log2_32(4);
118 RelocType = unsigned(MachO::ARM64_RELOC_BRANCH26);
119 return true;
120 }
121 }
122
canUseLocalRelocation(const MCSectionMachO & Section,const MCSymbol & Symbol,unsigned Log2Size)123 static bool canUseLocalRelocation(const MCSectionMachO &Section,
124 const MCSymbol &Symbol, unsigned Log2Size) {
125 // Debug info sections can use local relocations.
126 if (Section.hasAttribute(MachO::S_ATTR_DEBUG))
127 return true;
128
129 // Otherwise, only pointer sized relocations are supported.
130 if (Log2Size != 3)
131 return false;
132
133 // But only if they don't point to a few forbidden sections.
134 if (!Symbol.isInSection())
135 return true;
136 const MCSectionMachO &RefSec = cast<MCSectionMachO>(Symbol.getSection());
137 if (RefSec.getType() == MachO::S_CSTRING_LITERALS)
138 return false;
139
140 if (RefSec.getSegmentName() == "__DATA" &&
141 (RefSec.getName() == "__cfstring" ||
142 RefSec.getName() == "__objc_classrefs"))
143 return false;
144
145 return true;
146 }
147
recordRelocation(MachObjectWriter * Writer,MCAssembler & Asm,const MCFragment * Fragment,const MCFixup & Fixup,MCValue Target,uint64_t & FixedValue)148 void AArch64MachObjectWriter::recordRelocation(
149 MachObjectWriter *Writer, MCAssembler &Asm, const MCFragment *Fragment,
150 const MCFixup &Fixup, MCValue Target, uint64_t &FixedValue) {
151 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
152
153 // See <reloc.h>.
154 uint32_t FixupOffset = Asm.getFragmentOffset(*Fragment);
155 unsigned Log2Size = 0;
156 int64_t Value = 0;
157 unsigned Index = 0;
158 unsigned Type = 0;
159 unsigned Kind = Fixup.getKind();
160 const MCSymbol *RelSymbol = nullptr;
161
162 FixupOffset += Fixup.getOffset();
163
164 // AArch64 pcrel relocation addends do not include the section offset.
165 if (IsPCRel)
166 FixedValue += FixupOffset;
167
168 // ADRP fixups use relocations for the whole symbol value and only
169 // put the addend in the instruction itself. Clear out any value the
170 // generic code figured out from the sybmol definition.
171 if (Kind == AArch64::fixup_aarch64_pcrel_adrp_imm21)
172 FixedValue = 0;
173
174 // imm19 relocations are for conditional branches, which require
175 // assembler local symbols. If we got here, that's not what we have,
176 // so complain loudly.
177 if (Kind == AArch64::fixup_aarch64_pcrel_branch19) {
178 Asm.getContext().reportError(Fixup.getLoc(),
179 "conditional branch requires assembler-local"
180 " label. '" +
181 Target.getSymA()->getSymbol().getName() +
182 "' is external.");
183 return;
184 }
185
186 // 14-bit branch relocations should only target internal labels, and so
187 // should never get here.
188 if (Kind == AArch64::fixup_aarch64_pcrel_branch14) {
189 Asm.getContext().reportError(Fixup.getLoc(),
190 "Invalid relocation on conditional branch!");
191 return;
192 }
193
194 if (!getAArch64FixupKindMachOInfo(Fixup, Type, Target.getSymA(), Log2Size,
195 Asm)) {
196 Asm.getContext().reportError(Fixup.getLoc(), "unknown AArch64 fixup kind!");
197 return;
198 }
199
200 Value = Target.getConstant();
201
202 if (Target.isAbsolute()) { // constant
203 // FIXME: Should this always be extern?
204 // SymbolNum of 0 indicates the absolute section.
205 Type = MachO::ARM64_RELOC_UNSIGNED;
206
207 if (IsPCRel) {
208 Asm.getContext().reportError(Fixup.getLoc(),
209 "PC relative absolute relocation!");
210 return;
211
212 // FIXME: x86_64 sets the type to a branch reloc here. Should we do
213 // something similar?
214 }
215 } else if (Target.getSymB()) { // A - B + constant
216 const MCSymbol *A = &Target.getSymA()->getSymbol();
217 const MCSymbol *A_Base = Writer->getAtom(*A);
218
219 const MCSymbol *B = &Target.getSymB()->getSymbol();
220 const MCSymbol *B_Base = Writer->getAtom(*B);
221
222 // Check for "_foo@got - .", which comes through here as:
223 // Ltmp0:
224 // ... _foo@got - Ltmp0
225 if (Target.getSymA()->getKind() == MCSymbolRefExpr::VK_GOT &&
226 Target.getSymB()->getKind() == MCSymbolRefExpr::VK_None &&
227 Asm.getSymbolOffset(*B) ==
228 Asm.getFragmentOffset(*Fragment) + Fixup.getOffset()) {
229 // SymB is the PC, so use a PC-rel pointer-to-GOT relocation.
230 Type = MachO::ARM64_RELOC_POINTER_TO_GOT;
231 IsPCRel = 1;
232 MachO::any_relocation_info MRE;
233 MRE.r_word0 = FixupOffset;
234 MRE.r_word1 = (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
235 Writer->addRelocation(A_Base, Fragment->getParent(), MRE);
236 return;
237 } else if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
238 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None) {
239 // Otherwise, neither symbol can be modified.
240 Asm.getContext().reportError(Fixup.getLoc(),
241 "unsupported relocation of modified symbol");
242 return;
243 }
244
245 // We don't support PCrel relocations of differences.
246 if (IsPCRel) {
247 Asm.getContext().reportError(Fixup.getLoc(),
248 "unsupported pc-relative relocation of "
249 "difference");
250 return;
251 }
252
253 // AArch64 always uses external relocations. If there is no symbol to use as
254 // a base address (a local symbol with no preceding non-local symbol),
255 // error out.
256 //
257 // FIXME: We should probably just synthesize an external symbol and use
258 // that.
259 if (!A_Base) {
260 Asm.getContext().reportError(
261 Fixup.getLoc(),
262 "unsupported relocation of local symbol '" + A->getName() +
263 "'. Must have non-local symbol earlier in section.");
264 return;
265 }
266 if (!B_Base) {
267 Asm.getContext().reportError(
268 Fixup.getLoc(),
269 "unsupported relocation of local symbol '" + B->getName() +
270 "'. Must have non-local symbol earlier in section.");
271 return;
272 }
273
274 if (A_Base == B_Base && A_Base) {
275 Asm.getContext().reportError(
276 Fixup.getLoc(), "unsupported relocation with identical base");
277 return;
278 }
279
280 Value += (!A->getFragment() ? 0 : Writer->getSymbolAddress(*A, Asm)) -
281 (!A_Base || !A_Base->getFragment()
282 ? 0
283 : Writer->getSymbolAddress(*A_Base, Asm));
284 Value -= (!B->getFragment() ? 0 : Writer->getSymbolAddress(*B, Asm)) -
285 (!B_Base || !B_Base->getFragment()
286 ? 0
287 : Writer->getSymbolAddress(*B_Base, Asm));
288
289 Type = MachO::ARM64_RELOC_UNSIGNED;
290
291 MachO::any_relocation_info MRE;
292 MRE.r_word0 = FixupOffset;
293 MRE.r_word1 = (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
294 Writer->addRelocation(A_Base, Fragment->getParent(), MRE);
295
296 RelSymbol = B_Base;
297 Type = MachO::ARM64_RELOC_SUBTRACTOR;
298 } else { // A + constant
299 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
300 const MCSectionMachO &Section =
301 static_cast<const MCSectionMachO &>(*Fragment->getParent());
302
303 bool CanUseLocalRelocation =
304 canUseLocalRelocation(Section, *Symbol, Log2Size);
305 if (Symbol->isTemporary() && (Value || !CanUseLocalRelocation)) {
306 // Make sure that the symbol is actually in a section here. If it isn't,
307 // emit an error and exit.
308 if (!Symbol->isInSection()) {
309 Asm.getContext().reportError(
310 Fixup.getLoc(),
311 "unsupported relocation of local symbol '" + Symbol->getName() +
312 "'. Must have non-local symbol earlier in section.");
313 return;
314 }
315 const MCSection &Sec = Symbol->getSection();
316 if (!MCAsmInfoDarwin::isSectionAtomizableBySymbols(Sec))
317 Symbol->setUsedInReloc();
318 }
319
320 const MCSymbol *Base = Writer->getAtom(*Symbol);
321
322 // If the symbol is a variable it can either be in a section and
323 // we have a base or it is absolute and should have been expanded.
324 assert(!Symbol->isVariable() || Base);
325
326 // Relocations inside debug sections always use local relocations when
327 // possible. This seems to be done because the debugger doesn't fully
328 // understand relocation entries and expects to find values that
329 // have already been fixed up.
330 if (Symbol->isInSection()) {
331 if (Section.hasAttribute(MachO::S_ATTR_DEBUG))
332 Base = nullptr;
333 }
334
335 // AArch64 uses external relocations as much as possible. For debug
336 // sections, and for pointer-sized relocations (.quad), we allow section
337 // relocations. It's code sections that run into trouble.
338 if (Base) {
339 RelSymbol = Base;
340
341 // Add the local offset, if needed.
342 if (Base != Symbol)
343 Value += Asm.getSymbolOffset(*Symbol) - Asm.getSymbolOffset(*Base);
344 } else if (Symbol->isInSection()) {
345 if (!CanUseLocalRelocation) {
346 Asm.getContext().reportError(
347 Fixup.getLoc(),
348 "unsupported relocation of local symbol '" + Symbol->getName() +
349 "'. Must have non-local symbol earlier in section.");
350 return;
351 }
352 // Adjust the relocation to be section-relative.
353 // The index is the section ordinal (1-based).
354 const MCSection &Sec = Symbol->getSection();
355 Index = Sec.getOrdinal() + 1;
356 Value += Writer->getSymbolAddress(*Symbol, Asm);
357
358 if (IsPCRel)
359 Value -= Writer->getFragmentAddress(Asm, Fragment) + Fixup.getOffset() +
360 (1ULL << Log2Size);
361 } else {
362 llvm_unreachable(
363 "This constant variable should have been expanded during evaluation");
364 }
365 }
366
367 // If the relocation kind is Branch26, Page21, or Pageoff12, any addend
368 // is represented via an Addend relocation, not encoded directly into
369 // the instruction.
370 if ((Type == MachO::ARM64_RELOC_BRANCH26 ||
371 Type == MachO::ARM64_RELOC_PAGE21 ||
372 Type == MachO::ARM64_RELOC_PAGEOFF12) &&
373 Value) {
374 if (!isInt<24>(Value)) {
375 Asm.getContext().reportError(Fixup.getLoc(),
376 "addend too big for relocation");
377 return;
378 }
379
380 MachO::any_relocation_info MRE;
381 MRE.r_word0 = FixupOffset;
382 MRE.r_word1 =
383 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
384 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
385
386 // Now set up the Addend relocation.
387 Type = MachO::ARM64_RELOC_ADDEND;
388 Index = Value;
389 RelSymbol = nullptr;
390 IsPCRel = 0;
391 Log2Size = 2;
392
393 // Put zero into the instruction itself. The addend is in the relocation.
394 Value = 0;
395 }
396
397 // If there's any addend left to handle, encode it in the instruction.
398 FixedValue = Value;
399
400 // struct relocation_info (8 bytes)
401 MachO::any_relocation_info MRE;
402 MRE.r_word0 = FixupOffset;
403 MRE.r_word1 =
404 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
405 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
406 }
407
408 std::unique_ptr<MCObjectTargetWriter>
createAArch64MachObjectWriter(uint32_t CPUType,uint32_t CPUSubtype,bool IsILP32)409 llvm::createAArch64MachObjectWriter(uint32_t CPUType, uint32_t CPUSubtype,
410 bool IsILP32) {
411 return std::make_unique<AArch64MachObjectWriter>(CPUType, CPUSubtype,
412 IsILP32);
413 }
414