xref: /freebsd/contrib/llvm-project/lld/MachO/Arch/ARM64.cpp (revision f81cdf24ba5436367377f7c8e8f51f6df2a75ca7)
1 //===- ARM64.cpp ----------------------------------------------------------===//
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 "Arch/ARM64Common.h"
10 #include "InputFiles.h"
11 #include "Symbols.h"
12 #include "SyntheticSections.h"
13 #include "Target.h"
14 
15 #include "lld/Common/ErrorHandler.h"
16 #include "mach-o/compact_unwind_encoding.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/BinaryFormat/MachO.h"
20 #include "llvm/Support/Endian.h"
21 #include "llvm/Support/LEB128.h"
22 #include "llvm/Support/MathExtras.h"
23 
24 using namespace llvm;
25 using namespace llvm::MachO;
26 using namespace llvm::support::endian;
27 using namespace lld;
28 using namespace lld::macho;
29 
30 namespace {
31 
32 struct ARM64 : ARM64Common {
33   ARM64();
34   void writeStub(uint8_t *buf, const Symbol &, uint64_t) const override;
35   void writeStubHelperHeader(uint8_t *buf) const override;
36   void writeStubHelperEntry(uint8_t *buf, const Symbol &,
37                             uint64_t entryAddr) const override;
38 
39   void writeObjCMsgSendStub(uint8_t *buf, Symbol *sym, uint64_t stubsAddr,
40                             uint64_t stubOffset, uint64_t selrefsVA,
41                             uint64_t selectorIndex, uint64_t gotAddr,
42                             uint64_t msgSendIndex) const override;
43   void populateThunk(InputSection *thunk, Symbol *funcSym) override;
44   void applyOptimizationHints(uint8_t *, const ObjFile &) const override;
45 };
46 
47 } // namespace
48 
49 // Random notes on reloc types:
50 // ADDEND always pairs with BRANCH26, PAGE21, or PAGEOFF12
51 // POINTER_TO_GOT: ld64 supports a 4-byte pc-relative form as well as an 8-byte
52 // absolute version of this relocation. The semantics of the absolute relocation
53 // are weird -- it results in the value of the GOT slot being written, instead
54 // of the address. Let's not support it unless we find a real-world use case.
55 static constexpr std::array<RelocAttrs, 11> relocAttrsArray{{
56 #define B(x) RelocAttrBits::x
57     {"UNSIGNED",
58      B(UNSIGNED) | B(ABSOLUTE) | B(EXTERN) | B(LOCAL) | B(BYTE4) | B(BYTE8)},
59     {"SUBTRACTOR", B(SUBTRAHEND) | B(EXTERN) | B(BYTE4) | B(BYTE8)},
60     {"BRANCH26", B(PCREL) | B(EXTERN) | B(BRANCH) | B(BYTE4)},
61     {"PAGE21", B(PCREL) | B(EXTERN) | B(BYTE4)},
62     {"PAGEOFF12", B(ABSOLUTE) | B(EXTERN) | B(BYTE4)},
63     {"GOT_LOAD_PAGE21", B(PCREL) | B(EXTERN) | B(GOT) | B(BYTE4)},
64     {"GOT_LOAD_PAGEOFF12",
65      B(ABSOLUTE) | B(EXTERN) | B(GOT) | B(LOAD) | B(BYTE4)},
66     {"POINTER_TO_GOT", B(PCREL) | B(EXTERN) | B(GOT) | B(POINTER) | B(BYTE4)},
67     {"TLVP_LOAD_PAGE21", B(PCREL) | B(EXTERN) | B(TLV) | B(BYTE4)},
68     {"TLVP_LOAD_PAGEOFF12",
69      B(ABSOLUTE) | B(EXTERN) | B(TLV) | B(LOAD) | B(BYTE4)},
70     {"ADDEND", B(ADDEND)},
71 #undef B
72 }};
73 
74 static constexpr uint32_t stubCode[] = {
75     0x90000010, // 00: adrp  x16, __la_symbol_ptr@page
76     0xf9400210, // 04: ldr   x16, [x16, __la_symbol_ptr@pageoff]
77     0xd61f0200, // 08: br    x16
78 };
79 
80 void ARM64::writeStub(uint8_t *buf8, const Symbol &sym,
81                       uint64_t pointerVA) const {
82   ::writeStub(buf8, stubCode, sym, pointerVA);
83 }
84 
85 static constexpr uint32_t stubHelperHeaderCode[] = {
86     0x90000011, // 00: adrp  x17, _dyld_private@page
87     0x91000231, // 04: add   x17, x17, _dyld_private@pageoff
88     0xa9bf47f0, // 08: stp   x16/x17, [sp, #-16]!
89     0x90000010, // 0c: adrp  x16, dyld_stub_binder@page
90     0xf9400210, // 10: ldr   x16, [x16, dyld_stub_binder@pageoff]
91     0xd61f0200, // 14: br    x16
92 };
93 
94 void ARM64::writeStubHelperHeader(uint8_t *buf8) const {
95   ::writeStubHelperHeader<LP64>(buf8, stubHelperHeaderCode);
96 }
97 
98 static constexpr uint32_t stubHelperEntryCode[] = {
99     0x18000050, // 00: ldr  w16, l0
100     0x14000000, // 04: b    stubHelperHeader
101     0x00000000, // 08: l0: .long 0
102 };
103 
104 void ARM64::writeStubHelperEntry(uint8_t *buf8, const Symbol &sym,
105                                  uint64_t entryVA) const {
106   ::writeStubHelperEntry(buf8, stubHelperEntryCode, sym, entryVA);
107 }
108 
109 static constexpr uint32_t objcStubsFastCode[] = {
110     0x90000001, // adrp  x1, __objc_selrefs@page
111     0xf9400021, // ldr   x1, [x1, @selector("foo")@pageoff]
112     0x90000010, // adrp  x16, _got@page
113     0xf9400210, // ldr   x16, [x16, _objc_msgSend@pageoff]
114     0xd61f0200, // br    x16
115     0xd4200020, // brk   #0x1
116     0xd4200020, // brk   #0x1
117     0xd4200020, // brk   #0x1
118 };
119 
120 void ARM64::writeObjCMsgSendStub(uint8_t *buf, Symbol *sym, uint64_t stubsAddr,
121                                  uint64_t stubOffset, uint64_t selrefsVA,
122                                  uint64_t selectorIndex, uint64_t gotAddr,
123                                  uint64_t msgSendIndex) const {
124   ::writeObjCMsgSendStub<LP64>(buf, objcStubsFastCode, sym, stubsAddr,
125                                stubOffset, selrefsVA, selectorIndex, gotAddr,
126                                msgSendIndex);
127 }
128 
129 // A thunk is the relaxed variation of stubCode. We don't need the
130 // extra indirection through a lazy pointer because the target address
131 // is known at link time.
132 static constexpr uint32_t thunkCode[] = {
133     0x90000010, // 00: adrp  x16, <thunk.ptr>@page
134     0x91000210, // 04: add   x16, [x16,<thunk.ptr>@pageoff]
135     0xd61f0200, // 08: br    x16
136 };
137 
138 void ARM64::populateThunk(InputSection *thunk, Symbol *funcSym) {
139   thunk->align = 4;
140   thunk->data = {reinterpret_cast<const uint8_t *>(thunkCode),
141                  sizeof(thunkCode)};
142   thunk->relocs.emplace_back(/*type=*/ARM64_RELOC_PAGEOFF12,
143                              /*pcrel=*/false, /*length=*/2,
144                              /*offset=*/4, /*addend=*/0,
145                              /*referent=*/funcSym);
146   thunk->relocs.emplace_back(/*type=*/ARM64_RELOC_PAGE21,
147                              /*pcrel=*/true, /*length=*/2,
148                              /*offset=*/0, /*addend=*/0,
149                              /*referent=*/funcSym);
150 }
151 
152 ARM64::ARM64() : ARM64Common(LP64()) {
153   cpuType = CPU_TYPE_ARM64;
154   cpuSubtype = CPU_SUBTYPE_ARM64_ALL;
155 
156   stubSize = sizeof(stubCode);
157   thunkSize = sizeof(thunkCode);
158 
159   objcStubsFastSize = sizeof(objcStubsFastCode);
160   objcStubsAlignment = 32;
161 
162   // Branch immediate is two's complement 26 bits, which is implicitly
163   // multiplied by 4 (since all functions are 4-aligned: The branch range
164   // is -4*(2**(26-1))..4*(2**(26-1) - 1).
165   backwardBranchRange = 128 * 1024 * 1024;
166   forwardBranchRange = backwardBranchRange - 4;
167 
168   modeDwarfEncoding = UNWIND_ARM64_MODE_DWARF;
169   subtractorRelocType = ARM64_RELOC_SUBTRACTOR;
170   unsignedRelocType = ARM64_RELOC_UNSIGNED;
171 
172   stubHelperHeaderSize = sizeof(stubHelperHeaderCode);
173   stubHelperEntrySize = sizeof(stubHelperEntryCode);
174 
175   relocAttrs = {relocAttrsArray.data(), relocAttrsArray.size()};
176 }
177 
178 namespace {
179 struct Adrp {
180   uint32_t destRegister;
181   int64_t addend;
182 };
183 
184 struct Add {
185   uint8_t destRegister;
186   uint8_t srcRegister;
187   uint32_t addend;
188 };
189 
190 enum ExtendType { ZeroExtend = 1, Sign64 = 2, Sign32 = 3 };
191 
192 struct Ldr {
193   uint8_t destRegister;
194   uint8_t baseRegister;
195   uint8_t p2Size;
196   bool isFloat;
197   ExtendType extendType;
198   int64_t offset;
199 };
200 } // namespace
201 
202 static bool parseAdrp(uint32_t insn, Adrp &adrp) {
203   if ((insn & 0x9f000000) != 0x90000000)
204     return false;
205   adrp.destRegister = insn & 0x1f;
206   uint64_t immHi = (insn >> 5) & 0x7ffff;
207   uint64_t immLo = (insn >> 29) & 0x3;
208   adrp.addend = SignExtend64<21>(immLo | (immHi << 2)) * 4096;
209   return true;
210 }
211 
212 static bool parseAdd(uint32_t insn, Add &add) {
213   if ((insn & 0xffc00000) != 0x91000000)
214     return false;
215   add.destRegister = insn & 0x1f;
216   add.srcRegister = (insn >> 5) & 0x1f;
217   add.addend = (insn >> 10) & 0xfff;
218   return true;
219 }
220 
221 static bool parseLdr(uint32_t insn, Ldr &ldr) {
222   ldr.destRegister = insn & 0x1f;
223   ldr.baseRegister = (insn >> 5) & 0x1f;
224   uint8_t size = insn >> 30;
225   uint8_t opc = (insn >> 22) & 3;
226 
227   if ((insn & 0x3fc00000) == 0x39400000) {
228     // LDR (immediate), LDRB (immediate), LDRH (immediate)
229     ldr.p2Size = size;
230     ldr.extendType = ZeroExtend;
231     ldr.isFloat = false;
232   } else if ((insn & 0x3f800000) == 0x39800000) {
233     // LDRSB (immediate), LDRSH (immediate), LDRSW (immediate)
234     ldr.p2Size = size;
235     ldr.extendType = static_cast<ExtendType>(opc);
236     ldr.isFloat = false;
237   } else if ((insn & 0x3f400000) == 0x3d400000) {
238     // LDR (immediate, SIMD&FP)
239     ldr.extendType = ZeroExtend;
240     ldr.isFloat = true;
241     if (opc == 1)
242       ldr.p2Size = size;
243     else if (size == 0 && opc == 3)
244       ldr.p2Size = 4;
245     else
246       return false;
247   } else {
248     return false;
249   }
250   ldr.offset = ((insn >> 10) & 0xfff) << ldr.p2Size;
251   return true;
252 }
253 
254 static bool isValidAdrOffset(int32_t delta) { return isInt<21>(delta); }
255 
256 static void writeAdr(void *loc, uint32_t dest, int32_t delta) {
257   assert(isValidAdrOffset(delta));
258   uint32_t opcode = 0x10000000;
259   uint32_t immHi = (delta & 0x001ffffc) << 3;
260   uint32_t immLo = (delta & 0x00000003) << 29;
261   write32le(loc, opcode | immHi | immLo | dest);
262 }
263 
264 static void writeNop(void *loc) { write32le(loc, 0xd503201f); }
265 
266 static bool isLiteralLdrEligible(const Ldr &ldr) {
267   return ldr.p2Size > 1 && isShiftedInt<19, 2>(ldr.offset);
268 }
269 
270 static void writeLiteralLdr(void *loc, const Ldr &ldr) {
271   assert(isLiteralLdrEligible(ldr));
272   uint32_t imm19 = (ldr.offset / 4 & maskTrailingOnes<uint32_t>(19)) << 5;
273   uint32_t opcode;
274   switch (ldr.p2Size) {
275   case 2:
276     if (ldr.isFloat)
277       opcode = 0x1c000000;
278     else
279       opcode = ldr.extendType == Sign64 ? 0x98000000 : 0x18000000;
280     break;
281   case 3:
282     opcode = ldr.isFloat ? 0x5c000000 : 0x58000000;
283     break;
284   case 4:
285     opcode = 0x9c000000;
286     break;
287   default:
288     llvm_unreachable("Invalid literal ldr size");
289   }
290   write32le(loc, opcode | imm19 | ldr.destRegister);
291 }
292 
293 static bool isImmediateLdrEligible(const Ldr &ldr) {
294   // Note: We deviate from ld64's behavior, which converts to immediate loads
295   // only if ldr.offset < 4096, even though the offset is divided by the load's
296   // size in the 12-bit immediate operand. Only the unsigned offset variant is
297   // supported.
298 
299   uint32_t size = 1 << ldr.p2Size;
300   return ldr.offset >= 0 && (ldr.offset % size) == 0 &&
301          isUInt<12>(ldr.offset >> ldr.p2Size);
302 }
303 
304 static void writeImmediateLdr(void *loc, const Ldr &ldr) {
305   assert(isImmediateLdrEligible(ldr));
306   uint32_t opcode = 0x39000000;
307   if (ldr.isFloat) {
308     opcode |= 0x04000000;
309     assert(ldr.extendType == ZeroExtend);
310   }
311   opcode |= ldr.destRegister;
312   opcode |= ldr.baseRegister << 5;
313   uint8_t size, opc;
314   if (ldr.p2Size == 4) {
315     size = 0;
316     opc = 3;
317   } else {
318     opc = ldr.extendType;
319     size = ldr.p2Size;
320   }
321   uint32_t immBits = ldr.offset >> ldr.p2Size;
322   write32le(loc, opcode | (immBits << 10) | (opc << 22) | (size << 30));
323 }
324 
325 // Transforms a pair of adrp+add instructions into an adr instruction if the
326 // target is within the +/- 1 MiB range allowed by the adr's 21 bit signed
327 // immediate offset.
328 //
329 //   adrp xN, _foo@PAGE
330 //   add  xM, xN, _foo@PAGEOFF
331 // ->
332 //   adr  xM, _foo
333 //   nop
334 static void applyAdrpAdd(uint8_t *buf, const ConcatInputSection *isec,
335                          uint64_t offset1, uint64_t offset2) {
336   uint32_t ins1 = read32le(buf + offset1);
337   uint32_t ins2 = read32le(buf + offset2);
338   Adrp adrp;
339   Add add;
340   if (!parseAdrp(ins1, adrp) || !parseAdd(ins2, add))
341     return;
342   if (adrp.destRegister != add.srcRegister)
343     return;
344 
345   uint64_t addr1 = isec->getVA() + offset1;
346   uint64_t referent = pageBits(addr1) + adrp.addend + add.addend;
347   int64_t delta = referent - addr1;
348   if (!isValidAdrOffset(delta))
349     return;
350 
351   writeAdr(buf + offset1, add.destRegister, delta);
352   writeNop(buf + offset2);
353 }
354 
355 // Transforms two adrp instructions into a single adrp if their referent
356 // addresses are located on the same 4096 byte page.
357 //
358 //   adrp xN, _foo@PAGE
359 //   adrp xN, _bar@PAGE
360 // ->
361 //   adrp xN, _foo@PAGE
362 //   nop
363 static void applyAdrpAdrp(uint8_t *buf, const ConcatInputSection *isec,
364                           uint64_t offset1, uint64_t offset2) {
365   uint32_t ins1 = read32le(buf + offset1);
366   uint32_t ins2 = read32le(buf + offset2);
367   Adrp adrp1, adrp2;
368   if (!parseAdrp(ins1, adrp1) || !parseAdrp(ins2, adrp2))
369     return;
370   if (adrp1.destRegister != adrp2.destRegister)
371     return;
372 
373   uint64_t page1 = pageBits(offset1 + isec->getVA()) + adrp1.addend;
374   uint64_t page2 = pageBits(offset2 + isec->getVA()) + adrp2.addend;
375   if (page1 != page2)
376     return;
377 
378   writeNop(buf + offset2);
379 }
380 
381 // Transforms a pair of adrp+ldr (immediate) instructions into an ldr (literal)
382 // load from a PC-relative address if it is 4-byte aligned and within +/- 1 MiB,
383 // as ldr can encode a signed 19-bit offset that gets multiplied by 4.
384 //
385 //   adrp xN, _foo@PAGE
386 //   ldr  xM, [xN, _foo@PAGEOFF]
387 // ->
388 //   nop
389 //   ldr  xM, _foo
390 static void applyAdrpLdr(uint8_t *buf, const ConcatInputSection *isec,
391                          uint64_t offset1, uint64_t offset2) {
392   uint32_t ins1 = read32le(buf + offset1);
393   uint32_t ins2 = read32le(buf + offset2);
394   Adrp adrp;
395   Ldr ldr;
396   if (!parseAdrp(ins1, adrp) || !parseLdr(ins2, ldr))
397     return;
398   if (adrp.destRegister != ldr.baseRegister)
399     return;
400 
401   uint64_t addr1 = isec->getVA() + offset1;
402   uint64_t addr2 = isec->getVA() + offset2;
403   uint64_t referent = pageBits(addr1) + adrp.addend + ldr.offset;
404   ldr.offset = referent - addr2;
405   if (!isLiteralLdrEligible(ldr))
406     return;
407 
408   writeNop(buf + offset1);
409   writeLiteralLdr(buf + offset2, ldr);
410 }
411 
412 // GOT loads are emitted by the compiler as a pair of adrp and ldr instructions,
413 // but they may be changed to adrp+add by relaxGotLoad(). This hint performs
414 // the AdrpLdr or AdrpAdd transformation depending on whether it was relaxed.
415 static void applyAdrpLdrGot(uint8_t *buf, const ConcatInputSection *isec,
416                             uint64_t offset1, uint64_t offset2) {
417   uint32_t ins2 = read32le(buf + offset2);
418   Add add;
419   Ldr ldr;
420   if (parseAdd(ins2, add))
421     applyAdrpAdd(buf, isec, offset1, offset2);
422   else if (parseLdr(ins2, ldr))
423     applyAdrpLdr(buf, isec, offset1, offset2);
424 }
425 
426 // Optimizes an adrp+add+ldr sequence used for loading from a local symbol's
427 // address by loading directly if it's close enough, or to an adrp(p)+ldr
428 // sequence if it's not.
429 //
430 //   adrp x0, _foo@PAGE
431 //   add  x1, x0, _foo@PAGEOFF
432 //   ldr  x2, [x1, #off]
433 static void applyAdrpAddLdr(uint8_t *buf, const ConcatInputSection *isec,
434                             uint64_t offset1, uint64_t offset2,
435                             uint64_t offset3) {
436   uint32_t ins1 = read32le(buf + offset1);
437   Adrp adrp;
438   if (!parseAdrp(ins1, adrp))
439     return;
440   uint32_t ins2 = read32le(buf + offset2);
441   Add add;
442   if (!parseAdd(ins2, add))
443     return;
444   uint32_t ins3 = read32le(buf + offset3);
445   Ldr ldr;
446   if (!parseLdr(ins3, ldr))
447     return;
448   if (adrp.destRegister != add.srcRegister)
449     return;
450   if (add.destRegister != ldr.baseRegister)
451     return;
452 
453   // Load from the target address directly.
454   //   nop
455   //   nop
456   //   ldr x2, [_foo + #off]
457   uint64_t addr1 = isec->getVA() + offset1;
458   uint64_t addr3 = isec->getVA() + offset3;
459   uint64_t referent = pageBits(addr1) + adrp.addend + add.addend;
460   Ldr literalLdr = ldr;
461   literalLdr.offset += referent - addr3;
462   if (isLiteralLdrEligible(literalLdr)) {
463     writeNop(buf + offset1);
464     writeNop(buf + offset2);
465     writeLiteralLdr(buf + offset3, literalLdr);
466     return;
467   }
468 
469   // Load the target address into a register and load from there indirectly.
470   //   adr x1, _foo
471   //   nop
472   //   ldr x2, [x1, #off]
473   int64_t adrOffset = referent - addr1;
474   if (isValidAdrOffset(adrOffset)) {
475     writeAdr(buf + offset1, ldr.baseRegister, adrOffset);
476     // Note: ld64 moves the offset into the adr instruction for AdrpAddLdr, but
477     // not for AdrpLdrGotLdr. Its effect is the same either way.
478     writeNop(buf + offset2);
479     return;
480   }
481 
482   // Move the target's page offset into the ldr's immediate offset.
483   //   adrp x0, _foo@PAGE
484   //   nop
485   //   ldr x2, [x0, _foo@PAGEOFF + #off]
486   Ldr immediateLdr = ldr;
487   immediateLdr.baseRegister = adrp.destRegister;
488   immediateLdr.offset += add.addend;
489   if (isImmediateLdrEligible(immediateLdr)) {
490     writeNop(buf + offset2);
491     writeImmediateLdr(buf + offset3, immediateLdr);
492     return;
493   }
494 }
495 
496 // Relaxes a GOT-indirect load.
497 // If the referenced symbol is external and its GOT entry is within +/- 1 MiB,
498 // the GOT entry can be loaded with a single literal ldr instruction.
499 // If the referenced symbol is local and thus has been relaxed to adrp+add+ldr,
500 // we perform the AdrpAddLdr transformation.
501 static void applyAdrpLdrGotLdr(uint8_t *buf, const ConcatInputSection *isec,
502                                uint64_t offset1, uint64_t offset2,
503                                uint64_t offset3) {
504   uint32_t ins2 = read32le(buf + offset2);
505   Add add;
506   Ldr ldr2;
507 
508   if (parseAdd(ins2, add)) {
509     applyAdrpAddLdr(buf, isec, offset1, offset2, offset3);
510   } else if (parseLdr(ins2, ldr2)) {
511     // adrp x1, _foo@GOTPAGE
512     // ldr  x2, [x1, _foo@GOTPAGEOFF]
513     // ldr  x3, [x2, #off]
514 
515     uint32_t ins1 = read32le(buf + offset1);
516     Adrp adrp;
517     if (!parseAdrp(ins1, adrp))
518       return;
519     uint32_t ins3 = read32le(buf + offset3);
520     Ldr ldr3;
521     if (!parseLdr(ins3, ldr3))
522       return;
523 
524     if (ldr2.baseRegister != adrp.destRegister)
525       return;
526     if (ldr3.baseRegister != ldr2.destRegister)
527       return;
528     // Loads from the GOT must be pointer sized.
529     if (ldr2.p2Size != 3 || ldr2.isFloat)
530       return;
531 
532     uint64_t addr1 = isec->getVA() + offset1;
533     uint64_t addr2 = isec->getVA() + offset2;
534     uint64_t referent = pageBits(addr1) + adrp.addend + ldr2.offset;
535     // Load the GOT entry's address directly.
536     //   nop
537     //   ldr x2, _foo@GOTPAGE + _foo@GOTPAGEOFF
538     //   ldr x3, [x2, #off]
539     Ldr literalLdr = ldr2;
540     literalLdr.offset = referent - addr2;
541     if (isLiteralLdrEligible(literalLdr)) {
542       writeNop(buf + offset1);
543       writeLiteralLdr(buf + offset2, literalLdr);
544     }
545   }
546 }
547 
548 static uint64_t readValue(const uint8_t *&ptr, const uint8_t *end) {
549   unsigned int n = 0;
550   uint64_t value = decodeULEB128(ptr, &n, end);
551   ptr += n;
552   return value;
553 }
554 
555 template <typename Callback>
556 static void forEachHint(ArrayRef<uint8_t> data, Callback callback) {
557   std::array<uint64_t, 3> args;
558 
559   for (const uint8_t *p = data.begin(), *end = data.end(); p < end;) {
560     uint64_t type = readValue(p, end);
561     if (type == 0)
562       break;
563 
564     uint64_t argCount = readValue(p, end);
565     // All known LOH types as of 2022-09 have 3 or fewer arguments; skip others.
566     if (argCount > 3) {
567       for (unsigned i = 0; i < argCount; ++i)
568         readValue(p, end);
569       continue;
570     }
571 
572     for (unsigned i = 0; i < argCount; ++i)
573       args[i] = readValue(p, end);
574     callback(type, ArrayRef<uint64_t>(args.data(), argCount));
575   }
576 }
577 
578 // On RISC architectures like arm64, materializing a memory address generally
579 // takes multiple instructions. If the referenced symbol is located close enough
580 // in memory, fewer instructions are needed.
581 //
582 // Linker optimization hints record where addresses are computed. After
583 // addresses have been assigned, if possible, we change them to a shorter
584 // sequence of instructions. The size of the binary is not modified; the
585 // eliminated instructions are replaced with NOPs. This still leads to faster
586 // code as the CPU can skip over NOPs quickly.
587 //
588 // LOHs are specified by the LC_LINKER_OPTIMIZATION_HINTS load command, which
589 // points to a sequence of ULEB128-encoded numbers. Each entry specifies a
590 // transformation kind, and 2 or 3 addresses where the instructions are located.
591 void ARM64::applyOptimizationHints(uint8_t *outBuf, const ObjFile &obj) const {
592   ArrayRef<uint8_t> data = obj.getOptimizationHints();
593   if (data.empty())
594     return;
595 
596   const ConcatInputSection *section = nullptr;
597   uint64_t sectionAddr = 0;
598   uint8_t *buf = nullptr;
599 
600   auto findSection = [&](uint64_t addr) {
601     if (section && addr >= sectionAddr &&
602         addr < sectionAddr + section->getSize())
603       return true;
604 
605     auto secIt = std::prev(llvm::upper_bound(
606         obj.sections, addr,
607         [](uint64_t off, const Section *sec) { return off < sec->addr; }));
608     const Section *sec = *secIt;
609 
610     auto subsecIt = std::prev(llvm::upper_bound(
611         sec->subsections, addr - sec->addr,
612         [](uint64_t off, Subsection subsec) { return off < subsec.offset; }));
613     const Subsection &subsec = *subsecIt;
614     const ConcatInputSection *isec =
615         dyn_cast_or_null<ConcatInputSection>(subsec.isec);
616     if (!isec || isec->shouldOmitFromOutput())
617       return false;
618 
619     section = isec;
620     sectionAddr = subsec.offset + sec->addr;
621     buf = outBuf + section->outSecOff + section->parent->fileOff;
622     return true;
623   };
624 
625   auto isValidOffset = [&](uint64_t offset) {
626     if (offset < sectionAddr || offset >= sectionAddr + section->getSize()) {
627       error(toString(&obj) +
628             ": linker optimization hint spans multiple sections");
629       return false;
630     }
631     return true;
632   };
633 
634   bool hasAdrpAdrp = false;
635   forEachHint(data, [&](uint64_t kind, ArrayRef<uint64_t> args) {
636     if (kind == LOH_ARM64_ADRP_ADRP) {
637       hasAdrpAdrp = true;
638       return;
639     }
640 
641     if (!findSection(args[0]))
642       return;
643     switch (kind) {
644     case LOH_ARM64_ADRP_ADD:
645       if (isValidOffset(args[1]))
646         applyAdrpAdd(buf, section, args[0] - sectionAddr,
647                      args[1] - sectionAddr);
648       break;
649     case LOH_ARM64_ADRP_LDR:
650       if (isValidOffset(args[1]))
651         applyAdrpLdr(buf, section, args[0] - sectionAddr,
652                      args[1] - sectionAddr);
653       break;
654     case LOH_ARM64_ADRP_LDR_GOT:
655       if (isValidOffset(args[1]))
656         applyAdrpLdrGot(buf, section, args[0] - sectionAddr,
657                         args[1] - sectionAddr);
658       break;
659     case LOH_ARM64_ADRP_ADD_LDR:
660       if (isValidOffset(args[1]) && isValidOffset(args[2]))
661         applyAdrpAddLdr(buf, section, args[0] - sectionAddr,
662                         args[1] - sectionAddr, args[2] - sectionAddr);
663       break;
664     case LOH_ARM64_ADRP_LDR_GOT_LDR:
665       if (isValidOffset(args[1]) && isValidOffset(args[2]))
666         applyAdrpLdrGotLdr(buf, section, args[0] - sectionAddr,
667                            args[1] - sectionAddr, args[2] - sectionAddr);
668       break;
669     case LOH_ARM64_ADRP_ADD_STR:
670     case LOH_ARM64_ADRP_LDR_GOT_STR:
671       // TODO: Implement these
672       break;
673     }
674   });
675 
676   if (!hasAdrpAdrp)
677     return;
678 
679   // AdrpAdrp optimization hints are performed in a second pass because they
680   // might interfere with other transformations. For instance, consider the
681   // following input:
682   //
683   //   adrp x0, _foo@PAGE
684   //   add  x1, x0, _foo@PAGEOFF
685   //   adrp x0, _bar@PAGE
686   //   add  x2, x0, _bar@PAGEOFF
687   //
688   // If we perform the AdrpAdrp relaxation first, we get:
689   //
690   //   adrp x0, _foo@PAGE
691   //   add  x1, x0, _foo@PAGEOFF
692   //   nop
693   //   add x2, x0, _bar@PAGEOFF
694   //
695   // If we then apply AdrpAdd to the first two instructions, the add will have a
696   // garbage value in x0:
697   //
698   //   adr  x1, _foo
699   //   nop
700   //   nop
701   //   add  x2, x0, _bar@PAGEOFF
702   forEachHint(data, [&](uint64_t kind, ArrayRef<uint64_t> args) {
703     if (kind != LOH_ARM64_ADRP_ADRP)
704       return;
705     if (!findSection(args[0]))
706       return;
707     if (isValidOffset(args[1]))
708       applyAdrpAdrp(buf, section, args[0] - sectionAddr, args[1] - sectionAddr);
709   });
710 }
711 
712 TargetInfo *macho::createARM64TargetInfo() {
713   static ARM64 t;
714   return &t;
715 }
716