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/MathExtras.h"
22 
23 using namespace llvm;
24 using namespace llvm::MachO;
25 using namespace llvm::support::endian;
26 using namespace lld;
27 using namespace lld::macho;
28 
29 namespace {
30 
31 struct ARM64 : ARM64Common {
32   ARM64();
33   void writeStub(uint8_t *buf, const Symbol &) const override;
34   void writeStubHelperHeader(uint8_t *buf) const override;
35   void writeStubHelperEntry(uint8_t *buf, const Symbol &,
36                             uint64_t entryAddr) const override;
37   const RelocAttrs &getRelocAttrs(uint8_t type) const override;
38   void populateThunk(InputSection *thunk, Symbol *funcSym) override;
39   void applyOptimizationHints(uint8_t *, const ConcatInputSection *,
40                               ArrayRef<uint64_t>) const override;
41 };
42 
43 } // namespace
44 
45 // Random notes on reloc types:
46 // ADDEND always pairs with BRANCH26, PAGE21, or PAGEOFF12
47 // POINTER_TO_GOT: ld64 supports a 4-byte pc-relative form as well as an 8-byte
48 // absolute version of this relocation. The semantics of the absolute relocation
49 // are weird -- it results in the value of the GOT slot being written, instead
50 // of the address. Let's not support it unless we find a real-world use case.
51 
52 const RelocAttrs &ARM64::getRelocAttrs(uint8_t type) const {
53   static const std::array<RelocAttrs, 11> relocAttrsArray{{
54 #define B(x) RelocAttrBits::x
55       {"UNSIGNED",
56        B(UNSIGNED) | B(ABSOLUTE) | B(EXTERN) | B(LOCAL) | B(BYTE4) | B(BYTE8)},
57       {"SUBTRACTOR", B(SUBTRAHEND) | B(EXTERN) | B(BYTE4) | B(BYTE8)},
58       {"BRANCH26", B(PCREL) | B(EXTERN) | B(BRANCH) | B(BYTE4)},
59       {"PAGE21", B(PCREL) | B(EXTERN) | B(BYTE4)},
60       {"PAGEOFF12", B(ABSOLUTE) | B(EXTERN) | B(BYTE4)},
61       {"GOT_LOAD_PAGE21", B(PCREL) | B(EXTERN) | B(GOT) | B(BYTE4)},
62       {"GOT_LOAD_PAGEOFF12",
63        B(ABSOLUTE) | B(EXTERN) | B(GOT) | B(LOAD) | B(BYTE4)},
64       {"POINTER_TO_GOT", B(PCREL) | B(EXTERN) | B(GOT) | B(POINTER) | B(BYTE4)},
65       {"TLVP_LOAD_PAGE21", B(PCREL) | B(EXTERN) | B(TLV) | B(BYTE4)},
66       {"TLVP_LOAD_PAGEOFF12",
67        B(ABSOLUTE) | B(EXTERN) | B(TLV) | B(LOAD) | B(BYTE4)},
68       {"ADDEND", B(ADDEND)},
69 #undef B
70   }};
71   assert(type < relocAttrsArray.size() && "invalid relocation type");
72   if (type >= relocAttrsArray.size())
73     return invalidRelocAttrs;
74   return relocAttrsArray[type];
75 }
76 
77 static constexpr uint32_t stubCode[] = {
78     0x90000010, // 00: adrp  x16, __la_symbol_ptr@page
79     0xf9400210, // 04: ldr   x16, [x16, __la_symbol_ptr@pageoff]
80     0xd61f0200, // 08: br    x16
81 };
82 
83 void ARM64::writeStub(uint8_t *buf8, const Symbol &sym) const {
84   ::writeStub<LP64>(buf8, stubCode, sym);
85 }
86 
87 static constexpr uint32_t stubHelperHeaderCode[] = {
88     0x90000011, // 00: adrp  x17, _dyld_private@page
89     0x91000231, // 04: add   x17, x17, _dyld_private@pageoff
90     0xa9bf47f0, // 08: stp   x16/x17, [sp, #-16]!
91     0x90000010, // 0c: adrp  x16, dyld_stub_binder@page
92     0xf9400210, // 10: ldr   x16, [x16, dyld_stub_binder@pageoff]
93     0xd61f0200, // 14: br    x16
94 };
95 
96 void ARM64::writeStubHelperHeader(uint8_t *buf8) const {
97   ::writeStubHelperHeader<LP64>(buf8, stubHelperHeaderCode);
98 }
99 
100 static constexpr uint32_t stubHelperEntryCode[] = {
101     0x18000050, // 00: ldr  w16, l0
102     0x14000000, // 04: b    stubHelperHeader
103     0x00000000, // 08: l0: .long 0
104 };
105 
106 void ARM64::writeStubHelperEntry(uint8_t *buf8, const Symbol &sym,
107                                  uint64_t entryVA) const {
108   ::writeStubHelperEntry(buf8, stubHelperEntryCode, sym, entryVA);
109 }
110 
111 // A thunk is the relaxed variation of stubCode. We don't need the
112 // extra indirection through a lazy pointer because the target address
113 // is known at link time.
114 static constexpr uint32_t thunkCode[] = {
115     0x90000010, // 00: adrp  x16, <thunk.ptr>@page
116     0x91000210, // 04: add   x16, [x16,<thunk.ptr>@pageoff]
117     0xd61f0200, // 08: br    x16
118 };
119 
120 void ARM64::populateThunk(InputSection *thunk, Symbol *funcSym) {
121   thunk->align = 4;
122   thunk->data = {reinterpret_cast<const uint8_t *>(thunkCode),
123                  sizeof(thunkCode)};
124   thunk->relocs.push_back({/*type=*/ARM64_RELOC_PAGEOFF12,
125                            /*pcrel=*/false, /*length=*/2,
126                            /*offset=*/4, /*addend=*/0,
127                            /*referent=*/funcSym});
128   thunk->relocs.push_back({/*type=*/ARM64_RELOC_PAGE21,
129                            /*pcrel=*/true, /*length=*/2,
130                            /*offset=*/0, /*addend=*/0,
131                            /*referent=*/funcSym});
132 }
133 
134 ARM64::ARM64() : ARM64Common(LP64()) {
135   cpuType = CPU_TYPE_ARM64;
136   cpuSubtype = CPU_SUBTYPE_ARM64_ALL;
137 
138   stubSize = sizeof(stubCode);
139   thunkSize = sizeof(thunkCode);
140 
141   // Branch immediate is two's complement 26 bits, which is implicitly
142   // multiplied by 4 (since all functions are 4-aligned: The branch range
143   // is -4*(2**(26-1))..4*(2**(26-1) - 1).
144   backwardBranchRange = 128 * 1024 * 1024;
145   forwardBranchRange = backwardBranchRange - 4;
146 
147   modeDwarfEncoding = UNWIND_ARM64_MODE_DWARF;
148   subtractorRelocType = ARM64_RELOC_SUBTRACTOR;
149   unsignedRelocType = ARM64_RELOC_UNSIGNED;
150 
151   stubHelperHeaderSize = sizeof(stubHelperHeaderCode);
152   stubHelperEntrySize = sizeof(stubHelperEntryCode);
153 }
154 
155 namespace {
156 struct Adrp {
157   uint32_t destRegister;
158 };
159 
160 struct Add {
161   uint8_t destRegister;
162   uint8_t srcRegister;
163   uint32_t addend;
164 };
165 
166 enum ExtendType { ZeroExtend = 1, Sign64 = 2, Sign32 = 3 };
167 
168 struct Ldr {
169   uint8_t destRegister;
170   uint8_t baseRegister;
171   uint8_t size;
172   bool isFloat;
173   ExtendType extendType;
174   uint64_t offset;
175 };
176 
177 struct PerformedReloc {
178   const Reloc &rel;
179   uint64_t referentVA;
180 };
181 
182 class OptimizationHintContext {
183 public:
184   OptimizationHintContext(uint8_t *buf, const ConcatInputSection *isec,
185                           ArrayRef<uint64_t> relocTargets)
186       : buf(buf), isec(isec), relocTargets(relocTargets),
187         relocIt(isec->relocs.rbegin()) {}
188 
189   void applyAdrpAdd(const OptimizationHint &);
190   void applyAdrpAdrp(const OptimizationHint &);
191   void applyAdrpLdr(const OptimizationHint &);
192   void applyAdrpLdrGot(const OptimizationHint &);
193 
194 private:
195   uint8_t *buf;
196   const ConcatInputSection *isec;
197   ArrayRef<uint64_t> relocTargets;
198   std::vector<Reloc>::const_reverse_iterator relocIt;
199 
200   uint64_t getRelocTarget(const Reloc &);
201 
202   Optional<PerformedReloc> findPrimaryReloc(uint64_t offset);
203   Optional<PerformedReloc> findReloc(uint64_t offset);
204 };
205 } // namespace
206 
207 static bool parseAdrp(uint32_t insn, Adrp &adrp) {
208   if ((insn & 0x9f000000) != 0x90000000)
209     return false;
210   adrp.destRegister = insn & 0x1f;
211   return true;
212 }
213 
214 static bool parseAdd(uint32_t insn, Add &add) {
215   if ((insn & 0xffc00000) != 0x91000000)
216     return false;
217   add.destRegister = insn & 0x1f;
218   add.srcRegister = (insn >> 5) & 0x1f;
219   add.addend = (insn >> 10) & 0xfff;
220   return true;
221 }
222 
223 static bool parseLdr(uint32_t insn, Ldr &ldr) {
224   ldr.destRegister = insn & 0x1f;
225   ldr.baseRegister = (insn >> 5) & 0x1f;
226   uint8_t size = insn >> 30;
227   uint8_t opc = (insn >> 22) & 3;
228 
229   if ((insn & 0x3fc00000) == 0x39400000) {
230     // LDR (immediate), LDRB (immediate), LDRH (immediate)
231     ldr.size = 1 << size;
232     ldr.extendType = ZeroExtend;
233     ldr.isFloat = false;
234   } else if ((insn & 0x3f800000) == 0x39800000) {
235     // LDRSB (immediate), LDRSH (immediate), LDRSW (immediate)
236     ldr.size = 1 << size;
237     ldr.extendType = static_cast<ExtendType>(opc);
238     ldr.isFloat = false;
239   } else if ((insn & 0x3f400000) == 0x3d400000) {
240     // LDR (immediate, SIMD&FP)
241     ldr.extendType = ZeroExtend;
242     ldr.isFloat = true;
243     if (size == 2 && opc == 1)
244       ldr.size = 4;
245     else if (size == 3 && opc == 1)
246       ldr.size = 8;
247     else if (size == 0 && opc == 3)
248       ldr.size = 16;
249     else
250       return false;
251   } else {
252     return false;
253   }
254   ldr.offset = ((insn >> 10) & 0xfff) * ldr.size;
255   return true;
256 }
257 
258 static void writeAdr(void *loc, uint32_t dest, int32_t delta) {
259   uint32_t opcode = 0x10000000;
260   uint32_t immHi = (delta & 0x001ffffc) << 3;
261   uint32_t immLo = (delta & 0x00000003) << 29;
262   write32le(loc, opcode | immHi | immLo | dest);
263 }
264 
265 static void writeNop(void *loc) { write32le(loc, 0xd503201f); }
266 
267 static void writeLiteralLdr(void *loc, Ldr original, int32_t delta) {
268   uint32_t imm19 = (delta & 0x001ffffc) << 3;
269   uint32_t opcode = 0;
270   switch (original.size) {
271   case 4:
272     if (original.isFloat)
273       opcode = 0x1c000000;
274     else
275       opcode = original.extendType == Sign64 ? 0x98000000 : 0x18000000;
276     break;
277   case 8:
278     opcode = original.isFloat ? 0x5c000000 : 0x58000000;
279     break;
280   case 16:
281     opcode = 0x9c000000;
282     break;
283   default:
284     assert(false && "Invalid size for literal ldr");
285   }
286   write32le(loc, opcode | imm19 | original.destRegister);
287 }
288 
289 uint64_t OptimizationHintContext::getRelocTarget(const Reloc &reloc) {
290   size_t relocIdx = &reloc - isec->relocs.data();
291   return relocTargets[relocIdx];
292 }
293 
294 // Optimization hints are sorted in a monotonically increasing order by their
295 // first address as are relocations (albeit in decreasing order), so if we keep
296 // a pointer around to the last found relocation, we don't have to do a full
297 // binary search every time.
298 Optional<PerformedReloc>
299 OptimizationHintContext::findPrimaryReloc(uint64_t offset) {
300   const auto end = isec->relocs.rend();
301   while (relocIt != end && relocIt->offset < offset)
302     ++relocIt;
303   if (relocIt == end || relocIt->offset != offset)
304     return None;
305   return PerformedReloc{*relocIt, getRelocTarget(*relocIt)};
306 }
307 
308 // The second and third addresses of optimization hints have no such
309 // monotonicity as the first, so we search the entire range of relocations.
310 Optional<PerformedReloc> OptimizationHintContext::findReloc(uint64_t offset) {
311   // Optimization hints often apply to successive relocations, so we check for
312   // that first before doing a full binary search.
313   auto end = isec->relocs.rend();
314   if (relocIt < end - 1 && (relocIt + 1)->offset == offset)
315     return PerformedReloc{*(relocIt + 1), getRelocTarget(*(relocIt + 1))};
316 
317   auto reloc = lower_bound(isec->relocs, offset,
318                            [](const Reloc &reloc, uint64_t offset) {
319                              return offset < reloc.offset;
320                            });
321 
322   if (reloc == isec->relocs.end() || reloc->offset != offset)
323     return None;
324   return PerformedReloc{*reloc, getRelocTarget(*reloc)};
325 }
326 
327 // Transforms a pair of adrp+add instructions into an adr instruction if the
328 // target is within the +/- 1 MiB range allowed by the adr's 21 bit signed
329 // immediate offset.
330 //
331 //   adrp xN, _foo@PAGE
332 //   add  xM, xN, _foo@PAGEOFF
333 // ->
334 //   adr  xM, _foo
335 //   nop
336 void OptimizationHintContext::applyAdrpAdd(const OptimizationHint &hint) {
337   uint32_t ins1 = read32le(buf + hint.offset0);
338   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
339   Adrp adrp;
340   if (!parseAdrp(ins1, adrp))
341     return;
342   Add add;
343   if (!parseAdd(ins2, add))
344     return;
345   if (adrp.destRegister != add.srcRegister)
346     return;
347 
348   Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
349   Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
350   if (!rel1 || !rel2)
351     return;
352   if (rel1->referentVA != rel2->referentVA)
353     return;
354   int64_t delta = rel1->referentVA - rel1->rel.offset - isec->getVA();
355   if (delta >= (1 << 20) || delta < -(1 << 20))
356     return;
357 
358   writeAdr(buf + hint.offset0, add.destRegister, delta);
359   writeNop(buf + hint.offset0 + hint.delta[0]);
360 }
361 
362 // Transforms two adrp instructions into a single adrp if their referent
363 // addresses are located on the same 4096 byte page.
364 //
365 //   adrp xN, _foo@PAGE
366 //   adrp xN, _bar@PAGE
367 // ->
368 //   adrp xN, _foo@PAGE
369 //   nop
370 void OptimizationHintContext::applyAdrpAdrp(const OptimizationHint &hint) {
371   uint32_t ins1 = read32le(buf + hint.offset0);
372   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
373   Adrp adrp1, adrp2;
374   if (!parseAdrp(ins1, adrp1) || !parseAdrp(ins2, adrp2))
375     return;
376   if (adrp1.destRegister != adrp2.destRegister)
377     return;
378 
379   Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
380   Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
381   if (!rel1 || !rel2)
382     return;
383   if ((rel1->referentVA & ~0xfffULL) != (rel2->referentVA & ~0xfffULL))
384     return;
385 
386   writeNop(buf + hint.offset0 + hint.delta[0]);
387 }
388 
389 // Transforms a pair of adrp+ldr (immediate) instructions into an ldr (literal)
390 // load from a PC-relative address if it is 4-byte aligned and within +/- 1 MiB,
391 // as ldr can encode a signed 19-bit offset that gets multiplied by 4.
392 //
393 //   adrp xN, _foo@PAGE
394 //   ldr  xM, [xN, _foo@PAGEOFF]
395 // ->
396 //   nop
397 //   ldr  xM, _foo
398 void OptimizationHintContext::applyAdrpLdr(const OptimizationHint &hint) {
399   uint32_t ins1 = read32le(buf + hint.offset0);
400   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
401   Adrp adrp;
402   if (!parseAdrp(ins1, adrp))
403     return;
404   Ldr ldr;
405   if (!parseLdr(ins2, ldr))
406     return;
407   if (adrp.destRegister != ldr.baseRegister)
408     return;
409 
410   Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
411   Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
412   if (!rel1 || !rel2)
413     return;
414   if (ldr.offset != (rel1->referentVA & 0xfff))
415     return;
416   if ((rel1->referentVA & 3) != 0)
417     return;
418   if (ldr.size == 1 || ldr.size == 2)
419     return;
420   int64_t delta = rel1->referentVA - rel2->rel.offset - isec->getVA();
421   if (delta >= (1 << 20) || delta < -(1 << 20))
422     return;
423 
424   writeNop(buf + hint.offset0);
425   writeLiteralLdr(buf + hint.offset0 + hint.delta[0], ldr, delta);
426 }
427 
428 // GOT loads are emitted by the compiler as a pair of adrp and ldr instructions,
429 // but they may be changed to adrp+add by relaxGotLoad(). This hint performs
430 // the AdrpLdr or AdrpAdd transformation depending on whether it was relaxed.
431 void OptimizationHintContext::applyAdrpLdrGot(const OptimizationHint &hint) {
432   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
433   Add add;
434   Ldr ldr;
435   if (parseAdd(ins2, add))
436     applyAdrpAdd(hint);
437   else if (parseLdr(ins2, ldr))
438     applyAdrpLdr(hint);
439 }
440 
441 void ARM64::applyOptimizationHints(uint8_t *buf, const ConcatInputSection *isec,
442                                    ArrayRef<uint64_t> relocTargets) const {
443   assert(isec);
444   assert(relocTargets.size() == isec->relocs.size());
445 
446   // Note: Some of these optimizations might not be valid when shared regions
447   // are in use. Will need to revisit this if splitSegInfo is added.
448 
449   OptimizationHintContext ctx1(buf, isec, relocTargets);
450   for (const OptimizationHint &hint : isec->optimizationHints) {
451     switch (hint.type) {
452     case LOH_ARM64_ADRP_ADRP:
453       // This is done in another pass because the other optimization hints
454       // might cause its targets to be turned into NOPs.
455       break;
456     case LOH_ARM64_ADRP_LDR:
457       ctx1.applyAdrpLdr(hint);
458       break;
459     case LOH_ARM64_ADRP_ADD_LDR:
460     case LOH_ARM64_ADRP_LDR_GOT_LDR:
461     case LOH_ARM64_ADRP_ADD_STR:
462     case LOH_ARM64_ADRP_LDR_GOT_STR:
463       // TODO: Implement these
464       break;
465     case LOH_ARM64_ADRP_ADD:
466       ctx1.applyAdrpAdd(hint);
467       break;
468     case LOH_ARM64_ADRP_LDR_GOT:
469       ctx1.applyAdrpLdrGot(hint);
470       break;
471     }
472   }
473 
474   OptimizationHintContext ctx2(buf, isec, relocTargets);
475   for (const OptimizationHint &hint : isec->optimizationHints)
476     if (hint.type == LOH_ARM64_ADRP_ADRP)
477       ctx2.applyAdrpAdrp(hint);
478 }
479 
480 TargetInfo *macho::createARM64TargetInfo() {
481   static ARM64 t;
482   return &t;
483 }
484