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 struct PerformedReloc {
167   const Reloc &rel;
168   uint64_t referentVA;
169 };
170 
171 class OptimizationHintContext {
172 public:
173   OptimizationHintContext(uint8_t *buf, const ConcatInputSection *isec,
174                           ArrayRef<uint64_t> relocTargets)
175       : buf(buf), isec(isec), relocTargets(relocTargets),
176         relocIt(isec->relocs.rbegin()) {}
177 
178   void applyAdrpAdd(const OptimizationHint &);
179   void applyAdrpAdrp(const OptimizationHint &);
180 
181 private:
182   uint8_t *buf;
183   const ConcatInputSection *isec;
184   ArrayRef<uint64_t> relocTargets;
185   std::vector<Reloc>::const_reverse_iterator relocIt;
186 
187   uint64_t getRelocTarget(const Reloc &);
188 
189   Optional<PerformedReloc> findPrimaryReloc(uint64_t offset);
190   Optional<PerformedReloc> findReloc(uint64_t offset);
191 };
192 } // namespace
193 
194 static bool parseAdrp(uint32_t insn, Adrp &adrp) {
195   if ((insn & 0x9f000000) != 0x90000000)
196     return false;
197   adrp.destRegister = insn & 0x1f;
198   return true;
199 }
200 
201 static bool parseAdd(uint32_t insn, Add &add) {
202   if ((insn & 0xffc00000) != 0x91000000)
203     return false;
204   add.destRegister = insn & 0x1f;
205   add.srcRegister = (insn >> 5) & 0x1f;
206   add.addend = (insn >> 10) & 0xfff;
207   return true;
208 }
209 
210 static void writeAdr(void *loc, uint32_t dest, int32_t delta) {
211   uint32_t opcode = 0x10000000;
212   uint32_t immHi = (delta & 0x001ffffc) << 3;
213   uint32_t immLo = (delta & 0x00000003) << 29;
214   write32le(loc, opcode | immHi | immLo | dest);
215 }
216 
217 static void writeNop(void *loc) { write32le(loc, 0xd503201f); }
218 
219 uint64_t OptimizationHintContext::getRelocTarget(const Reloc &reloc) {
220   size_t relocIdx = &reloc - isec->relocs.data();
221   return relocTargets[relocIdx];
222 }
223 
224 // Optimization hints are sorted in a monotonically increasing order by their
225 // first address as are relocations (albeit in decreasing order), so if we keep
226 // a pointer around to the last found relocation, we don't have to do a full
227 // binary search every time.
228 Optional<PerformedReloc>
229 OptimizationHintContext::findPrimaryReloc(uint64_t offset) {
230   const auto end = isec->relocs.rend();
231   while (relocIt != end && relocIt->offset < offset)
232     ++relocIt;
233   if (relocIt == end || relocIt->offset != offset)
234     return None;
235   return PerformedReloc{*relocIt, getRelocTarget(*relocIt)};
236 }
237 
238 // The second and third addresses of optimization hints have no such
239 // monotonicity as the first, so we search the entire range of relocations.
240 Optional<PerformedReloc> OptimizationHintContext::findReloc(uint64_t offset) {
241   // Optimization hints often apply to successive relocations, so we check for
242   // that first before doing a full binary search.
243   auto end = isec->relocs.rend();
244   if (relocIt < end - 1 && (relocIt + 1)->offset == offset)
245     return PerformedReloc{*(relocIt + 1), getRelocTarget(*(relocIt + 1))};
246 
247   auto reloc = lower_bound(isec->relocs, offset,
248                            [](const Reloc &reloc, uint64_t offset) {
249                              return offset < reloc.offset;
250                            });
251 
252   if (reloc == isec->relocs.end() || reloc->offset != offset)
253     return None;
254   return PerformedReloc{*reloc, getRelocTarget(*reloc)};
255 }
256 
257 // Transforms a pair of adrp+add instructions into an adr instruction if the
258 // target is within the +/- 1 MiB range allowed by the adr's 21 bit signed
259 // immediate offset.
260 //
261 //   adrp xN, _foo@PAGE
262 //   add  xM, xN, _foo@PAGEOFF
263 // ->
264 //   adr  xM, _foo
265 //   nop
266 void OptimizationHintContext::applyAdrpAdd(const OptimizationHint &hint) {
267   uint32_t ins1 = read32le(buf + hint.offset0);
268   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
269   Adrp adrp;
270   if (!parseAdrp(ins1, adrp))
271     return;
272   Add add;
273   if (!parseAdd(ins2, add))
274     return;
275   if (adrp.destRegister != add.srcRegister)
276     return;
277 
278   Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
279   Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
280   if (!rel1 || !rel2)
281     return;
282   if (rel1->referentVA != rel2->referentVA)
283     return;
284   int64_t delta = rel1->referentVA - rel1->rel.offset - isec->getVA();
285   if (delta >= (1 << 20) || delta < -(1 << 20))
286     return;
287 
288   writeAdr(buf + hint.offset0, add.destRegister, delta);
289   writeNop(buf + hint.offset0 + hint.delta[0]);
290 }
291 
292 // Transforms two adrp instructions into a single adrp if their referent
293 // addresses are located on the same 4096 byte page.
294 //
295 //   adrp xN, _foo@PAGE
296 //   adrp xN, _bar@PAGE
297 // ->
298 //   adrp xN, _foo@PAGE
299 //   nop
300 void OptimizationHintContext::applyAdrpAdrp(const OptimizationHint &hint) {
301   uint32_t ins1 = read32le(buf + hint.offset0);
302   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
303   Adrp adrp1, adrp2;
304   if (!parseAdrp(ins1, adrp1) || !parseAdrp(ins2, adrp2))
305     return;
306   if (adrp1.destRegister != adrp2.destRegister)
307     return;
308 
309   Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
310   Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
311   if (!rel1 || !rel2)
312     return;
313   if ((rel1->referentVA & ~0xfffULL) != (rel2->referentVA & ~0xfffULL))
314     return;
315 
316   writeNop(buf + hint.offset0 + hint.delta[0]);
317 }
318 
319 void ARM64::applyOptimizationHints(uint8_t *buf, const ConcatInputSection *isec,
320                                    ArrayRef<uint64_t> relocTargets) const {
321   assert(isec);
322   assert(relocTargets.size() == isec->relocs.size());
323 
324   // Note: Some of these optimizations might not be valid when shared regions
325   // are in use. Will need to revisit this if splitSegInfo is added.
326 
327   OptimizationHintContext ctx1(buf, isec, relocTargets);
328   for (const OptimizationHint &hint : isec->optimizationHints) {
329     switch (hint.type) {
330     case LOH_ARM64_ADRP_ADRP:
331       // This is done in another pass because the other optimization hints
332       // might cause its targets to be turned into NOPs.
333       break;
334     case LOH_ARM64_ADRP_LDR:
335     case LOH_ARM64_ADRP_ADD_LDR:
336     case LOH_ARM64_ADRP_LDR_GOT_LDR:
337     case LOH_ARM64_ADRP_ADD_STR:
338     case LOH_ARM64_ADRP_LDR_GOT_STR:
339       // TODO: Implement these
340       break;
341     case LOH_ARM64_ADRP_ADD:
342       ctx1.applyAdrpAdd(hint);
343       break;
344     case LOH_ARM64_ADRP_LDR_GOT:
345       // TODO: Implement this as well
346       break;
347     }
348   }
349 
350   OptimizationHintContext ctx2(buf, isec, relocTargets);
351   for (const OptimizationHint &hint : isec->optimizationHints)
352     if (hint.type == LOH_ARM64_ADRP_ADRP)
353       ctx2.applyAdrpAdrp(hint);
354 }
355 
356 TargetInfo *macho::createARM64TargetInfo() {
357   static ARM64 t;
358   return &t;
359 }
360