xref: /freebsd/contrib/llvm-project/lld/MachO/ICF.cpp (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
1 //===- ICF.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 "ICF.h"
10 #include "ConcatOutputSection.h"
11 #include "Config.h"
12 #include "InputSection.h"
13 #include "SymbolTable.h"
14 #include "Symbols.h"
15 #include "UnwindInfoSection.h"
16 
17 #include "lld/Common/CommonLinkerContext.h"
18 #include "llvm/Support/LEB128.h"
19 #include "llvm/Support/Parallel.h"
20 #include "llvm/Support/TimeProfiler.h"
21 #include "llvm/Support/xxhash.h"
22 
23 #include <atomic>
24 
25 using namespace llvm;
26 using namespace lld;
27 using namespace lld::macho;
28 
29 static constexpr bool verboseDiagnostics = false;
30 
31 class ICF {
32 public:
33   ICF(std::vector<ConcatInputSection *> &inputs);
34   void run();
35 
36   using EqualsFn = bool (ICF::*)(const ConcatInputSection *,
37                                  const ConcatInputSection *);
38   void segregate(size_t begin, size_t end, EqualsFn);
39   size_t findBoundary(size_t begin, size_t end);
40   void forEachClassRange(size_t begin, size_t end,
41                          llvm::function_ref<void(size_t, size_t)> func);
42   void forEachClass(llvm::function_ref<void(size_t, size_t)> func);
43 
44   bool equalsConstant(const ConcatInputSection *ia,
45                       const ConcatInputSection *ib);
46   bool equalsVariable(const ConcatInputSection *ia,
47                       const ConcatInputSection *ib);
48 
49   // ICF needs a copy of the inputs vector because its equivalence-class
50   // segregation algorithm destroys the proper sequence.
51   std::vector<ConcatInputSection *> icfInputs;
52 
53   unsigned icfPass = 0;
54   std::atomic<bool> icfRepeat{false};
55   std::atomic<uint64_t> equalsConstantCount{0};
56   std::atomic<uint64_t> equalsVariableCount{0};
57 };
58 
59 ICF::ICF(std::vector<ConcatInputSection *> &inputs) {
60   icfInputs.assign(inputs.begin(), inputs.end());
61 }
62 
63 // ICF = Identical Code Folding
64 //
65 // We only fold __TEXT,__text, so this is really "code" folding, and not
66 // "COMDAT" folding. String and scalar constant literals are deduplicated
67 // elsewhere.
68 //
69 // Summary of segments & sections:
70 //
71 // The __TEXT segment is readonly at the MMU. Some sections are already
72 // deduplicated elsewhere (__TEXT,__cstring & __TEXT,__literal*) and some are
73 // synthetic and inherently free of duplicates (__TEXT,__stubs &
74 // __TEXT,__unwind_info). Note that we don't yet run ICF on __TEXT,__const,
75 // because doing so induces many test failures.
76 //
77 // The __LINKEDIT segment is readonly at the MMU, yet entirely synthetic, and
78 // thus ineligible for ICF.
79 //
80 // The __DATA_CONST segment is read/write at the MMU, but is logically const to
81 // the application after dyld applies fixups to pointer data. We currently
82 // fold only the __DATA_CONST,__cfstring section.
83 //
84 // The __DATA segment is read/write at the MMU, and as application-writeable
85 // data, none of its sections are eligible for ICF.
86 //
87 // Please see the large block comment in lld/ELF/ICF.cpp for an explanation
88 // of the segregation algorithm.
89 //
90 // FIXME(gkm): implement keep-unique attributes
91 // FIXME(gkm): implement address-significance tables for MachO object files
92 
93 // Compare "non-moving" parts of two ConcatInputSections, namely everything
94 // except references to other ConcatInputSections.
95 bool ICF::equalsConstant(const ConcatInputSection *ia,
96                          const ConcatInputSection *ib) {
97   if (verboseDiagnostics)
98     ++equalsConstantCount;
99   // We can only fold within the same OutputSection.
100   if (ia->parent != ib->parent)
101     return false;
102   if (ia->data.size() != ib->data.size())
103     return false;
104   if (ia->data != ib->data)
105     return false;
106   if (ia->relocs.size() != ib->relocs.size())
107     return false;
108   auto f = [](const Reloc &ra, const Reloc &rb) {
109     if (ra.type != rb.type)
110       return false;
111     if (ra.pcrel != rb.pcrel)
112       return false;
113     if (ra.length != rb.length)
114       return false;
115     if (ra.offset != rb.offset)
116       return false;
117     if (ra.referent.is<Symbol *>() != rb.referent.is<Symbol *>())
118       return false;
119 
120     InputSection *isecA, *isecB;
121 
122     uint64_t valueA = 0;
123     uint64_t valueB = 0;
124     if (ra.referent.is<Symbol *>()) {
125       const auto *sa = ra.referent.get<Symbol *>();
126       const auto *sb = rb.referent.get<Symbol *>();
127       if (sa->kind() != sb->kind())
128         return false;
129       // ICF runs before Undefineds are treated (and potentially converted into
130       // DylibSymbols).
131       if (isa<DylibSymbol>(sa) || isa<Undefined>(sa))
132         return sa == sb && ra.addend == rb.addend;
133       assert(isa<Defined>(sa));
134       const auto *da = cast<Defined>(sa);
135       const auto *db = cast<Defined>(sb);
136       if (!da->isec || !db->isec) {
137         assert(da->isAbsolute() && db->isAbsolute());
138         return da->value + ra.addend == db->value + rb.addend;
139       }
140       isecA = da->isec;
141       valueA = da->value;
142       isecB = db->isec;
143       valueB = db->value;
144     } else {
145       isecA = ra.referent.get<InputSection *>();
146       isecB = rb.referent.get<InputSection *>();
147     }
148 
149     if (isecA->parent != isecB->parent)
150       return false;
151     // Sections with identical parents should be of the same kind.
152     assert(isecA->kind() == isecB->kind());
153     // We will compare ConcatInputSection contents in equalsVariable.
154     if (isa<ConcatInputSection>(isecA))
155       return ra.addend == rb.addend;
156     // Else we have two literal sections. References to them are equal iff their
157     // offsets in the output section are equal.
158     if (ra.referent.is<Symbol *>())
159       // For symbol relocs, we compare the contents at the symbol address. We
160       // don't do `getOffset(value + addend)` because value + addend may not be
161       // a valid offset in the literal section.
162       return isecA->getOffset(valueA) == isecB->getOffset(valueB) &&
163              ra.addend == rb.addend;
164     else {
165       assert(valueA == 0 && valueB == 0);
166       // For section relocs, we compare the content at the section offset.
167       return isecA->getOffset(ra.addend) == isecB->getOffset(rb.addend);
168     }
169   };
170   return std::equal(ia->relocs.begin(), ia->relocs.end(), ib->relocs.begin(),
171                     f);
172 }
173 
174 // Compare the "moving" parts of two ConcatInputSections -- i.e. everything not
175 // handled by equalsConstant().
176 bool ICF::equalsVariable(const ConcatInputSection *ia,
177                          const ConcatInputSection *ib) {
178   if (verboseDiagnostics)
179     ++equalsVariableCount;
180   assert(ia->relocs.size() == ib->relocs.size());
181   auto f = [this](const Reloc &ra, const Reloc &rb) {
182     // We already filtered out mismatching values/addends in equalsConstant.
183     if (ra.referent == rb.referent)
184       return true;
185     const ConcatInputSection *isecA, *isecB;
186     if (ra.referent.is<Symbol *>()) {
187       // Matching DylibSymbols are already filtered out by the
188       // identical-referent check above. Non-matching DylibSymbols were filtered
189       // out in equalsConstant(). So we can safely cast to Defined here.
190       const auto *da = cast<Defined>(ra.referent.get<Symbol *>());
191       const auto *db = cast<Defined>(rb.referent.get<Symbol *>());
192       if (da->isAbsolute())
193         return true;
194       isecA = dyn_cast<ConcatInputSection>(da->isec);
195       if (!isecA)
196         return true; // literal sections were checked in equalsConstant.
197       isecB = cast<ConcatInputSection>(db->isec);
198     } else {
199       const auto *sa = ra.referent.get<InputSection *>();
200       const auto *sb = rb.referent.get<InputSection *>();
201       isecA = dyn_cast<ConcatInputSection>(sa);
202       if (!isecA)
203         return true;
204       isecB = cast<ConcatInputSection>(sb);
205     }
206     return isecA->icfEqClass[icfPass % 2] == isecB->icfEqClass[icfPass % 2];
207   };
208   if (!std::equal(ia->relocs.begin(), ia->relocs.end(), ib->relocs.begin(), f))
209     return false;
210 
211   // If there are symbols with associated unwind info, check that the unwind
212   // info matches. For simplicity, we only handle the case where there are only
213   // symbols at offset zero within the section (which is typically the case with
214   // .subsections_via_symbols.)
215   auto hasUnwind = [](Defined *d) { return d->unwindEntry != nullptr; };
216   auto itA = std::find_if(ia->symbols.begin(), ia->symbols.end(), hasUnwind);
217   auto itB = std::find_if(ib->symbols.begin(), ib->symbols.end(), hasUnwind);
218   if (itA == ia->symbols.end())
219     return itB == ib->symbols.end();
220   if (itB == ib->symbols.end())
221     return false;
222   const Defined *da = *itA;
223   const Defined *db = *itB;
224   if (da->unwindEntry->icfEqClass[icfPass % 2] !=
225           db->unwindEntry->icfEqClass[icfPass % 2] ||
226       da->value != 0 || db->value != 0)
227     return false;
228   auto isZero = [](Defined *d) { return d->value == 0; };
229   return std::find_if_not(std::next(itA), ia->symbols.end(), isZero) ==
230              ia->symbols.end() &&
231          std::find_if_not(std::next(itB), ib->symbols.end(), isZero) ==
232              ib->symbols.end();
233 }
234 
235 // Find the first InputSection after BEGIN whose equivalence class differs
236 size_t ICF::findBoundary(size_t begin, size_t end) {
237   uint64_t beginHash = icfInputs[begin]->icfEqClass[icfPass % 2];
238   for (size_t i = begin + 1; i < end; ++i)
239     if (beginHash != icfInputs[i]->icfEqClass[icfPass % 2])
240       return i;
241   return end;
242 }
243 
244 // Invoke FUNC on subranges with matching equivalence class
245 void ICF::forEachClassRange(size_t begin, size_t end,
246                             llvm::function_ref<void(size_t, size_t)> func) {
247   while (begin < end) {
248     size_t mid = findBoundary(begin, end);
249     func(begin, mid);
250     begin = mid;
251   }
252 }
253 
254 // Split icfInputs into shards, then parallelize invocation of FUNC on subranges
255 // with matching equivalence class
256 void ICF::forEachClass(llvm::function_ref<void(size_t, size_t)> func) {
257   // Only use threads when the benefits outweigh the overhead.
258   const size_t threadingThreshold = 1024;
259   if (icfInputs.size() < threadingThreshold) {
260     forEachClassRange(0, icfInputs.size(), func);
261     ++icfPass;
262     return;
263   }
264 
265   // Shard into non-overlapping intervals, and call FUNC in parallel.  The
266   // sharding must be completed before any calls to FUNC are made so that FUNC
267   // can modify the InputSection in its shard without causing data races.
268   const size_t shards = 256;
269   size_t step = icfInputs.size() / shards;
270   size_t boundaries[shards + 1];
271   boundaries[0] = 0;
272   boundaries[shards] = icfInputs.size();
273   parallelFor(1, shards, [&](size_t i) {
274     boundaries[i] = findBoundary((i - 1) * step, icfInputs.size());
275   });
276   parallelFor(1, shards + 1, [&](size_t i) {
277     if (boundaries[i - 1] < boundaries[i]) {
278       forEachClassRange(boundaries[i - 1], boundaries[i], func);
279     }
280   });
281   ++icfPass;
282 }
283 
284 void ICF::run() {
285   // Into each origin-section hash, combine all reloc referent section hashes.
286   for (icfPass = 0; icfPass < 2; ++icfPass) {
287     parallelForEach(icfInputs, [&](ConcatInputSection *isec) {
288       uint32_t hash = isec->icfEqClass[icfPass % 2];
289       for (const Reloc &r : isec->relocs) {
290         if (auto *sym = r.referent.dyn_cast<Symbol *>()) {
291           if (auto *defined = dyn_cast<Defined>(sym)) {
292             if (defined->isec) {
293               if (auto referentIsec =
294                       dyn_cast<ConcatInputSection>(defined->isec))
295                 hash += defined->value + referentIsec->icfEqClass[icfPass % 2];
296               else
297                 hash += defined->isec->kind() +
298                         defined->isec->getOffset(defined->value);
299             } else {
300               hash += defined->value;
301             }
302           } else {
303             // ICF runs before Undefined diags
304             assert(isa<Undefined>(sym) || isa<DylibSymbol>(sym));
305           }
306         }
307       }
308       // Set MSB to 1 to avoid collisions with non-hashed classes.
309       isec->icfEqClass[(icfPass + 1) % 2] = hash | (1ull << 31);
310     });
311   }
312 
313   llvm::stable_sort(
314       icfInputs, [](const ConcatInputSection *a, const ConcatInputSection *b) {
315         return a->icfEqClass[0] < b->icfEqClass[0];
316       });
317   forEachClass([&](size_t begin, size_t end) {
318     segregate(begin, end, &ICF::equalsConstant);
319   });
320 
321   // Split equivalence groups by comparing relocations until convergence
322   do {
323     icfRepeat = false;
324     forEachClass([&](size_t begin, size_t end) {
325       segregate(begin, end, &ICF::equalsVariable);
326     });
327   } while (icfRepeat);
328   log("ICF needed " + Twine(icfPass) + " iterations");
329   if (verboseDiagnostics) {
330     log("equalsConstant() called " + Twine(equalsConstantCount) + " times");
331     log("equalsVariable() called " + Twine(equalsVariableCount) + " times");
332   }
333 
334   // Fold sections within equivalence classes
335   forEachClass([&](size_t begin, size_t end) {
336     if (end - begin < 2)
337       return;
338     ConcatInputSection *beginIsec = icfInputs[begin];
339     for (size_t i = begin + 1; i < end; ++i)
340       beginIsec->foldIdentical(icfInputs[i]);
341   });
342 }
343 
344 // Split an equivalence class into smaller classes.
345 void ICF::segregate(size_t begin, size_t end, EqualsFn equals) {
346   while (begin < end) {
347     // Divide [begin, end) into two. Let mid be the start index of the
348     // second group.
349     auto bound = std::stable_partition(
350         icfInputs.begin() + begin + 1, icfInputs.begin() + end,
351         [&](ConcatInputSection *isec) {
352           return (this->*equals)(icfInputs[begin], isec);
353         });
354     size_t mid = bound - icfInputs.begin();
355 
356     // Split [begin, end) into [begin, mid) and [mid, end). We use mid as an
357     // equivalence class ID because every group ends with a unique index.
358     for (size_t i = begin; i < mid; ++i)
359       icfInputs[i]->icfEqClass[(icfPass + 1) % 2] = mid;
360 
361     // If we created a group, we need to iterate the main loop again.
362     if (mid != end)
363       icfRepeat = true;
364 
365     begin = mid;
366   }
367 }
368 
369 void macho::markSymAsAddrSig(Symbol *s) {
370   if (auto *d = dyn_cast_or_null<Defined>(s))
371     if (d->isec)
372       d->isec->keepUnique = true;
373 }
374 
375 void macho::markAddrSigSymbols() {
376   TimeTraceScope timeScope("Mark addrsig symbols");
377   for (InputFile *file : inputFiles) {
378     ObjFile *obj = dyn_cast<ObjFile>(file);
379     if (!obj)
380       continue;
381 
382     Section *addrSigSection = obj->addrSigSection;
383     if (!addrSigSection)
384       continue;
385     assert(addrSigSection->subsections.size() == 1);
386 
387     const InputSection *isec = addrSigSection->subsections[0].isec;
388 
389     for (const Reloc &r : isec->relocs) {
390       if (auto *sym = r.referent.dyn_cast<Symbol *>())
391         markSymAsAddrSig(sym);
392       else
393         error(toString(isec) + ": unexpected section relocation");
394     }
395   }
396 }
397 
398 void macho::foldIdenticalSections(bool onlyCfStrings) {
399   TimeTraceScope timeScope("Fold Identical Code Sections");
400   // The ICF equivalence-class segregation algorithm relies on pre-computed
401   // hashes of InputSection::data for the ConcatOutputSection::inputs and all
402   // sections referenced by their relocs. We could recursively traverse the
403   // relocs to find every referenced InputSection, but that precludes easy
404   // parallelization. Therefore, we hash every InputSection here where we have
405   // them all accessible as simple vectors.
406 
407   // If an InputSection is ineligible for ICF, we give it a unique ID to force
408   // it into an unfoldable singleton equivalence class.  Begin the unique-ID
409   // space at inputSections.size(), so that it will never intersect with
410   // equivalence-class IDs which begin at 0. Since hashes & unique IDs never
411   // coexist with equivalence-class IDs, this is not necessary, but might help
412   // someone keep the numbers straight in case we ever need to debug the
413   // ICF::segregate()
414   std::vector<ConcatInputSection *> hashable;
415   uint64_t icfUniqueID = inputSections.size();
416   for (ConcatInputSection *isec : inputSections) {
417     // FIXME: consider non-code __text sections as hashable?
418     bool isHashable =
419         (!onlyCfStrings || isCfStringSection(isec)) &&
420         (isCodeSection(isec) || isCfStringSection(isec) ||
421          isClassRefsSection(isec) || isGccExceptTabSection(isec)) &&
422         !isec->keepUnique && !isec->shouldOmitFromOutput() &&
423         sectionType(isec->getFlags()) == MachO::S_REGULAR;
424     if (isHashable) {
425       hashable.push_back(isec);
426       for (Defined *d : isec->symbols)
427         if (d->unwindEntry)
428           hashable.push_back(d->unwindEntry);
429 
430       // __cfstring has embedded addends that foil ICF's hashing / equality
431       // checks. (We can ignore embedded addends when doing ICF because the same
432       // information gets recorded in our Reloc structs.) We therefore create a
433       // mutable copy of the CFString and zero out the embedded addends before
434       // performing any hashing / equality checks.
435       if (isCfStringSection(isec) || isClassRefsSection(isec)) {
436         // We have to do this copying serially as the BumpPtrAllocator is not
437         // thread-safe. FIXME: Make a thread-safe allocator.
438         MutableArrayRef<uint8_t> copy = isec->data.copy(bAlloc());
439         for (const Reloc &r : isec->relocs)
440           target->relocateOne(copy.data() + r.offset, r, /*va=*/0,
441                               /*relocVA=*/0);
442         isec->data = copy;
443       }
444     } else if (!isEhFrameSection(isec)) {
445       // EH frames are gathered as hashables from unwindEntry above; give a
446       // unique ID to everything else.
447       isec->icfEqClass[0] = ++icfUniqueID;
448     }
449   }
450   parallelForEach(hashable, [](ConcatInputSection *isec) {
451     assert(isec->icfEqClass[0] == 0); // don't overwrite a unique ID!
452     // Turn-on the top bit to guarantee that valid hashes have no collisions
453     // with the small-integer unique IDs for ICF-ineligible sections
454     isec->icfEqClass[0] = xxHash64(isec->data) | (1ull << 31);
455   });
456   // Now that every input section is either hashed or marked as unique, run the
457   // segregation algorithm to detect foldable subsections.
458   ICF(hashable).run();
459 }
460