1 //===- verify-uselistorder.cpp - The LLVM Modular Optimizer ---------------===//
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 // Verify that use-list order can be serialized correctly. After reading the
10 // provided IR, this tool shuffles the use-lists and then writes and reads to a
11 // separate Module whose use-list orders are compared to the original.
12 //
13 // The shuffles are deterministic, but guarantee that use-lists will change.
14 // The algorithm per iteration is as follows:
15 //
16 // 1. Seed the random number generator. The seed is different for each
17 // shuffle. Shuffle 0 uses default+0, shuffle 1 uses default+1, and so on.
18 //
19 // 2. Visit every Value in a deterministic order.
20 //
21 // 3. Assign a random number to each Use in the Value's use-list in order.
22 //
23 // 4. If the numbers are already in order, reassign numbers until they aren't.
24 //
25 // 5. Sort the use-list using Value::sortUseList(), which is a stable sort.
26 //
27 //===----------------------------------------------------------------------===//
28
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/DenseSet.h"
31 #include "llvm/AsmParser/Parser.h"
32 #include "llvm/Bitcode/BitcodeReader.h"
33 #include "llvm/Bitcode/BitcodeWriter.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/UseListOrder.h"
37 #include "llvm/IR/Verifier.h"
38 #include "llvm/IRReader/IRReader.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/FileSystem.h"
43 #include "llvm/Support/FileUtilities.h"
44 #include "llvm/Support/InitLLVM.h"
45 #include "llvm/Support/MemoryBuffer.h"
46 #include "llvm/Support/SourceMgr.h"
47 #include "llvm/Support/SystemUtils.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include <random>
50 #include <vector>
51
52 using namespace llvm;
53
54 #define DEBUG_TYPE "uselistorder"
55
56 static cl::opt<std::string> InputFilename(cl::Positional,
57 cl::desc("<input bitcode file>"),
58 cl::init("-"),
59 cl::value_desc("filename"));
60
61 static cl::opt<bool> SaveTemps("save-temps", cl::desc("Save temp files"),
62 cl::init(false));
63
64 static cl::opt<unsigned>
65 NumShuffles("num-shuffles",
66 cl::desc("Number of times to shuffle and verify use-lists"),
67 cl::init(1));
68
69 namespace {
70
71 struct TempFile {
72 std::string Filename;
73 FileRemover Remover;
74 bool init(const std::string &Ext);
75 bool writeBitcode(const Module &M) const;
76 bool writeAssembly(const Module &M) const;
77 std::unique_ptr<Module> readBitcode(LLVMContext &Context) const;
78 std::unique_ptr<Module> readAssembly(LLVMContext &Context) const;
79 };
80
81 struct ValueMapping {
82 DenseMap<const Value *, unsigned> IDs;
83 std::vector<const Value *> Values;
84
85 /// Construct a value mapping for module.
86 ///
87 /// Creates mapping from every value in \c M to an ID. This mapping includes
88 /// un-referencable values.
89 ///
90 /// Every \a Value that gets serialized in some way should be represented
91 /// here. The order needs to be deterministic, but it's unnecessary to match
92 /// the value-ids in the bitcode writer.
93 ///
94 /// All constants that are referenced by other values are included in the
95 /// mapping, but others -- which wouldn't be serialized -- are not.
96 ValueMapping(const Module &M);
97
98 /// Map a value.
99 ///
100 /// Maps a value. If it's a constant, maps all of its operands first.
101 void map(const Value *V);
lookup__anon112481120111::ValueMapping102 unsigned lookup(const Value *V) const { return IDs.lookup(V); }
103 };
104
105 } // end namespace
106
init(const std::string & Ext)107 bool TempFile::init(const std::string &Ext) {
108 SmallVector<char, 64> Vector;
109 LLVM_DEBUG(dbgs() << " - create-temp-file\n");
110 if (auto EC = sys::fs::createTemporaryFile("uselistorder", Ext, Vector)) {
111 errs() << "verify-uselistorder: error: " << EC.message() << "\n";
112 return true;
113 }
114 assert(!Vector.empty());
115
116 Filename.assign(Vector.data(), Vector.data() + Vector.size());
117 Remover.setFile(Filename, !SaveTemps);
118 if (SaveTemps)
119 outs() << " - filename = " << Filename << "\n";
120 return false;
121 }
122
writeBitcode(const Module & M) const123 bool TempFile::writeBitcode(const Module &M) const {
124 LLVM_DEBUG(dbgs() << " - write bitcode\n");
125 std::error_code EC;
126 raw_fd_ostream OS(Filename, EC, sys::fs::OF_None);
127 if (EC) {
128 errs() << "verify-uselistorder: error: " << EC.message() << "\n";
129 return true;
130 }
131
132 WriteBitcodeToFile(M, OS, /* ShouldPreserveUseListOrder */ true);
133 return false;
134 }
135
writeAssembly(const Module & M) const136 bool TempFile::writeAssembly(const Module &M) const {
137 LLVM_DEBUG(dbgs() << " - write assembly\n");
138 std::error_code EC;
139 raw_fd_ostream OS(Filename, EC, sys::fs::OF_TextWithCRLF);
140 if (EC) {
141 errs() << "verify-uselistorder: error: " << EC.message() << "\n";
142 return true;
143 }
144
145 M.print(OS, nullptr, /* ShouldPreserveUseListOrder */ true);
146 return false;
147 }
148
readBitcode(LLVMContext & Context) const149 std::unique_ptr<Module> TempFile::readBitcode(LLVMContext &Context) const {
150 LLVM_DEBUG(dbgs() << " - read bitcode\n");
151 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOr =
152 MemoryBuffer::getFile(Filename);
153 if (!BufferOr) {
154 errs() << "verify-uselistorder: error: " << BufferOr.getError().message()
155 << "\n";
156 return nullptr;
157 }
158
159 MemoryBuffer *Buffer = BufferOr.get().get();
160 Expected<std::unique_ptr<Module>> ModuleOr =
161 parseBitcodeFile(Buffer->getMemBufferRef(), Context);
162 if (!ModuleOr) {
163 logAllUnhandledErrors(ModuleOr.takeError(), errs(),
164 "verify-uselistorder: error: ");
165 return nullptr;
166 }
167 return std::move(ModuleOr.get());
168 }
169
readAssembly(LLVMContext & Context) const170 std::unique_ptr<Module> TempFile::readAssembly(LLVMContext &Context) const {
171 LLVM_DEBUG(dbgs() << " - read assembly\n");
172 SMDiagnostic Err;
173 std::unique_ptr<Module> M = parseAssemblyFile(Filename, Err, Context);
174 if (!M.get())
175 Err.print("verify-uselistorder", errs());
176 return M;
177 }
178
ValueMapping(const Module & M)179 ValueMapping::ValueMapping(const Module &M) {
180 // Every value should be mapped, including things like void instructions and
181 // basic blocks that are kept out of the ValueEnumerator.
182 //
183 // The current mapping order makes it easier to debug the tables. It happens
184 // to be similar to the ID mapping when writing ValueEnumerator, but they
185 // aren't (and needn't be) in sync.
186
187 // Globals.
188 for (const GlobalVariable &G : M.globals())
189 map(&G);
190 for (const GlobalAlias &A : M.aliases())
191 map(&A);
192 for (const GlobalIFunc &IF : M.ifuncs())
193 map(&IF);
194 for (const Function &F : M)
195 map(&F);
196
197 // Constants used by globals.
198 for (const GlobalVariable &G : M.globals())
199 if (G.hasInitializer())
200 map(G.getInitializer());
201 for (const GlobalAlias &A : M.aliases())
202 map(A.getAliasee());
203 for (const GlobalIFunc &IF : M.ifuncs())
204 map(IF.getResolver());
205 for (const Function &F : M) {
206 if (F.hasPrefixData())
207 map(F.getPrefixData());
208 if (F.hasPrologueData())
209 map(F.getPrologueData());
210 if (F.hasPersonalityFn())
211 map(F.getPersonalityFn());
212 }
213
214 // Function bodies.
215 for (const Function &F : M) {
216 for (const Argument &A : F.args())
217 map(&A);
218 for (const BasicBlock &BB : F)
219 map(&BB);
220 for (const BasicBlock &BB : F)
221 for (const Instruction &I : BB)
222 map(&I);
223
224 // Constants used by instructions.
225 for (const BasicBlock &BB : F)
226 for (const Instruction &I : BB)
227 for (const Value *Op : I.operands()) {
228 // Look through a metadata wrapper.
229 if (const auto *MAV = dyn_cast<MetadataAsValue>(Op))
230 if (const auto *VAM = dyn_cast<ValueAsMetadata>(MAV->getMetadata()))
231 Op = VAM->getValue();
232
233 if ((isa<Constant>(Op) && !isa<GlobalValue>(*Op)) ||
234 isa<InlineAsm>(Op))
235 map(Op);
236 }
237 }
238 }
239
map(const Value * V)240 void ValueMapping::map(const Value *V) {
241 if (IDs.lookup(V))
242 return;
243
244 if (auto *C = dyn_cast<Constant>(V))
245 if (!isa<GlobalValue>(C))
246 for (const Value *Op : C->operands())
247 map(Op);
248
249 Values.push_back(V);
250 IDs[V] = Values.size();
251 }
252
253 #ifndef NDEBUG
dumpMapping(const ValueMapping & VM)254 static void dumpMapping(const ValueMapping &VM) {
255 dbgs() << "value-mapping (size = " << VM.Values.size() << "):\n";
256 for (unsigned I = 0, E = VM.Values.size(); I != E; ++I) {
257 dbgs() << " - id = " << I << ", value = ";
258 VM.Values[I]->dump();
259 }
260 }
261
debugValue(const ValueMapping & M,unsigned I,StringRef Desc)262 static void debugValue(const ValueMapping &M, unsigned I, StringRef Desc) {
263 const Value *V = M.Values[I];
264 dbgs() << " - " << Desc << " value = ";
265 V->dump();
266 for (const Use &U : V->uses()) {
267 dbgs() << " => use: op = " << U.getOperandNo()
268 << ", user-id = " << M.IDs.lookup(U.getUser()) << ", user = ";
269 U.getUser()->dump();
270 }
271 }
272
debugUserMismatch(const ValueMapping & L,const ValueMapping & R,unsigned I)273 static void debugUserMismatch(const ValueMapping &L, const ValueMapping &R,
274 unsigned I) {
275 dbgs() << " - fail: user mismatch: ID = " << I << "\n";
276 debugValue(L, I, "LHS");
277 debugValue(R, I, "RHS");
278
279 dbgs() << "\nlhs-";
280 dumpMapping(L);
281 dbgs() << "\nrhs-";
282 dumpMapping(R);
283 }
284
debugSizeMismatch(const ValueMapping & L,const ValueMapping & R)285 static void debugSizeMismatch(const ValueMapping &L, const ValueMapping &R) {
286 dbgs() << " - fail: map size: " << L.Values.size()
287 << " != " << R.Values.size() << "\n";
288 dbgs() << "\nlhs-";
289 dumpMapping(L);
290 dbgs() << "\nrhs-";
291 dumpMapping(R);
292 }
293 #endif
294
matches(const ValueMapping & LM,const ValueMapping & RM)295 static bool matches(const ValueMapping &LM, const ValueMapping &RM) {
296 LLVM_DEBUG(dbgs() << "compare value maps\n");
297 if (LM.Values.size() != RM.Values.size()) {
298 LLVM_DEBUG(debugSizeMismatch(LM, RM));
299 return false;
300 }
301
302 // This mapping doesn't include dangling constant users, since those don't
303 // get serialized. However, checking if users are constant and calling
304 // isConstantUsed() on every one is very expensive. Instead, just check if
305 // the user is mapped.
306 auto skipUnmappedUsers =
307 [&](Value::const_use_iterator &U, Value::const_use_iterator E,
308 const ValueMapping &M) {
309 while (U != E && !M.lookup(U->getUser()))
310 ++U;
311 };
312
313 // Iterate through all values, and check that both mappings have the same
314 // users.
315 for (unsigned I = 0, E = LM.Values.size(); I != E; ++I) {
316 const Value *L = LM.Values[I];
317 const Value *R = RM.Values[I];
318 auto LU = L->use_begin(), LE = L->use_end();
319 auto RU = R->use_begin(), RE = R->use_end();
320 skipUnmappedUsers(LU, LE, LM);
321 skipUnmappedUsers(RU, RE, RM);
322
323 while (LU != LE) {
324 if (RU == RE) {
325 LLVM_DEBUG(debugUserMismatch(LM, RM, I));
326 return false;
327 }
328 if (LM.lookup(LU->getUser()) != RM.lookup(RU->getUser())) {
329 LLVM_DEBUG(debugUserMismatch(LM, RM, I));
330 return false;
331 }
332 if (LU->getOperandNo() != RU->getOperandNo()) {
333 LLVM_DEBUG(debugUserMismatch(LM, RM, I));
334 return false;
335 }
336 skipUnmappedUsers(++LU, LE, LM);
337 skipUnmappedUsers(++RU, RE, RM);
338 }
339 if (RU != RE) {
340 LLVM_DEBUG(debugUserMismatch(LM, RM, I));
341 return false;
342 }
343 }
344
345 return true;
346 }
347
verifyAfterRoundTrip(const Module & M,std::unique_ptr<Module> OtherM)348 static void verifyAfterRoundTrip(const Module &M,
349 std::unique_ptr<Module> OtherM) {
350 if (!OtherM)
351 report_fatal_error("parsing failed");
352 if (verifyModule(*OtherM, &errs()))
353 report_fatal_error("verification failed");
354 if (!matches(ValueMapping(M), ValueMapping(*OtherM)))
355 report_fatal_error("use-list order changed");
356 }
357
verifyBitcodeUseListOrder(const Module & M)358 static void verifyBitcodeUseListOrder(const Module &M) {
359 TempFile F;
360 if (F.init("bc"))
361 report_fatal_error("failed to initialize bitcode file");
362
363 if (F.writeBitcode(M))
364 report_fatal_error("failed to write bitcode");
365
366 LLVMContext Context;
367 verifyAfterRoundTrip(M, F.readBitcode(Context));
368 }
369
verifyAssemblyUseListOrder(const Module & M)370 static void verifyAssemblyUseListOrder(const Module &M) {
371 TempFile F;
372 if (F.init("ll"))
373 report_fatal_error("failed to initialize assembly file");
374
375 if (F.writeAssembly(M))
376 report_fatal_error("failed to write assembly");
377
378 LLVMContext Context;
379 verifyAfterRoundTrip(M, F.readAssembly(Context));
380 }
381
verifyUseListOrder(const Module & M)382 static void verifyUseListOrder(const Module &M) {
383 outs() << "verify bitcode\n";
384 verifyBitcodeUseListOrder(M);
385 outs() << "verify assembly\n";
386 verifyAssemblyUseListOrder(M);
387 }
388
shuffleValueUseLists(Value * V,std::minstd_rand0 & Gen,DenseSet<Value * > & Seen)389 static void shuffleValueUseLists(Value *V, std::minstd_rand0 &Gen,
390 DenseSet<Value *> &Seen) {
391 if (!Seen.insert(V).second)
392 return;
393
394 if (auto *C = dyn_cast<Constant>(V))
395 if (!isa<GlobalValue>(C))
396 for (Value *Op : C->operands())
397 shuffleValueUseLists(Op, Gen, Seen);
398
399 if (V->use_empty() || std::next(V->use_begin()) == V->use_end())
400 // Nothing to shuffle for 0 or 1 users.
401 return;
402
403 // Generate random numbers between 10 and 99, which will line up nicely in
404 // debug output. We're not worried about collisons here.
405 LLVM_DEBUG(dbgs() << "V = "; V->dump());
406 std::uniform_int_distribution<short> Dist(10, 99);
407 SmallDenseMap<const Use *, short, 16> Order;
408 auto compareUses =
409 [&Order](const Use &L, const Use &R) { return Order[&L] < Order[&R]; };
410 do {
411 for (const Use &U : V->uses()) {
412 auto I = Dist(Gen);
413 Order[&U] = I;
414 LLVM_DEBUG(dbgs() << " - order: " << I << ", op = " << U.getOperandNo()
415 << ", U = ";
416 U.getUser()->dump());
417 }
418 } while (std::is_sorted(V->use_begin(), V->use_end(), compareUses));
419
420 LLVM_DEBUG(dbgs() << " => shuffle\n");
421 V->sortUseList(compareUses);
422
423 LLVM_DEBUG({
424 for (const Use &U : V->uses()) {
425 dbgs() << " - order: " << Order.lookup(&U)
426 << ", op = " << U.getOperandNo() << ", U = ";
427 U.getUser()->dump();
428 }
429 });
430 }
431
reverseValueUseLists(Value * V,DenseSet<Value * > & Seen)432 static void reverseValueUseLists(Value *V, DenseSet<Value *> &Seen) {
433 if (!Seen.insert(V).second)
434 return;
435
436 if (auto *C = dyn_cast<Constant>(V))
437 if (!isa<GlobalValue>(C))
438 for (Value *Op : C->operands())
439 reverseValueUseLists(Op, Seen);
440
441 if (V->use_empty() || std::next(V->use_begin()) == V->use_end())
442 // Nothing to shuffle for 0 or 1 users.
443 return;
444
445 LLVM_DEBUG({
446 dbgs() << "V = ";
447 V->dump();
448 for (const Use &U : V->uses()) {
449 dbgs() << " - order: op = " << U.getOperandNo() << ", U = ";
450 U.getUser()->dump();
451 }
452 dbgs() << " => reverse\n";
453 });
454
455 V->reverseUseList();
456
457 LLVM_DEBUG({
458 for (const Use &U : V->uses()) {
459 dbgs() << " - order: op = " << U.getOperandNo() << ", U = ";
460 U.getUser()->dump();
461 }
462 });
463 }
464
465 template <class Changer>
changeUseLists(Module & M,Changer changeValueUseList)466 static void changeUseLists(Module &M, Changer changeValueUseList) {
467 // Visit every value that would be serialized to an IR file.
468 //
469 // Globals.
470 for (GlobalVariable &G : M.globals())
471 changeValueUseList(&G);
472 for (GlobalAlias &A : M.aliases())
473 changeValueUseList(&A);
474 for (GlobalIFunc &IF : M.ifuncs())
475 changeValueUseList(&IF);
476 for (Function &F : M)
477 changeValueUseList(&F);
478
479 // Constants used by globals.
480 for (GlobalVariable &G : M.globals())
481 if (G.hasInitializer())
482 changeValueUseList(G.getInitializer());
483 for (GlobalAlias &A : M.aliases())
484 changeValueUseList(A.getAliasee());
485 for (GlobalIFunc &IF : M.ifuncs())
486 changeValueUseList(IF.getResolver());
487 for (Function &F : M) {
488 if (F.hasPrefixData())
489 changeValueUseList(F.getPrefixData());
490 if (F.hasPrologueData())
491 changeValueUseList(F.getPrologueData());
492 if (F.hasPersonalityFn())
493 changeValueUseList(F.getPersonalityFn());
494 }
495
496 // Function bodies.
497 for (Function &F : M) {
498 for (Argument &A : F.args())
499 changeValueUseList(&A);
500 for (BasicBlock &BB : F)
501 changeValueUseList(&BB);
502 for (BasicBlock &BB : F)
503 for (Instruction &I : BB)
504 changeValueUseList(&I);
505
506 // Constants used by instructions.
507 for (BasicBlock &BB : F)
508 for (Instruction &I : BB)
509 for (Value *Op : I.operands()) {
510 // Look through a metadata wrapper.
511 if (auto *MAV = dyn_cast<MetadataAsValue>(Op))
512 if (auto *VAM = dyn_cast<ValueAsMetadata>(MAV->getMetadata()))
513 Op = VAM->getValue();
514 if ((isa<Constant>(Op) && !isa<GlobalValue>(*Op)) ||
515 isa<InlineAsm>(Op))
516 changeValueUseList(Op);
517 }
518 }
519
520 if (verifyModule(M, &errs()))
521 report_fatal_error("verification failed");
522 }
523
shuffleUseLists(Module & M,unsigned SeedOffset)524 static void shuffleUseLists(Module &M, unsigned SeedOffset) {
525 std::minstd_rand0 Gen(std::minstd_rand0::default_seed + SeedOffset);
526 DenseSet<Value *> Seen;
527 changeUseLists(M, [&](Value *V) { shuffleValueUseLists(V, Gen, Seen); });
528 LLVM_DEBUG(dbgs() << "\n");
529 }
530
reverseUseLists(Module & M)531 static void reverseUseLists(Module &M) {
532 DenseSet<Value *> Seen;
533 changeUseLists(M, [&](Value *V) { reverseValueUseLists(V, Seen); });
534 LLVM_DEBUG(dbgs() << "\n");
535 }
536
main(int argc,char ** argv)537 int main(int argc, char **argv) {
538 InitLLVM X(argc, argv);
539
540 // Enable debug stream buffering.
541 EnableDebugBuffering = true;
542
543 LLVMContext Context;
544
545 cl::ParseCommandLineOptions(argc, argv,
546 "llvm tool to verify use-list order\n");
547
548 SMDiagnostic Err;
549
550 // Load the input module...
551 std::unique_ptr<Module> M = parseIRFile(InputFilename, Err, Context);
552
553 if (!M.get()) {
554 Err.print(argv[0], errs());
555 return 1;
556 }
557 if (verifyModule(*M, &errs())) {
558 errs() << argv[0] << ": " << InputFilename
559 << ": error: input module is broken!\n";
560 return 1;
561 }
562
563 // Verify the use lists now and after reversing them.
564 outs() << "*** verify-uselistorder ***\n";
565 verifyUseListOrder(*M);
566 outs() << "reverse\n";
567 reverseUseLists(*M);
568 verifyUseListOrder(*M);
569
570 for (unsigned I = 0, E = NumShuffles; I != E; ++I) {
571 outs() << "\n";
572
573 // Shuffle with a different (deterministic) seed each time.
574 outs() << "shuffle (" << I + 1 << " of " << E << ")\n";
575 shuffleUseLists(*M, I);
576
577 // Verify again before and after reversing.
578 verifyUseListOrder(*M);
579 outs() << "reverse\n";
580 reverseUseLists(*M);
581 verifyUseListOrder(*M);
582 }
583
584 return 0;
585 }
586