c0bd2446930365fe8a40d0ca8d35489876e0967f
[c11llvm.git] / CDSPass.cpp
1 //===-- CDSPass.cpp - xxx -------------------------------===//
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 // This file is distributed under the University of Illinois Open Source
7 // License. See LICENSE.TXT for details.
8 //
9 //===----------------------------------------------------------------------===//
10 //
11 // This file is a modified version of ThreadSanitizer.cpp, a part of a race detector.
12 //
13 // The tool is under development, for the details about previous versions see
14 // http://code.google.com/p/data-race-test
15 //
16 // The instrumentation phase is quite simple:
17 //   - Insert calls to run-time library before every memory access.
18 //      - Optimizations may apply to avoid instrumenting some of the accesses.
19 //   - Insert calls at function entry/exit.
20 // The rest is handled by the run-time library.
21 //===----------------------------------------------------------------------===//
22
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/SmallString.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/Analysis/CaptureTracking.h"
28 #include "llvm/IR/BasicBlock.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/LegacyPassManager.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/PassManager.h"
36 #include "llvm/Pass.h"
37 #include "llvm/ProfileData/InstrProf.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Support/AtomicOrdering.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
45 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
46 #include <vector>
47
48 using namespace llvm;
49
50 #define DEBUG_TYPE "CDS"
51 #include <llvm/IR/DebugLoc.h>
52
53 Value *getPosition( Instruction * I, IRBuilder <> IRB, bool print = false)
54 {
55         const DebugLoc & debug_location = I->getDebugLoc ();
56         std::string position_string;
57         {
58                 llvm::raw_string_ostream position_stream (position_string);
59                 debug_location . print (position_stream);
60         }
61
62         if (print) {
63                 errs() << position_string << "\n";
64         }
65
66         return IRB.CreateGlobalStringPtr (position_string);
67 }
68
69 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
70 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
71 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
72 // STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
73 // STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
74
75 STATISTIC(NumOmittedReadsBeforeWrite,
76           "Number of reads ignored due to following writes");
77 STATISTIC(NumOmittedReadsFromConstantGlobals,
78           "Number of reads from constant globals");
79 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
80 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
81
82 Type * OrdTy;
83
84 Type * Int8PtrTy;
85 Type * Int16PtrTy;
86 Type * Int32PtrTy;
87 Type * Int64PtrTy;
88
89 Type * VoidTy;
90
91 static const size_t kNumberOfAccessSizes = 4;
92
93 int getAtomicOrderIndex(AtomicOrdering order){
94         switch (order) {
95                 case AtomicOrdering::Monotonic: 
96                         return (int)AtomicOrderingCABI::relaxed;
97                 //  case AtomicOrdering::Consume:         // not specified yet
98                 //    return AtomicOrderingCABI::consume;
99                 case AtomicOrdering::Acquire: 
100                         return (int)AtomicOrderingCABI::acquire;
101                 case AtomicOrdering::Release: 
102                         return (int)AtomicOrderingCABI::release;
103                 case AtomicOrdering::AcquireRelease: 
104                         return (int)AtomicOrderingCABI::acq_rel;
105                 case AtomicOrdering::SequentiallyConsistent: 
106                         return (int)AtomicOrderingCABI::seq_cst;
107                 default:
108                         // unordered or Not Atomic
109                         return -1;
110         }
111 }
112
113 namespace {
114         struct CDSPass : public FunctionPass {
115                 static char ID;
116                 CDSPass() : FunctionPass(ID) {}
117                 bool runOnFunction(Function &F) override; 
118
119         private:
120                 void initializeCallbacks(Module &M);
121                 bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
122                 bool isAtomicCall(Instruction *I);
123                 bool instrumentAtomic(Instruction *I, const DataLayout &DL);
124                 bool instrumentAtomicCall(CallInst *CI, const DataLayout &DL);
125                 void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
126                                                                                         SmallVectorImpl<Instruction *> &All,
127                                                                                         const DataLayout &DL);
128                 bool addrPointsToConstantData(Value *Addr);
129                 int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
130
131                 // Callbacks to run-time library are computed in doInitialization.
132                 Constant * CDSFuncEntry;
133                 Constant * CDSFuncExit;
134
135                 Constant * CDSLoad[kNumberOfAccessSizes];
136                 Constant * CDSStore[kNumberOfAccessSizes];
137                 Constant * CDSAtomicInit[kNumberOfAccessSizes];
138                 Constant * CDSAtomicLoad[kNumberOfAccessSizes];
139                 Constant * CDSAtomicStore[kNumberOfAccessSizes];
140                 Constant * CDSAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
141                 Constant * CDSAtomicCAS_V1[kNumberOfAccessSizes];
142                 Constant * CDSAtomicCAS_V2[kNumberOfAccessSizes];
143                 Constant * CDSAtomicThreadFence;
144
145                 std::vector<StringRef> AtomicFuncNames;
146                 std::vector<StringRef> PartialAtomicFuncNames;
147         };
148 }
149
150 static bool isVtableAccess(Instruction *I) {
151         if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
152                 return Tag->isTBAAVtableAccess();
153         return false;
154 }
155
156 void CDSPass::initializeCallbacks(Module &M) {
157         LLVMContext &Ctx = M.getContext();
158
159         Type * Int1Ty = Type::getInt1Ty(Ctx);
160         OrdTy = Type::getInt32Ty(Ctx);
161
162         Int8PtrTy  = Type::getInt8PtrTy(Ctx);
163         Int16PtrTy = Type::getInt16PtrTy(Ctx);
164         Int32PtrTy = Type::getInt32PtrTy(Ctx);
165         Int64PtrTy = Type::getInt64PtrTy(Ctx);
166
167         VoidTy = Type::getVoidTy(Ctx);
168
169         CDSFuncEntry = M.getOrInsertFunction("cds_func_entry", 
170                                                                 VoidTy, Int8PtrTy);
171         CDSFuncExit = M.getOrInsertFunction("cds_func_exit", 
172                                                                 VoidTy, Int8PtrTy);
173
174         // Get the function to call from our untime library.
175         for (unsigned i = 0; i < kNumberOfAccessSizes; i++) {
176                 const unsigned ByteSize = 1U << i;
177                 const unsigned BitSize = ByteSize * 8;
178
179                 std::string ByteSizeStr = utostr(ByteSize);
180                 std::string BitSizeStr = utostr(BitSize);
181
182                 Type *Ty = Type::getIntNTy(Ctx, BitSize);
183                 Type *PtrTy = Ty->getPointerTo();
184
185                 // uint8_t cds_atomic_load8 (void * obj, int atomic_index)
186                 // void cds_atomic_store8 (void * obj, int atomic_index, uint8_t val)
187                 SmallString<32> LoadName("cds_load" + BitSizeStr);
188                 SmallString<32> StoreName("cds_store" + BitSizeStr);
189                 SmallString<32> AtomicInitName("cds_atomic_init" + BitSizeStr);
190                 SmallString<32> AtomicLoadName("cds_atomic_load" + BitSizeStr);
191                 SmallString<32> AtomicStoreName("cds_atomic_store" + BitSizeStr);
192
193                 CDSLoad[i]  = M.getOrInsertFunction(LoadName, VoidTy, PtrTy);
194                 CDSStore[i] = M.getOrInsertFunction(StoreName, VoidTy, PtrTy);
195                 CDSAtomicInit[i] = M.getOrInsertFunction(AtomicInitName, 
196                                                                 VoidTy, PtrTy, Ty, Int8PtrTy);
197                 CDSAtomicLoad[i]  = M.getOrInsertFunction(AtomicLoadName, 
198                                                                 Ty, PtrTy, OrdTy, Int8PtrTy);
199                 CDSAtomicStore[i] = M.getOrInsertFunction(AtomicStoreName, 
200                                                                 VoidTy, PtrTy, Ty, OrdTy, Int8PtrTy);
201
202                 for (int op = AtomicRMWInst::FIRST_BINOP; 
203                         op <= AtomicRMWInst::LAST_BINOP; ++op) {
204                         CDSAtomicRMW[op][i] = nullptr;
205                         std::string NamePart;
206
207                         if (op == AtomicRMWInst::Xchg)
208                                 NamePart = "_exchange";
209                         else if (op == AtomicRMWInst::Add) 
210                                 NamePart = "_fetch_add";
211                         else if (op == AtomicRMWInst::Sub)
212                                 NamePart = "_fetch_sub";
213                         else if (op == AtomicRMWInst::And)
214                                 NamePart = "_fetch_and";
215                         else if (op == AtomicRMWInst::Or)
216                                 NamePart = "_fetch_or";
217                         else if (op == AtomicRMWInst::Xor)
218                                 NamePart = "_fetch_xor";
219                         else
220                                 continue;
221
222                         SmallString<32> AtomicRMWName("cds_atomic" + NamePart + BitSizeStr);
223                         CDSAtomicRMW[op][i] = M.getOrInsertFunction(AtomicRMWName, 
224                                                                                 Ty, PtrTy, Ty, OrdTy, Int8PtrTy);
225                 }
226
227                 // only supportes strong version
228                 SmallString<32> AtomicCASName_V1("cds_atomic_compare_exchange" + BitSizeStr + "_v1");
229                 SmallString<32> AtomicCASName_V2("cds_atomic_compare_exchange" + BitSizeStr + "_v2");
230                 CDSAtomicCAS_V1[i] = M.getOrInsertFunction(AtomicCASName_V1, 
231                                                                 Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, Int8PtrTy);
232                 CDSAtomicCAS_V2[i] = M.getOrInsertFunction(AtomicCASName_V2, 
233                                                                 Int1Ty, PtrTy, PtrTy, Ty, OrdTy, OrdTy, Int8PtrTy);
234         }
235
236         CDSAtomicThreadFence = M.getOrInsertFunction("cds_atomic_thread_fence", 
237                                                                                                         VoidTy, OrdTy, Int8PtrTy);
238 }
239
240 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
241         // Peel off GEPs and BitCasts.
242         Addr = Addr->stripInBoundsOffsets();
243
244         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
245                 if (GV->hasSection()) {
246                         StringRef SectionName = GV->getSection();
247                         // Check if the global is in the PGO counters section.
248                         auto OF = Triple(M->getTargetTriple()).getObjectFormat();
249                         if (SectionName.endswith(
250                               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
251                                 return false;
252                 }
253
254                 // Check if the global is private gcov data.
255                 if (GV->getName().startswith("__llvm_gcov") ||
256                 GV->getName().startswith("__llvm_gcda"))
257                 return false;
258         }
259
260         // Do not instrument acesses from different address spaces; we cannot deal
261         // with them.
262         if (Addr) {
263                 Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
264                 if (PtrTy->getPointerAddressSpace() != 0)
265                         return false;
266         }
267
268         return true;
269 }
270
271 bool CDSPass::addrPointsToConstantData(Value *Addr) {
272         // If this is a GEP, just analyze its pointer operand.
273         if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
274                 Addr = GEP->getPointerOperand();
275
276         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
277                 if (GV->isConstant()) {
278                         // Reads from constant globals can not race with any writes.
279                         NumOmittedReadsFromConstantGlobals++;
280                         return true;
281                 }
282         } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
283                 if (isVtableAccess(L)) {
284                         // Reads from a vtable pointer can not race with any writes.
285                         NumOmittedReadsFromVtable++;
286                         return true;
287                 }
288         }
289         return false;
290 }
291
292 bool CDSPass::runOnFunction(Function &F) {
293         if (F.getName() == "main") {
294                 F.setName("user_main");
295                 errs() << "main replaced by user_main\n";
296         }
297
298         if (true) {
299                 initializeCallbacks( *F.getParent() );
300
301                 AtomicFuncNames = 
302                 {
303                         "atomic_init", "atomic_load", "atomic_store", 
304                         "atomic_fetch_", "atomic_exchange", "atomic_compare_exchange_"
305                 };
306
307                 PartialAtomicFuncNames = 
308                 { 
309                         "load", "store", "fetch", "exchange", "compare_exchange_" 
310                 };
311
312                 SmallVector<Instruction*, 8> AllLoadsAndStores;
313                 SmallVector<Instruction*, 8> LocalLoadsAndStores;
314                 SmallVector<Instruction*, 8> AtomicAccesses;
315
316                 std::vector<Instruction *> worklist;
317
318                 bool Res = false;
319                 bool HasAtomic = false;
320                 const DataLayout &DL = F.getParent()->getDataLayout();
321
322                 // errs() << "--- " << F.getName() << "---\n";
323
324                 for (auto &B : F) {
325                         for (auto &I : B) {
326                                 if ( (&I)->isAtomic() || isAtomicCall(&I) ) {
327                                         AtomicAccesses.push_back(&I);
328                                         HasAtomic = true;
329                                 } else if (isa<LoadInst>(I) || isa<StoreInst>(I)) {
330                                         LocalLoadsAndStores.push_back(&I);
331                                 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
332                                         // not implemented yet
333                                 }
334                         }
335
336                         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
337                 }
338
339                 for (auto Inst : AllLoadsAndStores) {
340                         Res |= instrumentLoadOrStore(Inst, DL);
341                 }
342
343                 for (auto Inst : AtomicAccesses) {
344                         Res |= instrumentAtomic(Inst, DL);
345                 }
346
347                 // only instrument functions that contain atomics
348                 if (Res && HasAtomic) {
349                         IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
350                         /* Unused for now
351                         Value *ReturnAddress = IRB.CreateCall(
352                                 Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
353                                 IRB.getInt32(0));
354                         */
355
356                         Value * FuncName = IRB.CreateGlobalStringPtr(F.getName());
357                         IRB.CreateCall(CDSFuncEntry, FuncName);
358
359                         EscapeEnumerator EE(F, "cds_cleanup", true);
360                         while (IRBuilder<> *AtExit = EE.Next()) {
361                           AtExit->CreateCall(CDSFuncExit, FuncName);
362                         }
363
364                         Res = true;
365                 }
366         }
367
368         return false;
369 }
370
371 void CDSPass::chooseInstructionsToInstrument(
372         SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
373         const DataLayout &DL) {
374         SmallPtrSet<Value*, 8> WriteTargets;
375         // Iterate from the end.
376         for (Instruction *I : reverse(Local)) {
377                 if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
378                         Value *Addr = Store->getPointerOperand();
379                         if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
380                                 continue;
381                         WriteTargets.insert(Addr);
382                 } else {
383                         LoadInst *Load = cast<LoadInst>(I);
384                         Value *Addr = Load->getPointerOperand();
385                         if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
386                                 continue;
387                         if (WriteTargets.count(Addr)) {
388                                 // We will write to this temp, so no reason to analyze the read.
389                                 NumOmittedReadsBeforeWrite++;
390                                 continue;
391                         }
392                         if (addrPointsToConstantData(Addr)) {
393                                 // Addr points to some constant data -- it can not race with any writes.
394                                 continue;
395                         }
396                 }
397                 Value *Addr = isa<StoreInst>(*I)
398                         ? cast<StoreInst>(I)->getPointerOperand()
399                         : cast<LoadInst>(I)->getPointerOperand();
400                 if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
401                                 !PointerMayBeCaptured(Addr, true, true)) {
402                         // The variable is addressable but not captured, so it cannot be
403                         // referenced from a different thread and participate in a data race
404                         // (see llvm/Analysis/CaptureTracking.h for details).
405                         NumOmittedNonCaptured++;
406                         continue;
407                 }
408                 All.push_back(I);
409         }
410         Local.clear();
411 }
412
413
414 bool CDSPass::instrumentLoadOrStore(Instruction *I,
415                                                                         const DataLayout &DL) {
416         IRBuilder<> IRB(I);
417         bool IsWrite = isa<StoreInst>(*I);
418         Value *Addr = IsWrite
419                 ? cast<StoreInst>(I)->getPointerOperand()
420                 : cast<LoadInst>(I)->getPointerOperand();
421
422         // swifterror memory addresses are mem2reg promoted by instruction selection.
423         // As such they cannot have regular uses like an instrumentation function and
424         // it makes no sense to track them as memory.
425         if (Addr->isSwiftError())
426         return false;
427
428         int Idx = getMemoryAccessFuncIndex(Addr, DL);
429         if (Idx < 0)
430                 return false;
431
432 //  not supported by CDS yet
433 /*  if (IsWrite && isVtableAccess(I)) {
434     LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
435     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
436     // StoredValue may be a vector type if we are storing several vptrs at once.
437     // In this case, just take the first element of the vector since this is
438     // enough to find vptr races.
439     if (isa<VectorType>(StoredValue->getType()))
440       StoredValue = IRB.CreateExtractElement(
441           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
442     if (StoredValue->getType()->isIntegerTy())
443       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
444     // Call TsanVptrUpdate.
445     IRB.CreateCall(TsanVptrUpdate,
446                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
447                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
448     NumInstrumentedVtableWrites++;
449     return true;
450   }
451
452   if (!IsWrite && isVtableAccess(I)) {
453     IRB.CreateCall(TsanVptrLoad,
454                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
455     NumInstrumentedVtableReads++;
456     return true;
457   }
458 */
459
460         Value *OnAccessFunc = nullptr;
461         OnAccessFunc = IsWrite ? CDSStore[Idx] : CDSLoad[Idx];
462
463         Type *ArgType = IRB.CreatePointerCast(Addr, Addr->getType())->getType();
464
465         if ( ArgType != Int8PtrTy && ArgType != Int16PtrTy && 
466                         ArgType != Int32PtrTy && ArgType != Int64PtrTy ) {
467                 // if other types of load or stores are passed in
468                 return false;   
469         }
470         IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, Addr->getType()));
471         if (IsWrite) NumInstrumentedWrites++;
472         else         NumInstrumentedReads++;
473         return true;
474 }
475
476 bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
477         IRBuilder<> IRB(I);
478
479         // errs() << "instrumenting: " << *I << "\n";
480
481         if (auto *CI = dyn_cast<CallInst>(I)) {
482                 return instrumentAtomicCall(CI, DL);
483         }
484
485         Value *position = getPosition(I, IRB);
486
487         if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
488                 Value *Addr = LI->getPointerOperand();
489                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
490                 if (Idx < 0)
491                         return false;
492
493                 int atomic_order_index = getAtomicOrderIndex(LI->getOrdering());
494                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
495                 Value *args[] = {Addr, order, position};
496                 Instruction* funcInst=CallInst::Create(CDSAtomicLoad[Idx], args);
497                 ReplaceInstWithInst(LI, funcInst);
498         } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
499                 Value *Addr = SI->getPointerOperand();
500                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
501                 if (Idx < 0)
502                         return false;
503
504                 int atomic_order_index = getAtomicOrderIndex(SI->getOrdering());
505                 Value *val = SI->getValueOperand();
506                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
507                 Value *args[] = {Addr, val, order, position};
508                 Instruction* funcInst=CallInst::Create(CDSAtomicStore[Idx], args);
509                 ReplaceInstWithInst(SI, funcInst);
510         } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
511                 Value *Addr = RMWI->getPointerOperand();
512                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
513                 if (Idx < 0)
514                         return false;
515
516                 int atomic_order_index = getAtomicOrderIndex(RMWI->getOrdering());
517                 Value *val = RMWI->getValOperand();
518                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
519                 Value *args[] = {Addr, val, order, position};
520                 Instruction* funcInst = CallInst::Create(CDSAtomicRMW[RMWI->getOperation()][Idx], args);
521                 ReplaceInstWithInst(RMWI, funcInst);
522         } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
523                 IRBuilder<> IRB(CASI);
524
525                 Value *Addr = CASI->getPointerOperand();
526                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
527                 if (Idx < 0)
528                         return false;
529
530                 const unsigned ByteSize = 1U << Idx;
531                 const unsigned BitSize = ByteSize * 8;
532                 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
533                 Type *PtrTy = Ty->getPointerTo();
534
535                 Value *CmpOperand = IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
536                 Value *NewOperand = IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
537
538                 int atomic_order_index_succ = getAtomicOrderIndex(CASI->getSuccessOrdering());
539                 int atomic_order_index_fail = getAtomicOrderIndex(CASI->getFailureOrdering());
540                 Value *order_succ = ConstantInt::get(OrdTy, atomic_order_index_succ);
541                 Value *order_fail = ConstantInt::get(OrdTy, atomic_order_index_fail);
542
543                 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
544                                                  CmpOperand, NewOperand,
545                                                  order_succ, order_fail, position};
546
547                 CallInst *funcInst = IRB.CreateCall(CDSAtomicCAS_V1[Idx], Args);
548                 Value *Success = IRB.CreateICmpEQ(funcInst, CmpOperand);
549
550                 Value *OldVal = funcInst;
551                 Type *OrigOldValTy = CASI->getNewValOperand()->getType();
552                 if (Ty != OrigOldValTy) {
553                         // The value is a pointer, so we need to cast the return value.
554                         OldVal = IRB.CreateIntToPtr(funcInst, OrigOldValTy);
555                 }
556
557                 Value *Res =
558                   IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
559                 Res = IRB.CreateInsertValue(Res, Success, 1);
560
561                 I->replaceAllUsesWith(Res);
562                 I->eraseFromParent();
563         } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
564                 int atomic_order_index = getAtomicOrderIndex(FI->getOrdering());
565                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
566                 Value *Args[] = {order, position};
567
568                 CallInst *funcInst = CallInst::Create(CDSAtomicThreadFence, Args);
569                 ReplaceInstWithInst(FI, funcInst);
570                 // errs() << "Thread Fences replaced\n";
571         }
572         return true;
573 }
574
575 bool CDSPass::isAtomicCall(Instruction *I) {
576         if ( auto *CI = dyn_cast<CallInst>(I) ) {
577                 Function *fun = CI->getCalledFunction();
578                 if (fun == NULL)
579                         return false;
580
581                 StringRef funName = fun->getName();
582
583                 // todo: come up with better rules for function name checking
584                 for (StringRef name : AtomicFuncNames) {
585                         if ( funName.contains(name) ) 
586                                 return true;
587                 }
588                 
589                 for (StringRef PartialName : PartialAtomicFuncNames) {
590                         if (funName.contains(PartialName) && 
591                                         funName.contains("atomic") )
592                                 return true;
593                 }
594         }
595
596         return false;
597 }
598
599 bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
600         IRBuilder<> IRB(CI);
601         Function *fun = CI->getCalledFunction();
602         StringRef funName = fun->getName();
603         std::vector<Value *> parameters;
604
605         User::op_iterator begin = CI->arg_begin();
606         User::op_iterator end = CI->arg_end();
607         for (User::op_iterator it = begin; it != end; ++it) {
608                 Value *param = *it;
609                 parameters.push_back(param);
610         }
611
612         // obtain source line number of the CallInst
613         Value *position = getPosition(CI, IRB);
614
615         // the pointer to the address is always the first argument
616         Value *OrigPtr = parameters[0];
617
618         int Idx = getMemoryAccessFuncIndex(OrigPtr, DL);
619         if (Idx < 0)
620                 return false;
621
622         const unsigned ByteSize = 1U << Idx;
623         const unsigned BitSize = ByteSize * 8;
624         Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
625         Type *PtrTy = Ty->getPointerTo();
626
627         // atomic_init; args = {obj, order}
628         if (funName.contains("atomic_init")) {
629                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
630                 Value *val = IRB.CreateBitOrPointerCast(parameters[1], Ty);
631                 Value *args[] = {ptr, val, position};
632
633                 Instruction* funcInst = CallInst::Create(CDSAtomicInit[Idx], args);
634                 ReplaceInstWithInst(CI, funcInst);
635
636                 return true;
637         }
638
639         // atomic_load; args = {obj, order}
640         if (funName.contains("atomic_load")) {
641                 bool isExplicit = funName.contains("atomic_load_explicit");
642
643                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
644                 Value *order;
645                 if (isExplicit)
646                         order = IRB.CreateBitOrPointerCast(parameters[1], OrdTy);
647                 else 
648                         order = ConstantInt::get(OrdTy, 
649                                                         (int) AtomicOrderingCABI::seq_cst);
650                 Value *args[] = {ptr, order, position};
651                 
652                 Instruction* funcInst = CallInst::Create(CDSAtomicLoad[Idx], args);
653                 ReplaceInstWithInst(CI, funcInst);
654
655                 return true;
656         } else if (funName.contains("atomic") && 
657                                         funName.contains("load") ) {
658                 // does this version of call always have an atomic order as an argument?
659                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
660                 Value *order = IRB.CreateBitOrPointerCast(parameters[1], OrdTy);
661                 Value *args[] = {ptr, order, position};
662
663                 if (!CI->getType()->isPointerTy()) {
664                         return false;   
665                 } 
666
667                 CallInst *funcInst = IRB.CreateCall(CDSAtomicLoad[Idx], args);
668                 Value *RetVal = IRB.CreateIntToPtr(funcInst, CI->getType());
669
670                 CI->replaceAllUsesWith(RetVal);
671                 CI->eraseFromParent();
672
673                 return true;
674         }
675
676         // atomic_store; args = {obj, val, order}
677         if (funName.contains("atomic_store")) {
678                 bool isExplicit = funName.contains("atomic_store_explicit");
679                 Value *OrigVal = parameters[1];
680
681                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
682                 Value *val = IRB.CreatePointerCast(OrigVal, Ty);
683                 Value *order;
684                 if (isExplicit)
685                         order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
686                 else 
687                         order = ConstantInt::get(OrdTy, 
688                                                         (int) AtomicOrderingCABI::seq_cst);
689                 Value *args[] = {ptr, val, order, position};
690                 
691                 Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
692                 ReplaceInstWithInst(CI, funcInst);
693
694                 return true;
695         } else if (funName.contains("atomic") && 
696                                         funName.contains("EEEE5store") ) {
697                 // does this version of call always have an atomic order as an argument?
698                 Value *OrigVal = parameters[1];
699
700                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
701                 Value *val = IRB.CreatePointerCast(OrigVal, Ty);
702                 Value *order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
703                 Value *args[] = {ptr, val, order, position};
704
705                 Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
706                 ReplaceInstWithInst(CI, funcInst);
707
708                 return true;
709         }
710
711         // atomic_fetch_*; args = {obj, val, order}
712         if (funName.contains("atomic_fetch_") || 
713                         funName.contains("atomic_exchange") ) {
714                 bool isExplicit = funName.contains("_explicit");
715                 Value *OrigVal = parameters[1];
716
717                 int op;
718                 if ( funName.contains("_fetch_add") )
719                         op = AtomicRMWInst::Add;
720                 else if ( funName.contains("_fetch_sub") )
721                         op = AtomicRMWInst::Sub;
722                 else if ( funName.contains("_fetch_and") )
723                         op = AtomicRMWInst::And;
724                 else if ( funName.contains("_fetch_or") )
725                         op = AtomicRMWInst::Or;
726                 else if ( funName.contains("_fetch_xor") )
727                         op = AtomicRMWInst::Xor;
728                 else if ( funName.contains("atomic_exchange") )
729                         op = AtomicRMWInst::Xchg;
730                 else {
731                         errs() << "Unknown atomic read-modify-write operation\n";
732                         return false;
733                 }
734
735                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
736                 Value *val = IRB.CreatePointerCast(OrigVal, Ty);
737                 Value *order;
738                 if (isExplicit)
739                         order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
740                 else 
741                         order = ConstantInt::get(OrdTy, 
742                                                         (int) AtomicOrderingCABI::seq_cst);
743                 Value *args[] = {ptr, val, order, position};
744                 
745                 Instruction* funcInst = CallInst::Create(CDSAtomicRMW[op][Idx], args);
746                 ReplaceInstWithInst(CI, funcInst);
747
748                 return true;
749         } else if (funName.contains("fetch")) {
750                 errs() << "atomic exchange captured. Not implemented yet. ";
751                 errs() << "See source file :";
752                 getPosition(CI, IRB, true);
753         } else if (funName.contains("exchange") &&
754                         !funName.contains("compare_exchange") ) {
755                 errs() << "atomic exchange captured. Not implemented yet. ";
756                 errs() << "See source file :";
757                 getPosition(CI, IRB, true);
758         }
759
760         /* atomic_compare_exchange_*; 
761            args = {obj, expected, new value, order1, order2}
762         */
763         if ( funName.contains("atomic_compare_exchange_") ) {
764                 bool isExplicit = funName.contains("_explicit");
765
766                 Value *Addr = IRB.CreatePointerCast(OrigPtr, PtrTy);
767                 Value *CmpOperand = IRB.CreatePointerCast(parameters[1], PtrTy);
768                 Value *NewOperand = IRB.CreateBitOrPointerCast(parameters[2], Ty);
769
770                 Value *order_succ, *order_fail;
771                 if (isExplicit) {
772                         order_succ = IRB.CreateBitOrPointerCast(parameters[3], OrdTy);
773                         order_fail = IRB.CreateBitOrPointerCast(parameters[4], OrdTy);
774                 } else  {
775                         order_succ = ConstantInt::get(OrdTy, 
776                                                         (int) AtomicOrderingCABI::seq_cst);
777                         order_fail = ConstantInt::get(OrdTy, 
778                                                         (int) AtomicOrderingCABI::seq_cst);
779                 }
780
781                 Value *args[] = {Addr, CmpOperand, NewOperand, 
782                                                         order_succ, order_fail, position};
783                 
784                 Instruction* funcInst = CallInst::Create(CDSAtomicCAS_V2[Idx], args);
785                 ReplaceInstWithInst(CI, funcInst);
786
787                 return true;
788         } else if ( funName.contains("compare_exchange_strong") ||
789                                 funName.contains("compare_exchange_weak") ) {
790                 Value *Addr = IRB.CreatePointerCast(OrigPtr, PtrTy);
791                 Value *CmpOperand = IRB.CreatePointerCast(parameters[1], PtrTy);
792                 Value *NewOperand = IRB.CreateBitOrPointerCast(parameters[2], Ty);
793
794                 Value *order_succ, *order_fail;
795                 order_succ = IRB.CreateBitOrPointerCast(parameters[3], OrdTy);
796                 order_fail = IRB.CreateBitOrPointerCast(parameters[4], OrdTy);
797
798                 Value *args[] = {Addr, CmpOperand, NewOperand, 
799                                                         order_succ, order_fail, position};
800                 Instruction* funcInst = CallInst::Create(CDSAtomicCAS_V2[Idx], args);
801                 ReplaceInstWithInst(CI, funcInst);
802
803                 return true;
804         }
805
806         return false;
807 }
808
809 int CDSPass::getMemoryAccessFuncIndex(Value *Addr,
810                                                                                 const DataLayout &DL) {
811         Type *OrigPtrTy = Addr->getType();
812         Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
813         assert(OrigTy->isSized());
814         uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
815         if (TypeSize != 8  && TypeSize != 16 &&
816                 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
817                 NumAccessesWithBadSize++;
818                 // Ignore all unusual sizes.
819                 return -1;
820         }
821         size_t Idx = countTrailingZeros(TypeSize / 8);
822         //assert(Idx < kNumberOfAccessSizes);
823         if (Idx >= kNumberOfAccessSizes) {
824                 return -1;
825         }
826         return Idx;
827 }
828
829
830 char CDSPass::ID = 0;
831
832 // Automatically enable the pass.
833 static void registerCDSPass(const PassManagerBuilder &,
834                                                         legacy::PassManagerBase &PM) {
835         PM.add(new CDSPass());
836 }
837
838 /* Enable the pass when opt level is greater than 0 */
839 static RegisterStandardPasses 
840         RegisterMyPass1(PassManagerBuilder::EP_OptimizerLast,
841 registerCDSPass);
842
843 /* Enable the pass when opt level is 0 */
844 static RegisterStandardPasses 
845         RegisterMyPass2(PassManagerBuilder::EP_EnabledOnOptLevel0,
846 registerCDSPass);