xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/Analysis.cpp (revision 734e82fe33aa764367791a7d603b383996c6b40b)
1 //===-- Analysis.cpp - CodeGen LLVM IR Analysis Utilities -----------------===//
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 // This file defines several CodeGen-specific LLVM IR analysis utilities.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/CodeGen/Analysis.h"
14 #include "llvm/Analysis/ValueTracking.h"
15 #include "llvm/CodeGen/MachineFunction.h"
16 #include "llvm/CodeGen/TargetInstrInfo.h"
17 #include "llvm/CodeGen/TargetLowering.h"
18 #include "llvm/CodeGen/TargetSubtargetInfo.h"
19 #include "llvm/IR/DataLayout.h"
20 #include "llvm/IR/DerivedTypes.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/IR/Instructions.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/Support/ErrorHandling.h"
26 #include "llvm/Target/TargetMachine.h"
27 
28 using namespace llvm;
29 
30 /// Compute the linearized index of a member in a nested aggregate/struct/array
31 /// by recursing and accumulating CurIndex as long as there are indices in the
32 /// index list.
33 unsigned llvm::ComputeLinearIndex(Type *Ty,
34                                   const unsigned *Indices,
35                                   const unsigned *IndicesEnd,
36                                   unsigned CurIndex) {
37   // Base case: We're done.
38   if (Indices && Indices == IndicesEnd)
39     return CurIndex;
40 
41   // Given a struct type, recursively traverse the elements.
42   if (StructType *STy = dyn_cast<StructType>(Ty)) {
43     for (auto I : llvm::enumerate(STy->elements())) {
44       Type *ET = I.value();
45       if (Indices && *Indices == I.index())
46         return ComputeLinearIndex(ET, Indices + 1, IndicesEnd, CurIndex);
47       CurIndex = ComputeLinearIndex(ET, nullptr, nullptr, CurIndex);
48     }
49     assert(!Indices && "Unexpected out of bound");
50     return CurIndex;
51   }
52   // Given an array type, recursively traverse the elements.
53   else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
54     Type *EltTy = ATy->getElementType();
55     unsigned NumElts = ATy->getNumElements();
56     // Compute the Linear offset when jumping one element of the array
57     unsigned EltLinearOffset = ComputeLinearIndex(EltTy, nullptr, nullptr, 0);
58     if (Indices) {
59       assert(*Indices < NumElts && "Unexpected out of bound");
60       // If the indice is inside the array, compute the index to the requested
61       // elt and recurse inside the element with the end of the indices list
62       CurIndex += EltLinearOffset* *Indices;
63       return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex);
64     }
65     CurIndex += EltLinearOffset*NumElts;
66     return CurIndex;
67   }
68   // We haven't found the type we're looking for, so keep searching.
69   return CurIndex + 1;
70 }
71 
72 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
73 /// EVTs that represent all the individual underlying
74 /// non-aggregate types that comprise it.
75 ///
76 /// If Offsets is non-null, it points to a vector to be filled in
77 /// with the in-memory offsets of each of the individual values.
78 ///
79 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
80                            Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
81                            SmallVectorImpl<EVT> *MemVTs,
82                            SmallVectorImpl<uint64_t> *Offsets,
83                            uint64_t StartingOffset) {
84   // Given a struct type, recursively traverse the elements.
85   if (StructType *STy = dyn_cast<StructType>(Ty)) {
86     // If the Offsets aren't needed, don't query the struct layout. This allows
87     // us to support structs with scalable vectors for operations that don't
88     // need offsets.
89     const StructLayout *SL = Offsets ? DL.getStructLayout(STy) : nullptr;
90     for (StructType::element_iterator EB = STy->element_begin(),
91                                       EI = EB,
92                                       EE = STy->element_end();
93          EI != EE; ++EI) {
94       // Don't compute the element offset if we didn't get a StructLayout above.
95       uint64_t EltOffset = SL ? SL->getElementOffset(EI - EB) : 0;
96       ComputeValueVTs(TLI, DL, *EI, ValueVTs, MemVTs, Offsets,
97                       StartingOffset + EltOffset);
98     }
99     return;
100   }
101   // Given an array type, recursively traverse the elements.
102   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
103     Type *EltTy = ATy->getElementType();
104     uint64_t EltSize = DL.getTypeAllocSize(EltTy).getFixedValue();
105     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
106       ComputeValueVTs(TLI, DL, EltTy, ValueVTs, MemVTs, Offsets,
107                       StartingOffset + i * EltSize);
108     return;
109   }
110   // Interpret void as zero return values.
111   if (Ty->isVoidTy())
112     return;
113   // Base case: we can get an EVT for this LLVM IR type.
114   ValueVTs.push_back(TLI.getValueType(DL, Ty));
115   if (MemVTs)
116     MemVTs->push_back(TLI.getMemValueType(DL, Ty));
117   if (Offsets)
118     Offsets->push_back(StartingOffset);
119 }
120 
121 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
122                            Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
123                            SmallVectorImpl<uint64_t> *Offsets,
124                            uint64_t StartingOffset) {
125   return ComputeValueVTs(TLI, DL, Ty, ValueVTs, /*MemVTs=*/nullptr, Offsets,
126                          StartingOffset);
127 }
128 
129 void llvm::computeValueLLTs(const DataLayout &DL, Type &Ty,
130                             SmallVectorImpl<LLT> &ValueTys,
131                             SmallVectorImpl<uint64_t> *Offsets,
132                             uint64_t StartingOffset) {
133   // Given a struct type, recursively traverse the elements.
134   if (StructType *STy = dyn_cast<StructType>(&Ty)) {
135     // If the Offsets aren't needed, don't query the struct layout. This allows
136     // us to support structs with scalable vectors for operations that don't
137     // need offsets.
138     const StructLayout *SL = Offsets ? DL.getStructLayout(STy) : nullptr;
139     for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I) {
140       uint64_t EltOffset = SL ? SL->getElementOffset(I) : 0;
141       computeValueLLTs(DL, *STy->getElementType(I), ValueTys, Offsets,
142                        StartingOffset + EltOffset);
143     }
144     return;
145   }
146   // Given an array type, recursively traverse the elements.
147   if (ArrayType *ATy = dyn_cast<ArrayType>(&Ty)) {
148     Type *EltTy = ATy->getElementType();
149     uint64_t EltSize = DL.getTypeAllocSize(EltTy).getFixedValue();
150     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
151       computeValueLLTs(DL, *EltTy, ValueTys, Offsets,
152                        StartingOffset + i * EltSize);
153     return;
154   }
155   // Interpret void as zero return values.
156   if (Ty.isVoidTy())
157     return;
158   // Base case: we can get an LLT for this LLVM IR type.
159   ValueTys.push_back(getLLTForType(Ty, DL));
160   if (Offsets != nullptr)
161     Offsets->push_back(StartingOffset * 8);
162 }
163 
164 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
165 GlobalValue *llvm::ExtractTypeInfo(Value *V) {
166   V = V->stripPointerCasts();
167   GlobalValue *GV = dyn_cast<GlobalValue>(V);
168   GlobalVariable *Var = dyn_cast<GlobalVariable>(V);
169 
170   if (Var && Var->getName() == "llvm.eh.catch.all.value") {
171     assert(Var->hasInitializer() &&
172            "The EH catch-all value must have an initializer");
173     Value *Init = Var->getInitializer();
174     GV = dyn_cast<GlobalValue>(Init);
175     if (!GV) V = cast<ConstantPointerNull>(Init);
176   }
177 
178   assert((GV || isa<ConstantPointerNull>(V)) &&
179          "TypeInfo must be a global variable or NULL");
180   return GV;
181 }
182 
183 /// getFCmpCondCode - Return the ISD condition code corresponding to
184 /// the given LLVM IR floating-point condition code.  This includes
185 /// consideration of global floating-point math flags.
186 ///
187 ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) {
188   switch (Pred) {
189   case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
190   case FCmpInst::FCMP_OEQ:   return ISD::SETOEQ;
191   case FCmpInst::FCMP_OGT:   return ISD::SETOGT;
192   case FCmpInst::FCMP_OGE:   return ISD::SETOGE;
193   case FCmpInst::FCMP_OLT:   return ISD::SETOLT;
194   case FCmpInst::FCMP_OLE:   return ISD::SETOLE;
195   case FCmpInst::FCMP_ONE:   return ISD::SETONE;
196   case FCmpInst::FCMP_ORD:   return ISD::SETO;
197   case FCmpInst::FCMP_UNO:   return ISD::SETUO;
198   case FCmpInst::FCMP_UEQ:   return ISD::SETUEQ;
199   case FCmpInst::FCMP_UGT:   return ISD::SETUGT;
200   case FCmpInst::FCMP_UGE:   return ISD::SETUGE;
201   case FCmpInst::FCMP_ULT:   return ISD::SETULT;
202   case FCmpInst::FCMP_ULE:   return ISD::SETULE;
203   case FCmpInst::FCMP_UNE:   return ISD::SETUNE;
204   case FCmpInst::FCMP_TRUE:  return ISD::SETTRUE;
205   default: llvm_unreachable("Invalid FCmp predicate opcode!");
206   }
207 }
208 
209 ISD::CondCode llvm::getFCmpCodeWithoutNaN(ISD::CondCode CC) {
210   switch (CC) {
211     case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
212     case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
213     case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
214     case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
215     case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
216     case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
217     default: return CC;
218   }
219 }
220 
221 ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) {
222   switch (Pred) {
223   case ICmpInst::ICMP_EQ:  return ISD::SETEQ;
224   case ICmpInst::ICMP_NE:  return ISD::SETNE;
225   case ICmpInst::ICMP_SLE: return ISD::SETLE;
226   case ICmpInst::ICMP_ULE: return ISD::SETULE;
227   case ICmpInst::ICMP_SGE: return ISD::SETGE;
228   case ICmpInst::ICMP_UGE: return ISD::SETUGE;
229   case ICmpInst::ICMP_SLT: return ISD::SETLT;
230   case ICmpInst::ICMP_ULT: return ISD::SETULT;
231   case ICmpInst::ICMP_SGT: return ISD::SETGT;
232   case ICmpInst::ICMP_UGT: return ISD::SETUGT;
233   default:
234     llvm_unreachable("Invalid ICmp predicate opcode!");
235   }
236 }
237 
238 ICmpInst::Predicate llvm::getICmpCondCode(ISD::CondCode Pred) {
239   switch (Pred) {
240   case ISD::SETEQ:
241     return ICmpInst::ICMP_EQ;
242   case ISD::SETNE:
243     return ICmpInst::ICMP_NE;
244   case ISD::SETLE:
245     return ICmpInst::ICMP_SLE;
246   case ISD::SETULE:
247     return ICmpInst::ICMP_ULE;
248   case ISD::SETGE:
249     return ICmpInst::ICMP_SGE;
250   case ISD::SETUGE:
251     return ICmpInst::ICMP_UGE;
252   case ISD::SETLT:
253     return ICmpInst::ICMP_SLT;
254   case ISD::SETULT:
255     return ICmpInst::ICMP_ULT;
256   case ISD::SETGT:
257     return ICmpInst::ICMP_SGT;
258   case ISD::SETUGT:
259     return ICmpInst::ICMP_UGT;
260   default:
261     llvm_unreachable("Invalid ISD integer condition code!");
262   }
263 }
264 
265 static bool isNoopBitcast(Type *T1, Type *T2,
266                           const TargetLoweringBase& TLI) {
267   return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
268          (isa<VectorType>(T1) && isa<VectorType>(T2) &&
269           TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
270 }
271 
272 /// Look through operations that will be free to find the earliest source of
273 /// this value.
274 ///
275 /// @param ValLoc If V has aggregate type, we will be interested in a particular
276 /// scalar component. This records its address; the reverse of this list gives a
277 /// sequence of indices appropriate for an extractvalue to locate the important
278 /// value. This value is updated during the function and on exit will indicate
279 /// similar information for the Value returned.
280 ///
281 /// @param DataBits If this function looks through truncate instructions, this
282 /// will record the smallest size attained.
283 static const Value *getNoopInput(const Value *V,
284                                  SmallVectorImpl<unsigned> &ValLoc,
285                                  unsigned &DataBits,
286                                  const TargetLoweringBase &TLI,
287                                  const DataLayout &DL) {
288   while (true) {
289     // Try to look through V1; if V1 is not an instruction, it can't be looked
290     // through.
291     const Instruction *I = dyn_cast<Instruction>(V);
292     if (!I || I->getNumOperands() == 0) return V;
293     const Value *NoopInput = nullptr;
294 
295     Value *Op = I->getOperand(0);
296     if (isa<BitCastInst>(I)) {
297       // Look through truly no-op bitcasts.
298       if (isNoopBitcast(Op->getType(), I->getType(), TLI))
299         NoopInput = Op;
300     } else if (isa<GetElementPtrInst>(I)) {
301       // Look through getelementptr
302       if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
303         NoopInput = Op;
304     } else if (isa<IntToPtrInst>(I)) {
305       // Look through inttoptr.
306       // Make sure this isn't a truncating or extending cast.  We could
307       // support this eventually, but don't bother for now.
308       if (!isa<VectorType>(I->getType()) &&
309           DL.getPointerSizeInBits() ==
310               cast<IntegerType>(Op->getType())->getBitWidth())
311         NoopInput = Op;
312     } else if (isa<PtrToIntInst>(I)) {
313       // Look through ptrtoint.
314       // Make sure this isn't a truncating or extending cast.  We could
315       // support this eventually, but don't bother for now.
316       if (!isa<VectorType>(I->getType()) &&
317           DL.getPointerSizeInBits() ==
318               cast<IntegerType>(I->getType())->getBitWidth())
319         NoopInput = Op;
320     } else if (isa<TruncInst>(I) &&
321                TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
322       DataBits =
323           std::min((uint64_t)DataBits,
324                    I->getType()->getPrimitiveSizeInBits().getFixedValue());
325       NoopInput = Op;
326     } else if (auto *CB = dyn_cast<CallBase>(I)) {
327       const Value *ReturnedOp = CB->getReturnedArgOperand();
328       if (ReturnedOp && isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI))
329         NoopInput = ReturnedOp;
330     } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
331       // Value may come from either the aggregate or the scalar
332       ArrayRef<unsigned> InsertLoc = IVI->getIndices();
333       if (ValLoc.size() >= InsertLoc.size() &&
334           std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) {
335         // The type being inserted is a nested sub-type of the aggregate; we
336         // have to remove those initial indices to get the location we're
337         // interested in for the operand.
338         ValLoc.resize(ValLoc.size() - InsertLoc.size());
339         NoopInput = IVI->getInsertedValueOperand();
340       } else {
341         // The struct we're inserting into has the value we're interested in, no
342         // change of address.
343         NoopInput = Op;
344       }
345     } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
346       // The part we're interested in will inevitably be some sub-section of the
347       // previous aggregate. Combine the two paths to obtain the true address of
348       // our element.
349       ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
350       ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
351       NoopInput = Op;
352     }
353     // Terminate if we couldn't find anything to look through.
354     if (!NoopInput)
355       return V;
356 
357     V = NoopInput;
358   }
359 }
360 
361 /// Return true if this scalar return value only has bits discarded on its path
362 /// from the "tail call" to the "ret". This includes the obvious noop
363 /// instructions handled by getNoopInput above as well as free truncations (or
364 /// extensions prior to the call).
365 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
366                                  SmallVectorImpl<unsigned> &RetIndices,
367                                  SmallVectorImpl<unsigned> &CallIndices,
368                                  bool AllowDifferingSizes,
369                                  const TargetLoweringBase &TLI,
370                                  const DataLayout &DL) {
371 
372   // Trace the sub-value needed by the return value as far back up the graph as
373   // possible, in the hope that it will intersect with the value produced by the
374   // call. In the simple case with no "returned" attribute, the hope is actually
375   // that we end up back at the tail call instruction itself.
376   unsigned BitsRequired = UINT_MAX;
377   RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL);
378 
379   // If this slot in the value returned is undef, it doesn't matter what the
380   // call puts there, it'll be fine.
381   if (isa<UndefValue>(RetVal))
382     return true;
383 
384   // Now do a similar search up through the graph to find where the value
385   // actually returned by the "tail call" comes from. In the simple case without
386   // a "returned" attribute, the search will be blocked immediately and the loop
387   // a Noop.
388   unsigned BitsProvided = UINT_MAX;
389   CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL);
390 
391   // There's no hope if we can't actually trace them to (the same part of!) the
392   // same value.
393   if (CallVal != RetVal || CallIndices != RetIndices)
394     return false;
395 
396   // However, intervening truncates may have made the call non-tail. Make sure
397   // all the bits that are needed by the "ret" have been provided by the "tail
398   // call". FIXME: with sufficiently cunning bit-tracking, we could look through
399   // extensions too.
400   if (BitsProvided < BitsRequired ||
401       (!AllowDifferingSizes && BitsProvided != BitsRequired))
402     return false;
403 
404   return true;
405 }
406 
407 /// For an aggregate type, determine whether a given index is within bounds or
408 /// not.
409 static bool indexReallyValid(Type *T, unsigned Idx) {
410   if (ArrayType *AT = dyn_cast<ArrayType>(T))
411     return Idx < AT->getNumElements();
412 
413   return Idx < cast<StructType>(T)->getNumElements();
414 }
415 
416 /// Move the given iterators to the next leaf type in depth first traversal.
417 ///
418 /// Performs a depth-first traversal of the type as specified by its arguments,
419 /// stopping at the next leaf node (which may be a legitimate scalar type or an
420 /// empty struct or array).
421 ///
422 /// @param SubTypes List of the partial components making up the type from
423 /// outermost to innermost non-empty aggregate. The element currently
424 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
425 ///
426 /// @param Path Set of extractvalue indices leading from the outermost type
427 /// (SubTypes[0]) to the leaf node currently represented.
428 ///
429 /// @returns true if a new type was found, false otherwise. Calling this
430 /// function again on a finished iterator will repeatedly return
431 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
432 /// aggregate or a non-aggregate
433 static bool advanceToNextLeafType(SmallVectorImpl<Type *> &SubTypes,
434                                   SmallVectorImpl<unsigned> &Path) {
435   // First march back up the tree until we can successfully increment one of the
436   // coordinates in Path.
437   while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
438     Path.pop_back();
439     SubTypes.pop_back();
440   }
441 
442   // If we reached the top, then the iterator is done.
443   if (Path.empty())
444     return false;
445 
446   // We know there's *some* valid leaf now, so march back down the tree picking
447   // out the left-most element at each node.
448   ++Path.back();
449   Type *DeeperType =
450       ExtractValueInst::getIndexedType(SubTypes.back(), Path.back());
451   while (DeeperType->isAggregateType()) {
452     if (!indexReallyValid(DeeperType, 0))
453       return true;
454 
455     SubTypes.push_back(DeeperType);
456     Path.push_back(0);
457 
458     DeeperType = ExtractValueInst::getIndexedType(DeeperType, 0);
459   }
460 
461   return true;
462 }
463 
464 /// Find the first non-empty, scalar-like type in Next and setup the iterator
465 /// components.
466 ///
467 /// Assuming Next is an aggregate of some kind, this function will traverse the
468 /// tree from left to right (i.e. depth-first) looking for the first
469 /// non-aggregate type which will play a role in function return.
470 ///
471 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
472 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
473 /// i32 in that type.
474 static bool firstRealType(Type *Next, SmallVectorImpl<Type *> &SubTypes,
475                           SmallVectorImpl<unsigned> &Path) {
476   // First initialise the iterator components to the first "leaf" node
477   // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
478   // despite nominally being an aggregate).
479   while (Type *FirstInner = ExtractValueInst::getIndexedType(Next, 0)) {
480     SubTypes.push_back(Next);
481     Path.push_back(0);
482     Next = FirstInner;
483   }
484 
485   // If there's no Path now, Next was originally scalar already (or empty
486   // leaf). We're done.
487   if (Path.empty())
488     return true;
489 
490   // Otherwise, use normal iteration to keep looking through the tree until we
491   // find a non-aggregate type.
492   while (ExtractValueInst::getIndexedType(SubTypes.back(), Path.back())
493              ->isAggregateType()) {
494     if (!advanceToNextLeafType(SubTypes, Path))
495       return false;
496   }
497 
498   return true;
499 }
500 
501 /// Set the iterator data-structures to the next non-empty, non-aggregate
502 /// subtype.
503 static bool nextRealType(SmallVectorImpl<Type *> &SubTypes,
504                          SmallVectorImpl<unsigned> &Path) {
505   do {
506     if (!advanceToNextLeafType(SubTypes, Path))
507       return false;
508 
509     assert(!Path.empty() && "found a leaf but didn't set the path?");
510   } while (ExtractValueInst::getIndexedType(SubTypes.back(), Path.back())
511                ->isAggregateType());
512 
513   return true;
514 }
515 
516 
517 /// Test if the given instruction is in a position to be optimized
518 /// with a tail-call. This roughly means that it's in a block with
519 /// a return and there's nothing that needs to be scheduled
520 /// between it and the return.
521 ///
522 /// This function only tests target-independent requirements.
523 bool llvm::isInTailCallPosition(const CallBase &Call, const TargetMachine &TM) {
524   const BasicBlock *ExitBB = Call.getParent();
525   const Instruction *Term = ExitBB->getTerminator();
526   const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
527 
528   // The block must end in a return statement or unreachable.
529   //
530   // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
531   // an unreachable, for now. The way tailcall optimization is currently
532   // implemented means it will add an epilogue followed by a jump. That is
533   // not profitable. Also, if the callee is a special function (e.g.
534   // longjmp on x86), it can end up causing miscompilation that has not
535   // been fully understood.
536   if (!Ret && ((!TM.Options.GuaranteedTailCallOpt &&
537                 Call.getCallingConv() != CallingConv::Tail &&
538                 Call.getCallingConv() != CallingConv::SwiftTail) ||
539                !isa<UnreachableInst>(Term)))
540     return false;
541 
542   // If I will have a chain, make sure no other instruction that will have a
543   // chain interposes between I and the return.
544   // Check for all calls including speculatable functions.
545   for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) {
546     if (&*BBI == &Call)
547       break;
548     // Debug info intrinsics do not get in the way of tail call optimization.
549     // Pseudo probe intrinsics do not block tail call optimization either.
550     if (BBI->isDebugOrPseudoInst())
551       continue;
552     // A lifetime end, assume or noalias.decl intrinsic should not stop tail
553     // call optimization.
554     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
555       if (II->getIntrinsicID() == Intrinsic::lifetime_end ||
556           II->getIntrinsicID() == Intrinsic::assume ||
557           II->getIntrinsicID() == Intrinsic::experimental_noalias_scope_decl)
558         continue;
559     if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
560         !isSafeToSpeculativelyExecute(&*BBI))
561       return false;
562   }
563 
564   const Function *F = ExitBB->getParent();
565   return returnTypeIsEligibleForTailCall(
566       F, &Call, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering());
567 }
568 
569 bool llvm::attributesPermitTailCall(const Function *F, const Instruction *I,
570                                     const ReturnInst *Ret,
571                                     const TargetLoweringBase &TLI,
572                                     bool *AllowDifferingSizes) {
573   // ADS may be null, so don't write to it directly.
574   bool DummyADS;
575   bool &ADS = AllowDifferingSizes ? *AllowDifferingSizes : DummyADS;
576   ADS = true;
577 
578   AttrBuilder CallerAttrs(F->getContext(), F->getAttributes().getRetAttrs());
579   AttrBuilder CalleeAttrs(F->getContext(),
580                           cast<CallInst>(I)->getAttributes().getRetAttrs());
581 
582   // Following attributes are completely benign as far as calling convention
583   // goes, they shouldn't affect whether the call is a tail call.
584   for (const auto &Attr : {Attribute::Alignment, Attribute::Dereferenceable,
585                            Attribute::DereferenceableOrNull, Attribute::NoAlias,
586                            Attribute::NonNull, Attribute::NoUndef}) {
587     CallerAttrs.removeAttribute(Attr);
588     CalleeAttrs.removeAttribute(Attr);
589   }
590 
591   if (CallerAttrs.contains(Attribute::ZExt)) {
592     if (!CalleeAttrs.contains(Attribute::ZExt))
593       return false;
594 
595     ADS = false;
596     CallerAttrs.removeAttribute(Attribute::ZExt);
597     CalleeAttrs.removeAttribute(Attribute::ZExt);
598   } else if (CallerAttrs.contains(Attribute::SExt)) {
599     if (!CalleeAttrs.contains(Attribute::SExt))
600       return false;
601 
602     ADS = false;
603     CallerAttrs.removeAttribute(Attribute::SExt);
604     CalleeAttrs.removeAttribute(Attribute::SExt);
605   }
606 
607   // Drop sext and zext return attributes if the result is not used.
608   // This enables tail calls for code like:
609   //
610   // define void @caller() {
611   // entry:
612   //   %unused_result = tail call zeroext i1 @callee()
613   //   br label %retlabel
614   // retlabel:
615   //   ret void
616   // }
617   if (I->use_empty()) {
618     CalleeAttrs.removeAttribute(Attribute::SExt);
619     CalleeAttrs.removeAttribute(Attribute::ZExt);
620   }
621 
622   // If they're still different, there's some facet we don't understand
623   // (currently only "inreg", but in future who knows). It may be OK but the
624   // only safe option is to reject the tail call.
625   return CallerAttrs == CalleeAttrs;
626 }
627 
628 /// Check whether B is a bitcast of a pointer type to another pointer type,
629 /// which is equal to A.
630 static bool isPointerBitcastEqualTo(const Value *A, const Value *B) {
631   assert(A && B && "Expected non-null inputs!");
632 
633   auto *BitCastIn = dyn_cast<BitCastInst>(B);
634 
635   if (!BitCastIn)
636     return false;
637 
638   if (!A->getType()->isPointerTy() || !B->getType()->isPointerTy())
639     return false;
640 
641   return A == BitCastIn->getOperand(0);
642 }
643 
644 bool llvm::returnTypeIsEligibleForTailCall(const Function *F,
645                                            const Instruction *I,
646                                            const ReturnInst *Ret,
647                                            const TargetLoweringBase &TLI) {
648   // If the block ends with a void return or unreachable, it doesn't matter
649   // what the call's return type is.
650   if (!Ret || Ret->getNumOperands() == 0) return true;
651 
652   // If the return value is undef, it doesn't matter what the call's
653   // return type is.
654   if (isa<UndefValue>(Ret->getOperand(0))) return true;
655 
656   // Make sure the attributes attached to each return are compatible.
657   bool AllowDifferingSizes;
658   if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
659     return false;
660 
661   const Value *RetVal = Ret->getOperand(0), *CallVal = I;
662   // Intrinsic like llvm.memcpy has no return value, but the expanded
663   // libcall may or may not have return value. On most platforms, it
664   // will be expanded as memcpy in libc, which returns the first
665   // argument. On other platforms like arm-none-eabi, memcpy may be
666   // expanded as library call without return value, like __aeabi_memcpy.
667   const CallInst *Call = cast<CallInst>(I);
668   if (Function *F = Call->getCalledFunction()) {
669     Intrinsic::ID IID = F->getIntrinsicID();
670     if (((IID == Intrinsic::memcpy &&
671           TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")) ||
672          (IID == Intrinsic::memmove &&
673           TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")) ||
674          (IID == Intrinsic::memset &&
675           TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset"))) &&
676         (RetVal == Call->getArgOperand(0) ||
677          isPointerBitcastEqualTo(RetVal, Call->getArgOperand(0))))
678       return true;
679   }
680 
681   SmallVector<unsigned, 4> RetPath, CallPath;
682   SmallVector<Type *, 4> RetSubTypes, CallSubTypes;
683 
684   bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
685   bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
686 
687   // Nothing's actually returned, it doesn't matter what the callee put there
688   // it's a valid tail call.
689   if (RetEmpty)
690     return true;
691 
692   // Iterate pairwise through each of the value types making up the tail call
693   // and the corresponding return. For each one we want to know whether it's
694   // essentially going directly from the tail call to the ret, via operations
695   // that end up not generating any code.
696   //
697   // We allow a certain amount of covariance here. For example it's permitted
698   // for the tail call to define more bits than the ret actually cares about
699   // (e.g. via a truncate).
700   do {
701     if (CallEmpty) {
702       // We've exhausted the values produced by the tail call instruction, the
703       // rest are essentially undef. The type doesn't really matter, but we need
704       // *something*.
705       Type *SlotType =
706           ExtractValueInst::getIndexedType(RetSubTypes.back(), RetPath.back());
707       CallVal = UndefValue::get(SlotType);
708     }
709 
710     // The manipulations performed when we're looking through an insertvalue or
711     // an extractvalue would happen at the front of the RetPath list, so since
712     // we have to copy it anyway it's more efficient to create a reversed copy.
713     SmallVector<unsigned, 4> TmpRetPath(llvm::reverse(RetPath));
714     SmallVector<unsigned, 4> TmpCallPath(llvm::reverse(CallPath));
715 
716     // Finally, we can check whether the value produced by the tail call at this
717     // index is compatible with the value we return.
718     if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
719                               AllowDifferingSizes, TLI,
720                               F->getParent()->getDataLayout()))
721       return false;
722 
723     CallEmpty  = !nextRealType(CallSubTypes, CallPath);
724   } while(nextRealType(RetSubTypes, RetPath));
725 
726   return true;
727 }
728 
729 static void collectEHScopeMembers(
730     DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope,
731     const MachineBasicBlock *MBB) {
732   SmallVector<const MachineBasicBlock *, 16> Worklist = {MBB};
733   while (!Worklist.empty()) {
734     const MachineBasicBlock *Visiting = Worklist.pop_back_val();
735     // Don't follow blocks which start new scopes.
736     if (Visiting->isEHPad() && Visiting != MBB)
737       continue;
738 
739     // Add this MBB to our scope.
740     auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope));
741 
742     // Don't revisit blocks.
743     if (!P.second) {
744       assert(P.first->second == EHScope && "MBB is part of two scopes!");
745       continue;
746     }
747 
748     // Returns are boundaries where scope transfer can occur, don't follow
749     // successors.
750     if (Visiting->isEHScopeReturnBlock())
751       continue;
752 
753     append_range(Worklist, Visiting->successors());
754   }
755 }
756 
757 DenseMap<const MachineBasicBlock *, int>
758 llvm::getEHScopeMembership(const MachineFunction &MF) {
759   DenseMap<const MachineBasicBlock *, int> EHScopeMembership;
760 
761   // We don't have anything to do if there aren't any EH pads.
762   if (!MF.hasEHScopes())
763     return EHScopeMembership;
764 
765   int EntryBBNumber = MF.front().getNumber();
766   bool IsSEH = isAsynchronousEHPersonality(
767       classifyEHPersonality(MF.getFunction().getPersonalityFn()));
768 
769   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
770   SmallVector<const MachineBasicBlock *, 16> EHScopeBlocks;
771   SmallVector<const MachineBasicBlock *, 16> UnreachableBlocks;
772   SmallVector<const MachineBasicBlock *, 16> SEHCatchPads;
773   SmallVector<std::pair<const MachineBasicBlock *, int>, 16> CatchRetSuccessors;
774   for (const MachineBasicBlock &MBB : MF) {
775     if (MBB.isEHScopeEntry()) {
776       EHScopeBlocks.push_back(&MBB);
777     } else if (IsSEH && MBB.isEHPad()) {
778       SEHCatchPads.push_back(&MBB);
779     } else if (MBB.pred_empty()) {
780       UnreachableBlocks.push_back(&MBB);
781     }
782 
783     MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator();
784 
785     // CatchPads are not scopes for SEH so do not consider CatchRet to
786     // transfer control to another scope.
787     if (MBBI == MBB.end() || MBBI->getOpcode() != TII->getCatchReturnOpcode())
788       continue;
789 
790     // FIXME: SEH CatchPads are not necessarily in the parent function:
791     // they could be inside a finally block.
792     const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB();
793     const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB();
794     CatchRetSuccessors.push_back(
795         {Successor, IsSEH ? EntryBBNumber : SuccessorColor->getNumber()});
796   }
797 
798   // We don't have anything to do if there aren't any EH pads.
799   if (EHScopeBlocks.empty())
800     return EHScopeMembership;
801 
802   // Identify all the basic blocks reachable from the function entry.
803   collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front());
804   // All blocks not part of a scope are in the parent function.
805   for (const MachineBasicBlock *MBB : UnreachableBlocks)
806     collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
807   // Next, identify all the blocks inside the scopes.
808   for (const MachineBasicBlock *MBB : EHScopeBlocks)
809     collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB);
810   // SEH CatchPads aren't really scopes, handle them separately.
811   for (const MachineBasicBlock *MBB : SEHCatchPads)
812     collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
813   // Finally, identify all the targets of a catchret.
814   for (std::pair<const MachineBasicBlock *, int> CatchRetPair :
815        CatchRetSuccessors)
816     collectEHScopeMembers(EHScopeMembership, CatchRetPair.second,
817                           CatchRetPair.first);
818   return EHScopeMembership;
819 }
820