xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/Analysis.cpp (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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/LLVMContext.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/MathExtras.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Transforms/Utils/GlobalStatus.h"
30 
31 using namespace llvm;
32 
33 /// Compute the linearized index of a member in a nested aggregate/struct/array
34 /// by recursing and accumulating CurIndex as long as there are indices in the
35 /// index list.
36 unsigned llvm::ComputeLinearIndex(Type *Ty,
37                                   const unsigned *Indices,
38                                   const unsigned *IndicesEnd,
39                                   unsigned CurIndex) {
40   // Base case: We're done.
41   if (Indices && Indices == IndicesEnd)
42     return CurIndex;
43 
44   // Given a struct type, recursively traverse the elements.
45   if (StructType *STy = dyn_cast<StructType>(Ty)) {
46     for (StructType::element_iterator EB = STy->element_begin(),
47                                       EI = EB,
48                                       EE = STy->element_end();
49         EI != EE; ++EI) {
50       if (Indices && *Indices == unsigned(EI - EB))
51         return ComputeLinearIndex(*EI, Indices+1, IndicesEnd, CurIndex);
52       CurIndex = ComputeLinearIndex(*EI, nullptr, nullptr, CurIndex);
53     }
54     assert(!Indices && "Unexpected out of bound");
55     return CurIndex;
56   }
57   // Given an array type, recursively traverse the elements.
58   else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
59     Type *EltTy = ATy->getElementType();
60     unsigned NumElts = ATy->getNumElements();
61     // Compute the Linear offset when jumping one element of the array
62     unsigned EltLinearOffset = ComputeLinearIndex(EltTy, nullptr, nullptr, 0);
63     if (Indices) {
64       assert(*Indices < NumElts && "Unexpected out of bound");
65       // If the indice is inside the array, compute the index to the requested
66       // elt and recurse inside the element with the end of the indices list
67       CurIndex += EltLinearOffset* *Indices;
68       return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex);
69     }
70     CurIndex += EltLinearOffset*NumElts;
71     return CurIndex;
72   }
73   // We haven't found the type we're looking for, so keep searching.
74   return CurIndex + 1;
75 }
76 
77 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
78 /// EVTs that represent all the individual underlying
79 /// non-aggregate types that comprise it.
80 ///
81 /// If Offsets is non-null, it points to a vector to be filled in
82 /// with the in-memory offsets of each of the individual values.
83 ///
84 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
85                            Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
86                            SmallVectorImpl<EVT> *MemVTs,
87                            SmallVectorImpl<uint64_t> *Offsets,
88                            uint64_t StartingOffset) {
89   // Given a struct type, recursively traverse the elements.
90   if (StructType *STy = dyn_cast<StructType>(Ty)) {
91     const StructLayout *SL = DL.getStructLayout(STy);
92     for (StructType::element_iterator EB = STy->element_begin(),
93                                       EI = EB,
94                                       EE = STy->element_end();
95          EI != EE; ++EI)
96       ComputeValueVTs(TLI, DL, *EI, ValueVTs, MemVTs, Offsets,
97                       StartingOffset + SL->getElementOffset(EI - EB));
98     return;
99   }
100   // Given an array type, recursively traverse the elements.
101   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
102     Type *EltTy = ATy->getElementType();
103     uint64_t EltSize = DL.getTypeAllocSize(EltTy);
104     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
105       ComputeValueVTs(TLI, DL, EltTy, ValueVTs, MemVTs, Offsets,
106                       StartingOffset + i * EltSize);
107     return;
108   }
109   // Interpret void as zero return values.
110   if (Ty->isVoidTy())
111     return;
112   // Base case: we can get an EVT for this LLVM IR type.
113   ValueVTs.push_back(TLI.getValueType(DL, Ty));
114   if (MemVTs)
115     MemVTs->push_back(TLI.getMemValueType(DL, Ty));
116   if (Offsets)
117     Offsets->push_back(StartingOffset);
118 }
119 
120 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
121                            Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
122                            SmallVectorImpl<uint64_t> *Offsets,
123                            uint64_t StartingOffset) {
124   return ComputeValueVTs(TLI, DL, Ty, ValueVTs, /*MemVTs=*/nullptr, Offsets,
125                          StartingOffset);
126 }
127 
128 void llvm::computeValueLLTs(const DataLayout &DL, Type &Ty,
129                             SmallVectorImpl<LLT> &ValueTys,
130                             SmallVectorImpl<uint64_t> *Offsets,
131                             uint64_t StartingOffset) {
132   // Given a struct type, recursively traverse the elements.
133   if (StructType *STy = dyn_cast<StructType>(&Ty)) {
134     const StructLayout *SL = DL.getStructLayout(STy);
135     for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I)
136       computeValueLLTs(DL, *STy->getElementType(I), ValueTys, Offsets,
137                        StartingOffset + SL->getElementOffset(I));
138     return;
139   }
140   // Given an array type, recursively traverse the elements.
141   if (ArrayType *ATy = dyn_cast<ArrayType>(&Ty)) {
142     Type *EltTy = ATy->getElementType();
143     uint64_t EltSize = DL.getTypeAllocSize(EltTy);
144     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
145       computeValueLLTs(DL, *EltTy, ValueTys, Offsets,
146                        StartingOffset + i * EltSize);
147     return;
148   }
149   // Interpret void as zero return values.
150   if (Ty.isVoidTy())
151     return;
152   // Base case: we can get an LLT for this LLVM IR type.
153   ValueTys.push_back(getLLTForType(Ty, DL));
154   if (Offsets != nullptr)
155     Offsets->push_back(StartingOffset * 8);
156 }
157 
158 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
159 GlobalValue *llvm::ExtractTypeInfo(Value *V) {
160   V = V->stripPointerCasts();
161   GlobalValue *GV = dyn_cast<GlobalValue>(V);
162   GlobalVariable *Var = dyn_cast<GlobalVariable>(V);
163 
164   if (Var && Var->getName() == "llvm.eh.catch.all.value") {
165     assert(Var->hasInitializer() &&
166            "The EH catch-all value must have an initializer");
167     Value *Init = Var->getInitializer();
168     GV = dyn_cast<GlobalValue>(Init);
169     if (!GV) V = cast<ConstantPointerNull>(Init);
170   }
171 
172   assert((GV || isa<ConstantPointerNull>(V)) &&
173          "TypeInfo must be a global variable or NULL");
174   return GV;
175 }
176 
177 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
178 /// processed uses a memory 'm' constraint.
179 bool
180 llvm::hasInlineAsmMemConstraint(InlineAsm::ConstraintInfoVector &CInfos,
181                                 const TargetLowering &TLI) {
182   for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
183     InlineAsm::ConstraintInfo &CI = CInfos[i];
184     for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
185       TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
186       if (CType == TargetLowering::C_Memory)
187         return true;
188     }
189 
190     // Indirect operand accesses access memory.
191     if (CI.isIndirect)
192       return true;
193   }
194 
195   return false;
196 }
197 
198 /// getFCmpCondCode - Return the ISD condition code corresponding to
199 /// the given LLVM IR floating-point condition code.  This includes
200 /// consideration of global floating-point math flags.
201 ///
202 ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) {
203   switch (Pred) {
204   case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
205   case FCmpInst::FCMP_OEQ:   return ISD::SETOEQ;
206   case FCmpInst::FCMP_OGT:   return ISD::SETOGT;
207   case FCmpInst::FCMP_OGE:   return ISD::SETOGE;
208   case FCmpInst::FCMP_OLT:   return ISD::SETOLT;
209   case FCmpInst::FCMP_OLE:   return ISD::SETOLE;
210   case FCmpInst::FCMP_ONE:   return ISD::SETONE;
211   case FCmpInst::FCMP_ORD:   return ISD::SETO;
212   case FCmpInst::FCMP_UNO:   return ISD::SETUO;
213   case FCmpInst::FCMP_UEQ:   return ISD::SETUEQ;
214   case FCmpInst::FCMP_UGT:   return ISD::SETUGT;
215   case FCmpInst::FCMP_UGE:   return ISD::SETUGE;
216   case FCmpInst::FCMP_ULT:   return ISD::SETULT;
217   case FCmpInst::FCMP_ULE:   return ISD::SETULE;
218   case FCmpInst::FCMP_UNE:   return ISD::SETUNE;
219   case FCmpInst::FCMP_TRUE:  return ISD::SETTRUE;
220   default: llvm_unreachable("Invalid FCmp predicate opcode!");
221   }
222 }
223 
224 ISD::CondCode llvm::getFCmpCodeWithoutNaN(ISD::CondCode CC) {
225   switch (CC) {
226     case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
227     case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
228     case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
229     case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
230     case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
231     case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
232     default: return CC;
233   }
234 }
235 
236 /// getICmpCondCode - Return the ISD condition code corresponding to
237 /// the given LLVM IR integer condition code.
238 ///
239 ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) {
240   switch (Pred) {
241   case ICmpInst::ICMP_EQ:  return ISD::SETEQ;
242   case ICmpInst::ICMP_NE:  return ISD::SETNE;
243   case ICmpInst::ICMP_SLE: return ISD::SETLE;
244   case ICmpInst::ICMP_ULE: return ISD::SETULE;
245   case ICmpInst::ICMP_SGE: return ISD::SETGE;
246   case ICmpInst::ICMP_UGE: return ISD::SETUGE;
247   case ICmpInst::ICMP_SLT: return ISD::SETLT;
248   case ICmpInst::ICMP_ULT: return ISD::SETULT;
249   case ICmpInst::ICMP_SGT: return ISD::SETGT;
250   case ICmpInst::ICMP_UGT: return ISD::SETUGT;
251   default:
252     llvm_unreachable("Invalid ICmp predicate opcode!");
253   }
254 }
255 
256 static bool isNoopBitcast(Type *T1, Type *T2,
257                           const TargetLoweringBase& TLI) {
258   return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
259          (isa<VectorType>(T1) && isa<VectorType>(T2) &&
260           TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
261 }
262 
263 /// Look through operations that will be free to find the earliest source of
264 /// this value.
265 ///
266 /// @param ValLoc If V has aggregate type, we will be interested in a particular
267 /// scalar component. This records its address; the reverse of this list gives a
268 /// sequence of indices appropriate for an extractvalue to locate the important
269 /// value. This value is updated during the function and on exit will indicate
270 /// similar information for the Value returned.
271 ///
272 /// @param DataBits If this function looks through truncate instructions, this
273 /// will record the smallest size attained.
274 static const Value *getNoopInput(const Value *V,
275                                  SmallVectorImpl<unsigned> &ValLoc,
276                                  unsigned &DataBits,
277                                  const TargetLoweringBase &TLI,
278                                  const DataLayout &DL) {
279   while (true) {
280     // Try to look through V1; if V1 is not an instruction, it can't be looked
281     // through.
282     const Instruction *I = dyn_cast<Instruction>(V);
283     if (!I || I->getNumOperands() == 0) return V;
284     const Value *NoopInput = nullptr;
285 
286     Value *Op = I->getOperand(0);
287     if (isa<BitCastInst>(I)) {
288       // Look through truly no-op bitcasts.
289       if (isNoopBitcast(Op->getType(), I->getType(), TLI))
290         NoopInput = Op;
291     } else if (isa<GetElementPtrInst>(I)) {
292       // Look through getelementptr
293       if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
294         NoopInput = Op;
295     } else if (isa<IntToPtrInst>(I)) {
296       // Look through inttoptr.
297       // Make sure this isn't a truncating or extending cast.  We could
298       // support this eventually, but don't bother for now.
299       if (!isa<VectorType>(I->getType()) &&
300           DL.getPointerSizeInBits() ==
301               cast<IntegerType>(Op->getType())->getBitWidth())
302         NoopInput = Op;
303     } else if (isa<PtrToIntInst>(I)) {
304       // Look through ptrtoint.
305       // Make sure this isn't a truncating or extending cast.  We could
306       // support this eventually, but don't bother for now.
307       if (!isa<VectorType>(I->getType()) &&
308           DL.getPointerSizeInBits() ==
309               cast<IntegerType>(I->getType())->getBitWidth())
310         NoopInput = Op;
311     } else if (isa<TruncInst>(I) &&
312                TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
313       DataBits = std::min((uint64_t)DataBits,
314                          I->getType()->getPrimitiveSizeInBits().getFixedSize());
315       NoopInput = Op;
316     } else if (auto *CB = dyn_cast<CallBase>(I)) {
317       const Value *ReturnedOp = CB->getReturnedArgOperand();
318       if (ReturnedOp && isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI))
319         NoopInput = ReturnedOp;
320     } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
321       // Value may come from either the aggregate or the scalar
322       ArrayRef<unsigned> InsertLoc = IVI->getIndices();
323       if (ValLoc.size() >= InsertLoc.size() &&
324           std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) {
325         // The type being inserted is a nested sub-type of the aggregate; we
326         // have to remove those initial indices to get the location we're
327         // interested in for the operand.
328         ValLoc.resize(ValLoc.size() - InsertLoc.size());
329         NoopInput = IVI->getInsertedValueOperand();
330       } else {
331         // The struct we're inserting into has the value we're interested in, no
332         // change of address.
333         NoopInput = Op;
334       }
335     } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
336       // The part we're interested in will inevitably be some sub-section of the
337       // previous aggregate. Combine the two paths to obtain the true address of
338       // our element.
339       ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
340       ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
341       NoopInput = Op;
342     }
343     // Terminate if we couldn't find anything to look through.
344     if (!NoopInput)
345       return V;
346 
347     V = NoopInput;
348   }
349 }
350 
351 /// Return true if this scalar return value only has bits discarded on its path
352 /// from the "tail call" to the "ret". This includes the obvious noop
353 /// instructions handled by getNoopInput above as well as free truncations (or
354 /// extensions prior to the call).
355 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
356                                  SmallVectorImpl<unsigned> &RetIndices,
357                                  SmallVectorImpl<unsigned> &CallIndices,
358                                  bool AllowDifferingSizes,
359                                  const TargetLoweringBase &TLI,
360                                  const DataLayout &DL) {
361 
362   // Trace the sub-value needed by the return value as far back up the graph as
363   // possible, in the hope that it will intersect with the value produced by the
364   // call. In the simple case with no "returned" attribute, the hope is actually
365   // that we end up back at the tail call instruction itself.
366   unsigned BitsRequired = UINT_MAX;
367   RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL);
368 
369   // If this slot in the value returned is undef, it doesn't matter what the
370   // call puts there, it'll be fine.
371   if (isa<UndefValue>(RetVal))
372     return true;
373 
374   // Now do a similar search up through the graph to find where the value
375   // actually returned by the "tail call" comes from. In the simple case without
376   // a "returned" attribute, the search will be blocked immediately and the loop
377   // a Noop.
378   unsigned BitsProvided = UINT_MAX;
379   CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL);
380 
381   // There's no hope if we can't actually trace them to (the same part of!) the
382   // same value.
383   if (CallVal != RetVal || CallIndices != RetIndices)
384     return false;
385 
386   // However, intervening truncates may have made the call non-tail. Make sure
387   // all the bits that are needed by the "ret" have been provided by the "tail
388   // call". FIXME: with sufficiently cunning bit-tracking, we could look through
389   // extensions too.
390   if (BitsProvided < BitsRequired ||
391       (!AllowDifferingSizes && BitsProvided != BitsRequired))
392     return false;
393 
394   return true;
395 }
396 
397 /// For an aggregate type, determine whether a given index is within bounds or
398 /// not.
399 static bool indexReallyValid(Type *T, unsigned Idx) {
400   if (ArrayType *AT = dyn_cast<ArrayType>(T))
401     return Idx < AT->getNumElements();
402 
403   return Idx < cast<StructType>(T)->getNumElements();
404 }
405 
406 /// Move the given iterators to the next leaf type in depth first traversal.
407 ///
408 /// Performs a depth-first traversal of the type as specified by its arguments,
409 /// stopping at the next leaf node (which may be a legitimate scalar type or an
410 /// empty struct or array).
411 ///
412 /// @param SubTypes List of the partial components making up the type from
413 /// outermost to innermost non-empty aggregate. The element currently
414 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
415 ///
416 /// @param Path Set of extractvalue indices leading from the outermost type
417 /// (SubTypes[0]) to the leaf node currently represented.
418 ///
419 /// @returns true if a new type was found, false otherwise. Calling this
420 /// function again on a finished iterator will repeatedly return
421 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
422 /// aggregate or a non-aggregate
423 static bool advanceToNextLeafType(SmallVectorImpl<Type *> &SubTypes,
424                                   SmallVectorImpl<unsigned> &Path) {
425   // First march back up the tree until we can successfully increment one of the
426   // coordinates in Path.
427   while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
428     Path.pop_back();
429     SubTypes.pop_back();
430   }
431 
432   // If we reached the top, then the iterator is done.
433   if (Path.empty())
434     return false;
435 
436   // We know there's *some* valid leaf now, so march back down the tree picking
437   // out the left-most element at each node.
438   ++Path.back();
439   Type *DeeperType =
440       ExtractValueInst::getIndexedType(SubTypes.back(), Path.back());
441   while (DeeperType->isAggregateType()) {
442     if (!indexReallyValid(DeeperType, 0))
443       return true;
444 
445     SubTypes.push_back(DeeperType);
446     Path.push_back(0);
447 
448     DeeperType = ExtractValueInst::getIndexedType(DeeperType, 0);
449   }
450 
451   return true;
452 }
453 
454 /// Find the first non-empty, scalar-like type in Next and setup the iterator
455 /// components.
456 ///
457 /// Assuming Next is an aggregate of some kind, this function will traverse the
458 /// tree from left to right (i.e. depth-first) looking for the first
459 /// non-aggregate type which will play a role in function return.
460 ///
461 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
462 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
463 /// i32 in that type.
464 static bool firstRealType(Type *Next, SmallVectorImpl<Type *> &SubTypes,
465                           SmallVectorImpl<unsigned> &Path) {
466   // First initialise the iterator components to the first "leaf" node
467   // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
468   // despite nominally being an aggregate).
469   while (Type *FirstInner = ExtractValueInst::getIndexedType(Next, 0)) {
470     SubTypes.push_back(Next);
471     Path.push_back(0);
472     Next = FirstInner;
473   }
474 
475   // If there's no Path now, Next was originally scalar already (or empty
476   // leaf). We're done.
477   if (Path.empty())
478     return true;
479 
480   // Otherwise, use normal iteration to keep looking through the tree until we
481   // find a non-aggregate type.
482   while (ExtractValueInst::getIndexedType(SubTypes.back(), Path.back())
483              ->isAggregateType()) {
484     if (!advanceToNextLeafType(SubTypes, Path))
485       return false;
486   }
487 
488   return true;
489 }
490 
491 /// Set the iterator data-structures to the next non-empty, non-aggregate
492 /// subtype.
493 static bool nextRealType(SmallVectorImpl<Type *> &SubTypes,
494                          SmallVectorImpl<unsigned> &Path) {
495   do {
496     if (!advanceToNextLeafType(SubTypes, Path))
497       return false;
498 
499     assert(!Path.empty() && "found a leaf but didn't set the path?");
500   } while (ExtractValueInst::getIndexedType(SubTypes.back(), Path.back())
501                ->isAggregateType());
502 
503   return true;
504 }
505 
506 
507 /// Test if the given instruction is in a position to be optimized
508 /// with a tail-call. This roughly means that it's in a block with
509 /// a return and there's nothing that needs to be scheduled
510 /// between it and the return.
511 ///
512 /// This function only tests target-independent requirements.
513 bool llvm::isInTailCallPosition(const CallBase &Call, const TargetMachine &TM) {
514   const BasicBlock *ExitBB = Call.getParent();
515   const Instruction *Term = ExitBB->getTerminator();
516   const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
517 
518   // The block must end in a return statement or unreachable.
519   //
520   // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
521   // an unreachable, for now. The way tailcall optimization is currently
522   // implemented means it will add an epilogue followed by a jump. That is
523   // not profitable. Also, if the callee is a special function (e.g.
524   // longjmp on x86), it can end up causing miscompilation that has not
525   // been fully understood.
526   if (!Ret &&
527       ((!TM.Options.GuaranteedTailCallOpt &&
528         Call.getCallingConv() != CallingConv::Tail) || !isa<UnreachableInst>(Term)))
529     return false;
530 
531   // If I will have a chain, make sure no other instruction that will have a
532   // chain interposes between I and the return.
533   // Check for all calls including speculatable functions.
534   for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) {
535     if (&*BBI == &Call)
536       break;
537     // Debug info intrinsics do not get in the way of tail call optimization.
538     if (isa<DbgInfoIntrinsic>(BBI))
539       continue;
540     // A lifetime end or assume intrinsic should not stop tail call
541     // optimization.
542     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
543       if (II->getIntrinsicID() == Intrinsic::lifetime_end ||
544           II->getIntrinsicID() == Intrinsic::assume)
545         continue;
546     if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
547         !isSafeToSpeculativelyExecute(&*BBI))
548       return false;
549   }
550 
551   const Function *F = ExitBB->getParent();
552   return returnTypeIsEligibleForTailCall(
553       F, &Call, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering());
554 }
555 
556 bool llvm::attributesPermitTailCall(const Function *F, const Instruction *I,
557                                     const ReturnInst *Ret,
558                                     const TargetLoweringBase &TLI,
559                                     bool *AllowDifferingSizes) {
560   // ADS may be null, so don't write to it directly.
561   bool DummyADS;
562   bool &ADS = AllowDifferingSizes ? *AllowDifferingSizes : DummyADS;
563   ADS = true;
564 
565   AttrBuilder CallerAttrs(F->getAttributes(), AttributeList::ReturnIndex);
566   AttrBuilder CalleeAttrs(cast<CallInst>(I)->getAttributes(),
567                           AttributeList::ReturnIndex);
568 
569   // Following attributes are completely benign as far as calling convention
570   // goes, they shouldn't affect whether the call is a tail call.
571   CallerAttrs.removeAttribute(Attribute::NoAlias);
572   CalleeAttrs.removeAttribute(Attribute::NoAlias);
573   CallerAttrs.removeAttribute(Attribute::NonNull);
574   CalleeAttrs.removeAttribute(Attribute::NonNull);
575   CallerAttrs.removeAttribute(Attribute::Dereferenceable);
576   CalleeAttrs.removeAttribute(Attribute::Dereferenceable);
577   CallerAttrs.removeAttribute(Attribute::DereferenceableOrNull);
578   CalleeAttrs.removeAttribute(Attribute::DereferenceableOrNull);
579 
580   if (CallerAttrs.contains(Attribute::ZExt)) {
581     if (!CalleeAttrs.contains(Attribute::ZExt))
582       return false;
583 
584     ADS = false;
585     CallerAttrs.removeAttribute(Attribute::ZExt);
586     CalleeAttrs.removeAttribute(Attribute::ZExt);
587   } else if (CallerAttrs.contains(Attribute::SExt)) {
588     if (!CalleeAttrs.contains(Attribute::SExt))
589       return false;
590 
591     ADS = false;
592     CallerAttrs.removeAttribute(Attribute::SExt);
593     CalleeAttrs.removeAttribute(Attribute::SExt);
594   }
595 
596   // Drop sext and zext return attributes if the result is not used.
597   // This enables tail calls for code like:
598   //
599   // define void @caller() {
600   // entry:
601   //   %unused_result = tail call zeroext i1 @callee()
602   //   br label %retlabel
603   // retlabel:
604   //   ret void
605   // }
606   if (I->use_empty()) {
607     CalleeAttrs.removeAttribute(Attribute::SExt);
608     CalleeAttrs.removeAttribute(Attribute::ZExt);
609   }
610 
611   // If they're still different, there's some facet we don't understand
612   // (currently only "inreg", but in future who knows). It may be OK but the
613   // only safe option is to reject the tail call.
614   return CallerAttrs == CalleeAttrs;
615 }
616 
617 /// Check whether B is a bitcast of a pointer type to another pointer type,
618 /// which is equal to A.
619 static bool isPointerBitcastEqualTo(const Value *A, const Value *B) {
620   assert(A && B && "Expected non-null inputs!");
621 
622   auto *BitCastIn = dyn_cast<BitCastInst>(B);
623 
624   if (!BitCastIn)
625     return false;
626 
627   if (!A->getType()->isPointerTy() || !B->getType()->isPointerTy())
628     return false;
629 
630   return A == BitCastIn->getOperand(0);
631 }
632 
633 bool llvm::returnTypeIsEligibleForTailCall(const Function *F,
634                                            const Instruction *I,
635                                            const ReturnInst *Ret,
636                                            const TargetLoweringBase &TLI) {
637   // If the block ends with a void return or unreachable, it doesn't matter
638   // what the call's return type is.
639   if (!Ret || Ret->getNumOperands() == 0) return true;
640 
641   // If the return value is undef, it doesn't matter what the call's
642   // return type is.
643   if (isa<UndefValue>(Ret->getOperand(0))) return true;
644 
645   // Make sure the attributes attached to each return are compatible.
646   bool AllowDifferingSizes;
647   if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
648     return false;
649 
650   const Value *RetVal = Ret->getOperand(0), *CallVal = I;
651   // Intrinsic like llvm.memcpy has no return value, but the expanded
652   // libcall may or may not have return value. On most platforms, it
653   // will be expanded as memcpy in libc, which returns the first
654   // argument. On other platforms like arm-none-eabi, memcpy may be
655   // expanded as library call without return value, like __aeabi_memcpy.
656   const CallInst *Call = cast<CallInst>(I);
657   if (Function *F = Call->getCalledFunction()) {
658     Intrinsic::ID IID = F->getIntrinsicID();
659     if (((IID == Intrinsic::memcpy &&
660           TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")) ||
661          (IID == Intrinsic::memmove &&
662           TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")) ||
663          (IID == Intrinsic::memset &&
664           TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset"))) &&
665         (RetVal == Call->getArgOperand(0) ||
666          isPointerBitcastEqualTo(RetVal, Call->getArgOperand(0))))
667       return true;
668   }
669 
670   SmallVector<unsigned, 4> RetPath, CallPath;
671   SmallVector<Type *, 4> RetSubTypes, CallSubTypes;
672 
673   bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
674   bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
675 
676   // Nothing's actually returned, it doesn't matter what the callee put there
677   // it's a valid tail call.
678   if (RetEmpty)
679     return true;
680 
681   // Iterate pairwise through each of the value types making up the tail call
682   // and the corresponding return. For each one we want to know whether it's
683   // essentially going directly from the tail call to the ret, via operations
684   // that end up not generating any code.
685   //
686   // We allow a certain amount of covariance here. For example it's permitted
687   // for the tail call to define more bits than the ret actually cares about
688   // (e.g. via a truncate).
689   do {
690     if (CallEmpty) {
691       // We've exhausted the values produced by the tail call instruction, the
692       // rest are essentially undef. The type doesn't really matter, but we need
693       // *something*.
694       Type *SlotType =
695           ExtractValueInst::getIndexedType(RetSubTypes.back(), RetPath.back());
696       CallVal = UndefValue::get(SlotType);
697     }
698 
699     // The manipulations performed when we're looking through an insertvalue or
700     // an extractvalue would happen at the front of the RetPath list, so since
701     // we have to copy it anyway it's more efficient to create a reversed copy.
702     SmallVector<unsigned, 4> TmpRetPath(RetPath.rbegin(), RetPath.rend());
703     SmallVector<unsigned, 4> TmpCallPath(CallPath.rbegin(), CallPath.rend());
704 
705     // Finally, we can check whether the value produced by the tail call at this
706     // index is compatible with the value we return.
707     if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
708                               AllowDifferingSizes, TLI,
709                               F->getParent()->getDataLayout()))
710       return false;
711 
712     CallEmpty  = !nextRealType(CallSubTypes, CallPath);
713   } while(nextRealType(RetSubTypes, RetPath));
714 
715   return true;
716 }
717 
718 static void collectEHScopeMembers(
719     DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope,
720     const MachineBasicBlock *MBB) {
721   SmallVector<const MachineBasicBlock *, 16> Worklist = {MBB};
722   while (!Worklist.empty()) {
723     const MachineBasicBlock *Visiting = Worklist.pop_back_val();
724     // Don't follow blocks which start new scopes.
725     if (Visiting->isEHPad() && Visiting != MBB)
726       continue;
727 
728     // Add this MBB to our scope.
729     auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope));
730 
731     // Don't revisit blocks.
732     if (!P.second) {
733       assert(P.first->second == EHScope && "MBB is part of two scopes!");
734       continue;
735     }
736 
737     // Returns are boundaries where scope transfer can occur, don't follow
738     // successors.
739     if (Visiting->isEHScopeReturnBlock())
740       continue;
741 
742     for (const MachineBasicBlock *Succ : Visiting->successors())
743       Worklist.push_back(Succ);
744   }
745 }
746 
747 DenseMap<const MachineBasicBlock *, int>
748 llvm::getEHScopeMembership(const MachineFunction &MF) {
749   DenseMap<const MachineBasicBlock *, int> EHScopeMembership;
750 
751   // We don't have anything to do if there aren't any EH pads.
752   if (!MF.hasEHScopes())
753     return EHScopeMembership;
754 
755   int EntryBBNumber = MF.front().getNumber();
756   bool IsSEH = isAsynchronousEHPersonality(
757       classifyEHPersonality(MF.getFunction().getPersonalityFn()));
758 
759   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
760   SmallVector<const MachineBasicBlock *, 16> EHScopeBlocks;
761   SmallVector<const MachineBasicBlock *, 16> UnreachableBlocks;
762   SmallVector<const MachineBasicBlock *, 16> SEHCatchPads;
763   SmallVector<std::pair<const MachineBasicBlock *, int>, 16> CatchRetSuccessors;
764   for (const MachineBasicBlock &MBB : MF) {
765     if (MBB.isEHScopeEntry()) {
766       EHScopeBlocks.push_back(&MBB);
767     } else if (IsSEH && MBB.isEHPad()) {
768       SEHCatchPads.push_back(&MBB);
769     } else if (MBB.pred_empty()) {
770       UnreachableBlocks.push_back(&MBB);
771     }
772 
773     MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator();
774 
775     // CatchPads are not scopes for SEH so do not consider CatchRet to
776     // transfer control to another scope.
777     if (MBBI == MBB.end() || MBBI->getOpcode() != TII->getCatchReturnOpcode())
778       continue;
779 
780     // FIXME: SEH CatchPads are not necessarily in the parent function:
781     // they could be inside a finally block.
782     const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB();
783     const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB();
784     CatchRetSuccessors.push_back(
785         {Successor, IsSEH ? EntryBBNumber : SuccessorColor->getNumber()});
786   }
787 
788   // We don't have anything to do if there aren't any EH pads.
789   if (EHScopeBlocks.empty())
790     return EHScopeMembership;
791 
792   // Identify all the basic blocks reachable from the function entry.
793   collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front());
794   // All blocks not part of a scope are in the parent function.
795   for (const MachineBasicBlock *MBB : UnreachableBlocks)
796     collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
797   // Next, identify all the blocks inside the scopes.
798   for (const MachineBasicBlock *MBB : EHScopeBlocks)
799     collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB);
800   // SEH CatchPads aren't really scopes, handle them separately.
801   for (const MachineBasicBlock *MBB : SEHCatchPads)
802     collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
803   // Finally, identify all the targets of a catchret.
804   for (std::pair<const MachineBasicBlock *, int> CatchRetPair :
805        CatchRetSuccessors)
806     collectEHScopeMembers(EHScopeMembership, CatchRetPair.second,
807                           CatchRetPair.first);
808   return EHScopeMembership;
809 }
810