LLVM  9.0.0svn
Analysis.cpp
Go to the documentation of this file.
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"
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"
29 
30 using namespace llvm;
31 
32 /// Compute the linearized index of a member in a nested aggregate/struct/array
33 /// by recursing and accumulating CurIndex as long as there are indices in the
34 /// index list.
36  const unsigned *Indices,
37  const unsigned *IndicesEnd,
38  unsigned CurIndex) {
39  // Base case: We're done.
40  if (Indices && Indices == IndicesEnd)
41  return CurIndex;
42 
43  // Given a struct type, recursively traverse the elements.
44  if (StructType *STy = dyn_cast<StructType>(Ty)) {
45  for (StructType::element_iterator EB = STy->element_begin(),
46  EI = EB,
47  EE = STy->element_end();
48  EI != EE; ++EI) {
49  if (Indices && *Indices == unsigned(EI - EB))
50  return ComputeLinearIndex(*EI, Indices+1, IndicesEnd, CurIndex);
51  CurIndex = ComputeLinearIndex(*EI, nullptr, nullptr, CurIndex);
52  }
53  assert(!Indices && "Unexpected out of bound");
54  return CurIndex;
55  }
56  // Given an array type, recursively traverse the elements.
57  else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
58  Type *EltTy = ATy->getElementType();
59  unsigned NumElts = ATy->getNumElements();
60  // Compute the Linear offset when jumping one element of the array
61  unsigned EltLinearOffset = ComputeLinearIndex(EltTy, nullptr, nullptr, 0);
62  if (Indices) {
63  assert(*Indices < NumElts && "Unexpected out of bound");
64  // If the indice is inside the array, compute the index to the requested
65  // elt and recurse inside the element with the end of the indices list
66  CurIndex += EltLinearOffset* *Indices;
67  return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex);
68  }
69  CurIndex += EltLinearOffset*NumElts;
70  return CurIndex;
71  }
72  // We haven't found the type we're looking for, so keep searching.
73  return CurIndex + 1;
74 }
75 
76 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
77 /// EVTs that represent all the individual underlying
78 /// non-aggregate types that comprise it.
79 ///
80 /// If Offsets is non-null, it points to a vector to be filled in
81 /// with the in-memory offsets of each of the individual values.
82 ///
83 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
84  Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
86  uint64_t StartingOffset) {
87  // Given a struct type, recursively traverse the elements.
88  if (StructType *STy = dyn_cast<StructType>(Ty)) {
89  const StructLayout *SL = DL.getStructLayout(STy);
90  for (StructType::element_iterator EB = STy->element_begin(),
91  EI = EB,
92  EE = STy->element_end();
93  EI != EE; ++EI)
94  ComputeValueVTs(TLI, DL, *EI, ValueVTs, Offsets,
95  StartingOffset + SL->getElementOffset(EI - EB));
96  return;
97  }
98  // Given an array type, recursively traverse the elements.
99  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
100  Type *EltTy = ATy->getElementType();
101  uint64_t EltSize = DL.getTypeAllocSize(EltTy);
102  for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
103  ComputeValueVTs(TLI, DL, EltTy, ValueVTs, Offsets,
104  StartingOffset + i * EltSize);
105  return;
106  }
107  // Interpret void as zero return values.
108  if (Ty->isVoidTy())
109  return;
110  // Base case: we can get an EVT for this LLVM IR type.
111  ValueVTs.push_back(TLI.getValueType(DL, Ty));
112  if (Offsets)
113  Offsets->push_back(StartingOffset);
114 }
115 
117  SmallVectorImpl<LLT> &ValueTys,
119  uint64_t StartingOffset) {
120  // Given a struct type, recursively traverse the elements.
121  if (StructType *STy = dyn_cast<StructType>(&Ty)) {
122  const StructLayout *SL = DL.getStructLayout(STy);
123  for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I)
124  computeValueLLTs(DL, *STy->getElementType(I), ValueTys, Offsets,
125  StartingOffset + SL->getElementOffset(I));
126  return;
127  }
128  // Given an array type, recursively traverse the elements.
129  if (ArrayType *ATy = dyn_cast<ArrayType>(&Ty)) {
130  Type *EltTy = ATy->getElementType();
131  uint64_t EltSize = DL.getTypeAllocSize(EltTy);
132  for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
133  computeValueLLTs(DL, *EltTy, ValueTys, Offsets,
134  StartingOffset + i * EltSize);
135  return;
136  }
137  // Interpret void as zero return values.
138  if (Ty.isVoidTy())
139  return;
140  // Base case: we can get an LLT for this LLVM IR type.
141  ValueTys.push_back(getLLTForType(Ty, DL));
142  if (Offsets != nullptr)
143  Offsets->push_back(StartingOffset * 8);
144 }
145 
146 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
148  V = V->stripPointerCasts();
149  GlobalValue *GV = dyn_cast<GlobalValue>(V);
151 
152  if (Var && Var->getName() == "llvm.eh.catch.all.value") {
153  assert(Var->hasInitializer() &&
154  "The EH catch-all value must have an initializer");
155  Value *Init = Var->getInitializer();
156  GV = dyn_cast<GlobalValue>(Init);
157  if (!GV) V = cast<ConstantPointerNull>(Init);
158  }
159 
160  assert((GV || isa<ConstantPointerNull>(V)) &&
161  "TypeInfo must be a global variable or NULL");
162  return GV;
163 }
164 
165 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
166 /// processed uses a memory 'm' constraint.
167 bool
169  const TargetLowering &TLI) {
170  for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
171  InlineAsm::ConstraintInfo &CI = CInfos[i];
172  for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
174  if (CType == TargetLowering::C_Memory)
175  return true;
176  }
177 
178  // Indirect operand accesses access memory.
179  if (CI.isIndirect)
180  return true;
181  }
182 
183  return false;
184 }
185 
186 /// getFCmpCondCode - Return the ISD condition code corresponding to
187 /// the given LLVM IR floating-point condition code. This includes
188 /// consideration of global floating-point math flags.
189 ///
191  switch (Pred) {
192  case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
193  case FCmpInst::FCMP_OEQ: return ISD::SETOEQ;
194  case FCmpInst::FCMP_OGT: return ISD::SETOGT;
195  case FCmpInst::FCMP_OGE: return ISD::SETOGE;
196  case FCmpInst::FCMP_OLT: return ISD::SETOLT;
197  case FCmpInst::FCMP_OLE: return ISD::SETOLE;
198  case FCmpInst::FCMP_ONE: return ISD::SETONE;
199  case FCmpInst::FCMP_ORD: return ISD::SETO;
200  case FCmpInst::FCMP_UNO: return ISD::SETUO;
201  case FCmpInst::FCMP_UEQ: return ISD::SETUEQ;
202  case FCmpInst::FCMP_UGT: return ISD::SETUGT;
203  case FCmpInst::FCMP_UGE: return ISD::SETUGE;
204  case FCmpInst::FCMP_ULT: return ISD::SETULT;
205  case FCmpInst::FCMP_ULE: return ISD::SETULE;
206  case FCmpInst::FCMP_UNE: return ISD::SETUNE;
207  case FCmpInst::FCMP_TRUE: return ISD::SETTRUE;
208  default: llvm_unreachable("Invalid FCmp predicate opcode!");
209  }
210 }
211 
213  switch (CC) {
214  case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
215  case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
216  case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
217  case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
218  case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
219  case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
220  default: return CC;
221  }
222 }
223 
224 /// getICmpCondCode - Return the ISD condition code corresponding to
225 /// the given LLVM IR integer condition code.
226 ///
228  switch (Pred) {
229  case ICmpInst::ICMP_EQ: return ISD::SETEQ;
230  case ICmpInst::ICMP_NE: return ISD::SETNE;
231  case ICmpInst::ICMP_SLE: return ISD::SETLE;
232  case ICmpInst::ICMP_ULE: return ISD::SETULE;
233  case ICmpInst::ICMP_SGE: return ISD::SETGE;
234  case ICmpInst::ICMP_UGE: return ISD::SETUGE;
235  case ICmpInst::ICMP_SLT: return ISD::SETLT;
236  case ICmpInst::ICMP_ULT: return ISD::SETULT;
237  case ICmpInst::ICMP_SGT: return ISD::SETGT;
238  case ICmpInst::ICMP_UGT: return ISD::SETUGT;
239  default:
240  llvm_unreachable("Invalid ICmp predicate opcode!");
241  }
242 }
243 
244 static bool isNoopBitcast(Type *T1, Type *T2,
245  const TargetLoweringBase& TLI) {
246  return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
247  (isa<VectorType>(T1) && isa<VectorType>(T2) &&
248  TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
249 }
250 
251 /// Look through operations that will be free to find the earliest source of
252 /// this value.
253 ///
254 /// @param ValLoc If V has aggegate type, we will be interested in a particular
255 /// scalar component. This records its address; the reverse of this list gives a
256 /// sequence of indices appropriate for an extractvalue to locate the important
257 /// value. This value is updated during the function and on exit will indicate
258 /// similar information for the Value returned.
259 ///
260 /// @param DataBits If this function looks through truncate instructions, this
261 /// will record the smallest size attained.
262 static const Value *getNoopInput(const Value *V,
264  unsigned &DataBits,
265  const TargetLoweringBase &TLI,
266  const DataLayout &DL) {
267  while (true) {
268  // Try to look through V1; if V1 is not an instruction, it can't be looked
269  // through.
270  const Instruction *I = dyn_cast<Instruction>(V);
271  if (!I || I->getNumOperands() == 0) return V;
272  const Value *NoopInput = nullptr;
273 
274  Value *Op = I->getOperand(0);
275  if (isa<BitCastInst>(I)) {
276  // Look through truly no-op bitcasts.
277  if (isNoopBitcast(Op->getType(), I->getType(), TLI))
278  NoopInput = Op;
279  } else if (isa<GetElementPtrInst>(I)) {
280  // Look through getelementptr
281  if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
282  NoopInput = Op;
283  } else if (isa<IntToPtrInst>(I)) {
284  // Look through inttoptr.
285  // Make sure this isn't a truncating or extending cast. We could
286  // support this eventually, but don't bother for now.
287  if (!isa<VectorType>(I->getType()) &&
288  DL.getPointerSizeInBits() ==
289  cast<IntegerType>(Op->getType())->getBitWidth())
290  NoopInput = Op;
291  } else if (isa<PtrToIntInst>(I)) {
292  // Look through ptrtoint.
293  // Make sure this isn't a truncating or extending cast. We could
294  // support this eventually, but don't bother for now.
295  if (!isa<VectorType>(I->getType()) &&
296  DL.getPointerSizeInBits() ==
297  cast<IntegerType>(I->getType())->getBitWidth())
298  NoopInput = Op;
299  } else if (isa<TruncInst>(I) &&
300  TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
301  DataBits = std::min(DataBits, I->getType()->getPrimitiveSizeInBits());
302  NoopInput = Op;
303  } else if (auto CS = ImmutableCallSite(I)) {
304  const Value *ReturnedOp = CS.getReturnedArgOperand();
305  if (ReturnedOp && isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI))
306  NoopInput = ReturnedOp;
307  } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
308  // Value may come from either the aggregate or the scalar
309  ArrayRef<unsigned> InsertLoc = IVI->getIndices();
310  if (ValLoc.size() >= InsertLoc.size() &&
311  std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) {
312  // The type being inserted is a nested sub-type of the aggregate; we
313  // have to remove those initial indices to get the location we're
314  // interested in for the operand.
315  ValLoc.resize(ValLoc.size() - InsertLoc.size());
316  NoopInput = IVI->getInsertedValueOperand();
317  } else {
318  // The struct we're inserting into has the value we're interested in, no
319  // change of address.
320  NoopInput = Op;
321  }
322  } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
323  // The part we're interested in will inevitably be some sub-section of the
324  // previous aggregate. Combine the two paths to obtain the true address of
325  // our element.
326  ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
327  ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
328  NoopInput = Op;
329  }
330  // Terminate if we couldn't find anything to look through.
331  if (!NoopInput)
332  return V;
333 
334  V = NoopInput;
335  }
336 }
337 
338 /// Return true if this scalar return value only has bits discarded on its path
339 /// from the "tail call" to the "ret". This includes the obvious noop
340 /// instructions handled by getNoopInput above as well as free truncations (or
341 /// extensions prior to the call).
342 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
343  SmallVectorImpl<unsigned> &RetIndices,
344  SmallVectorImpl<unsigned> &CallIndices,
345  bool AllowDifferingSizes,
346  const TargetLoweringBase &TLI,
347  const DataLayout &DL) {
348 
349  // Trace the sub-value needed by the return value as far back up the graph as
350  // possible, in the hope that it will intersect with the value produced by the
351  // call. In the simple case with no "returned" attribute, the hope is actually
352  // that we end up back at the tail call instruction itself.
353  unsigned BitsRequired = UINT_MAX;
354  RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL);
355 
356  // If this slot in the value returned is undef, it doesn't matter what the
357  // call puts there, it'll be fine.
358  if (isa<UndefValue>(RetVal))
359  return true;
360 
361  // Now do a similar search up through the graph to find where the value
362  // actually returned by the "tail call" comes from. In the simple case without
363  // a "returned" attribute, the search will be blocked immediately and the loop
364  // a Noop.
365  unsigned BitsProvided = UINT_MAX;
366  CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL);
367 
368  // There's no hope if we can't actually trace them to (the same part of!) the
369  // same value.
370  if (CallVal != RetVal || CallIndices != RetIndices)
371  return false;
372 
373  // However, intervening truncates may have made the call non-tail. Make sure
374  // all the bits that are needed by the "ret" have been provided by the "tail
375  // call". FIXME: with sufficiently cunning bit-tracking, we could look through
376  // extensions too.
377  if (BitsProvided < BitsRequired ||
378  (!AllowDifferingSizes && BitsProvided != BitsRequired))
379  return false;
380 
381  return true;
382 }
383 
384 /// For an aggregate type, determine whether a given index is within bounds or
385 /// not.
386 static bool indexReallyValid(CompositeType *T, unsigned Idx) {
387  if (ArrayType *AT = dyn_cast<ArrayType>(T))
388  return Idx < AT->getNumElements();
389 
390  return Idx < cast<StructType>(T)->getNumElements();
391 }
392 
393 /// Move the given iterators to the next leaf type in depth first traversal.
394 ///
395 /// Performs a depth-first traversal of the type as specified by its arguments,
396 /// stopping at the next leaf node (which may be a legitimate scalar type or an
397 /// empty struct or array).
398 ///
399 /// @param SubTypes List of the partial components making up the type from
400 /// outermost to innermost non-empty aggregate. The element currently
401 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
402 ///
403 /// @param Path Set of extractvalue indices leading from the outermost type
404 /// (SubTypes[0]) to the leaf node currently represented.
405 ///
406 /// @returns true if a new type was found, false otherwise. Calling this
407 /// function again on a finished iterator will repeatedly return
408 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
409 /// aggregate or a non-aggregate
412  // First march back up the tree until we can successfully increment one of the
413  // coordinates in Path.
414  while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
415  Path.pop_back();
416  SubTypes.pop_back();
417  }
418 
419  // If we reached the top, then the iterator is done.
420  if (Path.empty())
421  return false;
422 
423  // We know there's *some* valid leaf now, so march back down the tree picking
424  // out the left-most element at each node.
425  ++Path.back();
426  Type *DeeperType = SubTypes.back()->getTypeAtIndex(Path.back());
427  while (DeeperType->isAggregateType()) {
428  CompositeType *CT = cast<CompositeType>(DeeperType);
429  if (!indexReallyValid(CT, 0))
430  return true;
431 
432  SubTypes.push_back(CT);
433  Path.push_back(0);
434 
435  DeeperType = CT->getTypeAtIndex(0U);
436  }
437 
438  return true;
439 }
440 
441 /// Find the first non-empty, scalar-like type in Next and setup the iterator
442 /// components.
443 ///
444 /// Assuming Next is an aggregate of some kind, this function will traverse the
445 /// tree from left to right (i.e. depth-first) looking for the first
446 /// non-aggregate type which will play a role in function return.
447 ///
448 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
449 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
450 /// i32 in that type.
451 static bool firstRealType(Type *Next,
454  // First initialise the iterator components to the first "leaf" node
455  // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
456  // despite nominally being an aggregate).
457  while (Next->isAggregateType() &&
458  indexReallyValid(cast<CompositeType>(Next), 0)) {
459  SubTypes.push_back(cast<CompositeType>(Next));
460  Path.push_back(0);
461  Next = cast<CompositeType>(Next)->getTypeAtIndex(0U);
462  }
463 
464  // If there's no Path now, Next was originally scalar already (or empty
465  // leaf). We're done.
466  if (Path.empty())
467  return true;
468 
469  // Otherwise, use normal iteration to keep looking through the tree until we
470  // find a non-aggregate type.
471  while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType()) {
472  if (!advanceToNextLeafType(SubTypes, Path))
473  return false;
474  }
475 
476  return true;
477 }
478 
479 /// Set the iterator data-structures to the next non-empty, non-aggregate
480 /// subtype.
483  do {
484  if (!advanceToNextLeafType(SubTypes, Path))
485  return false;
486 
487  assert(!Path.empty() && "found a leaf but didn't set the path?");
488  } while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType());
489 
490  return true;
491 }
492 
493 
494 /// Test if the given instruction is in a position to be optimized
495 /// with a tail-call. This roughly means that it's in a block with
496 /// a return and there's nothing that needs to be scheduled
497 /// between it and the return.
498 ///
499 /// This function only tests target-independent requirements.
501  const Instruction *I = CS.getInstruction();
502  const BasicBlock *ExitBB = I->getParent();
503  const Instruction *Term = ExitBB->getTerminator();
504  const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
505 
506  // The block must end in a return statement or unreachable.
507  //
508  // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
509  // an unreachable, for now. The way tailcall optimization is currently
510  // implemented means it will add an epilogue followed by a jump. That is
511  // not profitable. Also, if the callee is a special function (e.g.
512  // longjmp on x86), it can end up causing miscompilation that has not
513  // been fully understood.
514  if (!Ret &&
515  (!TM.Options.GuaranteedTailCallOpt || !isa<UnreachableInst>(Term)))
516  return false;
517 
518  // If I will have a chain, make sure no other instruction that will have a
519  // chain interposes between I and the return.
520  if (I->mayHaveSideEffects() || I->mayReadFromMemory() ||
522  for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) {
523  if (&*BBI == I)
524  break;
525  // Debug info intrinsics do not get in the way of tail call optimization.
526  if (isa<DbgInfoIntrinsic>(BBI))
527  continue;
528  // A lifetime end intrinsic should not stop tail call optimization.
529  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
530  if (II->getIntrinsicID() == Intrinsic::lifetime_end)
531  continue;
532  if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
534  return false;
535  }
536 
537  const Function *F = ExitBB->getParent();
539  F, I, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering());
540 }
541 
543  const ReturnInst *Ret,
544  const TargetLoweringBase &TLI,
545  bool *AllowDifferingSizes) {
546  // ADS may be null, so don't write to it directly.
547  bool DummyADS;
548  bool &ADS = AllowDifferingSizes ? *AllowDifferingSizes : DummyADS;
549  ADS = true;
550 
552  AttrBuilder CalleeAttrs(cast<CallInst>(I)->getAttributes(),
554 
555  // NoAlias and NonNull are completely benign as far as calling convention
556  // goes, they shouldn't affect whether the call is a tail call.
557  CallerAttrs.removeAttribute(Attribute::NoAlias);
558  CalleeAttrs.removeAttribute(Attribute::NoAlias);
559  CallerAttrs.removeAttribute(Attribute::NonNull);
560  CalleeAttrs.removeAttribute(Attribute::NonNull);
561 
562  if (CallerAttrs.contains(Attribute::ZExt)) {
563  if (!CalleeAttrs.contains(Attribute::ZExt))
564  return false;
565 
566  ADS = false;
567  CallerAttrs.removeAttribute(Attribute::ZExt);
568  CalleeAttrs.removeAttribute(Attribute::ZExt);
569  } else if (CallerAttrs.contains(Attribute::SExt)) {
570  if (!CalleeAttrs.contains(Attribute::SExt))
571  return false;
572 
573  ADS = false;
574  CallerAttrs.removeAttribute(Attribute::SExt);
575  CalleeAttrs.removeAttribute(Attribute::SExt);
576  }
577 
578  // Drop sext and zext return attributes if the result is not used.
579  // This enables tail calls for code like:
580  //
581  // define void @caller() {
582  // entry:
583  // %unused_result = tail call zeroext i1 @callee()
584  // br label %retlabel
585  // retlabel:
586  // ret void
587  // }
588  if (I->use_empty()) {
589  CalleeAttrs.removeAttribute(Attribute::SExt);
590  CalleeAttrs.removeAttribute(Attribute::ZExt);
591  }
592 
593  // If they're still different, there's some facet we don't understand
594  // (currently only "inreg", but in future who knows). It may be OK but the
595  // only safe option is to reject the tail call.
596  return CallerAttrs == CalleeAttrs;
597 }
598 
600  const Instruction *I,
601  const ReturnInst *Ret,
602  const TargetLoweringBase &TLI) {
603  // If the block ends with a void return or unreachable, it doesn't matter
604  // what the call's return type is.
605  if (!Ret || Ret->getNumOperands() == 0) return true;
606 
607  // If the return value is undef, it doesn't matter what the call's
608  // return type is.
609  if (isa<UndefValue>(Ret->getOperand(0))) return true;
610 
611  // Make sure the attributes attached to each return are compatible.
612  bool AllowDifferingSizes;
613  if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
614  return false;
615 
616  const Value *RetVal = Ret->getOperand(0), *CallVal = I;
617  // Intrinsic like llvm.memcpy has no return value, but the expanded
618  // libcall may or may not have return value. On most platforms, it
619  // will be expanded as memcpy in libc, which returns the first
620  // argument. On other platforms like arm-none-eabi, memcpy may be
621  // expanded as library call without return value, like __aeabi_memcpy.
622  const CallInst *Call = cast<CallInst>(I);
623  if (Function *F = Call->getCalledFunction()) {
624  Intrinsic::ID IID = F->getIntrinsicID();
625  if (((IID == Intrinsic::memcpy &&
626  TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")) ||
627  (IID == Intrinsic::memmove &&
628  TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")) ||
629  (IID == Intrinsic::memset &&
630  TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset"))) &&
631  RetVal == Call->getArgOperand(0))
632  return true;
633  }
634 
635  SmallVector<unsigned, 4> RetPath, CallPath;
636  SmallVector<CompositeType *, 4> RetSubTypes, CallSubTypes;
637 
638  bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
639  bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
640 
641  // Nothing's actually returned, it doesn't matter what the callee put there
642  // it's a valid tail call.
643  if (RetEmpty)
644  return true;
645 
646  // Iterate pairwise through each of the value types making up the tail call
647  // and the corresponding return. For each one we want to know whether it's
648  // essentially going directly from the tail call to the ret, via operations
649  // that end up not generating any code.
650  //
651  // We allow a certain amount of covariance here. For example it's permitted
652  // for the tail call to define more bits than the ret actually cares about
653  // (e.g. via a truncate).
654  do {
655  if (CallEmpty) {
656  // We've exhausted the values produced by the tail call instruction, the
657  // rest are essentially undef. The type doesn't really matter, but we need
658  // *something*.
659  Type *SlotType = RetSubTypes.back()->getTypeAtIndex(RetPath.back());
660  CallVal = UndefValue::get(SlotType);
661  }
662 
663  // The manipulations performed when we're looking through an insertvalue or
664  // an extractvalue would happen at the front of the RetPath list, so since
665  // we have to copy it anyway it's more efficient to create a reversed copy.
666  SmallVector<unsigned, 4> TmpRetPath(RetPath.rbegin(), RetPath.rend());
667  SmallVector<unsigned, 4> TmpCallPath(CallPath.rbegin(), CallPath.rend());
668 
669  // Finally, we can check whether the value produced by the tail call at this
670  // index is compatible with the value we return.
671  if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
672  AllowDifferingSizes, TLI,
673  F->getParent()->getDataLayout()))
674  return false;
675 
676  CallEmpty = !nextRealType(CallSubTypes, CallPath);
677  } while(nextRealType(RetSubTypes, RetPath));
678 
679  return true;
680 }
681 
683  DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope,
684  const MachineBasicBlock *MBB) {
686  while (!Worklist.empty()) {
687  const MachineBasicBlock *Visiting = Worklist.pop_back_val();
688  // Don't follow blocks which start new scopes.
689  if (Visiting->isEHPad() && Visiting != MBB)
690  continue;
691 
692  // Add this MBB to our scope.
693  auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope));
694 
695  // Don't revisit blocks.
696  if (!P.second) {
697  assert(P.first->second == EHScope && "MBB is part of two scopes!");
698  continue;
699  }
700 
701  // Returns are boundaries where scope transfer can occur, don't follow
702  // successors.
703  if (Visiting->isEHScopeReturnBlock())
704  continue;
705 
706  for (const MachineBasicBlock *Succ : Visiting->successors())
707  Worklist.push_back(Succ);
708  }
709 }
710 
714 
715  // We don't have anything to do if there aren't any EH pads.
716  if (!MF.hasEHScopes())
717  return EHScopeMembership;
718 
719  int EntryBBNumber = MF.front().getNumber();
720  bool IsSEH = isAsynchronousEHPersonality(
722 
728  for (const MachineBasicBlock &MBB : MF) {
729  if (MBB.isEHScopeEntry()) {
730  EHScopeBlocks.push_back(&MBB);
731  } else if (IsSEH && MBB.isEHPad()) {
732  SEHCatchPads.push_back(&MBB);
733  } else if (MBB.pred_empty()) {
734  UnreachableBlocks.push_back(&MBB);
735  }
736 
737  MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator();
738 
739  // CatchPads are not scopes for SEH so do not consider CatchRet to
740  // transfer control to another scope.
741  if (MBBI == MBB.end() || MBBI->getOpcode() != TII->getCatchReturnOpcode())
742  continue;
743 
744  // FIXME: SEH CatchPads are not necessarily in the parent function:
745  // they could be inside a finally block.
746  const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB();
747  const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB();
748  CatchRetSuccessors.push_back(
749  {Successor, IsSEH ? EntryBBNumber : SuccessorColor->getNumber()});
750  }
751 
752  // We don't have anything to do if there aren't any EH pads.
753  if (EHScopeBlocks.empty())
754  return EHScopeMembership;
755 
756  // Identify all the basic blocks reachable from the function entry.
757  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front());
758  // All blocks not part of a scope are in the parent function.
759  for (const MachineBasicBlock *MBB : UnreachableBlocks)
760  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
761  // Next, identify all the blocks inside the scopes.
762  for (const MachineBasicBlock *MBB : EHScopeBlocks)
763  collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB);
764  // SEH CatchPads aren't really scopes, handle them separately.
765  for (const MachineBasicBlock *MBB : SEHCatchPads)
766  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
767  // Finally, identify all the targets of a catchret.
768  for (std::pair<const MachineBasicBlock *, int> CatchRetPair :
769  CatchRetSuccessors)
770  collectEHScopeMembers(EHScopeMembership, CatchRetPair.second,
771  CatchRetPair.first);
772  return EHScopeMembership;
773 }
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
Return a value (possibly void), from a function.
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
ISD::CondCode getICmpCondCode(ICmpInst::Predicate Pred)
getICmpCondCode - Return the ISD condition code corresponding to the given LLVM IR integer condition ...
Definition: Analysis.cpp:227
This instruction extracts a struct member or array element value from an aggregate value...
bool isEHScopeReturnBlock() const
Convenience function that returns true if the bock ends in a EH scope return instruction.
This class represents lattice values for constants.
Definition: AllocatorList.h:23
iterator begin() const
Definition: ArrayRef.h:136
const StructLayout * getStructLayout(StructType *Ty) const
Returns a StructLayout object, indicating the alignment of the struct, its size, and the offsets of i...
Definition: DataLayout.cpp:607
This class represents a function call, abstracting a target machine&#39;s calling convention.
virtual const TargetLowering * getTargetLowering() const
reverse_iterator rbegin() const
Definition: ArrayRef.h:139
unsigned less or equal
Definition: InstrTypes.h:672
static const Value * getNoopInput(const Value *V, SmallVectorImpl< unsigned > &ValLoc, unsigned &DataBits, const TargetLoweringBase &TLI, const DataLayout &DL)
Look through operations that will be free to find the earliest source of this value.
Definition: Analysis.cpp:262
The two locations do not alias at all.
Definition: AliasAnalysis.h:84
unsigned less than
Definition: InstrTypes.h:671
0 1 0 0 True if ordered and less than
Definition: InstrTypes.h:652
Offsets
Offsets in bytes from the start of the input buffer.
Definition: SIInstrInfo.h:1034
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:662
unsigned getPointerSizeInBits(unsigned AS=0) const
Layout pointer size, in bits FIXME: The defaults need to be removed once all of the backends/clients ...
Definition: DataLayout.h:388
F(f)
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:137
GlobalValue * ExtractTypeInfo(Value *V)
ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
Definition: Analysis.cpp:147
bool returnTypeIsEligibleForTailCall(const Function *F, const Instruction *I, const ReturnInst *Ret, const TargetLoweringBase &TLI)
Test if given that the input instruction is in the tail call position if the return type or any attri...
Definition: Analysis.cpp:599
ConstraintCodeVector Codes
Code - The constraint code, either the register name (in braces) or the constraint letter/number...
Definition: InlineAsm.h:148
iterator_range< succ_iterator > successors()
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1155
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
1 0 0 1 True if unordered or equal
Definition: InstrTypes.h:657
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:554
1 0 0 0 True if unordered: isnan(X) | isnan(Y)
Definition: InstrTypes.h:656
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
const HexagonInstrInfo * TII
Class to represent struct types.
Definition: DerivedTypes.h:232
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
0 1 0 1 True if ordered and less than or equal
Definition: InstrTypes.h:653
unsigned getCatchReturnOpcode() const
This class defines information used to lower LLVM code to legal SelectionDAG operators that the targe...
AttributeList getAttributes(LLVMContext &C, ID id)
Return the attributes for an intrinsic.
InstrTy * getInstruction() const
Definition: CallSite.h:96
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
#define T
Class to represent array types.
Definition: DerivedTypes.h:400
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:223
CondCode
ISD::CondCode enum - These are ordered carefully to make the bitfields below work out...
Definition: ISDOpcodes.h:968
int getNumber() const
MachineBasicBlocks are uniquely numbered at the function level, unless they&#39;re not in a MachineFuncti...
virtual const TargetInstrInfo * getInstrInfo() const
ISD::CondCode getFCmpCondCode(FCmpInst::Predicate Pred)
getFCmpCondCode - Return the ISD condition code corresponding to the given LLVM IR floating-point con...
Definition: Analysis.cpp:190
Type::subtype_iterator element_iterator
Definition: DerivedTypes.h:332
void ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL, Type *Ty, SmallVectorImpl< EVT > &ValueVTs, SmallVectorImpl< uint64_t > *Offsets=nullptr, uint64_t StartingOffset=0)
ComputeValueVTs - Given an LLVM IR type, compute a sequence of EVTs that represent all the individual...
Definition: Analysis.cpp:83
Value * getOperand(unsigned i) const
Definition: User.h:169
bool hasInlineAsmMemConstraint(InlineAsm::ConstraintInfoVector &CInfos, const TargetLowering &TLI)
hasInlineAsmMemConstraint - Return true if the inline asm instruction being processed uses a memory &#39;...
Definition: Analysis.cpp:168
TargetInstrInfo - Interface to description of machine instruction set.
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:140
static bool indexReallyValid(CompositeType *T, unsigned Idx)
For an aggregate type, determine whether a given index is within bounds or not.
Definition: Analysis.cpp:386
static bool nextRealType(SmallVectorImpl< CompositeType *> &SubTypes, SmallVectorImpl< unsigned > &Path)
Set the iterator data-structures to the next non-empty, non-aggregate subtype.
Definition: Analysis.cpp:481
#define P(N)
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
virtual ConstraintType getConstraintType(StringRef Constraint) const
Given a constraint, return the type of constraint it is for this target.
bool isIndirect
isIndirect - True if this operand is an indirect operand.
Definition: InlineAsm.h:144
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:148
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:223
void computeValueLLTs(const DataLayout &DL, Type &Ty, SmallVectorImpl< LLT > &ValueTys, SmallVectorImpl< uint64_t > *Offsets=nullptr, uint64_t StartingOffset=0)
computeValueLLTs - Given an LLVM IR type, compute a sequence of LLTs that represent all the individua...
Definition: Analysis.cpp:116
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:572
EHPersonality classifyEHPersonality(const Value *Pers)
See if the given exception handling personality function is one that we understand.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:646
unsigned GuaranteedTailCallOpt
GuaranteedTailCallOpt - This flag is enabled when -tailcallopt is specified on the commandline...
static bool firstRealType(Type *Next, SmallVectorImpl< CompositeType *> &SubTypes, SmallVectorImpl< unsigned > &Path)
Find the first non-empty, scalar-like type in Next and setup the iterator components.
Definition: Analysis.cpp:451
0 1 1 1 True if ordered (no nans)
Definition: InstrTypes.h:655
1 1 1 1 Always true (always folded)
Definition: InstrTypes.h:663
virtual bool allowTruncateForTailCall(Type *FromTy, Type *ToTy) const
Return true if a truncation from FromTy to ToTy is permitted when deciding whether a call is in tail ...
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1424
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:529
static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal, SmallVectorImpl< unsigned > &RetIndices, SmallVectorImpl< unsigned > &CallIndices, bool AllowDifferingSizes, const TargetLoweringBase &TLI, const DataLayout &DL)
Return true if this scalar return value only has bits discarded on its path from the "tail call" to t...
Definition: Analysis.cpp:342
const MachineBasicBlock & front() const
size_t size() const
Definition: SmallVector.h:52
1 1 0 1 True if unordered, less than, or equal
Definition: InstrTypes.h:661
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
EVT getValueType(const DataLayout &DL, Type *Ty, bool AllowUnknown=false) const
Return the EVT corresponding to this LLVM type.
bool isInTailCallPosition(ImmutableCallSite CS, const TargetMachine &TM)
Test if the given instruction is in a position to be optimized with a tail-call.
Definition: Analysis.cpp:500
signed greater than
Definition: InstrTypes.h:673
std::vector< ConstraintInfo > ConstraintInfoVector
Definition: InlineAsm.h:115
This base class for TargetLowering contains the SelectionDAG-independent parts that can be used from ...
LLT getLLTForType(Type &Ty, const DataLayout &DL)
Construct a low-level type based on an LLVM type.
0 0 1 0 True if ordered and greater than
Definition: InstrTypes.h:650
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:191
static bool isNoopBitcast(Type *T1, Type *T2, const TargetLoweringBase &TLI)
Definition: Analysis.cpp:244
iterator end()
Definition: BasicBlock.h:270
1 1 0 0 True if unordered or less than
Definition: InstrTypes.h:660
Module.h This file contains the declarations for the Module class.
iterator end() const
Definition: ArrayRef.h:137
bool isAggregateType() const
Return true if the type is an aggregate type.
Definition: Type.h:257
signed less than
Definition: InstrTypes.h:675
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
DenseMap< const MachineBasicBlock *, int > getEHScopeMembership(const MachineFunction &MF)
Definition: Analysis.cpp:712
static bool advanceToNextLeafType(SmallVectorImpl< CompositeType *> &SubTypes, SmallVectorImpl< unsigned > &Path)
Move the given iterators to the next leaf type in depth first traversal.
Definition: Analysis.cpp:410
ISD::CondCode getFCmpCodeWithoutNaN(ISD::CondCode CC)
getFCmpCodeWithoutNaN - Given an ISD condition code comparing floats, return the equivalent code if w...
Definition: Analysis.cpp:212
Intrinsic::ID getIntrinsicID() const LLVM_READONLY
getIntrinsicID - This method returns the ID number of the specified function, or Intrinsic::not_intri...
Definition: Function.h:193
const Function & getFunction() const
Return the LLVM function that this machine code represents.
virtual const TargetSubtargetInfo * getSubtargetImpl(const Function &) const
Virtual method implemented by subclasses that returns a reference to that target&#39;s TargetSubtargetInf...
signed less or equal
Definition: InstrTypes.h:676
bool isTypeLegal(EVT VT) const
Return true if the target has native support for the specified value type.
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:387
Common super class of ArrayType, StructType and VectorType.
Definition: DerivedTypes.h:193
static void collectEHScopeMembers(DenseMap< const MachineBasicBlock *, int > &EHScopeMembership, int EHScope, const MachineBasicBlock *MBB)
Definition: Analysis.cpp:682
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:461
reverse_iterator rend() const
Definition: ArrayRef.h:140
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:576
bool isEHPad() const
Returns true if the block is a landing pad.
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
unsigned greater or equal
Definition: InstrTypes.h:670
TargetOptions Options
Establish a view to a call site for examination.
Definition: CallSite.h:892
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation.
Definition: InstrTypes.h:1201
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
#define I(x, y, z)
Definition: MD5.cpp:58
bool mayReadFromMemory() const
Return true if this instruction may read memory.
0 1 1 0 True if ordered and operands are unequal
Definition: InstrTypes.h:654
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:332
1 0 1 0 True if unordered or greater than
Definition: InstrTypes.h:658
static EVT getEVT(Type *Ty, bool HandleUnknown=false)
Return the value type corresponding to the specified type.
Definition: ValueTypes.cpp:316
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
bool isAsynchronousEHPersonality(EHPersonality Pers)
Returns true if this personality function catches asynchronous exceptions.
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:114
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:649
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
LLVM Value Representation.
Definition: Value.h:72
Constant * getPersonalityFn() const
Get the personality function associated with this function.
Definition: Function.cpp:1302
1 0 1 1 True if unordered, greater than, or equal
Definition: InstrTypes.h:659
Primary interface to the complete machine description for the target machine.
Definition: TargetMachine.h:65
unsigned greater than
Definition: InstrTypes.h:669
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
bool attributesPermitTailCall(const Function *F, const Instruction *I, const ReturnInst *Ret, const TargetLoweringBase &TLI, bool *AllowDifferingSizes=nullptr)
Test if given that the input instruction is in the tail call position, if there is an attribute misma...
Definition: Analysis.cpp:542
0 0 1 1 True if ordered and greater than or equal
Definition: InstrTypes.h:651
unsigned ComputeLinearIndex(Type *Ty, const unsigned *Indices, const unsigned *IndicesEnd, unsigned CurIndex=0)
Compute the linearized index of a member in a nested aggregate/struct/array.
Definition: Analysis.cpp:35
bool use_empty() const
Definition: Value.h:322
#define T1
0 0 0 0 Always false (always folded)
Definition: InstrTypes.h:648
signed greater or equal
Definition: InstrTypes.h:674
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
This file describes how to lower LLVM code to machine code.
Type * getTypeAtIndex(const Value *V) const
Given an index value into the type, return the type of the element.
Definition: Type.cpp:529
const BasicBlock * getParent() const
Definition: Instruction.h:66
const char * getLibcallName(RTLIB::Libcall Call) const
Get the libcall routine name for the specified libcall.
This instruction inserts a struct field of array element value into an aggregate value.
void resize(size_type N)
Definition: SmallVector.h:344