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 
116 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
118  V = V->stripPointerCasts();
119  GlobalValue *GV = dyn_cast<GlobalValue>(V);
121 
122  if (Var && Var->getName() == "llvm.eh.catch.all.value") {
123  assert(Var->hasInitializer() &&
124  "The EH catch-all value must have an initializer");
125  Value *Init = Var->getInitializer();
126  GV = dyn_cast<GlobalValue>(Init);
127  if (!GV) V = cast<ConstantPointerNull>(Init);
128  }
129 
130  assert((GV || isa<ConstantPointerNull>(V)) &&
131  "TypeInfo must be a global variable or NULL");
132  return GV;
133 }
134 
135 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
136 /// processed uses a memory 'm' constraint.
137 bool
139  const TargetLowering &TLI) {
140  for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
141  InlineAsm::ConstraintInfo &CI = CInfos[i];
142  for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
144  if (CType == TargetLowering::C_Memory)
145  return true;
146  }
147 
148  // Indirect operand accesses access memory.
149  if (CI.isIndirect)
150  return true;
151  }
152 
153  return false;
154 }
155 
156 /// getFCmpCondCode - Return the ISD condition code corresponding to
157 /// the given LLVM IR floating-point condition code. This includes
158 /// consideration of global floating-point math flags.
159 ///
161  switch (Pred) {
162  case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
163  case FCmpInst::FCMP_OEQ: return ISD::SETOEQ;
164  case FCmpInst::FCMP_OGT: return ISD::SETOGT;
165  case FCmpInst::FCMP_OGE: return ISD::SETOGE;
166  case FCmpInst::FCMP_OLT: return ISD::SETOLT;
167  case FCmpInst::FCMP_OLE: return ISD::SETOLE;
168  case FCmpInst::FCMP_ONE: return ISD::SETONE;
169  case FCmpInst::FCMP_ORD: return ISD::SETO;
170  case FCmpInst::FCMP_UNO: return ISD::SETUO;
171  case FCmpInst::FCMP_UEQ: return ISD::SETUEQ;
172  case FCmpInst::FCMP_UGT: return ISD::SETUGT;
173  case FCmpInst::FCMP_UGE: return ISD::SETUGE;
174  case FCmpInst::FCMP_ULT: return ISD::SETULT;
175  case FCmpInst::FCMP_ULE: return ISD::SETULE;
176  case FCmpInst::FCMP_UNE: return ISD::SETUNE;
177  case FCmpInst::FCMP_TRUE: return ISD::SETTRUE;
178  default: llvm_unreachable("Invalid FCmp predicate opcode!");
179  }
180 }
181 
183  switch (CC) {
184  case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
185  case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
186  case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
187  case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
188  case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
189  case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
190  default: return CC;
191  }
192 }
193 
194 /// getICmpCondCode - Return the ISD condition code corresponding to
195 /// the given LLVM IR integer condition code.
196 ///
198  switch (Pred) {
199  case ICmpInst::ICMP_EQ: return ISD::SETEQ;
200  case ICmpInst::ICMP_NE: return ISD::SETNE;
201  case ICmpInst::ICMP_SLE: return ISD::SETLE;
202  case ICmpInst::ICMP_ULE: return ISD::SETULE;
203  case ICmpInst::ICMP_SGE: return ISD::SETGE;
204  case ICmpInst::ICMP_UGE: return ISD::SETUGE;
205  case ICmpInst::ICMP_SLT: return ISD::SETLT;
206  case ICmpInst::ICMP_ULT: return ISD::SETULT;
207  case ICmpInst::ICMP_SGT: return ISD::SETGT;
208  case ICmpInst::ICMP_UGT: return ISD::SETUGT;
209  default:
210  llvm_unreachable("Invalid ICmp predicate opcode!");
211  }
212 }
213 
214 static bool isNoopBitcast(Type *T1, Type *T2,
215  const TargetLoweringBase& TLI) {
216  return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
217  (isa<VectorType>(T1) && isa<VectorType>(T2) &&
218  TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
219 }
220 
221 /// Look through operations that will be free to find the earliest source of
222 /// this value.
223 ///
224 /// @param ValLoc If V has aggegate type, we will be interested in a particular
225 /// scalar component. This records its address; the reverse of this list gives a
226 /// sequence of indices appropriate for an extractvalue to locate the important
227 /// value. This value is updated during the function and on exit will indicate
228 /// similar information for the Value returned.
229 ///
230 /// @param DataBits If this function looks through truncate instructions, this
231 /// will record the smallest size attained.
232 static const Value *getNoopInput(const Value *V,
234  unsigned &DataBits,
235  const TargetLoweringBase &TLI,
236  const DataLayout &DL) {
237  while (true) {
238  // Try to look through V1; if V1 is not an instruction, it can't be looked
239  // through.
240  const Instruction *I = dyn_cast<Instruction>(V);
241  if (!I || I->getNumOperands() == 0) return V;
242  const Value *NoopInput = nullptr;
243 
244  Value *Op = I->getOperand(0);
245  if (isa<BitCastInst>(I)) {
246  // Look through truly no-op bitcasts.
247  if (isNoopBitcast(Op->getType(), I->getType(), TLI))
248  NoopInput = Op;
249  } else if (isa<GetElementPtrInst>(I)) {
250  // Look through getelementptr
251  if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
252  NoopInput = Op;
253  } else if (isa<IntToPtrInst>(I)) {
254  // Look through inttoptr.
255  // Make sure this isn't a truncating or extending cast. We could
256  // support this eventually, but don't bother for now.
257  if (!isa<VectorType>(I->getType()) &&
258  DL.getPointerSizeInBits() ==
259  cast<IntegerType>(Op->getType())->getBitWidth())
260  NoopInput = Op;
261  } else if (isa<PtrToIntInst>(I)) {
262  // Look through ptrtoint.
263  // Make sure this isn't a truncating or extending cast. We could
264  // support this eventually, but don't bother for now.
265  if (!isa<VectorType>(I->getType()) &&
266  DL.getPointerSizeInBits() ==
267  cast<IntegerType>(I->getType())->getBitWidth())
268  NoopInput = Op;
269  } else if (isa<TruncInst>(I) &&
270  TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
271  DataBits = std::min(DataBits, I->getType()->getPrimitiveSizeInBits());
272  NoopInput = Op;
273  } else if (auto CS = ImmutableCallSite(I)) {
274  const Value *ReturnedOp = CS.getReturnedArgOperand();
275  if (ReturnedOp && isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI))
276  NoopInput = ReturnedOp;
277  } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
278  // Value may come from either the aggregate or the scalar
279  ArrayRef<unsigned> InsertLoc = IVI->getIndices();
280  if (ValLoc.size() >= InsertLoc.size() &&
281  std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) {
282  // The type being inserted is a nested sub-type of the aggregate; we
283  // have to remove those initial indices to get the location we're
284  // interested in for the operand.
285  ValLoc.resize(ValLoc.size() - InsertLoc.size());
286  NoopInput = IVI->getInsertedValueOperand();
287  } else {
288  // The struct we're inserting into has the value we're interested in, no
289  // change of address.
290  NoopInput = Op;
291  }
292  } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
293  // The part we're interested in will inevitably be some sub-section of the
294  // previous aggregate. Combine the two paths to obtain the true address of
295  // our element.
296  ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
297  ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
298  NoopInput = Op;
299  }
300  // Terminate if we couldn't find anything to look through.
301  if (!NoopInput)
302  return V;
303 
304  V = NoopInput;
305  }
306 }
307 
308 /// Return true if this scalar return value only has bits discarded on its path
309 /// from the "tail call" to the "ret". This includes the obvious noop
310 /// instructions handled by getNoopInput above as well as free truncations (or
311 /// extensions prior to the call).
312 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
313  SmallVectorImpl<unsigned> &RetIndices,
314  SmallVectorImpl<unsigned> &CallIndices,
315  bool AllowDifferingSizes,
316  const TargetLoweringBase &TLI,
317  const DataLayout &DL) {
318 
319  // Trace the sub-value needed by the return value as far back up the graph as
320  // possible, in the hope that it will intersect with the value produced by the
321  // call. In the simple case with no "returned" attribute, the hope is actually
322  // that we end up back at the tail call instruction itself.
323  unsigned BitsRequired = UINT_MAX;
324  RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL);
325 
326  // If this slot in the value returned is undef, it doesn't matter what the
327  // call puts there, it'll be fine.
328  if (isa<UndefValue>(RetVal))
329  return true;
330 
331  // Now do a similar search up through the graph to find where the value
332  // actually returned by the "tail call" comes from. In the simple case without
333  // a "returned" attribute, the search will be blocked immediately and the loop
334  // a Noop.
335  unsigned BitsProvided = UINT_MAX;
336  CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL);
337 
338  // There's no hope if we can't actually trace them to (the same part of!) the
339  // same value.
340  if (CallVal != RetVal || CallIndices != RetIndices)
341  return false;
342 
343  // However, intervening truncates may have made the call non-tail. Make sure
344  // all the bits that are needed by the "ret" have been provided by the "tail
345  // call". FIXME: with sufficiently cunning bit-tracking, we could look through
346  // extensions too.
347  if (BitsProvided < BitsRequired ||
348  (!AllowDifferingSizes && BitsProvided != BitsRequired))
349  return false;
350 
351  return true;
352 }
353 
354 /// For an aggregate type, determine whether a given index is within bounds or
355 /// not.
356 static bool indexReallyValid(CompositeType *T, unsigned Idx) {
357  if (ArrayType *AT = dyn_cast<ArrayType>(T))
358  return Idx < AT->getNumElements();
359 
360  return Idx < cast<StructType>(T)->getNumElements();
361 }
362 
363 /// Move the given iterators to the next leaf type in depth first traversal.
364 ///
365 /// Performs a depth-first traversal of the type as specified by its arguments,
366 /// stopping at the next leaf node (which may be a legitimate scalar type or an
367 /// empty struct or array).
368 ///
369 /// @param SubTypes List of the partial components making up the type from
370 /// outermost to innermost non-empty aggregate. The element currently
371 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
372 ///
373 /// @param Path Set of extractvalue indices leading from the outermost type
374 /// (SubTypes[0]) to the leaf node currently represented.
375 ///
376 /// @returns true if a new type was found, false otherwise. Calling this
377 /// function again on a finished iterator will repeatedly return
378 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
379 /// aggregate or a non-aggregate
382  // First march back up the tree until we can successfully increment one of the
383  // coordinates in Path.
384  while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
385  Path.pop_back();
386  SubTypes.pop_back();
387  }
388 
389  // If we reached the top, then the iterator is done.
390  if (Path.empty())
391  return false;
392 
393  // We know there's *some* valid leaf now, so march back down the tree picking
394  // out the left-most element at each node.
395  ++Path.back();
396  Type *DeeperType = SubTypes.back()->getTypeAtIndex(Path.back());
397  while (DeeperType->isAggregateType()) {
398  CompositeType *CT = cast<CompositeType>(DeeperType);
399  if (!indexReallyValid(CT, 0))
400  return true;
401 
402  SubTypes.push_back(CT);
403  Path.push_back(0);
404 
405  DeeperType = CT->getTypeAtIndex(0U);
406  }
407 
408  return true;
409 }
410 
411 /// Find the first non-empty, scalar-like type in Next and setup the iterator
412 /// components.
413 ///
414 /// Assuming Next is an aggregate of some kind, this function will traverse the
415 /// tree from left to right (i.e. depth-first) looking for the first
416 /// non-aggregate type which will play a role in function return.
417 ///
418 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
419 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
420 /// i32 in that type.
421 static bool firstRealType(Type *Next,
424  // First initialise the iterator components to the first "leaf" node
425  // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
426  // despite nominally being an aggregate).
427  while (Next->isAggregateType() &&
428  indexReallyValid(cast<CompositeType>(Next), 0)) {
429  SubTypes.push_back(cast<CompositeType>(Next));
430  Path.push_back(0);
431  Next = cast<CompositeType>(Next)->getTypeAtIndex(0U);
432  }
433 
434  // If there's no Path now, Next was originally scalar already (or empty
435  // leaf). We're done.
436  if (Path.empty())
437  return true;
438 
439  // Otherwise, use normal iteration to keep looking through the tree until we
440  // find a non-aggregate type.
441  while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType()) {
442  if (!advanceToNextLeafType(SubTypes, Path))
443  return false;
444  }
445 
446  return true;
447 }
448 
449 /// Set the iterator data-structures to the next non-empty, non-aggregate
450 /// subtype.
453  do {
454  if (!advanceToNextLeafType(SubTypes, Path))
455  return false;
456 
457  assert(!Path.empty() && "found a leaf but didn't set the path?");
458  } while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType());
459 
460  return true;
461 }
462 
463 
464 /// Test if the given instruction is in a position to be optimized
465 /// with a tail-call. This roughly means that it's in a block with
466 /// a return and there's nothing that needs to be scheduled
467 /// between it and the return.
468 ///
469 /// This function only tests target-independent requirements.
471  const Instruction *I = CS.getInstruction();
472  const BasicBlock *ExitBB = I->getParent();
473  const Instruction *Term = ExitBB->getTerminator();
474  const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
475 
476  // The block must end in a return statement or unreachable.
477  //
478  // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
479  // an unreachable, for now. The way tailcall optimization is currently
480  // implemented means it will add an epilogue followed by a jump. That is
481  // not profitable. Also, if the callee is a special function (e.g.
482  // longjmp on x86), it can end up causing miscompilation that has not
483  // been fully understood.
484  if (!Ret &&
485  (!TM.Options.GuaranteedTailCallOpt || !isa<UnreachableInst>(Term)))
486  return false;
487 
488  // If I will have a chain, make sure no other instruction that will have a
489  // chain interposes between I and the return.
490  if (I->mayHaveSideEffects() || I->mayReadFromMemory() ||
492  for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) {
493  if (&*BBI == I)
494  break;
495  // Debug info intrinsics do not get in the way of tail call optimization.
496  if (isa<DbgInfoIntrinsic>(BBI))
497  continue;
498  // A lifetime end intrinsic should not stop tail call optimization.
499  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
500  if (II->getIntrinsicID() == Intrinsic::lifetime_end)
501  continue;
502  if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
504  return false;
505  }
506 
507  const Function *F = ExitBB->getParent();
509  F, I, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering());
510 }
511 
513  const ReturnInst *Ret,
514  const TargetLoweringBase &TLI,
515  bool *AllowDifferingSizes) {
516  // ADS may be null, so don't write to it directly.
517  bool DummyADS;
518  bool &ADS = AllowDifferingSizes ? *AllowDifferingSizes : DummyADS;
519  ADS = true;
520 
522  AttrBuilder CalleeAttrs(cast<CallInst>(I)->getAttributes(),
524 
525  // NoAlias and NonNull are completely benign as far as calling convention
526  // goes, they shouldn't affect whether the call is a tail call.
527  CallerAttrs.removeAttribute(Attribute::NoAlias);
528  CalleeAttrs.removeAttribute(Attribute::NoAlias);
529  CallerAttrs.removeAttribute(Attribute::NonNull);
530  CalleeAttrs.removeAttribute(Attribute::NonNull);
531 
532  if (CallerAttrs.contains(Attribute::ZExt)) {
533  if (!CalleeAttrs.contains(Attribute::ZExt))
534  return false;
535 
536  ADS = false;
537  CallerAttrs.removeAttribute(Attribute::ZExt);
538  CalleeAttrs.removeAttribute(Attribute::ZExt);
539  } else if (CallerAttrs.contains(Attribute::SExt)) {
540  if (!CalleeAttrs.contains(Attribute::SExt))
541  return false;
542 
543  ADS = false;
544  CallerAttrs.removeAttribute(Attribute::SExt);
545  CalleeAttrs.removeAttribute(Attribute::SExt);
546  }
547 
548  // Drop sext and zext return attributes if the result is not used.
549  // This enables tail calls for code like:
550  //
551  // define void @caller() {
552  // entry:
553  // %unused_result = tail call zeroext i1 @callee()
554  // br label %retlabel
555  // retlabel:
556  // ret void
557  // }
558  if (I->use_empty()) {
559  CalleeAttrs.removeAttribute(Attribute::SExt);
560  CalleeAttrs.removeAttribute(Attribute::ZExt);
561  }
562 
563  // If they're still different, there's some facet we don't understand
564  // (currently only "inreg", but in future who knows). It may be OK but the
565  // only safe option is to reject the tail call.
566  return CallerAttrs == CalleeAttrs;
567 }
568 
570  const Instruction *I,
571  const ReturnInst *Ret,
572  const TargetLoweringBase &TLI) {
573  // If the block ends with a void return or unreachable, it doesn't matter
574  // what the call's return type is.
575  if (!Ret || Ret->getNumOperands() == 0) return true;
576 
577  // If the return value is undef, it doesn't matter what the call's
578  // return type is.
579  if (isa<UndefValue>(Ret->getOperand(0))) return true;
580 
581  // Make sure the attributes attached to each return are compatible.
582  bool AllowDifferingSizes;
583  if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
584  return false;
585 
586  const Value *RetVal = Ret->getOperand(0), *CallVal = I;
587  // Intrinsic like llvm.memcpy has no return value, but the expanded
588  // libcall may or may not have return value. On most platforms, it
589  // will be expanded as memcpy in libc, which returns the first
590  // argument. On other platforms like arm-none-eabi, memcpy may be
591  // expanded as library call without return value, like __aeabi_memcpy.
592  const CallInst *Call = cast<CallInst>(I);
593  if (Function *F = Call->getCalledFunction()) {
594  Intrinsic::ID IID = F->getIntrinsicID();
595  if (((IID == Intrinsic::memcpy &&
596  TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")) ||
597  (IID == Intrinsic::memmove &&
598  TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")) ||
599  (IID == Intrinsic::memset &&
600  TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset"))) &&
601  RetVal == Call->getArgOperand(0))
602  return true;
603  }
604 
605  SmallVector<unsigned, 4> RetPath, CallPath;
606  SmallVector<CompositeType *, 4> RetSubTypes, CallSubTypes;
607 
608  bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
609  bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
610 
611  // Nothing's actually returned, it doesn't matter what the callee put there
612  // it's a valid tail call.
613  if (RetEmpty)
614  return true;
615 
616  // Iterate pairwise through each of the value types making up the tail call
617  // and the corresponding return. For each one we want to know whether it's
618  // essentially going directly from the tail call to the ret, via operations
619  // that end up not generating any code.
620  //
621  // We allow a certain amount of covariance here. For example it's permitted
622  // for the tail call to define more bits than the ret actually cares about
623  // (e.g. via a truncate).
624  do {
625  if (CallEmpty) {
626  // We've exhausted the values produced by the tail call instruction, the
627  // rest are essentially undef. The type doesn't really matter, but we need
628  // *something*.
629  Type *SlotType = RetSubTypes.back()->getTypeAtIndex(RetPath.back());
630  CallVal = UndefValue::get(SlotType);
631  }
632 
633  // The manipulations performed when we're looking through an insertvalue or
634  // an extractvalue would happen at the front of the RetPath list, so since
635  // we have to copy it anyway it's more efficient to create a reversed copy.
636  SmallVector<unsigned, 4> TmpRetPath(RetPath.rbegin(), RetPath.rend());
637  SmallVector<unsigned, 4> TmpCallPath(CallPath.rbegin(), CallPath.rend());
638 
639  // Finally, we can check whether the value produced by the tail call at this
640  // index is compatible with the value we return.
641  if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
642  AllowDifferingSizes, TLI,
643  F->getParent()->getDataLayout()))
644  return false;
645 
646  CallEmpty = !nextRealType(CallSubTypes, CallPath);
647  } while(nextRealType(RetSubTypes, RetPath));
648 
649  return true;
650 }
651 
653  DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope,
654  const MachineBasicBlock *MBB) {
656  while (!Worklist.empty()) {
657  const MachineBasicBlock *Visiting = Worklist.pop_back_val();
658  // Don't follow blocks which start new scopes.
659  if (Visiting->isEHPad() && Visiting != MBB)
660  continue;
661 
662  // Add this MBB to our scope.
663  auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope));
664 
665  // Don't revisit blocks.
666  if (!P.second) {
667  assert(P.first->second == EHScope && "MBB is part of two scopes!");
668  continue;
669  }
670 
671  // Returns are boundaries where scope transfer can occur, don't follow
672  // successors.
673  if (Visiting->isEHScopeReturnBlock())
674  continue;
675 
676  for (const MachineBasicBlock *Succ : Visiting->successors())
677  Worklist.push_back(Succ);
678  }
679 }
680 
684 
685  // We don't have anything to do if there aren't any EH pads.
686  if (!MF.hasEHScopes())
687  return EHScopeMembership;
688 
689  int EntryBBNumber = MF.front().getNumber();
690  bool IsSEH = isAsynchronousEHPersonality(
692 
698  for (const MachineBasicBlock &MBB : MF) {
699  if (MBB.isEHScopeEntry()) {
700  EHScopeBlocks.push_back(&MBB);
701  } else if (IsSEH && MBB.isEHPad()) {
702  SEHCatchPads.push_back(&MBB);
703  } else if (MBB.pred_empty()) {
704  UnreachableBlocks.push_back(&MBB);
705  }
706 
707  MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator();
708 
709  // CatchPads are not scopes for SEH so do not consider CatchRet to
710  // transfer control to another scope.
711  if (MBBI == MBB.end() || MBBI->getOpcode() != TII->getCatchReturnOpcode())
712  continue;
713 
714  // FIXME: SEH CatchPads are not necessarily in the parent function:
715  // they could be inside a finally block.
716  const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB();
717  const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB();
718  CatchRetSuccessors.push_back(
719  {Successor, IsSEH ? EntryBBNumber : SuccessorColor->getNumber()});
720  }
721 
722  // We don't have anything to do if there aren't any EH pads.
723  if (EHScopeBlocks.empty())
724  return EHScopeMembership;
725 
726  // Identify all the basic blocks reachable from the function entry.
727  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front());
728  // All blocks not part of a scope are in the parent function.
729  for (const MachineBasicBlock *MBB : UnreachableBlocks)
730  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
731  // Next, identify all the blocks inside the scopes.
732  for (const MachineBasicBlock *MBB : EHScopeBlocks)
733  collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB);
734  // SEH CatchPads aren't really scopes, handle them separately.
735  for (const MachineBasicBlock *MBB : SEHCatchPads)
736  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
737  // Finally, identify all the targets of a catchret.
738  for (std::pair<const MachineBasicBlock *, int> CatchRetPair :
739  CatchRetSuccessors)
740  collectEHScopeMembers(EHScopeMembership, CatchRetPair.second,
741  CatchRetPair.first);
742  return EHScopeMembership;
743 }
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:197
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:587
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:671
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:232
The two locations do not alias at all.
Definition: AliasAnalysis.h:83
unsigned less than
Definition: InstrTypes.h:670
0 1 0 0 True if ordered and less than
Definition: InstrTypes.h:651
Offsets
Offsets in bytes from the start of the input buffer.
Definition: SIInstrInfo.h:1024
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:661
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:362
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:117
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:569
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:1134
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:656
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:528
1 0 0 0 True if unordered: isnan(X) | isnan(Y)
Definition: InstrTypes.h:655
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:370
const HexagonInstrInfo * TII
Class to represent struct types.
Definition: DerivedTypes.h:200
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:652
InstrTy * getInstruction() const
Definition: CallSite.h:91
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.
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:368
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:958
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:160
Type::subtype_iterator element_iterator
Definition: DerivedTypes.h:300
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:138
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:356
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:451
#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
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:223
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:561
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:645
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:421
0 1 1 1 True if ordered (no nans)
Definition: InstrTypes.h:654
1 1 1 1 Always true (always folded)
Definition: InstrTypes.h:662
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:1414
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:528
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:312
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:660
#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:470
signed greater than
Definition: InstrTypes.h:672
std::vector< ConstraintInfo > ConstraintInfoVector
Definition: InlineAsm.h:115
This base class for TargetLowering contains the SelectionDAG-independent parts that can be used from ...
0 0 1 0 True if ordered and greater than
Definition: InstrTypes.h:649
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:214
iterator end()
Definition: BasicBlock.h:270
1 1 0 0 True if unordered or less than
Definition: InstrTypes.h:659
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:674
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:379
DenseMap< const MachineBasicBlock *, int > getEHScopeMembership(const MachineFunction &MF)
Definition: Analysis.cpp:682
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:380
ISD::CondCode getFCmpCodeWithoutNaN(ISD::CondCode CC)
getFCmpCodeWithoutNaN - Given an ISD condition code comparing floats, return the equivalent code if w...
Definition: Analysis.cpp:182
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:675
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:392
Common super class of ArrayType, StructType and VectorType.
Definition: DerivedTypes.h:161
static void collectEHScopeMembers(DenseMap< const MachineBasicBlock *, int > &EHScopeMembership, int EHScope, const MachineBasicBlock *MBB)
Definition: Analysis.cpp:652
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:435
reverse_iterator rend() const
Definition: ArrayRef.h:140
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:550
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:669
TargetOptions Options
Definition: TargetMachine.h:96
Establish a view to a call site for examination.
Definition: CallSite.h:886
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation.
Definition: InstrTypes.h:1180
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:653
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:322
1 0 1 0 True if unordered or greater than
Definition: InstrTypes.h:657
static EVT getEVT(Type *Ty, bool HandleUnknown=false)
Return the value type corresponding to the specified type.
Definition: ValueTypes.cpp:308
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:648
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:1298
1 0 1 1 True if unordered, greater than, or equal
Definition: InstrTypes.h:658
Primary interface to the complete machine description for the target machine.
Definition: TargetMachine.h:58
unsigned greater than
Definition: InstrTypes.h:668
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:512
0 0 1 1 True if ordered and greater than or equal
Definition: InstrTypes.h:650
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:647
signed greater or equal
Definition: InstrTypes.h:673
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:349