LLVM  8.0.0svn
Analysis.cpp
Go to the documentation of this file.
1 //===-- Analysis.cpp - CodeGen LLVM IR Analysis Utilities -----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines several CodeGen-specific LLVM IR analysis utilities.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/DerivedTypes.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Module.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.
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,
87  uint64_t StartingOffset) {
88  // Given a struct type, recursively traverse the elements.
89  if (StructType *STy = dyn_cast<StructType>(Ty)) {
90  const StructLayout *SL = DL.getStructLayout(STy);
91  for (StructType::element_iterator EB = STy->element_begin(),
92  EI = EB,
93  EE = STy->element_end();
94  EI != EE; ++EI)
95  ComputeValueVTs(TLI, DL, *EI, ValueVTs, Offsets,
96  StartingOffset + SL->getElementOffset(EI - EB));
97  return;
98  }
99  // Given an array type, recursively traverse the elements.
100  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
101  Type *EltTy = ATy->getElementType();
102  uint64_t EltSize = DL.getTypeAllocSize(EltTy);
103  for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
104  ComputeValueVTs(TLI, DL, EltTy, ValueVTs, Offsets,
105  StartingOffset + i * EltSize);
106  return;
107  }
108  // Interpret void as zero return values.
109  if (Ty->isVoidTy())
110  return;
111  // Base case: we can get an EVT for this LLVM IR type.
112  ValueVTs.push_back(TLI.getValueType(DL, Ty));
113  if (Offsets)
114  Offsets->push_back(StartingOffset);
115 }
116 
117 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
119  V = V->stripPointerCasts();
120  GlobalValue *GV = dyn_cast<GlobalValue>(V);
122 
123  if (Var && Var->getName() == "llvm.eh.catch.all.value") {
124  assert(Var->hasInitializer() &&
125  "The EH catch-all value must have an initializer");
126  Value *Init = Var->getInitializer();
127  GV = dyn_cast<GlobalValue>(Init);
128  if (!GV) V = cast<ConstantPointerNull>(Init);
129  }
130 
131  assert((GV || isa<ConstantPointerNull>(V)) &&
132  "TypeInfo must be a global variable or NULL");
133  return GV;
134 }
135 
136 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
137 /// processed uses a memory 'm' constraint.
138 bool
140  const TargetLowering &TLI) {
141  for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
142  InlineAsm::ConstraintInfo &CI = CInfos[i];
143  for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
145  if (CType == TargetLowering::C_Memory)
146  return true;
147  }
148 
149  // Indirect operand accesses access memory.
150  if (CI.isIndirect)
151  return true;
152  }
153 
154  return false;
155 }
156 
157 /// getFCmpCondCode - Return the ISD condition code corresponding to
158 /// the given LLVM IR floating-point condition code. This includes
159 /// consideration of global floating-point math flags.
160 ///
162  switch (Pred) {
163  case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
164  case FCmpInst::FCMP_OEQ: return ISD::SETOEQ;
165  case FCmpInst::FCMP_OGT: return ISD::SETOGT;
166  case FCmpInst::FCMP_OGE: return ISD::SETOGE;
167  case FCmpInst::FCMP_OLT: return ISD::SETOLT;
168  case FCmpInst::FCMP_OLE: return ISD::SETOLE;
169  case FCmpInst::FCMP_ONE: return ISD::SETONE;
170  case FCmpInst::FCMP_ORD: return ISD::SETO;
171  case FCmpInst::FCMP_UNO: return ISD::SETUO;
172  case FCmpInst::FCMP_UEQ: return ISD::SETUEQ;
173  case FCmpInst::FCMP_UGT: return ISD::SETUGT;
174  case FCmpInst::FCMP_UGE: return ISD::SETUGE;
175  case FCmpInst::FCMP_ULT: return ISD::SETULT;
176  case FCmpInst::FCMP_ULE: return ISD::SETULE;
177  case FCmpInst::FCMP_UNE: return ISD::SETUNE;
178  case FCmpInst::FCMP_TRUE: return ISD::SETTRUE;
179  default: llvm_unreachable("Invalid FCmp predicate opcode!");
180  }
181 }
182 
184  switch (CC) {
185  case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
186  case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
187  case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
188  case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
189  case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
190  case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
191  default: return CC;
192  }
193 }
194 
195 /// getICmpCondCode - Return the ISD condition code corresponding to
196 /// the given LLVM IR integer condition code.
197 ///
199  switch (Pred) {
200  case ICmpInst::ICMP_EQ: return ISD::SETEQ;
201  case ICmpInst::ICMP_NE: return ISD::SETNE;
202  case ICmpInst::ICMP_SLE: return ISD::SETLE;
203  case ICmpInst::ICMP_ULE: return ISD::SETULE;
204  case ICmpInst::ICMP_SGE: return ISD::SETGE;
205  case ICmpInst::ICMP_UGE: return ISD::SETUGE;
206  case ICmpInst::ICMP_SLT: return ISD::SETLT;
207  case ICmpInst::ICMP_ULT: return ISD::SETULT;
208  case ICmpInst::ICMP_SGT: return ISD::SETGT;
209  case ICmpInst::ICMP_UGT: return ISD::SETUGT;
210  default:
211  llvm_unreachable("Invalid ICmp predicate opcode!");
212  }
213 }
214 
215 static bool isNoopBitcast(Type *T1, Type *T2,
216  const TargetLoweringBase& TLI) {
217  return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
218  (isa<VectorType>(T1) && isa<VectorType>(T2) &&
219  TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
220 }
221 
222 /// Look through operations that will be free to find the earliest source of
223 /// this value.
224 ///
225 /// @param ValLoc If V has aggegate type, we will be interested in a particular
226 /// scalar component. This records its address; the reverse of this list gives a
227 /// sequence of indices appropriate for an extractvalue to locate the important
228 /// value. This value is updated during the function and on exit will indicate
229 /// similar information for the Value returned.
230 ///
231 /// @param DataBits If this function looks through truncate instructions, this
232 /// will record the smallest size attained.
233 static const Value *getNoopInput(const Value *V,
235  unsigned &DataBits,
236  const TargetLoweringBase &TLI,
237  const DataLayout &DL) {
238  while (true) {
239  // Try to look through V1; if V1 is not an instruction, it can't be looked
240  // through.
241  const Instruction *I = dyn_cast<Instruction>(V);
242  if (!I || I->getNumOperands() == 0) return V;
243  const Value *NoopInput = nullptr;
244 
245  Value *Op = I->getOperand(0);
246  if (isa<BitCastInst>(I)) {
247  // Look through truly no-op bitcasts.
248  if (isNoopBitcast(Op->getType(), I->getType(), TLI))
249  NoopInput = Op;
250  } else if (isa<GetElementPtrInst>(I)) {
251  // Look through getelementptr
252  if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
253  NoopInput = Op;
254  } else if (isa<IntToPtrInst>(I)) {
255  // Look through inttoptr.
256  // Make sure this isn't a truncating or extending cast. We could
257  // support this eventually, but don't bother for now.
258  if (!isa<VectorType>(I->getType()) &&
259  DL.getPointerSizeInBits() ==
260  cast<IntegerType>(Op->getType())->getBitWidth())
261  NoopInput = Op;
262  } else if (isa<PtrToIntInst>(I)) {
263  // Look through ptrtoint.
264  // Make sure this isn't a truncating or extending cast. We could
265  // support this eventually, but don't bother for now.
266  if (!isa<VectorType>(I->getType()) &&
267  DL.getPointerSizeInBits() ==
268  cast<IntegerType>(I->getType())->getBitWidth())
269  NoopInput = Op;
270  } else if (isa<TruncInst>(I) &&
271  TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
272  DataBits = std::min(DataBits, I->getType()->getPrimitiveSizeInBits());
273  NoopInput = Op;
274  } else if (auto CS = ImmutableCallSite(I)) {
275  const Value *ReturnedOp = CS.getReturnedArgOperand();
276  if (ReturnedOp && isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI))
277  NoopInput = ReturnedOp;
278  } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
279  // Value may come from either the aggregate or the scalar
280  ArrayRef<unsigned> InsertLoc = IVI->getIndices();
281  if (ValLoc.size() >= InsertLoc.size() &&
282  std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) {
283  // The type being inserted is a nested sub-type of the aggregate; we
284  // have to remove those initial indices to get the location we're
285  // interested in for the operand.
286  ValLoc.resize(ValLoc.size() - InsertLoc.size());
287  NoopInput = IVI->getInsertedValueOperand();
288  } else {
289  // The struct we're inserting into has the value we're interested in, no
290  // change of address.
291  NoopInput = Op;
292  }
293  } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
294  // The part we're interested in will inevitably be some sub-section of the
295  // previous aggregate. Combine the two paths to obtain the true address of
296  // our element.
297  ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
298  ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
299  NoopInput = Op;
300  }
301  // Terminate if we couldn't find anything to look through.
302  if (!NoopInput)
303  return V;
304 
305  V = NoopInput;
306  }
307 }
308 
309 /// Return true if this scalar return value only has bits discarded on its path
310 /// from the "tail call" to the "ret". This includes the obvious noop
311 /// instructions handled by getNoopInput above as well as free truncations (or
312 /// extensions prior to the call).
313 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
314  SmallVectorImpl<unsigned> &RetIndices,
315  SmallVectorImpl<unsigned> &CallIndices,
316  bool AllowDifferingSizes,
317  const TargetLoweringBase &TLI,
318  const DataLayout &DL) {
319 
320  // Trace the sub-value needed by the return value as far back up the graph as
321  // possible, in the hope that it will intersect with the value produced by the
322  // call. In the simple case with no "returned" attribute, the hope is actually
323  // that we end up back at the tail call instruction itself.
324  unsigned BitsRequired = UINT_MAX;
325  RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL);
326 
327  // If this slot in the value returned is undef, it doesn't matter what the
328  // call puts there, it'll be fine.
329  if (isa<UndefValue>(RetVal))
330  return true;
331 
332  // Now do a similar search up through the graph to find where the value
333  // actually returned by the "tail call" comes from. In the simple case without
334  // a "returned" attribute, the search will be blocked immediately and the loop
335  // a Noop.
336  unsigned BitsProvided = UINT_MAX;
337  CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL);
338 
339  // There's no hope if we can't actually trace them to (the same part of!) the
340  // same value.
341  if (CallVal != RetVal || CallIndices != RetIndices)
342  return false;
343 
344  // However, intervening truncates may have made the call non-tail. Make sure
345  // all the bits that are needed by the "ret" have been provided by the "tail
346  // call". FIXME: with sufficiently cunning bit-tracking, we could look through
347  // extensions too.
348  if (BitsProvided < BitsRequired ||
349  (!AllowDifferingSizes && BitsProvided != BitsRequired))
350  return false;
351 
352  return true;
353 }
354 
355 /// For an aggregate type, determine whether a given index is within bounds or
356 /// not.
357 static bool indexReallyValid(CompositeType *T, unsigned Idx) {
358  if (ArrayType *AT = dyn_cast<ArrayType>(T))
359  return Idx < AT->getNumElements();
360 
361  return Idx < cast<StructType>(T)->getNumElements();
362 }
363 
364 /// Move the given iterators to the next leaf type in depth first traversal.
365 ///
366 /// Performs a depth-first traversal of the type as specified by its arguments,
367 /// stopping at the next leaf node (which may be a legitimate scalar type or an
368 /// empty struct or array).
369 ///
370 /// @param SubTypes List of the partial components making up the type from
371 /// outermost to innermost non-empty aggregate. The element currently
372 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
373 ///
374 /// @param Path Set of extractvalue indices leading from the outermost type
375 /// (SubTypes[0]) to the leaf node currently represented.
376 ///
377 /// @returns true if a new type was found, false otherwise. Calling this
378 /// function again on a finished iterator will repeatedly return
379 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
380 /// aggregate or a non-aggregate
383  // First march back up the tree until we can successfully increment one of the
384  // coordinates in Path.
385  while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
386  Path.pop_back();
387  SubTypes.pop_back();
388  }
389 
390  // If we reached the top, then the iterator is done.
391  if (Path.empty())
392  return false;
393 
394  // We know there's *some* valid leaf now, so march back down the tree picking
395  // out the left-most element at each node.
396  ++Path.back();
397  Type *DeeperType = SubTypes.back()->getTypeAtIndex(Path.back());
398  while (DeeperType->isAggregateType()) {
399  CompositeType *CT = cast<CompositeType>(DeeperType);
400  if (!indexReallyValid(CT, 0))
401  return true;
402 
403  SubTypes.push_back(CT);
404  Path.push_back(0);
405 
406  DeeperType = CT->getTypeAtIndex(0U);
407  }
408 
409  return true;
410 }
411 
412 /// Find the first non-empty, scalar-like type in Next and setup the iterator
413 /// components.
414 ///
415 /// Assuming Next is an aggregate of some kind, this function will traverse the
416 /// tree from left to right (i.e. depth-first) looking for the first
417 /// non-aggregate type which will play a role in function return.
418 ///
419 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
420 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
421 /// i32 in that type.
422 static bool firstRealType(Type *Next,
425  // First initialise the iterator components to the first "leaf" node
426  // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
427  // despite nominally being an aggregate).
428  while (Next->isAggregateType() &&
429  indexReallyValid(cast<CompositeType>(Next), 0)) {
430  SubTypes.push_back(cast<CompositeType>(Next));
431  Path.push_back(0);
432  Next = cast<CompositeType>(Next)->getTypeAtIndex(0U);
433  }
434 
435  // If there's no Path now, Next was originally scalar already (or empty
436  // leaf). We're done.
437  if (Path.empty())
438  return true;
439 
440  // Otherwise, use normal iteration to keep looking through the tree until we
441  // find a non-aggregate type.
442  while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType()) {
443  if (!advanceToNextLeafType(SubTypes, Path))
444  return false;
445  }
446 
447  return true;
448 }
449 
450 /// Set the iterator data-structures to the next non-empty, non-aggregate
451 /// subtype.
454  do {
455  if (!advanceToNextLeafType(SubTypes, Path))
456  return false;
457 
458  assert(!Path.empty() && "found a leaf but didn't set the path?");
459  } while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType());
460 
461  return true;
462 }
463 
464 
465 /// Test if the given instruction is in a position to be optimized
466 /// with a tail-call. This roughly means that it's in a block with
467 /// a return and there's nothing that needs to be scheduled
468 /// between it and the return.
469 ///
470 /// This function only tests target-independent requirements.
472  const Instruction *I = CS.getInstruction();
473  const BasicBlock *ExitBB = I->getParent();
474  const Instruction *Term = ExitBB->getTerminator();
475  const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
476 
477  // The block must end in a return statement or unreachable.
478  //
479  // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
480  // an unreachable, for now. The way tailcall optimization is currently
481  // implemented means it will add an epilogue followed by a jump. That is
482  // not profitable. Also, if the callee is a special function (e.g.
483  // longjmp on x86), it can end up causing miscompilation that has not
484  // been fully understood.
485  if (!Ret &&
486  (!TM.Options.GuaranteedTailCallOpt || !isa<UnreachableInst>(Term)))
487  return false;
488 
489  // If I will have a chain, make sure no other instruction that will have a
490  // chain interposes between I and the return.
491  if (I->mayHaveSideEffects() || I->mayReadFromMemory() ||
493  for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) {
494  if (&*BBI == I)
495  break;
496  // Debug info intrinsics do not get in the way of tail call optimization.
497  if (isa<DbgInfoIntrinsic>(BBI))
498  continue;
499  // A lifetime end intrinsic should not stop tail call optimization.
500  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
501  if (II->getIntrinsicID() == Intrinsic::lifetime_end)
502  continue;
503  if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
505  return false;
506  }
507 
508  const Function *F = ExitBB->getParent();
510  F, I, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering());
511 }
512 
514  const ReturnInst *Ret,
515  const TargetLoweringBase &TLI,
516  bool *AllowDifferingSizes) {
517  // ADS may be null, so don't write to it directly.
518  bool DummyADS;
519  bool &ADS = AllowDifferingSizes ? *AllowDifferingSizes : DummyADS;
520  ADS = true;
521 
523  AttrBuilder CalleeAttrs(cast<CallInst>(I)->getAttributes(),
525 
526  // NoAlias and NonNull are completely benign as far as calling convention
527  // goes, they shouldn't affect whether the call is a tail call.
528  CallerAttrs.removeAttribute(Attribute::NoAlias);
529  CalleeAttrs.removeAttribute(Attribute::NoAlias);
530  CallerAttrs.removeAttribute(Attribute::NonNull);
531  CalleeAttrs.removeAttribute(Attribute::NonNull);
532 
533  if (CallerAttrs.contains(Attribute::ZExt)) {
534  if (!CalleeAttrs.contains(Attribute::ZExt))
535  return false;
536 
537  ADS = false;
538  CallerAttrs.removeAttribute(Attribute::ZExt);
539  CalleeAttrs.removeAttribute(Attribute::ZExt);
540  } else if (CallerAttrs.contains(Attribute::SExt)) {
541  if (!CalleeAttrs.contains(Attribute::SExt))
542  return false;
543 
544  ADS = false;
545  CallerAttrs.removeAttribute(Attribute::SExt);
546  CalleeAttrs.removeAttribute(Attribute::SExt);
547  }
548 
549  // If they're still different, there's some facet we don't understand
550  // (currently only "inreg", but in future who knows). It may be OK but the
551  // only safe option is to reject the tail call.
552  return CallerAttrs == CalleeAttrs;
553 }
554 
556  const Instruction *I,
557  const ReturnInst *Ret,
558  const TargetLoweringBase &TLI) {
559  // If the block ends with a void return or unreachable, it doesn't matter
560  // what the call's return type is.
561  if (!Ret || Ret->getNumOperands() == 0) return true;
562 
563  // If the return value is undef, it doesn't matter what the call's
564  // return type is.
565  if (isa<UndefValue>(Ret->getOperand(0))) return true;
566 
567  // Make sure the attributes attached to each return are compatible.
568  bool AllowDifferingSizes;
569  if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
570  return false;
571 
572  const Value *RetVal = Ret->getOperand(0), *CallVal = I;
573  // Intrinsic like llvm.memcpy has no return value, but the expanded
574  // libcall may or may not have return value. On most platforms, it
575  // will be expanded as memcpy in libc, which returns the first
576  // argument. On other platforms like arm-none-eabi, memcpy may be
577  // expanded as library call without return value, like __aeabi_memcpy.
578  const CallInst *Call = cast<CallInst>(I);
579  if (Function *F = Call->getCalledFunction()) {
580  Intrinsic::ID IID = F->getIntrinsicID();
581  if (((IID == Intrinsic::memcpy &&
582  TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")) ||
583  (IID == Intrinsic::memmove &&
584  TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")) ||
585  (IID == Intrinsic::memset &&
586  TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset"))) &&
587  RetVal == Call->getArgOperand(0))
588  return true;
589  }
590 
591  SmallVector<unsigned, 4> RetPath, CallPath;
592  SmallVector<CompositeType *, 4> RetSubTypes, CallSubTypes;
593 
594  bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
595  bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
596 
597  // Nothing's actually returned, it doesn't matter what the callee put there
598  // it's a valid tail call.
599  if (RetEmpty)
600  return true;
601 
602  // Iterate pairwise through each of the value types making up the tail call
603  // and the corresponding return. For each one we want to know whether it's
604  // essentially going directly from the tail call to the ret, via operations
605  // that end up not generating any code.
606  //
607  // We allow a certain amount of covariance here. For example it's permitted
608  // for the tail call to define more bits than the ret actually cares about
609  // (e.g. via a truncate).
610  do {
611  if (CallEmpty) {
612  // We've exhausted the values produced by the tail call instruction, the
613  // rest are essentially undef. The type doesn't really matter, but we need
614  // *something*.
615  Type *SlotType = RetSubTypes.back()->getTypeAtIndex(RetPath.back());
616  CallVal = UndefValue::get(SlotType);
617  }
618 
619  // The manipulations performed when we're looking through an insertvalue or
620  // an extractvalue would happen at the front of the RetPath list, so since
621  // we have to copy it anyway it's more efficient to create a reversed copy.
622  SmallVector<unsigned, 4> TmpRetPath(RetPath.rbegin(), RetPath.rend());
623  SmallVector<unsigned, 4> TmpCallPath(CallPath.rbegin(), CallPath.rend());
624 
625  // Finally, we can check whether the value produced by the tail call at this
626  // index is compatible with the value we return.
627  if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
628  AllowDifferingSizes, TLI,
629  F->getParent()->getDataLayout()))
630  return false;
631 
632  CallEmpty = !nextRealType(CallSubTypes, CallPath);
633  } while(nextRealType(RetSubTypes, RetPath));
634 
635  return true;
636 }
637 
639  DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope,
640  const MachineBasicBlock *MBB) {
642  while (!Worklist.empty()) {
643  const MachineBasicBlock *Visiting = Worklist.pop_back_val();
644  // Don't follow blocks which start new scopes.
645  if (Visiting->isEHPad() && Visiting != MBB)
646  continue;
647 
648  // Add this MBB to our scope.
649  auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope));
650 
651  // Don't revisit blocks.
652  if (!P.second) {
653  assert(P.first->second == EHScope && "MBB is part of two scopes!");
654  continue;
655  }
656 
657  // Returns are boundaries where scope transfer can occur, don't follow
658  // successors.
659  if (Visiting->isEHScopeReturnBlock())
660  continue;
661 
662  for (const MachineBasicBlock *Succ : Visiting->successors())
663  Worklist.push_back(Succ);
664  }
665 }
666 
670 
671  // We don't have anything to do if there aren't any EH pads.
672  if (!MF.hasEHScopes())
673  return EHScopeMembership;
674 
675  int EntryBBNumber = MF.front().getNumber();
676  bool IsSEH = isAsynchronousEHPersonality(
678 
684  for (const MachineBasicBlock &MBB : MF) {
685  if (MBB.isEHScopeEntry()) {
686  EHScopeBlocks.push_back(&MBB);
687  } else if (IsSEH && MBB.isEHPad()) {
688  SEHCatchPads.push_back(&MBB);
689  } else if (MBB.pred_empty()) {
690  UnreachableBlocks.push_back(&MBB);
691  }
692 
693  MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator();
694 
695  // CatchPads are not scopes for SEH so do not consider CatchRet to
696  // transfer control to another scope.
697  if (MBBI == MBB.end() || MBBI->getOpcode() != TII->getCatchReturnOpcode())
698  continue;
699 
700  // FIXME: SEH CatchPads are not necessarily in the parent function:
701  // they could be inside a finally block.
702  const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB();
703  const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB();
704  CatchRetSuccessors.push_back(
705  {Successor, IsSEH ? EntryBBNumber : SuccessorColor->getNumber()});
706  }
707 
708  // We don't have anything to do if there aren't any EH pads.
709  if (EHScopeBlocks.empty())
710  return EHScopeMembership;
711 
712  // Identify all the basic blocks reachable from the function entry.
713  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front());
714  // All blocks not part of a scope are in the parent function.
715  for (const MachineBasicBlock *MBB : UnreachableBlocks)
716  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
717  // Next, identify all the blocks inside the scopes.
718  for (const MachineBasicBlock *MBB : EHScopeBlocks)
719  collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB);
720  // SEH CatchPads aren't really scopes, handle them separately.
721  for (const MachineBasicBlock *MBB : SEHCatchPads)
722  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
723  // Finally, identify all the targets of a catchret.
724  for (std::pair<const MachineBasicBlock *, int> CatchRetPair :
725  CatchRetSuccessors)
726  collectEHScopeMembers(EHScopeMembership, CatchRetPair.second,
727  CatchRetPair.first);
728  return EHScopeMembership;
729 }
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:111
ISD::CondCode getICmpCondCode(ICmpInst::Predicate Pred)
getICmpCondCode - Return the ISD condition code corresponding to the given LLVM IR integer condition ...
Definition: Analysis.cpp:198
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:24
iterator begin() const
Definition: ArrayRef.h:137
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:588
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:140
unsigned less or equal
Definition: InstrTypes.h:668
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:233
The two locations do not alias at all.
Definition: AliasAnalysis.h:85
unsigned less than
Definition: InstrTypes.h:667
0 1 0 0 True if ordered and less than
Definition: InstrTypes.h:648
Offsets
Offsets in bytes from the start of the input buffer.
Definition: SIInstrInfo.h:1015
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:658
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:360
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:138
GlobalValue * ExtractTypeInfo(Value *V)
ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
Definition: Analysis.cpp:118
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:555
ConstraintCodeVector Codes
Code - The constraint code, either the register name (in braces) or the constraint letter/number...
Definition: InlineAsm.h:149
iterator_range< succ_iterator > successors()
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1079
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:653
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:521
1 0 0 0 True if unordered: isnan(X) | isnan(Y)
Definition: InstrTypes.h:652
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:364
const HexagonInstrInfo * TII
Class to represent struct types.
Definition: DerivedTypes.h:201
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
0 1 0 1 True if ordered and less than or equal
Definition: InstrTypes.h:649
InstrTy * getInstruction() const
Definition: CallSite.h:92
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:245
#define T
Class to represent array types.
Definition: DerivedTypes.h:369
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:224
CondCode
ISD::CondCode enum - These are ordered carefully to make the bitfields below work out...
Definition: ISDOpcodes.h:952
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:161
Type::subtype_iterator element_iterator
Definition: DerivedTypes.h:301
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:84
Value * getOperand(unsigned i) const
Definition: User.h:170
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:139
TargetInstrInfo - Interface to description of machine instruction set.
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:141
static bool indexReallyValid(CompositeType *T, unsigned Idx)
For an aggregate type, determine whether a given index is within bounds or not.
Definition: Analysis.cpp:357
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:452
#define P(N)
LLVM Basic Block Representation.
Definition: BasicBlock.h:58
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:46
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:145
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:149
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:224
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:562
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:642
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:422
0 1 1 1 True if ordered (no nans)
Definition: InstrTypes.h:651
1 1 1 1 Always true (always folded)
Definition: InstrTypes.h:659
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:1389
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:530
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:313
const MachineBasicBlock & front() const
size_t size() const
Definition: SmallVector.h:53
1 1 0 1 True if unordered, less than, or equal
Definition: InstrTypes.h:657
#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:471
signed greater than
Definition: InstrTypes.h:669
std::vector< ConstraintInfo > ConstraintInfoVector
Definition: InlineAsm.h:116
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:646
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:192
static bool isNoopBitcast(Type *T1, Type *T2, const TargetLoweringBase &TLI)
Definition: Analysis.cpp:215
iterator end()
Definition: BasicBlock.h:271
1 1 0 0 True if unordered or less than
Definition: InstrTypes.h:656
Module.h This file contains the declarations for the Module class.
iterator end() const
Definition: ArrayRef.h:138
bool isAggregateType() const
Return true if the type is an aggregate type.
Definition: Type.h:258
signed less than
Definition: InstrTypes.h:671
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:381
DenseMap< const MachineBasicBlock *, int > getEHScopeMembership(const MachineFunction &MF)
Definition: Analysis.cpp:668
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:381
ISD::CondCode getFCmpCodeWithoutNaN(ISD::CondCode CC)
getFCmpCodeWithoutNaN - Given an ISD condition code comparing floats, return the equivalent code if w...
Definition: Analysis.cpp:183
Intrinsic::ID getIntrinsicID() const LLVM_READONLY
getIntrinsicID - This method returns the ID number of the specified function, or Intrinsic::not_intri...
Definition: Function.h:194
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:672
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:394
Common super class of ArrayType, StructType and VectorType.
Definition: DerivedTypes.h:162
static void collectEHScopeMembers(DenseMap< const MachineBasicBlock *, int > &EHScopeMembership, int EHScope, const MachineBasicBlock *MBB)
Definition: Analysis.cpp:638
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:428
reverse_iterator rend() const
Definition: ArrayRef.h:141
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:543
bool isEHPad() const
Returns true if the block is a landing pad.
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
unsigned greater or equal
Definition: InstrTypes.h:666
TargetOptions Options
Definition: TargetMachine.h:97
Establish a view to a call site for examination.
Definition: CallSite.h:711
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation.
Definition: InstrTypes.h:1112
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:107
#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:650
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:323
1 0 1 0 True if unordered or greater than
Definition: InstrTypes.h:654
static EVT getEVT(Type *Ty, bool HandleUnknown=false)
Return the value type corresponding to the specified type.
Definition: ValueTypes.cpp:309
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:115
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:645
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:566
LLVM Value Representation.
Definition: Value.h:73
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:655
Primary interface to the complete machine description for the target machine.
Definition: TargetMachine.h:59
unsigned greater than
Definition: InstrTypes.h:665
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
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:513
0 0 1 1 True if ordered and greater than or equal
Definition: InstrTypes.h:647
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:36
#define T1
0 0 0 0 Always false (always folded)
Definition: InstrTypes.h:644
signed greater or equal
Definition: InstrTypes.h:670
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
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:530
const BasicBlock * getParent() const
Definition: Instruction.h:67
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:351