Coverage Report

Created: 2019-07-24 05:18

/Users/buildslave/jenkins/workspace/clang-stage2-coverage-R/llvm/lib/CodeGen/Analysis.cpp
Line
Count
Source (jump to first uncovered line)
1
//===-- Analysis.cpp - CodeGen LLVM IR Analysis Utilities -----------------===//
2
//
3
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4
// See https://llvm.org/LICENSE.txt for license information.
5
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6
//
7
//===----------------------------------------------------------------------===//
8
//
9
// This file defines several CodeGen-specific LLVM IR analysis utilities.
10
//
11
//===----------------------------------------------------------------------===//
12
13
#include "llvm/CodeGen/Analysis.h"
14
#include "llvm/Analysis/ValueTracking.h"
15
#include "llvm/CodeGen/MachineFunction.h"
16
#include "llvm/CodeGen/TargetInstrInfo.h"
17
#include "llvm/CodeGen/TargetLowering.h"
18
#include "llvm/CodeGen/TargetSubtargetInfo.h"
19
#include "llvm/IR/DataLayout.h"
20
#include "llvm/IR/DerivedTypes.h"
21
#include "llvm/IR/Function.h"
22
#include "llvm/IR/Instructions.h"
23
#include "llvm/IR/IntrinsicInst.h"
24
#include "llvm/IR/LLVMContext.h"
25
#include "llvm/IR/Module.h"
26
#include "llvm/Support/ErrorHandling.h"
27
#include "llvm/Support/MathExtras.h"
28
#include "llvm/Transforms/Utils/GlobalStatus.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.
35
unsigned llvm::ComputeLinearIndex(Type *Ty,
36
                                  const unsigned *Indices,
37
                                  const unsigned *IndicesEnd,
38
79.0k
                                  unsigned CurIndex) {
39
79.0k
  // Base case: We're done.
40
79.0k
  if (Indices && 
Indices == IndicesEnd55.8k
)
41
27.8k
    return CurIndex;
42
51.2k
43
51.2k
  // Given a struct type, recursively traverse the elements.
44
51.2k
  if (StructType *STy = dyn_cast<StructType>(Ty)) {
45
26.3k
    for (StructType::element_iterator EB = STy->element_begin(),
46
26.3k
                                      EI = EB,
47
26.3k
                                      EE = STy->element_end();
48
47.8k
        EI != EE; 
++EI21.5k
) {
49
47.8k
      if (Indices && 
*Indices == unsigned(EI - EB)47.8k
)
50
26.3k
        return ComputeLinearIndex(*EI, Indices+1, IndicesEnd, CurIndex);
51
21.5k
      CurIndex = ComputeLinearIndex(*EI, nullptr, nullptr, CurIndex);
52
21.5k
    }
53
26.3k
    assert(!Indices && "Unexpected out of bound");
54
13
    return CurIndex;
55
24.9k
  }
56
24.9k
  // Given an array type, recursively traverse the elements.
57
24.9k
  else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
58
1.71k
    Type *EltTy = ATy->getElementType();
59
1.71k
    unsigned NumElts = ATy->getNumElements();
60
1.71k
    // Compute the Linear offset when jumping one element of the array
61
1.71k
    unsigned EltLinearOffset = ComputeLinearIndex(EltTy, nullptr, nullptr, 0);
62
1.71k
    if (Indices) {
63
1.70k
      assert(*Indices < NumElts && "Unexpected out of bound");
64
1.70k
      // If the indice is inside the array, compute the index to the requested
65
1.70k
      // elt and recurse inside the element with the end of the indices list
66
1.70k
      CurIndex += EltLinearOffset* *Indices;
67
1.70k
      return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex);
68
1.70k
    }
69
8
    CurIndex += EltLinearOffset*NumElts;
70
8
    return CurIndex;
71
8
  }
72
23.2k
  // We haven't found the type we're looking for, so keep searching.
73
23.2k
  return CurIndex + 1;
74
23.2k
}
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,
85
                           SmallVectorImpl<EVT> *MemVTs,
86
                           SmallVectorImpl<uint64_t> *Offsets,
87
14.4M
                           uint64_t StartingOffset) {
88
14.4M
  // Given a struct type, recursively traverse the elements.
89
14.4M
  if (StructType *STy = dyn_cast<StructType>(Ty)) {
90
23.7k
    const StructLayout *SL = DL.getStructLayout(STy);
91
23.7k
    for (StructType::element_iterator EB = STy->element_begin(),
92
23.7k
                                      EI = EB,
93
23.7k
                                      EE = STy->element_end();
94
90.8k
         EI != EE; 
++EI67.1k
)
95
67.1k
      ComputeValueVTs(TLI, DL, *EI, ValueVTs, MemVTs, Offsets,
96
67.1k
                      StartingOffset + SL->getElementOffset(EI - EB));
97
23.7k
    return;
98
23.7k
  }
99
14.3M
  // Given an array type, recursively traverse the elements.
100
14.3M
  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
101
8.07k
    Type *EltTy = ATy->getElementType();
102
8.07k
    uint64_t EltSize = DL.getTypeAllocSize(EltTy);
103
52.7k
    for (unsigned i = 0, e = ATy->getNumElements(); i != e; 
++i44.6k
)
104
44.6k
      ComputeValueVTs(TLI, DL, EltTy, ValueVTs, MemVTs, Offsets,
105
44.6k
                      StartingOffset + i * EltSize);
106
8.07k
    return;
107
8.07k
  }
108
14.3M
  // Interpret void as zero return values.
109
14.3M
  if (Ty->isVoidTy())
110
504k
    return;
111
13.8M
  // Base case: we can get an EVT for this LLVM IR type.
112
13.8M
  ValueVTs.push_back(TLI.getValueType(DL, Ty));
113
13.8M
  if (MemVTs)
114
1.16M
    MemVTs->push_back(TLI.getMemValueType(DL, Ty));
115
13.8M
  if (Offsets)
116
5.99M
    Offsets->push_back(StartingOffset);
117
13.8M
}
118
119
void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
120
                           Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
121
                           SmallVectorImpl<uint64_t> *Offsets,
122
13.1M
                           uint64_t StartingOffset) {
123
13.1M
  return ComputeValueVTs(TLI, DL, Ty, ValueVTs, /*MemVTs=*/nullptr, Offsets,
124
13.1M
                         StartingOffset);
125
13.1M
}
126
127
void llvm::computeValueLLTs(const DataLayout &DL, Type &Ty,
128
                            SmallVectorImpl<LLT> &ValueTys,
129
                            SmallVectorImpl<uint64_t> *Offsets,
130
9.95M
                            uint64_t StartingOffset) {
131
9.95M
  // Given a struct type, recursively traverse the elements.
132
9.95M
  if (StructType *STy = dyn_cast<StructType>(&Ty)) {
133
19.0k
    const StructLayout *SL = DL.getStructLayout(STy);
134
57.6k
    for (unsigned I = 0, E = STy->getNumElements(); I != E; 
++I38.5k
)
135
38.5k
      computeValueLLTs(DL, *STy->getElementType(I), ValueTys, Offsets,
136
38.5k
                       StartingOffset + SL->getElementOffset(I));
137
19.0k
    return;
138
19.0k
  }
139
9.93M
  // Given an array type, recursively traverse the elements.
140
9.93M
  if (ArrayType *ATy = dyn_cast<ArrayType>(&Ty)) {
141
2.02k
    Type *EltTy = ATy->getElementType();
142
2.02k
    uint64_t EltSize = DL.getTypeAllocSize(EltTy);
143
7.80k
    for (unsigned i = 0, e = ATy->getNumElements(); i != e; 
++i5.77k
)
144
5.77k
      computeValueLLTs(DL, *EltTy, ValueTys, Offsets,
145
5.77k
                       StartingOffset + i * EltSize);
146
2.02k
    return;
147
2.02k
  }
148
9.93M
  // Interpret void as zero return values.
149
9.93M
  if (Ty.isVoidTy())
150
0
    return;
151
9.93M
  // Base case: we can get an LLT for this LLVM IR type.
152
9.93M
  ValueTys.push_back(getLLTForType(Ty, DL));
153
9.93M
  if (Offsets != nullptr)
154
1.60M
    Offsets->push_back(StartingOffset * 8);
155
9.93M
}
156
157
/// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
158
235
GlobalValue *llvm::ExtractTypeInfo(Value *V) {
159
235
  V = V->stripPointerCasts();
160
235
  GlobalValue *GV = dyn_cast<GlobalValue>(V);
161
235
  GlobalVariable *Var = dyn_cast<GlobalVariable>(V);
162
235
163
235
  if (Var && 
Var->getName() == "llvm.eh.catch.all.value"232
) {
164
0
    assert(Var->hasInitializer() &&
165
0
           "The EH catch-all value must have an initializer");
166
0
    Value *Init = Var->getInitializer();
167
0
    GV = dyn_cast<GlobalValue>(Init);
168
0
    if (!GV) V = cast<ConstantPointerNull>(Init);
169
0
  }
170
235
171
235
  assert((GV || isa<ConstantPointerNull>(V)) &&
172
235
         "TypeInfo must be a global variable or NULL");
173
235
  return GV;
174
235
}
175
176
/// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
177
/// processed uses a memory 'm' constraint.
178
bool
179
llvm::hasInlineAsmMemConstraint(InlineAsm::ConstraintInfoVector &CInfos,
180
0
                                const TargetLowering &TLI) {
181
0
  for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
182
0
    InlineAsm::ConstraintInfo &CI = CInfos[i];
183
0
    for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
184
0
      TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
185
0
      if (CType == TargetLowering::C_Memory)
186
0
        return true;
187
0
    }
188
0
189
0
    // Indirect operand accesses access memory.
190
0
    if (CI.isIndirect)
191
0
      return true;
192
0
  }
193
0
194
0
  return false;
195
0
}
196
197
/// getFCmpCondCode - Return the ISD condition code corresponding to
198
/// the given LLVM IR floating-point condition code.  This includes
199
/// consideration of global floating-point math flags.
200
///
201
19.4k
ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) {
202
19.4k
  switch (Pred) {
203
19.4k
  
case FCmpInst::FCMP_FALSE: return ISD::SETFALSE49
;
204
19.4k
  
case FCmpInst::FCMP_OEQ: return ISD::SETOEQ4.45k
;
205
19.4k
  
case FCmpInst::FCMP_OGT: return ISD::SETOGT4.04k
;
206
19.4k
  
case FCmpInst::FCMP_OGE: return ISD::SETOGE643
;
207
19.4k
  
case FCmpInst::FCMP_OLT: return ISD::SETOLT3.86k
;
208
19.4k
  
case FCmpInst::FCMP_OLE: return ISD::SETOLE472
;
209
19.4k
  
case FCmpInst::FCMP_ONE: return ISD::SETONE295
;
210
19.4k
  
case FCmpInst::FCMP_ORD: return ISD::SETO257
;
211
19.4k
  
case FCmpInst::FCMP_UNO: return ISD::SETUO1.79k
;
212
19.4k
  
case FCmpInst::FCMP_UEQ: return ISD::SETUEQ926
;
213
19.4k
  
case FCmpInst::FCMP_UGT: return ISD::SETUGT526
;
214
19.4k
  
case FCmpInst::FCMP_UGE: return ISD::SETUGE417
;
215
19.4k
  
case FCmpInst::FCMP_ULT: return ISD::SETULT461
;
216
19.4k
  
case FCmpInst::FCMP_ULE: return ISD::SETULE473
;
217
19.4k
  
case FCmpInst::FCMP_UNE: return ISD::SETUNE690
;
218
19.4k
  
case FCmpInst::FCMP_TRUE: return ISD::SETTRUE49
;
219
19.4k
  
default: 0
llvm_unreachable0
("Invalid FCmp predicate opcode!");
220
19.4k
  }
221
19.4k
}
222
223
1.37k
ISD::CondCode llvm::getFCmpCodeWithoutNaN(ISD::CondCode CC) {
224
1.37k
  switch (CC) {
225
1.37k
    
case ISD::SETOEQ: 74
case ISD::SETUEQ: return ISD::SETEQ74
;
226
74
    
case ISD::SETONE: 30
case ISD::SETUNE: return ISD::SETNE30
;
227
434
    case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
228
434
    
case ISD::SETOLE: 170
case ISD::SETULE: return ISD::SETLE170
;
229
445
    case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
230
445
    
case ISD::SETOGE: 188
case ISD::SETUGE: return ISD::SETGE188
;
231
188
    
default: return CC30
;
232
1.37k
  }
233
1.37k
}
234
235
/// getICmpCondCode - Return the ISD condition code corresponding to
236
/// the given LLVM IR integer condition code.
237
///
238
672k
ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) {
239
672k
  switch (Pred) {
240
672k
  
case ICmpInst::ICMP_EQ: return ISD::SETEQ395k
;
241
672k
  
case ICmpInst::ICMP_NE: return ISD::SETNE52.1k
;
242
672k
  
case ICmpInst::ICMP_SLE: return ISD::SETLE10.3k
;
243
672k
  
case ICmpInst::ICMP_ULE: return ISD::SETULE12.8k
;
244
672k
  
case ICmpInst::ICMP_SGE: return ISD::SETGE5.71k
;
245
672k
  
case ICmpInst::ICMP_UGE: return ISD::SETUGE2.97k
;
246
672k
  
case ICmpInst::ICMP_SLT: return ISD::SETLT52.1k
;
247
672k
  
case ICmpInst::ICMP_ULT: return ISD::SETULT64.0k
;
248
672k
  
case ICmpInst::ICMP_SGT: return ISD::SETGT40.2k
;
249
672k
  
case ICmpInst::ICMP_UGT: return ISD::SETUGT36.4k
;
250
672k
  default:
251
0
    llvm_unreachable("Invalid ICmp predicate opcode!");
252
672k
  }
253
672k
}
254
255
static bool isNoopBitcast(Type *T1, Type *T2,
256
316
                          const TargetLoweringBase& TLI) {
257
316
  return T1 == T2 || 
(117
T1->isPointerTy()117
&&
T2->isPointerTy()116
) ||
258
316
         
(1
isa<VectorType>(T1)1
&&
isa<VectorType>(T2)1
&&
259
1
          TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
260
316
}
261
262
/// Look through operations that will be free to find the earliest source of
263
/// this value.
264
///
265
/// @param ValLoc If V has aggegate type, we will be interested in a particular
266
/// scalar component. This records its address; the reverse of this list gives a
267
/// sequence of indices appropriate for an extractvalue to locate the important
268
/// value. This value is updated during the function and on exit will indicate
269
/// similar information for the Value returned.
270
///
271
/// @param DataBits If this function looks through truncate instructions, this
272
/// will record the smallest size attained.
273
static const Value *getNoopInput(const Value *V,
274
                                 SmallVectorImpl<unsigned> &ValLoc,
275
                                 unsigned &DataBits,
276
                                 const TargetLoweringBase &TLI,
277
112k
                                 const DataLayout &DL) {
278
113k
  while (true) {
279
113k
    // Try to look through V1; if V1 is not an instruction, it can't be looked
280
113k
    // through.
281
113k
    const Instruction *I = dyn_cast<Instruction>(V);
282
113k
    if (!I || 
I->getNumOperands() == 0112k
)
return V724
;
283
112k
    const Value *NoopInput = nullptr;
284
112k
285
112k
    Value *Op = I->getOperand(0);
286
112k
    if (isa<BitCastInst>(I)) {
287
116
      // Look through truly no-op bitcasts.
288
116
      if (isNoopBitcast(Op->getType(), I->getType(), TLI))
289
116
        NoopInput = Op;
290
112k
    } else if (isa<GetElementPtrInst>(I)) {
291
157
      // Look through getelementptr
292
157
      if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
293
0
        NoopInput = Op;
294
112k
    } else if (isa<IntToPtrInst>(I)) {
295
42
      // Look through inttoptr.
296
42
      // Make sure this isn't a truncating or extending cast.  We could
297
42
      // support this eventually, but don't bother for now.
298
42
      if (!isa<VectorType>(I->getType()) &&
299
42
          DL.getPointerSizeInBits() ==
300
42
              cast<IntegerType>(Op->getType())->getBitWidth())
301
42
        NoopInput = Op;
302
112k
    } else if (isa<PtrToIntInst>(I)) {
303
48
      // Look through ptrtoint.
304
48
      // Make sure this isn't a truncating or extending cast.  We could
305
48
      // support this eventually, but don't bother for now.
306
48
      if (!isa<VectorType>(I->getType()) &&
307
48
          DL.getPointerSizeInBits() ==
308
48
              cast<IntegerType>(I->getType())->getBitWidth())
309
48
        NoopInput = Op;
310
112k
    } else if (isa<TruncInst>(I) &&
311
112k
               
TLI.allowTruncateForTailCall(Op->getType(), I->getType())43
) {
312
31
      DataBits = std::min(DataBits, I->getType()->getPrimitiveSizeInBits());
313
31
      NoopInput = Op;
314
112k
    } else if (auto CS = ImmutableCallSite(I)) {
315
111k
      const Value *ReturnedOp = CS.getReturnedArgOperand();
316
111k
      if (ReturnedOp && 
isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI)200
)
317
200
        NoopInput = ReturnedOp;
318
111k
    } else 
if (const InsertValueInst *1.00k
IVI1.00k
= dyn_cast<InsertValueInst>(V)) {
319
85
      // Value may come from either the aggregate or the scalar
320
85
      ArrayRef<unsigned> InsertLoc = IVI->getIndices();
321
85
      if (ValLoc.size() >= InsertLoc.size() &&
322
85
          
std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())83
) {
323
49
        // The type being inserted is a nested sub-type of the aggregate; we
324
49
        // have to remove those initial indices to get the location we're
325
49
        // interested in for the operand.
326
49
        ValLoc.resize(ValLoc.size() - InsertLoc.size());
327
49
        NoopInput = IVI->getInsertedValueOperand();
328
49
      } else {
329
36
        // The struct we're inserting into has the value we're interested in, no
330
36
        // change of address.
331
36
        NoopInput = Op;
332
36
      }
333
917
    } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
334
40
      // The part we're interested in will inevitably be some sub-section of the
335
40
      // previous aggregate. Combine the two paths to obtain the true address of
336
40
      // our element.
337
40
      ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
338
40
      ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
339
40
      NoopInput = Op;
340
40
    }
341
112k
    // Terminate if we couldn't find anything to look through.
342
112k
    if (!NoopInput)
343
112k
      return V;
344
564
345
564
    V = NoopInput;
346
564
  }
347
112k
}
348
349
/// Return true if this scalar return value only has bits discarded on its path
350
/// from the "tail call" to the "ret". This includes the obvious noop
351
/// instructions handled by getNoopInput above as well as free truncations (or
352
/// extensions prior to the call).
353
static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
354
                                 SmallVectorImpl<unsigned> &RetIndices,
355
                                 SmallVectorImpl<unsigned> &CallIndices,
356
                                 bool AllowDifferingSizes,
357
                                 const TargetLoweringBase &TLI,
358
56.3k
                                 const DataLayout &DL) {
359
56.3k
360
56.3k
  // Trace the sub-value needed by the return value as far back up the graph as
361
56.3k
  // possible, in the hope that it will intersect with the value produced by the
362
56.3k
  // call. In the simple case with no "returned" attribute, the hope is actually
363
56.3k
  // that we end up back at the tail call instruction itself.
364
56.3k
  unsigned BitsRequired = UINT_MAX;
365
56.3k
  RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL);
366
56.3k
367
56.3k
  // If this slot in the value returned is undef, it doesn't matter what the
368
56.3k
  // call puts there, it'll be fine.
369
56.3k
  if (isa<UndefValue>(RetVal))
370
2
    return true;
371
56.3k
372
56.3k
  // Now do a similar search up through the graph to find where the value
373
56.3k
  // actually returned by the "tail call" comes from. In the simple case without
374
56.3k
  // a "returned" attribute, the search will be blocked immediately and the loop
375
56.3k
  // a Noop.
376
56.3k
  unsigned BitsProvided = UINT_MAX;
377
56.3k
  CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL);
378
56.3k
379
56.3k
  // There's no hope if we can't actually trace them to (the same part of!) the
380
56.3k
  // same value.
381
56.3k
  if (CallVal != RetVal || 
CallIndices != RetIndices54.7k
)
382
1.58k
    return false;
383
54.7k
384
54.7k
  // However, intervening truncates may have made the call non-tail. Make sure
385
54.7k
  // all the bits that are needed by the "ret" have been provided by the "tail
386
54.7k
  // call". FIXME: with sufficiently cunning bit-tracking, we could look through
387
54.7k
  // extensions too.
388
54.7k
  if (BitsProvided < BitsRequired ||
389
54.7k
      
(54.7k
!AllowDifferingSizes54.7k
&&
BitsProvided != BitsRequired210
))
390
3
    return false;
391
54.7k
392
54.7k
  return true;
393
54.7k
}
394
395
/// For an aggregate type, determine whether a given index is within bounds or
396
/// not.
397
352
static bool indexReallyValid(CompositeType *T, unsigned Idx) {
398
352
  if (ArrayType *AT = dyn_cast<ArrayType>(T))
399
34
    return Idx < AT->getNumElements();
400
318
401
318
  return Idx < cast<StructType>(T)->getNumElements();
402
318
}
403
404
/// Move the given iterators to the next leaf type in depth first traversal.
405
///
406
/// Performs a depth-first traversal of the type as specified by its arguments,
407
/// stopping at the next leaf node (which may be a legitimate scalar type or an
408
/// empty struct or array).
409
///
410
/// @param SubTypes List of the partial components making up the type from
411
/// outermost to innermost non-empty aggregate. The element currently
412
/// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
413
///
414
/// @param Path Set of extractvalue indices leading from the outermost type
415
/// (SubTypes[0]) to the leaf node currently represented.
416
///
417
/// @returns true if a new type was found, false otherwise. Calling this
418
/// function again on a finished iterator will repeatedly return
419
/// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
420
/// aggregate or a non-aggregate
421
static bool advanceToNextLeafType(SmallVectorImpl<CompositeType *> &SubTypes,
422
109k
                                  SmallVectorImpl<unsigned> &Path) {
423
109k
  // First march back up the tree until we can successfully increment one of the
424
109k
  // coordinates in Path.
425
109k
  while (!Path.empty() && 
!indexReallyValid(SubTypes.back(), Path.back() + 1)223
) {
426
90
    Path.pop_back();
427
90
    SubTypes.pop_back();
428
90
  }
429
109k
430
109k
  // If we reached the top, then the iterator is done.
431
109k
  if (Path.empty())
432
109k
    return false;
433
133
434
133
  // We know there's *some* valid leaf now, so march back down the tree picking
435
133
  // out the left-most element at each node.
436
133
  ++Path.back();
437
133
  Type *DeeperType = SubTypes.back()->getTypeAtIndex(Path.back());
438
155
  while (DeeperType->isAggregateType()) {
439
26
    CompositeType *CT = cast<CompositeType>(DeeperType);
440
26
    if (!indexReallyValid(CT, 0))
441
4
      return true;
442
22
443
22
    SubTypes.push_back(CT);
444
22
    Path.push_back(0);
445
22
446
22
    DeeperType = CT->getTypeAtIndex(0U);
447
22
  }
448
133
449
133
  
return true129
;
450
133
}
451
452
/// Find the first non-empty, scalar-like type in Next and setup the iterator
453
/// components.
454
///
455
/// Assuming Next is an aggregate of some kind, this function will traverse the
456
/// tree from left to right (i.e. depth-first) looking for the first
457
/// non-aggregate type which will play a role in function return.
458
///
459
/// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
460
/// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
461
/// i32 in that type.
462
static bool firstRealType(Type *Next,
463
                          SmallVectorImpl<CompositeType *> &SubTypes,
464
112k
                          SmallVectorImpl<unsigned> &Path) {
465
112k
  // First initialise the iterator components to the first "leaf" node
466
112k
  // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
467
112k
  // despite nominally being an aggregate).
468
112k
  while (Next->isAggregateType() &&
469
112k
         
indexReallyValid(cast<CompositeType>(Next), 0)103
) {
470
101
    SubTypes.push_back(cast<CompositeType>(Next));
471
101
    Path.push_back(0);
472
101
    Next = cast<CompositeType>(Next)->getTypeAtIndex(0U);
473
101
  }
474
112k
475
112k
  // If there's no Path now, Next was originally scalar already (or empty
476
112k
  // leaf). We're done.
477
112k
  if (Path.empty())
478
112k
    return true;
479
95
480
95
  // Otherwise, use normal iteration to keep looking through the tree until we
481
95
  // find a non-aggregate type.
482
99
  
while (95
SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType()) {
483
4
    if (!advanceToNextLeafType(SubTypes, Path))
484
0
      return false;
485
4
  }
486
95
487
95
  return true;
488
95
}
489
490
/// Set the iterator data-structures to the next non-empty, non-aggregate
491
/// subtype.
492
static bool nextRealType(SmallVectorImpl<CompositeType *> &SubTypes,
493
109k
                         SmallVectorImpl<unsigned> &Path) {
494
109k
  do {
495
109k
    if (!advanceToNextLeafType(SubTypes, Path))
496
109k
      return false;
497
130
498
130
    assert(!Path.empty() && "found a leaf but didn't set the path?");
499
130
  } while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType());
500
109k
501
109k
  
return true128
;
502
109k
}
503
504
505
/// Test if the given instruction is in a position to be optimized
506
/// with a tail-call. This roughly means that it's in a block with
507
/// a return and there's nothing that needs to be scheduled
508
/// between it and the return.
509
///
510
/// This function only tests target-independent requirements.
511
285k
bool llvm::isInTailCallPosition(ImmutableCallSite CS, const TargetMachine &TM) {
512
285k
  const Instruction *I = CS.getInstruction();
513
285k
  const BasicBlock *ExitBB = I->getParent();
514
285k
  const Instruction *Term = ExitBB->getTerminator();
515
285k
  const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
516
285k
517
285k
  // The block must end in a return statement or unreachable.
518
285k
  //
519
285k
  // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
520
285k
  // an unreachable, for now. The way tailcall optimization is currently
521
285k
  // implemented means it will add an epilogue followed by a jump. That is
522
285k
  // not profitable. Also, if the callee is a special function (e.g.
523
285k
  // longjmp on x86), it can end up causing miscompilation that has not
524
285k
  // been fully understood.
525
285k
  if (!Ret &&
526
285k
      
(198k
!TM.Options.GuaranteedTailCallOpt198k
||
!isa<UnreachableInst>(Term)1
))
527
198k
    return false;
528
86.7k
529
86.7k
  // If I will have a chain, make sure no other instruction that will have a
530
86.7k
  // chain interposes between I and the return.
531
86.7k
  if (I->mayHaveSideEffects() || 
I->mayReadFromMemory()226
||
532
86.7k
      
!isSafeToSpeculativelyExecute(I)198
)
533
90.9k
    
for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; 86.7k
--BBI4.22k
) {
534
90.9k
      if (&*BBI == I)
535
68.9k
        break;
536
21.9k
      // Debug info intrinsics do not get in the way of tail call optimization.
537
21.9k
      if (isa<DbgInfoIntrinsic>(BBI))
538
33
        continue;
539
21.9k
      // A lifetime end intrinsic should not stop tail call optimization.
540
21.9k
      if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
541
470
        if (II->getIntrinsicID() == Intrinsic::lifetime_end)
542
407
          continue;
543
21.5k
      if (BBI->mayHaveSideEffects() || 
BBI->mayReadFromMemory()4.06k
||
544
21.5k
          
!isSafeToSpeculativelyExecute(&*BBI)3.88k
)
545
17.7k
        return false;
546
21.5k
    }
547
86.7k
548
86.7k
  const Function *F = ExitBB->getParent();
549
68.9k
  return returnTypeIsEligibleForTailCall(
550
68.9k
      F, I, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering());
551
86.7k
}
552
553
bool llvm::attributesPermitTailCall(const Function *F, const Instruction *I,
554
                                    const ReturnInst *Ret,
555
                                    const TargetLoweringBase &TLI,
556
174k
                                    bool *AllowDifferingSizes) {
557
174k
  // ADS may be null, so don't write to it directly.
558
174k
  bool DummyADS;
559
174k
  bool &ADS = AllowDifferingSizes ? 
*AllowDifferingSizes56.5k
:
DummyADS117k
;
560
174k
  ADS = true;
561
174k
562
174k
  AttrBuilder CallerAttrs(F->getAttributes(), AttributeList::ReturnIndex);
563
174k
  AttrBuilder CalleeAttrs(cast<CallInst>(I)->getAttributes(),
564
174k
                          AttributeList::ReturnIndex);
565
174k
566
174k
  // NoAlias and NonNull are completely benign as far as calling convention
567
174k
  // goes, they shouldn't affect whether the call is a tail call.
568
174k
  CallerAttrs.removeAttribute(Attribute::NoAlias);
569
174k
  CalleeAttrs.removeAttribute(Attribute::NoAlias);
570
174k
  CallerAttrs.removeAttribute(Attribute::NonNull);
571
174k
  CalleeAttrs.removeAttribute(Attribute::NonNull);
572
174k
573
174k
  if (CallerAttrs.contains(Attribute::ZExt)) {
574
499
    if (!CalleeAttrs.contains(Attribute::ZExt))
575
159
      return false;
576
340
577
340
    ADS = false;
578
340
    CallerAttrs.removeAttribute(Attribute::ZExt);
579
340
    CalleeAttrs.removeAttribute(Attribute::ZExt);
580
173k
  } else if (CallerAttrs.contains(Attribute::SExt)) {
581
59
    if (!CalleeAttrs.contains(Attribute::SExt))
582
5
      return false;
583
54
584
54
    ADS = false;
585
54
    CallerAttrs.removeAttribute(Attribute::SExt);
586
54
    CalleeAttrs.removeAttribute(Attribute::SExt);
587
54
  }
588
174k
589
174k
  // Drop sext and zext return attributes if the result is not used.
590
174k
  // This enables tail calls for code like:
591
174k
  //
592
174k
  // define void @caller() {
593
174k
  // entry:
594
174k
  //   %unused_result = tail call zeroext i1 @callee()
595
174k
  //   br label %retlabel
596
174k
  // retlabel:
597
174k
  //   ret void
598
174k
  // }
599
174k
  
if (173k
I->use_empty()173k
) {
600
15.2k
    CalleeAttrs.removeAttribute(Attribute::SExt);
601
15.2k
    CalleeAttrs.removeAttribute(Attribute::ZExt);
602
15.2k
  }
603
173k
604
173k
  // If they're still different, there's some facet we don't understand
605
173k
  // (currently only "inreg", but in future who knows). It may be OK but the
606
173k
  // only safe option is to reject the tail call.
607
173k
  return CallerAttrs == CalleeAttrs;
608
174k
}
609
610
bool llvm::returnTypeIsEligibleForTailCall(const Function *F,
611
                                           const Instruction *I,
612
                                           const ReturnInst *Ret,
613
68.9k
                                           const TargetLoweringBase &TLI) {
614
68.9k
  // If the block ends with a void return or unreachable, it doesn't matter
615
68.9k
  // what the call's return type is.
616
68.9k
  if (!Ret || 
Ret->getNumOperands() == 068.9k
)
return true12.3k
;
617
56.6k
618
56.6k
  // If the return value is undef, it doesn't matter what the call's
619
56.6k
  // return type is.
620
56.6k
  if (isa<UndefValue>(Ret->getOperand(0))) 
return true48
;
621
56.5k
622
56.5k
  // Make sure the attributes attached to each return are compatible.
623
56.5k
  bool AllowDifferingSizes;
624
56.5k
  if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
625
238
    return false;
626
56.3k
627
56.3k
  const Value *RetVal = Ret->getOperand(0), *CallVal = I;
628
56.3k
  // Intrinsic like llvm.memcpy has no return value, but the expanded
629
56.3k
  // libcall may or may not have return value. On most platforms, it
630
56.3k
  // will be expanded as memcpy in libc, which returns the first
631
56.3k
  // argument. On other platforms like arm-none-eabi, memcpy may be
632
56.3k
  // expanded as library call without return value, like __aeabi_memcpy.
633
56.3k
  const CallInst *Call = cast<CallInst>(I);
634
56.3k
  if (Function *F = Call->getCalledFunction()) {
635
55.9k
    Intrinsic::ID IID = F->getIntrinsicID();
636
55.9k
    if (((IID == Intrinsic::memcpy &&
637
55.9k
          
TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")15
) ||
638
55.9k
         
(55.9k
IID == Intrinsic::memmove55.9k
&&
639
55.9k
          
TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")2
) ||
640
55.9k
         
(55.9k
IID == Intrinsic::memset55.9k
&&
641
55.9k
          
TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset")2
)) &&
642
55.9k
        
RetVal == Call->getArgOperand(0)16
)
643
10
      return true;
644
56.3k
  }
645
56.3k
646
56.3k
  SmallVector<unsigned, 4> RetPath, CallPath;
647
56.3k
  SmallVector<CompositeType *, 4> RetSubTypes, CallSubTypes;
648
56.3k
649
56.3k
  bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
650
56.3k
  bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
651
56.3k
652
56.3k
  // Nothing's actually returned, it doesn't matter what the callee put there
653
56.3k
  // it's a valid tail call.
654
56.3k
  if (RetEmpty)
655
0
    return true;
656
56.3k
657
56.3k
  // Iterate pairwise through each of the value types making up the tail call
658
56.3k
  // and the corresponding return. For each one we want to know whether it's
659
56.3k
  // essentially going directly from the tail call to the ret, via operations
660
56.3k
  // that end up not generating any code.
661
56.3k
  //
662
56.3k
  // We allow a certain amount of covariance here. For example it's permitted
663
56.3k
  // for the tail call to define more bits than the ret actually cares about
664
56.3k
  // (e.g. via a truncate).
665
56.3k
  
do 56.3k
{
666
56.3k
    if (CallEmpty) {
667
2
      // We've exhausted the values produced by the tail call instruction, the
668
2
      // rest are essentially undef. The type doesn't really matter, but we need
669
2
      // *something*.
670
2
      Type *SlotType = RetSubTypes.back()->getTypeAtIndex(RetPath.back());
671
2
      CallVal = UndefValue::get(SlotType);
672
2
    }
673
56.3k
674
56.3k
    // The manipulations performed when we're looking through an insertvalue or
675
56.3k
    // an extractvalue would happen at the front of the RetPath list, so since
676
56.3k
    // we have to copy it anyway it's more efficient to create a reversed copy.
677
56.3k
    SmallVector<unsigned, 4> TmpRetPath(RetPath.rbegin(), RetPath.rend());
678
56.3k
    SmallVector<unsigned, 4> TmpCallPath(CallPath.rbegin(), CallPath.rend());
679
56.3k
680
56.3k
    // Finally, we can check whether the value produced by the tail call at this
681
56.3k
    // index is compatible with the value we return.
682
56.3k
    if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
683
56.3k
                              AllowDifferingSizes, TLI,
684
56.3k
                              F->getParent()->getDataLayout()))
685
1.58k
      return false;
686
54.7k
687
54.7k
    CallEmpty  = !nextRealType(CallSubTypes, CallPath);
688
54.7k
  } while(nextRealType(RetSubTypes, RetPath));
689
56.3k
690
56.3k
  
return true54.7k
;
691
56.3k
}
692
693
static void collectEHScopeMembers(
694
    DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope,
695
2.09k
    const MachineBasicBlock *MBB) {
696
2.09k
  SmallVector<const MachineBasicBlock *, 16> Worklist = {MBB};
697
6.49k
  while (!Worklist.empty()) {
698
4.39k
    const MachineBasicBlock *Visiting = Worklist.pop_back_val();
699
4.39k
    // Don't follow blocks which start new scopes.
700
4.39k
    if (Visiting->isEHPad() && 
Visiting != MBB1.69k
)
701
924
      continue;
702
3.46k
703
3.46k
    // Add this MBB to our scope.
704
3.46k
    auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope));
705
3.46k
706
3.46k
    // Don't revisit blocks.
707
3.46k
    if (!P.second) {
708
975
      assert(P.first->second == EHScope && "MBB is part of two scopes!");
709
975
      continue;
710
975
    }
711
2.49k
712
2.49k
    // Returns are boundaries where scope transfer can occur, don't follow
713
2.49k
    // successors.
714
2.49k
    if (Visiting->isEHScopeReturnBlock())
715
627
      continue;
716
1.86k
717
1.86k
    for (const MachineBasicBlock *Succ : Visiting->successors())
718
2.29k
      Worklist.push_back(Succ);
719
1.86k
  }
720
2.09k
}
721
722
DenseMap<const MachineBasicBlock *, int>
723
1.79M
llvm::getEHScopeMembership(const MachineFunction &MF) {
724
1.79M
  DenseMap<const MachineBasicBlock *, int> EHScopeMembership;
725
1.79M
726
1.79M
  // We don't have anything to do if there aren't any EH pads.
727
1.79M
  if (!MF.hasEHScopes())
728
1.79M
    return EHScopeMembership;
729
508
730
508
  int EntryBBNumber = MF.front().getNumber();
731
508
  bool IsSEH = isAsynchronousEHPersonality(
732
508
      classifyEHPersonality(MF.getFunction().getPersonalityFn()));
733
508
734
508
  const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
735
508
  SmallVector<const MachineBasicBlock *, 16> EHScopeBlocks;
736
508
  SmallVector<const MachineBasicBlock *, 16> UnreachableBlocks;
737
508
  SmallVector<const MachineBasicBlock *, 16> SEHCatchPads;
738
508
  SmallVector<std::pair<const MachineBasicBlock *, int>, 16> CatchRetSuccessors;
739
2.77k
  for (const MachineBasicBlock &MBB : MF) {
740
2.77k
    if (MBB.isEHScopeEntry()) {
741
737
      EHScopeBlocks.push_back(&MBB);
742
2.04k
    } else if (IsSEH && 
MBB.isEHPad()446
) {
743
131
      SEHCatchPads.push_back(&MBB);
744
1.91k
    } else if (MBB.pred_empty()) {
745
511
      UnreachableBlocks.push_back(&MBB);
746
511
    }
747
2.77k
748
2.77k
    MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator();
749
2.77k
750
2.77k
    // CatchPads are not scopes for SEH so do not consider CatchRet to
751
2.77k
    // transfer control to another scope.
752
2.77k
    if (MBBI == MBB.end() || 
MBBI->getOpcode() != TII->getCatchReturnOpcode()1.89k
)
753
2.32k
      continue;
754
449
755
449
    // FIXME: SEH CatchPads are not necessarily in the parent function:
756
449
    // they could be inside a finally block.
757
449
    const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB();
758
449
    const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB();
759
449
    CatchRetSuccessors.push_back(
760
449
        {Successor, IsSEH ? 
EntryBBNumber0
: SuccessorColor->getNumber()});
761
449
  }
762
508
763
508
  // We don't have anything to do if there aren't any EH pads.
764
508
  if (EHScopeBlocks.empty())
765
69
    return EHScopeMembership;
766
439
767
439
  // Identify all the basic blocks reachable from the function entry.
768
439
  collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front());
769
439
  // All blocks not part of a scope are in the parent function.
770
439
  for (const MachineBasicBlock *MBB : UnreachableBlocks)
771
442
    collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
772
439
  // Next, identify all the blocks inside the scopes.
773
439
  for (const MachineBasicBlock *MBB : EHScopeBlocks)
774
737
    collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB);
775
439
  // SEH CatchPads aren't really scopes, handle them separately.
776
439
  for (const MachineBasicBlock *MBB : SEHCatchPads)
777
29
    collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB);
778
439
  // Finally, identify all the targets of a catchret.
779
439
  for (std::pair<const MachineBasicBlock *, int> CatchRetPair :
780
439
       CatchRetSuccessors)
781
449
    collectEHScopeMembers(EHScopeMembership, CatchRetPair.second,
782
449
                          CatchRetPair.first);
783
439
  return EHScopeMembership;
784
439
}