LLVM  8.0.0svn
ValueTracking.cpp
Go to the documentation of this file.
1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
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 contains routines that help analyze properties that chains of
11 // computations have.
12 //
13 //===----------------------------------------------------------------------===//
14 
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/StringRef.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/LoopInfo.h"
35 #include "llvm/IR/Argument.h"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CallSite.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/DiagnosticInfo.h"
45 #include "llvm/IR/Dominators.h"
46 #include "llvm/IR/Function.h"
48 #include "llvm/IR/GlobalAlias.h"
49 #include "llvm/IR/GlobalValue.h"
50 #include "llvm/IR/GlobalVariable.h"
51 #include "llvm/IR/InstrTypes.h"
52 #include "llvm/IR/Instruction.h"
53 #include "llvm/IR/Instructions.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Intrinsics.h"
56 #include "llvm/IR/LLVMContext.h"
57 #include "llvm/IR/Metadata.h"
58 #include "llvm/IR/Module.h"
59 #include "llvm/IR/Operator.h"
60 #include "llvm/IR/PatternMatch.h"
61 #include "llvm/IR/Type.h"
62 #include "llvm/IR/User.h"
63 #include "llvm/IR/Value.h"
64 #include "llvm/Support/Casting.h"
66 #include "llvm/Support/Compiler.h"
68 #include "llvm/Support/KnownBits.h"
70 #include <algorithm>
71 #include <array>
72 #include <cassert>
73 #include <cstdint>
74 #include <iterator>
75 #include <utility>
76 
77 using namespace llvm;
78 using namespace llvm::PatternMatch;
79 
80 const unsigned MaxDepth = 6;
81 
82 // Controls the number of uses of the value searched for possible
83 // dominating comparisons.
84 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
85  cl::Hidden, cl::init(20));
86 
87 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
88 /// returns the element type's bitwidth.
89 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
90  if (unsigned BitWidth = Ty->getScalarSizeInBits())
91  return BitWidth;
92 
93  return DL.getIndexTypeSizeInBits(Ty);
94 }
95 
96 namespace {
97 
98 // Simplifying using an assume can only be done in a particular control-flow
99 // context (the context instruction provides that context). If an assume and
100 // the context instruction are not in the same block then the DT helps in
101 // figuring out if we can use it.
102 struct Query {
103  const DataLayout &DL;
104  AssumptionCache *AC;
105  const Instruction *CxtI;
106  const DominatorTree *DT;
107 
108  // Unlike the other analyses, this may be a nullptr because not all clients
109  // provide it currently.
111 
112  /// Set of assumptions that should be excluded from further queries.
113  /// This is because of the potential for mutual recursion to cause
114  /// computeKnownBits to repeatedly visit the same assume intrinsic. The
115  /// classic case of this is assume(x = y), which will attempt to determine
116  /// bits in x from bits in y, which will attempt to determine bits in y from
117  /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
118  /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo
119  /// (all of which can call computeKnownBits), and so on.
120  std::array<const Value *, MaxDepth> Excluded;
121 
122  /// If true, it is safe to use metadata during simplification.
123  InstrInfoQuery IIQ;
124 
125  unsigned NumExcluded = 0;
126 
127  Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
128  const DominatorTree *DT, bool UseInstrInfo,
129  OptimizationRemarkEmitter *ORE = nullptr)
130  : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {}
131 
132  Query(const Query &Q, const Value *NewExcl)
133  : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE), IIQ(Q.IIQ),
134  NumExcluded(Q.NumExcluded) {
135  Excluded = Q.Excluded;
136  Excluded[NumExcluded++] = NewExcl;
137  assert(NumExcluded <= Excluded.size());
138  }
139 
140  bool isExcluded(const Value *Value) const {
141  if (NumExcluded == 0)
142  return false;
143  auto End = Excluded.begin() + NumExcluded;
144  return std::find(Excluded.begin(), End, Value) != End;
145  }
146 };
147 
148 } // end anonymous namespace
149 
150 // Given the provided Value and, potentially, a context instruction, return
151 // the preferred context instruction (if any).
152 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
153  // If we've been provided with a context instruction, then use that (provided
154  // it has been inserted).
155  if (CxtI && CxtI->getParent())
156  return CxtI;
157 
158  // If the value is really an already-inserted instruction, then use that.
159  CxtI = dyn_cast<Instruction>(V);
160  if (CxtI && CxtI->getParent())
161  return CxtI;
162 
163  return nullptr;
164 }
165 
166 static void computeKnownBits(const Value *V, KnownBits &Known,
167  unsigned Depth, const Query &Q);
168 
169 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
170  const DataLayout &DL, unsigned Depth,
171  AssumptionCache *AC, const Instruction *CxtI,
172  const DominatorTree *DT,
173  OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
174  ::computeKnownBits(V, Known, Depth,
175  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
176 }
177 
178 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
179  const Query &Q);
180 
182  unsigned Depth, AssumptionCache *AC,
183  const Instruction *CxtI,
184  const DominatorTree *DT,
186  bool UseInstrInfo) {
188  V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
189 }
190 
191 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
192  const DataLayout &DL, AssumptionCache *AC,
193  const Instruction *CxtI, const DominatorTree *DT,
194  bool UseInstrInfo) {
195  assert(LHS->getType() == RHS->getType() &&
196  "LHS and RHS should have the same type");
197  assert(LHS->getType()->isIntOrIntVectorTy() &&
198  "LHS and RHS should be integers");
199  // Look for an inverted mask: (X & ~M) op (Y & M).
200  Value *M;
201  if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
202  match(RHS, m_c_And(m_Specific(M), m_Value())))
203  return true;
204  if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
205  match(LHS, m_c_And(m_Specific(M), m_Value())))
206  return true;
207  IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
208  KnownBits LHSKnown(IT->getBitWidth());
209  KnownBits RHSKnown(IT->getBitWidth());
210  computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
211  computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
212  return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
213 }
214 
216  for (const User *U : CxtI->users()) {
217  if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
218  if (IC->isEquality())
219  if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
220  if (C->isNullValue())
221  continue;
222  return false;
223  }
224  return true;
225 }
226 
227 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
228  const Query &Q);
229 
231  bool OrZero, unsigned Depth,
232  AssumptionCache *AC, const Instruction *CxtI,
233  const DominatorTree *DT, bool UseInstrInfo) {
235  V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
236 }
237 
238 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
239 
240 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
241  AssumptionCache *AC, const Instruction *CxtI,
242  const DominatorTree *DT, bool UseInstrInfo) {
243  return ::isKnownNonZero(V, Depth,
244  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
245 }
246 
247 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
248  unsigned Depth, AssumptionCache *AC,
249  const Instruction *CxtI, const DominatorTree *DT,
250  bool UseInstrInfo) {
251  KnownBits Known =
252  computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
253  return Known.isNonNegative();
254 }
255 
256 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
257  AssumptionCache *AC, const Instruction *CxtI,
258  const DominatorTree *DT, bool UseInstrInfo) {
259  if (auto *CI = dyn_cast<ConstantInt>(V))
260  return CI->getValue().isStrictlyPositive();
261 
262  // TODO: We'd doing two recursive queries here. We should factor this such
263  // that only a single query is needed.
264  return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) &&
265  isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo);
266 }
267 
268 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
269  AssumptionCache *AC, const Instruction *CxtI,
270  const DominatorTree *DT, bool UseInstrInfo) {
271  KnownBits Known =
272  computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
273  return Known.isNegative();
274 }
275 
276 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
277 
278 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
279  const DataLayout &DL, AssumptionCache *AC,
280  const Instruction *CxtI, const DominatorTree *DT,
281  bool UseInstrInfo) {
283  Query(DL, AC, safeCxtI(V1, safeCxtI(V2, CxtI)), DT,
284  UseInstrInfo, /*ORE=*/nullptr));
285 }
286 
287 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
288  const Query &Q);
289 
290 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
291  const DataLayout &DL, unsigned Depth,
292  AssumptionCache *AC, const Instruction *CxtI,
293  const DominatorTree *DT, bool UseInstrInfo) {
295  V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
296 }
297 
298 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
299  const Query &Q);
300 
301 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
302  unsigned Depth, AssumptionCache *AC,
303  const Instruction *CxtI,
304  const DominatorTree *DT, bool UseInstrInfo) {
306  V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
307 }
308 
309 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
310  bool NSW,
311  KnownBits &KnownOut, KnownBits &Known2,
312  unsigned Depth, const Query &Q) {
313  unsigned BitWidth = KnownOut.getBitWidth();
314 
315  // If an initial sequence of bits in the result is not needed, the
316  // corresponding bits in the operands are not needed.
317  KnownBits LHSKnown(BitWidth);
318  computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
319  computeKnownBits(Op1, Known2, Depth + 1, Q);
320 
321  KnownOut = KnownBits::computeForAddSub(Add, NSW, LHSKnown, Known2);
322 }
323 
324 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
325  KnownBits &Known, KnownBits &Known2,
326  unsigned Depth, const Query &Q) {
327  unsigned BitWidth = Known.getBitWidth();
328  computeKnownBits(Op1, Known, Depth + 1, Q);
329  computeKnownBits(Op0, Known2, Depth + 1, Q);
330 
331  bool isKnownNegative = false;
332  bool isKnownNonNegative = false;
333  // If the multiplication is known not to overflow, compute the sign bit.
334  if (NSW) {
335  if (Op0 == Op1) {
336  // The product of a number with itself is non-negative.
337  isKnownNonNegative = true;
338  } else {
339  bool isKnownNonNegativeOp1 = Known.isNonNegative();
340  bool isKnownNonNegativeOp0 = Known2.isNonNegative();
341  bool isKnownNegativeOp1 = Known.isNegative();
342  bool isKnownNegativeOp0 = Known2.isNegative();
343  // The product of two numbers with the same sign is non-negative.
344  isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
345  (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
346  // The product of a negative number and a non-negative number is either
347  // negative or zero.
348  if (!isKnownNonNegative)
349  isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
350  isKnownNonZero(Op0, Depth, Q)) ||
351  (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
352  isKnownNonZero(Op1, Depth, Q));
353  }
354  }
355 
356  assert(!Known.hasConflict() && !Known2.hasConflict());
357  // Compute a conservative estimate for high known-0 bits.
358  unsigned LeadZ = std::max(Known.countMinLeadingZeros() +
359  Known2.countMinLeadingZeros(),
360  BitWidth) - BitWidth;
361  LeadZ = std::min(LeadZ, BitWidth);
362 
363  // The result of the bottom bits of an integer multiply can be
364  // inferred by looking at the bottom bits of both operands and
365  // multiplying them together.
366  // We can infer at least the minimum number of known trailing bits
367  // of both operands. Depending on number of trailing zeros, we can
368  // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming
369  // a and b are divisible by m and n respectively.
370  // We then calculate how many of those bits are inferrable and set
371  // the output. For example, the i8 mul:
372  // a = XXXX1100 (12)
373  // b = XXXX1110 (14)
374  // We know the bottom 3 bits are zero since the first can be divided by
375  // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4).
376  // Applying the multiplication to the trimmed arguments gets:
377  // XX11 (3)
378  // X111 (7)
379  // -------
380  // XX11
381  // XX11
382  // XX11
383  // XX11
384  // -------
385  // XXXXX01
386  // Which allows us to infer the 2 LSBs. Since we're multiplying the result
387  // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits.
388  // The proof for this can be described as:
389  // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) &&
390  // (C7 == (1 << (umin(countTrailingZeros(C1), C5) +
391  // umin(countTrailingZeros(C2), C6) +
392  // umin(C5 - umin(countTrailingZeros(C1), C5),
393  // C6 - umin(countTrailingZeros(C2), C6)))) - 1)
394  // %aa = shl i8 %a, C5
395  // %bb = shl i8 %b, C6
396  // %aaa = or i8 %aa, C1
397  // %bbb = or i8 %bb, C2
398  // %mul = mul i8 %aaa, %bbb
399  // %mask = and i8 %mul, C7
400  // =>
401  // %mask = i8 ((C1*C2)&C7)
402  // Where C5, C6 describe the known bits of %a, %b
403  // C1, C2 describe the known bottom bits of %a, %b.
404  // C7 describes the mask of the known bits of the result.
405  APInt Bottom0 = Known.One;
406  APInt Bottom1 = Known2.One;
407 
408  // How many times we'd be able to divide each argument by 2 (shr by 1).
409  // This gives us the number of trailing zeros on the multiplication result.
410  unsigned TrailBitsKnown0 = (Known.Zero | Known.One).countTrailingOnes();
411  unsigned TrailBitsKnown1 = (Known2.Zero | Known2.One).countTrailingOnes();
412  unsigned TrailZero0 = Known.countMinTrailingZeros();
413  unsigned TrailZero1 = Known2.countMinTrailingZeros();
414  unsigned TrailZ = TrailZero0 + TrailZero1;
415 
416  // Figure out the fewest known-bits operand.
417  unsigned SmallestOperand = std::min(TrailBitsKnown0 - TrailZero0,
418  TrailBitsKnown1 - TrailZero1);
419  unsigned ResultBitsKnown = std::min(SmallestOperand + TrailZ, BitWidth);
420 
421  APInt BottomKnown = Bottom0.getLoBits(TrailBitsKnown0) *
422  Bottom1.getLoBits(TrailBitsKnown1);
423 
424  Known.resetAll();
425  Known.Zero.setHighBits(LeadZ);
426  Known.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
427  Known.One |= BottomKnown.getLoBits(ResultBitsKnown);
428 
429  // Only make use of no-wrap flags if we failed to compute the sign bit
430  // directly. This matters if the multiplication always overflows, in
431  // which case we prefer to follow the result of the direct computation,
432  // though as the program is invoking undefined behaviour we can choose
433  // whatever we like here.
434  if (isKnownNonNegative && !Known.isNegative())
435  Known.makeNonNegative();
436  else if (isKnownNegative && !Known.isNonNegative())
437  Known.makeNegative();
438 }
439 
441  KnownBits &Known) {
442  unsigned BitWidth = Known.getBitWidth();
443  unsigned NumRanges = Ranges.getNumOperands() / 2;
444  assert(NumRanges >= 1);
445 
446  Known.Zero.setAllBits();
447  Known.One.setAllBits();
448 
449  for (unsigned i = 0; i < NumRanges; ++i) {
450  ConstantInt *Lower =
451  mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
452  ConstantInt *Upper =
453  mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
454  ConstantRange Range(Lower->getValue(), Upper->getValue());
455 
456  // The first CommonPrefixBits of all values in Range are equal.
457  unsigned CommonPrefixBits =
458  (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
459 
460  APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
461  Known.One &= Range.getUnsignedMax() & Mask;
462  Known.Zero &= ~Range.getUnsignedMax() & Mask;
463  }
464 }
465 
466 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
467  SmallVector<const Value *, 16> WorkSet(1, I);
470 
471  // The instruction defining an assumption's condition itself is always
472  // considered ephemeral to that assumption (even if it has other
473  // non-ephemeral users). See r246696's test case for an example.
474  if (is_contained(I->operands(), E))
475  return true;
476 
477  while (!WorkSet.empty()) {
478  const Value *V = WorkSet.pop_back_val();
479  if (!Visited.insert(V).second)
480  continue;
481 
482  // If all uses of this value are ephemeral, then so is this value.
483  if (llvm::all_of(V->users(), [&](const User *U) {
484  return EphValues.count(U);
485  })) {
486  if (V == E)
487  return true;
488 
489  if (V == I || isSafeToSpeculativelyExecute(V)) {
490  EphValues.insert(V);
491  if (const User *U = dyn_cast<User>(V))
492  for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
493  J != JE; ++J)
494  WorkSet.push_back(*J);
495  }
496  }
497  }
498 
499  return false;
500 }
501 
502 // Is this an intrinsic that cannot be speculated but also cannot trap?
504  if (const CallInst *CI = dyn_cast<CallInst>(I))
505  if (Function *F = CI->getCalledFunction())
506  switch (F->getIntrinsicID()) {
507  default: break;
508  // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
509  case Intrinsic::assume:
510  case Intrinsic::sideeffect:
511  case Intrinsic::dbg_declare:
512  case Intrinsic::dbg_value:
513  case Intrinsic::dbg_label:
514  case Intrinsic::invariant_start:
515  case Intrinsic::invariant_end:
516  case Intrinsic::lifetime_start:
517  case Intrinsic::lifetime_end:
518  case Intrinsic::objectsize:
519  case Intrinsic::ptr_annotation:
520  case Intrinsic::var_annotation:
521  return true;
522  }
523 
524  return false;
525 }
526 
528  const Instruction *CxtI,
529  const DominatorTree *DT) {
530  // There are two restrictions on the use of an assume:
531  // 1. The assume must dominate the context (or the control flow must
532  // reach the assume whenever it reaches the context).
533  // 2. The context must not be in the assume's set of ephemeral values
534  // (otherwise we will use the assume to prove that the condition
535  // feeding the assume is trivially true, thus causing the removal of
536  // the assume).
537 
538  if (DT) {
539  if (DT->dominates(Inv, CxtI))
540  return true;
541  } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
542  // We don't have a DT, but this trivially dominates.
543  return true;
544  }
545 
546  // With or without a DT, the only remaining case we will check is if the
547  // instructions are in the same BB. Give up if that is not the case.
548  if (Inv->getParent() != CxtI->getParent())
549  return false;
550 
551  // If we have a dom tree, then we now know that the assume doesn't dominate
552  // the other instruction. If we don't have a dom tree then we can check if
553  // the assume is first in the BB.
554  if (!DT) {
555  // Search forward from the assume until we reach the context (or the end
556  // of the block); the common case is that the assume will come first.
557  for (auto I = std::next(BasicBlock::const_iterator(Inv)),
558  IE = Inv->getParent()->end(); I != IE; ++I)
559  if (&*I == CxtI)
560  return true;
561  }
562 
563  // The context comes first, but they're both in the same block. Make sure
564  // there is nothing in between that might interrupt the control flow.
566  std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
567  I != IE; ++I)
569  return false;
570 
571  return !isEphemeralValueOf(Inv, CxtI);
572 }
573 
574 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
575  unsigned Depth, const Query &Q) {
576  // Use of assumptions is context-sensitive. If we don't have a context, we
577  // cannot use them!
578  if (!Q.AC || !Q.CxtI)
579  return;
580 
581  unsigned BitWidth = Known.getBitWidth();
582 
583  // Note that the patterns below need to be kept in sync with the code
584  // in AssumptionCache::updateAffectedValues.
585 
586  for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
587  if (!AssumeVH)
588  continue;
589  CallInst *I = cast<CallInst>(AssumeVH);
590  assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
591  "Got assumption for the wrong function!");
592  if (Q.isExcluded(I))
593  continue;
594 
595  // Warning: This loop can end up being somewhat performance sensitive.
596  // We're running this loop for once for each value queried resulting in a
597  // runtime of ~O(#assumes * #values).
598 
599  assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
600  "must be an assume intrinsic");
601 
602  Value *Arg = I->getArgOperand(0);
603 
604  if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
605  assert(BitWidth == 1 && "assume operand is not i1?");
606  Known.setAllOnes();
607  return;
608  }
609  if (match(Arg, m_Not(m_Specific(V))) &&
610  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
611  assert(BitWidth == 1 && "assume operand is not i1?");
612  Known.setAllZero();
613  return;
614  }
615 
616  // The remaining tests are all recursive, so bail out if we hit the limit.
617  if (Depth == MaxDepth)
618  continue;
619 
620  Value *A, *B;
621  auto m_V = m_CombineOr(m_Specific(V),
623  m_BitCast(m_Specific(V))));
624 
625  CmpInst::Predicate Pred;
626  uint64_t C;
627  // assume(v = a)
628  if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
629  Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
630  KnownBits RHSKnown(BitWidth);
631  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
632  Known.Zero |= RHSKnown.Zero;
633  Known.One |= RHSKnown.One;
634  // assume(v & b = a)
635  } else if (match(Arg,
636  m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
637  Pred == ICmpInst::ICMP_EQ &&
638  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
639  KnownBits RHSKnown(BitWidth);
640  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
641  KnownBits MaskKnown(BitWidth);
642  computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
643 
644  // For those bits in the mask that are known to be one, we can propagate
645  // known bits from the RHS to V.
646  Known.Zero |= RHSKnown.Zero & MaskKnown.One;
647  Known.One |= RHSKnown.One & MaskKnown.One;
648  // assume(~(v & b) = a)
649  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
650  m_Value(A))) &&
651  Pred == ICmpInst::ICMP_EQ &&
652  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
653  KnownBits RHSKnown(BitWidth);
654  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
655  KnownBits MaskKnown(BitWidth);
656  computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
657 
658  // For those bits in the mask that are known to be one, we can propagate
659  // inverted known bits from the RHS to V.
660  Known.Zero |= RHSKnown.One & MaskKnown.One;
661  Known.One |= RHSKnown.Zero & MaskKnown.One;
662  // assume(v | b = a)
663  } else if (match(Arg,
664  m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
665  Pred == ICmpInst::ICMP_EQ &&
666  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
667  KnownBits RHSKnown(BitWidth);
668  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
669  KnownBits BKnown(BitWidth);
670  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
671 
672  // For those bits in B that are known to be zero, we can propagate known
673  // bits from the RHS to V.
674  Known.Zero |= RHSKnown.Zero & BKnown.Zero;
675  Known.One |= RHSKnown.One & BKnown.Zero;
676  // assume(~(v | b) = a)
677  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
678  m_Value(A))) &&
679  Pred == ICmpInst::ICMP_EQ &&
680  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
681  KnownBits RHSKnown(BitWidth);
682  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
683  KnownBits BKnown(BitWidth);
684  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
685 
686  // For those bits in B that are known to be zero, we can propagate
687  // inverted known bits from the RHS to V.
688  Known.Zero |= RHSKnown.One & BKnown.Zero;
689  Known.One |= RHSKnown.Zero & BKnown.Zero;
690  // assume(v ^ b = a)
691  } else if (match(Arg,
692  m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
693  Pred == ICmpInst::ICMP_EQ &&
694  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
695  KnownBits RHSKnown(BitWidth);
696  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
697  KnownBits BKnown(BitWidth);
698  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
699 
700  // For those bits in B that are known to be zero, we can propagate known
701  // bits from the RHS to V. For those bits in B that are known to be one,
702  // we can propagate inverted known bits from the RHS to V.
703  Known.Zero |= RHSKnown.Zero & BKnown.Zero;
704  Known.One |= RHSKnown.One & BKnown.Zero;
705  Known.Zero |= RHSKnown.One & BKnown.One;
706  Known.One |= RHSKnown.Zero & BKnown.One;
707  // assume(~(v ^ b) = a)
708  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
709  m_Value(A))) &&
710  Pred == ICmpInst::ICMP_EQ &&
711  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
712  KnownBits RHSKnown(BitWidth);
713  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
714  KnownBits BKnown(BitWidth);
715  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
716 
717  // For those bits in B that are known to be zero, we can propagate
718  // inverted known bits from the RHS to V. For those bits in B that are
719  // known to be one, we can propagate known bits from the RHS to V.
720  Known.Zero |= RHSKnown.One & BKnown.Zero;
721  Known.One |= RHSKnown.Zero & BKnown.Zero;
722  Known.Zero |= RHSKnown.Zero & BKnown.One;
723  Known.One |= RHSKnown.One & BKnown.One;
724  // assume(v << c = a)
725  } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
726  m_Value(A))) &&
727  Pred == ICmpInst::ICMP_EQ &&
728  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
729  C < BitWidth) {
730  KnownBits RHSKnown(BitWidth);
731  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
732  // For those bits in RHS that are known, we can propagate them to known
733  // bits in V shifted to the right by C.
734  RHSKnown.Zero.lshrInPlace(C);
735  Known.Zero |= RHSKnown.Zero;
736  RHSKnown.One.lshrInPlace(C);
737  Known.One |= RHSKnown.One;
738  // assume(~(v << c) = a)
739  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
740  m_Value(A))) &&
741  Pred == ICmpInst::ICMP_EQ &&
742  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
743  C < BitWidth) {
744  KnownBits RHSKnown(BitWidth);
745  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
746  // For those bits in RHS that are known, we can propagate them inverted
747  // to known bits in V shifted to the right by C.
748  RHSKnown.One.lshrInPlace(C);
749  Known.Zero |= RHSKnown.One;
750  RHSKnown.Zero.lshrInPlace(C);
751  Known.One |= RHSKnown.Zero;
752  // assume(v >> c = a)
753  } else if (match(Arg,
754  m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
755  m_Value(A))) &&
756  Pred == ICmpInst::ICMP_EQ &&
757  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
758  C < BitWidth) {
759  KnownBits RHSKnown(BitWidth);
760  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
761  // For those bits in RHS that are known, we can propagate them to known
762  // bits in V shifted to the right by C.
763  Known.Zero |= RHSKnown.Zero << C;
764  Known.One |= RHSKnown.One << C;
765  // assume(~(v >> c) = a)
766  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
767  m_Value(A))) &&
768  Pred == ICmpInst::ICMP_EQ &&
769  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
770  C < BitWidth) {
771  KnownBits RHSKnown(BitWidth);
772  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
773  // For those bits in RHS that are known, we can propagate them inverted
774  // to known bits in V shifted to the right by C.
775  Known.Zero |= RHSKnown.One << C;
776  Known.One |= RHSKnown.Zero << C;
777  // assume(v >=_s c) where c is non-negative
778  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
779  Pred == ICmpInst::ICMP_SGE &&
780  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
781  KnownBits RHSKnown(BitWidth);
782  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
783 
784  if (RHSKnown.isNonNegative()) {
785  // We know that the sign bit is zero.
786  Known.makeNonNegative();
787  }
788  // assume(v >_s c) where c is at least -1.
789  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
790  Pred == ICmpInst::ICMP_SGT &&
791  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
792  KnownBits RHSKnown(BitWidth);
793  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
794 
795  if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
796  // We know that the sign bit is zero.
797  Known.makeNonNegative();
798  }
799  // assume(v <=_s c) where c is negative
800  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
801  Pred == ICmpInst::ICMP_SLE &&
802  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
803  KnownBits RHSKnown(BitWidth);
804  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
805 
806  if (RHSKnown.isNegative()) {
807  // We know that the sign bit is one.
808  Known.makeNegative();
809  }
810  // assume(v <_s c) where c is non-positive
811  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
812  Pred == ICmpInst::ICMP_SLT &&
813  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
814  KnownBits RHSKnown(BitWidth);
815  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
816 
817  if (RHSKnown.isZero() || RHSKnown.isNegative()) {
818  // We know that the sign bit is one.
819  Known.makeNegative();
820  }
821  // assume(v <=_u c)
822  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
823  Pred == ICmpInst::ICMP_ULE &&
824  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
825  KnownBits RHSKnown(BitWidth);
826  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
827 
828  // Whatever high bits in c are zero are known to be zero.
829  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
830  // assume(v <_u c)
831  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
832  Pred == ICmpInst::ICMP_ULT &&
833  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
834  KnownBits RHSKnown(BitWidth);
835  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
836 
837  // If the RHS is known zero, then this assumption must be wrong (nothing
838  // is unsigned less than zero). Signal a conflict and get out of here.
839  if (RHSKnown.isZero()) {
840  Known.Zero.setAllBits();
841  Known.One.setAllBits();
842  break;
843  }
844 
845  // Whatever high bits in c are zero are known to be zero (if c is a power
846  // of 2, then one more).
847  if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
848  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
849  else
850  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
851  }
852  }
853 
854  // If assumptions conflict with each other or previous known bits, then we
855  // have a logical fallacy. It's possible that the assumption is not reachable,
856  // so this isn't a real bug. On the other hand, the program may have undefined
857  // behavior, or we might have a bug in the compiler. We can't assert/crash, so
858  // clear out the known bits, try to warn the user, and hope for the best.
859  if (Known.Zero.intersects(Known.One)) {
860  Known.resetAll();
861 
862  if (Q.ORE)
863  Q.ORE->emit([&]() {
864  auto *CxtI = const_cast<Instruction *>(Q.CxtI);
865  return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
866  CxtI)
867  << "Detected conflicting code assumptions. Program may "
868  "have undefined behavior, or compiler may have "
869  "internal error.";
870  });
871  }
872 }
873 
874 /// Compute known bits from a shift operator, including those with a
875 /// non-constant shift amount. Known is the output of this function. Known2 is a
876 /// pre-allocated temporary with the same bit width as Known. KZF and KOF are
877 /// operator-specific functions that, given the known-zero or known-one bits
878 /// respectively, and a shift amount, compute the implied known-zero or
879 /// known-one bits of the shift operator's result respectively for that shift
880 /// amount. The results from calling KZF and KOF are conservatively combined for
881 /// all permitted shift amounts.
883  const Operator *I, KnownBits &Known, KnownBits &Known2,
884  unsigned Depth, const Query &Q,
885  function_ref<APInt(const APInt &, unsigned)> KZF,
886  function_ref<APInt(const APInt &, unsigned)> KOF) {
887  unsigned BitWidth = Known.getBitWidth();
888 
889  if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
890  unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
891 
892  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
893  Known.Zero = KZF(Known.Zero, ShiftAmt);
894  Known.One = KOF(Known.One, ShiftAmt);
895  // If the known bits conflict, this must be an overflowing left shift, so
896  // the shift result is poison. We can return anything we want. Choose 0 for
897  // the best folding opportunity.
898  if (Known.hasConflict())
899  Known.setAllZero();
900 
901  return;
902  }
903 
904  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
905 
906  // If the shift amount could be greater than or equal to the bit-width of the
907  // LHS, the value could be poison, but bail out because the check below is
908  // expensive. TODO: Should we just carry on?
909  if ((~Known.Zero).uge(BitWidth)) {
910  Known.resetAll();
911  return;
912  }
913 
914  // Note: We cannot use Known.Zero.getLimitedValue() here, because if
915  // BitWidth > 64 and any upper bits are known, we'll end up returning the
916  // limit value (which implies all bits are known).
917  uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
918  uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
919 
920  // It would be more-clearly correct to use the two temporaries for this
921  // calculation. Reusing the APInts here to prevent unnecessary allocations.
922  Known.resetAll();
923 
924  // If we know the shifter operand is nonzero, we can sometimes infer more
925  // known bits. However this is expensive to compute, so be lazy about it and
926  // only compute it when absolutely necessary.
927  Optional<bool> ShifterOperandIsNonZero;
928 
929  // Early exit if we can't constrain any well-defined shift amount.
930  if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
931  !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
932  ShifterOperandIsNonZero = isKnownNonZero(I->getOperand(1), Depth + 1, Q);
933  if (!*ShifterOperandIsNonZero)
934  return;
935  }
936 
937  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
938 
939  Known.Zero.setAllBits();
940  Known.One.setAllBits();
941  for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
942  // Combine the shifted known input bits only for those shift amounts
943  // compatible with its known constraints.
944  if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
945  continue;
946  if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
947  continue;
948  // If we know the shifter is nonzero, we may be able to infer more known
949  // bits. This check is sunk down as far as possible to avoid the expensive
950  // call to isKnownNonZero if the cheaper checks above fail.
951  if (ShiftAmt == 0) {
952  if (!ShifterOperandIsNonZero.hasValue())
953  ShifterOperandIsNonZero =
954  isKnownNonZero(I->getOperand(1), Depth + 1, Q);
955  if (*ShifterOperandIsNonZero)
956  continue;
957  }
958 
959  Known.Zero &= KZF(Known2.Zero, ShiftAmt);
960  Known.One &= KOF(Known2.One, ShiftAmt);
961  }
962 
963  // If the known bits conflict, the result is poison. Return a 0 and hope the
964  // caller can further optimize that.
965  if (Known.hasConflict())
966  Known.setAllZero();
967 }
968 
969 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
970  unsigned Depth, const Query &Q) {
971  unsigned BitWidth = Known.getBitWidth();
972 
973  KnownBits Known2(Known);
974  switch (I->getOpcode()) {
975  default: break;
976  case Instruction::Load:
977  if (MDNode *MD =
978  Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range))
980  break;
981  case Instruction::And: {
982  // If either the LHS or the RHS are Zero, the result is zero.
983  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
984  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
985 
986  // Output known-1 bits are only known if set in both the LHS & RHS.
987  Known.One &= Known2.One;
988  // Output known-0 are known to be clear if zero in either the LHS | RHS.
989  Known.Zero |= Known2.Zero;
990 
991  // and(x, add (x, -1)) is a common idiom that always clears the low bit;
992  // here we handle the more general case of adding any odd number by
993  // matching the form add(x, add(x, y)) where y is odd.
994  // TODO: This could be generalized to clearing any bit set in y where the
995  // following bit is known to be unset in y.
996  Value *X = nullptr, *Y = nullptr;
997  if (!Known.Zero[0] && !Known.One[0] &&
998  match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) {
999  Known2.resetAll();
1000  computeKnownBits(Y, Known2, Depth + 1, Q);
1001  if (Known2.countMinTrailingOnes() > 0)
1002  Known.Zero.setBit(0);
1003  }
1004  break;
1005  }
1006  case Instruction::Or:
1007  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1008  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1009 
1010  // Output known-0 bits are only known if clear in both the LHS & RHS.
1011  Known.Zero &= Known2.Zero;
1012  // Output known-1 are known to be set if set in either the LHS | RHS.
1013  Known.One |= Known2.One;
1014  break;
1015  case Instruction::Xor: {
1016  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1017  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1018 
1019  // Output known-0 bits are known if clear or set in both the LHS & RHS.
1020  APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
1021  // Output known-1 are known to be set if set in only one of the LHS, RHS.
1022  Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
1023  Known.Zero = std::move(KnownZeroOut);
1024  break;
1025  }
1026  case Instruction::Mul: {
1027  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1028  computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
1029  Known2, Depth, Q);
1030  break;
1031  }
1032  case Instruction::UDiv: {
1033  // For the purposes of computing leading zeros we can conservatively
1034  // treat a udiv as a logical right shift by the power of 2 known to
1035  // be less than the denominator.
1036  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1037  unsigned LeadZ = Known2.countMinLeadingZeros();
1038 
1039  Known2.resetAll();
1040  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1041  unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
1042  if (RHSMaxLeadingZeros != BitWidth)
1043  LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
1044 
1045  Known.Zero.setHighBits(LeadZ);
1046  break;
1047  }
1048  case Instruction::Select: {
1049  const Value *LHS, *RHS;
1050  SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
1052  computeKnownBits(RHS, Known, Depth + 1, Q);
1053  computeKnownBits(LHS, Known2, Depth + 1, Q);
1054  } else {
1055  computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1056  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1057  }
1058 
1059  unsigned MaxHighOnes = 0;
1060  unsigned MaxHighZeros = 0;
1061  if (SPF == SPF_SMAX) {
1062  // If both sides are negative, the result is negative.
1063  if (Known.isNegative() && Known2.isNegative())
1064  // We can derive a lower bound on the result by taking the max of the
1065  // leading one bits.
1066  MaxHighOnes =
1068  // If either side is non-negative, the result is non-negative.
1069  else if (Known.isNonNegative() || Known2.isNonNegative())
1070  MaxHighZeros = 1;
1071  } else if (SPF == SPF_SMIN) {
1072  // If both sides are non-negative, the result is non-negative.
1073  if (Known.isNonNegative() && Known2.isNonNegative())
1074  // We can derive an upper bound on the result by taking the max of the
1075  // leading zero bits.
1076  MaxHighZeros = std::max(Known.countMinLeadingZeros(),
1077  Known2.countMinLeadingZeros());
1078  // If either side is negative, the result is negative.
1079  else if (Known.isNegative() || Known2.isNegative())
1080  MaxHighOnes = 1;
1081  } else if (SPF == SPF_UMAX) {
1082  // We can derive a lower bound on the result by taking the max of the
1083  // leading one bits.
1084  MaxHighOnes =
1086  } else if (SPF == SPF_UMIN) {
1087  // We can derive an upper bound on the result by taking the max of the
1088  // leading zero bits.
1089  MaxHighZeros =
1091  } else if (SPF == SPF_ABS) {
1092  // RHS from matchSelectPattern returns the negation part of abs pattern.
1093  // If the negate has an NSW flag we can assume the sign bit of the result
1094  // will be 0 because that makes abs(INT_MIN) undefined.
1095  if (Q.IIQ.hasNoSignedWrap(cast<Instruction>(RHS)))
1096  MaxHighZeros = 1;
1097  }
1098 
1099  // Only known if known in both the LHS and RHS.
1100  Known.One &= Known2.One;
1101  Known.Zero &= Known2.Zero;
1102  if (MaxHighOnes > 0)
1103  Known.One.setHighBits(MaxHighOnes);
1104  if (MaxHighZeros > 0)
1105  Known.Zero.setHighBits(MaxHighZeros);
1106  break;
1107  }
1108  case Instruction::FPTrunc:
1109  case Instruction::FPExt:
1110  case Instruction::FPToUI:
1111  case Instruction::FPToSI:
1112  case Instruction::SIToFP:
1113  case Instruction::UIToFP:
1114  break; // Can't work with floating point.
1115  case Instruction::PtrToInt:
1116  case Instruction::IntToPtr:
1117  // Fall through and handle them the same as zext/trunc.
1119  case Instruction::ZExt:
1120  case Instruction::Trunc: {
1121  Type *SrcTy = I->getOperand(0)->getType();
1122 
1123  unsigned SrcBitWidth;
1124  // Note that we handle pointer operands here because of inttoptr/ptrtoint
1125  // which fall through here.
1126  Type *ScalarTy = SrcTy->getScalarType();
1127  SrcBitWidth = ScalarTy->isPointerTy() ?
1128  Q.DL.getIndexTypeSizeInBits(ScalarTy) :
1129  Q.DL.getTypeSizeInBits(ScalarTy);
1130 
1131  assert(SrcBitWidth && "SrcBitWidth can't be zero");
1132  Known = Known.zextOrTrunc(SrcBitWidth);
1133  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1134  Known = Known.zextOrTrunc(BitWidth);
1135  // Any top bits are known to be zero.
1136  if (BitWidth > SrcBitWidth)
1137  Known.Zero.setBitsFrom(SrcBitWidth);
1138  break;
1139  }
1140  case Instruction::BitCast: {
1141  Type *SrcTy = I->getOperand(0)->getType();
1142  if (SrcTy->isIntOrPtrTy() &&
1143  // TODO: For now, not handling conversions like:
1144  // (bitcast i64 %x to <2 x i32>)
1145  !I->getType()->isVectorTy()) {
1146  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1147  break;
1148  }
1149  break;
1150  }
1151  case Instruction::SExt: {
1152  // Compute the bits in the result that are not present in the input.
1153  unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1154 
1155  Known = Known.trunc(SrcBitWidth);
1156  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1157  // If the sign bit of the input is known set or clear, then we know the
1158  // top bits of the result.
1159  Known = Known.sext(BitWidth);
1160  break;
1161  }
1162  case Instruction::Shl: {
1163  // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1164  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1165  auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1166  APInt KZResult = KnownZero << ShiftAmt;
1167  KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1168  // If this shift has "nsw" keyword, then the result is either a poison
1169  // value or has the same sign bit as the first operand.
1170  if (NSW && KnownZero.isSignBitSet())
1171  KZResult.setSignBit();
1172  return KZResult;
1173  };
1174 
1175  auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1176  APInt KOResult = KnownOne << ShiftAmt;
1177  if (NSW && KnownOne.isSignBitSet())
1178  KOResult.setSignBit();
1179  return KOResult;
1180  };
1181 
1182  computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1183  break;
1184  }
1185  case Instruction::LShr: {
1186  // (lshr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1187  auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1188  APInt KZResult = KnownZero.lshr(ShiftAmt);
1189  // High bits known zero.
1190  KZResult.setHighBits(ShiftAmt);
1191  return KZResult;
1192  };
1193 
1194  auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1195  return KnownOne.lshr(ShiftAmt);
1196  };
1197 
1198  computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1199  break;
1200  }
1201  case Instruction::AShr: {
1202  // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1203  auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1204  return KnownZero.ashr(ShiftAmt);
1205  };
1206 
1207  auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1208  return KnownOne.ashr(ShiftAmt);
1209  };
1210 
1211  computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1212  break;
1213  }
1214  case Instruction::Sub: {
1215  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1216  computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1217  Known, Known2, Depth, Q);
1218  break;
1219  }
1220  case Instruction::Add: {
1221  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1222  computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1223  Known, Known2, Depth, Q);
1224  break;
1225  }
1226  case Instruction::SRem:
1227  if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1228  APInt RA = Rem->getValue().abs();
1229  if (RA.isPowerOf2()) {
1230  APInt LowBits = RA - 1;
1231  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1232 
1233  // The low bits of the first operand are unchanged by the srem.
1234  Known.Zero = Known2.Zero & LowBits;
1235  Known.One = Known2.One & LowBits;
1236 
1237  // If the first operand is non-negative or has all low bits zero, then
1238  // the upper bits are all zero.
1239  if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1240  Known.Zero |= ~LowBits;
1241 
1242  // If the first operand is negative and not all low bits are zero, then
1243  // the upper bits are all one.
1244  if (Known2.isNegative() && LowBits.intersects(Known2.One))
1245  Known.One |= ~LowBits;
1246 
1247  assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1248  break;
1249  }
1250  }
1251 
1252  // The sign bit is the LHS's sign bit, except when the result of the
1253  // remainder is zero.
1254  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1255  // If it's known zero, our sign bit is also zero.
1256  if (Known2.isNonNegative())
1257  Known.makeNonNegative();
1258 
1259  break;
1260  case Instruction::URem: {
1261  if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1262  const APInt &RA = Rem->getValue();
1263  if (RA.isPowerOf2()) {
1264  APInt LowBits = (RA - 1);
1265  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1266  Known.Zero |= ~LowBits;
1267  Known.One &= LowBits;
1268  break;
1269  }
1270  }
1271 
1272  // Since the result is less than or equal to either operand, any leading
1273  // zero bits in either operand must also exist in the result.
1274  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1275  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1276 
1277  unsigned Leaders =
1279  Known.resetAll();
1280  Known.Zero.setHighBits(Leaders);
1281  break;
1282  }
1283 
1284  case Instruction::Alloca: {
1285  const AllocaInst *AI = cast<AllocaInst>(I);
1286  unsigned Align = AI->getAlignment();
1287  if (Align == 0)
1288  Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1289 
1290  if (Align > 0)
1291  Known.Zero.setLowBits(countTrailingZeros(Align));
1292  break;
1293  }
1294  case Instruction::GetElementPtr: {
1295  // Analyze all of the subscripts of this getelementptr instruction
1296  // to determine if we can prove known low zero bits.
1297  KnownBits LocalKnown(BitWidth);
1298  computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1299  unsigned TrailZ = LocalKnown.countMinTrailingZeros();
1300 
1302  for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1303  Value *Index = I->getOperand(i);
1304  if (StructType *STy = GTI.getStructTypeOrNull()) {
1305  // Handle struct member offset arithmetic.
1306 
1307  // Handle case when index is vector zeroinitializer
1308  Constant *CIndex = cast<Constant>(Index);
1309  if (CIndex->isZeroValue())
1310  continue;
1311 
1312  if (CIndex->getType()->isVectorTy())
1313  Index = CIndex->getSplatValue();
1314 
1315  unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1316  const StructLayout *SL = Q.DL.getStructLayout(STy);
1317  uint64_t Offset = SL->getElementOffset(Idx);
1318  TrailZ = std::min<unsigned>(TrailZ,
1319  countTrailingZeros(Offset));
1320  } else {
1321  // Handle array index arithmetic.
1322  Type *IndexedTy = GTI.getIndexedType();
1323  if (!IndexedTy->isSized()) {
1324  TrailZ = 0;
1325  break;
1326  }
1327  unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1328  uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1329  LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1330  computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1331  TrailZ = std::min(TrailZ,
1332  unsigned(countTrailingZeros(TypeSize) +
1333  LocalKnown.countMinTrailingZeros()));
1334  }
1335  }
1336 
1337  Known.Zero.setLowBits(TrailZ);
1338  break;
1339  }
1340  case Instruction::PHI: {
1341  const PHINode *P = cast<PHINode>(I);
1342  // Handle the case of a simple two-predecessor recurrence PHI.
1343  // There's a lot more that could theoretically be done here, but
1344  // this is sufficient to catch some interesting cases.
1345  if (P->getNumIncomingValues() == 2) {
1346  for (unsigned i = 0; i != 2; ++i) {
1347  Value *L = P->getIncomingValue(i);
1348  Value *R = P->getIncomingValue(!i);
1349  Operator *LU = dyn_cast<Operator>(L);
1350  if (!LU)
1351  continue;
1352  unsigned Opcode = LU->getOpcode();
1353  // Check for operations that have the property that if
1354  // both their operands have low zero bits, the result
1355  // will have low zero bits.
1356  if (Opcode == Instruction::Add ||
1357  Opcode == Instruction::Sub ||
1358  Opcode == Instruction::And ||
1359  Opcode == Instruction::Or ||
1360  Opcode == Instruction::Mul) {
1361  Value *LL = LU->getOperand(0);
1362  Value *LR = LU->getOperand(1);
1363  // Find a recurrence.
1364  if (LL == I)
1365  L = LR;
1366  else if (LR == I)
1367  L = LL;
1368  else
1369  break;
1370  // Ok, we have a PHI of the form L op= R. Check for low
1371  // zero bits.
1372  computeKnownBits(R, Known2, Depth + 1, Q);
1373 
1374  // We need to take the minimum number of known bits
1375  KnownBits Known3(Known);
1376  computeKnownBits(L, Known3, Depth + 1, Q);
1377 
1378  Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1379  Known3.countMinTrailingZeros()));
1380 
1381  auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1382  if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) {
1383  // If initial value of recurrence is nonnegative, and we are adding
1384  // a nonnegative number with nsw, the result can only be nonnegative
1385  // or poison value regardless of the number of times we execute the
1386  // add in phi recurrence. If initial value is negative and we are
1387  // adding a negative number with nsw, the result can only be
1388  // negative or poison value. Similar arguments apply to sub and mul.
1389  //
1390  // (add non-negative, non-negative) --> non-negative
1391  // (add negative, negative) --> negative
1392  if (Opcode == Instruction::Add) {
1393  if (Known2.isNonNegative() && Known3.isNonNegative())
1394  Known.makeNonNegative();
1395  else if (Known2.isNegative() && Known3.isNegative())
1396  Known.makeNegative();
1397  }
1398 
1399  // (sub nsw non-negative, negative) --> non-negative
1400  // (sub nsw negative, non-negative) --> negative
1401  else if (Opcode == Instruction::Sub && LL == I) {
1402  if (Known2.isNonNegative() && Known3.isNegative())
1403  Known.makeNonNegative();
1404  else if (Known2.isNegative() && Known3.isNonNegative())
1405  Known.makeNegative();
1406  }
1407 
1408  // (mul nsw non-negative, non-negative) --> non-negative
1409  else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1410  Known3.isNonNegative())
1411  Known.makeNonNegative();
1412  }
1413 
1414  break;
1415  }
1416  }
1417  }
1418 
1419  // Unreachable blocks may have zero-operand PHI nodes.
1420  if (P->getNumIncomingValues() == 0)
1421  break;
1422 
1423  // Otherwise take the unions of the known bit sets of the operands,
1424  // taking conservative care to avoid excessive recursion.
1425  if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1426  // Skip if every incoming value references to ourself.
1427  if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1428  break;
1429 
1430  Known.Zero.setAllBits();
1431  Known.One.setAllBits();
1432  for (Value *IncValue : P->incoming_values()) {
1433  // Skip direct self references.
1434  if (IncValue == P) continue;
1435 
1436  Known2 = KnownBits(BitWidth);
1437  // Recurse, but cap the recursion to one level, because we don't
1438  // want to waste time spinning around in loops.
1439  computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1440  Known.Zero &= Known2.Zero;
1441  Known.One &= Known2.One;
1442  // If all bits have been ruled out, there's no need to check
1443  // more operands.
1444  if (!Known.Zero && !Known.One)
1445  break;
1446  }
1447  }
1448  break;
1449  }
1450  case Instruction::Call:
1451  case Instruction::Invoke:
1452  // If range metadata is attached to this call, set known bits from that,
1453  // and then intersect with known bits based on other properties of the
1454  // function.
1455  if (MDNode *MD =
1456  Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range))
1458  if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1459  computeKnownBits(RV, Known2, Depth + 1, Q);
1460  Known.Zero |= Known2.Zero;
1461  Known.One |= Known2.One;
1462  }
1463  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1464  switch (II->getIntrinsicID()) {
1465  default: break;
1466  case Intrinsic::bitreverse:
1467  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1468  Known.Zero |= Known2.Zero.reverseBits();
1469  Known.One |= Known2.One.reverseBits();
1470  break;
1471  case Intrinsic::bswap:
1472  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1473  Known.Zero |= Known2.Zero.byteSwap();
1474  Known.One |= Known2.One.byteSwap();
1475  break;
1476  case Intrinsic::ctlz: {
1477  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1478  // If we have a known 1, its position is our upper bound.
1479  unsigned PossibleLZ = Known2.One.countLeadingZeros();
1480  // If this call is undefined for 0, the result will be less than 2^n.
1481  if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1482  PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1483  unsigned LowBits = Log2_32(PossibleLZ)+1;
1484  Known.Zero.setBitsFrom(LowBits);
1485  break;
1486  }
1487  case Intrinsic::cttz: {
1488  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1489  // If we have a known 1, its position is our upper bound.
1490  unsigned PossibleTZ = Known2.One.countTrailingZeros();
1491  // If this call is undefined for 0, the result will be less than 2^n.
1492  if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1493  PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1494  unsigned LowBits = Log2_32(PossibleTZ)+1;
1495  Known.Zero.setBitsFrom(LowBits);
1496  break;
1497  }
1498  case Intrinsic::ctpop: {
1499  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1500  // We can bound the space the count needs. Also, bits known to be zero
1501  // can't contribute to the population.
1502  unsigned BitsPossiblySet = Known2.countMaxPopulation();
1503  unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1504  Known.Zero.setBitsFrom(LowBits);
1505  // TODO: we could bound KnownOne using the lower bound on the number
1506  // of bits which might be set provided by popcnt KnownOne2.
1507  break;
1508  }
1509  case Intrinsic::x86_sse42_crc32_64_64:
1510  Known.Zero.setBitsFrom(32);
1511  break;
1512  }
1513  }
1514  break;
1515  case Instruction::ExtractElement:
1516  // Look through extract element. At the moment we keep this simple and skip
1517  // tracking the specific element. But at least we might find information
1518  // valid for all elements of the vector (for example if vector is sign
1519  // extended, shifted, etc).
1520  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1521  break;
1522  case Instruction::ExtractValue:
1523  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1524  const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1525  if (EVI->getNumIndices() != 1) break;
1526  if (EVI->getIndices()[0] == 0) {
1527  switch (II->getIntrinsicID()) {
1528  default: break;
1529  case Intrinsic::uadd_with_overflow:
1530  case Intrinsic::sadd_with_overflow:
1531  computeKnownBitsAddSub(true, II->getArgOperand(0),
1532  II->getArgOperand(1), false, Known, Known2,
1533  Depth, Q);
1534  break;
1535  case Intrinsic::usub_with_overflow:
1536  case Intrinsic::ssub_with_overflow:
1537  computeKnownBitsAddSub(false, II->getArgOperand(0),
1538  II->getArgOperand(1), false, Known, Known2,
1539  Depth, Q);
1540  break;
1541  case Intrinsic::umul_with_overflow:
1542  case Intrinsic::smul_with_overflow:
1543  computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1544  Known, Known2, Depth, Q);
1545  break;
1546  }
1547  }
1548  }
1549  }
1550 }
1551 
1552 /// Determine which bits of V are known to be either zero or one and return
1553 /// them.
1554 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1555  KnownBits Known(getBitWidth(V->getType(), Q.DL));
1556  computeKnownBits(V, Known, Depth, Q);
1557  return Known;
1558 }
1559 
1560 /// Determine which bits of V are known to be either zero or one and return
1561 /// them in the Known bit set.
1562 ///
1563 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1564 /// we cannot optimize based on the assumption that it is zero without changing
1565 /// it to be an explicit zero. If we don't change it to zero, other code could
1566 /// optimized based on the contradictory assumption that it is non-zero.
1567 /// Because instcombine aggressively folds operations with undef args anyway,
1568 /// this won't lose us code quality.
1569 ///
1570 /// This function is defined on values with integer type, values with pointer
1571 /// type, and vectors of integers. In the case
1572 /// where V is a vector, known zero, and known one values are the
1573 /// same width as the vector element, and the bit is set only if it is true
1574 /// for all of the elements in the vector.
1575 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1576  const Query &Q) {
1577  assert(V && "No Value?");
1578  assert(Depth <= MaxDepth && "Limit Search Depth");
1579  unsigned BitWidth = Known.getBitWidth();
1580 
1581  assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||
1582  V->getType()->isPtrOrPtrVectorTy()) &&
1583  "Not integer or pointer type!");
1584 
1585  Type *ScalarTy = V->getType()->getScalarType();
1586  unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
1587  Q.DL.getIndexTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
1588  assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth");
1589  (void)BitWidth;
1590  (void)ExpectedWidth;
1591 
1592  const APInt *C;
1593  if (match(V, m_APInt(C))) {
1594  // We know all of the bits for a scalar constant or a splat vector constant!
1595  Known.One = *C;
1596  Known.Zero = ~Known.One;
1597  return;
1598  }
1599  // Null and aggregate-zero are all-zeros.
1600  if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1601  Known.setAllZero();
1602  return;
1603  }
1604  // Handle a constant vector by taking the intersection of the known bits of
1605  // each element.
1606  if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1607  // We know that CDS must be a vector of integers. Take the intersection of
1608  // each element.
1609  Known.Zero.setAllBits(); Known.One.setAllBits();
1610  for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1611  APInt Elt = CDS->getElementAsAPInt(i);
1612  Known.Zero &= ~Elt;
1613  Known.One &= Elt;
1614  }
1615  return;
1616  }
1617 
1618  if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1619  // We know that CV must be a vector of integers. Take the intersection of
1620  // each element.
1621  Known.Zero.setAllBits(); Known.One.setAllBits();
1622  for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1623  Constant *Element = CV->getAggregateElement(i);
1624  auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1625  if (!ElementCI) {
1626  Known.resetAll();
1627  return;
1628  }
1629  const APInt &Elt = ElementCI->getValue();
1630  Known.Zero &= ~Elt;
1631  Known.One &= Elt;
1632  }
1633  return;
1634  }
1635 
1636  // Start out not knowing anything.
1637  Known.resetAll();
1638 
1639  // We can't imply anything about undefs.
1640  if (isa<UndefValue>(V))
1641  return;
1642 
1643  // There's no point in looking through other users of ConstantData for
1644  // assumptions. Confirm that we've handled them all.
1645  assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1646 
1647  // Limit search depth.
1648  // All recursive calls that increase depth must come after this.
1649  if (Depth == MaxDepth)
1650  return;
1651 
1652  // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1653  // the bits of its aliasee.
1654  if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1655  if (!GA->isInterposable())
1656  computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1657  return;
1658  }
1659 
1660  if (const Operator *I = dyn_cast<Operator>(V))
1661  computeKnownBitsFromOperator(I, Known, Depth, Q);
1662 
1663  // Aligned pointers have trailing zeros - refine Known.Zero set
1664  if (V->getType()->isPointerTy()) {
1665  unsigned Align = V->getPointerAlignment(Q.DL);
1666  if (Align)
1667  Known.Zero.setLowBits(countTrailingZeros(Align));
1668  }
1669 
1670  // computeKnownBitsFromAssume strictly refines Known.
1671  // Therefore, we run them after computeKnownBitsFromOperator.
1672 
1673  // Check whether a nearby assume intrinsic can determine some known bits.
1674  computeKnownBitsFromAssume(V, Known, Depth, Q);
1675 
1676  assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1677 }
1678 
1679 /// Return true if the given value is known to have exactly one
1680 /// bit set when defined. For vectors return true if every element is known to
1681 /// be a power of two when defined. Supports values with integer or pointer
1682 /// types and vectors of integers.
1683 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1684  const Query &Q) {
1685  assert(Depth <= MaxDepth && "Limit Search Depth");
1686 
1687  // Attempt to match against constants.
1688  if (OrZero && match(V, m_Power2OrZero()))
1689  return true;
1690  if (match(V, m_Power2()))
1691  return true;
1692 
1693  // 1 << X is clearly a power of two if the one is not shifted off the end. If
1694  // it is shifted off the end then the result is undefined.
1695  if (match(V, m_Shl(m_One(), m_Value())))
1696  return true;
1697 
1698  // (signmask) >>l X is clearly a power of two if the one is not shifted off
1699  // the bottom. If it is shifted off the bottom then the result is undefined.
1700  if (match(V, m_LShr(m_SignMask(), m_Value())))
1701  return true;
1702 
1703  // The remaining tests are all recursive, so bail out if we hit the limit.
1704  if (Depth++ == MaxDepth)
1705  return false;
1706 
1707  Value *X = nullptr, *Y = nullptr;
1708  // A shift left or a logical shift right of a power of two is a power of two
1709  // or zero.
1710  if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1711  match(V, m_LShr(m_Value(X), m_Value()))))
1712  return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1713 
1714  if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1715  return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1716 
1717  if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1718  return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1719  isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1720 
1721  if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1722  // A power of two and'd with anything is a power of two or zero.
1723  if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1724  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1725  return true;
1726  // X & (-X) is always a power of two or zero.
1727  if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1728  return true;
1729  return false;
1730  }
1731 
1732  // Adding a power-of-two or zero to the same power-of-two or zero yields
1733  // either the original power-of-two, a larger power-of-two or zero.
1734  if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1735  const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1736  if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) ||
1737  Q.IIQ.hasNoSignedWrap(VOBO)) {
1738  if (match(X, m_And(m_Specific(Y), m_Value())) ||
1739  match(X, m_And(m_Value(), m_Specific(Y))))
1740  if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1741  return true;
1742  if (match(Y, m_And(m_Specific(X), m_Value())) ||
1743  match(Y, m_And(m_Value(), m_Specific(X))))
1744  if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1745  return true;
1746 
1747  unsigned BitWidth = V->getType()->getScalarSizeInBits();
1748  KnownBits LHSBits(BitWidth);
1749  computeKnownBits(X, LHSBits, Depth, Q);
1750 
1751  KnownBits RHSBits(BitWidth);
1752  computeKnownBits(Y, RHSBits, Depth, Q);
1753  // If i8 V is a power of two or zero:
1754  // ZeroBits: 1 1 1 0 1 1 1 1
1755  // ~ZeroBits: 0 0 0 1 0 0 0 0
1756  if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1757  // If OrZero isn't set, we cannot give back a zero result.
1758  // Make sure either the LHS or RHS has a bit set.
1759  if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1760  return true;
1761  }
1762  }
1763 
1764  // An exact divide or right shift can only shift off zero bits, so the result
1765  // is a power of two only if the first operand is a power of two and not
1766  // copying a sign bit (sdiv int_min, 2).
1767  if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1768  match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1769  return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1770  Depth, Q);
1771  }
1772 
1773  return false;
1774 }
1775 
1776 /// Test whether a GEP's result is known to be non-null.
1777 ///
1778 /// Uses properties inherent in a GEP to try to determine whether it is known
1779 /// to be non-null.
1780 ///
1781 /// Currently this routine does not support vector GEPs.
1782 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1783  const Query &Q) {
1784  const Function *F = nullptr;
1785  if (const Instruction *I = dyn_cast<Instruction>(GEP))
1786  F = I->getFunction();
1787 
1788  if (!GEP->isInBounds() ||
1790  return false;
1791 
1792  // FIXME: Support vector-GEPs.
1793  assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1794 
1795  // If the base pointer is non-null, we cannot walk to a null address with an
1796  // inbounds GEP in address space zero.
1797  if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1798  return true;
1799 
1800  // Walk the GEP operands and see if any operand introduces a non-zero offset.
1801  // If so, then the GEP cannot produce a null pointer, as doing so would
1802  // inherently violate the inbounds contract within address space zero.
1803  for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1804  GTI != GTE; ++GTI) {
1805  // Struct types are easy -- they must always be indexed by a constant.
1806  if (StructType *STy = GTI.getStructTypeOrNull()) {
1807  ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1808  unsigned ElementIdx = OpC->getZExtValue();
1809  const StructLayout *SL = Q.DL.getStructLayout(STy);
1810  uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1811  if (ElementOffset > 0)
1812  return true;
1813  continue;
1814  }
1815 
1816  // If we have a zero-sized type, the index doesn't matter. Keep looping.
1817  if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1818  continue;
1819 
1820  // Fast path the constant operand case both for efficiency and so we don't
1821  // increment Depth when just zipping down an all-constant GEP.
1822  if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1823  if (!OpC->isZero())
1824  return true;
1825  continue;
1826  }
1827 
1828  // We post-increment Depth here because while isKnownNonZero increments it
1829  // as well, when we pop back up that increment won't persist. We don't want
1830  // to recurse 10k times just because we have 10k GEP operands. We don't
1831  // bail completely out because we want to handle constant GEPs regardless
1832  // of depth.
1833  if (Depth++ >= MaxDepth)
1834  continue;
1835 
1836  if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1837  return true;
1838  }
1839 
1840  return false;
1841 }
1842 
1844  const Instruction *CtxI,
1845  const DominatorTree *DT) {
1846  assert(V->getType()->isPointerTy() && "V must be pointer type");
1847  assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
1848 
1849  if (!CtxI || !DT)
1850  return false;
1851 
1852  unsigned NumUsesExplored = 0;
1853  for (auto *U : V->users()) {
1854  // Avoid massive lists
1855  if (NumUsesExplored >= DomConditionsMaxUses)
1856  break;
1857  NumUsesExplored++;
1858 
1859  // If the value is used as an argument to a call or invoke, then argument
1860  // attributes may provide an answer about null-ness.
1861  if (auto CS = ImmutableCallSite(U))
1862  if (auto *CalledFunc = CS.getCalledFunction())
1863  for (const Argument &Arg : CalledFunc->args())
1864  if (CS.getArgOperand(Arg.getArgNo()) == V &&
1865  Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
1866  return true;
1867 
1868  // Consider only compare instructions uniquely controlling a branch
1869  CmpInst::Predicate Pred;
1870  if (!match(const_cast<User *>(U),
1871  m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
1872  (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
1873  continue;
1874 
1877  for (auto *CmpU : U->users()) {
1878  assert(WorkList.empty() && "Should be!");
1879  if (Visited.insert(CmpU).second)
1880  WorkList.push_back(CmpU);
1881 
1882  while (!WorkList.empty()) {
1883  auto *Curr = WorkList.pop_back_val();
1884 
1885  // If a user is an AND, add all its users to the work list. We only
1886  // propagate "pred != null" condition through AND because it is only
1887  // correct to assume that all conditions of AND are met in true branch.
1888  // TODO: Support similar logic of OR and EQ predicate?
1889  if (Pred == ICmpInst::ICMP_NE)
1890  if (auto *BO = dyn_cast<BinaryOperator>(Curr))
1891  if (BO->getOpcode() == Instruction::And) {
1892  for (auto *BOU : BO->users())
1893  if (Visited.insert(BOU).second)
1894  WorkList.push_back(BOU);
1895  continue;
1896  }
1897 
1898  if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) {
1899  assert(BI->isConditional() && "uses a comparison!");
1900 
1901  BasicBlock *NonNullSuccessor =
1902  BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
1903  BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
1904  if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
1905  return true;
1906  } else if (Pred == ICmpInst::ICMP_NE && isGuard(Curr) &&
1907  DT->dominates(cast<Instruction>(Curr), CtxI)) {
1908  return true;
1909  }
1910  }
1911  }
1912  }
1913 
1914  return false;
1915 }
1916 
1917 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1918 /// ensure that the value it's attached to is never Value? 'RangeType' is
1919 /// is the type of the value described by the range.
1920 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1921  const unsigned NumRanges = Ranges->getNumOperands() / 2;
1922  assert(NumRanges >= 1);
1923  for (unsigned i = 0; i < NumRanges; ++i) {
1924  ConstantInt *Lower =
1925  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1926  ConstantInt *Upper =
1927  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1928  ConstantRange Range(Lower->getValue(), Upper->getValue());
1929  if (Range.contains(Value))
1930  return false;
1931  }
1932  return true;
1933 }
1934 
1935 /// Return true if the given value is known to be non-zero when defined. For
1936 /// vectors, return true if every element is known to be non-zero when
1937 /// defined. For pointers, if the context instruction and dominator tree are
1938 /// specified, perform context-sensitive analysis and return true if the
1939 /// pointer couldn't possibly be null at the specified instruction.
1940 /// Supports values with integer or pointer type and vectors of integers.
1941 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1942  if (auto *C = dyn_cast<Constant>(V)) {
1943  if (C->isNullValue())
1944  return false;
1945  if (isa<ConstantInt>(C))
1946  // Must be non-zero due to null test above.
1947  return true;
1948 
1949  // For constant vectors, check that all elements are undefined or known
1950  // non-zero to determine that the whole vector is known non-zero.
1951  if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1952  for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1953  Constant *Elt = C->getAggregateElement(i);
1954  if (!Elt || Elt->isNullValue())
1955  return false;
1956  if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1957  return false;
1958  }
1959  return true;
1960  }
1961 
1962  // A global variable in address space 0 is non null unless extern weak
1963  // or an absolute symbol reference. Other address spaces may have null as a
1964  // valid address for a global, so we can't assume anything.
1965  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
1966  if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
1967  GV->getType()->getAddressSpace() == 0)
1968  return true;
1969  } else
1970  return false;
1971  }
1972 
1973  if (auto *I = dyn_cast<Instruction>(V)) {
1974  if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) {
1975  // If the possible ranges don't contain zero, then the value is
1976  // definitely non-zero.
1977  if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1978  const APInt ZeroValue(Ty->getBitWidth(), 0);
1979  if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1980  return true;
1981  }
1982  }
1983  }
1984 
1985  // Some of the tests below are recursive, so bail out if we hit the limit.
1986  if (Depth++ >= MaxDepth)
1987  return false;
1988 
1989  // Check for pointer simplifications.
1990  if (V->getType()->isPointerTy()) {
1991  // Alloca never returns null, malloc might.
1992  if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
1993  return true;
1994 
1995  // A byval, inalloca, or nonnull argument is never null.
1996  if (const Argument *A = dyn_cast<Argument>(V))
1997  if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
1998  return true;
1999 
2000  // A Load tagged with nonnull metadata is never null.
2001  if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2002  if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull))
2003  return true;
2004 
2005  if (auto CS = ImmutableCallSite(V)) {
2006  if (CS.isReturnNonNull())
2007  return true;
2008  if (const auto *RP = getArgumentAliasingToReturnedPointer(CS))
2009  return isKnownNonZero(RP, Depth, Q);
2010  }
2011  }
2012 
2013 
2014  // Check for recursive pointer simplifications.
2015  if (V->getType()->isPointerTy()) {
2016  if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
2017  return true;
2018 
2019  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
2020  if (isGEPKnownNonNull(GEP, Depth, Q))
2021  return true;
2022  }
2023 
2024  unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
2025 
2026  // X | Y != 0 if X != 0 or Y != 0.
2027  Value *X = nullptr, *Y = nullptr;
2028  if (match(V, m_Or(m_Value(X), m_Value(Y))))
2029  return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
2030 
2031  // ext X != 0 if X != 0.
2032  if (isa<SExtInst>(V) || isa<ZExtInst>(V))
2033  return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2034 
2035  // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2036  // if the lowest bit is shifted off the end.
2037  if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2038  // shl nuw can't remove any non-zero bits.
2039  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2040  if (Q.IIQ.hasNoUnsignedWrap(BO))
2041  return isKnownNonZero(X, Depth, Q);
2042 
2043  KnownBits Known(BitWidth);
2044  computeKnownBits(X, Known, Depth, Q);
2045  if (Known.One[0])
2046  return true;
2047  }
2048  // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2049  // defined if the sign bit is shifted off the end.
2050  else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2051  // shr exact can only shift out zero bits.
2052  const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2053  if (BO->isExact())
2054  return isKnownNonZero(X, Depth, Q);
2055 
2056  KnownBits Known = computeKnownBits(X, Depth, Q);
2057  if (Known.isNegative())
2058  return true;
2059 
2060  // If the shifter operand is a constant, and all of the bits shifted
2061  // out are known to be zero, and X is known non-zero then at least one
2062  // non-zero bit must remain.
2063  if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2064  auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2065  // Is there a known one in the portion not shifted out?
2066  if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2067  return true;
2068  // Are all the bits to be shifted out known zero?
2069  if (Known.countMinTrailingZeros() >= ShiftVal)
2070  return isKnownNonZero(X, Depth, Q);
2071  }
2072  }
2073  // div exact can only produce a zero if the dividend is zero.
2074  else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2075  return isKnownNonZero(X, Depth, Q);
2076  }
2077  // X + Y.
2078  else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2079  KnownBits XKnown = computeKnownBits(X, Depth, Q);
2080  KnownBits YKnown = computeKnownBits(Y, Depth, Q);
2081 
2082  // If X and Y are both non-negative (as signed values) then their sum is not
2083  // zero unless both X and Y are zero.
2084  if (XKnown.isNonNegative() && YKnown.isNonNegative())
2085  if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
2086  return true;
2087 
2088  // If X and Y are both negative (as signed values) then their sum is not
2089  // zero unless both X and Y equal INT_MIN.
2090  if (XKnown.isNegative() && YKnown.isNegative()) {
2091  APInt Mask = APInt::getSignedMaxValue(BitWidth);
2092  // The sign bit of X is set. If some other bit is set then X is not equal
2093  // to INT_MIN.
2094  if (XKnown.One.intersects(Mask))
2095  return true;
2096  // The sign bit of Y is set. If some other bit is set then Y is not equal
2097  // to INT_MIN.
2098  if (YKnown.One.intersects(Mask))
2099  return true;
2100  }
2101 
2102  // The sum of a non-negative number and a power of two is not zero.
2103  if (XKnown.isNonNegative() &&
2104  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2105  return true;
2106  if (YKnown.isNonNegative() &&
2107  isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2108  return true;
2109  }
2110  // X * Y.
2111  else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2112  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2113  // If X and Y are non-zero then so is X * Y as long as the multiplication
2114  // does not overflow.
2115  if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) &&
2116  isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
2117  return true;
2118  }
2119  // (C ? X : Y) != 0 if X != 0 and Y != 0.
2120  else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2121  if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
2122  isKnownNonZero(SI->getFalseValue(), Depth, Q))
2123  return true;
2124  }
2125  // PHI
2126  else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2127  // Try and detect a recurrence that monotonically increases from a
2128  // starting value, as these are common as induction variables.
2129  if (PN->getNumIncomingValues() == 2) {
2130  Value *Start = PN->getIncomingValue(0);
2131  Value *Induction = PN->getIncomingValue(1);
2132  if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2133  std::swap(Start, Induction);
2134  if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2135  if (!C->isZero() && !C->isNegative()) {
2136  ConstantInt *X;
2137  if (Q.IIQ.UseInstrInfo &&
2138  (match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2139  match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2140  !X->isNegative())
2141  return true;
2142  }
2143  }
2144  }
2145  // Check if all incoming values are non-zero constant.
2146  bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
2147  return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
2148  });
2149  if (AllNonZeroConstants)
2150  return true;
2151  }
2152 
2153  KnownBits Known(BitWidth);
2154  computeKnownBits(V, Known, Depth, Q);
2155  return Known.One != 0;
2156 }
2157 
2158 /// Return true if V2 == V1 + X, where X is known non-zero.
2159 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
2160  const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2161  if (!BO || BO->getOpcode() != Instruction::Add)
2162  return false;
2163  Value *Op = nullptr;
2164  if (V2 == BO->getOperand(0))
2165  Op = BO->getOperand(1);
2166  else if (V2 == BO->getOperand(1))
2167  Op = BO->getOperand(0);
2168  else
2169  return false;
2170  return isKnownNonZero(Op, 0, Q);
2171 }
2172 
2173 /// Return true if it is known that V1 != V2.
2174 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
2175  if (V1 == V2)
2176  return false;
2177  if (V1->getType() != V2->getType())
2178  // We can't look through casts yet.
2179  return false;
2180  if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2181  return true;
2182 
2183  if (V1->getType()->isIntOrIntVectorTy()) {
2184  // Are any known bits in V1 contradictory to known bits in V2? If V1
2185  // has a known zero where V2 has a known one, they must not be equal.
2186  KnownBits Known1 = computeKnownBits(V1, 0, Q);
2187  KnownBits Known2 = computeKnownBits(V2, 0, Q);
2188 
2189  if (Known1.Zero.intersects(Known2.One) ||
2190  Known2.Zero.intersects(Known1.One))
2191  return true;
2192  }
2193  return false;
2194 }
2195 
2196 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2197 /// simplify operations downstream. Mask is known to be zero for bits that V
2198 /// cannot have.
2199 ///
2200 /// This function is defined on values with integer type, values with pointer
2201 /// type, and vectors of integers. In the case
2202 /// where V is a vector, the mask, known zero, and known one values are the
2203 /// same width as the vector element, and the bit is set only if it is true
2204 /// for all of the elements in the vector.
2205 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2206  const Query &Q) {
2207  KnownBits Known(Mask.getBitWidth());
2208  computeKnownBits(V, Known, Depth, Q);
2209  return Mask.isSubsetOf(Known.Zero);
2210 }
2211 
2212 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow).
2213 // Returns the input and lower/upper bounds.
2214 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In,
2215  const APInt *&CLow, const APInt *&CHigh) {
2216  assert(isa<Operator>(Select) &&
2217  cast<Operator>(Select)->getOpcode() == Instruction::Select &&
2218  "Input should be a Select!");
2219 
2220  const Value *LHS, *RHS, *LHS2, *RHS2;
2221  SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor;
2222  if (SPF != SPF_SMAX && SPF != SPF_SMIN)
2223  return false;
2224 
2225  if (!match(RHS, m_APInt(CLow)))
2226  return false;
2227 
2228  SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor;
2229  if (getInverseMinMaxFlavor(SPF) != SPF2)
2230  return false;
2231 
2232  if (!match(RHS2, m_APInt(CHigh)))
2233  return false;
2234 
2235  if (SPF == SPF_SMIN)
2236  std::swap(CLow, CHigh);
2237 
2238  In = LHS2;
2239  return CLow->sle(*CHigh);
2240 }
2241 
2242 /// For vector constants, loop over the elements and find the constant with the
2243 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2244 /// or if any element was not analyzed; otherwise, return the count for the
2245 /// element with the minimum number of sign bits.
2246 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2247  unsigned TyBits) {
2248  const auto *CV = dyn_cast<Constant>(V);
2249  if (!CV || !CV->getType()->isVectorTy())
2250  return 0;
2251 
2252  unsigned MinSignBits = TyBits;
2253  unsigned NumElts = CV->getType()->getVectorNumElements();
2254  for (unsigned i = 0; i != NumElts; ++i) {
2255  // If we find a non-ConstantInt, bail out.
2256  auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2257  if (!Elt)
2258  return 0;
2259 
2260  MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2261  }
2262 
2263  return MinSignBits;
2264 }
2265 
2266 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2267  const Query &Q);
2268 
2269 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2270  const Query &Q) {
2271  unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2272  assert(Result > 0 && "At least one sign bit needs to be present!");
2273  return Result;
2274 }
2275 
2276 /// Return the number of times the sign bit of the register is replicated into
2277 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2278 /// (itself), but other cases can give us information. For example, immediately
2279 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2280 /// other, so we return 3. For vectors, return the number of sign bits for the
2281 /// vector element with the minimum number of known sign bits.
2282 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2283  const Query &Q) {
2284  assert(Depth <= MaxDepth && "Limit Search Depth");
2285 
2286  // We return the minimum number of sign bits that are guaranteed to be present
2287  // in V, so for undef we have to conservatively return 1. We don't have the
2288  // same behavior for poison though -- that's a FIXME today.
2289 
2290  Type *ScalarTy = V->getType()->getScalarType();
2291  unsigned TyBits = ScalarTy->isPointerTy() ?
2292  Q.DL.getIndexTypeSizeInBits(ScalarTy) :
2293  Q.DL.getTypeSizeInBits(ScalarTy);
2294 
2295  unsigned Tmp, Tmp2;
2296  unsigned FirstAnswer = 1;
2297 
2298  // Note that ConstantInt is handled by the general computeKnownBits case
2299  // below.
2300 
2301  if (Depth == MaxDepth)
2302  return 1; // Limit search depth.
2303 
2304  const Operator *U = dyn_cast<Operator>(V);
2305  switch (Operator::getOpcode(V)) {
2306  default: break;
2307  case Instruction::SExt:
2308  Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2309  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2310 
2311  case Instruction::SDiv: {
2312  const APInt *Denominator;
2313  // sdiv X, C -> adds log(C) sign bits.
2314  if (match(U->getOperand(1), m_APInt(Denominator))) {
2315 
2316  // Ignore non-positive denominator.
2317  if (!Denominator->isStrictlyPositive())
2318  break;
2319 
2320  // Calculate the incoming numerator bits.
2321  unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2322 
2323  // Add floor(log(C)) bits to the numerator bits.
2324  return std::min(TyBits, NumBits + Denominator->logBase2());
2325  }
2326  break;
2327  }
2328 
2329  case Instruction::SRem: {
2330  const APInt *Denominator;
2331  // srem X, C -> we know that the result is within [-C+1,C) when C is a
2332  // positive constant. This let us put a lower bound on the number of sign
2333  // bits.
2334  if (match(U->getOperand(1), m_APInt(Denominator))) {
2335 
2336  // Ignore non-positive denominator.
2337  if (!Denominator->isStrictlyPositive())
2338  break;
2339 
2340  // Calculate the incoming numerator bits. SRem by a positive constant
2341  // can't lower the number of sign bits.
2342  unsigned NumrBits =
2343  ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2344 
2345  // Calculate the leading sign bit constraints by examining the
2346  // denominator. Given that the denominator is positive, there are two
2347  // cases:
2348  //
2349  // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2350  // (1 << ceilLogBase2(C)).
2351  //
2352  // 2. the numerator is negative. Then the result range is (-C,0] and
2353  // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2354  //
2355  // Thus a lower bound on the number of sign bits is `TyBits -
2356  // ceilLogBase2(C)`.
2357 
2358  unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2359  return std::max(NumrBits, ResBits);
2360  }
2361  break;
2362  }
2363 
2364  case Instruction::AShr: {
2365  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2366  // ashr X, C -> adds C sign bits. Vectors too.
2367  const APInt *ShAmt;
2368  if (match(U->getOperand(1), m_APInt(ShAmt))) {
2369  if (ShAmt->uge(TyBits))
2370  break; // Bad shift.
2371  unsigned ShAmtLimited = ShAmt->getZExtValue();
2372  Tmp += ShAmtLimited;
2373  if (Tmp > TyBits) Tmp = TyBits;
2374  }
2375  return Tmp;
2376  }
2377  case Instruction::Shl: {
2378  const APInt *ShAmt;
2379  if (match(U->getOperand(1), m_APInt(ShAmt))) {
2380  // shl destroys sign bits.
2381  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2382  if (ShAmt->uge(TyBits) || // Bad shift.
2383  ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
2384  Tmp2 = ShAmt->getZExtValue();
2385  return Tmp - Tmp2;
2386  }
2387  break;
2388  }
2389  case Instruction::And:
2390  case Instruction::Or:
2391  case Instruction::Xor: // NOT is handled here.
2392  // Logical binary ops preserve the number of sign bits at the worst.
2393  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2394  if (Tmp != 1) {
2395  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2396  FirstAnswer = std::min(Tmp, Tmp2);
2397  // We computed what we know about the sign bits as our first
2398  // answer. Now proceed to the generic code that uses
2399  // computeKnownBits, and pick whichever answer is better.
2400  }
2401  break;
2402 
2403  case Instruction::Select: {
2404  // If we have a clamp pattern, we know that the number of sign bits will be
2405  // the minimum of the clamp min/max range.
2406  const Value *X;
2407  const APInt *CLow, *CHigh;
2408  if (isSignedMinMaxClamp(U, X, CLow, CHigh))
2409  return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
2410 
2411  Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2412  if (Tmp == 1) break;
2413  Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2414  return std::min(Tmp, Tmp2);
2415  }
2416 
2417  case Instruction::Add:
2418  // Add can have at most one carry bit. Thus we know that the output
2419  // is, at worst, one more bit than the inputs.
2420  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2421  if (Tmp == 1) break;
2422 
2423  // Special case decrementing a value (ADD X, -1):
2424  if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2425  if (CRHS->isAllOnesValue()) {
2426  KnownBits Known(TyBits);
2427  computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2428 
2429  // If the input is known to be 0 or 1, the output is 0/-1, which is all
2430  // sign bits set.
2431  if ((Known.Zero | 1).isAllOnesValue())
2432  return TyBits;
2433 
2434  // If we are subtracting one from a positive number, there is no carry
2435  // out of the result.
2436  if (Known.isNonNegative())
2437  return Tmp;
2438  }
2439 
2440  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2441  if (Tmp2 == 1) break;
2442  return std::min(Tmp, Tmp2)-1;
2443 
2444  case Instruction::Sub:
2445  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2446  if (Tmp2 == 1) break;
2447 
2448  // Handle NEG.
2449  if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2450  if (CLHS->isNullValue()) {
2451  KnownBits Known(TyBits);
2452  computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2453  // If the input is known to be 0 or 1, the output is 0/-1, which is all
2454  // sign bits set.
2455  if ((Known.Zero | 1).isAllOnesValue())
2456  return TyBits;
2457 
2458  // If the input is known to be positive (the sign bit is known clear),
2459  // the output of the NEG has the same number of sign bits as the input.
2460  if (Known.isNonNegative())
2461  return Tmp2;
2462 
2463  // Otherwise, we treat this like a SUB.
2464  }
2465 
2466  // Sub can have at most one carry bit. Thus we know that the output
2467  // is, at worst, one more bit than the inputs.
2468  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2469  if (Tmp == 1) break;
2470  return std::min(Tmp, Tmp2)-1;
2471 
2472  case Instruction::Mul: {
2473  // The output of the Mul can be at most twice the valid bits in the inputs.
2474  unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2475  if (SignBitsOp0 == 1) break;
2476  unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2477  if (SignBitsOp1 == 1) break;
2478  unsigned OutValidBits =
2479  (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
2480  return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
2481  }
2482 
2483  case Instruction::PHI: {
2484  const PHINode *PN = cast<PHINode>(U);
2485  unsigned NumIncomingValues = PN->getNumIncomingValues();
2486  // Don't analyze large in-degree PHIs.
2487  if (NumIncomingValues > 4) break;
2488  // Unreachable blocks may have zero-operand PHI nodes.
2489  if (NumIncomingValues == 0) break;
2490 
2491  // Take the minimum of all incoming values. This can't infinitely loop
2492  // because of our depth threshold.
2493  Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2494  for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2495  if (Tmp == 1) return Tmp;
2496  Tmp = std::min(
2497  Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2498  }
2499  return Tmp;
2500  }
2501 
2502  case Instruction::Trunc:
2503  // FIXME: it's tricky to do anything useful for this, but it is an important
2504  // case for targets like X86.
2505  break;
2506 
2507  case Instruction::ExtractElement:
2508  // Look through extract element. At the moment we keep this simple and skip
2509  // tracking the specific element. But at least we might find information
2510  // valid for all elements of the vector (for example if vector is sign
2511  // extended, shifted, etc).
2512  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2513 
2514  case Instruction::ShuffleVector: {
2515  // TODO: This is copied almost directly from the SelectionDAG version of
2516  // ComputeNumSignBits. It would be better if we could share common
2517  // code. If not, make sure that changes are translated to the DAG.
2518 
2519  // Collect the minimum number of sign bits that are shared by every vector
2520  // element referenced by the shuffle.
2521  auto *Shuf = cast<ShuffleVectorInst>(U);
2522  int NumElts = Shuf->getOperand(0)->getType()->getVectorNumElements();
2523  int NumMaskElts = Shuf->getMask()->getType()->getVectorNumElements();
2524  APInt DemandedLHS(NumElts, 0), DemandedRHS(NumElts, 0);
2525  for (int i = 0; i != NumMaskElts; ++i) {
2526  int M = Shuf->getMaskValue(i);
2527  assert(M < NumElts * 2 && "Invalid shuffle mask constant");
2528  // For undef elements, we don't know anything about the common state of
2529  // the shuffle result.
2530  if (M == -1)
2531  return 1;
2532  if (M < NumElts)
2533  DemandedLHS.setBit(M % NumElts);
2534  else
2535  DemandedRHS.setBit(M % NumElts);
2536  }
2538  if (!!DemandedLHS)
2539  Tmp = ComputeNumSignBits(Shuf->getOperand(0), Depth + 1, Q);
2540  if (!!DemandedRHS) {
2541  Tmp2 = ComputeNumSignBits(Shuf->getOperand(1), Depth + 1, Q);
2542  Tmp = std::min(Tmp, Tmp2);
2543  }
2544  // If we don't know anything, early out and try computeKnownBits fall-back.
2545  if (Tmp == 1)
2546  break;
2547  assert(Tmp <= V->getType()->getScalarSizeInBits() &&
2548  "Failed to determine minimum sign bits");
2549  return Tmp;
2550  }
2551  }
2552 
2553  // Finally, if we can prove that the top bits of the result are 0's or 1's,
2554  // use this information.
2555 
2556  // If we can examine all elements of a vector constant successfully, we're
2557  // done (we can't do any better than that). If not, keep trying.
2558  if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2559  return VecSignBits;
2560 
2561  KnownBits Known(TyBits);
2562  computeKnownBits(V, Known, Depth, Q);
2563 
2564  // If we know that the sign bit is either zero or one, determine the number of
2565  // identical bits in the top of the input value.
2566  return std::max(FirstAnswer, Known.countMinSignBits());
2567 }
2568 
2569 /// This function computes the integer multiple of Base that equals V.
2570 /// If successful, it returns true and returns the multiple in
2571 /// Multiple. If unsuccessful, it returns false. It looks
2572 /// through SExt instructions only if LookThroughSExt is true.
2573 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2574  bool LookThroughSExt, unsigned Depth) {
2575  const unsigned MaxDepth = 6;
2576 
2577  assert(V && "No Value?");
2578  assert(Depth <= MaxDepth && "Limit Search Depth");
2579  assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2580 
2581  Type *T = V->getType();
2582 
2583  ConstantInt *CI = dyn_cast<ConstantInt>(V);
2584 
2585  if (Base == 0)
2586  return false;
2587 
2588  if (Base == 1) {
2589  Multiple = V;
2590  return true;
2591  }
2592 
2594  Constant *BaseVal = ConstantInt::get(T, Base);
2595  if (CO && CO == BaseVal) {
2596  // Multiple is 1.
2597  Multiple = ConstantInt::get(T, 1);
2598  return true;
2599  }
2600 
2601  if (CI && CI->getZExtValue() % Base == 0) {
2602  Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2603  return true;
2604  }
2605 
2606  if (Depth == MaxDepth) return false; // Limit search depth.
2607 
2608  Operator *I = dyn_cast<Operator>(V);
2609  if (!I) return false;
2610 
2611  switch (I->getOpcode()) {
2612  default: break;
2613  case Instruction::SExt:
2614  if (!LookThroughSExt) return false;
2615  // otherwise fall through to ZExt
2617  case Instruction::ZExt:
2618  return ComputeMultiple(I->getOperand(0), Base, Multiple,
2619  LookThroughSExt, Depth+1);
2620  case Instruction::Shl:
2621  case Instruction::Mul: {
2622  Value *Op0 = I->getOperand(0);
2623  Value *Op1 = I->getOperand(1);
2624 
2625  if (I->getOpcode() == Instruction::Shl) {
2626  ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2627  if (!Op1CI) return false;
2628  // Turn Op0 << Op1 into Op0 * 2^Op1
2629  APInt Op1Int = Op1CI->getValue();
2630  uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2631  APInt API(Op1Int.getBitWidth(), 0);
2632  API.setBit(BitToSet);
2633  Op1 = ConstantInt::get(V->getContext(), API);
2634  }
2635 
2636  Value *Mul0 = nullptr;
2637  if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2638  if (Constant *Op1C = dyn_cast<Constant>(Op1))
2639  if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2640  if (Op1C->getType()->getPrimitiveSizeInBits() <
2641  MulC->getType()->getPrimitiveSizeInBits())
2642  Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2643  if (Op1C->getType()->getPrimitiveSizeInBits() >
2644  MulC->getType()->getPrimitiveSizeInBits())
2645  MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2646 
2647  // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2648  Multiple = ConstantExpr::getMul(MulC, Op1C);
2649  return true;
2650  }
2651 
2652  if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2653  if (Mul0CI->getValue() == 1) {
2654  // V == Base * Op1, so return Op1
2655  Multiple = Op1;
2656  return true;
2657  }
2658  }
2659 
2660  Value *Mul1 = nullptr;
2661  if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2662  if (Constant *Op0C = dyn_cast<Constant>(Op0))
2663  if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2664  if (Op0C->getType()->getPrimitiveSizeInBits() <
2665  MulC->getType()->getPrimitiveSizeInBits())
2666  Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2667  if (Op0C->getType()->getPrimitiveSizeInBits() >
2668  MulC->getType()->getPrimitiveSizeInBits())
2669  MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2670 
2671  // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2672  Multiple = ConstantExpr::getMul(MulC, Op0C);
2673  return true;
2674  }
2675 
2676  if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2677  if (Mul1CI->getValue() == 1) {
2678  // V == Base * Op0, so return Op0
2679  Multiple = Op0;
2680  return true;
2681  }
2682  }
2683  }
2684  }
2685 
2686  // We could not determine if V is a multiple of Base.
2687  return false;
2688 }
2689 
2691  const TargetLibraryInfo *TLI) {
2692  const Function *F = ICS.getCalledFunction();
2693  if (!F)
2694  return Intrinsic::not_intrinsic;
2695 
2696  if (F->isIntrinsic())
2697  return F->getIntrinsicID();
2698 
2699  if (!TLI)
2700  return Intrinsic::not_intrinsic;
2701 
2702  LibFunc Func;
2703  // We're going to make assumptions on the semantics of the functions, check
2704  // that the target knows that it's available in this environment and it does
2705  // not have local linkage.
2706  if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2707  return Intrinsic::not_intrinsic;
2708 
2709  if (!ICS.onlyReadsMemory())
2710  return Intrinsic::not_intrinsic;
2711 
2712  // Otherwise check if we have a call to a function that can be turned into a
2713  // vector intrinsic.
2714  switch (Func) {
2715  default:
2716  break;
2717  case LibFunc_sin:
2718  case LibFunc_sinf:
2719  case LibFunc_sinl:
2720  return Intrinsic::sin;
2721  case LibFunc_cos:
2722  case LibFunc_cosf:
2723  case LibFunc_cosl:
2724  return Intrinsic::cos;
2725  case LibFunc_exp:
2726  case LibFunc_expf:
2727  case LibFunc_expl:
2728  return Intrinsic::exp;
2729  case LibFunc_exp2:
2730  case LibFunc_exp2f:
2731  case LibFunc_exp2l:
2732  return Intrinsic::exp2;
2733  case LibFunc_log:
2734  case LibFunc_logf:
2735  case LibFunc_logl:
2736  return Intrinsic::log;
2737  case LibFunc_log10:
2738  case LibFunc_log10f:
2739  case LibFunc_log10l:
2740  return Intrinsic::log10;
2741  case LibFunc_log2:
2742  case LibFunc_log2f:
2743  case LibFunc_log2l:
2744  return Intrinsic::log2;
2745  case LibFunc_fabs:
2746  case LibFunc_fabsf:
2747  case LibFunc_fabsl:
2748  return Intrinsic::fabs;
2749  case LibFunc_fmin:
2750  case LibFunc_fminf:
2751  case LibFunc_fminl:
2752  return Intrinsic::minnum;
2753  case LibFunc_fmax:
2754  case LibFunc_fmaxf:
2755  case LibFunc_fmaxl:
2756  return Intrinsic::maxnum;
2757  case LibFunc_copysign:
2758  case LibFunc_copysignf:
2759  case LibFunc_copysignl:
2760  return Intrinsic::copysign;
2761  case LibFunc_floor:
2762  case LibFunc_floorf:
2763  case LibFunc_floorl:
2764  return Intrinsic::floor;
2765  case LibFunc_ceil:
2766  case LibFunc_ceilf:
2767  case LibFunc_ceill:
2768  return Intrinsic::ceil;
2769  case LibFunc_trunc:
2770  case LibFunc_truncf:
2771  case LibFunc_truncl:
2772  return Intrinsic::trunc;
2773  case LibFunc_rint:
2774  case LibFunc_rintf:
2775  case LibFunc_rintl:
2776  return Intrinsic::rint;
2777  case LibFunc_nearbyint:
2778  case LibFunc_nearbyintf:
2779  case LibFunc_nearbyintl:
2780  return Intrinsic::nearbyint;
2781  case LibFunc_round:
2782  case LibFunc_roundf:
2783  case LibFunc_roundl:
2784  return Intrinsic::round;
2785  case LibFunc_pow:
2786  case LibFunc_powf:
2787  case LibFunc_powl:
2788  return Intrinsic::pow;
2789  case LibFunc_sqrt:
2790  case LibFunc_sqrtf:
2791  case LibFunc_sqrtl:
2792  return Intrinsic::sqrt;
2793  }
2794 
2795  return Intrinsic::not_intrinsic;
2796 }
2797 
2798 /// Return true if we can prove that the specified FP value is never equal to
2799 /// -0.0.
2800 ///
2801 /// NOTE: this function will need to be revisited when we support non-default
2802 /// rounding modes!
2804  unsigned Depth) {
2805  if (auto *CFP = dyn_cast<ConstantFP>(V))
2806  return !CFP->getValueAPF().isNegZero();
2807 
2808  // Limit search depth.
2809  if (Depth == MaxDepth)
2810  return false;
2811 
2812  auto *Op = dyn_cast<Operator>(V);
2813  if (!Op)
2814  return false;
2815 
2816  // Check if the nsz fast-math flag is set.
2817  if (auto *FPO = dyn_cast<FPMathOperator>(Op))
2818  if (FPO->hasNoSignedZeros())
2819  return true;
2820 
2821  // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
2822  if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
2823  return true;
2824 
2825  // sitofp and uitofp turn into +0.0 for zero.
2826  if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
2827  return true;
2828 
2829  if (auto *Call = dyn_cast<CallInst>(Op)) {
2830  Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
2831  switch (IID) {
2832  default:
2833  break;
2834  // sqrt(-0.0) = -0.0, no other negative results are possible.
2835  case Intrinsic::sqrt:
2836  case Intrinsic::canonicalize:
2837  return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
2838  // fabs(x) != -0.0
2839  case Intrinsic::fabs:
2840  return true;
2841  }
2842  }
2843 
2844  return false;
2845 }
2846 
2847 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2848 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2849 /// bit despite comparing equal.
2851  const TargetLibraryInfo *TLI,
2852  bool SignBitOnly,
2853  unsigned Depth) {
2854  // TODO: This function does not do the right thing when SignBitOnly is true
2855  // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2856  // which flips the sign bits of NaNs. See
2857  // https://llvm.org/bugs/show_bug.cgi?id=31702.
2858 
2859  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2860  return !CFP->getValueAPF().isNegative() ||
2861  (!SignBitOnly && CFP->getValueAPF().isZero());
2862  }
2863 
2864  // Handle vector of constants.
2865  if (auto *CV = dyn_cast<Constant>(V)) {
2866  if (CV->getType()->isVectorTy()) {
2867  unsigned NumElts = CV->getType()->getVectorNumElements();
2868  for (unsigned i = 0; i != NumElts; ++i) {
2869  auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
2870  if (!CFP)
2871  return false;
2872  if (CFP->getValueAPF().isNegative() &&
2873  (SignBitOnly || !CFP->getValueAPF().isZero()))
2874  return false;
2875  }
2876 
2877  // All non-negative ConstantFPs.
2878  return true;
2879  }
2880  }
2881 
2882  if (Depth == MaxDepth)
2883  return false; // Limit search depth.
2884 
2885  const Operator *I = dyn_cast<Operator>(V);
2886  if (!I)
2887  return false;
2888 
2889  switch (I->getOpcode()) {
2890  default:
2891  break;
2892  // Unsigned integers are always nonnegative.
2893  case Instruction::UIToFP:
2894  return true;
2895  case Instruction::FMul:
2896  // x*x is always non-negative or a NaN.
2897  if (I->getOperand(0) == I->getOperand(1) &&
2898  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2899  return true;
2900 
2902  case Instruction::FAdd:
2903  case Instruction::FDiv:
2904  case Instruction::FRem:
2905  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2906  Depth + 1) &&
2907  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2908  Depth + 1);
2909  case Instruction::Select:
2910  return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2911  Depth + 1) &&
2912  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2913  Depth + 1);
2914  case Instruction::FPExt:
2915  case Instruction::FPTrunc:
2916  // Widening/narrowing never change sign.
2917  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2918  Depth + 1);
2919  case Instruction::ExtractElement:
2920  // Look through extract element. At the moment we keep this simple and skip
2921  // tracking the specific element. But at least we might find information
2922  // valid for all elements of the vector.
2923  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2924  Depth + 1);
2925  case Instruction::Call:
2926  const auto *CI = cast<CallInst>(I);
2927  Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2928  switch (IID) {
2929  default:
2930  break;
2931  case Intrinsic::maxnum:
2932  return (isKnownNeverNaN(I->getOperand(0), TLI) &&
2934  SignBitOnly, Depth + 1)) ||
2935  (isKnownNeverNaN(I->getOperand(1), TLI) &&
2937  SignBitOnly, Depth + 1));
2938 
2939  case Intrinsic::maximum:
2940  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2941  Depth + 1) ||
2942  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2943  Depth + 1);
2944  case Intrinsic::minnum:
2945  case Intrinsic::minimum:
2946  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2947  Depth + 1) &&
2948  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2949  Depth + 1);
2950  case Intrinsic::exp:
2951  case Intrinsic::exp2:
2952  case Intrinsic::fabs:
2953  return true;
2954 
2955  case Intrinsic::sqrt:
2956  // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2957  if (!SignBitOnly)
2958  return true;
2959  return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2960  CannotBeNegativeZero(CI->getOperand(0), TLI));
2961 
2962  case Intrinsic::powi:
2963  if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2964  // powi(x,n) is non-negative if n is even.
2965  if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2966  return true;
2967  }
2968  // TODO: This is not correct. Given that exp is an integer, here are the
2969  // ways that pow can return a negative value:
2970  //
2971  // pow(x, exp) --> negative if exp is odd and x is negative.
2972  // pow(-0, exp) --> -inf if exp is negative odd.
2973  // pow(-0, exp) --> -0 if exp is positive odd.
2974  // pow(-inf, exp) --> -0 if exp is negative odd.
2975  // pow(-inf, exp) --> -inf if exp is positive odd.
2976  //
2977  // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2978  // but we must return false if x == -0. Unfortunately we do not currently
2979  // have a way of expressing this constraint. See details in
2980  // https://llvm.org/bugs/show_bug.cgi?id=31702.
2981  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2982  Depth + 1);
2983 
2984  case Intrinsic::fma:
2985  case Intrinsic::fmuladd:
2986  // x*x+y is non-negative if y is non-negative.
2987  return I->getOperand(0) == I->getOperand(1) &&
2988  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
2989  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2990  Depth + 1);
2991  }
2992  break;
2993  }
2994  return false;
2995 }
2996 
2998  const TargetLibraryInfo *TLI) {
2999  return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3000 }
3001 
3003  return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3004 }
3005 
3007  unsigned Depth) {
3008  assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
3009 
3010  // If we're told that NaNs won't happen, assume they won't.
3011  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3012  if (FPMathOp->hasNoNaNs())
3013  return true;
3014 
3015  // Handle scalar constants.
3016  if (auto *CFP = dyn_cast<ConstantFP>(V))
3017  return !CFP->isNaN();
3018 
3019  if (Depth == MaxDepth)
3020  return false;
3021 
3022  if (auto *Inst = dyn_cast<Instruction>(V)) {
3023  switch (Inst->getOpcode()) {
3024  case Instruction::FAdd:
3025  case Instruction::FMul:
3026  case Instruction::FSub:
3027  case Instruction::FDiv:
3028  case Instruction::FRem: {
3029  // TODO: Need isKnownNeverInfinity
3030  return false;
3031  }
3032  case Instruction::Select: {
3033  return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3034  isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3035  }
3036  case Instruction::SIToFP:
3037  case Instruction::UIToFP:
3038  return true;
3039  case Instruction::FPTrunc:
3040  case Instruction::FPExt:
3041  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3042  default:
3043  break;
3044  }
3045  }
3046 
3047  if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3048  switch (II->getIntrinsicID()) {
3049  case Intrinsic::canonicalize:
3050  case Intrinsic::fabs:
3051  case Intrinsic::copysign:
3052  case Intrinsic::exp:
3053  case Intrinsic::exp2:
3054  case Intrinsic::floor:
3055  case Intrinsic::ceil:
3056  case Intrinsic::trunc:
3057  case Intrinsic::rint:
3058  case Intrinsic::nearbyint:
3059  case Intrinsic::round:
3060  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3061  case Intrinsic::sqrt:
3062  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3063  CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3064  default:
3065  return false;
3066  }
3067  }
3068 
3069  // Bail out for constant expressions, but try to handle vector constants.
3070  if (!V->getType()->isVectorTy() || !isa<Constant>(V))
3071  return false;
3072 
3073  // For vectors, verify that each element is not NaN.
3074  unsigned NumElts = V->getType()->getVectorNumElements();
3075  for (unsigned i = 0; i != NumElts; ++i) {
3076  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3077  if (!Elt)
3078  return false;
3079  if (isa<UndefValue>(Elt))
3080  continue;
3081  auto *CElt = dyn_cast<ConstantFP>(Elt);
3082  if (!CElt || CElt->isNaN())
3083  return false;
3084  }
3085  // All elements were confirmed not-NaN or undefined.
3086  return true;
3087 }
3088 
3090 
3091  // All byte-wide stores are splatable, even of arbitrary variables.
3092  if (V->getType()->isIntegerTy(8))
3093  return V;
3094 
3095  LLVMContext &Ctx = V->getContext();
3096 
3097  // Undef don't care.
3098  auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
3099  if (isa<UndefValue>(V))
3100  return UndefInt8;
3101 
3102  Constant *C = dyn_cast<Constant>(V);
3103  if (!C) {
3104  // Conceptually, we could handle things like:
3105  // %a = zext i8 %X to i16
3106  // %b = shl i16 %a, 8
3107  // %c = or i16 %a, %b
3108  // but until there is an example that actually needs this, it doesn't seem
3109  // worth worrying about.
3110  return nullptr;
3111  }
3112 
3113  // Handle 'null' ConstantArrayZero etc.
3114  if (C->isNullValue())
3116 
3117  // Constant floating-point values can be handled as integer values if the
3118  // corresponding integer value is "byteable". An important case is 0.0.
3119  if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3120  Type *Ty = nullptr;
3121  if (CFP->getType()->isHalfTy())
3122  Ty = Type::getInt16Ty(Ctx);
3123  else if (CFP->getType()->isFloatTy())
3124  Ty = Type::getInt32Ty(Ctx);
3125  else if (CFP->getType()->isDoubleTy())
3126  Ty = Type::getInt64Ty(Ctx);
3127  // Don't handle long double formats, which have strange constraints.
3128  return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty)) : nullptr;
3129  }
3130 
3131  // We can handle constant integers that are multiple of 8 bits.
3132  if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
3133  if (CI->getBitWidth() % 8 == 0) {
3134  assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
3135  if (!CI->getValue().isSplat(8))
3136  return nullptr;
3137  return ConstantInt::get(Ctx, CI->getValue().trunc(8));
3138  }
3139  }
3140 
3141  auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
3142  if (LHS == RHS)
3143  return LHS;
3144  if (!LHS || !RHS)
3145  return nullptr;
3146  if (LHS == UndefInt8)
3147  return RHS;
3148  if (RHS == UndefInt8)
3149  return LHS;
3150  return nullptr;
3151  };
3152 
3153  if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
3154  Value *Val = UndefInt8;
3155  for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
3156  if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I)))))
3157  return nullptr;
3158  return Val;
3159  }
3160 
3161  if (isa<ConstantVector>(C)) {
3162  Constant *Splat = cast<ConstantVector>(C)->getSplatValue();
3163  return Splat ? isBytewiseValue(Splat) : nullptr;
3164  }
3165 
3166  if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
3167  Value *Val = UndefInt8;
3168  for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
3169  if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I)))))
3170  return nullptr;
3171  return Val;
3172  }
3173 
3174  // Don't try to handle the handful of other constants.
3175  return nullptr;
3176 }
3177 
3178 // This is the recursive version of BuildSubAggregate. It takes a few different
3179 // arguments. Idxs is the index within the nested struct From that we are
3180 // looking at now (which is of type IndexedType). IdxSkip is the number of
3181 // indices from Idxs that should be left out when inserting into the resulting
3182 // struct. To is the result struct built so far, new insertvalue instructions
3183 // build on that.
3184 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
3186  unsigned IdxSkip,
3187  Instruction *InsertBefore) {
3188  StructType *STy = dyn_cast<StructType>(IndexedType);
3189  if (STy) {
3190  // Save the original To argument so we can modify it
3191  Value *OrigTo = To;
3192  // General case, the type indexed by Idxs is a struct
3193  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3194  // Process each struct element recursively
3195  Idxs.push_back(i);
3196  Value *PrevTo = To;
3197  To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3198  InsertBefore);
3199  Idxs.pop_back();
3200  if (!To) {
3201  // Couldn't find any inserted value for this index? Cleanup
3202  while (PrevTo != OrigTo) {
3203  InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3204  PrevTo = Del->getAggregateOperand();
3205  Del->eraseFromParent();
3206  }
3207  // Stop processing elements
3208  break;
3209  }
3210  }
3211  // If we successfully found a value for each of our subaggregates
3212  if (To)
3213  return To;
3214  }
3215  // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3216  // the struct's elements had a value that was inserted directly. In the latter
3217  // case, perhaps we can't determine each of the subelements individually, but
3218  // we might be able to find the complete struct somewhere.
3219 
3220  // Find the value that is at that particular spot
3221  Value *V = FindInsertedValue(From, Idxs);
3222 
3223  if (!V)
3224  return nullptr;
3225 
3226  // Insert the value in the new (sub) aggregate
3227  return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
3228  "tmp", InsertBefore);
3229 }
3230 
3231 // This helper takes a nested struct and extracts a part of it (which is again a
3232 // struct) into a new value. For example, given the struct:
3233 // { a, { b, { c, d }, e } }
3234 // and the indices "1, 1" this returns
3235 // { c, d }.
3236 //
3237 // It does this by inserting an insertvalue for each element in the resulting
3238 // struct, as opposed to just inserting a single struct. This will only work if
3239 // each of the elements of the substruct are known (ie, inserted into From by an
3240 // insertvalue instruction somewhere).
3241 //
3242 // All inserted insertvalue instructions are inserted before InsertBefore
3244  Instruction *InsertBefore) {
3245  assert(InsertBefore && "Must have someplace to insert!");
3246  Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3247  idx_range);
3248  Value *To = UndefValue::get(IndexedType);
3249  SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3250  unsigned IdxSkip = Idxs.size();
3251 
3252  return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3253 }
3254 
3255 /// Given an aggregate and a sequence of indices, see if the scalar value
3256 /// indexed is already around as a register, for example if it was inserted
3257 /// directly into the aggregate.
3258 ///
3259 /// If InsertBefore is not null, this function will duplicate (modified)
3260 /// insertvalues when a part of a nested struct is extracted.
3262  Instruction *InsertBefore) {
3263  // Nothing to index? Just return V then (this is useful at the end of our
3264  // recursion).
3265  if (idx_range.empty())
3266  return V;
3267  // We have indices, so V should have an indexable type.
3268  assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
3269  "Not looking at a struct or array?");
3270  assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
3271  "Invalid indices for type?");
3272 
3273  if (Constant *C = dyn_cast<Constant>(V)) {
3274  C = C->getAggregateElement(idx_range[0]);
3275  if (!C) return nullptr;
3276  return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3277  }
3278 
3279  if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3280  // Loop the indices for the insertvalue instruction in parallel with the
3281  // requested indices
3282  const unsigned *req_idx = idx_range.begin();
3283  for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3284  i != e; ++i, ++req_idx) {
3285  if (req_idx == idx_range.end()) {
3286  // We can't handle this without inserting insertvalues
3287  if (!InsertBefore)
3288  return nullptr;
3289 
3290  // The requested index identifies a part of a nested aggregate. Handle
3291  // this specially. For example,
3292  // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3293  // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3294  // %C = extractvalue {i32, { i32, i32 } } %B, 1
3295  // This can be changed into
3296  // %A = insertvalue {i32, i32 } undef, i32 10, 0
3297  // %C = insertvalue {i32, i32 } %A, i32 11, 1
3298  // which allows the unused 0,0 element from the nested struct to be
3299  // removed.
3300  return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3301  InsertBefore);
3302  }
3303 
3304  // This insert value inserts something else than what we are looking for.
3305  // See if the (aggregate) value inserted into has the value we are
3306  // looking for, then.
3307  if (*req_idx != *i)
3308  return FindInsertedValue(I->getAggregateOperand(), idx_range,
3309  InsertBefore);
3310  }
3311  // If we end up here, the indices of the insertvalue match with those
3312  // requested (though possibly only partially). Now we recursively look at
3313  // the inserted value, passing any remaining indices.
3314  return FindInsertedValue(I->getInsertedValueOperand(),
3315  makeArrayRef(req_idx, idx_range.end()),
3316  InsertBefore);
3317  }
3318 
3319  if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3320  // If we're extracting a value from an aggregate that was extracted from
3321  // something else, we can extract from that something else directly instead.
3322  // However, we will need to chain I's indices with the requested indices.
3323 
3324  // Calculate the number of indices required
3325  unsigned size = I->getNumIndices() + idx_range.size();
3326  // Allocate some space to put the new indices in
3328  Idxs.reserve(size);
3329  // Add indices from the extract value instruction
3330  Idxs.append(I->idx_begin(), I->idx_end());
3331 
3332  // Add requested indices
3333  Idxs.append(idx_range.begin(), idx_range.end());
3334 
3335  assert(Idxs.size() == size
3336  && "Number of indices added not correct?");
3337 
3338  return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
3339  }
3340  // Otherwise, we don't know (such as, extracting from a function return value
3341  // or load instruction)
3342  return nullptr;
3343 }
3344 
3345 /// Analyze the specified pointer to see if it can be expressed as a base
3346 /// pointer plus a constant offset. Return the base and offset to the caller.
3348  const DataLayout &DL) {
3349  unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
3350  APInt ByteOffset(BitWidth, 0);
3351 
3352  // We walk up the defs but use a visited set to handle unreachable code. In
3353  // that case, we stop after accumulating the cycle once (not that it
3354  // matters).
3355  SmallPtrSet<Value *, 16> Visited;
3356  while (Visited.insert(Ptr).second) {
3357  if (Ptr->getType()->isVectorTy())
3358  break;
3359 
3360  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
3361  // If one of the values we have visited is an addrspacecast, then
3362  // the pointer type of this GEP may be different from the type
3363  // of the Ptr parameter which was passed to this function. This
3364  // means when we construct GEPOffset, we need to use the size
3365  // of GEP's pointer type rather than the size of the original
3366  // pointer type.
3367  APInt GEPOffset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
3368  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
3369  break;
3370 
3371  ByteOffset += GEPOffset.getSExtValue();
3372 
3373  Ptr = GEP->getPointerOperand();
3374  } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
3375  Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
3376  Ptr = cast<Operator>(Ptr)->getOperand(0);
3377  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
3378  if (GA->isInterposable())
3379  break;
3380  Ptr = GA->getAliasee();
3381  } else {
3382  break;
3383  }
3384  }
3385  Offset = ByteOffset.getSExtValue();
3386  return Ptr;
3387 }
3388 
3390  unsigned CharSize) {
3391  // Make sure the GEP has exactly three arguments.
3392  if (GEP->getNumOperands() != 3)
3393  return false;
3394 
3395  // Make sure the index-ee is a pointer to array of \p CharSize integers.
3396  // CharSize.
3398  if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
3399  return false;
3400 
3401  // Check to make sure that the first operand of the GEP is an integer and
3402  // has value 0 so that we are sure we're indexing into the initializer.
3403  const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3404  if (!FirstIdx || !FirstIdx->isZero())
3405  return false;
3406 
3407  return true;
3408 }
3409 
3411  ConstantDataArraySlice &Slice,
3412  unsigned ElementSize, uint64_t Offset) {
3413  assert(V);
3414 
3415  // Look through bitcast instructions and geps.
3416  V = V->stripPointerCasts();
3417 
3418  // If the value is a GEP instruction or constant expression, treat it as an
3419  // offset.
3420  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3421  // The GEP operator should be based on a pointer to string constant, and is
3422  // indexing into the string constant.
3423  if (!isGEPBasedOnPointerToString(GEP, ElementSize))
3424  return false;
3425 
3426  // If the second index isn't a ConstantInt, then this is a variable index
3427  // into the array. If this occurs, we can't say anything meaningful about
3428  // the string.
3429  uint64_t StartIdx = 0;
3430  if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3431  StartIdx = CI->getZExtValue();
3432  else
3433  return false;
3434  return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3435  StartIdx + Offset);
3436  }
3437 
3438  // The GEP instruction, constant or instruction, must reference a global
3439  // variable that is a constant and is initialized. The referenced constant
3440  // initializer is the array that we'll use for optimization.
3441  const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3442  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3443  return false;
3444 
3445  const ConstantDataArray *Array;
3446  ArrayType *ArrayTy;
3447  if (GV->getInitializer()->isNullValue()) {
3448  Type *GVTy = GV->getValueType();
3449  if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3450  // A zeroinitializer for the array; there is no ConstantDataArray.
3451  Array = nullptr;
3452  } else {
3453  const DataLayout &DL = GV->getParent()->getDataLayout();
3454  uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3455  uint64_t Length = SizeInBytes / (ElementSize / 8);
3456  if (Length <= Offset)
3457  return false;
3458 
3459  Slice.Array = nullptr;
3460  Slice.Offset = 0;
3461  Slice.Length = Length - Offset;
3462  return true;
3463  }
3464  } else {
3465  // This must be a ConstantDataArray.
3466  Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3467  if (!Array)
3468  return false;
3469  ArrayTy = Array->getType();
3470  }
3471  if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3472  return false;
3473 
3474  uint64_t NumElts = ArrayTy->getArrayNumElements();
3475  if (Offset > NumElts)
3476  return false;
3477 
3478  Slice.Array = Array;
3479  Slice.Offset = Offset;
3480  Slice.Length = NumElts - Offset;
3481  return true;
3482 }
3483 
3484 /// This function computes the length of a null-terminated C string pointed to
3485 /// by V. If successful, it returns true and returns the string in Str.
3486 /// If unsuccessful, it returns false.
3488  uint64_t Offset, bool TrimAtNul) {
3489  ConstantDataArraySlice Slice;
3490  if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3491  return false;
3492 
3493  if (Slice.Array == nullptr) {
3494  if (TrimAtNul) {
3495  Str = StringRef();
3496  return true;
3497  }
3498  if (Slice.Length == 1) {
3499  Str = StringRef("", 1);
3500  return true;
3501  }
3502  // We cannot instantiate a StringRef as we do not have an appropriate string
3503  // of 0s at hand.
3504  return false;
3505  }
3506 
3507  // Start out with the entire array in the StringRef.
3508  Str = Slice.Array->getAsString();
3509  // Skip over 'offset' bytes.
3510  Str = Str.substr(Slice.Offset);
3511 
3512  if (TrimAtNul) {
3513  // Trim off the \0 and anything after it. If the array is not nul
3514  // terminated, we just return the whole end of string. The client may know
3515  // some other way that the string is length-bound.
3516  Str = Str.substr(0, Str.find('\0'));
3517  }
3518  return true;
3519 }
3520 
3521 // These next two are very similar to the above, but also look through PHI
3522 // nodes.
3523 // TODO: See if we can integrate these two together.
3524 
3525 /// If we can compute the length of the string pointed to by
3526 /// the specified pointer, return 'len+1'. If we can't, return 0.
3527 static uint64_t GetStringLengthH(const Value *V,
3529  unsigned CharSize) {
3530  // Look through noop bitcast instructions.
3531  V = V->stripPointerCasts();
3532 
3533  // If this is a PHI node, there are two cases: either we have already seen it
3534  // or we haven't.
3535  if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3536  if (!PHIs.insert(PN).second)
3537  return ~0ULL; // already in the set.
3538 
3539  // If it was new, see if all the input strings are the same length.
3540  uint64_t LenSoFar = ~0ULL;
3541  for (Value *IncValue : PN->incoming_values()) {
3542  uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3543  if (Len == 0) return 0; // Unknown length -> unknown.
3544 
3545  if (Len == ~0ULL) continue;
3546 
3547  if (Len != LenSoFar && LenSoFar != ~0ULL)
3548  return 0; // Disagree -> unknown.
3549  LenSoFar = Len;
3550  }
3551 
3552  // Success, all agree.
3553  return LenSoFar;
3554  }
3555 
3556  // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3557  if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3558  uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3559  if (Len1 == 0) return 0;
3560  uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3561  if (Len2 == 0) return 0;
3562  if (Len1 == ~0ULL) return Len2;
3563  if (Len2 == ~0ULL) return Len1;
3564  if (Len1 != Len2) return 0;
3565  return Len1;
3566  }
3567 
3568  // Otherwise, see if we can read the string.
3569  ConstantDataArraySlice Slice;
3570  if (!getConstantDataArrayInfo(V, Slice, CharSize))
3571  return 0;
3572 
3573  if (Slice.Array == nullptr)
3574  return 1;
3575 
3576  // Search for nul characters
3577  unsigned NullIndex = 0;
3578  for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3579  if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3580  break;
3581  }
3582 
3583  return NullIndex + 1;
3584 }
3585 
3586 /// If we can compute the length of the string pointed to by
3587 /// the specified pointer, return 'len+1'. If we can't, return 0.
3588 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3589  if (!V->getType()->isPointerTy())
3590  return 0;
3591 
3593  uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3594  // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3595  // an empty string as a length.
3596  return Len == ~0ULL ? 1 : Len;
3597 }
3598 
3600  assert(CS &&
3601  "getArgumentAliasingToReturnedPointer only works on nonnull CallSite");
3602  if (const Value *RV = CS.getReturnedArgOperand())
3603  return RV;
3604  // This can be used only as a aliasing property.
3606  return CS.getArgOperand(0);
3607  return nullptr;
3608 }
3609 
3611  ImmutableCallSite CS) {
3612  return CS.getIntrinsicID() == Intrinsic::launder_invariant_group ||
3613  CS.getIntrinsicID() == Intrinsic::strip_invariant_group;
3614 }
3615 
3616 /// \p PN defines a loop-variant pointer to an object. Check if the
3617 /// previous iteration of the loop was referring to the same object as \p PN.
3619  const LoopInfo *LI) {
3620  // Find the loop-defined value.
3621  Loop *L = LI->getLoopFor(PN->getParent());
3622  if (PN->getNumIncomingValues() != 2)
3623  return true;
3624 
3625  // Find the value from previous iteration.
3626  auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3627  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3628  PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3629  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3630  return true;
3631 
3632  // If a new pointer is loaded in the loop, the pointer references a different
3633  // object in every iteration. E.g.:
3634  // for (i)
3635  // int *p = a[i];
3636  // ...
3637  if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3638  if (!L->isLoopInvariant(Load->getPointerOperand()))
3639  return false;
3640  return true;
3641 }
3642 
3644  unsigned MaxLookup) {
3645  if (!V->getType()->isPointerTy())
3646  return V;
3647  for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3648  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3649  V = GEP->getPointerOperand();
3650  } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3651  Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3652  V = cast<Operator>(V)->getOperand(0);
3653  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3654  if (GA->isInterposable())
3655  return V;
3656  V = GA->getAliasee();
3657  } else if (isa<AllocaInst>(V)) {
3658  // An alloca can't be further simplified.
3659  return V;
3660  } else {
3661  if (auto CS = CallSite(V)) {
3662  // CaptureTracking can know about special capturing properties of some
3663  // intrinsics like launder.invariant.group, that can't be expressed with
3664  // the attributes, but have properties like returning aliasing pointer.
3665  // Because some analysis may assume that nocaptured pointer is not
3666  // returned from some special intrinsic (because function would have to
3667  // be marked with returns attribute), it is crucial to use this function
3668  // because it should be in sync with CaptureTracking. Not using it may
3669  // cause weird miscompilations where 2 aliasing pointers are assumed to
3670  // noalias.
3671  if (auto *RP = getArgumentAliasingToReturnedPointer(CS)) {
3672  V = RP;
3673  continue;
3674  }
3675  }
3676 
3677  // See if InstructionSimplify knows any relevant tricks.
3678  if (Instruction *I = dyn_cast<Instruction>(V))
3679  // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3680  if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3681  V = Simplified;
3682  continue;
3683  }
3684 
3685  return V;
3686  }
3687  assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3688  }
3689  return V;
3690 }
3691 
3693  const DataLayout &DL, LoopInfo *LI,
3694  unsigned MaxLookup) {
3695  SmallPtrSet<Value *, 4> Visited;
3696  SmallVector<Value *, 4> Worklist;
3697  Worklist.push_back(V);
3698  do {
3699  Value *P = Worklist.pop_back_val();
3700  P = GetUnderlyingObject(P, DL, MaxLookup);
3701 
3702  if (!Visited.insert(P).second)
3703  continue;
3704 
3705  if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3706  Worklist.push_back(SI->getTrueValue());
3707  Worklist.push_back(SI->getFalseValue());
3708  continue;
3709  }
3710 
3711  if (PHINode *PN = dyn_cast<PHINode>(P)) {
3712  // If this PHI changes the underlying object in every iteration of the
3713  // loop, don't look through it. Consider:
3714  // int **A;
3715  // for (i) {
3716  // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3717  // Curr = A[i];
3718  // *Prev, *Curr;
3719  //
3720  // Prev is tracking Curr one iteration behind so they refer to different
3721  // underlying objects.
3722  if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3724  for (Value *IncValue : PN->incoming_values())
3725  Worklist.push_back(IncValue);
3726  continue;
3727  }
3728 
3729  Objects.push_back(P);
3730  } while (!Worklist.empty());
3731 }
3732 
3733 /// This is the function that does the work of looking through basic
3734 /// ptrtoint+arithmetic+inttoptr sequences.
3735 static const Value *getUnderlyingObjectFromInt(const Value *V) {
3736  do {
3737  if (const Operator *U = dyn_cast<Operator>(V)) {
3738  // If we find a ptrtoint, we can transfer control back to the
3739  // regular getUnderlyingObjectFromInt.
3740  if (U->getOpcode() == Instruction::PtrToInt)
3741  return U->getOperand(0);
3742  // If we find an add of a constant, a multiplied value, or a phi, it's
3743  // likely that the other operand will lead us to the base
3744  // object. We don't have to worry about the case where the
3745  // object address is somehow being computed by the multiply,
3746  // because our callers only care when the result is an
3747  // identifiable object.
3748  if (U->getOpcode() != Instruction::Add ||
3749  (!isa<ConstantInt>(U->getOperand(1)) &&
3750  Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
3751  !isa<PHINode>(U->getOperand(1))))
3752  return V;
3753  V = U->getOperand(0);
3754  } else {
3755  return V;
3756  }
3757  assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
3758  } while (true);
3759 }
3760 
3761 /// This is a wrapper around GetUnderlyingObjects and adds support for basic
3762 /// ptrtoint+arithmetic+inttoptr sequences.
3763 /// It returns false if unidentified object is found in GetUnderlyingObjects.
3765  SmallVectorImpl<Value *> &Objects,
3766  const DataLayout &DL) {
3768  SmallVector<const Value *, 4> Working(1, V);
3769  do {
3770  V = Working.pop_back_val();
3771 
3773  GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
3774 
3775  for (Value *V : Objs) {
3776  if (!Visited.insert(V).second)
3777  continue;
3778  if (Operator::getOpcode(V) == Instruction::IntToPtr) {
3779  const Value *O =
3780  getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
3781  if (O->getType()->isPointerTy()) {
3782  Working.push_back(O);
3783  continue;
3784  }
3785  }
3786  // If GetUnderlyingObjects fails to find an identifiable object,
3787  // getUnderlyingObjectsForCodeGen also fails for safety.
3788  if (!isIdentifiedObject(V)) {
3789  Objects.clear();
3790  return false;
3791  }
3792  Objects.push_back(const_cast<Value *>(V));
3793  }
3794  } while (!Working.empty());
3795  return true;
3796 }
3797 
3798 /// Return true if the only users of this pointer are lifetime markers.
3800  for (const User *U : V->users()) {
3801  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3802  if (!II) return false;
3803 
3804  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3805  II->getIntrinsicID() != Intrinsic::lifetime_end)
3806  return false;
3807  }
3808  return true;
3809 }
3810 
3812  const Instruction *CtxI,
3813  const DominatorTree *DT) {
3814  const Operator *Inst = dyn_cast<Operator>(V);
3815  if (!Inst)
3816  return false;
3817 
3818  for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3819  if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3820  if (C->canTrap())
3821  return false;
3822 
3823  switch (Inst->getOpcode()) {
3824  default:
3825  return true;
3826  case Instruction::UDiv:
3827  case Instruction::URem: {
3828  // x / y is undefined if y == 0.
3829  const APInt *V;
3830  if (match(Inst->getOperand(1), m_APInt(V)))
3831  return *V != 0;
3832  return false;
3833  }
3834  case Instruction::SDiv:
3835  case Instruction::SRem: {
3836  // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3837  const APInt *Numerator, *Denominator;
3838  if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3839  return false;
3840  // We cannot hoist this division if the denominator is 0.
3841  if (*Denominator == 0)
3842  return false;
3843  // It's safe to hoist if the denominator is not 0 or -1.
3844  if (*Denominator != -1)
3845  return true;
3846  // At this point we know that the denominator is -1. It is safe to hoist as
3847  // long we know that the numerator is not INT_MIN.
3848  if (match(Inst->getOperand(0), m_APInt(Numerator)))
3849  return !Numerator->isMinSignedValue();
3850  // The numerator *might* be MinSignedValue.
3851  return false;
3852  }
3853  case Instruction::Load: {
3854  const LoadInst *LI = cast<LoadInst>(Inst);
3855  if (!LI->isUnordered() ||
3856  // Speculative load may create a race that did not exist in the source.
3857  LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3858  // Speculative load may load data from dirty regions.
3859  LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3860  LI->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3861  return false;
3862  const DataLayout &DL = LI->getModule()->getDataLayout();
3864  LI->getAlignment(), DL, CtxI, DT);
3865  }
3866  case Instruction::Call: {
3867  auto *CI = cast<const CallInst>(Inst);
3868  const Function *Callee = CI->getCalledFunction();
3869 
3870  // The called function could have undefined behavior or side-effects, even
3871  // if marked readnone nounwind.
3872  return Callee && Callee->isSpeculatable();
3873  }
3874  case Instruction::VAArg:
3875  case Instruction::Alloca:
3876  case Instruction::Invoke:
3877  case Instruction::PHI:
3878  case Instruction::Store:
3879  case Instruction::Ret:
3880  case Instruction::Br:
3881  case Instruction::IndirectBr:
3882  case Instruction::Switch:
3883  case Instruction::Unreachable:
3884  case Instruction::Fence:
3885  case Instruction::AtomicRMW:
3886  case Instruction::AtomicCmpXchg:
3887  case Instruction::LandingPad:
3888  case Instruction::Resume:
3889  case Instruction::CatchSwitch:
3890  case Instruction::CatchPad:
3891  case Instruction::CatchRet:
3892  case Instruction::CleanupPad:
3893  case Instruction::CleanupRet:
3894  return false; // Misc instructions which have effects
3895  }
3896 }
3897 
3900 }
3901 
3903  const Value *LHS, const Value *RHS, const DataLayout &DL,
3904  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
3905  bool UseInstrInfo) {
3906  // Multiplying n * m significant bits yields a result of n + m significant
3907  // bits. If the total number of significant bits does not exceed the
3908  // result bit width (minus 1), there is no overflow.
3909  // This means if we have enough leading zero bits in the operands
3910  // we can guarantee that the result does not overflow.
3911  // Ref: "Hacker's Delight" by Henry Warren
3912  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3913  KnownBits LHSKnown(BitWidth);
3914  KnownBits RHSKnown(BitWidth);
3915  computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT, nullptr,
3916  UseInstrInfo);
3917  computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT, nullptr,
3918  UseInstrInfo);
3919  // Note that underestimating the number of zero bits gives a more
3920  // conservative answer.
3921  unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
3922  RHSKnown.countMinLeadingZeros();
3923  // First handle the easy case: if we have enough zero bits there's
3924  // definitely no overflow.
3925  if (ZeroBits >= BitWidth)
3927 
3928  // Get the largest possible values for each operand.
3929  APInt LHSMax = ~LHSKnown.Zero;
3930  APInt RHSMax = ~RHSKnown.Zero;
3931 
3932  // We know the multiply operation doesn't overflow if the maximum values for
3933  // each operand will not overflow after we multiply them together.
3934  bool MaxOverflow;
3935  (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3936  if (!MaxOverflow)
3938 
3939  // We know it always overflows if multiplying the smallest possible values for
3940  // the operands also results in overflow.
3941  bool MinOverflow;
3942  (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3943  if (MinOverflow)
3945 
3947 }
3948 
3951  const DataLayout &DL, AssumptionCache *AC,
3952  const Instruction *CxtI,
3953  const DominatorTree *DT, bool UseInstrInfo) {
3954  // Multiplying n * m significant bits yields a result of n + m significant
3955  // bits. If the total number of significant bits does not exceed the
3956  // result bit width (minus 1), there is no overflow.
3957  // This means if we have enough leading sign bits in the operands
3958  // we can guarantee that the result does not overflow.
3959  // Ref: "Hacker's Delight" by Henry Warren
3960  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3961 
3962  // Note that underestimating the number of sign bits gives a more
3963  // conservative answer.
3964  unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
3965  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
3966 
3967  // First handle the easy case: if we have enough sign bits there's
3968  // definitely no overflow.
3969  if (SignBits > BitWidth + 1)
3971 
3972  // There are two ambiguous cases where there can be no overflow:
3973  // SignBits == BitWidth + 1 and
3974  // SignBits == BitWidth
3975  // The second case is difficult to check, therefore we only handle the
3976  // first case.
3977  if (SignBits == BitWidth + 1) {
3978  // It overflows only when both arguments are negative and the true
3979  // product is exactly the minimum negative number.
3980  // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
3981  // For simplicity we just check if at least one side is not negative.
3982  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
3983  nullptr, UseInstrInfo);
3984  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
3985  nullptr, UseInstrInfo);
3986  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
3988  }
3990 }
3991 
3993  const Value *LHS, const Value *RHS, const DataLayout &DL,
3994  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
3995  bool UseInstrInfo) {
3996  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
3997  nullptr, UseInstrInfo);
3998  if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
3999  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4000  nullptr, UseInstrInfo);
4001 
4002  if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
4003  // The sign bit is set in both cases: this MUST overflow.
4004  // Create a simple add instruction, and insert it into the struct.
4006  }
4007 
4008  if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
4009  // The sign bit is clear in both cases: this CANNOT overflow.
4010  // Create a simple add instruction, and insert it into the struct.
4012  }
4013  }
4014 
4016 }
4017 
4018 /// Return true if we can prove that adding the two values of the
4019 /// knownbits will not overflow.
4020 /// Otherwise return false.
4021 static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
4022  const KnownBits &RHSKnown) {
4023  // Addition of two 2's complement numbers having opposite signs will never
4024  // overflow.
4025  if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
4026  (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
4027  return true;
4028 
4029  // If either of the values is known to be non-negative, adding them can only
4030  // overflow if the second is also non-negative, so we can assume that.
4031  // Two non-negative numbers will only overflow if there is a carry to the
4032  // sign bit, so we can check if even when the values are as big as possible
4033  // there is no overflow to the sign bit.
4034  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
4035  APInt MaxLHS = ~LHSKnown.Zero;
4036  MaxLHS.clearSignBit();
4037  APInt MaxRHS = ~RHSKnown.Zero;
4038  MaxRHS.clearSignBit();
4039  APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
4040  return Result.isSignBitClear();
4041  }
4042 
4043  // If either of the values is known to be negative, adding them can only
4044  // overflow if the second is also negative, so we can assume that.
4045  // Two negative number will only overflow if there is no carry to the sign
4046  // bit, so we can check if even when the values are as small as possible
4047  // there is overflow to the sign bit.
4048  if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
4049  APInt MinLHS = LHSKnown.One;
4050  MinLHS.clearSignBit();
4051  APInt MinRHS = RHSKnown.One;
4052  MinRHS.clearSignBit();
4053  APInt Result = std::move(MinLHS) + std::move(MinRHS);
4054  return Result.isSignBitSet();
4055  }
4056 
4057  // If we reached here it means that we know nothing about the sign bits.
4058  // In this case we can't know if there will be an overflow, since by
4059  // changing the sign bits any two values can be made to overflow.
4060  return false;
4061 }
4062 
4064  const Value *RHS,
4065  const AddOperator *Add,
4066  const DataLayout &DL,
4067  AssumptionCache *AC,
4068  const Instruction *CxtI,
4069  const DominatorTree *DT) {
4070  if (Add && Add->hasNoSignedWrap()) {
4072  }
4073 
4074  // If LHS and RHS each have at least two sign bits, the addition will look
4075  // like
4076  //
4077  // XX..... +
4078  // YY.....
4079  //
4080  // If the carry into the most significant position is 0, X and Y can't both
4081  // be 1 and therefore the carry out of the addition is also 0.
4082  //
4083  // If the carry into the most significant position is 1, X and Y can't both
4084  // be 0 and therefore the carry out of the addition is also 1.
4085  //
4086  // Since the carry into the most significant position is always equal to
4087  // the carry out of the addition, there is no signed overflow.
4088  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4089  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4091 
4092  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
4093  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
4094 
4095  if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
4097 
4098  // The remaining code needs Add to be available. Early returns if not so.
4099  if (!Add)
4101 
4102  // If the sign of Add is the same as at least one of the operands, this add
4103  // CANNOT overflow. This is particularly useful when the sum is
4104  // @llvm.assume'ed non-negative rather than proved so from analyzing its
4105  // operands.
4106  bool LHSOrRHSKnownNonNegative =
4107  (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
4108  bool LHSOrRHSKnownNegative =
4109  (LHSKnown.isNegative() || RHSKnown.isNegative());
4110  if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
4111  KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
4112  if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
4113  (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
4115  }
4116  }
4117 
4119 }
4120 
4122  const Value *RHS,
4123  const DataLayout &DL,
4124  AssumptionCache *AC,
4125  const Instruction *CxtI,
4126  const DominatorTree *DT) {
4127  // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
4128  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
4129  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
4130  if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
4132 
4134 }
4135 
4137  const Value *RHS,
4138  const DataLayout &DL,
4139  AssumptionCache *AC,
4140  const Instruction *CxtI,
4141  const DominatorTree *DT) {
4142  // If LHS and RHS each have at least two sign bits, the subtraction
4143  // cannot overflow.
4144  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4145  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4147 
4148  KnownBits LHSKnown = computeKnownBits(LHS, DL, 0, AC, CxtI, DT);
4149 
4150  KnownBits RHSKnown = computeKnownBits(RHS, DL, 0, AC, CxtI, DT);
4151 
4152  // Subtraction of two 2's complement numbers having identical signs will
4153  // never overflow.
4154  if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
4155  (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
4157 
4158  // TODO: implement logic similar to checkRippleForAdd
4160 }
4161 
4163  const DominatorTree &DT) {
4164 #ifndef NDEBUG
4165  auto IID = II->getIntrinsicID();
4166  assert((IID == Intrinsic::sadd_with_overflow ||
4167  IID == Intrinsic::uadd_with_overflow ||
4168  IID == Intrinsic::ssub_with_overflow ||
4169  IID == Intrinsic::usub_with_overflow ||
4170  IID == Intrinsic::smul_with_overflow ||
4171  IID == Intrinsic::umul_with_overflow) &&
4172  "Not an overflow intrinsic!");
4173 #endif
4174 
4175  SmallVector<const BranchInst *, 2> GuardingBranches;
4177 
4178  for (const User *U : II->users()) {
4179  if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
4180  assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
4181 
4182  if (EVI->getIndices()[0] == 0)
4183  Results.push_back(EVI);
4184  else {
4185  assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
4186 
4187  for (const auto *U : EVI->users())
4188  if (const auto *B = dyn_cast<BranchInst>(U)) {
4189  assert(B->isConditional() && "How else is it using an i1?");
4190  GuardingBranches.push_back(B);
4191  }
4192  }
4193  } else {
4194  // We are using the aggregate directly in a way we don't want to analyze
4195  // here (storing it to a global, say).
4196  return false;
4197  }
4198  }
4199 
4200  auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4201  BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4202  if (!NoWrapEdge.isSingleEdge())
4203  return false;
4204 
4205  // Check if all users of the add are provably no-wrap.
4206  for (const auto *Result : Results) {
4207  // If the extractvalue itself is not executed on overflow, the we don't
4208  // need to check each use separately, since domination is transitive.
4209  if (DT.dominates(NoWrapEdge, Result->getParent()))
4210  continue;
4211 
4212  for (auto &RU : Result->uses())
4213  if (!DT.dominates(NoWrapEdge, RU))
4214  return false;
4215  }
4216 
4217  return true;
4218  };
4219 
4220  return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4221 }
4222 
4223 
4225  const DataLayout &DL,
4226  AssumptionCache *AC,
4227  const Instruction *CxtI,
4228  const DominatorTree *DT) {
4230  Add, DL, AC, CxtI, DT);
4231 }
4232 
4234  const Value *RHS,
4235  const DataLayout &DL,
4236  AssumptionCache *AC,
4237  const Instruction *CxtI,
4238  const DominatorTree *DT) {
4239  return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
4240 }
4241 
4243  // A memory operation returns normally if it isn't volatile. A volatile
4244  // operation is allowed to trap.
4245  //
4246  // An atomic operation isn't guaranteed to return in a reasonable amount of
4247  // time because it's possible for another thread to interfere with it for an
4248  // arbitrary length of time, but programs aren't allowed to rely on that.
4249  if (const LoadInst *LI = dyn_cast<LoadInst>(I))
4250  return !LI->isVolatile();
4251  if (const StoreInst *SI = dyn_cast<StoreInst>(I))
4252  return !SI->isVolatile();
4253  if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
4254  return !CXI->isVolatile();
4255  if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
4256  return !RMWI->isVolatile();
4257  if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
4258  return !MII->isVolatile();
4259 
4260  // If there is no successor, then execution can't transfer to it.
4261  if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
4262  return !CRI->unwindsToCaller();
4263  if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
4264  return !CatchSwitch->unwindsToCaller();
4265  if (isa<ResumeInst>(I))
4266  return false;
4267  if (isa<ReturnInst>(I))
4268  return false;
4269  if (isa<UnreachableInst>(I))
4270  return false;
4271 
4272  // Calls can throw, or contain an infinite loop, or kill the process.
4273  if (auto CS = ImmutableCallSite(I)) {
4274  // Call sites that throw have implicit non-local control flow.
4275  if (!CS.doesNotThrow())
4276  return false;
4277 
4278  // Non-throwing call sites can loop infinitely, call exit/pthread_exit
4279  // etc. and thus not return. However, LLVM already assumes that
4280  //
4281  // - Thread exiting actions are modeled as writes to memory invisible to
4282  // the program.
4283  //
4284  // - Loops that don't have side effects (side effects are volatile/atomic
4285  // stores and IO) always terminate (see http://llvm.org/PR965).
4286  // Furthermore IO itself is also modeled as writes to memory invisible to
4287  // the program.
4288  //
4289  // We rely on those assumptions here, and use the memory effects of the call
4290  // target as a proxy for checking that it always returns.
4291 
4292  // FIXME: This isn't aggressive enough; a call which only writes to a global
4293  // is guaranteed to return.
4294  return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
4295  match(I, m_Intrinsic<Intrinsic::assume>()) ||
4296  match(I, m_Intrinsic<Intrinsic::sideeffect>());
4297  }
4298 
4299  // Other instructions return normally.
4300  return true;
4301 }
4302 
4304  // TODO: This is slightly consdervative for invoke instruction since exiting
4305  // via an exception *is* normal control for them.
4306  for (auto I = BB->begin(), E = BB->end(); I != E; ++I)
4308  return false;
4309  return true;
4310 }
4311 
4313  const Loop *L) {
4314  // The loop header is guaranteed to be executed for every iteration.
4315  //
4316  // FIXME: Relax this constraint to cover all basic blocks that are
4317  // guaranteed to be executed at every iteration.
4318  if (I->getParent() != L->getHeader()) return false;
4319 
4320  for (const Instruction &LI : *L->getHeader()) {
4321  if (&LI == I) return true;
4322  if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
4323  }
4324  llvm_unreachable("Instruction not contained in its own parent basic block.");
4325 }
4326 
4328  switch (I->getOpcode()) {
4329  case Instruction::Add:
4330  case Instruction::Sub:
4331  case Instruction::Xor:
4332  case Instruction::Trunc:
4333  case Instruction::BitCast:
4334  case Instruction::AddrSpaceCast:
4335  case Instruction::Mul:
4336  case Instruction::Shl:
4337  case Instruction::GetElementPtr:
4338  // These operations all propagate poison unconditionally. Note that poison
4339  // is not any particular value, so xor or subtraction of poison with
4340  // itself still yields poison, not zero.
4341  return true;
4342 
4343  case Instruction::AShr:
4344  case Instruction::SExt:
4345  // For these operations, one bit of the input is replicated across
4346  // multiple output bits. A replicated poison bit is still poison.
4347  return true;
4348 
4349  case Instruction::ICmp:
4350  // Comparing poison with any value yields poison. This is why, for
4351  // instance, x s< (x +nsw 1) can be folded to true.
4352  return true;
4353 
4354  default:
4355  return false;
4356  }
4357 }
4358 
4360  switch (I->getOpcode()) {
4361  case Instruction::Store:
4362  return cast<StoreInst>(I)->getPointerOperand();
4363 
4364  case Instruction::Load:
4365  return cast<LoadInst>(I)->getPointerOperand();
4366 
4367  case Instruction::AtomicCmpXchg:
4368  return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
4369 
4370  case Instruction::AtomicRMW:
4371  return cast<AtomicRMWInst>(I)->getPointerOperand();
4372 
4373  case Instruction::UDiv:
4374  case Instruction::SDiv:
4375  case Instruction::URem:
4376  case Instruction::SRem:
4377  return I->getOperand(1);
4378 
4379  default:
4380  return nullptr;
4381  }
4382 }
4383 
4385  // We currently only look for uses of poison values within the same basic
4386  // block, as that makes it easier to guarantee that the uses will be
4387  // executed given that PoisonI is executed.
4388  //
4389  // FIXME: Expand this to consider uses beyond the same basic block. To do
4390  // this, look out for the distinction between post-dominance and strong
4391  // post-dominance.
4392  const BasicBlock *BB = PoisonI->getParent();
4393 
4394  // Set of instructions that we have proved will yield poison if PoisonI
4395  // does.
4396  SmallSet<const Value *, 16> YieldsPoison;
4398  YieldsPoison.insert(PoisonI);
4399  Visited.insert(PoisonI->getParent());
4400 
4401  BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
4402 
4403  unsigned Iter = 0;
4404  while (Iter++ < MaxDepth) {
4405  for (auto &I : make_range(Begin, End)) {
4406  if (&I != PoisonI) {
4407  const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
4408  if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
4409  return true;
4411  return false;
4412  }
4413 
4414  // Mark poison that propagates from I through uses of I.
4415  if (YieldsPoison.count(&I)) {
4416  for (const User *User : I.users()) {
4417  const Instruction *UserI = cast<Instruction>(User);
4418  if (propagatesFullPoison(UserI))
4419  YieldsPoison.insert(User);
4420  }
4421  }
4422  }
4423 
4424  if (auto *NextBB = BB->getSingleSuccessor()) {
4425  if (Visited.insert(NextBB).second) {
4426  BB = NextBB;
4427  Begin = BB->getFirstNonPHI()->getIterator();
4428  End = BB->end();
4429  continue;
4430  }
4431  }
4432 
4433  break;
4434  }
4435  return false;
4436 }
4437 
4438 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
4439  if (FMF.noNaNs())
4440  return true;
4441 
4442  if (auto *C = dyn_cast<ConstantFP>(V))
4443  return !C->isNaN();
4444 
4445  if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4446  if (!C->getElementType()->isFloatingPointTy())
4447  return false;
4448  for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4449  if (C->getElementAsAPFloat(I).isNaN())
4450  return false;
4451  }
4452  return true;
4453  }
4454 
4455  return false;
4456 }
4457 
4458 static bool isKnownNonZero(const Value *V) {
4459  if (auto *C = dyn_cast<ConstantFP>(V))
4460  return !C->isZero();
4461 
4462  if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4463  if (!C->getElementType()->isFloatingPointTy())
4464  return false;
4465  for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4466  if (C->getElementAsAPFloat(I).isZero())
4467  return false;
4468  }
4469  return true;
4470  }
4471 
4472  return false;
4473 }
4474 
4475 /// Match clamp pattern for float types without care about NaNs or signed zeros.
4476 /// Given non-min/max outer cmp/select from the clamp pattern this
4477 /// function recognizes if it can be substitued by a "canonical" min/max
4478 /// pattern.
4480  Value *CmpLHS, Value *CmpRHS,
4481  Value *TrueVal, Value *FalseVal,
4482  Value *&LHS, Value *&RHS) {
4483  // Try to match
4484  // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
4485  // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
4486  // and return description of the outer Max/Min.
4487 
4488  // First, check if select has inverse order:
4489  if (CmpRHS == FalseVal) {
4490  std::swap(TrueVal, FalseVal);
4491  Pred = CmpInst::getInversePredicate(Pred);
4492  }
4493 
4494  // Assume success now. If there's no match, callers should not use these anyway.
4495  LHS = TrueVal;
4496  RHS = FalseVal;
4497 
4498  const APFloat *FC1;
4499  if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
4500  return {SPF_UNKNOWN, SPNB_NA, false};
4501 
4502  const APFloat *FC2;
4503  switch (Pred) {
4504  case CmpInst::FCMP_OLT:
4505  case CmpInst::FCMP_OLE:
4506  case CmpInst::FCMP_ULT:
4507  case CmpInst::FCMP_ULE:
4508  if (match(FalseVal,
4509  m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
4510  m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4511  FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
4512  return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
4513  break;
4514  case CmpInst::FCMP_OGT:
4515  case CmpInst::FCMP_OGE:
4516  case CmpInst::FCMP_UGT:
4517  case CmpInst::FCMP_UGE:
4518  if (match(FalseVal,
4519  m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
4520  m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4521  FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
4522  return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
4523  break;
4524  default:
4525  break;
4526  }
4527 
4528  return {SPF_UNKNOWN, SPNB_NA, false};
4529 }
4530 
4531 /// Recognize variations of:
4532 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
4534  Value *CmpLHS, Value *CmpRHS,
4535  Value *TrueVal, Value *FalseVal) {
4536  // Swap the select operands and predicate to match the patterns below.
4537  if (CmpRHS != TrueVal) {
4538  Pred = ICmpInst::getSwappedPredicate(Pred);
4539  std::swap(TrueVal, FalseVal);
4540  }
4541  const APInt *C1;
4542  if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
4543  const APInt *C2;
4544  // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
4545  if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4546  C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
4547  return {SPF_SMAX, SPNB_NA, false};
4548 
4549  // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
4550  if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4551  C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
4552  return {SPF_SMIN, SPNB_NA, false};
4553 
4554  // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
4555  if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4556  C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
4557  return {SPF_UMAX, SPNB_NA, false};
4558 
4559  // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
4560  if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4561  C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
4562  return {SPF_UMIN, SPNB_NA, false};
4563  }
4564  return {SPF_UNKNOWN, SPNB_NA, false};
4565 }
4566 
4567 /// Recognize variations of:
4568 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
4570  Value *CmpLHS, Value *CmpRHS,
4571  Value *TVal, Value *FVal,
4572  unsigned Depth) {
4573  // TODO: Allow FP min/max with nnan/nsz.
4574  assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
4575 
4576  Value *A, *B;
4577  SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
4579  return {SPF_UNKNOWN, SPNB_NA, false};
4580 
4581  Value *C, *D;
4582  SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
4583  if (L.Flavor != R.Flavor)
4584  return {SPF_UNKNOWN, SPNB_NA, false};
4585 
4586  // We have something like: x Pred y ? min(a, b) : min(c, d).
4587  // Try to match the compare to the min/max operations of the select operands.
4588  // First, make sure we have the right compare predicate.
4589  switch (L.Flavor) {
4590  case SPF_SMIN:
4591  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
4592  Pred = ICmpInst::getSwappedPredicate(Pred);
4593  std::swap(CmpLHS, CmpRHS);
4594  }
4595  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
4596  break;
4597  return {SPF_UNKNOWN, SPNB_NA, false};
4598  case SPF_SMAX:
4599  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
4600  Pred = ICmpInst::getSwappedPredicate(Pred);
4601  std::swap(CmpLHS, CmpRHS);
4602  }
4603  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
4604  break;
4605  return {SPF_UNKNOWN, SPNB_NA, false};
4606  case SPF_UMIN:
4607  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
4608  Pred = ICmpInst::getSwappedPredicate(Pred);
4609  std::swap(CmpLHS, CmpRHS);
4610  }
4611  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
4612  break;
4613  return {SPF_UNKNOWN, SPNB_NA, false};
4614  case SPF_UMAX:
4615  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
4616  Pred = ICmpInst::getSwappedPredicate(Pred);
4617  std::swap(CmpLHS, CmpRHS);
4618  }
4619  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
4620  break;
4621  return {SPF_UNKNOWN, SPNB_NA, false};
4622  default:
4623  return {SPF_UNKNOWN, SPNB_NA, false};
4624  }
4625 
4626  // If there is a common operand in the already matched min/max and the other
4627  // min/max operands match the compare operands (either directly or inverted),
4628  // then this is min/max of the same flavor.
4629 
4630  // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4631  // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4632  if (D == B) {
4633  if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4634  match(A, m_Not(m_Specific(CmpRHS)))))
4635  return {L.Flavor, SPNB_NA, false};
4636  }
4637  // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4638  // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4639  if (C == B) {
4640  if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4641  match(A, m_Not(m_Specific(CmpRHS)))))
4642  return {L.Flavor, SPNB_NA, false};
4643  }
4644  // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4645  // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4646  if (D == A) {
4647  if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4648  match(B, m_Not(m_Specific(CmpRHS)))))
4649  return {L.Flavor, SPNB_NA, false};
4650  }
4651  // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4652  // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4653  if (C == A) {
4654  if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4655  match(B, m_Not(m_Specific(CmpRHS)))))
4656  return {L.Flavor, SPNB_NA, false};
4657  }
4658 
4659  return {SPF_UNKNOWN, SPNB_NA, false};
4660 }
4661 
4662 /// Match non-obvious integer minimum and maximum sequences.
4664  Value *CmpLHS, Value *CmpRHS,
4665  Value *TrueVal, Value *FalseVal,
4666  Value *&LHS, Value *&RHS,
4667  unsigned Depth) {
4668  // Assume success. If there's no match, callers should not use these anyway.
4669  LHS = TrueVal;
4670  RHS = FalseVal;
4671 
4672  SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
4674  return SPR;
4675 
4676  SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
4678  return SPR;
4679 
4680  if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
4681  return {SPF_UNKNOWN, SPNB_NA, false};
4682 
4683  // Z = X -nsw Y
4684  // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
4685  // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
4686  if (match(TrueVal, m_Zero()) &&
4687  match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4688  return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4689 
4690  // Z = X -nsw Y
4691  // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
4692  // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
4693  if (match(FalseVal, m_Zero()) &&
4694  match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4695  return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4696 
4697  const APInt *C1;
4698  if (!match(CmpRHS, m_APInt(C1)))
4699  return {SPF_UNKNOWN, SPNB_NA, false};
4700 
4701  // An unsigned min/max can be written with a signed compare.
4702  const APInt *C2;
4703  if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
4704  (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
4705  // Is the sign bit set?
4706  // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
4707  // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
4708  if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
4709  C2->isMaxSignedValue())
4710  return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4711 
4712  // Is the sign bit clear?
4713  // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
4714  // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
4715  if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
4716  C2->isMinSignedValue())
4717  return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4718  }
4719 
4720  // Look through 'not' ops to find disguised signed min/max.
4721  // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4722  // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4723  if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4724  match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4725  return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4726 
4727  // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4728  // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4729  if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4730  match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4731  return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4732 
4733  return {SPF_UNKNOWN, SPNB_NA, false};
4734 }
4735 
4736 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
4737  assert(X && Y && "Invalid operand");
4738 
4739  // X = sub (0, Y) || X = sub nsw (0, Y)
4740  if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
4741  (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
4742  return true;
4743 
4744  // Y = sub (0, X) || Y = sub nsw (0, X)
4745  if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
4746  (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
4747  return true;
4748 
4749  // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
4750  Value *A, *B;
4751  return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
4752  match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
4753  (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
4754  match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
4755 }
4756 
4758  FastMathFlags FMF,
4759  Value *CmpLHS, Value *CmpRHS,
4760  Value *TrueVal, Value *FalseVal,
4761  Value *&LHS, Value *&RHS,
4762  unsigned Depth) {
4763  if (CmpInst::isFPPredicate(Pred)) {
4764  // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one
4765  // 0.0 operand, set the compare's 0.0 operands to that same value for the
4766  // purpose of identifying min/max. Disregard vector constants with undefined
4767  // elements because those can not be back-propagated for analysis.
4768  Value *OutputZeroVal = nullptr;
4769  if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) &&
4770  !cast<Constant>(TrueVal)->containsUndefElement())
4771  OutputZeroVal = TrueVal;
4772  else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) &&
4773  !cast<Constant>(FalseVal)->containsUndefElement())
4774  OutputZeroVal = FalseVal;
4775 
4776  if (OutputZeroVal) {
4777  if (match(CmpLHS, m_AnyZeroFP()))
4778  CmpLHS = OutputZeroVal;
4779  if (match(CmpRHS, m_AnyZeroFP()))
4780  CmpRHS = OutputZeroVal;
4781  }
4782  }
4783 
4784  LHS = CmpLHS;
4785  RHS = CmpRHS;
4786 
4787  // Signed zero may return inconsistent results between implementations.
4788  // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4789  // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4790  // Therefore, we behave conservatively and only proceed if at least one of the
4791  // operands is known to not be zero or if we don't care about signed zero.
4792  switch (Pred) {
4793  default: break;
4794  // FIXME: Include OGT/OLT/UGT/ULT.
4797  if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4798  !isKnownNonZero(CmpRHS))
4799  return {SPF_UNKNOWN, SPNB_NA, false};
4800  }
4801 
4802  SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4803  bool Ordered = false;
4804 
4805  // When given one NaN and one non-NaN input:
4806  // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4807  // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4808  // ordered comparison fails), which could be NaN or non-NaN.
4809  // so here we discover exactly what NaN behavior is required/accepted.
4810  if (CmpInst::isFPPredicate(Pred)) {
4811  bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4812  bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4813 
4814  if (LHSSafe && RHSSafe) {
4815  // Both operands are known non-NaN.
4816  NaNBehavior = SPNB_RETURNS_ANY;
4817  } else if (CmpInst::isOrdered(Pred)) {
4818  // An ordered comparison will return false when given a NaN, so it
4819  // returns the RHS.
4820  Ordered = true;
4821  if (LHSSafe)
4822  // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4823  NaNBehavior = SPNB_RETURNS_NAN;
4824  else if (RHSSafe)
4825  NaNBehavior = SPNB_RETURNS_OTHER;
4826  else
4827  // Completely unsafe.
4828  return {SPF_UNKNOWN, SPNB_NA, false};
4829  } else {
4830  Ordered = false;
4831  // An unordered comparison will return true when given a NaN, so it
4832  // returns the LHS.
4833  if (LHSSafe)
4834  // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4835  NaNBehavior = SPNB_RETURNS_OTHER;
4836  else if (RHSSafe)
4837  NaNBehavior = SPNB_RETURNS_NAN;
4838  else
4839  // Completely unsafe.
4840  return {SPF_UNKNOWN, SPNB_NA, false};
4841  }
4842  }
4843 
4844  if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4845  std::swap(CmpLHS, CmpRHS);
4846  Pred = CmpInst::getSwappedPredicate(Pred);
4847  if (NaNBehavior == SPNB_RETURNS_NAN)
4848  NaNBehavior = SPNB_RETURNS_OTHER;
4849  else if (NaNBehavior == SPNB_RETURNS_OTHER)
4850  NaNBehavior = SPNB_RETURNS_NAN;
4851  Ordered = !Ordered;
4852  }
4853 
4854  // ([if]cmp X, Y) ? X : Y
4855  if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4856  switch (Pred) {
4857  default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4858  case ICmpInst::ICMP_UGT:
4859  case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4860  case ICmpInst::ICMP_SGT:
4861  case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4862  case ICmpInst::ICMP_ULT:
4863  case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4864  case ICmpInst::ICMP_SLT:
4865  case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4866  case FCmpInst::FCMP_UGT:
4867  case FCmpInst::FCMP_UGE:
4868  case FCmpInst::FCMP_OGT:
4869  case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4870  case FCmpInst::FCMP_ULT:
4871  case FCmpInst::FCMP_ULE:
4872  case FCmpInst::FCMP_OLT:
4873  case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4874  }
4875  }
4876 
4877  if (isKnownNegation(TrueVal, FalseVal)) {
4878  // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
4879  // match against either LHS or sext(LHS).
4880  auto MaybeSExtCmpLHS =
4881  m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
4882  auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
4883  auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
4884  if (match(TrueVal, MaybeSExtCmpLHS)) {
4885  // Set the return values. If the compare uses the negated value (-X >s 0),
4886  // swap the return values because the negated value is always 'RHS'.
4887  LHS = TrueVal;
4888  RHS = FalseVal;
4889  if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
4890  std::swap(LHS, RHS);
4891 
4892  // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
4893  // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
4894  if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4895  return {SPF_ABS, SPNB_NA, false};
4896 
4897  // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
4898  // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
4899  if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4900  return {SPF_NABS, SPNB_NA, false};
4901  }
4902  else if (match(FalseVal, MaybeSExtCmpLHS)) {
4903  // Set the return values. If the compare uses the negated value (-X >s 0),
4904  // swap the return values because the negated value is always 'RHS'.
4905  LHS = FalseVal;
4906  RHS = TrueVal;
4907  if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
4908  std::swap(LHS, RHS);
4909 
4910  // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
4911  // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
4912  if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4913  return {SPF_NABS, SPNB_NA, false};
4914 
4915  // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
4916  // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
4917  if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4918  return {SPF_ABS, SPNB_NA, false};
4919  }
4920  }
4921 
4922  if (CmpInst::isIntPredicate(Pred))
4923  return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
4924 
4925  // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
4926  // may return either -0.0 or 0.0, so fcmp/select pair has stricter
4927  // semantics than minNum. Be conservative in such case.
4928  if (NaNBehavior != SPNB_RETURNS_ANY ||
4929  (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4930  !isKnownNonZero(CmpRHS)))
4931  return {SPF_UNKNOWN, SPNB_NA, false};
4932 
4933  return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4934 }
4935 
4936 /// Helps to match a select pattern in case of a type mismatch.
4937 ///
4938 /// The function processes the case when type of true and false values of a
4939 /// select instruction differs from type of the cmp instruction operands because
4940 /// of a cast instruction. The function checks if it is legal to move the cast
4941 /// operation after "select". If yes, it returns the new second value of
4942 /// "select" (with the assumption that cast is moved):
4943 /// 1. As operand of cast instruction when both values of "select" are same cast
4944 /// instructions.
4945 /// 2. As restored constant (by applying reverse cast operation) when the first
4946 /// value of the "select" is a cast operation and the second value is a
4947 /// constant.
4948 /// NOTE: We return only the new second value because the first value could be
4949 /// accessed as operand of cast instruction.
4950 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4951  Instruction::CastOps *CastOp) {
4952  auto *Cast1 = dyn_cast<CastInst>(V1);
4953  if (!Cast1)
4954  return nullptr;
4955 
4956  *CastOp = Cast1->getOpcode();
4957  Type *SrcTy = Cast1->getSrcTy();
4958  if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4959  // If V1 and V2 are both the same cast from the same type, look through V1.
4960  if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4961  return Cast2->getOperand(0);
4962  return nullptr;
4963  }
4964 
4965  auto *C = dyn_cast<Constant>(V2);
4966  if (!C)
4967  return nullptr;
4968 
4969  Constant *CastedTo = nullptr;
4970  switch (*CastOp) {
4971  case Instruction::ZExt:
4972  if (CmpI->isUnsigned())
4973  CastedTo = ConstantExpr::getTrunc(C, SrcTy);
4974  break;
4975  case Instruction::SExt:
4976  if (CmpI->isSigned())
4977  CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
4978  break;
4979  case Instruction::Trunc:
4980  Constant *CmpConst;
4981  if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
4982  CmpConst->getType() == SrcTy) {
4983  // Here we have the following case:
4984  //
4985  // %cond = cmp iN %x, CmpConst
4986  // %tr = trunc iN %x to iK
4987  // %narrowsel = select i1 %cond, iK %t, iK C
4988  //
4989  // We can always move trunc after select operation:
4990  //
4991  // %cond = cmp iN %x, CmpConst
4992  // %widesel = select i1 %cond, iN %x, iN CmpConst
4993  // %tr = trunc iN %widesel to iK
4994  //
4995  // Note that C could be extended in any way because we don't care about
4996  // upper bits after truncation. It can't be abs pattern, because it would
4997  // look like:
4998  //
4999  // select i1 %cond, x, -x.
5000  //
5001  // So only min/max pattern could be matched. Such match requires widened C
5002  // == CmpConst. That is why set widened C = CmpConst, condition trunc
5003  // CmpConst == C is checked below.
5004  CastedTo = CmpConst;
5005  } else {
5006  CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
5007  }
5008  break;
5009  case Instruction::FPTrunc:
5010  CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
5011  break;
5012  case Instruction::FPExt:
5013  CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
5014  break;
5015  case Instruction::FPToUI:
5016  CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
5017  break;
5018  case Instruction::FPToSI:
5019  CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
5020  break;
5021  case Instruction::UIToFP:
5022  CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
5023  break;
5024  case Instruction::SIToFP:
5025  CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
5026  break;
5027  default:
5028  break;
5029  }
5030 
5031  if (!CastedTo)
5032  return nullptr;
5033 
5034  // Make sure the cast doesn't lose any information.
5035  Constant *CastedBack =
5036  ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
5037  if (CastedBack != C)
5038  return nullptr;
5039 
5040  return CastedTo;
5041 }
5042 
5044  Instruction::CastOps *CastOp,
5045  unsigned Depth) {
5046  if (Depth >= MaxDepth)
5047  return {SPF_UNKNOWN, SPNB_NA, false};
5048 
5050  if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
5051 
5052  CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
5053  if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
5054 
5055  CmpInst::Predicate Pred = CmpI->getPredicate();
5056  Value *CmpLHS = CmpI->getOperand(0);
5057  Value *CmpRHS = CmpI->getOperand(1);
5058  Value *TrueVal = SI->getTrueValue();
5059  Value *FalseVal = SI->getFalseValue();
5060  FastMathFlags FMF;
5061  if (isa<FPMathOperator>(CmpI))
5062  FMF = CmpI->getFastMathFlags();
5063 
5064  // Bail out early.
5065  if (CmpI->isEquality())
5066  return {SPF_UNKNOWN, SPNB_NA, false};
5067 
5068  // Deal with type mismatches.
5069  if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
5070  if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
5071  // If this is a potential fmin/fmax with a cast to integer, then ignore
5072  // -0.0 because there is no corresponding integer value.
5073  if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5074  FMF.setNoSignedZeros();
5075  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5076  cast<CastInst>(TrueVal)->getOperand(0), C,
5077  LHS, RHS, Depth);
5078  }
5079  if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
5080  // If this is a potential fmin/fmax with a cast to integer, then ignore
5081  // -0.0 because there is no corresponding integer value.
5082  if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5083  FMF.setNoSignedZeros();
5084  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5085  C, cast<CastInst>(FalseVal)->getOperand(0),
5086  LHS, RHS, Depth);
5087  }
5088  }
5089  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
5090  LHS, RHS, Depth);
5091 }
5092 
5094  if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
5095  if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
5096  if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
5097  if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
5098  if (SPF == SPF_FMINNUM)
5099  return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
5100  if (SPF == SPF_FMAXNUM)
5101  return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
5102  llvm_unreachable("unhandled!");
5103 }
5104 
5106  if (SPF == SPF_SMIN) return SPF_SMAX;
5107  if (SPF == SPF_UMIN) return SPF_UMAX;
5108  if (SPF == SPF_SMAX) return SPF_SMIN;
5109  if (SPF == SPF_UMAX) return SPF_UMIN;
5110  llvm_unreachable("unhandled!");
5111 }
5112 
5115 }
5116 
5117 /// Return true if "icmp Pred LHS RHS" is always true.
5118 static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS,
5119  const Value *RHS, const DataLayout &DL,
5120  unsigned Depth) {
5121  assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
5122  if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
5123  return true;
5124 
5125  switch (Pred) {
5126  default:
5127  return false;
5128 
5129  case CmpInst::ICMP_SLE: {
5130  const APInt *C;
5131 
5132  // LHS s<= LHS +_{nsw} C if C >= 0
5133  if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
5134  return !C->isNegative();
5135  return false;
5136  }
5137 
5138  case CmpInst::ICMP_ULE: {
5139  const APInt *C;
5140 
5141  // LHS u<= LHS +_{nuw} C for any C
5142  if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
5143  return true;
5144 
5145  // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
5146  auto MatchNUWAddsToSameValue = [&](const Value *A, const