LLVM  9.0.0svn
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::fshr:
1510  case Intrinsic::fshl: {
1511  const APInt *SA;
1512  if (!match(I->getOperand(2), m_APInt(SA)))
1513  break;
1514 
1515  // Normalize to funnel shift left.
1516  uint64_t ShiftAmt = SA->urem(BitWidth);
1517  if (II->getIntrinsicID() == Intrinsic::fshr)
1518  ShiftAmt = BitWidth - ShiftAmt;
1519 
1520  KnownBits Known3(Known);
1521  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1522  computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q);
1523 
1524  Known.Zero =
1525  Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt);
1526  Known.One =
1527  Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt);
1528  break;
1529  }
1530  case Intrinsic::x86_sse42_crc32_64_64:
1531  Known.Zero.setBitsFrom(32);
1532  break;
1533  }
1534  }
1535  break;
1536  case Instruction::ExtractElement:
1537  // Look through extract element. At the moment we keep this simple and skip
1538  // tracking the specific element. But at least we might find information
1539  // valid for all elements of the vector (for example if vector is sign
1540  // extended, shifted, etc).
1541  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1542  break;
1543  case Instruction::ExtractValue:
1544  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1545  const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1546  if (EVI->getNumIndices() != 1) break;
1547  if (EVI->getIndices()[0] == 0) {
1548  switch (II->getIntrinsicID()) {
1549  default: break;
1550  case Intrinsic::uadd_with_overflow:
1551  case Intrinsic::sadd_with_overflow:
1552  computeKnownBitsAddSub(true, II->getArgOperand(0),
1553  II->getArgOperand(1), false, Known, Known2,
1554  Depth, Q);
1555  break;
1556  case Intrinsic::usub_with_overflow:
1557  case Intrinsic::ssub_with_overflow:
1558  computeKnownBitsAddSub(false, II->getArgOperand(0),
1559  II->getArgOperand(1), false, Known, Known2,
1560  Depth, Q);
1561  break;
1562  case Intrinsic::umul_with_overflow:
1563  case Intrinsic::smul_with_overflow:
1564  computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1565  Known, Known2, Depth, Q);
1566  break;
1567  }
1568  }
1569  }
1570  }
1571 }
1572 
1573 /// Determine which bits of V are known to be either zero or one and return
1574 /// them.
1575 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1576  KnownBits Known(getBitWidth(V->getType(), Q.DL));
1577  computeKnownBits(V, Known, Depth, Q);
1578  return Known;
1579 }
1580 
1581 /// Determine which bits of V are known to be either zero or one and return
1582 /// them in the Known bit set.
1583 ///
1584 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1585 /// we cannot optimize based on the assumption that it is zero without changing
1586 /// it to be an explicit zero. If we don't change it to zero, other code could
1587 /// optimized based on the contradictory assumption that it is non-zero.
1588 /// Because instcombine aggressively folds operations with undef args anyway,
1589 /// this won't lose us code quality.
1590 ///
1591 /// This function is defined on values with integer type, values with pointer
1592 /// type, and vectors of integers. In the case
1593 /// where V is a vector, known zero, and known one values are the
1594 /// same width as the vector element, and the bit is set only if it is true
1595 /// for all of the elements in the vector.
1596 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1597  const Query &Q) {
1598  assert(V && "No Value?");
1599  assert(Depth <= MaxDepth && "Limit Search Depth");
1600  unsigned BitWidth = Known.getBitWidth();
1601 
1602  assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||
1603  V->getType()->isPtrOrPtrVectorTy()) &&
1604  "Not integer or pointer type!");
1605 
1606  Type *ScalarTy = V->getType()->getScalarType();
1607  unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
1608  Q.DL.getIndexTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
1609  assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth");
1610  (void)BitWidth;
1611  (void)ExpectedWidth;
1612 
1613  const APInt *C;
1614  if (match(V, m_APInt(C))) {
1615  // We know all of the bits for a scalar constant or a splat vector constant!
1616  Known.One = *C;
1617  Known.Zero = ~Known.One;
1618  return;
1619  }
1620  // Null and aggregate-zero are all-zeros.
1621  if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1622  Known.setAllZero();
1623  return;
1624  }
1625  // Handle a constant vector by taking the intersection of the known bits of
1626  // each element.
1627  if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1628  // We know that CDS must be a vector of integers. Take the intersection of
1629  // each element.
1630  Known.Zero.setAllBits(); Known.One.setAllBits();
1631  for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1632  APInt Elt = CDS->getElementAsAPInt(i);
1633  Known.Zero &= ~Elt;
1634  Known.One &= Elt;
1635  }
1636  return;
1637  }
1638 
1639  if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1640  // We know that CV must be a vector of integers. Take the intersection of
1641  // each element.
1642  Known.Zero.setAllBits(); Known.One.setAllBits();
1643  for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1644  Constant *Element = CV->getAggregateElement(i);
1645  auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1646  if (!ElementCI) {
1647  Known.resetAll();
1648  return;
1649  }
1650  const APInt &Elt = ElementCI->getValue();
1651  Known.Zero &= ~Elt;
1652  Known.One &= Elt;
1653  }
1654  return;
1655  }
1656 
1657  // Start out not knowing anything.
1658  Known.resetAll();
1659 
1660  // We can't imply anything about undefs.
1661  if (isa<UndefValue>(V))
1662  return;
1663 
1664  // There's no point in looking through other users of ConstantData for
1665  // assumptions. Confirm that we've handled them all.
1666  assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1667 
1668  // Limit search depth.
1669  // All recursive calls that increase depth must come after this.
1670  if (Depth == MaxDepth)
1671  return;
1672 
1673  // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1674  // the bits of its aliasee.
1675  if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1676  if (!GA->isInterposable())
1677  computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1678  return;
1679  }
1680 
1681  if (const Operator *I = dyn_cast<Operator>(V))
1682  computeKnownBitsFromOperator(I, Known, Depth, Q);
1683 
1684  // Aligned pointers have trailing zeros - refine Known.Zero set
1685  if (V->getType()->isPointerTy()) {
1686  unsigned Align = V->getPointerAlignment(Q.DL);
1687  if (Align)
1688  Known.Zero.setLowBits(countTrailingZeros(Align));
1689  }
1690 
1691  // computeKnownBitsFromAssume strictly refines Known.
1692  // Therefore, we run them after computeKnownBitsFromOperator.
1693 
1694  // Check whether a nearby assume intrinsic can determine some known bits.
1695  computeKnownBitsFromAssume(V, Known, Depth, Q);
1696 
1697  assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1698 }
1699 
1700 /// Return true if the given value is known to have exactly one
1701 /// bit set when defined. For vectors return true if every element is known to
1702 /// be a power of two when defined. Supports values with integer or pointer
1703 /// types and vectors of integers.
1704 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1705  const Query &Q) {
1706  assert(Depth <= MaxDepth && "Limit Search Depth");
1707 
1708  // Attempt to match against constants.
1709  if (OrZero && match(V, m_Power2OrZero()))
1710  return true;
1711  if (match(V, m_Power2()))
1712  return true;
1713 
1714  // 1 << X is clearly a power of two if the one is not shifted off the end. If
1715  // it is shifted off the end then the result is undefined.
1716  if (match(V, m_Shl(m_One(), m_Value())))
1717  return true;
1718 
1719  // (signmask) >>l X is clearly a power of two if the one is not shifted off
1720  // the bottom. If it is shifted off the bottom then the result is undefined.
1721  if (match(V, m_LShr(m_SignMask(), m_Value())))
1722  return true;
1723 
1724  // The remaining tests are all recursive, so bail out if we hit the limit.
1725  if (Depth++ == MaxDepth)
1726  return false;
1727 
1728  Value *X = nullptr, *Y = nullptr;
1729  // A shift left or a logical shift right of a power of two is a power of two
1730  // or zero.
1731  if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1732  match(V, m_LShr(m_Value(X), m_Value()))))
1733  return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1734 
1735  if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1736  return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1737 
1738  if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1739  return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1740  isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1741 
1742  if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1743  // A power of two and'd with anything is a power of two or zero.
1744  if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1745  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1746  return true;
1747  // X & (-X) is always a power of two or zero.
1748  if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1749  return true;
1750  return false;
1751  }
1752 
1753  // Adding a power-of-two or zero to the same power-of-two or zero yields
1754  // either the original power-of-two, a larger power-of-two or zero.
1755  if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1756  const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1757  if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) ||
1758  Q.IIQ.hasNoSignedWrap(VOBO)) {
1759  if (match(X, m_And(m_Specific(Y), m_Value())) ||
1760  match(X, m_And(m_Value(), m_Specific(Y))))
1761  if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1762  return true;
1763  if (match(Y, m_And(m_Specific(X), m_Value())) ||
1764  match(Y, m_And(m_Value(), m_Specific(X))))
1765  if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1766  return true;
1767 
1768  unsigned BitWidth = V->getType()->getScalarSizeInBits();
1769  KnownBits LHSBits(BitWidth);
1770  computeKnownBits(X, LHSBits, Depth, Q);
1771 
1772  KnownBits RHSBits(BitWidth);
1773  computeKnownBits(Y, RHSBits, Depth, Q);
1774  // If i8 V is a power of two or zero:
1775  // ZeroBits: 1 1 1 0 1 1 1 1
1776  // ~ZeroBits: 0 0 0 1 0 0 0 0
1777  if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1778  // If OrZero isn't set, we cannot give back a zero result.
1779  // Make sure either the LHS or RHS has a bit set.
1780  if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1781  return true;
1782  }
1783  }
1784 
1785  // An exact divide or right shift can only shift off zero bits, so the result
1786  // is a power of two only if the first operand is a power of two and not
1787  // copying a sign bit (sdiv int_min, 2).
1788  if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1789  match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1790  return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1791  Depth, Q);
1792  }
1793 
1794  return false;
1795 }
1796 
1797 /// Test whether a GEP's result is known to be non-null.
1798 ///
1799 /// Uses properties inherent in a GEP to try to determine whether it is known
1800 /// to be non-null.
1801 ///
1802 /// Currently this routine does not support vector GEPs.
1803 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1804  const Query &Q) {
1805  const Function *F = nullptr;
1806  if (const Instruction *I = dyn_cast<Instruction>(GEP))
1807  F = I->getFunction();
1808 
1809  if (!GEP->isInBounds() ||
1811  return false;
1812 
1813  // FIXME: Support vector-GEPs.
1814  assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1815 
1816  // If the base pointer is non-null, we cannot walk to a null address with an
1817  // inbounds GEP in address space zero.
1818  if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1819  return true;
1820 
1821  // Walk the GEP operands and see if any operand introduces a non-zero offset.
1822  // If so, then the GEP cannot produce a null pointer, as doing so would
1823  // inherently violate the inbounds contract within address space zero.
1824  for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1825  GTI != GTE; ++GTI) {
1826  // Struct types are easy -- they must always be indexed by a constant.
1827  if (StructType *STy = GTI.getStructTypeOrNull()) {
1828  ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1829  unsigned ElementIdx = OpC->getZExtValue();
1830  const StructLayout *SL = Q.DL.getStructLayout(STy);
1831  uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1832  if (ElementOffset > 0)
1833  return true;
1834  continue;
1835  }
1836 
1837  // If we have a zero-sized type, the index doesn't matter. Keep looping.
1838  if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1839  continue;
1840 
1841  // Fast path the constant operand case both for efficiency and so we don't
1842  // increment Depth when just zipping down an all-constant GEP.
1843  if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1844  if (!OpC->isZero())
1845  return true;
1846  continue;
1847  }
1848 
1849  // We post-increment Depth here because while isKnownNonZero increments it
1850  // as well, when we pop back up that increment won't persist. We don't want
1851  // to recurse 10k times just because we have 10k GEP operands. We don't
1852  // bail completely out because we want to handle constant GEPs regardless
1853  // of depth.
1854  if (Depth++ >= MaxDepth)
1855  continue;
1856 
1857  if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1858  return true;
1859  }
1860 
1861  return false;
1862 }
1863 
1865  const Instruction *CtxI,
1866  const DominatorTree *DT) {
1867  assert(V->getType()->isPointerTy() && "V must be pointer type");
1868  assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
1869 
1870  if (!CtxI || !DT)
1871  return false;
1872 
1873  unsigned NumUsesExplored = 0;
1874  for (auto *U : V->users()) {
1875  // Avoid massive lists
1876  if (NumUsesExplored >= DomConditionsMaxUses)
1877  break;
1878  NumUsesExplored++;
1879 
1880  // If the value is used as an argument to a call or invoke, then argument
1881  // attributes may provide an answer about null-ness.
1882  if (auto CS = ImmutableCallSite(U))
1883  if (auto *CalledFunc = CS.getCalledFunction())
1884  for (const Argument &Arg : CalledFunc->args())
1885  if (CS.getArgOperand(Arg.getArgNo()) == V &&
1886  Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
1887  return true;
1888 
1889  // Consider only compare instructions uniquely controlling a branch
1890  CmpInst::Predicate Pred;
1891  if (!match(const_cast<User *>(U),
1892  m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
1893  (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
1894  continue;
1895 
1898  for (auto *CmpU : U->users()) {
1899  assert(WorkList.empty() && "Should be!");
1900  if (Visited.insert(CmpU).second)
1901  WorkList.push_back(CmpU);
1902 
1903  while (!WorkList.empty()) {
1904  auto *Curr = WorkList.pop_back_val();
1905 
1906  // If a user is an AND, add all its users to the work list. We only
1907  // propagate "pred != null" condition through AND because it is only
1908  // correct to assume that all conditions of AND are met in true branch.
1909  // TODO: Support similar logic of OR and EQ predicate?
1910  if (Pred == ICmpInst::ICMP_NE)
1911  if (auto *BO = dyn_cast<BinaryOperator>(Curr))
1912  if (BO->getOpcode() == Instruction::And) {
1913  for (auto *BOU : BO->users())
1914  if (Visited.insert(BOU).second)
1915  WorkList.push_back(BOU);
1916  continue;
1917  }
1918 
1919  if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) {
1920  assert(BI->isConditional() && "uses a comparison!");
1921 
1922  BasicBlock *NonNullSuccessor =
1923  BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
1924  BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
1925  if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
1926  return true;
1927  } else if (Pred == ICmpInst::ICMP_NE && isGuard(Curr) &&
1928  DT->dominates(cast<Instruction>(Curr), CtxI)) {
1929  return true;
1930  }
1931  }
1932  }
1933  }
1934 
1935  return false;
1936 }
1937 
1938 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1939 /// ensure that the value it's attached to is never Value? 'RangeType' is
1940 /// is the type of the value described by the range.
1941 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1942  const unsigned NumRanges = Ranges->getNumOperands() / 2;
1943  assert(NumRanges >= 1);
1944  for (unsigned i = 0; i < NumRanges; ++i) {
1945  ConstantInt *Lower =
1946  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1947  ConstantInt *Upper =
1948  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1949  ConstantRange Range(Lower->getValue(), Upper->getValue());
1950  if (Range.contains(Value))
1951  return false;
1952  }
1953  return true;
1954 }
1955 
1956 /// Return true if the given value is known to be non-zero when defined. For
1957 /// vectors, return true if every element is known to be non-zero when
1958 /// defined. For pointers, if the context instruction and dominator tree are
1959 /// specified, perform context-sensitive analysis and return true if the
1960 /// pointer couldn't possibly be null at the specified instruction.
1961 /// Supports values with integer or pointer type and vectors of integers.
1962 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1963  if (auto *C = dyn_cast<Constant>(V)) {
1964  if (C->isNullValue())
1965  return false;
1966  if (isa<ConstantInt>(C))
1967  // Must be non-zero due to null test above.
1968  return true;
1969 
1970  // For constant vectors, check that all elements are undefined or known
1971  // non-zero to determine that the whole vector is known non-zero.
1972  if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1973  for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1974  Constant *Elt = C->getAggregateElement(i);
1975  if (!Elt || Elt->isNullValue())
1976  return false;
1977  if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1978  return false;
1979  }
1980  return true;
1981  }
1982 
1983  // A global variable in address space 0 is non null unless extern weak
1984  // or an absolute symbol reference. Other address spaces may have null as a
1985  // valid address for a global, so we can't assume anything.
1986  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
1987  if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
1988  GV->getType()->getAddressSpace() == 0)
1989  return true;
1990  } else
1991  return false;
1992  }
1993 
1994  if (auto *I = dyn_cast<Instruction>(V)) {
1995  if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) {
1996  // If the possible ranges don't contain zero, then the value is
1997  // definitely non-zero.
1998  if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1999  const APInt ZeroValue(Ty->getBitWidth(), 0);
2000  if (rangeMetadataExcludesValue(Ranges, ZeroValue))
2001  return true;
2002  }
2003  }
2004  }
2005 
2006  // Some of the tests below are recursive, so bail out if we hit the limit.
2007  if (Depth++ >= MaxDepth)
2008  return false;
2009 
2010  // Check for pointer simplifications.
2011  if (V->getType()->isPointerTy()) {
2012  // Alloca never returns null, malloc might.
2013  if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
2014  return true;
2015 
2016  // A byval, inalloca, or nonnull argument is never null.
2017  if (const Argument *A = dyn_cast<Argument>(V))
2018  if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
2019  return true;
2020 
2021  // A Load tagged with nonnull metadata is never null.
2022  if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2023  if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull))
2024  return true;
2025 
2026  if (const auto *Call = dyn_cast<CallBase>(V)) {
2027  if (Call->isReturnNonNull())
2028  return true;
2029  if (const auto *RP = getArgumentAliasingToReturnedPointer(Call))
2030  return isKnownNonZero(RP, Depth, Q);
2031  }
2032  }
2033 
2034 
2035  // Check for recursive pointer simplifications.
2036  if (V->getType()->isPointerTy()) {
2037  if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
2038  return true;
2039 
2040  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
2041  if (isGEPKnownNonNull(GEP, Depth, Q))
2042  return true;
2043  }
2044 
2045  unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
2046 
2047  // X | Y != 0 if X != 0 or Y != 0.
2048  Value *X = nullptr, *Y = nullptr;
2049  if (match(V, m_Or(m_Value(X), m_Value(Y))))
2050  return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
2051 
2052  // ext X != 0 if X != 0.
2053  if (isa<SExtInst>(V) || isa<ZExtInst>(V))
2054  return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2055 
2056  // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2057  // if the lowest bit is shifted off the end.
2058  if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2059  // shl nuw can't remove any non-zero bits.
2060  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2061  if (Q.IIQ.hasNoUnsignedWrap(BO))
2062  return isKnownNonZero(X, Depth, Q);
2063 
2064  KnownBits Known(BitWidth);
2065  computeKnownBits(X, Known, Depth, Q);
2066  if (Known.One[0])
2067  return true;
2068  }
2069  // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2070  // defined if the sign bit is shifted off the end.
2071  else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2072  // shr exact can only shift out zero bits.
2073  const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2074  if (BO->isExact())
2075  return isKnownNonZero(X, Depth, Q);
2076 
2077  KnownBits Known = computeKnownBits(X, Depth, Q);
2078  if (Known.isNegative())
2079  return true;
2080 
2081  // If the shifter operand is a constant, and all of the bits shifted
2082  // out are known to be zero, and X is known non-zero then at least one
2083  // non-zero bit must remain.
2084  if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2085  auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2086  // Is there a known one in the portion not shifted out?
2087  if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2088  return true;
2089  // Are all the bits to be shifted out known zero?
2090  if (Known.countMinTrailingZeros() >= ShiftVal)
2091  return isKnownNonZero(X, Depth, Q);
2092  }
2093  }
2094  // div exact can only produce a zero if the dividend is zero.
2095  else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2096  return isKnownNonZero(X, Depth, Q);
2097  }
2098  // X + Y.
2099  else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2100  KnownBits XKnown = computeKnownBits(X, Depth, Q);
2101  KnownBits YKnown = computeKnownBits(Y, Depth, Q);
2102 
2103  // If X and Y are both non-negative (as signed values) then their sum is not
2104  // zero unless both X and Y are zero.
2105  if (XKnown.isNonNegative() && YKnown.isNonNegative())
2106  if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
2107  return true;
2108 
2109  // If X and Y are both negative (as signed values) then their sum is not
2110  // zero unless both X and Y equal INT_MIN.
2111  if (XKnown.isNegative() && YKnown.isNegative()) {
2112  APInt Mask = APInt::getSignedMaxValue(BitWidth);
2113  // The sign bit of X is set. If some other bit is set then X is not equal
2114  // to INT_MIN.
2115  if (XKnown.One.intersects(Mask))
2116  return true;
2117  // The sign bit of Y is set. If some other bit is set then Y is not equal
2118  // to INT_MIN.
2119  if (YKnown.One.intersects(Mask))
2120  return true;
2121  }
2122 
2123  // The sum of a non-negative number and a power of two is not zero.
2124  if (XKnown.isNonNegative() &&
2125  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2126  return true;
2127  if (YKnown.isNonNegative() &&
2128  isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2129  return true;
2130  }
2131  // X * Y.
2132  else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2133  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2134  // If X and Y are non-zero then so is X * Y as long as the multiplication
2135  // does not overflow.
2136  if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) &&
2137  isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
2138  return true;
2139  }
2140  // (C ? X : Y) != 0 if X != 0 and Y != 0.
2141  else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2142  if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
2143  isKnownNonZero(SI->getFalseValue(), Depth, Q))
2144  return true;
2145  }
2146  // PHI
2147  else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2148  // Try and detect a recurrence that monotonically increases from a
2149  // starting value, as these are common as induction variables.
2150  if (PN->getNumIncomingValues() == 2) {
2151  Value *Start = PN->getIncomingValue(0);
2152  Value *Induction = PN->getIncomingValue(1);
2153  if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2154  std::swap(Start, Induction);
2155  if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2156  if (!C->isZero() && !C->isNegative()) {
2157  ConstantInt *X;
2158  if (Q.IIQ.UseInstrInfo &&
2159  (match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2160  match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2161  !X->isNegative())
2162  return true;
2163  }
2164  }
2165  }
2166  // Check if all incoming values are non-zero constant.
2167  bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
2168  return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
2169  });
2170  if (AllNonZeroConstants)
2171  return true;
2172  }
2173 
2174  KnownBits Known(BitWidth);
2175  computeKnownBits(V, Known, Depth, Q);
2176  return Known.One != 0;
2177 }
2178 
2179 /// Return true if V2 == V1 + X, where X is known non-zero.
2180 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
2181  const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2182  if (!BO || BO->getOpcode() != Instruction::Add)
2183  return false;
2184  Value *Op = nullptr;
2185  if (V2 == BO->getOperand(0))
2186  Op = BO->getOperand(1);
2187  else if (V2 == BO->getOperand(1))
2188  Op = BO->getOperand(0);
2189  else
2190  return false;
2191  return isKnownNonZero(Op, 0, Q);
2192 }
2193 
2194 /// Return true if it is known that V1 != V2.
2195 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
2196  if (V1 == V2)
2197  return false;
2198  if (V1->getType() != V2->getType())
2199  // We can't look through casts yet.
2200  return false;
2201  if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2202  return true;
2203 
2204  if (V1->getType()->isIntOrIntVectorTy()) {
2205  // Are any known bits in V1 contradictory to known bits in V2? If V1
2206  // has a known zero where V2 has a known one, they must not be equal.
2207  KnownBits Known1 = computeKnownBits(V1, 0, Q);
2208  KnownBits Known2 = computeKnownBits(V2, 0, Q);
2209 
2210  if (Known1.Zero.intersects(Known2.One) ||
2211  Known2.Zero.intersects(Known1.One))
2212  return true;
2213  }
2214  return false;
2215 }
2216 
2217 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2218 /// simplify operations downstream. Mask is known to be zero for bits that V
2219 /// cannot have.
2220 ///
2221 /// This function is defined on values with integer type, values with pointer
2222 /// type, and vectors of integers. In the case
2223 /// where V is a vector, the mask, known zero, and known one values are the
2224 /// same width as the vector element, and the bit is set only if it is true
2225 /// for all of the elements in the vector.
2226 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2227  const Query &Q) {
2228  KnownBits Known(Mask.getBitWidth());
2229  computeKnownBits(V, Known, Depth, Q);
2230  return Mask.isSubsetOf(Known.Zero);
2231 }
2232 
2233 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow).
2234 // Returns the input and lower/upper bounds.
2235 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In,
2236  const APInt *&CLow, const APInt *&CHigh) {
2237  assert(isa<Operator>(Select) &&
2238  cast<Operator>(Select)->getOpcode() == Instruction::Select &&
2239  "Input should be a Select!");
2240 
2241  const Value *LHS, *RHS, *LHS2, *RHS2;
2242  SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor;
2243  if (SPF != SPF_SMAX && SPF != SPF_SMIN)
2244  return false;
2245 
2246  if (!match(RHS, m_APInt(CLow)))
2247  return false;
2248 
2249  SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor;
2250  if (getInverseMinMaxFlavor(SPF) != SPF2)
2251  return false;
2252 
2253  if (!match(RHS2, m_APInt(CHigh)))
2254  return false;
2255 
2256  if (SPF == SPF_SMIN)
2257  std::swap(CLow, CHigh);
2258 
2259  In = LHS2;
2260  return CLow->sle(*CHigh);
2261 }
2262 
2263 /// For vector constants, loop over the elements and find the constant with the
2264 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2265 /// or if any element was not analyzed; otherwise, return the count for the
2266 /// element with the minimum number of sign bits.
2267 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2268  unsigned TyBits) {
2269  const auto *CV = dyn_cast<Constant>(V);
2270  if (!CV || !CV->getType()->isVectorTy())
2271  return 0;
2272 
2273  unsigned MinSignBits = TyBits;
2274  unsigned NumElts = CV->getType()->getVectorNumElements();
2275  for (unsigned i = 0; i != NumElts; ++i) {
2276  // If we find a non-ConstantInt, bail out.
2277  auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2278  if (!Elt)
2279  return 0;
2280 
2281  MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2282  }
2283 
2284  return MinSignBits;
2285 }
2286 
2287 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2288  const Query &Q);
2289 
2290 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2291  const Query &Q) {
2292  unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2293  assert(Result > 0 && "At least one sign bit needs to be present!");
2294  return Result;
2295 }
2296 
2297 /// Return the number of times the sign bit of the register is replicated into
2298 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2299 /// (itself), but other cases can give us information. For example, immediately
2300 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2301 /// other, so we return 3. For vectors, return the number of sign bits for the
2302 /// vector element with the minimum number of known sign bits.
2303 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2304  const Query &Q) {
2305  assert(Depth <= MaxDepth && "Limit Search Depth");
2306 
2307  // We return the minimum number of sign bits that are guaranteed to be present
2308  // in V, so for undef we have to conservatively return 1. We don't have the
2309  // same behavior for poison though -- that's a FIXME today.
2310 
2311  Type *ScalarTy = V->getType()->getScalarType();
2312  unsigned TyBits = ScalarTy->isPointerTy() ?
2313  Q.DL.getIndexTypeSizeInBits(ScalarTy) :
2314  Q.DL.getTypeSizeInBits(ScalarTy);
2315 
2316  unsigned Tmp, Tmp2;
2317  unsigned FirstAnswer = 1;
2318 
2319  // Note that ConstantInt is handled by the general computeKnownBits case
2320  // below.
2321 
2322  if (Depth == MaxDepth)
2323  return 1; // Limit search depth.
2324 
2325  const Operator *U = dyn_cast<Operator>(V);
2326  switch (Operator::getOpcode(V)) {
2327  default: break;
2328  case Instruction::SExt:
2329  Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2330  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2331 
2332  case Instruction::SDiv: {
2333  const APInt *Denominator;
2334  // sdiv X, C -> adds log(C) sign bits.
2335  if (match(U->getOperand(1), m_APInt(Denominator))) {
2336 
2337  // Ignore non-positive denominator.
2338  if (!Denominator->isStrictlyPositive())
2339  break;
2340 
2341  // Calculate the incoming numerator bits.
2342  unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2343 
2344  // Add floor(log(C)) bits to the numerator bits.
2345  return std::min(TyBits, NumBits + Denominator->logBase2());
2346  }
2347  break;
2348  }
2349 
2350  case Instruction::SRem: {
2351  const APInt *Denominator;
2352  // srem X, C -> we know that the result is within [-C+1,C) when C is a
2353  // positive constant. This let us put a lower bound on the number of sign
2354  // bits.
2355  if (match(U->getOperand(1), m_APInt(Denominator))) {
2356 
2357  // Ignore non-positive denominator.
2358  if (!Denominator->isStrictlyPositive())
2359  break;
2360 
2361  // Calculate the incoming numerator bits. SRem by a positive constant
2362  // can't lower the number of sign bits.
2363  unsigned NumrBits =
2364  ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2365 
2366  // Calculate the leading sign bit constraints by examining the
2367  // denominator. Given that the denominator is positive, there are two
2368  // cases:
2369  //
2370  // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2371  // (1 << ceilLogBase2(C)).
2372  //
2373  // 2. the numerator is negative. Then the result range is (-C,0] and
2374  // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2375  //
2376  // Thus a lower bound on the number of sign bits is `TyBits -
2377  // ceilLogBase2(C)`.
2378 
2379  unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2380  return std::max(NumrBits, ResBits);
2381  }
2382  break;
2383  }
2384 
2385  case Instruction::AShr: {
2386  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2387  // ashr X, C -> adds C sign bits. Vectors too.
2388  const APInt *ShAmt;
2389  if (match(U->getOperand(1), m_APInt(ShAmt))) {
2390  if (ShAmt->uge(TyBits))
2391  break; // Bad shift.
2392  unsigned ShAmtLimited = ShAmt->getZExtValue();
2393  Tmp += ShAmtLimited;
2394  if (Tmp > TyBits) Tmp = TyBits;
2395  }
2396  return Tmp;
2397  }
2398  case Instruction::Shl: {
2399  const APInt *ShAmt;
2400  if (match(U->getOperand(1), m_APInt(ShAmt))) {
2401  // shl destroys sign bits.
2402  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2403  if (ShAmt->uge(TyBits) || // Bad shift.
2404  ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
2405  Tmp2 = ShAmt->getZExtValue();
2406  return Tmp - Tmp2;
2407  }
2408  break;
2409  }
2410  case Instruction::And:
2411  case Instruction::Or:
2412  case Instruction::Xor: // NOT is handled here.
2413  // Logical binary ops preserve the number of sign bits at the worst.
2414  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2415  if (Tmp != 1) {
2416  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2417  FirstAnswer = std::min(Tmp, Tmp2);
2418  // We computed what we know about the sign bits as our first
2419  // answer. Now proceed to the generic code that uses
2420  // computeKnownBits, and pick whichever answer is better.
2421  }
2422  break;
2423 
2424  case Instruction::Select: {
2425  // If we have a clamp pattern, we know that the number of sign bits will be
2426  // the minimum of the clamp min/max range.
2427  const Value *X;
2428  const APInt *CLow, *CHigh;
2429  if (isSignedMinMaxClamp(U, X, CLow, CHigh))
2430  return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
2431 
2432  Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2433  if (Tmp == 1) break;
2434  Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2435  return std::min(Tmp, Tmp2);
2436  }
2437 
2438  case Instruction::Add:
2439  // Add can have at most one carry bit. Thus we know that the output
2440  // is, at worst, one more bit than the inputs.
2441  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2442  if (Tmp == 1) break;
2443 
2444  // Special case decrementing a value (ADD X, -1):
2445  if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2446  if (CRHS->isAllOnesValue()) {
2447  KnownBits Known(TyBits);
2448  computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2449 
2450  // If the input is known to be 0 or 1, the output is 0/-1, which is all
2451  // sign bits set.
2452  if ((Known.Zero | 1).isAllOnesValue())
2453  return TyBits;
2454 
2455  // If we are subtracting one from a positive number, there is no carry
2456  // out of the result.
2457  if (Known.isNonNegative())
2458  return Tmp;
2459  }
2460 
2461  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2462  if (Tmp2 == 1) break;
2463  return std::min(Tmp, Tmp2)-1;
2464 
2465  case Instruction::Sub:
2466  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2467  if (Tmp2 == 1) break;
2468 
2469  // Handle NEG.
2470  if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2471  if (CLHS->isNullValue()) {
2472  KnownBits Known(TyBits);
2473  computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2474  // If the input is known to be 0 or 1, the output is 0/-1, which is all
2475  // sign bits set.
2476  if ((Known.Zero | 1).isAllOnesValue())
2477  return TyBits;
2478 
2479  // If the input is known to be positive (the sign bit is known clear),
2480  // the output of the NEG has the same number of sign bits as the input.
2481  if (Known.isNonNegative())
2482  return Tmp2;
2483 
2484  // Otherwise, we treat this like a SUB.
2485  }
2486 
2487  // Sub can have at most one carry bit. Thus we know that the output
2488  // is, at worst, one more bit than the inputs.
2489  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2490  if (Tmp == 1) break;
2491  return std::min(Tmp, Tmp2)-1;
2492 
2493  case Instruction::Mul: {
2494  // The output of the Mul can be at most twice the valid bits in the inputs.
2495  unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2496  if (SignBitsOp0 == 1) break;
2497  unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2498  if (SignBitsOp1 == 1) break;
2499  unsigned OutValidBits =
2500  (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
2501  return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
2502  }
2503 
2504  case Instruction::PHI: {
2505  const PHINode *PN = cast<PHINode>(U);
2506  unsigned NumIncomingValues = PN->getNumIncomingValues();
2507  // Don't analyze large in-degree PHIs.
2508  if (NumIncomingValues > 4) break;
2509  // Unreachable blocks may have zero-operand PHI nodes.
2510  if (NumIncomingValues == 0) break;
2511 
2512  // Take the minimum of all incoming values. This can't infinitely loop
2513  // because of our depth threshold.
2514  Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2515  for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2516  if (Tmp == 1) return Tmp;
2517  Tmp = std::min(
2518  Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2519  }
2520  return Tmp;
2521  }
2522 
2523  case Instruction::Trunc:
2524  // FIXME: it's tricky to do anything useful for this, but it is an important
2525  // case for targets like X86.
2526  break;
2527 
2528  case Instruction::ExtractElement:
2529  // Look through extract element. At the moment we keep this simple and skip
2530  // tracking the specific element. But at least we might find information
2531  // valid for all elements of the vector (for example if vector is sign
2532  // extended, shifted, etc).
2533  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2534 
2535  case Instruction::ShuffleVector: {
2536  // TODO: This is copied almost directly from the SelectionDAG version of
2537  // ComputeNumSignBits. It would be better if we could share common
2538  // code. If not, make sure that changes are translated to the DAG.
2539 
2540  // Collect the minimum number of sign bits that are shared by every vector
2541  // element referenced by the shuffle.
2542  auto *Shuf = cast<ShuffleVectorInst>(U);
2543  int NumElts = Shuf->getOperand(0)->getType()->getVectorNumElements();
2544  int NumMaskElts = Shuf->getMask()->getType()->getVectorNumElements();
2545  APInt DemandedLHS(NumElts, 0), DemandedRHS(NumElts, 0);
2546  for (int i = 0; i != NumMaskElts; ++i) {
2547  int M = Shuf->getMaskValue(i);
2548  assert(M < NumElts * 2 && "Invalid shuffle mask constant");
2549  // For undef elements, we don't know anything about the common state of
2550  // the shuffle result.
2551  if (M == -1)
2552  return 1;
2553  if (M < NumElts)
2554  DemandedLHS.setBit(M % NumElts);
2555  else
2556  DemandedRHS.setBit(M % NumElts);
2557  }
2559  if (!!DemandedLHS)
2560  Tmp = ComputeNumSignBits(Shuf->getOperand(0), Depth + 1, Q);
2561  if (!!DemandedRHS) {
2562  Tmp2 = ComputeNumSignBits(Shuf->getOperand(1), Depth + 1, Q);
2563  Tmp = std::min(Tmp, Tmp2);
2564  }
2565  // If we don't know anything, early out and try computeKnownBits fall-back.
2566  if (Tmp == 1)
2567  break;
2568  assert(Tmp <= V->getType()->getScalarSizeInBits() &&
2569  "Failed to determine minimum sign bits");
2570  return Tmp;
2571  }
2572  }
2573 
2574  // Finally, if we can prove that the top bits of the result are 0's or 1's,
2575  // use this information.
2576 
2577  // If we can examine all elements of a vector constant successfully, we're
2578  // done (we can't do any better than that). If not, keep trying.
2579  if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2580  return VecSignBits;
2581 
2582  KnownBits Known(TyBits);
2583  computeKnownBits(V, Known, Depth, Q);
2584 
2585  // If we know that the sign bit is either zero or one, determine the number of
2586  // identical bits in the top of the input value.
2587  return std::max(FirstAnswer, Known.countMinSignBits());
2588 }
2589 
2590 /// This function computes the integer multiple of Base that equals V.
2591 /// If successful, it returns true and returns the multiple in
2592 /// Multiple. If unsuccessful, it returns false. It looks
2593 /// through SExt instructions only if LookThroughSExt is true.
2594 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2595  bool LookThroughSExt, unsigned Depth) {
2596  const unsigned MaxDepth = 6;
2597 
2598  assert(V && "No Value?");
2599  assert(Depth <= MaxDepth && "Limit Search Depth");
2600  assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2601 
2602  Type *T = V->getType();
2603 
2604  ConstantInt *CI = dyn_cast<ConstantInt>(V);
2605 
2606  if (Base == 0)
2607  return false;
2608 
2609  if (Base == 1) {
2610  Multiple = V;
2611  return true;
2612  }
2613 
2615  Constant *BaseVal = ConstantInt::get(T, Base);
2616  if (CO && CO == BaseVal) {
2617  // Multiple is 1.
2618  Multiple = ConstantInt::get(T, 1);
2619  return true;
2620  }
2621 
2622  if (CI && CI->getZExtValue() % Base == 0) {
2623  Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2624  return true;
2625  }
2626 
2627  if (Depth == MaxDepth) return false; // Limit search depth.
2628 
2629  Operator *I = dyn_cast<Operator>(V);
2630  if (!I) return false;
2631 
2632  switch (I->getOpcode()) {
2633  default: break;
2634  case Instruction::SExt:
2635  if (!LookThroughSExt) return false;
2636  // otherwise fall through to ZExt
2638  case Instruction::ZExt:
2639  return ComputeMultiple(I->getOperand(0), Base, Multiple,
2640  LookThroughSExt, Depth+1);
2641  case Instruction::Shl:
2642  case Instruction::Mul: {
2643  Value *Op0 = I->getOperand(0);
2644  Value *Op1 = I->getOperand(1);
2645 
2646  if (I->getOpcode() == Instruction::Shl) {
2647  ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2648  if (!Op1CI) return false;
2649  // Turn Op0 << Op1 into Op0 * 2^Op1
2650  APInt Op1Int = Op1CI->getValue();
2651  uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2652  APInt API(Op1Int.getBitWidth(), 0);
2653  API.setBit(BitToSet);
2654  Op1 = ConstantInt::get(V->getContext(), API);
2655  }
2656 
2657  Value *Mul0 = nullptr;
2658  if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2659  if (Constant *Op1C = dyn_cast<Constant>(Op1))
2660  if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2661  if (Op1C->getType()->getPrimitiveSizeInBits() <
2662  MulC->getType()->getPrimitiveSizeInBits())
2663  Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2664  if (Op1C->getType()->getPrimitiveSizeInBits() >
2665  MulC->getType()->getPrimitiveSizeInBits())
2666  MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2667 
2668  // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2669  Multiple = ConstantExpr::getMul(MulC, Op1C);
2670  return true;
2671  }
2672 
2673  if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2674  if (Mul0CI->getValue() == 1) {
2675  // V == Base * Op1, so return Op1
2676  Multiple = Op1;
2677  return true;
2678  }
2679  }
2680 
2681  Value *Mul1 = nullptr;
2682  if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2683  if (Constant *Op0C = dyn_cast<Constant>(Op0))
2684  if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2685  if (Op0C->getType()->getPrimitiveSizeInBits() <
2686  MulC->getType()->getPrimitiveSizeInBits())
2687  Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2688  if (Op0C->getType()->getPrimitiveSizeInBits() >
2689  MulC->getType()->getPrimitiveSizeInBits())
2690  MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2691 
2692  // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2693  Multiple = ConstantExpr::getMul(MulC, Op0C);
2694  return true;
2695  }
2696 
2697  if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2698  if (Mul1CI->getValue() == 1) {
2699  // V == Base * Op0, so return Op0
2700  Multiple = Op0;
2701  return true;
2702  }
2703  }
2704  }
2705  }
2706 
2707  // We could not determine if V is a multiple of Base.
2708  return false;
2709 }
2710 
2712  const TargetLibraryInfo *TLI) {
2713  const Function *F = ICS.getCalledFunction();
2714  if (!F)
2715  return Intrinsic::not_intrinsic;
2716 
2717  if (F->isIntrinsic())
2718  return F->getIntrinsicID();
2719 
2720  if (!TLI)
2721  return Intrinsic::not_intrinsic;
2722 
2723  LibFunc Func;
2724  // We're going to make assumptions on the semantics of the functions, check
2725  // that the target knows that it's available in this environment and it does
2726  // not have local linkage.
2727  if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2728  return Intrinsic::not_intrinsic;
2729 
2730  if (!ICS.onlyReadsMemory())
2731  return Intrinsic::not_intrinsic;
2732 
2733  // Otherwise check if we have a call to a function that can be turned into a
2734  // vector intrinsic.
2735  switch (Func) {
2736  default:
2737  break;
2738  case LibFunc_sin:
2739  case LibFunc_sinf:
2740  case LibFunc_sinl:
2741  return Intrinsic::sin;
2742  case LibFunc_cos:
2743  case LibFunc_cosf:
2744  case LibFunc_cosl:
2745  return Intrinsic::cos;
2746  case LibFunc_exp:
2747  case LibFunc_expf:
2748  case LibFunc_expl:
2749  return Intrinsic::exp;
2750  case LibFunc_exp2:
2751  case LibFunc_exp2f:
2752  case LibFunc_exp2l:
2753  return Intrinsic::exp2;
2754  case LibFunc_log:
2755  case LibFunc_logf:
2756  case LibFunc_logl:
2757  return Intrinsic::log;
2758  case LibFunc_log10:
2759  case LibFunc_log10f:
2760  case LibFunc_log10l:
2761  return Intrinsic::log10;
2762  case LibFunc_log2:
2763  case LibFunc_log2f:
2764  case LibFunc_log2l:
2765  return Intrinsic::log2;
2766  case LibFunc_fabs:
2767  case LibFunc_fabsf:
2768  case LibFunc_fabsl:
2769  return Intrinsic::fabs;
2770  case LibFunc_fmin:
2771  case LibFunc_fminf:
2772  case LibFunc_fminl:
2773  return Intrinsic::minnum;
2774  case LibFunc_fmax:
2775  case LibFunc_fmaxf:
2776  case LibFunc_fmaxl:
2777  return Intrinsic::maxnum;
2778  case LibFunc_copysign:
2779  case LibFunc_copysignf:
2780  case LibFunc_copysignl:
2781  return Intrinsic::copysign;
2782  case LibFunc_floor:
2783  case LibFunc_floorf:
2784  case LibFunc_floorl:
2785  return Intrinsic::floor;
2786  case LibFunc_ceil:
2787  case LibFunc_ceilf:
2788  case LibFunc_ceill:
2789  return Intrinsic::ceil;
2790  case LibFunc_trunc:
2791  case LibFunc_truncf:
2792  case LibFunc_truncl:
2793  return Intrinsic::trunc;
2794  case LibFunc_rint:
2795  case LibFunc_rintf:
2796  case LibFunc_rintl:
2797  return Intrinsic::rint;
2798  case LibFunc_nearbyint:
2799  case LibFunc_nearbyintf:
2800  case LibFunc_nearbyintl:
2801  return Intrinsic::nearbyint;
2802  case LibFunc_round:
2803  case LibFunc_roundf:
2804  case LibFunc_roundl:
2805  return Intrinsic::round;
2806  case LibFunc_pow:
2807  case LibFunc_powf:
2808  case LibFunc_powl:
2809  return Intrinsic::pow;
2810  case LibFunc_sqrt:
2811  case LibFunc_sqrtf:
2812  case LibFunc_sqrtl:
2813  return Intrinsic::sqrt;
2814  }
2815 
2816  return Intrinsic::not_intrinsic;
2817 }
2818 
2819 /// Return true if we can prove that the specified FP value is never equal to
2820 /// -0.0.
2821 ///
2822 /// NOTE: this function will need to be revisited when we support non-default
2823 /// rounding modes!
2825  unsigned Depth) {
2826  if (auto *CFP = dyn_cast<ConstantFP>(V))
2827  return !CFP->getValueAPF().isNegZero();
2828 
2829  // Limit search depth.
2830  if (Depth == MaxDepth)
2831  return false;
2832 
2833  auto *Op = dyn_cast<Operator>(V);
2834  if (!Op)
2835  return false;
2836 
2837  // Check if the nsz fast-math flag is set.
2838  if (auto *FPO = dyn_cast<FPMathOperator>(Op))
2839  if (FPO->hasNoSignedZeros())
2840  return true;
2841 
2842  // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
2843  if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
2844  return true;
2845 
2846  // sitofp and uitofp turn into +0.0 for zero.
2847  if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
2848  return true;
2849 
2850  if (auto *Call = dyn_cast<CallInst>(Op)) {
2851  Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
2852  switch (IID) {
2853  default:
2854  break;
2855  // sqrt(-0.0) = -0.0, no other negative results are possible.
2856  case Intrinsic::sqrt:
2857  case Intrinsic::canonicalize:
2858  return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
2859  // fabs(x) != -0.0
2860  case Intrinsic::fabs:
2861  return true;
2862  }
2863  }
2864 
2865  return false;
2866 }
2867 
2868 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2869 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2870 /// bit despite comparing equal.
2872  const TargetLibraryInfo *TLI,
2873  bool SignBitOnly,
2874  unsigned Depth) {
2875  // TODO: This function does not do the right thing when SignBitOnly is true
2876  // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2877  // which flips the sign bits of NaNs. See
2878  // https://llvm.org/bugs/show_bug.cgi?id=31702.
2879 
2880  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2881  return !CFP->getValueAPF().isNegative() ||
2882  (!SignBitOnly && CFP->getValueAPF().isZero());
2883  }
2884 
2885  // Handle vector of constants.
2886  if (auto *CV = dyn_cast<Constant>(V)) {
2887  if (CV->getType()->isVectorTy()) {
2888  unsigned NumElts = CV->getType()->getVectorNumElements();
2889  for (unsigned i = 0; i != NumElts; ++i) {
2890  auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
2891  if (!CFP)
2892  return false;
2893  if (CFP->getValueAPF().isNegative() &&
2894  (SignBitOnly || !CFP->getValueAPF().isZero()))
2895  return false;
2896  }
2897 
2898  // All non-negative ConstantFPs.
2899  return true;
2900  }
2901  }
2902 
2903  if (Depth == MaxDepth)
2904  return false; // Limit search depth.
2905 
2906  const Operator *I = dyn_cast<Operator>(V);
2907  if (!I)
2908  return false;
2909 
2910  switch (I->getOpcode()) {
2911  default:
2912  break;
2913  // Unsigned integers are always nonnegative.
2914  case Instruction::UIToFP:
2915  return true;
2916  case Instruction::FMul:
2917  // x*x is always non-negative or a NaN.
2918  if (I->getOperand(0) == I->getOperand(1) &&
2919  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2920  return true;
2921 
2923  case Instruction::FAdd:
2924  case Instruction::FDiv:
2925  case Instruction::FRem:
2926  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2927  Depth + 1) &&
2928  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2929  Depth + 1);
2930  case Instruction::Select:
2931  return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2932  Depth + 1) &&
2933  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2934  Depth + 1);
2935  case Instruction::FPExt:
2936  case Instruction::FPTrunc:
2937  // Widening/narrowing never change sign.
2938  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2939  Depth + 1);
2940  case Instruction::ExtractElement:
2941  // Look through extract element. At the moment we keep this simple and skip
2942  // tracking the specific element. But at least we might find information
2943  // valid for all elements of the vector.
2944  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2945  Depth + 1);
2946  case Instruction::Call:
2947  const auto *CI = cast<CallInst>(I);
2948  Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2949  switch (IID) {
2950  default:
2951  break;
2952  case Intrinsic::maxnum:
2953  return (isKnownNeverNaN(I->getOperand(0), TLI) &&
2955  SignBitOnly, Depth + 1)) ||
2956  (isKnownNeverNaN(I->getOperand(1), TLI) &&
2958  SignBitOnly, Depth + 1));
2959 
2960  case Intrinsic::maximum:
2961  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2962  Depth + 1) ||
2963  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2964  Depth + 1);
2965  case Intrinsic::minnum:
2966  case Intrinsic::minimum:
2967  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2968  Depth + 1) &&
2969  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2970  Depth + 1);
2971  case Intrinsic::exp:
2972  case Intrinsic::exp2:
2973  case Intrinsic::fabs:
2974  return true;
2975 
2976  case Intrinsic::sqrt:
2977  // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2978  if (!SignBitOnly)
2979  return true;
2980  return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2981  CannotBeNegativeZero(CI->getOperand(0), TLI));
2982 
2983  case Intrinsic::powi:
2984  if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2985  // powi(x,n) is non-negative if n is even.
2986  if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2987  return true;
2988  }
2989  // TODO: This is not correct. Given that exp is an integer, here are the
2990  // ways that pow can return a negative value:
2991  //
2992  // pow(x, exp) --> negative if exp is odd and x is negative.
2993  // pow(-0, exp) --> -inf if exp is negative odd.
2994  // pow(-0, exp) --> -0 if exp is positive odd.
2995  // pow(-inf, exp) --> -0 if exp is negative odd.
2996  // pow(-inf, exp) --> -inf if exp is positive odd.
2997  //
2998  // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2999  // but we must return false if x == -0. Unfortunately we do not currently
3000  // have a way of expressing this constraint. See details in
3001  // https://llvm.org/bugs/show_bug.cgi?id=31702.
3002  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3003  Depth + 1);
3004 
3005  case Intrinsic::fma:
3006  case Intrinsic::fmuladd:
3007  // x*x+y is non-negative if y is non-negative.
3008  return I->getOperand(0) == I->getOperand(1) &&
3009  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
3010  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3011  Depth + 1);
3012  }
3013  break;
3014  }
3015  return false;
3016 }
3017 
3019  const TargetLibraryInfo *TLI) {
3020  return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3021 }
3022 
3024  return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3025 }
3026 
3028  unsigned Depth) {
3029  assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
3030 
3031  // If we're told that NaNs won't happen, assume they won't.
3032  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3033  if (FPMathOp->hasNoNaNs())
3034  return true;
3035 
3036  // Handle scalar constants.
3037  if (auto *CFP = dyn_cast<ConstantFP>(V))
3038  return !CFP->isNaN();
3039 
3040  if (Depth == MaxDepth)
3041  return false;
3042 
3043  if (auto *Inst = dyn_cast<Instruction>(V)) {
3044  switch (Inst->getOpcode()) {
3045  case Instruction::FAdd:
3046  case Instruction::FMul:
3047  case Instruction::FSub:
3048  case Instruction::FDiv:
3049  case Instruction::FRem: {
3050  // TODO: Need isKnownNeverInfinity
3051  return false;
3052  }
3053  case Instruction::Select: {
3054  return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3055  isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3056  }
3057  case Instruction::SIToFP:
3058  case Instruction::UIToFP:
3059  return true;
3060  case Instruction::FPTrunc:
3061  case Instruction::FPExt:
3062  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3063  default:
3064  break;
3065  }
3066  }
3067 
3068  if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3069  switch (II->getIntrinsicID()) {
3070  case Intrinsic::canonicalize:
3071  case Intrinsic::fabs:
3072  case Intrinsic::copysign:
3073  case Intrinsic::exp:
3074  case Intrinsic::exp2:
3075  case Intrinsic::floor:
3076  case Intrinsic::ceil:
3077  case Intrinsic::trunc:
3078  case Intrinsic::rint:
3079  case Intrinsic::nearbyint:
3080  case Intrinsic::round:
3081  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3082  case Intrinsic::sqrt:
3083  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3084  CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3085  default:
3086  return false;
3087  }
3088  }
3089 
3090  // Bail out for constant expressions, but try to handle vector constants.
3091  if (!V->getType()->isVectorTy() || !isa<Constant>(V))
3092  return false;
3093 
3094  // For vectors, verify that each element is not NaN.
3095  unsigned NumElts = V->getType()->getVectorNumElements();
3096  for (unsigned i = 0; i != NumElts; ++i) {
3097  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3098  if (!Elt)
3099  return false;
3100  if (isa<UndefValue>(Elt))
3101  continue;
3102  auto *CElt = dyn_cast<ConstantFP>(Elt);
3103  if (!CElt || CElt->isNaN())
3104  return false;
3105  }
3106  // All elements were confirmed not-NaN or undefined.
3107  return true;
3108 }
3109 
3111 
3112  // All byte-wide stores are splatable, even of arbitrary variables.
3113  if (V->getType()->isIntegerTy(8))
3114  return V;
3115 
3116  LLVMContext &Ctx = V->getContext();
3117 
3118  // Undef don't care.
3119  auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
3120  if (isa<UndefValue>(V))
3121  return UndefInt8;
3122 
3123  Constant *C = dyn_cast<Constant>(V);
3124  if (!C) {
3125  // Conceptually, we could handle things like:
3126  // %a = zext i8 %X to i16
3127  // %b = shl i16 %a, 8
3128  // %c = or i16 %a, %b
3129  // but until there is an example that actually needs this, it doesn't seem
3130  // worth worrying about.
3131  return nullptr;
3132  }
3133 
3134  // Handle 'null' ConstantArrayZero etc.
3135  if (C->isNullValue())
3137 
3138  // Constant floating-point values can be handled as integer values if the
3139  // corresponding integer value is "byteable". An important case is 0.0.
3140  if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3141  Type *Ty = nullptr;
3142  if (CFP->getType()->isHalfTy())
3143  Ty = Type::getInt16Ty(Ctx);
3144  else if (CFP->getType()->isFloatTy())
3145  Ty = Type::getInt32Ty(Ctx);
3146  else if (CFP->getType()->isDoubleTy())
3147  Ty = Type::getInt64Ty(Ctx);
3148  // Don't handle long double formats, which have strange constraints.
3149  return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty)) : nullptr;
3150  }
3151 
3152  // We can handle constant integers that are multiple of 8 bits.
3153  if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
3154  if (CI->getBitWidth() % 8 == 0) {
3155  assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
3156  if (!CI->getValue().isSplat(8))
3157  return nullptr;
3158  return ConstantInt::get(Ctx, CI->getValue().trunc(8));
3159  }
3160  }
3161 
3162  auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
3163  if (LHS == RHS)
3164  return LHS;
3165  if (!LHS || !RHS)
3166  return nullptr;
3167  if (LHS == UndefInt8)
3168  return RHS;
3169  if (RHS == UndefInt8)
3170  return LHS;
3171  return nullptr;
3172  };
3173 
3174  if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
3175  Value *Val = UndefInt8;
3176  for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
3177  if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I)))))
3178  return nullptr;
3179  return Val;
3180  }
3181 
3182  if (isa<ConstantVector>(C)) {
3183  Constant *Splat = cast<ConstantVector>(C)->getSplatValue();
3184  return Splat ? isBytewiseValue(Splat) : nullptr;
3185  }
3186 
3187  if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
3188  Value *Val = UndefInt8;
3189  for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
3190  if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I)))))
3191  return nullptr;
3192  return Val;
3193  }
3194 
3195  // Don't try to handle the handful of other constants.
3196  return nullptr;
3197 }
3198 
3199 // This is the recursive version of BuildSubAggregate. It takes a few different
3200 // arguments. Idxs is the index within the nested struct From that we are
3201 // looking at now (which is of type IndexedType). IdxSkip is the number of
3202 // indices from Idxs that should be left out when inserting into the resulting
3203 // struct. To is the result struct built so far, new insertvalue instructions
3204 // build on that.
3205 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
3207  unsigned IdxSkip,
3208  Instruction *InsertBefore) {
3209  StructType *STy = dyn_cast<StructType>(IndexedType);
3210  if (STy) {
3211  // Save the original To argument so we can modify it
3212  Value *OrigTo = To;
3213  // General case, the type indexed by Idxs is a struct
3214  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3215  // Process each struct element recursively
3216  Idxs.push_back(i);
3217  Value *PrevTo = To;
3218  To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3219  InsertBefore);
3220  Idxs.pop_back();
3221  if (!To) {
3222  // Couldn't find any inserted value for this index? Cleanup
3223  while (PrevTo != OrigTo) {
3224  InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3225  PrevTo = Del->getAggregateOperand();
3226  Del->eraseFromParent();
3227  }
3228  // Stop processing elements
3229  break;
3230  }
3231  }
3232  // If we successfully found a value for each of our subaggregates
3233  if (To)
3234  return To;
3235  }
3236  // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3237  // the struct's elements had a value that was inserted directly. In the latter
3238  // case, perhaps we can't determine each of the subelements individually, but
3239  // we might be able to find the complete struct somewhere.
3240 
3241  // Find the value that is at that particular spot
3242  Value *V = FindInsertedValue(From, Idxs);
3243 
3244  if (!V)
3245  return nullptr;
3246 
3247  // Insert the value in the new (sub) aggregate
3248  return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
3249  "tmp", InsertBefore);
3250 }
3251 
3252 // This helper takes a nested struct and extracts a part of it (which is again a
3253 // struct) into a new value. For example, given the struct:
3254 // { a, { b, { c, d }, e } }
3255 // and the indices "1, 1" this returns
3256 // { c, d }.
3257 //
3258 // It does this by inserting an insertvalue for each element in the resulting
3259 // struct, as opposed to just inserting a single struct. This will only work if
3260 // each of the elements of the substruct are known (ie, inserted into From by an
3261 // insertvalue instruction somewhere).
3262 //
3263 // All inserted insertvalue instructions are inserted before InsertBefore
3265  Instruction *InsertBefore) {
3266  assert(InsertBefore && "Must have someplace to insert!");
3267  Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3268  idx_range);
3269  Value *To = UndefValue::get(IndexedType);
3270  SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3271  unsigned IdxSkip = Idxs.size();
3272 
3273  return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3274 }
3275 
3276 /// Given an aggregate and a sequence of indices, see if the scalar value
3277 /// indexed is already around as a register, for example if it was inserted
3278 /// directly into the aggregate.
3279 ///
3280 /// If InsertBefore is not null, this function will duplicate (modified)
3281 /// insertvalues when a part of a nested struct is extracted.
3283  Instruction *InsertBefore) {
3284  // Nothing to index? Just return V then (this is useful at the end of our
3285  // recursion).
3286  if (idx_range.empty())
3287  return V;
3288  // We have indices, so V should have an indexable type.
3289  assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
3290  "Not looking at a struct or array?");
3291  assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
3292  "Invalid indices for type?");
3293 
3294  if (Constant *C = dyn_cast<Constant>(V)) {
3295  C = C->getAggregateElement(idx_range[0]);
3296  if (!C) return nullptr;
3297  return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3298  }
3299 
3300  if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3301  // Loop the indices for the insertvalue instruction in parallel with the
3302  // requested indices
3303  const unsigned *req_idx = idx_range.begin();
3304  for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3305  i != e; ++i, ++req_idx) {
3306  if (req_idx == idx_range.end()) {
3307  // We can't handle this without inserting insertvalues
3308  if (!InsertBefore)
3309  return nullptr;
3310 
3311  // The requested index identifies a part of a nested aggregate. Handle
3312  // this specially. For example,
3313  // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3314  // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3315  // %C = extractvalue {i32, { i32, i32 } } %B, 1
3316  // This can be changed into
3317  // %A = insertvalue {i32, i32 } undef, i32 10, 0
3318  // %C = insertvalue {i32, i32 } %A, i32 11, 1
3319  // which allows the unused 0,0 element from the nested struct to be
3320  // removed.
3321  return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3322  InsertBefore);
3323  }
3324 
3325  // This insert value inserts something else than what we are looking for.
3326  // See if the (aggregate) value inserted into has the value we are
3327  // looking for, then.
3328  if (*req_idx != *i)
3329  return FindInsertedValue(I->getAggregateOperand(), idx_range,
3330  InsertBefore);
3331  }
3332  // If we end up here, the indices of the insertvalue match with those
3333  // requested (though possibly only partially). Now we recursively look at
3334  // the inserted value, passing any remaining indices.
3335  return FindInsertedValue(I->getInsertedValueOperand(),
3336  makeArrayRef(req_idx, idx_range.end()),
3337  InsertBefore);
3338  }
3339 
3340  if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3341  // If we're extracting a value from an aggregate that was extracted from
3342  // something else, we can extract from that something else directly instead.
3343  // However, we will need to chain I's indices with the requested indices.
3344 
3345  // Calculate the number of indices required
3346  unsigned size = I->getNumIndices() + idx_range.size();
3347  // Allocate some space to put the new indices in
3349  Idxs.reserve(size);
3350  // Add indices from the extract value instruction
3351  Idxs.append(I->idx_begin(), I->idx_end());
3352 
3353  // Add requested indices
3354  Idxs.append(idx_range.begin(), idx_range.end());
3355 
3356  assert(Idxs.size() == size
3357  && "Number of indices added not correct?");
3358 
3359  return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
3360  }
3361  // Otherwise, we don't know (such as, extracting from a function return value
3362  // or load instruction)
3363  return nullptr;
3364 }
3365 
3366 /// Analyze the specified pointer to see if it can be expressed as a base
3367 /// pointer plus a constant offset. Return the base and offset to the caller.
3369  const DataLayout &DL) {
3370  unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
3371  APInt ByteOffset(BitWidth, 0);
3372 
3373  // We walk up the defs but use a visited set to handle unreachable code. In
3374  // that case, we stop after accumulating the cycle once (not that it
3375  // matters).
3376  SmallPtrSet<Value *, 16> Visited;
3377  while (Visited.insert(Ptr).second) {
3378  if (Ptr->getType()->isVectorTy())
3379  break;
3380 
3381  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
3382  // If one of the values we have visited is an addrspacecast, then
3383  // the pointer type of this GEP may be different from the type
3384  // of the Ptr parameter which was passed to this function. This
3385  // means when we construct GEPOffset, we need to use the size
3386  // of GEP's pointer type rather than the size of the original
3387  // pointer type.
3388  APInt GEPOffset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
3389  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
3390  break;
3391 
3392  APInt OrigByteOffset(ByteOffset);
3393  ByteOffset += GEPOffset.sextOrTrunc(ByteOffset.getBitWidth());
3394  if (ByteOffset.getMinSignedBits() > 64) {
3395  // Stop traversal if the pointer offset wouldn't fit into int64_t
3396  // (this should be removed if Offset is updated to an APInt)
3397  ByteOffset = OrigByteOffset;
3398  break;
3399  }
3400 
3401  Ptr = GEP->getPointerOperand();
3402  } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
3403  Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
3404  Ptr = cast<Operator>(Ptr)->getOperand(0);
3405  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
3406  if (GA->isInterposable())
3407  break;
3408  Ptr = GA->getAliasee();
3409  } else {
3410  break;
3411  }
3412  }
3413  Offset = ByteOffset.getSExtValue();
3414  return Ptr;
3415 }
3416 
3418  unsigned CharSize) {
3419  // Make sure the GEP has exactly three arguments.
3420  if (GEP->getNumOperands() != 3)
3421  return false;
3422 
3423  // Make sure the index-ee is a pointer to array of \p CharSize integers.
3424  // CharSize.
3426  if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
3427  return false;
3428 
3429  // Check to make sure that the first operand of the GEP is an integer and
3430  // has value 0 so that we are sure we're indexing into the initializer.
3431  const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3432  if (!FirstIdx || !FirstIdx->isZero())
3433  return false;
3434 
3435  return true;
3436 }
3437 
3439  ConstantDataArraySlice &Slice,
3440  unsigned ElementSize, uint64_t Offset) {
3441  assert(V);
3442 
3443  // Look through bitcast instructions and geps.
3444  V = V->stripPointerCasts();
3445 
3446  // If the value is a GEP instruction or constant expression, treat it as an
3447  // offset.
3448  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3449  // The GEP operator should be based on a pointer to string constant, and is
3450  // indexing into the string constant.
3451  if (!isGEPBasedOnPointerToString(GEP, ElementSize))
3452  return false;
3453 
3454  // If the second index isn't a ConstantInt, then this is a variable index
3455  // into the array. If this occurs, we can't say anything meaningful about
3456  // the string.
3457  uint64_t StartIdx = 0;
3458  if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3459  StartIdx = CI->getZExtValue();
3460  else
3461  return false;
3462  return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3463  StartIdx + Offset);
3464  }
3465 
3466  // The GEP instruction, constant or instruction, must reference a global
3467  // variable that is a constant and is initialized. The referenced constant
3468  // initializer is the array that we'll use for optimization.
3469  const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3470  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3471  return false;
3472 
3473  const ConstantDataArray *Array;
3474  ArrayType *ArrayTy;
3475  if (GV->getInitializer()->isNullValue()) {
3476  Type *GVTy = GV->getValueType();
3477  if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3478  // A zeroinitializer for the array; there is no ConstantDataArray.
3479  Array = nullptr;
3480  } else {
3481  const DataLayout &DL = GV->getParent()->getDataLayout();
3482  uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3483  uint64_t Length = SizeInBytes / (ElementSize / 8);
3484  if (Length <= Offset)
3485  return false;
3486 
3487  Slice.Array = nullptr;
3488  Slice.Offset = 0;
3489  Slice.Length = Length - Offset;
3490  return true;
3491  }
3492  } else {
3493  // This must be a ConstantDataArray.
3494  Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3495  if (!Array)
3496  return false;
3497  ArrayTy = Array->getType();
3498  }
3499  if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3500  return false;
3501 
3502  uint64_t NumElts = ArrayTy->getArrayNumElements();
3503  if (Offset > NumElts)
3504  return false;
3505 
3506  Slice.Array = Array;
3507  Slice.Offset = Offset;
3508  Slice.Length = NumElts - Offset;
3509  return true;
3510 }
3511 
3512 /// This function computes the length of a null-terminated C string pointed to
3513 /// by V. If successful, it returns true and returns the string in Str.
3514 /// If unsuccessful, it returns false.
3516  uint64_t Offset, bool TrimAtNul) {
3517  ConstantDataArraySlice Slice;
3518  if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3519  return false;
3520 
3521  if (Slice.Array == nullptr) {
3522  if (TrimAtNul) {
3523  Str = StringRef();
3524  return true;
3525  }
3526  if (Slice.Length == 1) {
3527  Str = StringRef("", 1);
3528  return true;
3529  }
3530  // We cannot instantiate a StringRef as we do not have an appropriate string
3531  // of 0s at hand.
3532  return false;
3533  }
3534 
3535  // Start out with the entire array in the StringRef.
3536  Str = Slice.Array->getAsString();
3537  // Skip over 'offset' bytes.
3538  Str = Str.substr(Slice.Offset);
3539 
3540  if (TrimAtNul) {
3541  // Trim off the \0 and anything after it. If the array is not nul
3542  // terminated, we just return the whole end of string. The client may know
3543  // some other way that the string is length-bound.
3544  Str = Str.substr(0, Str.find('\0'));
3545  }
3546  return true;
3547 }
3548 
3549 // These next two are very similar to the above, but also look through PHI
3550 // nodes.
3551 // TODO: See if we can integrate these two together.
3552 
3553 /// If we can compute the length of the string pointed to by
3554 /// the specified pointer, return 'len+1'. If we can't, return 0.
3555 static uint64_t GetStringLengthH(const Value *V,
3557  unsigned CharSize) {
3558  // Look through noop bitcast instructions.
3559  V = V->stripPointerCasts();
3560 
3561  // If this is a PHI node, there are two cases: either we have already seen it
3562  // or we haven't.
3563  if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3564  if (!PHIs.insert(PN).second)
3565  return ~0ULL; // already in the set.
3566 
3567  // If it was new, see if all the input strings are the same length.
3568  uint64_t LenSoFar = ~0ULL;
3569  for (Value *IncValue : PN->incoming_values()) {
3570  uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3571  if (Len == 0) return 0; // Unknown length -> unknown.
3572 
3573  if (Len == ~0ULL) continue;
3574 
3575  if (Len != LenSoFar && LenSoFar != ~0ULL)
3576  return 0; // Disagree -> unknown.
3577  LenSoFar = Len;
3578  }
3579 
3580  // Success, all agree.
3581  return LenSoFar;
3582  }
3583 
3584  // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3585  if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3586  uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3587  if (Len1 == 0) return 0;
3588  uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3589  if (Len2 == 0) return 0;
3590  if (Len1 == ~0ULL) return Len2;
3591  if (Len2 == ~0ULL) return Len1;
3592  if (Len1 != Len2) return 0;
3593  return Len1;
3594  }
3595 
3596  // Otherwise, see if we can read the string.
3597  ConstantDataArraySlice Slice;
3598  if (!getConstantDataArrayInfo(V, Slice, CharSize))
3599  return 0;
3600 
3601  if (Slice.Array == nullptr)
3602  return 1;
3603 
3604  // Search for nul characters
3605  unsigned NullIndex = 0;
3606  for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3607  if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3608  break;
3609  }
3610 
3611  return NullIndex + 1;
3612 }
3613 
3614 /// If we can compute the length of the string pointed to by
3615 /// the specified pointer, return 'len+1'. If we can't, return 0.
3616 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3617  if (!V->getType()->isPointerTy())
3618  return 0;
3619 
3621  uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3622  // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3623  // an empty string as a length.
3624  return Len == ~0ULL ? 1 : Len;
3625 }
3626 
3628  assert(Call &&
3629  "getArgumentAliasingToReturnedPointer only works on nonnull calls");
3630  if (const Value *RV = Call->getReturnedArgOperand())
3631  return RV;
3632  // This can be used only as a aliasing property.
3634  return Call->getArgOperand(0);
3635  return nullptr;
3636 }
3637 
3639  const CallBase *Call) {
3640  return Call->getIntrinsicID() == Intrinsic::launder_invariant_group ||
3641  Call->getIntrinsicID() == Intrinsic::strip_invariant_group;
3642 }
3643 
3644 /// \p PN defines a loop-variant pointer to an object. Check if the
3645 /// previous iteration of the loop was referring to the same object as \p PN.
3647  const LoopInfo *LI) {
3648  // Find the loop-defined value.
3649  Loop *L = LI->getLoopFor(PN->getParent());
3650  if (PN->getNumIncomingValues() != 2)
3651  return true;
3652 
3653  // Find the value from previous iteration.
3654  auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3655  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3656  PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3657  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3658  return true;
3659 
3660  // If a new pointer is loaded in the loop, the pointer references a different
3661  // object in every iteration. E.g.:
3662  // for (i)
3663  // int *p = a[i];
3664  // ...
3665  if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3666  if (!L->isLoopInvariant(Load->getPointerOperand()))
3667  return false;
3668  return true;
3669 }
3670 
3672  unsigned MaxLookup) {
3673  if (!V->getType()->isPointerTy())
3674  return V;
3675  for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3676  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3677  V = GEP->getPointerOperand();
3678  } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3679  Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3680  V = cast<Operator>(V)->getOperand(0);
3681  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3682  if (GA->isInterposable())
3683  return V;
3684  V = GA->getAliasee();
3685  } else if (isa<AllocaInst>(V)) {
3686  // An alloca can't be further simplified.
3687  return V;
3688  } else {
3689  if (auto *Call = dyn_cast<CallBase>(V)) {
3690  // CaptureTracking can know about special capturing properties of some
3691  // intrinsics like launder.invariant.group, that can't be expressed with
3692  // the attributes, but have properties like returning aliasing pointer.
3693  // Because some analysis may assume that nocaptured pointer is not
3694  // returned from some special intrinsic (because function would have to
3695  // be marked with returns attribute), it is crucial to use this function
3696  // because it should be in sync with CaptureTracking. Not using it may
3697  // cause weird miscompilations where 2 aliasing pointers are assumed to
3698  // noalias.
3699  if (auto *RP = getArgumentAliasingToReturnedPointer(Call)) {
3700  V = RP;
3701  continue;
3702  }
3703  }
3704 
3705  // See if InstructionSimplify knows any relevant tricks.
3706  if (Instruction *I = dyn_cast<Instruction>(V))
3707  // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3708  if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3709  V = Simplified;
3710  continue;
3711  }
3712 
3713  return V;
3714  }
3715  assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3716  }
3717  return V;
3718 }
3719 
3721  const DataLayout &DL, LoopInfo *LI,
3722  unsigned MaxLookup) {
3723  SmallPtrSet<Value *, 4> Visited;
3724  SmallVector<Value *, 4> Worklist;
3725  Worklist.push_back(V);
3726  do {
3727  Value *P = Worklist.pop_back_val();
3728  P = GetUnderlyingObject(P, DL, MaxLookup);
3729 
3730  if (!Visited.insert(P).second)
3731  continue;
3732 
3733  if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3734  Worklist.push_back(SI->getTrueValue());
3735  Worklist.push_back(SI->getFalseValue());
3736  continue;
3737  }
3738 
3739  if (PHINode *PN = dyn_cast<PHINode>(P)) {
3740  // If this PHI changes the underlying object in every iteration of the
3741  // loop, don't look through it. Consider:
3742  // int **A;
3743  // for (i) {
3744  // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3745  // Curr = A[i];
3746  // *Prev, *Curr;
3747  //
3748  // Prev is tracking Curr one iteration behind so they refer to different
3749  // underlying objects.
3750  if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3752  for (Value *IncValue : PN->incoming_values())
3753  Worklist.push_back(IncValue);
3754  continue;
3755  }
3756 
3757  Objects.push_back(P);
3758  } while (!Worklist.empty());
3759 }
3760 
3761 /// This is the function that does the work of looking through basic
3762 /// ptrtoint+arithmetic+inttoptr sequences.
3763 static const Value *getUnderlyingObjectFromInt(const Value *V) {
3764  do {
3765  if (const Operator *U = dyn_cast<Operator>(V)) {
3766  // If we find a ptrtoint, we can transfer control back to the
3767  // regular getUnderlyingObjectFromInt.
3768  if (U->getOpcode() == Instruction::PtrToInt)
3769  return U->getOperand(0);
3770  // If we find an add of a constant, a multiplied value, or a phi, it's
3771  // likely that the other operand will lead us to the base
3772  // object. We don't have to worry about the case where the
3773  // object address is somehow being computed by the multiply,
3774  // because our callers only care when the result is an
3775  // identifiable object.
3776  if (U->getOpcode() != Instruction::Add ||
3777  (!isa<ConstantInt>(U->getOperand(1)) &&
3778  Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
3779  !isa<PHINode>(U->getOperand(1))))
3780  return V;
3781  V = U->getOperand(0);
3782  } else {
3783  return V;
3784  }
3785  assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
3786  } while (true);
3787 }
3788 
3789 /// This is a wrapper around GetUnderlyingObjects and adds support for basic
3790 /// ptrtoint+arithmetic+inttoptr sequences.
3791 /// It returns false if unidentified object is found in GetUnderlyingObjects.
3793  SmallVectorImpl<Value *> &Objects,
3794  const DataLayout &DL) {
3796  SmallVector<const Value *, 4> Working(1, V);
3797  do {
3798  V = Working.pop_back_val();
3799 
3801  GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
3802 
3803  for (Value *V : Objs) {
3804  if (!Visited.insert(V).second)
3805  continue;
3806  if (Operator::getOpcode(V) == Instruction::IntToPtr) {
3807  const Value *O =
3808  getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
3809  if (O->getType()->isPointerTy()) {
3810  Working.push_back(O);
3811  continue;
3812  }
3813  }
3814  // If GetUnderlyingObjects fails to find an identifiable object,
3815  // getUnderlyingObjectsForCodeGen also fails for safety.
3816  if (!isIdentifiedObject(V)) {
3817  Objects.clear();
3818  return false;
3819  }
3820  Objects.push_back(const_cast<Value *>(V));
3821  }
3822  } while (!Working.empty());
3823  return true;
3824 }
3825 
3826 /// Return true if the only users of this pointer are lifetime markers.
3828  for (const User *U : V->users()) {
3829  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3830  if (!II) return false;
3831 
3832  if (!II->isLifetimeStartOrEnd())
3833  return false;
3834  }
3835  return true;
3836 }
3837 
3839  const Instruction *CtxI,
3840  const DominatorTree *DT) {
3841  const Operator *Inst = dyn_cast<Operator>(V);
3842  if (!Inst)
3843  return false;
3844 
3845  for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3846  if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3847  if (C->canTrap())
3848  return false;
3849 
3850  switch (Inst->getOpcode()) {
3851  default:
3852  return true;
3853  case Instruction::UDiv:
3854  case Instruction::URem: {
3855  // x / y is undefined if y == 0.
3856  const APInt *V;
3857  if (match(Inst->getOperand(1), m_APInt(V)))
3858  return *V != 0;
3859  return false;
3860  }
3861  case Instruction::SDiv:
3862  case Instruction::SRem: {
3863  // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3864  const APInt *Numerator, *Denominator;
3865  if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3866  return false;
3867  // We cannot hoist this division if the denominator is 0.
3868  if (*Denominator == 0)
3869  return false;
3870  // It's safe to hoist if the denominator is not 0 or -1.
3871  if (*Denominator != -1)
3872  return true;
3873  // At this point we know that the denominator is -1. It is safe to hoist as
3874  // long we know that the numerator is not INT_MIN.
3875  if (match(Inst->getOperand(0), m_APInt(Numerator)))
3876  return !Numerator->isMinSignedValue();
3877  // The numerator *might* be MinSignedValue.
3878  return false;
3879  }
3880  case Instruction::Load: {
3881  const LoadInst *LI = cast<LoadInst>(Inst);
3882  if (!LI->isUnordered() ||
3883  // Speculative load may create a race that did not exist in the source.
3884  LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3885  // Speculative load may load data from dirty regions.
3886  LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3887  LI->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3888  return false;
3889  const DataLayout &DL = LI->getModule()->getDataLayout();
3891  LI->getAlignment(), DL, CtxI, DT);
3892  }
3893  case Instruction::Call: {
3894  auto *CI = cast<const CallInst>(Inst);
3895  const Function *Callee = CI->getCalledFunction();
3896 
3897  // The called function could have undefined behavior or side-effects, even
3898  // if marked readnone nounwind.
3899  return Callee && Callee->isSpeculatable();
3900  }
3901  case Instruction::VAArg:
3902  case Instruction::Alloca:
3903  case Instruction::Invoke:
3904  case Instruction::PHI:
3905  case Instruction::Store:
3906  case Instruction::Ret:
3907  case Instruction::Br:
3908  case Instruction::IndirectBr:
3909  case Instruction::Switch:
3910  case Instruction::Unreachable:
3911  case Instruction::Fence:
3912  case Instruction::AtomicRMW:
3913  case Instruction::AtomicCmpXchg:
3914  case Instruction::LandingPad:
3915  case Instruction::Resume:
3916  case Instruction::CatchSwitch:
3917  case Instruction::CatchPad:
3918  case Instruction::CatchRet:
3919  case Instruction::CleanupPad:
3920  case Instruction::CleanupRet:
3921  return false; // Misc instructions which have effects
3922  }
3923 }
3924 
3927 }
3928 
3930  const Value *LHS, const Value *RHS, const DataLayout &DL,
3931  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
3932  bool UseInstrInfo) {
3933  // Multiplying n * m significant bits yields a result of n + m significant
3934  // bits. If the total number of significant bits does not exceed the
3935  // result bit width (minus 1), there is no overflow.
3936  // This means if we have enough leading zero bits in the operands
3937  // we can guarantee that the result does not overflow.
3938  // Ref: "Hacker's Delight" by Henry Warren
3939  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3940  KnownBits LHSKnown(BitWidth);
3941  KnownBits RHSKnown(BitWidth);
3942  computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT, nullptr,
3943  UseInstrInfo);
3944  computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT, nullptr,
3945  UseInstrInfo);
3946  // Note that underestimating the number of zero bits gives a more
3947  // conservative answer.
3948  unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
3949  RHSKnown.countMinLeadingZeros();
3950  // First handle the easy case: if we have enough zero bits there's
3951  // definitely no overflow.
3952  if (ZeroBits >= BitWidth)
3954 
3955  // Get the largest possible values for each operand.
3956  APInt LHSMax = ~LHSKnown.Zero;
3957  APInt RHSMax = ~RHSKnown.Zero;
3958 
3959  // We know the multiply operation doesn't overflow if the maximum values for
3960  // each operand will not overflow after we multiply them together.
3961  bool MaxOverflow;
3962  (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3963  if (!MaxOverflow)
3965 
3966  // We know it always overflows if multiplying the smallest possible values for
3967  // the operands also results in overflow.
3968  bool MinOverflow;
3969  (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3970  if (MinOverflow)
3972 
3974 }
3975 
3978  const DataLayout &DL, AssumptionCache *AC,
3979  const Instruction *CxtI,
3980  const DominatorTree *DT, bool UseInstrInfo) {
3981  // Multiplying n * m significant bits yields a result of n + m significant
3982  // bits. If the total number of significant bits does not exceed the
3983  // result bit width (minus 1), there is no overflow.
3984  // This means if we have enough leading sign bits in the operands
3985  // we can guarantee that the result does not overflow.
3986  // Ref: "Hacker's Delight" by Henry Warren
3987  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3988 
3989  // Note that underestimating the number of sign bits gives a more
3990  // conservative answer.
3991  unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
3992  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
3993 
3994  // First handle the easy case: if we have enough sign bits there's
3995  // definitely no overflow.
3996  if (SignBits > BitWidth + 1)
3998 
3999  // There are two ambiguous cases where there can be no overflow:
4000  // SignBits == BitWidth + 1 and
4001  // SignBits == BitWidth
4002  // The second case is difficult to check, therefore we only handle the
4003  // first case.
4004  if (SignBits == BitWidth + 1) {
4005  // It overflows only when both arguments are negative and the true
4006  // product is exactly the minimum negative number.
4007  // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
4008  // For simplicity we just check if at least one side is not negative.
4009  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4010  nullptr, UseInstrInfo);
4011  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4012  nullptr, UseInstrInfo);
4013  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
4015  }
4017 }
4018 
4020  const Value *LHS, const Value *RHS, const DataLayout &DL,
4021  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4022  bool UseInstrInfo) {
4023  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4024  nullptr, UseInstrInfo);
4025  if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
4026  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4027  nullptr, UseInstrInfo);
4028 
4029  if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
4030  // The sign bit is set in both cases: this MUST overflow.
4032  }
4033 
4034  if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
4035  // The sign bit is clear in both cases: this CANNOT overflow.
4037  }
4038  }
4039 
4041 }
4042 
4043 /// Return true if we can prove that adding the two values of the
4044 /// knownbits will not overflow.
4045 /// Otherwise return false.
4046 static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
4047  const KnownBits &RHSKnown) {
4048  // Addition of two 2's complement numbers having opposite signs will never
4049  // overflow.
4050  if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
4051  (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
4052  return true;
4053 
4054  // If either of the values is known to be non-negative, adding them can only
4055  // overflow if the second is also non-negative, so we can assume that.
4056  // Two non-negative numbers will only overflow if there is a carry to the
4057  // sign bit, so we can check if even when the values are as big as possible
4058  // there is no overflow to the sign bit.
4059  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
4060  APInt MaxLHS = ~LHSKnown.Zero;
4061  MaxLHS.clearSignBit();
4062  APInt MaxRHS = ~RHSKnown.Zero;
4063  MaxRHS.clearSignBit();
4064  APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
4065  return Result.isSignBitClear();
4066  }
4067 
4068  // If either of the values is known to be negative, adding them can only
4069  // overflow if the second is also negative, so we can assume that.
4070  // Two negative number will only overflow if there is no carry to the sign
4071  // bit, so we can check if even when the values are as small as possible
4072  // there is overflow to the sign bit.
4073  if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
4074  APInt MinLHS = LHSKnown.One;
4075  MinLHS.clearSignBit();
4076  APInt MinRHS = RHSKnown.One;
4077  MinRHS.clearSignBit();
4078  APInt Result = std::move(MinLHS) + std::move(MinRHS);
4079  return Result.isSignBitSet();
4080  }
4081 
4082  // If we reached here it means that we know nothing about the sign bits.
4083  // In this case we can't know if there will be an overflow, since by
4084  // changing the sign bits any two values can be made to overflow.
4085  return false;
4086 }
4087 
4089  const Value *RHS,
4090  const AddOperator *Add,
4091  const DataLayout &DL,
4092  AssumptionCache *AC,
4093  const Instruction *CxtI,
4094  const DominatorTree *DT) {
4095  if (Add && Add->hasNoSignedWrap()) {
4097  }
4098 
4099  // If LHS and RHS each have at least two sign bits, the addition will look
4100  // like
4101  //
4102  // XX..... +
4103  // YY.....
4104  //
4105  // If the carry into the most significant position is 0, X and Y can't both
4106  // be 1 and therefore the carry out of the addition is also 0.
4107  //
4108  // If the carry into the most significant position is 1, X and Y can't both
4109  // be 0 and therefore the carry out of the addition is also 1.
4110  //
4111  // Since the carry into the most significant position is always equal to
4112  // the carry out of the addition, there is no signed overflow.
4113  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4114  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4116 
4117  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
4118  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
4119 
4120  if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
4122 
4123  // The remaining code needs Add to be available. Early returns if not so.
4124  if (!Add)
4126 
4127  // If the sign of Add is the same as at least one of the operands, this add
4128  // CANNOT overflow. This is particularly useful when the sum is
4129  // @llvm.assume'ed non-negative rather than proved so from analyzing its
4130  // operands.
4131  bool LHSOrRHSKnownNonNegative =
4132  (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
4133  bool LHSOrRHSKnownNegative =
4134  (LHSKnown.isNegative() || RHSKnown.isNegative());
4135  if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
4136  KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
4137  if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
4138  (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
4140  }
4141  }
4142 
4144 }
4145 
4147  const Value *RHS,
4148  const DataLayout &DL,
4149  AssumptionCache *AC,
4150  const Instruction *CxtI,
4151  const DominatorTree *DT) {
4152  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
4153  if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
4154  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
4155 
4156  // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
4157  if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
4159 
4160  // If the LHS is non-negative and the RHS negative, we always wrap.
4161  if (LHSKnown.isNonNegative() && RHSKnown.isNegative())
4163  }
4164 
4166 }
4167 
4169  const Value *RHS,
4170  const DataLayout &DL,
4171  AssumptionCache *AC,
4172  const Instruction *CxtI,
4173  const DominatorTree *DT) {
4174  // If LHS and RHS each have at least two sign bits, the subtraction
4175  // cannot overflow.
4176  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4177  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4179 
4180  KnownBits LHSKnown = computeKnownBits(LHS, DL, 0, AC, CxtI, DT);
4181 
4182  KnownBits RHSKnown = computeKnownBits(RHS, DL, 0, AC, CxtI, DT);
4183 
4184  // Subtraction of two 2's complement numbers having identical signs will
4185  // never overflow.
4186  if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
4187  (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
4189 
4190  // TODO: implement logic similar to checkRippleForAdd
4192 }
4193 
4195  const DominatorTree &DT) {
4196 #ifndef NDEBUG
4197  auto IID = II->getIntrinsicID();
4198  assert((IID == Intrinsic::sadd_with_overflow ||
4199  IID == Intrinsic::uadd_with_overflow ||
4200  IID == Intrinsic::ssub_with_overflow ||
4201  IID == Intrinsic::usub_with_overflow ||
4202  IID == Intrinsic::smul_with_overflow ||
4203  IID == Intrinsic::umul_with_overflow) &&
4204  "Not an overflow intrinsic!");
4205 #endif
4206 
4207  SmallVector<const BranchInst *, 2> GuardingBranches;
4209 
4210  for (const User *U : II->users()) {
4211  if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
4212  assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
4213 
4214  if (EVI->getIndices()[0] == 0)
4215  Results.push_back(EVI);
4216  else {
4217  assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
4218 
4219  for (const auto *U : EVI->users())
4220  if (const auto *B = dyn_cast<BranchInst>(U)) {
4221  assert(B->isConditional() && "How else is it using an i1?");
4222  GuardingBranches.push_back(B);
4223  }
4224  }
4225  } else {
4226  // We are using the aggregate directly in a way we don't want to analyze
4227  // here (storing it to a global, say).
4228  return false;
4229  }
4230  }
4231 
4232  auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4233  BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4234  if (!NoWrapEdge.isSingleEdge())
4235  return false;
4236 
4237  // Check if all users of the add are provably no-wrap.
4238  for (const auto *Result : Results) {
4239  // If the extractvalue itself is not executed on overflow, the we don't
4240  // need to check each use separately, since domination is transitive.
4241  if (DT.dominates(NoWrapEdge, Result->getParent()))
4242  continue;
4243 
4244  for (auto &RU : Result->uses())
4245  if (!DT.dominates(NoWrapEdge, RU))
4246  return false;
4247  }
4248 
4249  return true;
4250  };
4251 
4252  return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4253 }
4254 
4255 
4257  const DataLayout &DL,
4258  AssumptionCache *AC,
4259  const Instruction *CxtI,
4260  const DominatorTree *DT) {
4262  Add, DL, AC, CxtI, DT);
4263 }
4264 
4266  const Value *RHS,
4267  const DataLayout &DL,
4268  AssumptionCache *AC,
4269  const Instruction *CxtI,
4270  const DominatorTree *DT) {
4271  return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
4272 }
4273 
4275  // A memory operation returns normally if it isn't volatile. A volatile
4276  // operation is allowed to trap.
4277  //
4278  // An atomic operation isn't guaranteed to return in a reasonable amount of
4279  // time because it's possible for another thread to interfere with it for an
4280  // arbitrary length of time, but programs aren't allowed to rely on that.
4281  if (const LoadInst *LI = dyn_cast<LoadInst>(I))
4282  return !LI->isVolatile();
4283  if (const StoreInst *SI = dyn_cast<StoreInst>(I))
4284  return !SI->isVolatile();
4285  if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
4286  return !CXI->isVolatile();
4287  if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
4288  return !RMWI->isVolatile();
4289  if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
4290  return !MII->isVolatile();
4291 
4292  // If there is no successor, then execution can't transfer to it.
4293  if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
4294  return !CRI->unwindsToCaller();
4295  if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
4296  return !CatchSwitch->unwindsToCaller();
4297  if (isa<ResumeInst>(I))
4298  return false;
4299  if (isa<ReturnInst>(I))
4300  return false;
4301  if (isa<UnreachableInst>(I))
4302  return false;
4303 
4304  // Calls can throw, or contain an infinite loop, or kill the process.
4305  if (auto CS = ImmutableCallSite(I)) {
4306  // Call sites that throw have implicit non-local control flow.
4307  if (!CS.doesNotThrow())
4308  return false;
4309 
4310  // Non-throwing call sites can loop infinitely, call exit/pthread_exit
4311  // etc. and thus not return. However, LLVM already assumes that
4312  //
4313  // - Thread exiting actions are modeled as writes to memory invisible to
4314  // the program.
4315  //
4316  // - Loops that don't have side effects (side effects are volatile/atomic
4317  // stores and IO) always terminate (see http://llvm.org/PR965).
4318  // Furthermore IO itself is also modeled as writes to memory invisible to
4319  // the program.
4320  //
4321  // We rely on those assumptions here, and use the memory effects of the call
4322  // target as a proxy for checking that it always returns.
4323 
4324  // FIXME: This isn't aggressive enough; a call which only writes to a global
4325  // is guaranteed to return.
4326  return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
4327  match(I, m_Intrinsic<Intrinsic::assume>()) ||
4328  match(I, m_Intrinsic<Intrinsic::sideeffect>());
4329  }
4330 
4331  // Other instructions return normally.
4332  return true;
4333 }
4334 
4336  // TODO: This is slightly consdervative for invoke instruction since exiting
4337  // via an exception *is* normal control for them.
4338  for (auto I = BB->begin(), E = BB->end(); I != E; ++I)
4340  return false;
4341  return true;
4342 }
4343 
4345  const Loop *L) {
4346  // The loop header is guaranteed to be executed for every iteration.
4347  //
4348  // FIXME: Relax this constraint to cover all basic blocks that are
4349  // guaranteed to be executed at every iteration.
4350  if (I->getParent() != L->getHeader()) return false;
4351 
4352  for (const Instruction &LI : *L->getHeader()) {
4353  if (&LI == I) return true;
4354  if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
4355  }
4356  llvm_unreachable("Instruction not contained in its own parent basic block.");
4357 }
4358 
4360  switch (I->getOpcode()) {
4361  case Instruction::Add:
4362  case Instruction::Sub:
4363  case Instruction::Xor:
4364  case Instruction::Trunc:
4365  case Instruction::BitCast:
4366  case Instruction::AddrSpaceCast:
4367  case Instruction::Mul:
4368  case Instruction::Shl:
4369  case Instruction::GetElementPtr:
4370  // These operations all propagate poison unconditionally. Note that poison
4371  // is not any particular value, so xor or subtraction of poison with
4372  // itself still yields poison, not zero.
4373  return true;
4374 
4375  case Instruction::AShr:
4376  case Instruction::SExt:
4377  // For these operations, one bit of the input is replicated across
4378  // multiple output bits. A replicated poison bit is still poison.
4379  return true;
4380 
4381  case Instruction::ICmp:
4382  // Comparing poison with any value yields poison. This is why, for
4383  // instance, x s< (x +nsw 1) can be folded to true.
4384  return true;
4385 
4386  default:
4387  return false;
4388  }
4389 }
4390 
4392  switch (I->getOpcode()) {
4393  case Instruction::Store:
4394  return cast<StoreInst>(I)->getPointerOperand();
4395 
4396  case Instruction::Load:
4397  return cast<LoadInst>(I)->getPointerOperand();
4398 
4399  case Instruction::AtomicCmpXchg:
4400  return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
4401 
4402  case Instruction::AtomicRMW:
4403  return cast<AtomicRMWInst>(I)->getPointerOperand();
4404 
4405  case Instruction::UDiv:
4406  case Instruction::SDiv:
4407  case Instruction::URem:
4408  case Instruction::SRem:
4409  return I->getOperand(1);
4410 
4411  default:
4412  return nullptr;
4413  }
4414 }
4415 
4417  // We currently only look for uses of poison values within the same basic
4418  // block, as that makes it easier to guarantee that the uses will be
4419  // executed given that PoisonI is executed.
4420  //
4421  // FIXME: Expand this to consider uses beyond the same basic block. To do
4422  // this, look out for the distinction between post-dominance and strong
4423  // post-dominance.
4424  const BasicBlock *BB = PoisonI->getParent();
4425 
4426  // Set of instructions that we have proved will yield poison if PoisonI
4427  // does.
4428  SmallSet<const Value *, 16> YieldsPoison;
4430  YieldsPoison.insert(PoisonI);
4431  Visited.insert(PoisonI->getParent());
4432 
4433  BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
4434 
4435  unsigned Iter = 0;
4436  while (Iter++ < MaxDepth) {
4437  for (auto &I : make_range(Begin, End)) {
4438  if (&I != PoisonI) {
4439  const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
4440  if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
4441  return true;
4443  return false;
4444  }
4445 
4446  // Mark poison that propagates from I through uses of I.
4447  if (YieldsPoison.count(&I)) {
4448  for (const User *User : I.users()) {
4449  const Instruction *UserI = cast<Instruction>(User);
4450  if (propagatesFullPoison(UserI))
4451  YieldsPoison.insert(User);
4452  }
4453  }
4454  }
4455 
4456  if (auto *NextBB = BB->getSingleSuccessor()) {
4457  if (Visited.insert(NextBB).second) {
4458  BB = NextBB;
4459  Begin = BB->getFirstNonPHI()->getIterator();
4460  End = BB->end();
4461  continue;
4462  }
4463  }
4464 
4465  break;
4466  }
4467  return false;
4468 }
4469 
4470 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
4471  if (FMF.noNaNs())
4472  return true;
4473 
4474  if (auto *C = dyn_cast<ConstantFP>(V))
4475  return !C->isNaN();
4476 
4477  if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4478  if (!C->getElementType()->isFloatingPointTy())
4479  return false;
4480  for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4481  if (C->getElementAsAPFloat(I).isNaN())
4482  return false;
4483  }
4484  return true;
4485  }
4486 
4487  return false;
4488 }
4489 
4490 static bool isKnownNonZero(const Value *V) {
4491  if (auto *C = dyn_cast<ConstantFP>(V))
4492  return !C->isZero();
4493 
4494  if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4495  if (!C->getElementType()->isFloatingPointTy())
4496  return false;
4497  for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4498  if (C->getElementAsAPFloat(I).isZero())
4499  return false;
4500  }
4501  return true;
4502  }
4503 
4504  return false;
4505 }
4506 
4507 /// Match clamp pattern for float types without care about NaNs or signed zeros.
4508 /// Given non-min/max outer cmp/select from the clamp pattern this
4509 /// function recognizes if it can be substitued by a "canonical" min/max
4510 /// pattern.
4512  Value *CmpLHS, Value *CmpRHS,
4513  Value *TrueVal, Value *FalseVal,
4514  Value *&LHS, Value *&RHS) {
4515  // Try to match
4516  // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
4517  // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
4518  // and return description of the outer Max/Min.
4519 
4520  // First, check if select has inverse order:
4521  if (CmpRHS == FalseVal) {
4522  std::swap(TrueVal, FalseVal);
4523  Pred = CmpInst::getInversePredicate(Pred);
4524  }
4525 
4526  // Assume success now. If there's no match, callers should not use these anyway.
4527  LHS = TrueVal;
4528  RHS = FalseVal;
4529 
4530  const APFloat *FC1;
4531  if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
4532  return {SPF_UNKNOWN, SPNB_NA, false};
4533 
4534  const APFloat *FC2;
4535  switch (Pred) {
4536  case CmpInst::FCMP_OLT:
4537  case CmpInst::FCMP_OLE:
4538  case CmpInst::FCMP_ULT:
4539  case CmpInst::FCMP_ULE:
4540  if (match(FalseVal,
4541  m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
4542  m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4543  FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
4544  return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
4545  break;
4546  case CmpInst::FCMP_OGT:
4547  case CmpInst::FCMP_OGE:
4548  case CmpInst::FCMP_UGT:
4549  case CmpInst::FCMP_UGE:
4550  if (match(FalseVal,
4551  m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
4552  m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4553  FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
4554  return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
4555  break;
4556  default:
4557  break;
4558  }
4559 
4560  return {SPF_UNKNOWN, SPNB_NA, false};
4561 }
4562 
4563 /// Recognize variations of:
4564 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
4566  Value *CmpLHS, Value *CmpRHS,
4567  Value *TrueVal, Value *FalseVal) {
4568  // Swap the select operands and predicate to match the patterns below.
4569  if (CmpRHS != TrueVal) {
4570  Pred = ICmpInst::getSwappedPredicate(Pred);
4571  std::swap(TrueVal, FalseVal);
4572  }
4573  const APInt *C1;
4574  if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
4575  const APInt *C2;
4576  // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
4577  if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4578  C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
4579  return {SPF_SMAX, SPNB_NA, false};
4580 
4581  // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
4582  if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4583  C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
4584  return {SPF_SMIN, SPNB_NA, false};
4585 
4586  // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
4587  if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4588  C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
4589  return {SPF_UMAX, SPNB_NA, false};
4590 
4591  // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
4592  if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4593  C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
4594  return {SPF_UMIN, SPNB_NA, false};
4595  }
4596  return {SPF_UNKNOWN, SPNB_NA, false};
4597 }
4598 
4599 /// Recognize variations of:
4600 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
4602  Value *CmpLHS, Value *CmpRHS,
4603  Value *TVal, Value *FVal,
4604  unsigned Depth) {
4605  // TODO: Allow FP min/max with nnan/nsz.
4606  assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
4607 
4608  Value *A, *B;
4609  SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
4611  return {SPF_UNKNOWN, SPNB_NA, false};
4612 
4613  Value *C, *D;
4614  SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
4615  if (L.Flavor != R.Flavor)
4616  return {SPF_UNKNOWN, SPNB_NA, false};
4617 
4618  // We have something like: x Pred y ? min(a, b) : min(c, d).
4619  // Try to match the compare to the min/max operations of the select operands.
4620  // First, make sure we have the right compare predicate.
4621  switch (L.Flavor) {
4622  case SPF_SMIN:
4623  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
4624  Pred = ICmpInst::getSwappedPredicate(Pred);
4625  std::swap(CmpLHS, CmpRHS);
4626  }
4627  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
4628  break;
4629  return {SPF_UNKNOWN, SPNB_NA, false};
4630  case SPF_SMAX:
4631  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
4632  Pred = ICmpInst::getSwappedPredicate(Pred);
4633  std::swap(CmpLHS, CmpRHS);
4634  }
4635  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
4636  break;
4637  return {SPF_UNKNOWN, SPNB_NA, false};
4638  case SPF_UMIN:
4639  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
4640  Pred = ICmpInst::getSwappedPredicate(Pred);
4641  std::swap(CmpLHS, CmpRHS);
4642  }
4643  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
4644  break;
4645  return {SPF_UNKNOWN, SPNB_NA, false};
4646  case SPF_UMAX:
4647  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
4648  Pred = ICmpInst::getSwappedPredicate(Pred);
4649  std::swap(CmpLHS, CmpRHS);
4650  }
4651  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
4652  break;
4653  return {SPF_UNKNOWN, SPNB_NA, false};
4654  default:
4655  return {SPF_UNKNOWN, SPNB_NA, false};
4656  }
4657 
4658  // If there is a common operand in the already matched min/max and the other
4659  // min/max operands match the compare operands (either directly or inverted),
4660  // then this is min/max of the same flavor.
4661 
4662  // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4663  // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4664  if (D == B) {
4665  if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4666  match(A, m_Not(m_Specific(CmpRHS)))))
4667  return {L.Flavor, SPNB_NA, false};
4668  }
4669  // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4670  // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4671  if (C == B) {
4672  if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4673  match(A, m_Not(m_Specific(CmpRHS)))))
4674  return {L.Flavor, SPNB_NA, false};
4675  }
4676  // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4677  // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4678  if (D == A) {
4679  if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4680  match(B, m_Not(m_Specific(CmpRHS)))))
4681  return {L.Flavor, SPNB_NA, false};
4682  }
4683  // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4684  // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4685  if (C == A) {
4686  if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4687  match(B, m_Not(m_Specific(CmpRHS)))))
4688  return {L.Flavor, SPNB_NA, false};
4689  }
4690 
4691  return {SPF_UNKNOWN, SPNB_NA, false};
4692 }
4693 
4694 /// Match non-obvious integer minimum and maximum sequences.
4696  Value *CmpLHS, Value *CmpRHS,
4697  Value *TrueVal, Value *FalseVal,
4698  Value *&LHS, Value *&RHS,
4699  unsigned Depth) {
4700  // Assume success. If there's no match, callers should not use these anyway.
4701  LHS = TrueVal;
4702  RHS = FalseVal;
4703 
4704  SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
4706  return SPR;
4707 
4708  SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
4710  return SPR;
4711 
4712  if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
4713  return {SPF_UNKNOWN, SPNB_NA, false};
4714 
4715  // Z = X -nsw Y
4716  // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
4717  // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
4718  if (match(TrueVal, m_Zero()) &&
4719  match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4720  return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4721 
4722  // Z = X -nsw Y
4723  // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
4724  // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
4725  if (match(FalseVal, m_Zero()) &&
4726  match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4727  return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4728 
4729  const APInt *C1;
4730  if (!match(CmpRHS, m_APInt(C1)))
4731  return {SPF_UNKNOWN, SPNB_NA, false};
4732 
4733  // An unsigned min/max can be written with a signed compare.
4734  const APInt *C2;
4735  if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
4736  (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
4737  // Is the sign bit set?
4738  // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
4739  // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
4740  if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
4741  C2->isMaxSignedValue())
4742  return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4743 
4744  // Is the sign bit clear?
4745  // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
4746  // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
4747  if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
4748  C2->isMinSignedValue())
4749  return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4750  }
4751 
4752  // Look through 'not' ops to find disguised signed min/max.
4753  // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4754  // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4755  if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4756  match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4757  return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4758 
4759  // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4760  // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4761  if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4762  match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4763  return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4764 
4765  return {SPF_UNKNOWN, SPNB_NA, false};
4766 }
4767 
4768 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
4769  assert(X && Y && "Invalid operand");
4770 
4771  // X = sub (0, Y) || X = sub nsw (0, Y)
4772  if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
4773  (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
4774  return true;
4775 
4776  // Y = sub (0, X) || Y = sub nsw (0, X)
4777  if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
4778  (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
4779  return true;
4780 
4781  // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
4782  Value *A, *B;
4783  return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
4784  match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
4785  (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
4786  match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
4787 }
4788 
4790  FastMathFlags FMF,
4791  Value *CmpLHS, Value *CmpRHS,
4792  Value *TrueVal, Value *FalseVal,
4793  Value *&LHS, Value *&RHS,
4794  unsigned Depth) {
4795  if (CmpInst::isFPPredicate(Pred)) {
4796  // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one
4797  // 0.0 operand, set the compare's 0.0 operands to that same value for the
4798  // purpose of identifying min/max. Disregard vector constants with undefined
4799  // elements because those can not be back-propagated for analysis.
4800  Value *OutputZeroVal = nullptr;
4801  if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) &&
4802  !cast<Constant>(TrueVal)->containsUndefElement())
4803  OutputZeroVal = TrueVal;
4804  else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) &&
4805  !cast<Constant>(FalseVal)->containsUndefElement())
4806  OutputZeroVal = FalseVal;
4807 
4808  if (OutputZeroVal) {
4809  if (match(CmpLHS, m_AnyZeroFP()))
4810  CmpLHS = OutputZeroVal;
4811  if (match(CmpRHS, m_AnyZeroFP()))
4812  CmpRHS = OutputZeroVal;
4813  }
4814  }
4815 
4816  LHS = CmpLHS;
4817  RHS = CmpRHS;
4818 
4819  // Signed zero may return inconsistent results between implementations.
4820  // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4821  // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4822  // Therefore, we behave conservatively and only proceed if at least one of the
4823  // operands is known to not be zero or if we don't care about signed zero.
4824  switch (Pred) {
4825  default: break;
4826  // FIXME: Include OGT/OLT/UGT/ULT.
4829  if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4830  !isKnownNonZero(CmpRHS))
4831  return {SPF_UNKNOWN, SPNB_NA, false};
4832  }
4833 
4834  SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4835  bool Ordered = false;
4836 
4837  // When given one NaN and one non-NaN input:
4838  // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4839  // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4840  // ordered comparison fails), which could be NaN or non-NaN.
4841  // so here we discover exactly what NaN behavior is required/accepted.
4842  if (CmpInst::isFPPredicate(Pred)) {
4843  bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4844  bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4845 
4846  if (LHSSafe && RHSSafe) {
4847  // Both operands are known non-NaN.
4848  NaNBehavior = SPNB_RETURNS_ANY;
4849  } else if (CmpInst::isOrdered(Pred)) {
4850  // An ordered comparison will return false when given a NaN, so it
4851  // returns the RHS.
4852  Ordered = true;
4853  if (LHSSafe)
4854  // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4855  NaNBehavior = SPNB_RETURNS_NAN;
4856  else if (RHSSafe)
4857  NaNBehavior = SPNB_RETURNS_OTHER;
4858  else
4859  // Completely unsafe.
4860  return {SPF_UNKNOWN, SPNB_NA, false};
4861  } else {
4862  Ordered = false;
4863  // An unordered comparison will return true when given a NaN, so it
4864  // returns the LHS.
4865  if (LHSSafe)
4866  // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4867  NaNBehavior = SPNB_RETURNS_OTHER;
4868  else if (RHSSafe)
4869  NaNBehavior = SPNB_RETURNS_NAN;
4870  else
4871  // Completely unsafe.
4872  return {SPF_UNKNOWN, SPNB_NA, false};
4873  }
4874  }
4875 
4876  if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4877  std::swap(CmpLHS, CmpRHS);
4878  Pred = CmpInst::getSwappedPredicate(Pred);
4879  if (NaNBehavior == SPNB_RETURNS_NAN)
4880  NaNBehavior = SPNB_RETURNS_OTHER;
4881  else if (NaNBehavior == SPNB_RETURNS_OTHER)
4882  NaNBehavior = SPNB_RETURNS_NAN;
4883  Ordered = !Ordered;
4884  }
4885 
4886  // ([if]cmp X, Y) ? X : Y
4887  if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4888  switch (Pred) {
4889  default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4890  case ICmpInst::ICMP_UGT:
4891  case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4892  case ICmpInst::ICMP_SGT:
4893  case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4894  case ICmpInst::ICMP_ULT:
4895  case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4896  case ICmpInst::ICMP_SLT:
4897  case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4898  case FCmpInst::FCMP_UGT:
4899  case FCmpInst::FCMP_UGE:
4900  case FCmpInst::FCMP_OGT:
4901  case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4902  case FCmpInst::FCMP_ULT:
4903  case FCmpInst::FCMP_ULE:
4904  case FCmpInst::FCMP_OLT:
4905  case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4906  }
4907  }
4908 
4909  if (isKnownNegation(TrueVal, FalseVal)) {
4910  // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
4911  // match against either LHS or sext(LHS).
4912  auto MaybeSExtCmpLHS =
4913  m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
4914  auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
4915  auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
4916  if (match(TrueVal, MaybeSExtCmpLHS)) {
4917  // Set the return values. If the compare uses the negated value (-X >s 0),
4918  // swap the return values because the negated value is always 'RHS'.
4919  LHS = TrueVal;
4920  RHS = FalseVal;
4921  if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
4922  std::swap(LHS, RHS);
4923 
4924  // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
4925  // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
4926  if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4927  return {SPF_ABS, SPNB_NA, false};
4928 
4929  // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
4930  // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
4931  if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4932  return {SPF_NABS, SPNB_NA, false};
4933  }
4934  else if (match(FalseVal, MaybeSExtCmpLHS)) {
4935  // Set the return values. If the compare uses the negated value (-X >s 0),
4936  // swap the return values because the negated value is always 'RHS'.
4937  LHS = FalseVal;
4938  RHS = TrueVal;
4939  if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
4940  std::swap(LHS, RHS);
4941 
4942  // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
4943  // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
4944  if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4945  return {SPF_NABS, SPNB_NA, false};
4946 
4947  // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
4948  // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
4949  if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4950  return {SPF_ABS, SPNB_NA, false};
4951  }
4952  }
4953 
4954  if (CmpInst::isIntPredicate(Pred))
4955  return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
4956 
4957  // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
4958  // may return either -0.0 or 0.0, so fcmp/select pair has stricter
4959  // semantics than minNum. Be conservative in such case.
4960  if (NaNBehavior != SPNB_RETURNS_ANY ||
4961  (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4962  !isKnownNonZero(CmpRHS)))
4963  return {SPF_UNKNOWN, SPNB_NA, false};
4964 
4965  return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4966 }
4967 
4968 /// Helps to match a select pattern in case of a type mismatch.
4969 ///
4970 /// The function processes the case when type of true and false values of a
4971 /// select instruction differs from type of the cmp instruction operands because
4972 /// of a cast instruction. The function checks if it is legal to move the cast
4973 /// operation after "select". If yes, it returns the new second value of
4974 /// "select" (with the assumption that cast is moved):
4975 /// 1. As operand of cast instruction when both values of "select" are same cast
4976 /// instructions.
4977 /// 2. As restored constant (by applying reverse cast operation) when the first
4978 /// value of the "select" is a cast operation and the second value is a
4979 /// constant.
4980 /// NOTE: We return only the new second value because the first value could be
4981 /// accessed as operand of cast instruction.
4982 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4983  Instruction::CastOps *CastOp) {
4984  auto *Cast1 = dyn_cast<CastInst>(V1);
4985  if (!Cast1)
4986  return nullptr;
4987 
4988  *CastOp = Cast1->getOpcode();
4989  Type *SrcTy = Cast1->getSrcTy();
4990  if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4991  // If V1 and V2 are both the same cast from the same type, look through V1.
4992  if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4993  return Cast2->getOperand(0);
4994  return nullptr;
4995  }
4996 
4997  auto *C = dyn_cast<Constant>(V2);
4998  if (!C)
4999  return nullptr;
5000 
5001  Constant *CastedTo = nullptr;
5002  switch (*CastOp) {
5003  case Instruction::ZExt:
5004  if (CmpI->isUnsigned())
5005  CastedTo = ConstantExpr::getTrunc(C, SrcTy);
5006  break;
5007  case Instruction::SExt:
5008  if (CmpI->isSigned())
5009  CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
5010  break;
5011  case Instruction::Trunc:
5012  Constant *CmpConst;
5013  if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
5014  CmpConst->getType() == SrcTy) {
5015  // Here we have the following case:
5016  //
5017  // %cond = cmp iN %x, CmpConst
5018  // %tr = trunc iN %x to iK
5019  // %narrowsel = select i1 %cond, iK %t, iK C
5020  //
5021  // We can always move trunc after select operation:
5022  //
5023  // %cond = cmp iN %x, CmpConst
5024  // %widesel = select i1 %cond, iN %x, iN CmpConst
5025  // %tr = trunc iN %widesel to iK
5026  //
5027  // Note that C could be extended in any way because we don't care about
5028  // upper bits after truncation. It can't be abs pattern, because it would
5029  // look like:
5030  //
5031  // select i1 %cond, x, -x.
5032  //
5033  // So only min/max pattern could be matched. Such match requires widened C
5034  // == CmpConst. That is why set widened C = CmpConst, condition trunc
5035  // CmpConst == C is checked below.
5036  CastedTo = CmpConst;
5037  } else {
5038  CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
5039  }
5040  break;
5041  case Instruction::FPTrunc:
5042  CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
5043  break;
5044  case Instruction::FPExt:
5045  CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
5046  break;
5047  case Instruction::FPToUI:
5048  CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
5049  break;
5050  case Instruction::FPToSI:
5051  CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
5052  break;
5053  case Instruction::UIToFP:
5054  CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
5055  break;
5056  case Instruction::SIToFP:
5057  CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
5058  break;
5059  default:
5060  break;
5061  }
5062 
5063  if (!CastedTo)
5064  return nullptr;
5065 
5066  // Make sure the cast doesn't lose any information.
5067  Constant *CastedBack =
5068  ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
5069  if (CastedBack != C)
5070  return nullptr;
5071 
5072  return CastedTo;
5073 }
5074 
5076  Instruction::CastOps *CastOp,
5077  unsigned Depth) {
5078  if (Depth >= MaxDepth)
5079  return {SPF_UNKNOWN, SPNB_NA, false};
5080 
5082  if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
5083 
5084  CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
5085  if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
5086 
5087  CmpInst::Predicate Pred = CmpI->getPredicate();
5088  Value *CmpLHS = CmpI->getOperand(0);
5089  Value *CmpRHS = CmpI->getOperand(1);
5090  Value *TrueVal = SI->getTrueValue();
5091  Value *FalseVal = SI->getFalseValue();
5092  FastMathFlags FMF;
5093  if (isa<FPMathOperator>(CmpI))
5094  FMF = CmpI->getFastMathFlags();
5095 
5096  // Bail out early.
5097  if (CmpI->isEquality())
5098  return {SPF_UNKNOWN, SPNB_NA, false};
5099 
5100  // Deal with type mismatches.
5101  if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
5102  if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
5103  // If this is a potential fmin/fmax with a cast to integer, then ignore
5104  // -0.0 because there is no corresponding integer value.
5105  if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5106  FMF.setNoSignedZeros();
5107  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5108  cast<CastInst>(TrueVal)->getOperand(0), C,
5109  LHS, RHS, Depth);
5110  }
5111  if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
5112  // If this is a potential fmin/fmax with a cast to integer, then ignore
5113  // -0.0 because there is no corresponding integer value.
5114  if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5115  FMF.setNoSignedZeros();
5116  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5117  C, cast<CastInst>(FalseVal)->getOperand(0),
5118  LHS, RHS, Depth);
5119  }
5120  }
5121  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
5122  LHS, RHS, Depth);
5123 }
5124 
5126  if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
5127  if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
5128  if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
5129  if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
5130  if (SPF == SPF_FMINNUM)
5131  return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
5132  if (SPF == SPF_FMAXNUM)
5133  return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
5134  llvm_unreachable("unhandled!");
5135 }
5136 
5138  if (SPF == SPF_SMIN) return SPF_SMAX;
5139  if (SPF == SPF_UMIN) return SPF_UMAX;
5140  if (SPF == SPF_SMAX) return SPF_SMIN;
5141  if (SPF == SPF_UMAX) return SPF_UMIN;
5142  llvm_unreachable("unhandled!");
5143 }
5144 
5147 }
5148 
5149 /// Return true if "icmp Pred LHS RHS" is always true.