LLVM  9.0.0svn
ValueTracking.cpp
Go to the documentation of this file.
1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains routines that help analyze properties that chains of
10 // computations have.
11 //
12 //===----------------------------------------------------------------------===//
13 
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/StringRef.h"
30 #include "llvm/Analysis/Loads.h"
31 #include "llvm/Analysis/LoopInfo.h"
34 #include "llvm/IR/Argument.h"
35 #include "llvm/IR/Attributes.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/ConstantRange.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
47 #include "llvm/IR/GlobalAlias.h"
48 #include "llvm/IR/GlobalValue.h"
49 #include "llvm/IR/GlobalVariable.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/LLVMContext.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/IR/Operator.h"
59 #include "llvm/IR/PatternMatch.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/User.h"
62 #include "llvm/IR/Value.h"
63 #include "llvm/Support/Casting.h"
65 #include "llvm/Support/Compiler.h"
67 #include "llvm/Support/KnownBits.h"
69 #include <algorithm>
70 #include <array>
71 #include <cassert>
72 #include <cstdint>
73 #include <iterator>
74 #include <utility>
75 
76 using namespace llvm;
77 using namespace llvm::PatternMatch;
78 
79 const unsigned MaxDepth = 6;
80 
81 // Controls the number of uses of the value searched for possible
82 // dominating comparisons.
83 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
84  cl::Hidden, cl::init(20));
85 
86 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
87 /// returns the element type's bitwidth.
88 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
89  if (unsigned BitWidth = Ty->getScalarSizeInBits())
90  return BitWidth;
91 
92  return DL.getIndexTypeSizeInBits(Ty);
93 }
94 
95 namespace {
96 
97 // Simplifying using an assume can only be done in a particular control-flow
98 // context (the context instruction provides that context). If an assume and
99 // the context instruction are not in the same block then the DT helps in
100 // figuring out if we can use it.
101 struct Query {
102  const DataLayout &DL;
103  AssumptionCache *AC;
104  const Instruction *CxtI;
105  const DominatorTree *DT;
106 
107  // Unlike the other analyses, this may be a nullptr because not all clients
108  // provide it currently.
110 
111  /// Set of assumptions that should be excluded from further queries.
112  /// This is because of the potential for mutual recursion to cause
113  /// computeKnownBits to repeatedly visit the same assume intrinsic. The
114  /// classic case of this is assume(x = y), which will attempt to determine
115  /// bits in x from bits in y, which will attempt to determine bits in y from
116  /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
117  /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo
118  /// (all of which can call computeKnownBits), and so on.
119  std::array<const Value *, MaxDepth> Excluded;
120 
121  /// If true, it is safe to use metadata during simplification.
122  InstrInfoQuery IIQ;
123 
124  unsigned NumExcluded = 0;
125 
126  Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
127  const DominatorTree *DT, bool UseInstrInfo,
128  OptimizationRemarkEmitter *ORE = nullptr)
129  : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {}
130 
131  Query(const Query &Q, const Value *NewExcl)
132  : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE), IIQ(Q.IIQ),
133  NumExcluded(Q.NumExcluded) {
134  Excluded = Q.Excluded;
135  Excluded[NumExcluded++] = NewExcl;
136  assert(NumExcluded <= Excluded.size());
137  }
138 
139  bool isExcluded(const Value *Value) const {
140  if (NumExcluded == 0)
141  return false;
142  auto End = Excluded.begin() + NumExcluded;
143  return std::find(Excluded.begin(), End, Value) != End;
144  }
145 };
146 
147 } // end anonymous namespace
148 
149 // Given the provided Value and, potentially, a context instruction, return
150 // the preferred context instruction (if any).
151 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
152  // If we've been provided with a context instruction, then use that (provided
153  // it has been inserted).
154  if (CxtI && CxtI->getParent())
155  return CxtI;
156 
157  // If the value is really an already-inserted instruction, then use that.
158  CxtI = dyn_cast<Instruction>(V);
159  if (CxtI && CxtI->getParent())
160  return CxtI;
161 
162  return nullptr;
163 }
164 
165 static void computeKnownBits(const Value *V, KnownBits &Known,
166  unsigned Depth, const Query &Q);
167 
168 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
169  const DataLayout &DL, unsigned Depth,
170  AssumptionCache *AC, const Instruction *CxtI,
171  const DominatorTree *DT,
172  OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
173  ::computeKnownBits(V, Known, Depth,
174  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
175 }
176 
177 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
178  const Query &Q);
179 
181  unsigned Depth, AssumptionCache *AC,
182  const Instruction *CxtI,
183  const DominatorTree *DT,
185  bool UseInstrInfo) {
187  V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
188 }
189 
190 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
191  const DataLayout &DL, AssumptionCache *AC,
192  const Instruction *CxtI, const DominatorTree *DT,
193  bool UseInstrInfo) {
194  assert(LHS->getType() == RHS->getType() &&
195  "LHS and RHS should have the same type");
196  assert(LHS->getType()->isIntOrIntVectorTy() &&
197  "LHS and RHS should be integers");
198  // Look for an inverted mask: (X & ~M) op (Y & M).
199  Value *M;
200  if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
201  match(RHS, m_c_And(m_Specific(M), m_Value())))
202  return true;
203  if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
204  match(LHS, m_c_And(m_Specific(M), m_Value())))
205  return true;
206  IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
207  KnownBits LHSKnown(IT->getBitWidth());
208  KnownBits RHSKnown(IT->getBitWidth());
209  computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
210  computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
211  return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
212 }
213 
215  for (const User *U : CxtI->users()) {
216  if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
217  if (IC->isEquality())
218  if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
219  if (C->isNullValue())
220  continue;
221  return false;
222  }
223  return true;
224 }
225 
226 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
227  const Query &Q);
228 
230  bool OrZero, unsigned Depth,
231  AssumptionCache *AC, const Instruction *CxtI,
232  const DominatorTree *DT, bool UseInstrInfo) {
234  V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
235 }
236 
237 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
238 
239 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
240  AssumptionCache *AC, const Instruction *CxtI,
241  const DominatorTree *DT, bool UseInstrInfo) {
242  return ::isKnownNonZero(V, Depth,
243  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
244 }
245 
246 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
247  unsigned Depth, AssumptionCache *AC,
248  const Instruction *CxtI, const DominatorTree *DT,
249  bool UseInstrInfo) {
250  KnownBits Known =
251  computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
252  return Known.isNonNegative();
253 }
254 
255 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
256  AssumptionCache *AC, const Instruction *CxtI,
257  const DominatorTree *DT, bool UseInstrInfo) {
258  if (auto *CI = dyn_cast<ConstantInt>(V))
259  return CI->getValue().isStrictlyPositive();
260 
261  // TODO: We'd doing two recursive queries here. We should factor this such
262  // that only a single query is needed.
263  return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) &&
264  isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo);
265 }
266 
267 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
268  AssumptionCache *AC, const Instruction *CxtI,
269  const DominatorTree *DT, bool UseInstrInfo) {
270  KnownBits Known =
271  computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
272  return Known.isNegative();
273 }
274 
275 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
276 
277 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
278  const DataLayout &DL, AssumptionCache *AC,
279  const Instruction *CxtI, const DominatorTree *DT,
280  bool UseInstrInfo) {
282  Query(DL, AC, safeCxtI(V1, safeCxtI(V2, CxtI)), DT,
283  UseInstrInfo, /*ORE=*/nullptr));
284 }
285 
286 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
287  const Query &Q);
288 
289 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
290  const DataLayout &DL, unsigned Depth,
291  AssumptionCache *AC, const Instruction *CxtI,
292  const DominatorTree *DT, bool UseInstrInfo) {
294  V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
295 }
296 
297 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
298  const Query &Q);
299 
300 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
301  unsigned Depth, AssumptionCache *AC,
302  const Instruction *CxtI,
303  const DominatorTree *DT, bool UseInstrInfo) {
305  V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
306 }
307 
308 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
309  bool NSW,
310  KnownBits &KnownOut, KnownBits &Known2,
311  unsigned Depth, const Query &Q) {
312  unsigned BitWidth = KnownOut.getBitWidth();
313 
314  // If an initial sequence of bits in the result is not needed, the
315  // corresponding bits in the operands are not needed.
316  KnownBits LHSKnown(BitWidth);
317  computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
318  computeKnownBits(Op1, Known2, Depth + 1, Q);
319 
320  KnownOut = KnownBits::computeForAddSub(Add, NSW, LHSKnown, Known2);
321 }
322 
323 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
324  KnownBits &Known, KnownBits &Known2,
325  unsigned Depth, const Query &Q) {
326  unsigned BitWidth = Known.getBitWidth();
327  computeKnownBits(Op1, Known, Depth + 1, Q);
328  computeKnownBits(Op0, Known2, Depth + 1, Q);
329 
330  bool isKnownNegative = false;
331  bool isKnownNonNegative = false;
332  // If the multiplication is known not to overflow, compute the sign bit.
333  if (NSW) {
334  if (Op0 == Op1) {
335  // The product of a number with itself is non-negative.
336  isKnownNonNegative = true;
337  } else {
338  bool isKnownNonNegativeOp1 = Known.isNonNegative();
339  bool isKnownNonNegativeOp0 = Known2.isNonNegative();
340  bool isKnownNegativeOp1 = Known.isNegative();
341  bool isKnownNegativeOp0 = Known2.isNegative();
342  // The product of two numbers with the same sign is non-negative.
343  isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
344  (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
345  // The product of a negative number and a non-negative number is either
346  // negative or zero.
347  if (!isKnownNonNegative)
348  isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
349  isKnownNonZero(Op0, Depth, Q)) ||
350  (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
351  isKnownNonZero(Op1, Depth, Q));
352  }
353  }
354 
355  assert(!Known.hasConflict() && !Known2.hasConflict());
356  // Compute a conservative estimate for high known-0 bits.
357  unsigned LeadZ = std::max(Known.countMinLeadingZeros() +
358  Known2.countMinLeadingZeros(),
359  BitWidth) - BitWidth;
360  LeadZ = std::min(LeadZ, BitWidth);
361 
362  // The result of the bottom bits of an integer multiply can be
363  // inferred by looking at the bottom bits of both operands and
364  // multiplying them together.
365  // We can infer at least the minimum number of known trailing bits
366  // of both operands. Depending on number of trailing zeros, we can
367  // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming
368  // a and b are divisible by m and n respectively.
369  // We then calculate how many of those bits are inferrable and set
370  // the output. For example, the i8 mul:
371  // a = XXXX1100 (12)
372  // b = XXXX1110 (14)
373  // We know the bottom 3 bits are zero since the first can be divided by
374  // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4).
375  // Applying the multiplication to the trimmed arguments gets:
376  // XX11 (3)
377  // X111 (7)
378  // -------
379  // XX11
380  // XX11
381  // XX11
382  // XX11
383  // -------
384  // XXXXX01
385  // Which allows us to infer the 2 LSBs. Since we're multiplying the result
386  // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits.
387  // The proof for this can be described as:
388  // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) &&
389  // (C7 == (1 << (umin(countTrailingZeros(C1), C5) +
390  // umin(countTrailingZeros(C2), C6) +
391  // umin(C5 - umin(countTrailingZeros(C1), C5),
392  // C6 - umin(countTrailingZeros(C2), C6)))) - 1)
393  // %aa = shl i8 %a, C5
394  // %bb = shl i8 %b, C6
395  // %aaa = or i8 %aa, C1
396  // %bbb = or i8 %bb, C2
397  // %mul = mul i8 %aaa, %bbb
398  // %mask = and i8 %mul, C7
399  // =>
400  // %mask = i8 ((C1*C2)&C7)
401  // Where C5, C6 describe the known bits of %a, %b
402  // C1, C2 describe the known bottom bits of %a, %b.
403  // C7 describes the mask of the known bits of the result.
404  APInt Bottom0 = Known.One;
405  APInt Bottom1 = Known2.One;
406 
407  // How many times we'd be able to divide each argument by 2 (shr by 1).
408  // This gives us the number of trailing zeros on the multiplication result.
409  unsigned TrailBitsKnown0 = (Known.Zero | Known.One).countTrailingOnes();
410  unsigned TrailBitsKnown1 = (Known2.Zero | Known2.One).countTrailingOnes();
411  unsigned TrailZero0 = Known.countMinTrailingZeros();
412  unsigned TrailZero1 = Known2.countMinTrailingZeros();
413  unsigned TrailZ = TrailZero0 + TrailZero1;
414 
415  // Figure out the fewest known-bits operand.
416  unsigned SmallestOperand = std::min(TrailBitsKnown0 - TrailZero0,
417  TrailBitsKnown1 - TrailZero1);
418  unsigned ResultBitsKnown = std::min(SmallestOperand + TrailZ, BitWidth);
419 
420  APInt BottomKnown = Bottom0.getLoBits(TrailBitsKnown0) *
421  Bottom1.getLoBits(TrailBitsKnown1);
422 
423  Known.resetAll();
424  Known.Zero.setHighBits(LeadZ);
425  Known.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
426  Known.One |= BottomKnown.getLoBits(ResultBitsKnown);
427 
428  // Only make use of no-wrap flags if we failed to compute the sign bit
429  // directly. This matters if the multiplication always overflows, in
430  // which case we prefer to follow the result of the direct computation,
431  // though as the program is invoking undefined behaviour we can choose
432  // whatever we like here.
433  if (isKnownNonNegative && !Known.isNegative())
434  Known.makeNonNegative();
435  else if (isKnownNegative && !Known.isNonNegative())
436  Known.makeNegative();
437 }
438 
440  KnownBits &Known) {
441  unsigned BitWidth = Known.getBitWidth();
442  unsigned NumRanges = Ranges.getNumOperands() / 2;
443  assert(NumRanges >= 1);
444 
445  Known.Zero.setAllBits();
446  Known.One.setAllBits();
447 
448  for (unsigned i = 0; i < NumRanges; ++i) {
449  ConstantInt *Lower =
450  mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
451  ConstantInt *Upper =
452  mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
453  ConstantRange Range(Lower->getValue(), Upper->getValue());
454 
455  // The first CommonPrefixBits of all values in Range are equal.
456  unsigned CommonPrefixBits =
457  (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
458 
459  APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
460  Known.One &= Range.getUnsignedMax() & Mask;
461  Known.Zero &= ~Range.getUnsignedMax() & Mask;
462  }
463 }
464 
465 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
466  SmallVector<const Value *, 16> WorkSet(1, I);
469 
470  // The instruction defining an assumption's condition itself is always
471  // considered ephemeral to that assumption (even if it has other
472  // non-ephemeral users). See r246696's test case for an example.
473  if (is_contained(I->operands(), E))
474  return true;
475 
476  while (!WorkSet.empty()) {
477  const Value *V = WorkSet.pop_back_val();
478  if (!Visited.insert(V).second)
479  continue;
480 
481  // If all uses of this value are ephemeral, then so is this value.
482  if (llvm::all_of(V->users(), [&](const User *U) {
483  return EphValues.count(U);
484  })) {
485  if (V == E)
486  return true;
487 
488  if (V == I || isSafeToSpeculativelyExecute(V)) {
489  EphValues.insert(V);
490  if (const User *U = dyn_cast<User>(V))
491  for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
492  J != JE; ++J)
493  WorkSet.push_back(*J);
494  }
495  }
496  }
497 
498  return false;
499 }
500 
501 // Is this an intrinsic that cannot be speculated but also cannot trap?
503  if (const CallInst *CI = dyn_cast<CallInst>(I))
504  if (Function *F = CI->getCalledFunction())
505  switch (F->getIntrinsicID()) {
506  default: break;
507  // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
508  case Intrinsic::assume:
509  case Intrinsic::sideeffect:
510  case Intrinsic::dbg_declare:
511  case Intrinsic::dbg_value:
512  case Intrinsic::dbg_label:
513  case Intrinsic::invariant_start:
514  case Intrinsic::invariant_end:
515  case Intrinsic::lifetime_start:
516  case Intrinsic::lifetime_end:
517  case Intrinsic::objectsize:
518  case Intrinsic::ptr_annotation:
519  case Intrinsic::var_annotation:
520  return true;
521  }
522 
523  return false;
524 }
525 
527  const Instruction *CxtI,
528  const DominatorTree *DT) {
529  // There are two restrictions on the use of an assume:
530  // 1. The assume must dominate the context (or the control flow must
531  // reach the assume whenever it reaches the context).
532  // 2. The context must not be in the assume's set of ephemeral values
533  // (otherwise we will use the assume to prove that the condition
534  // feeding the assume is trivially true, thus causing the removal of
535  // the assume).
536 
537  if (DT) {
538  if (DT->dominates(Inv, CxtI))
539  return true;
540  } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
541  // We don't have a DT, but this trivially dominates.
542  return true;
543  }
544 
545  // With or without a DT, the only remaining case we will check is if the
546  // instructions are in the same BB. Give up if that is not the case.
547  if (Inv->getParent() != CxtI->getParent())
548  return false;
549 
550  // If we have a dom tree, then we now know that the assume doesn't dominate
551  // the other instruction. If we don't have a dom tree then we can check if
552  // the assume is first in the BB.
553  if (!DT) {
554  // Search forward from the assume until we reach the context (or the end
555  // of the block); the common case is that the assume will come first.
556  for (auto I = std::next(BasicBlock::const_iterator(Inv)),
557  IE = Inv->getParent()->end(); I != IE; ++I)
558  if (&*I == CxtI)
559  return true;
560  }
561 
562  // The context comes first, but they're both in the same block. Make sure
563  // there is nothing in between that might interrupt the control flow.
565  std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
566  I != IE; ++I)
568  return false;
569 
570  return !isEphemeralValueOf(Inv, CxtI);
571 }
572 
573 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
574  unsigned Depth, const Query &Q) {
575  // Use of assumptions is context-sensitive. If we don't have a context, we
576  // cannot use them!
577  if (!Q.AC || !Q.CxtI)
578  return;
579 
580  unsigned BitWidth = Known.getBitWidth();
581 
582  // Note that the patterns below need to be kept in sync with the code
583  // in AssumptionCache::updateAffectedValues.
584 
585  for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
586  if (!AssumeVH)
587  continue;
588  CallInst *I = cast<CallInst>(AssumeVH);
589  assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
590  "Got assumption for the wrong function!");
591  if (Q.isExcluded(I))
592  continue;
593 
594  // Warning: This loop can end up being somewhat performance sensitive.
595  // We're running this loop for once for each value queried resulting in a
596  // runtime of ~O(#assumes * #values).
597 
598  assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
599  "must be an assume intrinsic");
600 
601  Value *Arg = I->getArgOperand(0);
602 
603  if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
604  assert(BitWidth == 1 && "assume operand is not i1?");
605  Known.setAllOnes();
606  return;
607  }
608  if (match(Arg, m_Not(m_Specific(V))) &&
609  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
610  assert(BitWidth == 1 && "assume operand is not i1?");
611  Known.setAllZero();
612  return;
613  }
614 
615  // The remaining tests are all recursive, so bail out if we hit the limit.
616  if (Depth == MaxDepth)
617  continue;
618 
619  Value *A, *B;
620  auto m_V = m_CombineOr(m_Specific(V),
622  m_BitCast(m_Specific(V))));
623 
624  CmpInst::Predicate Pred;
625  uint64_t C;
626  // assume(v = a)
627  if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
628  Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
629  KnownBits RHSKnown(BitWidth);
630  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
631  Known.Zero |= RHSKnown.Zero;
632  Known.One |= RHSKnown.One;
633  // assume(v & b = a)
634  } else if (match(Arg,
635  m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
636  Pred == ICmpInst::ICMP_EQ &&
637  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
638  KnownBits RHSKnown(BitWidth);
639  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
640  KnownBits MaskKnown(BitWidth);
641  computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
642 
643  // For those bits in the mask that are known to be one, we can propagate
644  // known bits from the RHS to V.
645  Known.Zero |= RHSKnown.Zero & MaskKnown.One;
646  Known.One |= RHSKnown.One & MaskKnown.One;
647  // assume(~(v & b) = a)
648  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
649  m_Value(A))) &&
650  Pred == ICmpInst::ICMP_EQ &&
651  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
652  KnownBits RHSKnown(BitWidth);
653  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
654  KnownBits MaskKnown(BitWidth);
655  computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
656 
657  // For those bits in the mask that are known to be one, we can propagate
658  // inverted known bits from the RHS to V.
659  Known.Zero |= RHSKnown.One & MaskKnown.One;
660  Known.One |= RHSKnown.Zero & MaskKnown.One;
661  // assume(v | b = a)
662  } else if (match(Arg,
663  m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
664  Pred == ICmpInst::ICMP_EQ &&
665  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
666  KnownBits RHSKnown(BitWidth);
667  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
668  KnownBits BKnown(BitWidth);
669  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
670 
671  // For those bits in B that are known to be zero, we can propagate known
672  // bits from the RHS to V.
673  Known.Zero |= RHSKnown.Zero & BKnown.Zero;
674  Known.One |= RHSKnown.One & BKnown.Zero;
675  // assume(~(v | b) = a)
676  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
677  m_Value(A))) &&
678  Pred == ICmpInst::ICMP_EQ &&
679  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
680  KnownBits RHSKnown(BitWidth);
681  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
682  KnownBits BKnown(BitWidth);
683  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
684 
685  // For those bits in B that are known to be zero, we can propagate
686  // inverted known bits from the RHS to V.
687  Known.Zero |= RHSKnown.One & BKnown.Zero;
688  Known.One |= RHSKnown.Zero & BKnown.Zero;
689  // assume(v ^ b = a)
690  } else if (match(Arg,
691  m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
692  Pred == ICmpInst::ICMP_EQ &&
693  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
694  KnownBits RHSKnown(BitWidth);
695  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
696  KnownBits BKnown(BitWidth);
697  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
698 
699  // For those bits in B that are known to be zero, we can propagate known
700  // bits from the RHS to V. For those bits in B that are known to be one,
701  // we can propagate inverted known bits from the RHS to V.
702  Known.Zero |= RHSKnown.Zero & BKnown.Zero;
703  Known.One |= RHSKnown.One & BKnown.Zero;
704  Known.Zero |= RHSKnown.One & BKnown.One;
705  Known.One |= RHSKnown.Zero & BKnown.One;
706  // assume(~(v ^ b) = a)
707  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
708  m_Value(A))) &&
709  Pred == ICmpInst::ICMP_EQ &&
710  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
711  KnownBits RHSKnown(BitWidth);
712  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
713  KnownBits BKnown(BitWidth);
714  computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
715 
716  // For those bits in B that are known to be zero, we can propagate
717  // inverted known bits from the RHS to V. For those bits in B that are
718  // known to be one, we can propagate known bits from the RHS to V.
719  Known.Zero |= RHSKnown.One & BKnown.Zero;
720  Known.One |= RHSKnown.Zero & BKnown.Zero;
721  Known.Zero |= RHSKnown.Zero & BKnown.One;
722  Known.One |= RHSKnown.One & BKnown.One;
723  // assume(v << c = a)
724  } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
725  m_Value(A))) &&
726  Pred == ICmpInst::ICMP_EQ &&
727  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
728  C < BitWidth) {
729  KnownBits RHSKnown(BitWidth);
730  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
731  // For those bits in RHS that are known, we can propagate them to known
732  // bits in V shifted to the right by C.
733  RHSKnown.Zero.lshrInPlace(C);
734  Known.Zero |= RHSKnown.Zero;
735  RHSKnown.One.lshrInPlace(C);
736  Known.One |= RHSKnown.One;
737  // assume(~(v << c) = a)
738  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
739  m_Value(A))) &&
740  Pred == ICmpInst::ICMP_EQ &&
741  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
742  C < BitWidth) {
743  KnownBits RHSKnown(BitWidth);
744  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
745  // For those bits in RHS that are known, we can propagate them inverted
746  // to known bits in V shifted to the right by C.
747  RHSKnown.One.lshrInPlace(C);
748  Known.Zero |= RHSKnown.One;
749  RHSKnown.Zero.lshrInPlace(C);
750  Known.One |= RHSKnown.Zero;
751  // assume(v >> c = a)
752  } else if (match(Arg,
753  m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
754  m_Value(A))) &&
755  Pred == ICmpInst::ICMP_EQ &&
756  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
757  C < BitWidth) {
758  KnownBits RHSKnown(BitWidth);
759  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
760  // For those bits in RHS that are known, we can propagate them to known
761  // bits in V shifted to the right by C.
762  Known.Zero |= RHSKnown.Zero << C;
763  Known.One |= RHSKnown.One << C;
764  // assume(~(v >> c) = a)
765  } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
766  m_Value(A))) &&
767  Pred == ICmpInst::ICMP_EQ &&
768  isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
769  C < BitWidth) {
770  KnownBits RHSKnown(BitWidth);
771  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
772  // For those bits in RHS that are known, we can propagate them inverted
773  // to known bits in V shifted to the right by C.
774  Known.Zero |= RHSKnown.One << C;
775  Known.One |= RHSKnown.Zero << C;
776  // assume(v >=_s c) where c is non-negative
777  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
778  Pred == ICmpInst::ICMP_SGE &&
779  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
780  KnownBits RHSKnown(BitWidth);
781  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
782 
783  if (RHSKnown.isNonNegative()) {
784  // We know that the sign bit is zero.
785  Known.makeNonNegative();
786  }
787  // assume(v >_s c) where c is at least -1.
788  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
789  Pred == ICmpInst::ICMP_SGT &&
790  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
791  KnownBits RHSKnown(BitWidth);
792  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
793 
794  if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
795  // We know that the sign bit is zero.
796  Known.makeNonNegative();
797  }
798  // assume(v <=_s c) where c is negative
799  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
800  Pred == ICmpInst::ICMP_SLE &&
801  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
802  KnownBits RHSKnown(BitWidth);
803  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
804 
805  if (RHSKnown.isNegative()) {
806  // We know that the sign bit is one.
807  Known.makeNegative();
808  }
809  // assume(v <_s c) where c is non-positive
810  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
811  Pred == ICmpInst::ICMP_SLT &&
812  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
813  KnownBits RHSKnown(BitWidth);
814  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
815 
816  if (RHSKnown.isZero() || RHSKnown.isNegative()) {
817  // We know that the sign bit is one.
818  Known.makeNegative();
819  }
820  // assume(v <=_u c)
821  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
822  Pred == ICmpInst::ICMP_ULE &&
823  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
824  KnownBits RHSKnown(BitWidth);
825  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
826 
827  // Whatever high bits in c are zero are known to be zero.
828  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
829  // assume(v <_u c)
830  } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
831  Pred == ICmpInst::ICMP_ULT &&
832  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
833  KnownBits RHSKnown(BitWidth);
834  computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
835 
836  // If the RHS is known zero, then this assumption must be wrong (nothing
837  // is unsigned less than zero). Signal a conflict and get out of here.
838  if (RHSKnown.isZero()) {
839  Known.Zero.setAllBits();
840  Known.One.setAllBits();
841  break;
842  }
843 
844  // Whatever high bits in c are zero are known to be zero (if c is a power
845  // of 2, then one more).
846  if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
847  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
848  else
849  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
850  }
851  }
852 
853  // If assumptions conflict with each other or previous known bits, then we
854  // have a logical fallacy. It's possible that the assumption is not reachable,
855  // so this isn't a real bug. On the other hand, the program may have undefined
856  // behavior, or we might have a bug in the compiler. We can't assert/crash, so
857  // clear out the known bits, try to warn the user, and hope for the best.
858  if (Known.Zero.intersects(Known.One)) {
859  Known.resetAll();
860 
861  if (Q.ORE)
862  Q.ORE->emit([&]() {
863  auto *CxtI = const_cast<Instruction *>(Q.CxtI);
864  return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
865  CxtI)
866  << "Detected conflicting code assumptions. Program may "
867  "have undefined behavior, or compiler may have "
868  "internal error.";
869  });
870  }
871 }
872 
873 /// Compute known bits from a shift operator, including those with a
874 /// non-constant shift amount. Known is the output of this function. Known2 is a
875 /// pre-allocated temporary with the same bit width as Known. KZF and KOF are
876 /// operator-specific functions that, given the known-zero or known-one bits
877 /// respectively, and a shift amount, compute the implied known-zero or
878 /// known-one bits of the shift operator's result respectively for that shift
879 /// amount. The results from calling KZF and KOF are conservatively combined for
880 /// all permitted shift amounts.
882  const Operator *I, KnownBits &Known, KnownBits &Known2,
883  unsigned Depth, const Query &Q,
884  function_ref<APInt(const APInt &, unsigned)> KZF,
885  function_ref<APInt(const APInt &, unsigned)> KOF) {
886  unsigned BitWidth = Known.getBitWidth();
887 
888  if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
889  unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
890 
891  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
892  Known.Zero = KZF(Known.Zero, ShiftAmt);
893  Known.One = KOF(Known.One, ShiftAmt);
894  // If the known bits conflict, this must be an overflowing left shift, so
895  // the shift result is poison. We can return anything we want. Choose 0 for
896  // the best folding opportunity.
897  if (Known.hasConflict())
898  Known.setAllZero();
899 
900  return;
901  }
902 
903  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
904 
905  // If the shift amount could be greater than or equal to the bit-width of the
906  // LHS, the value could be poison, but bail out because the check below is
907  // expensive. TODO: Should we just carry on?
908  if ((~Known.Zero).uge(BitWidth)) {
909  Known.resetAll();
910  return;
911  }
912 
913  // Note: We cannot use Known.Zero.getLimitedValue() here, because if
914  // BitWidth > 64 and any upper bits are known, we'll end up returning the
915  // limit value (which implies all bits are known).
916  uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
917  uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
918 
919  // It would be more-clearly correct to use the two temporaries for this
920  // calculation. Reusing the APInts here to prevent unnecessary allocations.
921  Known.resetAll();
922 
923  // If we know the shifter operand is nonzero, we can sometimes infer more
924  // known bits. However this is expensive to compute, so be lazy about it and
925  // only compute it when absolutely necessary.
926  Optional<bool> ShifterOperandIsNonZero;
927 
928  // Early exit if we can't constrain any well-defined shift amount.
929  if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
930  !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
931  ShifterOperandIsNonZero = isKnownNonZero(I->getOperand(1), Depth + 1, Q);
932  if (!*ShifterOperandIsNonZero)
933  return;
934  }
935 
936  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
937 
938  Known.Zero.setAllBits();
939  Known.One.setAllBits();
940  for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
941  // Combine the shifted known input bits only for those shift amounts
942  // compatible with its known constraints.
943  if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
944  continue;
945  if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
946  continue;
947  // If we know the shifter is nonzero, we may be able to infer more known
948  // bits. This check is sunk down as far as possible to avoid the expensive
949  // call to isKnownNonZero if the cheaper checks above fail.
950  if (ShiftAmt == 0) {
951  if (!ShifterOperandIsNonZero.hasValue())
952  ShifterOperandIsNonZero =
953  isKnownNonZero(I->getOperand(1), Depth + 1, Q);
954  if (*ShifterOperandIsNonZero)
955  continue;
956  }
957 
958  Known.Zero &= KZF(Known2.Zero, ShiftAmt);
959  Known.One &= KOF(Known2.One, ShiftAmt);
960  }
961 
962  // If the known bits conflict, the result is poison. Return a 0 and hope the
963  // caller can further optimize that.
964  if (Known.hasConflict())
965  Known.setAllZero();
966 }
967 
968 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
969  unsigned Depth, const Query &Q) {
970  unsigned BitWidth = Known.getBitWidth();
971 
972  KnownBits Known2(Known);
973  switch (I->getOpcode()) {
974  default: break;
975  case Instruction::Load:
976  if (MDNode *MD =
977  Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range))
979  break;
980  case Instruction::And: {
981  // If either the LHS or the RHS are Zero, the result is zero.
982  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
983  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
984 
985  // Output known-1 bits are only known if set in both the LHS & RHS.
986  Known.One &= Known2.One;
987  // Output known-0 are known to be clear if zero in either the LHS | RHS.
988  Known.Zero |= Known2.Zero;
989 
990  // and(x, add (x, -1)) is a common idiom that always clears the low bit;
991  // here we handle the more general case of adding any odd number by
992  // matching the form add(x, add(x, y)) where y is odd.
993  // TODO: This could be generalized to clearing any bit set in y where the
994  // following bit is known to be unset in y.
995  Value *X = nullptr, *Y = nullptr;
996  if (!Known.Zero[0] && !Known.One[0] &&
997  match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) {
998  Known2.resetAll();
999  computeKnownBits(Y, Known2, Depth + 1, Q);
1000  if (Known2.countMinTrailingOnes() > 0)
1001  Known.Zero.setBit(0);
1002  }
1003  break;
1004  }
1005  case Instruction::Or:
1006  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1007  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1008 
1009  // Output known-0 bits are only known if clear in both the LHS & RHS.
1010  Known.Zero &= Known2.Zero;
1011  // Output known-1 are known to be set if set in either the LHS | RHS.
1012  Known.One |= Known2.One;
1013  break;
1014  case Instruction::Xor: {
1015  computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1016  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1017 
1018  // Output known-0 bits are known if clear or set in both the LHS & RHS.
1019  APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
1020  // Output known-1 are known to be set if set in only one of the LHS, RHS.
1021  Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
1022  Known.Zero = std::move(KnownZeroOut);
1023  break;
1024  }
1025  case Instruction::Mul: {
1026  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1027  computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
1028  Known2, Depth, Q);
1029  break;
1030  }
1031  case Instruction::UDiv: {
1032  // For the purposes of computing leading zeros we can conservatively
1033  // treat a udiv as a logical right shift by the power of 2 known to
1034  // be less than the denominator.
1035  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1036  unsigned LeadZ = Known2.countMinLeadingZeros();
1037 
1038  Known2.resetAll();
1039  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1040  unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
1041  if (RHSMaxLeadingZeros != BitWidth)
1042  LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
1043 
1044  Known.Zero.setHighBits(LeadZ);
1045  break;
1046  }
1047  case Instruction::Select: {
1048  const Value *LHS, *RHS;
1049  SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
1051  computeKnownBits(RHS, Known, Depth + 1, Q);
1052  computeKnownBits(LHS, Known2, Depth + 1, Q);
1053  } else {
1054  computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1055  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1056  }
1057 
1058  unsigned MaxHighOnes = 0;
1059  unsigned MaxHighZeros = 0;
1060  if (SPF == SPF_SMAX) {
1061  // If both sides are negative, the result is negative.
1062  if (Known.isNegative() && Known2.isNegative())
1063  // We can derive a lower bound on the result by taking the max of the
1064  // leading one bits.
1065  MaxHighOnes =
1067  // If either side is non-negative, the result is non-negative.
1068  else if (Known.isNonNegative() || Known2.isNonNegative())
1069  MaxHighZeros = 1;
1070  } else if (SPF == SPF_SMIN) {
1071  // If both sides are non-negative, the result is non-negative.
1072  if (Known.isNonNegative() && Known2.isNonNegative())
1073  // We can derive an upper bound on the result by taking the max of the
1074  // leading zero bits.
1075  MaxHighZeros = std::max(Known.countMinLeadingZeros(),
1076  Known2.countMinLeadingZeros());
1077  // If either side is negative, the result is negative.
1078  else if (Known.isNegative() || Known2.isNegative())
1079  MaxHighOnes = 1;
1080  } else if (SPF == SPF_UMAX) {
1081  // We can derive a lower bound on the result by taking the max of the
1082  // leading one bits.
1083  MaxHighOnes =
1085  } else if (SPF == SPF_UMIN) {
1086  // We can derive an upper bound on the result by taking the max of the
1087  // leading zero bits.
1088  MaxHighZeros =
1090  } else if (SPF == SPF_ABS) {
1091  // RHS from matchSelectPattern returns the negation part of abs pattern.
1092  // If the negate has an NSW flag we can assume the sign bit of the result
1093  // will be 0 because that makes abs(INT_MIN) undefined.
1094  if (Q.IIQ.hasNoSignedWrap(cast<Instruction>(RHS)))
1095  MaxHighZeros = 1;
1096  }
1097 
1098  // Only known if known in both the LHS and RHS.
1099  Known.One &= Known2.One;
1100  Known.Zero &= Known2.Zero;
1101  if (MaxHighOnes > 0)
1102  Known.One.setHighBits(MaxHighOnes);
1103  if (MaxHighZeros > 0)
1104  Known.Zero.setHighBits(MaxHighZeros);
1105  break;
1106  }
1107  case Instruction::FPTrunc:
1108  case Instruction::FPExt:
1109  case Instruction::FPToUI:
1110  case Instruction::FPToSI:
1111  case Instruction::SIToFP:
1112  case Instruction::UIToFP:
1113  break; // Can't work with floating point.
1114  case Instruction::PtrToInt:
1115  case Instruction::IntToPtr:
1116  // Fall through and handle them the same as zext/trunc.
1118  case Instruction::ZExt:
1119  case Instruction::Trunc: {
1120  Type *SrcTy = I->getOperand(0)->getType();
1121 
1122  unsigned SrcBitWidth;
1123  // Note that we handle pointer operands here because of inttoptr/ptrtoint
1124  // which fall through here.
1125  Type *ScalarTy = SrcTy->getScalarType();
1126  SrcBitWidth = ScalarTy->isPointerTy() ?
1127  Q.DL.getIndexTypeSizeInBits(ScalarTy) :
1128  Q.DL.getTypeSizeInBits(ScalarTy);
1129 
1130  assert(SrcBitWidth && "SrcBitWidth can't be zero");
1131  Known = Known.zextOrTrunc(SrcBitWidth);
1132  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1133  Known = Known.zextOrTrunc(BitWidth);
1134  // Any top bits are known to be zero.
1135  if (BitWidth > SrcBitWidth)
1136  Known.Zero.setBitsFrom(SrcBitWidth);
1137  break;
1138  }
1139  case Instruction::BitCast: {
1140  Type *SrcTy = I->getOperand(0)->getType();
1141  if (SrcTy->isIntOrPtrTy() &&
1142  // TODO: For now, not handling conversions like:
1143  // (bitcast i64 %x to <2 x i32>)
1144  !I->getType()->isVectorTy()) {
1145  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1146  break;
1147  }
1148  break;
1149  }
1150  case Instruction::SExt: {
1151  // Compute the bits in the result that are not present in the input.
1152  unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1153 
1154  Known = Known.trunc(SrcBitWidth);
1155  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1156  // If the sign bit of the input is known set or clear, then we know the
1157  // top bits of the result.
1158  Known = Known.sext(BitWidth);
1159  break;
1160  }
1161  case Instruction::Shl: {
1162  // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1163  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1164  auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1165  APInt KZResult = KnownZero << ShiftAmt;
1166  KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1167  // If this shift has "nsw" keyword, then the result is either a poison
1168  // value or has the same sign bit as the first operand.
1169  if (NSW && KnownZero.isSignBitSet())
1170  KZResult.setSignBit();
1171  return KZResult;
1172  };
1173 
1174  auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1175  APInt KOResult = KnownOne << ShiftAmt;
1176  if (NSW && KnownOne.isSignBitSet())
1177  KOResult.setSignBit();
1178  return KOResult;
1179  };
1180 
1181  computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1182  break;
1183  }
1184  case Instruction::LShr: {
1185  // (lshr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1186  auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1187  APInt KZResult = KnownZero.lshr(ShiftAmt);
1188  // High bits known zero.
1189  KZResult.setHighBits(ShiftAmt);
1190  return KZResult;
1191  };
1192 
1193  auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1194  return KnownOne.lshr(ShiftAmt);
1195  };
1196 
1197  computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1198  break;
1199  }
1200  case Instruction::AShr: {
1201  // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1202  auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1203  return KnownZero.ashr(ShiftAmt);
1204  };
1205 
1206  auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1207  return KnownOne.ashr(ShiftAmt);
1208  };
1209 
1210  computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1211  break;
1212  }
1213  case Instruction::Sub: {
1214  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1215  computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1216  Known, Known2, Depth, Q);
1217  break;
1218  }
1219  case Instruction::Add: {
1220  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1221  computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1222  Known, Known2, Depth, Q);
1223  break;
1224  }
1225  case Instruction::SRem:
1226  if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1227  APInt RA = Rem->getValue().abs();
1228  if (RA.isPowerOf2()) {
1229  APInt LowBits = RA - 1;
1230  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1231 
1232  // The low bits of the first operand are unchanged by the srem.
1233  Known.Zero = Known2.Zero & LowBits;
1234  Known.One = Known2.One & LowBits;
1235 
1236  // If the first operand is non-negative or has all low bits zero, then
1237  // the upper bits are all zero.
1238  if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1239  Known.Zero |= ~LowBits;
1240 
1241  // If the first operand is negative and not all low bits are zero, then
1242  // the upper bits are all one.
1243  if (Known2.isNegative() && LowBits.intersects(Known2.One))
1244  Known.One |= ~LowBits;
1245 
1246  assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1247  break;
1248  }
1249  }
1250 
1251  // The sign bit is the LHS's sign bit, except when the result of the
1252  // remainder is zero.
1253  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1254  // If it's known zero, our sign bit is also zero.
1255  if (Known2.isNonNegative())
1256  Known.makeNonNegative();
1257 
1258  break;
1259  case Instruction::URem: {
1260  if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1261  const APInt &RA = Rem->getValue();
1262  if (RA.isPowerOf2()) {
1263  APInt LowBits = (RA - 1);
1264  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1265  Known.Zero |= ~LowBits;
1266  Known.One &= LowBits;
1267  break;
1268  }
1269  }
1270 
1271  // Since the result is less than or equal to either operand, any leading
1272  // zero bits in either operand must also exist in the result.
1273  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1274  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1275 
1276  unsigned Leaders =
1278  Known.resetAll();
1279  Known.Zero.setHighBits(Leaders);
1280  break;
1281  }
1282 
1283  case Instruction::Alloca: {
1284  const AllocaInst *AI = cast<AllocaInst>(I);
1285  unsigned Align = AI->getAlignment();
1286  if (Align == 0)
1287  Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1288 
1289  if (Align > 0)
1290  Known.Zero.setLowBits(countTrailingZeros(Align));
1291  break;
1292  }
1293  case Instruction::GetElementPtr: {
1294  // Analyze all of the subscripts of this getelementptr instruction
1295  // to determine if we can prove known low zero bits.
1296  KnownBits LocalKnown(BitWidth);
1297  computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1298  unsigned TrailZ = LocalKnown.countMinTrailingZeros();
1299 
1301  for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1302  Value *Index = I->getOperand(i);
1303  if (StructType *STy = GTI.getStructTypeOrNull()) {
1304  // Handle struct member offset arithmetic.
1305 
1306  // Handle case when index is vector zeroinitializer
1307  Constant *CIndex = cast<Constant>(Index);
1308  if (CIndex->isZeroValue())
1309  continue;
1310 
1311  if (CIndex->getType()->isVectorTy())
1312  Index = CIndex->getSplatValue();
1313 
1314  unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1315  const StructLayout *SL = Q.DL.getStructLayout(STy);
1316  uint64_t Offset = SL->getElementOffset(Idx);
1317  TrailZ = std::min<unsigned>(TrailZ,
1318  countTrailingZeros(Offset));
1319  } else {
1320  // Handle array index arithmetic.
1321  Type *IndexedTy = GTI.getIndexedType();
1322  if (!IndexedTy->isSized()) {
1323  TrailZ = 0;
1324  break;
1325  }
1326  unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1327  uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1328  LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1329  computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1330  TrailZ = std::min(TrailZ,
1331  unsigned(countTrailingZeros(TypeSize) +
1332  LocalKnown.countMinTrailingZeros()));
1333  }
1334  }
1335 
1336  Known.Zero.setLowBits(TrailZ);
1337  break;
1338  }
1339  case Instruction::PHI: {
1340  const PHINode *P = cast<PHINode>(I);
1341  // Handle the case of a simple two-predecessor recurrence PHI.
1342  // There's a lot more that could theoretically be done here, but
1343  // this is sufficient to catch some interesting cases.
1344  if (P->getNumIncomingValues() == 2) {
1345  for (unsigned i = 0; i != 2; ++i) {
1346  Value *L = P->getIncomingValue(i);
1347  Value *R = P->getIncomingValue(!i);
1348  Operator *LU = dyn_cast<Operator>(L);
1349  if (!LU)
1350  continue;
1351  unsigned Opcode = LU->getOpcode();
1352  // Check for operations that have the property that if
1353  // both their operands have low zero bits, the result
1354  // will have low zero bits.
1355  if (Opcode == Instruction::Add ||
1356  Opcode == Instruction::Sub ||
1357  Opcode == Instruction::And ||
1358  Opcode == Instruction::Or ||
1359  Opcode == Instruction::Mul) {
1360  Value *LL = LU->getOperand(0);
1361  Value *LR = LU->getOperand(1);
1362  // Find a recurrence.
1363  if (LL == I)
1364  L = LR;
1365  else if (LR == I)
1366  L = LL;
1367  else
1368  break;
1369  // Ok, we have a PHI of the form L op= R. Check for low
1370  // zero bits.
1371  computeKnownBits(R, Known2, Depth + 1, Q);
1372 
1373  // We need to take the minimum number of known bits
1374  KnownBits Known3(Known);
1375  computeKnownBits(L, Known3, Depth + 1, Q);
1376 
1377  Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1378  Known3.countMinTrailingZeros()));
1379 
1380  auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1381  if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) {
1382  // If initial value of recurrence is nonnegative, and we are adding
1383  // a nonnegative number with nsw, the result can only be nonnegative
1384  // or poison value regardless of the number of times we execute the
1385  // add in phi recurrence. If initial value is negative and we are
1386  // adding a negative number with nsw, the result can only be
1387  // negative or poison value. Similar arguments apply to sub and mul.
1388  //
1389  // (add non-negative, non-negative) --> non-negative
1390  // (add negative, negative) --> negative
1391  if (Opcode == Instruction::Add) {
1392  if (Known2.isNonNegative() && Known3.isNonNegative())
1393  Known.makeNonNegative();
1394  else if (Known2.isNegative() && Known3.isNegative())
1395  Known.makeNegative();
1396  }
1397 
1398  // (sub nsw non-negative, negative) --> non-negative
1399  // (sub nsw negative, non-negative) --> negative
1400  else if (Opcode == Instruction::Sub && LL == I) {
1401  if (Known2.isNonNegative() && Known3.isNegative())
1402  Known.makeNonNegative();
1403  else if (Known2.isNegative() && Known3.isNonNegative())
1404  Known.makeNegative();
1405  }
1406 
1407  // (mul nsw non-negative, non-negative) --> non-negative
1408  else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1409  Known3.isNonNegative())
1410  Known.makeNonNegative();
1411  }
1412 
1413  break;
1414  }
1415  }
1416  }
1417 
1418  // Unreachable blocks may have zero-operand PHI nodes.
1419  if (P->getNumIncomingValues() == 0)
1420  break;
1421 
1422  // Otherwise take the unions of the known bit sets of the operands,
1423  // taking conservative care to avoid excessive recursion.
1424  if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1425  // Skip if every incoming value references to ourself.
1426  if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1427  break;
1428 
1429  Known.Zero.setAllBits();
1430  Known.One.setAllBits();
1431  for (Value *IncValue : P->incoming_values()) {
1432  // Skip direct self references.
1433  if (IncValue == P) continue;
1434 
1435  Known2 = KnownBits(BitWidth);
1436  // Recurse, but cap the recursion to one level, because we don't
1437  // want to waste time spinning around in loops.
1438  computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1439  Known.Zero &= Known2.Zero;
1440  Known.One &= Known2.One;
1441  // If all bits have been ruled out, there's no need to check
1442  // more operands.
1443  if (!Known.Zero && !Known.One)
1444  break;
1445  }
1446  }
1447  break;
1448  }
1449  case Instruction::Call:
1450  case Instruction::Invoke:
1451  // If range metadata is attached to this call, set known bits from that,
1452  // and then intersect with known bits based on other properties of the
1453  // function.
1454  if (MDNode *MD =
1455  Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range))
1457  if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1458  computeKnownBits(RV, Known2, Depth + 1, Q);
1459  Known.Zero |= Known2.Zero;
1460  Known.One |= Known2.One;
1461  }
1462  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1463  switch (II->getIntrinsicID()) {
1464  default: break;
1465  case Intrinsic::bitreverse:
1466  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1467  Known.Zero |= Known2.Zero.reverseBits();
1468  Known.One |= Known2.One.reverseBits();
1469  break;
1470  case Intrinsic::bswap:
1471  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1472  Known.Zero |= Known2.Zero.byteSwap();
1473  Known.One |= Known2.One.byteSwap();
1474  break;
1475  case Intrinsic::ctlz: {
1476  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1477  // If we have a known 1, its position is our upper bound.
1478  unsigned PossibleLZ = Known2.One.countLeadingZeros();
1479  // If this call is undefined for 0, the result will be less than 2^n.
1480  if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1481  PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1482  unsigned LowBits = Log2_32(PossibleLZ)+1;
1483  Known.Zero.setBitsFrom(LowBits);
1484  break;
1485  }
1486  case Intrinsic::cttz: {
1487  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1488  // If we have a known 1, its position is our upper bound.
1489  unsigned PossibleTZ = Known2.One.countTrailingZeros();
1490  // If this call is undefined for 0, the result will be less than 2^n.
1491  if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1492  PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1493  unsigned LowBits = Log2_32(PossibleTZ)+1;
1494  Known.Zero.setBitsFrom(LowBits);
1495  break;
1496  }
1497  case Intrinsic::ctpop: {
1498  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1499  // We can bound the space the count needs. Also, bits known to be zero
1500  // can't contribute to the population.
1501  unsigned BitsPossiblySet = Known2.countMaxPopulation();
1502  unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1503  Known.Zero.setBitsFrom(LowBits);
1504  // TODO: we could bound KnownOne using the lower bound on the number
1505  // of bits which might be set provided by popcnt KnownOne2.
1506  break;
1507  }
1508  case Intrinsic::fshr:
1509  case Intrinsic::fshl: {
1510  const APInt *SA;
1511  if (!match(I->getOperand(2), m_APInt(SA)))
1512  break;
1513 
1514  // Normalize to funnel shift left.
1515  uint64_t ShiftAmt = SA->urem(BitWidth);
1516  if (II->getIntrinsicID() == Intrinsic::fshr)
1517  ShiftAmt = BitWidth - ShiftAmt;
1518 
1519  KnownBits Known3(Known);
1520  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1521  computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q);
1522 
1523  Known.Zero =
1524  Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt);
1525  Known.One =
1526  Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt);
1527  break;
1528  }
1529  case Intrinsic::x86_sse42_crc32_64_64:
1530  Known.Zero.setBitsFrom(32);
1531  break;
1532  }
1533  }
1534  break;
1535  case Instruction::ExtractElement:
1536  // Look through extract element. At the moment we keep this simple and skip
1537  // tracking the specific element. But at least we might find information
1538  // valid for all elements of the vector (for example if vector is sign
1539  // extended, shifted, etc).
1540  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1541  break;
1542  case Instruction::ExtractValue:
1543  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1544  const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1545  if (EVI->getNumIndices() != 1) break;
1546  if (EVI->getIndices()[0] == 0) {
1547  switch (II->getIntrinsicID()) {
1548  default: break;
1549  case Intrinsic::uadd_with_overflow:
1550  case Intrinsic::sadd_with_overflow:
1551  computeKnownBitsAddSub(true, II->getArgOperand(0),
1552  II->getArgOperand(1), false, Known, Known2,
1553  Depth, Q);
1554  break;
1555  case Intrinsic::usub_with_overflow:
1556  case Intrinsic::ssub_with_overflow:
1557  computeKnownBitsAddSub(false, II->getArgOperand(0),
1558  II->getArgOperand(1), false, Known, Known2,
1559  Depth, Q);
1560  break;
1561  case Intrinsic::umul_with_overflow:
1562  case Intrinsic::smul_with_overflow:
1563  computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1564  Known, Known2, Depth, Q);
1565  break;
1566  }
1567  }
1568  }
1569  }
1570 }
1571 
1572 /// Determine which bits of V are known to be either zero or one and return
1573 /// them.
1574 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1575  KnownBits Known(getBitWidth(V->getType(), Q.DL));
1576  computeKnownBits(V, Known, Depth, Q);
1577  return Known;
1578 }
1579 
1580 /// Determine which bits of V are known to be either zero or one and return
1581 /// them in the Known bit set.
1582 ///
1583 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1584 /// we cannot optimize based on the assumption that it is zero without changing
1585 /// it to be an explicit zero. If we don't change it to zero, other code could
1586 /// optimized based on the contradictory assumption that it is non-zero.
1587 /// Because instcombine aggressively folds operations with undef args anyway,
1588 /// this won't lose us code quality.
1589 ///
1590 /// This function is defined on values with integer type, values with pointer
1591 /// type, and vectors of integers. In the case
1592 /// where V is a vector, known zero, and known one values are the
1593 /// same width as the vector element, and the bit is set only if it is true
1594 /// for all of the elements in the vector.
1595 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1596  const Query &Q) {
1597  assert(V && "No Value?");
1598  assert(Depth <= MaxDepth && "Limit Search Depth");
1599  unsigned BitWidth = Known.getBitWidth();
1600 
1601  assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||
1602  V->getType()->isPtrOrPtrVectorTy()) &&
1603  "Not integer or pointer type!");
1604 
1605  Type *ScalarTy = V->getType()->getScalarType();
1606  unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
1607  Q.DL.getIndexTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
1608  assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth");
1609  (void)BitWidth;
1610  (void)ExpectedWidth;
1611 
1612  const APInt *C;
1613  if (match(V, m_APInt(C))) {
1614  // We know all of the bits for a scalar constant or a splat vector constant!
1615  Known.One = *C;
1616  Known.Zero = ~Known.One;
1617  return;
1618  }
1619  // Null and aggregate-zero are all-zeros.
1620  if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1621  Known.setAllZero();
1622  return;
1623  }
1624  // Handle a constant vector by taking the intersection of the known bits of
1625  // each element.
1626  if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1627  // We know that CDS must be a vector of integers. Take the intersection of
1628  // each element.
1629  Known.Zero.setAllBits(); Known.One.setAllBits();
1630  for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1631  APInt Elt = CDS->getElementAsAPInt(i);
1632  Known.Zero &= ~Elt;
1633  Known.One &= Elt;
1634  }
1635  return;
1636  }
1637 
1638  if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1639  // We know that CV must be a vector of integers. Take the intersection of
1640  // each element.
1641  Known.Zero.setAllBits(); Known.One.setAllBits();
1642  for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1643  Constant *Element = CV->getAggregateElement(i);
1644  auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1645  if (!ElementCI) {
1646  Known.resetAll();
1647  return;
1648  }
1649  const APInt &Elt = ElementCI->getValue();
1650  Known.Zero &= ~Elt;
1651  Known.One &= Elt;
1652  }
1653  return;
1654  }
1655 
1656  // Start out not knowing anything.
1657  Known.resetAll();
1658 
1659  // We can't imply anything about undefs.
1660  if (isa<UndefValue>(V))
1661  return;
1662 
1663  // There's no point in looking through other users of ConstantData for
1664  // assumptions. Confirm that we've handled them all.
1665  assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1666 
1667  // Limit search depth.
1668  // All recursive calls that increase depth must come after this.
1669  if (Depth == MaxDepth)
1670  return;
1671 
1672  // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1673  // the bits of its aliasee.
1674  if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1675  if (!GA->isInterposable())
1676  computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1677  return;
1678  }
1679 
1680  if (const Operator *I = dyn_cast<Operator>(V))
1681  computeKnownBitsFromOperator(I, Known, Depth, Q);
1682 
1683  // Aligned pointers have trailing zeros - refine Known.Zero set
1684  if (V->getType()->isPointerTy()) {
1685  unsigned Align = V->getPointerAlignment(Q.DL);
1686  if (Align)
1687  Known.Zero.setLowBits(countTrailingZeros(Align));
1688  }
1689 
1690  // computeKnownBitsFromAssume strictly refines Known.
1691  // Therefore, we run them after computeKnownBitsFromOperator.
1692 
1693  // Check whether a nearby assume intrinsic can determine some known bits.
1694  computeKnownBitsFromAssume(V, Known, Depth, Q);
1695 
1696  assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1697 }
1698 
1699 /// Return true if the given value is known to have exactly one
1700 /// bit set when defined. For vectors return true if every element is known to
1701 /// be a power of two when defined. Supports values with integer or pointer
1702 /// types and vectors of integers.
1703 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1704  const Query &Q) {
1705  assert(Depth <= MaxDepth && "Limit Search Depth");
1706 
1707  // Attempt to match against constants.
1708  if (OrZero && match(V, m_Power2OrZero()))
1709  return true;
1710  if (match(V, m_Power2()))
1711  return true;
1712 
1713  // 1 << X is clearly a power of two if the one is not shifted off the end. If
1714  // it is shifted off the end then the result is undefined.
1715  if (match(V, m_Shl(m_One(), m_Value())))
1716  return true;
1717 
1718  // (signmask) >>l X is clearly a power of two if the one is not shifted off
1719  // the bottom. If it is shifted off the bottom then the result is undefined.
1720  if (match(V, m_LShr(m_SignMask(), m_Value())))
1721  return true;
1722 
1723  // The remaining tests are all recursive, so bail out if we hit the limit.
1724  if (Depth++ == MaxDepth)
1725  return false;
1726 
1727  Value *X = nullptr, *Y = nullptr;
1728  // A shift left or a logical shift right of a power of two is a power of two
1729  // or zero.
1730  if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1731  match(V, m_LShr(m_Value(X), m_Value()))))
1732  return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1733 
1734  if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1735  return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1736 
1737  if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1738  return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1739  isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1740 
1741  if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1742  // A power of two and'd with anything is a power of two or zero.
1743  if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1744  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1745  return true;
1746  // X & (-X) is always a power of two or zero.
1747  if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1748  return true;
1749  return false;
1750  }
1751 
1752  // Adding a power-of-two or zero to the same power-of-two or zero yields
1753  // either the original power-of-two, a larger power-of-two or zero.
1754  if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1755  const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1756  if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) ||
1757  Q.IIQ.hasNoSignedWrap(VOBO)) {
1758  if (match(X, m_And(m_Specific(Y), m_Value())) ||
1759  match(X, m_And(m_Value(), m_Specific(Y))))
1760  if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1761  return true;
1762  if (match(Y, m_And(m_Specific(X), m_Value())) ||
1763  match(Y, m_And(m_Value(), m_Specific(X))))
1764  if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1765  return true;
1766 
1767  unsigned BitWidth = V->getType()->getScalarSizeInBits();
1768  KnownBits LHSBits(BitWidth);
1769  computeKnownBits(X, LHSBits, Depth, Q);
1770 
1771  KnownBits RHSBits(BitWidth);
1772  computeKnownBits(Y, RHSBits, Depth, Q);
1773  // If i8 V is a power of two or zero:
1774  // ZeroBits: 1 1 1 0 1 1 1 1
1775  // ~ZeroBits: 0 0 0 1 0 0 0 0
1776  if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1777  // If OrZero isn't set, we cannot give back a zero result.
1778  // Make sure either the LHS or RHS has a bit set.
1779  if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1780  return true;
1781  }
1782  }
1783 
1784  // An exact divide or right shift can only shift off zero bits, so the result
1785  // is a power of two only if the first operand is a power of two and not
1786  // copying a sign bit (sdiv int_min, 2).
1787  if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1788  match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1789  return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1790  Depth, Q);
1791  }
1792 
1793  return false;
1794 }
1795 
1796 /// Test whether a GEP's result is known to be non-null.
1797 ///
1798 /// Uses properties inherent in a GEP to try to determine whether it is known
1799 /// to be non-null.
1800 ///
1801 /// Currently this routine does not support vector GEPs.
1802 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1803  const Query &Q) {
1804  const Function *F = nullptr;
1805  if (const Instruction *I = dyn_cast<Instruction>(GEP))
1806  F = I->getFunction();
1807 
1808  if (!GEP->isInBounds() ||
1810  return false;
1811 
1812  // FIXME: Support vector-GEPs.
1813  assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1814 
1815  // If the base pointer is non-null, we cannot walk to a null address with an
1816  // inbounds GEP in address space zero.
1817  if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1818  return true;
1819 
1820  // Walk the GEP operands and see if any operand introduces a non-zero offset.
1821  // If so, then the GEP cannot produce a null pointer, as doing so would
1822  // inherently violate the inbounds contract within address space zero.
1823  for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1824  GTI != GTE; ++GTI) {
1825  // Struct types are easy -- they must always be indexed by a constant.
1826  if (StructType *STy = GTI.getStructTypeOrNull()) {
1827  ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1828  unsigned ElementIdx = OpC->getZExtValue();
1829  const StructLayout *SL = Q.DL.getStructLayout(STy);
1830  uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1831  if (ElementOffset > 0)
1832  return true;
1833  continue;
1834  }
1835 
1836  // If we have a zero-sized type, the index doesn't matter. Keep looping.
1837  if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1838  continue;
1839 
1840  // Fast path the constant operand case both for efficiency and so we don't
1841  // increment Depth when just zipping down an all-constant GEP.
1842  if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1843  if (!OpC->isZero())
1844  return true;
1845  continue;
1846  }
1847 
1848  // We post-increment Depth here because while isKnownNonZero increments it
1849  // as well, when we pop back up that increment won't persist. We don't want
1850  // to recurse 10k times just because we have 10k GEP operands. We don't
1851  // bail completely out because we want to handle constant GEPs regardless
1852  // of depth.
1853  if (Depth++ >= MaxDepth)
1854  continue;
1855 
1856  if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1857  return true;
1858  }
1859 
1860  return false;
1861 }
1862 
1864  const Instruction *CtxI,
1865  const DominatorTree *DT) {
1866  assert(V->getType()->isPointerTy() && "V must be pointer type");
1867  assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
1868 
1869  if (!CtxI || !DT)
1870  return false;
1871 
1872  unsigned NumUsesExplored = 0;
1873  for (auto *U : V->users()) {
1874  // Avoid massive lists
1875  if (NumUsesExplored >= DomConditionsMaxUses)
1876  break;
1877  NumUsesExplored++;
1878 
1879  // If the value is used as an argument to a call or invoke, then argument
1880  // attributes may provide an answer about null-ness.
1881  if (auto CS = ImmutableCallSite(U))
1882  if (auto *CalledFunc = CS.getCalledFunction())
1883  for (const Argument &Arg : CalledFunc->args())
1884  if (CS.getArgOperand(Arg.getArgNo()) == V &&
1885  Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
1886  return true;
1887 
1888  // Consider only compare instructions uniquely controlling a branch
1889  CmpInst::Predicate Pred;
1890  if (!match(const_cast<User *>(U),
1891  m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
1892  (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
1893  continue;
1894 
1897  for (auto *CmpU : U->users()) {
1898  assert(WorkList.empty() && "Should be!");
1899  if (Visited.insert(CmpU).second)
1900  WorkList.push_back(CmpU);
1901 
1902  while (!WorkList.empty()) {
1903  auto *Curr = WorkList.pop_back_val();
1904 
1905  // If a user is an AND, add all its users to the work list. We only
1906  // propagate "pred != null" condition through AND because it is only
1907  // correct to assume that all conditions of AND are met in true branch.
1908  // TODO: Support similar logic of OR and EQ predicate?
1909  if (Pred == ICmpInst::ICMP_NE)
1910  if (auto *BO = dyn_cast<BinaryOperator>(Curr))
1911  if (BO->getOpcode() == Instruction::And) {
1912  for (auto *BOU : BO->users())
1913  if (Visited.insert(BOU).second)
1914  WorkList.push_back(BOU);
1915  continue;
1916  }
1917 
1918  if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) {
1919  assert(BI->isConditional() && "uses a comparison!");
1920 
1921  BasicBlock *NonNullSuccessor =
1922  BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
1923  BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
1924  if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
1925  return true;
1926  } else if (Pred == ICmpInst::ICMP_NE && isGuard(Curr) &&
1927  DT->dominates(cast<Instruction>(Curr), CtxI)) {
1928  return true;
1929  }
1930  }
1931  }
1932  }
1933 
1934  return false;
1935 }
1936 
1937 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1938 /// ensure that the value it's attached to is never Value? 'RangeType' is
1939 /// is the type of the value described by the range.
1940 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1941  const unsigned NumRanges = Ranges->getNumOperands() / 2;
1942  assert(NumRanges >= 1);
1943  for (unsigned i = 0; i < NumRanges; ++i) {
1944  ConstantInt *Lower =
1945  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1946  ConstantInt *Upper =
1947  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1948  ConstantRange Range(Lower->getValue(), Upper->getValue());
1949  if (Range.contains(Value))
1950  return false;
1951  }
1952  return true;
1953 }
1954 
1955 /// Return true if the given value is known to be non-zero when defined. For
1956 /// vectors, return true if every element is known to be non-zero when
1957 /// defined. For pointers, if the context instruction and dominator tree are
1958 /// specified, perform context-sensitive analysis and return true if the
1959 /// pointer couldn't possibly be null at the specified instruction.
1960 /// Supports values with integer or pointer type and vectors of integers.
1961 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1962  if (auto *C = dyn_cast<Constant>(V)) {
1963  if (C->isNullValue())
1964  return false;
1965  if (isa<ConstantInt>(C))
1966  // Must be non-zero due to null test above.
1967  return true;
1968 
1969  // For constant vectors, check that all elements are undefined or known
1970  // non-zero to determine that the whole vector is known non-zero.
1971  if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1972  for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1973  Constant *Elt = C->getAggregateElement(i);
1974  if (!Elt || Elt->isNullValue())
1975  return false;
1976  if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1977  return false;
1978  }
1979  return true;
1980  }
1981 
1982  // A global variable in address space 0 is non null unless extern weak
1983  // or an absolute symbol reference. Other address spaces may have null as a
1984  // valid address for a global, so we can't assume anything.
1985  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
1986  if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
1987  GV->getType()->getAddressSpace() == 0)
1988  return true;
1989  } else
1990  return false;
1991  }
1992 
1993  if (auto *I = dyn_cast<Instruction>(V)) {
1994  if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) {
1995  // If the possible ranges don't contain zero, then the value is
1996  // definitely non-zero.
1997  if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1998  const APInt ZeroValue(Ty->getBitWidth(), 0);
1999  if (rangeMetadataExcludesValue(Ranges, ZeroValue))
2000  return true;
2001  }
2002  }
2003  }
2004 
2005  // Some of the tests below are recursive, so bail out if we hit the limit.
2006  if (Depth++ >= MaxDepth)
2007  return false;
2008 
2009  // Check for pointer simplifications.
2010  if (V->getType()->isPointerTy()) {
2011  // Alloca never returns null, malloc might.
2012  if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
2013  return true;
2014 
2015  // A byval, inalloca, or nonnull argument is never null.
2016  if (const Argument *A = dyn_cast<Argument>(V))
2017  if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
2018  return true;
2019 
2020  // A Load tagged with nonnull metadata is never null.
2021  if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2022  if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull))
2023  return true;
2024 
2025  if (const auto *Call = dyn_cast<CallBase>(V)) {
2026  if (Call->isReturnNonNull())
2027  return true;
2028  if (const auto *RP = getArgumentAliasingToReturnedPointer(Call))
2029  return isKnownNonZero(RP, Depth, Q);
2030  }
2031  }
2032 
2033 
2034  // Check for recursive pointer simplifications.
2035  if (V->getType()->isPointerTy()) {
2036  if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
2037  return true;
2038 
2039  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
2040  if (isGEPKnownNonNull(GEP, Depth, Q))
2041  return true;
2042  }
2043 
2044  unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
2045 
2046  // X | Y != 0 if X != 0 or Y != 0.
2047  Value *X = nullptr, *Y = nullptr;
2048  if (match(V, m_Or(m_Value(X), m_Value(Y))))
2049  return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
2050 
2051  // ext X != 0 if X != 0.
2052  if (isa<SExtInst>(V) || isa<ZExtInst>(V))
2053  return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2054 
2055  // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2056  // if the lowest bit is shifted off the end.
2057  if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2058  // shl nuw can't remove any non-zero bits.
2059  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2060  if (Q.IIQ.hasNoUnsignedWrap(BO))
2061  return isKnownNonZero(X, Depth, Q);
2062 
2063  KnownBits Known(BitWidth);
2064  computeKnownBits(X, Known, Depth, Q);
2065  if (Known.One[0])
2066  return true;
2067  }
2068  // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2069  // defined if the sign bit is shifted off the end.
2070  else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2071  // shr exact can only shift out zero bits.
2072  const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2073  if (BO->isExact())
2074  return isKnownNonZero(X, Depth, Q);
2075 
2076  KnownBits Known = computeKnownBits(X, Depth, Q);
2077  if (Known.isNegative())
2078  return true;
2079 
2080  // If the shifter operand is a constant, and all of the bits shifted
2081  // out are known to be zero, and X is known non-zero then at least one
2082  // non-zero bit must remain.
2083  if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2084  auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2085  // Is there a known one in the portion not shifted out?
2086  if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2087  return true;
2088  // Are all the bits to be shifted out known zero?
2089  if (Known.countMinTrailingZeros() >= ShiftVal)
2090  return isKnownNonZero(X, Depth, Q);
2091  }
2092  }
2093  // div exact can only produce a zero if the dividend is zero.
2094  else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2095  return isKnownNonZero(X, Depth, Q);
2096  }
2097  // X + Y.
2098  else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2099  KnownBits XKnown = computeKnownBits(X, Depth, Q);
2100  KnownBits YKnown = computeKnownBits(Y, Depth, Q);
2101 
2102  // If X and Y are both non-negative (as signed values) then their sum is not
2103  // zero unless both X and Y are zero.
2104  if (XKnown.isNonNegative() && YKnown.isNonNegative())
2105  if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
2106  return true;
2107 
2108  // If X and Y are both negative (as signed values) then their sum is not
2109  // zero unless both X and Y equal INT_MIN.
2110  if (XKnown.isNegative() && YKnown.isNegative()) {
2111  APInt Mask = APInt::getSignedMaxValue(BitWidth);
2112  // The sign bit of X is set. If some other bit is set then X is not equal
2113  // to INT_MIN.
2114  if (XKnown.One.intersects(Mask))
2115  return true;
2116  // The sign bit of Y is set. If some other bit is set then Y is not equal
2117  // to INT_MIN.
2118  if (YKnown.One.intersects(Mask))
2119  return true;
2120  }
2121 
2122  // The sum of a non-negative number and a power of two is not zero.
2123  if (XKnown.isNonNegative() &&
2124  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2125  return true;
2126  if (YKnown.isNonNegative() &&
2127  isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2128  return true;
2129  }
2130  // X * Y.
2131  else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2132  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2133  // If X and Y are non-zero then so is X * Y as long as the multiplication
2134  // does not overflow.
2135  if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) &&
2136  isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
2137  return true;
2138  }
2139  // (C ? X : Y) != 0 if X != 0 and Y != 0.
2140  else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2141  if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
2142  isKnownNonZero(SI->getFalseValue(), Depth, Q))
2143  return true;
2144  }
2145  // PHI
2146  else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2147  // Try and detect a recurrence that monotonically increases from a
2148  // starting value, as these are common as induction variables.
2149  if (PN->getNumIncomingValues() == 2) {
2150  Value *Start = PN->getIncomingValue(0);
2151  Value *Induction = PN->getIncomingValue(1);
2152  if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2153  std::swap(Start, Induction);
2154  if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2155  if (!C->isZero() && !C->isNegative()) {
2156  ConstantInt *X;
2157  if (Q.IIQ.UseInstrInfo &&
2158  (match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2159  match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2160  !X->isNegative())
2161  return true;
2162  }
2163  }
2164  }
2165  // Check if all incoming values are non-zero constant.
2166  bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
2167  return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
2168  });
2169  if (AllNonZeroConstants)
2170  return true;
2171  }
2172 
2173  KnownBits Known(BitWidth);
2174  computeKnownBits(V, Known, Depth, Q);
2175  return Known.One != 0;
2176 }
2177 
2178 /// Return true if V2 == V1 + X, where X is known non-zero.
2179 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
2180  const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2181  if (!BO || BO->getOpcode() != Instruction::Add)
2182  return false;
2183  Value *Op = nullptr;
2184  if (V2 == BO->getOperand(0))
2185  Op = BO->getOperand(1);
2186  else if (V2 == BO->getOperand(1))
2187  Op = BO->getOperand(0);
2188  else
2189  return false;
2190  return isKnownNonZero(Op, 0, Q);
2191 }
2192 
2193 /// Return true if it is known that V1 != V2.
2194 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
2195  if (V1 == V2)
2196  return false;
2197  if (V1->getType() != V2->getType())
2198  // We can't look through casts yet.
2199  return false;
2200  if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2201  return true;
2202 
2203  if (V1->getType()->isIntOrIntVectorTy()) {
2204  // Are any known bits in V1 contradictory to known bits in V2? If V1
2205  // has a known zero where V2 has a known one, they must not be equal.
2206  KnownBits Known1 = computeKnownBits(V1, 0, Q);
2207  KnownBits Known2 = computeKnownBits(V2, 0, Q);
2208 
2209  if (Known1.Zero.intersects(Known2.One) ||
2210  Known2.Zero.intersects(Known1.One))
2211  return true;
2212  }
2213  return false;
2214 }
2215 
2216 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2217 /// simplify operations downstream. Mask is known to be zero for bits that V
2218 /// cannot have.
2219 ///
2220 /// This function is defined on values with integer type, values with pointer
2221 /// type, and vectors of integers. In the case
2222 /// where V is a vector, the mask, known zero, and known one values are the
2223 /// same width as the vector element, and the bit is set only if it is true
2224 /// for all of the elements in the vector.
2225 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2226  const Query &Q) {
2227  KnownBits Known(Mask.getBitWidth());
2228  computeKnownBits(V, Known, Depth, Q);
2229  return Mask.isSubsetOf(Known.Zero);
2230 }
2231 
2232 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow).
2233 // Returns the input and lower/upper bounds.
2234 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In,
2235  const APInt *&CLow, const APInt *&CHigh) {
2236  assert(isa<Operator>(Select) &&
2237  cast<Operator>(Select)->getOpcode() == Instruction::Select &&
2238  "Input should be a Select!");
2239 
2240  const Value *LHS, *RHS, *LHS2, *RHS2;
2241  SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor;
2242  if (SPF != SPF_SMAX && SPF != SPF_SMIN)
2243  return false;
2244 
2245  if (!match(RHS, m_APInt(CLow)))
2246  return false;
2247 
2248  SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor;
2249  if (getInverseMinMaxFlavor(SPF) != SPF2)
2250  return false;
2251 
2252  if (!match(RHS2, m_APInt(CHigh)))
2253  return false;
2254 
2255  if (SPF == SPF_SMIN)
2256  std::swap(CLow, CHigh);
2257 
2258  In = LHS2;
2259  return CLow->sle(*CHigh);
2260 }
2261 
2262 /// For vector constants, loop over the elements and find the constant with the
2263 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2264 /// or if any element was not analyzed; otherwise, return the count for the
2265 /// element with the minimum number of sign bits.
2266 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2267  unsigned TyBits) {
2268  const auto *CV = dyn_cast<Constant>(V);
2269  if (!CV || !CV->getType()->isVectorTy())
2270  return 0;
2271 
2272  unsigned MinSignBits = TyBits;
2273  unsigned NumElts = CV->getType()->getVectorNumElements();
2274  for (unsigned i = 0; i != NumElts; ++i) {
2275  // If we find a non-ConstantInt, bail out.
2276  auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2277  if (!Elt)
2278  return 0;
2279 
2280  MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2281  }
2282 
2283  return MinSignBits;
2284 }
2285 
2286 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2287  const Query &Q);
2288 
2289 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2290  const Query &Q) {
2291  unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2292  assert(Result > 0 && "At least one sign bit needs to be present!");
2293  return Result;
2294 }
2295 
2296 /// Return the number of times the sign bit of the register is replicated into
2297 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2298 /// (itself), but other cases can give us information. For example, immediately
2299 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2300 /// other, so we return 3. For vectors, return the number of sign bits for the
2301 /// vector element with the minimum number of known sign bits.
2302 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2303  const Query &Q) {
2304  assert(Depth <= MaxDepth && "Limit Search Depth");
2305 
2306  // We return the minimum number of sign bits that are guaranteed to be present
2307  // in V, so for undef we have to conservatively return 1. We don't have the
2308  // same behavior for poison though -- that's a FIXME today.
2309 
2310  Type *ScalarTy = V->getType()->getScalarType();
2311  unsigned TyBits = ScalarTy->isPointerTy() ?
2312  Q.DL.getIndexTypeSizeInBits(ScalarTy) :
2313  Q.DL.getTypeSizeInBits(ScalarTy);
2314 
2315  unsigned Tmp, Tmp2;
2316  unsigned FirstAnswer = 1;
2317 
2318  // Note that ConstantInt is handled by the general computeKnownBits case
2319  // below.
2320 
2321  if (Depth == MaxDepth)
2322  return 1; // Limit search depth.
2323 
2324  const Operator *U = dyn_cast<Operator>(V);
2325  switch (Operator::getOpcode(V)) {
2326  default: break;
2327  case Instruction::SExt:
2328  Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2329  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2330 
2331  case Instruction::SDiv: {
2332  const APInt *Denominator;
2333  // sdiv X, C -> adds log(C) sign bits.
2334  if (match(U->getOperand(1), m_APInt(Denominator))) {
2335 
2336  // Ignore non-positive denominator.
2337  if (!Denominator->isStrictlyPositive())
2338  break;
2339 
2340  // Calculate the incoming numerator bits.
2341  unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2342 
2343  // Add floor(log(C)) bits to the numerator bits.
2344  return std::min(TyBits, NumBits + Denominator->logBase2());
2345  }
2346  break;
2347  }
2348 
2349  case Instruction::SRem: {
2350  const APInt *Denominator;
2351  // srem X, C -> we know that the result is within [-C+1,C) when C is a
2352  // positive constant. This let us put a lower bound on the number of sign
2353  // bits.
2354  if (match(U->getOperand(1), m_APInt(Denominator))) {
2355 
2356  // Ignore non-positive denominator.
2357  if (!Denominator->isStrictlyPositive())
2358  break;
2359 
2360  // Calculate the incoming numerator bits. SRem by a positive constant
2361  // can't lower the number of sign bits.
2362  unsigned NumrBits =
2363  ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2364 
2365  // Calculate the leading sign bit constraints by examining the
2366  // denominator. Given that the denominator is positive, there are two
2367  // cases:
2368  //
2369  // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2370  // (1 << ceilLogBase2(C)).
2371  //
2372  // 2. the numerator is negative. Then the result range is (-C,0] and
2373  // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2374  //
2375  // Thus a lower bound on the number of sign bits is `TyBits -
2376  // ceilLogBase2(C)`.
2377 
2378  unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2379  return std::max(NumrBits, ResBits);
2380  }
2381  break;
2382  }
2383 
2384  case Instruction::AShr: {
2385  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2386  // ashr X, C -> adds C sign bits. Vectors too.
2387  const APInt *ShAmt;
2388  if (match(U->getOperand(1), m_APInt(ShAmt))) {
2389  if (ShAmt->uge(TyBits))
2390  break; // Bad shift.
2391  unsigned ShAmtLimited = ShAmt->getZExtValue();
2392  Tmp += ShAmtLimited;
2393  if (Tmp > TyBits) Tmp = TyBits;
2394  }
2395  return Tmp;
2396  }
2397  case Instruction::Shl: {
2398  const APInt *ShAmt;
2399  if (match(U->getOperand(1), m_APInt(ShAmt))) {
2400  // shl destroys sign bits.
2401  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2402  if (ShAmt->uge(TyBits) || // Bad shift.
2403  ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
2404  Tmp2 = ShAmt->getZExtValue();
2405  return Tmp - Tmp2;
2406  }
2407  break;
2408  }
2409  case Instruction::And:
2410  case Instruction::Or:
2411  case Instruction::Xor: // NOT is handled here.
2412  // Logical binary ops preserve the number of sign bits at the worst.
2413  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2414  if (Tmp != 1) {
2415  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2416  FirstAnswer = std::min(Tmp, Tmp2);
2417  // We computed what we know about the sign bits as our first
2418  // answer. Now proceed to the generic code that uses
2419  // computeKnownBits, and pick whichever answer is better.
2420  }
2421  break;
2422 
2423  case Instruction::Select: {
2424  // If we have a clamp pattern, we know that the number of sign bits will be
2425  // the minimum of the clamp min/max range.
2426  const Value *X;
2427  const APInt *CLow, *CHigh;
2428  if (isSignedMinMaxClamp(U, X, CLow, CHigh))
2429  return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
2430 
2431  Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2432  if (Tmp == 1) break;
2433  Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2434  return std::min(Tmp, Tmp2);
2435  }
2436 
2437  case Instruction::Add:
2438  // Add can have at most one carry bit. Thus we know that the output
2439  // is, at worst, one more bit than the inputs.
2440  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2441  if (Tmp == 1) break;
2442 
2443  // Special case decrementing a value (ADD X, -1):
2444  if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2445  if (CRHS->isAllOnesValue()) {
2446  KnownBits Known(TyBits);
2447  computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2448 
2449  // If the input is known to be 0 or 1, the output is 0/-1, which is all
2450  // sign bits set.
2451  if ((Known.Zero | 1).isAllOnesValue())
2452  return TyBits;
2453 
2454  // If we are subtracting one from a positive number, there is no carry
2455  // out of the result.
2456  if (Known.isNonNegative())
2457  return Tmp;
2458  }
2459 
2460  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2461  if (Tmp2 == 1) break;
2462  return std::min(Tmp, Tmp2)-1;
2463 
2464  case Instruction::Sub:
2465  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2466  if (Tmp2 == 1) break;
2467 
2468  // Handle NEG.
2469  if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2470  if (CLHS->isNullValue()) {
2471  KnownBits Known(TyBits);
2472  computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2473  // If the input is known to be 0 or 1, the output is 0/-1, which is all
2474  // sign bits set.
2475  if ((Known.Zero | 1).isAllOnesValue())
2476  return TyBits;
2477 
2478  // If the input is known to be positive (the sign bit is known clear),
2479  // the output of the NEG has the same number of sign bits as the input.
2480  if (Known.isNonNegative())
2481  return Tmp2;
2482 
2483  // Otherwise, we treat this like a SUB.
2484  }
2485 
2486  // Sub can have at most one carry bit. Thus we know that the output
2487  // is, at worst, one more bit than the inputs.
2488  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2489  if (Tmp == 1) break;
2490  return std::min(Tmp, Tmp2)-1;
2491 
2492  case Instruction::Mul: {
2493  // The output of the Mul can be at most twice the valid bits in the inputs.
2494  unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2495  if (SignBitsOp0 == 1) break;
2496  unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2497  if (SignBitsOp1 == 1) break;
2498  unsigned OutValidBits =
2499  (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
2500  return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
2501  }
2502 
2503  case Instruction::PHI: {
2504  const PHINode *PN = cast<PHINode>(U);
2505  unsigned NumIncomingValues = PN->getNumIncomingValues();
2506  // Don't analyze large in-degree PHIs.
2507  if (NumIncomingValues > 4) break;
2508  // Unreachable blocks may have zero-operand PHI nodes.
2509  if (NumIncomingValues == 0) break;
2510 
2511  // Take the minimum of all incoming values. This can't infinitely loop
2512  // because of our depth threshold.
2513  Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2514  for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2515  if (Tmp == 1) return Tmp;
2516  Tmp = std::min(
2517  Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2518  }
2519  return Tmp;
2520  }
2521 
2522  case Instruction::Trunc:
2523  // FIXME: it's tricky to do anything useful for this, but it is an important
2524  // case for targets like X86.
2525  break;
2526 
2527  case Instruction::ExtractElement:
2528  // Look through extract element. At the moment we keep this simple and skip
2529  // tracking the specific element. But at least we might find information
2530  // valid for all elements of the vector (for example if vector is sign
2531  // extended, shifted, etc).
2532  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2533 
2534  case Instruction::ShuffleVector: {
2535  // TODO: This is copied almost directly from the SelectionDAG version of
2536  // ComputeNumSignBits. It would be better if we could share common
2537  // code. If not, make sure that changes are translated to the DAG.
2538 
2539  // Collect the minimum number of sign bits that are shared by every vector
2540  // element referenced by the shuffle.
2541  auto *Shuf = cast<ShuffleVectorInst>(U);
2542  int NumElts = Shuf->getOperand(0)->getType()->getVectorNumElements();
2543  int NumMaskElts = Shuf->getMask()->getType()->getVectorNumElements();
2544  APInt DemandedLHS(NumElts, 0), DemandedRHS(NumElts, 0);
2545  for (int i = 0; i != NumMaskElts; ++i) {
2546  int M = Shuf->getMaskValue(i);
2547  assert(M < NumElts * 2 && "Invalid shuffle mask constant");
2548  // For undef elements, we don't know anything about the common state of
2549  // the shuffle result.
2550  if (M == -1)
2551  return 1;
2552  if (M < NumElts)
2553  DemandedLHS.setBit(M % NumElts);
2554  else
2555  DemandedRHS.setBit(M % NumElts);
2556  }
2558  if (!!DemandedLHS)
2559  Tmp = ComputeNumSignBits(Shuf->getOperand(0), Depth + 1, Q);
2560  if (!!DemandedRHS) {
2561  Tmp2 = ComputeNumSignBits(Shuf->getOperand(1), Depth + 1, Q);
2562  Tmp = std::min(Tmp, Tmp2);
2563  }
2564  // If we don't know anything, early out and try computeKnownBits fall-back.
2565  if (Tmp == 1)
2566  break;
2567  assert(Tmp <= V->getType()->getScalarSizeInBits() &&
2568  "Failed to determine minimum sign bits");
2569  return Tmp;
2570  }
2571  }
2572 
2573  // Finally, if we can prove that the top bits of the result are 0's or 1's,
2574  // use this information.
2575 
2576  // If we can examine all elements of a vector constant successfully, we're
2577  // done (we can't do any better than that). If not, keep trying.
2578  if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2579  return VecSignBits;
2580 
2581  KnownBits Known(TyBits);
2582  computeKnownBits(V, Known, Depth, Q);
2583 
2584  // If we know that the sign bit is either zero or one, determine the number of
2585  // identical bits in the top of the input value.
2586  return std::max(FirstAnswer, Known.countMinSignBits());
2587 }
2588 
2589 /// This function computes the integer multiple of Base that equals V.
2590 /// If successful, it returns true and returns the multiple in
2591 /// Multiple. If unsuccessful, it returns false. It looks
2592 /// through SExt instructions only if LookThroughSExt is true.
2593 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2594  bool LookThroughSExt, unsigned Depth) {
2595  const unsigned MaxDepth = 6;
2596 
2597  assert(V && "No Value?");
2598  assert(Depth <= MaxDepth && "Limit Search Depth");
2599  assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2600 
2601  Type *T = V->getType();
2602 
2603  ConstantInt *CI = dyn_cast<ConstantInt>(V);
2604 
2605  if (Base == 0)
2606  return false;
2607 
2608  if (Base == 1) {
2609  Multiple = V;
2610  return true;
2611  }
2612 
2614  Constant *BaseVal = ConstantInt::get(T, Base);
2615  if (CO && CO == BaseVal) {
2616  // Multiple is 1.
2617  Multiple = ConstantInt::get(T, 1);
2618  return true;
2619  }
2620 
2621  if (CI && CI->getZExtValue() % Base == 0) {
2622  Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2623  return true;
2624  }
2625 
2626  if (Depth == MaxDepth) return false; // Limit search depth.
2627 
2628  Operator *I = dyn_cast<Operator>(V);
2629  if (!I) return false;
2630 
2631  switch (I->getOpcode()) {
2632  default: break;
2633  case Instruction::SExt:
2634  if (!LookThroughSExt) return false;
2635  // otherwise fall through to ZExt
2637  case Instruction::ZExt:
2638  return ComputeMultiple(I->getOperand(0), Base, Multiple,
2639  LookThroughSExt, Depth+1);
2640  case Instruction::Shl:
2641  case Instruction::Mul: {
2642  Value *Op0 = I->getOperand(0);
2643  Value *Op1 = I->getOperand(1);
2644 
2645  if (I->getOpcode() == Instruction::Shl) {
2646  ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2647  if (!Op1CI) return false;
2648  // Turn Op0 << Op1 into Op0 * 2^Op1
2649  APInt Op1Int = Op1CI->getValue();
2650  uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2651  APInt API(Op1Int.getBitWidth(), 0);
2652  API.setBit(BitToSet);
2653  Op1 = ConstantInt::get(V->getContext(), API);
2654  }
2655 
2656  Value *Mul0 = nullptr;
2657  if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2658  if (Constant *Op1C = dyn_cast<Constant>(Op1))
2659  if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2660  if (Op1C->getType()->getPrimitiveSizeInBits() <
2661  MulC->getType()->getPrimitiveSizeInBits())
2662  Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2663  if (Op1C->getType()->getPrimitiveSizeInBits() >
2664  MulC->getType()->getPrimitiveSizeInBits())
2665  MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2666 
2667  // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2668  Multiple = ConstantExpr::getMul(MulC, Op1C);
2669  return true;
2670  }
2671 
2672  if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2673  if (Mul0CI->getValue() == 1) {
2674  // V == Base * Op1, so return Op1
2675  Multiple = Op1;
2676  return true;
2677  }
2678  }
2679 
2680  Value *Mul1 = nullptr;
2681  if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2682  if (Constant *Op0C = dyn_cast<Constant>(Op0))
2683  if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2684  if (Op0C->getType()->getPrimitiveSizeInBits() <
2685  MulC->getType()->getPrimitiveSizeInBits())
2686  Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2687  if (Op0C->getType()->getPrimitiveSizeInBits() >
2688  MulC->getType()->getPrimitiveSizeInBits())
2689  MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2690 
2691  // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2692  Multiple = ConstantExpr::getMul(MulC, Op0C);
2693  return true;
2694  }
2695 
2696  if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2697  if (Mul1CI->getValue() == 1) {
2698  // V == Base * Op0, so return Op0
2699  Multiple = Op0;
2700  return true;
2701  }
2702  }
2703  }
2704  }
2705 
2706  // We could not determine if V is a multiple of Base.
2707  return false;
2708 }
2709 
2711  const TargetLibraryInfo *TLI) {
2712  const Function *F = ICS.getCalledFunction();
2713  if (!F)
2714  return Intrinsic::not_intrinsic;
2715 
2716  if (F->isIntrinsic())
2717  return F->getIntrinsicID();
2718 
2719  if (!TLI)
2720  return Intrinsic::not_intrinsic;
2721 
2722  LibFunc Func;
2723  // We're going to make assumptions on the semantics of the functions, check
2724  // that the target knows that it's available in this environment and it does
2725  // not have local linkage.
2726  if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2727  return Intrinsic::not_intrinsic;
2728 
2729  if (!ICS.onlyReadsMemory())
2730  return Intrinsic::not_intrinsic;
2731 
2732  // Otherwise check if we have a call to a function that can be turned into a
2733  // vector intrinsic.
2734  switch (Func) {
2735  default:
2736  break;
2737  case LibFunc_sin:
2738  case LibFunc_sinf:
2739  case LibFunc_sinl:
2740  return Intrinsic::sin;
2741  case LibFunc_cos:
2742  case LibFunc_cosf:
2743  case LibFunc_cosl:
2744  return Intrinsic::cos;
2745  case LibFunc_exp:
2746  case LibFunc_expf:
2747  case LibFunc_expl:
2748  return Intrinsic::exp;
2749  case LibFunc_exp2:
2750  case LibFunc_exp2f:
2751  case LibFunc_exp2l:
2752  return Intrinsic::exp2;
2753  case LibFunc_log:
2754  case LibFunc_logf:
2755  case LibFunc_logl:
2756  return Intrinsic::log;
2757  case LibFunc_log10:
2758  case LibFunc_log10f:
2759  case LibFunc_log10l:
2760  return Intrinsic::log10;
2761  case LibFunc_log2:
2762  case LibFunc_log2f:
2763  case LibFunc_log2l:
2764  return Intrinsic::log2;
2765  case LibFunc_fabs:
2766  case LibFunc_fabsf:
2767  case LibFunc_fabsl:
2768  return Intrinsic::fabs;
2769  case LibFunc_fmin:
2770  case LibFunc_fminf:
2771  case LibFunc_fminl:
2772  return Intrinsic::minnum;
2773  case LibFunc_fmax:
2774  case LibFunc_fmaxf:
2775  case LibFunc_fmaxl:
2776  return Intrinsic::maxnum;
2777  case LibFunc_copysign:
2778  case LibFunc_copysignf:
2779  case LibFunc_copysignl:
2780  return Intrinsic::copysign;
2781  case LibFunc_floor:
2782  case LibFunc_floorf:
2783  case LibFunc_floorl:
2784  return Intrinsic::floor;
2785  case LibFunc_ceil:
2786  case LibFunc_ceilf:
2787  case LibFunc_ceill:
2788  return Intrinsic::ceil;
2789  case LibFunc_trunc:
2790  case LibFunc_truncf:
2791  case LibFunc_truncl:
2792  return Intrinsic::trunc;
2793  case LibFunc_rint:
2794  case LibFunc_rintf:
2795  case LibFunc_rintl:
2796  return Intrinsic::rint;
2797  case LibFunc_nearbyint:
2798  case LibFunc_nearbyintf:
2799  case LibFunc_nearbyintl:
2800  return Intrinsic::nearbyint;
2801  case LibFunc_round:
2802  case LibFunc_roundf:
2803  case LibFunc_roundl:
2804  return Intrinsic::round;
2805  case LibFunc_pow:
2806  case LibFunc_powf:
2807  case LibFunc_powl:
2808  return Intrinsic::pow;
2809  case LibFunc_sqrt:
2810  case LibFunc_sqrtf:
2811  case LibFunc_sqrtl:
2812  return Intrinsic::sqrt;
2813  }
2814 
2815  return Intrinsic::not_intrinsic;
2816 }
2817 
2818 /// Return true if we can prove that the specified FP value is never equal to
2819 /// -0.0.
2820 ///
2821 /// NOTE: this function will need to be revisited when we support non-default
2822 /// rounding modes!
2824  unsigned Depth) {
2825  if (auto *CFP = dyn_cast<ConstantFP>(V))
2826  return !CFP->getValueAPF().isNegZero();
2827 
2828  // Limit search depth.
2829  if (Depth == MaxDepth)
2830  return false;
2831 
2832  auto *Op = dyn_cast<Operator>(V);
2833  if (!Op)
2834  return false;
2835 
2836  // Check if the nsz fast-math flag is set.
2837  if (auto *FPO = dyn_cast<FPMathOperator>(Op))
2838  if (FPO->hasNoSignedZeros())
2839  return true;
2840 
2841  // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
2842  if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
2843  return true;
2844 
2845  // sitofp and uitofp turn into +0.0 for zero.
2846  if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
2847  return true;
2848 
2849  if (auto *Call = dyn_cast<CallInst>(Op)) {
2850  Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
2851  switch (IID) {
2852  default:
2853  break;
2854  // sqrt(-0.0) = -0.0, no other negative results are possible.
2855  case Intrinsic::sqrt:
2856  case Intrinsic::canonicalize:
2857  return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
2858  // fabs(x) != -0.0
2859  case Intrinsic::fabs:
2860  return true;
2861  }
2862  }
2863 
2864  return false;
2865 }
2866 
2867 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2868 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2869 /// bit despite comparing equal.
2871  const TargetLibraryInfo *TLI,
2872  bool SignBitOnly,
2873  unsigned Depth) {
2874  // TODO: This function does not do the right thing when SignBitOnly is true
2875  // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2876  // which flips the sign bits of NaNs. See
2877  // https://llvm.org/bugs/show_bug.cgi?id=31702.
2878 
2879  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2880  return !CFP->getValueAPF().isNegative() ||
2881  (!SignBitOnly && CFP->getValueAPF().isZero());
2882  }
2883 
2884  // Handle vector of constants.
2885  if (auto *CV = dyn_cast<Constant>(V)) {
2886  if (CV->getType()->isVectorTy()) {
2887  unsigned NumElts = CV->getType()->getVectorNumElements();
2888  for (unsigned i = 0; i != NumElts; ++i) {
2889  auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
2890  if (!CFP)
2891  return false;
2892  if (CFP->getValueAPF().isNegative() &&
2893  (SignBitOnly || !CFP->getValueAPF().isZero()))
2894  return false;
2895  }
2896 
2897  // All non-negative ConstantFPs.
2898  return true;
2899  }
2900  }
2901 
2902  if (Depth == MaxDepth)
2903  return false; // Limit search depth.
2904 
2905  const Operator *I = dyn_cast<Operator>(V);
2906  if (!I)
2907  return false;
2908 
2909  switch (I->getOpcode()) {
2910  default:
2911  break;
2912  // Unsigned integers are always nonnegative.
2913  case Instruction::UIToFP:
2914  return true;
2915  case Instruction::FMul:
2916  // x*x is always non-negative or a NaN.
2917  if (I->getOperand(0) == I->getOperand(1) &&
2918  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2919  return true;
2920 
2922  case Instruction::FAdd:
2923  case Instruction::FDiv:
2924  case Instruction::FRem:
2925  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2926  Depth + 1) &&
2927  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2928  Depth + 1);
2929  case Instruction::Select:
2930  return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2931  Depth + 1) &&
2932  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2933  Depth + 1);
2934  case Instruction::FPExt:
2935  case Instruction::FPTrunc:
2936  // Widening/narrowing never change sign.
2937  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2938  Depth + 1);
2939  case Instruction::ExtractElement:
2940  // Look through extract element. At the moment we keep this simple and skip
2941  // tracking the specific element. But at least we might find information
2942  // valid for all elements of the vector.
2943  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2944  Depth + 1);
2945  case Instruction::Call:
2946  const auto *CI = cast<CallInst>(I);
2947  Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2948  switch (IID) {
2949  default:
2950  break;
2951  case Intrinsic::maxnum:
2952  return (isKnownNeverNaN(I->getOperand(0), TLI) &&
2954  SignBitOnly, Depth + 1)) ||
2955  (isKnownNeverNaN(I->getOperand(1), TLI) &&
2957  SignBitOnly, Depth + 1));
2958 
2959  case Intrinsic::maximum:
2960  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2961  Depth + 1) ||
2962  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2963  Depth + 1);
2964  case Intrinsic::minnum:
2965  case Intrinsic::minimum:
2966  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2967  Depth + 1) &&
2968  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2969  Depth + 1);
2970  case Intrinsic::exp:
2971  case Intrinsic::exp2:
2972  case Intrinsic::fabs:
2973  return true;
2974 
2975  case Intrinsic::sqrt:
2976  // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2977  if (!SignBitOnly)
2978  return true;
2979  return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2980  CannotBeNegativeZero(CI->getOperand(0), TLI));
2981 
2982  case Intrinsic::powi:
2983  if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2984  // powi(x,n) is non-negative if n is even.
2985  if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2986  return true;
2987  }
2988  // TODO: This is not correct. Given that exp is an integer, here are the
2989  // ways that pow can return a negative value:
2990  //
2991  // pow(x, exp) --> negative if exp is odd and x is negative.
2992  // pow(-0, exp) --> -inf if exp is negative odd.
2993  // pow(-0, exp) --> -0 if exp is positive odd.
2994  // pow(-inf, exp) --> -0 if exp is negative odd.
2995  // pow(-inf, exp) --> -inf if exp is positive odd.
2996  //
2997  // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2998  // but we must return false if x == -0. Unfortunately we do not currently
2999  // have a way of expressing this constraint. See details in
3000  // https://llvm.org/bugs/show_bug.cgi?id=31702.
3001  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3002  Depth + 1);
3003 
3004  case Intrinsic::fma:
3005  case Intrinsic::fmuladd:
3006  // x*x+y is non-negative if y is non-negative.
3007  return I->getOperand(0) == I->getOperand(1) &&
3008  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
3009  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3010  Depth + 1);
3011  }
3012  break;
3013  }
3014  return false;
3015 }
3016 
3018  const TargetLibraryInfo *TLI) {
3019  return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3020 }
3021 
3023  return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3024 }
3025 
3027  unsigned Depth) {
3028  assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
3029 
3030  // If we're told that NaNs won't happen, assume they won't.
3031  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3032  if (FPMathOp->hasNoNaNs())
3033  return true;
3034 
3035  // Handle scalar constants.
3036  if (auto *CFP = dyn_cast<ConstantFP>(V))
3037  return !CFP->isNaN();
3038 
3039  if (Depth == MaxDepth)
3040  return false;
3041 
3042  if (auto *Inst = dyn_cast<Instruction>(V)) {
3043  switch (Inst->getOpcode()) {
3044  case Instruction::FAdd:
3045  case Instruction::FMul:
3046  case Instruction::FSub:
3047  case Instruction::FDiv:
3048  case Instruction::FRem: {
3049  // TODO: Need isKnownNeverInfinity
3050  return false;
3051  }
3052  case Instruction::Select: {
3053  return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3054  isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3055  }
3056  case Instruction::SIToFP:
3057  case Instruction::UIToFP:
3058  return true;
3059  case Instruction::FPTrunc:
3060  case Instruction::FPExt:
3061  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3062  default:
3063  break;
3064  }
3065  }
3066 
3067  if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3068  switch (II->getIntrinsicID()) {
3069  case Intrinsic::canonicalize:
3070  case Intrinsic::fabs:
3071  case Intrinsic::copysign:
3072  case Intrinsic::exp:
3073  case Intrinsic::exp2:
3074  case Intrinsic::floor:
3075  case Intrinsic::ceil:
3076  case Intrinsic::trunc:
3077  case Intrinsic::rint:
3078  case Intrinsic::nearbyint:
3079  case Intrinsic::round:
3080  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3081  case Intrinsic::sqrt:
3082  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3083  CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3084  default:
3085  return false;
3086  }
3087  }
3088 
3089  // Bail out for constant expressions, but try to handle vector constants.
3090  if (!V->getType()->isVectorTy() || !isa<Constant>(V))
3091  return false;
3092 
3093  // For vectors, verify that each element is not NaN.
3094  unsigned NumElts = V->getType()->getVectorNumElements();
3095  for (unsigned i = 0; i != NumElts; ++i) {
3096  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3097  if (!Elt)
3098  return false;
3099  if (isa<UndefValue>(Elt))
3100  continue;
3101  auto *CElt = dyn_cast<ConstantFP>(Elt);
3102  if (!CElt || CElt->isNaN())
3103  return false;
3104  }
3105  // All elements were confirmed not-NaN or undefined.
3106  return true;
3107 }
3108 
3110 
3111  // All byte-wide stores are splatable, even of arbitrary variables.
3112  if (V->getType()->isIntegerTy(8))
3113  return V;
3114 
3115  LLVMContext &Ctx = V->getContext();
3116 
3117  // Undef don't care.
3118  auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
3119  if (isa<UndefValue>(V))
3120  return UndefInt8;
3121 
3122  Constant *C = dyn_cast<Constant>(V);
3123  if (!C) {
3124  // Conceptually, we could handle things like:
3125  // %a = zext i8 %X to i16
3126  // %b = shl i16 %a, 8
3127  // %c = or i16 %a, %b
3128  // but until there is an example that actually needs this, it doesn't seem
3129  // worth worrying about.
3130  return nullptr;
3131  }
3132 
3133  // Handle 'null' ConstantArrayZero etc.
3134  if (C->isNullValue())
3136 
3137  // Constant floating-point values can be handled as integer values if the
3138  // corresponding integer value is "byteable". An important case is 0.0.
3139  if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3140  Type *Ty = nullptr;
3141  if (CFP->getType()->isHalfTy())
3142  Ty = Type::getInt16Ty(Ctx);
3143  else if (CFP->getType()->isFloatTy())
3144  Ty = Type::getInt32Ty(Ctx);
3145  else if (CFP->getType()->isDoubleTy())
3146  Ty = Type::getInt64Ty(Ctx);
3147  // Don't handle long double formats, which have strange constraints.
3148  return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty)) : nullptr;
3149  }
3150 
3151  // We can handle constant integers that are multiple of 8 bits.
3152  if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
3153  if (CI->getBitWidth() % 8 == 0) {
3154  assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
3155  if (!CI->getValue().isSplat(8))
3156  return nullptr;
3157  return ConstantInt::get(Ctx, CI->getValue().trunc(8));
3158  }
3159  }
3160 
3161  auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
3162  if (LHS == RHS)
3163  return LHS;
3164  if (!LHS || !RHS)
3165  return nullptr;
3166  if (LHS == UndefInt8)
3167  return RHS;
3168  if (RHS == UndefInt8)
3169  return LHS;
3170  return nullptr;
3171  };
3172 
3173  if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
3174  Value *Val = UndefInt8;
3175  for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
3176  if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I)))))
3177  return nullptr;
3178  return Val;
3179  }
3180 
3181  if (isa<ConstantVector>(C)) {
3182  Constant *Splat = cast<ConstantVector>(C)->getSplatValue();
3183  return Splat ? isBytewiseValue(Splat) : nullptr;
3184  }
3185 
3186  if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
3187  Value *Val = UndefInt8;
3188  for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
3189  if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I)))))
3190  return nullptr;
3191  return Val;
3192  }
3193 
3194  // Don't try to handle the handful of other constants.
3195  return nullptr;
3196 }
3197 
3198 // This is the recursive version of BuildSubAggregate. It takes a few different
3199 // arguments. Idxs is the index within the nested struct From that we are
3200 // looking at now (which is of type IndexedType). IdxSkip is the number of
3201 // indices from Idxs that should be left out when inserting into the resulting
3202 // struct. To is the result struct built so far, new insertvalue instructions
3203 // build on that.
3204 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
3206  unsigned IdxSkip,
3207  Instruction *InsertBefore) {
3208  StructType *STy = dyn_cast<StructType>(IndexedType);
3209  if (STy) {
3210  // Save the original To argument so we can modify it
3211  Value *OrigTo = To;
3212  // General case, the type indexed by Idxs is a struct
3213  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3214  // Process each struct element recursively
3215  Idxs.push_back(i);
3216  Value *PrevTo = To;
3217  To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3218  InsertBefore);
3219  Idxs.pop_back();
3220  if (!To) {
3221  // Couldn't find any inserted value for this index? Cleanup
3222  while (PrevTo != OrigTo) {
3223  InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3224  PrevTo = Del->getAggregateOperand();
3225  Del->eraseFromParent();
3226  }
3227  // Stop processing elements
3228  break;
3229  }
3230  }
3231  // If we successfully found a value for each of our subaggregates
3232  if (To)
3233  return To;
3234  }
3235  // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3236  // the struct's elements had a value that was inserted directly. In the latter
3237  // case, perhaps we can't determine each of the subelements individually, but
3238  // we might be able to find the complete struct somewhere.
3239 
3240  // Find the value that is at that particular spot
3241  Value *V = FindInsertedValue(From, Idxs);
3242 
3243  if (!V)
3244  return nullptr;
3245 
3246  // Insert the value in the new (sub) aggregate
3247  return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
3248  "tmp", InsertBefore);
3249 }
3250 
3251 // This helper takes a nested struct and extracts a part of it (which is again a
3252 // struct) into a new value. For example, given the struct:
3253 // { a, { b, { c, d }, e } }
3254 // and the indices "1, 1" this returns
3255 // { c, d }.
3256 //
3257 // It does this by inserting an insertvalue for each element in the resulting
3258 // struct, as opposed to just inserting a single struct. This will only work if
3259 // each of the elements of the substruct are known (ie, inserted into From by an
3260 // insertvalue instruction somewhere).
3261 //
3262 // All inserted insertvalue instructions are inserted before InsertBefore
3264  Instruction *InsertBefore) {
3265  assert(InsertBefore && "Must have someplace to insert!");
3266  Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3267  idx_range);
3268  Value *To = UndefValue::get(IndexedType);
3269  SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3270  unsigned IdxSkip = Idxs.size();
3271 
3272  return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3273 }
3274 
3275 /// Given an aggregate and a sequence of indices, see if the scalar value
3276 /// indexed is already around as a register, for example if it was inserted
3277 /// directly into the aggregate.
3278 ///
3279 /// If InsertBefore is not null, this function will duplicate (modified)
3280 /// insertvalues when a part of a nested struct is extracted.
3282  Instruction *InsertBefore) {
3283  // Nothing to index? Just return V then (this is useful at the end of our
3284  // recursion).
3285  if (idx_range.empty())
3286  return V;
3287  // We have indices, so V should have an indexable type.
3288  assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
3289  "Not looking at a struct or array?");
3290  assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
3291  "Invalid indices for type?");
3292 
3293  if (Constant *C = dyn_cast<Constant>(V)) {
3294  C = C->getAggregateElement(idx_range[0]);
3295  if (!C) return nullptr;
3296  return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3297  }
3298 
3299  if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3300  // Loop the indices for the insertvalue instruction in parallel with the
3301  // requested indices
3302  const unsigned *req_idx = idx_range.begin();
3303  for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3304  i != e; ++i, ++req_idx) {
3305  if (req_idx == idx_range.end()) {
3306  // We can't handle this without inserting insertvalues
3307  if (!InsertBefore)
3308  return nullptr;
3309 
3310  // The requested index identifies a part of a nested aggregate. Handle
3311  // this specially. For example,
3312  // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3313  // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3314  // %C = extractvalue {i32, { i32, i32 } } %B, 1
3315  // This can be changed into
3316  // %A = insertvalue {i32, i32 } undef, i32 10, 0
3317  // %C = insertvalue {i32, i32 } %A, i32 11, 1
3318  // which allows the unused 0,0 element from the nested struct to be
3319  // removed.
3320  return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3321  InsertBefore);
3322  }
3323 
3324  // This insert value inserts something else than what we are looking for.
3325  // See if the (aggregate) value inserted into has the value we are
3326  // looking for, then.
3327  if (*req_idx != *i)
3328  return FindInsertedValue(I->getAggregateOperand(), idx_range,
3329  InsertBefore);
3330  }
3331  // If we end up here, the indices of the insertvalue match with those
3332  // requested (though possibly only partially). Now we recursively look at
3333  // the inserted value, passing any remaining indices.
3334  return FindInsertedValue(I->getInsertedValueOperand(),
3335  makeArrayRef(req_idx, idx_range.end()),
3336  InsertBefore);
3337  }
3338 
3339  if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3340  // If we're extracting a value from an aggregate that was extracted from
3341  // something else, we can extract from that something else directly instead.
3342  // However, we will need to chain I's indices with the requested indices.
3343 
3344  // Calculate the number of indices required
3345  unsigned size = I->getNumIndices() + idx_range.size();
3346  // Allocate some space to put the new indices in
3348  Idxs.reserve(size);
3349  // Add indices from the extract value instruction
3350  Idxs.append(I->idx_begin(), I->idx_end());
3351 
3352  // Add requested indices
3353  Idxs.append(idx_range.begin(), idx_range.end());
3354 
3355  assert(Idxs.size() == size
3356  && "Number of indices added not correct?");
3357 
3358  return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
3359  }
3360  // Otherwise, we don't know (such as, extracting from a function return value
3361  // or load instruction)
3362  return nullptr;
3363 }
3364 
3365 /// Analyze the specified pointer to see if it can be expressed as a base
3366 /// pointer plus a constant offset. Return the base and offset to the caller.
3368  const DataLayout &DL) {
3369  unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
3370  APInt ByteOffset(BitWidth, 0);
3371 
3372  // We walk up the defs but use a visited set to handle unreachable code. In
3373  // that case, we stop after accumulating the cycle once (not that it
3374  // matters).
3375  SmallPtrSet<Value *, 16> Visited;
3376  while (Visited.insert(Ptr).second) {
3377  if (Ptr->getType()->isVectorTy())
3378  break;
3379 
3380  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
3381  // If one of the values we have visited is an addrspacecast, then
3382  // the pointer type of this GEP may be different from the type
3383  // of the Ptr parameter which was passed to this function. This
3384  // means when we construct GEPOffset, we need to use the size
3385  // of GEP's pointer type rather than the size of the original
3386  // pointer type.
3387  APInt GEPOffset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
3388  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
3389  break;
3390 
3391  APInt OrigByteOffset(ByteOffset);
3392  ByteOffset += GEPOffset.sextOrTrunc(ByteOffset.getBitWidth());
3393  if (ByteOffset.getMinSignedBits() > 64) {
3394  // Stop traversal if the pointer offset wouldn't fit into int64_t
3395  // (this should be removed if Offset is updated to an APInt)
3396  ByteOffset = OrigByteOffset;
3397  break;
3398  }
3399 
3400  Ptr = GEP->getPointerOperand();
3401  } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
3402  Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
3403  Ptr = cast<Operator>(Ptr)->getOperand(0);
3404  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
3405  if (GA->isInterposable())
3406  break;
3407  Ptr = GA->getAliasee();
3408  } else {
3409  break;
3410  }
3411  }
3412  Offset = ByteOffset.getSExtValue();
3413  return Ptr;
3414 }
3415 
3417  unsigned CharSize) {
3418  // Make sure the GEP has exactly three arguments.
3419  if (GEP->getNumOperands() != 3)
3420  return false;
3421 
3422  // Make sure the index-ee is a pointer to array of \p CharSize integers.
3423  // CharSize.
3425  if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
3426  return false;
3427 
3428  // Check to make sure that the first operand of the GEP is an integer and
3429  // has value 0 so that we are sure we're indexing into the initializer.
3430  const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3431  if (!FirstIdx || !FirstIdx->isZero())
3432  return false;
3433 
3434  return true;
3435 }
3436 
3438  ConstantDataArraySlice &Slice,
3439  unsigned ElementSize, uint64_t Offset) {
3440  assert(V);
3441 
3442  // Look through bitcast instructions and geps.
3443  V = V->stripPointerCasts();
3444 
3445  // If the value is a GEP instruction or constant expression, treat it as an
3446  // offset.
3447  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3448  // The GEP operator should be based on a pointer to string constant, and is
3449  // indexing into the string constant.
3450  if (!isGEPBasedOnPointerToString(GEP, ElementSize))
3451  return false;
3452 
3453  // If the second index isn't a ConstantInt, then this is a variable index
3454  // into the array. If this occurs, we can't say anything meaningful about
3455  // the string.
3456  uint64_t StartIdx = 0;
3457  if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3458  StartIdx = CI->getZExtValue();
3459  else
3460  return false;
3461  return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3462  StartIdx + Offset);
3463  }
3464 
3465  // The GEP instruction, constant or instruction, must reference a global
3466  // variable that is a constant and is initialized. The referenced constant
3467  // initializer is the array that we'll use for optimization.
3468  const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3469  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3470  return false;
3471 
3472  const ConstantDataArray *Array;
3473  ArrayType *ArrayTy;
3474  if (GV->getInitializer()->isNullValue()) {
3475  Type *GVTy = GV->getValueType();
3476  if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3477  // A zeroinitializer for the array; there is no ConstantDataArray.
3478  Array = nullptr;
3479  } else {
3480  const DataLayout &DL = GV->getParent()->getDataLayout();
3481  uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3482  uint64_t Length = SizeInBytes / (ElementSize / 8);
3483  if (Length <= Offset)
3484  return false;
3485 
3486  Slice.Array = nullptr;
3487  Slice.Offset = 0;
3488  Slice.Length = Length - Offset;
3489  return true;
3490  }
3491  } else {
3492  // This must be a ConstantDataArray.
3493  Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3494  if (!Array)
3495  return false;
3496  ArrayTy = Array->getType();
3497  }
3498  if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3499  return false;
3500 
3501  uint64_t NumElts = ArrayTy->getArrayNumElements();
3502  if (Offset > NumElts)
3503  return false;
3504 
3505  Slice.Array = Array;
3506  Slice.Offset = Offset;
3507  Slice.Length = NumElts - Offset;
3508  return true;
3509 }
3510 
3511 /// This function computes the length of a null-terminated C string pointed to
3512 /// by V. If successful, it returns true and returns the string in Str.
3513 /// If unsuccessful, it returns false.
3515  uint64_t Offset, bool TrimAtNul) {
3516  ConstantDataArraySlice Slice;
3517  if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3518  return false;
3519 
3520  if (Slice.Array == nullptr) {
3521  if (TrimAtNul) {
3522  Str = StringRef();
3523  return true;
3524  }
3525  if (Slice.Length == 1) {
3526  Str = StringRef("", 1);
3527  return true;
3528  }
3529  // We cannot instantiate a StringRef as we do not have an appropriate string
3530  // of 0s at hand.
3531  return false;
3532  }
3533 
3534  // Start out with the entire array in the StringRef.
3535  Str = Slice.Array->getAsString();
3536  // Skip over 'offset' bytes.
3537  Str = Str.substr(Slice.Offset);
3538 
3539  if (TrimAtNul) {
3540  // Trim off the \0 and anything after it. If the array is not nul
3541  // terminated, we just return the whole end of string. The client may know
3542  // some other way that the string is length-bound.
3543  Str = Str.substr(0, Str.find('\0'));
3544  }
3545  return true;
3546 }
3547 
3548 // These next two are very similar to the above, but also look through PHI
3549 // nodes.
3550 // TODO: See if we can integrate these two together.
3551 
3552 /// If we can compute the length of the string pointed to by
3553 /// the specified pointer, return 'len+1'. If we can't, return 0.
3554 static uint64_t GetStringLengthH(const Value *V,
3556  unsigned CharSize) {
3557  // Look through noop bitcast instructions.
3558  V = V->stripPointerCasts();
3559 
3560  // If this is a PHI node, there are two cases: either we have already seen it
3561  // or we haven't.
3562  if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3563  if (!PHIs.insert(PN).second)
3564  return ~0ULL; // already in the set.
3565 
3566  // If it was new, see if all the input strings are the same length.
3567  uint64_t LenSoFar = ~0ULL;
3568  for (Value *IncValue : PN->incoming_values()) {
3569  uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3570  if (Len == 0) return 0; // Unknown length -> unknown.
3571 
3572  if (Len == ~0ULL) continue;
3573 
3574  if (Len != LenSoFar && LenSoFar != ~0ULL)
3575  return 0; // Disagree -> unknown.
3576  LenSoFar = Len;
3577  }
3578 
3579  // Success, all agree.
3580  return LenSoFar;
3581  }
3582 
3583  // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3584  if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3585  uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3586  if (Len1 == 0) return 0;
3587  uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3588  if (Len2 == 0) return 0;
3589  if (Len1 == ~0ULL) return Len2;
3590  if (Len2 == ~0ULL) return Len1;
3591  if (Len1 != Len2) return 0;
3592  return Len1;
3593  }
3594 
3595  // Otherwise, see if we can read the string.
3596  ConstantDataArraySlice Slice;
3597  if (!getConstantDataArrayInfo(V, Slice, CharSize))
3598  return 0;
3599 
3600  if (Slice.Array == nullptr)
3601  return 1;
3602 
3603  // Search for nul characters
3604  unsigned NullIndex = 0;
3605  for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3606  if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3607  break;
3608  }
3609 
3610  return NullIndex + 1;
3611 }
3612 
3613 /// If we can compute the length of the string pointed to by
3614 /// the specified pointer, return 'len+1'. If we can't, return 0.
3615 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3616  if (!V->getType()->isPointerTy())
3617  return 0;
3618 
3620  uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3621  // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3622  // an empty string as a length.
3623  return Len == ~0ULL ? 1 : Len;
3624 }
3625 
3627  assert(Call &&
3628  "getArgumentAliasingToReturnedPointer only works on nonnull calls");
3629  if (const Value *RV = Call->getReturnedArgOperand())
3630  return RV;
3631  // This can be used only as a aliasing property.
3633  return Call->getArgOperand(0);
3634  return nullptr;
3635 }
3636 
3638  const CallBase *Call) {
3639  return Call->getIntrinsicID() == Intrinsic::launder_invariant_group ||
3640  Call->getIntrinsicID() == Intrinsic::strip_invariant_group;
3641 }
3642 
3643 /// \p PN defines a loop-variant pointer to an object. Check if the
3644 /// previous iteration of the loop was referring to the same object as \p PN.
3646  const LoopInfo *LI) {
3647  // Find the loop-defined value.
3648  Loop *L = LI->getLoopFor(PN->getParent());
3649  if (PN->getNumIncomingValues() != 2)
3650  return true;
3651 
3652  // Find the value from previous iteration.
3653  auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3654  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3655  PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3656  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3657  return true;
3658 
3659  // If a new pointer is loaded in the loop, the pointer references a different
3660  // object in every iteration. E.g.:
3661  // for (i)
3662  // int *p = a[i];
3663  // ...
3664  if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3665  if (!L->isLoopInvariant(Load->getPointerOperand()))
3666  return false;
3667  return true;
3668 }
3669 
3671  unsigned MaxLookup) {
3672  if (!V->getType()->isPointerTy())
3673  return V;
3674  for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3675  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3676  V = GEP->getPointerOperand();
3677  } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3678  Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3679  V = cast<Operator>(V)->getOperand(0);
3680  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3681  if (GA->isInterposable())
3682  return V;
3683  V = GA->getAliasee();
3684  } else if (isa<AllocaInst>(V)) {
3685  // An alloca can't be further simplified.
3686  return V;
3687  } else {
3688  if (auto *Call = dyn_cast<CallBase>(V)) {
3689  // CaptureTracking can know about special capturing properties of some
3690  // intrinsics like launder.invariant.group, that can't be expressed with
3691  // the attributes, but have properties like returning aliasing pointer.
3692  // Because some analysis may assume that nocaptured pointer is not
3693  // returned from some special intrinsic (because function would have to
3694  // be marked with returns attribute), it is crucial to use this function
3695  // because it should be in sync with CaptureTracking. Not using it may
3696  // cause weird miscompilations where 2 aliasing pointers are assumed to
3697  // noalias.
3698  if (auto *RP = getArgumentAliasingToReturnedPointer(Call)) {
3699  V = RP;
3700  continue;
3701  }
3702  }
3703 
3704  // See if InstructionSimplify knows any relevant tricks.
3705  if (Instruction *I = dyn_cast<Instruction>(V))
3706  // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3707  if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3708  V = Simplified;
3709  continue;
3710  }
3711 
3712  return V;
3713  }
3714  assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3715  }
3716  return V;
3717 }
3718 
3720  const DataLayout &DL, LoopInfo *LI,
3721  unsigned MaxLookup) {
3722  SmallPtrSet<Value *, 4> Visited;
3723  SmallVector<Value *, 4> Worklist;
3724  Worklist.push_back(V);
3725  do {
3726  Value *P = Worklist.pop_back_val();
3727  P = GetUnderlyingObject(P, DL, MaxLookup);
3728 
3729  if (!Visited.insert(P).second)
3730  continue;
3731 
3732  if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3733  Worklist.push_back(SI->getTrueValue());
3734  Worklist.push_back(SI->getFalseValue());
3735  continue;
3736  }
3737 
3738  if (PHINode *PN = dyn_cast<PHINode>(P)) {
3739  // If this PHI changes the underlying object in every iteration of the
3740  // loop, don't look through it. Consider:
3741  // int **A;
3742  // for (i) {
3743  // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3744  // Curr = A[i];
3745  // *Prev, *Curr;
3746  //
3747  // Prev is tracking Curr one iteration behind so they refer to different
3748  // underlying objects.
3749  if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3751  for (Value *IncValue : PN->incoming_values())
3752  Worklist.push_back(IncValue);
3753  continue;
3754  }
3755 
3756  Objects.push_back(P);
3757  } while (!Worklist.empty());
3758 }
3759 
3760 /// This is the function that does the work of looking through basic
3761 /// ptrtoint+arithmetic+inttoptr sequences.
3762 static const Value *getUnderlyingObjectFromInt(const Value *V) {
3763  do {
3764  if (const Operator *U = dyn_cast<Operator>(V)) {
3765  // If we find a ptrtoint, we can transfer control back to the
3766  // regular getUnderlyingObjectFromInt.
3767  if (U->getOpcode() == Instruction::PtrToInt)
3768  return U->getOperand(0);
3769  // If we find an add of a constant, a multiplied value, or a phi, it's
3770  // likely that the other operand will lead us to the base
3771  // object. We don't have to worry about the case where the
3772  // object address is somehow being computed by the multiply,
3773  // because our callers only care when the result is an
3774  // identifiable object.
3775  if (U->getOpcode() != Instruction::Add ||
3776  (!isa<ConstantInt>(U->getOperand(1)) &&
3777  Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
3778  !isa<PHINode>(U->getOperand(1))))
3779  return V;
3780  V = U->getOperand(0);
3781  } else {
3782  return V;
3783  }
3784  assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
3785  } while (true);
3786 }
3787 
3788 /// This is a wrapper around GetUnderlyingObjects and adds support for basic
3789 /// ptrtoint+arithmetic+inttoptr sequences.
3790 /// It returns false if unidentified object is found in GetUnderlyingObjects.
3792  SmallVectorImpl<Value *> &Objects,
3793  const DataLayout &DL) {
3795  SmallVector<const Value *, 4> Working(1, V);
3796  do {
3797  V = Working.pop_back_val();
3798 
3800  GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
3801 
3802  for (Value *V : Objs) {
3803  if (!Visited.insert(V).second)
3804  continue;
3805  if (Operator::getOpcode(V) == Instruction::IntToPtr) {
3806  const Value *O =
3807  getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
3808  if (O->getType()->isPointerTy()) {
3809  Working.push_back(O);
3810  continue;
3811  }
3812  }
3813  // If GetUnderlyingObjects fails to find an identifiable object,
3814  // getUnderlyingObjectsForCodeGen also fails for safety.
3815  if (!isIdentifiedObject(V)) {
3816  Objects.clear();
3817  return false;
3818  }
3819  Objects.push_back(const_cast<Value *>(V));
3820  }
3821  } while (!Working.empty());
3822  return true;
3823 }
3824 
3825 /// Return true if the only users of this pointer are lifetime markers.
3827  for (const User *U : V->users()) {
3828  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3829  if (!II) return false;
3830 
3831  if (!II->isLifetimeStartOrEnd())
3832  return false;
3833  }
3834  return true;
3835 }
3836 
3838  const Instruction *CtxI,
3839  const DominatorTree *DT) {
3840  const Operator *Inst = dyn_cast<Operator>(V);
3841  if (!Inst)
3842  return false;
3843 
3844  for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3845  if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3846  if (C->canTrap())
3847  return false;
3848 
3849  switch (Inst->getOpcode()) {
3850  default:
3851  return true;
3852  case Instruction::UDiv:
3853  case Instruction::URem: {
3854  // x / y is undefined if y == 0.
3855  const APInt *V;
3856  if (match(Inst->getOperand(1), m_APInt(V)))
3857  return *V != 0;
3858  return false;
3859  }
3860  case Instruction::SDiv:
3861  case Instruction::SRem: {
3862  // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3863  const APInt *Numerator, *Denominator;
3864  if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3865  return false;
3866  // We cannot hoist this division if the denominator is 0.
3867  if (*Denominator == 0)
3868  return false;
3869  // It's safe to hoist if the denominator is not 0 or -1.
3870  if (*Denominator != -1)
3871  return true;
3872  // At this point we know that the denominator is -1. It is safe to hoist as
3873  // long we know that the numerator is not INT_MIN.
3874  if (match(Inst->getOperand(0), m_APInt(Numerator)))
3875  return !Numerator->isMinSignedValue();
3876  // The numerator *might* be MinSignedValue.
3877  return false;
3878  }
3879  case Instruction::Load: {
3880  const LoadInst *LI = cast<LoadInst>(Inst);
3881  if (!LI->isUnordered() ||
3882  // Speculative load may create a race that did not exist in the source.
3883  LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3884  // Speculative load may load data from dirty regions.
3885  LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3886  LI->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3887  return false;
3888  const DataLayout &DL = LI->getModule()->getDataLayout();
3890  LI->getAlignment(), DL, CtxI, DT);
3891  }
3892  case Instruction::Call: {
3893  auto *CI = cast<const CallInst>(Inst);
3894  const Function *Callee = CI->getCalledFunction();
3895 
3896  // The called function could have undefined behavior or side-effects, even
3897  // if marked readnone nounwind.
3898  return Callee && Callee->isSpeculatable();
3899  }
3900  case Instruction::VAArg:
3901  case Instruction::Alloca:
3902  case Instruction::Invoke:
3903  case Instruction::PHI:
3904  case Instruction::Store:
3905  case Instruction::Ret:
3906  case Instruction::Br:
3907  case Instruction::IndirectBr:
3908  case Instruction::Switch:
3909  case Instruction::Unreachable:
3910  case Instruction::Fence:
3911  case Instruction::AtomicRMW:
3912  case Instruction::AtomicCmpXchg:
3913  case Instruction::LandingPad:
3914  case Instruction::Resume:
3915  case Instruction::CatchSwitch:
3916  case Instruction::CatchPad:
3917  case Instruction::CatchRet:
3918  case Instruction::CleanupPad:
3919  case Instruction::CleanupRet:
3920  return false; // Misc instructions which have effects
3921  }
3922 }
3923 
3926 }
3927 
3929  const Value *LHS, const Value *RHS, const DataLayout &DL,
3930  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
3931  bool UseInstrInfo) {
3932  // Multiplying n * m significant bits yields a result of n + m significant
3933  // bits. If the total number of significant bits does not exceed the
3934  // result bit width (minus 1), there is no overflow.
3935  // This means if we have enough leading zero bits in the operands
3936  // we can guarantee that the result does not overflow.
3937  // Ref: "Hacker's Delight" by Henry Warren
3938  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3939  KnownBits LHSKnown(BitWidth);
3940  KnownBits RHSKnown(BitWidth);
3941  computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT, nullptr,
3942  UseInstrInfo);
3943  computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT, nullptr,
3944  UseInstrInfo);
3945  // Note that underestimating the number of zero bits gives a more
3946  // conservative answer.
3947  unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
3948  RHSKnown.countMinLeadingZeros();
3949  // First handle the easy case: if we have enough zero bits there's
3950  // definitely no overflow.
3951  if (ZeroBits >= BitWidth)
3953 
3954  // Get the largest possible values for each operand.
3955  APInt LHSMax = ~LHSKnown.Zero;
3956  APInt RHSMax = ~RHSKnown.Zero;
3957 
3958  // We know the multiply operation doesn't overflow if the maximum values for
3959  // each operand will not overflow after we multiply them together.
3960  bool MaxOverflow;
3961  (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3962  if (!MaxOverflow)
3964 
3965  // We know it always overflows if multiplying the smallest possible values for
3966  // the operands also results in overflow.
3967  bool MinOverflow;
3968  (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3969  if (MinOverflow)
3971 
3973 }
3974 
3977  const DataLayout &DL, AssumptionCache *AC,
3978  const Instruction *CxtI,
3979  const DominatorTree *DT, bool UseInstrInfo) {
3980  // Multiplying n * m significant bits yields a result of n + m significant
3981  // bits. If the total number of significant bits does not exceed the
3982  // result bit width (minus 1), there is no overflow.
3983  // This means if we have enough leading sign bits in the operands
3984  // we can guarantee that the result does not overflow.
3985  // Ref: "Hacker's Delight" by Henry Warren
3986  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3987 
3988  // Note that underestimating the number of sign bits gives a more
3989  // conservative answer.
3990  unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
3991  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
3992 
3993  // First handle the easy case: if we have enough sign bits there's
3994  // definitely no overflow.
3995  if (SignBits > BitWidth + 1)
3997 
3998  // There are two ambiguous cases where there can be no overflow:
3999  // SignBits == BitWidth + 1 and
4000  // SignBits == BitWidth
4001  // The second case is difficult to check, therefore we only handle the
4002  // first case.
4003  if (SignBits == BitWidth + 1) {
4004  // It overflows only when both arguments are negative and the true
4005  // product is exactly the minimum negative number.
4006  // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
4007  // For simplicity we just check if at least one side is not negative.
4008  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4009  nullptr, UseInstrInfo);
4010  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4011  nullptr, UseInstrInfo);
4012  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
4014  }
4016 }
4017 
4019  const Value *LHS, const Value *RHS, const DataLayout &DL,
4020  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4021  bool UseInstrInfo) {
4022  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4023  nullptr, UseInstrInfo);
4024  if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
4025  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4026  nullptr, UseInstrInfo);
4027 
4028  if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
4029  // The sign bit is set in both cases: this MUST overflow.
4031  }
4032 
4033  if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
4034  // The sign bit is clear in both cases: this CANNOT overflow.
4036  }
4037  }
4038 
4040 }
4041 
4042 /// Return true if we can prove that adding the two values of the
4043 /// knownbits will not overflow.
4044 /// Otherwise return false.
4045 static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
4046  const KnownBits &RHSKnown) {
4047  // Addition of two 2's complement numbers having opposite signs will never
4048  // overflow.
4049  if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
4050  (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
4051  return true;
4052 
4053  // If either of the values is known to be non-negative, adding them can only
4054  // overflow if the second is also non-negative, so we can assume that.
4055  // Two non-negative numbers will only overflow if there is a carry to the
4056  // sign bit, so we can check if even when the values are as big as possible
4057  // there is no overflow to the sign bit.
4058  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
4059  APInt MaxLHS = ~LHSKnown.Zero;
4060  MaxLHS.clearSignBit();
4061  APInt MaxRHS = ~RHSKnown.Zero;
4062  MaxRHS.clearSignBit();
4063  APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
4064  return Result.isSignBitClear();
4065  }
4066 
4067  // If either of the values is known to be negative, adding them can only
4068  // overflow if the second is also negative, so we can assume that.
4069  // Two negative number will only overflow if there is no carry to the sign
4070  // bit, so we can check if even when the values are as small as possible
4071  // there is overflow to the sign bit.
4072  if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
4073  APInt MinLHS = LHSKnown.One;
4074  MinLHS.clearSignBit();
4075  APInt MinRHS = RHSKnown.One;
4076  MinRHS.clearSignBit();
4077  APInt Result = std::move(MinLHS) + std::move(MinRHS);
4078  return Result.isSignBitSet();
4079  }
4080 
4081  // If we reached here it means that we know nothing about the sign bits.
4082  // In this case we can't know if there will be an overflow, since by
4083  // changing the sign bits any two values can be made to overflow.
4084  return false;
4085 }
4086 
4088  const Value *RHS,
4089  const AddOperator *Add,
4090  const DataLayout &DL,
4091  AssumptionCache *AC,
4092  const Instruction *CxtI,
4093  const DominatorTree *DT) {
4094  if (Add && Add->hasNoSignedWrap()) {
4096  }
4097 
4098  // If LHS and RHS each have at least two sign bits, the addition will look
4099  // like
4100  //
4101  // XX..... +
4102  // YY.....
4103  //
4104  // If the carry into the most significant position is 0, X and Y can't both
4105  // be 1 and therefore the carry out of the addition is also 0.
4106  //
4107  // If the carry into the most significant position is 1, X and Y can't both
4108  // be 0 and therefore the carry out of the addition is also 1.
4109  //
4110  // Since the carry into the most significant position is always equal to
4111  // the carry out of the addition, there is no signed overflow.
4112  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4113  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4115 
4116  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
4117  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
4118 
4119  if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
4121 
4122  // The remaining code needs Add to be available. Early returns if not so.
4123  if (!Add)
4125 
4126  // If the sign of Add is the same as at least one of the operands, this add
4127  // CANNOT overflow. This is particularly useful when the sum is
4128  // @llvm.assume'ed non-negative rather than proved so from analyzing its
4129  // operands.
4130  bool LHSOrRHSKnownNonNegative =
4131  (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
4132  bool LHSOrRHSKnownNegative =
4133  (LHSKnown.isNegative() || RHSKnown.isNegative());
4134  if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
4135  KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
4136  if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
4137  (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
4139  }
4140  }
4141 
4143 }
4144 
4146  const Value *RHS,
4147  const DataLayout &DL,
4148  AssumptionCache *AC,
4149  const Instruction *CxtI,
4150  const DominatorTree *DT) {
4151  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
4152  if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
4153  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
4154 
4155  // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
4156  if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
4158 
4159  // If the LHS is non-negative and the RHS negative, we always wrap.
4160  if (LHSKnown.isNonNegative() && RHSKnown.isNegative())
4162  }
4163 
4165 }
4166 
4168  const Value *RHS,
4169  const DataLayout &DL,
4170  AssumptionCache *AC,
4171  const Instruction *CxtI,
4172  const DominatorTree *DT) {
4173  // If LHS and RHS each have at least two sign bits, the subtraction
4174  // cannot overflow.
4175  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4176  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4178 
4179  KnownBits LHSKnown = computeKnownBits(LHS, DL, 0, AC, CxtI, DT);
4180 
4181  KnownBits RHSKnown = computeKnownBits(RHS, DL, 0, AC, CxtI, DT);
4182 
4183  // Subtraction of two 2's complement numbers having identical signs will
4184  // never overflow.
4185  if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
4186  (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
4188 
4189  // TODO: implement logic similar to checkRippleForAdd
4191 }
4192 
4194  const DominatorTree &DT) {
4195 #ifndef NDEBUG
4196  auto IID = II->getIntrinsicID();
4197  assert((IID == Intrinsic::sadd_with_overflow ||
4198  IID == Intrinsic::uadd_with_overflow ||
4199  IID == Intrinsic::ssub_with_overflow ||
4200  IID == Intrinsic::usub_with_overflow ||
4201  IID == Intrinsic::smul_with_overflow ||
4202  IID == Intrinsic::umul_with_overflow) &&
4203  "Not an overflow intrinsic!");
4204 #endif
4205 
4206  SmallVector<const BranchInst *, 2> GuardingBranches;
4208 
4209  for (const User *U : II->users()) {
4210  if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
4211  assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
4212 
4213  if (EVI->getIndices()[0] == 0)
4214  Results.push_back(EVI);
4215  else {
4216  assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
4217 
4218  for (const auto *U : EVI->users())
4219  if (const auto *B = dyn_cast<BranchInst>(U)) {
4220  assert(B->isConditional() && "How else is it using an i1?");
4221  GuardingBranches.push_back(B);
4222  }
4223  }
4224  } else {
4225  // We are using the aggregate directly in a way we don't want to analyze
4226  // here (storing it to a global, say).
4227  return false;
4228  }
4229  }
4230 
4231  auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4232  BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4233  if (!NoWrapEdge.isSingleEdge())
4234  return false;
4235 
4236  // Check if all users of the add are provably no-wrap.
4237  for (const auto *Result : Results) {
4238  // If the extractvalue itself is not executed on overflow, the we don't
4239  // need to check each use separately, since domination is transitive.
4240  if (DT.dominates(NoWrapEdge, Result->getParent()))
4241  continue;
4242 
4243  for (auto &RU : Result->uses())
4244  if (!DT.dominates(NoWrapEdge, RU))
4245  return false;
4246  }
4247 
4248  return true;
4249  };
4250 
4251  return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4252 }
4253 
4254 
4256  const DataLayout &DL,
4257  AssumptionCache *AC,
4258  const Instruction *CxtI,
4259  const DominatorTree *DT) {
4261  Add, DL, AC, CxtI, DT);
4262 }
4263 
4265  const Value *RHS,
4266  const DataLayout &DL,
4267  AssumptionCache *AC,
4268  const Instruction *CxtI,
4269  const DominatorTree *DT) {
4270  return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
4271 }
4272 
4274  // A memory operation returns normally if it isn't volatile. A volatile
4275  // operation is allowed to trap.
4276  //
4277  // An atomic operation isn't guaranteed to return in a reasonable amount of
4278  // time because it's possible for another thread to interfere with it for an
4279  // arbitrary length of time, but programs aren't allowed to rely on that.
4280  if (const LoadInst *LI = dyn_cast<LoadInst>(I))
4281  return !LI->isVolatile();
4282  if (const StoreInst *SI = dyn_cast<StoreInst>(I))
4283  return !SI->isVolatile();
4284  if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
4285  return !CXI->isVolatile();
4286  if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
4287  return !RMWI->isVolatile();
4288  if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
4289  return !MII->isVolatile();
4290 
4291  // If there is no successor, then execution can't transfer to it.
4292  if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
4293  return !CRI->unwindsToCaller();
4294  if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
4295  return !CatchSwitch->unwindsToCaller();
4296  if (isa<ResumeInst>(I))
4297  return false;
4298  if (isa<ReturnInst>(I))
4299  return false;
4300  if (isa<UnreachableInst>(I))
4301  return false;
4302 
4303  // Calls can throw, or contain an infinite loop, or kill the process.
4304  if (auto CS = ImmutableCallSite(I)) {
4305  // Call sites that throw have implicit non-local control flow.
4306  if (!CS.doesNotThrow())
4307  return false;
4308 
4309  // Non-throwing call sites can loop infinitely, call exit/pthread_exit
4310  // etc. and thus not return. However, LLVM already assumes that
4311  //
4312  // - Thread exiting actions are modeled as writes to memory invisible to
4313  // the program.
4314  //
4315  // - Loops that don't have side effects (side effects are volatile/atomic
4316  // stores and IO) always terminate (see http://llvm.org/PR965).
4317  // Furthermore IO itself is also modeled as writes to memory invisible to
4318  // the program.
4319  //
4320  // We rely on those assumptions here, and use the memory effects of the call
4321  // target as a proxy for checking that it always returns.
4322 
4323  // FIXME: This isn't aggressive enough; a call which only writes to a global
4324  // is guaranteed to return.
4325  return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
4326  match(I, m_Intrinsic<Intrinsic::assume>()) ||
4327  match(I, m_Intrinsic<Intrinsic::sideeffect>());
4328  }
4329 
4330  // Other instructions return normally.
4331  return true;
4332 }
4333 
4335  // TODO: This is slightly consdervative for invoke instruction since exiting
4336  // via an exception *is* normal control for them.
4337  for (auto I = BB->begin(), E = BB->end(); I != E; ++I)
4339  return false;
4340  return true;
4341 }
4342 
4344  const Loop *L) {
4345  // The loop header is guaranteed to be executed for every iteration.
4346  //
4347  // FIXME: Relax this constraint to cover all basic blocks that are
4348  // guaranteed to be executed at every iteration.
4349  if (I->getParent() != L->getHeader()) return false;
4350 
4351  for (const Instruction &LI : *L->getHeader()) {
4352  if (&LI == I) return true;
4353  if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
4354  }
4355  llvm_unreachable("Instruction not contained in its own parent basic block.");
4356 }
4357 
4359  switch (I->getOpcode()) {
4360  case Instruction::Add:
4361  case Instruction::Sub:
4362  case Instruction::Xor:
4363  case Instruction::Trunc:
4364  case Instruction::BitCast:
4365  case Instruction::AddrSpaceCast:
4366  case Instruction::Mul:
4367  case Instruction::Shl:
4368  case Instruction::GetElementPtr:
4369  // These operations all propagate poison unconditionally. Note that poison
4370  // is not any particular value, so xor or subtraction of poison with
4371  // itself still yields poison, not zero.
4372  return true;
4373 
4374  case Instruction::AShr:
4375  case Instruction::SExt:
4376  // For these operations, one bit of the input is replicated across
4377  // multiple output bits. A replicated poison bit is still poison.
4378  return true;
4379 
4380  case Instruction::ICmp:
4381  // Comparing poison with any value yields poison. This is why, for
4382  // instance, x s< (x +nsw 1) can be folded to true.
4383  return true;
4384 
4385  default:
4386  return false;
4387  }
4388 }
4389 
4391  switch (I->getOpcode()) {
4392  case Instruction::Store:
4393  return cast<StoreInst>(I)->getPointerOperand();
4394 
4395  case Instruction::Load:
4396  return cast<LoadInst>(I)->getPointerOperand();
4397 
4398  case Instruction::AtomicCmpXchg:
4399  return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
4400 
4401  case Instruction::AtomicRMW:
4402  return cast<AtomicRMWInst>(I)->getPointerOperand();
4403 
4404  case Instruction::UDiv:
4405  case Instruction::SDiv:
4406  case Instruction::URem:
4407  case Instruction::SRem:
4408  return I->getOperand(1);
4409 
4410  default:
4411  return nullptr;
4412  }
4413 }
4414 
4416  // We currently only look for uses of poison values within the same basic
4417  // block, as that makes it easier to guarantee that the uses will be
4418  // executed given that PoisonI is executed.
4419  //
4420  // FIXME: Expand this to consider uses beyond the same basic block. To do
4421  // this, look out for the distinction between post-dominance and strong
4422  // post-dominance.
4423  const BasicBlock *BB = PoisonI->getParent();
4424 
4425  // Set of instructions that we have proved will yield poison if PoisonI
4426  // does.
4427  SmallSet<const Value *, 16> YieldsPoison;
4429  YieldsPoison.insert(PoisonI);
4430  Visited.insert(PoisonI->getParent());
4431 
4432  BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
4433 
4434  unsigned Iter = 0;
4435  while (Iter++ < MaxDepth) {
4436  for (auto &I : make_range(Begin, End)) {
4437  if (&I != PoisonI) {
4438  const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
4439  if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
4440  return true;
4442  return false;
4443  }
4444 
4445  // Mark poison that propagates from I through uses of I.
4446  if (YieldsPoison.count(&I)) {
4447  for (const User *User : I.users()) {
4448  const Instruction *UserI = cast<Instruction>(User);
4449  if (propagatesFullPoison(UserI))
4450  YieldsPoison.insert(User);
4451  }
4452  }
4453  }
4454 
4455  if (auto *NextBB = BB->getSingleSuccessor()) {
4456  if (Visited.insert(NextBB).second) {
4457  BB = NextBB;
4458  Begin = BB->getFirstNonPHI()->getIterator();
4459  End = BB->end();
4460  continue;
4461  }
4462  }
4463 
4464  break;
4465  }
4466  return false;
4467 }
4468 
4469 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
4470  if (FMF.noNaNs())
4471  return true;
4472 
4473  if (auto *C = dyn_cast<ConstantFP>(V))
4474  return !C->isNaN();
4475 
4476  if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4477  if (!C->getElementType()->isFloatingPointTy())
4478  return false;
4479  for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4480  if (C->getElementAsAPFloat(I).isNaN())
4481  return false;
4482  }
4483  return true;
4484  }
4485 
4486  return false;
4487 }
4488 
4489 static bool isKnownNonZero(const Value *V) {
4490  if (auto *C = dyn_cast<ConstantFP>(V))
4491  return !C->isZero();
4492 
4493  if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4494  if (!C->getElementType()->isFloatingPointTy())
4495  return false;
4496  for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4497  if (C->getElementAsAPFloat(I).isZero())
4498  return false;
4499  }
4500  return true;
4501  }
4502 
4503  return false;
4504 }
4505 
4506 /// Match clamp pattern for float types without care about NaNs or signed zeros.
4507 /// Given non-min/max outer cmp/select from the clamp pattern this
4508 /// function recognizes if it can be substitued by a "canonical" min/max
4509 /// pattern.
4511  Value *CmpLHS, Value *CmpRHS,
4512  Value *TrueVal, Value *FalseVal,
4513  Value *&LHS, Value *&RHS) {
4514  // Try to match
4515  // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
4516  // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
4517  // and return description of the outer Max/Min.
4518 
4519  // First, check if select has inverse order:
4520  if (CmpRHS == FalseVal) {
4521  std::swap(TrueVal, FalseVal);
4522  Pred = CmpInst::getInversePredicate(Pred);
4523  }
4524 
4525  // Assume success now. If there's no match, callers should not use these anyway.
4526  LHS = TrueVal;
4527  RHS = FalseVal;
4528 
4529  const APFloat *FC1;
4530  if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
4531  return {SPF_UNKNOWN, SPNB_NA, false};
4532 
4533  const APFloat *FC2;
4534  switch (Pred) {
4535  case CmpInst::FCMP_OLT:
4536  case CmpInst::FCMP_OLE:
4537  case CmpInst::FCMP_ULT:
4538  case CmpInst::FCMP_ULE:
4539  if (match(FalseVal,
4540  m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
4541  m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4542  FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
4543  return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
4544  break;
4545  case CmpInst::FCMP_OGT:
4546  case CmpInst::FCMP_OGE:
4547  case CmpInst::FCMP_UGT:
4548  case CmpInst::FCMP_UGE:
4549  if (match(FalseVal,
4550  m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
4551  m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4552  FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
4553  return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
4554  break;
4555  default:
4556  break;
4557  }
4558 
4559  return {SPF_UNKNOWN, SPNB_NA, false};
4560 }
4561 
4562 /// Recognize variations of:
4563 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
4565  Value *CmpLHS, Value *CmpRHS,
4566  Value *TrueVal, Value *FalseVal) {
4567  // Swap the select operands and predicate to match the patterns below.
4568  if (CmpRHS != TrueVal) {
4569  Pred = ICmpInst::getSwappedPredicate(Pred);
4570  std::swap(TrueVal, FalseVal);
4571  }
4572  const APInt *C1;
4573  if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
4574  const APInt *C2;
4575  // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
4576  if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4577  C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
4578  return {SPF_SMAX, SPNB_NA, false};
4579 
4580  // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
4581  if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4582  C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
4583  return {SPF_SMIN, SPNB_NA, false};
4584 
4585  // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
4586  if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4587  C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
4588  return {SPF_UMAX, SPNB_NA, false};
4589 
4590  // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
4591  if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4592  C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
4593  return {SPF_UMIN, SPNB_NA, false};
4594  }
4595  return {SPF_UNKNOWN, SPNB_NA, false};
4596 }
4597 
4598 /// Recognize variations of:
4599 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
4601  Value *CmpLHS, Value *CmpRHS,
4602  Value *TVal, Value *FVal,
4603  unsigned Depth) {
4604  // TODO: Allow FP min/max with nnan/nsz.
4605  assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
4606 
4607  Value *A, *B;
4608  SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
4610  return {SPF_UNKNOWN, SPNB_NA, false};
4611 
4612  Value *C, *D;
4613  SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
4614  if (L.Flavor != R.Flavor)
4615  return {SPF_UNKNOWN, SPNB_NA, false};
4616 
4617  // We have something like: x Pred y ? min(a, b) : min(c, d).
4618  // Try to match the compare to the min/max operations of the select operands.
4619  // First, make sure we have the right compare predicate.
4620  switch (L.Flavor) {
4621  case SPF_SMIN:
4622  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
4623  Pred = ICmpInst::getSwappedPredicate(Pred);
4624  std::swap(CmpLHS, CmpRHS);
4625  }
4626  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
4627  break;
4628  return {SPF_UNKNOWN, SPNB_NA, false};
4629  case SPF_SMAX:
4630  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
4631  Pred = ICmpInst::getSwappedPredicate(Pred);
4632  std::swap(CmpLHS, CmpRHS);
4633  }
4634  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
4635  break;
4636  return {SPF_UNKNOWN, SPNB_NA, false};
4637  case SPF_UMIN:
4638  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
4639  Pred = ICmpInst::getSwappedPredicate(Pred);
4640  std::swap(CmpLHS, CmpRHS);
4641  }
4642  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
4643  break;
4644  return {SPF_UNKNOWN, SPNB_NA, false};
4645  case SPF_UMAX:
4646  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
4647  Pred = ICmpInst::getSwappedPredicate(Pred);
4648  std::swap(CmpLHS, CmpRHS);
4649  }
4650  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
4651  break;
4652  return {SPF_UNKNOWN, SPNB_NA, false};
4653  default:
4654  return {SPF_UNKNOWN, SPNB_NA, false};
4655  }
4656 
4657  // If there is a common operand in the already matched min/max and the other
4658  // min/max operands match the compare operands (either directly or inverted),
4659  // then this is min/max of the same flavor.
4660 
4661  // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4662  // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4663  if (D == B) {
4664  if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4665  match(A, m_Not(m_Specific(CmpRHS)))))
4666  return {L.Flavor, SPNB_NA, false};
4667  }
4668  // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4669  // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4670  if (C == B) {
4671  if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4672  match(A, m_Not(m_Specific(CmpRHS)))))
4673  return {L.Flavor, SPNB_NA, false};
4674  }
4675  // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4676  // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4677  if (D == A) {
4678  if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4679  match(B, m_Not(m_Specific(CmpRHS)))))
4680  return {L.Flavor, SPNB_NA, false};
4681  }
4682  // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4683  // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4684  if (C == A) {
4685  if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4686  match(B, m_Not(m_Specific(CmpRHS)))))
4687  return {L.Flavor, SPNB_NA, false};
4688  }
4689 
4690  return {SPF_UNKNOWN, SPNB_NA, false};
4691 }
4692 
4693 /// Match non-obvious integer minimum and maximum sequences.
4695  Value *CmpLHS, Value *CmpRHS,
4696  Value *TrueVal, Value *FalseVal,
4697  Value *&LHS, Value *&RHS,
4698  unsigned Depth) {
4699  // Assume success. If there's no match, callers should not use these anyway.
4700  LHS = TrueVal;
4701  RHS = FalseVal;
4702 
4703  SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
4705  return SPR;
4706 
4707  SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
4709  return SPR;
4710 
4711  if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
4712  return {SPF_UNKNOWN, SPNB_NA, false};
4713 
4714  // Z = X -nsw Y
4715  // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
4716  // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
4717  if (match(TrueVal, m_Zero()) &&
4718  match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4719  return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4720 
4721  // Z = X -nsw Y
4722  // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
4723  // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
4724  if (match(FalseVal, m_Zero()) &&
4725  match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4726  return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4727 
4728  const APInt *C1;
4729  if (!match(CmpRHS, m_APInt(C1)))
4730  return {SPF_UNKNOWN, SPNB_NA, false};
4731 
4732  // An unsigned min/max can be written with a signed compare.
4733  const APInt *C2;
4734  if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
4735  (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
4736  // Is the sign bit set?
4737  // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
4738  // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
4739  if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
4740  C2->isMaxSignedValue())
4741  return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4742 
4743  // Is the sign bit clear?
4744  // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
4745  // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
4746  if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
4747  C2->isMinSignedValue())
4748  return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4749  }
4750 
4751  // Look through 'not' ops to find disguised signed min/max.
4752  // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4753  // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4754  if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4755  match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4756  return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4757 
4758  // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4759  // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4760  if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4761  match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4762  return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4763 
4764  return {SPF_UNKNOWN, SPNB_NA, false};
4765 }
4766 
4767 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
4768  assert(X && Y && "Invalid operand");
4769 
4770  // X = sub (0, Y) || X = sub nsw (0, Y)
4771  if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
4772  (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
4773  return true;
4774 
4775  // Y = sub (0, X) || Y = sub nsw (0, X)
4776  if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
4777  (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
4778  return true;
4779 
4780  // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
4781  Value *A, *B;
4782  return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
4783  match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
4784  (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
4785  match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
4786 }
4787 
4789  FastMathFlags FMF,
4790  Value *CmpLHS, Value *CmpRHS,
4791  Value *TrueVal, Value *FalseVal,
4792  Value *&LHS, Value *&RHS,
4793  unsigned Depth) {
4794  if (CmpInst::isFPPredicate(Pred)) {
4795  // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one
4796  // 0.0 operand, set the compare's 0.0 operands to that same value for the
4797  // purpose of identifying min/max. Disregard vector constants with undefined
4798  // elements because those can not be back-propagated for analysis.
4799  Value *OutputZeroVal = nullptr;
4800  if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) &&
4801  !cast<Constant>(TrueVal)->containsUndefElement())
4802  OutputZeroVal = TrueVal;
4803  else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) &&
4804  !cast<Constant>(FalseVal)->containsUndefElement())
4805  OutputZeroVal = FalseVal;
4806 
4807  if (OutputZeroVal) {
4808  if (match(CmpLHS, m_AnyZeroFP()))
4809  CmpLHS = OutputZeroVal;
4810  if (match(CmpRHS, m_AnyZeroFP()))
4811  CmpRHS = OutputZeroVal;
4812  }
4813  }
4814 
4815  LHS = CmpLHS;
4816  RHS = CmpRHS;
4817 
4818  // Signed zero may return inconsistent results between implementations.
4819  // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4820  // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4821  // Therefore, we behave conservatively and only proceed if at least one of the
4822  // operands is known to not be zero or if we don't care about signed zero.
4823  switch (Pred) {
4824  default: break;
4825  // FIXME: Include OGT/OLT/UGT/ULT.
4828  if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4829  !isKnownNonZero(CmpRHS))
4830  return {SPF_UNKNOWN, SPNB_NA, false};
4831  }
4832 
4833  SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4834  bool Ordered = false;
4835 
4836  // When given one NaN and one non-NaN input:
4837  // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4838  // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4839  // ordered comparison fails), which could be NaN or non-NaN.
4840  // so here we discover exactly what NaN behavior is required/accepted.
4841  if (CmpInst::isFPPredicate(Pred)) {
4842  bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4843  bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4844 
4845  if (LHSSafe && RHSSafe) {
4846  // Both operands are known non-NaN.
4847  NaNBehavior = SPNB_RETURNS_ANY;
4848  } else if (CmpInst::isOrdered(Pred)) {
4849  // An ordered comparison will return false when given a NaN, so it
4850  // returns the RHS.
4851  Ordered = true;
4852  if (LHSSafe)
4853  // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4854  NaNBehavior = SPNB_RETURNS_NAN;
4855  else if (RHSSafe)
4856  NaNBehavior = SPNB_RETURNS_OTHER;
4857  else
4858  // Completely unsafe.
4859  return {SPF_UNKNOWN, SPNB_NA, false};
4860  } else {
4861  Ordered = false;
4862  // An unordered comparison will return true when given a NaN, so it
4863  // returns the LHS.
4864  if (LHSSafe)
4865  // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4866  NaNBehavior = SPNB_RETURNS_OTHER;
4867  else if (RHSSafe)
4868  NaNBehavior = SPNB_RETURNS_NAN;
4869  else
4870  // Completely unsafe.
4871  return {SPF_UNKNOWN, SPNB_NA, false};
4872  }
4873  }
4874 
4875  if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4876  std::swap(CmpLHS, CmpRHS);
4877  Pred = CmpInst::getSwappedPredicate(Pred);
4878  if (NaNBehavior == SPNB_RETURNS_NAN)
4879  NaNBehavior = SPNB_RETURNS_OTHER;
4880  else if (NaNBehavior == SPNB_RETURNS_OTHER)
4881  NaNBehavior = SPNB_RETURNS_NAN;
4882  Ordered = !Ordered;
4883  }
4884 
4885  // ([if]cmp X, Y) ? X : Y
4886  if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4887  switch (Pred) {
4888  default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4889  case ICmpInst::ICMP_UGT:
4890  case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4891  case ICmpInst::ICMP_SGT:
4892  case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4893  case ICmpInst::ICMP_ULT:
4894  case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4895  case ICmpInst::ICMP_SLT:
4896  case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4897  case FCmpInst::FCMP_UGT:
4898  case FCmpInst::FCMP_UGE:
4899  case FCmpInst::FCMP_OGT:
4900  case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4901  case FCmpInst::FCMP_ULT:
4902  case FCmpInst::FCMP_ULE:
4903  case FCmpInst::FCMP_OLT:
4904  case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4905  }
4906  }
4907 
4908  if (isKnownNegation(TrueVal, FalseVal)) {
4909  // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
4910  // match against either LHS or sext(LHS).
4911  auto MaybeSExtCmpLHS =
4912  m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
4913  auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
4914  auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
4915  if (match(TrueVal, MaybeSExtCmpLHS)) {
4916  // Set the return values. If the compare uses the negated value (-X >s 0),
4917  // swap the return values because the negated value is always 'RHS'.
4918  LHS = TrueVal;
4919  RHS = FalseVal;
4920  if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
4921  std::swap(LHS, RHS);
4922 
4923  // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
4924  // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
4925  if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4926  return {SPF_ABS, SPNB_NA, false};
4927 
4928  // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
4929  // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
4930  if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4931  return {SPF_NABS, SPNB_NA, false};
4932  }
4933  else if (match(FalseVal, MaybeSExtCmpLHS)) {
4934  // Set the return values. If the compare uses the negated value (-X >s 0),
4935  // swap the return values because the negated value is always 'RHS'.
4936  LHS = FalseVal;
4937  RHS = TrueVal;
4938  if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
4939  std::swap(LHS, RHS);
4940 
4941  // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
4942  // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
4943  if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4944  return {SPF_NABS, SPNB_NA, false};
4945 
4946  // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
4947  // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
4948  if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4949  return {SPF_ABS, SPNB_NA, false};
4950  }
4951  }
4952 
4953  if (CmpInst::isIntPredicate(Pred))
4954  return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
4955 
4956  // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
4957  // may return either -0.0 or 0.0, so fcmp/select pair has stricter
4958  // semantics than minNum. Be conservative in such case.
4959  if (NaNBehavior != SPNB_RETURNS_ANY ||
4960  (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4961  !isKnownNonZero(CmpRHS)))
4962  return {SPF_UNKNOWN, SPNB_NA, false};
4963 
4964  return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4965 }
4966 
4967 /// Helps to match a select pattern in case of a type mismatch.
4968 ///
4969 /// The function processes the case when type of true and false values of a
4970 /// select instruction differs from type of the cmp instruction operands because
4971 /// of a cast instruction. The function checks if it is legal to move the cast
4972 /// operation after "select". If yes, it returns the new second value of
4973 /// "select" (with the assumption that cast is moved):
4974 /// 1. As operand of cast instruction when both values of "select" are same cast
4975 /// instructions.
4976 /// 2. As restored constant (by applying reverse cast operation) when the first
4977 /// value of the "select" is a cast operation and the second value is a
4978 /// constant.
4979 /// NOTE: We return only the new second value because the first value could be
4980 /// accessed as operand of cast instruction.
4981 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4982  Instruction::CastOps *CastOp) {
4983  auto *Cast1 = dyn_cast<CastInst>(V1);
4984  if (!Cast1)
4985  return nullptr;
4986 
4987  *CastOp = Cast1->getOpcode();
4988  Type *SrcTy = Cast1->getSrcTy();
4989  if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4990  // If V1 and V2 are both the same cast from the same type, look through V1.
4991  if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4992  return Cast2->getOperand(0);
4993  return nullptr;
4994  }
4995 
4996  auto *C = dyn_cast<Constant>(V2);
4997  if (!C)
4998  return nullptr;
4999 
5000  Constant *CastedTo = nullptr;
5001  switch (*CastOp) {
5002  case Instruction::ZExt:
5003  if (CmpI->isUnsigned())
5004  CastedTo = ConstantExpr::getTrunc(C, SrcTy);
5005  break;
5006  case Instruction::SExt:
5007  if (CmpI->isSigned())
5008  CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
5009  break;
5010  case Instruction::Trunc:
5011  Constant *CmpConst;
5012  if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
5013  CmpConst->getType() == SrcTy) {
5014  // Here we have the following case:
5015  //
5016  // %cond = cmp iN %x, CmpConst
5017  // %tr = trunc iN %x to iK
5018  // %narrowsel = select i1 %cond, iK %t, iK C
5019  //
5020  // We can always move trunc after select operation:
5021  //
5022  // %cond = cmp iN %x, CmpConst
5023  // %widesel = select i1 %cond, iN %x, iN CmpConst
5024  // %tr = trunc iN %widesel to iK
5025  //
5026  // Note that C could be extended in any way because we don't care about
5027  // upper bits after truncation. It can't be abs pattern, because it would
5028  // look like:
5029  //
5030  // select i1 %cond, x, -x.
5031  //
5032  // So only min/max pattern could be matched. Such match requires widened C
5033  // == CmpConst. That is why set widened C = CmpConst, condition trunc
5034  // CmpConst == C is checked below.
5035  CastedTo = CmpConst;
5036  } else {
5037  CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
5038  }
5039  break;
5040  case Instruction::FPTrunc:
5041  CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
5042  break;
5043  case Instruction::FPExt:
5044  CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
5045  break;
5046  case Instruction::FPToUI:
5047  CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
5048  break;
5049  case Instruction::FPToSI:
5050  CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
5051  break;
5052  case Instruction::UIToFP:
5053  CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
5054  break;
5055  case Instruction::SIToFP:
5056  CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
5057  break;
5058  default:
5059  break;
5060  }
5061 
5062  if (!CastedTo)
5063  return nullptr;
5064 
5065  // Make sure the cast doesn't lose any information.
5066  Constant *CastedBack =
5067  ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
5068  if (CastedBack != C)
5069  return nullptr;
5070 
5071  return CastedTo;
5072 }
5073 
5075  Instruction::CastOps *CastOp,
5076  unsigned Depth) {
5077  if (Depth >= MaxDepth)
5078  return {SPF_UNKNOWN, SPNB_NA, false};
5079 
5081  if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
5082 
5083  CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
5084  if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
5085 
5086  CmpInst::Predicate Pred = CmpI->getPredicate();
5087  Value *CmpLHS = CmpI->getOperand(0);
5088  Value *CmpRHS = CmpI->getOperand(1);
5089  Value *TrueVal = SI->getTrueValue();
5090  Value *FalseVal = SI->getFalseValue();
5091  FastMathFlags FMF;
5092  if (isa<FPMathOperator>(CmpI))
5093  FMF = CmpI->getFastMathFlags();
5094 
5095  // Bail out early.
5096  if (CmpI->isEquality())
5097  return {SPF_UNKNOWN, SPNB_NA, false};
5098 
5099  // Deal with type mismatches.
5100  if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
5101  if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
5102  // If this is a potential fmin/fmax with a cast to integer, then ignore
5103  // -0.0 because there is no corresponding integer value.
5104  if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5105  FMF.setNoSignedZeros();
5106  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5107  cast<CastInst>(TrueVal)->getOperand(0), C,
5108  LHS, RHS, Depth);
5109  }
5110  if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
5111  // If this is a potential fmin/fmax with a cast to integer, then ignore
5112  // -0.0 because there is no corresponding integer value.
5113  if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5114  FMF.setNoSignedZeros();
5115  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5116  C, cast<CastInst>(FalseVal)->getOperand(0),
5117  LHS, RHS, Depth);
5118  }
5119  }
5120  return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
5121  LHS, RHS, Depth);
5122 }
5123 
5125  if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
5126  if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
5127  if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
5128  if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
5129  if (SPF == SPF_FMINNUM)
5130  return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
5131  if (SPF == SPF_FMAXNUM)
5132  return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
5133  llvm_unreachable("unhandled!");
5134 }
5135 
5137  if (SPF == SPF_SMIN) return SPF_SMAX;
5138  if (SPF == SPF_UMIN) return SPF_UMAX;
5139  if (SPF == SPF_SMAX) return SPF_SMIN;
5140  if (SPF == SPF_UMAX) return SPF_UMIN;
5141  llvm_unreachable("unhandled!");
5142 }
5143 
5146 }
5147 
5148 /// Return true if "icmp Pred LHS RHS" is always true.