LLVM  6.0.0svn
Reassociate.cpp
Go to the documentation of this file.
1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass reassociates commutative expressions in an order that is designed
11 // to promote better constant propagation, GCSE, LICM, PRE, etc.
12 //
13 // For example: 4 + (x + 5) -> x + (4 + 5)
14 //
15 // In the implementation of this algorithm, constants are assigned rank = 0,
16 // function arguments are rank = 1, and other values are assigned ranks
17 // corresponding to the reverse post order traversal of current function
18 // (starting at 2), which effectively gives values in deep loops higher rank
19 // than values not in loops.
20 //
21 //===----------------------------------------------------------------------===//
22 
24 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/Statistic.h"
31 #include "llvm/IR/CFG.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/Debug.h"
43 #include "llvm/Transforms/Scalar.h"
45 #include <algorithm>
46 using namespace llvm;
47 using namespace reassociate;
48 
49 #define DEBUG_TYPE "reassociate"
50 
51 STATISTIC(NumChanged, "Number of insts reassociated");
52 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
53 STATISTIC(NumFactor , "Number of multiplies factored");
54 
55 #ifndef NDEBUG
56 /// Print out the expression identified in the Ops list.
57 ///
59  Module *M = I->getModule();
60  dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
61  << *Ops[0].Op->getType() << '\t';
62  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
63  dbgs() << "[ ";
64  Ops[i].Op->printAsOperand(dbgs(), false, M);
65  dbgs() << ", #" << Ops[i].Rank << "] ";
66  }
67 }
68 #endif
69 
70 /// Utility class representing a non-constant Xor-operand. We classify
71 /// non-constant Xor-Operands into two categories:
72 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
73 /// C2)
74 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
75 /// constant.
76 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
77 /// operand as "E | 0"
79 public:
80  XorOpnd(Value *V);
81 
82  bool isInvalid() const { return SymbolicPart == nullptr; }
83  bool isOrExpr() const { return isOr; }
84  Value *getValue() const { return OrigVal; }
85  Value *getSymbolicPart() const { return SymbolicPart; }
86  unsigned getSymbolicRank() const { return SymbolicRank; }
87  const APInt &getConstPart() const { return ConstPart; }
88 
89  void Invalidate() { SymbolicPart = OrigVal = nullptr; }
90  void setSymbolicRank(unsigned R) { SymbolicRank = R; }
91 
92 private:
93  Value *OrigVal;
94  Value *SymbolicPart;
95  APInt ConstPart;
96  unsigned SymbolicRank;
97  bool isOr;
98 };
99 
101  assert(!isa<ConstantInt>(V) && "No ConstantInt");
102  OrigVal = V;
104  SymbolicRank = 0;
105 
106  if (I && (I->getOpcode() == Instruction::Or ||
107  I->getOpcode() == Instruction::And)) {
108  Value *V0 = I->getOperand(0);
109  Value *V1 = I->getOperand(1);
110  const APInt *C;
111  if (match(V0, PatternMatch::m_APInt(C)))
112  std::swap(V0, V1);
113 
114  if (match(V1, PatternMatch::m_APInt(C))) {
115  ConstPart = *C;
116  SymbolicPart = V0;
117  isOr = (I->getOpcode() == Instruction::Or);
118  return;
119  }
120  }
121 
122  // view the operand as "V | 0"
123  SymbolicPart = V;
124  ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits());
125  isOr = true;
126 }
127 
128 /// Return true if V is an instruction of the specified opcode and if it
129 /// only has one use.
130 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
131  if (V->hasOneUse() && isa<Instruction>(V) &&
132  cast<Instruction>(V)->getOpcode() == Opcode &&
133  (!isa<FPMathOperator>(V) ||
134  cast<Instruction>(V)->hasUnsafeAlgebra()))
135  return cast<BinaryOperator>(V);
136  return nullptr;
137 }
138 
139 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
140  unsigned Opcode2) {
141  if (V->hasOneUse() && isa<Instruction>(V) &&
142  (cast<Instruction>(V)->getOpcode() == Opcode1 ||
143  cast<Instruction>(V)->getOpcode() == Opcode2) &&
144  (!isa<FPMathOperator>(V) ||
145  cast<Instruction>(V)->hasUnsafeAlgebra()))
146  return cast<BinaryOperator>(V);
147  return nullptr;
148 }
149 
150 void ReassociatePass::BuildRankMap(Function &F,
152  unsigned Rank = 2;
153 
154  // Assign distinct ranks to function arguments.
155  for (auto &Arg : F.args()) {
156  ValueRankMap[&Arg] = ++Rank;
157  DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
158  << "\n");
159  }
160 
161  // Traverse basic blocks in ReversePostOrder
162  for (BasicBlock *BB : RPOT) {
163  unsigned BBRank = RankMap[BB] = ++Rank << 16;
164 
165  // Walk the basic block, adding precomputed ranks for any instructions that
166  // we cannot move. This ensures that the ranks for these instructions are
167  // all different in the block.
168  for (Instruction &I : *BB)
169  if (mayBeMemoryDependent(I))
170  ValueRankMap[&I] = ++BBRank;
171  }
172 }
173 
174 unsigned ReassociatePass::getRank(Value *V) {
176  if (!I) {
177  if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
178  return 0; // Otherwise it's a global or constant, rank 0.
179  }
180 
181  if (unsigned Rank = ValueRankMap[I])
182  return Rank; // Rank already known?
183 
184  // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
185  // we can reassociate expressions for code motion! Since we do not recurse
186  // for PHI nodes, we cannot have infinite recursion here, because there
187  // cannot be loops in the value graph that do not go through PHI nodes.
188  unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
189  for (unsigned i = 0, e = I->getNumOperands();
190  i != e && Rank != MaxRank; ++i)
191  Rank = std::max(Rank, getRank(I->getOperand(i)));
192 
193  // If this is a not or neg instruction, do not count it for rank. This
194  // assures us that X and ~X will have the same rank.
197  ++Rank;
198 
199  DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
200 
201  return ValueRankMap[I] = Rank;
202 }
203 
204 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
205 void ReassociatePass::canonicalizeOperands(Instruction *I) {
206  assert(isa<BinaryOperator>(I) && "Expected binary operator.");
207  assert(I->isCommutative() && "Expected commutative operator.");
208 
209  Value *LHS = I->getOperand(0);
210  Value *RHS = I->getOperand(1);
211  if (LHS == RHS || isa<Constant>(RHS))
212  return;
213  if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
214  cast<BinaryOperator>(I)->swapOperands();
215 }
216 
217 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
218  Instruction *InsertBefore, Value *FlagsOp) {
219  if (S1->getType()->isIntOrIntVectorTy())
220  return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
221  else {
222  BinaryOperator *Res =
223  BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
224  Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
225  return Res;
226  }
227 }
228 
229 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
230  Instruction *InsertBefore, Value *FlagsOp) {
231  if (S1->getType()->isIntOrIntVectorTy())
232  return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
233  else {
234  BinaryOperator *Res =
235  BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
236  Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
237  return Res;
238  }
239 }
240 
242  Instruction *InsertBefore, Value *FlagsOp) {
243  if (S1->getType()->isIntOrIntVectorTy())
244  return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
245  else {
246  BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
247  Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
248  return Res;
249  }
250 }
251 
252 /// Replace 0-X with X*-1.
254  Type *Ty = Neg->getType();
255  Constant *NegOne = Ty->isIntOrIntVectorTy() ?
257 
258  BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
259  Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
260  Res->takeName(Neg);
261  Neg->replaceAllUsesWith(Res);
262  Res->setDebugLoc(Neg->getDebugLoc());
263  return Res;
264 }
265 
266 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
267 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
268 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
269 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
270 /// even x in Bitwidth-bit arithmetic.
271 static unsigned CarmichaelShift(unsigned Bitwidth) {
272  if (Bitwidth < 3)
273  return Bitwidth - 1;
274  return Bitwidth - 2;
275 }
276 
277 /// Add the extra weight 'RHS' to the existing weight 'LHS',
278 /// reducing the combined weight using any special properties of the operation.
279 /// The existing weight LHS represents the computation X op X op ... op X where
280 /// X occurs LHS times. The combined weight represents X op X op ... op X with
281 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
282 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
283 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
284 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
285  // If we were working with infinite precision arithmetic then the combined
286  // weight would be LHS + RHS. But we are using finite precision arithmetic,
287  // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
288  // for nilpotent operations and addition, but not for idempotent operations
289  // and multiplication), so it is important to correctly reduce the combined
290  // weight back into range if wrapping would be wrong.
291 
292  // If RHS is zero then the weight didn't change.
293  if (RHS.isMinValue())
294  return;
295  // If LHS is zero then the combined weight is RHS.
296  if (LHS.isMinValue()) {
297  LHS = RHS;
298  return;
299  }
300  // From this point on we know that neither LHS nor RHS is zero.
301 
302  if (Instruction::isIdempotent(Opcode)) {
303  // Idempotent means X op X === X, so any non-zero weight is equivalent to a
304  // weight of 1. Keeping weights at zero or one also means that wrapping is
305  // not a problem.
306  assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
307  return; // Return a weight of 1.
308  }
309  if (Instruction::isNilpotent(Opcode)) {
310  // Nilpotent means X op X === 0, so reduce weights modulo 2.
311  assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
312  LHS = 0; // 1 + 1 === 0 modulo 2.
313  return;
314  }
315  if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
316  // TODO: Reduce the weight by exploiting nsw/nuw?
317  LHS += RHS;
318  return;
319  }
320 
321  assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
322  "Unknown associative operation!");
323  unsigned Bitwidth = LHS.getBitWidth();
324  // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
325  // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
326  // bit number x, since either x is odd in which case x^CM = 1, or x is even in
327  // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
328  // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
329  // which by a happy accident means that they can always be represented using
330  // Bitwidth bits.
331  // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
332  // the Carmichael number).
333  if (Bitwidth > 3) {
334  /// CM - The value of Carmichael's lambda function.
335  APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
336  // Any weight W >= Threshold can be replaced with W - CM.
337  APInt Threshold = CM + Bitwidth;
338  assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
339  // For Bitwidth 4 or more the following sum does not overflow.
340  LHS += RHS;
341  while (LHS.uge(Threshold))
342  LHS -= CM;
343  } else {
344  // To avoid problems with overflow do everything the same as above but using
345  // a larger type.
346  unsigned CM = 1U << CarmichaelShift(Bitwidth);
347  unsigned Threshold = CM + Bitwidth;
348  assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
349  "Weights not reduced!");
350  unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
351  while (Total >= Threshold)
352  Total -= CM;
353  LHS = Total;
354  }
355 }
356 
357 typedef std::pair<Value*, APInt> RepeatedValue;
358 
359 /// Given an associative binary expression, return the leaf
360 /// nodes in Ops along with their weights (how many times the leaf occurs). The
361 /// original expression is the same as
362 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
363 /// op
364 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
365 /// op
366 /// ...
367 /// op
368 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
369 ///
370 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
371 ///
372 /// This routine may modify the function, in which case it returns 'true'. The
373 /// changes it makes may well be destructive, changing the value computed by 'I'
374 /// to something completely different. Thus if the routine returns 'true' then
375 /// you MUST either replace I with a new expression computed from the Ops array,
376 /// or use RewriteExprTree to put the values back in.
377 ///
378 /// A leaf node is either not a binary operation of the same kind as the root
379 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
380 /// opcode), or is the same kind of binary operator but has a use which either
381 /// does not belong to the expression, or does belong to the expression but is
382 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
383 /// of the expression, while for non-leaf nodes (except for the root 'I') every
384 /// use is a non-leaf node of the expression.
385 ///
386 /// For example:
387 /// expression graph node names
388 ///
389 /// + | I
390 /// / \ |
391 /// + + | A, B
392 /// / \ / \ |
393 /// * + * | C, D, E
394 /// / \ / \ / \ |
395 /// + * | F, G
396 ///
397 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
398 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
399 ///
400 /// The expression is maximal: if some instruction is a binary operator of the
401 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
402 /// then the instruction also belongs to the expression, is not a leaf node of
403 /// it, and its operands also belong to the expression (but may be leaf nodes).
404 ///
405 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
406 /// order to ensure that every non-root node in the expression has *exactly one*
407 /// use by a non-leaf node of the expression. This destruction means that the
408 /// caller MUST either replace 'I' with a new expression or use something like
409 /// RewriteExprTree to put the values back in if the routine indicates that it
410 /// made a change by returning 'true'.
411 ///
412 /// In the above example either the right operand of A or the left operand of B
413 /// will be replaced by undef. If it is B's operand then this gives:
414 ///
415 /// + | I
416 /// / \ |
417 /// + + | A, B - operand of B replaced with undef
418 /// / \ \ |
419 /// * + * | C, D, E
420 /// / \ / \ / \ |
421 /// + * | F, G
422 ///
423 /// Note that such undef operands can only be reached by passing through 'I'.
424 /// For example, if you visit operands recursively starting from a leaf node
425 /// then you will never see such an undef operand unless you get back to 'I',
426 /// which requires passing through a phi node.
427 ///
428 /// Note that this routine may also mutate binary operators of the wrong type
429 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
430 /// of the expression) if it can turn them into binary operators of the right
431 /// type and thus make the expression bigger.
432 
435  DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
436  unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
437  unsigned Opcode = I->getOpcode();
438  assert(I->isAssociative() && I->isCommutative() &&
439  "Expected an associative and commutative operation!");
440 
441  // Visit all operands of the expression, keeping track of their weight (the
442  // number of paths from the expression root to the operand, or if you like
443  // the number of times that operand occurs in the linearized expression).
444  // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
445  // while A has weight two.
446 
447  // Worklist of non-leaf nodes (their operands are in the expression too) along
448  // with their weights, representing a certain number of paths to the operator.
449  // If an operator occurs in the worklist multiple times then we found multiple
450  // ways to get to it.
451  SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
452  Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
453  bool Changed = false;
454 
455  // Leaves of the expression are values that either aren't the right kind of
456  // operation (eg: a constant, or a multiply in an add tree), or are, but have
457  // some uses that are not inside the expression. For example, in I = X + X,
458  // X = A + B, the value X has two uses (by I) that are in the expression. If
459  // X has any other uses, for example in a return instruction, then we consider
460  // X to be a leaf, and won't analyze it further. When we first visit a value,
461  // if it has more than one use then at first we conservatively consider it to
462  // be a leaf. Later, as the expression is explored, we may discover some more
463  // uses of the value from inside the expression. If all uses turn out to be
464  // from within the expression (and the value is a binary operator of the right
465  // kind) then the value is no longer considered to be a leaf, and its operands
466  // are explored.
467 
468  // Leaves - Keeps track of the set of putative leaves as well as the number of
469  // paths to each leaf seen so far.
470  typedef DenseMap<Value*, APInt> LeafMap;
471  LeafMap Leaves; // Leaf -> Total weight so far.
472  SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
473 
474 #ifndef NDEBUG
475  SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
476 #endif
477  while (!Worklist.empty()) {
478  std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
479  I = P.first; // We examine the operands of this binary operator.
480 
481  for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
482  Value *Op = I->getOperand(OpIdx);
483  APInt Weight = P.second; // Number of paths to this operand.
484  DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
485  assert(!Op->use_empty() && "No uses, so how did we get to it?!");
486 
487  // If this is a binary operation of the right kind with only one use then
488  // add its operands to the expression.
489  if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
490  assert(Visited.insert(Op).second && "Not first visit!");
491  DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
492  Worklist.push_back(std::make_pair(BO, Weight));
493  continue;
494  }
495 
496  // Appears to be a leaf. Is the operand already in the set of leaves?
497  LeafMap::iterator It = Leaves.find(Op);
498  if (It == Leaves.end()) {
499  // Not in the leaf map. Must be the first time we saw this operand.
500  assert(Visited.insert(Op).second && "Not first visit!");
501  if (!Op->hasOneUse()) {
502  // This value has uses not accounted for by the expression, so it is
503  // not safe to modify. Mark it as being a leaf.
504  DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
505  LeafOrder.push_back(Op);
506  Leaves[Op] = Weight;
507  continue;
508  }
509  // No uses outside the expression, try morphing it.
510  } else {
511  // Already in the leaf map.
512  assert(It != Leaves.end() && Visited.count(Op) &&
513  "In leaf map but not visited!");
514 
515  // Update the number of paths to the leaf.
516  IncorporateWeight(It->second, Weight, Opcode);
517 
518 #if 0 // TODO: Re-enable once PR13021 is fixed.
519  // The leaf already has one use from inside the expression. As we want
520  // exactly one such use, drop this new use of the leaf.
521  assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
522  I->setOperand(OpIdx, UndefValue::get(I->getType()));
523  Changed = true;
524 
525  // If the leaf is a binary operation of the right kind and we now see
526  // that its multiple original uses were in fact all by nodes belonging
527  // to the expression, then no longer consider it to be a leaf and add
528  // its operands to the expression.
529  if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
530  DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
531  Worklist.push_back(std::make_pair(BO, It->second));
532  Leaves.erase(It);
533  continue;
534  }
535 #endif
536 
537  // If we still have uses that are not accounted for by the expression
538  // then it is not safe to modify the value.
539  if (!Op->hasOneUse())
540  continue;
541 
542  // No uses outside the expression, try morphing it.
543  Weight = It->second;
544  Leaves.erase(It); // Since the value may be morphed below.
545  }
546 
547  // At this point we have a value which, first of all, is not a binary
548  // expression of the right kind, and secondly, is only used inside the
549  // expression. This means that it can safely be modified. See if we
550  // can usefully morph it into an expression of the right kind.
551  assert((!isa<Instruction>(Op) ||
552  cast<Instruction>(Op)->getOpcode() != Opcode
553  || (isa<FPMathOperator>(Op) &&
554  !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
555  "Should have been handled above!");
556  assert(Op->hasOneUse() && "Has uses outside the expression tree!");
557 
558  // If this is a multiply expression, turn any internal negations into
559  // multiplies by -1 so they can be reassociated.
560  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
561  if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
562  (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
563  DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
564  BO = LowerNegateToMultiply(BO);
565  DEBUG(dbgs() << *BO << '\n');
566  Worklist.push_back(std::make_pair(BO, Weight));
567  Changed = true;
568  continue;
569  }
570 
571  // Failed to morph into an expression of the right type. This really is
572  // a leaf.
573  DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
574  assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
575  LeafOrder.push_back(Op);
576  Leaves[Op] = Weight;
577  }
578  }
579 
580  // The leaves, repeated according to their weights, represent the linearized
581  // form of the expression.
582  for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
583  Value *V = LeafOrder[i];
584  LeafMap::iterator It = Leaves.find(V);
585  if (It == Leaves.end())
586  // Node initially thought to be a leaf wasn't.
587  continue;
588  assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
589  APInt Weight = It->second;
590  if (Weight.isMinValue())
591  // Leaf already output or weight reduction eliminated it.
592  continue;
593  // Ensure the leaf is only output once.
594  It->second = 0;
595  Ops.push_back(std::make_pair(V, Weight));
596  }
597 
598  // For nilpotent operations or addition there may be no operands, for example
599  // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
600  // in both cases the weight reduces to 0 causing the value to be skipped.
601  if (Ops.empty()) {
602  Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
603  assert(Identity && "Associative operation without identity!");
604  Ops.emplace_back(Identity, APInt(Bitwidth, 1));
605  }
606 
607  return Changed;
608 }
609 
610 /// Now that the operands for this expression tree are
611 /// linearized and optimized, emit them in-order.
612 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
614  assert(Ops.size() > 1 && "Single values should be used directly!");
615 
616  // Since our optimizations should never increase the number of operations, the
617  // new expression can usually be written reusing the existing binary operators
618  // from the original expression tree, without creating any new instructions,
619  // though the rewritten expression may have a completely different topology.
620  // We take care to not change anything if the new expression will be the same
621  // as the original. If more than trivial changes (like commuting operands)
622  // were made then we are obliged to clear out any optional subclass data like
623  // nsw flags.
624 
625  /// NodesToRewrite - Nodes from the original expression available for writing
626  /// the new expression into.
627  SmallVector<BinaryOperator*, 8> NodesToRewrite;
628  unsigned Opcode = I->getOpcode();
629  BinaryOperator *Op = I;
630 
631  /// NotRewritable - The operands being written will be the leaves of the new
632  /// expression and must not be used as inner nodes (via NodesToRewrite) by
633  /// mistake. Inner nodes are always reassociable, and usually leaves are not
634  /// (if they were they would have been incorporated into the expression and so
635  /// would not be leaves), so most of the time there is no danger of this. But
636  /// in rare cases a leaf may become reassociable if an optimization kills uses
637  /// of it, or it may momentarily become reassociable during rewriting (below)
638  /// due it being removed as an operand of one of its uses. Ensure that misuse
639  /// of leaf nodes as inner nodes cannot occur by remembering all of the future
640  /// leaves and refusing to reuse any of them as inner nodes.
641  SmallPtrSet<Value*, 8> NotRewritable;
642  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
643  NotRewritable.insert(Ops[i].Op);
644 
645  // ExpressionChanged - Non-null if the rewritten expression differs from the
646  // original in some non-trivial way, requiring the clearing of optional flags.
647  // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
648  BinaryOperator *ExpressionChanged = nullptr;
649  for (unsigned i = 0; ; ++i) {
650  // The last operation (which comes earliest in the IR) is special as both
651  // operands will come from Ops, rather than just one with the other being
652  // a subexpression.
653  if (i+2 == Ops.size()) {
654  Value *NewLHS = Ops[i].Op;
655  Value *NewRHS = Ops[i+1].Op;
656  Value *OldLHS = Op->getOperand(0);
657  Value *OldRHS = Op->getOperand(1);
658 
659  if (NewLHS == OldLHS && NewRHS == OldRHS)
660  // Nothing changed, leave it alone.
661  break;
662 
663  if (NewLHS == OldRHS && NewRHS == OldLHS) {
664  // The order of the operands was reversed. Swap them.
665  DEBUG(dbgs() << "RA: " << *Op << '\n');
666  Op->swapOperands();
667  DEBUG(dbgs() << "TO: " << *Op << '\n');
668  MadeChange = true;
669  ++NumChanged;
670  break;
671  }
672 
673  // The new operation differs non-trivially from the original. Overwrite
674  // the old operands with the new ones.
675  DEBUG(dbgs() << "RA: " << *Op << '\n');
676  if (NewLHS != OldLHS) {
677  BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
678  if (BO && !NotRewritable.count(BO))
679  NodesToRewrite.push_back(BO);
680  Op->setOperand(0, NewLHS);
681  }
682  if (NewRHS != OldRHS) {
683  BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
684  if (BO && !NotRewritable.count(BO))
685  NodesToRewrite.push_back(BO);
686  Op->setOperand(1, NewRHS);
687  }
688  DEBUG(dbgs() << "TO: " << *Op << '\n');
689 
690  ExpressionChanged = Op;
691  MadeChange = true;
692  ++NumChanged;
693 
694  break;
695  }
696 
697  // Not the last operation. The left-hand side will be a sub-expression
698  // while the right-hand side will be the current element of Ops.
699  Value *NewRHS = Ops[i].Op;
700  if (NewRHS != Op->getOperand(1)) {
701  DEBUG(dbgs() << "RA: " << *Op << '\n');
702  if (NewRHS == Op->getOperand(0)) {
703  // The new right-hand side was already present as the left operand. If
704  // we are lucky then swapping the operands will sort out both of them.
705  Op->swapOperands();
706  } else {
707  // Overwrite with the new right-hand side.
708  BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
709  if (BO && !NotRewritable.count(BO))
710  NodesToRewrite.push_back(BO);
711  Op->setOperand(1, NewRHS);
712  ExpressionChanged = Op;
713  }
714  DEBUG(dbgs() << "TO: " << *Op << '\n');
715  MadeChange = true;
716  ++NumChanged;
717  }
718 
719  // Now deal with the left-hand side. If this is already an operation node
720  // from the original expression then just rewrite the rest of the expression
721  // into it.
722  BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
723  if (BO && !NotRewritable.count(BO)) {
724  Op = BO;
725  continue;
726  }
727 
728  // Otherwise, grab a spare node from the original expression and use that as
729  // the left-hand side. If there are no nodes left then the optimizers made
730  // an expression with more nodes than the original! This usually means that
731  // they did something stupid but it might mean that the problem was just too
732  // hard (finding the mimimal number of multiplications needed to realize a
733  // multiplication expression is NP-complete). Whatever the reason, smart or
734  // stupid, create a new node if there are none left.
735  BinaryOperator *NewOp;
736  if (NodesToRewrite.empty()) {
739  Undef, Undef, "", I);
740  if (NewOp->getType()->isFPOrFPVectorTy())
741  NewOp->setFastMathFlags(I->getFastMathFlags());
742  } else {
743  NewOp = NodesToRewrite.pop_back_val();
744  }
745 
746  DEBUG(dbgs() << "RA: " << *Op << '\n');
747  Op->setOperand(0, NewOp);
748  DEBUG(dbgs() << "TO: " << *Op << '\n');
749  ExpressionChanged = Op;
750  MadeChange = true;
751  ++NumChanged;
752  Op = NewOp;
753  }
754 
755  // If the expression changed non-trivially then clear out all subclass data
756  // starting from the operator specified in ExpressionChanged, and compactify
757  // the operators to just before the expression root to guarantee that the
758  // expression tree is dominated by all of Ops.
759  if (ExpressionChanged)
760  do {
761  // Preserve FastMathFlags.
762  if (isa<FPMathOperator>(I)) {
764  ExpressionChanged->clearSubclassOptionalData();
765  ExpressionChanged->setFastMathFlags(Flags);
766  } else
767  ExpressionChanged->clearSubclassOptionalData();
768 
769  if (ExpressionChanged == I)
770  break;
771  ExpressionChanged->moveBefore(I);
772  ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
773  } while (1);
774 
775  // Throw away any left over nodes from the original expression.
776  for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
777  RedoInsts.insert(NodesToRewrite[i]);
778 }
779 
780 /// Insert instructions before the instruction pointed to by BI,
781 /// that computes the negative version of the value specified. The negative
782 /// version of the value is returned, and BI is left pointing at the instruction
783 /// that should be processed next by the reassociation pass.
784 /// Also add intermediate instructions to the redo list that are modified while
785 /// pushing the negates through adds. These will be revisited to see if
786 /// additional opportunities have been exposed.
789  if (Constant *C = dyn_cast<Constant>(V)) {
790  if (C->getType()->isFPOrFPVectorTy()) {
791  return ConstantExpr::getFNeg(C);
792  }
793  return ConstantExpr::getNeg(C);
794  }
795 
796 
797  // We are trying to expose opportunity for reassociation. One of the things
798  // that we want to do to achieve this is to push a negation as deep into an
799  // expression chain as possible, to expose the add instructions. In practice,
800  // this means that we turn this:
801  // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
802  // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
803  // the constants. We assume that instcombine will clean up the mess later if
804  // we introduce tons of unnecessary negation instructions.
805  //
806  if (BinaryOperator *I =
807  isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
808  // Push the negates through the add.
809  I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
810  I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
811  if (I->getOpcode() == Instruction::Add) {
812  I->setHasNoUnsignedWrap(false);
813  I->setHasNoSignedWrap(false);
814  }
815 
816  // We must move the add instruction here, because the neg instructions do
817  // not dominate the old add instruction in general. By moving it, we are
818  // assured that the neg instructions we just inserted dominate the
819  // instruction we are about to insert after them.
820  //
821  I->moveBefore(BI);
822  I->setName(I->getName()+".neg");
823 
824  // Add the intermediate negates to the redo list as processing them later
825  // could expose more reassociating opportunities.
826  ToRedo.insert(I);
827  return I;
828  }
829 
830  // Okay, we need to materialize a negated version of V with an instruction.
831  // Scan the use lists of V to see if we have one already.
832  for (User *U : V->users()) {
834  continue;
835 
836  // We found one! Now we have to make sure that the definition dominates
837  // this use. We do this by moving it to the entry block (if it is a
838  // non-instruction value) or right after the definition. These negates will
839  // be zapped by reassociate later, so we don't need much finesse here.
840  BinaryOperator *TheNeg = cast<BinaryOperator>(U);
841 
842  // Verify that the negate is in this function, V might be a constant expr.
843  if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
844  continue;
845 
846  BasicBlock::iterator InsertPt;
847  if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
848  if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
849  InsertPt = II->getNormalDest()->begin();
850  } else {
851  InsertPt = ++InstInput->getIterator();
852  }
853  while (isa<PHINode>(InsertPt)) ++InsertPt;
854  } else {
855  InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
856  }
857  TheNeg->moveBefore(&*InsertPt);
858  if (TheNeg->getOpcode() == Instruction::Sub) {
859  TheNeg->setHasNoUnsignedWrap(false);
860  TheNeg->setHasNoSignedWrap(false);
861  } else {
862  TheNeg->andIRFlags(BI);
863  }
864  ToRedo.insert(TheNeg);
865  return TheNeg;
866  }
867 
868  // Insert a 'neg' instruction that subtracts the value from zero to get the
869  // negation.
870  BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
871  ToRedo.insert(NewNeg);
872  return NewNeg;
873 }
874 
875 /// Return true if we should break up this subtract of X-Y into (X + -Y).
877  // If this is a negation, we can't split it up!
879  return false;
880 
881  // Don't breakup X - undef.
882  if (isa<UndefValue>(Sub->getOperand(1)))
883  return false;
884 
885  // Don't bother to break this up unless either the LHS is an associable add or
886  // subtract or if this is only used by one.
887  Value *V0 = Sub->getOperand(0);
888  if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
889  isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
890  return true;
891  Value *V1 = Sub->getOperand(1);
892  if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
893  isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
894  return true;
895  Value *VB = Sub->user_back();
896  if (Sub->hasOneUse() &&
897  (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
898  isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
899  return true;
900 
901  return false;
902 }
903 
904 /// If we have (X-Y), and if either X is an add, or if this is only used by an
905 /// add, transform this into (X+(0-Y)) to promote better reassociation.
906 static BinaryOperator *
908  // Convert a subtract into an add and a neg instruction. This allows sub
909  // instructions to be commuted with other add instructions.
910  //
911  // Calculate the negative value of Operand 1 of the sub instruction,
912  // and set it as the RHS of the add instruction we just made.
913  //
914  Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
915  BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
916  Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
917  Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
918  New->takeName(Sub);
919 
920  // Everyone now refers to the add instruction.
921  Sub->replaceAllUsesWith(New);
922  New->setDebugLoc(Sub->getDebugLoc());
923 
924  DEBUG(dbgs() << "Negated: " << *New << '\n');
925  return New;
926 }
927 
928 /// If this is a shift of a reassociable multiply or is used by one, change
929 /// this into a multiply by a constant to assist with further reassociation.
931  Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
932  MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
933 
934  BinaryOperator *Mul =
935  BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
936  Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
937  Mul->takeName(Shl);
938 
939  // Everyone now refers to the mul instruction.
940  Shl->replaceAllUsesWith(Mul);
941  Mul->setDebugLoc(Shl->getDebugLoc());
942 
943  // We can safely preserve the nuw flag in all cases. It's also safe to turn a
944  // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
945  // handling.
946  bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
947  bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
948  if (NSW && NUW)
949  Mul->setHasNoSignedWrap(true);
950  Mul->setHasNoUnsignedWrap(NUW);
951  return Mul;
952 }
953 
954 /// Scan backwards and forwards among values with the same rank as element i
955 /// to see if X exists. If X does not exist, return i. This is useful when
956 /// scanning for 'x' when we see '-x' because they both get the same rank.
958  unsigned i, Value *X) {
959  unsigned XRank = Ops[i].Rank;
960  unsigned e = Ops.size();
961  for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
962  if (Ops[j].Op == X)
963  return j;
964  if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
965  if (Instruction *I2 = dyn_cast<Instruction>(X))
966  if (I1->isIdenticalTo(I2))
967  return j;
968  }
969  // Scan backwards.
970  for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
971  if (Ops[j].Op == X)
972  return j;
973  if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
974  if (Instruction *I2 = dyn_cast<Instruction>(X))
975  if (I1->isIdenticalTo(I2))
976  return j;
977  }
978  return i;
979 }
980 
981 /// Emit a tree of add instructions, summing Ops together
982 /// and returning the result. Insert the tree before I.
985  if (Ops.size() == 1) return Ops.back();
986 
987  Value *V1 = Ops.back();
988  Ops.pop_back();
989  Value *V2 = EmitAddTreeOfValues(I, Ops);
990  return CreateAdd(V2, V1, "tmp", I, I);
991 }
992 
993 /// If V is an expression tree that is a multiplication sequence,
994 /// and if this sequence contains a multiply by Factor,
995 /// remove Factor from the tree and return the new tree.
996 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
997  BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
998  if (!BO)
999  return nullptr;
1000 
1002  MadeChange |= LinearizeExprTree(BO, Tree);
1004  Factors.reserve(Tree.size());
1005  for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1006  RepeatedValue E = Tree[i];
1007  Factors.append(E.second.getZExtValue(),
1008  ValueEntry(getRank(E.first), E.first));
1009  }
1010 
1011  bool FoundFactor = false;
1012  bool NeedsNegate = false;
1013  for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1014  if (Factors[i].Op == Factor) {
1015  FoundFactor = true;
1016  Factors.erase(Factors.begin()+i);
1017  break;
1018  }
1019 
1020  // If this is a negative version of this factor, remove it.
1021  if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1022  if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1023  if (FC1->getValue() == -FC2->getValue()) {
1024  FoundFactor = NeedsNegate = true;
1025  Factors.erase(Factors.begin()+i);
1026  break;
1027  }
1028  } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1029  if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1030  const APFloat &F1 = FC1->getValueAPF();
1031  APFloat F2(FC2->getValueAPF());
1032  F2.changeSign();
1033  if (F1.compare(F2) == APFloat::cmpEqual) {
1034  FoundFactor = NeedsNegate = true;
1035  Factors.erase(Factors.begin() + i);
1036  break;
1037  }
1038  }
1039  }
1040  }
1041 
1042  if (!FoundFactor) {
1043  // Make sure to restore the operands to the expression tree.
1044  RewriteExprTree(BO, Factors);
1045  return nullptr;
1046  }
1047 
1048  BasicBlock::iterator InsertPt = ++BO->getIterator();
1049 
1050  // If this was just a single multiply, remove the multiply and return the only
1051  // remaining operand.
1052  if (Factors.size() == 1) {
1053  RedoInsts.insert(BO);
1054  V = Factors[0].Op;
1055  } else {
1056  RewriteExprTree(BO, Factors);
1057  V = BO;
1058  }
1059 
1060  if (NeedsNegate)
1061  V = CreateNeg(V, "neg", &*InsertPt, BO);
1062 
1063  return V;
1064 }
1065 
1066 /// If V is a single-use multiply, recursively add its operands as factors,
1067 /// otherwise add V to the list of factors.
1068 ///
1069 /// Ops is the top-level list of add operands we're trying to factor.
1071  SmallVectorImpl<Value*> &Factors) {
1072  BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1073  if (!BO) {
1074  Factors.push_back(V);
1075  return;
1076  }
1077 
1078  // Otherwise, add the LHS and RHS to the list of factors.
1079  FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1080  FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1081 }
1082 
1083 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1084 /// This optimizes based on identities. If it can be reduced to a single Value,
1085 /// it is returned, otherwise the Ops list is mutated as necessary.
1086 static Value *OptimizeAndOrXor(unsigned Opcode,
1088  // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1089  // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1090  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1091  // First, check for X and ~X in the operand list.
1092  assert(i < Ops.size());
1093  if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1094  Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1095  unsigned FoundX = FindInOperandList(Ops, i, X);
1096  if (FoundX != i) {
1097  if (Opcode == Instruction::And) // ...&X&~X = 0
1098  return Constant::getNullValue(X->getType());
1099 
1100  if (Opcode == Instruction::Or) // ...|X|~X = -1
1101  return Constant::getAllOnesValue(X->getType());
1102  }
1103  }
1104 
1105  // Next, check for duplicate pairs of values, which we assume are next to
1106  // each other, due to our sorting criteria.
1107  assert(i < Ops.size());
1108  if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1109  if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1110  // Drop duplicate values for And and Or.
1111  Ops.erase(Ops.begin()+i);
1112  --i; --e;
1113  ++NumAnnihil;
1114  continue;
1115  }
1116 
1117  // Drop pairs of values for Xor.
1118  assert(Opcode == Instruction::Xor);
1119  if (e == 2)
1120  return Constant::getNullValue(Ops[0].Op->getType());
1121 
1122  // Y ^ X^X -> Y
1123  Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1124  i -= 1; e -= 2;
1125  ++NumAnnihil;
1126  }
1127  }
1128  return nullptr;
1129 }
1130 
1131 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1132 /// instruction with the given two operands, and return the resulting
1133 /// instruction. There are two special cases: 1) if the constant operand is 0,
1134 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1135 /// be returned.
1136 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1137  const APInt &ConstOpnd) {
1138  if (ConstOpnd.isNullValue())
1139  return nullptr;
1140 
1141  if (ConstOpnd.isAllOnesValue())
1142  return Opnd;
1143 
1144  Instruction *I = BinaryOperator::CreateAnd(
1145  Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1146  InsertBefore);
1147  I->setDebugLoc(InsertBefore->getDebugLoc());
1148  return I;
1149 }
1150 
1151 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1152 // into "R ^ C", where C would be 0, and R is a symbolic value.
1153 //
1154 // If it was successful, true is returned, and the "R" and "C" is returned
1155 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1156 // and both "Res" and "ConstOpnd" remain unchanged.
1157 //
1158 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1159  APInt &ConstOpnd, Value *&Res) {
1160  // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1161  // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1162  // = (x & ~c1) ^ (c1 ^ c2)
1163  // It is useful only when c1 == c2.
1164  if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue())
1165  return false;
1166 
1167  if (!Opnd1->getValue()->hasOneUse())
1168  return false;
1169 
1170  const APInt &C1 = Opnd1->getConstPart();
1171  if (C1 != ConstOpnd)
1172  return false;
1173 
1174  Value *X = Opnd1->getSymbolicPart();
1175  Res = createAndInstr(I, X, ~C1);
1176  // ConstOpnd was C2, now C1 ^ C2.
1177  ConstOpnd ^= C1;
1178 
1179  if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1180  RedoInsts.insert(T);
1181  return true;
1182 }
1183 
1184 
1185 // Helper function of OptimizeXor(). It tries to simplify
1186 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1187 // symbolic value.
1188 //
1189 // If it was successful, true is returned, and the "R" and "C" is returned
1190 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1191 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1192 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1193 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1194  XorOpnd *Opnd2, APInt &ConstOpnd,
1195  Value *&Res) {
1196  Value *X = Opnd1->getSymbolicPart();
1197  if (X != Opnd2->getSymbolicPart())
1198  return false;
1199 
1200  // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1201  int DeadInstNum = 1;
1202  if (Opnd1->getValue()->hasOneUse())
1203  DeadInstNum++;
1204  if (Opnd2->getValue()->hasOneUse())
1205  DeadInstNum++;
1206 
1207  // Xor-Rule 2:
1208  // (x | c1) ^ (x & c2)
1209  // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1210  // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1211  // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1212  //
1213  if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1214  if (Opnd2->isOrExpr())
1215  std::swap(Opnd1, Opnd2);
1216 
1217  const APInt &C1 = Opnd1->getConstPart();
1218  const APInt &C2 = Opnd2->getConstPart();
1219  APInt C3((~C1) ^ C2);
1220 
1221  // Do not increase code size!
1222  if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1223  int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1224  if (NewInstNum > DeadInstNum)
1225  return false;
1226  }
1227 
1228  Res = createAndInstr(I, X, C3);
1229  ConstOpnd ^= C1;
1230 
1231  } else if (Opnd1->isOrExpr()) {
1232  // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1233  //
1234  const APInt &C1 = Opnd1->getConstPart();
1235  const APInt &C2 = Opnd2->getConstPart();
1236  APInt C3 = C1 ^ C2;
1237 
1238  // Do not increase code size
1239  if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1240  int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1241  if (NewInstNum > DeadInstNum)
1242  return false;
1243  }
1244 
1245  Res = createAndInstr(I, X, C3);
1246  ConstOpnd ^= C3;
1247  } else {
1248  // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1249  //
1250  const APInt &C1 = Opnd1->getConstPart();
1251  const APInt &C2 = Opnd2->getConstPart();
1252  APInt C3 = C1 ^ C2;
1253  Res = createAndInstr(I, X, C3);
1254  }
1255 
1256  // Put the original operands in the Redo list; hope they will be deleted
1257  // as dead code.
1258  if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1259  RedoInsts.insert(T);
1260  if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1261  RedoInsts.insert(T);
1262 
1263  return true;
1264 }
1265 
1266 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1267 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1268 /// necessary.
1269 Value *ReassociatePass::OptimizeXor(Instruction *I,
1271  if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1272  return V;
1273 
1274  if (Ops.size() == 1)
1275  return nullptr;
1276 
1278  SmallVector<XorOpnd*, 8> OpndPtrs;
1279  Type *Ty = Ops[0].Op->getType();
1280  APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1281 
1282  // Step 1: Convert ValueEntry to XorOpnd
1283  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1284  Value *V = Ops[i].Op;
1285  const APInt *C;
1286  // TODO: Support non-splat vectors.
1287  if (match(V, PatternMatch::m_APInt(C))) {
1288  ConstOpnd ^= *C;
1289  } else {
1290  XorOpnd O(V);
1291  O.setSymbolicRank(getRank(O.getSymbolicPart()));
1292  Opnds.push_back(O);
1293  }
1294  }
1295 
1296  // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1297  // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1298  // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1299  // with the previous loop --- the iterator of the "Opnds" may be invalidated
1300  // when new elements are added to the vector.
1301  for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1302  OpndPtrs.push_back(&Opnds[i]);
1303 
1304  // Step 2: Sort the Xor-Operands in a way such that the operands containing
1305  // the same symbolic value cluster together. For instance, the input operand
1306  // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1307  // ("x | 123", "x & 789", "y & 456").
1308  //
1309  // The purpose is twofold:
1310  // 1) Cluster together the operands sharing the same symbolic-value.
1311  // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1312  // could potentially shorten crital path, and expose more loop-invariants.
1313  // Note that values' rank are basically defined in RPO order (FIXME).
1314  // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1315  // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1316  // "z" in the order of X-Y-Z is better than any other orders.
1317  std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1318  [](XorOpnd *LHS, XorOpnd *RHS) {
1319  return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1320  });
1321 
1322  // Step 3: Combine adjacent operands
1323  XorOpnd *PrevOpnd = nullptr;
1324  bool Changed = false;
1325  for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1326  XorOpnd *CurrOpnd = OpndPtrs[i];
1327  // The combined value
1328  Value *CV;
1329 
1330  // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1331  if (!ConstOpnd.isNullValue() &&
1332  CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1333  Changed = true;
1334  if (CV)
1335  *CurrOpnd = XorOpnd(CV);
1336  else {
1337  CurrOpnd->Invalidate();
1338  continue;
1339  }
1340  }
1341 
1342  if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1343  PrevOpnd = CurrOpnd;
1344  continue;
1345  }
1346 
1347  // step 3.2: When previous and current operands share the same symbolic
1348  // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1349  //
1350  if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1351  // Remove previous operand
1352  PrevOpnd->Invalidate();
1353  if (CV) {
1354  *CurrOpnd = XorOpnd(CV);
1355  PrevOpnd = CurrOpnd;
1356  } else {
1357  CurrOpnd->Invalidate();
1358  PrevOpnd = nullptr;
1359  }
1360  Changed = true;
1361  }
1362  }
1363 
1364  // Step 4: Reassemble the Ops
1365  if (Changed) {
1366  Ops.clear();
1367  for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1368  XorOpnd &O = Opnds[i];
1369  if (O.isInvalid())
1370  continue;
1371  ValueEntry VE(getRank(O.getValue()), O.getValue());
1372  Ops.push_back(VE);
1373  }
1374  if (!ConstOpnd.isNullValue()) {
1375  Value *C = ConstantInt::get(Ty, ConstOpnd);
1376  ValueEntry VE(getRank(C), C);
1377  Ops.push_back(VE);
1378  }
1379  unsigned Sz = Ops.size();
1380  if (Sz == 1)
1381  return Ops.back().Op;
1382  if (Sz == 0) {
1383  assert(ConstOpnd.isNullValue());
1384  return ConstantInt::get(Ty, ConstOpnd);
1385  }
1386  }
1387 
1388  return nullptr;
1389 }
1390 
1391 /// Optimize a series of operands to an 'add' instruction. This
1392 /// optimizes based on identities. If it can be reduced to a single Value, it
1393 /// is returned, otherwise the Ops list is mutated as necessary.
1394 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1396  // Scan the operand lists looking for X and -X pairs. If we find any, we
1397  // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1398  // scan for any
1399  // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1400 
1401  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1402  Value *TheOp = Ops[i].Op;
1403  // Check to see if we've seen this operand before. If so, we factor all
1404  // instances of the operand together. Due to our sorting criteria, we know
1405  // that these need to be next to each other in the vector.
1406  if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1407  // Rescan the list, remove all instances of this operand from the expr.
1408  unsigned NumFound = 0;
1409  do {
1410  Ops.erase(Ops.begin()+i);
1411  ++NumFound;
1412  } while (i != Ops.size() && Ops[i].Op == TheOp);
1413 
1414  DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1415  ++NumFactor;
1416 
1417  // Insert a new multiply.
1418  Type *Ty = TheOp->getType();
1419  Constant *C = Ty->isIntOrIntVectorTy() ?
1420  ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1421  Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1422 
1423  // Now that we have inserted a multiply, optimize it. This allows us to
1424  // handle cases that require multiple factoring steps, such as this:
1425  // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1426  RedoInsts.insert(Mul);
1427 
1428  // If every add operand was a duplicate, return the multiply.
1429  if (Ops.empty())
1430  return Mul;
1431 
1432  // Otherwise, we had some input that didn't have the dupe, such as
1433  // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1434  // things being added by this operation.
1435  Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1436 
1437  --i;
1438  e = Ops.size();
1439  continue;
1440  }
1441 
1442  // Check for X and -X or X and ~X in the operand list.
1443  if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1444  !BinaryOperator::isNot(TheOp))
1445  continue;
1446 
1447  Value *X = nullptr;
1448  if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1449  X = BinaryOperator::getNegArgument(TheOp);
1450  else if (BinaryOperator::isNot(TheOp))
1451  X = BinaryOperator::getNotArgument(TheOp);
1452 
1453  unsigned FoundX = FindInOperandList(Ops, i, X);
1454  if (FoundX == i)
1455  continue;
1456 
1457  // Remove X and -X from the operand list.
1458  if (Ops.size() == 2 &&
1459  (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1460  return Constant::getNullValue(X->getType());
1461 
1462  // Remove X and ~X from the operand list.
1463  if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1464  return Constant::getAllOnesValue(X->getType());
1465 
1466  Ops.erase(Ops.begin()+i);
1467  if (i < FoundX)
1468  --FoundX;
1469  else
1470  --i; // Need to back up an extra one.
1471  Ops.erase(Ops.begin()+FoundX);
1472  ++NumAnnihil;
1473  --i; // Revisit element.
1474  e -= 2; // Removed two elements.
1475 
1476  // if X and ~X we append -1 to the operand list.
1477  if (BinaryOperator::isNot(TheOp)) {
1479  Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1480  e += 1;
1481  }
1482  }
1483 
1484  // Scan the operand list, checking to see if there are any common factors
1485  // between operands. Consider something like A*A+A*B*C+D. We would like to
1486  // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1487  // To efficiently find this, we count the number of times a factor occurs
1488  // for any ADD operands that are MULs.
1489  DenseMap<Value*, unsigned> FactorOccurrences;
1490 
1491  // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1492  // where they are actually the same multiply.
1493  unsigned MaxOcc = 0;
1494  Value *MaxOccVal = nullptr;
1495  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1496  BinaryOperator *BOp =
1497  isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1498  if (!BOp)
1499  continue;
1500 
1501  // Compute all of the factors of this added value.
1502  SmallVector<Value*, 8> Factors;
1503  FindSingleUseMultiplyFactors(BOp, Factors);
1504  assert(Factors.size() > 1 && "Bad linearize!");
1505 
1506  // Add one to FactorOccurrences for each unique factor in this op.
1507  SmallPtrSet<Value*, 8> Duplicates;
1508  for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1509  Value *Factor = Factors[i];
1510  if (!Duplicates.insert(Factor).second)
1511  continue;
1512 
1513  unsigned Occ = ++FactorOccurrences[Factor];
1514  if (Occ > MaxOcc) {
1515  MaxOcc = Occ;
1516  MaxOccVal = Factor;
1517  }
1518 
1519  // If Factor is a negative constant, add the negated value as a factor
1520  // because we can percolate the negate out. Watch for minint, which
1521  // cannot be positivified.
1522  if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1523  if (CI->isNegative() && !CI->isMinValue(true)) {
1524  Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1525  if (!Duplicates.insert(Factor).second)
1526  continue;
1527  unsigned Occ = ++FactorOccurrences[Factor];
1528  if (Occ > MaxOcc) {
1529  MaxOcc = Occ;
1530  MaxOccVal = Factor;
1531  }
1532  }
1533  } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1534  if (CF->isNegative()) {
1535  APFloat F(CF->getValueAPF());
1536  F.changeSign();
1537  Factor = ConstantFP::get(CF->getContext(), F);
1538  if (!Duplicates.insert(Factor).second)
1539  continue;
1540  unsigned Occ = ++FactorOccurrences[Factor];
1541  if (Occ > MaxOcc) {
1542  MaxOcc = Occ;
1543  MaxOccVal = Factor;
1544  }
1545  }
1546  }
1547  }
1548  }
1549 
1550  // If any factor occurred more than one time, we can pull it out.
1551  if (MaxOcc > 1) {
1552  DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1553  ++NumFactor;
1554 
1555  // Create a new instruction that uses the MaxOccVal twice. If we don't do
1556  // this, we could otherwise run into situations where removing a factor
1557  // from an expression will drop a use of maxocc, and this can cause
1558  // RemoveFactorFromExpression on successive values to behave differently.
1559  Instruction *DummyInst =
1560  I->getType()->isIntOrIntVectorTy()
1561  ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1562  : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1563 
1565  for (unsigned i = 0; i != Ops.size(); ++i) {
1566  // Only try to remove factors from expressions we're allowed to.
1567  BinaryOperator *BOp =
1568  isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1569  if (!BOp)
1570  continue;
1571 
1572  if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1573  // The factorized operand may occur several times. Convert them all in
1574  // one fell swoop.
1575  for (unsigned j = Ops.size(); j != i;) {
1576  --j;
1577  if (Ops[j].Op == Ops[i].Op) {
1578  NewMulOps.push_back(V);
1579  Ops.erase(Ops.begin()+j);
1580  }
1581  }
1582  --i;
1583  }
1584  }
1585 
1586  // No need for extra uses anymore.
1587  DummyInst->deleteValue();
1588 
1589  unsigned NumAddedValues = NewMulOps.size();
1590  Value *V = EmitAddTreeOfValues(I, NewMulOps);
1591 
1592  // Now that we have inserted the add tree, optimize it. This allows us to
1593  // handle cases that require multiple factoring steps, such as this:
1594  // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1595  assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1596  (void)NumAddedValues;
1597  if (Instruction *VI = dyn_cast<Instruction>(V))
1598  RedoInsts.insert(VI);
1599 
1600  // Create the multiply.
1601  Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1602 
1603  // Rerun associate on the multiply in case the inner expression turned into
1604  // a multiply. We want to make sure that we keep things in canonical form.
1605  RedoInsts.insert(V2);
1606 
1607  // If every add operand included the factor (e.g. "A*B + A*C"), then the
1608  // entire result expression is just the multiply "A*(B+C)".
1609  if (Ops.empty())
1610  return V2;
1611 
1612  // Otherwise, we had some input that didn't have the factor, such as
1613  // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1614  // things being added by this operation.
1615  Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1616  }
1617 
1618  return nullptr;
1619 }
1620 
1621 /// \brief Build up a vector of value/power pairs factoring a product.
1622 ///
1623 /// Given a series of multiplication operands, build a vector of factors and
1624 /// the powers each is raised to when forming the final product. Sort them in
1625 /// the order of descending power.
1626 ///
1627 /// (x*x) -> [(x, 2)]
1628 /// ((x*x)*x) -> [(x, 3)]
1629 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1630 ///
1631 /// \returns Whether any factors have a power greater than one.
1633  SmallVectorImpl<Factor> &Factors) {
1634  // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1635  // Compute the sum of powers of simplifiable factors.
1636  unsigned FactorPowerSum = 0;
1637  for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1638  Value *Op = Ops[Idx-1].Op;
1639 
1640  // Count the number of occurrences of this value.
1641  unsigned Count = 1;
1642  for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1643  ++Count;
1644  // Track for simplification all factors which occur 2 or more times.
1645  if (Count > 1)
1646  FactorPowerSum += Count;
1647  }
1648 
1649  // We can only simplify factors if the sum of the powers of our simplifiable
1650  // factors is 4 or higher. When that is the case, we will *always* have
1651  // a simplification. This is an important invariant to prevent cyclicly
1652  // trying to simplify already minimal formations.
1653  if (FactorPowerSum < 4)
1654  return false;
1655 
1656  // Now gather the simplifiable factors, removing them from Ops.
1657  FactorPowerSum = 0;
1658  for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1659  Value *Op = Ops[Idx-1].Op;
1660 
1661  // Count the number of occurrences of this value.
1662  unsigned Count = 1;
1663  for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1664  ++Count;
1665  if (Count == 1)
1666  continue;
1667  // Move an even number of occurrences to Factors.
1668  Count &= ~1U;
1669  Idx -= Count;
1670  FactorPowerSum += Count;
1671  Factors.push_back(Factor(Op, Count));
1672  Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1673  }
1674 
1675  // None of the adjustments above should have reduced the sum of factor powers
1676  // below our mininum of '4'.
1677  assert(FactorPowerSum >= 4);
1678 
1679  std::stable_sort(Factors.begin(), Factors.end(),
1680  [](const Factor &LHS, const Factor &RHS) {
1681  return LHS.Power > RHS.Power;
1682  });
1683  return true;
1684 }
1685 
1686 /// \brief Build a tree of multiplies, computing the product of Ops.
1688  SmallVectorImpl<Value*> &Ops) {
1689  if (Ops.size() == 1)
1690  return Ops.back();
1691 
1692  Value *LHS = Ops.pop_back_val();
1693  do {
1694  if (LHS->getType()->isIntOrIntVectorTy())
1695  LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1696  else
1697  LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1698  } while (!Ops.empty());
1699 
1700  return LHS;
1701 }
1702 
1703 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1704 ///
1705 /// Given a vector of values raised to various powers, where no two values are
1706 /// equal and the powers are sorted in decreasing order, compute the minimal
1707 /// DAG of multiplies to compute the final product, and return that product
1708 /// value.
1709 Value *
1710 ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1711  SmallVectorImpl<Factor> &Factors) {
1712  assert(Factors[0].Power);
1713  SmallVector<Value *, 4> OuterProduct;
1714  for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1715  Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1716  if (Factors[Idx].Power != Factors[LastIdx].Power) {
1717  LastIdx = Idx;
1718  continue;
1719  }
1720 
1721  // We want to multiply across all the factors with the same power so that
1722  // we can raise them to that power as a single entity. Build a mini tree
1723  // for that.
1724  SmallVector<Value *, 4> InnerProduct;
1725  InnerProduct.push_back(Factors[LastIdx].Base);
1726  do {
1727  InnerProduct.push_back(Factors[Idx].Base);
1728  ++Idx;
1729  } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1730 
1731  // Reset the base value of the first factor to the new expression tree.
1732  // We'll remove all the factors with the same power in a second pass.
1733  Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1734  if (Instruction *MI = dyn_cast<Instruction>(M))
1735  RedoInsts.insert(MI);
1736 
1737  LastIdx = Idx;
1738  }
1739  // Unique factors with equal powers -- we've folded them into the first one's
1740  // base.
1741  Factors.erase(std::unique(Factors.begin(), Factors.end(),
1742  [](const Factor &LHS, const Factor &RHS) {
1743  return LHS.Power == RHS.Power;
1744  }),
1745  Factors.end());
1746 
1747  // Iteratively collect the base of each factor with an add power into the
1748  // outer product, and halve each power in preparation for squaring the
1749  // expression.
1750  for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1751  if (Factors[Idx].Power & 1)
1752  OuterProduct.push_back(Factors[Idx].Base);
1753  Factors[Idx].Power >>= 1;
1754  }
1755  if (Factors[0].Power) {
1756  Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1757  OuterProduct.push_back(SquareRoot);
1758  OuterProduct.push_back(SquareRoot);
1759  }
1760  if (OuterProduct.size() == 1)
1761  return OuterProduct.front();
1762 
1763  Value *V = buildMultiplyTree(Builder, OuterProduct);
1764  return V;
1765 }
1766 
1767 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1769  // We can only optimize the multiplies when there is a chain of more than
1770  // three, such that a balanced tree might require fewer total multiplies.
1771  if (Ops.size() < 4)
1772  return nullptr;
1773 
1774  // Try to turn linear trees of multiplies without other uses of the
1775  // intermediate stages into minimal multiply DAGs with perfect sub-expression
1776  // re-use.
1777  SmallVector<Factor, 4> Factors;
1778  if (!collectMultiplyFactors(Ops, Factors))
1779  return nullptr; // All distinct factors, so nothing left for us to do.
1780 
1781  IRBuilder<> Builder(I);
1782  // The reassociate transformation for FP operations is performed only
1783  // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1784  // to the newly generated operations.
1785  if (auto FPI = dyn_cast<FPMathOperator>(I))
1786  Builder.setFastMathFlags(FPI->getFastMathFlags());
1787 
1788  Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1789  if (Ops.empty())
1790  return V;
1791 
1792  ValueEntry NewEntry = ValueEntry(getRank(V), V);
1793  Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1794  return nullptr;
1795 }
1796 
1797 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1799  // Now that we have the linearized expression tree, try to optimize it.
1800  // Start by folding any constants that we found.
1801  Constant *Cst = nullptr;
1802  unsigned Opcode = I->getOpcode();
1803  while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1804  Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1805  Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1806  }
1807  // If there was nothing but constants then we are done.
1808  if (Ops.empty())
1809  return Cst;
1810 
1811  // Put the combined constant back at the end of the operand list, except if
1812  // there is no point. For example, an add of 0 gets dropped here, while a
1813  // multiplication by zero turns the whole expression into zero.
1814  if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1815  if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1816  return Cst;
1817  Ops.push_back(ValueEntry(0, Cst));
1818  }
1819 
1820  if (Ops.size() == 1) return Ops[0].Op;
1821 
1822  // Handle destructive annihilation due to identities between elements in the
1823  // argument list here.
1824  unsigned NumOps = Ops.size();
1825  switch (Opcode) {
1826  default: break;
1827  case Instruction::And:
1828  case Instruction::Or:
1829  if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1830  return Result;
1831  break;
1832 
1833  case Instruction::Xor:
1834  if (Value *Result = OptimizeXor(I, Ops))
1835  return Result;
1836  break;
1837 
1838  case Instruction::Add:
1839  case Instruction::FAdd:
1840  if (Value *Result = OptimizeAdd(I, Ops))
1841  return Result;
1842  break;
1843 
1844  case Instruction::Mul:
1845  case Instruction::FMul:
1846  if (Value *Result = OptimizeMul(I, Ops))
1847  return Result;
1848  break;
1849  }
1850 
1851  if (Ops.size() != NumOps)
1852  return OptimizeExpression(I, Ops);
1853  return nullptr;
1854 }
1855 
1856 // Remove dead instructions and if any operands are trivially dead add them to
1857 // Insts so they will be removed as well.
1858 void ReassociatePass::RecursivelyEraseDeadInsts(
1860  assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1861  SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1862  ValueRankMap.erase(I);
1863  Insts.remove(I);
1864  RedoInsts.remove(I);
1865  I->eraseFromParent();
1866  for (auto Op : Ops)
1867  if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1868  if (OpInst->use_empty())
1869  Insts.insert(OpInst);
1870 }
1871 
1872 /// Zap the given instruction, adding interesting operands to the work list.
1873 void ReassociatePass::EraseInst(Instruction *I) {
1874  assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1875  DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
1876 
1877  SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1878  // Erase the dead instruction.
1879  ValueRankMap.erase(I);
1880  RedoInsts.remove(I);
1881  I->eraseFromParent();
1882  // Optimize its operands.
1883  SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1884  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1885  if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1886  // If this is a node in an expression tree, climb to the expression root
1887  // and add that since that's where optimization actually happens.
1888  unsigned Opcode = Op->getOpcode();
1889  while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1890  Visited.insert(Op).second)
1891  Op = Op->user_back();
1892  RedoInsts.insert(Op);
1893  }
1894 
1895  MadeChange = true;
1896 }
1897 
1898 // Canonicalize expressions of the following form:
1899 // x + (-Constant * y) -> x - (Constant * y)
1900 // x - (-Constant * y) -> x + (Constant * y)
1901 Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1902  if (!I->hasOneUse() || I->getType()->isVectorTy())
1903  return nullptr;
1904 
1905  // Must be a fmul or fdiv instruction.
1906  unsigned Opcode = I->getOpcode();
1907  if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1908  return nullptr;
1909 
1910  auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1911  auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1912 
1913  // Both operands are constant, let it get constant folded away.
1914  if (C0 && C1)
1915  return nullptr;
1916 
1917  ConstantFP *CF = C0 ? C0 : C1;
1918 
1919  // Must have one constant operand.
1920  if (!CF)
1921  return nullptr;
1922 
1923  // Must be a negative ConstantFP.
1924  if (!CF->isNegative())
1925  return nullptr;
1926 
1927  // User must be a binary operator with one or more uses.
1928  Instruction *User = I->user_back();
1929  if (!isa<BinaryOperator>(User) || User->use_empty())
1930  return nullptr;
1931 
1932  unsigned UserOpcode = User->getOpcode();
1933  if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1934  return nullptr;
1935 
1936  // Subtraction is not commutative. Explicitly, the following transform is
1937  // not valid: (-Constant * y) - x -> x + (Constant * y)
1938  if (!User->isCommutative() && User->getOperand(1) != I)
1939  return nullptr;
1940 
1941  // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
1942  // resulting subtract will be broken up later. This can get us into an
1943  // infinite loop during reassociation.
1944  if (UserOpcode == Instruction::FAdd && ShouldBreakUpSubtract(User))
1945  return nullptr;
1946 
1947  // Change the sign of the constant.
1948  APFloat Val = CF->getValueAPF();
1949  Val.changeSign();
1950  I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1951 
1952  // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1953  // ((-Const*y) + x) -> (x + (-Const*y)).
1954  if (User->getOperand(0) == I && User->isCommutative())
1955  cast<BinaryOperator>(User)->swapOperands();
1956 
1957  Value *Op0 = User->getOperand(0);
1958  Value *Op1 = User->getOperand(1);
1959  BinaryOperator *NI;
1960  switch (UserOpcode) {
1961  default:
1962  llvm_unreachable("Unexpected Opcode!");
1963  case Instruction::FAdd:
1964  NI = BinaryOperator::CreateFSub(Op0, Op1);
1965  NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1966  break;
1967  case Instruction::FSub:
1968  NI = BinaryOperator::CreateFAdd(Op0, Op1);
1969  NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1970  break;
1971  }
1972 
1973  NI->insertBefore(User);
1974  NI->setName(User->getName());
1975  User->replaceAllUsesWith(NI);
1976  NI->setDebugLoc(I->getDebugLoc());
1977  RedoInsts.insert(I);
1978  MadeChange = true;
1979  return NI;
1980 }
1981 
1982 /// Inspect and optimize the given instruction. Note that erasing
1983 /// instructions is not allowed.
1984 void ReassociatePass::OptimizeInst(Instruction *I) {
1985  // Only consider operations that we understand.
1986  if (!isa<BinaryOperator>(I))
1987  return;
1988 
1989  if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1990  // If an operand of this shift is a reassociable multiply, or if the shift
1991  // is used by a reassociable multiply or add, turn into a multiply.
1992  if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1993  (I->hasOneUse() &&
1994  (isReassociableOp(I->user_back(), Instruction::Mul) ||
1996  Instruction *NI = ConvertShiftToMul(I);
1997  RedoInsts.insert(I);
1998  MadeChange = true;
1999  I = NI;
2000  }
2001 
2002  // Canonicalize negative constants out of expressions.
2003  if (Instruction *Res = canonicalizeNegConstExpr(I))
2004  I = Res;
2005 
2006  // Commute binary operators, to canonicalize the order of their operands.
2007  // This can potentially expose more CSE opportunities, and makes writing other
2008  // transformations simpler.
2009  if (I->isCommutative())
2010  canonicalizeOperands(I);
2011 
2012  // Don't optimize floating point instructions that don't have unsafe algebra.
2013  if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra())
2014  return;
2015 
2016  // Do not reassociate boolean (i1) expressions. We want to preserve the
2017  // original order of evaluation for short-circuited comparisons that
2018  // SimplifyCFG has folded to AND/OR expressions. If the expression
2019  // is not further optimized, it is likely to be transformed back to a
2020  // short-circuited form for code gen, and the source order may have been
2021  // optimized for the most likely conditions.
2022  if (I->getType()->isIntegerTy(1))
2023  return;
2024 
2025  // If this is a subtract instruction which is not already in negate form,
2026  // see if we can convert it to X+-Y.
2027  if (I->getOpcode() == Instruction::Sub) {
2028  if (ShouldBreakUpSubtract(I)) {
2029  Instruction *NI = BreakUpSubtract(I, RedoInsts);
2030  RedoInsts.insert(I);
2031  MadeChange = true;
2032  I = NI;
2033  } else if (BinaryOperator::isNeg(I)) {
2034  // Otherwise, this is a negation. See if the operand is a multiply tree
2035  // and if this is not an inner node of a multiply tree.
2036  if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2037  (!I->hasOneUse() ||
2038  !isReassociableOp(I->user_back(), Instruction::Mul))) {
2040  // If the negate was simplified, revisit the users to see if we can
2041  // reassociate further.
2042  for (User *U : NI->users()) {
2043  if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2044  RedoInsts.insert(Tmp);
2045  }
2046  RedoInsts.insert(I);
2047  MadeChange = true;
2048  I = NI;
2049  }
2050  }
2051  } else if (I->getOpcode() == Instruction::FSub) {
2052  if (ShouldBreakUpSubtract(I)) {
2053  Instruction *NI = BreakUpSubtract(I, RedoInsts);
2054  RedoInsts.insert(I);
2055  MadeChange = true;
2056  I = NI;
2057  } else if (BinaryOperator::isFNeg(I)) {
2058  // Otherwise, this is a negation. See if the operand is a multiply tree
2059  // and if this is not an inner node of a multiply tree.
2060  if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2061  (!I->hasOneUse() ||
2062  !isReassociableOp(I->user_back(), Instruction::FMul))) {
2063  // If the negate was simplified, revisit the users to see if we can
2064  // reassociate further.
2066  for (User *U : NI->users()) {
2067  if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2068  RedoInsts.insert(Tmp);
2069  }
2070  RedoInsts.insert(I);
2071  MadeChange = true;
2072  I = NI;
2073  }
2074  }
2075  }
2076 
2077  // If this instruction is an associative binary operator, process it.
2078  if (!I->isAssociative()) return;
2079  BinaryOperator *BO = cast<BinaryOperator>(I);
2080 
2081  // If this is an interior node of a reassociable tree, ignore it until we
2082  // get to the root of the tree, to avoid N^2 analysis.
2083  unsigned Opcode = BO->getOpcode();
2084  if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2085  // During the initial run we will get to the root of the tree.
2086  // But if we get here while we are redoing instructions, there is no
2087  // guarantee that the root will be visited. So Redo later
2088  if (BO->user_back() != BO &&
2089  BO->getParent() == BO->user_back()->getParent())
2090  RedoInsts.insert(BO->user_back());
2091  return;
2092  }
2093 
2094  // If this is an add tree that is used by a sub instruction, ignore it
2095  // until we process the subtract.
2096  if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2097  cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2098  return;
2099  if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2100  cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2101  return;
2102 
2103  ReassociateExpression(BO);
2104 }
2105 
2106 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2107  // First, walk the expression tree, linearizing the tree, collecting the
2108  // operand information.
2110  MadeChange |= LinearizeExprTree(I, Tree);
2112  Ops.reserve(Tree.size());
2113  for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2114  RepeatedValue E = Tree[i];
2115  Ops.append(E.second.getZExtValue(),
2116  ValueEntry(getRank(E.first), E.first));
2117  }
2118 
2119  DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2120 
2121  // Now that we have linearized the tree to a list and have gathered all of
2122  // the operands and their ranks, sort the operands by their rank. Use a
2123  // stable_sort so that values with equal ranks will have their relative
2124  // positions maintained (and so the compiler is deterministic). Note that
2125  // this sorts so that the highest ranking values end up at the beginning of
2126  // the vector.
2127  std::stable_sort(Ops.begin(), Ops.end());
2128 
2129  // Now that we have the expression tree in a convenient
2130  // sorted form, optimize it globally if possible.
2131  if (Value *V = OptimizeExpression(I, Ops)) {
2132  if (V == I)
2133  // Self-referential expression in unreachable code.
2134  return;
2135  // This expression tree simplified to something that isn't a tree,
2136  // eliminate it.
2137  DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2138  I->replaceAllUsesWith(V);
2139  if (Instruction *VI = dyn_cast<Instruction>(V))
2140  if (I->getDebugLoc())
2141  VI->setDebugLoc(I->getDebugLoc());
2142  RedoInsts.insert(I);
2143  ++NumAnnihil;
2144  return;
2145  }
2146 
2147  // We want to sink immediates as deeply as possible except in the case where
2148  // this is a multiply tree used only by an add, and the immediate is a -1.
2149  // In this case we reassociate to put the negation on the outside so that we
2150  // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2151  if (I->hasOneUse()) {
2152  if (I->getOpcode() == Instruction::Mul &&
2153  cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2154  isa<ConstantInt>(Ops.back().Op) &&
2155  cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2156  ValueEntry Tmp = Ops.pop_back_val();
2157  Ops.insert(Ops.begin(), Tmp);
2158  } else if (I->getOpcode() == Instruction::FMul &&
2159  cast<Instruction>(I->user_back())->getOpcode() ==
2160  Instruction::FAdd &&
2161  isa<ConstantFP>(Ops.back().Op) &&
2162  cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2163  ValueEntry Tmp = Ops.pop_back_val();
2164  Ops.insert(Ops.begin(), Tmp);
2165  }
2166  }
2167 
2168  DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2169 
2170  if (Ops.size() == 1) {
2171  if (Ops[0].Op == I)
2172  // Self-referential expression in unreachable code.
2173  return;
2174 
2175  // This expression tree simplified to something that isn't a tree,
2176  // eliminate it.
2177  I->replaceAllUsesWith(Ops[0].Op);
2178  if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2179  OI->setDebugLoc(I->getDebugLoc());
2180  RedoInsts.insert(I);
2181  return;
2182  }
2183 
2184  // Now that we ordered and optimized the expressions, splat them back into
2185  // the expression tree, removing any unneeded nodes.
2186  RewriteExprTree(I, Ops);
2187 }
2188 
2190  // Get the functions basic blocks in Reverse Post Order. This order is used by
2191  // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2192  // blocks (it has been seen that the analysis in this pass could hang when
2193  // analysing dead basic blocks).
2195 
2196  // Calculate the rank map for F.
2197  BuildRankMap(F, RPOT);
2198 
2199  MadeChange = false;
2200  // Traverse the same blocks that was analysed by BuildRankMap.
2201  for (BasicBlock *BI : RPOT) {
2202  assert(RankMap.count(&*BI) && "BB should be ranked.");
2203  // Optimize every instruction in the basic block.
2204  for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2205  if (isInstructionTriviallyDead(&*II)) {
2206  EraseInst(&*II++);
2207  } else {
2208  OptimizeInst(&*II);
2209  assert(II->getParent() == &*BI && "Moved to a different block!");
2210  ++II;
2211  }
2212 
2213  // Make a copy of all the instructions to be redone so we can remove dead
2214  // instructions.
2215  SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts);
2216  // Iterate over all instructions to be reevaluated and remove trivially dead
2217  // instructions. If any operand of the trivially dead instruction becomes
2218  // dead mark it for deletion as well. Continue this process until all
2219  // trivially dead instructions have been removed.
2220  while (!ToRedo.empty()) {
2221  Instruction *I = ToRedo.pop_back_val();
2222  if (isInstructionTriviallyDead(I)) {
2223  RecursivelyEraseDeadInsts(I, ToRedo);
2224  MadeChange = true;
2225  }
2226  }
2227 
2228  // Now that we have removed dead instructions, we can reoptimize the
2229  // remaining instructions.
2230  while (!RedoInsts.empty()) {
2231  Instruction *I = RedoInsts.pop_back_val();
2233  EraseInst(I);
2234  else
2235  OptimizeInst(I);
2236  }
2237  }
2238 
2239  // We are done with the rank map.
2240  RankMap.clear();
2241  ValueRankMap.clear();
2242 
2243  if (MadeChange) {
2244  PreservedAnalyses PA;
2245  PA.preserveSet<CFGAnalyses>();
2246  PA.preserve<GlobalsAA>();
2247  return PA;
2248  }
2249 
2250  return PreservedAnalyses::all();
2251 }
2252 
2253 namespace {
2254  class ReassociateLegacyPass : public FunctionPass {
2255  ReassociatePass Impl;
2256  public:
2257  static char ID; // Pass identification, replacement for typeid
2258  ReassociateLegacyPass() : FunctionPass(ID) {
2260  }
2261 
2262  bool runOnFunction(Function &F) override {
2263  if (skipFunction(F))
2264  return false;
2265 
2266  FunctionAnalysisManager DummyFAM;
2267  auto PA = Impl.run(F, DummyFAM);
2268  return !PA.areAllPreserved();
2269  }
2270 
2271  void getAnalysisUsage(AnalysisUsage &AU) const override {
2272  AU.setPreservesCFG();
2274  }
2275  };
2276 }
2277 
2278 char ReassociateLegacyPass::ID = 0;
2279 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2280  "Reassociate expressions", false, false)
2281 
2282 // Public interface to the Reassociate pass
2284  return new ReassociateLegacyPass();
2285 }
Legacy wrapper pass to provide the GlobalsAAResult object.
uint64_t CallInst * C
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
void push_back(const T &Elt)
Definition: SmallVector.h:212
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
void setFastMathFlags(FastMathFlags FMF)
Convenience function for setting multiple fast-math flags on this instruction, which must be an opera...
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1542
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
struct fuzzer::@309 Flags
LLVM_NODISCARD T pop_back_val()
Definition: SetVector.h:228
static BinaryOperator * LowerNegateToMultiply(Instruction *Neg)
Replace 0-X with X*-1.
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:523
INITIALIZE_PASS(ReassociateLegacyPass, "reassociate", "Reassociate expressions", false, false) FunctionPass *llvm
This is the interface for a simple mod/ref and alias analysis over globals.
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:63
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
bool isIdempotent() const
Return true if the instruction is idempotent:
Definition: Instruction.h:438
Utility class representing a non-constant Xor-operand.
Definition: Reassociate.cpp:78
static Constant * getBinOpIdentity(unsigned Opcode, Type *Ty)
Return the identity for the given binary operation, i.e.
Definition: Constants.cpp:2203
nary reassociate
Utility class representing a base and exponent pair which form one factor of some product...
Definition: Reassociate.h:48
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:697
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:93
STATISTIC(NumFunctions, "Total number of functions")
F(f)
static unsigned FindInOperandList(const SmallVectorImpl< ValueEntry > &Ops, unsigned i, Value *X)
Scan backwards and forwards among values with the same rank as element i to see if X exists...
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
void changeSign()
Definition: APFloat.h:1050
void reserve(size_type N)
Definition: SmallVector.h:380
Reassociate commutative expressions.
Definition: Reassociate.h:58
op_iterator op_begin()
Definition: User.h:214
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1488
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:207
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
bool swapOperands()
Exchange the two operands to this instruction.
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:3627
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
static const Value * getNegArgument(const Value *BinOp)
Helper functions to extract the unary argument of a NEG, FNEG or NOT operation implemented via Sub...
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
bool isNilpotent() const
Return true if the instruction is nilpotent:
Definition: Instruction.h:452
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
static bool collectMultiplyFactors(SmallVectorImpl< ValueEntry > &Ops, SmallVectorImpl< Factor > &Factors)
Build up a vector of value/power pairs factoring a product.
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:664
FunctionPass * createReassociatePass()
bool hasUnsafeAlgebra() const
Determine whether the unsafe-algebra flag is set.
static BinaryOperator * BreakUpSubtract(Instruction *Sub, SetVector< AssertingVH< Instruction >> &ToRedo)
If we have (X-Y), and if either X is an add, or if this is only used by an add, transform this into (...
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:284
static Constant * get(unsigned Opcode, Constant *C1, Constant *C2, unsigned Flags=0, Type *OnlyIfReducedTy=nullptr)
get - Return a binary or shift operator constant expression, folding if possible. ...
Definition: Constants.cpp:1711
static Value * OptimizeAndOrXor(unsigned Opcode, SmallVectorImpl< ValueEntry > &Ops)
Optimize a series of operands to an &#39;and&#39;, &#39;or&#39;, or &#39;xor&#39; instruction.
FastMathFlags getFastMathFlags() const
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
bool getBoolValue() const
Convert APInt to a boolean value.
Definition: APInt.h:471
std::pair< Value *, APInt > RepeatedValue
bool isNegative() const
Return true if the sign bit is set.
Definition: Constants.h:300
void andIRFlags(const Value *V)
Logical &#39;and&#39; of any supported wrapping, exact, and fast-math flags of V and this instruction...
static Value * createAndInstr(Instruction *InsertBefore, Value *Opnd, const APInt &ConstOpnd)
Helper function of CombineXorOpnd().
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:428
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:290
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
static const Value * getNotArgument(const Value *BinOp)
Value * getOperand(unsigned i) const
Definition: User.h:154
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:301
const BasicBlock & getEntryBlock() const
Definition: Function.h:572
#define P(N)
static Constant * getFNeg(Constant *C)
Definition: Constants.cpp:2103
bool isAllOnesValue() const
Determine if all bits are set.
Definition: APInt.h:389
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:260
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:281
bool areAllPreserved() const
Test whether all analyses are preserved (and none are abandoned).
Definition: PassManager.h:321
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:75
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1164
const char * getOpcodeName() const
Definition: Instruction.h:127
static Value * EmitAddTreeOfValues(Instruction *I, SmallVectorImpl< WeakTrackingVH > &Ops)
Emit a tree of add instructions, summing Ops together and returning the result.
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
This file contains the declarations for the subclasses of Constant, which represent the different fla...
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:264
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:363
static APInt getOneBitSet(unsigned numBits, unsigned BitNo)
Return an APInt with exactly one bit set in the result.
Definition: APInt.h:581
bool isAssociative() const LLVM_READONLY
Return true if the instruction is associative:
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:216
Analysis pass providing a never-invalidated alias analysis result.
Value * getSymbolicPart() const
Definition: Reassociate.cpp:85
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
static bool isNot(const Value *V)
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:374
self_iterator getIterator()
Definition: ilist_node.h:82
static Constant * getAllOnesValue(Type *Ty)
Get the all ones value.
Definition: Constants.cpp:261
NUW NUW NUW NUW Exact static Exact BinaryOperator * CreateNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
iterator erase(const_iterator CI)
Definition: SmallVector.h:449
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
static unsigned CarmichaelShift(unsigned Bitwidth)
Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael function.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:933
const APInt & getConstPart() const
Definition: Reassociate.cpp:87
const APFloat & getValueAPF() const
Definition: Constants.h:294
static BinaryOperator * CreateFNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
Value * getValue() const
Definition: Reassociate.cpp:84
Iterator for intrusive lists based on ilist_node.
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:410
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type...
Definition: Type.cpp:130
static Value * NegateValue(Value *V, Instruction *BI, SetVector< AssertingVH< Instruction >> &ToRedo)
Insert instructions before the instruction pointed to by BI, that computes the negative version of th...
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
unsigned getSymbolicRank() const
Definition: Reassociate.cpp:86
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:63
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:560
void setSymbolicRank(unsigned R)
Definition: Reassociate.cpp:90
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:285
static Constant * get(Type *Ty, double V)
This returns a ConstantFP, or a vector containing a splat of a ConstantFP, for the specified value in...
Definition: Constants.cpp:623
Value handle that asserts if the Value is deleted.
Definition: ValueHandle.h:238
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:420
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:57
static bool isNeg(const Value *V)
Check if the given Value is a NEG, FNeg, or NOT instruction.
Class for arbitrary precision integers.
Definition: APInt.h:69
static Value * buildMultiplyTree(IRBuilder<> &Builder, SmallVectorImpl< Value *> &Ops)
Build a tree of multiplies, computing the product of Ops.
static BinaryOperator * Create(BinaryOps Op, Value *S1, Value *S2, const Twine &Name=Twine(), Instruction *InsertBefore=nullptr)
Construct a binary instruction, given the opcode and the two operands.
iterator_range< user_iterator > users()
Definition: Value.h:395
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:482
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:398
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
amdgpu Simplify well known AMD library false Value Value * Arg
static void FindSingleUseMultiplyFactors(Value *V, SmallVectorImpl< Value *> &Factors)
If V is a single-use multiply, recursively add its operands as factors, otherwise add V to the list o...
static Constant * getNeg(Constant *C, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2096
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
iterator insert(iterator where, pointer New)
Definition: ilist.h:241
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:284
bool mayBeMemoryDependent(const Instruction &I)
Returns true if the result or effects of the given instructions I depend on or influence global memor...
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:656
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
static bool isFNeg(const Value *V, bool IgnoreZeroSign=false)
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:189
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
static void PrintOps(Instruction *I, const SmallVectorImpl< ValueEntry > &Ops)
Print out the expression identified in the Ops list.
Definition: Reassociate.cpp:58
static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode)
Add the extra weight &#39;RHS&#39; to the existing weight &#39;LHS&#39;, reducing the combined weight using any speci...
#define I(x, y, z)
Definition: MD5.cpp:58
bool empty() const
Determine if the SetVector is empty or not.
Definition: SetVector.h:73
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
static Constant * getShl(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2186
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
static BinaryOperator * CreateNeg(Value *S1, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
bool isMinValue() const
Determine if this is the smallest unsigned value.
Definition: APInt.h:430
bool isFPOrFPVectorTy() const
Return true if this is a FP type or a vector of FP.
Definition: Type.h:185
static int const Threshold
TODO: Write a new FunctionPass AliasAnalysis so that it can keep a cache.
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
static BinaryOperator * ConvertShiftToMul(Instruction *Shl)
If this is a shift of a reassociable multiply or is used by one, change this into a multiply by a con...
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
user_iterator user_begin()
Definition: Value.h:371
void initializeReassociateLegacyPassPass(PassRegistry &)
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:115
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:293
LLVM Value Representation.
Definition: Value.h:73
A vector that has set insertion semantics.
Definition: SetVector.h:41
void clearSubclassOptionalData()
Clear the optional flags contained in this value.
Definition: Value.h:471
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:221
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:88
static BinaryOperator * isReassociableOp(Value *V, unsigned Opcode)
Return true if V is an instruction of the specified opcode and if it only has one use...
Invoke instruction.
#define DEBUG(X)
Definition: Debug.h:118
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:408
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
A container for analyses that lazily runs them and caches their results.
static bool ShouldBreakUpSubtract(Instruction *Sub)
Return true if we should break up this subtract of X-Y into (X + -Y).
static APInt getNullValue(unsigned numBits)
Get the &#39;0&#39; value.
Definition: APInt.h:562
static BinaryOperator * CreateMul(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
Value * CreateFMul(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:947
static Constant * getBinOpAbsorber(unsigned Opcode, Type *Ty)
Return the absorbing element for the given binary operation, i.e.
Definition: Constants.cpp:2222
static bool LinearizeExprTree(BinaryOperator *I, SmallVectorImpl< RepeatedValue > &Ops)
Given an associative binary expression, return the leaf nodes in Ops along with their weights (how ma...
bool use_empty() const
Definition: Value.h:322
iterator_range< arg_iterator > args()
Definition: Function.h:621
bool isNullValue() const
Determine if all bits are clear.
Definition: APInt.h:399
cmpResult compare(const APFloat &RHS) const
Definition: APFloat.h:1102
const BasicBlock * getParent() const
Definition: Instruction.h:66
PreservedAnalyses run(Function &F, FunctionAnalysisManager &)