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/APFloat.h"
25 #include "llvm/ADT/APInt.h"
26 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
34 #include "llvm/IR/Argument.h"
35 #include "llvm/IR/BasicBlock.h"
36 #include "llvm/IR/CFG.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/InstrTypes.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Operator.h"
45 #include "llvm/IR/PassManager.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/User.h"
49 #include "llvm/IR/Value.h"
50 #include "llvm/IR/ValueHandle.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/Casting.h"
53 #include "llvm/Support/Debug.h"
56 #include "llvm/Transforms/Scalar.h"
58 #include <algorithm>
59 #include <cassert>
60 #include <utility>
61 
62 using namespace llvm;
63 using namespace reassociate;
64 
65 #define DEBUG_TYPE "reassociate"
66 
67 STATISTIC(NumChanged, "Number of insts reassociated");
68 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
69 STATISTIC(NumFactor , "Number of multiplies factored");
70 
71 #ifndef NDEBUG
72 /// Print out the expression identified in the Ops list.
74  Module *M = I->getModule();
75  dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
76  << *Ops[0].Op->getType() << '\t';
77  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
78  dbgs() << "[ ";
79  Ops[i].Op->printAsOperand(dbgs(), false, M);
80  dbgs() << ", #" << Ops[i].Rank << "] ";
81  }
82 }
83 #endif
84 
85 /// Utility class representing a non-constant Xor-operand. We classify
86 /// non-constant Xor-Operands into two categories:
87 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
88 /// C2)
89 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
90 /// constant.
91 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
92 /// operand as "E | 0"
94 public:
95  XorOpnd(Value *V);
96 
97  bool isInvalid() const { return SymbolicPart == nullptr; }
98  bool isOrExpr() const { return isOr; }
99  Value *getValue() const { return OrigVal; }
100  Value *getSymbolicPart() const { return SymbolicPart; }
101  unsigned getSymbolicRank() const { return SymbolicRank; }
102  const APInt &getConstPart() const { return ConstPart; }
103 
104  void Invalidate() { SymbolicPart = OrigVal = nullptr; }
105  void setSymbolicRank(unsigned R) { SymbolicRank = R; }
106 
107 private:
108  Value *OrigVal;
109  Value *SymbolicPart;
110  APInt ConstPart;
111  unsigned SymbolicRank;
112  bool isOr;
113 };
114 
116  assert(!isa<ConstantInt>(V) && "No ConstantInt");
117  OrigVal = V;
119  SymbolicRank = 0;
120 
121  if (I && (I->getOpcode() == Instruction::Or ||
122  I->getOpcode() == Instruction::And)) {
123  Value *V0 = I->getOperand(0);
124  Value *V1 = I->getOperand(1);
125  const APInt *C;
126  if (match(V0, PatternMatch::m_APInt(C)))
127  std::swap(V0, V1);
128 
129  if (match(V1, PatternMatch::m_APInt(C))) {
130  ConstPart = *C;
131  SymbolicPart = V0;
132  isOr = (I->getOpcode() == Instruction::Or);
133  return;
134  }
135  }
136 
137  // view the operand as "V | 0"
138  SymbolicPart = V;
139  ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits());
140  isOr = true;
141 }
142 
143 /// Return true if V is an instruction of the specified opcode and if it
144 /// only has one use.
145 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
146  auto *I = dyn_cast<Instruction>(V);
147  if (I && I->hasOneUse() && I->getOpcode() == Opcode)
148  if (!isa<FPMathOperator>(I) || I->isFast())
149  return cast<BinaryOperator>(I);
150  return nullptr;
151 }
152 
153 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
154  unsigned Opcode2) {
155  auto *I = dyn_cast<Instruction>(V);
156  if (I && I->hasOneUse() &&
157  (I->getOpcode() == Opcode1 || I->getOpcode() == Opcode2))
158  if (!isa<FPMathOperator>(I) || I->isFast())
159  return cast<BinaryOperator>(I);
160  return nullptr;
161 }
162 
163 void ReassociatePass::BuildRankMap(Function &F,
165  unsigned Rank = 2;
166 
167  // Assign distinct ranks to function arguments.
168  for (auto &Arg : F.args()) {
169  ValueRankMap[&Arg] = ++Rank;
170  DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
171  << "\n");
172  }
173 
174  // Traverse basic blocks in ReversePostOrder
175  for (BasicBlock *BB : RPOT) {
176  unsigned BBRank = RankMap[BB] = ++Rank << 16;
177 
178  // Walk the basic block, adding precomputed ranks for any instructions that
179  // we cannot move. This ensures that the ranks for these instructions are
180  // all different in the block.
181  for (Instruction &I : *BB)
182  if (mayBeMemoryDependent(I))
183  ValueRankMap[&I] = ++BBRank;
184  }
185 }
186 
187 unsigned ReassociatePass::getRank(Value *V) {
189  if (!I) {
190  if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
191  return 0; // Otherwise it's a global or constant, rank 0.
192  }
193 
194  if (unsigned Rank = ValueRankMap[I])
195  return Rank; // Rank already known?
196 
197  // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
198  // we can reassociate expressions for code motion! Since we do not recurse
199  // for PHI nodes, we cannot have infinite recursion here, because there
200  // cannot be loops in the value graph that do not go through PHI nodes.
201  unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
202  for (unsigned i = 0, e = I->getNumOperands();
203  i != e && Rank != MaxRank; ++i)
204  Rank = std::max(Rank, getRank(I->getOperand(i)));
205 
206  // If this is a not or neg instruction, do not count it for rank. This
207  // assures us that X and ~X will have the same rank.
210  ++Rank;
211 
212  DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
213 
214  return ValueRankMap[I] = Rank;
215 }
216 
217 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
218 void ReassociatePass::canonicalizeOperands(Instruction *I) {
219  assert(isa<BinaryOperator>(I) && "Expected binary operator.");
220  assert(I->isCommutative() && "Expected commutative operator.");
221 
222  Value *LHS = I->getOperand(0);
223  Value *RHS = I->getOperand(1);
224  if (LHS == RHS || isa<Constant>(RHS))
225  return;
226  if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
227  cast<BinaryOperator>(I)->swapOperands();
228 }
229 
230 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
231  Instruction *InsertBefore, Value *FlagsOp) {
232  if (S1->getType()->isIntOrIntVectorTy())
233  return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
234  else {
235  BinaryOperator *Res =
236  BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
237  Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
238  return Res;
239  }
240 }
241 
242 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
243  Instruction *InsertBefore, Value *FlagsOp) {
244  if (S1->getType()->isIntOrIntVectorTy())
245  return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
246  else {
247  BinaryOperator *Res =
248  BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
249  Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
250  return Res;
251  }
252 }
253 
255  Instruction *InsertBefore, Value *FlagsOp) {
256  if (S1->getType()->isIntOrIntVectorTy())
257  return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
258  else {
259  BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
260  Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
261  return Res;
262  }
263 }
264 
265 /// Replace 0-X with X*-1.
267  Type *Ty = Neg->getType();
268  Constant *NegOne = Ty->isIntOrIntVectorTy() ?
270 
271  BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
272  Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
273  Res->takeName(Neg);
274  Neg->replaceAllUsesWith(Res);
275  Res->setDebugLoc(Neg->getDebugLoc());
276  return Res;
277 }
278 
279 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
280 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
281 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
282 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
283 /// even x in Bitwidth-bit arithmetic.
284 static unsigned CarmichaelShift(unsigned Bitwidth) {
285  if (Bitwidth < 3)
286  return Bitwidth - 1;
287  return Bitwidth - 2;
288 }
289 
290 /// Add the extra weight 'RHS' to the existing weight 'LHS',
291 /// reducing the combined weight using any special properties of the operation.
292 /// The existing weight LHS represents the computation X op X op ... op X where
293 /// X occurs LHS times. The combined weight represents X op X op ... op X with
294 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
295 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
296 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
297 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
298  // If we were working with infinite precision arithmetic then the combined
299  // weight would be LHS + RHS. But we are using finite precision arithmetic,
300  // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
301  // for nilpotent operations and addition, but not for idempotent operations
302  // and multiplication), so it is important to correctly reduce the combined
303  // weight back into range if wrapping would be wrong.
304 
305  // If RHS is zero then the weight didn't change.
306  if (RHS.isMinValue())
307  return;
308  // If LHS is zero then the combined weight is RHS.
309  if (LHS.isMinValue()) {
310  LHS = RHS;
311  return;
312  }
313  // From this point on we know that neither LHS nor RHS is zero.
314 
315  if (Instruction::isIdempotent(Opcode)) {
316  // Idempotent means X op X === X, so any non-zero weight is equivalent to a
317  // weight of 1. Keeping weights at zero or one also means that wrapping is
318  // not a problem.
319  assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
320  return; // Return a weight of 1.
321  }
322  if (Instruction::isNilpotent(Opcode)) {
323  // Nilpotent means X op X === 0, so reduce weights modulo 2.
324  assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
325  LHS = 0; // 1 + 1 === 0 modulo 2.
326  return;
327  }
328  if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
329  // TODO: Reduce the weight by exploiting nsw/nuw?
330  LHS += RHS;
331  return;
332  }
333 
334  assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
335  "Unknown associative operation!");
336  unsigned Bitwidth = LHS.getBitWidth();
337  // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
338  // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
339  // bit number x, since either x is odd in which case x^CM = 1, or x is even in
340  // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
341  // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
342  // which by a happy accident means that they can always be represented using
343  // Bitwidth bits.
344  // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
345  // the Carmichael number).
346  if (Bitwidth > 3) {
347  /// CM - The value of Carmichael's lambda function.
348  APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
349  // Any weight W >= Threshold can be replaced with W - CM.
350  APInt Threshold = CM + Bitwidth;
351  assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
352  // For Bitwidth 4 or more the following sum does not overflow.
353  LHS += RHS;
354  while (LHS.uge(Threshold))
355  LHS -= CM;
356  } else {
357  // To avoid problems with overflow do everything the same as above but using
358  // a larger type.
359  unsigned CM = 1U << CarmichaelShift(Bitwidth);
360  unsigned Threshold = CM + Bitwidth;
361  assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
362  "Weights not reduced!");
363  unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
364  while (Total >= Threshold)
365  Total -= CM;
366  LHS = Total;
367  }
368 }
369 
370 using RepeatedValue = std::pair<Value*, APInt>;
371 
372 /// Given an associative binary expression, return the leaf
373 /// nodes in Ops along with their weights (how many times the leaf occurs). The
374 /// original expression is the same as
375 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
376 /// op
377 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
378 /// op
379 /// ...
380 /// op
381 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
382 ///
383 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
384 ///
385 /// This routine may modify the function, in which case it returns 'true'. The
386 /// changes it makes may well be destructive, changing the value computed by 'I'
387 /// to something completely different. Thus if the routine returns 'true' then
388 /// you MUST either replace I with a new expression computed from the Ops array,
389 /// or use RewriteExprTree to put the values back in.
390 ///
391 /// A leaf node is either not a binary operation of the same kind as the root
392 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
393 /// opcode), or is the same kind of binary operator but has a use which either
394 /// does not belong to the expression, or does belong to the expression but is
395 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
396 /// of the expression, while for non-leaf nodes (except for the root 'I') every
397 /// use is a non-leaf node of the expression.
398 ///
399 /// For example:
400 /// expression graph node names
401 ///
402 /// + | I
403 /// / \ |
404 /// + + | A, B
405 /// / \ / \ |
406 /// * + * | C, D, E
407 /// / \ / \ / \ |
408 /// + * | F, G
409 ///
410 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
411 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
412 ///
413 /// The expression is maximal: if some instruction is a binary operator of the
414 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
415 /// then the instruction also belongs to the expression, is not a leaf node of
416 /// it, and its operands also belong to the expression (but may be leaf nodes).
417 ///
418 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
419 /// order to ensure that every non-root node in the expression has *exactly one*
420 /// use by a non-leaf node of the expression. This destruction means that the
421 /// caller MUST either replace 'I' with a new expression or use something like
422 /// RewriteExprTree to put the values back in if the routine indicates that it
423 /// made a change by returning 'true'.
424 ///
425 /// In the above example either the right operand of A or the left operand of B
426 /// will be replaced by undef. If it is B's operand then this gives:
427 ///
428 /// + | I
429 /// / \ |
430 /// + + | A, B - operand of B replaced with undef
431 /// / \ \ |
432 /// * + * | C, D, E
433 /// / \ / \ / \ |
434 /// + * | F, G
435 ///
436 /// Note that such undef operands can only be reached by passing through 'I'.
437 /// For example, if you visit operands recursively starting from a leaf node
438 /// then you will never see such an undef operand unless you get back to 'I',
439 /// which requires passing through a phi node.
440 ///
441 /// Note that this routine may also mutate binary operators of the wrong type
442 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
443 /// of the expression) if it can turn them into binary operators of the right
444 /// type and thus make the expression bigger.
447  DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
448  unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
449  unsigned Opcode = I->getOpcode();
450  assert(I->isAssociative() && I->isCommutative() &&
451  "Expected an associative and commutative operation!");
452 
453  // Visit all operands of the expression, keeping track of their weight (the
454  // number of paths from the expression root to the operand, or if you like
455  // the number of times that operand occurs in the linearized expression).
456  // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
457  // while A has weight two.
458 
459  // Worklist of non-leaf nodes (their operands are in the expression too) along
460  // with their weights, representing a certain number of paths to the operator.
461  // If an operator occurs in the worklist multiple times then we found multiple
462  // ways to get to it.
463  SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
464  Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
465  bool Changed = false;
466 
467  // Leaves of the expression are values that either aren't the right kind of
468  // operation (eg: a constant, or a multiply in an add tree), or are, but have
469  // some uses that are not inside the expression. For example, in I = X + X,
470  // X = A + B, the value X has two uses (by I) that are in the expression. If
471  // X has any other uses, for example in a return instruction, then we consider
472  // X to be a leaf, and won't analyze it further. When we first visit a value,
473  // if it has more than one use then at first we conservatively consider it to
474  // be a leaf. Later, as the expression is explored, we may discover some more
475  // uses of the value from inside the expression. If all uses turn out to be
476  // from within the expression (and the value is a binary operator of the right
477  // kind) then the value is no longer considered to be a leaf, and its operands
478  // are explored.
479 
480  // Leaves - Keeps track of the set of putative leaves as well as the number of
481  // paths to each leaf seen so far.
482  using LeafMap = DenseMap<Value *, APInt>;
483  LeafMap Leaves; // Leaf -> Total weight so far.
484  SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
485 
486 #ifndef NDEBUG
487  SmallPtrSet<Value *, 8> Visited; // For sanity checking the iteration scheme.
488 #endif
489  while (!Worklist.empty()) {
490  std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
491  I = P.first; // We examine the operands of this binary operator.
492 
493  for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
494  Value *Op = I->getOperand(OpIdx);
495  APInt Weight = P.second; // Number of paths to this operand.
496  DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
497  assert(!Op->use_empty() && "No uses, so how did we get to it?!");
498 
499  // If this is a binary operation of the right kind with only one use then
500  // add its operands to the expression.
501  if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
502  assert(Visited.insert(Op).second && "Not first visit!");
503  DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
504  Worklist.push_back(std::make_pair(BO, Weight));
505  continue;
506  }
507 
508  // Appears to be a leaf. Is the operand already in the set of leaves?
509  LeafMap::iterator It = Leaves.find(Op);
510  if (It == Leaves.end()) {
511  // Not in the leaf map. Must be the first time we saw this operand.
512  assert(Visited.insert(Op).second && "Not first visit!");
513  if (!Op->hasOneUse()) {
514  // This value has uses not accounted for by the expression, so it is
515  // not safe to modify. Mark it as being a leaf.
516  DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
517  LeafOrder.push_back(Op);
518  Leaves[Op] = Weight;
519  continue;
520  }
521  // No uses outside the expression, try morphing it.
522  } else {
523  // Already in the leaf map.
524  assert(It != Leaves.end() && Visited.count(Op) &&
525  "In leaf map but not visited!");
526 
527  // Update the number of paths to the leaf.
528  IncorporateWeight(It->second, Weight, Opcode);
529 
530 #if 0 // TODO: Re-enable once PR13021 is fixed.
531  // The leaf already has one use from inside the expression. As we want
532  // exactly one such use, drop this new use of the leaf.
533  assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
534  I->setOperand(OpIdx, UndefValue::get(I->getType()));
535  Changed = true;
536 
537  // If the leaf is a binary operation of the right kind and we now see
538  // that its multiple original uses were in fact all by nodes belonging
539  // to the expression, then no longer consider it to be a leaf and add
540  // its operands to the expression.
541  if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
542  DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
543  Worklist.push_back(std::make_pair(BO, It->second));
544  Leaves.erase(It);
545  continue;
546  }
547 #endif
548 
549  // If we still have uses that are not accounted for by the expression
550  // then it is not safe to modify the value.
551  if (!Op->hasOneUse())
552  continue;
553 
554  // No uses outside the expression, try morphing it.
555  Weight = It->second;
556  Leaves.erase(It); // Since the value may be morphed below.
557  }
558 
559  // At this point we have a value which, first of all, is not a binary
560  // expression of the right kind, and secondly, is only used inside the
561  // expression. This means that it can safely be modified. See if we
562  // can usefully morph it into an expression of the right kind.
563  assert((!isa<Instruction>(Op) ||
564  cast<Instruction>(Op)->getOpcode() != Opcode
565  || (isa<FPMathOperator>(Op) &&
566  !cast<Instruction>(Op)->isFast())) &&
567  "Should have been handled above!");
568  assert(Op->hasOneUse() && "Has uses outside the expression tree!");
569 
570  // If this is a multiply expression, turn any internal negations into
571  // multiplies by -1 so they can be reassociated.
572  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
573  if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
574  (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
575  DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
576  BO = LowerNegateToMultiply(BO);
577  DEBUG(dbgs() << *BO << '\n');
578  Worklist.push_back(std::make_pair(BO, Weight));
579  Changed = true;
580  continue;
581  }
582 
583  // Failed to morph into an expression of the right type. This really is
584  // a leaf.
585  DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
586  assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
587  LeafOrder.push_back(Op);
588  Leaves[Op] = Weight;
589  }
590  }
591 
592  // The leaves, repeated according to their weights, represent the linearized
593  // form of the expression.
594  for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
595  Value *V = LeafOrder[i];
596  LeafMap::iterator It = Leaves.find(V);
597  if (It == Leaves.end())
598  // Node initially thought to be a leaf wasn't.
599  continue;
600  assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
601  APInt Weight = It->second;
602  if (Weight.isMinValue())
603  // Leaf already output or weight reduction eliminated it.
604  continue;
605  // Ensure the leaf is only output once.
606  It->second = 0;
607  Ops.push_back(std::make_pair(V, Weight));
608  }
609 
610  // For nilpotent operations or addition there may be no operands, for example
611  // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
612  // in both cases the weight reduces to 0 causing the value to be skipped.
613  if (Ops.empty()) {
614  Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
615  assert(Identity && "Associative operation without identity!");
616  Ops.emplace_back(Identity, APInt(Bitwidth, 1));
617  }
618 
619  return Changed;
620 }
621 
622 /// Now that the operands for this expression tree are
623 /// linearized and optimized, emit them in-order.
624 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
626  assert(Ops.size() > 1 && "Single values should be used directly!");
627 
628  // Since our optimizations should never increase the number of operations, the
629  // new expression can usually be written reusing the existing binary operators
630  // from the original expression tree, without creating any new instructions,
631  // though the rewritten expression may have a completely different topology.
632  // We take care to not change anything if the new expression will be the same
633  // as the original. If more than trivial changes (like commuting operands)
634  // were made then we are obliged to clear out any optional subclass data like
635  // nsw flags.
636 
637  /// NodesToRewrite - Nodes from the original expression available for writing
638  /// the new expression into.
639  SmallVector<BinaryOperator*, 8> NodesToRewrite;
640  unsigned Opcode = I->getOpcode();
641  BinaryOperator *Op = I;
642 
643  /// NotRewritable - The operands being written will be the leaves of the new
644  /// expression and must not be used as inner nodes (via NodesToRewrite) by
645  /// mistake. Inner nodes are always reassociable, and usually leaves are not
646  /// (if they were they would have been incorporated into the expression and so
647  /// would not be leaves), so most of the time there is no danger of this. But
648  /// in rare cases a leaf may become reassociable if an optimization kills uses
649  /// of it, or it may momentarily become reassociable during rewriting (below)
650  /// due it being removed as an operand of one of its uses. Ensure that misuse
651  /// of leaf nodes as inner nodes cannot occur by remembering all of the future
652  /// leaves and refusing to reuse any of them as inner nodes.
653  SmallPtrSet<Value*, 8> NotRewritable;
654  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
655  NotRewritable.insert(Ops[i].Op);
656 
657  // ExpressionChanged - Non-null if the rewritten expression differs from the
658  // original in some non-trivial way, requiring the clearing of optional flags.
659  // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
660  BinaryOperator *ExpressionChanged = nullptr;
661  for (unsigned i = 0; ; ++i) {
662  // The last operation (which comes earliest in the IR) is special as both
663  // operands will come from Ops, rather than just one with the other being
664  // a subexpression.
665  if (i+2 == Ops.size()) {
666  Value *NewLHS = Ops[i].Op;
667  Value *NewRHS = Ops[i+1].Op;
668  Value *OldLHS = Op->getOperand(0);
669  Value *OldRHS = Op->getOperand(1);
670 
671  if (NewLHS == OldLHS && NewRHS == OldRHS)
672  // Nothing changed, leave it alone.
673  break;
674 
675  if (NewLHS == OldRHS && NewRHS == OldLHS) {
676  // The order of the operands was reversed. Swap them.
677  DEBUG(dbgs() << "RA: " << *Op << '\n');
678  Op->swapOperands();
679  DEBUG(dbgs() << "TO: " << *Op << '\n');
680  MadeChange = true;
681  ++NumChanged;
682  break;
683  }
684 
685  // The new operation differs non-trivially from the original. Overwrite
686  // the old operands with the new ones.
687  DEBUG(dbgs() << "RA: " << *Op << '\n');
688  if (NewLHS != OldLHS) {
689  BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
690  if (BO && !NotRewritable.count(BO))
691  NodesToRewrite.push_back(BO);
692  Op->setOperand(0, NewLHS);
693  }
694  if (NewRHS != OldRHS) {
695  BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
696  if (BO && !NotRewritable.count(BO))
697  NodesToRewrite.push_back(BO);
698  Op->setOperand(1, NewRHS);
699  }
700  DEBUG(dbgs() << "TO: " << *Op << '\n');
701 
702  ExpressionChanged = Op;
703  MadeChange = true;
704  ++NumChanged;
705 
706  break;
707  }
708 
709  // Not the last operation. The left-hand side will be a sub-expression
710  // while the right-hand side will be the current element of Ops.
711  Value *NewRHS = Ops[i].Op;
712  if (NewRHS != Op->getOperand(1)) {
713  DEBUG(dbgs() << "RA: " << *Op << '\n');
714  if (NewRHS == Op->getOperand(0)) {
715  // The new right-hand side was already present as the left operand. If
716  // we are lucky then swapping the operands will sort out both of them.
717  Op->swapOperands();
718  } else {
719  // Overwrite with the new right-hand side.
720  BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
721  if (BO && !NotRewritable.count(BO))
722  NodesToRewrite.push_back(BO);
723  Op->setOperand(1, NewRHS);
724  ExpressionChanged = Op;
725  }
726  DEBUG(dbgs() << "TO: " << *Op << '\n');
727  MadeChange = true;
728  ++NumChanged;
729  }
730 
731  // Now deal with the left-hand side. If this is already an operation node
732  // from the original expression then just rewrite the rest of the expression
733  // into it.
734  BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
735  if (BO && !NotRewritable.count(BO)) {
736  Op = BO;
737  continue;
738  }
739 
740  // Otherwise, grab a spare node from the original expression and use that as
741  // the left-hand side. If there are no nodes left then the optimizers made
742  // an expression with more nodes than the original! This usually means that
743  // they did something stupid but it might mean that the problem was just too
744  // hard (finding the mimimal number of multiplications needed to realize a
745  // multiplication expression is NP-complete). Whatever the reason, smart or
746  // stupid, create a new node if there are none left.
747  BinaryOperator *NewOp;
748  if (NodesToRewrite.empty()) {
751  Undef, Undef, "", I);
752  if (NewOp->getType()->isFPOrFPVectorTy())
753  NewOp->setFastMathFlags(I->getFastMathFlags());
754  } else {
755  NewOp = NodesToRewrite.pop_back_val();
756  }
757 
758  DEBUG(dbgs() << "RA: " << *Op << '\n');
759  Op->setOperand(0, NewOp);
760  DEBUG(dbgs() << "TO: " << *Op << '\n');
761  ExpressionChanged = Op;
762  MadeChange = true;
763  ++NumChanged;
764  Op = NewOp;
765  }
766 
767  // If the expression changed non-trivially then clear out all subclass data
768  // starting from the operator specified in ExpressionChanged, and compactify
769  // the operators to just before the expression root to guarantee that the
770  // expression tree is dominated by all of Ops.
771  if (ExpressionChanged)
772  do {
773  // Preserve FastMathFlags.
774  if (isa<FPMathOperator>(I)) {
775  FastMathFlags Flags = I->getFastMathFlags();
776  ExpressionChanged->clearSubclassOptionalData();
777  ExpressionChanged->setFastMathFlags(Flags);
778  } else
779  ExpressionChanged->clearSubclassOptionalData();
780 
781  if (ExpressionChanged == I)
782  break;
783  ExpressionChanged->moveBefore(I);
784  ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
785  } while (true);
786 
787  // Throw away any left over nodes from the original expression.
788  for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
789  RedoInsts.insert(NodesToRewrite[i]);
790 }
791 
792 /// Insert instructions before the instruction pointed to by BI,
793 /// that computes the negative version of the value specified. The negative
794 /// version of the value is returned, and BI is left pointing at the instruction
795 /// that should be processed next by the reassociation pass.
796 /// Also add intermediate instructions to the redo list that are modified while
797 /// pushing the negates through adds. These will be revisited to see if
798 /// additional opportunities have been exposed.
801  if (auto *C = dyn_cast<Constant>(V))
804 
805  // We are trying to expose opportunity for reassociation. One of the things
806  // that we want to do to achieve this is to push a negation as deep into an
807  // expression chain as possible, to expose the add instructions. In practice,
808  // this means that we turn this:
809  // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
810  // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
811  // the constants. We assume that instcombine will clean up the mess later if
812  // we introduce tons of unnecessary negation instructions.
813  //
814  if (BinaryOperator *I =
815  isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
816  // Push the negates through the add.
817  I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
818  I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
819  if (I->getOpcode() == Instruction::Add) {
820  I->setHasNoUnsignedWrap(false);
821  I->setHasNoSignedWrap(false);
822  }
823 
824  // We must move the add instruction here, because the neg instructions do
825  // not dominate the old add instruction in general. By moving it, we are
826  // assured that the neg instructions we just inserted dominate the
827  // instruction we are about to insert after them.
828  //
829  I->moveBefore(BI);
830  I->setName(I->getName()+".neg");
831 
832  // Add the intermediate negates to the redo list as processing them later
833  // could expose more reassociating opportunities.
834  ToRedo.insert(I);
835  return I;
836  }
837 
838  // Okay, we need to materialize a negated version of V with an instruction.
839  // Scan the use lists of V to see if we have one already.
840  for (User *U : V->users()) {
842  continue;
843 
844  // We found one! Now we have to make sure that the definition dominates
845  // this use. We do this by moving it to the entry block (if it is a
846  // non-instruction value) or right after the definition. These negates will
847  // be zapped by reassociate later, so we don't need much finesse here.
848  BinaryOperator *TheNeg = cast<BinaryOperator>(U);
849 
850  // Verify that the negate is in this function, V might be a constant expr.
851  if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
852  continue;
853 
854  BasicBlock::iterator InsertPt;
855  if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
856  if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
857  InsertPt = II->getNormalDest()->begin();
858  } else {
859  InsertPt = ++InstInput->getIterator();
860  }
861  while (isa<PHINode>(InsertPt)) ++InsertPt;
862  } else {
863  InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
864  }
865  TheNeg->moveBefore(&*InsertPt);
866  if (TheNeg->getOpcode() == Instruction::Sub) {
867  TheNeg->setHasNoUnsignedWrap(false);
868  TheNeg->setHasNoSignedWrap(false);
869  } else {
870  TheNeg->andIRFlags(BI);
871  }
872  ToRedo.insert(TheNeg);
873  return TheNeg;
874  }
875 
876  // Insert a 'neg' instruction that subtracts the value from zero to get the
877  // negation.
878  BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
879  ToRedo.insert(NewNeg);
880  return NewNeg;
881 }
882 
883 /// Return true if we should break up this subtract of X-Y into (X + -Y).
885  // If this is a negation, we can't split it up!
887  return false;
888 
889  // Don't breakup X - undef.
890  if (isa<UndefValue>(Sub->getOperand(1)))
891  return false;
892 
893  // Don't bother to break this up unless either the LHS is an associable add or
894  // subtract or if this is only used by one.
895  Value *V0 = Sub->getOperand(0);
896  if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
897  isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
898  return true;
899  Value *V1 = Sub->getOperand(1);
900  if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
901  isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
902  return true;
903  Value *VB = Sub->user_back();
904  if (Sub->hasOneUse() &&
905  (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
906  isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
907  return true;
908 
909  return false;
910 }
911 
912 /// If we have (X-Y), and if either X is an add, or if this is only used by an
913 /// add, transform this into (X+(0-Y)) to promote better reassociation.
914 static BinaryOperator *
916  // Convert a subtract into an add and a neg instruction. This allows sub
917  // instructions to be commuted with other add instructions.
918  //
919  // Calculate the negative value of Operand 1 of the sub instruction,
920  // and set it as the RHS of the add instruction we just made.
921  Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
922  BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
923  Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
924  Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
925  New->takeName(Sub);
926 
927  // Everyone now refers to the add instruction.
928  Sub->replaceAllUsesWith(New);
929  New->setDebugLoc(Sub->getDebugLoc());
930 
931  DEBUG(dbgs() << "Negated: " << *New << '\n');
932  return New;
933 }
934 
935 /// If this is a shift of a reassociable multiply or is used by one, change
936 /// this into a multiply by a constant to assist with further reassociation.
938  Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
939  MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
940 
941  BinaryOperator *Mul =
942  BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
943  Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
944  Mul->takeName(Shl);
945 
946  // Everyone now refers to the mul instruction.
947  Shl->replaceAllUsesWith(Mul);
948  Mul->setDebugLoc(Shl->getDebugLoc());
949 
950  // We can safely preserve the nuw flag in all cases. It's also safe to turn a
951  // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
952  // handling.
953  bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
954  bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
955  if (NSW && NUW)
956  Mul->setHasNoSignedWrap(true);
957  Mul->setHasNoUnsignedWrap(NUW);
958  return Mul;
959 }
960 
961 /// Scan backwards and forwards among values with the same rank as element i
962 /// to see if X exists. If X does not exist, return i. This is useful when
963 /// scanning for 'x' when we see '-x' because they both get the same rank.
965  unsigned i, Value *X) {
966  unsigned XRank = Ops[i].Rank;
967  unsigned e = Ops.size();
968  for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
969  if (Ops[j].Op == X)
970  return j;
971  if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
972  if (Instruction *I2 = dyn_cast<Instruction>(X))
973  if (I1->isIdenticalTo(I2))
974  return j;
975  }
976  // Scan backwards.
977  for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
978  if (Ops[j].Op == X)
979  return j;
980  if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
981  if (Instruction *I2 = dyn_cast<Instruction>(X))
982  if (I1->isIdenticalTo(I2))
983  return j;
984  }
985  return i;
986 }
987 
988 /// Emit a tree of add instructions, summing Ops together
989 /// and returning the result. Insert the tree before I.
992  if (Ops.size() == 1) return Ops.back();
993 
994  Value *V1 = Ops.back();
995  Ops.pop_back();
996  Value *V2 = EmitAddTreeOfValues(I, Ops);
997  return CreateAdd(V2, V1, "reass.add", I, I);
998 }
999 
1000 /// If V is an expression tree that is a multiplication sequence,
1001 /// and if this sequence contains a multiply by Factor,
1002 /// remove Factor from the tree and return the new tree.
1003 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1004  BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1005  if (!BO)
1006  return nullptr;
1007 
1009  MadeChange |= LinearizeExprTree(BO, Tree);
1011  Factors.reserve(Tree.size());
1012  for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1013  RepeatedValue E = Tree[i];
1014  Factors.append(E.second.getZExtValue(),
1015  ValueEntry(getRank(E.first), E.first));
1016  }
1017 
1018  bool FoundFactor = false;
1019  bool NeedsNegate = false;
1020  for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1021  if (Factors[i].Op == Factor) {
1022  FoundFactor = true;
1023  Factors.erase(Factors.begin()+i);
1024  break;
1025  }
1026 
1027  // If this is a negative version of this factor, remove it.
1028  if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1029  if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1030  if (FC1->getValue() == -FC2->getValue()) {
1031  FoundFactor = NeedsNegate = true;
1032  Factors.erase(Factors.begin()+i);
1033  break;
1034  }
1035  } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1036  if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1037  const APFloat &F1 = FC1->getValueAPF();
1038  APFloat F2(FC2->getValueAPF());
1039  F2.changeSign();
1040  if (F1.compare(F2) == APFloat::cmpEqual) {
1041  FoundFactor = NeedsNegate = true;
1042  Factors.erase(Factors.begin() + i);
1043  break;
1044  }
1045  }
1046  }
1047  }
1048 
1049  if (!FoundFactor) {
1050  // Make sure to restore the operands to the expression tree.
1051  RewriteExprTree(BO, Factors);
1052  return nullptr;
1053  }
1054 
1055  BasicBlock::iterator InsertPt = ++BO->getIterator();
1056 
1057  // If this was just a single multiply, remove the multiply and return the only
1058  // remaining operand.
1059  if (Factors.size() == 1) {
1060  RedoInsts.insert(BO);
1061  V = Factors[0].Op;
1062  } else {
1063  RewriteExprTree(BO, Factors);
1064  V = BO;
1065  }
1066 
1067  if (NeedsNegate)
1068  V = CreateNeg(V, "neg", &*InsertPt, BO);
1069 
1070  return V;
1071 }
1072 
1073 /// If V is a single-use multiply, recursively add its operands as factors,
1074 /// otherwise add V to the list of factors.
1075 ///
1076 /// Ops is the top-level list of add operands we're trying to factor.
1078  SmallVectorImpl<Value*> &Factors) {
1079  BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1080  if (!BO) {
1081  Factors.push_back(V);
1082  return;
1083  }
1084 
1085  // Otherwise, add the LHS and RHS to the list of factors.
1086  FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1087  FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1088 }
1089 
1090 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1091 /// This optimizes based on identities. If it can be reduced to a single Value,
1092 /// it is returned, otherwise the Ops list is mutated as necessary.
1093 static Value *OptimizeAndOrXor(unsigned Opcode,
1095  // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1096  // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1097  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1098  // First, check for X and ~X in the operand list.
1099  assert(i < Ops.size());
1100  if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1101  Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1102  unsigned FoundX = FindInOperandList(Ops, i, X);
1103  if (FoundX != i) {
1104  if (Opcode == Instruction::And) // ...&X&~X = 0
1105  return Constant::getNullValue(X->getType());
1106 
1107  if (Opcode == Instruction::Or) // ...|X|~X = -1
1108  return Constant::getAllOnesValue(X->getType());
1109  }
1110  }
1111 
1112  // Next, check for duplicate pairs of values, which we assume are next to
1113  // each other, due to our sorting criteria.
1114  assert(i < Ops.size());
1115  if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1116  if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1117  // Drop duplicate values for And and Or.
1118  Ops.erase(Ops.begin()+i);
1119  --i; --e;
1120  ++NumAnnihil;
1121  continue;
1122  }
1123 
1124  // Drop pairs of values for Xor.
1125  assert(Opcode == Instruction::Xor);
1126  if (e == 2)
1127  return Constant::getNullValue(Ops[0].Op->getType());
1128 
1129  // Y ^ X^X -> Y
1130  Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1131  i -= 1; e -= 2;
1132  ++NumAnnihil;
1133  }
1134  }
1135  return nullptr;
1136 }
1137 
1138 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1139 /// instruction with the given two operands, and return the resulting
1140 /// instruction. There are two special cases: 1) if the constant operand is 0,
1141 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1142 /// be returned.
1143 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1144  const APInt &ConstOpnd) {
1145  if (ConstOpnd.isNullValue())
1146  return nullptr;
1147 
1148  if (ConstOpnd.isAllOnesValue())
1149  return Opnd;
1150 
1151  Instruction *I = BinaryOperator::CreateAnd(
1152  Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1153  InsertBefore);
1154  I->setDebugLoc(InsertBefore->getDebugLoc());
1155  return I;
1156 }
1157 
1158 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1159 // into "R ^ C", where C would be 0, and R is a symbolic value.
1160 //
1161 // If it was successful, true is returned, and the "R" and "C" is returned
1162 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1163 // and both "Res" and "ConstOpnd" remain unchanged.
1164 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1165  APInt &ConstOpnd, Value *&Res) {
1166  // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1167  // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1168  // = (x & ~c1) ^ (c1 ^ c2)
1169  // It is useful only when c1 == c2.
1170  if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue())
1171  return false;
1172 
1173  if (!Opnd1->getValue()->hasOneUse())
1174  return false;
1175 
1176  const APInt &C1 = Opnd1->getConstPart();
1177  if (C1 != ConstOpnd)
1178  return false;
1179 
1180  Value *X = Opnd1->getSymbolicPart();
1181  Res = createAndInstr(I, X, ~C1);
1182  // ConstOpnd was C2, now C1 ^ C2.
1183  ConstOpnd ^= C1;
1184 
1185  if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1186  RedoInsts.insert(T);
1187  return true;
1188 }
1189 
1190 // Helper function of OptimizeXor(). It tries to simplify
1191 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1192 // symbolic value.
1193 //
1194 // If it was successful, true is returned, and the "R" and "C" is returned
1195 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1196 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1197 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1198 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1199  XorOpnd *Opnd2, APInt &ConstOpnd,
1200  Value *&Res) {
1201  Value *X = Opnd1->getSymbolicPart();
1202  if (X != Opnd2->getSymbolicPart())
1203  return false;
1204 
1205  // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1206  int DeadInstNum = 1;
1207  if (Opnd1->getValue()->hasOneUse())
1208  DeadInstNum++;
1209  if (Opnd2->getValue()->hasOneUse())
1210  DeadInstNum++;
1211 
1212  // Xor-Rule 2:
1213  // (x | c1) ^ (x & c2)
1214  // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1215  // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1216  // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1217  //
1218  if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1219  if (Opnd2->isOrExpr())
1220  std::swap(Opnd1, Opnd2);
1221 
1222  const APInt &C1 = Opnd1->getConstPart();
1223  const APInt &C2 = Opnd2->getConstPart();
1224  APInt C3((~C1) ^ C2);
1225 
1226  // Do not increase code size!
1227  if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1228  int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1229  if (NewInstNum > DeadInstNum)
1230  return false;
1231  }
1232 
1233  Res = createAndInstr(I, X, C3);
1234  ConstOpnd ^= C1;
1235  } else if (Opnd1->isOrExpr()) {
1236  // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1237  //
1238  const APInt &C1 = Opnd1->getConstPart();
1239  const APInt &C2 = Opnd2->getConstPart();
1240  APInt C3 = C1 ^ C2;
1241 
1242  // Do not increase code size
1243  if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1244  int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1245  if (NewInstNum > DeadInstNum)
1246  return false;
1247  }
1248 
1249  Res = createAndInstr(I, X, C3);
1250  ConstOpnd ^= C3;
1251  } else {
1252  // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1253  //
1254  const APInt &C1 = Opnd1->getConstPart();
1255  const APInt &C2 = Opnd2->getConstPart();
1256  APInt C3 = C1 ^ C2;
1257  Res = createAndInstr(I, X, C3);
1258  }
1259 
1260  // Put the original operands in the Redo list; hope they will be deleted
1261  // as dead code.
1262  if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1263  RedoInsts.insert(T);
1264  if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1265  RedoInsts.insert(T);
1266 
1267  return true;
1268 }
1269 
1270 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1271 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1272 /// necessary.
1273 Value *ReassociatePass::OptimizeXor(Instruction *I,
1275  if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1276  return V;
1277 
1278  if (Ops.size() == 1)
1279  return nullptr;
1280 
1282  SmallVector<XorOpnd*, 8> OpndPtrs;
1283  Type *Ty = Ops[0].Op->getType();
1284  APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1285 
1286  // Step 1: Convert ValueEntry to XorOpnd
1287  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1288  Value *V = Ops[i].Op;
1289  const APInt *C;
1290  // TODO: Support non-splat vectors.
1291  if (match(V, PatternMatch::m_APInt(C))) {
1292  ConstOpnd ^= *C;
1293  } else {
1294  XorOpnd O(V);
1295  O.setSymbolicRank(getRank(O.getSymbolicPart()));
1296  Opnds.push_back(O);
1297  }
1298  }
1299 
1300  // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1301  // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1302  // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1303  // with the previous loop --- the iterator of the "Opnds" may be invalidated
1304  // when new elements are added to the vector.
1305  for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1306  OpndPtrs.push_back(&Opnds[i]);
1307 
1308  // Step 2: Sort the Xor-Operands in a way such that the operands containing
1309  // the same symbolic value cluster together. For instance, the input operand
1310  // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1311  // ("x | 123", "x & 789", "y & 456").
1312  //
1313  // The purpose is twofold:
1314  // 1) Cluster together the operands sharing the same symbolic-value.
1315  // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1316  // could potentially shorten crital path, and expose more loop-invariants.
1317  // Note that values' rank are basically defined in RPO order (FIXME).
1318  // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1319  // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1320  // "z" in the order of X-Y-Z is better than any other orders.
1321  std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1322  [](XorOpnd *LHS, XorOpnd *RHS) {
1323  return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1324  });
1325 
1326  // Step 3: Combine adjacent operands
1327  XorOpnd *PrevOpnd = nullptr;
1328  bool Changed = false;
1329  for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1330  XorOpnd *CurrOpnd = OpndPtrs[i];
1331  // The combined value
1332  Value *CV;
1333 
1334  // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1335  if (!ConstOpnd.isNullValue() &&
1336  CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1337  Changed = true;
1338  if (CV)
1339  *CurrOpnd = XorOpnd(CV);
1340  else {
1341  CurrOpnd->Invalidate();
1342  continue;
1343  }
1344  }
1345 
1346  if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1347  PrevOpnd = CurrOpnd;
1348  continue;
1349  }
1350 
1351  // step 3.2: When previous and current operands share the same symbolic
1352  // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1353  if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1354  // Remove previous operand
1355  PrevOpnd->Invalidate();
1356  if (CV) {
1357  *CurrOpnd = XorOpnd(CV);
1358  PrevOpnd = CurrOpnd;
1359  } else {
1360  CurrOpnd->Invalidate();
1361  PrevOpnd = nullptr;
1362  }
1363  Changed = true;
1364  }
1365  }
1366 
1367  // Step 4: Reassemble the Ops
1368  if (Changed) {
1369  Ops.clear();
1370  for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1371  XorOpnd &O = Opnds[i];
1372  if (O.isInvalid())
1373  continue;
1374  ValueEntry VE(getRank(O.getValue()), O.getValue());
1375  Ops.push_back(VE);
1376  }
1377  if (!ConstOpnd.isNullValue()) {
1378  Value *C = ConstantInt::get(Ty, ConstOpnd);
1379  ValueEntry VE(getRank(C), C);
1380  Ops.push_back(VE);
1381  }
1382  unsigned Sz = Ops.size();
1383  if (Sz == 1)
1384  return Ops.back().Op;
1385  if (Sz == 0) {
1386  assert(ConstOpnd.isNullValue());
1387  return ConstantInt::get(Ty, ConstOpnd);
1388  }
1389  }
1390 
1391  return nullptr;
1392 }
1393 
1394 /// Optimize a series of operands to an 'add' instruction. This
1395 /// optimizes based on identities. If it can be reduced to a single Value, it
1396 /// is returned, otherwise the Ops list is mutated as necessary.
1397 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1399  // Scan the operand lists looking for X and -X pairs. If we find any, we
1400  // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1401  // scan for any
1402  // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1403 
1404  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1405  Value *TheOp = Ops[i].Op;
1406  // Check to see if we've seen this operand before. If so, we factor all
1407  // instances of the operand together. Due to our sorting criteria, we know
1408  // that these need to be next to each other in the vector.
1409  if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1410  // Rescan the list, remove all instances of this operand from the expr.
1411  unsigned NumFound = 0;
1412  do {
1413  Ops.erase(Ops.begin()+i);
1414  ++NumFound;
1415  } while (i != Ops.size() && Ops[i].Op == TheOp);
1416 
1417  DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1418  ++NumFactor;
1419 
1420  // Insert a new multiply.
1421  Type *Ty = TheOp->getType();
1422  Constant *C = Ty->isIntOrIntVectorTy() ?
1423  ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1424  Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1425 
1426  // Now that we have inserted a multiply, optimize it. This allows us to
1427  // handle cases that require multiple factoring steps, such as this:
1428  // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1429  RedoInsts.insert(Mul);
1430 
1431  // If every add operand was a duplicate, return the multiply.
1432  if (Ops.empty())
1433  return Mul;
1434 
1435  // Otherwise, we had some input that didn't have the dupe, such as
1436  // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1437  // things being added by this operation.
1438  Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1439 
1440  --i;
1441  e = Ops.size();
1442  continue;
1443  }
1444 
1445  // Check for X and -X or X and ~X in the operand list.
1446  if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1447  !BinaryOperator::isNot(TheOp))
1448  continue;
1449 
1450  Value *X = nullptr;
1451  if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1452  X = BinaryOperator::getNegArgument(TheOp);
1453  else if (BinaryOperator::isNot(TheOp))
1454  X = BinaryOperator::getNotArgument(TheOp);
1455 
1456  unsigned FoundX = FindInOperandList(Ops, i, X);
1457  if (FoundX == i)
1458  continue;
1459 
1460  // Remove X and -X from the operand list.
1461  if (Ops.size() == 2 &&
1462  (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1463  return Constant::getNullValue(X->getType());
1464 
1465  // Remove X and ~X from the operand list.
1466  if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1467  return Constant::getAllOnesValue(X->getType());
1468 
1469  Ops.erase(Ops.begin()+i);
1470  if (i < FoundX)
1471  --FoundX;
1472  else
1473  --i; // Need to back up an extra one.
1474  Ops.erase(Ops.begin()+FoundX);
1475  ++NumAnnihil;
1476  --i; // Revisit element.
1477  e -= 2; // Removed two elements.
1478 
1479  // if X and ~X we append -1 to the operand list.
1480  if (BinaryOperator::isNot(TheOp)) {
1482  Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1483  e += 1;
1484  }
1485  }
1486 
1487  // Scan the operand list, checking to see if there are any common factors
1488  // between operands. Consider something like A*A+A*B*C+D. We would like to
1489  // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1490  // To efficiently find this, we count the number of times a factor occurs
1491  // for any ADD operands that are MULs.
1492  DenseMap<Value*, unsigned> FactorOccurrences;
1493 
1494  // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1495  // where they are actually the same multiply.
1496  unsigned MaxOcc = 0;
1497  Value *MaxOccVal = nullptr;
1498  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1499  BinaryOperator *BOp =
1500  isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1501  if (!BOp)
1502  continue;
1503 
1504  // Compute all of the factors of this added value.
1505  SmallVector<Value*, 8> Factors;
1506  FindSingleUseMultiplyFactors(BOp, Factors);
1507  assert(Factors.size() > 1 && "Bad linearize!");
1508 
1509  // Add one to FactorOccurrences for each unique factor in this op.
1510  SmallPtrSet<Value*, 8> Duplicates;
1511  for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1512  Value *Factor = Factors[i];
1513  if (!Duplicates.insert(Factor).second)
1514  continue;
1515 
1516  unsigned Occ = ++FactorOccurrences[Factor];
1517  if (Occ > MaxOcc) {
1518  MaxOcc = Occ;
1519  MaxOccVal = Factor;
1520  }
1521 
1522  // If Factor is a negative constant, add the negated value as a factor
1523  // because we can percolate the negate out. Watch for minint, which
1524  // cannot be positivified.
1525  if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1526  if (CI->isNegative() && !CI->isMinValue(true)) {
1527  Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1528  if (!Duplicates.insert(Factor).second)
1529  continue;
1530  unsigned Occ = ++FactorOccurrences[Factor];
1531  if (Occ > MaxOcc) {
1532  MaxOcc = Occ;
1533  MaxOccVal = Factor;
1534  }
1535  }
1536  } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1537  if (CF->isNegative()) {
1538  APFloat F(CF->getValueAPF());
1539  F.changeSign();
1540  Factor = ConstantFP::get(CF->getContext(), F);
1541  if (!Duplicates.insert(Factor).second)
1542  continue;
1543  unsigned Occ = ++FactorOccurrences[Factor];
1544  if (Occ > MaxOcc) {
1545  MaxOcc = Occ;
1546  MaxOccVal = Factor;
1547  }
1548  }
1549  }
1550  }
1551  }
1552 
1553  // If any factor occurred more than one time, we can pull it out.
1554  if (MaxOcc > 1) {
1555  DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1556  ++NumFactor;
1557 
1558  // Create a new instruction that uses the MaxOccVal twice. If we don't do
1559  // this, we could otherwise run into situations where removing a factor
1560  // from an expression will drop a use of maxocc, and this can cause
1561  // RemoveFactorFromExpression on successive values to behave differently.
1562  Instruction *DummyInst =
1563  I->getType()->isIntOrIntVectorTy()
1564  ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1565  : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1566 
1568  for (unsigned i = 0; i != Ops.size(); ++i) {
1569  // Only try to remove factors from expressions we're allowed to.
1570  BinaryOperator *BOp =
1571  isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1572  if (!BOp)
1573  continue;
1574 
1575  if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1576  // The factorized operand may occur several times. Convert them all in
1577  // one fell swoop.
1578  for (unsigned j = Ops.size(); j != i;) {
1579  --j;
1580  if (Ops[j].Op == Ops[i].Op) {
1581  NewMulOps.push_back(V);
1582  Ops.erase(Ops.begin()+j);
1583  }
1584  }
1585  --i;
1586  }
1587  }
1588 
1589  // No need for extra uses anymore.
1590  DummyInst->deleteValue();
1591 
1592  unsigned NumAddedValues = NewMulOps.size();
1593  Value *V = EmitAddTreeOfValues(I, NewMulOps);
1594 
1595  // Now that we have inserted the add tree, optimize it. This allows us to
1596  // handle cases that require multiple factoring steps, such as this:
1597  // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1598  assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1599  (void)NumAddedValues;
1600  if (Instruction *VI = dyn_cast<Instruction>(V))
1601  RedoInsts.insert(VI);
1602 
1603  // Create the multiply.
1604  Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I);
1605 
1606  // Rerun associate on the multiply in case the inner expression turned into
1607  // a multiply. We want to make sure that we keep things in canonical form.
1608  RedoInsts.insert(V2);
1609 
1610  // If every add operand included the factor (e.g. "A*B + A*C"), then the
1611  // entire result expression is just the multiply "A*(B+C)".
1612  if (Ops.empty())
1613  return V2;
1614 
1615  // Otherwise, we had some input that didn't have the factor, such as
1616  // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1617  // things being added by this operation.
1618  Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1619  }
1620 
1621  return nullptr;
1622 }
1623 
1624 /// \brief Build up a vector of value/power pairs factoring a product.
1625 ///
1626 /// Given a series of multiplication operands, build a vector of factors and
1627 /// the powers each is raised to when forming the final product. Sort them in
1628 /// the order of descending power.
1629 ///
1630 /// (x*x) -> [(x, 2)]
1631 /// ((x*x)*x) -> [(x, 3)]
1632 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1633 ///
1634 /// \returns Whether any factors have a power greater than one.
1636  SmallVectorImpl<Factor> &Factors) {
1637  // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1638  // Compute the sum of powers of simplifiable factors.
1639  unsigned FactorPowerSum = 0;
1640  for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1641  Value *Op = Ops[Idx-1].Op;
1642 
1643  // Count the number of occurrences of this value.
1644  unsigned Count = 1;
1645  for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1646  ++Count;
1647  // Track for simplification all factors which occur 2 or more times.
1648  if (Count > 1)
1649  FactorPowerSum += Count;
1650  }
1651 
1652  // We can only simplify factors if the sum of the powers of our simplifiable
1653  // factors is 4 or higher. When that is the case, we will *always* have
1654  // a simplification. This is an important invariant to prevent cyclicly
1655  // trying to simplify already minimal formations.
1656  if (FactorPowerSum < 4)
1657  return false;
1658 
1659  // Now gather the simplifiable factors, removing them from Ops.
1660  FactorPowerSum = 0;
1661  for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1662  Value *Op = Ops[Idx-1].Op;
1663 
1664  // Count the number of occurrences of this value.
1665  unsigned Count = 1;
1666  for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1667  ++Count;
1668  if (Count == 1)
1669  continue;
1670  // Move an even number of occurrences to Factors.
1671  Count &= ~1U;
1672  Idx -= Count;
1673  FactorPowerSum += Count;
1674  Factors.push_back(Factor(Op, Count));
1675  Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1676  }
1677 
1678  // None of the adjustments above should have reduced the sum of factor powers
1679  // below our mininum of '4'.
1680  assert(FactorPowerSum >= 4);
1681 
1682  std::stable_sort(Factors.begin(), Factors.end(),
1683  [](const Factor &LHS, const Factor &RHS) {
1684  return LHS.Power > RHS.Power;
1685  });
1686  return true;
1687 }
1688 
1689 /// \brief Build a tree of multiplies, computing the product of Ops.
1691  SmallVectorImpl<Value*> &Ops) {
1692  if (Ops.size() == 1)
1693  return Ops.back();
1694 
1695  Value *LHS = Ops.pop_back_val();
1696  do {
1697  if (LHS->getType()->isIntOrIntVectorTy())
1698  LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1699  else
1700  LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1701  } while (!Ops.empty());
1702 
1703  return LHS;
1704 }
1705 
1706 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1707 ///
1708 /// Given a vector of values raised to various powers, where no two values are
1709 /// equal and the powers are sorted in decreasing order, compute the minimal
1710 /// DAG of multiplies to compute the final product, and return that product
1711 /// value.
1712 Value *
1713 ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1714  SmallVectorImpl<Factor> &Factors) {
1715  assert(Factors[0].Power);
1716  SmallVector<Value *, 4> OuterProduct;
1717  for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1718  Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1719  if (Factors[Idx].Power != Factors[LastIdx].Power) {
1720  LastIdx = Idx;
1721  continue;
1722  }
1723 
1724  // We want to multiply across all the factors with the same power so that
1725  // we can raise them to that power as a single entity. Build a mini tree
1726  // for that.
1727  SmallVector<Value *, 4> InnerProduct;
1728  InnerProduct.push_back(Factors[LastIdx].Base);
1729  do {
1730  InnerProduct.push_back(Factors[Idx].Base);
1731  ++Idx;
1732  } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1733 
1734  // Reset the base value of the first factor to the new expression tree.
1735  // We'll remove all the factors with the same power in a second pass.
1736  Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1737  if (Instruction *MI = dyn_cast<Instruction>(M))
1738  RedoInsts.insert(MI);
1739 
1740  LastIdx = Idx;
1741  }
1742  // Unique factors with equal powers -- we've folded them into the first one's
1743  // base.
1744  Factors.erase(std::unique(Factors.begin(), Factors.end(),
1745  [](const Factor &LHS, const Factor &RHS) {
1746  return LHS.Power == RHS.Power;
1747  }),
1748  Factors.end());
1749 
1750  // Iteratively collect the base of each factor with an add power into the
1751  // outer product, and halve each power in preparation for squaring the
1752  // expression.
1753  for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1754  if (Factors[Idx].Power & 1)
1755  OuterProduct.push_back(Factors[Idx].Base);
1756  Factors[Idx].Power >>= 1;
1757  }
1758  if (Factors[0].Power) {
1759  Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1760  OuterProduct.push_back(SquareRoot);
1761  OuterProduct.push_back(SquareRoot);
1762  }
1763  if (OuterProduct.size() == 1)
1764  return OuterProduct.front();
1765 
1766  Value *V = buildMultiplyTree(Builder, OuterProduct);
1767  return V;
1768 }
1769 
1770 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1772  // We can only optimize the multiplies when there is a chain of more than
1773  // three, such that a balanced tree might require fewer total multiplies.
1774  if (Ops.size() < 4)
1775  return nullptr;
1776 
1777  // Try to turn linear trees of multiplies without other uses of the
1778  // intermediate stages into minimal multiply DAGs with perfect sub-expression
1779  // re-use.
1780  SmallVector<Factor, 4> Factors;
1781  if (!collectMultiplyFactors(Ops, Factors))
1782  return nullptr; // All distinct factors, so nothing left for us to do.
1783 
1784  IRBuilder<> Builder(I);
1785  // The reassociate transformation for FP operations is performed only
1786  // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1787  // to the newly generated operations.
1788  if (auto FPI = dyn_cast<FPMathOperator>(I))
1789  Builder.setFastMathFlags(FPI->getFastMathFlags());
1790 
1791  Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1792  if (Ops.empty())
1793  return V;
1794 
1795  ValueEntry NewEntry = ValueEntry(getRank(V), V);
1796  Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1797  return nullptr;
1798 }
1799 
1800 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1802  // Now that we have the linearized expression tree, try to optimize it.
1803  // Start by folding any constants that we found.
1804  Constant *Cst = nullptr;
1805  unsigned Opcode = I->getOpcode();
1806  while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1807  Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1808  Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1809  }
1810  // If there was nothing but constants then we are done.
1811  if (Ops.empty())
1812  return Cst;
1813 
1814  // Put the combined constant back at the end of the operand list, except if
1815  // there is no point. For example, an add of 0 gets dropped here, while a
1816  // multiplication by zero turns the whole expression into zero.
1817  if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1818  if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1819  return Cst;
1820  Ops.push_back(ValueEntry(0, Cst));
1821  }
1822 
1823  if (Ops.size() == 1) return Ops[0].Op;
1824 
1825  // Handle destructive annihilation due to identities between elements in the
1826  // argument list here.
1827  unsigned NumOps = Ops.size();
1828  switch (Opcode) {
1829  default: break;
1830  case Instruction::And:
1831  case Instruction::Or:
1832  if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1833  return Result;
1834  break;
1835 
1836  case Instruction::Xor:
1837  if (Value *Result = OptimizeXor(I, Ops))
1838  return Result;
1839  break;
1840 
1841  case Instruction::Add:
1842  case Instruction::FAdd:
1843  if (Value *Result = OptimizeAdd(I, Ops))
1844  return Result;
1845  break;
1846 
1847  case Instruction::Mul:
1848  case Instruction::FMul:
1849  if (Value *Result = OptimizeMul(I, Ops))
1850  return Result;
1851  break;
1852  }
1853 
1854  if (Ops.size() != NumOps)
1855  return OptimizeExpression(I, Ops);
1856  return nullptr;
1857 }
1858 
1859 // Remove dead instructions and if any operands are trivially dead add them to
1860 // Insts so they will be removed as well.
1861 void ReassociatePass::RecursivelyEraseDeadInsts(
1863  assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1864  SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1865  ValueRankMap.erase(I);
1866  Insts.remove(I);
1867  RedoInsts.remove(I);
1868  I->eraseFromParent();
1869  for (auto Op : Ops)
1870  if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1871  if (OpInst->use_empty())
1872  Insts.insert(OpInst);
1873 }
1874 
1875 /// Zap the given instruction, adding interesting operands to the work list.
1876 void ReassociatePass::EraseInst(Instruction *I) {
1877  assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1878  DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
1879 
1880  SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1881  // Erase the dead instruction.
1882  ValueRankMap.erase(I);
1883  RedoInsts.remove(I);
1884  I->eraseFromParent();
1885  // Optimize its operands.
1886  SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1887  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1888  if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1889  // If this is a node in an expression tree, climb to the expression root
1890  // and add that since that's where optimization actually happens.
1891  unsigned Opcode = Op->getOpcode();
1892  while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1893  Visited.insert(Op).second)
1894  Op = Op->user_back();
1895  RedoInsts.insert(Op);
1896  }
1897 
1898  MadeChange = true;
1899 }
1900 
1901 // Canonicalize expressions of the following form:
1902 // x + (-Constant * y) -> x - (Constant * y)
1903 // x - (-Constant * y) -> x + (Constant * y)
1904 Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1905  if (!I->hasOneUse() || I->getType()->isVectorTy())
1906  return nullptr;
1907 
1908  // Must be a fmul or fdiv instruction.
1909  unsigned Opcode = I->getOpcode();
1910  if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1911  return nullptr;
1912 
1913  auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1914  auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1915 
1916  // Both operands are constant, let it get constant folded away.
1917  if (C0 && C1)
1918  return nullptr;
1919 
1920  ConstantFP *CF = C0 ? C0 : C1;
1921 
1922  // Must have one constant operand.
1923  if (!CF)
1924  return nullptr;
1925 
1926  // Must be a negative ConstantFP.
1927  if (!CF->isNegative())
1928  return nullptr;
1929 
1930  // User must be a binary operator with one or more uses.
1931  Instruction *User = I->user_back();
1932  if (!isa<BinaryOperator>(User) || User->use_empty())
1933  return nullptr;
1934 
1935  unsigned UserOpcode = User->getOpcode();
1936  if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1937  return nullptr;
1938 
1939  // Subtraction is not commutative. Explicitly, the following transform is
1940  // not valid: (-Constant * y) - x -> x + (Constant * y)
1941  if (!User->isCommutative() && User->getOperand(1) != I)
1942  return nullptr;
1943 
1944  // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
1945  // resulting subtract will be broken up later. This can get us into an
1946  // infinite loop during reassociation.
1947  if (UserOpcode == Instruction::FAdd && ShouldBreakUpSubtract(User))
1948  return nullptr;
1949 
1950  // Change the sign of the constant.
1951  APFloat Val = CF->getValueAPF();
1952  Val.changeSign();
1953  I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1954 
1955  // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1956  // ((-Const*y) + x) -> (x + (-Const*y)).
1957  if (User->getOperand(0) == I && User->isCommutative())
1958  cast<BinaryOperator>(User)->swapOperands();
1959 
1960  Value *Op0 = User->getOperand(0);
1961  Value *Op1 = User->getOperand(1);
1962  BinaryOperator *NI;
1963  switch (UserOpcode) {
1964  default:
1965  llvm_unreachable("Unexpected Opcode!");
1966  case Instruction::FAdd:
1967  NI = BinaryOperator::CreateFSub(Op0, Op1);
1968  NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1969  break;
1970  case Instruction::FSub:
1971  NI = BinaryOperator::CreateFAdd(Op0, Op1);
1972  NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1973  break;
1974  }
1975 
1976  NI->insertBefore(User);
1977  NI->setName(User->getName());
1978  User->replaceAllUsesWith(NI);
1979  NI->setDebugLoc(I->getDebugLoc());
1980  RedoInsts.insert(I);
1981  MadeChange = true;
1982  return NI;
1983 }
1984 
1985 /// Inspect and optimize the given instruction. Note that erasing
1986 /// instructions is not allowed.
1987 void ReassociatePass::OptimizeInst(Instruction *I) {
1988  // Only consider operations that we understand.
1989  if (!isa<BinaryOperator>(I))
1990  return;
1991 
1992  if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1993  // If an operand of this shift is a reassociable multiply, or if the shift
1994  // is used by a reassociable multiply or add, turn into a multiply.
1995  if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1996  (I->hasOneUse() &&
1997  (isReassociableOp(I->user_back(), Instruction::Mul) ||
1999  Instruction *NI = ConvertShiftToMul(I);
2000  RedoInsts.insert(I);
2001  MadeChange = true;
2002  I = NI;
2003  }
2004 
2005  // Canonicalize negative constants out of expressions.
2006  if (Instruction *Res = canonicalizeNegConstExpr(I))
2007  I = Res;
2008 
2009  // Commute binary operators, to canonicalize the order of their operands.
2010  // This can potentially expose more CSE opportunities, and makes writing other
2011  // transformations simpler.
2012  if (I->isCommutative())
2013  canonicalizeOperands(I);
2014 
2015  // Don't optimize floating-point instructions unless they are 'fast'.
2016  if (I->getType()->isFPOrFPVectorTy() && !I->isFast())
2017  return;
2018 
2019  // Do not reassociate boolean (i1) expressions. We want to preserve the
2020  // original order of evaluation for short-circuited comparisons that
2021  // SimplifyCFG has folded to AND/OR expressions. If the expression
2022  // is not further optimized, it is likely to be transformed back to a
2023  // short-circuited form for code gen, and the source order may have been
2024  // optimized for the most likely conditions.
2025  if (I->getType()->isIntegerTy(1))
2026  return;
2027 
2028  // If this is a subtract instruction which is not already in negate form,
2029  // see if we can convert it to X+-Y.
2030  if (I->getOpcode() == Instruction::Sub) {
2031  if (ShouldBreakUpSubtract(I)) {
2032  Instruction *NI = BreakUpSubtract(I, RedoInsts);
2033  RedoInsts.insert(I);
2034  MadeChange = true;
2035  I = NI;
2036  } else if (BinaryOperator::isNeg(I)) {
2037  // Otherwise, this is a negation. See if the operand is a multiply tree
2038  // and if this is not an inner node of a multiply tree.
2039  if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2040  (!I->hasOneUse() ||
2041  !isReassociableOp(I->user_back(), Instruction::Mul))) {
2043  // If the negate was simplified, revisit the users to see if we can
2044  // reassociate further.
2045  for (User *U : NI->users()) {
2046  if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2047  RedoInsts.insert(Tmp);
2048  }
2049  RedoInsts.insert(I);
2050  MadeChange = true;
2051  I = NI;
2052  }
2053  }
2054  } else if (I->getOpcode() == Instruction::FSub) {
2055  if (ShouldBreakUpSubtract(I)) {
2056  Instruction *NI = BreakUpSubtract(I, RedoInsts);
2057  RedoInsts.insert(I);
2058  MadeChange = true;
2059  I = NI;
2060  } else if (BinaryOperator::isFNeg(I)) {
2061  // Otherwise, this is a negation. See if the operand is a multiply tree
2062  // and if this is not an inner node of a multiply tree.
2063  if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2064  (!I->hasOneUse() ||
2065  !isReassociableOp(I->user_back(), Instruction::FMul))) {
2066  // If the negate was simplified, revisit the users to see if we can
2067  // reassociate further.
2069  for (User *U : NI->users()) {
2070  if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2071  RedoInsts.insert(Tmp);
2072  }
2073  RedoInsts.insert(I);
2074  MadeChange = true;
2075  I = NI;
2076  }
2077  }
2078  }
2079 
2080  // If this instruction is an associative binary operator, process it.
2081  if (!I->isAssociative()) return;
2082  BinaryOperator *BO = cast<BinaryOperator>(I);
2083 
2084  // If this is an interior node of a reassociable tree, ignore it until we
2085  // get to the root of the tree, to avoid N^2 analysis.
2086  unsigned Opcode = BO->getOpcode();
2087  if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2088  // During the initial run we will get to the root of the tree.
2089  // But if we get here while we are redoing instructions, there is no
2090  // guarantee that the root will be visited. So Redo later
2091  if (BO->user_back() != BO &&
2092  BO->getParent() == BO->user_back()->getParent())
2093  RedoInsts.insert(BO->user_back());
2094  return;
2095  }
2096 
2097  // If this is an add tree that is used by a sub instruction, ignore it
2098  // until we process the subtract.
2099  if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2100  cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2101  return;
2102  if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2103  cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2104  return;
2105 
2106  ReassociateExpression(BO);
2107 }
2108 
2109 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2110  // First, walk the expression tree, linearizing the tree, collecting the
2111  // operand information.
2113  MadeChange |= LinearizeExprTree(I, Tree);
2115  Ops.reserve(Tree.size());
2116  for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2117  RepeatedValue E = Tree[i];
2118  Ops.append(E.second.getZExtValue(),
2119  ValueEntry(getRank(E.first), E.first));
2120  }
2121 
2122  DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2123 
2124  // Now that we have linearized the tree to a list and have gathered all of
2125  // the operands and their ranks, sort the operands by their rank. Use a
2126  // stable_sort so that values with equal ranks will have their relative
2127  // positions maintained (and so the compiler is deterministic). Note that
2128  // this sorts so that the highest ranking values end up at the beginning of
2129  // the vector.
2130  std::stable_sort(Ops.begin(), Ops.end());
2131 
2132  // Now that we have the expression tree in a convenient
2133  // sorted form, optimize it globally if possible.
2134  if (Value *V = OptimizeExpression(I, Ops)) {
2135  if (V == I)
2136  // Self-referential expression in unreachable code.
2137  return;
2138  // This expression tree simplified to something that isn't a tree,
2139  // eliminate it.
2140  DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2141  I->replaceAllUsesWith(V);
2142  if (Instruction *VI = dyn_cast<Instruction>(V))
2143  if (I->getDebugLoc())
2144  VI->setDebugLoc(I->getDebugLoc());
2145  RedoInsts.insert(I);
2146  ++NumAnnihil;
2147  return;
2148  }
2149 
2150  // We want to sink immediates as deeply as possible except in the case where
2151  // this is a multiply tree used only by an add, and the immediate is a -1.
2152  // In this case we reassociate to put the negation on the outside so that we
2153  // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2154  if (I->hasOneUse()) {
2155  if (I->getOpcode() == Instruction::Mul &&
2156  cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2157  isa<ConstantInt>(Ops.back().Op) &&
2158  cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2159  ValueEntry Tmp = Ops.pop_back_val();
2160  Ops.insert(Ops.begin(), Tmp);
2161  } else if (I->getOpcode() == Instruction::FMul &&
2162  cast<Instruction>(I->user_back())->getOpcode() ==
2163  Instruction::FAdd &&
2164  isa<ConstantFP>(Ops.back().Op) &&
2165  cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2166  ValueEntry Tmp = Ops.pop_back_val();
2167  Ops.insert(Ops.begin(), Tmp);
2168  }
2169  }
2170 
2171  DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2172 
2173  if (Ops.size() == 1) {
2174  if (Ops[0].Op == I)
2175  // Self-referential expression in unreachable code.
2176  return;
2177 
2178  // This expression tree simplified to something that isn't a tree,
2179  // eliminate it.
2180  I->replaceAllUsesWith(Ops[0].Op);
2181  if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2182  OI->setDebugLoc(I->getDebugLoc());
2183  RedoInsts.insert(I);
2184  return;
2185  }
2186 
2187  // Now that we ordered and optimized the expressions, splat them back into
2188  // the expression tree, removing any unneeded nodes.
2189  RewriteExprTree(I, Ops);
2190 }
2191 
2193  // Get the functions basic blocks in Reverse Post Order. This order is used by
2194  // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2195  // blocks (it has been seen that the analysis in this pass could hang when
2196  // analysing dead basic blocks).
2198 
2199  // Calculate the rank map for F.
2200  BuildRankMap(F, RPOT);
2201 
2202  MadeChange = false;
2203  // Traverse the same blocks that was analysed by BuildRankMap.
2204  for (BasicBlock *BI : RPOT) {
2205  assert(RankMap.count(&*BI) && "BB should be ranked.");
2206  // Optimize every instruction in the basic block.
2207  for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2208  if (isInstructionTriviallyDead(&*II)) {
2209  EraseInst(&*II++);
2210  } else {
2211  OptimizeInst(&*II);
2212  assert(II->getParent() == &*BI && "Moved to a different block!");
2213  ++II;
2214  }
2215 
2216  // Make a copy of all the instructions to be redone so we can remove dead
2217  // instructions.
2218  SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts);
2219  // Iterate over all instructions to be reevaluated and remove trivially dead
2220  // instructions. If any operand of the trivially dead instruction becomes
2221  // dead mark it for deletion as well. Continue this process until all
2222  // trivially dead instructions have been removed.
2223  while (!ToRedo.empty()) {
2224  Instruction *I = ToRedo.pop_back_val();
2225  if (isInstructionTriviallyDead(I)) {
2226  RecursivelyEraseDeadInsts(I, ToRedo);
2227  MadeChange = true;
2228  }
2229  }
2230 
2231  // Now that we have removed dead instructions, we can reoptimize the
2232  // remaining instructions.
2233  while (!RedoInsts.empty()) {
2234  Instruction *I = RedoInsts.pop_back_val();
2236  EraseInst(I);
2237  else
2238  OptimizeInst(I);
2239  }
2240  }
2241 
2242  // We are done with the rank map.
2243  RankMap.clear();
2244  ValueRankMap.clear();
2245 
2246  if (MadeChange) {
2247  PreservedAnalyses PA;
2248  PA.preserveSet<CFGAnalyses>();
2249  PA.preserve<GlobalsAA>();
2250  return PA;
2251  }
2252 
2253  return PreservedAnalyses::all();
2254 }
2255 
2256 namespace {
2257 
2258  class ReassociateLegacyPass : public FunctionPass {
2259  ReassociatePass Impl;
2260 
2261  public:
2262  static char ID; // Pass identification, replacement for typeid
2263 
2264  ReassociateLegacyPass() : FunctionPass(ID) {
2266  }
2267 
2268  bool runOnFunction(Function &F) override {
2269  if (skipFunction(F))
2270  return false;
2271 
2272  FunctionAnalysisManager DummyFAM;
2273  auto PA = Impl.run(F, DummyFAM);
2274  return !PA.areAllPreserved();
2275  }
2276 
2277  void getAnalysisUsage(AnalysisUsage &AU) const override {
2278  AU.setPreservesCFG();
2280  }
2281  };
2282 
2283 } // end anonymous namespace
2284 
2285 char ReassociateLegacyPass::ID = 0;
2286 
2287 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2288  "Reassociate expressions", false, false)
2289 
2290 // Public interface to the Reassociate pass
2292  return new ReassociateLegacyPass();
2293 }
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)
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:469
Utility class representing a non-constant Xor-operand.
Definition: Reassociate.cpp:93
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:59
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:728
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:95
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...
std::pair< Value *, APInt > RepeatedValue
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:71
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:3641
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:483
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:668
FunctionPass * createReassociatePass()
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:286
This file implements a class to represent arbitrary precision integral constant values and operations...
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
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
auto lower_bound(R &&Range, ForwardIt I) -> decltype(adl_begin(Range))
Provide wrappers to std::lower_bound which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:904
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:430
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:292
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
static bool runOnFunction(Function &F, bool PostInlining)
#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:371
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
This file declares a class to represent arbitrary precision floating point values and provide a varie...
bool isFast() const
Determine whether all fast-math-flags are set.
Analysis pass providing a never-invalidated alias analysis result.
Value * getSymbolicPart() const
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:382
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:937
const APInt & getConstPart() const
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:99
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:418
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
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)
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:451
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:401
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
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:220
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:73
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:377
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:324
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:477
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
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:414
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
This header defines various interfaces for pass management in LLVM.
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:951
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:328
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 &)