1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass reassociates commutative expressions in an order that is designed
11 // to promote better constant propagation, GCSE, LICM, PRE, etc.
13 // For example: 4 + (x + 5) -> x + (4 + 5)
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.
21 //===----------------------------------------------------------------------===//
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/PostOrderIterator.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/IR/CFG.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/ValueHandle.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Transforms/Utils/Local.h"
44 #define DEBUG_TYPE "reassociate"
46 STATISTIC(NumChanged
, "Number of insts reassociated");
47 STATISTIC(NumAnnihil
, "Number of expr tree annihilated");
48 STATISTIC(NumFactor
, "Number of multiplies factored");
54 ValueEntry(unsigned R
, Value
*O
) : Rank(R
), Op(O
) {}
56 inline bool operator<(const ValueEntry
&LHS
, const ValueEntry
&RHS
) {
57 return LHS
.Rank
> RHS
.Rank
; // Sort so that highest rank goes to start.
62 /// PrintOps - Print out the expression identified in the Ops list.
64 static void PrintOps(Instruction
*I
, const SmallVectorImpl
<ValueEntry
> &Ops
) {
65 Module
*M
= I
->getParent()->getParent()->getParent();
66 dbgs() << Instruction::getOpcodeName(I
->getOpcode()) << " "
67 << *Ops
[0].Op
->getType() << '\t';
68 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
) {
70 Ops
[i
].Op
->printAsOperand(dbgs(), false, M
);
71 dbgs() << ", #" << Ops
[i
].Rank
<< "] ";
77 /// \brief Utility class representing a base and exponent pair which form one
78 /// factor of some product.
83 Factor(Value
*Base
, unsigned Power
) : Base(Base
), Power(Power
) {}
85 /// \brief Sort factors by their Base.
87 bool operator()(const Factor
&LHS
, const Factor
&RHS
) {
88 return LHS
.Base
< RHS
.Base
;
92 /// \brief Compare factors for equal bases.
94 bool operator()(const Factor
&LHS
, const Factor
&RHS
) {
95 return LHS
.Base
== RHS
.Base
;
99 /// \brief Sort factors in descending order by their power.
100 struct PowerDescendingSorter
{
101 bool operator()(const Factor
&LHS
, const Factor
&RHS
) {
102 return LHS
.Power
> RHS
.Power
;
106 /// \brief Compare factors for equal powers.
108 bool operator()(const Factor
&LHS
, const Factor
&RHS
) {
109 return LHS
.Power
== RHS
.Power
;
114 /// Utility class representing a non-constant Xor-operand. We classify
115 /// non-constant Xor-Operands into two categories:
116 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
118 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
120 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
121 /// operand as "E | 0"
126 bool isInvalid() const { return SymbolicPart
== nullptr; }
127 bool isOrExpr() const { return isOr
; }
128 Value
*getValue() const { return OrigVal
; }
129 Value
*getSymbolicPart() const { return SymbolicPart
; }
130 unsigned getSymbolicRank() const { return SymbolicRank
; }
131 const APInt
&getConstPart() const { return ConstPart
; }
133 void Invalidate() { SymbolicPart
= OrigVal
= nullptr; }
134 void setSymbolicRank(unsigned R
) { SymbolicRank
= R
; }
136 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank.
137 // The purpose is twofold:
138 // 1) Cluster together the operands sharing the same symbolic-value.
139 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
140 // could potentially shorten crital path, and expose more loop-invariants.
141 // Note that values' rank are basically defined in RPO order (FIXME).
142 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
143 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
144 // "z" in the order of X-Y-Z is better than any other orders.
145 struct PtrSortFunctor
{
146 bool operator()(XorOpnd
* const &LHS
, XorOpnd
* const &RHS
) {
147 return LHS
->getSymbolicRank() < RHS
->getSymbolicRank();
154 unsigned SymbolicRank
;
160 class Reassociate
: public FunctionPass
{
161 DenseMap
<BasicBlock
*, unsigned> RankMap
;
162 DenseMap
<AssertingVH
<Value
>, unsigned> ValueRankMap
;
163 SetVector
<AssertingVH
<Instruction
> > RedoInsts
;
166 static char ID
; // Pass identification, replacement for typeid
167 Reassociate() : FunctionPass(ID
) {
168 initializeReassociatePass(*PassRegistry::getPassRegistry());
171 bool runOnFunction(Function
&F
) override
;
173 void getAnalysisUsage(AnalysisUsage
&AU
) const override
{
174 AU
.setPreservesCFG();
177 void BuildRankMap(Function
&F
);
178 unsigned getRank(Value
*V
);
179 void canonicalizeOperands(Instruction
*I
);
180 void ReassociateExpression(BinaryOperator
*I
);
181 void RewriteExprTree(BinaryOperator
*I
, SmallVectorImpl
<ValueEntry
> &Ops
);
182 Value
*OptimizeExpression(BinaryOperator
*I
,
183 SmallVectorImpl
<ValueEntry
> &Ops
);
184 Value
*OptimizeAdd(Instruction
*I
, SmallVectorImpl
<ValueEntry
> &Ops
);
185 Value
*OptimizeXor(Instruction
*I
, SmallVectorImpl
<ValueEntry
> &Ops
);
186 bool CombineXorOpnd(Instruction
*I
, XorOpnd
*Opnd1
, APInt
&ConstOpnd
,
188 bool CombineXorOpnd(Instruction
*I
, XorOpnd
*Opnd1
, XorOpnd
*Opnd2
,
189 APInt
&ConstOpnd
, Value
*&Res
);
190 bool collectMultiplyFactors(SmallVectorImpl
<ValueEntry
> &Ops
,
191 SmallVectorImpl
<Factor
> &Factors
);
192 Value
*buildMinimalMultiplyDAG(IRBuilder
<> &Builder
,
193 SmallVectorImpl
<Factor
> &Factors
);
194 Value
*OptimizeMul(BinaryOperator
*I
, SmallVectorImpl
<ValueEntry
> &Ops
);
195 Value
*RemoveFactorFromExpression(Value
*V
, Value
*Factor
);
196 void EraseInst(Instruction
*I
);
197 void OptimizeInst(Instruction
*I
);
198 Instruction
*canonicalizeNegConstExpr(Instruction
*I
);
202 XorOpnd::XorOpnd(Value
*V
) {
203 assert(!isa
<ConstantInt
>(V
) && "No ConstantInt");
205 Instruction
*I
= dyn_cast
<Instruction
>(V
);
208 if (I
&& (I
->getOpcode() == Instruction::Or
||
209 I
->getOpcode() == Instruction::And
)) {
210 Value
*V0
= I
->getOperand(0);
211 Value
*V1
= I
->getOperand(1);
212 if (isa
<ConstantInt
>(V0
))
215 if (ConstantInt
*C
= dyn_cast
<ConstantInt
>(V1
)) {
216 ConstPart
= C
->getValue();
218 isOr
= (I
->getOpcode() == Instruction::Or
);
223 // view the operand as "V | 0"
225 ConstPart
= APInt::getNullValue(V
->getType()->getIntegerBitWidth());
229 char Reassociate::ID
= 0;
230 INITIALIZE_PASS(Reassociate
, "reassociate",
231 "Reassociate expressions", false, false)
233 // Public interface to the Reassociate pass
234 FunctionPass
*llvm::createReassociatePass() { return new Reassociate(); }
236 /// isReassociableOp - Return true if V is an instruction of the specified
237 /// opcode and if it only has one use.
238 static BinaryOperator
*isReassociableOp(Value
*V
, unsigned Opcode
) {
239 if (V
->hasOneUse() && isa
<Instruction
>(V
) &&
240 cast
<Instruction
>(V
)->getOpcode() == Opcode
&&
241 (!isa
<FPMathOperator
>(V
) ||
242 cast
<Instruction
>(V
)->hasUnsafeAlgebra()))
243 return cast
<BinaryOperator
>(V
);
247 static BinaryOperator
*isReassociableOp(Value
*V
, unsigned Opcode1
,
249 if (V
->hasOneUse() && isa
<Instruction
>(V
) &&
250 (cast
<Instruction
>(V
)->getOpcode() == Opcode1
||
251 cast
<Instruction
>(V
)->getOpcode() == Opcode2
) &&
252 (!isa
<FPMathOperator
>(V
) ||
253 cast
<Instruction
>(V
)->hasUnsafeAlgebra()))
254 return cast
<BinaryOperator
>(V
);
258 static bool isUnmovableInstruction(Instruction
*I
) {
259 switch (I
->getOpcode()) {
260 case Instruction::PHI
:
261 case Instruction::LandingPad
:
262 case Instruction::Alloca
:
263 case Instruction::Load
:
264 case Instruction::Invoke
:
265 case Instruction::UDiv
:
266 case Instruction::SDiv
:
267 case Instruction::FDiv
:
268 case Instruction::URem
:
269 case Instruction::SRem
:
270 case Instruction::FRem
:
272 case Instruction::Call
:
273 return !isa
<DbgInfoIntrinsic
>(I
);
279 void Reassociate::BuildRankMap(Function
&F
) {
282 // Assign distinct ranks to function arguments.
283 for (Function::arg_iterator I
= F
.arg_begin(), E
= F
.arg_end(); I
!= E
; ++I
) {
284 ValueRankMap
[&*I
] = ++i
;
285 DEBUG(dbgs() << "Calculated Rank[" << I
->getName() << "] = " << i
<< "\n");
288 ReversePostOrderTraversal
<Function
*> RPOT(&F
);
289 for (ReversePostOrderTraversal
<Function
*>::rpo_iterator I
= RPOT
.begin(),
290 E
= RPOT
.end(); I
!= E
; ++I
) {
292 unsigned BBRank
= RankMap
[BB
] = ++i
<< 16;
294 // Walk the basic block, adding precomputed ranks for any instructions that
295 // we cannot move. This ensures that the ranks for these instructions are
296 // all different in the block.
297 for (BasicBlock::iterator I
= BB
->begin(), E
= BB
->end(); I
!= E
; ++I
)
298 if (isUnmovableInstruction(I
))
299 ValueRankMap
[&*I
] = ++BBRank
;
303 unsigned Reassociate::getRank(Value
*V
) {
304 Instruction
*I
= dyn_cast
<Instruction
>(V
);
306 if (isa
<Argument
>(V
)) return ValueRankMap
[V
]; // Function argument.
307 return 0; // Otherwise it's a global or constant, rank 0.
310 if (unsigned Rank
= ValueRankMap
[I
])
311 return Rank
; // Rank already known?
313 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
314 // we can reassociate expressions for code motion! Since we do not recurse
315 // for PHI nodes, we cannot have infinite recursion here, because there
316 // cannot be loops in the value graph that do not go through PHI nodes.
317 unsigned Rank
= 0, MaxRank
= RankMap
[I
->getParent()];
318 for (unsigned i
= 0, e
= I
->getNumOperands();
319 i
!= e
&& Rank
!= MaxRank
; ++i
)
320 Rank
= std::max(Rank
, getRank(I
->getOperand(i
)));
322 // If this is a not or neg instruction, do not count it for rank. This
323 // assures us that X and ~X will have the same rank.
324 Type
*Ty
= V
->getType();
325 if ((!Ty
->isIntegerTy() && !Ty
->isFloatingPointTy()) ||
326 (!BinaryOperator::isNot(I
) && !BinaryOperator::isNeg(I
) &&
327 !BinaryOperator::isFNeg(I
)))
330 DEBUG(dbgs() << "Calculated Rank[" << V
->getName() << "] = " << Rank
<< "\n");
332 return ValueRankMap
[I
] = Rank
;
335 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
336 void Reassociate::canonicalizeOperands(Instruction
*I
) {
337 assert(isa
<BinaryOperator
>(I
) && "Expected binary operator.");
338 assert(I
->isCommutative() && "Expected commutative operator.");
340 Value
*LHS
= I
->getOperand(0);
341 Value
*RHS
= I
->getOperand(1);
342 unsigned LHSRank
= getRank(LHS
);
343 unsigned RHSRank
= getRank(RHS
);
345 if (isa
<Constant
>(RHS
))
348 if (isa
<Constant
>(LHS
) || RHSRank
< LHSRank
)
349 cast
<BinaryOperator
>(I
)->swapOperands();
352 static BinaryOperator
*CreateAdd(Value
*S1
, Value
*S2
, const Twine
&Name
,
353 Instruction
*InsertBefore
, Value
*FlagsOp
) {
354 if (S1
->getType()->isIntegerTy())
355 return BinaryOperator::CreateAdd(S1
, S2
, Name
, InsertBefore
);
357 BinaryOperator
*Res
=
358 BinaryOperator::CreateFAdd(S1
, S2
, Name
, InsertBefore
);
359 Res
->setFastMathFlags(cast
<FPMathOperator
>(FlagsOp
)->getFastMathFlags());
364 static BinaryOperator
*CreateMul(Value
*S1
, Value
*S2
, const Twine
&Name
,
365 Instruction
*InsertBefore
, Value
*FlagsOp
) {
366 if (S1
->getType()->isIntegerTy())
367 return BinaryOperator::CreateMul(S1
, S2
, Name
, InsertBefore
);
369 BinaryOperator
*Res
=
370 BinaryOperator::CreateFMul(S1
, S2
, Name
, InsertBefore
);
371 Res
->setFastMathFlags(cast
<FPMathOperator
>(FlagsOp
)->getFastMathFlags());
376 static BinaryOperator
*CreateNeg(Value
*S1
, const Twine
&Name
,
377 Instruction
*InsertBefore
, Value
*FlagsOp
) {
378 if (S1
->getType()->isIntegerTy())
379 return BinaryOperator::CreateNeg(S1
, Name
, InsertBefore
);
381 BinaryOperator
*Res
= BinaryOperator::CreateFNeg(S1
, Name
, InsertBefore
);
382 Res
->setFastMathFlags(cast
<FPMathOperator
>(FlagsOp
)->getFastMathFlags());
387 /// LowerNegateToMultiply - Replace 0-X with X*-1.
389 static BinaryOperator
*LowerNegateToMultiply(Instruction
*Neg
) {
390 Type
*Ty
= Neg
->getType();
391 Constant
*NegOne
= Ty
->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty
)
392 : ConstantFP::get(Ty
, -1.0);
394 BinaryOperator
*Res
= CreateMul(Neg
->getOperand(1), NegOne
, "", Neg
, Neg
);
395 Neg
->setOperand(1, Constant::getNullValue(Ty
)); // Drop use of op.
397 Neg
->replaceAllUsesWith(Res
);
398 Res
->setDebugLoc(Neg
->getDebugLoc());
402 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
403 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
404 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
405 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
406 /// even x in Bitwidth-bit arithmetic.
407 static unsigned CarmichaelShift(unsigned Bitwidth
) {
413 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
414 /// reducing the combined weight using any special properties of the operation.
415 /// The existing weight LHS represents the computation X op X op ... op X where
416 /// X occurs LHS times. The combined weight represents X op X op ... op X with
417 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
418 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
419 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
420 static void IncorporateWeight(APInt
&LHS
, const APInt
&RHS
, unsigned Opcode
) {
421 // If we were working with infinite precision arithmetic then the combined
422 // weight would be LHS + RHS. But we are using finite precision arithmetic,
423 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
424 // for nilpotent operations and addition, but not for idempotent operations
425 // and multiplication), so it is important to correctly reduce the combined
426 // weight back into range if wrapping would be wrong.
428 // If RHS is zero then the weight didn't change.
429 if (RHS
.isMinValue())
431 // If LHS is zero then the combined weight is RHS.
432 if (LHS
.isMinValue()) {
436 // From this point on we know that neither LHS nor RHS is zero.
438 if (Instruction::isIdempotent(Opcode
)) {
439 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
440 // weight of 1. Keeping weights at zero or one also means that wrapping is
442 assert(LHS
== 1 && RHS
== 1 && "Weights not reduced!");
443 return; // Return a weight of 1.
445 if (Instruction::isNilpotent(Opcode
)) {
446 // Nilpotent means X op X === 0, so reduce weights modulo 2.
447 assert(LHS
== 1 && RHS
== 1 && "Weights not reduced!");
448 LHS
= 0; // 1 + 1 === 0 modulo 2.
451 if (Opcode
== Instruction::Add
|| Opcode
== Instruction::FAdd
) {
452 // TODO: Reduce the weight by exploiting nsw/nuw?
457 assert((Opcode
== Instruction::Mul
|| Opcode
== Instruction::FMul
) &&
458 "Unknown associative operation!");
459 unsigned Bitwidth
= LHS
.getBitWidth();
460 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
461 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
462 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
463 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
464 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
465 // which by a happy accident means that they can always be represented using
467 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
468 // the Carmichael number).
470 /// CM - The value of Carmichael's lambda function.
471 APInt CM
= APInt::getOneBitSet(Bitwidth
, CarmichaelShift(Bitwidth
));
472 // Any weight W >= Threshold can be replaced with W - CM.
473 APInt Threshold
= CM
+ Bitwidth
;
474 assert(LHS
.ult(Threshold
) && RHS
.ult(Threshold
) && "Weights not reduced!");
475 // For Bitwidth 4 or more the following sum does not overflow.
477 while (LHS
.uge(Threshold
))
480 // To avoid problems with overflow do everything the same as above but using
482 unsigned CM
= 1U << CarmichaelShift(Bitwidth
);
483 unsigned Threshold
= CM
+ Bitwidth
;
484 assert(LHS
.getZExtValue() < Threshold
&& RHS
.getZExtValue() < Threshold
&&
485 "Weights not reduced!");
486 unsigned Total
= LHS
.getZExtValue() + RHS
.getZExtValue();
487 while (Total
>= Threshold
)
493 typedef std::pair
<Value
*, APInt
> RepeatedValue
;
495 /// LinearizeExprTree - Given an associative binary expression, return the leaf
496 /// nodes in Ops along with their weights (how many times the leaf occurs). The
497 /// original expression is the same as
498 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
500 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
504 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
506 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
508 /// This routine may modify the function, in which case it returns 'true'. The
509 /// changes it makes may well be destructive, changing the value computed by 'I'
510 /// to something completely different. Thus if the routine returns 'true' then
511 /// you MUST either replace I with a new expression computed from the Ops array,
512 /// or use RewriteExprTree to put the values back in.
514 /// A leaf node is either not a binary operation of the same kind as the root
515 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
516 /// opcode), or is the same kind of binary operator but has a use which either
517 /// does not belong to the expression, or does belong to the expression but is
518 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
519 /// of the expression, while for non-leaf nodes (except for the root 'I') every
520 /// use is a non-leaf node of the expression.
523 /// expression graph node names
533 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
534 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
536 /// The expression is maximal: if some instruction is a binary operator of the
537 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
538 /// then the instruction also belongs to the expression, is not a leaf node of
539 /// it, and its operands also belong to the expression (but may be leaf nodes).
541 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
542 /// order to ensure that every non-root node in the expression has *exactly one*
543 /// use by a non-leaf node of the expression. This destruction means that the
544 /// caller MUST either replace 'I' with a new expression or use something like
545 /// RewriteExprTree to put the values back in if the routine indicates that it
546 /// made a change by returning 'true'.
548 /// In the above example either the right operand of A or the left operand of B
549 /// will be replaced by undef. If it is B's operand then this gives:
553 /// + + | A, B - operand of B replaced with undef
559 /// Note that such undef operands can only be reached by passing through 'I'.
560 /// For example, if you visit operands recursively starting from a leaf node
561 /// then you will never see such an undef operand unless you get back to 'I',
562 /// which requires passing through a phi node.
564 /// Note that this routine may also mutate binary operators of the wrong type
565 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
566 /// of the expression) if it can turn them into binary operators of the right
567 /// type and thus make the expression bigger.
569 static bool LinearizeExprTree(BinaryOperator
*I
,
570 SmallVectorImpl
<RepeatedValue
> &Ops
) {
571 DEBUG(dbgs() << "LINEARIZE: " << *I
<< '\n');
572 unsigned Bitwidth
= I
->getType()->getScalarType()->getPrimitiveSizeInBits();
573 unsigned Opcode
= I
->getOpcode();
574 assert(I
->isAssociative() && I
->isCommutative() &&
575 "Expected an associative and commutative operation!");
577 // Visit all operands of the expression, keeping track of their weight (the
578 // number of paths from the expression root to the operand, or if you like
579 // the number of times that operand occurs in the linearized expression).
580 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
581 // while A has weight two.
583 // Worklist of non-leaf nodes (their operands are in the expression too) along
584 // with their weights, representing a certain number of paths to the operator.
585 // If an operator occurs in the worklist multiple times then we found multiple
586 // ways to get to it.
587 SmallVector
<std::pair
<BinaryOperator
*, APInt
>, 8> Worklist
; // (Op, Weight)
588 Worklist
.push_back(std::make_pair(I
, APInt(Bitwidth
, 1)));
589 bool Changed
= false;
591 // Leaves of the expression are values that either aren't the right kind of
592 // operation (eg: a constant, or a multiply in an add tree), or are, but have
593 // some uses that are not inside the expression. For example, in I = X + X,
594 // X = A + B, the value X has two uses (by I) that are in the expression. If
595 // X has any other uses, for example in a return instruction, then we consider
596 // X to be a leaf, and won't analyze it further. When we first visit a value,
597 // if it has more than one use then at first we conservatively consider it to
598 // be a leaf. Later, as the expression is explored, we may discover some more
599 // uses of the value from inside the expression. If all uses turn out to be
600 // from within the expression (and the value is a binary operator of the right
601 // kind) then the value is no longer considered to be a leaf, and its operands
604 // Leaves - Keeps track of the set of putative leaves as well as the number of
605 // paths to each leaf seen so far.
606 typedef DenseMap
<Value
*, APInt
> LeafMap
;
607 LeafMap Leaves
; // Leaf -> Total weight so far.
608 SmallVector
<Value
*, 8> LeafOrder
; // Ensure deterministic leaf output order.
611 SmallPtrSet
<Value
*, 8> Visited
; // For sanity checking the iteration scheme.
613 while (!Worklist
.empty()) {
614 std::pair
<BinaryOperator
*, APInt
> P
= Worklist
.pop_back_val();
615 I
= P
.first
; // We examine the operands of this binary operator.
617 for (unsigned OpIdx
= 0; OpIdx
< 2; ++OpIdx
) { // Visit operands.
618 Value
*Op
= I
->getOperand(OpIdx
);
619 APInt Weight
= P
.second
; // Number of paths to this operand.
620 DEBUG(dbgs() << "OPERAND: " << *Op
<< " (" << Weight
<< ")\n");
621 assert(!Op
->use_empty() && "No uses, so how did we get to it?!");
623 // If this is a binary operation of the right kind with only one use then
624 // add its operands to the expression.
625 if (BinaryOperator
*BO
= isReassociableOp(Op
, Opcode
)) {
626 assert(Visited
.insert(Op
).second
&& "Not first visit!");
627 DEBUG(dbgs() << "DIRECT ADD: " << *Op
<< " (" << Weight
<< ")\n");
628 Worklist
.push_back(std::make_pair(BO
, Weight
));
632 // Appears to be a leaf. Is the operand already in the set of leaves?
633 LeafMap::iterator It
= Leaves
.find(Op
);
634 if (It
== Leaves
.end()) {
635 // Not in the leaf map. Must be the first time we saw this operand.
636 assert(Visited
.insert(Op
).second
&& "Not first visit!");
637 if (!Op
->hasOneUse()) {
638 // This value has uses not accounted for by the expression, so it is
639 // not safe to modify. Mark it as being a leaf.
640 DEBUG(dbgs() << "ADD USES LEAF: " << *Op
<< " (" << Weight
<< ")\n");
641 LeafOrder
.push_back(Op
);
645 // No uses outside the expression, try morphing it.
646 } else if (It
!= Leaves
.end()) {
647 // Already in the leaf map.
648 assert(Visited
.count(Op
) && "In leaf map but not visited!");
650 // Update the number of paths to the leaf.
651 IncorporateWeight(It
->second
, Weight
, Opcode
);
653 #if 0 // TODO: Re-enable once PR13021 is fixed.
654 // The leaf already has one use from inside the expression. As we want
655 // exactly one such use, drop this new use of the leaf.
656 assert(!Op
->hasOneUse() && "Only one use, but we got here twice!");
657 I
->setOperand(OpIdx
, UndefValue::get(I
->getType()));
660 // If the leaf is a binary operation of the right kind and we now see
661 // that its multiple original uses were in fact all by nodes belonging
662 // to the expression, then no longer consider it to be a leaf and add
663 // its operands to the expression.
664 if (BinaryOperator
*BO
= isReassociableOp(Op
, Opcode
)) {
665 DEBUG(dbgs() << "UNLEAF: " << *Op
<< " (" << It
->second
<< ")\n");
666 Worklist
.push_back(std::make_pair(BO
, It
->second
));
672 // If we still have uses that are not accounted for by the expression
673 // then it is not safe to modify the value.
674 if (!Op
->hasOneUse())
677 // No uses outside the expression, try morphing it.
679 Leaves
.erase(It
); // Since the value may be morphed below.
682 // At this point we have a value which, first of all, is not a binary
683 // expression of the right kind, and secondly, is only used inside the
684 // expression. This means that it can safely be modified. See if we
685 // can usefully morph it into an expression of the right kind.
686 assert((!isa
<Instruction
>(Op
) ||
687 cast
<Instruction
>(Op
)->getOpcode() != Opcode
688 || (isa
<FPMathOperator
>(Op
) &&
689 !cast
<Instruction
>(Op
)->hasUnsafeAlgebra())) &&
690 "Should have been handled above!");
691 assert(Op
->hasOneUse() && "Has uses outside the expression tree!");
693 // If this is a multiply expression, turn any internal negations into
694 // multiplies by -1 so they can be reassociated.
695 if (BinaryOperator
*BO
= dyn_cast
<BinaryOperator
>(Op
))
696 if ((Opcode
== Instruction::Mul
&& BinaryOperator::isNeg(BO
)) ||
697 (Opcode
== Instruction::FMul
&& BinaryOperator::isFNeg(BO
))) {
698 DEBUG(dbgs() << "MORPH LEAF: " << *Op
<< " (" << Weight
<< ") TO ");
699 BO
= LowerNegateToMultiply(BO
);
700 DEBUG(dbgs() << *BO
<< '\n');
701 Worklist
.push_back(std::make_pair(BO
, Weight
));
706 // Failed to morph into an expression of the right type. This really is
708 DEBUG(dbgs() << "ADD LEAF: " << *Op
<< " (" << Weight
<< ")\n");
709 assert(!isReassociableOp(Op
, Opcode
) && "Value was morphed?");
710 LeafOrder
.push_back(Op
);
715 // The leaves, repeated according to their weights, represent the linearized
716 // form of the expression.
717 for (unsigned i
= 0, e
= LeafOrder
.size(); i
!= e
; ++i
) {
718 Value
*V
= LeafOrder
[i
];
719 LeafMap::iterator It
= Leaves
.find(V
);
720 if (It
== Leaves
.end())
721 // Node initially thought to be a leaf wasn't.
723 assert(!isReassociableOp(V
, Opcode
) && "Shouldn't be a leaf!");
724 APInt Weight
= It
->second
;
725 if (Weight
.isMinValue())
726 // Leaf already output or weight reduction eliminated it.
728 // Ensure the leaf is only output once.
730 Ops
.push_back(std::make_pair(V
, Weight
));
733 // For nilpotent operations or addition there may be no operands, for example
734 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
735 // in both cases the weight reduces to 0 causing the value to be skipped.
737 Constant
*Identity
= ConstantExpr::getBinOpIdentity(Opcode
, I
->getType());
738 assert(Identity
&& "Associative operation without identity!");
739 Ops
.push_back(std::make_pair(Identity
, APInt(Bitwidth
, 1)));
745 // RewriteExprTree - Now that the operands for this expression tree are
746 // linearized and optimized, emit them in-order.
747 void Reassociate::RewriteExprTree(BinaryOperator
*I
,
748 SmallVectorImpl
<ValueEntry
> &Ops
) {
749 assert(Ops
.size() > 1 && "Single values should be used directly!");
751 // Since our optimizations should never increase the number of operations, the
752 // new expression can usually be written reusing the existing binary operators
753 // from the original expression tree, without creating any new instructions,
754 // though the rewritten expression may have a completely different topology.
755 // We take care to not change anything if the new expression will be the same
756 // as the original. If more than trivial changes (like commuting operands)
757 // were made then we are obliged to clear out any optional subclass data like
760 /// NodesToRewrite - Nodes from the original expression available for writing
761 /// the new expression into.
762 SmallVector
<BinaryOperator
*, 8> NodesToRewrite
;
763 unsigned Opcode
= I
->getOpcode();
764 BinaryOperator
*Op
= I
;
766 /// NotRewritable - The operands being written will be the leaves of the new
767 /// expression and must not be used as inner nodes (via NodesToRewrite) by
768 /// mistake. Inner nodes are always reassociable, and usually leaves are not
769 /// (if they were they would have been incorporated into the expression and so
770 /// would not be leaves), so most of the time there is no danger of this. But
771 /// in rare cases a leaf may become reassociable if an optimization kills uses
772 /// of it, or it may momentarily become reassociable during rewriting (below)
773 /// due it being removed as an operand of one of its uses. Ensure that misuse
774 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
775 /// leaves and refusing to reuse any of them as inner nodes.
776 SmallPtrSet
<Value
*, 8> NotRewritable
;
777 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
)
778 NotRewritable
.insert(Ops
[i
].Op
);
780 // ExpressionChanged - Non-null if the rewritten expression differs from the
781 // original in some non-trivial way, requiring the clearing of optional flags.
782 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
783 BinaryOperator
*ExpressionChanged
= nullptr;
784 for (unsigned i
= 0; ; ++i
) {
785 // The last operation (which comes earliest in the IR) is special as both
786 // operands will come from Ops, rather than just one with the other being
788 if (i
+2 == Ops
.size()) {
789 Value
*NewLHS
= Ops
[i
].Op
;
790 Value
*NewRHS
= Ops
[i
+1].Op
;
791 Value
*OldLHS
= Op
->getOperand(0);
792 Value
*OldRHS
= Op
->getOperand(1);
794 if (NewLHS
== OldLHS
&& NewRHS
== OldRHS
)
795 // Nothing changed, leave it alone.
798 if (NewLHS
== OldRHS
&& NewRHS
== OldLHS
) {
799 // The order of the operands was reversed. Swap them.
800 DEBUG(dbgs() << "RA: " << *Op
<< '\n');
802 DEBUG(dbgs() << "TO: " << *Op
<< '\n');
808 // The new operation differs non-trivially from the original. Overwrite
809 // the old operands with the new ones.
810 DEBUG(dbgs() << "RA: " << *Op
<< '\n');
811 if (NewLHS
!= OldLHS
) {
812 BinaryOperator
*BO
= isReassociableOp(OldLHS
, Opcode
);
813 if (BO
&& !NotRewritable
.count(BO
))
814 NodesToRewrite
.push_back(BO
);
815 Op
->setOperand(0, NewLHS
);
817 if (NewRHS
!= OldRHS
) {
818 BinaryOperator
*BO
= isReassociableOp(OldRHS
, Opcode
);
819 if (BO
&& !NotRewritable
.count(BO
))
820 NodesToRewrite
.push_back(BO
);
821 Op
->setOperand(1, NewRHS
);
823 DEBUG(dbgs() << "TO: " << *Op
<< '\n');
825 ExpressionChanged
= Op
;
832 // Not the last operation. The left-hand side will be a sub-expression
833 // while the right-hand side will be the current element of Ops.
834 Value
*NewRHS
= Ops
[i
].Op
;
835 if (NewRHS
!= Op
->getOperand(1)) {
836 DEBUG(dbgs() << "RA: " << *Op
<< '\n');
837 if (NewRHS
== Op
->getOperand(0)) {
838 // The new right-hand side was already present as the left operand. If
839 // we are lucky then swapping the operands will sort out both of them.
842 // Overwrite with the new right-hand side.
843 BinaryOperator
*BO
= isReassociableOp(Op
->getOperand(1), Opcode
);
844 if (BO
&& !NotRewritable
.count(BO
))
845 NodesToRewrite
.push_back(BO
);
846 Op
->setOperand(1, NewRHS
);
847 ExpressionChanged
= Op
;
849 DEBUG(dbgs() << "TO: " << *Op
<< '\n');
854 // Now deal with the left-hand side. If this is already an operation node
855 // from the original expression then just rewrite the rest of the expression
857 BinaryOperator
*BO
= isReassociableOp(Op
->getOperand(0), Opcode
);
858 if (BO
&& !NotRewritable
.count(BO
)) {
863 // Otherwise, grab a spare node from the original expression and use that as
864 // the left-hand side. If there are no nodes left then the optimizers made
865 // an expression with more nodes than the original! This usually means that
866 // they did something stupid but it might mean that the problem was just too
867 // hard (finding the mimimal number of multiplications needed to realize a
868 // multiplication expression is NP-complete). Whatever the reason, smart or
869 // stupid, create a new node if there are none left.
870 BinaryOperator
*NewOp
;
871 if (NodesToRewrite
.empty()) {
872 Constant
*Undef
= UndefValue::get(I
->getType());
873 NewOp
= BinaryOperator::Create(Instruction::BinaryOps(Opcode
),
874 Undef
, Undef
, "", I
);
875 if (NewOp
->getType()->isFloatingPointTy())
876 NewOp
->setFastMathFlags(I
->getFastMathFlags());
878 NewOp
= NodesToRewrite
.pop_back_val();
881 DEBUG(dbgs() << "RA: " << *Op
<< '\n');
882 Op
->setOperand(0, NewOp
);
883 DEBUG(dbgs() << "TO: " << *Op
<< '\n');
884 ExpressionChanged
= Op
;
890 // If the expression changed non-trivially then clear out all subclass data
891 // starting from the operator specified in ExpressionChanged, and compactify
892 // the operators to just before the expression root to guarantee that the
893 // expression tree is dominated by all of Ops.
894 if (ExpressionChanged
)
896 // Preserve FastMathFlags.
897 if (isa
<FPMathOperator
>(I
)) {
898 FastMathFlags Flags
= I
->getFastMathFlags();
899 ExpressionChanged
->clearSubclassOptionalData();
900 ExpressionChanged
->setFastMathFlags(Flags
);
902 ExpressionChanged
->clearSubclassOptionalData();
904 if (ExpressionChanged
== I
)
906 ExpressionChanged
->moveBefore(I
);
907 ExpressionChanged
= cast
<BinaryOperator
>(*ExpressionChanged
->user_begin());
910 // Throw away any left over nodes from the original expression.
911 for (unsigned i
= 0, e
= NodesToRewrite
.size(); i
!= e
; ++i
)
912 RedoInsts
.insert(NodesToRewrite
[i
]);
915 /// NegateValue - Insert instructions before the instruction pointed to by BI,
916 /// that computes the negative version of the value specified. The negative
917 /// version of the value is returned, and BI is left pointing at the instruction
918 /// that should be processed next by the reassociation pass.
919 static Value
*NegateValue(Value
*V
, Instruction
*BI
) {
920 if (ConstantFP
*C
= dyn_cast
<ConstantFP
>(V
))
921 return ConstantExpr::getFNeg(C
);
922 if (Constant
*C
= dyn_cast
<Constant
>(V
))
923 return ConstantExpr::getNeg(C
);
925 // We are trying to expose opportunity for reassociation. One of the things
926 // that we want to do to achieve this is to push a negation as deep into an
927 // expression chain as possible, to expose the add instructions. In practice,
928 // this means that we turn this:
929 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
930 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
931 // the constants. We assume that instcombine will clean up the mess later if
932 // we introduce tons of unnecessary negation instructions.
934 if (BinaryOperator
*I
=
935 isReassociableOp(V
, Instruction::Add
, Instruction::FAdd
)) {
936 // Push the negates through the add.
937 I
->setOperand(0, NegateValue(I
->getOperand(0), BI
));
938 I
->setOperand(1, NegateValue(I
->getOperand(1), BI
));
940 // We must move the add instruction here, because the neg instructions do
941 // not dominate the old add instruction in general. By moving it, we are
942 // assured that the neg instructions we just inserted dominate the
943 // instruction we are about to insert after them.
946 I
->setName(I
->getName()+".neg");
950 // Okay, we need to materialize a negated version of V with an instruction.
951 // Scan the use lists of V to see if we have one already.
952 for (User
*U
: V
->users()) {
953 if (!BinaryOperator::isNeg(U
) && !BinaryOperator::isFNeg(U
))
956 // We found one! Now we have to make sure that the definition dominates
957 // this use. We do this by moving it to the entry block (if it is a
958 // non-instruction value) or right after the definition. These negates will
959 // be zapped by reassociate later, so we don't need much finesse here.
960 BinaryOperator
*TheNeg
= cast
<BinaryOperator
>(U
);
962 // Verify that the negate is in this function, V might be a constant expr.
963 if (TheNeg
->getParent()->getParent() != BI
->getParent()->getParent())
966 BasicBlock::iterator InsertPt
;
967 if (Instruction
*InstInput
= dyn_cast
<Instruction
>(V
)) {
968 if (InvokeInst
*II
= dyn_cast
<InvokeInst
>(InstInput
)) {
969 InsertPt
= II
->getNormalDest()->begin();
971 InsertPt
= InstInput
;
974 while (isa
<PHINode
>(InsertPt
)) ++InsertPt
;
976 InsertPt
= TheNeg
->getParent()->getParent()->getEntryBlock().begin();
978 TheNeg
->moveBefore(InsertPt
);
982 // Insert a 'neg' instruction that subtracts the value from zero to get the
984 return CreateNeg(V
, V
->getName() + ".neg", BI
, BI
);
987 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
988 /// X-Y into (X + -Y).
989 static bool ShouldBreakUpSubtract(Instruction
*Sub
) {
990 // If this is a negation, we can't split it up!
991 if (BinaryOperator::isNeg(Sub
) || BinaryOperator::isFNeg(Sub
))
994 // Don't breakup X - undef.
995 if (isa
<UndefValue
>(Sub
->getOperand(1)))
998 // Don't bother to break this up unless either the LHS is an associable add or
999 // subtract or if this is only used by one.
1000 Value
*V0
= Sub
->getOperand(0);
1001 if (isReassociableOp(V0
, Instruction::Add
, Instruction::FAdd
) ||
1002 isReassociableOp(V0
, Instruction::Sub
, Instruction::FSub
))
1004 Value
*V1
= Sub
->getOperand(1);
1005 if (isReassociableOp(V1
, Instruction::Add
, Instruction::FAdd
) ||
1006 isReassociableOp(V1
, Instruction::Sub
, Instruction::FSub
))
1008 Value
*VB
= Sub
->user_back();
1009 if (Sub
->hasOneUse() &&
1010 (isReassociableOp(VB
, Instruction::Add
, Instruction::FAdd
) ||
1011 isReassociableOp(VB
, Instruction::Sub
, Instruction::FSub
)))
1017 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
1018 /// only used by an add, transform this into (X+(0-Y)) to promote better
1020 static BinaryOperator
*BreakUpSubtract(Instruction
*Sub
) {
1021 // Convert a subtract into an add and a neg instruction. This allows sub
1022 // instructions to be commuted with other add instructions.
1024 // Calculate the negative value of Operand 1 of the sub instruction,
1025 // and set it as the RHS of the add instruction we just made.
1027 Value
*NegVal
= NegateValue(Sub
->getOperand(1), Sub
);
1028 BinaryOperator
*New
= CreateAdd(Sub
->getOperand(0), NegVal
, "", Sub
, Sub
);
1029 Sub
->setOperand(0, Constant::getNullValue(Sub
->getType())); // Drop use of op.
1030 Sub
->setOperand(1, Constant::getNullValue(Sub
->getType())); // Drop use of op.
1033 // Everyone now refers to the add instruction.
1034 Sub
->replaceAllUsesWith(New
);
1035 New
->setDebugLoc(Sub
->getDebugLoc());
1037 DEBUG(dbgs() << "Negated: " << *New
<< '\n');
1041 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
1042 /// by one, change this into a multiply by a constant to assist with further
1044 static BinaryOperator
*ConvertShiftToMul(Instruction
*Shl
) {
1045 Constant
*MulCst
= ConstantInt::get(Shl
->getType(), 1);
1046 MulCst
= ConstantExpr::getShl(MulCst
, cast
<Constant
>(Shl
->getOperand(1)));
1048 BinaryOperator
*Mul
=
1049 BinaryOperator::CreateMul(Shl
->getOperand(0), MulCst
, "", Shl
);
1050 Shl
->setOperand(0, UndefValue::get(Shl
->getType())); // Drop use of op.
1053 // Everyone now refers to the mul instruction.
1054 Shl
->replaceAllUsesWith(Mul
);
1055 Mul
->setDebugLoc(Shl
->getDebugLoc());
1057 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1058 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1060 bool NSW
= cast
<BinaryOperator
>(Shl
)->hasNoSignedWrap();
1061 bool NUW
= cast
<BinaryOperator
>(Shl
)->hasNoUnsignedWrap();
1063 Mul
->setHasNoSignedWrap(true);
1064 Mul
->setHasNoUnsignedWrap(NUW
);
1068 /// FindInOperandList - Scan backwards and forwards among values with the same
1069 /// rank as element i to see if X exists. If X does not exist, return i. This
1070 /// is useful when scanning for 'x' when we see '-x' because they both get the
1072 static unsigned FindInOperandList(SmallVectorImpl
<ValueEntry
> &Ops
, unsigned i
,
1074 unsigned XRank
= Ops
[i
].Rank
;
1075 unsigned e
= Ops
.size();
1076 for (unsigned j
= i
+1; j
!= e
&& Ops
[j
].Rank
== XRank
; ++j
) {
1079 if (Instruction
*I1
= dyn_cast
<Instruction
>(Ops
[j
].Op
))
1080 if (Instruction
*I2
= dyn_cast
<Instruction
>(X
))
1081 if (I1
->isIdenticalTo(I2
))
1085 for (unsigned j
= i
-1; j
!= ~0U && Ops
[j
].Rank
== XRank
; --j
) {
1088 if (Instruction
*I1
= dyn_cast
<Instruction
>(Ops
[j
].Op
))
1089 if (Instruction
*I2
= dyn_cast
<Instruction
>(X
))
1090 if (I1
->isIdenticalTo(I2
))
1096 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
1097 /// and returning the result. Insert the tree before I.
1098 static Value
*EmitAddTreeOfValues(Instruction
*I
,
1099 SmallVectorImpl
<WeakVH
> &Ops
){
1100 if (Ops
.size() == 1) return Ops
.back();
1102 Value
*V1
= Ops
.back();
1104 Value
*V2
= EmitAddTreeOfValues(I
, Ops
);
1105 return CreateAdd(V2
, V1
, "tmp", I
, I
);
1108 /// RemoveFactorFromExpression - If V is an expression tree that is a
1109 /// multiplication sequence, and if this sequence contains a multiply by Factor,
1110 /// remove Factor from the tree and return the new tree.
1111 Value
*Reassociate::RemoveFactorFromExpression(Value
*V
, Value
*Factor
) {
1112 BinaryOperator
*BO
= isReassociableOp(V
, Instruction::Mul
, Instruction::FMul
);
1116 SmallVector
<RepeatedValue
, 8> Tree
;
1117 MadeChange
|= LinearizeExprTree(BO
, Tree
);
1118 SmallVector
<ValueEntry
, 8> Factors
;
1119 Factors
.reserve(Tree
.size());
1120 for (unsigned i
= 0, e
= Tree
.size(); i
!= e
; ++i
) {
1121 RepeatedValue E
= Tree
[i
];
1122 Factors
.append(E
.second
.getZExtValue(),
1123 ValueEntry(getRank(E
.first
), E
.first
));
1126 bool FoundFactor
= false;
1127 bool NeedsNegate
= false;
1128 for (unsigned i
= 0, e
= Factors
.size(); i
!= e
; ++i
) {
1129 if (Factors
[i
].Op
== Factor
) {
1131 Factors
.erase(Factors
.begin()+i
);
1135 // If this is a negative version of this factor, remove it.
1136 if (ConstantInt
*FC1
= dyn_cast
<ConstantInt
>(Factor
)) {
1137 if (ConstantInt
*FC2
= dyn_cast
<ConstantInt
>(Factors
[i
].Op
))
1138 if (FC1
->getValue() == -FC2
->getValue()) {
1139 FoundFactor
= NeedsNegate
= true;
1140 Factors
.erase(Factors
.begin()+i
);
1143 } else if (ConstantFP
*FC1
= dyn_cast
<ConstantFP
>(Factor
)) {
1144 if (ConstantFP
*FC2
= dyn_cast
<ConstantFP
>(Factors
[i
].Op
)) {
1145 APFloat
F1(FC1
->getValueAPF());
1146 APFloat
F2(FC2
->getValueAPF());
1148 if (F1
.compare(F2
) == APFloat::cmpEqual
) {
1149 FoundFactor
= NeedsNegate
= true;
1150 Factors
.erase(Factors
.begin() + i
);
1158 // Make sure to restore the operands to the expression tree.
1159 RewriteExprTree(BO
, Factors
);
1163 BasicBlock::iterator InsertPt
= BO
; ++InsertPt
;
1165 // If this was just a single multiply, remove the multiply and return the only
1166 // remaining operand.
1167 if (Factors
.size() == 1) {
1168 RedoInsts
.insert(BO
);
1171 RewriteExprTree(BO
, Factors
);
1176 V
= CreateNeg(V
, "neg", InsertPt
, BO
);
1181 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1182 /// add its operands as factors, otherwise add V to the list of factors.
1184 /// Ops is the top-level list of add operands we're trying to factor.
1185 static void FindSingleUseMultiplyFactors(Value
*V
,
1186 SmallVectorImpl
<Value
*> &Factors
,
1187 const SmallVectorImpl
<ValueEntry
> &Ops
) {
1188 BinaryOperator
*BO
= isReassociableOp(V
, Instruction::Mul
, Instruction::FMul
);
1190 Factors
.push_back(V
);
1194 // Otherwise, add the LHS and RHS to the list of factors.
1195 FindSingleUseMultiplyFactors(BO
->getOperand(1), Factors
, Ops
);
1196 FindSingleUseMultiplyFactors(BO
->getOperand(0), Factors
, Ops
);
1199 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1200 /// instruction. This optimizes based on identities. If it can be reduced to
1201 /// a single Value, it is returned, otherwise the Ops list is mutated as
1203 static Value
*OptimizeAndOrXor(unsigned Opcode
,
1204 SmallVectorImpl
<ValueEntry
> &Ops
) {
1205 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1206 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1207 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
) {
1208 // First, check for X and ~X in the operand list.
1209 assert(i
< Ops
.size());
1210 if (BinaryOperator::isNot(Ops
[i
].Op
)) { // Cannot occur for ^.
1211 Value
*X
= BinaryOperator::getNotArgument(Ops
[i
].Op
);
1212 unsigned FoundX
= FindInOperandList(Ops
, i
, X
);
1214 if (Opcode
== Instruction::And
) // ...&X&~X = 0
1215 return Constant::getNullValue(X
->getType());
1217 if (Opcode
== Instruction::Or
) // ...|X|~X = -1
1218 return Constant::getAllOnesValue(X
->getType());
1222 // Next, check for duplicate pairs of values, which we assume are next to
1223 // each other, due to our sorting criteria.
1224 assert(i
< Ops
.size());
1225 if (i
+1 != Ops
.size() && Ops
[i
+1].Op
== Ops
[i
].Op
) {
1226 if (Opcode
== Instruction::And
|| Opcode
== Instruction::Or
) {
1227 // Drop duplicate values for And and Or.
1228 Ops
.erase(Ops
.begin()+i
);
1234 // Drop pairs of values for Xor.
1235 assert(Opcode
== Instruction::Xor
);
1237 return Constant::getNullValue(Ops
[0].Op
->getType());
1240 Ops
.erase(Ops
.begin()+i
, Ops
.begin()+i
+2);
1248 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1249 /// instruction with the given two operands, and return the resulting
1250 /// instruction. There are two special cases: 1) if the constant operand is 0,
1251 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1253 static Value
*createAndInstr(Instruction
*InsertBefore
, Value
*Opnd
,
1254 const APInt
&ConstOpnd
) {
1255 if (ConstOpnd
!= 0) {
1256 if (!ConstOpnd
.isAllOnesValue()) {
1257 LLVMContext
&Ctx
= Opnd
->getType()->getContext();
1259 I
= BinaryOperator::CreateAnd(Opnd
, ConstantInt::get(Ctx
, ConstOpnd
),
1260 "and.ra", InsertBefore
);
1261 I
->setDebugLoc(InsertBefore
->getDebugLoc());
1269 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1270 // into "R ^ C", where C would be 0, and R is a symbolic value.
1272 // If it was successful, true is returned, and the "R" and "C" is returned
1273 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1274 // and both "Res" and "ConstOpnd" remain unchanged.
1276 bool Reassociate::CombineXorOpnd(Instruction
*I
, XorOpnd
*Opnd1
,
1277 APInt
&ConstOpnd
, Value
*&Res
) {
1278 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1279 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1280 // = (x & ~c1) ^ (c1 ^ c2)
1281 // It is useful only when c1 == c2.
1282 if (Opnd1
->isOrExpr() && Opnd1
->getConstPart() != 0) {
1283 if (!Opnd1
->getValue()->hasOneUse())
1286 const APInt
&C1
= Opnd1
->getConstPart();
1287 if (C1
!= ConstOpnd
)
1290 Value
*X
= Opnd1
->getSymbolicPart();
1291 Res
= createAndInstr(I
, X
, ~C1
);
1292 // ConstOpnd was C2, now C1 ^ C2.
1295 if (Instruction
*T
= dyn_cast
<Instruction
>(Opnd1
->getValue()))
1296 RedoInsts
.insert(T
);
1303 // Helper function of OptimizeXor(). It tries to simplify
1304 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1307 // If it was successful, true is returned, and the "R" and "C" is returned
1308 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1309 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1310 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1311 bool Reassociate::CombineXorOpnd(Instruction
*I
, XorOpnd
*Opnd1
, XorOpnd
*Opnd2
,
1312 APInt
&ConstOpnd
, Value
*&Res
) {
1313 Value
*X
= Opnd1
->getSymbolicPart();
1314 if (X
!= Opnd2
->getSymbolicPart())
1317 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1318 int DeadInstNum
= 1;
1319 if (Opnd1
->getValue()->hasOneUse())
1321 if (Opnd2
->getValue()->hasOneUse())
1325 // (x | c1) ^ (x & c2)
1326 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1327 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1328 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1330 if (Opnd1
->isOrExpr() != Opnd2
->isOrExpr()) {
1331 if (Opnd2
->isOrExpr())
1332 std::swap(Opnd1
, Opnd2
);
1334 const APInt
&C1
= Opnd1
->getConstPart();
1335 const APInt
&C2
= Opnd2
->getConstPart();
1336 APInt
C3((~C1
) ^ C2
);
1338 // Do not increase code size!
1339 if (C3
!= 0 && !C3
.isAllOnesValue()) {
1340 int NewInstNum
= ConstOpnd
!= 0 ? 1 : 2;
1341 if (NewInstNum
> DeadInstNum
)
1345 Res
= createAndInstr(I
, X
, C3
);
1348 } else if (Opnd1
->isOrExpr()) {
1349 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1351 const APInt
&C1
= Opnd1
->getConstPart();
1352 const APInt
&C2
= Opnd2
->getConstPart();
1355 // Do not increase code size
1356 if (C3
!= 0 && !C3
.isAllOnesValue()) {
1357 int NewInstNum
= ConstOpnd
!= 0 ? 1 : 2;
1358 if (NewInstNum
> DeadInstNum
)
1362 Res
= createAndInstr(I
, X
, C3
);
1365 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1367 const APInt
&C1
= Opnd1
->getConstPart();
1368 const APInt
&C2
= Opnd2
->getConstPart();
1370 Res
= createAndInstr(I
, X
, C3
);
1373 // Put the original operands in the Redo list; hope they will be deleted
1375 if (Instruction
*T
= dyn_cast
<Instruction
>(Opnd1
->getValue()))
1376 RedoInsts
.insert(T
);
1377 if (Instruction
*T
= dyn_cast
<Instruction
>(Opnd2
->getValue()))
1378 RedoInsts
.insert(T
);
1383 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1384 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1386 Value
*Reassociate::OptimizeXor(Instruction
*I
,
1387 SmallVectorImpl
<ValueEntry
> &Ops
) {
1388 if (Value
*V
= OptimizeAndOrXor(Instruction::Xor
, Ops
))
1391 if (Ops
.size() == 1)
1394 SmallVector
<XorOpnd
, 8> Opnds
;
1395 SmallVector
<XorOpnd
*, 8> OpndPtrs
;
1396 Type
*Ty
= Ops
[0].Op
->getType();
1397 APInt
ConstOpnd(Ty
->getIntegerBitWidth(), 0);
1399 // Step 1: Convert ValueEntry to XorOpnd
1400 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
) {
1401 Value
*V
= Ops
[i
].Op
;
1402 if (!isa
<ConstantInt
>(V
)) {
1404 O
.setSymbolicRank(getRank(O
.getSymbolicPart()));
1407 ConstOpnd
^= cast
<ConstantInt
>(V
)->getValue();
1410 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1411 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1412 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1413 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1414 // when new elements are added to the vector.
1415 for (unsigned i
= 0, e
= Opnds
.size(); i
!= e
; ++i
)
1416 OpndPtrs
.push_back(&Opnds
[i
]);
1418 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1419 // the same symbolic value cluster together. For instance, the input operand
1420 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1421 // ("x | 123", "x & 789", "y & 456").
1422 std::stable_sort(OpndPtrs
.begin(), OpndPtrs
.end(), XorOpnd::PtrSortFunctor());
1424 // Step 3: Combine adjacent operands
1425 XorOpnd
*PrevOpnd
= nullptr;
1426 bool Changed
= false;
1427 for (unsigned i
= 0, e
= Opnds
.size(); i
< e
; i
++) {
1428 XorOpnd
*CurrOpnd
= OpndPtrs
[i
];
1429 // The combined value
1432 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1433 if (ConstOpnd
!= 0 && CombineXorOpnd(I
, CurrOpnd
, ConstOpnd
, CV
)) {
1436 *CurrOpnd
= XorOpnd(CV
);
1438 CurrOpnd
->Invalidate();
1443 if (!PrevOpnd
|| CurrOpnd
->getSymbolicPart() != PrevOpnd
->getSymbolicPart()) {
1444 PrevOpnd
= CurrOpnd
;
1448 // step 3.2: When previous and current operands share the same symbolic
1449 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1451 if (CombineXorOpnd(I
, CurrOpnd
, PrevOpnd
, ConstOpnd
, CV
)) {
1452 // Remove previous operand
1453 PrevOpnd
->Invalidate();
1455 *CurrOpnd
= XorOpnd(CV
);
1456 PrevOpnd
= CurrOpnd
;
1458 CurrOpnd
->Invalidate();
1465 // Step 4: Reassemble the Ops
1468 for (unsigned int i
= 0, e
= Opnds
.size(); i
< e
; i
++) {
1469 XorOpnd
&O
= Opnds
[i
];
1472 ValueEntry
VE(getRank(O
.getValue()), O
.getValue());
1475 if (ConstOpnd
!= 0) {
1476 Value
*C
= ConstantInt::get(Ty
->getContext(), ConstOpnd
);
1477 ValueEntry
VE(getRank(C
), C
);
1480 int Sz
= Ops
.size();
1482 return Ops
.back().Op
;
1484 assert(ConstOpnd
== 0);
1485 return ConstantInt::get(Ty
->getContext(), ConstOpnd
);
1492 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1493 /// optimizes based on identities. If it can be reduced to a single Value, it
1494 /// is returned, otherwise the Ops list is mutated as necessary.
1495 Value
*Reassociate::OptimizeAdd(Instruction
*I
,
1496 SmallVectorImpl
<ValueEntry
> &Ops
) {
1497 // Scan the operand lists looking for X and -X pairs. If we find any, we
1498 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1500 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1502 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
) {
1503 Value
*TheOp
= Ops
[i
].Op
;
1504 // Check to see if we've seen this operand before. If so, we factor all
1505 // instances of the operand together. Due to our sorting criteria, we know
1506 // that these need to be next to each other in the vector.
1507 if (i
+1 != Ops
.size() && Ops
[i
+1].Op
== TheOp
) {
1508 // Rescan the list, remove all instances of this operand from the expr.
1509 unsigned NumFound
= 0;
1511 Ops
.erase(Ops
.begin()+i
);
1513 } while (i
!= Ops
.size() && Ops
[i
].Op
== TheOp
);
1515 DEBUG(dbgs() << "\nFACTORING [" << NumFound
<< "]: " << *TheOp
<< '\n');
1518 // Insert a new multiply.
1519 Type
*Ty
= TheOp
->getType();
1520 Constant
*C
= Ty
->isIntegerTy() ? ConstantInt::get(Ty
, NumFound
)
1521 : ConstantFP::get(Ty
, NumFound
);
1522 Instruction
*Mul
= CreateMul(TheOp
, C
, "factor", I
, I
);
1524 // Now that we have inserted a multiply, optimize it. This allows us to
1525 // handle cases that require multiple factoring steps, such as this:
1526 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1527 RedoInsts
.insert(Mul
);
1529 // If every add operand was a duplicate, return the multiply.
1533 // Otherwise, we had some input that didn't have the dupe, such as
1534 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1535 // things being added by this operation.
1536 Ops
.insert(Ops
.begin(), ValueEntry(getRank(Mul
), Mul
));
1543 // Check for X and -X or X and ~X in the operand list.
1544 if (!BinaryOperator::isNeg(TheOp
) && !BinaryOperator::isFNeg(TheOp
) &&
1545 !BinaryOperator::isNot(TheOp
))
1549 if (BinaryOperator::isNeg(TheOp
) || BinaryOperator::isFNeg(TheOp
))
1550 X
= BinaryOperator::getNegArgument(TheOp
);
1551 else if (BinaryOperator::isNot(TheOp
))
1552 X
= BinaryOperator::getNotArgument(TheOp
);
1554 unsigned FoundX
= FindInOperandList(Ops
, i
, X
);
1558 // Remove X and -X from the operand list.
1559 if (Ops
.size() == 2 &&
1560 (BinaryOperator::isNeg(TheOp
) || BinaryOperator::isFNeg(TheOp
)))
1561 return Constant::getNullValue(X
->getType());
1563 // Remove X and ~X from the operand list.
1564 if (Ops
.size() == 2 && BinaryOperator::isNot(TheOp
))
1565 return Constant::getAllOnesValue(X
->getType());
1567 Ops
.erase(Ops
.begin()+i
);
1571 --i
; // Need to back up an extra one.
1572 Ops
.erase(Ops
.begin()+FoundX
);
1574 --i
; // Revisit element.
1575 e
-= 2; // Removed two elements.
1577 // if X and ~X we append -1 to the operand list.
1578 if (BinaryOperator::isNot(TheOp
)) {
1579 Value
*V
= Constant::getAllOnesValue(X
->getType());
1580 Ops
.insert(Ops
.end(), ValueEntry(getRank(V
), V
));
1585 // Scan the operand list, checking to see if there are any common factors
1586 // between operands. Consider something like A*A+A*B*C+D. We would like to
1587 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1588 // To efficiently find this, we count the number of times a factor occurs
1589 // for any ADD operands that are MULs.
1590 DenseMap
<Value
*, unsigned> FactorOccurrences
;
1592 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1593 // where they are actually the same multiply.
1594 unsigned MaxOcc
= 0;
1595 Value
*MaxOccVal
= nullptr;
1596 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
) {
1597 BinaryOperator
*BOp
=
1598 isReassociableOp(Ops
[i
].Op
, Instruction::Mul
, Instruction::FMul
);
1602 // Compute all of the factors of this added value.
1603 SmallVector
<Value
*, 8> Factors
;
1604 FindSingleUseMultiplyFactors(BOp
, Factors
, Ops
);
1605 assert(Factors
.size() > 1 && "Bad linearize!");
1607 // Add one to FactorOccurrences for each unique factor in this op.
1608 SmallPtrSet
<Value
*, 8> Duplicates
;
1609 for (unsigned i
= 0, e
= Factors
.size(); i
!= e
; ++i
) {
1610 Value
*Factor
= Factors
[i
];
1611 if (!Duplicates
.insert(Factor
).second
)
1614 unsigned Occ
= ++FactorOccurrences
[Factor
];
1620 // If Factor is a negative constant, add the negated value as a factor
1621 // because we can percolate the negate out. Watch for minint, which
1622 // cannot be positivified.
1623 if (ConstantInt
*CI
= dyn_cast
<ConstantInt
>(Factor
)) {
1624 if (CI
->isNegative() && !CI
->isMinValue(true)) {
1625 Factor
= ConstantInt::get(CI
->getContext(), -CI
->getValue());
1626 assert(!Duplicates
.count(Factor
) &&
1627 "Shouldn't have two constant factors, missed a canonicalize");
1628 unsigned Occ
= ++FactorOccurrences
[Factor
];
1634 } else if (ConstantFP
*CF
= dyn_cast
<ConstantFP
>(Factor
)) {
1635 if (CF
->isNegative()) {
1636 APFloat
F(CF
->getValueAPF());
1638 Factor
= ConstantFP::get(CF
->getContext(), F
);
1639 assert(!Duplicates
.count(Factor
) &&
1640 "Shouldn't have two constant factors, missed a canonicalize");
1641 unsigned Occ
= ++FactorOccurrences
[Factor
];
1651 // If any factor occurred more than one time, we can pull it out.
1653 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc
<< "]: " << *MaxOccVal
<< '\n');
1656 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1657 // this, we could otherwise run into situations where removing a factor
1658 // from an expression will drop a use of maxocc, and this can cause
1659 // RemoveFactorFromExpression on successive values to behave differently.
1660 Instruction
*DummyInst
=
1661 I
->getType()->isIntegerTy()
1662 ? BinaryOperator::CreateAdd(MaxOccVal
, MaxOccVal
)
1663 : BinaryOperator::CreateFAdd(MaxOccVal
, MaxOccVal
);
1665 SmallVector
<WeakVH
, 4> NewMulOps
;
1666 for (unsigned i
= 0; i
!= Ops
.size(); ++i
) {
1667 // Only try to remove factors from expressions we're allowed to.
1668 BinaryOperator
*BOp
=
1669 isReassociableOp(Ops
[i
].Op
, Instruction::Mul
, Instruction::FMul
);
1673 if (Value
*V
= RemoveFactorFromExpression(Ops
[i
].Op
, MaxOccVal
)) {
1674 // The factorized operand may occur several times. Convert them all in
1676 for (unsigned j
= Ops
.size(); j
!= i
;) {
1678 if (Ops
[j
].Op
== Ops
[i
].Op
) {
1679 NewMulOps
.push_back(V
);
1680 Ops
.erase(Ops
.begin()+j
);
1687 // No need for extra uses anymore.
1690 unsigned NumAddedValues
= NewMulOps
.size();
1691 Value
*V
= EmitAddTreeOfValues(I
, NewMulOps
);
1693 // Now that we have inserted the add tree, optimize it. This allows us to
1694 // handle cases that require multiple factoring steps, such as this:
1695 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1696 assert(NumAddedValues
> 1 && "Each occurrence should contribute a value");
1697 (void)NumAddedValues
;
1698 if (Instruction
*VI
= dyn_cast
<Instruction
>(V
))
1699 RedoInsts
.insert(VI
);
1701 // Create the multiply.
1702 Instruction
*V2
= CreateMul(V
, MaxOccVal
, "tmp", I
, I
);
1704 // Rerun associate on the multiply in case the inner expression turned into
1705 // a multiply. We want to make sure that we keep things in canonical form.
1706 RedoInsts
.insert(V2
);
1708 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1709 // entire result expression is just the multiply "A*(B+C)".
1713 // Otherwise, we had some input that didn't have the factor, such as
1714 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1715 // things being added by this operation.
1716 Ops
.insert(Ops
.begin(), ValueEntry(getRank(V2
), V2
));
1722 /// \brief Build up a vector of value/power pairs factoring a product.
1724 /// Given a series of multiplication operands, build a vector of factors and
1725 /// the powers each is raised to when forming the final product. Sort them in
1726 /// the order of descending power.
1728 /// (x*x) -> [(x, 2)]
1729 /// ((x*x)*x) -> [(x, 3)]
1730 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1732 /// \returns Whether any factors have a power greater than one.
1733 bool Reassociate::collectMultiplyFactors(SmallVectorImpl
<ValueEntry
> &Ops
,
1734 SmallVectorImpl
<Factor
> &Factors
) {
1735 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1736 // Compute the sum of powers of simplifiable factors.
1737 unsigned FactorPowerSum
= 0;
1738 for (unsigned Idx
= 1, Size
= Ops
.size(); Idx
< Size
; ++Idx
) {
1739 Value
*Op
= Ops
[Idx
-1].Op
;
1741 // Count the number of occurrences of this value.
1743 for (; Idx
< Size
&& Ops
[Idx
].Op
== Op
; ++Idx
)
1745 // Track for simplification all factors which occur 2 or more times.
1747 FactorPowerSum
+= Count
;
1750 // We can only simplify factors if the sum of the powers of our simplifiable
1751 // factors is 4 or higher. When that is the case, we will *always* have
1752 // a simplification. This is an important invariant to prevent cyclicly
1753 // trying to simplify already minimal formations.
1754 if (FactorPowerSum
< 4)
1757 // Now gather the simplifiable factors, removing them from Ops.
1759 for (unsigned Idx
= 1; Idx
< Ops
.size(); ++Idx
) {
1760 Value
*Op
= Ops
[Idx
-1].Op
;
1762 // Count the number of occurrences of this value.
1764 for (; Idx
< Ops
.size() && Ops
[Idx
].Op
== Op
; ++Idx
)
1768 // Move an even number of occurrences to Factors.
1771 FactorPowerSum
+= Count
;
1772 Factors
.push_back(Factor(Op
, Count
));
1773 Ops
.erase(Ops
.begin()+Idx
, Ops
.begin()+Idx
+Count
);
1776 // None of the adjustments above should have reduced the sum of factor powers
1777 // below our mininum of '4'.
1778 assert(FactorPowerSum
>= 4);
1780 std::stable_sort(Factors
.begin(), Factors
.end(), Factor::PowerDescendingSorter());
1784 /// \brief Build a tree of multiplies, computing the product of Ops.
1785 static Value
*buildMultiplyTree(IRBuilder
<> &Builder
,
1786 SmallVectorImpl
<Value
*> &Ops
) {
1787 if (Ops
.size() == 1)
1790 Value
*LHS
= Ops
.pop_back_val();
1792 if (LHS
->getType()->isIntegerTy())
1793 LHS
= Builder
.CreateMul(LHS
, Ops
.pop_back_val());
1795 LHS
= Builder
.CreateFMul(LHS
, Ops
.pop_back_val());
1796 } while (!Ops
.empty());
1801 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1803 /// Given a vector of values raised to various powers, where no two values are
1804 /// equal and the powers are sorted in decreasing order, compute the minimal
1805 /// DAG of multiplies to compute the final product, and return that product
1807 Value
*Reassociate::buildMinimalMultiplyDAG(IRBuilder
<> &Builder
,
1808 SmallVectorImpl
<Factor
> &Factors
) {
1809 assert(Factors
[0].Power
);
1810 SmallVector
<Value
*, 4> OuterProduct
;
1811 for (unsigned LastIdx
= 0, Idx
= 1, Size
= Factors
.size();
1812 Idx
< Size
&& Factors
[Idx
].Power
> 0; ++Idx
) {
1813 if (Factors
[Idx
].Power
!= Factors
[LastIdx
].Power
) {
1818 // We want to multiply across all the factors with the same power so that
1819 // we can raise them to that power as a single entity. Build a mini tree
1821 SmallVector
<Value
*, 4> InnerProduct
;
1822 InnerProduct
.push_back(Factors
[LastIdx
].Base
);
1824 InnerProduct
.push_back(Factors
[Idx
].Base
);
1826 } while (Idx
< Size
&& Factors
[Idx
].Power
== Factors
[LastIdx
].Power
);
1828 // Reset the base value of the first factor to the new expression tree.
1829 // We'll remove all the factors with the same power in a second pass.
1830 Value
*M
= Factors
[LastIdx
].Base
= buildMultiplyTree(Builder
, InnerProduct
);
1831 if (Instruction
*MI
= dyn_cast
<Instruction
>(M
))
1832 RedoInsts
.insert(MI
);
1836 // Unique factors with equal powers -- we've folded them into the first one's
1838 Factors
.erase(std::unique(Factors
.begin(), Factors
.end(),
1839 Factor::PowerEqual()),
1842 // Iteratively collect the base of each factor with an add power into the
1843 // outer product, and halve each power in preparation for squaring the
1845 for (unsigned Idx
= 0, Size
= Factors
.size(); Idx
!= Size
; ++Idx
) {
1846 if (Factors
[Idx
].Power
& 1)
1847 OuterProduct
.push_back(Factors
[Idx
].Base
);
1848 Factors
[Idx
].Power
>>= 1;
1850 if (Factors
[0].Power
) {
1851 Value
*SquareRoot
= buildMinimalMultiplyDAG(Builder
, Factors
);
1852 OuterProduct
.push_back(SquareRoot
);
1853 OuterProduct
.push_back(SquareRoot
);
1855 if (OuterProduct
.size() == 1)
1856 return OuterProduct
.front();
1858 Value
*V
= buildMultiplyTree(Builder
, OuterProduct
);
1862 Value
*Reassociate::OptimizeMul(BinaryOperator
*I
,
1863 SmallVectorImpl
<ValueEntry
> &Ops
) {
1864 // We can only optimize the multiplies when there is a chain of more than
1865 // three, such that a balanced tree might require fewer total multiplies.
1869 // Try to turn linear trees of multiplies without other uses of the
1870 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1872 SmallVector
<Factor
, 4> Factors
;
1873 if (!collectMultiplyFactors(Ops
, Factors
))
1874 return nullptr; // All distinct factors, so nothing left for us to do.
1876 IRBuilder
<> Builder(I
);
1877 Value
*V
= buildMinimalMultiplyDAG(Builder
, Factors
);
1881 ValueEntry NewEntry
= ValueEntry(getRank(V
), V
);
1882 Ops
.insert(std::lower_bound(Ops
.begin(), Ops
.end(), NewEntry
), NewEntry
);
1886 Value
*Reassociate::OptimizeExpression(BinaryOperator
*I
,
1887 SmallVectorImpl
<ValueEntry
> &Ops
) {
1888 // Now that we have the linearized expression tree, try to optimize it.
1889 // Start by folding any constants that we found.
1890 Constant
*Cst
= nullptr;
1891 unsigned Opcode
= I
->getOpcode();
1892 while (!Ops
.empty() && isa
<Constant
>(Ops
.back().Op
)) {
1893 Constant
*C
= cast
<Constant
>(Ops
.pop_back_val().Op
);
1894 Cst
= Cst
? ConstantExpr::get(Opcode
, C
, Cst
) : C
;
1896 // If there was nothing but constants then we are done.
1900 // Put the combined constant back at the end of the operand list, except if
1901 // there is no point. For example, an add of 0 gets dropped here, while a
1902 // multiplication by zero turns the whole expression into zero.
1903 if (Cst
&& Cst
!= ConstantExpr::getBinOpIdentity(Opcode
, I
->getType())) {
1904 if (Cst
== ConstantExpr::getBinOpAbsorber(Opcode
, I
->getType()))
1906 Ops
.push_back(ValueEntry(0, Cst
));
1909 if (Ops
.size() == 1) return Ops
[0].Op
;
1911 // Handle destructive annihilation due to identities between elements in the
1912 // argument list here.
1913 unsigned NumOps
= Ops
.size();
1916 case Instruction::And
:
1917 case Instruction::Or
:
1918 if (Value
*Result
= OptimizeAndOrXor(Opcode
, Ops
))
1922 case Instruction::Xor
:
1923 if (Value
*Result
= OptimizeXor(I
, Ops
))
1927 case Instruction::Add
:
1928 case Instruction::FAdd
:
1929 if (Value
*Result
= OptimizeAdd(I
, Ops
))
1933 case Instruction::Mul
:
1934 case Instruction::FMul
:
1935 if (Value
*Result
= OptimizeMul(I
, Ops
))
1940 if (Ops
.size() != NumOps
)
1941 return OptimizeExpression(I
, Ops
);
1945 /// EraseInst - Zap the given instruction, adding interesting operands to the
1947 void Reassociate::EraseInst(Instruction
*I
) {
1948 assert(isInstructionTriviallyDead(I
) && "Trivially dead instructions only!");
1949 SmallVector
<Value
*, 8> Ops(I
->op_begin(), I
->op_end());
1950 // Erase the dead instruction.
1951 ValueRankMap
.erase(I
);
1952 RedoInsts
.remove(I
);
1953 I
->eraseFromParent();
1954 // Optimize its operands.
1955 SmallPtrSet
<Instruction
*, 8> Visited
; // Detect self-referential nodes.
1956 for (unsigned i
= 0, e
= Ops
.size(); i
!= e
; ++i
)
1957 if (Instruction
*Op
= dyn_cast
<Instruction
>(Ops
[i
])) {
1958 // If this is a node in an expression tree, climb to the expression root
1959 // and add that since that's where optimization actually happens.
1960 unsigned Opcode
= Op
->getOpcode();
1961 while (Op
->hasOneUse() && Op
->user_back()->getOpcode() == Opcode
&&
1962 Visited
.insert(Op
).second
)
1963 Op
= Op
->user_back();
1964 RedoInsts
.insert(Op
);
1968 // Canonicalize expressions of the following form:
1969 // x + (-Constant * y) -> x - (Constant * y)
1970 // x - (-Constant * y) -> x + (Constant * y)
1971 Instruction
*Reassociate::canonicalizeNegConstExpr(Instruction
*I
) {
1972 if (!I
->hasOneUse() || I
->getType()->isVectorTy())
1975 // Must be a mul, fmul, or fdiv instruction.
1976 unsigned Opcode
= I
->getOpcode();
1977 if (Opcode
!= Instruction::Mul
&& Opcode
!= Instruction::FMul
&&
1978 Opcode
!= Instruction::FDiv
)
1981 // Must have at least one constant operand.
1982 Constant
*C0
= dyn_cast
<Constant
>(I
->getOperand(0));
1983 Constant
*C1
= dyn_cast
<Constant
>(I
->getOperand(1));
1987 // Must be a negative ConstantInt or ConstantFP.
1988 Constant
*C
= C0
? C0
: C1
;
1989 unsigned ConstIdx
= C0
? 0 : 1;
1990 if (auto *CI
= dyn_cast
<ConstantInt
>(C
)) {
1991 if (!CI
->isNegative())
1993 } else if (auto *CF
= dyn_cast
<ConstantFP
>(C
)) {
1994 if (!CF
->isNegative())
1999 // User must be a binary operator with one or more uses.
2000 Instruction
*User
= I
->user_back();
2001 if (!isa
<BinaryOperator
>(User
) || !User
->getNumUses())
2004 unsigned UserOpcode
= User
->getOpcode();
2005 if (UserOpcode
!= Instruction::Add
&& UserOpcode
!= Instruction::FAdd
&&
2006 UserOpcode
!= Instruction::Sub
&& UserOpcode
!= Instruction::FSub
)
2009 // Subtraction is not commutative. Explicitly, the following transform is
2010 // not valid: (-Constant * y) - x -> x + (Constant * y)
2011 if (!User
->isCommutative() && User
->getOperand(1) != I
)
2014 // Change the sign of the constant.
2015 if (ConstantInt
*CI
= dyn_cast
<ConstantInt
>(C
))
2016 I
->setOperand(ConstIdx
, ConstantInt::get(CI
->getContext(), -CI
->getValue()));
2018 ConstantFP
*CF
= cast
<ConstantFP
>(C
);
2019 APFloat Val
= CF
->getValueAPF();
2021 I
->setOperand(ConstIdx
, ConstantFP::get(CF
->getContext(), Val
));
2024 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
2025 // ((-Const*y) + x) -> (x + (-Const*y)).
2026 if (User
->getOperand(0) == I
&& User
->isCommutative())
2027 cast
<BinaryOperator
>(User
)->swapOperands();
2029 Value
*Op0
= User
->getOperand(0);
2030 Value
*Op1
= User
->getOperand(1);
2032 switch(UserOpcode
) {
2034 llvm_unreachable("Unexpected Opcode!");
2035 case Instruction::Add
:
2036 NI
= BinaryOperator::CreateSub(Op0
, Op1
);
2038 case Instruction::Sub
:
2039 NI
= BinaryOperator::CreateAdd(Op0
, Op1
);
2041 case Instruction::FAdd
:
2042 NI
= BinaryOperator::CreateFSub(Op0
, Op1
);
2043 NI
->setFastMathFlags(cast
<FPMathOperator
>(User
)->getFastMathFlags());
2045 case Instruction::FSub
:
2046 NI
= BinaryOperator::CreateFAdd(Op0
, Op1
);
2047 NI
->setFastMathFlags(cast
<FPMathOperator
>(User
)->getFastMathFlags());
2051 NI
->insertBefore(User
);
2052 NI
->setName(User
->getName());
2053 User
->replaceAllUsesWith(NI
);
2054 NI
->setDebugLoc(I
->getDebugLoc());
2055 RedoInsts
.insert(I
);
2060 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
2061 /// instructions is not allowed.
2062 void Reassociate::OptimizeInst(Instruction
*I
) {
2063 // Only consider operations that we understand.
2064 if (!isa
<BinaryOperator
>(I
))
2067 if (I
->getOpcode() == Instruction::Shl
&& isa
<ConstantInt
>(I
->getOperand(1)))
2068 // If an operand of this shift is a reassociable multiply, or if the shift
2069 // is used by a reassociable multiply or add, turn into a multiply.
2070 if (isReassociableOp(I
->getOperand(0), Instruction::Mul
) ||
2072 (isReassociableOp(I
->user_back(), Instruction::Mul
) ||
2073 isReassociableOp(I
->user_back(), Instruction::Add
)))) {
2074 Instruction
*NI
= ConvertShiftToMul(I
);
2075 RedoInsts
.insert(I
);
2080 // Canonicalize negative constants out of expressions.
2081 if (Instruction
*Res
= canonicalizeNegConstExpr(I
))
2084 // Commute binary operators, to canonicalize the order of their operands.
2085 // This can potentially expose more CSE opportunities, and makes writing other
2086 // transformations simpler.
2087 if (I
->isCommutative())
2088 canonicalizeOperands(I
);
2090 // Don't optimize vector instructions.
2091 if (I
->getType()->isVectorTy())
2094 // Don't optimize floating point instructions that don't have unsafe algebra.
2095 if (I
->getType()->isFloatingPointTy() && !I
->hasUnsafeAlgebra())
2098 // Do not reassociate boolean (i1) expressions. We want to preserve the
2099 // original order of evaluation for short-circuited comparisons that
2100 // SimplifyCFG has folded to AND/OR expressions. If the expression
2101 // is not further optimized, it is likely to be transformed back to a
2102 // short-circuited form for code gen, and the source order may have been
2103 // optimized for the most likely conditions.
2104 if (I
->getType()->isIntegerTy(1))
2107 // If this is a subtract instruction which is not already in negate form,
2108 // see if we can convert it to X+-Y.
2109 if (I
->getOpcode() == Instruction::Sub
) {
2110 if (ShouldBreakUpSubtract(I
)) {
2111 Instruction
*NI
= BreakUpSubtract(I
);
2112 RedoInsts
.insert(I
);
2115 } else if (BinaryOperator::isNeg(I
)) {
2116 // Otherwise, this is a negation. See if the operand is a multiply tree
2117 // and if this is not an inner node of a multiply tree.
2118 if (isReassociableOp(I
->getOperand(1), Instruction::Mul
) &&
2120 !isReassociableOp(I
->user_back(), Instruction::Mul
))) {
2121 Instruction
*NI
= LowerNegateToMultiply(I
);
2122 RedoInsts
.insert(I
);
2127 } else if (I
->getOpcode() == Instruction::FSub
) {
2128 if (ShouldBreakUpSubtract(I
)) {
2129 Instruction
*NI
= BreakUpSubtract(I
);
2130 RedoInsts
.insert(I
);
2133 } else if (BinaryOperator::isFNeg(I
)) {
2134 // Otherwise, this is a negation. See if the operand is a multiply tree
2135 // and if this is not an inner node of a multiply tree.
2136 if (isReassociableOp(I
->getOperand(1), Instruction::FMul
) &&
2138 !isReassociableOp(I
->user_back(), Instruction::FMul
))) {
2139 Instruction
*NI
= LowerNegateToMultiply(I
);
2140 RedoInsts
.insert(I
);
2147 // If this instruction is an associative binary operator, process it.
2148 if (!I
->isAssociative()) return;
2149 BinaryOperator
*BO
= cast
<BinaryOperator
>(I
);
2151 // If this is an interior node of a reassociable tree, ignore it until we
2152 // get to the root of the tree, to avoid N^2 analysis.
2153 unsigned Opcode
= BO
->getOpcode();
2154 if (BO
->hasOneUse() && BO
->user_back()->getOpcode() == Opcode
)
2157 // If this is an add tree that is used by a sub instruction, ignore it
2158 // until we process the subtract.
2159 if (BO
->hasOneUse() && BO
->getOpcode() == Instruction::Add
&&
2160 cast
<Instruction
>(BO
->user_back())->getOpcode() == Instruction::Sub
)
2162 if (BO
->hasOneUse() && BO
->getOpcode() == Instruction::FAdd
&&
2163 cast
<Instruction
>(BO
->user_back())->getOpcode() == Instruction::FSub
)
2166 ReassociateExpression(BO
);
2169 void Reassociate::ReassociateExpression(BinaryOperator
*I
) {
2170 assert(!I
->getType()->isVectorTy() &&
2171 "Reassociation of vector instructions is not supported.");
2173 // First, walk the expression tree, linearizing the tree, collecting the
2174 // operand information.
2175 SmallVector
<RepeatedValue
, 8> Tree
;
2176 MadeChange
|= LinearizeExprTree(I
, Tree
);
2177 SmallVector
<ValueEntry
, 8> Ops
;
2178 Ops
.reserve(Tree
.size());
2179 for (unsigned i
= 0, e
= Tree
.size(); i
!= e
; ++i
) {
2180 RepeatedValue E
= Tree
[i
];
2181 Ops
.append(E
.second
.getZExtValue(),
2182 ValueEntry(getRank(E
.first
), E
.first
));
2185 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I
, Ops
); dbgs() << '\n');
2187 // Now that we have linearized the tree to a list and have gathered all of
2188 // the operands and their ranks, sort the operands by their rank. Use a
2189 // stable_sort so that values with equal ranks will have their relative
2190 // positions maintained (and so the compiler is deterministic). Note that
2191 // this sorts so that the highest ranking values end up at the beginning of
2193 std::stable_sort(Ops
.begin(), Ops
.end());
2195 // OptimizeExpression - Now that we have the expression tree in a convenient
2196 // sorted form, optimize it globally if possible.
2197 if (Value
*V
= OptimizeExpression(I
, Ops
)) {
2199 // Self-referential expression in unreachable code.
2201 // This expression tree simplified to something that isn't a tree,
2203 DEBUG(dbgs() << "Reassoc to scalar: " << *V
<< '\n');
2204 I
->replaceAllUsesWith(V
);
2205 if (Instruction
*VI
= dyn_cast
<Instruction
>(V
))
2206 VI
->setDebugLoc(I
->getDebugLoc());
2207 RedoInsts
.insert(I
);
2212 // We want to sink immediates as deeply as possible except in the case where
2213 // this is a multiply tree used only by an add, and the immediate is a -1.
2214 // In this case we reassociate to put the negation on the outside so that we
2215 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2216 if (I
->hasOneUse()) {
2217 if (I
->getOpcode() == Instruction::Mul
&&
2218 cast
<Instruction
>(I
->user_back())->getOpcode() == Instruction::Add
&&
2219 isa
<ConstantInt
>(Ops
.back().Op
) &&
2220 cast
<ConstantInt
>(Ops
.back().Op
)->isAllOnesValue()) {
2221 ValueEntry Tmp
= Ops
.pop_back_val();
2222 Ops
.insert(Ops
.begin(), Tmp
);
2223 } else if (I
->getOpcode() == Instruction::FMul
&&
2224 cast
<Instruction
>(I
->user_back())->getOpcode() ==
2225 Instruction::FAdd
&&
2226 isa
<ConstantFP
>(Ops
.back().Op
) &&
2227 cast
<ConstantFP
>(Ops
.back().Op
)->isExactlyValue(-1.0)) {
2228 ValueEntry Tmp
= Ops
.pop_back_val();
2229 Ops
.insert(Ops
.begin(), Tmp
);
2233 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I
, Ops
); dbgs() << '\n');
2235 if (Ops
.size() == 1) {
2237 // Self-referential expression in unreachable code.
2240 // This expression tree simplified to something that isn't a tree,
2242 I
->replaceAllUsesWith(Ops
[0].Op
);
2243 if (Instruction
*OI
= dyn_cast
<Instruction
>(Ops
[0].Op
))
2244 OI
->setDebugLoc(I
->getDebugLoc());
2245 RedoInsts
.insert(I
);
2249 // Now that we ordered and optimized the expressions, splat them back into
2250 // the expression tree, removing any unneeded nodes.
2251 RewriteExprTree(I
, Ops
);
2254 bool Reassociate::runOnFunction(Function
&F
) {
2255 if (skipOptnoneFunction(F
))
2258 // Calculate the rank map for F
2262 for (Function::iterator BI
= F
.begin(), BE
= F
.end(); BI
!= BE
; ++BI
) {
2263 // Optimize every instruction in the basic block.
2264 for (BasicBlock::iterator II
= BI
->begin(), IE
= BI
->end(); II
!= IE
; )
2265 if (isInstructionTriviallyDead(II
)) {
2269 assert(II
->getParent() == BI
&& "Moved to a different block!");
2273 // If this produced extra instructions to optimize, handle them now.
2274 while (!RedoInsts
.empty()) {
2275 Instruction
*I
= RedoInsts
.pop_back_val();
2276 if (isInstructionTriviallyDead(I
))
2283 // We are done with the rank map.
2285 ValueRankMap
.clear();