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
23 #define DEBUG_TYPE "reassociate"
24 #include "llvm/Transforms/Scalar.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/PostOrderIterator.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Assembly/Writer.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ValueHandle.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include <algorithm>
44 using namespace llvm;
45
46 STATISTIC(NumChanged, "Number of insts reassociated");
47 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
48 STATISTIC(NumFactor , "Number of multiplies factored");
49
50 namespace {
51 struct ValueEntry {
52 unsigned Rank;
53 Value *Op;
ValueEntry__anon81b135320111::ValueEntry54 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
55 };
operator <(const ValueEntry & LHS,const ValueEntry & RHS)56 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
57 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
58 }
59 }
60
61 #ifndef NDEBUG
62 /// PrintOps - Print out the expression identified in the Ops list.
63 ///
PrintOps(Instruction * I,const SmallVectorImpl<ValueEntry> & Ops)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) {
69 dbgs() << "[ ";
70 WriteAsOperand(dbgs(), Ops[i].Op, false, M);
71 dbgs() << ", #" << Ops[i].Rank << "] ";
72 }
73 }
74 #endif
75
76 namespace {
77 /// \brief Utility class representing a base and exponent pair which form one
78 /// factor of some product.
79 struct Factor {
80 Value *Base;
81 unsigned Power;
82
Factor__anon81b135320211::Factor83 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
84
85 /// \brief Sort factors by their Base.
86 struct BaseSorter {
operator ()__anon81b135320211::Factor::BaseSorter87 bool operator()(const Factor &LHS, const Factor &RHS) {
88 return LHS.Base < RHS.Base;
89 }
90 };
91
92 /// \brief Compare factors for equal bases.
93 struct BaseEqual {
operator ()__anon81b135320211::Factor::BaseEqual94 bool operator()(const Factor &LHS, const Factor &RHS) {
95 return LHS.Base == RHS.Base;
96 }
97 };
98
99 /// \brief Sort factors in descending order by their power.
100 struct PowerDescendingSorter {
operator ()__anon81b135320211::Factor::PowerDescendingSorter101 bool operator()(const Factor &LHS, const Factor &RHS) {
102 return LHS.Power > RHS.Power;
103 }
104 };
105
106 /// \brief Compare factors for equal powers.
107 struct PowerEqual {
operator ()__anon81b135320211::Factor::PowerEqual108 bool operator()(const Factor &LHS, const Factor &RHS) {
109 return LHS.Power == RHS.Power;
110 }
111 };
112 };
113
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
117 /// C2)
118 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
119 /// constant.
120 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
121 /// operand as "E | 0"
122 class XorOpnd {
123 public:
124 XorOpnd(Value *V);
125
isInvalid() const126 bool isInvalid() const { return SymbolicPart == 0; }
isOrExpr() const127 bool isOrExpr() const { return isOr; }
getValue() const128 Value *getValue() const { return OrigVal; }
getSymbolicPart() const129 Value *getSymbolicPart() const { return SymbolicPart; }
getSymbolicRank() const130 unsigned getSymbolicRank() const { return SymbolicRank; }
getConstPart() const131 const APInt &getConstPart() const { return ConstPart; }
132
Invalidate()133 void Invalidate() { SymbolicPart = OrigVal = 0; }
setSymbolicRank(unsigned R)134 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
135
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 {
operator ()__anon81b135320211::XorOpnd::PtrSortFunctor146 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) {
147 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
148 }
149 };
150 private:
151 Value *OrigVal;
152 Value *SymbolicPart;
153 APInt ConstPart;
154 unsigned SymbolicRank;
155 bool isOr;
156 };
157 }
158
159 namespace {
160 class Reassociate : public FunctionPass {
161 DenseMap<BasicBlock*, unsigned> RankMap;
162 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
163 SetVector<AssertingVH<Instruction> > RedoInsts;
164 bool MadeChange;
165 public:
166 static char ID; // Pass identification, replacement for typeid
Reassociate()167 Reassociate() : FunctionPass(ID) {
168 initializeReassociatePass(*PassRegistry::getPassRegistry());
169 }
170
171 bool runOnFunction(Function &F);
172
getAnalysisUsage(AnalysisUsage & AU) const173 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
174 AU.setPreservesCFG();
175 }
176 private:
177 void BuildRankMap(Function &F);
178 unsigned getRank(Value *V);
179 void ReassociateExpression(BinaryOperator *I);
180 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
181 Value *OptimizeExpression(BinaryOperator *I,
182 SmallVectorImpl<ValueEntry> &Ops);
183 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
184 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
185 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd,
186 Value *&Res);
187 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
188 APInt &ConstOpnd, Value *&Res);
189 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
190 SmallVectorImpl<Factor> &Factors);
191 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
192 SmallVectorImpl<Factor> &Factors);
193 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
194 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
195 void EraseInst(Instruction *I);
196 void OptimizeInst(Instruction *I);
197 };
198 }
199
XorOpnd(Value * V)200 XorOpnd::XorOpnd(Value *V) {
201 assert(!isa<ConstantInt>(V) && "No ConstantInt");
202 OrigVal = V;
203 Instruction *I = dyn_cast<Instruction>(V);
204 SymbolicRank = 0;
205
206 if (I && (I->getOpcode() == Instruction::Or ||
207 I->getOpcode() == Instruction::And)) {
208 Value *V0 = I->getOperand(0);
209 Value *V1 = I->getOperand(1);
210 if (isa<ConstantInt>(V0))
211 std::swap(V0, V1);
212
213 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
214 ConstPart = C->getValue();
215 SymbolicPart = V0;
216 isOr = (I->getOpcode() == Instruction::Or);
217 return;
218 }
219 }
220
221 // view the operand as "V | 0"
222 SymbolicPart = V;
223 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
224 isOr = true;
225 }
226
227 char Reassociate::ID = 0;
228 INITIALIZE_PASS(Reassociate, "reassociate",
229 "Reassociate expressions", false, false)
230
231 // Public interface to the Reassociate pass
createReassociatePass()232 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
233
234 /// isReassociableOp - Return true if V is an instruction of the specified
235 /// opcode and if it only has one use.
isReassociableOp(Value * V,unsigned Opcode)236 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
237 if (V->hasOneUse() && isa<Instruction>(V) &&
238 cast<Instruction>(V)->getOpcode() == Opcode)
239 return cast<BinaryOperator>(V);
240 return 0;
241 }
242
isUnmovableInstruction(Instruction * I)243 static bool isUnmovableInstruction(Instruction *I) {
244 switch (I->getOpcode()) {
245 case Instruction::PHI:
246 case Instruction::LandingPad:
247 case Instruction::Alloca:
248 case Instruction::Load:
249 case Instruction::Invoke:
250 case Instruction::UDiv:
251 case Instruction::SDiv:
252 case Instruction::FDiv:
253 case Instruction::URem:
254 case Instruction::SRem:
255 case Instruction::FRem:
256 return true;
257 case Instruction::Call:
258 return !isa<DbgInfoIntrinsic>(I);
259 default:
260 return false;
261 }
262 }
263
BuildRankMap(Function & F)264 void Reassociate::BuildRankMap(Function &F) {
265 unsigned i = 2;
266
267 // Assign distinct ranks to function arguments
268 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
269 ValueRankMap[&*I] = ++i;
270
271 ReversePostOrderTraversal<Function*> RPOT(&F);
272 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
273 E = RPOT.end(); I != E; ++I) {
274 BasicBlock *BB = *I;
275 unsigned BBRank = RankMap[BB] = ++i << 16;
276
277 // Walk the basic block, adding precomputed ranks for any instructions that
278 // we cannot move. This ensures that the ranks for these instructions are
279 // all different in the block.
280 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
281 if (isUnmovableInstruction(I))
282 ValueRankMap[&*I] = ++BBRank;
283 }
284 }
285
getRank(Value * V)286 unsigned Reassociate::getRank(Value *V) {
287 Instruction *I = dyn_cast<Instruction>(V);
288 if (I == 0) {
289 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
290 return 0; // Otherwise it's a global or constant, rank 0.
291 }
292
293 if (unsigned Rank = ValueRankMap[I])
294 return Rank; // Rank already known?
295
296 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
297 // we can reassociate expressions for code motion! Since we do not recurse
298 // for PHI nodes, we cannot have infinite recursion here, because there
299 // cannot be loops in the value graph that do not go through PHI nodes.
300 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
301 for (unsigned i = 0, e = I->getNumOperands();
302 i != e && Rank != MaxRank; ++i)
303 Rank = std::max(Rank, getRank(I->getOperand(i)));
304
305 // If this is a not or neg instruction, do not count it for rank. This
306 // assures us that X and ~X will have the same rank.
307 if (!I->getType()->isIntegerTy() ||
308 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
309 ++Rank;
310
311 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
312 // << Rank << "\n");
313
314 return ValueRankMap[I] = Rank;
315 }
316
317 /// LowerNegateToMultiply - Replace 0-X with X*-1.
318 ///
LowerNegateToMultiply(Instruction * Neg)319 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
320 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
321
322 BinaryOperator *Res =
323 BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
324 Neg->setOperand(1, Constant::getNullValue(Neg->getType())); // Drop use of op.
325 Res->takeName(Neg);
326 Neg->replaceAllUsesWith(Res);
327 Res->setDebugLoc(Neg->getDebugLoc());
328 return Res;
329 }
330
331 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
332 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
333 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
334 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
335 /// even x in Bitwidth-bit arithmetic.
CarmichaelShift(unsigned Bitwidth)336 static unsigned CarmichaelShift(unsigned Bitwidth) {
337 if (Bitwidth < 3)
338 return Bitwidth - 1;
339 return Bitwidth - 2;
340 }
341
342 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
343 /// reducing the combined weight using any special properties of the operation.
344 /// The existing weight LHS represents the computation X op X op ... op X where
345 /// X occurs LHS times. The combined weight represents X op X op ... op X with
346 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
347 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
348 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
IncorporateWeight(APInt & LHS,const APInt & RHS,unsigned Opcode)349 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
350 // If we were working with infinite precision arithmetic then the combined
351 // weight would be LHS + RHS. But we are using finite precision arithmetic,
352 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
353 // for nilpotent operations and addition, but not for idempotent operations
354 // and multiplication), so it is important to correctly reduce the combined
355 // weight back into range if wrapping would be wrong.
356
357 // If RHS is zero then the weight didn't change.
358 if (RHS.isMinValue())
359 return;
360 // If LHS is zero then the combined weight is RHS.
361 if (LHS.isMinValue()) {
362 LHS = RHS;
363 return;
364 }
365 // From this point on we know that neither LHS nor RHS is zero.
366
367 if (Instruction::isIdempotent(Opcode)) {
368 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
369 // weight of 1. Keeping weights at zero or one also means that wrapping is
370 // not a problem.
371 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
372 return; // Return a weight of 1.
373 }
374 if (Instruction::isNilpotent(Opcode)) {
375 // Nilpotent means X op X === 0, so reduce weights modulo 2.
376 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
377 LHS = 0; // 1 + 1 === 0 modulo 2.
378 return;
379 }
380 if (Opcode == Instruction::Add) {
381 // TODO: Reduce the weight by exploiting nsw/nuw?
382 LHS += RHS;
383 return;
384 }
385
386 assert(Opcode == Instruction::Mul && "Unknown associative operation!");
387 unsigned Bitwidth = LHS.getBitWidth();
388 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
389 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
390 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
391 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
392 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
393 // which by a happy accident means that they can always be represented using
394 // Bitwidth bits.
395 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
396 // the Carmichael number).
397 if (Bitwidth > 3) {
398 /// CM - The value of Carmichael's lambda function.
399 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
400 // Any weight W >= Threshold can be replaced with W - CM.
401 APInt Threshold = CM + Bitwidth;
402 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
403 // For Bitwidth 4 or more the following sum does not overflow.
404 LHS += RHS;
405 while (LHS.uge(Threshold))
406 LHS -= CM;
407 } else {
408 // To avoid problems with overflow do everything the same as above but using
409 // a larger type.
410 unsigned CM = 1U << CarmichaelShift(Bitwidth);
411 unsigned Threshold = CM + Bitwidth;
412 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
413 "Weights not reduced!");
414 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
415 while (Total >= Threshold)
416 Total -= CM;
417 LHS = Total;
418 }
419 }
420
421 typedef std::pair<Value*, APInt> RepeatedValue;
422
423 /// LinearizeExprTree - Given an associative binary expression, return the leaf
424 /// nodes in Ops along with their weights (how many times the leaf occurs). The
425 /// original expression is the same as
426 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
427 /// op
428 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
429 /// op
430 /// ...
431 /// op
432 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
433 ///
434 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
435 ///
436 /// This routine may modify the function, in which case it returns 'true'. The
437 /// changes it makes may well be destructive, changing the value computed by 'I'
438 /// to something completely different. Thus if the routine returns 'true' then
439 /// you MUST either replace I with a new expression computed from the Ops array,
440 /// or use RewriteExprTree to put the values back in.
441 ///
442 /// A leaf node is either not a binary operation of the same kind as the root
443 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
444 /// opcode), or is the same kind of binary operator but has a use which either
445 /// does not belong to the expression, or does belong to the expression but is
446 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
447 /// of the expression, while for non-leaf nodes (except for the root 'I') every
448 /// use is a non-leaf node of the expression.
449 ///
450 /// For example:
451 /// expression graph node names
452 ///
453 /// + | I
454 /// / \ |
455 /// + + | A, B
456 /// / \ / \ |
457 /// * + * | C, D, E
458 /// / \ / \ / \ |
459 /// + * | F, G
460 ///
461 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
462 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
463 ///
464 /// The expression is maximal: if some instruction is a binary operator of the
465 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
466 /// then the instruction also belongs to the expression, is not a leaf node of
467 /// it, and its operands also belong to the expression (but may be leaf nodes).
468 ///
469 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
470 /// order to ensure that every non-root node in the expression has *exactly one*
471 /// use by a non-leaf node of the expression. This destruction means that the
472 /// caller MUST either replace 'I' with a new expression or use something like
473 /// RewriteExprTree to put the values back in if the routine indicates that it
474 /// made a change by returning 'true'.
475 ///
476 /// In the above example either the right operand of A or the left operand of B
477 /// will be replaced by undef. If it is B's operand then this gives:
478 ///
479 /// + | I
480 /// / \ |
481 /// + + | A, B - operand of B replaced with undef
482 /// / \ \ |
483 /// * + * | C, D, E
484 /// / \ / \ / \ |
485 /// + * | F, G
486 ///
487 /// Note that such undef operands can only be reached by passing through 'I'.
488 /// For example, if you visit operands recursively starting from a leaf node
489 /// then you will never see such an undef operand unless you get back to 'I',
490 /// which requires passing through a phi node.
491 ///
492 /// Note that this routine may also mutate binary operators of the wrong type
493 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
494 /// of the expression) if it can turn them into binary operators of the right
495 /// type and thus make the expression bigger.
496
LinearizeExprTree(BinaryOperator * I,SmallVectorImpl<RepeatedValue> & Ops)497 static bool LinearizeExprTree(BinaryOperator *I,
498 SmallVectorImpl<RepeatedValue> &Ops) {
499 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
500 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
501 unsigned Opcode = I->getOpcode();
502 assert(Instruction::isAssociative(Opcode) &&
503 Instruction::isCommutative(Opcode) &&
504 "Expected an associative and commutative operation!");
505
506 // Visit all operands of the expression, keeping track of their weight (the
507 // number of paths from the expression root to the operand, or if you like
508 // the number of times that operand occurs in the linearized expression).
509 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
510 // while A has weight two.
511
512 // Worklist of non-leaf nodes (their operands are in the expression too) along
513 // with their weights, representing a certain number of paths to the operator.
514 // If an operator occurs in the worklist multiple times then we found multiple
515 // ways to get to it.
516 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
517 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
518 bool MadeChange = false;
519
520 // Leaves of the expression are values that either aren't the right kind of
521 // operation (eg: a constant, or a multiply in an add tree), or are, but have
522 // some uses that are not inside the expression. For example, in I = X + X,
523 // X = A + B, the value X has two uses (by I) that are in the expression. If
524 // X has any other uses, for example in a return instruction, then we consider
525 // X to be a leaf, and won't analyze it further. When we first visit a value,
526 // if it has more than one use then at first we conservatively consider it to
527 // be a leaf. Later, as the expression is explored, we may discover some more
528 // uses of the value from inside the expression. If all uses turn out to be
529 // from within the expression (and the value is a binary operator of the right
530 // kind) then the value is no longer considered to be a leaf, and its operands
531 // are explored.
532
533 // Leaves - Keeps track of the set of putative leaves as well as the number of
534 // paths to each leaf seen so far.
535 typedef DenseMap<Value*, APInt> LeafMap;
536 LeafMap Leaves; // Leaf -> Total weight so far.
537 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
538
539 #ifndef NDEBUG
540 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
541 #endif
542 while (!Worklist.empty()) {
543 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
544 I = P.first; // We examine the operands of this binary operator.
545
546 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
547 Value *Op = I->getOperand(OpIdx);
548 APInt Weight = P.second; // Number of paths to this operand.
549 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
550 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
551
552 // If this is a binary operation of the right kind with only one use then
553 // add its operands to the expression.
554 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
555 assert(Visited.insert(Op) && "Not first visit!");
556 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
557 Worklist.push_back(std::make_pair(BO, Weight));
558 continue;
559 }
560
561 // Appears to be a leaf. Is the operand already in the set of leaves?
562 LeafMap::iterator It = Leaves.find(Op);
563 if (It == Leaves.end()) {
564 // Not in the leaf map. Must be the first time we saw this operand.
565 assert(Visited.insert(Op) && "Not first visit!");
566 if (!Op->hasOneUse()) {
567 // This value has uses not accounted for by the expression, so it is
568 // not safe to modify. Mark it as being a leaf.
569 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
570 LeafOrder.push_back(Op);
571 Leaves[Op] = Weight;
572 continue;
573 }
574 // No uses outside the expression, try morphing it.
575 } else if (It != Leaves.end()) {
576 // Already in the leaf map.
577 assert(Visited.count(Op) && "In leaf map but not visited!");
578
579 // Update the number of paths to the leaf.
580 IncorporateWeight(It->second, Weight, Opcode);
581
582 #if 0 // TODO: Re-enable once PR13021 is fixed.
583 // The leaf already has one use from inside the expression. As we want
584 // exactly one such use, drop this new use of the leaf.
585 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
586 I->setOperand(OpIdx, UndefValue::get(I->getType()));
587 MadeChange = true;
588
589 // If the leaf is a binary operation of the right kind and we now see
590 // that its multiple original uses were in fact all by nodes belonging
591 // to the expression, then no longer consider it to be a leaf and add
592 // its operands to the expression.
593 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
594 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
595 Worklist.push_back(std::make_pair(BO, It->second));
596 Leaves.erase(It);
597 continue;
598 }
599 #endif
600
601 // If we still have uses that are not accounted for by the expression
602 // then it is not safe to modify the value.
603 if (!Op->hasOneUse())
604 continue;
605
606 // No uses outside the expression, try morphing it.
607 Weight = It->second;
608 Leaves.erase(It); // Since the value may be morphed below.
609 }
610
611 // At this point we have a value which, first of all, is not a binary
612 // expression of the right kind, and secondly, is only used inside the
613 // expression. This means that it can safely be modified. See if we
614 // can usefully morph it into an expression of the right kind.
615 assert((!isa<Instruction>(Op) ||
616 cast<Instruction>(Op)->getOpcode() != Opcode) &&
617 "Should have been handled above!");
618 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
619
620 // If this is a multiply expression, turn any internal negations into
621 // multiplies by -1 so they can be reassociated.
622 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op);
623 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) {
624 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
625 BO = LowerNegateToMultiply(BO);
626 DEBUG(dbgs() << *BO << 'n');
627 Worklist.push_back(std::make_pair(BO, Weight));
628 MadeChange = true;
629 continue;
630 }
631
632 // Failed to morph into an expression of the right type. This really is
633 // a leaf.
634 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
635 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
636 LeafOrder.push_back(Op);
637 Leaves[Op] = Weight;
638 }
639 }
640
641 // The leaves, repeated according to their weights, represent the linearized
642 // form of the expression.
643 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
644 Value *V = LeafOrder[i];
645 LeafMap::iterator It = Leaves.find(V);
646 if (It == Leaves.end())
647 // Node initially thought to be a leaf wasn't.
648 continue;
649 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
650 APInt Weight = It->second;
651 if (Weight.isMinValue())
652 // Leaf already output or weight reduction eliminated it.
653 continue;
654 // Ensure the leaf is only output once.
655 It->second = 0;
656 Ops.push_back(std::make_pair(V, Weight));
657 }
658
659 // For nilpotent operations or addition there may be no operands, for example
660 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
661 // in both cases the weight reduces to 0 causing the value to be skipped.
662 if (Ops.empty()) {
663 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
664 assert(Identity && "Associative operation without identity!");
665 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
666 }
667
668 return MadeChange;
669 }
670
671 // RewriteExprTree - Now that the operands for this expression tree are
672 // linearized and optimized, emit them in-order.
RewriteExprTree(BinaryOperator * I,SmallVectorImpl<ValueEntry> & Ops)673 void Reassociate::RewriteExprTree(BinaryOperator *I,
674 SmallVectorImpl<ValueEntry> &Ops) {
675 assert(Ops.size() > 1 && "Single values should be used directly!");
676
677 // Since our optimizations should never increase the number of operations, the
678 // new expression can usually be written reusing the existing binary operators
679 // from the original expression tree, without creating any new instructions,
680 // though the rewritten expression may have a completely different topology.
681 // We take care to not change anything if the new expression will be the same
682 // as the original. If more than trivial changes (like commuting operands)
683 // were made then we are obliged to clear out any optional subclass data like
684 // nsw flags.
685
686 /// NodesToRewrite - Nodes from the original expression available for writing
687 /// the new expression into.
688 SmallVector<BinaryOperator*, 8> NodesToRewrite;
689 unsigned Opcode = I->getOpcode();
690 BinaryOperator *Op = I;
691
692 /// NotRewritable - The operands being written will be the leaves of the new
693 /// expression and must not be used as inner nodes (via NodesToRewrite) by
694 /// mistake. Inner nodes are always reassociable, and usually leaves are not
695 /// (if they were they would have been incorporated into the expression and so
696 /// would not be leaves), so most of the time there is no danger of this. But
697 /// in rare cases a leaf may become reassociable if an optimization kills uses
698 /// of it, or it may momentarily become reassociable during rewriting (below)
699 /// due it being removed as an operand of one of its uses. Ensure that misuse
700 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
701 /// leaves and refusing to reuse any of them as inner nodes.
702 SmallPtrSet<Value*, 8> NotRewritable;
703 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
704 NotRewritable.insert(Ops[i].Op);
705
706 // ExpressionChanged - Non-null if the rewritten expression differs from the
707 // original in some non-trivial way, requiring the clearing of optional flags.
708 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
709 BinaryOperator *ExpressionChanged = 0;
710 for (unsigned i = 0; ; ++i) {
711 // The last operation (which comes earliest in the IR) is special as both
712 // operands will come from Ops, rather than just one with the other being
713 // a subexpression.
714 if (i+2 == Ops.size()) {
715 Value *NewLHS = Ops[i].Op;
716 Value *NewRHS = Ops[i+1].Op;
717 Value *OldLHS = Op->getOperand(0);
718 Value *OldRHS = Op->getOperand(1);
719
720 if (NewLHS == OldLHS && NewRHS == OldRHS)
721 // Nothing changed, leave it alone.
722 break;
723
724 if (NewLHS == OldRHS && NewRHS == OldLHS) {
725 // The order of the operands was reversed. Swap them.
726 DEBUG(dbgs() << "RA: " << *Op << '\n');
727 Op->swapOperands();
728 DEBUG(dbgs() << "TO: " << *Op << '\n');
729 MadeChange = true;
730 ++NumChanged;
731 break;
732 }
733
734 // The new operation differs non-trivially from the original. Overwrite
735 // the old operands with the new ones.
736 DEBUG(dbgs() << "RA: " << *Op << '\n');
737 if (NewLHS != OldLHS) {
738 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
739 if (BO && !NotRewritable.count(BO))
740 NodesToRewrite.push_back(BO);
741 Op->setOperand(0, NewLHS);
742 }
743 if (NewRHS != OldRHS) {
744 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
745 if (BO && !NotRewritable.count(BO))
746 NodesToRewrite.push_back(BO);
747 Op->setOperand(1, NewRHS);
748 }
749 DEBUG(dbgs() << "TO: " << *Op << '\n');
750
751 ExpressionChanged = Op;
752 MadeChange = true;
753 ++NumChanged;
754
755 break;
756 }
757
758 // Not the last operation. The left-hand side will be a sub-expression
759 // while the right-hand side will be the current element of Ops.
760 Value *NewRHS = Ops[i].Op;
761 if (NewRHS != Op->getOperand(1)) {
762 DEBUG(dbgs() << "RA: " << *Op << '\n');
763 if (NewRHS == Op->getOperand(0)) {
764 // The new right-hand side was already present as the left operand. If
765 // we are lucky then swapping the operands will sort out both of them.
766 Op->swapOperands();
767 } else {
768 // Overwrite with the new right-hand side.
769 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
770 if (BO && !NotRewritable.count(BO))
771 NodesToRewrite.push_back(BO);
772 Op->setOperand(1, NewRHS);
773 ExpressionChanged = Op;
774 }
775 DEBUG(dbgs() << "TO: " << *Op << '\n');
776 MadeChange = true;
777 ++NumChanged;
778 }
779
780 // Now deal with the left-hand side. If this is already an operation node
781 // from the original expression then just rewrite the rest of the expression
782 // into it.
783 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
784 if (BO && !NotRewritable.count(BO)) {
785 Op = BO;
786 continue;
787 }
788
789 // Otherwise, grab a spare node from the original expression and use that as
790 // the left-hand side. If there are no nodes left then the optimizers made
791 // an expression with more nodes than the original! This usually means that
792 // they did something stupid but it might mean that the problem was just too
793 // hard (finding the mimimal number of multiplications needed to realize a
794 // multiplication expression is NP-complete). Whatever the reason, smart or
795 // stupid, create a new node if there are none left.
796 BinaryOperator *NewOp;
797 if (NodesToRewrite.empty()) {
798 Constant *Undef = UndefValue::get(I->getType());
799 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
800 Undef, Undef, "", I);
801 } else {
802 NewOp = NodesToRewrite.pop_back_val();
803 }
804
805 DEBUG(dbgs() << "RA: " << *Op << '\n');
806 Op->setOperand(0, NewOp);
807 DEBUG(dbgs() << "TO: " << *Op << '\n');
808 ExpressionChanged = Op;
809 MadeChange = true;
810 ++NumChanged;
811 Op = NewOp;
812 }
813
814 // If the expression changed non-trivially then clear out all subclass data
815 // starting from the operator specified in ExpressionChanged, and compactify
816 // the operators to just before the expression root to guarantee that the
817 // expression tree is dominated by all of Ops.
818 if (ExpressionChanged)
819 do {
820 ExpressionChanged->clearSubclassOptionalData();
821 if (ExpressionChanged == I)
822 break;
823 ExpressionChanged->moveBefore(I);
824 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->use_begin());
825 } while (1);
826
827 // Throw away any left over nodes from the original expression.
828 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
829 RedoInsts.insert(NodesToRewrite[i]);
830 }
831
832 /// NegateValue - Insert instructions before the instruction pointed to by BI,
833 /// that computes the negative version of the value specified. The negative
834 /// version of the value is returned, and BI is left pointing at the instruction
835 /// that should be processed next by the reassociation pass.
NegateValue(Value * V,Instruction * BI)836 static Value *NegateValue(Value *V, Instruction *BI) {
837 if (Constant *C = dyn_cast<Constant>(V))
838 return ConstantExpr::getNeg(C);
839
840 // We are trying to expose opportunity for reassociation. One of the things
841 // that we want to do to achieve this is to push a negation as deep into an
842 // expression chain as possible, to expose the add instructions. In practice,
843 // this means that we turn this:
844 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
845 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
846 // the constants. We assume that instcombine will clean up the mess later if
847 // we introduce tons of unnecessary negation instructions.
848 //
849 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) {
850 // Push the negates through the add.
851 I->setOperand(0, NegateValue(I->getOperand(0), BI));
852 I->setOperand(1, NegateValue(I->getOperand(1), BI));
853
854 // We must move the add instruction here, because the neg instructions do
855 // not dominate the old add instruction in general. By moving it, we are
856 // assured that the neg instructions we just inserted dominate the
857 // instruction we are about to insert after them.
858 //
859 I->moveBefore(BI);
860 I->setName(I->getName()+".neg");
861 return I;
862 }
863
864 // Okay, we need to materialize a negated version of V with an instruction.
865 // Scan the use lists of V to see if we have one already.
866 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
867 User *U = *UI;
868 if (!BinaryOperator::isNeg(U)) continue;
869
870 // We found one! Now we have to make sure that the definition dominates
871 // this use. We do this by moving it to the entry block (if it is a
872 // non-instruction value) or right after the definition. These negates will
873 // be zapped by reassociate later, so we don't need much finesse here.
874 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
875
876 // Verify that the negate is in this function, V might be a constant expr.
877 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
878 continue;
879
880 BasicBlock::iterator InsertPt;
881 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
882 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
883 InsertPt = II->getNormalDest()->begin();
884 } else {
885 InsertPt = InstInput;
886 ++InsertPt;
887 }
888 while (isa<PHINode>(InsertPt)) ++InsertPt;
889 } else {
890 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
891 }
892 TheNeg->moveBefore(InsertPt);
893 return TheNeg;
894 }
895
896 // Insert a 'neg' instruction that subtracts the value from zero to get the
897 // negation.
898 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
899 }
900
901 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
902 /// X-Y into (X + -Y).
ShouldBreakUpSubtract(Instruction * Sub)903 static bool ShouldBreakUpSubtract(Instruction *Sub) {
904 // If this is a negation, we can't split it up!
905 if (BinaryOperator::isNeg(Sub))
906 return false;
907
908 // Don't bother to break this up unless either the LHS is an associable add or
909 // subtract or if this is only used by one.
910 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
911 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
912 return true;
913 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
914 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
915 return true;
916 if (Sub->hasOneUse() &&
917 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
918 isReassociableOp(Sub->use_back(), Instruction::Sub)))
919 return true;
920
921 return false;
922 }
923
924 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
925 /// only used by an add, transform this into (X+(0-Y)) to promote better
926 /// reassociation.
BreakUpSubtract(Instruction * Sub)927 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
928 // Convert a subtract into an add and a neg instruction. This allows sub
929 // instructions to be commuted with other add instructions.
930 //
931 // Calculate the negative value of Operand 1 of the sub instruction,
932 // and set it as the RHS of the add instruction we just made.
933 //
934 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
935 BinaryOperator *New =
936 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
937 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
938 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
939 New->takeName(Sub);
940
941 // Everyone now refers to the add instruction.
942 Sub->replaceAllUsesWith(New);
943 New->setDebugLoc(Sub->getDebugLoc());
944
945 DEBUG(dbgs() << "Negated: " << *New << '\n');
946 return New;
947 }
948
949 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
950 /// by one, change this into a multiply by a constant to assist with further
951 /// reassociation.
ConvertShiftToMul(Instruction * Shl)952 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
953 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
954 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
955
956 BinaryOperator *Mul =
957 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
958 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
959 Mul->takeName(Shl);
960 Shl->replaceAllUsesWith(Mul);
961 Mul->setDebugLoc(Shl->getDebugLoc());
962 return Mul;
963 }
964
965 /// FindInOperandList - Scan backwards and forwards among values with the same
966 /// rank as element i to see if X exists. If X does not exist, return i. This
967 /// is useful when scanning for 'x' when we see '-x' because they both get the
968 /// same rank.
FindInOperandList(SmallVectorImpl<ValueEntry> & Ops,unsigned i,Value * X)969 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
970 Value *X) {
971 unsigned XRank = Ops[i].Rank;
972 unsigned e = Ops.size();
973 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
974 if (Ops[j].Op == X)
975 return j;
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 return i;
981 }
982
983 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
984 /// and returning the result. Insert the tree before I.
EmitAddTreeOfValues(Instruction * I,SmallVectorImpl<WeakVH> & Ops)985 static Value *EmitAddTreeOfValues(Instruction *I,
986 SmallVectorImpl<WeakVH> &Ops){
987 if (Ops.size() == 1) return Ops.back();
988
989 Value *V1 = Ops.back();
990 Ops.pop_back();
991 Value *V2 = EmitAddTreeOfValues(I, Ops);
992 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
993 }
994
995 /// RemoveFactorFromExpression - If V is an expression tree that is a
996 /// multiplication sequence, and if this sequence contains a multiply by Factor,
997 /// remove Factor from the tree and return the new tree.
RemoveFactorFromExpression(Value * V,Value * Factor)998 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
999 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
1000 if (!BO) return 0;
1001
1002 SmallVector<RepeatedValue, 8> Tree;
1003 MadeChange |= LinearizeExprTree(BO, Tree);
1004 SmallVector<ValueEntry, 8> Factors;
1005 Factors.reserve(Tree.size());
1006 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1007 RepeatedValue E = Tree[i];
1008 Factors.append(E.second.getZExtValue(),
1009 ValueEntry(getRank(E.first), E.first));
1010 }
1011
1012 bool FoundFactor = false;
1013 bool NeedsNegate = false;
1014 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1015 if (Factors[i].Op == Factor) {
1016 FoundFactor = true;
1017 Factors.erase(Factors.begin()+i);
1018 break;
1019 }
1020
1021 // If this is a negative version of this factor, remove it.
1022 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
1023 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1024 if (FC1->getValue() == -FC2->getValue()) {
1025 FoundFactor = NeedsNegate = true;
1026 Factors.erase(Factors.begin()+i);
1027 break;
1028 }
1029 }
1030
1031 if (!FoundFactor) {
1032 // Make sure to restore the operands to the expression tree.
1033 RewriteExprTree(BO, Factors);
1034 return 0;
1035 }
1036
1037 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1038
1039 // If this was just a single multiply, remove the multiply and return the only
1040 // remaining operand.
1041 if (Factors.size() == 1) {
1042 RedoInsts.insert(BO);
1043 V = Factors[0].Op;
1044 } else {
1045 RewriteExprTree(BO, Factors);
1046 V = BO;
1047 }
1048
1049 if (NeedsNegate)
1050 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
1051
1052 return V;
1053 }
1054
1055 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1056 /// add its operands as factors, otherwise add V to the list of factors.
1057 ///
1058 /// Ops is the top-level list of add operands we're trying to factor.
FindSingleUseMultiplyFactors(Value * V,SmallVectorImpl<Value * > & Factors,const SmallVectorImpl<ValueEntry> & Ops)1059 static void FindSingleUseMultiplyFactors(Value *V,
1060 SmallVectorImpl<Value*> &Factors,
1061 const SmallVectorImpl<ValueEntry> &Ops) {
1062 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
1063 if (!BO) {
1064 Factors.push_back(V);
1065 return;
1066 }
1067
1068 // Otherwise, add the LHS and RHS to the list of factors.
1069 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1070 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1071 }
1072
1073 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1074 /// instruction. This optimizes based on identities. If it can be reduced to
1075 /// a single Value, it is returned, otherwise the Ops list is mutated as
1076 /// necessary.
OptimizeAndOrXor(unsigned Opcode,SmallVectorImpl<ValueEntry> & Ops)1077 static Value *OptimizeAndOrXor(unsigned Opcode,
1078 SmallVectorImpl<ValueEntry> &Ops) {
1079 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1080 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1081 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1082 // First, check for X and ~X in the operand list.
1083 assert(i < Ops.size());
1084 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1085 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1086 unsigned FoundX = FindInOperandList(Ops, i, X);
1087 if (FoundX != i) {
1088 if (Opcode == Instruction::And) // ...&X&~X = 0
1089 return Constant::getNullValue(X->getType());
1090
1091 if (Opcode == Instruction::Or) // ...|X|~X = -1
1092 return Constant::getAllOnesValue(X->getType());
1093 }
1094 }
1095
1096 // Next, check for duplicate pairs of values, which we assume are next to
1097 // each other, due to our sorting criteria.
1098 assert(i < Ops.size());
1099 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1100 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1101 // Drop duplicate values for And and Or.
1102 Ops.erase(Ops.begin()+i);
1103 --i; --e;
1104 ++NumAnnihil;
1105 continue;
1106 }
1107
1108 // Drop pairs of values for Xor.
1109 assert(Opcode == Instruction::Xor);
1110 if (e == 2)
1111 return Constant::getNullValue(Ops[0].Op->getType());
1112
1113 // Y ^ X^X -> Y
1114 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1115 i -= 1; e -= 2;
1116 ++NumAnnihil;
1117 }
1118 }
1119 return 0;
1120 }
1121
1122 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1123 /// instruction with the given two operands, and return the resulting
1124 /// instruction. There are two special cases: 1) if the constant operand is 0,
1125 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1126 /// be returned.
createAndInstr(Instruction * InsertBefore,Value * Opnd,const APInt & ConstOpnd)1127 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1128 const APInt &ConstOpnd) {
1129 if (ConstOpnd != 0) {
1130 if (!ConstOpnd.isAllOnesValue()) {
1131 LLVMContext &Ctx = Opnd->getType()->getContext();
1132 Instruction *I;
1133 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1134 "and.ra", InsertBefore);
1135 I->setDebugLoc(InsertBefore->getDebugLoc());
1136 return I;
1137 }
1138 return Opnd;
1139 }
1140 return 0;
1141 }
1142
1143 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1144 // into "R ^ C", where C would be 0, and R is a symbolic value.
1145 //
1146 // If it was successful, true is returned, and the "R" and "C" is returned
1147 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1148 // and both "Res" and "ConstOpnd" remain unchanged.
1149 //
CombineXorOpnd(Instruction * I,XorOpnd * Opnd1,APInt & ConstOpnd,Value * & Res)1150 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1151 APInt &ConstOpnd, Value *&Res) {
1152 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1153 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1154 // = (x & ~c1) ^ (c1 ^ c2)
1155 // It is useful only when c1 == c2.
1156 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1157 if (!Opnd1->getValue()->hasOneUse())
1158 return false;
1159
1160 const APInt &C1 = Opnd1->getConstPart();
1161 if (C1 != ConstOpnd)
1162 return false;
1163
1164 Value *X = Opnd1->getSymbolicPart();
1165 Res = createAndInstr(I, X, ~C1);
1166 // ConstOpnd was C2, now C1 ^ C2.
1167 ConstOpnd ^= C1;
1168
1169 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1170 RedoInsts.insert(T);
1171 return true;
1172 }
1173 return false;
1174 }
1175
1176
1177 // Helper function of OptimizeXor(). It tries to simplify
1178 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1179 // symbolic value.
1180 //
1181 // If it was successful, true is returned, and the "R" and "C" is returned
1182 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1183 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1184 // returned, and both "Res" and "ConstOpnd" remain unchanged.
CombineXorOpnd(Instruction * I,XorOpnd * Opnd1,XorOpnd * Opnd2,APInt & ConstOpnd,Value * & Res)1185 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1186 APInt &ConstOpnd, Value *&Res) {
1187 Value *X = Opnd1->getSymbolicPart();
1188 if (X != Opnd2->getSymbolicPart())
1189 return false;
1190
1191 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1192 int DeadInstNum = 1;
1193 if (Opnd1->getValue()->hasOneUse())
1194 DeadInstNum++;
1195 if (Opnd2->getValue()->hasOneUse())
1196 DeadInstNum++;
1197
1198 // Xor-Rule 2:
1199 // (x | c1) ^ (x & c2)
1200 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1201 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1202 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1203 //
1204 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1205 if (Opnd2->isOrExpr())
1206 std::swap(Opnd1, Opnd2);
1207
1208 const APInt &C1 = Opnd1->getConstPart();
1209 const APInt &C2 = Opnd2->getConstPart();
1210 APInt C3((~C1) ^ C2);
1211
1212 // Do not increase code size!
1213 if (C3 != 0 && !C3.isAllOnesValue()) {
1214 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1215 if (NewInstNum > DeadInstNum)
1216 return false;
1217 }
1218
1219 Res = createAndInstr(I, X, C3);
1220 ConstOpnd ^= C1;
1221
1222 } else if (Opnd1->isOrExpr()) {
1223 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1224 //
1225 const APInt &C1 = Opnd1->getConstPart();
1226 const APInt &C2 = Opnd2->getConstPart();
1227 APInt C3 = C1 ^ C2;
1228
1229 // Do not increase code size
1230 if (C3 != 0 && !C3.isAllOnesValue()) {
1231 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1232 if (NewInstNum > DeadInstNum)
1233 return false;
1234 }
1235
1236 Res = createAndInstr(I, X, C3);
1237 ConstOpnd ^= C3;
1238 } else {
1239 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1240 //
1241 const APInt &C1 = Opnd1->getConstPart();
1242 const APInt &C2 = Opnd2->getConstPart();
1243 APInt C3 = C1 ^ C2;
1244 Res = createAndInstr(I, X, C3);
1245 }
1246
1247 // Put the original operands in the Redo list; hope they will be deleted
1248 // as dead code.
1249 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1250 RedoInsts.insert(T);
1251 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1252 RedoInsts.insert(T);
1253
1254 return true;
1255 }
1256
1257 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1258 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1259 /// necessary.
OptimizeXor(Instruction * I,SmallVectorImpl<ValueEntry> & Ops)1260 Value *Reassociate::OptimizeXor(Instruction *I,
1261 SmallVectorImpl<ValueEntry> &Ops) {
1262 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1263 return V;
1264
1265 if (Ops.size() == 1)
1266 return 0;
1267
1268 SmallVector<XorOpnd, 8> Opnds;
1269 SmallVector<XorOpnd*, 8> OpndPtrs;
1270 Type *Ty = Ops[0].Op->getType();
1271 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1272
1273 // Step 1: Convert ValueEntry to XorOpnd
1274 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1275 Value *V = Ops[i].Op;
1276 if (!isa<ConstantInt>(V)) {
1277 XorOpnd O(V);
1278 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1279 Opnds.push_back(O);
1280 } else
1281 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1282 }
1283
1284 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1285 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1286 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1287 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1288 // when new elements are added to the vector.
1289 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1290 OpndPtrs.push_back(&Opnds[i]);
1291
1292 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1293 // the same symbolic value cluster together. For instance, the input operand
1294 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1295 // ("x | 123", "x & 789", "y & 456").
1296 std::sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1297
1298 // Step 3: Combine adjacent operands
1299 XorOpnd *PrevOpnd = 0;
1300 bool Changed = false;
1301 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1302 XorOpnd *CurrOpnd = OpndPtrs[i];
1303 // The combined value
1304 Value *CV;
1305
1306 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1307 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1308 Changed = true;
1309 if (CV)
1310 *CurrOpnd = XorOpnd(CV);
1311 else {
1312 CurrOpnd->Invalidate();
1313 continue;
1314 }
1315 }
1316
1317 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1318 PrevOpnd = CurrOpnd;
1319 continue;
1320 }
1321
1322 // step 3.2: When previous and current operands share the same symbolic
1323 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1324 //
1325 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1326 // Remove previous operand
1327 PrevOpnd->Invalidate();
1328 if (CV) {
1329 *CurrOpnd = XorOpnd(CV);
1330 PrevOpnd = CurrOpnd;
1331 } else {
1332 CurrOpnd->Invalidate();
1333 PrevOpnd = 0;
1334 }
1335 Changed = true;
1336 }
1337 }
1338
1339 // Step 4: Reassemble the Ops
1340 if (Changed) {
1341 Ops.clear();
1342 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1343 XorOpnd &O = Opnds[i];
1344 if (O.isInvalid())
1345 continue;
1346 ValueEntry VE(getRank(O.getValue()), O.getValue());
1347 Ops.push_back(VE);
1348 }
1349 if (ConstOpnd != 0) {
1350 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1351 ValueEntry VE(getRank(C), C);
1352 Ops.push_back(VE);
1353 }
1354 int Sz = Ops.size();
1355 if (Sz == 1)
1356 return Ops.back().Op;
1357 else if (Sz == 0) {
1358 assert(ConstOpnd == 0);
1359 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1360 }
1361 }
1362
1363 return 0;
1364 }
1365
1366 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1367 /// optimizes based on identities. If it can be reduced to a single Value, it
1368 /// is returned, otherwise the Ops list is mutated as necessary.
OptimizeAdd(Instruction * I,SmallVectorImpl<ValueEntry> & Ops)1369 Value *Reassociate::OptimizeAdd(Instruction *I,
1370 SmallVectorImpl<ValueEntry> &Ops) {
1371 // Scan the operand lists looking for X and -X pairs. If we find any, we
1372 // can simplify the expression. X+-X == 0. While we're at it, scan for any
1373 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1374 //
1375 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
1376 //
1377 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1378 Value *TheOp = Ops[i].Op;
1379 // Check to see if we've seen this operand before. If so, we factor all
1380 // instances of the operand together. Due to our sorting criteria, we know
1381 // that these need to be next to each other in the vector.
1382 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1383 // Rescan the list, remove all instances of this operand from the expr.
1384 unsigned NumFound = 0;
1385 do {
1386 Ops.erase(Ops.begin()+i);
1387 ++NumFound;
1388 } while (i != Ops.size() && Ops[i].Op == TheOp);
1389
1390 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1391 ++NumFactor;
1392
1393 // Insert a new multiply.
1394 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
1395 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
1396
1397 // Now that we have inserted a multiply, optimize it. This allows us to
1398 // handle cases that require multiple factoring steps, such as this:
1399 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1400 RedoInsts.insert(cast<Instruction>(Mul));
1401
1402 // If every add operand was a duplicate, return the multiply.
1403 if (Ops.empty())
1404 return Mul;
1405
1406 // Otherwise, we had some input that didn't have the dupe, such as
1407 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1408 // things being added by this operation.
1409 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1410
1411 --i;
1412 e = Ops.size();
1413 continue;
1414 }
1415
1416 // Check for X and -X in the operand list.
1417 if (!BinaryOperator::isNeg(TheOp))
1418 continue;
1419
1420 Value *X = BinaryOperator::getNegArgument(TheOp);
1421 unsigned FoundX = FindInOperandList(Ops, i, X);
1422 if (FoundX == i)
1423 continue;
1424
1425 // Remove X and -X from the operand list.
1426 if (Ops.size() == 2)
1427 return Constant::getNullValue(X->getType());
1428
1429 Ops.erase(Ops.begin()+i);
1430 if (i < FoundX)
1431 --FoundX;
1432 else
1433 --i; // Need to back up an extra one.
1434 Ops.erase(Ops.begin()+FoundX);
1435 ++NumAnnihil;
1436 --i; // Revisit element.
1437 e -= 2; // Removed two elements.
1438 }
1439
1440 // Scan the operand list, checking to see if there are any common factors
1441 // between operands. Consider something like A*A+A*B*C+D. We would like to
1442 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1443 // To efficiently find this, we count the number of times a factor occurs
1444 // for any ADD operands that are MULs.
1445 DenseMap<Value*, unsigned> FactorOccurrences;
1446
1447 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1448 // where they are actually the same multiply.
1449 unsigned MaxOcc = 0;
1450 Value *MaxOccVal = 0;
1451 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1452 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1453 if (!BOp)
1454 continue;
1455
1456 // Compute all of the factors of this added value.
1457 SmallVector<Value*, 8> Factors;
1458 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1459 assert(Factors.size() > 1 && "Bad linearize!");
1460
1461 // Add one to FactorOccurrences for each unique factor in this op.
1462 SmallPtrSet<Value*, 8> Duplicates;
1463 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1464 Value *Factor = Factors[i];
1465 if (!Duplicates.insert(Factor)) continue;
1466
1467 unsigned Occ = ++FactorOccurrences[Factor];
1468 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1469
1470 // If Factor is a negative constant, add the negated value as a factor
1471 // because we can percolate the negate out. Watch for minint, which
1472 // cannot be positivified.
1473 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
1474 if (CI->isNegative() && !CI->isMinValue(true)) {
1475 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1476 assert(!Duplicates.count(Factor) &&
1477 "Shouldn't have two constant factors, missed a canonicalize");
1478
1479 unsigned Occ = ++FactorOccurrences[Factor];
1480 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1481 }
1482 }
1483 }
1484
1485 // If any factor occurred more than one time, we can pull it out.
1486 if (MaxOcc > 1) {
1487 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1488 ++NumFactor;
1489
1490 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1491 // this, we could otherwise run into situations where removing a factor
1492 // from an expression will drop a use of maxocc, and this can cause
1493 // RemoveFactorFromExpression on successive values to behave differently.
1494 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
1495 SmallVector<WeakVH, 4> NewMulOps;
1496 for (unsigned i = 0; i != Ops.size(); ++i) {
1497 // Only try to remove factors from expressions we're allowed to.
1498 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1499 if (!BOp)
1500 continue;
1501
1502 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1503 // The factorized operand may occur several times. Convert them all in
1504 // one fell swoop.
1505 for (unsigned j = Ops.size(); j != i;) {
1506 --j;
1507 if (Ops[j].Op == Ops[i].Op) {
1508 NewMulOps.push_back(V);
1509 Ops.erase(Ops.begin()+j);
1510 }
1511 }
1512 --i;
1513 }
1514 }
1515
1516 // No need for extra uses anymore.
1517 delete DummyInst;
1518
1519 unsigned NumAddedValues = NewMulOps.size();
1520 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1521
1522 // Now that we have inserted the add tree, optimize it. This allows us to
1523 // handle cases that require multiple factoring steps, such as this:
1524 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1525 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1526 (void)NumAddedValues;
1527 if (Instruction *VI = dyn_cast<Instruction>(V))
1528 RedoInsts.insert(VI);
1529
1530 // Create the multiply.
1531 Instruction *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
1532
1533 // Rerun associate on the multiply in case the inner expression turned into
1534 // a multiply. We want to make sure that we keep things in canonical form.
1535 RedoInsts.insert(V2);
1536
1537 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1538 // entire result expression is just the multiply "A*(B+C)".
1539 if (Ops.empty())
1540 return V2;
1541
1542 // Otherwise, we had some input that didn't have the factor, such as
1543 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1544 // things being added by this operation.
1545 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1546 }
1547
1548 return 0;
1549 }
1550
1551 namespace {
1552 /// \brief Predicate tests whether a ValueEntry's op is in a map.
1553 struct IsValueInMap {
1554 const DenseMap<Value *, unsigned> ⤅
1555
IsValueInMap__anon81b135320411::IsValueInMap1556 IsValueInMap(const DenseMap<Value *, unsigned> &Map) : Map(Map) {}
1557
operator ()__anon81b135320411::IsValueInMap1558 bool operator()(const ValueEntry &Entry) {
1559 return Map.find(Entry.Op) != Map.end();
1560 }
1561 };
1562 }
1563
1564 /// \brief Build up a vector of value/power pairs factoring a product.
1565 ///
1566 /// Given a series of multiplication operands, build a vector of factors and
1567 /// the powers each is raised to when forming the final product. Sort them in
1568 /// the order of descending power.
1569 ///
1570 /// (x*x) -> [(x, 2)]
1571 /// ((x*x)*x) -> [(x, 3)]
1572 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1573 ///
1574 /// \returns Whether any factors have a power greater than one.
collectMultiplyFactors(SmallVectorImpl<ValueEntry> & Ops,SmallVectorImpl<Factor> & Factors)1575 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1576 SmallVectorImpl<Factor> &Factors) {
1577 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1578 // Compute the sum of powers of simplifiable factors.
1579 unsigned FactorPowerSum = 0;
1580 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1581 Value *Op = Ops[Idx-1].Op;
1582
1583 // Count the number of occurrences of this value.
1584 unsigned Count = 1;
1585 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1586 ++Count;
1587 // Track for simplification all factors which occur 2 or more times.
1588 if (Count > 1)
1589 FactorPowerSum += Count;
1590 }
1591
1592 // We can only simplify factors if the sum of the powers of our simplifiable
1593 // factors is 4 or higher. When that is the case, we will *always* have
1594 // a simplification. This is an important invariant to prevent cyclicly
1595 // trying to simplify already minimal formations.
1596 if (FactorPowerSum < 4)
1597 return false;
1598
1599 // Now gather the simplifiable factors, removing them from Ops.
1600 FactorPowerSum = 0;
1601 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1602 Value *Op = Ops[Idx-1].Op;
1603
1604 // Count the number of occurrences of this value.
1605 unsigned Count = 1;
1606 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1607 ++Count;
1608 if (Count == 1)
1609 continue;
1610 // Move an even number of occurrences to Factors.
1611 Count &= ~1U;
1612 Idx -= Count;
1613 FactorPowerSum += Count;
1614 Factors.push_back(Factor(Op, Count));
1615 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1616 }
1617
1618 // None of the adjustments above should have reduced the sum of factor powers
1619 // below our mininum of '4'.
1620 assert(FactorPowerSum >= 4);
1621
1622 std::sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1623 return true;
1624 }
1625
1626 /// \brief Build a tree of multiplies, computing the product of Ops.
buildMultiplyTree(IRBuilder<> & Builder,SmallVectorImpl<Value * > & Ops)1627 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1628 SmallVectorImpl<Value*> &Ops) {
1629 if (Ops.size() == 1)
1630 return Ops.back();
1631
1632 Value *LHS = Ops.pop_back_val();
1633 do {
1634 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1635 } while (!Ops.empty());
1636
1637 return LHS;
1638 }
1639
1640 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1641 ///
1642 /// Given a vector of values raised to various powers, where no two values are
1643 /// equal and the powers are sorted in decreasing order, compute the minimal
1644 /// DAG of multiplies to compute the final product, and return that product
1645 /// value.
buildMinimalMultiplyDAG(IRBuilder<> & Builder,SmallVectorImpl<Factor> & Factors)1646 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1647 SmallVectorImpl<Factor> &Factors) {
1648 assert(Factors[0].Power);
1649 SmallVector<Value *, 4> OuterProduct;
1650 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1651 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1652 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1653 LastIdx = Idx;
1654 continue;
1655 }
1656
1657 // We want to multiply across all the factors with the same power so that
1658 // we can raise them to that power as a single entity. Build a mini tree
1659 // for that.
1660 SmallVector<Value *, 4> InnerProduct;
1661 InnerProduct.push_back(Factors[LastIdx].Base);
1662 do {
1663 InnerProduct.push_back(Factors[Idx].Base);
1664 ++Idx;
1665 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1666
1667 // Reset the base value of the first factor to the new expression tree.
1668 // We'll remove all the factors with the same power in a second pass.
1669 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1670 if (Instruction *MI = dyn_cast<Instruction>(M))
1671 RedoInsts.insert(MI);
1672
1673 LastIdx = Idx;
1674 }
1675 // Unique factors with equal powers -- we've folded them into the first one's
1676 // base.
1677 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1678 Factor::PowerEqual()),
1679 Factors.end());
1680
1681 // Iteratively collect the base of each factor with an add power into the
1682 // outer product, and halve each power in preparation for squaring the
1683 // expression.
1684 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1685 if (Factors[Idx].Power & 1)
1686 OuterProduct.push_back(Factors[Idx].Base);
1687 Factors[Idx].Power >>= 1;
1688 }
1689 if (Factors[0].Power) {
1690 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1691 OuterProduct.push_back(SquareRoot);
1692 OuterProduct.push_back(SquareRoot);
1693 }
1694 if (OuterProduct.size() == 1)
1695 return OuterProduct.front();
1696
1697 Value *V = buildMultiplyTree(Builder, OuterProduct);
1698 return V;
1699 }
1700
OptimizeMul(BinaryOperator * I,SmallVectorImpl<ValueEntry> & Ops)1701 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1702 SmallVectorImpl<ValueEntry> &Ops) {
1703 // We can only optimize the multiplies when there is a chain of more than
1704 // three, such that a balanced tree might require fewer total multiplies.
1705 if (Ops.size() < 4)
1706 return 0;
1707
1708 // Try to turn linear trees of multiplies without other uses of the
1709 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1710 // re-use.
1711 SmallVector<Factor, 4> Factors;
1712 if (!collectMultiplyFactors(Ops, Factors))
1713 return 0; // All distinct factors, so nothing left for us to do.
1714
1715 IRBuilder<> Builder(I);
1716 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1717 if (Ops.empty())
1718 return V;
1719
1720 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1721 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1722 return 0;
1723 }
1724
OptimizeExpression(BinaryOperator * I,SmallVectorImpl<ValueEntry> & Ops)1725 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1726 SmallVectorImpl<ValueEntry> &Ops) {
1727 // Now that we have the linearized expression tree, try to optimize it.
1728 // Start by folding any constants that we found.
1729 Constant *Cst = 0;
1730 unsigned Opcode = I->getOpcode();
1731 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1732 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1733 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1734 }
1735 // If there was nothing but constants then we are done.
1736 if (Ops.empty())
1737 return Cst;
1738
1739 // Put the combined constant back at the end of the operand list, except if
1740 // there is no point. For example, an add of 0 gets dropped here, while a
1741 // multiplication by zero turns the whole expression into zero.
1742 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1743 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1744 return Cst;
1745 Ops.push_back(ValueEntry(0, Cst));
1746 }
1747
1748 if (Ops.size() == 1) return Ops[0].Op;
1749
1750 // Handle destructive annihilation due to identities between elements in the
1751 // argument list here.
1752 unsigned NumOps = Ops.size();
1753 switch (Opcode) {
1754 default: break;
1755 case Instruction::And:
1756 case Instruction::Or:
1757 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1758 return Result;
1759 break;
1760
1761 case Instruction::Xor:
1762 if (Value *Result = OptimizeXor(I, Ops))
1763 return Result;
1764 break;
1765
1766 case Instruction::Add:
1767 if (Value *Result = OptimizeAdd(I, Ops))
1768 return Result;
1769 break;
1770
1771 case Instruction::Mul:
1772 if (Value *Result = OptimizeMul(I, Ops))
1773 return Result;
1774 break;
1775 }
1776
1777 if (Ops.size() != NumOps)
1778 return OptimizeExpression(I, Ops);
1779 return 0;
1780 }
1781
1782 /// EraseInst - Zap the given instruction, adding interesting operands to the
1783 /// work list.
EraseInst(Instruction * I)1784 void Reassociate::EraseInst(Instruction *I) {
1785 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1786 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1787 // Erase the dead instruction.
1788 ValueRankMap.erase(I);
1789 RedoInsts.remove(I);
1790 I->eraseFromParent();
1791 // Optimize its operands.
1792 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1793 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1794 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1795 // If this is a node in an expression tree, climb to the expression root
1796 // and add that since that's where optimization actually happens.
1797 unsigned Opcode = Op->getOpcode();
1798 while (Op->hasOneUse() && Op->use_back()->getOpcode() == Opcode &&
1799 Visited.insert(Op))
1800 Op = Op->use_back();
1801 RedoInsts.insert(Op);
1802 }
1803 }
1804
1805 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
1806 /// instructions is not allowed.
OptimizeInst(Instruction * I)1807 void Reassociate::OptimizeInst(Instruction *I) {
1808 // Only consider operations that we understand.
1809 if (!isa<BinaryOperator>(I))
1810 return;
1811
1812 if (I->getOpcode() == Instruction::Shl &&
1813 isa<ConstantInt>(I->getOperand(1)))
1814 // If an operand of this shift is a reassociable multiply, or if the shift
1815 // is used by a reassociable multiply or add, turn into a multiply.
1816 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1817 (I->hasOneUse() &&
1818 (isReassociableOp(I->use_back(), Instruction::Mul) ||
1819 isReassociableOp(I->use_back(), Instruction::Add)))) {
1820 Instruction *NI = ConvertShiftToMul(I);
1821 RedoInsts.insert(I);
1822 MadeChange = true;
1823 I = NI;
1824 }
1825
1826 // Floating point binary operators are not associative, but we can still
1827 // commute (some) of them, to canonicalize the order of their operands.
1828 // This can potentially expose more CSE opportunities, and makes writing
1829 // other transformations simpler.
1830 if ((I->getType()->isFloatingPointTy() || I->getType()->isVectorTy())) {
1831 // FAdd and FMul can be commuted.
1832 if (I->getOpcode() != Instruction::FMul &&
1833 I->getOpcode() != Instruction::FAdd)
1834 return;
1835
1836 Value *LHS = I->getOperand(0);
1837 Value *RHS = I->getOperand(1);
1838 unsigned LHSRank = getRank(LHS);
1839 unsigned RHSRank = getRank(RHS);
1840
1841 // Sort the operands by rank.
1842 if (RHSRank < LHSRank) {
1843 I->setOperand(0, RHS);
1844 I->setOperand(1, LHS);
1845 }
1846
1847 return;
1848 }
1849
1850 // Do not reassociate boolean (i1) expressions. We want to preserve the
1851 // original order of evaluation for short-circuited comparisons that
1852 // SimplifyCFG has folded to AND/OR expressions. If the expression
1853 // is not further optimized, it is likely to be transformed back to a
1854 // short-circuited form for code gen, and the source order may have been
1855 // optimized for the most likely conditions.
1856 if (I->getType()->isIntegerTy(1))
1857 return;
1858
1859 // If this is a subtract instruction which is not already in negate form,
1860 // see if we can convert it to X+-Y.
1861 if (I->getOpcode() == Instruction::Sub) {
1862 if (ShouldBreakUpSubtract(I)) {
1863 Instruction *NI = BreakUpSubtract(I);
1864 RedoInsts.insert(I);
1865 MadeChange = true;
1866 I = NI;
1867 } else if (BinaryOperator::isNeg(I)) {
1868 // Otherwise, this is a negation. See if the operand is a multiply tree
1869 // and if this is not an inner node of a multiply tree.
1870 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
1871 (!I->hasOneUse() ||
1872 !isReassociableOp(I->use_back(), Instruction::Mul))) {
1873 Instruction *NI = LowerNegateToMultiply(I);
1874 RedoInsts.insert(I);
1875 MadeChange = true;
1876 I = NI;
1877 }
1878 }
1879 }
1880
1881 // If this instruction is an associative binary operator, process it.
1882 if (!I->isAssociative()) return;
1883 BinaryOperator *BO = cast<BinaryOperator>(I);
1884
1885 // If this is an interior node of a reassociable tree, ignore it until we
1886 // get to the root of the tree, to avoid N^2 analysis.
1887 unsigned Opcode = BO->getOpcode();
1888 if (BO->hasOneUse() && BO->use_back()->getOpcode() == Opcode)
1889 return;
1890
1891 // If this is an add tree that is used by a sub instruction, ignore it
1892 // until we process the subtract.
1893 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
1894 cast<Instruction>(BO->use_back())->getOpcode() == Instruction::Sub)
1895 return;
1896
1897 ReassociateExpression(BO);
1898 }
1899
ReassociateExpression(BinaryOperator * I)1900 void Reassociate::ReassociateExpression(BinaryOperator *I) {
1901
1902 // First, walk the expression tree, linearizing the tree, collecting the
1903 // operand information.
1904 SmallVector<RepeatedValue, 8> Tree;
1905 MadeChange |= LinearizeExprTree(I, Tree);
1906 SmallVector<ValueEntry, 8> Ops;
1907 Ops.reserve(Tree.size());
1908 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1909 RepeatedValue E = Tree[i];
1910 Ops.append(E.second.getZExtValue(),
1911 ValueEntry(getRank(E.first), E.first));
1912 }
1913
1914 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1915
1916 // Now that we have linearized the tree to a list and have gathered all of
1917 // the operands and their ranks, sort the operands by their rank. Use a
1918 // stable_sort so that values with equal ranks will have their relative
1919 // positions maintained (and so the compiler is deterministic). Note that
1920 // this sorts so that the highest ranking values end up at the beginning of
1921 // the vector.
1922 std::stable_sort(Ops.begin(), Ops.end());
1923
1924 // OptimizeExpression - Now that we have the expression tree in a convenient
1925 // sorted form, optimize it globally if possible.
1926 if (Value *V = OptimizeExpression(I, Ops)) {
1927 if (V == I)
1928 // Self-referential expression in unreachable code.
1929 return;
1930 // This expression tree simplified to something that isn't a tree,
1931 // eliminate it.
1932 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1933 I->replaceAllUsesWith(V);
1934 if (Instruction *VI = dyn_cast<Instruction>(V))
1935 VI->setDebugLoc(I->getDebugLoc());
1936 RedoInsts.insert(I);
1937 ++NumAnnihil;
1938 return;
1939 }
1940
1941 // We want to sink immediates as deeply as possible except in the case where
1942 // this is a multiply tree used only by an add, and the immediate is a -1.
1943 // In this case we reassociate to put the negation on the outside so that we
1944 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1945 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1946 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1947 isa<ConstantInt>(Ops.back().Op) &&
1948 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1949 ValueEntry Tmp = Ops.pop_back_val();
1950 Ops.insert(Ops.begin(), Tmp);
1951 }
1952
1953 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1954
1955 if (Ops.size() == 1) {
1956 if (Ops[0].Op == I)
1957 // Self-referential expression in unreachable code.
1958 return;
1959
1960 // This expression tree simplified to something that isn't a tree,
1961 // eliminate it.
1962 I->replaceAllUsesWith(Ops[0].Op);
1963 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
1964 OI->setDebugLoc(I->getDebugLoc());
1965 RedoInsts.insert(I);
1966 return;
1967 }
1968
1969 // Now that we ordered and optimized the expressions, splat them back into
1970 // the expression tree, removing any unneeded nodes.
1971 RewriteExprTree(I, Ops);
1972 }
1973
runOnFunction(Function & F)1974 bool Reassociate::runOnFunction(Function &F) {
1975 // Calculate the rank map for F
1976 BuildRankMap(F);
1977
1978 MadeChange = false;
1979 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
1980 // Optimize every instruction in the basic block.
1981 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
1982 if (isInstructionTriviallyDead(II)) {
1983 EraseInst(II++);
1984 } else {
1985 OptimizeInst(II);
1986 assert(II->getParent() == BI && "Moved to a different block!");
1987 ++II;
1988 }
1989
1990 // If this produced extra instructions to optimize, handle them now.
1991 while (!RedoInsts.empty()) {
1992 Instruction *I = RedoInsts.pop_back_val();
1993 if (isInstructionTriviallyDead(I))
1994 EraseInst(I);
1995 else
1996 OptimizeInst(I);
1997 }
1998 }
1999
2000 // We are done with the rank map.
2001 RankMap.clear();
2002 ValueRankMap.clear();
2003
2004 return MadeChange;
2005 }
2006