Bug Summary

File:lib/Transforms/Scalar/Reassociate.cpp
Warning:line 171, column 8
Dereference of null pointer (loaded from variable '__begin1')

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name Reassociate.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn345461/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/lib/Transforms/Scalar -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-10-27-211344-32123-1 -x c++ /build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp -faddrsig
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#include "llvm/Transforms/Scalar/Reassociate.h"
24#include "llvm/ADT/APFloat.h"
25#include "llvm/ADT/APInt.h"
26#include "llvm/ADT/DenseMap.h"
27#include "llvm/ADT/PostOrderIterator.h"
28#include "llvm/ADT/SetVector.h"
29#include "llvm/ADT/SmallPtrSet.h"
30#include "llvm/ADT/SmallSet.h"
31#include "llvm/ADT/SmallVector.h"
32#include "llvm/ADT/Statistic.h"
33#include "llvm/Analysis/GlobalsModRef.h"
34#include "llvm/Transforms/Utils/Local.h"
35#include "llvm/Analysis/ValueTracking.h"
36#include "llvm/IR/Argument.h"
37#include "llvm/IR/BasicBlock.h"
38#include "llvm/IR/CFG.h"
39#include "llvm/IR/Constant.h"
40#include "llvm/IR/Constants.h"
41#include "llvm/IR/Function.h"
42#include "llvm/IR/IRBuilder.h"
43#include "llvm/IR/InstrTypes.h"
44#include "llvm/IR/Instruction.h"
45#include "llvm/IR/Instructions.h"
46#include "llvm/IR/IntrinsicInst.h"
47#include "llvm/IR/Operator.h"
48#include "llvm/IR/PassManager.h"
49#include "llvm/IR/PatternMatch.h"
50#include "llvm/IR/Type.h"
51#include "llvm/IR/User.h"
52#include "llvm/IR/Value.h"
53#include "llvm/IR/ValueHandle.h"
54#include "llvm/Pass.h"
55#include "llvm/Support/Casting.h"
56#include "llvm/Support/Debug.h"
57#include "llvm/Support/ErrorHandling.h"
58#include "llvm/Support/raw_ostream.h"
59#include "llvm/Transforms/Scalar.h"
60#include <algorithm>
61#include <cassert>
62#include <utility>
63
64using namespace llvm;
65using namespace reassociate;
66using namespace PatternMatch;
67
68#define DEBUG_TYPE"reassociate" "reassociate"
69
70STATISTIC(NumChanged, "Number of insts reassociated")static llvm::Statistic NumChanged = {"reassociate", "NumChanged"
, "Number of insts reassociated", {0}, {false}}
;
71STATISTIC(NumAnnihil, "Number of expr tree annihilated")static llvm::Statistic NumAnnihil = {"reassociate", "NumAnnihil"
, "Number of expr tree annihilated", {0}, {false}}
;
72STATISTIC(NumFactor , "Number of multiplies factored")static llvm::Statistic NumFactor = {"reassociate", "NumFactor"
, "Number of multiplies factored", {0}, {false}}
;
73
74#ifndef NDEBUG
75/// Print out the expression identified in the Ops list.
76static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
77 Module *M = I->getModule();
78 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
79 << *Ops[0].Op->getType() << '\t';
80 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
81 dbgs() << "[ ";
82 Ops[i].Op->printAsOperand(dbgs(), false, M);
83 dbgs() << ", #" << Ops[i].Rank << "] ";
84 }
85}
86#endif
87
88/// Utility class representing a non-constant Xor-operand. We classify
89/// non-constant Xor-Operands into two categories:
90/// C1) The operand is in the form "X & C", where C is a constant and C != ~0
91/// C2)
92/// C2.1) The operand is in the form of "X | C", where C is a non-zero
93/// constant.
94/// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
95/// operand as "E | 0"
96class llvm::reassociate::XorOpnd {
97public:
98 XorOpnd(Value *V);
99
100 bool isInvalid() const { return SymbolicPart == nullptr; }
101 bool isOrExpr() const { return isOr; }
102 Value *getValue() const { return OrigVal; }
103 Value *getSymbolicPart() const { return SymbolicPart; }
104 unsigned getSymbolicRank() const { return SymbolicRank; }
105 const APInt &getConstPart() const { return ConstPart; }
106
107 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
108 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
109
110private:
111 Value *OrigVal;
112 Value *SymbolicPart;
113 APInt ConstPart;
114 unsigned SymbolicRank;
115 bool isOr;
116};
117
118XorOpnd::XorOpnd(Value *V) {
119 assert(!isa<ConstantInt>(V) && "No ConstantInt")((!isa<ConstantInt>(V) && "No ConstantInt") ? static_cast
<void> (0) : __assert_fail ("!isa<ConstantInt>(V) && \"No ConstantInt\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 119, __PRETTY_FUNCTION__))
;
120 OrigVal = V;
121 Instruction *I = dyn_cast<Instruction>(V);
122 SymbolicRank = 0;
123
124 if (I && (I->getOpcode() == Instruction::Or ||
125 I->getOpcode() == Instruction::And)) {
126 Value *V0 = I->getOperand(0);
127 Value *V1 = I->getOperand(1);
128 const APInt *C;
129 if (match(V0, m_APInt(C)))
130 std::swap(V0, V1);
131
132 if (match(V1, m_APInt(C))) {
133 ConstPart = *C;
134 SymbolicPart = V0;
135 isOr = (I->getOpcode() == Instruction::Or);
136 return;
137 }
138 }
139
140 // view the operand as "V | 0"
141 SymbolicPart = V;
142 ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits());
143 isOr = true;
144}
145
146/// Return true if V is an instruction of the specified opcode and if it
147/// only has one use.
148static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
149 auto *I = dyn_cast<Instruction>(V);
150 if (I && I->hasOneUse() && I->getOpcode() == Opcode)
151 if (!isa<FPMathOperator>(I) || I->isFast())
152 return cast<BinaryOperator>(I);
153 return nullptr;
154}
155
156static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
157 unsigned Opcode2) {
158 auto *I = dyn_cast<Instruction>(V);
159 if (I && I->hasOneUse() &&
160 (I->getOpcode() == Opcode1 || I->getOpcode() == Opcode2))
161 if (!isa<FPMathOperator>(I) || I->isFast())
162 return cast<BinaryOperator>(I);
163 return nullptr;
164}
165
166void ReassociatePass::BuildRankMap(Function &F,
167 ReversePostOrderTraversal<Function*> &RPOT) {
168 unsigned Rank = 2;
169
170 // Assign distinct ranks to function arguments.
171 for (auto &Arg : F.args()) {
5
Assuming '__begin1' is not equal to '__end1'
6
Null pointer value stored to '__begin1'
7
Assuming '__begin1' is not equal to '__end1'
8
Dereference of null pointer (loaded from variable '__begin1')
172 ValueRankMap[&Arg] = ++Rank;
173 LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rankdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
Arg.getName() << "] = " << Rank << "\n"; }
} while (false)
174 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
Arg.getName() << "] = " << Rank << "\n"; }
} while (false)
;
175 }
176
177 // Traverse basic blocks in ReversePostOrder
178 for (BasicBlock *BB : RPOT) {
179 unsigned BBRank = RankMap[BB] = ++Rank << 16;
180
181 // Walk the basic block, adding precomputed ranks for any instructions that
182 // we cannot move. This ensures that the ranks for these instructions are
183 // all different in the block.
184 for (Instruction &I : *BB)
185 if (mayBeMemoryDependent(I))
186 ValueRankMap[&I] = ++BBRank;
187 }
188}
189
190unsigned ReassociatePass::getRank(Value *V) {
191 Instruction *I = dyn_cast<Instruction>(V);
192 if (!I) {
193 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
194 return 0; // Otherwise it's a global or constant, rank 0.
195 }
196
197 if (unsigned Rank = ValueRankMap[I])
198 return Rank; // Rank already known?
199
200 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
201 // we can reassociate expressions for code motion! Since we do not recurse
202 // for PHI nodes, we cannot have infinite recursion here, because there
203 // cannot be loops in the value graph that do not go through PHI nodes.
204 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
205 for (unsigned i = 0, e = I->getNumOperands(); i != e && Rank != MaxRank; ++i)
206 Rank = std::max(Rank, getRank(I->getOperand(i)));
207
208 // If this is a 'not' or 'neg' instruction, do not count it for rank. This
209 // assures us that X and ~X will have the same rank.
210 if (!match(I, m_Not(m_Value())) && !match(I, m_Neg(m_Value())) &&
211 !match(I, m_FNeg(m_Value())))
212 ++Rank;
213
214 LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rankdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
V->getName() << "] = " << Rank << "\n";
} } while (false)
215 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
V->getName() << "] = " << Rank << "\n";
} } while (false)
;
216
217 return ValueRankMap[I] = Rank;
218}
219
220// Canonicalize constants to RHS. Otherwise, sort the operands by rank.
221void ReassociatePass::canonicalizeOperands(Instruction *I) {
222 assert(isa<BinaryOperator>(I) && "Expected binary operator.")((isa<BinaryOperator>(I) && "Expected binary operator."
) ? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(I) && \"Expected binary operator.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 222, __PRETTY_FUNCTION__))
;
223 assert(I->isCommutative() && "Expected commutative operator.")((I->isCommutative() && "Expected commutative operator."
) ? static_cast<void> (0) : __assert_fail ("I->isCommutative() && \"Expected commutative operator.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 223, __PRETTY_FUNCTION__))
;
224
225 Value *LHS = I->getOperand(0);
226 Value *RHS = I->getOperand(1);
227 if (LHS == RHS || isa<Constant>(RHS))
228 return;
229 if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
230 cast<BinaryOperator>(I)->swapOperands();
231}
232
233static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
234 Instruction *InsertBefore, Value *FlagsOp) {
235 if (S1->getType()->isIntOrIntVectorTy())
236 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
237 else {
238 BinaryOperator *Res =
239 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
240 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
241 return Res;
242 }
243}
244
245static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
246 Instruction *InsertBefore, Value *FlagsOp) {
247 if (S1->getType()->isIntOrIntVectorTy())
248 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
249 else {
250 BinaryOperator *Res =
251 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
252 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
253 return Res;
254 }
255}
256
257static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
258 Instruction *InsertBefore, Value *FlagsOp) {
259 if (S1->getType()->isIntOrIntVectorTy())
260 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
261 else {
262 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
263 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
264 return Res;
265 }
266}
267
268/// Replace 0-X with X*-1.
269static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
270 Type *Ty = Neg->getType();
271 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
272 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
273
274 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
275 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
276 Res->takeName(Neg);
277 Neg->replaceAllUsesWith(Res);
278 Res->setDebugLoc(Neg->getDebugLoc());
279 return Res;
280}
281
282/// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
283/// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
284/// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
285/// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
286/// even x in Bitwidth-bit arithmetic.
287static unsigned CarmichaelShift(unsigned Bitwidth) {
288 if (Bitwidth < 3)
289 return Bitwidth - 1;
290 return Bitwidth - 2;
291}
292
293/// Add the extra weight 'RHS' to the existing weight 'LHS',
294/// reducing the combined weight using any special properties of the operation.
295/// The existing weight LHS represents the computation X op X op ... op X where
296/// X occurs LHS times. The combined weight represents X op X op ... op X with
297/// X occurring LHS + RHS times. If op is "Xor" for example then the combined
298/// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
299/// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
300static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
301 // If we were working with infinite precision arithmetic then the combined
302 // weight would be LHS + RHS. But we are using finite precision arithmetic,
303 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
304 // for nilpotent operations and addition, but not for idempotent operations
305 // and multiplication), so it is important to correctly reduce the combined
306 // weight back into range if wrapping would be wrong.
307
308 // If RHS is zero then the weight didn't change.
309 if (RHS.isMinValue())
310 return;
311 // If LHS is zero then the combined weight is RHS.
312 if (LHS.isMinValue()) {
313 LHS = RHS;
314 return;
315 }
316 // From this point on we know that neither LHS nor RHS is zero.
317
318 if (Instruction::isIdempotent(Opcode)) {
319 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
320 // weight of 1. Keeping weights at zero or one also means that wrapping is
321 // not a problem.
322 assert(LHS == 1 && RHS == 1 && "Weights not reduced!")((LHS == 1 && RHS == 1 && "Weights not reduced!"
) ? static_cast<void> (0) : __assert_fail ("LHS == 1 && RHS == 1 && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 322, __PRETTY_FUNCTION__))
;
323 return; // Return a weight of 1.
324 }
325 if (Instruction::isNilpotent(Opcode)) {
326 // Nilpotent means X op X === 0, so reduce weights modulo 2.
327 assert(LHS == 1 && RHS == 1 && "Weights not reduced!")((LHS == 1 && RHS == 1 && "Weights not reduced!"
) ? static_cast<void> (0) : __assert_fail ("LHS == 1 && RHS == 1 && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 327, __PRETTY_FUNCTION__))
;
328 LHS = 0; // 1 + 1 === 0 modulo 2.
329 return;
330 }
331 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
332 // TODO: Reduce the weight by exploiting nsw/nuw?
333 LHS += RHS;
334 return;
335 }
336
337 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&(((Opcode == Instruction::Mul || Opcode == Instruction::FMul)
&& "Unknown associative operation!") ? static_cast<
void> (0) : __assert_fail ("(Opcode == Instruction::Mul || Opcode == Instruction::FMul) && \"Unknown associative operation!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 338, __PRETTY_FUNCTION__))
338 "Unknown associative operation!")(((Opcode == Instruction::Mul || Opcode == Instruction::FMul)
&& "Unknown associative operation!") ? static_cast<
void> (0) : __assert_fail ("(Opcode == Instruction::Mul || Opcode == Instruction::FMul) && \"Unknown associative operation!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 338, __PRETTY_FUNCTION__))
;
339 unsigned Bitwidth = LHS.getBitWidth();
340 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
341 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
342 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
343 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
344 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
345 // which by a happy accident means that they can always be represented using
346 // Bitwidth bits.
347 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
348 // the Carmichael number).
349 if (Bitwidth > 3) {
350 /// CM - The value of Carmichael's lambda function.
351 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
352 // Any weight W >= Threshold can be replaced with W - CM.
353 APInt Threshold = CM + Bitwidth;
354 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!")((LHS.ult(Threshold) && RHS.ult(Threshold) &&
"Weights not reduced!") ? static_cast<void> (0) : __assert_fail
("LHS.ult(Threshold) && RHS.ult(Threshold) && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 354, __PRETTY_FUNCTION__))
;
355 // For Bitwidth 4 or more the following sum does not overflow.
356 LHS += RHS;
357 while (LHS.uge(Threshold))
358 LHS -= CM;
359 } else {
360 // To avoid problems with overflow do everything the same as above but using
361 // a larger type.
362 unsigned CM = 1U << CarmichaelShift(Bitwidth);
363 unsigned Threshold = CM + Bitwidth;
364 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&((LHS.getZExtValue() < Threshold && RHS.getZExtValue
() < Threshold && "Weights not reduced!") ? static_cast
<void> (0) : __assert_fail ("LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 365, __PRETTY_FUNCTION__))
365 "Weights not reduced!")((LHS.getZExtValue() < Threshold && RHS.getZExtValue
() < Threshold && "Weights not reduced!") ? static_cast
<void> (0) : __assert_fail ("LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 365, __PRETTY_FUNCTION__))
;
366 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
367 while (Total >= Threshold)
368 Total -= CM;
369 LHS = Total;
370 }
371}
372
373using RepeatedValue = std::pair<Value*, APInt>;
374
375/// Given an associative binary expression, return the leaf
376/// nodes in Ops along with their weights (how many times the leaf occurs). The
377/// original expression is the same as
378/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
379/// op
380/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
381/// op
382/// ...
383/// op
384/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
385///
386/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
387///
388/// This routine may modify the function, in which case it returns 'true'. The
389/// changes it makes may well be destructive, changing the value computed by 'I'
390/// to something completely different. Thus if the routine returns 'true' then
391/// you MUST either replace I with a new expression computed from the Ops array,
392/// or use RewriteExprTree to put the values back in.
393///
394/// A leaf node is either not a binary operation of the same kind as the root
395/// node 'I' (i.e. is not a binary operator at all, or is, but with a different
396/// opcode), or is the same kind of binary operator but has a use which either
397/// does not belong to the expression, or does belong to the expression but is
398/// a leaf node. Every leaf node has at least one use that is a non-leaf node
399/// of the expression, while for non-leaf nodes (except for the root 'I') every
400/// use is a non-leaf node of the expression.
401///
402/// For example:
403/// expression graph node names
404///
405/// + | I
406/// / \ |
407/// + + | A, B
408/// / \ / \ |
409/// * + * | C, D, E
410/// / \ / \ / \ |
411/// + * | F, G
412///
413/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
414/// that order) (C, 1), (E, 1), (F, 2), (G, 2).
415///
416/// The expression is maximal: if some instruction is a binary operator of the
417/// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
418/// then the instruction also belongs to the expression, is not a leaf node of
419/// it, and its operands also belong to the expression (but may be leaf nodes).
420///
421/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
422/// order to ensure that every non-root node in the expression has *exactly one*
423/// use by a non-leaf node of the expression. This destruction means that the
424/// caller MUST either replace 'I' with a new expression or use something like
425/// RewriteExprTree to put the values back in if the routine indicates that it
426/// made a change by returning 'true'.
427///
428/// In the above example either the right operand of A or the left operand of B
429/// will be replaced by undef. If it is B's operand then this gives:
430///
431/// + | I
432/// / \ |
433/// + + | A, B - operand of B replaced with undef
434/// / \ \ |
435/// * + * | C, D, E
436/// / \ / \ / \ |
437/// + * | F, G
438///
439/// Note that such undef operands can only be reached by passing through 'I'.
440/// For example, if you visit operands recursively starting from a leaf node
441/// then you will never see such an undef operand unless you get back to 'I',
442/// which requires passing through a phi node.
443///
444/// Note that this routine may also mutate binary operators of the wrong type
445/// that have all uses inside the expression (i.e. only used by non-leaf nodes
446/// of the expression) if it can turn them into binary operators of the right
447/// type and thus make the expression bigger.
448static bool LinearizeExprTree(BinaryOperator *I,
449 SmallVectorImpl<RepeatedValue> &Ops) {
450 LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "LINEARIZE: " << *I <<
'\n'; } } while (false)
;
451 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
452 unsigned Opcode = I->getOpcode();
453 assert(I->isAssociative() && I->isCommutative() &&((I->isAssociative() && I->isCommutative() &&
"Expected an associative and commutative operation!") ? static_cast
<void> (0) : __assert_fail ("I->isAssociative() && I->isCommutative() && \"Expected an associative and commutative operation!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 454, __PRETTY_FUNCTION__))
454 "Expected an associative and commutative operation!")((I->isAssociative() && I->isCommutative() &&
"Expected an associative and commutative operation!") ? static_cast
<void> (0) : __assert_fail ("I->isAssociative() && I->isCommutative() && \"Expected an associative and commutative operation!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 454, __PRETTY_FUNCTION__))
;
455
456 // Visit all operands of the expression, keeping track of their weight (the
457 // number of paths from the expression root to the operand, or if you like
458 // the number of times that operand occurs in the linearized expression).
459 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
460 // while A has weight two.
461
462 // Worklist of non-leaf nodes (their operands are in the expression too) along
463 // with their weights, representing a certain number of paths to the operator.
464 // If an operator occurs in the worklist multiple times then we found multiple
465 // ways to get to it.
466 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
467 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
468 bool Changed = false;
469
470 // Leaves of the expression are values that either aren't the right kind of
471 // operation (eg: a constant, or a multiply in an add tree), or are, but have
472 // some uses that are not inside the expression. For example, in I = X + X,
473 // X = A + B, the value X has two uses (by I) that are in the expression. If
474 // X has any other uses, for example in a return instruction, then we consider
475 // X to be a leaf, and won't analyze it further. When we first visit a value,
476 // if it has more than one use then at first we conservatively consider it to
477 // be a leaf. Later, as the expression is explored, we may discover some more
478 // uses of the value from inside the expression. If all uses turn out to be
479 // from within the expression (and the value is a binary operator of the right
480 // kind) then the value is no longer considered to be a leaf, and its operands
481 // are explored.
482
483 // Leaves - Keeps track of the set of putative leaves as well as the number of
484 // paths to each leaf seen so far.
485 using LeafMap = DenseMap<Value *, APInt>;
486 LeafMap Leaves; // Leaf -> Total weight so far.
487 SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
488
489#ifndef NDEBUG
490 SmallPtrSet<Value *, 8> Visited; // For sanity checking the iteration scheme.
491#endif
492 while (!Worklist.empty()) {
493 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
494 I = P.first; // We examine the operands of this binary operator.
495
496 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
497 Value *Op = I->getOperand(OpIdx);
498 APInt Weight = P.second; // Number of paths to this operand.
499 LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "OPERAND: " << *Op <<
" (" << Weight << ")\n"; } } while (false)
;
500 assert(!Op->use_empty() && "No uses, so how did we get to it?!")((!Op->use_empty() && "No uses, so how did we get to it?!"
) ? static_cast<void> (0) : __assert_fail ("!Op->use_empty() && \"No uses, so how did we get to it?!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 500, __PRETTY_FUNCTION__))
;
501
502 // If this is a binary operation of the right kind with only one use then
503 // add its operands to the expression.
504 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
505 assert(Visited.insert(Op).second && "Not first visit!")((Visited.insert(Op).second && "Not first visit!") ? static_cast
<void> (0) : __assert_fail ("Visited.insert(Op).second && \"Not first visit!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 505, __PRETTY_FUNCTION__))
;
506 LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "DIRECT ADD: " << *Op
<< " (" << Weight << ")\n"; } } while (false
)
;
507 Worklist.push_back(std::make_pair(BO, Weight));
508 continue;
509 }
510
511 // Appears to be a leaf. Is the operand already in the set of leaves?
512 LeafMap::iterator It = Leaves.find(Op);
513 if (It == Leaves.end()) {
514 // Not in the leaf map. Must be the first time we saw this operand.
515 assert(Visited.insert(Op).second && "Not first visit!")((Visited.insert(Op).second && "Not first visit!") ? static_cast
<void> (0) : __assert_fail ("Visited.insert(Op).second && \"Not first visit!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 515, __PRETTY_FUNCTION__))
;
516 if (!Op->hasOneUse()) {
517 // This value has uses not accounted for by the expression, so it is
518 // not safe to modify. Mark it as being a leaf.
519 LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "ADD USES LEAF: " <<
*Op << " (" << Weight << ")\n"; } } while (
false)
520 << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "ADD USES LEAF: " <<
*Op << " (" << Weight << ")\n"; } } while (
false)
;
521 LeafOrder.push_back(Op);
522 Leaves[Op] = Weight;
523 continue;
524 }
525 // No uses outside the expression, try morphing it.
526 } else {
527 // Already in the leaf map.
528 assert(It != Leaves.end() && Visited.count(Op) &&((It != Leaves.end() && Visited.count(Op) && "In leaf map but not visited!"
) ? static_cast<void> (0) : __assert_fail ("It != Leaves.end() && Visited.count(Op) && \"In leaf map but not visited!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 529, __PRETTY_FUNCTION__))
529 "In leaf map but not visited!")((It != Leaves.end() && Visited.count(Op) && "In leaf map but not visited!"
) ? static_cast<void> (0) : __assert_fail ("It != Leaves.end() && Visited.count(Op) && \"In leaf map but not visited!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 529, __PRETTY_FUNCTION__))
;
530
531 // Update the number of paths to the leaf.
532 IncorporateWeight(It->second, Weight, Opcode);
533
534#if 0 // TODO: Re-enable once PR13021 is fixed.
535 // The leaf already has one use from inside the expression. As we want
536 // exactly one such use, drop this new use of the leaf.
537 assert(!Op->hasOneUse() && "Only one use, but we got here twice!")((!Op->hasOneUse() && "Only one use, but we got here twice!"
) ? static_cast<void> (0) : __assert_fail ("!Op->hasOneUse() && \"Only one use, but we got here twice!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 537, __PRETTY_FUNCTION__))
;
538 I->setOperand(OpIdx, UndefValue::get(I->getType()));
539 Changed = true;
540
541 // If the leaf is a binary operation of the right kind and we now see
542 // that its multiple original uses were in fact all by nodes belonging
543 // to the expression, then no longer consider it to be a leaf and add
544 // its operands to the expression.
545 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
546 LLVM_DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "UNLEAF: " << *Op <<
" (" << It->second << ")\n"; } } while (false
)
;
547 Worklist.push_back(std::make_pair(BO, It->second));
548 Leaves.erase(It);
549 continue;
550 }
551#endif
552
553 // If we still have uses that are not accounted for by the expression
554 // then it is not safe to modify the value.
555 if (!Op->hasOneUse())
556 continue;
557
558 // No uses outside the expression, try morphing it.
559 Weight = It->second;
560 Leaves.erase(It); // Since the value may be morphed below.
561 }
562
563 // At this point we have a value which, first of all, is not a binary
564 // expression of the right kind, and secondly, is only used inside the
565 // expression. This means that it can safely be modified. See if we
566 // can usefully morph it into an expression of the right kind.
567 assert((!isa<Instruction>(Op) ||(((!isa<Instruction>(Op) || cast<Instruction>(Op)
->getOpcode() != Opcode || (isa<FPMathOperator>(Op) &&
!cast<Instruction>(Op)->isFast())) && "Should have been handled above!"
) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 571, __PRETTY_FUNCTION__))
568 cast<Instruction>(Op)->getOpcode() != Opcode(((!isa<Instruction>(Op) || cast<Instruction>(Op)
->getOpcode() != Opcode || (isa<FPMathOperator>(Op) &&
!cast<Instruction>(Op)->isFast())) && "Should have been handled above!"
) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 571, __PRETTY_FUNCTION__))
569 || (isa<FPMathOperator>(Op) &&(((!isa<Instruction>(Op) || cast<Instruction>(Op)
->getOpcode() != Opcode || (isa<FPMathOperator>(Op) &&
!cast<Instruction>(Op)->isFast())) && "Should have been handled above!"
) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 571, __PRETTY_FUNCTION__))
570 !cast<Instruction>(Op)->isFast())) &&(((!isa<Instruction>(Op) || cast<Instruction>(Op)
->getOpcode() != Opcode || (isa<FPMathOperator>(Op) &&
!cast<Instruction>(Op)->isFast())) && "Should have been handled above!"
) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 571, __PRETTY_FUNCTION__))
571 "Should have been handled above!")(((!isa<Instruction>(Op) || cast<Instruction>(Op)
->getOpcode() != Opcode || (isa<FPMathOperator>(Op) &&
!cast<Instruction>(Op)->isFast())) && "Should have been handled above!"
) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 571, __PRETTY_FUNCTION__))
;
572 assert(Op->hasOneUse() && "Has uses outside the expression tree!")((Op->hasOneUse() && "Has uses outside the expression tree!"
) ? static_cast<void> (0) : __assert_fail ("Op->hasOneUse() && \"Has uses outside the expression tree!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 572, __PRETTY_FUNCTION__))
;
573
574 // If this is a multiply expression, turn any internal negations into
575 // multiplies by -1 so they can be reassociated.
576 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
577 if ((Opcode == Instruction::Mul && match(BO, m_Neg(m_Value()))) ||
578 (Opcode == Instruction::FMul && match(BO, m_FNeg(m_Value())))) {
579 LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "MORPH LEAF: " << *Op
<< " (" << Weight << ") TO "; } } while (false
)
580 << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "MORPH LEAF: " << *Op
<< " (" << Weight << ") TO "; } } while (false
)
;
581 BO = LowerNegateToMultiply(BO);
582 LLVM_DEBUG(dbgs() << *BO << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << *BO << '\n'; } } while
(false)
;
583 Worklist.push_back(std::make_pair(BO, Weight));
584 Changed = true;
585 continue;
586 }
587
588 // Failed to morph into an expression of the right type. This really is
589 // a leaf.
590 LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "ADD LEAF: " << *Op <<
" (" << Weight << ")\n"; } } while (false)
;
591 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?")((!isReassociableOp(Op, Opcode) && "Value was morphed?"
) ? static_cast<void> (0) : __assert_fail ("!isReassociableOp(Op, Opcode) && \"Value was morphed?\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 591, __PRETTY_FUNCTION__))
;
592 LeafOrder.push_back(Op);
593 Leaves[Op] = Weight;
594 }
595 }
596
597 // The leaves, repeated according to their weights, represent the linearized
598 // form of the expression.
599 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
600 Value *V = LeafOrder[i];
601 LeafMap::iterator It = Leaves.find(V);
602 if (It == Leaves.end())
603 // Node initially thought to be a leaf wasn't.
604 continue;
605 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!")((!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"
) ? static_cast<void> (0) : __assert_fail ("!isReassociableOp(V, Opcode) && \"Shouldn't be a leaf!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 605, __PRETTY_FUNCTION__))
;
606 APInt Weight = It->second;
607 if (Weight.isMinValue())
608 // Leaf already output or weight reduction eliminated it.
609 continue;
610 // Ensure the leaf is only output once.
611 It->second = 0;
612 Ops.push_back(std::make_pair(V, Weight));
613 }
614
615 // For nilpotent operations or addition there may be no operands, for example
616 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
617 // in both cases the weight reduces to 0 causing the value to be skipped.
618 if (Ops.empty()) {
619 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
620 assert(Identity && "Associative operation without identity!")((Identity && "Associative operation without identity!"
) ? static_cast<void> (0) : __assert_fail ("Identity && \"Associative operation without identity!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 620, __PRETTY_FUNCTION__))
;
621 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
622 }
623
624 return Changed;
625}
626
627/// Now that the operands for this expression tree are
628/// linearized and optimized, emit them in-order.
629void ReassociatePass::RewriteExprTree(BinaryOperator *I,
630 SmallVectorImpl<ValueEntry> &Ops) {
631 assert(Ops.size() > 1 && "Single values should be used directly!")((Ops.size() > 1 && "Single values should be used directly!"
) ? static_cast<void> (0) : __assert_fail ("Ops.size() > 1 && \"Single values should be used directly!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 631, __PRETTY_FUNCTION__))
;
632
633 // Since our optimizations should never increase the number of operations, the
634 // new expression can usually be written reusing the existing binary operators
635 // from the original expression tree, without creating any new instructions,
636 // though the rewritten expression may have a completely different topology.
637 // We take care to not change anything if the new expression will be the same
638 // as the original. If more than trivial changes (like commuting operands)
639 // were made then we are obliged to clear out any optional subclass data like
640 // nsw flags.
641
642 /// NodesToRewrite - Nodes from the original expression available for writing
643 /// the new expression into.
644 SmallVector<BinaryOperator*, 8> NodesToRewrite;
645 unsigned Opcode = I->getOpcode();
646 BinaryOperator *Op = I;
647
648 /// NotRewritable - The operands being written will be the leaves of the new
649 /// expression and must not be used as inner nodes (via NodesToRewrite) by
650 /// mistake. Inner nodes are always reassociable, and usually leaves are not
651 /// (if they were they would have been incorporated into the expression and so
652 /// would not be leaves), so most of the time there is no danger of this. But
653 /// in rare cases a leaf may become reassociable if an optimization kills uses
654 /// of it, or it may momentarily become reassociable during rewriting (below)
655 /// due it being removed as an operand of one of its uses. Ensure that misuse
656 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
657 /// leaves and refusing to reuse any of them as inner nodes.
658 SmallPtrSet<Value*, 8> NotRewritable;
659 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
660 NotRewritable.insert(Ops[i].Op);
661
662 // ExpressionChanged - Non-null if the rewritten expression differs from the
663 // original in some non-trivial way, requiring the clearing of optional flags.
664 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
665 BinaryOperator *ExpressionChanged = nullptr;
666 for (unsigned i = 0; ; ++i) {
667 // The last operation (which comes earliest in the IR) is special as both
668 // operands will come from Ops, rather than just one with the other being
669 // a subexpression.
670 if (i+2 == Ops.size()) {
671 Value *NewLHS = Ops[i].Op;
672 Value *NewRHS = Ops[i+1].Op;
673 Value *OldLHS = Op->getOperand(0);
674 Value *OldRHS = Op->getOperand(1);
675
676 if (NewLHS == OldLHS && NewRHS == OldRHS)
677 // Nothing changed, leave it alone.
678 break;
679
680 if (NewLHS == OldRHS && NewRHS == OldLHS) {
681 // The order of the operands was reversed. Swap them.
682 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
683 Op->swapOperands();
684 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
685 MadeChange = true;
686 ++NumChanged;
687 break;
688 }
689
690 // The new operation differs non-trivially from the original. Overwrite
691 // the old operands with the new ones.
692 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
693 if (NewLHS != OldLHS) {
694 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
695 if (BO && !NotRewritable.count(BO))
696 NodesToRewrite.push_back(BO);
697 Op->setOperand(0, NewLHS);
698 }
699 if (NewRHS != OldRHS) {
700 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
701 if (BO && !NotRewritable.count(BO))
702 NodesToRewrite.push_back(BO);
703 Op->setOperand(1, NewRHS);
704 }
705 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
706
707 ExpressionChanged = Op;
708 MadeChange = true;
709 ++NumChanged;
710
711 break;
712 }
713
714 // Not the last operation. The left-hand side will be a sub-expression
715 // while the right-hand side will be the current element of Ops.
716 Value *NewRHS = Ops[i].Op;
717 if (NewRHS != Op->getOperand(1)) {
718 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
719 if (NewRHS == Op->getOperand(0)) {
720 // The new right-hand side was already present as the left operand. If
721 // we are lucky then swapping the operands will sort out both of them.
722 Op->swapOperands();
723 } else {
724 // Overwrite with the new right-hand side.
725 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
726 if (BO && !NotRewritable.count(BO))
727 NodesToRewrite.push_back(BO);
728 Op->setOperand(1, NewRHS);
729 ExpressionChanged = Op;
730 }
731 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
732 MadeChange = true;
733 ++NumChanged;
734 }
735
736 // Now deal with the left-hand side. If this is already an operation node
737 // from the original expression then just rewrite the rest of the expression
738 // into it.
739 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
740 if (BO && !NotRewritable.count(BO)) {
741 Op = BO;
742 continue;
743 }
744
745 // Otherwise, grab a spare node from the original expression and use that as
746 // the left-hand side. If there are no nodes left then the optimizers made
747 // an expression with more nodes than the original! This usually means that
748 // they did something stupid but it might mean that the problem was just too
749 // hard (finding the mimimal number of multiplications needed to realize a
750 // multiplication expression is NP-complete). Whatever the reason, smart or
751 // stupid, create a new node if there are none left.
752 BinaryOperator *NewOp;
753 if (NodesToRewrite.empty()) {
754 Constant *Undef = UndefValue::get(I->getType());
755 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
756 Undef, Undef, "", I);
757 if (NewOp->getType()->isFPOrFPVectorTy())
758 NewOp->setFastMathFlags(I->getFastMathFlags());
759 } else {
760 NewOp = NodesToRewrite.pop_back_val();
761 }
762
763 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
764 Op->setOperand(0, NewOp);
765 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
766 ExpressionChanged = Op;
767 MadeChange = true;
768 ++NumChanged;
769 Op = NewOp;
770 }
771
772 // If the expression changed non-trivially then clear out all subclass data
773 // starting from the operator specified in ExpressionChanged, and compactify
774 // the operators to just before the expression root to guarantee that the
775 // expression tree is dominated by all of Ops.
776 if (ExpressionChanged)
777 do {
778 // Preserve FastMathFlags.
779 if (isa<FPMathOperator>(I)) {
780 FastMathFlags Flags = I->getFastMathFlags();
781 ExpressionChanged->clearSubclassOptionalData();
782 ExpressionChanged->setFastMathFlags(Flags);
783 } else
784 ExpressionChanged->clearSubclassOptionalData();
785
786 if (ExpressionChanged == I)
787 break;
788
789 // Discard any debug info related to the expressions that has changed (we
790 // can leave debug infor related to the root, since the result of the
791 // expression tree should be the same even after reassociation).
792 SmallVector<DbgVariableIntrinsic *, 1> DbgUsers;
793 findDbgUsers(DbgUsers, ExpressionChanged);
794 for (auto *DII : DbgUsers) {
795 Value *Undef = UndefValue::get(ExpressionChanged->getType());
796 DII->setOperand(0, MetadataAsValue::get(DII->getContext(),
797 ValueAsMetadata::get(Undef)));
798 }
799
800 ExpressionChanged->moveBefore(I);
801 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
802 } while (true);
803
804 // Throw away any left over nodes from the original expression.
805 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
806 RedoInsts.insert(NodesToRewrite[i]);
807}
808
809/// Insert instructions before the instruction pointed to by BI,
810/// that computes the negative version of the value specified. The negative
811/// version of the value is returned, and BI is left pointing at the instruction
812/// that should be processed next by the reassociation pass.
813/// Also add intermediate instructions to the redo list that are modified while
814/// pushing the negates through adds. These will be revisited to see if
815/// additional opportunities have been exposed.
816static Value *NegateValue(Value *V, Instruction *BI,
817 ReassociatePass::OrderedSet &ToRedo) {
818 if (auto *C = dyn_cast<Constant>(V))
819 return C->getType()->isFPOrFPVectorTy() ? ConstantExpr::getFNeg(C) :
820 ConstantExpr::getNeg(C);
821
822 // We are trying to expose opportunity for reassociation. One of the things
823 // that we want to do to achieve this is to push a negation as deep into an
824 // expression chain as possible, to expose the add instructions. In practice,
825 // this means that we turn this:
826 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
827 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
828 // the constants. We assume that instcombine will clean up the mess later if
829 // we introduce tons of unnecessary negation instructions.
830 //
831 if (BinaryOperator *I =
832 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
833 // Push the negates through the add.
834 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
835 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
836 if (I->getOpcode() == Instruction::Add) {
837 I->setHasNoUnsignedWrap(false);
838 I->setHasNoSignedWrap(false);
839 }
840
841 // We must move the add instruction here, because the neg instructions do
842 // not dominate the old add instruction in general. By moving it, we are
843 // assured that the neg instructions we just inserted dominate the
844 // instruction we are about to insert after them.
845 //
846 I->moveBefore(BI);
847 I->setName(I->getName()+".neg");
848
849 // Add the intermediate negates to the redo list as processing them later
850 // could expose more reassociating opportunities.
851 ToRedo.insert(I);
852 return I;
853 }
854
855 // Okay, we need to materialize a negated version of V with an instruction.
856 // Scan the use lists of V to see if we have one already.
857 for (User *U : V->users()) {
858 if (!match(U, m_Neg(m_Value())) && !match(U, m_FNeg(m_Value())))
859 continue;
860
861 // We found one! Now we have to make sure that the definition dominates
862 // this use. We do this by moving it to the entry block (if it is a
863 // non-instruction value) or right after the definition. These negates will
864 // be zapped by reassociate later, so we don't need much finesse here.
865 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
866
867 // Verify that the negate is in this function, V might be a constant expr.
868 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
869 continue;
870
871 BasicBlock::iterator InsertPt;
872 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
873 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
874 InsertPt = II->getNormalDest()->begin();
875 } else {
876 InsertPt = ++InstInput->getIterator();
877 }
878 while (isa<PHINode>(InsertPt)) ++InsertPt;
879 } else {
880 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
881 }
882 TheNeg->moveBefore(&*InsertPt);
883 if (TheNeg->getOpcode() == Instruction::Sub) {
884 TheNeg->setHasNoUnsignedWrap(false);
885 TheNeg->setHasNoSignedWrap(false);
886 } else {
887 TheNeg->andIRFlags(BI);
888 }
889 ToRedo.insert(TheNeg);
890 return TheNeg;
891 }
892
893 // Insert a 'neg' instruction that subtracts the value from zero to get the
894 // negation.
895 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
896 ToRedo.insert(NewNeg);
897 return NewNeg;
898}
899
900/// Return true if we should break up this subtract of X-Y into (X + -Y).
901static bool ShouldBreakUpSubtract(Instruction *Sub) {
902 // If this is a negation, we can't split it up!
903 if (match(Sub, m_Neg(m_Value())) || match(Sub, m_FNeg(m_Value())))
904 return false;
905
906 // Don't breakup X - undef.
907 if (isa<UndefValue>(Sub->getOperand(1)))
908 return false;
909
910 // Don't bother to break this up unless either the LHS is an associable add or
911 // subtract or if this is only used by one.
912 Value *V0 = Sub->getOperand(0);
913 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
914 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
915 return true;
916 Value *V1 = Sub->getOperand(1);
917 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
918 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
919 return true;
920 Value *VB = Sub->user_back();
921 if (Sub->hasOneUse() &&
922 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
923 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
924 return true;
925
926 return false;
927}
928
929/// If we have (X-Y), and if either X is an add, or if this is only used by an
930/// add, transform this into (X+(0-Y)) to promote better reassociation.
931static BinaryOperator *BreakUpSubtract(Instruction *Sub,
932 ReassociatePass::OrderedSet &ToRedo) {
933 // Convert a subtract into an add and a neg instruction. This allows sub
934 // instructions to be commuted with other add instructions.
935 //
936 // Calculate the negative value of Operand 1 of the sub instruction,
937 // and set it as the RHS of the add instruction we just made.
938 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
939 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
940 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
941 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
942 New->takeName(Sub);
943
944 // Everyone now refers to the add instruction.
945 Sub->replaceAllUsesWith(New);
946 New->setDebugLoc(Sub->getDebugLoc());
947
948 LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Negated: " << *New <<
'\n'; } } while (false)
;
949 return New;
950}
951
952/// If this is a shift of a reassociable multiply or is used by one, change
953/// this into a multiply by a constant to assist with further reassociation.
954static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
955 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
956 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
957
958 BinaryOperator *Mul =
959 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
960 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
961 Mul->takeName(Shl);
962
963 // Everyone now refers to the mul instruction.
964 Shl->replaceAllUsesWith(Mul);
965 Mul->setDebugLoc(Shl->getDebugLoc());
966
967 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
968 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
969 // handling.
970 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
971 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
972 if (NSW && NUW)
973 Mul->setHasNoSignedWrap(true);
974 Mul->setHasNoUnsignedWrap(NUW);
975 return Mul;
976}
977
978/// Scan backwards and forwards among values with the same rank as element i
979/// to see if X exists. If X does not exist, return i. This is useful when
980/// scanning for 'x' when we see '-x' because they both get the same rank.
981static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
982 unsigned i, Value *X) {
983 unsigned XRank = Ops[i].Rank;
984 unsigned e = Ops.size();
985 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
986 if (Ops[j].Op == X)
987 return j;
988 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
989 if (Instruction *I2 = dyn_cast<Instruction>(X))
990 if (I1->isIdenticalTo(I2))
991 return j;
992 }
993 // Scan backwards.
994 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
995 if (Ops[j].Op == X)
996 return j;
997 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
998 if (Instruction *I2 = dyn_cast<Instruction>(X))
999 if (I1->isIdenticalTo(I2))
1000 return j;
1001 }
1002 return i;
1003}
1004
1005/// Emit a tree of add instructions, summing Ops together
1006/// and returning the result. Insert the tree before I.
1007static Value *EmitAddTreeOfValues(Instruction *I,
1008 SmallVectorImpl<WeakTrackingVH> &Ops) {
1009 if (Ops.size() == 1) return Ops.back();
1010
1011 Value *V1 = Ops.back();
1012 Ops.pop_back();
1013 Value *V2 = EmitAddTreeOfValues(I, Ops);
1014 return CreateAdd(V2, V1, "reass.add", I, I);
1015}
1016
1017/// If V is an expression tree that is a multiplication sequence,
1018/// and if this sequence contains a multiply by Factor,
1019/// remove Factor from the tree and return the new tree.
1020Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1021 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1022 if (!BO)
1023 return nullptr;
1024
1025 SmallVector<RepeatedValue, 8> Tree;
1026 MadeChange |= LinearizeExprTree(BO, Tree);
1027 SmallVector<ValueEntry, 8> Factors;
1028 Factors.reserve(Tree.size());
1029 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1030 RepeatedValue E = Tree[i];
1031 Factors.append(E.second.getZExtValue(),
1032 ValueEntry(getRank(E.first), E.first));
1033 }
1034
1035 bool FoundFactor = false;
1036 bool NeedsNegate = false;
1037 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1038 if (Factors[i].Op == Factor) {
1039 FoundFactor = true;
1040 Factors.erase(Factors.begin()+i);
1041 break;
1042 }
1043
1044 // If this is a negative version of this factor, remove it.
1045 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1046 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1047 if (FC1->getValue() == -FC2->getValue()) {
1048 FoundFactor = NeedsNegate = true;
1049 Factors.erase(Factors.begin()+i);
1050 break;
1051 }
1052 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1053 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1054 const APFloat &F1 = FC1->getValueAPF();
1055 APFloat F2(FC2->getValueAPF());
1056 F2.changeSign();
1057 if (F1.compare(F2) == APFloat::cmpEqual) {
1058 FoundFactor = NeedsNegate = true;
1059 Factors.erase(Factors.begin() + i);
1060 break;
1061 }
1062 }
1063 }
1064 }
1065
1066 if (!FoundFactor) {
1067 // Make sure to restore the operands to the expression tree.
1068 RewriteExprTree(BO, Factors);
1069 return nullptr;
1070 }
1071
1072 BasicBlock::iterator InsertPt = ++BO->getIterator();
1073
1074 // If this was just a single multiply, remove the multiply and return the only
1075 // remaining operand.
1076 if (Factors.size() == 1) {
1077 RedoInsts.insert(BO);
1078 V = Factors[0].Op;
1079 } else {
1080 RewriteExprTree(BO, Factors);
1081 V = BO;
1082 }
1083
1084 if (NeedsNegate)
1085 V = CreateNeg(V, "neg", &*InsertPt, BO);
1086
1087 return V;
1088}
1089
1090/// If V is a single-use multiply, recursively add its operands as factors,
1091/// otherwise add V to the list of factors.
1092///
1093/// Ops is the top-level list of add operands we're trying to factor.
1094static void FindSingleUseMultiplyFactors(Value *V,
1095 SmallVectorImpl<Value*> &Factors) {
1096 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1097 if (!BO) {
1098 Factors.push_back(V);
1099 return;
1100 }
1101
1102 // Otherwise, add the LHS and RHS to the list of factors.
1103 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1104 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1105}
1106
1107/// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1108/// This optimizes based on identities. If it can be reduced to a single Value,
1109/// it is returned, otherwise the Ops list is mutated as necessary.
1110static Value *OptimizeAndOrXor(unsigned Opcode,
1111 SmallVectorImpl<ValueEntry> &Ops) {
1112 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1113 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1114 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1115 // First, check for X and ~X in the operand list.
1116 assert(i < Ops.size())((i < Ops.size()) ? static_cast<void> (0) : __assert_fail
("i < Ops.size()", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1116, __PRETTY_FUNCTION__))
;
1117 Value *X;
1118 if (match(Ops[i].Op, m_Not(m_Value(X)))) { // Cannot occur for ^.
1119 unsigned FoundX = FindInOperandList(Ops, i, X);
1120 if (FoundX != i) {
1121 if (Opcode == Instruction::And) // ...&X&~X = 0
1122 return Constant::getNullValue(X->getType());
1123
1124 if (Opcode == Instruction::Or) // ...|X|~X = -1
1125 return Constant::getAllOnesValue(X->getType());
1126 }
1127 }
1128
1129 // Next, check for duplicate pairs of values, which we assume are next to
1130 // each other, due to our sorting criteria.
1131 assert(i < Ops.size())((i < Ops.size()) ? static_cast<void> (0) : __assert_fail
("i < Ops.size()", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1131, __PRETTY_FUNCTION__))
;
1132 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1133 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1134 // Drop duplicate values for And and Or.
1135 Ops.erase(Ops.begin()+i);
1136 --i; --e;
1137 ++NumAnnihil;
1138 continue;
1139 }
1140
1141 // Drop pairs of values for Xor.
1142 assert(Opcode == Instruction::Xor)((Opcode == Instruction::Xor) ? static_cast<void> (0) :
__assert_fail ("Opcode == Instruction::Xor", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1142, __PRETTY_FUNCTION__))
;
1143 if (e == 2)
1144 return Constant::getNullValue(Ops[0].Op->getType());
1145
1146 // Y ^ X^X -> Y
1147 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1148 i -= 1; e -= 2;
1149 ++NumAnnihil;
1150 }
1151 }
1152 return nullptr;
1153}
1154
1155/// Helper function of CombineXorOpnd(). It creates a bitwise-and
1156/// instruction with the given two operands, and return the resulting
1157/// instruction. There are two special cases: 1) if the constant operand is 0,
1158/// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1159/// be returned.
1160static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1161 const APInt &ConstOpnd) {
1162 if (ConstOpnd.isNullValue())
1163 return nullptr;
1164
1165 if (ConstOpnd.isAllOnesValue())
1166 return Opnd;
1167
1168 Instruction *I = BinaryOperator::CreateAnd(
1169 Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1170 InsertBefore);
1171 I->setDebugLoc(InsertBefore->getDebugLoc());
1172 return I;
1173}
1174
1175// Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1176// into "R ^ C", where C would be 0, and R is a symbolic value.
1177//
1178// If it was successful, true is returned, and the "R" and "C" is returned
1179// via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1180// and both "Res" and "ConstOpnd" remain unchanged.
1181bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1182 APInt &ConstOpnd, Value *&Res) {
1183 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1184 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1185 // = (x & ~c1) ^ (c1 ^ c2)
1186 // It is useful only when c1 == c2.
1187 if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue())
1188 return false;
1189
1190 if (!Opnd1->getValue()->hasOneUse())
1191 return false;
1192
1193 const APInt &C1 = Opnd1->getConstPart();
1194 if (C1 != ConstOpnd)
1195 return false;
1196
1197 Value *X = Opnd1->getSymbolicPart();
1198 Res = createAndInstr(I, X, ~C1);
1199 // ConstOpnd was C2, now C1 ^ C2.
1200 ConstOpnd ^= C1;
1201
1202 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1203 RedoInsts.insert(T);
1204 return true;
1205}
1206
1207// Helper function of OptimizeXor(). It tries to simplify
1208// "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1209// symbolic value.
1210//
1211// If it was successful, true is returned, and the "R" and "C" is returned
1212// via "Res" and "ConstOpnd", respectively (If the entire expression is
1213// evaluated to a constant, the Res is set to NULL); otherwise, false is
1214// returned, and both "Res" and "ConstOpnd" remain unchanged.
1215bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1216 XorOpnd *Opnd2, APInt &ConstOpnd,
1217 Value *&Res) {
1218 Value *X = Opnd1->getSymbolicPart();
1219 if (X != Opnd2->getSymbolicPart())
1220 return false;
1221
1222 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1223 int DeadInstNum = 1;
1224 if (Opnd1->getValue()->hasOneUse())
1225 DeadInstNum++;
1226 if (Opnd2->getValue()->hasOneUse())
1227 DeadInstNum++;
1228
1229 // Xor-Rule 2:
1230 // (x | c1) ^ (x & c2)
1231 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1232 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1233 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1234 //
1235 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1236 if (Opnd2->isOrExpr())
1237 std::swap(Opnd1, Opnd2);
1238
1239 const APInt &C1 = Opnd1->getConstPart();
1240 const APInt &C2 = Opnd2->getConstPart();
1241 APInt C3((~C1) ^ C2);
1242
1243 // Do not increase code size!
1244 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1245 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1246 if (NewInstNum > DeadInstNum)
1247 return false;
1248 }
1249
1250 Res = createAndInstr(I, X, C3);
1251 ConstOpnd ^= C1;
1252 } else if (Opnd1->isOrExpr()) {
1253 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1254 //
1255 const APInt &C1 = Opnd1->getConstPart();
1256 const APInt &C2 = Opnd2->getConstPart();
1257 APInt C3 = C1 ^ C2;
1258
1259 // Do not increase code size
1260 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1261 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1262 if (NewInstNum > DeadInstNum)
1263 return false;
1264 }
1265
1266 Res = createAndInstr(I, X, C3);
1267 ConstOpnd ^= C3;
1268 } else {
1269 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1270 //
1271 const APInt &C1 = Opnd1->getConstPart();
1272 const APInt &C2 = Opnd2->getConstPart();
1273 APInt C3 = C1 ^ C2;
1274 Res = createAndInstr(I, X, C3);
1275 }
1276
1277 // Put the original operands in the Redo list; hope they will be deleted
1278 // as dead code.
1279 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1280 RedoInsts.insert(T);
1281 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1282 RedoInsts.insert(T);
1283
1284 return true;
1285}
1286
1287/// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1288/// to a single Value, it is returned, otherwise the Ops list is mutated as
1289/// necessary.
1290Value *ReassociatePass::OptimizeXor(Instruction *I,
1291 SmallVectorImpl<ValueEntry> &Ops) {
1292 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1293 return V;
1294
1295 if (Ops.size() == 1)
1296 return nullptr;
1297
1298 SmallVector<XorOpnd, 8> Opnds;
1299 SmallVector<XorOpnd*, 8> OpndPtrs;
1300 Type *Ty = Ops[0].Op->getType();
1301 APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1302
1303 // Step 1: Convert ValueEntry to XorOpnd
1304 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1305 Value *V = Ops[i].Op;
1306 const APInt *C;
1307 // TODO: Support non-splat vectors.
1308 if (match(V, m_APInt(C))) {
1309 ConstOpnd ^= *C;
1310 } else {
1311 XorOpnd O(V);
1312 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1313 Opnds.push_back(O);
1314 }
1315 }
1316
1317 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1318 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1319 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1320 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1321 // when new elements are added to the vector.
1322 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1323 OpndPtrs.push_back(&Opnds[i]);
1324
1325 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1326 // the same symbolic value cluster together. For instance, the input operand
1327 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1328 // ("x | 123", "x & 789", "y & 456").
1329 //
1330 // The purpose is twofold:
1331 // 1) Cluster together the operands sharing the same symbolic-value.
1332 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1333 // could potentially shorten crital path, and expose more loop-invariants.
1334 // Note that values' rank are basically defined in RPO order (FIXME).
1335 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1336 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1337 // "z" in the order of X-Y-Z is better than any other orders.
1338 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1339 [](XorOpnd *LHS, XorOpnd *RHS) {
1340 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1341 });
1342
1343 // Step 3: Combine adjacent operands
1344 XorOpnd *PrevOpnd = nullptr;
1345 bool Changed = false;
1346 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1347 XorOpnd *CurrOpnd = OpndPtrs[i];
1348 // The combined value
1349 Value *CV;
1350
1351 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1352 if (!ConstOpnd.isNullValue() &&
1353 CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1354 Changed = true;
1355 if (CV)
1356 *CurrOpnd = XorOpnd(CV);
1357 else {
1358 CurrOpnd->Invalidate();
1359 continue;
1360 }
1361 }
1362
1363 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1364 PrevOpnd = CurrOpnd;
1365 continue;
1366 }
1367
1368 // step 3.2: When previous and current operands share the same symbolic
1369 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1370 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1371 // Remove previous operand
1372 PrevOpnd->Invalidate();
1373 if (CV) {
1374 *CurrOpnd = XorOpnd(CV);
1375 PrevOpnd = CurrOpnd;
1376 } else {
1377 CurrOpnd->Invalidate();
1378 PrevOpnd = nullptr;
1379 }
1380 Changed = true;
1381 }
1382 }
1383
1384 // Step 4: Reassemble the Ops
1385 if (Changed) {
1386 Ops.clear();
1387 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1388 XorOpnd &O = Opnds[i];
1389 if (O.isInvalid())
1390 continue;
1391 ValueEntry VE(getRank(O.getValue()), O.getValue());
1392 Ops.push_back(VE);
1393 }
1394 if (!ConstOpnd.isNullValue()) {
1395 Value *C = ConstantInt::get(Ty, ConstOpnd);
1396 ValueEntry VE(getRank(C), C);
1397 Ops.push_back(VE);
1398 }
1399 unsigned Sz = Ops.size();
1400 if (Sz == 1)
1401 return Ops.back().Op;
1402 if (Sz == 0) {
1403 assert(ConstOpnd.isNullValue())((ConstOpnd.isNullValue()) ? static_cast<void> (0) : __assert_fail
("ConstOpnd.isNullValue()", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1403, __PRETTY_FUNCTION__))
;
1404 return ConstantInt::get(Ty, ConstOpnd);
1405 }
1406 }
1407
1408 return nullptr;
1409}
1410
1411/// Optimize a series of operands to an 'add' instruction. This
1412/// optimizes based on identities. If it can be reduced to a single Value, it
1413/// is returned, otherwise the Ops list is mutated as necessary.
1414Value *ReassociatePass::OptimizeAdd(Instruction *I,
1415 SmallVectorImpl<ValueEntry> &Ops) {
1416 // Scan the operand lists looking for X and -X pairs. If we find any, we
1417 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1418 // scan for any
1419 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1420
1421 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1422 Value *TheOp = Ops[i].Op;
1423 // Check to see if we've seen this operand before. If so, we factor all
1424 // instances of the operand together. Due to our sorting criteria, we know
1425 // that these need to be next to each other in the vector.
1426 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1427 // Rescan the list, remove all instances of this operand from the expr.
1428 unsigned NumFound = 0;
1429 do {
1430 Ops.erase(Ops.begin()+i);
1431 ++NumFound;
1432 } while (i != Ops.size() && Ops[i].Op == TheOp);
1433
1434 LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOpdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << NumFound
<< "]: " << *TheOp << '\n'; } } while (false
)
1435 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << NumFound
<< "]: " << *TheOp << '\n'; } } while (false
)
;
1436 ++NumFactor;
1437
1438 // Insert a new multiply.
1439 Type *Ty = TheOp->getType();
1440 Constant *C = Ty->isIntOrIntVectorTy() ?
1441 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1442 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1443
1444 // Now that we have inserted a multiply, optimize it. This allows us to
1445 // handle cases that require multiple factoring steps, such as this:
1446 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1447 RedoInsts.insert(Mul);
1448
1449 // If every add operand was a duplicate, return the multiply.
1450 if (Ops.empty())
1451 return Mul;
1452
1453 // Otherwise, we had some input that didn't have the dupe, such as
1454 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1455 // things being added by this operation.
1456 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1457
1458 --i;
1459 e = Ops.size();
1460 continue;
1461 }
1462
1463 // Check for X and -X or X and ~X in the operand list.
1464 Value *X;
1465 if (!match(TheOp, m_Neg(m_Value(X))) && !match(TheOp, m_Not(m_Value(X))) &&
1466 !match(TheOp, m_FNeg(m_Value(X))))
1467 continue;
1468
1469 unsigned FoundX = FindInOperandList(Ops, i, X);
1470 if (FoundX == i)
1471 continue;
1472
1473 // Remove X and -X from the operand list.
1474 if (Ops.size() == 2 &&
1475 (match(TheOp, m_Neg(m_Value())) || match(TheOp, m_FNeg(m_Value()))))
1476 return Constant::getNullValue(X->getType());
1477
1478 // Remove X and ~X from the operand list.
1479 if (Ops.size() == 2 && match(TheOp, m_Not(m_Value())))
1480 return Constant::getAllOnesValue(X->getType());
1481
1482 Ops.erase(Ops.begin()+i);
1483 if (i < FoundX)
1484 --FoundX;
1485 else
1486 --i; // Need to back up an extra one.
1487 Ops.erase(Ops.begin()+FoundX);
1488 ++NumAnnihil;
1489 --i; // Revisit element.
1490 e -= 2; // Removed two elements.
1491
1492 // if X and ~X we append -1 to the operand list.
1493 if (match(TheOp, m_Not(m_Value()))) {
1494 Value *V = Constant::getAllOnesValue(X->getType());
1495 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1496 e += 1;
1497 }
1498 }
1499
1500 // Scan the operand list, checking to see if there are any common factors
1501 // between operands. Consider something like A*A+A*B*C+D. We would like to
1502 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1503 // To efficiently find this, we count the number of times a factor occurs
1504 // for any ADD operands that are MULs.
1505 DenseMap<Value*, unsigned> FactorOccurrences;
1506
1507 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1508 // where they are actually the same multiply.
1509 unsigned MaxOcc = 0;
1510 Value *MaxOccVal = nullptr;
1511 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1512 BinaryOperator *BOp =
1513 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1514 if (!BOp)
1515 continue;
1516
1517 // Compute all of the factors of this added value.
1518 SmallVector<Value*, 8> Factors;
1519 FindSingleUseMultiplyFactors(BOp, Factors);
1520 assert(Factors.size() > 1 && "Bad linearize!")((Factors.size() > 1 && "Bad linearize!") ? static_cast
<void> (0) : __assert_fail ("Factors.size() > 1 && \"Bad linearize!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1520, __PRETTY_FUNCTION__))
;
1521
1522 // Add one to FactorOccurrences for each unique factor in this op.
1523 SmallPtrSet<Value*, 8> Duplicates;
1524 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1525 Value *Factor = Factors[i];
1526 if (!Duplicates.insert(Factor).second)
1527 continue;
1528
1529 unsigned Occ = ++FactorOccurrences[Factor];
1530 if (Occ > MaxOcc) {
1531 MaxOcc = Occ;
1532 MaxOccVal = Factor;
1533 }
1534
1535 // If Factor is a negative constant, add the negated value as a factor
1536 // because we can percolate the negate out. Watch for minint, which
1537 // cannot be positivified.
1538 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1539 if (CI->isNegative() && !CI->isMinValue(true)) {
1540 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1541 if (!Duplicates.insert(Factor).second)
1542 continue;
1543 unsigned Occ = ++FactorOccurrences[Factor];
1544 if (Occ > MaxOcc) {
1545 MaxOcc = Occ;
1546 MaxOccVal = Factor;
1547 }
1548 }
1549 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1550 if (CF->isNegative()) {
1551 APFloat F(CF->getValueAPF());
1552 F.changeSign();
1553 Factor = ConstantFP::get(CF->getContext(), F);
1554 if (!Duplicates.insert(Factor).second)
1555 continue;
1556 unsigned Occ = ++FactorOccurrences[Factor];
1557 if (Occ > MaxOcc) {
1558 MaxOcc = Occ;
1559 MaxOccVal = Factor;
1560 }
1561 }
1562 }
1563 }
1564 }
1565
1566 // If any factor occurred more than one time, we can pull it out.
1567 if (MaxOcc > 1) {
1568 LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc
<< "]: " << *MaxOccVal << '\n'; } } while (
false)
1569 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc
<< "]: " << *MaxOccVal << '\n'; } } while (
false)
;
1570 ++NumFactor;
1571
1572 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1573 // this, we could otherwise run into situations where removing a factor
1574 // from an expression will drop a use of maxocc, and this can cause
1575 // RemoveFactorFromExpression on successive values to behave differently.
1576 Instruction *DummyInst =
1577 I->getType()->isIntOrIntVectorTy()
1578 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1579 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1580
1581 SmallVector<WeakTrackingVH, 4> NewMulOps;
1582 for (unsigned i = 0; i != Ops.size(); ++i) {
1583 // Only try to remove factors from expressions we're allowed to.
1584 BinaryOperator *BOp =
1585 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1586 if (!BOp)
1587 continue;
1588
1589 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1590 // The factorized operand may occur several times. Convert them all in
1591 // one fell swoop.
1592 for (unsigned j = Ops.size(); j != i;) {
1593 --j;
1594 if (Ops[j].Op == Ops[i].Op) {
1595 NewMulOps.push_back(V);
1596 Ops.erase(Ops.begin()+j);
1597 }
1598 }
1599 --i;
1600 }
1601 }
1602
1603 // No need for extra uses anymore.
1604 DummyInst->deleteValue();
1605
1606 unsigned NumAddedValues = NewMulOps.size();
1607 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1608
1609 // Now that we have inserted the add tree, optimize it. This allows us to
1610 // handle cases that require multiple factoring steps, such as this:
1611 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1612 assert(NumAddedValues > 1 && "Each occurrence should contribute a value")((NumAddedValues > 1 && "Each occurrence should contribute a value"
) ? static_cast<void> (0) : __assert_fail ("NumAddedValues > 1 && \"Each occurrence should contribute a value\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1612, __PRETTY_FUNCTION__))
;
1613 (void)NumAddedValues;
1614 if (Instruction *VI = dyn_cast<Instruction>(V))
1615 RedoInsts.insert(VI);
1616
1617 // Create the multiply.
1618 Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I);
1619
1620 // Rerun associate on the multiply in case the inner expression turned into
1621 // a multiply. We want to make sure that we keep things in canonical form.
1622 RedoInsts.insert(V2);
1623
1624 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1625 // entire result expression is just the multiply "A*(B+C)".
1626 if (Ops.empty())
1627 return V2;
1628
1629 // Otherwise, we had some input that didn't have the factor, such as
1630 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1631 // things being added by this operation.
1632 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1633 }
1634
1635 return nullptr;
1636}
1637
1638/// Build up a vector of value/power pairs factoring a product.
1639///
1640/// Given a series of multiplication operands, build a vector of factors and
1641/// the powers each is raised to when forming the final product. Sort them in
1642/// the order of descending power.
1643///
1644/// (x*x) -> [(x, 2)]
1645/// ((x*x)*x) -> [(x, 3)]
1646/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1647///
1648/// \returns Whether any factors have a power greater than one.
1649static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1650 SmallVectorImpl<Factor> &Factors) {
1651 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1652 // Compute the sum of powers of simplifiable factors.
1653 unsigned FactorPowerSum = 0;
1654 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1655 Value *Op = Ops[Idx-1].Op;
1656
1657 // Count the number of occurrences of this value.
1658 unsigned Count = 1;
1659 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1660 ++Count;
1661 // Track for simplification all factors which occur 2 or more times.
1662 if (Count > 1)
1663 FactorPowerSum += Count;
1664 }
1665
1666 // We can only simplify factors if the sum of the powers of our simplifiable
1667 // factors is 4 or higher. When that is the case, we will *always* have
1668 // a simplification. This is an important invariant to prevent cyclicly
1669 // trying to simplify already minimal formations.
1670 if (FactorPowerSum < 4)
1671 return false;
1672
1673 // Now gather the simplifiable factors, removing them from Ops.
1674 FactorPowerSum = 0;
1675 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1676 Value *Op = Ops[Idx-1].Op;
1677
1678 // Count the number of occurrences of this value.
1679 unsigned Count = 1;
1680 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1681 ++Count;
1682 if (Count == 1)
1683 continue;
1684 // Move an even number of occurrences to Factors.
1685 Count &= ~1U;
1686 Idx -= Count;
1687 FactorPowerSum += Count;
1688 Factors.push_back(Factor(Op, Count));
1689 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1690 }
1691
1692 // None of the adjustments above should have reduced the sum of factor powers
1693 // below our mininum of '4'.
1694 assert(FactorPowerSum >= 4)((FactorPowerSum >= 4) ? static_cast<void> (0) : __assert_fail
("FactorPowerSum >= 4", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1694, __PRETTY_FUNCTION__))
;
1695
1696 std::stable_sort(Factors.begin(), Factors.end(),
1697 [](const Factor &LHS, const Factor &RHS) {
1698 return LHS.Power > RHS.Power;
1699 });
1700 return true;
1701}
1702
1703/// Build a tree of multiplies, computing the product of Ops.
1704static Value *buildMultiplyTree(IRBuilder<> &Builder,
1705 SmallVectorImpl<Value*> &Ops) {
1706 if (Ops.size() == 1)
1707 return Ops.back();
1708
1709 Value *LHS = Ops.pop_back_val();
1710 do {
1711 if (LHS->getType()->isIntOrIntVectorTy())
1712 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1713 else
1714 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1715 } while (!Ops.empty());
1716
1717 return LHS;
1718}
1719
1720/// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1721///
1722/// Given a vector of values raised to various powers, where no two values are
1723/// equal and the powers are sorted in decreasing order, compute the minimal
1724/// DAG of multiplies to compute the final product, and return that product
1725/// value.
1726Value *
1727ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1728 SmallVectorImpl<Factor> &Factors) {
1729 assert(Factors[0].Power)((Factors[0].Power) ? static_cast<void> (0) : __assert_fail
("Factors[0].Power", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1729, __PRETTY_FUNCTION__))
;
1730 SmallVector<Value *, 4> OuterProduct;
1731 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1732 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1733 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1734 LastIdx = Idx;
1735 continue;
1736 }
1737
1738 // We want to multiply across all the factors with the same power so that
1739 // we can raise them to that power as a single entity. Build a mini tree
1740 // for that.
1741 SmallVector<Value *, 4> InnerProduct;
1742 InnerProduct.push_back(Factors[LastIdx].Base);
1743 do {
1744 InnerProduct.push_back(Factors[Idx].Base);
1745 ++Idx;
1746 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1747
1748 // Reset the base value of the first factor to the new expression tree.
1749 // We'll remove all the factors with the same power in a second pass.
1750 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1751 if (Instruction *MI = dyn_cast<Instruction>(M))
1752 RedoInsts.insert(MI);
1753
1754 LastIdx = Idx;
1755 }
1756 // Unique factors with equal powers -- we've folded them into the first one's
1757 // base.
1758 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1759 [](const Factor &LHS, const Factor &RHS) {
1760 return LHS.Power == RHS.Power;
1761 }),
1762 Factors.end());
1763
1764 // Iteratively collect the base of each factor with an add power into the
1765 // outer product, and halve each power in preparation for squaring the
1766 // expression.
1767 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1768 if (Factors[Idx].Power & 1)
1769 OuterProduct.push_back(Factors[Idx].Base);
1770 Factors[Idx].Power >>= 1;
1771 }
1772 if (Factors[0].Power) {
1773 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1774 OuterProduct.push_back(SquareRoot);
1775 OuterProduct.push_back(SquareRoot);
1776 }
1777 if (OuterProduct.size() == 1)
1778 return OuterProduct.front();
1779
1780 Value *V = buildMultiplyTree(Builder, OuterProduct);
1781 return V;
1782}
1783
1784Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1785 SmallVectorImpl<ValueEntry> &Ops) {
1786 // We can only optimize the multiplies when there is a chain of more than
1787 // three, such that a balanced tree might require fewer total multiplies.
1788 if (Ops.size() < 4)
1789 return nullptr;
1790
1791 // Try to turn linear trees of multiplies without other uses of the
1792 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1793 // re-use.
1794 SmallVector<Factor, 4> Factors;
1795 if (!collectMultiplyFactors(Ops, Factors))
1796 return nullptr; // All distinct factors, so nothing left for us to do.
1797
1798 IRBuilder<> Builder(I);
1799 // The reassociate transformation for FP operations is performed only
1800 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1801 // to the newly generated operations.
1802 if (auto FPI = dyn_cast<FPMathOperator>(I))
1803 Builder.setFastMathFlags(FPI->getFastMathFlags());
1804
1805 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1806 if (Ops.empty())
1807 return V;
1808
1809 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1810 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1811 return nullptr;
1812}
1813
1814Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1815 SmallVectorImpl<ValueEntry> &Ops) {
1816 // Now that we have the linearized expression tree, try to optimize it.
1817 // Start by folding any constants that we found.
1818 Constant *Cst = nullptr;
1819 unsigned Opcode = I->getOpcode();
1820 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1821 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1822 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1823 }
1824 // If there was nothing but constants then we are done.
1825 if (Ops.empty())
1826 return Cst;
1827
1828 // Put the combined constant back at the end of the operand list, except if
1829 // there is no point. For example, an add of 0 gets dropped here, while a
1830 // multiplication by zero turns the whole expression into zero.
1831 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1832 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1833 return Cst;
1834 Ops.push_back(ValueEntry(0, Cst));
1835 }
1836
1837 if (Ops.size() == 1) return Ops[0].Op;
1838
1839 // Handle destructive annihilation due to identities between elements in the
1840 // argument list here.
1841 unsigned NumOps = Ops.size();
1842 switch (Opcode) {
1843 default: break;
1844 case Instruction::And:
1845 case Instruction::Or:
1846 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1847 return Result;
1848 break;
1849
1850 case Instruction::Xor:
1851 if (Value *Result = OptimizeXor(I, Ops))
1852 return Result;
1853 break;
1854
1855 case Instruction::Add:
1856 case Instruction::FAdd:
1857 if (Value *Result = OptimizeAdd(I, Ops))
1858 return Result;
1859 break;
1860
1861 case Instruction::Mul:
1862 case Instruction::FMul:
1863 if (Value *Result = OptimizeMul(I, Ops))
1864 return Result;
1865 break;
1866 }
1867
1868 if (Ops.size() != NumOps)
1869 return OptimizeExpression(I, Ops);
1870 return nullptr;
1871}
1872
1873// Remove dead instructions and if any operands are trivially dead add them to
1874// Insts so they will be removed as well.
1875void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I,
1876 OrderedSet &Insts) {
1877 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!")((isInstructionTriviallyDead(I) && "Trivially dead instructions only!"
) ? static_cast<void> (0) : __assert_fail ("isInstructionTriviallyDead(I) && \"Trivially dead instructions only!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1877, __PRETTY_FUNCTION__))
;
1878 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1879 ValueRankMap.erase(I);
1880 Insts.remove(I);
1881 RedoInsts.remove(I);
1882 I->eraseFromParent();
1883 for (auto Op : Ops)
1884 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1885 if (OpInst->use_empty())
1886 Insts.insert(OpInst);
1887}
1888
1889/// Zap the given instruction, adding interesting operands to the work list.
1890void ReassociatePass::EraseInst(Instruction *I) {
1891 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!")((isInstructionTriviallyDead(I) && "Trivially dead instructions only!"
) ? static_cast<void> (0) : __assert_fail ("isInstructionTriviallyDead(I) && \"Trivially dead instructions only!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1891, __PRETTY_FUNCTION__))
;
1892 LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Erasing dead inst: "; I->
dump(); } } while (false)
;
1893
1894 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1895 // Erase the dead instruction.
1896 ValueRankMap.erase(I);
1897 RedoInsts.remove(I);
1898 I->eraseFromParent();
1899 // Optimize its operands.
1900 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1901 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1902 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1903 // If this is a node in an expression tree, climb to the expression root
1904 // and add that since that's where optimization actually happens.
1905 unsigned Opcode = Op->getOpcode();
1906 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1907 Visited.insert(Op).second)
1908 Op = Op->user_back();
1909
1910 // The instruction we're going to push may be coming from a
1911 // dead block, and Reassociate skips the processing of unreachable
1912 // blocks because it's a waste of time and also because it can
1913 // lead to infinite loop due to LLVM's non-standard definition
1914 // of dominance.
1915 if (ValueRankMap.find(Op) != ValueRankMap.end())
1916 RedoInsts.insert(Op);
1917 }
1918
1919 MadeChange = true;
1920}
1921
1922// Canonicalize expressions of the following form:
1923// x + (-Constant * y) -> x - (Constant * y)
1924// x - (-Constant * y) -> x + (Constant * y)
1925Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1926 if (!I->hasOneUse() || I->getType()->isVectorTy())
1927 return nullptr;
1928
1929 // Must be a fmul or fdiv instruction.
1930 unsigned Opcode = I->getOpcode();
1931 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1932 return nullptr;
1933
1934 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1935 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1936
1937 // Both operands are constant, let it get constant folded away.
1938 if (C0 && C1)
1939 return nullptr;
1940
1941 ConstantFP *CF = C0 ? C0 : C1;
1942
1943 // Must have one constant operand.
1944 if (!CF)
1945 return nullptr;
1946
1947 // Must be a negative ConstantFP.
1948 if (!CF->isNegative())
1949 return nullptr;
1950
1951 // User must be a binary operator with one or more uses.
1952 Instruction *User = I->user_back();
1953 if (!isa<BinaryOperator>(User) || User->use_empty())
1954 return nullptr;
1955
1956 unsigned UserOpcode = User->getOpcode();
1957 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1958 return nullptr;
1959
1960 // Subtraction is not commutative. Explicitly, the following transform is
1961 // not valid: (-Constant * y) - x -> x + (Constant * y)
1962 if (!User->isCommutative() && User->getOperand(1) != I)
1963 return nullptr;
1964
1965 // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
1966 // resulting subtract will be broken up later. This can get us into an
1967 // infinite loop during reassociation.
1968 if (UserOpcode == Instruction::FAdd && ShouldBreakUpSubtract(User))
1969 return nullptr;
1970
1971 // Change the sign of the constant.
1972 APFloat Val = CF->getValueAPF();
1973 Val.changeSign();
1974 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1975
1976 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1977 // ((-Const*y) + x) -> (x + (-Const*y)).
1978 if (User->getOperand(0) == I && User->isCommutative())
1979 cast<BinaryOperator>(User)->swapOperands();
1980
1981 Value *Op0 = User->getOperand(0);
1982 Value *Op1 = User->getOperand(1);
1983 BinaryOperator *NI;
1984 switch (UserOpcode) {
1985 default:
1986 llvm_unreachable("Unexpected Opcode!")::llvm::llvm_unreachable_internal("Unexpected Opcode!", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 1986)
;
1987 case Instruction::FAdd:
1988 NI = BinaryOperator::CreateFSub(Op0, Op1);
1989 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1990 break;
1991 case Instruction::FSub:
1992 NI = BinaryOperator::CreateFAdd(Op0, Op1);
1993 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1994 break;
1995 }
1996
1997 NI->insertBefore(User);
1998 NI->setName(User->getName());
1999 User->replaceAllUsesWith(NI);
2000 NI->setDebugLoc(I->getDebugLoc());
2001 RedoInsts.insert(I);
2002 MadeChange = true;
2003 return NI;
2004}
2005
2006/// Inspect and optimize the given instruction. Note that erasing
2007/// instructions is not allowed.
2008void ReassociatePass::OptimizeInst(Instruction *I) {
2009 // Only consider operations that we understand.
2010 if (!isa<BinaryOperator>(I))
2011 return;
2012
2013 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2014 // If an operand of this shift is a reassociable multiply, or if the shift
2015 // is used by a reassociable multiply or add, turn into a multiply.
2016 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2017 (I->hasOneUse() &&
2018 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2019 isReassociableOp(I->user_back(), Instruction::Add)))) {
2020 Instruction *NI = ConvertShiftToMul(I);
2021 RedoInsts.insert(I);
2022 MadeChange = true;
2023 I = NI;
2024 }
2025
2026 // Canonicalize negative constants out of expressions.
2027 if (Instruction *Res = canonicalizeNegConstExpr(I))
2028 I = Res;
2029
2030 // Commute binary operators, to canonicalize the order of their operands.
2031 // This can potentially expose more CSE opportunities, and makes writing other
2032 // transformations simpler.
2033 if (I->isCommutative())
2034 canonicalizeOperands(I);
2035
2036 // Don't optimize floating-point instructions unless they are 'fast'.
2037 if (I->getType()->isFPOrFPVectorTy() && !I->isFast())
2038 return;
2039
2040 // Do not reassociate boolean (i1) expressions. We want to preserve the
2041 // original order of evaluation for short-circuited comparisons that
2042 // SimplifyCFG has folded to AND/OR expressions. If the expression
2043 // is not further optimized, it is likely to be transformed back to a
2044 // short-circuited form for code gen, and the source order may have been
2045 // optimized for the most likely conditions.
2046 if (I->getType()->isIntegerTy(1))
2047 return;
2048
2049 // If this is a subtract instruction which is not already in negate form,
2050 // see if we can convert it to X+-Y.
2051 if (I->getOpcode() == Instruction::Sub) {
2052 if (ShouldBreakUpSubtract(I)) {
2053 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2054 RedoInsts.insert(I);
2055 MadeChange = true;
2056 I = NI;
2057 } else if (match(I, m_Neg(m_Value()))) {
2058 // Otherwise, this is a negation. See if the operand is a multiply tree
2059 // and if this is not an inner node of a multiply tree.
2060 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2061 (!I->hasOneUse() ||
2062 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2063 Instruction *NI = LowerNegateToMultiply(I);
2064 // If the negate was simplified, revisit the users to see if we can
2065 // reassociate further.
2066 for (User *U : NI->users()) {
2067 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2068 RedoInsts.insert(Tmp);
2069 }
2070 RedoInsts.insert(I);
2071 MadeChange = true;
2072 I = NI;
2073 }
2074 }
2075 } else if (I->getOpcode() == Instruction::FSub) {
2076 if (ShouldBreakUpSubtract(I)) {
2077 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2078 RedoInsts.insert(I);
2079 MadeChange = true;
2080 I = NI;
2081 } else if (match(I, m_FNeg(m_Value()))) {
2082 // Otherwise, this is a negation. See if the operand is a multiply tree
2083 // and if this is not an inner node of a multiply tree.
2084 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2085 (!I->hasOneUse() ||
2086 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2087 // If the negate was simplified, revisit the users to see if we can
2088 // reassociate further.
2089 Instruction *NI = LowerNegateToMultiply(I);
2090 for (User *U : NI->users()) {
2091 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2092 RedoInsts.insert(Tmp);
2093 }
2094 RedoInsts.insert(I);
2095 MadeChange = true;
2096 I = NI;
2097 }
2098 }
2099 }
2100
2101 // If this instruction is an associative binary operator, process it.
2102 if (!I->isAssociative()) return;
2103 BinaryOperator *BO = cast<BinaryOperator>(I);
2104
2105 // If this is an interior node of a reassociable tree, ignore it until we
2106 // get to the root of the tree, to avoid N^2 analysis.
2107 unsigned Opcode = BO->getOpcode();
2108 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2109 // During the initial run we will get to the root of the tree.
2110 // But if we get here while we are redoing instructions, there is no
2111 // guarantee that the root will be visited. So Redo later
2112 if (BO->user_back() != BO &&
2113 BO->getParent() == BO->user_back()->getParent())
2114 RedoInsts.insert(BO->user_back());
2115 return;
2116 }
2117
2118 // If this is an add tree that is used by a sub instruction, ignore it
2119 // until we process the subtract.
2120 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2121 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2122 return;
2123 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2124 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2125 return;
2126
2127 ReassociateExpression(BO);
2128}
2129
2130void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2131 // First, walk the expression tree, linearizing the tree, collecting the
2132 // operand information.
2133 SmallVector<RepeatedValue, 8> Tree;
2134 MadeChange |= LinearizeExprTree(I, Tree);
2135 SmallVector<ValueEntry, 8> Ops;
2136 Ops.reserve(Tree.size());
2137 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2138 RepeatedValue E = Tree[i];
2139 Ops.append(E.second.getZExtValue(),
2140 ValueEntry(getRank(E.first), E.first));
2141 }
2142
2143 LLVM_DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RAIn:\t"; PrintOps(I, Ops
); dbgs() << '\n'; } } while (false)
;
2144
2145 // Now that we have linearized the tree to a list and have gathered all of
2146 // the operands and their ranks, sort the operands by their rank. Use a
2147 // stable_sort so that values with equal ranks will have their relative
2148 // positions maintained (and so the compiler is deterministic). Note that
2149 // this sorts so that the highest ranking values end up at the beginning of
2150 // the vector.
2151 std::stable_sort(Ops.begin(), Ops.end());
2152
2153 // Now that we have the expression tree in a convenient
2154 // sorted form, optimize it globally if possible.
2155 if (Value *V = OptimizeExpression(I, Ops)) {
2156 if (V == I)
2157 // Self-referential expression in unreachable code.
2158 return;
2159 // This expression tree simplified to something that isn't a tree,
2160 // eliminate it.
2161 LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Reassoc to scalar: " <<
*V << '\n'; } } while (false)
;
2162 I->replaceAllUsesWith(V);
2163 if (Instruction *VI = dyn_cast<Instruction>(V))
2164 if (I->getDebugLoc())
2165 VI->setDebugLoc(I->getDebugLoc());
2166 RedoInsts.insert(I);
2167 ++NumAnnihil;
2168 return;
2169 }
2170
2171 // We want to sink immediates as deeply as possible except in the case where
2172 // this is a multiply tree used only by an add, and the immediate is a -1.
2173 // In this case we reassociate to put the negation on the outside so that we
2174 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2175 if (I->hasOneUse()) {
2176 if (I->getOpcode() == Instruction::Mul &&
2177 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2178 isa<ConstantInt>(Ops.back().Op) &&
2179 cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2180 ValueEntry Tmp = Ops.pop_back_val();
2181 Ops.insert(Ops.begin(), Tmp);
2182 } else if (I->getOpcode() == Instruction::FMul &&
2183 cast<Instruction>(I->user_back())->getOpcode() ==
2184 Instruction::FAdd &&
2185 isa<ConstantFP>(Ops.back().Op) &&
2186 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2187 ValueEntry Tmp = Ops.pop_back_val();
2188 Ops.insert(Ops.begin(), Tmp);
2189 }
2190 }
2191
2192 LLVM_DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RAOut:\t"; PrintOps(I, Ops
); dbgs() << '\n'; } } while (false)
;
2193
2194 if (Ops.size() == 1) {
2195 if (Ops[0].Op == I)
2196 // Self-referential expression in unreachable code.
2197 return;
2198
2199 // This expression tree simplified to something that isn't a tree,
2200 // eliminate it.
2201 I->replaceAllUsesWith(Ops[0].Op);
2202 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2203 OI->setDebugLoc(I->getDebugLoc());
2204 RedoInsts.insert(I);
2205 return;
2206 }
2207
2208 if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) {
2209 // Find the pair with the highest count in the pairmap and move it to the
2210 // back of the list so that it can later be CSE'd.
2211 // example:
2212 // a*b*c*d*e
2213 // if c*e is the most "popular" pair, we can express this as
2214 // (((c*e)*d)*b)*a
2215 unsigned Max = 1;
2216 unsigned BestRank = 0;
2217 std::pair<unsigned, unsigned> BestPair;
2218 unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin;
2219 for (unsigned i = 0; i < Ops.size() - 1; ++i)
2220 for (unsigned j = i + 1; j < Ops.size(); ++j) {
2221 unsigned Score = 0;
2222 Value *Op0 = Ops[i].Op;
2223 Value *Op1 = Ops[j].Op;
2224 if (std::less<Value *>()(Op1, Op0))
2225 std::swap(Op0, Op1);
2226 auto it = PairMap[Idx].find({Op0, Op1});
2227 if (it != PairMap[Idx].end())
2228 Score += it->second;
2229
2230 unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank);
2231 if (Score > Max || (Score == Max && MaxRank < BestRank)) {
2232 BestPair = {i, j};
2233 Max = Score;
2234 BestRank = MaxRank;
2235 }
2236 }
2237 if (Max > 1) {
2238 auto Op0 = Ops[BestPair.first];
2239 auto Op1 = Ops[BestPair.second];
2240 Ops.erase(&Ops[BestPair.second]);
2241 Ops.erase(&Ops[BestPair.first]);
2242 Ops.push_back(Op0);
2243 Ops.push_back(Op1);
2244 }
2245 }
2246 // Now that we ordered and optimized the expressions, splat them back into
2247 // the expression tree, removing any unneeded nodes.
2248 RewriteExprTree(I, Ops);
2249}
2250
2251void
2252ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) {
2253 // Make a "pairmap" of how often each operand pair occurs.
2254 for (BasicBlock *BI : RPOT) {
2255 for (Instruction &I : *BI) {
2256 if (!I.isAssociative())
2257 continue;
2258
2259 // Ignore nodes that aren't at the root of trees.
2260 if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode())
2261 continue;
2262
2263 // Collect all operands in a single reassociable expression.
2264 // Since Reassociate has already been run once, we can assume things
2265 // are already canonical according to Reassociation's regime.
2266 SmallVector<Value *, 8> Worklist = { I.getOperand(0), I.getOperand(1) };
2267 SmallVector<Value *, 8> Ops;
2268 while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) {
2269 Value *Op = Worklist.pop_back_val();
2270 Instruction *OpI = dyn_cast<Instruction>(Op);
2271 if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) {
2272 Ops.push_back(Op);
2273 continue;
2274 }
2275 // Be paranoid about self-referencing expressions in unreachable code.
2276 if (OpI->getOperand(0) != OpI)
2277 Worklist.push_back(OpI->getOperand(0));
2278 if (OpI->getOperand(1) != OpI)
2279 Worklist.push_back(OpI->getOperand(1));
2280 }
2281 // Skip extremely long expressions.
2282 if (Ops.size() > GlobalReassociateLimit)
2283 continue;
2284
2285 // Add all pairwise combinations of operands to the pair map.
2286 unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin;
2287 SmallSet<std::pair<Value *, Value*>, 32> Visited;
2288 for (unsigned i = 0; i < Ops.size() - 1; ++i) {
2289 for (unsigned j = i + 1; j < Ops.size(); ++j) {
2290 // Canonicalize operand orderings.
2291 Value *Op0 = Ops[i];
2292 Value *Op1 = Ops[j];
2293 if (std::less<Value *>()(Op1, Op0))
2294 std::swap(Op0, Op1);
2295 if (!Visited.insert({Op0, Op1}).second)
2296 continue;
2297 auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, 1});
2298 if (!res.second)
2299 ++res.first->second;
2300 }
2301 }
2302 }
2303 }
2304}
2305
2306PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2307 // Get the functions basic blocks in Reverse Post Order. This order is used by
2308 // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2309 // blocks (it has been seen that the analysis in this pass could hang when
2310 // analysing dead basic blocks).
2311 ReversePostOrderTraversal<Function *> RPOT(&F);
2312
2313 // Calculate the rank map for F.
2314 BuildRankMap(F, RPOT);
4
Calling 'ReassociatePass::BuildRankMap'
2315
2316 // Build the pair map before running reassociate.
2317 // Technically this would be more accurate if we did it after one round
2318 // of reassociation, but in practice it doesn't seem to help much on
2319 // real-world code, so don't waste the compile time running reassociate
2320 // twice.
2321 // If a user wants, they could expicitly run reassociate twice in their
2322 // pass pipeline for further potential gains.
2323 // It might also be possible to update the pair map during runtime, but the
2324 // overhead of that may be large if there's many reassociable chains.
2325 BuildPairMap(RPOT);
2326
2327 MadeChange = false;
2328
2329 // Traverse the same blocks that were analysed by BuildRankMap.
2330 for (BasicBlock *BI : RPOT) {
2331 assert(RankMap.count(&*BI) && "BB should be ranked.")((RankMap.count(&*BI) && "BB should be ranked.") ?
static_cast<void> (0) : __assert_fail ("RankMap.count(&*BI) && \"BB should be ranked.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 2331, __PRETTY_FUNCTION__))
;
2332 // Optimize every instruction in the basic block.
2333 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2334 if (isInstructionTriviallyDead(&*II)) {
2335 EraseInst(&*II++);
2336 } else {
2337 OptimizeInst(&*II);
2338 assert(II->getParent() == &*BI && "Moved to a different block!")((II->getParent() == &*BI && "Moved to a different block!"
) ? static_cast<void> (0) : __assert_fail ("II->getParent() == &*BI && \"Moved to a different block!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/Reassociate.cpp"
, 2338, __PRETTY_FUNCTION__))
;
2339 ++II;
2340 }
2341
2342 // Make a copy of all the instructions to be redone so we can remove dead
2343 // instructions.
2344 OrderedSet ToRedo(RedoInsts);
2345 // Iterate over all instructions to be reevaluated and remove trivially dead
2346 // instructions. If any operand of the trivially dead instruction becomes
2347 // dead mark it for deletion as well. Continue this process until all
2348 // trivially dead instructions have been removed.
2349 while (!ToRedo.empty()) {
2350 Instruction *I = ToRedo.pop_back_val();
2351 if (isInstructionTriviallyDead(I)) {
2352 RecursivelyEraseDeadInsts(I, ToRedo);
2353 MadeChange = true;
2354 }
2355 }
2356
2357 // Now that we have removed dead instructions, we can reoptimize the
2358 // remaining instructions.
2359 while (!RedoInsts.empty()) {
2360 Instruction *I = RedoInsts.front();
2361 RedoInsts.erase(RedoInsts.begin());
2362 if (isInstructionTriviallyDead(I))
2363 EraseInst(I);
2364 else
2365 OptimizeInst(I);
2366 }
2367 }
2368
2369 // We are done with the rank map and pair map.
2370 RankMap.clear();
2371 ValueRankMap.clear();
2372 for (auto &Entry : PairMap)
2373 Entry.clear();
2374
2375 if (MadeChange) {
2376 PreservedAnalyses PA;
2377 PA.preserveSet<CFGAnalyses>();
2378 PA.preserve<GlobalsAA>();
2379 return PA;
2380 }
2381
2382 return PreservedAnalyses::all();
2383}
2384
2385namespace {
2386
2387 class ReassociateLegacyPass : public FunctionPass {
2388 ReassociatePass Impl;
2389
2390 public:
2391 static char ID; // Pass identification, replacement for typeid
2392
2393 ReassociateLegacyPass() : FunctionPass(ID) {
2394 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2395 }
2396
2397 bool runOnFunction(Function &F) override {
2398 if (skipFunction(F))
1
Assuming the condition is false
2
Taking false branch
2399 return false;
2400
2401 FunctionAnalysisManager DummyFAM;
2402 auto PA = Impl.run(F, DummyFAM);
3
Calling 'ReassociatePass::run'
2403 return !PA.areAllPreserved();
2404 }
2405
2406 void getAnalysisUsage(AnalysisUsage &AU) const override {
2407 AU.setPreservesCFG();
2408 AU.addPreserved<GlobalsAAWrapperPass>();
2409 }
2410 };
2411
2412} // end anonymous namespace
2413
2414char ReassociateLegacyPass::ID = 0;
2415
2416INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",static void *initializeReassociateLegacyPassPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "Reassociate expressions"
, "reassociate", &ReassociateLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<ReassociateLegacyPass>), false, false)
; Registry.registerPass(*PI, true); return PI; } static llvm::
once_flag InitializeReassociateLegacyPassPassFlag; void llvm::
initializeReassociateLegacyPassPass(PassRegistry &Registry
) { llvm::call_once(InitializeReassociateLegacyPassPassFlag, initializeReassociateLegacyPassPassOnce
, std::ref(Registry)); }
2417 "Reassociate expressions", false, false)static void *initializeReassociateLegacyPassPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "Reassociate expressions"
, "reassociate", &ReassociateLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<ReassociateLegacyPass>), false, false)
; Registry.registerPass(*PI, true); return PI; } static llvm::
once_flag InitializeReassociateLegacyPassPassFlag; void llvm::
initializeReassociateLegacyPassPass(PassRegistry &Registry
) { llvm::call_once(InitializeReassociateLegacyPassPassFlag, initializeReassociateLegacyPassPassOnce
, std::ref(Registry)); }
2418
2419// Public interface to the Reassociate pass
2420FunctionPass *llvm::createReassociatePass() {
2421 return new ReassociateLegacyPass();
2422}