Bug Summary

File:lib/Transforms/Scalar/Reassociate.cpp
Warning:line 1480, column 37
Called C++ object pointer is null

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-eagerly-assume -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-7/lib/clang/7.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-7~svn338205/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-7~svn338205/build-llvm/include -I /build/llvm-toolchain-snapshot-7~svn338205/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/8/../../../../include/c++/8 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/8/../../../../include/x86_64-linux-gnu/c++/8 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/8/../../../../include/x86_64-linux-gnu/c++/8 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/8/../../../../include/c++/8/backward -internal-isystem /usr/include/clang/7.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-7/lib/clang/7.0.0/include -internal-externc-isystem /usr/lib/gcc/x86_64-linux-gnu/8/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-class-memaccess -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-7~svn338205/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-2018-07-29-043837-17923-1 -x c++ /build/llvm-toolchain-snapshot-7~svn338205/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;
66
67#define DEBUG_TYPE"reassociate" "reassociate"
68
69STATISTIC(NumChanged, "Number of insts reassociated")static llvm::Statistic NumChanged = {"reassociate", "NumChanged"
, "Number of insts reassociated", {0}, {false}}
;
70STATISTIC(NumAnnihil, "Number of expr tree annihilated")static llvm::Statistic NumAnnihil = {"reassociate", "NumAnnihil"
, "Number of expr tree annihilated", {0}, {false}}
;
71STATISTIC(NumFactor , "Number of multiplies factored")static llvm::Statistic NumFactor = {"reassociate", "NumFactor"
, "Number of multiplies factored", {0}, {false}}
;
72
73#ifndef NDEBUG
74/// Print out the expression identified in the Ops list.
75static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
76 Module *M = I->getModule();
77 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
78 << *Ops[0].Op->getType() << '\t';
79 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
80 dbgs() << "[ ";
81 Ops[i].Op->printAsOperand(dbgs(), false, M);
82 dbgs() << ", #" << Ops[i].Rank << "] ";
83 }
84}
85#endif
86
87/// Utility class representing a non-constant Xor-operand. We classify
88/// non-constant Xor-Operands into two categories:
89/// C1) The operand is in the form "X & C", where C is a constant and C != ~0
90/// C2)
91/// C2.1) The operand is in the form of "X | C", where C is a non-zero
92/// constant.
93/// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
94/// operand as "E | 0"
95class llvm::reassociate::XorOpnd {
96public:
97 XorOpnd(Value *V);
98
99 bool isInvalid() const { return SymbolicPart == nullptr; }
100 bool isOrExpr() const { return isOr; }
101 Value *getValue() const { return OrigVal; }
102 Value *getSymbolicPart() const { return SymbolicPart; }
103 unsigned getSymbolicRank() const { return SymbolicRank; }
104 const APInt &getConstPart() const { return ConstPart; }
105
106 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
107 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
108
109private:
110 Value *OrigVal;
111 Value *SymbolicPart;
112 APInt ConstPart;
113 unsigned SymbolicRank;
114 bool isOr;
115};
116
117XorOpnd::XorOpnd(Value *V) {
118 assert(!isa<ConstantInt>(V) && "No ConstantInt")(static_cast <bool> (!isa<ConstantInt>(V) &&
"No ConstantInt") ? void (0) : __assert_fail ("!isa<ConstantInt>(V) && \"No ConstantInt\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 118, __extension__ __PRETTY_FUNCTION__))
;
119 OrigVal = V;
120 Instruction *I = dyn_cast<Instruction>(V);
121 SymbolicRank = 0;
122
123 if (I && (I->getOpcode() == Instruction::Or ||
124 I->getOpcode() == Instruction::And)) {
125 Value *V0 = I->getOperand(0);
126 Value *V1 = I->getOperand(1);
127 const APInt *C;
128 if (match(V0, PatternMatch::m_APInt(C)))
129 std::swap(V0, V1);
130
131 if (match(V1, PatternMatch::m_APInt(C))) {
132 ConstPart = *C;
133 SymbolicPart = V0;
134 isOr = (I->getOpcode() == Instruction::Or);
135 return;
136 }
137 }
138
139 // view the operand as "V | 0"
140 SymbolicPart = V;
141 ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits());
142 isOr = true;
143}
144
145/// Return true if V is an instruction of the specified opcode and if it
146/// only has one use.
147static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
148 auto *I = dyn_cast<Instruction>(V);
149 if (I && I->hasOneUse() && I->getOpcode() == Opcode)
150 if (!isa<FPMathOperator>(I) || I->isFast())
151 return cast<BinaryOperator>(I);
152 return nullptr;
153}
154
155static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
156 unsigned Opcode2) {
157 auto *I = dyn_cast<Instruction>(V);
158 if (I && I->hasOneUse() &&
159 (I->getOpcode() == Opcode1 || I->getOpcode() == Opcode2))
160 if (!isa<FPMathOperator>(I) || I->isFast())
161 return cast<BinaryOperator>(I);
162 return nullptr;
163}
164
165void ReassociatePass::BuildRankMap(Function &F,
166 ReversePostOrderTraversal<Function*> &RPOT) {
167 unsigned Rank = 2;
168
169 // Assign distinct ranks to function arguments.
170 for (auto &Arg : F.args()) {
171 ValueRankMap[&Arg] = ++Rank;
172 LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rankdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
Arg.getName() << "] = " << Rank << "\n"; }
} while (false)
173 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
Arg.getName() << "] = " << Rank << "\n"; }
} while (false)
;
174 }
175
176 // Traverse basic blocks in ReversePostOrder
177 for (BasicBlock *BB : RPOT) {
178 unsigned BBRank = RankMap[BB] = ++Rank << 16;
179
180 // Walk the basic block, adding precomputed ranks for any instructions that
181 // we cannot move. This ensures that the ranks for these instructions are
182 // all different in the block.
183 for (Instruction &I : *BB)
184 if (mayBeMemoryDependent(I))
185 ValueRankMap[&I] = ++BBRank;
186 }
187}
188
189unsigned ReassociatePass::getRank(Value *V) {
190 Instruction *I = dyn_cast<Instruction>(V);
191 if (!I) {
192 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
193 return 0; // Otherwise it's a global or constant, rank 0.
194 }
195
196 if (unsigned Rank = ValueRankMap[I])
197 return Rank; // Rank already known?
198
199 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
200 // we can reassociate expressions for code motion! Since we do not recurse
201 // for PHI nodes, we cannot have infinite recursion here, because there
202 // cannot be loops in the value graph that do not go through PHI nodes.
203 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
204 for (unsigned i = 0, e = I->getNumOperands(); i != e && Rank != MaxRank; ++i)
205 Rank = std::max(Rank, getRank(I->getOperand(i)));
206
207 // If this is a not or neg instruction, do not count it for rank. This
208 // assures us that X and ~X will have the same rank.
209 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
210 !BinaryOperator::isFNeg(I))
211 ++Rank;
212
213 LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rankdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
V->getName() << "] = " << Rank << "\n";
} } while (false)
214 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Calculated Rank[" <<
V->getName() << "] = " << Rank << "\n";
} } while (false)
;
215
216 return ValueRankMap[I] = Rank;
217}
218
219// Canonicalize constants to RHS. Otherwise, sort the operands by rank.
220void ReassociatePass::canonicalizeOperands(Instruction *I) {
221 assert(isa<BinaryOperator>(I) && "Expected binary operator.")(static_cast <bool> (isa<BinaryOperator>(I) &&
"Expected binary operator.") ? void (0) : __assert_fail ("isa<BinaryOperator>(I) && \"Expected binary operator.\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 221, __extension__ __PRETTY_FUNCTION__))
;
222 assert(I->isCommutative() && "Expected commutative operator.")(static_cast <bool> (I->isCommutative() && "Expected commutative operator."
) ? void (0) : __assert_fail ("I->isCommutative() && \"Expected commutative operator.\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 222, __extension__ __PRETTY_FUNCTION__))
;
223
224 Value *LHS = I->getOperand(0);
225 Value *RHS = I->getOperand(1);
226 if (LHS == RHS || isa<Constant>(RHS))
227 return;
228 if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
229 cast<BinaryOperator>(I)->swapOperands();
230}
231
232static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
233 Instruction *InsertBefore, Value *FlagsOp) {
234 if (S1->getType()->isIntOrIntVectorTy())
235 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
236 else {
237 BinaryOperator *Res =
238 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
239 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
240 return Res;
241 }
242}
243
244static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
245 Instruction *InsertBefore, Value *FlagsOp) {
246 if (S1->getType()->isIntOrIntVectorTy())
247 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
248 else {
249 BinaryOperator *Res =
250 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
251 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
252 return Res;
253 }
254}
255
256static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
257 Instruction *InsertBefore, Value *FlagsOp) {
258 if (S1->getType()->isIntOrIntVectorTy())
259 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
260 else {
261 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
262 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
263 return Res;
264 }
265}
266
267/// Replace 0-X with X*-1.
268static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
269 Type *Ty = Neg->getType();
270 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
271 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
272
273 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
274 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
275 Res->takeName(Neg);
276 Neg->replaceAllUsesWith(Res);
277 Res->setDebugLoc(Neg->getDebugLoc());
278 return Res;
279}
280
281/// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
282/// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
283/// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
284/// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
285/// even x in Bitwidth-bit arithmetic.
286static unsigned CarmichaelShift(unsigned Bitwidth) {
287 if (Bitwidth < 3)
288 return Bitwidth - 1;
289 return Bitwidth - 2;
290}
291
292/// Add the extra weight 'RHS' to the existing weight 'LHS',
293/// reducing the combined weight using any special properties of the operation.
294/// The existing weight LHS represents the computation X op X op ... op X where
295/// X occurs LHS times. The combined weight represents X op X op ... op X with
296/// X occurring LHS + RHS times. If op is "Xor" for example then the combined
297/// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
298/// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
299static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
300 // If we were working with infinite precision arithmetic then the combined
301 // weight would be LHS + RHS. But we are using finite precision arithmetic,
302 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
303 // for nilpotent operations and addition, but not for idempotent operations
304 // and multiplication), so it is important to correctly reduce the combined
305 // weight back into range if wrapping would be wrong.
306
307 // If RHS is zero then the weight didn't change.
308 if (RHS.isMinValue())
309 return;
310 // If LHS is zero then the combined weight is RHS.
311 if (LHS.isMinValue()) {
312 LHS = RHS;
313 return;
314 }
315 // From this point on we know that neither LHS nor RHS is zero.
316
317 if (Instruction::isIdempotent(Opcode)) {
318 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
319 // weight of 1. Keeping weights at zero or one also means that wrapping is
320 // not a problem.
321 assert(LHS == 1 && RHS == 1 && "Weights not reduced!")(static_cast <bool> (LHS == 1 && RHS == 1 &&
"Weights not reduced!") ? void (0) : __assert_fail ("LHS == 1 && RHS == 1 && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 321, __extension__ __PRETTY_FUNCTION__))
;
322 return; // Return a weight of 1.
323 }
324 if (Instruction::isNilpotent(Opcode)) {
325 // Nilpotent means X op X === 0, so reduce weights modulo 2.
326 assert(LHS == 1 && RHS == 1 && "Weights not reduced!")(static_cast <bool> (LHS == 1 && RHS == 1 &&
"Weights not reduced!") ? void (0) : __assert_fail ("LHS == 1 && RHS == 1 && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 326, __extension__ __PRETTY_FUNCTION__))
;
327 LHS = 0; // 1 + 1 === 0 modulo 2.
328 return;
329 }
330 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
331 // TODO: Reduce the weight by exploiting nsw/nuw?
332 LHS += RHS;
333 return;
334 }
335
336 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&(static_cast <bool> ((Opcode == Instruction::Mul || Opcode
== Instruction::FMul) && "Unknown associative operation!"
) ? void (0) : __assert_fail ("(Opcode == Instruction::Mul || Opcode == Instruction::FMul) && \"Unknown associative operation!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 337, __extension__ __PRETTY_FUNCTION__))
337 "Unknown associative operation!")(static_cast <bool> ((Opcode == Instruction::Mul || Opcode
== Instruction::FMul) && "Unknown associative operation!"
) ? void (0) : __assert_fail ("(Opcode == Instruction::Mul || Opcode == Instruction::FMul) && \"Unknown associative operation!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 337, __extension__ __PRETTY_FUNCTION__))
;
338 unsigned Bitwidth = LHS.getBitWidth();
339 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
340 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
341 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
342 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
343 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
344 // which by a happy accident means that they can always be represented using
345 // Bitwidth bits.
346 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
347 // the Carmichael number).
348 if (Bitwidth > 3) {
349 /// CM - The value of Carmichael's lambda function.
350 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
351 // Any weight W >= Threshold can be replaced with W - CM.
352 APInt Threshold = CM + Bitwidth;
353 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!")(static_cast <bool> (LHS.ult(Threshold) && RHS.
ult(Threshold) && "Weights not reduced!") ? void (0) :
__assert_fail ("LHS.ult(Threshold) && RHS.ult(Threshold) && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 353, __extension__ __PRETTY_FUNCTION__))
;
354 // For Bitwidth 4 or more the following sum does not overflow.
355 LHS += RHS;
356 while (LHS.uge(Threshold))
357 LHS -= CM;
358 } else {
359 // To avoid problems with overflow do everything the same as above but using
360 // a larger type.
361 unsigned CM = 1U << CarmichaelShift(Bitwidth);
362 unsigned Threshold = CM + Bitwidth;
363 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&(static_cast <bool> (LHS.getZExtValue() < Threshold &&
RHS.getZExtValue() < Threshold && "Weights not reduced!"
) ? void (0) : __assert_fail ("LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 364, __extension__ __PRETTY_FUNCTION__))
364 "Weights not reduced!")(static_cast <bool> (LHS.getZExtValue() < Threshold &&
RHS.getZExtValue() < Threshold && "Weights not reduced!"
) ? void (0) : __assert_fail ("LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && \"Weights not reduced!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 364, __extension__ __PRETTY_FUNCTION__))
;
365 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
366 while (Total >= Threshold)
367 Total -= CM;
368 LHS = Total;
369 }
370}
371
372using RepeatedValue = std::pair<Value*, APInt>;
373
374/// Given an associative binary expression, return the leaf
375/// nodes in Ops along with their weights (how many times the leaf occurs). The
376/// original expression is the same as
377/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
378/// op
379/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
380/// op
381/// ...
382/// op
383/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
384///
385/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
386///
387/// This routine may modify the function, in which case it returns 'true'. The
388/// changes it makes may well be destructive, changing the value computed by 'I'
389/// to something completely different. Thus if the routine returns 'true' then
390/// you MUST either replace I with a new expression computed from the Ops array,
391/// or use RewriteExprTree to put the values back in.
392///
393/// A leaf node is either not a binary operation of the same kind as the root
394/// node 'I' (i.e. is not a binary operator at all, or is, but with a different
395/// opcode), or is the same kind of binary operator but has a use which either
396/// does not belong to the expression, or does belong to the expression but is
397/// a leaf node. Every leaf node has at least one use that is a non-leaf node
398/// of the expression, while for non-leaf nodes (except for the root 'I') every
399/// use is a non-leaf node of the expression.
400///
401/// For example:
402/// expression graph node names
403///
404/// + | I
405/// / \ |
406/// + + | A, B
407/// / \ / \ |
408/// * + * | C, D, E
409/// / \ / \ / \ |
410/// + * | F, G
411///
412/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
413/// that order) (C, 1), (E, 1), (F, 2), (G, 2).
414///
415/// The expression is maximal: if some instruction is a binary operator of the
416/// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
417/// then the instruction also belongs to the expression, is not a leaf node of
418/// it, and its operands also belong to the expression (but may be leaf nodes).
419///
420/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
421/// order to ensure that every non-root node in the expression has *exactly one*
422/// use by a non-leaf node of the expression. This destruction means that the
423/// caller MUST either replace 'I' with a new expression or use something like
424/// RewriteExprTree to put the values back in if the routine indicates that it
425/// made a change by returning 'true'.
426///
427/// In the above example either the right operand of A or the left operand of B
428/// will be replaced by undef. If it is B's operand then this gives:
429///
430/// + | I
431/// / \ |
432/// + + | A, B - operand of B replaced with undef
433/// / \ \ |
434/// * + * | C, D, E
435/// / \ / \ / \ |
436/// + * | F, G
437///
438/// Note that such undef operands can only be reached by passing through 'I'.
439/// For example, if you visit operands recursively starting from a leaf node
440/// then you will never see such an undef operand unless you get back to 'I',
441/// which requires passing through a phi node.
442///
443/// Note that this routine may also mutate binary operators of the wrong type
444/// that have all uses inside the expression (i.e. only used by non-leaf nodes
445/// of the expression) if it can turn them into binary operators of the right
446/// type and thus make the expression bigger.
447static bool LinearizeExprTree(BinaryOperator *I,
448 SmallVectorImpl<RepeatedValue> &Ops) {
449 LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "LINEARIZE: " << *I <<
'\n'; } } while (false)
;
450 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
451 unsigned Opcode = I->getOpcode();
452 assert(I->isAssociative() && I->isCommutative() &&(static_cast <bool> (I->isAssociative() && I
->isCommutative() && "Expected an associative and commutative operation!"
) ? void (0) : __assert_fail ("I->isAssociative() && I->isCommutative() && \"Expected an associative and commutative operation!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 453, __extension__ __PRETTY_FUNCTION__))
453 "Expected an associative and commutative operation!")(static_cast <bool> (I->isAssociative() && I
->isCommutative() && "Expected an associative and commutative operation!"
) ? void (0) : __assert_fail ("I->isAssociative() && I->isCommutative() && \"Expected an associative and commutative operation!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 453, __extension__ __PRETTY_FUNCTION__))
;
454
455 // Visit all operands of the expression, keeping track of their weight (the
456 // number of paths from the expression root to the operand, or if you like
457 // the number of times that operand occurs in the linearized expression).
458 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
459 // while A has weight two.
460
461 // Worklist of non-leaf nodes (their operands are in the expression too) along
462 // with their weights, representing a certain number of paths to the operator.
463 // If an operator occurs in the worklist multiple times then we found multiple
464 // ways to get to it.
465 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
466 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
467 bool Changed = false;
468
469 // Leaves of the expression are values that either aren't the right kind of
470 // operation (eg: a constant, or a multiply in an add tree), or are, but have
471 // some uses that are not inside the expression. For example, in I = X + X,
472 // X = A + B, the value X has two uses (by I) that are in the expression. If
473 // X has any other uses, for example in a return instruction, then we consider
474 // X to be a leaf, and won't analyze it further. When we first visit a value,
475 // if it has more than one use then at first we conservatively consider it to
476 // be a leaf. Later, as the expression is explored, we may discover some more
477 // uses of the value from inside the expression. If all uses turn out to be
478 // from within the expression (and the value is a binary operator of the right
479 // kind) then the value is no longer considered to be a leaf, and its operands
480 // are explored.
481
482 // Leaves - Keeps track of the set of putative leaves as well as the number of
483 // paths to each leaf seen so far.
484 using LeafMap = DenseMap<Value *, APInt>;
485 LeafMap Leaves; // Leaf -> Total weight so far.
486 SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
487
488#ifndef NDEBUG
489 SmallPtrSet<Value *, 8> Visited; // For sanity checking the iteration scheme.
490#endif
491 while (!Worklist.empty()) {
492 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
493 I = P.first; // We examine the operands of this binary operator.
494
495 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
496 Value *Op = I->getOperand(OpIdx);
497 APInt Weight = P.second; // Number of paths to this operand.
498 LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "OPERAND: " << *Op <<
" (" << Weight << ")\n"; } } while (false)
;
499 assert(!Op->use_empty() && "No uses, so how did we get to it?!")(static_cast <bool> (!Op->use_empty() && "No uses, so how did we get to it?!"
) ? void (0) : __assert_fail ("!Op->use_empty() && \"No uses, so how did we get to it?!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 499, __extension__ __PRETTY_FUNCTION__))
;
500
501 // If this is a binary operation of the right kind with only one use then
502 // add its operands to the expression.
503 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
504 assert(Visited.insert(Op).second && "Not first visit!")(static_cast <bool> (Visited.insert(Op).second &&
"Not first visit!") ? void (0) : __assert_fail ("Visited.insert(Op).second && \"Not first visit!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 504, __extension__ __PRETTY_FUNCTION__))
;
505 LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "DIRECT ADD: " << *Op
<< " (" << Weight << ")\n"; } } while (false
)
;
506 Worklist.push_back(std::make_pair(BO, Weight));
507 continue;
508 }
509
510 // Appears to be a leaf. Is the operand already in the set of leaves?
511 LeafMap::iterator It = Leaves.find(Op);
512 if (It == Leaves.end()) {
513 // Not in the leaf map. Must be the first time we saw this operand.
514 assert(Visited.insert(Op).second && "Not first visit!")(static_cast <bool> (Visited.insert(Op).second &&
"Not first visit!") ? void (0) : __assert_fail ("Visited.insert(Op).second && \"Not first visit!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 514, __extension__ __PRETTY_FUNCTION__))
;
515 if (!Op->hasOneUse()) {
516 // This value has uses not accounted for by the expression, so it is
517 // not safe to modify. Mark it as being a leaf.
518 LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "ADD USES LEAF: " <<
*Op << " (" << Weight << ")\n"; } } while (
false)
519 << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "ADD USES LEAF: " <<
*Op << " (" << Weight << ")\n"; } } while (
false)
;
520 LeafOrder.push_back(Op);
521 Leaves[Op] = Weight;
522 continue;
523 }
524 // No uses outside the expression, try morphing it.
525 } else {
526 // Already in the leaf map.
527 assert(It != Leaves.end() && Visited.count(Op) &&(static_cast <bool> (It != Leaves.end() && Visited
.count(Op) && "In leaf map but not visited!") ? void (
0) : __assert_fail ("It != Leaves.end() && Visited.count(Op) && \"In leaf map but not visited!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 528, __extension__ __PRETTY_FUNCTION__))
528 "In leaf map but not visited!")(static_cast <bool> (It != Leaves.end() && Visited
.count(Op) && "In leaf map but not visited!") ? void (
0) : __assert_fail ("It != Leaves.end() && Visited.count(Op) && \"In leaf map but not visited!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 528, __extension__ __PRETTY_FUNCTION__))
;
529
530 // Update the number of paths to the leaf.
531 IncorporateWeight(It->second, Weight, Opcode);
532
533#if 0 // TODO: Re-enable once PR13021 is fixed.
534 // The leaf already has one use from inside the expression. As we want
535 // exactly one such use, drop this new use of the leaf.
536 assert(!Op->hasOneUse() && "Only one use, but we got here twice!")(static_cast <bool> (!Op->hasOneUse() && "Only one use, but we got here twice!"
) ? void (0) : __assert_fail ("!Op->hasOneUse() && \"Only one use, but we got here twice!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 536, __extension__ __PRETTY_FUNCTION__))
;
537 I->setOperand(OpIdx, UndefValue::get(I->getType()));
538 Changed = true;
539
540 // If the leaf is a binary operation of the right kind and we now see
541 // that its multiple original uses were in fact all by nodes belonging
542 // to the expression, then no longer consider it to be a leaf and add
543 // its operands to the expression.
544 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
545 LLVM_DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "UNLEAF: " << *Op <<
" (" << It->second << ")\n"; } } while (false
)
;
546 Worklist.push_back(std::make_pair(BO, It->second));
547 Leaves.erase(It);
548 continue;
549 }
550#endif
551
552 // If we still have uses that are not accounted for by the expression
553 // then it is not safe to modify the value.
554 if (!Op->hasOneUse())
555 continue;
556
557 // No uses outside the expression, try morphing it.
558 Weight = It->second;
559 Leaves.erase(It); // Since the value may be morphed below.
560 }
561
562 // At this point we have a value which, first of all, is not a binary
563 // expression of the right kind, and secondly, is only used inside the
564 // expression. This means that it can safely be modified. See if we
565 // can usefully morph it into an expression of the right kind.
566 assert((!isa<Instruction>(Op) ||(static_cast <bool> ((!isa<Instruction>(Op) || cast
<Instruction>(Op)->getOpcode() != Opcode || (isa<
FPMathOperator>(Op) && !cast<Instruction>(Op
)->isFast())) && "Should have been handled above!"
) ? 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-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 570, __extension__ __PRETTY_FUNCTION__))
567 cast<Instruction>(Op)->getOpcode() != Opcode(static_cast <bool> ((!isa<Instruction>(Op) || cast
<Instruction>(Op)->getOpcode() != Opcode || (isa<
FPMathOperator>(Op) && !cast<Instruction>(Op
)->isFast())) && "Should have been handled above!"
) ? 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-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 570, __extension__ __PRETTY_FUNCTION__))
568 || (isa<FPMathOperator>(Op) &&(static_cast <bool> ((!isa<Instruction>(Op) || cast
<Instruction>(Op)->getOpcode() != Opcode || (isa<
FPMathOperator>(Op) && !cast<Instruction>(Op
)->isFast())) && "Should have been handled above!"
) ? 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-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 570, __extension__ __PRETTY_FUNCTION__))
569 !cast<Instruction>(Op)->isFast())) &&(static_cast <bool> ((!isa<Instruction>(Op) || cast
<Instruction>(Op)->getOpcode() != Opcode || (isa<
FPMathOperator>(Op) && !cast<Instruction>(Op
)->isFast())) && "Should have been handled above!"
) ? 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-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 570, __extension__ __PRETTY_FUNCTION__))
570 "Should have been handled above!")(static_cast <bool> ((!isa<Instruction>(Op) || cast
<Instruction>(Op)->getOpcode() != Opcode || (isa<
FPMathOperator>(Op) && !cast<Instruction>(Op
)->isFast())) && "Should have been handled above!"
) ? 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-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 570, __extension__ __PRETTY_FUNCTION__))
;
571 assert(Op->hasOneUse() && "Has uses outside the expression tree!")(static_cast <bool> (Op->hasOneUse() && "Has uses outside the expression tree!"
) ? void (0) : __assert_fail ("Op->hasOneUse() && \"Has uses outside the expression tree!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 571, __extension__ __PRETTY_FUNCTION__))
;
572
573 // If this is a multiply expression, turn any internal negations into
574 // multiplies by -1 so they can be reassociated.
575 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
576 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
577 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
578 LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "MORPH LEAF: " << *Op
<< " (" << Weight << ") TO "; } } while (false
)
579 << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "MORPH LEAF: " << *Op
<< " (" << Weight << ") TO "; } } while (false
)
;
580 BO = LowerNegateToMultiply(BO);
581 LLVM_DEBUG(dbgs() << *BO << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << *BO << '\n'; } } while
(false)
;
582 Worklist.push_back(std::make_pair(BO, Weight));
583 Changed = true;
584 continue;
585 }
586
587 // Failed to morph into an expression of the right type. This really is
588 // a leaf.
589 LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "ADD LEAF: " << *Op <<
" (" << Weight << ")\n"; } } while (false)
;
590 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?")(static_cast <bool> (!isReassociableOp(Op, Opcode) &&
"Value was morphed?") ? void (0) : __assert_fail ("!isReassociableOp(Op, Opcode) && \"Value was morphed?\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 590, __extension__ __PRETTY_FUNCTION__))
;
591 LeafOrder.push_back(Op);
592 Leaves[Op] = Weight;
593 }
594 }
595
596 // The leaves, repeated according to their weights, represent the linearized
597 // form of the expression.
598 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
599 Value *V = LeafOrder[i];
600 LeafMap::iterator It = Leaves.find(V);
601 if (It == Leaves.end())
602 // Node initially thought to be a leaf wasn't.
603 continue;
604 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!")(static_cast <bool> (!isReassociableOp(V, Opcode) &&
"Shouldn't be a leaf!") ? void (0) : __assert_fail ("!isReassociableOp(V, Opcode) && \"Shouldn't be a leaf!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 604, __extension__ __PRETTY_FUNCTION__))
;
605 APInt Weight = It->second;
606 if (Weight.isMinValue())
607 // Leaf already output or weight reduction eliminated it.
608 continue;
609 // Ensure the leaf is only output once.
610 It->second = 0;
611 Ops.push_back(std::make_pair(V, Weight));
612 }
613
614 // For nilpotent operations or addition there may be no operands, for example
615 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
616 // in both cases the weight reduces to 0 causing the value to be skipped.
617 if (Ops.empty()) {
618 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
619 assert(Identity && "Associative operation without identity!")(static_cast <bool> (Identity && "Associative operation without identity!"
) ? void (0) : __assert_fail ("Identity && \"Associative operation without identity!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 619, __extension__ __PRETTY_FUNCTION__))
;
620 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
621 }
622
623 return Changed;
624}
625
626/// Now that the operands for this expression tree are
627/// linearized and optimized, emit them in-order.
628void ReassociatePass::RewriteExprTree(BinaryOperator *I,
629 SmallVectorImpl<ValueEntry> &Ops) {
630 assert(Ops.size() > 1 && "Single values should be used directly!")(static_cast <bool> (Ops.size() > 1 && "Single values should be used directly!"
) ? void (0) : __assert_fail ("Ops.size() > 1 && \"Single values should be used directly!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 630, __extension__ __PRETTY_FUNCTION__))
;
631
632 // Since our optimizations should never increase the number of operations, the
633 // new expression can usually be written reusing the existing binary operators
634 // from the original expression tree, without creating any new instructions,
635 // though the rewritten expression may have a completely different topology.
636 // We take care to not change anything if the new expression will be the same
637 // as the original. If more than trivial changes (like commuting operands)
638 // were made then we are obliged to clear out any optional subclass data like
639 // nsw flags.
640
641 /// NodesToRewrite - Nodes from the original expression available for writing
642 /// the new expression into.
643 SmallVector<BinaryOperator*, 8> NodesToRewrite;
644 unsigned Opcode = I->getOpcode();
645 BinaryOperator *Op = I;
646
647 /// NotRewritable - The operands being written will be the leaves of the new
648 /// expression and must not be used as inner nodes (via NodesToRewrite) by
649 /// mistake. Inner nodes are always reassociable, and usually leaves are not
650 /// (if they were they would have been incorporated into the expression and so
651 /// would not be leaves), so most of the time there is no danger of this. But
652 /// in rare cases a leaf may become reassociable if an optimization kills uses
653 /// of it, or it may momentarily become reassociable during rewriting (below)
654 /// due it being removed as an operand of one of its uses. Ensure that misuse
655 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
656 /// leaves and refusing to reuse any of them as inner nodes.
657 SmallPtrSet<Value*, 8> NotRewritable;
658 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
659 NotRewritable.insert(Ops[i].Op);
660
661 // ExpressionChanged - Non-null if the rewritten expression differs from the
662 // original in some non-trivial way, requiring the clearing of optional flags.
663 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
664 BinaryOperator *ExpressionChanged = nullptr;
665 for (unsigned i = 0; ; ++i) {
666 // The last operation (which comes earliest in the IR) is special as both
667 // operands will come from Ops, rather than just one with the other being
668 // a subexpression.
669 if (i+2 == Ops.size()) {
670 Value *NewLHS = Ops[i].Op;
671 Value *NewRHS = Ops[i+1].Op;
672 Value *OldLHS = Op->getOperand(0);
673 Value *OldRHS = Op->getOperand(1);
674
675 if (NewLHS == OldLHS && NewRHS == OldRHS)
676 // Nothing changed, leave it alone.
677 break;
678
679 if (NewLHS == OldRHS && NewRHS == OldLHS) {
680 // The order of the operands was reversed. Swap them.
681 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
682 Op->swapOperands();
683 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
684 MadeChange = true;
685 ++NumChanged;
686 break;
687 }
688
689 // The new operation differs non-trivially from the original. Overwrite
690 // the old operands with the new ones.
691 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
692 if (NewLHS != OldLHS) {
693 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
694 if (BO && !NotRewritable.count(BO))
695 NodesToRewrite.push_back(BO);
696 Op->setOperand(0, NewLHS);
697 }
698 if (NewRHS != OldRHS) {
699 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
700 if (BO && !NotRewritable.count(BO))
701 NodesToRewrite.push_back(BO);
702 Op->setOperand(1, NewRHS);
703 }
704 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
705
706 ExpressionChanged = Op;
707 MadeChange = true;
708 ++NumChanged;
709
710 break;
711 }
712
713 // Not the last operation. The left-hand side will be a sub-expression
714 // while the right-hand side will be the current element of Ops.
715 Value *NewRHS = Ops[i].Op;
716 if (NewRHS != Op->getOperand(1)) {
717 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
718 if (NewRHS == Op->getOperand(0)) {
719 // The new right-hand side was already present as the left operand. If
720 // we are lucky then swapping the operands will sort out both of them.
721 Op->swapOperands();
722 } else {
723 // Overwrite with the new right-hand side.
724 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
725 if (BO && !NotRewritable.count(BO))
726 NodesToRewrite.push_back(BO);
727 Op->setOperand(1, NewRHS);
728 ExpressionChanged = Op;
729 }
730 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
731 MadeChange = true;
732 ++NumChanged;
733 }
734
735 // Now deal with the left-hand side. If this is already an operation node
736 // from the original expression then just rewrite the rest of the expression
737 // into it.
738 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
739 if (BO && !NotRewritable.count(BO)) {
740 Op = BO;
741 continue;
742 }
743
744 // Otherwise, grab a spare node from the original expression and use that as
745 // the left-hand side. If there are no nodes left then the optimizers made
746 // an expression with more nodes than the original! This usually means that
747 // they did something stupid but it might mean that the problem was just too
748 // hard (finding the mimimal number of multiplications needed to realize a
749 // multiplication expression is NP-complete). Whatever the reason, smart or
750 // stupid, create a new node if there are none left.
751 BinaryOperator *NewOp;
752 if (NodesToRewrite.empty()) {
753 Constant *Undef = UndefValue::get(I->getType());
754 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
755 Undef, Undef, "", I);
756 if (NewOp->getType()->isFPOrFPVectorTy())
757 NewOp->setFastMathFlags(I->getFastMathFlags());
758 } else {
759 NewOp = NodesToRewrite.pop_back_val();
760 }
761
762 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "RA: " << *Op <<
'\n'; } } while (false)
;
763 Op->setOperand(0, NewOp);
764 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "TO: " << *Op <<
'\n'; } } while (false)
;
765 ExpressionChanged = Op;
766 MadeChange = true;
767 ++NumChanged;
768 Op = NewOp;
769 }
770
771 // If the expression changed non-trivially then clear out all subclass data
772 // starting from the operator specified in ExpressionChanged, and compactify
773 // the operators to just before the expression root to guarantee that the
774 // expression tree is dominated by all of Ops.
775 if (ExpressionChanged)
776 do {
777 // Preserve FastMathFlags.
778 if (isa<FPMathOperator>(I)) {
779 FastMathFlags Flags = I->getFastMathFlags();
780 ExpressionChanged->clearSubclassOptionalData();
781 ExpressionChanged->setFastMathFlags(Flags);
782 } else
783 ExpressionChanged->clearSubclassOptionalData();
784
785 if (ExpressionChanged == I)
786 break;
787
788 // Discard any debug info related to the expressions that has changed (we
789 // can leave debug infor related to the root, since the result of the
790 // expression tree should be the same even after reassociation).
791 SmallVector<DbgInfoIntrinsic *, 1> DbgUsers;
792 findDbgUsers(DbgUsers, ExpressionChanged);
793 for (auto *DII : DbgUsers) {
794 Value *Undef = UndefValue::get(ExpressionChanged->getType());
795 DII->setOperand(0, MetadataAsValue::get(DII->getContext(),
796 ValueAsMetadata::get(Undef)));
797 }
798
799 ExpressionChanged->moveBefore(I);
800 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
801 } while (true);
802
803 // Throw away any left over nodes from the original expression.
804 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
805 RedoInsts.insert(NodesToRewrite[i]);
806}
807
808/// Insert instructions before the instruction pointed to by BI,
809/// that computes the negative version of the value specified. The negative
810/// version of the value is returned, and BI is left pointing at the instruction
811/// that should be processed next by the reassociation pass.
812/// Also add intermediate instructions to the redo list that are modified while
813/// pushing the negates through adds. These will be revisited to see if
814/// additional opportunities have been exposed.
815static Value *NegateValue(Value *V, Instruction *BI,
816 ReassociatePass::OrderedSet &ToRedo) {
817 if (auto *C = dyn_cast<Constant>(V))
818 return C->getType()->isFPOrFPVectorTy() ? ConstantExpr::getFNeg(C) :
819 ConstantExpr::getNeg(C);
820
821 // We are trying to expose opportunity for reassociation. One of the things
822 // that we want to do to achieve this is to push a negation as deep into an
823 // expression chain as possible, to expose the add instructions. In practice,
824 // this means that we turn this:
825 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
826 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
827 // the constants. We assume that instcombine will clean up the mess later if
828 // we introduce tons of unnecessary negation instructions.
829 //
830 if (BinaryOperator *I =
831 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
832 // Push the negates through the add.
833 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
834 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
835 if (I->getOpcode() == Instruction::Add) {
836 I->setHasNoUnsignedWrap(false);
837 I->setHasNoSignedWrap(false);
838 }
839
840 // We must move the add instruction here, because the neg instructions do
841 // not dominate the old add instruction in general. By moving it, we are
842 // assured that the neg instructions we just inserted dominate the
843 // instruction we are about to insert after them.
844 //
845 I->moveBefore(BI);
846 I->setName(I->getName()+".neg");
847
848 // Add the intermediate negates to the redo list as processing them later
849 // could expose more reassociating opportunities.
850 ToRedo.insert(I);
851 return I;
852 }
853
854 // Okay, we need to materialize a negated version of V with an instruction.
855 // Scan the use lists of V to see if we have one already.
856 for (User *U : V->users()) {
857 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
858 continue;
859
860 // We found one! Now we have to make sure that the definition dominates
861 // this use. We do this by moving it to the entry block (if it is a
862 // non-instruction value) or right after the definition. These negates will
863 // be zapped by reassociate later, so we don't need much finesse here.
864 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
865
866 // Verify that the negate is in this function, V might be a constant expr.
867 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
868 continue;
869
870 BasicBlock::iterator InsertPt;
871 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
872 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
873 InsertPt = II->getNormalDest()->begin();
874 } else {
875 InsertPt = ++InstInput->getIterator();
876 }
877 while (isa<PHINode>(InsertPt)) ++InsertPt;
878 } else {
879 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
880 }
881 TheNeg->moveBefore(&*InsertPt);
882 if (TheNeg->getOpcode() == Instruction::Sub) {
883 TheNeg->setHasNoUnsignedWrap(false);
884 TheNeg->setHasNoSignedWrap(false);
885 } else {
886 TheNeg->andIRFlags(BI);
887 }
888 ToRedo.insert(TheNeg);
889 return TheNeg;
890 }
891
892 // Insert a 'neg' instruction that subtracts the value from zero to get the
893 // negation.
894 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
895 ToRedo.insert(NewNeg);
896 return NewNeg;
897}
898
899/// Return true if we should break up this subtract of X-Y into (X + -Y).
900static bool ShouldBreakUpSubtract(Instruction *Sub) {
901 // If this is a negation, we can't split it up!
902 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
903 return false;
904
905 // Don't breakup X - undef.
906 if (isa<UndefValue>(Sub->getOperand(1)))
907 return false;
908
909 // Don't bother to break this up unless either the LHS is an associable add or
910 // subtract or if this is only used by one.
911 Value *V0 = Sub->getOperand(0);
912 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
913 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
914 return true;
915 Value *V1 = Sub->getOperand(1);
916 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
917 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
918 return true;
919 Value *VB = Sub->user_back();
920 if (Sub->hasOneUse() &&
921 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
922 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
923 return true;
924
925 return false;
926}
927
928/// If we have (X-Y), and if either X is an add, or if this is only used by an
929/// add, transform this into (X+(0-Y)) to promote better reassociation.
930static BinaryOperator *BreakUpSubtract(Instruction *Sub,
931 ReassociatePass::OrderedSet &ToRedo) {
932 // Convert a subtract into an add and a neg instruction. This allows sub
933 // instructions to be commuted with other add instructions.
934 //
935 // Calculate the negative value of Operand 1 of the sub instruction,
936 // and set it as the RHS of the add instruction we just made.
937 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
938 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
939 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
940 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
941 New->takeName(Sub);
942
943 // Everyone now refers to the add instruction.
944 Sub->replaceAllUsesWith(New);
945 New->setDebugLoc(Sub->getDebugLoc());
946
947 LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Negated: " << *New <<
'\n'; } } while (false)
;
948 return New;
949}
950
951/// If this is a shift of a reassociable multiply or is used by one, change
952/// this into a multiply by a constant to assist with further reassociation.
953static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
954 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
955 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
956
957 BinaryOperator *Mul =
958 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
959 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
960 Mul->takeName(Shl);
961
962 // Everyone now refers to the mul instruction.
963 Shl->replaceAllUsesWith(Mul);
964 Mul->setDebugLoc(Shl->getDebugLoc());
965
966 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
967 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
968 // handling.
969 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
970 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
971 if (NSW && NUW)
972 Mul->setHasNoSignedWrap(true);
973 Mul->setHasNoUnsignedWrap(NUW);
974 return Mul;
975}
976
977/// Scan backwards and forwards among values with the same rank as element i
978/// to see if X exists. If X does not exist, return i. This is useful when
979/// scanning for 'x' when we see '-x' because they both get the same rank.
980static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
981 unsigned i, Value *X) {
982 unsigned XRank = Ops[i].Rank;
983 unsigned e = Ops.size();
984 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
985 if (Ops[j].Op == X)
986 return j;
987 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
988 if (Instruction *I2 = dyn_cast<Instruction>(X))
989 if (I1->isIdenticalTo(I2))
990 return j;
991 }
992 // Scan backwards.
993 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
994 if (Ops[j].Op == X)
995 return j;
996 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
997 if (Instruction *I2 = dyn_cast<Instruction>(X))
998 if (I1->isIdenticalTo(I2))
999 return j;
1000 }
1001 return i;
1002}
1003
1004/// Emit a tree of add instructions, summing Ops together
1005/// and returning the result. Insert the tree before I.
1006static Value *EmitAddTreeOfValues(Instruction *I,
1007 SmallVectorImpl<WeakTrackingVH> &Ops) {
1008 if (Ops.size() == 1) return Ops.back();
1009
1010 Value *V1 = Ops.back();
1011 Ops.pop_back();
1012 Value *V2 = EmitAddTreeOfValues(I, Ops);
1013 return CreateAdd(V2, V1, "reass.add", I, I);
1014}
1015
1016/// If V is an expression tree that is a multiplication sequence,
1017/// and if this sequence contains a multiply by Factor,
1018/// remove Factor from the tree and return the new tree.
1019Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1020 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1021 if (!BO)
1022 return nullptr;
1023
1024 SmallVector<RepeatedValue, 8> Tree;
1025 MadeChange |= LinearizeExprTree(BO, Tree);
1026 SmallVector<ValueEntry, 8> Factors;
1027 Factors.reserve(Tree.size());
1028 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1029 RepeatedValue E = Tree[i];
1030 Factors.append(E.second.getZExtValue(),
1031 ValueEntry(getRank(E.first), E.first));
1032 }
1033
1034 bool FoundFactor = false;
1035 bool NeedsNegate = false;
1036 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1037 if (Factors[i].Op == Factor) {
1038 FoundFactor = true;
1039 Factors.erase(Factors.begin()+i);
1040 break;
1041 }
1042
1043 // If this is a negative version of this factor, remove it.
1044 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1045 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1046 if (FC1->getValue() == -FC2->getValue()) {
1047 FoundFactor = NeedsNegate = true;
1048 Factors.erase(Factors.begin()+i);
1049 break;
1050 }
1051 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1052 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1053 const APFloat &F1 = FC1->getValueAPF();
1054 APFloat F2(FC2->getValueAPF());
1055 F2.changeSign();
1056 if (F1.compare(F2) == APFloat::cmpEqual) {
1057 FoundFactor = NeedsNegate = true;
1058 Factors.erase(Factors.begin() + i);
1059 break;
1060 }
1061 }
1062 }
1063 }
1064
1065 if (!FoundFactor) {
1066 // Make sure to restore the operands to the expression tree.
1067 RewriteExprTree(BO, Factors);
1068 return nullptr;
1069 }
1070
1071 BasicBlock::iterator InsertPt = ++BO->getIterator();
1072
1073 // If this was just a single multiply, remove the multiply and return the only
1074 // remaining operand.
1075 if (Factors.size() == 1) {
1076 RedoInsts.insert(BO);
1077 V = Factors[0].Op;
1078 } else {
1079 RewriteExprTree(BO, Factors);
1080 V = BO;
1081 }
1082
1083 if (NeedsNegate)
1084 V = CreateNeg(V, "neg", &*InsertPt, BO);
1085
1086 return V;
1087}
1088
1089/// If V is a single-use multiply, recursively add its operands as factors,
1090/// otherwise add V to the list of factors.
1091///
1092/// Ops is the top-level list of add operands we're trying to factor.
1093static void FindSingleUseMultiplyFactors(Value *V,
1094 SmallVectorImpl<Value*> &Factors) {
1095 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1096 if (!BO) {
1097 Factors.push_back(V);
1098 return;
1099 }
1100
1101 // Otherwise, add the LHS and RHS to the list of factors.
1102 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1103 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1104}
1105
1106/// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1107/// This optimizes based on identities. If it can be reduced to a single Value,
1108/// it is returned, otherwise the Ops list is mutated as necessary.
1109static Value *OptimizeAndOrXor(unsigned Opcode,
1110 SmallVectorImpl<ValueEntry> &Ops) {
1111 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1112 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1113 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1114 // First, check for X and ~X in the operand list.
1115 assert(i < Ops.size())(static_cast <bool> (i < Ops.size()) ? void (0) : __assert_fail
("i < Ops.size()", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1115, __extension__ __PRETTY_FUNCTION__))
;
1116 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1117 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1118 unsigned FoundX = FindInOperandList(Ops, i, X);
1119 if (FoundX != i) {
1120 if (Opcode == Instruction::And) // ...&X&~X = 0
1121 return Constant::getNullValue(X->getType());
1122
1123 if (Opcode == Instruction::Or) // ...|X|~X = -1
1124 return Constant::getAllOnesValue(X->getType());
1125 }
1126 }
1127
1128 // Next, check for duplicate pairs of values, which we assume are next to
1129 // each other, due to our sorting criteria.
1130 assert(i < Ops.size())(static_cast <bool> (i < Ops.size()) ? void (0) : __assert_fail
("i < Ops.size()", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1130, __extension__ __PRETTY_FUNCTION__))
;
1131 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1132 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1133 // Drop duplicate values for And and Or.
1134 Ops.erase(Ops.begin()+i);
1135 --i; --e;
1136 ++NumAnnihil;
1137 continue;
1138 }
1139
1140 // Drop pairs of values for Xor.
1141 assert(Opcode == Instruction::Xor)(static_cast <bool> (Opcode == Instruction::Xor) ? void
(0) : __assert_fail ("Opcode == Instruction::Xor", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1141, __extension__ __PRETTY_FUNCTION__))
;
1142 if (e == 2)
1143 return Constant::getNullValue(Ops[0].Op->getType());
1144
1145 // Y ^ X^X -> Y
1146 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1147 i -= 1; e -= 2;
1148 ++NumAnnihil;
1149 }
1150 }
1151 return nullptr;
1152}
1153
1154/// Helper function of CombineXorOpnd(). It creates a bitwise-and
1155/// instruction with the given two operands, and return the resulting
1156/// instruction. There are two special cases: 1) if the constant operand is 0,
1157/// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1158/// be returned.
1159static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1160 const APInt &ConstOpnd) {
1161 if (ConstOpnd.isNullValue())
1162 return nullptr;
1163
1164 if (ConstOpnd.isAllOnesValue())
1165 return Opnd;
1166
1167 Instruction *I = BinaryOperator::CreateAnd(
1168 Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1169 InsertBefore);
1170 I->setDebugLoc(InsertBefore->getDebugLoc());
1171 return I;
1172}
1173
1174// Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1175// into "R ^ C", where C would be 0, and R is a symbolic value.
1176//
1177// If it was successful, true is returned, and the "R" and "C" is returned
1178// via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1179// and both "Res" and "ConstOpnd" remain unchanged.
1180bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1181 APInt &ConstOpnd, Value *&Res) {
1182 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1183 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1184 // = (x & ~c1) ^ (c1 ^ c2)
1185 // It is useful only when c1 == c2.
1186 if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue())
1187 return false;
1188
1189 if (!Opnd1->getValue()->hasOneUse())
1190 return false;
1191
1192 const APInt &C1 = Opnd1->getConstPart();
1193 if (C1 != ConstOpnd)
1194 return false;
1195
1196 Value *X = Opnd1->getSymbolicPart();
1197 Res = createAndInstr(I, X, ~C1);
1198 // ConstOpnd was C2, now C1 ^ C2.
1199 ConstOpnd ^= C1;
1200
1201 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1202 RedoInsts.insert(T);
1203 return true;
1204}
1205
1206// Helper function of OptimizeXor(). It tries to simplify
1207// "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1208// symbolic value.
1209//
1210// If it was successful, true is returned, and the "R" and "C" is returned
1211// via "Res" and "ConstOpnd", respectively (If the entire expression is
1212// evaluated to a constant, the Res is set to NULL); otherwise, false is
1213// returned, and both "Res" and "ConstOpnd" remain unchanged.
1214bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1215 XorOpnd *Opnd2, APInt &ConstOpnd,
1216 Value *&Res) {
1217 Value *X = Opnd1->getSymbolicPart();
1218 if (X != Opnd2->getSymbolicPart())
1219 return false;
1220
1221 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1222 int DeadInstNum = 1;
1223 if (Opnd1->getValue()->hasOneUse())
1224 DeadInstNum++;
1225 if (Opnd2->getValue()->hasOneUse())
1226 DeadInstNum++;
1227
1228 // Xor-Rule 2:
1229 // (x | c1) ^ (x & c2)
1230 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1231 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1232 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1233 //
1234 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1235 if (Opnd2->isOrExpr())
1236 std::swap(Opnd1, Opnd2);
1237
1238 const APInt &C1 = Opnd1->getConstPart();
1239 const APInt &C2 = Opnd2->getConstPart();
1240 APInt C3((~C1) ^ C2);
1241
1242 // Do not increase code size!
1243 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1244 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1245 if (NewInstNum > DeadInstNum)
1246 return false;
1247 }
1248
1249 Res = createAndInstr(I, X, C3);
1250 ConstOpnd ^= C1;
1251 } else if (Opnd1->isOrExpr()) {
1252 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1253 //
1254 const APInt &C1 = Opnd1->getConstPart();
1255 const APInt &C2 = Opnd2->getConstPart();
1256 APInt C3 = C1 ^ C2;
1257
1258 // Do not increase code size
1259 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1260 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1261 if (NewInstNum > DeadInstNum)
1262 return false;
1263 }
1264
1265 Res = createAndInstr(I, X, C3);
1266 ConstOpnd ^= C3;
1267 } else {
1268 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1269 //
1270 const APInt &C1 = Opnd1->getConstPart();
1271 const APInt &C2 = Opnd2->getConstPart();
1272 APInt C3 = C1 ^ C2;
1273 Res = createAndInstr(I, X, C3);
1274 }
1275
1276 // Put the original operands in the Redo list; hope they will be deleted
1277 // as dead code.
1278 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1279 RedoInsts.insert(T);
1280 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1281 RedoInsts.insert(T);
1282
1283 return true;
1284}
1285
1286/// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1287/// to a single Value, it is returned, otherwise the Ops list is mutated as
1288/// necessary.
1289Value *ReassociatePass::OptimizeXor(Instruction *I,
1290 SmallVectorImpl<ValueEntry> &Ops) {
1291 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1292 return V;
1293
1294 if (Ops.size() == 1)
1295 return nullptr;
1296
1297 SmallVector<XorOpnd, 8> Opnds;
1298 SmallVector<XorOpnd*, 8> OpndPtrs;
1299 Type *Ty = Ops[0].Op->getType();
1300 APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1301
1302 // Step 1: Convert ValueEntry to XorOpnd
1303 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1304 Value *V = Ops[i].Op;
1305 const APInt *C;
1306 // TODO: Support non-splat vectors.
1307 if (match(V, PatternMatch::m_APInt(C))) {
1308 ConstOpnd ^= *C;
1309 } else {
1310 XorOpnd O(V);
1311 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1312 Opnds.push_back(O);
1313 }
1314 }
1315
1316 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1317 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1318 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1319 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1320 // when new elements are added to the vector.
1321 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1322 OpndPtrs.push_back(&Opnds[i]);
1323
1324 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1325 // the same symbolic value cluster together. For instance, the input operand
1326 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1327 // ("x | 123", "x & 789", "y & 456").
1328 //
1329 // The purpose is twofold:
1330 // 1) Cluster together the operands sharing the same symbolic-value.
1331 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1332 // could potentially shorten crital path, and expose more loop-invariants.
1333 // Note that values' rank are basically defined in RPO order (FIXME).
1334 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1335 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1336 // "z" in the order of X-Y-Z is better than any other orders.
1337 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1338 [](XorOpnd *LHS, XorOpnd *RHS) {
1339 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1340 });
1341
1342 // Step 3: Combine adjacent operands
1343 XorOpnd *PrevOpnd = nullptr;
1344 bool Changed = false;
1345 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1346 XorOpnd *CurrOpnd = OpndPtrs[i];
1347 // The combined value
1348 Value *CV;
1349
1350 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1351 if (!ConstOpnd.isNullValue() &&
1352 CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1353 Changed = true;
1354 if (CV)
1355 *CurrOpnd = XorOpnd(CV);
1356 else {
1357 CurrOpnd->Invalidate();
1358 continue;
1359 }
1360 }
1361
1362 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1363 PrevOpnd = CurrOpnd;
1364 continue;
1365 }
1366
1367 // step 3.2: When previous and current operands share the same symbolic
1368 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1369 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1370 // Remove previous operand
1371 PrevOpnd->Invalidate();
1372 if (CV) {
1373 *CurrOpnd = XorOpnd(CV);
1374 PrevOpnd = CurrOpnd;
1375 } else {
1376 CurrOpnd->Invalidate();
1377 PrevOpnd = nullptr;
1378 }
1379 Changed = true;
1380 }
1381 }
1382
1383 // Step 4: Reassemble the Ops
1384 if (Changed) {
1385 Ops.clear();
1386 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1387 XorOpnd &O = Opnds[i];
1388 if (O.isInvalid())
1389 continue;
1390 ValueEntry VE(getRank(O.getValue()), O.getValue());
1391 Ops.push_back(VE);
1392 }
1393 if (!ConstOpnd.isNullValue()) {
1394 Value *C = ConstantInt::get(Ty, ConstOpnd);
1395 ValueEntry VE(getRank(C), C);
1396 Ops.push_back(VE);
1397 }
1398 unsigned Sz = Ops.size();
1399 if (Sz == 1)
1400 return Ops.back().Op;
1401 if (Sz == 0) {
1402 assert(ConstOpnd.isNullValue())(static_cast <bool> (ConstOpnd.isNullValue()) ? void (0
) : __assert_fail ("ConstOpnd.isNullValue()", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1402, __extension__ __PRETTY_FUNCTION__))
;
1403 return ConstantInt::get(Ty, ConstOpnd);
1404 }
1405 }
1406
1407 return nullptr;
1408}
1409
1410/// Optimize a series of operands to an 'add' instruction. This
1411/// optimizes based on identities. If it can be reduced to a single Value, it
1412/// is returned, otherwise the Ops list is mutated as necessary.
1413Value *ReassociatePass::OptimizeAdd(Instruction *I,
1414 SmallVectorImpl<ValueEntry> &Ops) {
1415 // Scan the operand lists looking for X and -X pairs. If we find any, we
1416 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1417 // scan for any
1418 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1419
1420 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1
Assuming 'i' is not equal to 'e'
2
Loop condition is true. Entering loop body
1421 Value *TheOp = Ops[i].Op;
1422 // Check to see if we've seen this operand before. If so, we factor all
1423 // instances of the operand together. Due to our sorting criteria, we know
1424 // that these need to be next to each other in the vector.
1425 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
3
Assuming the condition is true
4
Assuming the condition is false
5
Taking false branch
1426 // Rescan the list, remove all instances of this operand from the expr.
1427 unsigned NumFound = 0;
1428 do {
1429 Ops.erase(Ops.begin()+i);
1430 ++NumFound;
1431 } while (i != Ops.size() && Ops[i].Op == TheOp);
1432
1433 LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOpdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << NumFound
<< "]: " << *TheOp << '\n'; } } while (false
)
1434 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << NumFound
<< "]: " << *TheOp << '\n'; } } while (false
)
;
1435 ++NumFactor;
1436
1437 // Insert a new multiply.
1438 Type *Ty = TheOp->getType();
1439 Constant *C = Ty->isIntOrIntVectorTy() ?
1440 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1441 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1442
1443 // Now that we have inserted a multiply, optimize it. This allows us to
1444 // handle cases that require multiple factoring steps, such as this:
1445 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1446 RedoInsts.insert(Mul);
1447
1448 // If every add operand was a duplicate, return the multiply.
1449 if (Ops.empty())
1450 return Mul;
1451
1452 // Otherwise, we had some input that didn't have the dupe, such as
1453 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1454 // things being added by this operation.
1455 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1456
1457 --i;
1458 e = Ops.size();
1459 continue;
1460 }
1461
1462 // Check for X and -X or X and ~X in the operand list.
1463 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
6
Assuming the condition is false
1464 !BinaryOperator::isNot(TheOp))
1465 continue;
1466
1467 Value *X = nullptr;
7
'X' initialized to a null pointer value
1468 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
8
Assuming the condition is false
9
Assuming the condition is false
10
Taking false branch
1469 X = BinaryOperator::getNegArgument(TheOp);
1470 else if (BinaryOperator::isNot(TheOp))
11
Assuming the condition is false
12
Taking false branch
1471 X = BinaryOperator::getNotArgument(TheOp);
1472
1473 unsigned FoundX = FindInOperandList(Ops, i, X);
1474 if (FoundX == i)
13
Assuming 'FoundX' is not equal to 'i'
14
Taking false branch
1475 continue;
1476
1477 // Remove X and -X from the operand list.
1478 if (Ops.size() == 2 &&
15
Assuming the condition is true
1479 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
16
Assuming the condition is true
1480 return Constant::getNullValue(X->getType());
17
Called C++ object pointer is null
1481
1482 // Remove X and ~X from the operand list.
1483 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1484 return Constant::getAllOnesValue(X->getType());
1485
1486 Ops.erase(Ops.begin()+i);
1487 if (i < FoundX)
1488 --FoundX;
1489 else
1490 --i; // Need to back up an extra one.
1491 Ops.erase(Ops.begin()+FoundX);
1492 ++NumAnnihil;
1493 --i; // Revisit element.
1494 e -= 2; // Removed two elements.
1495
1496 // if X and ~X we append -1 to the operand list.
1497 if (BinaryOperator::isNot(TheOp)) {
1498 Value *V = Constant::getAllOnesValue(X->getType());
1499 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1500 e += 1;
1501 }
1502 }
1503
1504 // Scan the operand list, checking to see if there are any common factors
1505 // between operands. Consider something like A*A+A*B*C+D. We would like to
1506 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1507 // To efficiently find this, we count the number of times a factor occurs
1508 // for any ADD operands that are MULs.
1509 DenseMap<Value*, unsigned> FactorOccurrences;
1510
1511 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1512 // where they are actually the same multiply.
1513 unsigned MaxOcc = 0;
1514 Value *MaxOccVal = nullptr;
1515 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1516 BinaryOperator *BOp =
1517 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1518 if (!BOp)
1519 continue;
1520
1521 // Compute all of the factors of this added value.
1522 SmallVector<Value*, 8> Factors;
1523 FindSingleUseMultiplyFactors(BOp, Factors);
1524 assert(Factors.size() > 1 && "Bad linearize!")(static_cast <bool> (Factors.size() > 1 && "Bad linearize!"
) ? void (0) : __assert_fail ("Factors.size() > 1 && \"Bad linearize!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1524, __extension__ __PRETTY_FUNCTION__))
;
1525
1526 // Add one to FactorOccurrences for each unique factor in this op.
1527 SmallPtrSet<Value*, 8> Duplicates;
1528 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1529 Value *Factor = Factors[i];
1530 if (!Duplicates.insert(Factor).second)
1531 continue;
1532
1533 unsigned Occ = ++FactorOccurrences[Factor];
1534 if (Occ > MaxOcc) {
1535 MaxOcc = Occ;
1536 MaxOccVal = Factor;
1537 }
1538
1539 // If Factor is a negative constant, add the negated value as a factor
1540 // because we can percolate the negate out. Watch for minint, which
1541 // cannot be positivified.
1542 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1543 if (CI->isNegative() && !CI->isMinValue(true)) {
1544 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1545 if (!Duplicates.insert(Factor).second)
1546 continue;
1547 unsigned Occ = ++FactorOccurrences[Factor];
1548 if (Occ > MaxOcc) {
1549 MaxOcc = Occ;
1550 MaxOccVal = Factor;
1551 }
1552 }
1553 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1554 if (CF->isNegative()) {
1555 APFloat F(CF->getValueAPF());
1556 F.changeSign();
1557 Factor = ConstantFP::get(CF->getContext(), F);
1558 if (!Duplicates.insert(Factor).second)
1559 continue;
1560 unsigned Occ = ++FactorOccurrences[Factor];
1561 if (Occ > MaxOcc) {
1562 MaxOcc = Occ;
1563 MaxOccVal = Factor;
1564 }
1565 }
1566 }
1567 }
1568 }
1569
1570 // If any factor occurred more than one time, we can pull it out.
1571 if (MaxOcc > 1) {
1572 LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc
<< "]: " << *MaxOccVal << '\n'; } } while (
false)
1573 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc
<< "]: " << *MaxOccVal << '\n'; } } while (
false)
;
1574 ++NumFactor;
1575
1576 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1577 // this, we could otherwise run into situations where removing a factor
1578 // from an expression will drop a use of maxocc, and this can cause
1579 // RemoveFactorFromExpression on successive values to behave differently.
1580 Instruction *DummyInst =
1581 I->getType()->isIntOrIntVectorTy()
1582 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1583 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1584
1585 SmallVector<WeakTrackingVH, 4> NewMulOps;
1586 for (unsigned i = 0; i != Ops.size(); ++i) {
1587 // Only try to remove factors from expressions we're allowed to.
1588 BinaryOperator *BOp =
1589 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1590 if (!BOp)
1591 continue;
1592
1593 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1594 // The factorized operand may occur several times. Convert them all in
1595 // one fell swoop.
1596 for (unsigned j = Ops.size(); j != i;) {
1597 --j;
1598 if (Ops[j].Op == Ops[i].Op) {
1599 NewMulOps.push_back(V);
1600 Ops.erase(Ops.begin()+j);
1601 }
1602 }
1603 --i;
1604 }
1605 }
1606
1607 // No need for extra uses anymore.
1608 DummyInst->deleteValue();
1609
1610 unsigned NumAddedValues = NewMulOps.size();
1611 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1612
1613 // Now that we have inserted the add tree, optimize it. This allows us to
1614 // handle cases that require multiple factoring steps, such as this:
1615 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1616 assert(NumAddedValues > 1 && "Each occurrence should contribute a value")(static_cast <bool> (NumAddedValues > 1 && "Each occurrence should contribute a value"
) ? void (0) : __assert_fail ("NumAddedValues > 1 && \"Each occurrence should contribute a value\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1616, __extension__ __PRETTY_FUNCTION__))
;
1617 (void)NumAddedValues;
1618 if (Instruction *VI = dyn_cast<Instruction>(V))
1619 RedoInsts.insert(VI);
1620
1621 // Create the multiply.
1622 Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I);
1623
1624 // Rerun associate on the multiply in case the inner expression turned into
1625 // a multiply. We want to make sure that we keep things in canonical form.
1626 RedoInsts.insert(V2);
1627
1628 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1629 // entire result expression is just the multiply "A*(B+C)".
1630 if (Ops.empty())
1631 return V2;
1632
1633 // Otherwise, we had some input that didn't have the factor, such as
1634 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1635 // things being added by this operation.
1636 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1637 }
1638
1639 return nullptr;
1640}
1641
1642/// Build up a vector of value/power pairs factoring a product.
1643///
1644/// Given a series of multiplication operands, build a vector of factors and
1645/// the powers each is raised to when forming the final product. Sort them in
1646/// the order of descending power.
1647///
1648/// (x*x) -> [(x, 2)]
1649/// ((x*x)*x) -> [(x, 3)]
1650/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1651///
1652/// \returns Whether any factors have a power greater than one.
1653static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1654 SmallVectorImpl<Factor> &Factors) {
1655 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1656 // Compute the sum of powers of simplifiable factors.
1657 unsigned FactorPowerSum = 0;
1658 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1659 Value *Op = Ops[Idx-1].Op;
1660
1661 // Count the number of occurrences of this value.
1662 unsigned Count = 1;
1663 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1664 ++Count;
1665 // Track for simplification all factors which occur 2 or more times.
1666 if (Count > 1)
1667 FactorPowerSum += Count;
1668 }
1669
1670 // We can only simplify factors if the sum of the powers of our simplifiable
1671 // factors is 4 or higher. When that is the case, we will *always* have
1672 // a simplification. This is an important invariant to prevent cyclicly
1673 // trying to simplify already minimal formations.
1674 if (FactorPowerSum < 4)
1675 return false;
1676
1677 // Now gather the simplifiable factors, removing them from Ops.
1678 FactorPowerSum = 0;
1679 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1680 Value *Op = Ops[Idx-1].Op;
1681
1682 // Count the number of occurrences of this value.
1683 unsigned Count = 1;
1684 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1685 ++Count;
1686 if (Count == 1)
1687 continue;
1688 // Move an even number of occurrences to Factors.
1689 Count &= ~1U;
1690 Idx -= Count;
1691 FactorPowerSum += Count;
1692 Factors.push_back(Factor(Op, Count));
1693 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1694 }
1695
1696 // None of the adjustments above should have reduced the sum of factor powers
1697 // below our mininum of '4'.
1698 assert(FactorPowerSum >= 4)(static_cast <bool> (FactorPowerSum >= 4) ? void (0)
: __assert_fail ("FactorPowerSum >= 4", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1698, __extension__ __PRETTY_FUNCTION__))
;
1699
1700 std::stable_sort(Factors.begin(), Factors.end(),
1701 [](const Factor &LHS, const Factor &RHS) {
1702 return LHS.Power > RHS.Power;
1703 });
1704 return true;
1705}
1706
1707/// Build a tree of multiplies, computing the product of Ops.
1708static Value *buildMultiplyTree(IRBuilder<> &Builder,
1709 SmallVectorImpl<Value*> &Ops) {
1710 if (Ops.size() == 1)
1711 return Ops.back();
1712
1713 Value *LHS = Ops.pop_back_val();
1714 do {
1715 if (LHS->getType()->isIntOrIntVectorTy())
1716 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1717 else
1718 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1719 } while (!Ops.empty());
1720
1721 return LHS;
1722}
1723
1724/// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1725///
1726/// Given a vector of values raised to various powers, where no two values are
1727/// equal and the powers are sorted in decreasing order, compute the minimal
1728/// DAG of multiplies to compute the final product, and return that product
1729/// value.
1730Value *
1731ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1732 SmallVectorImpl<Factor> &Factors) {
1733 assert(Factors[0].Power)(static_cast <bool> (Factors[0].Power) ? void (0) : __assert_fail
("Factors[0].Power", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1733, __extension__ __PRETTY_FUNCTION__))
;
1734 SmallVector<Value *, 4> OuterProduct;
1735 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1736 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1737 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1738 LastIdx = Idx;
1739 continue;
1740 }
1741
1742 // We want to multiply across all the factors with the same power so that
1743 // we can raise them to that power as a single entity. Build a mini tree
1744 // for that.
1745 SmallVector<Value *, 4> InnerProduct;
1746 InnerProduct.push_back(Factors[LastIdx].Base);
1747 do {
1748 InnerProduct.push_back(Factors[Idx].Base);
1749 ++Idx;
1750 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1751
1752 // Reset the base value of the first factor to the new expression tree.
1753 // We'll remove all the factors with the same power in a second pass.
1754 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1755 if (Instruction *MI = dyn_cast<Instruction>(M))
1756 RedoInsts.insert(MI);
1757
1758 LastIdx = Idx;
1759 }
1760 // Unique factors with equal powers -- we've folded them into the first one's
1761 // base.
1762 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1763 [](const Factor &LHS, const Factor &RHS) {
1764 return LHS.Power == RHS.Power;
1765 }),
1766 Factors.end());
1767
1768 // Iteratively collect the base of each factor with an add power into the
1769 // outer product, and halve each power in preparation for squaring the
1770 // expression.
1771 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1772 if (Factors[Idx].Power & 1)
1773 OuterProduct.push_back(Factors[Idx].Base);
1774 Factors[Idx].Power >>= 1;
1775 }
1776 if (Factors[0].Power) {
1777 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1778 OuterProduct.push_back(SquareRoot);
1779 OuterProduct.push_back(SquareRoot);
1780 }
1781 if (OuterProduct.size() == 1)
1782 return OuterProduct.front();
1783
1784 Value *V = buildMultiplyTree(Builder, OuterProduct);
1785 return V;
1786}
1787
1788Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1789 SmallVectorImpl<ValueEntry> &Ops) {
1790 // We can only optimize the multiplies when there is a chain of more than
1791 // three, such that a balanced tree might require fewer total multiplies.
1792 if (Ops.size() < 4)
1793 return nullptr;
1794
1795 // Try to turn linear trees of multiplies without other uses of the
1796 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1797 // re-use.
1798 SmallVector<Factor, 4> Factors;
1799 if (!collectMultiplyFactors(Ops, Factors))
1800 return nullptr; // All distinct factors, so nothing left for us to do.
1801
1802 IRBuilder<> Builder(I);
1803 // The reassociate transformation for FP operations is performed only
1804 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1805 // to the newly generated operations.
1806 if (auto FPI = dyn_cast<FPMathOperator>(I))
1807 Builder.setFastMathFlags(FPI->getFastMathFlags());
1808
1809 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1810 if (Ops.empty())
1811 return V;
1812
1813 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1814 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1815 return nullptr;
1816}
1817
1818Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1819 SmallVectorImpl<ValueEntry> &Ops) {
1820 // Now that we have the linearized expression tree, try to optimize it.
1821 // Start by folding any constants that we found.
1822 Constant *Cst = nullptr;
1823 unsigned Opcode = I->getOpcode();
1824 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1825 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1826 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1827 }
1828 // If there was nothing but constants then we are done.
1829 if (Ops.empty())
1830 return Cst;
1831
1832 // Put the combined constant back at the end of the operand list, except if
1833 // there is no point. For example, an add of 0 gets dropped here, while a
1834 // multiplication by zero turns the whole expression into zero.
1835 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1836 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1837 return Cst;
1838 Ops.push_back(ValueEntry(0, Cst));
1839 }
1840
1841 if (Ops.size() == 1) return Ops[0].Op;
1842
1843 // Handle destructive annihilation due to identities between elements in the
1844 // argument list here.
1845 unsigned NumOps = Ops.size();
1846 switch (Opcode) {
1847 default: break;
1848 case Instruction::And:
1849 case Instruction::Or:
1850 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1851 return Result;
1852 break;
1853
1854 case Instruction::Xor:
1855 if (Value *Result = OptimizeXor(I, Ops))
1856 return Result;
1857 break;
1858
1859 case Instruction::Add:
1860 case Instruction::FAdd:
1861 if (Value *Result = OptimizeAdd(I, Ops))
1862 return Result;
1863 break;
1864
1865 case Instruction::Mul:
1866 case Instruction::FMul:
1867 if (Value *Result = OptimizeMul(I, Ops))
1868 return Result;
1869 break;
1870 }
1871
1872 if (Ops.size() != NumOps)
1873 return OptimizeExpression(I, Ops);
1874 return nullptr;
1875}
1876
1877// Remove dead instructions and if any operands are trivially dead add them to
1878// Insts so they will be removed as well.
1879void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I,
1880 OrderedSet &Insts) {
1881 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!")(static_cast <bool> (isInstructionTriviallyDead(I) &&
"Trivially dead instructions only!") ? void (0) : __assert_fail
("isInstructionTriviallyDead(I) && \"Trivially dead instructions only!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1881, __extension__ __PRETTY_FUNCTION__))
;
1882 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1883 ValueRankMap.erase(I);
1884 Insts.remove(I);
1885 RedoInsts.remove(I);
1886 I->eraseFromParent();
1887 for (auto Op : Ops)
1888 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1889 if (OpInst->use_empty())
1890 Insts.insert(OpInst);
1891}
1892
1893/// Zap the given instruction, adding interesting operands to the work list.
1894void ReassociatePass::EraseInst(Instruction *I) {
1895 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!")(static_cast <bool> (isInstructionTriviallyDead(I) &&
"Trivially dead instructions only!") ? void (0) : __assert_fail
("isInstructionTriviallyDead(I) && \"Trivially dead instructions only!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1895, __extension__ __PRETTY_FUNCTION__))
;
1896 LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Erasing dead inst: "; I->
dump(); } } while (false)
;
1897
1898 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1899 // Erase the dead instruction.
1900 ValueRankMap.erase(I);
1901 RedoInsts.remove(I);
1902 I->eraseFromParent();
1903 // Optimize its operands.
1904 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1905 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1906 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1907 // If this is a node in an expression tree, climb to the expression root
1908 // and add that since that's where optimization actually happens.
1909 unsigned Opcode = Op->getOpcode();
1910 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1911 Visited.insert(Op).second)
1912 Op = Op->user_back();
1913
1914 // The instruction we're going to push may be coming from a
1915 // dead block, and Reassociate skips the processing of unreachable
1916 // blocks because it's a waste of time and also because it can
1917 // lead to infinite loop due to LLVM's non-standard definition
1918 // of dominance.
1919 if (ValueRankMap.find(Op) != ValueRankMap.end())
1920 RedoInsts.insert(Op);
1921 }
1922
1923 MadeChange = true;
1924}
1925
1926// Canonicalize expressions of the following form:
1927// x + (-Constant * y) -> x - (Constant * y)
1928// x - (-Constant * y) -> x + (Constant * y)
1929Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1930 if (!I->hasOneUse() || I->getType()->isVectorTy())
1931 return nullptr;
1932
1933 // Must be a fmul or fdiv instruction.
1934 unsigned Opcode = I->getOpcode();
1935 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1936 return nullptr;
1937
1938 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1939 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1940
1941 // Both operands are constant, let it get constant folded away.
1942 if (C0 && C1)
1943 return nullptr;
1944
1945 ConstantFP *CF = C0 ? C0 : C1;
1946
1947 // Must have one constant operand.
1948 if (!CF)
1949 return nullptr;
1950
1951 // Must be a negative ConstantFP.
1952 if (!CF->isNegative())
1953 return nullptr;
1954
1955 // User must be a binary operator with one or more uses.
1956 Instruction *User = I->user_back();
1957 if (!isa<BinaryOperator>(User) || User->use_empty())
1958 return nullptr;
1959
1960 unsigned UserOpcode = User->getOpcode();
1961 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1962 return nullptr;
1963
1964 // Subtraction is not commutative. Explicitly, the following transform is
1965 // not valid: (-Constant * y) - x -> x + (Constant * y)
1966 if (!User->isCommutative() && User->getOperand(1) != I)
1967 return nullptr;
1968
1969 // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
1970 // resulting subtract will be broken up later. This can get us into an
1971 // infinite loop during reassociation.
1972 if (UserOpcode == Instruction::FAdd && ShouldBreakUpSubtract(User))
1973 return nullptr;
1974
1975 // Change the sign of the constant.
1976 APFloat Val = CF->getValueAPF();
1977 Val.changeSign();
1978 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1979
1980 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1981 // ((-Const*y) + x) -> (x + (-Const*y)).
1982 if (User->getOperand(0) == I && User->isCommutative())
1983 cast<BinaryOperator>(User)->swapOperands();
1984
1985 Value *Op0 = User->getOperand(0);
1986 Value *Op1 = User->getOperand(1);
1987 BinaryOperator *NI;
1988 switch (UserOpcode) {
1989 default:
1990 llvm_unreachable("Unexpected Opcode!")::llvm::llvm_unreachable_internal("Unexpected Opcode!", "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 1990)
;
1991 case Instruction::FAdd:
1992 NI = BinaryOperator::CreateFSub(Op0, Op1);
1993 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1994 break;
1995 case Instruction::FSub:
1996 NI = BinaryOperator::CreateFAdd(Op0, Op1);
1997 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1998 break;
1999 }
2000
2001 NI->insertBefore(User);
2002 NI->setName(User->getName());
2003 User->replaceAllUsesWith(NI);
2004 NI->setDebugLoc(I->getDebugLoc());
2005 RedoInsts.insert(I);
2006 MadeChange = true;
2007 return NI;
2008}
2009
2010/// Inspect and optimize the given instruction. Note that erasing
2011/// instructions is not allowed.
2012void ReassociatePass::OptimizeInst(Instruction *I) {
2013 // Only consider operations that we understand.
2014 if (!isa<BinaryOperator>(I))
2015 return;
2016
2017 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2018 // If an operand of this shift is a reassociable multiply, or if the shift
2019 // is used by a reassociable multiply or add, turn into a multiply.
2020 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2021 (I->hasOneUse() &&
2022 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2023 isReassociableOp(I->user_back(), Instruction::Add)))) {
2024 Instruction *NI = ConvertShiftToMul(I);
2025 RedoInsts.insert(I);
2026 MadeChange = true;
2027 I = NI;
2028 }
2029
2030 // Canonicalize negative constants out of expressions.
2031 if (Instruction *Res = canonicalizeNegConstExpr(I))
2032 I = Res;
2033
2034 // Commute binary operators, to canonicalize the order of their operands.
2035 // This can potentially expose more CSE opportunities, and makes writing other
2036 // transformations simpler.
2037 if (I->isCommutative())
2038 canonicalizeOperands(I);
2039
2040 // Don't optimize floating-point instructions unless they are 'fast'.
2041 if (I->getType()->isFPOrFPVectorTy() && !I->isFast())
2042 return;
2043
2044 // Do not reassociate boolean (i1) expressions. We want to preserve the
2045 // original order of evaluation for short-circuited comparisons that
2046 // SimplifyCFG has folded to AND/OR expressions. If the expression
2047 // is not further optimized, it is likely to be transformed back to a
2048 // short-circuited form for code gen, and the source order may have been
2049 // optimized for the most likely conditions.
2050 if (I->getType()->isIntegerTy(1))
2051 return;
2052
2053 // If this is a subtract instruction which is not already in negate form,
2054 // see if we can convert it to X+-Y.
2055 if (I->getOpcode() == Instruction::Sub) {
2056 if (ShouldBreakUpSubtract(I)) {
2057 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2058 RedoInsts.insert(I);
2059 MadeChange = true;
2060 I = NI;
2061 } else if (BinaryOperator::isNeg(I)) {
2062 // Otherwise, this is a negation. See if the operand is a multiply tree
2063 // and if this is not an inner node of a multiply tree.
2064 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2065 (!I->hasOneUse() ||
2066 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2067 Instruction *NI = LowerNegateToMultiply(I);
2068 // If the negate was simplified, revisit the users to see if we can
2069 // reassociate further.
2070 for (User *U : NI->users()) {
2071 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2072 RedoInsts.insert(Tmp);
2073 }
2074 RedoInsts.insert(I);
2075 MadeChange = true;
2076 I = NI;
2077 }
2078 }
2079 } else if (I->getOpcode() == Instruction::FSub) {
2080 if (ShouldBreakUpSubtract(I)) {
2081 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2082 RedoInsts.insert(I);
2083 MadeChange = true;
2084 I = NI;
2085 } else if (BinaryOperator::isFNeg(I)) {
2086 // Otherwise, this is a negation. See if the operand is a multiply tree
2087 // and if this is not an inner node of a multiply tree.
2088 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2089 (!I->hasOneUse() ||
2090 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2091 // If the negate was simplified, revisit the users to see if we can
2092 // reassociate further.
2093 Instruction *NI = LowerNegateToMultiply(I);
2094 for (User *U : NI->users()) {
2095 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2096 RedoInsts.insert(Tmp);
2097 }
2098 RedoInsts.insert(I);
2099 MadeChange = true;
2100 I = NI;
2101 }
2102 }
2103 }
2104
2105 // If this instruction is an associative binary operator, process it.
2106 if (!I->isAssociative()) return;
2107 BinaryOperator *BO = cast<BinaryOperator>(I);
2108
2109 // If this is an interior node of a reassociable tree, ignore it until we
2110 // get to the root of the tree, to avoid N^2 analysis.
2111 unsigned Opcode = BO->getOpcode();
2112 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2113 // During the initial run we will get to the root of the tree.
2114 // But if we get here while we are redoing instructions, there is no
2115 // guarantee that the root will be visited. So Redo later
2116 if (BO->user_back() != BO &&
2117 BO->getParent() == BO->user_back()->getParent())
2118 RedoInsts.insert(BO->user_back());
2119 return;
2120 }
2121
2122 // If this is an add tree that is used by a sub instruction, ignore it
2123 // until we process the subtract.
2124 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2125 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2126 return;
2127 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2128 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2129 return;
2130
2131 ReassociateExpression(BO);
2132}
2133
2134void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2135 // First, walk the expression tree, linearizing the tree, collecting the
2136 // operand information.
2137 SmallVector<RepeatedValue, 8> Tree;
2138 MadeChange |= LinearizeExprTree(I, Tree);
2139 SmallVector<ValueEntry, 8> Ops;
2140 Ops.reserve(Tree.size());
2141 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2142 RepeatedValue E = Tree[i];
2143 Ops.append(E.second.getZExtValue(),
2144 ValueEntry(getRank(E.first), E.first));
2145 }
2146
2147 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)
;
2148
2149 // Now that we have linearized the tree to a list and have gathered all of
2150 // the operands and their ranks, sort the operands by their rank. Use a
2151 // stable_sort so that values with equal ranks will have their relative
2152 // positions maintained (and so the compiler is deterministic). Note that
2153 // this sorts so that the highest ranking values end up at the beginning of
2154 // the vector.
2155 std::stable_sort(Ops.begin(), Ops.end());
2156
2157 // Now that we have the expression tree in a convenient
2158 // sorted form, optimize it globally if possible.
2159 if (Value *V = OptimizeExpression(I, Ops)) {
2160 if (V == I)
2161 // Self-referential expression in unreachable code.
2162 return;
2163 // This expression tree simplified to something that isn't a tree,
2164 // eliminate it.
2165 LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("reassociate")) { dbgs() << "Reassoc to scalar: " <<
*V << '\n'; } } while (false)
;
2166 I->replaceAllUsesWith(V);
2167 if (Instruction *VI = dyn_cast<Instruction>(V))
2168 if (I->getDebugLoc())
2169 VI->setDebugLoc(I->getDebugLoc());
2170 RedoInsts.insert(I);
2171 ++NumAnnihil;
2172 return;
2173 }
2174
2175 // We want to sink immediates as deeply as possible except in the case where
2176 // this is a multiply tree used only by an add, and the immediate is a -1.
2177 // In this case we reassociate to put the negation on the outside so that we
2178 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2179 if (I->hasOneUse()) {
2180 if (I->getOpcode() == Instruction::Mul &&
2181 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2182 isa<ConstantInt>(Ops.back().Op) &&
2183 cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2184 ValueEntry Tmp = Ops.pop_back_val();
2185 Ops.insert(Ops.begin(), Tmp);
2186 } else if (I->getOpcode() == Instruction::FMul &&
2187 cast<Instruction>(I->user_back())->getOpcode() ==
2188 Instruction::FAdd &&
2189 isa<ConstantFP>(Ops.back().Op) &&
2190 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2191 ValueEntry Tmp = Ops.pop_back_val();
2192 Ops.insert(Ops.begin(), Tmp);
2193 }
2194 }
2195
2196 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)
;
2197
2198 if (Ops.size() == 1) {
2199 if (Ops[0].Op == I)
2200 // Self-referential expression in unreachable code.
2201 return;
2202
2203 // This expression tree simplified to something that isn't a tree,
2204 // eliminate it.
2205 I->replaceAllUsesWith(Ops[0].Op);
2206 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2207 OI->setDebugLoc(I->getDebugLoc());
2208 RedoInsts.insert(I);
2209 return;
2210 }
2211
2212 if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) {
2213 // Find the pair with the highest count in the pairmap and move it to the
2214 // back of the list so that it can later be CSE'd.
2215 // example:
2216 // a*b*c*d*e
2217 // if c*e is the most "popular" pair, we can express this as
2218 // (((c*e)*d)*b)*a
2219 unsigned Max = 1;
2220 unsigned BestRank = 0;
2221 std::pair<unsigned, unsigned> BestPair;
2222 unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin;
2223 for (unsigned i = 0; i < Ops.size() - 1; ++i)
2224 for (unsigned j = i + 1; j < Ops.size(); ++j) {
2225 unsigned Score = 0;
2226 Value *Op0 = Ops[i].Op;
2227 Value *Op1 = Ops[j].Op;
2228 if (std::less<Value *>()(Op1, Op0))
2229 std::swap(Op0, Op1);
2230 auto it = PairMap[Idx].find({Op0, Op1});
2231 if (it != PairMap[Idx].end())
2232 Score += it->second;
2233
2234 unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank);
2235 if (Score > Max || (Score == Max && MaxRank < BestRank)) {
2236 BestPair = {i, j};
2237 Max = Score;
2238 BestRank = MaxRank;
2239 }
2240 }
2241 if (Max > 1) {
2242 auto Op0 = Ops[BestPair.first];
2243 auto Op1 = Ops[BestPair.second];
2244 Ops.erase(&Ops[BestPair.second]);
2245 Ops.erase(&Ops[BestPair.first]);
2246 Ops.push_back(Op0);
2247 Ops.push_back(Op1);
2248 }
2249 }
2250 // Now that we ordered and optimized the expressions, splat them back into
2251 // the expression tree, removing any unneeded nodes.
2252 RewriteExprTree(I, Ops);
2253}
2254
2255void
2256ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) {
2257 // Make a "pairmap" of how often each operand pair occurs.
2258 for (BasicBlock *BI : RPOT) {
2259 for (Instruction &I : *BI) {
2260 if (!I.isAssociative())
2261 continue;
2262
2263 // Ignore nodes that aren't at the root of trees.
2264 if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode())
2265 continue;
2266
2267 // Collect all operands in a single reassociable expression.
2268 // Since Reassociate has already been run once, we can assume things
2269 // are already canonical according to Reassociation's regime.
2270 SmallVector<Value *, 8> Worklist = { I.getOperand(0), I.getOperand(1) };
2271 SmallVector<Value *, 8> Ops;
2272 while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) {
2273 Value *Op = Worklist.pop_back_val();
2274 Instruction *OpI = dyn_cast<Instruction>(Op);
2275 if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) {
2276 Ops.push_back(Op);
2277 continue;
2278 }
2279 // Be paranoid about self-referencing expressions in unreachable code.
2280 if (OpI->getOperand(0) != OpI)
2281 Worklist.push_back(OpI->getOperand(0));
2282 if (OpI->getOperand(1) != OpI)
2283 Worklist.push_back(OpI->getOperand(1));
2284 }
2285 // Skip extremely long expressions.
2286 if (Ops.size() > GlobalReassociateLimit)
2287 continue;
2288
2289 // Add all pairwise combinations of operands to the pair map.
2290 unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin;
2291 SmallSet<std::pair<Value *, Value*>, 32> Visited;
2292 for (unsigned i = 0; i < Ops.size() - 1; ++i) {
2293 for (unsigned j = i + 1; j < Ops.size(); ++j) {
2294 // Canonicalize operand orderings.
2295 Value *Op0 = Ops[i];
2296 Value *Op1 = Ops[j];
2297 if (std::less<Value *>()(Op1, Op0))
2298 std::swap(Op0, Op1);
2299 if (!Visited.insert({Op0, Op1}).second)
2300 continue;
2301 auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, 1});
2302 if (!res.second)
2303 ++res.first->second;
2304 }
2305 }
2306 }
2307 }
2308}
2309
2310PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2311 // Get the functions basic blocks in Reverse Post Order. This order is used by
2312 // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2313 // blocks (it has been seen that the analysis in this pass could hang when
2314 // analysing dead basic blocks).
2315 ReversePostOrderTraversal<Function *> RPOT(&F);
2316
2317 // Calculate the rank map for F.
2318 BuildRankMap(F, RPOT);
2319
2320 // Build the pair map before running reassociate.
2321 // Technically this would be more accurate if we did it after one round
2322 // of reassociation, but in practice it doesn't seem to help much on
2323 // real-world code, so don't waste the compile time running reassociate
2324 // twice.
2325 // If a user wants, they could expicitly run reassociate twice in their
2326 // pass pipeline for further potential gains.
2327 // It might also be possible to update the pair map during runtime, but the
2328 // overhead of that may be large if there's many reassociable chains.
2329 BuildPairMap(RPOT);
2330
2331 MadeChange = false;
2332
2333 // Traverse the same blocks that were analysed by BuildRankMap.
2334 for (BasicBlock *BI : RPOT) {
2335 assert(RankMap.count(&*BI) && "BB should be ranked.")(static_cast <bool> (RankMap.count(&*BI) &&
"BB should be ranked.") ? void (0) : __assert_fail ("RankMap.count(&*BI) && \"BB should be ranked.\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 2335, __extension__ __PRETTY_FUNCTION__))
;
2336 // Optimize every instruction in the basic block.
2337 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2338 if (isInstructionTriviallyDead(&*II)) {
2339 EraseInst(&*II++);
2340 } else {
2341 OptimizeInst(&*II);
2342 assert(II->getParent() == &*BI && "Moved to a different block!")(static_cast <bool> (II->getParent() == &*BI &&
"Moved to a different block!") ? void (0) : __assert_fail ("II->getParent() == &*BI && \"Moved to a different block!\""
, "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/Reassociate.cpp"
, 2342, __extension__ __PRETTY_FUNCTION__))
;
2343 ++II;
2344 }
2345
2346 // Make a copy of all the instructions to be redone so we can remove dead
2347 // instructions.
2348 OrderedSet ToRedo(RedoInsts);
2349 // Iterate over all instructions to be reevaluated and remove trivially dead
2350 // instructions. If any operand of the trivially dead instruction becomes
2351 // dead mark it for deletion as well. Continue this process until all
2352 // trivially dead instructions have been removed.
2353 while (!ToRedo.empty()) {
2354 Instruction *I = ToRedo.pop_back_val();
2355 if (isInstructionTriviallyDead(I)) {
2356 RecursivelyEraseDeadInsts(I, ToRedo);
2357 MadeChange = true;
2358 }
2359 }
2360
2361 // Now that we have removed dead instructions, we can reoptimize the
2362 // remaining instructions.
2363 while (!RedoInsts.empty()) {
2364 Instruction *I = RedoInsts.front();
2365 RedoInsts.erase(RedoInsts.begin());
2366 if (isInstructionTriviallyDead(I))
2367 EraseInst(I);
2368 else
2369 OptimizeInst(I);
2370 }
2371 }
2372
2373 // We are done with the rank map and pair map.
2374 RankMap.clear();
2375 ValueRankMap.clear();
2376 for (auto &Entry : PairMap)
2377 Entry.clear();
2378
2379 if (MadeChange) {
2380 PreservedAnalyses PA;
2381 PA.preserveSet<CFGAnalyses>();
2382 PA.preserve<GlobalsAA>();
2383 return PA;
2384 }
2385
2386 return PreservedAnalyses::all();
2387}
2388
2389namespace {
2390
2391 class ReassociateLegacyPass : public FunctionPass {
2392 ReassociatePass Impl;
2393
2394 public:
2395 static char ID; // Pass identification, replacement for typeid
2396
2397 ReassociateLegacyPass() : FunctionPass(ID) {
2398 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2399 }
2400
2401 bool runOnFunction(Function &F) override {
2402 if (skipFunction(F))
2403 return false;
2404
2405 FunctionAnalysisManager DummyFAM;
2406 auto PA = Impl.run(F, DummyFAM);
2407 return !PA.areAllPreserved();
2408 }
2409
2410 void getAnalysisUsage(AnalysisUsage &AU) const override {
2411 AU.setPreservesCFG();
2412 AU.addPreserved<GlobalsAAWrapperPass>();
2413 }
2414 };
2415
2416} // end anonymous namespace
2417
2418char ReassociateLegacyPass::ID = 0;
2419
2420INITIALIZE_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)); }
2421 "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)); }
2422
2423// Public interface to the Reassociate pass
2424FunctionPass *llvm::createReassociatePass() {
2425 return new ReassociateLegacyPass();
2426}