1 /*
2 * Copyright (C) 2014 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "bounds_check_elimination.h"
18
19 #include <limits>
20
21 #include "base/scoped_arena_allocator.h"
22 #include "base/scoped_arena_containers.h"
23 #include "induction_var_range.h"
24 #include "nodes.h"
25 #include "side_effects_analysis.h"
26
27 namespace art {
28
29 class MonotonicValueRange;
30
31 /**
32 * A value bound is represented as a pair of value and constant,
33 * e.g. array.length - 1.
34 */
35 class ValueBound : public ValueObject {
36 public:
ValueBound(HInstruction * instruction,int32_t constant)37 ValueBound(HInstruction* instruction, int32_t constant) {
38 if (instruction != nullptr && instruction->IsIntConstant()) {
39 // Normalize ValueBound with constant instruction.
40 int32_t instr_const = instruction->AsIntConstant()->GetValue();
41 if (!WouldAddOverflowOrUnderflow(instr_const, constant)) {
42 instruction_ = nullptr;
43 constant_ = instr_const + constant;
44 return;
45 }
46 }
47 instruction_ = instruction;
48 constant_ = constant;
49 }
50
51 // Return whether (left + right) overflows or underflows.
WouldAddOverflowOrUnderflow(int32_t left,int32_t right)52 static bool WouldAddOverflowOrUnderflow(int32_t left, int32_t right) {
53 if (right == 0) {
54 return false;
55 }
56 if ((right > 0) && (left <= (std::numeric_limits<int32_t>::max() - right))) {
57 // No overflow.
58 return false;
59 }
60 if ((right < 0) && (left >= (std::numeric_limits<int32_t>::min() - right))) {
61 // No underflow.
62 return false;
63 }
64 return true;
65 }
66
67 // Return true if instruction can be expressed as "left_instruction + right_constant".
IsAddOrSubAConstant(HInstruction * instruction,HInstruction ** left_instruction,int32_t * right_constant)68 static bool IsAddOrSubAConstant(HInstruction* instruction,
69 /* out */ HInstruction** left_instruction,
70 /* out */ int32_t* right_constant) {
71 HInstruction* left_so_far = nullptr;
72 int32_t right_so_far = 0;
73 while (instruction->IsAdd() || instruction->IsSub()) {
74 HBinaryOperation* bin_op = instruction->AsBinaryOperation();
75 HInstruction* left = bin_op->GetLeft();
76 HInstruction* right = bin_op->GetRight();
77 if (right->IsIntConstant()) {
78 int32_t v = right->AsIntConstant()->GetValue();
79 int32_t c = instruction->IsAdd() ? v : -v;
80 if (!WouldAddOverflowOrUnderflow(right_so_far, c)) {
81 instruction = left;
82 left_so_far = left;
83 right_so_far += c;
84 continue;
85 }
86 }
87 break;
88 }
89 // Return result: either false and "null+0" or true and "instr+constant".
90 *left_instruction = left_so_far;
91 *right_constant = right_so_far;
92 return left_so_far != nullptr;
93 }
94
95 // Expresses any instruction as a value bound.
AsValueBound(HInstruction * instruction)96 static ValueBound AsValueBound(HInstruction* instruction) {
97 if (instruction->IsIntConstant()) {
98 return ValueBound(nullptr, instruction->AsIntConstant()->GetValue());
99 }
100 HInstruction *left;
101 int32_t right;
102 if (IsAddOrSubAConstant(instruction, &left, &right)) {
103 return ValueBound(left, right);
104 }
105 return ValueBound(instruction, 0);
106 }
107
108 // Try to detect useful value bound format from an instruction, e.g.
109 // a constant or array length related value.
DetectValueBoundFromValue(HInstruction * instruction,bool * found)110 static ValueBound DetectValueBoundFromValue(HInstruction* instruction, /* out */ bool* found) {
111 DCHECK(instruction != nullptr);
112 if (instruction->IsIntConstant()) {
113 *found = true;
114 return ValueBound(nullptr, instruction->AsIntConstant()->GetValue());
115 }
116
117 if (instruction->IsArrayLength()) {
118 *found = true;
119 return ValueBound(instruction, 0);
120 }
121 // Try to detect (array.length + c) format.
122 HInstruction *left;
123 int32_t right;
124 if (IsAddOrSubAConstant(instruction, &left, &right)) {
125 if (left->IsArrayLength()) {
126 *found = true;
127 return ValueBound(left, right);
128 }
129 }
130
131 // No useful bound detected.
132 *found = false;
133 return ValueBound::Max();
134 }
135
GetInstruction() const136 HInstruction* GetInstruction() const { return instruction_; }
GetConstant() const137 int32_t GetConstant() const { return constant_; }
138
IsRelatedToArrayLength() const139 bool IsRelatedToArrayLength() const {
140 // Some bounds are created with HNewArray* as the instruction instead
141 // of HArrayLength*. They are treated the same.
142 return (instruction_ != nullptr) &&
143 (instruction_->IsArrayLength() || instruction_->IsNewArray());
144 }
145
IsConstant() const146 bool IsConstant() const {
147 return instruction_ == nullptr;
148 }
149
Min()150 static ValueBound Min() { return ValueBound(nullptr, std::numeric_limits<int32_t>::min()); }
Max()151 static ValueBound Max() { return ValueBound(nullptr, std::numeric_limits<int32_t>::max()); }
152
Equals(ValueBound bound) const153 bool Equals(ValueBound bound) const {
154 return instruction_ == bound.instruction_ && constant_ == bound.constant_;
155 }
156
Equal(HInstruction * instruction1,HInstruction * instruction2)157 static bool Equal(HInstruction* instruction1, HInstruction* instruction2) {
158 if (instruction1 == instruction2) {
159 return true;
160 }
161 if (instruction1 == nullptr || instruction2 == nullptr) {
162 return false;
163 }
164 instruction1 = HuntForDeclaration(instruction1);
165 instruction2 = HuntForDeclaration(instruction2);
166 return instruction1 == instruction2;
167 }
168
169 // Returns if it's certain this->bound >= `bound`.
GreaterThanOrEqualTo(ValueBound bound) const170 bool GreaterThanOrEqualTo(ValueBound bound) const {
171 if (Equal(instruction_, bound.instruction_)) {
172 return constant_ >= bound.constant_;
173 }
174 // Not comparable. Just return false.
175 return false;
176 }
177
178 // Returns if it's certain this->bound <= `bound`.
LessThanOrEqualTo(ValueBound bound) const179 bool LessThanOrEqualTo(ValueBound bound) const {
180 if (Equal(instruction_, bound.instruction_)) {
181 return constant_ <= bound.constant_;
182 }
183 // Not comparable. Just return false.
184 return false;
185 }
186
187 // Returns if it's certain this->bound > `bound`.
GreaterThan(ValueBound bound) const188 bool GreaterThan(ValueBound bound) const {
189 if (Equal(instruction_, bound.instruction_)) {
190 return constant_ > bound.constant_;
191 }
192 // Not comparable. Just return false.
193 return false;
194 }
195
196 // Returns if it's certain this->bound < `bound`.
LessThan(ValueBound bound) const197 bool LessThan(ValueBound bound) const {
198 if (Equal(instruction_, bound.instruction_)) {
199 return constant_ < bound.constant_;
200 }
201 // Not comparable. Just return false.
202 return false;
203 }
204
205 // Try to narrow lower bound. Returns the greatest of the two if possible.
206 // Pick one if they are not comparable.
NarrowLowerBound(ValueBound bound1,ValueBound bound2)207 static ValueBound NarrowLowerBound(ValueBound bound1, ValueBound bound2) {
208 if (bound1.GreaterThanOrEqualTo(bound2)) {
209 return bound1;
210 }
211 if (bound2.GreaterThanOrEqualTo(bound1)) {
212 return bound2;
213 }
214
215 // Not comparable. Just pick one. We may lose some info, but that's ok.
216 // Favor constant as lower bound.
217 return bound1.IsConstant() ? bound1 : bound2;
218 }
219
220 // Try to narrow upper bound. Returns the lowest of the two if possible.
221 // Pick one if they are not comparable.
NarrowUpperBound(ValueBound bound1,ValueBound bound2)222 static ValueBound NarrowUpperBound(ValueBound bound1, ValueBound bound2) {
223 if (bound1.LessThanOrEqualTo(bound2)) {
224 return bound1;
225 }
226 if (bound2.LessThanOrEqualTo(bound1)) {
227 return bound2;
228 }
229
230 // Not comparable. Just pick one. We may lose some info, but that's ok.
231 // Favor array length as upper bound.
232 return bound1.IsRelatedToArrayLength() ? bound1 : bound2;
233 }
234
235 // Add a constant to a ValueBound.
236 // `overflow` or `underflow` will return whether the resulting bound may
237 // overflow or underflow an int.
Add(int32_t c,bool * overflow,bool * underflow) const238 ValueBound Add(int32_t c, /* out */ bool* overflow, /* out */ bool* underflow) const {
239 *overflow = *underflow = false;
240 if (c == 0) {
241 return *this;
242 }
243
244 int32_t new_constant;
245 if (c > 0) {
246 if (constant_ > (std::numeric_limits<int32_t>::max() - c)) {
247 *overflow = true;
248 return Max();
249 }
250
251 new_constant = constant_ + c;
252 // (array.length + non-positive-constant) won't overflow an int.
253 if (IsConstant() || (IsRelatedToArrayLength() && new_constant <= 0)) {
254 return ValueBound(instruction_, new_constant);
255 }
256 // Be conservative.
257 *overflow = true;
258 return Max();
259 } else {
260 if (constant_ < (std::numeric_limits<int32_t>::min() - c)) {
261 *underflow = true;
262 return Min();
263 }
264
265 new_constant = constant_ + c;
266 // Regardless of the value new_constant, (array.length+new_constant) will
267 // never underflow since array.length is no less than 0.
268 if (IsConstant() || IsRelatedToArrayLength()) {
269 return ValueBound(instruction_, new_constant);
270 }
271 // Be conservative.
272 *underflow = true;
273 return Min();
274 }
275 }
276
277 private:
278 HInstruction* instruction_;
279 int32_t constant_;
280 };
281
282 /**
283 * Represent a range of lower bound and upper bound, both being inclusive.
284 * Currently a ValueRange may be generated as a result of the following:
285 * comparisons related to array bounds, array bounds check, add/sub on top
286 * of an existing value range, NewArray or a loop phi corresponding to an
287 * incrementing/decrementing array index (MonotonicValueRange).
288 */
289 class ValueRange : public ArenaObject<kArenaAllocBoundsCheckElimination> {
290 public:
ValueRange(ScopedArenaAllocator * allocator,ValueBound lower,ValueBound upper)291 ValueRange(ScopedArenaAllocator* allocator, ValueBound lower, ValueBound upper)
292 : allocator_(allocator), lower_(lower), upper_(upper) {}
293
~ValueRange()294 virtual ~ValueRange() {}
295
AsMonotonicValueRange()296 virtual MonotonicValueRange* AsMonotonicValueRange() { return nullptr; }
IsMonotonicValueRange()297 bool IsMonotonicValueRange() {
298 return AsMonotonicValueRange() != nullptr;
299 }
300
GetAllocator() const301 ScopedArenaAllocator* GetAllocator() const { return allocator_; }
GetLower() const302 ValueBound GetLower() const { return lower_; }
GetUpper() const303 ValueBound GetUpper() const { return upper_; }
304
IsConstantValueRange() const305 bool IsConstantValueRange() const { return lower_.IsConstant() && upper_.IsConstant(); }
306
307 // If it's certain that this value range fits in other_range.
FitsIn(ValueRange * other_range) const308 virtual bool FitsIn(ValueRange* other_range) const {
309 if (other_range == nullptr) {
310 return true;
311 }
312 DCHECK(!other_range->IsMonotonicValueRange());
313 return lower_.GreaterThanOrEqualTo(other_range->lower_) &&
314 upper_.LessThanOrEqualTo(other_range->upper_);
315 }
316
317 // Returns the intersection of this and range.
318 // If it's not possible to do intersection because some
319 // bounds are not comparable, it's ok to pick either bound.
Narrow(ValueRange * range)320 virtual ValueRange* Narrow(ValueRange* range) {
321 if (range == nullptr) {
322 return this;
323 }
324
325 if (range->IsMonotonicValueRange()) {
326 return this;
327 }
328
329 return new (allocator_) ValueRange(
330 allocator_,
331 ValueBound::NarrowLowerBound(lower_, range->lower_),
332 ValueBound::NarrowUpperBound(upper_, range->upper_));
333 }
334
335 // Shift a range by a constant.
Add(int32_t constant) const336 ValueRange* Add(int32_t constant) const {
337 bool overflow, underflow;
338 ValueBound lower = lower_.Add(constant, &overflow, &underflow);
339 if (underflow) {
340 // Lower bound underflow will wrap around to positive values
341 // and invalidate the upper bound.
342 return nullptr;
343 }
344 ValueBound upper = upper_.Add(constant, &overflow, &underflow);
345 if (overflow) {
346 // Upper bound overflow will wrap around to negative values
347 // and invalidate the lower bound.
348 return nullptr;
349 }
350 return new (allocator_) ValueRange(allocator_, lower, upper);
351 }
352
353 private:
354 ScopedArenaAllocator* const allocator_;
355 const ValueBound lower_; // inclusive
356 const ValueBound upper_; // inclusive
357
358 DISALLOW_COPY_AND_ASSIGN(ValueRange);
359 };
360
361 /**
362 * A monotonically incrementing/decrementing value range, e.g.
363 * the variable i in "for (int i=0; i<array.length; i++)".
364 * Special care needs to be taken to account for overflow/underflow
365 * of such value ranges.
366 */
367 class MonotonicValueRange : public ValueRange {
368 public:
MonotonicValueRange(ScopedArenaAllocator * allocator,HPhi * induction_variable,HInstruction * initial,int32_t increment,ValueBound bound)369 MonotonicValueRange(ScopedArenaAllocator* allocator,
370 HPhi* induction_variable,
371 HInstruction* initial,
372 int32_t increment,
373 ValueBound bound)
374 // To be conservative, give it full range [Min(), Max()] in case it's
375 // used as a regular value range, due to possible overflow/underflow.
376 : ValueRange(allocator, ValueBound::Min(), ValueBound::Max()),
377 induction_variable_(induction_variable),
378 initial_(initial),
379 increment_(increment),
380 bound_(bound) {}
381
~MonotonicValueRange()382 virtual ~MonotonicValueRange() {}
383
GetIncrement() const384 int32_t GetIncrement() const { return increment_; }
GetBound() const385 ValueBound GetBound() const { return bound_; }
GetLoopHeader() const386 HBasicBlock* GetLoopHeader() const {
387 DCHECK(induction_variable_->GetBlock()->IsLoopHeader());
388 return induction_variable_->GetBlock();
389 }
390
AsMonotonicValueRange()391 MonotonicValueRange* AsMonotonicValueRange() override { return this; }
392
393 // If it's certain that this value range fits in other_range.
FitsIn(ValueRange * other_range) const394 bool FitsIn(ValueRange* other_range) const override {
395 if (other_range == nullptr) {
396 return true;
397 }
398 DCHECK(!other_range->IsMonotonicValueRange());
399 return false;
400 }
401
402 // Try to narrow this MonotonicValueRange given another range.
403 // Ideally it will return a normal ValueRange. But due to
404 // possible overflow/underflow, that may not be possible.
Narrow(ValueRange * range)405 ValueRange* Narrow(ValueRange* range) override {
406 if (range == nullptr) {
407 return this;
408 }
409 DCHECK(!range->IsMonotonicValueRange());
410
411 if (increment_ > 0) {
412 // Monotonically increasing.
413 ValueBound lower = ValueBound::NarrowLowerBound(bound_, range->GetLower());
414 if (!lower.IsConstant() || lower.GetConstant() == std::numeric_limits<int32_t>::min()) {
415 // Lower bound isn't useful. Leave it to deoptimization.
416 return this;
417 }
418
419 // We currently conservatively assume max array length is Max().
420 // If we can make assumptions about the max array length, e.g. due to the max heap size,
421 // divided by the element size (such as 4 bytes for each integer array), we can
422 // lower this number and rule out some possible overflows.
423 int32_t max_array_len = std::numeric_limits<int32_t>::max();
424
425 // max possible integer value of range's upper value.
426 int32_t upper = std::numeric_limits<int32_t>::max();
427 // Try to lower upper.
428 ValueBound upper_bound = range->GetUpper();
429 if (upper_bound.IsConstant()) {
430 upper = upper_bound.GetConstant();
431 } else if (upper_bound.IsRelatedToArrayLength() && upper_bound.GetConstant() <= 0) {
432 // Normal case. e.g. <= array.length - 1.
433 upper = max_array_len + upper_bound.GetConstant();
434 }
435
436 // If we can prove for the last number in sequence of initial_,
437 // initial_ + increment_, initial_ + 2 x increment_, ...
438 // that's <= upper, (last_num_in_sequence + increment_) doesn't trigger overflow,
439 // then this MonoticValueRange is narrowed to a normal value range.
440
441 // Be conservative first, assume last number in the sequence hits upper.
442 int32_t last_num_in_sequence = upper;
443 if (initial_->IsIntConstant()) {
444 int32_t initial_constant = initial_->AsIntConstant()->GetValue();
445 if (upper <= initial_constant) {
446 last_num_in_sequence = upper;
447 } else {
448 // Cast to int64_t for the substraction part to avoid int32_t overflow.
449 last_num_in_sequence = initial_constant +
450 ((int64_t)upper - (int64_t)initial_constant) / increment_ * increment_;
451 }
452 }
453 if (last_num_in_sequence <= (std::numeric_limits<int32_t>::max() - increment_)) {
454 // No overflow. The sequence will be stopped by the upper bound test as expected.
455 return new (GetAllocator()) ValueRange(GetAllocator(), lower, range->GetUpper());
456 }
457
458 // There might be overflow. Give up narrowing.
459 return this;
460 } else {
461 DCHECK_NE(increment_, 0);
462 // Monotonically decreasing.
463 ValueBound upper = ValueBound::NarrowUpperBound(bound_, range->GetUpper());
464 if ((!upper.IsConstant() || upper.GetConstant() == std::numeric_limits<int32_t>::max()) &&
465 !upper.IsRelatedToArrayLength()) {
466 // Upper bound isn't useful. Leave it to deoptimization.
467 return this;
468 }
469
470 // Need to take care of underflow. Try to prove underflow won't happen
471 // for common cases.
472 if (range->GetLower().IsConstant()) {
473 int32_t constant = range->GetLower().GetConstant();
474 if (constant >= (std::numeric_limits<int32_t>::min() - increment_)) {
475 return new (GetAllocator()) ValueRange(GetAllocator(), range->GetLower(), upper);
476 }
477 }
478
479 // For non-constant lower bound, just assume might be underflow. Give up narrowing.
480 return this;
481 }
482 }
483
484 private:
485 HPhi* const induction_variable_; // Induction variable for this monotonic value range.
486 HInstruction* const initial_; // Initial value.
487 const int32_t increment_; // Increment for each loop iteration.
488 const ValueBound bound_; // Additional value bound info for initial_.
489
490 DISALLOW_COPY_AND_ASSIGN(MonotonicValueRange);
491 };
492
493 class BCEVisitor : public HGraphVisitor {
494 public:
495 // The least number of bounds checks that should be eliminated by triggering
496 // the deoptimization technique.
497 static constexpr size_t kThresholdForAddingDeoptimize = 2;
498
499 // Very large lengths are considered an anomaly. This is a threshold beyond which we don't
500 // bother to apply the deoptimization technique since it's likely, or sometimes certain,
501 // an AIOOBE will be thrown.
502 static constexpr uint32_t kMaxLengthForAddingDeoptimize =
503 std::numeric_limits<int32_t>::max() - 1024 * 1024;
504
505 // Added blocks for loop body entry test.
IsAddedBlock(HBasicBlock * block) const506 bool IsAddedBlock(HBasicBlock* block) const {
507 return block->GetBlockId() >= initial_block_size_;
508 }
509
BCEVisitor(HGraph * graph,const SideEffectsAnalysis & side_effects,HInductionVarAnalysis * induction_analysis)510 BCEVisitor(HGraph* graph,
511 const SideEffectsAnalysis& side_effects,
512 HInductionVarAnalysis* induction_analysis)
513 : HGraphVisitor(graph),
514 allocator_(graph->GetArenaStack()),
515 maps_(graph->GetBlocks().size(),
516 ScopedArenaSafeMap<int, ValueRange*>(
517 std::less<int>(),
518 allocator_.Adapter(kArenaAllocBoundsCheckElimination)),
519 allocator_.Adapter(kArenaAllocBoundsCheckElimination)),
520 first_index_bounds_check_map_(std::less<int>(),
521 allocator_.Adapter(kArenaAllocBoundsCheckElimination)),
522 early_exit_loop_(std::less<uint32_t>(),
523 allocator_.Adapter(kArenaAllocBoundsCheckElimination)),
524 taken_test_loop_(std::less<uint32_t>(),
525 allocator_.Adapter(kArenaAllocBoundsCheckElimination)),
526 finite_loop_(allocator_.Adapter(kArenaAllocBoundsCheckElimination)),
527 has_dom_based_dynamic_bce_(false),
528 initial_block_size_(graph->GetBlocks().size()),
529 side_effects_(side_effects),
530 induction_range_(induction_analysis),
531 next_(nullptr) {}
532
VisitBasicBlock(HBasicBlock * block)533 void VisitBasicBlock(HBasicBlock* block) override {
534 DCHECK(!IsAddedBlock(block));
535 first_index_bounds_check_map_.clear();
536 // Visit phis and instructions using a safe iterator. The iteration protects
537 // against deleting the current instruction during iteration. However, it
538 // must advance next_ if that instruction is deleted during iteration.
539 for (HInstruction* instruction = block->GetFirstPhi(); instruction != nullptr;) {
540 DCHECK(instruction->IsInBlock());
541 next_ = instruction->GetNext();
542 instruction->Accept(this);
543 instruction = next_;
544 }
545 for (HInstruction* instruction = block->GetFirstInstruction(); instruction != nullptr;) {
546 DCHECK(instruction->IsInBlock());
547 next_ = instruction->GetNext();
548 instruction->Accept(this);
549 instruction = next_;
550 }
551 // We should never deoptimize from an osr method, otherwise we might wrongly optimize
552 // code dominated by the deoptimization.
553 if (!GetGraph()->IsCompilingOsr()) {
554 AddComparesWithDeoptimization(block);
555 }
556 }
557
Finish()558 void Finish() {
559 // Preserve SSA structure which may have been broken by adding one or more
560 // new taken-test structures (see TransformLoopForDeoptimizationIfNeeded()).
561 InsertPhiNodes();
562
563 // Clear the loop data structures.
564 early_exit_loop_.clear();
565 taken_test_loop_.clear();
566 finite_loop_.clear();
567 }
568
569 private:
570 // Return the map of proven value ranges at the beginning of a basic block.
GetValueRangeMap(HBasicBlock * basic_block)571 ScopedArenaSafeMap<int, ValueRange*>* GetValueRangeMap(HBasicBlock* basic_block) {
572 if (IsAddedBlock(basic_block)) {
573 // Added blocks don't keep value ranges.
574 return nullptr;
575 }
576 return &maps_[basic_block->GetBlockId()];
577 }
578
579 // Traverse up the dominator tree to look for value range info.
LookupValueRange(HInstruction * instruction,HBasicBlock * basic_block)580 ValueRange* LookupValueRange(HInstruction* instruction, HBasicBlock* basic_block) {
581 while (basic_block != nullptr) {
582 ScopedArenaSafeMap<int, ValueRange*>* map = GetValueRangeMap(basic_block);
583 if (map != nullptr) {
584 if (map->find(instruction->GetId()) != map->end()) {
585 return map->Get(instruction->GetId());
586 }
587 } else {
588 DCHECK(IsAddedBlock(basic_block));
589 }
590 basic_block = basic_block->GetDominator();
591 }
592 // Didn't find any.
593 return nullptr;
594 }
595
596 // Helper method to assign a new range to an instruction in given basic block.
AssignRange(HBasicBlock * basic_block,HInstruction * instruction,ValueRange * range)597 void AssignRange(HBasicBlock* basic_block, HInstruction* instruction, ValueRange* range) {
598 DCHECK(!range->IsMonotonicValueRange() || instruction->IsLoopHeaderPhi());
599 GetValueRangeMap(basic_block)->Overwrite(instruction->GetId(), range);
600 }
601
602 // Narrow the value range of `instruction` at the end of `basic_block` with `range`,
603 // and push the narrowed value range to `successor`.
ApplyRangeFromComparison(HInstruction * instruction,HBasicBlock * basic_block,HBasicBlock * successor,ValueRange * range)604 void ApplyRangeFromComparison(HInstruction* instruction, HBasicBlock* basic_block,
605 HBasicBlock* successor, ValueRange* range) {
606 ValueRange* existing_range = LookupValueRange(instruction, basic_block);
607 if (existing_range == nullptr) {
608 if (range != nullptr) {
609 AssignRange(successor, instruction, range);
610 }
611 return;
612 }
613 if (existing_range->IsMonotonicValueRange()) {
614 DCHECK(instruction->IsLoopHeaderPhi());
615 // Make sure the comparison is in the loop header so each increment is
616 // checked with a comparison.
617 if (instruction->GetBlock() != basic_block) {
618 return;
619 }
620 }
621 AssignRange(successor, instruction, existing_range->Narrow(range));
622 }
623
624 // Special case that we may simultaneously narrow two MonotonicValueRange's to
625 // regular value ranges.
HandleIfBetweenTwoMonotonicValueRanges(HIf * instruction,HInstruction * left,HInstruction * right,IfCondition cond,MonotonicValueRange * left_range,MonotonicValueRange * right_range)626 void HandleIfBetweenTwoMonotonicValueRanges(HIf* instruction,
627 HInstruction* left,
628 HInstruction* right,
629 IfCondition cond,
630 MonotonicValueRange* left_range,
631 MonotonicValueRange* right_range) {
632 DCHECK(left->IsLoopHeaderPhi());
633 DCHECK(right->IsLoopHeaderPhi());
634 if (instruction->GetBlock() != left->GetBlock()) {
635 // Comparison needs to be in loop header to make sure it's done after each
636 // increment/decrement.
637 return;
638 }
639
640 // Handle common cases which also don't have overflow/underflow concerns.
641 if (left_range->GetIncrement() == 1 &&
642 left_range->GetBound().IsConstant() &&
643 right_range->GetIncrement() == -1 &&
644 right_range->GetBound().IsRelatedToArrayLength() &&
645 right_range->GetBound().GetConstant() < 0) {
646 HBasicBlock* successor = nullptr;
647 int32_t left_compensation = 0;
648 int32_t right_compensation = 0;
649 if (cond == kCondLT) {
650 left_compensation = -1;
651 right_compensation = 1;
652 successor = instruction->IfTrueSuccessor();
653 } else if (cond == kCondLE) {
654 successor = instruction->IfTrueSuccessor();
655 } else if (cond == kCondGT) {
656 successor = instruction->IfFalseSuccessor();
657 } else if (cond == kCondGE) {
658 left_compensation = -1;
659 right_compensation = 1;
660 successor = instruction->IfFalseSuccessor();
661 } else {
662 // We don't handle '=='/'!=' test in case left and right can cross and
663 // miss each other.
664 return;
665 }
666
667 if (successor != nullptr) {
668 bool overflow;
669 bool underflow;
670 ValueRange* new_left_range = new (&allocator_) ValueRange(
671 &allocator_,
672 left_range->GetBound(),
673 right_range->GetBound().Add(left_compensation, &overflow, &underflow));
674 if (!overflow && !underflow) {
675 ApplyRangeFromComparison(left, instruction->GetBlock(), successor,
676 new_left_range);
677 }
678
679 ValueRange* new_right_range = new (&allocator_) ValueRange(
680 &allocator_,
681 left_range->GetBound().Add(right_compensation, &overflow, &underflow),
682 right_range->GetBound());
683 if (!overflow && !underflow) {
684 ApplyRangeFromComparison(right, instruction->GetBlock(), successor,
685 new_right_range);
686 }
687 }
688 }
689 }
690
691 // Handle "if (left cmp_cond right)".
HandleIf(HIf * instruction,HInstruction * left,HInstruction * right,IfCondition cond)692 void HandleIf(HIf* instruction, HInstruction* left, HInstruction* right, IfCondition cond) {
693 HBasicBlock* block = instruction->GetBlock();
694
695 HBasicBlock* true_successor = instruction->IfTrueSuccessor();
696 // There should be no critical edge at this point.
697 DCHECK_EQ(true_successor->GetPredecessors().size(), 1u);
698
699 HBasicBlock* false_successor = instruction->IfFalseSuccessor();
700 // There should be no critical edge at this point.
701 DCHECK_EQ(false_successor->GetPredecessors().size(), 1u);
702
703 ValueRange* left_range = LookupValueRange(left, block);
704 MonotonicValueRange* left_monotonic_range = nullptr;
705 if (left_range != nullptr) {
706 left_monotonic_range = left_range->AsMonotonicValueRange();
707 if (left_monotonic_range != nullptr) {
708 HBasicBlock* loop_head = left_monotonic_range->GetLoopHeader();
709 if (instruction->GetBlock() != loop_head) {
710 // For monotonic value range, don't handle `instruction`
711 // if it's not defined in the loop header.
712 return;
713 }
714 }
715 }
716
717 bool found;
718 ValueBound bound = ValueBound::DetectValueBoundFromValue(right, &found);
719 // Each comparison can establish a lower bound and an upper bound
720 // for the left hand side.
721 ValueBound lower = bound;
722 ValueBound upper = bound;
723 if (!found) {
724 // No constant or array.length+c format bound found.
725 // For i<j, we can still use j's upper bound as i's upper bound. Same for lower.
726 ValueRange* right_range = LookupValueRange(right, block);
727 if (right_range != nullptr) {
728 if (right_range->IsMonotonicValueRange()) {
729 if (left_range != nullptr && left_range->IsMonotonicValueRange()) {
730 HandleIfBetweenTwoMonotonicValueRanges(instruction, left, right, cond,
731 left_range->AsMonotonicValueRange(),
732 right_range->AsMonotonicValueRange());
733 return;
734 }
735 }
736 lower = right_range->GetLower();
737 upper = right_range->GetUpper();
738 } else {
739 lower = ValueBound::Min();
740 upper = ValueBound::Max();
741 }
742 }
743
744 bool overflow, underflow;
745 if (cond == kCondLT || cond == kCondLE) {
746 if (!upper.Equals(ValueBound::Max())) {
747 int32_t compensation = (cond == kCondLT) ? -1 : 0; // upper bound is inclusive
748 ValueBound new_upper = upper.Add(compensation, &overflow, &underflow);
749 if (overflow || underflow) {
750 return;
751 }
752 ValueRange* new_range = new (&allocator_) ValueRange(
753 &allocator_, ValueBound::Min(), new_upper);
754 ApplyRangeFromComparison(left, block, true_successor, new_range);
755 }
756
757 // array.length as a lower bound isn't considered useful.
758 if (!lower.Equals(ValueBound::Min()) && !lower.IsRelatedToArrayLength()) {
759 int32_t compensation = (cond == kCondLE) ? 1 : 0; // lower bound is inclusive
760 ValueBound new_lower = lower.Add(compensation, &overflow, &underflow);
761 if (overflow || underflow) {
762 return;
763 }
764 ValueRange* new_range = new (&allocator_) ValueRange(
765 &allocator_, new_lower, ValueBound::Max());
766 ApplyRangeFromComparison(left, block, false_successor, new_range);
767 }
768 } else if (cond == kCondGT || cond == kCondGE) {
769 // array.length as a lower bound isn't considered useful.
770 if (!lower.Equals(ValueBound::Min()) && !lower.IsRelatedToArrayLength()) {
771 int32_t compensation = (cond == kCondGT) ? 1 : 0; // lower bound is inclusive
772 ValueBound new_lower = lower.Add(compensation, &overflow, &underflow);
773 if (overflow || underflow) {
774 return;
775 }
776 ValueRange* new_range = new (&allocator_) ValueRange(
777 &allocator_, new_lower, ValueBound::Max());
778 ApplyRangeFromComparison(left, block, true_successor, new_range);
779 }
780
781 if (!upper.Equals(ValueBound::Max())) {
782 int32_t compensation = (cond == kCondGE) ? -1 : 0; // upper bound is inclusive
783 ValueBound new_upper = upper.Add(compensation, &overflow, &underflow);
784 if (overflow || underflow) {
785 return;
786 }
787 ValueRange* new_range = new (&allocator_) ValueRange(
788 &allocator_, ValueBound::Min(), new_upper);
789 ApplyRangeFromComparison(left, block, false_successor, new_range);
790 }
791 } else if (cond == kCondNE || cond == kCondEQ) {
792 if (left->IsArrayLength()) {
793 if (lower.IsConstant() && upper.IsConstant()) {
794 // Special case:
795 // length == [c,d] yields [c, d] along true
796 // length != [c,d] yields [c, d] along false
797 if (!lower.Equals(ValueBound::Min()) || !upper.Equals(ValueBound::Max())) {
798 ValueRange* new_range = new (&allocator_) ValueRange(&allocator_, lower, upper);
799 ApplyRangeFromComparison(
800 left, block, cond == kCondEQ ? true_successor : false_successor, new_range);
801 }
802 // In addition:
803 // length == 0 yields [1, max] along false
804 // length != 0 yields [1, max] along true
805 if (lower.GetConstant() == 0 && upper.GetConstant() == 0) {
806 ValueRange* new_range = new (&allocator_) ValueRange(
807 &allocator_, ValueBound(nullptr, 1), ValueBound::Max());
808 ApplyRangeFromComparison(
809 left, block, cond == kCondEQ ? false_successor : true_successor, new_range);
810 }
811 }
812 } else if (lower.IsRelatedToArrayLength() && lower.Equals(upper)) {
813 // Special aliasing case, with x not array length itself:
814 // x == [length,length] yields x == length along true
815 // x != [length,length] yields x == length along false
816 ValueRange* new_range = new (&allocator_) ValueRange(&allocator_, lower, upper);
817 ApplyRangeFromComparison(
818 left, block, cond == kCondEQ ? true_successor : false_successor, new_range);
819 }
820 }
821 }
822
VisitBoundsCheck(HBoundsCheck * bounds_check)823 void VisitBoundsCheck(HBoundsCheck* bounds_check) override {
824 HBasicBlock* block = bounds_check->GetBlock();
825 HInstruction* index = bounds_check->InputAt(0);
826 HInstruction* array_length = bounds_check->InputAt(1);
827 DCHECK(array_length->IsIntConstant() ||
828 array_length->IsArrayLength() ||
829 array_length->IsPhi());
830 bool try_dynamic_bce = true;
831 // Analyze index range.
832 if (!index->IsIntConstant()) {
833 // Non-constant index.
834 ValueBound lower = ValueBound(nullptr, 0); // constant 0
835 ValueBound upper = ValueBound(array_length, -1); // array_length - 1
836 ValueRange array_range(&allocator_, lower, upper);
837 // Try index range obtained by dominator-based analysis.
838 ValueRange* index_range = LookupValueRange(index, block);
839 if (index_range != nullptr) {
840 if (index_range->FitsIn(&array_range)) {
841 ReplaceInstruction(bounds_check, index);
842 return;
843 } else if (index_range->IsConstantValueRange()) {
844 // If the non-constant index turns out to have a constant range,
845 // make one more attempt to get a constant in the array range.
846 ValueRange* existing_range = LookupValueRange(array_length, block);
847 if (existing_range != nullptr &&
848 existing_range->IsConstantValueRange() &&
849 existing_range->GetLower().GetConstant() > 0) {
850 ValueBound constant_upper(nullptr, existing_range->GetLower().GetConstant() - 1);
851 ValueRange constant_array_range(&allocator_, lower, constant_upper);
852 if (index_range->FitsIn(&constant_array_range)) {
853 ReplaceInstruction(bounds_check, index);
854 return;
855 }
856 }
857 }
858 }
859 // Try index range obtained by induction variable analysis.
860 // Disables dynamic bce if OOB is certain.
861 if (InductionRangeFitsIn(&array_range, bounds_check, &try_dynamic_bce)) {
862 ReplaceInstruction(bounds_check, index);
863 return;
864 }
865 } else {
866 // Constant index.
867 int32_t constant = index->AsIntConstant()->GetValue();
868 if (constant < 0) {
869 // Will always throw exception.
870 return;
871 } else if (array_length->IsIntConstant()) {
872 if (constant < array_length->AsIntConstant()->GetValue()) {
873 ReplaceInstruction(bounds_check, index);
874 }
875 return;
876 }
877 // Analyze array length range.
878 DCHECK(array_length->IsArrayLength());
879 ValueRange* existing_range = LookupValueRange(array_length, block);
880 if (existing_range != nullptr) {
881 ValueBound lower = existing_range->GetLower();
882 DCHECK(lower.IsConstant());
883 if (constant < lower.GetConstant()) {
884 ReplaceInstruction(bounds_check, index);
885 return;
886 } else {
887 // Existing range isn't strong enough to eliminate the bounds check.
888 // Fall through to update the array_length range with info from this
889 // bounds check.
890 }
891 }
892 // Once we have an array access like 'array[5] = 1', we record array.length >= 6.
893 // We currently don't do it for non-constant index since a valid array[i] can't prove
894 // a valid array[i-1] yet due to the lower bound side.
895 if (constant == std::numeric_limits<int32_t>::max()) {
896 // Max() as an index will definitely throw AIOOBE.
897 return;
898 } else {
899 ValueBound lower = ValueBound(nullptr, constant + 1);
900 ValueBound upper = ValueBound::Max();
901 ValueRange* range = new (&allocator_) ValueRange(&allocator_, lower, upper);
902 AssignRange(block, array_length, range);
903 }
904 }
905
906 // If static analysis fails, and OOB is not certain, try dynamic elimination.
907 if (try_dynamic_bce) {
908 // Try loop-based dynamic elimination.
909 HLoopInformation* loop = bounds_check->GetBlock()->GetLoopInformation();
910 bool needs_finite_test = false;
911 bool needs_taken_test = false;
912 if (DynamicBCESeemsProfitable(loop, bounds_check->GetBlock()) &&
913 induction_range_.CanGenerateRange(
914 bounds_check, index, &needs_finite_test, &needs_taken_test) &&
915 CanHandleInfiniteLoop(loop, index, needs_finite_test) &&
916 // Do this test last, since it may generate code.
917 CanHandleLength(loop, array_length, needs_taken_test)) {
918 TransformLoopForDeoptimizationIfNeeded(loop, needs_taken_test);
919 TransformLoopForDynamicBCE(loop, bounds_check);
920 return;
921 }
922 // Otherwise, prepare dominator-based dynamic elimination.
923 if (first_index_bounds_check_map_.find(array_length->GetId()) ==
924 first_index_bounds_check_map_.end()) {
925 // Remember the first bounds check against each array_length. That bounds check
926 // instruction has an associated HEnvironment where we may add an HDeoptimize
927 // to eliminate subsequent bounds checks against the same array_length.
928 first_index_bounds_check_map_.Put(array_length->GetId(), bounds_check);
929 }
930 }
931 }
932
HasSameInputAtBackEdges(HPhi * phi)933 static bool HasSameInputAtBackEdges(HPhi* phi) {
934 DCHECK(phi->IsLoopHeaderPhi());
935 HConstInputsRef inputs = phi->GetInputs();
936 // Start with input 1. Input 0 is from the incoming block.
937 const HInstruction* input1 = inputs[1];
938 DCHECK(phi->GetBlock()->GetLoopInformation()->IsBackEdge(
939 *phi->GetBlock()->GetPredecessors()[1]));
940 for (size_t i = 2; i < inputs.size(); ++i) {
941 DCHECK(phi->GetBlock()->GetLoopInformation()->IsBackEdge(
942 *phi->GetBlock()->GetPredecessors()[i]));
943 if (input1 != inputs[i]) {
944 return false;
945 }
946 }
947 return true;
948 }
949
VisitPhi(HPhi * phi)950 void VisitPhi(HPhi* phi) override {
951 if (phi->IsLoopHeaderPhi()
952 && (phi->GetType() == DataType::Type::kInt32)
953 && HasSameInputAtBackEdges(phi)) {
954 HInstruction* instruction = phi->InputAt(1);
955 HInstruction *left;
956 int32_t increment;
957 if (ValueBound::IsAddOrSubAConstant(instruction, &left, &increment)) {
958 if (left == phi) {
959 HInstruction* initial_value = phi->InputAt(0);
960 ValueRange* range = nullptr;
961 if (increment == 0) {
962 // Add constant 0. It's really a fixed value.
963 range = new (&allocator_) ValueRange(
964 &allocator_,
965 ValueBound(initial_value, 0),
966 ValueBound(initial_value, 0));
967 } else {
968 // Monotonically increasing/decreasing.
969 bool found;
970 ValueBound bound = ValueBound::DetectValueBoundFromValue(
971 initial_value, &found);
972 if (!found) {
973 // No constant or array.length+c bound found.
974 // For i=j, we can still use j's upper bound as i's upper bound.
975 // Same for lower.
976 ValueRange* initial_range = LookupValueRange(initial_value, phi->GetBlock());
977 if (initial_range != nullptr) {
978 bound = increment > 0 ? initial_range->GetLower() :
979 initial_range->GetUpper();
980 } else {
981 bound = increment > 0 ? ValueBound::Min() : ValueBound::Max();
982 }
983 }
984 range = new (&allocator_) MonotonicValueRange(
985 &allocator_,
986 phi,
987 initial_value,
988 increment,
989 bound);
990 }
991 AssignRange(phi->GetBlock(), phi, range);
992 }
993 }
994 }
995 }
996
VisitIf(HIf * instruction)997 void VisitIf(HIf* instruction) override {
998 if (instruction->InputAt(0)->IsCondition()) {
999 HCondition* cond = instruction->InputAt(0)->AsCondition();
1000 HandleIf(instruction, cond->GetLeft(), cond->GetRight(), cond->GetCondition());
1001 }
1002 }
1003
VisitAdd(HAdd * add)1004 void VisitAdd(HAdd* add) override {
1005 HInstruction* right = add->GetRight();
1006 if (right->IsIntConstant()) {
1007 ValueRange* left_range = LookupValueRange(add->GetLeft(), add->GetBlock());
1008 if (left_range == nullptr) {
1009 return;
1010 }
1011 ValueRange* range = left_range->Add(right->AsIntConstant()->GetValue());
1012 if (range != nullptr) {
1013 AssignRange(add->GetBlock(), add, range);
1014 }
1015 }
1016 }
1017
VisitSub(HSub * sub)1018 void VisitSub(HSub* sub) override {
1019 HInstruction* left = sub->GetLeft();
1020 HInstruction* right = sub->GetRight();
1021 if (right->IsIntConstant()) {
1022 ValueRange* left_range = LookupValueRange(left, sub->GetBlock());
1023 if (left_range == nullptr) {
1024 return;
1025 }
1026 ValueRange* range = left_range->Add(-right->AsIntConstant()->GetValue());
1027 if (range != nullptr) {
1028 AssignRange(sub->GetBlock(), sub, range);
1029 return;
1030 }
1031 }
1032
1033 // Here we are interested in the typical triangular case of nested loops,
1034 // such as the inner loop 'for (int j=0; j<array.length-i; j++)' where i
1035 // is the index for outer loop. In this case, we know j is bounded by array.length-1.
1036
1037 // Try to handle (array.length - i) or (array.length + c - i) format.
1038 HInstruction* left_of_left; // left input of left.
1039 int32_t right_const = 0;
1040 if (ValueBound::IsAddOrSubAConstant(left, &left_of_left, &right_const)) {
1041 left = left_of_left;
1042 }
1043 // The value of left input of the sub equals (left + right_const).
1044
1045 if (left->IsArrayLength()) {
1046 HInstruction* array_length = left->AsArrayLength();
1047 ValueRange* right_range = LookupValueRange(right, sub->GetBlock());
1048 if (right_range != nullptr) {
1049 ValueBound lower = right_range->GetLower();
1050 ValueBound upper = right_range->GetUpper();
1051 if (lower.IsConstant() && upper.IsRelatedToArrayLength()) {
1052 HInstruction* upper_inst = upper.GetInstruction();
1053 // Make sure it's the same array.
1054 if (ValueBound::Equal(array_length, upper_inst)) {
1055 int32_t c0 = right_const;
1056 int32_t c1 = lower.GetConstant();
1057 int32_t c2 = upper.GetConstant();
1058 // (array.length + c0 - v) where v is in [c1, array.length + c2]
1059 // gets [c0 - c2, array.length + c0 - c1] as its value range.
1060 if (!ValueBound::WouldAddOverflowOrUnderflow(c0, -c2) &&
1061 !ValueBound::WouldAddOverflowOrUnderflow(c0, -c1)) {
1062 if ((c0 - c1) <= 0) {
1063 // array.length + (c0 - c1) won't overflow/underflow.
1064 ValueRange* range = new (&allocator_) ValueRange(
1065 &allocator_,
1066 ValueBound(nullptr, right_const - upper.GetConstant()),
1067 ValueBound(array_length, right_const - lower.GetConstant()));
1068 AssignRange(sub->GetBlock(), sub, range);
1069 }
1070 }
1071 }
1072 }
1073 }
1074 }
1075 }
1076
FindAndHandlePartialArrayLength(HBinaryOperation * instruction)1077 void FindAndHandlePartialArrayLength(HBinaryOperation* instruction) {
1078 DCHECK(instruction->IsDiv() || instruction->IsShr() || instruction->IsUShr());
1079 HInstruction* right = instruction->GetRight();
1080 int32_t right_const;
1081 if (right->IsIntConstant()) {
1082 right_const = right->AsIntConstant()->GetValue();
1083 // Detect division by two or more.
1084 if ((instruction->IsDiv() && right_const <= 1) ||
1085 (instruction->IsShr() && right_const < 1) ||
1086 (instruction->IsUShr() && right_const < 1)) {
1087 return;
1088 }
1089 } else {
1090 return;
1091 }
1092
1093 // Try to handle array.length/2 or (array.length-1)/2 format.
1094 HInstruction* left = instruction->GetLeft();
1095 HInstruction* left_of_left; // left input of left.
1096 int32_t c = 0;
1097 if (ValueBound::IsAddOrSubAConstant(left, &left_of_left, &c)) {
1098 left = left_of_left;
1099 }
1100 // The value of left input of instruction equals (left + c).
1101
1102 // (array_length + 1) or smaller divided by two or more
1103 // always generate a value in [Min(), array_length].
1104 // This is true even if array_length is Max().
1105 if (left->IsArrayLength() && c <= 1) {
1106 if (instruction->IsUShr() && c < 0) {
1107 // Make sure for unsigned shift, left side is not negative.
1108 // e.g. if array_length is 2, ((array_length - 3) >>> 2) is way bigger
1109 // than array_length.
1110 return;
1111 }
1112 ValueRange* range = new (&allocator_) ValueRange(
1113 &allocator_,
1114 ValueBound(nullptr, std::numeric_limits<int32_t>::min()),
1115 ValueBound(left, 0));
1116 AssignRange(instruction->GetBlock(), instruction, range);
1117 }
1118 }
1119
VisitDiv(HDiv * div)1120 void VisitDiv(HDiv* div) override {
1121 FindAndHandlePartialArrayLength(div);
1122 }
1123
VisitShr(HShr * shr)1124 void VisitShr(HShr* shr) override {
1125 FindAndHandlePartialArrayLength(shr);
1126 }
1127
VisitUShr(HUShr * ushr)1128 void VisitUShr(HUShr* ushr) override {
1129 FindAndHandlePartialArrayLength(ushr);
1130 }
1131
VisitAnd(HAnd * instruction)1132 void VisitAnd(HAnd* instruction) override {
1133 if (instruction->GetRight()->IsIntConstant()) {
1134 int32_t constant = instruction->GetRight()->AsIntConstant()->GetValue();
1135 if (constant > 0) {
1136 // constant serves as a mask so any number masked with it
1137 // gets a [0, constant] value range.
1138 ValueRange* range = new (&allocator_) ValueRange(
1139 &allocator_,
1140 ValueBound(nullptr, 0),
1141 ValueBound(nullptr, constant));
1142 AssignRange(instruction->GetBlock(), instruction, range);
1143 }
1144 }
1145 }
1146
VisitRem(HRem * instruction)1147 void VisitRem(HRem* instruction) override {
1148 HInstruction* left = instruction->GetLeft();
1149 HInstruction* right = instruction->GetRight();
1150
1151 // Handle 'i % CONST' format expression in array index, e.g:
1152 // array[i % 20];
1153 if (right->IsIntConstant()) {
1154 int32_t right_const = std::abs(right->AsIntConstant()->GetValue());
1155 if (right_const == 0) {
1156 return;
1157 }
1158 // The sign of divisor CONST doesn't affect the sign final value range.
1159 // For example:
1160 // if (i > 0) {
1161 // array[i % 10]; // index value range [0, 9]
1162 // array[i % -10]; // index value range [0, 9]
1163 // }
1164 ValueRange* right_range = new (&allocator_) ValueRange(
1165 &allocator_,
1166 ValueBound(nullptr, 1 - right_const),
1167 ValueBound(nullptr, right_const - 1));
1168
1169 ValueRange* left_range = LookupValueRange(left, instruction->GetBlock());
1170 if (left_range != nullptr) {
1171 right_range = right_range->Narrow(left_range);
1172 }
1173 AssignRange(instruction->GetBlock(), instruction, right_range);
1174 return;
1175 }
1176
1177 // Handle following pattern:
1178 // i0 NullCheck
1179 // i1 ArrayLength[i0]
1180 // i2 DivByZeroCheck [i1] <-- right
1181 // i3 Rem [i5, i2] <-- we are here.
1182 // i4 BoundsCheck [i3,i1]
1183 if (right->IsDivZeroCheck()) {
1184 // if array_length can pass div-by-zero check,
1185 // array_length must be > 0.
1186 right = right->AsDivZeroCheck()->InputAt(0);
1187 }
1188
1189 // Handle 'i % array.length' format expression in array index, e.g:
1190 // array[(i+7) % array.length];
1191 if (right->IsArrayLength()) {
1192 ValueBound lower = ValueBound::Min(); // ideally, lower should be '1-array_length'.
1193 ValueBound upper = ValueBound(right, -1); // array_length - 1
1194 ValueRange* right_range = new (&allocator_) ValueRange(
1195 &allocator_,
1196 lower,
1197 upper);
1198 ValueRange* left_range = LookupValueRange(left, instruction->GetBlock());
1199 if (left_range != nullptr) {
1200 right_range = right_range->Narrow(left_range);
1201 }
1202 AssignRange(instruction->GetBlock(), instruction, right_range);
1203 return;
1204 }
1205 }
1206
VisitNewArray(HNewArray * new_array)1207 void VisitNewArray(HNewArray* new_array) override {
1208 HInstruction* len = new_array->GetLength();
1209 if (!len->IsIntConstant()) {
1210 HInstruction *left;
1211 int32_t right_const;
1212 if (ValueBound::IsAddOrSubAConstant(len, &left, &right_const)) {
1213 // (left + right_const) is used as size to new the array.
1214 // We record "-right_const <= left <= new_array - right_const";
1215 ValueBound lower = ValueBound(nullptr, -right_const);
1216 // We use new_array for the bound instead of new_array.length,
1217 // which isn't available as an instruction yet. new_array will
1218 // be treated the same as new_array.length when it's used in a ValueBound.
1219 ValueBound upper = ValueBound(new_array, -right_const);
1220 ValueRange* range = new (&allocator_) ValueRange(&allocator_, lower, upper);
1221 ValueRange* existing_range = LookupValueRange(left, new_array->GetBlock());
1222 if (existing_range != nullptr) {
1223 range = existing_range->Narrow(range);
1224 }
1225 AssignRange(new_array->GetBlock(), left, range);
1226 }
1227 }
1228 }
1229
1230 /**
1231 * After null/bounds checks are eliminated, some invariant array references
1232 * may be exposed underneath which can be hoisted out of the loop to the
1233 * preheader or, in combination with dynamic bce, the deoptimization block.
1234 *
1235 * for (int i = 0; i < n; i++) {
1236 * <-------+
1237 * for (int j = 0; j < n; j++) |
1238 * a[i][j] = 0; --a[i]--+
1239 * }
1240 *
1241 * Note: this optimization is no longer applied after dominator-based dynamic deoptimization
1242 * has occurred (see AddCompareWithDeoptimization()), since in those cases it would be
1243 * unsafe to hoist array references across their deoptimization instruction inside a loop.
1244 */
VisitArrayGet(HArrayGet * array_get)1245 void VisitArrayGet(HArrayGet* array_get) override {
1246 if (!has_dom_based_dynamic_bce_ && array_get->IsInLoop()) {
1247 HLoopInformation* loop = array_get->GetBlock()->GetLoopInformation();
1248 if (loop->IsDefinedOutOfTheLoop(array_get->InputAt(0)) &&
1249 loop->IsDefinedOutOfTheLoop(array_get->InputAt(1))) {
1250 SideEffects loop_effects = side_effects_.GetLoopEffects(loop->GetHeader());
1251 if (!array_get->GetSideEffects().MayDependOn(loop_effects)) {
1252 // We can hoist ArrayGet only if its execution is guaranteed on every iteration.
1253 // In other words only if array_get_bb dominates all back branches.
1254 if (loop->DominatesAllBackEdges(array_get->GetBlock())) {
1255 HoistToPreHeaderOrDeoptBlock(loop, array_get);
1256 }
1257 }
1258 }
1259 }
1260 }
1261
1262 /** Performs dominator-based dynamic elimination on suitable set of bounds checks. */
AddCompareWithDeoptimization(HBasicBlock * block,HInstruction * array_length,HInstruction * base,int32_t min_c,int32_t max_c)1263 void AddCompareWithDeoptimization(HBasicBlock* block,
1264 HInstruction* array_length,
1265 HInstruction* base,
1266 int32_t min_c, int32_t max_c) {
1267 HBoundsCheck* bounds_check =
1268 first_index_bounds_check_map_.Get(array_length->GetId())->AsBoundsCheck();
1269 // Construct deoptimization on single or double bounds on range [base-min_c,base+max_c],
1270 // for example either for a[0]..a[3] just 3 or for a[base-1]..a[base+3] both base-1
1271 // and base+3, since we made the assumption any in between value may occur too.
1272 // In code, using unsigned comparisons:
1273 // (1) constants only
1274 // if (max_c >= a.length) deoptimize;
1275 // (2) general case
1276 // if (base-min_c > base+max_c) deoptimize;
1277 // if (base+max_c >= a.length ) deoptimize;
1278 static_assert(kMaxLengthForAddingDeoptimize < std::numeric_limits<int32_t>::max(),
1279 "Incorrect max length may be subject to arithmetic wrap-around");
1280 HInstruction* upper = GetGraph()->GetIntConstant(max_c);
1281 if (base == nullptr) {
1282 DCHECK_GE(min_c, 0);
1283 } else {
1284 HInstruction* lower = new (GetGraph()->GetAllocator())
1285 HAdd(DataType::Type::kInt32, base, GetGraph()->GetIntConstant(min_c));
1286 upper = new (GetGraph()->GetAllocator()) HAdd(DataType::Type::kInt32, base, upper);
1287 block->InsertInstructionBefore(lower, bounds_check);
1288 block->InsertInstructionBefore(upper, bounds_check);
1289 InsertDeoptInBlock(bounds_check, new (GetGraph()->GetAllocator()) HAbove(lower, upper));
1290 }
1291 InsertDeoptInBlock(
1292 bounds_check, new (GetGraph()->GetAllocator()) HAboveOrEqual(upper, array_length));
1293 // Flag that this kind of deoptimization has occurred.
1294 has_dom_based_dynamic_bce_ = true;
1295 }
1296
1297 /** Attempts dominator-based dynamic elimination on remaining candidates. */
AddComparesWithDeoptimization(HBasicBlock * block)1298 void AddComparesWithDeoptimization(HBasicBlock* block) {
1299 for (const auto& entry : first_index_bounds_check_map_) {
1300 HBoundsCheck* bounds_check = entry.second;
1301 HInstruction* index = bounds_check->InputAt(0);
1302 HInstruction* array_length = bounds_check->InputAt(1);
1303 if (!array_length->IsArrayLength()) {
1304 continue; // disregard phis and constants
1305 }
1306 // Collect all bounds checks that are still there and that are related as "a[base + constant]"
1307 // for a base instruction (possibly absent) and various constants. Note that no attempt
1308 // is made to partition the set into matching subsets (viz. a[0], a[1] and a[base+1] and
1309 // a[base+2] are considered as one set).
1310 // TODO: would such a partitioning be worthwhile?
1311 ValueBound value = ValueBound::AsValueBound(index);
1312 HInstruction* base = value.GetInstruction();
1313 int32_t min_c = base == nullptr ? 0 : value.GetConstant();
1314 int32_t max_c = value.GetConstant();
1315 ScopedArenaVector<HBoundsCheck*> candidates(
1316 allocator_.Adapter(kArenaAllocBoundsCheckElimination));
1317 ScopedArenaVector<HBoundsCheck*> standby(
1318 allocator_.Adapter(kArenaAllocBoundsCheckElimination));
1319 for (const HUseListNode<HInstruction*>& use : array_length->GetUses()) {
1320 // Another bounds check in same or dominated block?
1321 HInstruction* user = use.GetUser();
1322 HBasicBlock* other_block = user->GetBlock();
1323 if (user->IsBoundsCheck() && block->Dominates(other_block)) {
1324 HBoundsCheck* other_bounds_check = user->AsBoundsCheck();
1325 HInstruction* other_index = other_bounds_check->InputAt(0);
1326 HInstruction* other_array_length = other_bounds_check->InputAt(1);
1327 ValueBound other_value = ValueBound::AsValueBound(other_index);
1328 if (array_length == other_array_length && base == other_value.GetInstruction()) {
1329 // Reject certain OOB if BoundsCheck(l, l) occurs on considered subset.
1330 if (array_length == other_index) {
1331 candidates.clear();
1332 standby.clear();
1333 break;
1334 }
1335 // Since a subsequent dominated block could be under a conditional, only accept
1336 // the other bounds check if it is in same block or both blocks dominate the exit.
1337 // TODO: we could improve this by testing proper post-dominance, or even if this
1338 // constant is seen along *all* conditional paths that follow.
1339 HBasicBlock* exit = GetGraph()->GetExitBlock();
1340 if (block == user->GetBlock() ||
1341 (block->Dominates(exit) && other_block->Dominates(exit))) {
1342 int32_t other_c = other_value.GetConstant();
1343 min_c = std::min(min_c, other_c);
1344 max_c = std::max(max_c, other_c);
1345 candidates.push_back(other_bounds_check);
1346 } else {
1347 // Add this candidate later only if it falls into the range.
1348 standby.push_back(other_bounds_check);
1349 }
1350 }
1351 }
1352 }
1353 // Add standby candidates that fall in selected range.
1354 for (HBoundsCheck* other_bounds_check : standby) {
1355 HInstruction* other_index = other_bounds_check->InputAt(0);
1356 int32_t other_c = ValueBound::AsValueBound(other_index).GetConstant();
1357 if (min_c <= other_c && other_c <= max_c) {
1358 candidates.push_back(other_bounds_check);
1359 }
1360 }
1361 // Perform dominator-based deoptimization if it seems profitable, where we eliminate
1362 // bounds checks and replace these with deopt checks that guard against any possible
1363 // OOB. Note that we reject cases where the distance min_c:max_c range gets close to
1364 // the maximum possible array length, since those cases are likely to always deopt
1365 // (such situations do not necessarily go OOB, though, since the array could be really
1366 // large, or the programmer could rely on arithmetic wrap-around from max to min).
1367 size_t threshold = kThresholdForAddingDeoptimize + (base == nullptr ? 0 : 1); // extra test?
1368 uint32_t distance = static_cast<uint32_t>(max_c) - static_cast<uint32_t>(min_c);
1369 if (candidates.size() >= threshold &&
1370 (base != nullptr || min_c >= 0) && // reject certain OOB
1371 distance <= kMaxLengthForAddingDeoptimize) { // reject likely/certain deopt
1372 AddCompareWithDeoptimization(block, array_length, base, min_c, max_c);
1373 for (HBoundsCheck* other_bounds_check : candidates) {
1374 // Only replace if still in the graph. This avoids visiting the same
1375 // bounds check twice if it occurred multiple times in the use list.
1376 if (other_bounds_check->IsInBlock()) {
1377 ReplaceInstruction(other_bounds_check, other_bounds_check->InputAt(0));
1378 }
1379 }
1380 }
1381 }
1382 }
1383
1384 /**
1385 * Returns true if static range analysis based on induction variables can determine the bounds
1386 * check on the given array range is always satisfied with the computed index range. The output
1387 * parameter try_dynamic_bce is set to false if OOB is certain.
1388 */
InductionRangeFitsIn(ValueRange * array_range,HBoundsCheck * context,bool * try_dynamic_bce)1389 bool InductionRangeFitsIn(ValueRange* array_range,
1390 HBoundsCheck* context,
1391 bool* try_dynamic_bce) {
1392 InductionVarRange::Value v1;
1393 InductionVarRange::Value v2;
1394 bool needs_finite_test = false;
1395 HInstruction* index = context->InputAt(0);
1396 HInstruction* hint = HuntForDeclaration(context->InputAt(1));
1397 if (induction_range_.GetInductionRange(context, index, hint, &v1, &v2, &needs_finite_test)) {
1398 if (v1.is_known && (v1.a_constant == 0 || v1.a_constant == 1) &&
1399 v2.is_known && (v2.a_constant == 0 || v2.a_constant == 1)) {
1400 DCHECK(v1.a_constant == 1 || v1.instruction == nullptr);
1401 DCHECK(v2.a_constant == 1 || v2.instruction == nullptr);
1402 ValueRange index_range(&allocator_,
1403 ValueBound(v1.instruction, v1.b_constant),
1404 ValueBound(v2.instruction, v2.b_constant));
1405 // If analysis reveals a certain OOB, disable dynamic BCE. Otherwise,
1406 // use analysis for static bce only if loop is finite.
1407 if (index_range.GetLower().LessThan(array_range->GetLower()) ||
1408 index_range.GetUpper().GreaterThan(array_range->GetUpper())) {
1409 *try_dynamic_bce = false;
1410 } else if (!needs_finite_test && index_range.FitsIn(array_range)) {
1411 return true;
1412 }
1413 }
1414 }
1415 return false;
1416 }
1417
1418 /**
1419 * Performs loop-based dynamic elimination on a bounds check. In order to minimize the
1420 * number of eventually generated tests, related bounds checks with tests that can be
1421 * combined with tests for the given bounds check are collected first.
1422 */
TransformLoopForDynamicBCE(HLoopInformation * loop,HBoundsCheck * bounds_check)1423 void TransformLoopForDynamicBCE(HLoopInformation* loop, HBoundsCheck* bounds_check) {
1424 HInstruction* index = bounds_check->InputAt(0);
1425 HInstruction* array_length = bounds_check->InputAt(1);
1426 DCHECK(loop->IsDefinedOutOfTheLoop(array_length)); // pre-checked
1427 DCHECK(loop->DominatesAllBackEdges(bounds_check->GetBlock()));
1428 // Collect all bounds checks in the same loop that are related as "a[base + constant]"
1429 // for a base instruction (possibly absent) and various constants.
1430 ValueBound value = ValueBound::AsValueBound(index);
1431 HInstruction* base = value.GetInstruction();
1432 int32_t min_c = base == nullptr ? 0 : value.GetConstant();
1433 int32_t max_c = value.GetConstant();
1434 ScopedArenaVector<HBoundsCheck*> candidates(
1435 allocator_.Adapter(kArenaAllocBoundsCheckElimination));
1436 ScopedArenaVector<HBoundsCheck*> standby(
1437 allocator_.Adapter(kArenaAllocBoundsCheckElimination));
1438 for (const HUseListNode<HInstruction*>& use : array_length->GetUses()) {
1439 HInstruction* user = use.GetUser();
1440 if (user->IsBoundsCheck() && loop == user->GetBlock()->GetLoopInformation()) {
1441 HBoundsCheck* other_bounds_check = user->AsBoundsCheck();
1442 HInstruction* other_index = other_bounds_check->InputAt(0);
1443 HInstruction* other_array_length = other_bounds_check->InputAt(1);
1444 ValueBound other_value = ValueBound::AsValueBound(other_index);
1445 int32_t other_c = other_value.GetConstant();
1446 if (array_length == other_array_length && base == other_value.GetInstruction()) {
1447 // Ensure every candidate could be picked for code generation.
1448 bool b1 = false, b2 = false;
1449 if (!induction_range_.CanGenerateRange(other_bounds_check, other_index, &b1, &b2)) {
1450 continue;
1451 }
1452 // Does the current basic block dominate all back edges? If not,
1453 // add this candidate later only if it falls into the range.
1454 if (!loop->DominatesAllBackEdges(user->GetBlock())) {
1455 standby.push_back(other_bounds_check);
1456 continue;
1457 }
1458 min_c = std::min(min_c, other_c);
1459 max_c = std::max(max_c, other_c);
1460 candidates.push_back(other_bounds_check);
1461 }
1462 }
1463 }
1464 // Add standby candidates that fall in selected range.
1465 for (HBoundsCheck* other_bounds_check : standby) {
1466 HInstruction* other_index = other_bounds_check->InputAt(0);
1467 int32_t other_c = ValueBound::AsValueBound(other_index).GetConstant();
1468 if (min_c <= other_c && other_c <= max_c) {
1469 candidates.push_back(other_bounds_check);
1470 }
1471 }
1472 // Perform loop-based deoptimization if it seems profitable, where we eliminate bounds
1473 // checks and replace these with deopt checks that guard against any possible OOB.
1474 DCHECK_LT(0u, candidates.size());
1475 uint32_t distance = static_cast<uint32_t>(max_c) - static_cast<uint32_t>(min_c);
1476 if ((base != nullptr || min_c >= 0) && // reject certain OOB
1477 distance <= kMaxLengthForAddingDeoptimize) { // reject likely/certain deopt
1478 HBasicBlock* block = GetPreHeader(loop, bounds_check);
1479 HInstruction* min_lower = nullptr;
1480 HInstruction* min_upper = nullptr;
1481 HInstruction* max_lower = nullptr;
1482 HInstruction* max_upper = nullptr;
1483 // Iterate over all bounds checks.
1484 for (HBoundsCheck* other_bounds_check : candidates) {
1485 // Only handle if still in the graph. This avoids visiting the same
1486 // bounds check twice if it occurred multiple times in the use list.
1487 if (other_bounds_check->IsInBlock()) {
1488 HInstruction* other_index = other_bounds_check->InputAt(0);
1489 int32_t other_c = ValueBound::AsValueBound(other_index).GetConstant();
1490 // Generate code for either the maximum or minimum. Range analysis already was queried
1491 // whether code generation on the original and, thus, related bounds check was possible.
1492 // It handles either loop invariants (lower is not set) or unit strides.
1493 if (other_c == max_c) {
1494 induction_range_.GenerateRange(
1495 other_bounds_check, other_index, GetGraph(), block, &max_lower, &max_upper);
1496 } else if (other_c == min_c && base != nullptr) {
1497 induction_range_.GenerateRange(
1498 other_bounds_check, other_index, GetGraph(), block, &min_lower, &min_upper);
1499 }
1500 ReplaceInstruction(other_bounds_check, other_index);
1501 }
1502 }
1503 // In code, using unsigned comparisons:
1504 // (1) constants only
1505 // if (max_upper >= a.length ) deoptimize;
1506 // (2) two symbolic invariants
1507 // if (min_upper > max_upper) deoptimize; unless min_c == max_c
1508 // if (max_upper >= a.length ) deoptimize;
1509 // (3) general case, unit strides (where lower would exceed upper for arithmetic wrap-around)
1510 // if (min_lower > max_lower) deoptimize; unless min_c == max_c
1511 // if (max_lower > max_upper) deoptimize;
1512 // if (max_upper >= a.length ) deoptimize;
1513 if (base == nullptr) {
1514 // Constants only.
1515 DCHECK_GE(min_c, 0);
1516 DCHECK(min_lower == nullptr && min_upper == nullptr &&
1517 max_lower == nullptr && max_upper != nullptr);
1518 } else if (max_lower == nullptr) {
1519 // Two symbolic invariants.
1520 if (min_c != max_c) {
1521 DCHECK(min_lower == nullptr && min_upper != nullptr &&
1522 max_lower == nullptr && max_upper != nullptr);
1523 InsertDeoptInLoop(
1524 loop, block, new (GetGraph()->GetAllocator()) HAbove(min_upper, max_upper));
1525 } else {
1526 DCHECK(min_lower == nullptr && min_upper == nullptr &&
1527 max_lower == nullptr && max_upper != nullptr);
1528 }
1529 } else {
1530 // General case, unit strides.
1531 if (min_c != max_c) {
1532 DCHECK(min_lower != nullptr && min_upper != nullptr &&
1533 max_lower != nullptr && max_upper != nullptr);
1534 InsertDeoptInLoop(
1535 loop, block, new (GetGraph()->GetAllocator()) HAbove(min_lower, max_lower));
1536 } else {
1537 DCHECK(min_lower == nullptr && min_upper == nullptr &&
1538 max_lower != nullptr && max_upper != nullptr);
1539 }
1540 InsertDeoptInLoop(
1541 loop, block, new (GetGraph()->GetAllocator()) HAbove(max_lower, max_upper));
1542 }
1543 InsertDeoptInLoop(
1544 loop, block, new (GetGraph()->GetAllocator()) HAboveOrEqual(max_upper, array_length));
1545 } else {
1546 // TODO: if rejected, avoid doing this again for subsequent instructions in this set?
1547 }
1548 }
1549
1550 /**
1551 * Returns true if heuristics indicate that dynamic bce may be profitable.
1552 */
DynamicBCESeemsProfitable(HLoopInformation * loop,HBasicBlock * block)1553 bool DynamicBCESeemsProfitable(HLoopInformation* loop, HBasicBlock* block) {
1554 if (loop != nullptr) {
1555 // The loop preheader of an irreducible loop does not dominate all the blocks in
1556 // the loop. We would need to find the common dominator of all blocks in the loop.
1557 if (loop->IsIrreducible()) {
1558 return false;
1559 }
1560 // We should never deoptimize from an osr method, otherwise we might wrongly optimize
1561 // code dominated by the deoptimization.
1562 if (GetGraph()->IsCompilingOsr()) {
1563 return false;
1564 }
1565 // A try boundary preheader is hard to handle.
1566 // TODO: remove this restriction.
1567 if (loop->GetPreHeader()->GetLastInstruction()->IsTryBoundary()) {
1568 return false;
1569 }
1570 // Does loop have early-exits? If so, the full range may not be covered by the loop
1571 // at runtime and testing the range may apply deoptimization unnecessarily.
1572 if (IsEarlyExitLoop(loop)) {
1573 return false;
1574 }
1575 // Does the current basic block dominate all back edges? If not,
1576 // don't apply dynamic bce to something that may not be executed.
1577 return loop->DominatesAllBackEdges(block);
1578 }
1579 return false;
1580 }
1581
1582 /**
1583 * Returns true if the loop has early exits, which implies it may not cover
1584 * the full range computed by range analysis based on induction variables.
1585 */
IsEarlyExitLoop(HLoopInformation * loop)1586 bool IsEarlyExitLoop(HLoopInformation* loop) {
1587 const uint32_t loop_id = loop->GetHeader()->GetBlockId();
1588 // If loop has been analyzed earlier for early-exit, don't repeat the analysis.
1589 auto it = early_exit_loop_.find(loop_id);
1590 if (it != early_exit_loop_.end()) {
1591 return it->second;
1592 }
1593 // First time early-exit analysis for this loop. Since analysis requires scanning
1594 // the full loop-body, results of the analysis is stored for subsequent queries.
1595 HBlocksInLoopReversePostOrderIterator it_loop(*loop);
1596 for (it_loop.Advance(); !it_loop.Done(); it_loop.Advance()) {
1597 for (HBasicBlock* successor : it_loop.Current()->GetSuccessors()) {
1598 if (!loop->Contains(*successor)) {
1599 early_exit_loop_.Put(loop_id, true);
1600 return true;
1601 }
1602 }
1603 }
1604 early_exit_loop_.Put(loop_id, false);
1605 return false;
1606 }
1607
1608 /**
1609 * Returns true if the array length is already loop invariant, or can be made so
1610 * by handling the null check under the hood of the array length operation.
1611 */
CanHandleLength(HLoopInformation * loop,HInstruction * length,bool needs_taken_test)1612 bool CanHandleLength(HLoopInformation* loop, HInstruction* length, bool needs_taken_test) {
1613 if (loop->IsDefinedOutOfTheLoop(length)) {
1614 return true;
1615 } else if (length->IsArrayLength() && length->GetBlock()->GetLoopInformation() == loop) {
1616 if (CanHandleNullCheck(loop, length->InputAt(0), needs_taken_test)) {
1617 HoistToPreHeaderOrDeoptBlock(loop, length);
1618 return true;
1619 }
1620 }
1621 return false;
1622 }
1623
1624 /**
1625 * Returns true if the null check is already loop invariant, or can be made so
1626 * by generating a deoptimization test.
1627 */
CanHandleNullCheck(HLoopInformation * loop,HInstruction * check,bool needs_taken_test)1628 bool CanHandleNullCheck(HLoopInformation* loop, HInstruction* check, bool needs_taken_test) {
1629 if (loop->IsDefinedOutOfTheLoop(check)) {
1630 return true;
1631 } else if (check->IsNullCheck() && check->GetBlock()->GetLoopInformation() == loop) {
1632 HInstruction* array = check->InputAt(0);
1633 if (loop->IsDefinedOutOfTheLoop(array)) {
1634 // Generate: if (array == null) deoptimize;
1635 TransformLoopForDeoptimizationIfNeeded(loop, needs_taken_test);
1636 HBasicBlock* block = GetPreHeader(loop, check);
1637 HInstruction* cond =
1638 new (GetGraph()->GetAllocator()) HEqual(array, GetGraph()->GetNullConstant());
1639 InsertDeoptInLoop(loop, block, cond, /* is_null_check= */ true);
1640 ReplaceInstruction(check, array);
1641 return true;
1642 }
1643 }
1644 return false;
1645 }
1646
1647 /**
1648 * Returns true if compiler can apply dynamic bce to loops that may be infinite
1649 * (e.g. for (int i = 0; i <= U; i++) with U = MAX_INT), which would invalidate
1650 * the range analysis evaluation code by "overshooting" the computed range.
1651 * Since deoptimization would be a bad choice, and there is no other version
1652 * of the loop to use, dynamic bce in such cases is only allowed if other tests
1653 * ensure the loop is finite.
1654 */
CanHandleInfiniteLoop(HLoopInformation * loop,HInstruction * index,bool needs_infinite_test)1655 bool CanHandleInfiniteLoop(HLoopInformation* loop, HInstruction* index, bool needs_infinite_test) {
1656 if (needs_infinite_test) {
1657 // If we already forced the loop to be finite, allow directly.
1658 const uint32_t loop_id = loop->GetHeader()->GetBlockId();
1659 if (finite_loop_.find(loop_id) != finite_loop_.end()) {
1660 return true;
1661 }
1662 // Otherwise, allow dynamic bce if the index (which is necessarily an induction at
1663 // this point) is the direct loop index (viz. a[i]), since then the runtime tests
1664 // ensure upper bound cannot cause an infinite loop.
1665 HInstruction* control = loop->GetHeader()->GetLastInstruction();
1666 if (control->IsIf()) {
1667 HInstruction* if_expr = control->AsIf()->InputAt(0);
1668 if (if_expr->IsCondition()) {
1669 HCondition* condition = if_expr->AsCondition();
1670 if (index == condition->InputAt(0) ||
1671 index == condition->InputAt(1)) {
1672 finite_loop_.insert(loop_id);
1673 return true;
1674 }
1675 }
1676 }
1677 return false;
1678 }
1679 return true;
1680 }
1681
1682 /**
1683 * Returns appropriate preheader for the loop, depending on whether the
1684 * instruction appears in the loop header or proper loop-body.
1685 */
GetPreHeader(HLoopInformation * loop,HInstruction * instruction)1686 HBasicBlock* GetPreHeader(HLoopInformation* loop, HInstruction* instruction) {
1687 // Use preheader unless there is an earlier generated deoptimization block since
1688 // hoisted expressions may depend on and/or used by the deoptimization tests.
1689 HBasicBlock* header = loop->GetHeader();
1690 const uint32_t loop_id = header->GetBlockId();
1691 auto it = taken_test_loop_.find(loop_id);
1692 if (it != taken_test_loop_.end()) {
1693 HBasicBlock* block = it->second;
1694 // If always taken, keep it that way by returning the original preheader,
1695 // which can be found by following the predecessor of the true-block twice.
1696 if (instruction->GetBlock() == header) {
1697 return block->GetSinglePredecessor()->GetSinglePredecessor();
1698 }
1699 return block;
1700 }
1701 return loop->GetPreHeader();
1702 }
1703
1704 /** Inserts a deoptimization test in a loop preheader. */
InsertDeoptInLoop(HLoopInformation * loop,HBasicBlock * block,HInstruction * condition,bool is_null_check=false)1705 void InsertDeoptInLoop(HLoopInformation* loop,
1706 HBasicBlock* block,
1707 HInstruction* condition,
1708 bool is_null_check = false) {
1709 HInstruction* suspend = loop->GetSuspendCheck();
1710 block->InsertInstructionBefore(condition, block->GetLastInstruction());
1711 DeoptimizationKind kind =
1712 is_null_check ? DeoptimizationKind::kLoopNullBCE : DeoptimizationKind::kLoopBoundsBCE;
1713 HDeoptimize* deoptimize = new (GetGraph()->GetAllocator()) HDeoptimize(
1714 GetGraph()->GetAllocator(), condition, kind, suspend->GetDexPc());
1715 block->InsertInstructionBefore(deoptimize, block->GetLastInstruction());
1716 if (suspend->HasEnvironment()) {
1717 deoptimize->CopyEnvironmentFromWithLoopPhiAdjustment(
1718 suspend->GetEnvironment(), loop->GetHeader());
1719 }
1720 }
1721
1722 /** Inserts a deoptimization test right before a bounds check. */
InsertDeoptInBlock(HBoundsCheck * bounds_check,HInstruction * condition)1723 void InsertDeoptInBlock(HBoundsCheck* bounds_check, HInstruction* condition) {
1724 HBasicBlock* block = bounds_check->GetBlock();
1725 block->InsertInstructionBefore(condition, bounds_check);
1726 HDeoptimize* deoptimize = new (GetGraph()->GetAllocator()) HDeoptimize(
1727 GetGraph()->GetAllocator(),
1728 condition,
1729 DeoptimizationKind::kBlockBCE,
1730 bounds_check->GetDexPc());
1731 block->InsertInstructionBefore(deoptimize, bounds_check);
1732 deoptimize->CopyEnvironmentFrom(bounds_check->GetEnvironment());
1733 }
1734
1735 /** Hoists instruction out of the loop to preheader or deoptimization block. */
HoistToPreHeaderOrDeoptBlock(HLoopInformation * loop,HInstruction * instruction)1736 void HoistToPreHeaderOrDeoptBlock(HLoopInformation* loop, HInstruction* instruction) {
1737 HBasicBlock* block = GetPreHeader(loop, instruction);
1738 DCHECK(!instruction->HasEnvironment());
1739 instruction->MoveBefore(block->GetLastInstruction());
1740 }
1741
1742 /**
1743 * Adds a new taken-test structure to a loop if needed and not already done.
1744 * The taken-test protects range analysis evaluation code to avoid any
1745 * deoptimization caused by incorrect trip-count evaluation in non-taken loops.
1746 *
1747 * old_preheader
1748 * |
1749 * if_block <- taken-test protects deoptimization block
1750 * / \
1751 * true_block false_block <- deoptimizations/invariants are placed in true_block
1752 * \ /
1753 * new_preheader <- may require phi nodes to preserve SSA structure
1754 * |
1755 * header
1756 *
1757 * For example, this loop:
1758 *
1759 * for (int i = lower; i < upper; i++) {
1760 * array[i] = 0;
1761 * }
1762 *
1763 * will be transformed to:
1764 *
1765 * if (lower < upper) {
1766 * if (array == null) deoptimize;
1767 * array_length = array.length;
1768 * if (lower > upper) deoptimize; // unsigned
1769 * if (upper >= array_length) deoptimize; // unsigned
1770 * } else {
1771 * array_length = 0;
1772 * }
1773 * for (int i = lower; i < upper; i++) {
1774 * // Loop without null check and bounds check, and any array.length replaced with array_length.
1775 * array[i] = 0;
1776 * }
1777 */
TransformLoopForDeoptimizationIfNeeded(HLoopInformation * loop,bool needs_taken_test)1778 void TransformLoopForDeoptimizationIfNeeded(HLoopInformation* loop, bool needs_taken_test) {
1779 // Not needed (can use preheader) or already done (can reuse)?
1780 const uint32_t loop_id = loop->GetHeader()->GetBlockId();
1781 if (!needs_taken_test || taken_test_loop_.find(loop_id) != taken_test_loop_.end()) {
1782 return;
1783 }
1784
1785 // Generate top test structure.
1786 HBasicBlock* header = loop->GetHeader();
1787 GetGraph()->TransformLoopHeaderForBCE(header);
1788 HBasicBlock* new_preheader = loop->GetPreHeader();
1789 HBasicBlock* if_block = new_preheader->GetDominator();
1790 HBasicBlock* true_block = if_block->GetSuccessors()[0]; // True successor.
1791 HBasicBlock* false_block = if_block->GetSuccessors()[1]; // False successor.
1792
1793 // Goto instructions.
1794 true_block->AddInstruction(new (GetGraph()->GetAllocator()) HGoto());
1795 false_block->AddInstruction(new (GetGraph()->GetAllocator()) HGoto());
1796 new_preheader->AddInstruction(new (GetGraph()->GetAllocator()) HGoto());
1797
1798 // Insert the taken-test to see if the loop body is entered. If the
1799 // loop isn't entered at all, it jumps around the deoptimization block.
1800 if_block->AddInstruction(new (GetGraph()->GetAllocator()) HGoto()); // placeholder
1801 HInstruction* condition = induction_range_.GenerateTakenTest(
1802 header->GetLastInstruction(), GetGraph(), if_block);
1803 DCHECK(condition != nullptr);
1804 if_block->RemoveInstruction(if_block->GetLastInstruction());
1805 if_block->AddInstruction(new (GetGraph()->GetAllocator()) HIf(condition));
1806
1807 taken_test_loop_.Put(loop_id, true_block);
1808 }
1809
1810 /**
1811 * Inserts phi nodes that preserve SSA structure in generated top test structures.
1812 * All uses of instructions in the deoptimization block that reach the loop need
1813 * a phi node in the new loop preheader to fix the dominance relation.
1814 *
1815 * Example:
1816 * if_block
1817 * / \
1818 * x_0 = .. false_block
1819 * \ /
1820 * x_1 = phi(x_0, null) <- synthetic phi
1821 * |
1822 * new_preheader
1823 */
InsertPhiNodes()1824 void InsertPhiNodes() {
1825 // Scan all new deoptimization blocks.
1826 for (const auto& entry : taken_test_loop_) {
1827 HBasicBlock* true_block = entry.second;
1828 HBasicBlock* new_preheader = true_block->GetSingleSuccessor();
1829 // Scan all instructions in a new deoptimization block.
1830 for (HInstructionIterator it(true_block->GetInstructions()); !it.Done(); it.Advance()) {
1831 HInstruction* instruction = it.Current();
1832 DataType::Type type = instruction->GetType();
1833 HPhi* phi = nullptr;
1834 // Scan all uses of an instruction and replace each later use with a phi node.
1835 const HUseList<HInstruction*>& uses = instruction->GetUses();
1836 for (auto it2 = uses.begin(), end2 = uses.end(); it2 != end2; /* ++it2 below */) {
1837 HInstruction* user = it2->GetUser();
1838 size_t index = it2->GetIndex();
1839 // Increment `it2` now because `*it2` may disappear thanks to user->ReplaceInput().
1840 ++it2;
1841 if (user->GetBlock() != true_block) {
1842 if (phi == nullptr) {
1843 phi = NewPhi(new_preheader, instruction, type);
1844 }
1845 user->ReplaceInput(phi, index); // Removes the use node from the list.
1846 induction_range_.Replace(user, instruction, phi); // update induction
1847 }
1848 }
1849 // Scan all environment uses of an instruction and replace each later use with a phi node.
1850 const HUseList<HEnvironment*>& env_uses = instruction->GetEnvUses();
1851 for (auto it2 = env_uses.begin(), end2 = env_uses.end(); it2 != end2; /* ++it2 below */) {
1852 HEnvironment* user = it2->GetUser();
1853 size_t index = it2->GetIndex();
1854 // Increment `it2` now because `*it2` may disappear thanks to user->RemoveAsUserOfInput().
1855 ++it2;
1856 if (user->GetHolder()->GetBlock() != true_block) {
1857 if (phi == nullptr) {
1858 phi = NewPhi(new_preheader, instruction, type);
1859 }
1860 user->RemoveAsUserOfInput(index);
1861 user->SetRawEnvAt(index, phi);
1862 phi->AddEnvUseAt(user, index);
1863 }
1864 }
1865 }
1866 }
1867 }
1868
1869 /**
1870 * Construct a phi(instruction, 0) in the new preheader to fix the dominance relation.
1871 * These are synthetic phi nodes without a virtual register.
1872 */
NewPhi(HBasicBlock * new_preheader,HInstruction * instruction,DataType::Type type)1873 HPhi* NewPhi(HBasicBlock* new_preheader,
1874 HInstruction* instruction,
1875 DataType::Type type) {
1876 HGraph* graph = GetGraph();
1877 HInstruction* zero;
1878 switch (type) {
1879 case DataType::Type::kReference: zero = graph->GetNullConstant(); break;
1880 case DataType::Type::kFloat32: zero = graph->GetFloatConstant(0); break;
1881 case DataType::Type::kFloat64: zero = graph->GetDoubleConstant(0); break;
1882 default: zero = graph->GetConstant(type, 0); break;
1883 }
1884 HPhi* phi = new (graph->GetAllocator())
1885 HPhi(graph->GetAllocator(), kNoRegNumber, /*number_of_inputs*/ 2, HPhi::ToPhiType(type));
1886 phi->SetRawInputAt(0, instruction);
1887 phi->SetRawInputAt(1, zero);
1888 if (type == DataType::Type::kReference) {
1889 phi->SetReferenceTypeInfo(instruction->GetReferenceTypeInfo());
1890 }
1891 new_preheader->AddPhi(phi);
1892 return phi;
1893 }
1894
1895 /** Helper method to replace an instruction with another instruction. */
ReplaceInstruction(HInstruction * instruction,HInstruction * replacement)1896 void ReplaceInstruction(HInstruction* instruction, HInstruction* replacement) {
1897 // Safe iteration.
1898 if (instruction == next_) {
1899 next_ = next_->GetNext();
1900 }
1901 // Replace and remove.
1902 instruction->ReplaceWith(replacement);
1903 instruction->GetBlock()->RemoveInstruction(instruction);
1904 }
1905
1906 // Use local allocator for allocating memory.
1907 ScopedArenaAllocator allocator_;
1908
1909 // A set of maps, one per basic block, from instruction to range.
1910 ScopedArenaVector<ScopedArenaSafeMap<int, ValueRange*>> maps_;
1911
1912 // Map an HArrayLength instruction's id to the first HBoundsCheck instruction
1913 // in a block that checks an index against that HArrayLength.
1914 ScopedArenaSafeMap<int, HBoundsCheck*> first_index_bounds_check_map_;
1915
1916 // Early-exit loop bookkeeping.
1917 ScopedArenaSafeMap<uint32_t, bool> early_exit_loop_;
1918
1919 // Taken-test loop bookkeeping.
1920 ScopedArenaSafeMap<uint32_t, HBasicBlock*> taken_test_loop_;
1921
1922 // Finite loop bookkeeping.
1923 ScopedArenaSet<uint32_t> finite_loop_;
1924
1925 // Flag that denotes whether dominator-based dynamic elimination has occurred.
1926 bool has_dom_based_dynamic_bce_;
1927
1928 // Initial number of blocks.
1929 uint32_t initial_block_size_;
1930
1931 // Side effects.
1932 const SideEffectsAnalysis& side_effects_;
1933
1934 // Range analysis based on induction variables.
1935 InductionVarRange induction_range_;
1936
1937 // Safe iteration.
1938 HInstruction* next_;
1939
1940 DISALLOW_COPY_AND_ASSIGN(BCEVisitor);
1941 };
1942
Run()1943 bool BoundsCheckElimination::Run() {
1944 if (!graph_->HasBoundsChecks()) {
1945 return false;
1946 }
1947
1948 // Reverse post order guarantees a node's dominators are visited first.
1949 // We want to visit in the dominator-based order since if a value is known to
1950 // be bounded by a range at one instruction, it must be true that all uses of
1951 // that value dominated by that instruction fits in that range. Range of that
1952 // value can be narrowed further down in the dominator tree.
1953 BCEVisitor visitor(graph_, side_effects_, induction_analysis_);
1954 for (size_t i = 0, size = graph_->GetReversePostOrder().size(); i != size; ++i) {
1955 HBasicBlock* current = graph_->GetReversePostOrder()[i];
1956 if (visitor.IsAddedBlock(current)) {
1957 // Skip added blocks. Their effects are already taken care of.
1958 continue;
1959 }
1960 visitor.VisitBasicBlock(current);
1961 // Skip forward to the current block in case new basic blocks were inserted
1962 // (which always appear earlier in reverse post order) to avoid visiting the
1963 // same basic block twice.
1964 size_t new_size = graph_->GetReversePostOrder().size();
1965 DCHECK_GE(new_size, size);
1966 i += new_size - size;
1967 DCHECK_EQ(current, graph_->GetReversePostOrder()[i]);
1968 size = new_size;
1969 }
1970
1971 // Perform cleanup.
1972 visitor.Finish();
1973
1974 return true;
1975 }
1976
1977 } // namespace art
1978