1 /*
2 * Copyright (C) 2015 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 "load_store_elimination.h"
18
19 #include <algorithm>
20 #include <optional>
21 #include <sstream>
22 #include <variant>
23
24 #include "base/arena_allocator.h"
25 #include "base/arena_bit_vector.h"
26 #include "base/array_ref.h"
27 #include "base/bit_vector-inl.h"
28 #include "base/bit_vector.h"
29 #include "base/globals.h"
30 #include "base/indenter.h"
31 #include "base/iteration_range.h"
32 #include "base/scoped_arena_allocator.h"
33 #include "base/scoped_arena_containers.h"
34 #include "base/transform_iterator.h"
35 #include "escape.h"
36 #include "execution_subgraph.h"
37 #include "handle.h"
38 #include "load_store_analysis.h"
39 #include "mirror/class_loader.h"
40 #include "mirror/dex_cache.h"
41 #include "nodes.h"
42 #include "optimizing/execution_subgraph.h"
43 #include "optimizing_compiler_stats.h"
44 #include "reference_type_propagation.h"
45 #include "side_effects_analysis.h"
46 #include "stack_map.h"
47
48 /**
49 * The general algorithm of load-store elimination (LSE).
50 *
51 * We use load-store analysis to collect a list of heap locations and perform
52 * alias analysis of those heap locations. LSE then keeps track of a list of
53 * heap values corresponding to the heap locations and stores that put those
54 * values in these locations.
55 * - In phase 1, we visit basic blocks in reverse post order and for each basic
56 * block, visit instructions sequentially, recording heap values and looking
57 * for loads and stores to eliminate without relying on loop Phis.
58 * - In phase 2, we look for loads that can be replaced by creating loop Phis
59 * or using a loop-invariant value.
60 * - In phase 3, we determine which stores are dead and can be eliminated and
61 * based on that information we re-evaluate whether some kept stores are
62 * storing the same value as the value in the heap location; such stores are
63 * also marked for elimination.
64 * - In phase 4, we commit the changes, replacing loads marked for elimination
65 * in previous processing and removing stores not marked for keeping. We also
66 * remove allocations that are no longer needed.
67 * - In phase 5, we move allocations which only escape along some executions
68 * closer to their escape points and fixup non-escaping paths with their actual
69 * values, creating PHIs when needed.
70 *
71 * 1. Walk over blocks and their instructions.
72 *
73 * The initial set of heap values for a basic block is
74 * - For a loop header of an irreducible loop, all heap values are unknown.
75 * - For a loop header of a normal loop, all values unknown at the end of the
76 * preheader are initialized to unknown, other heap values are set to Phi
77 * placeholders as we cannot determine yet whether these values are known on
78 * all back-edges. We use Phi placeholders also for array heap locations with
79 * index defined inside the loop but this helps only when the value remains
80 * zero from the array allocation throughout the loop.
81 * - For other basic blocks, we merge incoming values from the end of all
82 * predecessors. If any incoming value is unknown, the start value for this
83 * block is also unknown. Otherwise, if all the incoming values are the same
84 * (including the case of a single predecessor), the incoming value is used.
85 * Otherwise, we use a Phi placeholder to indicate different incoming values.
86 * We record whether such Phi placeholder depends on a loop Phi placeholder.
87 *
88 * For each instruction in the block
89 * - If the instruction is a load from a heap location with a known value not
90 * dependent on a loop Phi placeholder, the load can be eliminated, either by
91 * using an existing instruction or by creating new Phi(s) instead. In order
92 * to maintain the validity of all heap locations during the optimization
93 * phase, we only record substitutes at this phase and the real elimination
94 * is delayed till the end of LSE. Loads that require a loop Phi placeholder
95 * replacement are recorded for processing later. We also keep track of the
96 * heap-value at the start load so that later partial-LSE can predicate the
97 * load.
98 * - If the instruction is a store, it updates the heap value for the heap
99 * location with the stored value and records the store itself so that we can
100 * mark it for keeping if the value becomes observable. Heap values are
101 * invalidated for heap locations that may alias with the store instruction's
102 * heap location and their recorded stores are marked for keeping as they are
103 * now potentially observable. The store instruction can be eliminated unless
104 * the value stored is later needed e.g. by a load from the same/aliased heap
105 * location or the heap location persists at method return/deoptimization.
106 * - A store that stores the same value as the heap value is eliminated.
107 * - For newly instantiated instances, their heap values are initialized to
108 * language defined default values.
109 * - Finalizable objects are considered as persisting at method
110 * return/deoptimization.
111 * - Some instructions such as invokes are treated as loading and invalidating
112 * all the heap values, depending on the instruction's side effects.
113 * - SIMD graphs (with VecLoad and VecStore instructions) are also handled. Any
114 * partial overlap access among ArrayGet/ArraySet/VecLoad/Store is seen as
115 * alias and no load/store is eliminated in such case.
116 * - Currently this LSE algorithm doesn't handle graph with try-catch, due to
117 * the special block merging structure.
118 *
119 * The time complexity of the initial phase has several components. The total
120 * time for the initialization of heap values for all blocks is
121 * O(heap_locations * edges)
122 * and the time complexity for simple instruction processing is
123 * O(instructions).
124 * See the description of phase 3 for additional complexity due to matching of
125 * existing Phis for replacing loads.
126 *
127 * 2. Process loads that depend on loop Phi placeholders.
128 *
129 * We go over these loads to determine whether they can be eliminated. We look
130 * for the set of all Phi placeholders that feed the load and depend on a loop
131 * Phi placeholder and, if we find no unknown value, we construct the necessary
132 * Phi(s) or, if all other inputs are identical, i.e. the location does not
133 * change in the loop, just use that input. If we do find an unknown input, this
134 * must be from a loop back-edge and we replace the loop Phi placeholder with
135 * unknown value and re-process loads and stores that previously depended on
136 * loop Phi placeholders. This shall find at least one load of an unknown value
137 * which is now known to be unreplaceable or a new unknown value on a back-edge
138 * and we repeat this process until each load is either marked for replacement
139 * or found to be unreplaceable. As we mark at least one additional loop Phi
140 * placeholder as unreplacable in each iteration, this process shall terminate.
141 *
142 * The depth-first search for Phi placeholders in FindLoopPhisToMaterialize()
143 * is limited by the number of Phi placeholders and their dependencies we need
144 * to search with worst-case time complexity
145 * O(phi_placeholder_dependencies) .
146 * The dependencies are usually just the Phi placeholders' potential inputs,
147 * but if we use TryReplacingLoopPhiPlaceholderWithDefault() for default value
148 * replacement search, there are additional dependencies to consider, see below.
149 *
150 * In the successful case (no unknown inputs found) we use the Floyd-Warshall
151 * algorithm to determine transitive closures for each found Phi placeholder,
152 * and then match or materialize Phis from the smallest transitive closure,
153 * so that we can determine if such subset has a single other input. This has
154 * time complexity
155 * O(phi_placeholders_found^3) .
156 * Note that successful TryReplacingLoopPhiPlaceholderWithDefault() does not
157 * contribute to this as such Phi placeholders are replaced immediately.
158 * The total time of all such successful cases has time complexity
159 * O(phi_placeholders^3)
160 * because the found sets are disjoint and `Sum(n_i^3) <= Sum(n_i)^3`. Similar
161 * argument applies to the searches used to find all successful cases, so their
162 * total contribution is also just an insignificant
163 * O(phi_placeholder_dependencies) .
164 * The materialization of Phis has an insignificant total time complexity
165 * O(phi_placeholders * edges) .
166 *
167 * If we find an unknown input, we re-process heap values and loads with a time
168 * complexity that's the same as the phase 1 in the worst case. Adding this to
169 * the depth-first search time complexity yields
170 * O(phi_placeholder_dependencies + heap_locations * edges + instructions)
171 * for a single iteration. We can ignore the middle term as it's proprotional
172 * to the number of Phi placeholder inputs included in the first term. Using
173 * the upper limit of number of such iterations, the total time complexity is
174 * O((phi_placeholder_dependencies + instructions) * phi_placeholders) .
175 *
176 * The upper bound of Phi placeholder inputs is
177 * heap_locations * edges
178 * but if we use TryReplacingLoopPhiPlaceholderWithDefault(), the dependencies
179 * include other heap locations in predecessor blocks with the upper bound of
180 * heap_locations^2 * edges .
181 * Using the estimate
182 * edges <= blocks^2
183 * and
184 * phi_placeholders <= heap_locations * blocks ,
185 * the worst-case time complexity of the
186 * O(phi_placeholder_dependencies * phi_placeholders)
187 * term from unknown input cases is actually
188 * O(heap_locations^3 * blocks^3) ,
189 * exactly as the estimate for the Floyd-Warshall parts of successful cases.
190 * Adding the other term from the unknown input cases (to account for the case
191 * with significantly more instructions than blocks and heap locations), the
192 * phase 2 time complexity is
193 * O(heap_locations^3 * blocks^3 + heap_locations * blocks * instructions) .
194 *
195 * See the description of phase 3 for additional complexity due to matching of
196 * existing Phis for replacing loads.
197 *
198 * 3. Determine which stores to keep and which to eliminate.
199 *
200 * During instruction processing in phase 1 and re-processing in phase 2, we are
201 * keeping a record of the stores and Phi placeholders that become observable
202 * and now propagate the observable Phi placeholders to all actual stores that
203 * feed them. Having determined observable stores, we look for stores that just
204 * overwrite the old value with the same. Since ignoring non-observable stores
205 * actually changes the old values in heap locations, we need to recalculate
206 * Phi placeholder replacements but we proceed similarly to the previous phase.
207 * We look for the set of all Phis that feed the old value replaced by the store
208 * (but ignoring whether they depend on a loop Phi) and, if we find no unknown
209 * value, we try to match existing Phis (we do not create new Phis anymore) or,
210 * if all other inputs are identical, i.e. the location does not change in the
211 * loop, just use that input. If this succeeds and the old value is identical to
212 * the value we're storing, such store shall be eliminated.
213 *
214 * The work is similar to the phase 2, except that we're not re-processing loads
215 * and stores anymore, so the time complexity of phase 3 is
216 * O(heap_locations^3 * blocks^3) .
217 *
218 * There is additional complexity in matching existing Phis shared between the
219 * phases 1, 2 and 3. We are never trying to match two or more Phis at the same
220 * time (this could be difficult and slow), so each matching attempt is just
221 * looking at Phis in the block (both old Phis and newly created Phis) and their
222 * inputs. As we create at most `heap_locations` Phis in each block, the upper
223 * bound on the number of Phis we look at is
224 * heap_locations * (old_phis + heap_locations)
225 * and the worst-case time complexity is
226 * O(heap_locations^2 * edges + heap_locations * old_phis * edges) .
227 * The first term is lower than one term in phase 2, so the relevant part is
228 * O(heap_locations * old_phis * edges) .
229 *
230 * 4. Replace loads and remove unnecessary stores and singleton allocations.
231 *
232 * A special type of objects called singletons are instantiated in the method
233 * and have a single name, i.e. no aliases. Singletons have exclusive heap
234 * locations since they have no aliases. Singletons are helpful in narrowing
235 * down the life span of a heap location such that they do not always need to
236 * participate in merging heap values. Allocation of a singleton can be
237 * eliminated if that singleton is not used and does not persist at method
238 * return/deoptimization.
239 *
240 * The time complexity of this phase is
241 * O(instructions + instruction_uses) .
242 *
243 * 5. Partial LSE
244 *
245 * Move allocations closer to their escapes and remove/predicate loads and
246 * stores as required.
247 *
248 * Partial singletons are objects which only escape from the function or have
249 * multiple names along certain execution paths. In cases where we recognize
250 * these partial singletons we can move the allocation and initialization
251 * closer to the actual escape(s). We can then perform a simplified version of
252 * LSE step 2 to determine the unescaped value of any reads performed after the
253 * object may have escaped. These are used to replace these reads with
254 * 'predicated-read' instructions where the value is only read if the object
255 * has actually escaped. We use the existence of the object itself as the
256 * marker of whether escape has occurred.
257 *
258 * There are several steps in this sub-pass
259 *
260 * 5.1 Group references
261 *
262 * Since all heap-locations for a single reference escape at the same time, we
263 * need to group the heap-locations by reference and process them at the same
264 * time.
265 *
266 * O(heap_locations).
267 *
268 * FIXME: The time complexity above assumes we can bucket the heap-locations in
269 * O(1) which is not true since we just perform a linear-scan of the heap-ref
270 * list. Since there are generally only a small number of heap-references which
271 * are partial-singletons this is fine and lower real overhead than a hash map.
272 *
273 * 5.2 Generate materializations
274 *
275 * Once we have the references we add new 'materialization blocks' on the edges
276 * where escape becomes inevitable. This information is calculated by the
277 * execution-subgraphs created during load-store-analysis. We create new
278 * 'materialization's in these blocks and initialize them with the value of
279 * each heap-location ignoring side effects (since the object hasn't escaped
280 * yet). Worst case this is the same time-complexity as step 3 since we may
281 * need to materialize phis.
282 *
283 * O(heap_locations^2 * materialization_edges)
284 *
285 * 5.3 Propagate materializations
286 *
287 * Since we use the materialization as the marker for escape we need to
288 * propagate it throughout the graph. Since the subgraph analysis considers any
289 * lifetime that escapes a loop (and hence would require a loop-phi) to be
290 * escaping at the loop-header we do not need to create any loop-phis to do
291 * this.
292 *
293 * O(edges)
294 *
295 * NB: Currently the subgraph analysis considers all objects to have their
296 * lifetimes start at the entry block. This simplifies that analysis enormously
297 * but means that we cannot distinguish between an escape in a loop where the
298 * lifetime does not escape the loop (in which case this pass could optimize)
299 * and one where it does escape the loop (in which case the whole loop is
300 * escaping). This is a shortcoming that would be good to fix at some point.
301 *
302 * 5.4 Propagate partial values
303 *
304 * We need to replace loads and stores to the partial reference with predicated
305 * ones that have default non-escaping values. Again this is the same as step 3.
306 *
307 * O(heap_locations^2 * edges)
308 *
309 * 5.5 Final fixup
310 *
311 * Now all we need to do is replace and remove uses of the old reference with the
312 * appropriate materialization.
313 *
314 * O(instructions + uses)
315 *
316 * FIXME: The time complexities described above assumes that the
317 * HeapLocationCollector finds a heap location for an instruction in O(1)
318 * time but it is currently O(heap_locations); this can be fixed by adding
319 * a hash map to the HeapLocationCollector.
320 */
321
322 namespace art {
323
324 #define LSE_VLOG \
325 if (::art::LoadStoreElimination::kVerboseLoggingMode && VLOG_IS_ON(compiler)) LOG(INFO)
326
327 class PartialLoadStoreEliminationHelper;
328 class HeapRefHolder;
329
330 // Use HGraphDelegateVisitor for which all VisitInvokeXXX() delegate to VisitInvoke().
331 class LSEVisitor final : private HGraphDelegateVisitor {
332 public:
333 LSEVisitor(HGraph* graph,
334 const HeapLocationCollector& heap_location_collector,
335 bool perform_partial_lse,
336 OptimizingCompilerStats* stats);
337
338 void Run();
339
340 private:
341 class PhiPlaceholder {
342 public:
PhiPlaceholder()343 constexpr PhiPlaceholder() : block_id_(-1), heap_location_(-1) {}
PhiPlaceholder(uint32_t block_id,size_t heap_location)344 constexpr PhiPlaceholder(uint32_t block_id, size_t heap_location)
345 : block_id_(block_id), heap_location_(dchecked_integral_cast<uint32_t>(heap_location)) {}
346
347 constexpr PhiPlaceholder(const PhiPlaceholder& p) = default;
348 constexpr PhiPlaceholder(PhiPlaceholder&& p) = default;
349 constexpr PhiPlaceholder& operator=(const PhiPlaceholder& p) = default;
350 constexpr PhiPlaceholder& operator=(PhiPlaceholder&& p) = default;
351
GetBlockId() const352 constexpr uint32_t GetBlockId() const {
353 return block_id_;
354 }
355
GetHeapLocation() const356 constexpr size_t GetHeapLocation() const {
357 return heap_location_;
358 }
359
Equals(const PhiPlaceholder & p2) const360 constexpr bool Equals(const PhiPlaceholder& p2) const {
361 return block_id_ == p2.block_id_ && heap_location_ == p2.heap_location_;
362 }
363
Dump(std::ostream & oss) const364 void Dump(std::ostream& oss) const {
365 oss << "PhiPlaceholder[blk: " << block_id_ << ", heap_location_: " << heap_location_ << "]";
366 }
367
368 private:
369 uint32_t block_id_;
370 uint32_t heap_location_;
371 };
372
373 struct Marker {};
374
375 class Value;
376
377 class PriorValueHolder {
378 public:
379 constexpr explicit PriorValueHolder(Value prior);
380
IsInstruction() const381 constexpr bool IsInstruction() const {
382 return std::holds_alternative<HInstruction*>(value_);
383 }
IsPhi() const384 constexpr bool IsPhi() const {
385 return std::holds_alternative<PhiPlaceholder>(value_);
386 }
IsDefault() const387 constexpr bool IsDefault() const {
388 return std::holds_alternative<Marker>(value_);
389 }
GetPhiPlaceholder() const390 constexpr PhiPlaceholder GetPhiPlaceholder() const {
391 DCHECK(IsPhi());
392 return std::get<PhiPlaceholder>(value_);
393 }
GetInstruction() const394 constexpr HInstruction* GetInstruction() const {
395 DCHECK(IsInstruction());
396 return std::get<HInstruction*>(value_);
397 }
398
399 Value ToValue() const;
400 void Dump(std::ostream& oss) const;
401
Equals(PriorValueHolder other) const402 constexpr bool Equals(PriorValueHolder other) const {
403 return value_ == other.value_;
404 }
405
406 private:
407 std::variant<Marker, HInstruction*, PhiPlaceholder> value_;
408 };
409
410 friend constexpr bool operator==(const Marker&, const Marker&);
411 friend constexpr bool operator==(const PriorValueHolder& p1, const PriorValueHolder& p2);
412 friend constexpr bool operator==(const PhiPlaceholder& p1, const PhiPlaceholder& p2);
413 friend std::ostream& operator<<(std::ostream& oss, const PhiPlaceholder& p2);
414
415 class Value {
416 public:
417 enum class ValuelessType {
418 kInvalid,
419 kPureUnknown,
420 kDefault,
421 };
422 struct MergedUnknownMarker {
423 PhiPlaceholder phi_;
424 };
425 struct NeedsNonLoopPhiMarker {
426 PhiPlaceholder phi_;
427 };
428 struct NeedsLoopPhiMarker {
429 PhiPlaceholder phi_;
430 };
431
Invalid()432 static constexpr Value Invalid() {
433 return Value(ValuelessType::kInvalid);
434 }
435
436 // An unknown heap value. Loads with such a value in the heap location cannot be eliminated.
437 // A heap location can be set to an unknown heap value when:
438 // - it is coming from outside the method,
439 // - it is killed due to aliasing, or side effects, or merging with an unknown value.
PureUnknown()440 static constexpr Value PureUnknown() {
441 return Value(ValuelessType::kPureUnknown);
442 }
443
PartialUnknown(Value old_value)444 static constexpr Value PartialUnknown(Value old_value) {
445 if (old_value.IsInvalid() || old_value.IsPureUnknown()) {
446 return PureUnknown();
447 } else {
448 return Value(PriorValueHolder(old_value));
449 }
450 }
451
MergedUnknown(PhiPlaceholder phi_placeholder)452 static constexpr Value MergedUnknown(PhiPlaceholder phi_placeholder) {
453 return Value(MergedUnknownMarker{phi_placeholder});
454 }
455
456 // Default heap value after an allocation.
457 // A heap location can be set to that value right after an allocation.
Default()458 static constexpr Value Default() {
459 return Value(ValuelessType::kDefault);
460 }
461
ForInstruction(HInstruction * instruction)462 static constexpr Value ForInstruction(HInstruction* instruction) {
463 return Value(instruction);
464 }
465
ForNonLoopPhiPlaceholder(PhiPlaceholder phi_placeholder)466 static constexpr Value ForNonLoopPhiPlaceholder(PhiPlaceholder phi_placeholder) {
467 return Value(NeedsNonLoopPhiMarker{phi_placeholder});
468 }
469
ForLoopPhiPlaceholder(PhiPlaceholder phi_placeholder)470 static constexpr Value ForLoopPhiPlaceholder(PhiPlaceholder phi_placeholder) {
471 return Value(NeedsLoopPhiMarker{phi_placeholder});
472 }
473
ForPhiPlaceholder(PhiPlaceholder phi_placeholder,bool needs_loop_phi)474 static constexpr Value ForPhiPlaceholder(PhiPlaceholder phi_placeholder, bool needs_loop_phi) {
475 return needs_loop_phi ? ForLoopPhiPlaceholder(phi_placeholder)
476 : ForNonLoopPhiPlaceholder(phi_placeholder);
477 }
478
IsValid() const479 constexpr bool IsValid() const {
480 return !IsInvalid();
481 }
482
IsInvalid() const483 constexpr bool IsInvalid() const {
484 return std::holds_alternative<ValuelessType>(value_) &&
485 GetValuelessType() == ValuelessType::kInvalid;
486 }
487
IsPartialUnknown() const488 bool IsPartialUnknown() const {
489 return std::holds_alternative<PriorValueHolder>(value_);
490 }
491
IsMergedUnknown() const492 bool IsMergedUnknown() const {
493 return std::holds_alternative<MergedUnknownMarker>(value_);
494 }
495
IsPureUnknown() const496 bool IsPureUnknown() const {
497 return std::holds_alternative<ValuelessType>(value_) &&
498 GetValuelessType() == ValuelessType::kPureUnknown;
499 }
500
IsUnknown() const501 bool IsUnknown() const {
502 return IsPureUnknown() || IsMergedUnknown() || IsPartialUnknown();
503 }
504
IsDefault() const505 bool IsDefault() const {
506 return std::holds_alternative<ValuelessType>(value_) &&
507 GetValuelessType() == ValuelessType::kDefault;
508 }
509
IsInstruction() const510 bool IsInstruction() const {
511 return std::holds_alternative<HInstruction*>(value_);
512 }
513
NeedsNonLoopPhi() const514 bool NeedsNonLoopPhi() const {
515 return std::holds_alternative<NeedsNonLoopPhiMarker>(value_);
516 }
517
NeedsLoopPhi() const518 bool NeedsLoopPhi() const {
519 return std::holds_alternative<NeedsLoopPhiMarker>(value_);
520 }
521
NeedsPhi() const522 bool NeedsPhi() const {
523 return NeedsNonLoopPhi() || NeedsLoopPhi();
524 }
525
GetInstruction() const526 HInstruction* GetInstruction() const {
527 DCHECK(IsInstruction()) << *this;
528 return std::get<HInstruction*>(value_);
529 }
530
GetPriorValue() const531 PriorValueHolder GetPriorValue() const {
532 DCHECK(IsPartialUnknown());
533 return std::get<PriorValueHolder>(value_);
534 }
535
GetPhiPlaceholder() const536 PhiPlaceholder GetPhiPlaceholder() const {
537 DCHECK(NeedsPhi() || IsMergedUnknown());
538 if (NeedsNonLoopPhi()) {
539 return std::get<NeedsNonLoopPhiMarker>(value_).phi_;
540 } else if (NeedsLoopPhi()) {
541 return std::get<NeedsLoopPhiMarker>(value_).phi_;
542 } else {
543 return std::get<MergedUnknownMarker>(value_).phi_;
544 }
545 }
546
GetMergeBlockId() const547 uint32_t GetMergeBlockId() const {
548 DCHECK(IsMergedUnknown()) << this;
549 return std::get<MergedUnknownMarker>(value_).phi_.GetBlockId();
550 }
551
GetMergeBlock(const HGraph * graph) const552 HBasicBlock* GetMergeBlock(const HGraph* graph) const {
553 DCHECK(IsMergedUnknown()) << *this;
554 return graph->GetBlocks()[GetMergeBlockId()];
555 }
556
GetHeapLocation() const557 size_t GetHeapLocation() const {
558 DCHECK(IsMergedUnknown() || NeedsPhi()) << this;
559 return GetPhiPlaceholder().GetHeapLocation();
560 }
561
562 constexpr bool ExactEquals(Value other) const;
563
564 constexpr bool Equals(Value other) const;
565
Equals(HInstruction * instruction) const566 constexpr bool Equals(HInstruction* instruction) const {
567 return Equals(ForInstruction(instruction));
568 }
569
570 std::ostream& Dump(std::ostream& os) const;
571
572 // Public for use with lists.
Value()573 constexpr Value() : value_(ValuelessType::kInvalid) {}
574
575 private:
576 using ValueHolder = std::variant<ValuelessType,
577 HInstruction*,
578 MergedUnknownMarker,
579 NeedsNonLoopPhiMarker,
580 NeedsLoopPhiMarker,
581 PriorValueHolder>;
GetValuelessType() const582 constexpr ValuelessType GetValuelessType() const {
583 return std::get<ValuelessType>(value_);
584 }
585
Value(ValueHolder v)586 constexpr explicit Value(ValueHolder v) : value_(v) {}
587
588 friend std::ostream& operator<<(std::ostream& os, const Value& v);
589
590 ValueHolder value_;
591
592 static_assert(std::is_move_assignable<PhiPlaceholder>::value);
593 };
594
595 friend constexpr bool operator==(const Value::NeedsLoopPhiMarker& p1,
596 const Value::NeedsLoopPhiMarker& p2);
597 friend constexpr bool operator==(const Value::NeedsNonLoopPhiMarker& p1,
598 const Value::NeedsNonLoopPhiMarker& p2);
599 friend constexpr bool operator==(const Value::MergedUnknownMarker& p1,
600 const Value::MergedUnknownMarker& p2);
601
602 // Get Phi placeholder index for access to `phi_placeholder_replacements_`
603 // and "visited" bit vectors during depth-first searches.
PhiPlaceholderIndex(PhiPlaceholder phi_placeholder) const604 size_t PhiPlaceholderIndex(PhiPlaceholder phi_placeholder) const {
605 size_t res =
606 phi_placeholder.GetBlockId() * heap_location_collector_.GetNumberOfHeapLocations() +
607 phi_placeholder.GetHeapLocation();
608 DCHECK_EQ(phi_placeholder, GetPhiPlaceholderAt(res))
609 << res << "blks: " << GetGraph()->GetBlocks().size()
610 << " hls: " << heap_location_collector_.GetNumberOfHeapLocations();
611 return res;
612 }
613
PhiPlaceholderIndex(Value phi_placeholder) const614 size_t PhiPlaceholderIndex(Value phi_placeholder) const {
615 return PhiPlaceholderIndex(phi_placeholder.GetPhiPlaceholder());
616 }
617
IsPartialNoEscape(HBasicBlock * blk,size_t idx)618 bool IsPartialNoEscape(HBasicBlock* blk, size_t idx) {
619 auto* ri = heap_location_collector_.GetHeapLocation(idx)->GetReferenceInfo();
620 if (!ri->IsPartialSingleton()) {
621 return false;
622 }
623 ArrayRef<const ExecutionSubgraph::ExcludedCohort> cohorts =
624 ri->GetNoEscapeSubgraph()->GetExcludedCohorts();
625 return std::none_of(cohorts.cbegin(),
626 cohorts.cend(),
627 [&](const ExecutionSubgraph::ExcludedCohort& ex) -> bool {
628 // Make sure we haven't yet and never will escape.
629 return ex.PrecedesBlock(blk) ||
630 ex.ContainsBlock(blk) ||
631 ex.SucceedsBlock(blk);
632 });
633 }
634
GetPhiPlaceholderAt(size_t off) const635 PhiPlaceholder GetPhiPlaceholderAt(size_t off) const {
636 DCHECK_LT(off, num_phi_placeholders_);
637 size_t id = off % heap_location_collector_.GetNumberOfHeapLocations();
638 // Technically this should be (off - id) / NumberOfHeapLocations
639 // but due to truncation it's all the same.
640 size_t blk_id = off / heap_location_collector_.GetNumberOfHeapLocations();
641 return GetPhiPlaceholder(blk_id, id);
642 }
643
GetPhiPlaceholder(uint32_t block_id,size_t idx) const644 PhiPlaceholder GetPhiPlaceholder(uint32_t block_id, size_t idx) const {
645 DCHECK(GetGraph()->GetBlocks()[block_id] != nullptr) << block_id;
646 return PhiPlaceholder(block_id, idx);
647 }
648
Replacement(Value value) const649 Value Replacement(Value value) const {
650 DCHECK(value.NeedsPhi() ||
651 (current_phase_ == Phase::kPartialElimination && value.IsMergedUnknown()))
652 << value << " phase: " << current_phase_;
653 Value replacement = phi_placeholder_replacements_[PhiPlaceholderIndex(value)];
654 DCHECK(replacement.IsUnknown() || replacement.IsInstruction());
655 DCHECK(replacement.IsUnknown() ||
656 FindSubstitute(replacement.GetInstruction()) == replacement.GetInstruction());
657 return replacement;
658 }
659
ReplacementOrValue(Value value) const660 Value ReplacementOrValue(Value value) const {
661 if (current_phase_ == Phase::kPartialElimination) {
662 // In this phase we are materializing the default values which are used
663 // only if the partial singleton did not escape, so we can replace
664 // a partial unknown with the prior value.
665 if (value.IsPartialUnknown()) {
666 value = value.GetPriorValue().ToValue();
667 }
668 if ((value.IsMergedUnknown() || value.NeedsPhi()) &&
669 phi_placeholder_replacements_[PhiPlaceholderIndex(value)].IsValid()) {
670 value = phi_placeholder_replacements_[PhiPlaceholderIndex(value)];
671 DCHECK(!value.IsMergedUnknown());
672 DCHECK(!value.NeedsPhi());
673 } else if (value.IsMergedUnknown()) {
674 return Value::ForLoopPhiPlaceholder(value.GetPhiPlaceholder());
675 }
676 if (value.IsInstruction() && value.GetInstruction()->IsInstanceFieldGet()) {
677 DCHECK_LT(static_cast<size_t>(value.GetInstruction()->GetId()),
678 substitute_instructions_for_loads_.size());
679 HInstruction* substitute =
680 substitute_instructions_for_loads_[value.GetInstruction()->GetId()];
681 if (substitute != nullptr) {
682 DCHECK(substitute->IsPredicatedInstanceFieldGet());
683 return Value::ForInstruction(substitute);
684 }
685 }
686 DCHECK(!value.IsInstruction() ||
687 FindSubstitute(value.GetInstruction()) == value.GetInstruction());
688 return value;
689 }
690 if (value.NeedsPhi() && phi_placeholder_replacements_[PhiPlaceholderIndex(value)].IsValid()) {
691 return Replacement(value);
692 } else {
693 DCHECK(!value.IsInstruction() ||
694 FindSubstitute(value.GetInstruction()) == value.GetInstruction());
695 return value;
696 }
697 }
698
699 // The record of a heap value and instruction(s) that feed that value.
700 struct ValueRecord {
701 Value value;
702 Value stored_by;
703 };
704
FindOrAddTypeConversionIfNecessary(HInstruction * instruction,HInstruction * value,DataType::Type expected_type)705 HTypeConversion* FindOrAddTypeConversionIfNecessary(HInstruction* instruction,
706 HInstruction* value,
707 DataType::Type expected_type) {
708 // Should never add type conversion into boolean value.
709 if (expected_type == DataType::Type::kBool ||
710 DataType::IsTypeConversionImplicit(value->GetType(), expected_type) ||
711 // TODO: This prevents type conversion of default values but we can still insert
712 // type conversion of other constants and there is no constant folding pass after LSE.
713 IsZeroBitPattern(value)) {
714 return nullptr;
715 }
716
717 // Check if there is already a suitable TypeConversion we can reuse.
718 for (const HUseListNode<HInstruction*>& use : value->GetUses()) {
719 if (use.GetUser()->IsTypeConversion() &&
720 use.GetUser()->GetType() == expected_type &&
721 // TODO: We could move the TypeConversion to a common dominator
722 // if it does not cross irreducible loop header.
723 use.GetUser()->GetBlock()->Dominates(instruction->GetBlock()) &&
724 // Don't share across irreducible loop headers.
725 // TODO: can be more fine-grained than this by testing each dominator.
726 (use.GetUser()->GetBlock() == instruction->GetBlock() ||
727 !GetGraph()->HasIrreducibleLoops())) {
728 if (use.GetUser()->GetBlock() == instruction->GetBlock() &&
729 use.GetUser()->GetBlock()->GetInstructions().FoundBefore(instruction, use.GetUser())) {
730 // Move the TypeConversion before the instruction.
731 use.GetUser()->MoveBefore(instruction);
732 }
733 DCHECK(use.GetUser()->StrictlyDominates(instruction));
734 return use.GetUser()->AsTypeConversion();
735 }
736 }
737
738 // We must create a new TypeConversion instruction.
739 HTypeConversion* type_conversion = new (GetGraph()->GetAllocator()) HTypeConversion(
740 expected_type, value, instruction->GetDexPc());
741 instruction->GetBlock()->InsertInstructionBefore(type_conversion, instruction);
742 return type_conversion;
743 }
744
745 // Find an instruction's substitute if it's a removed load.
746 // Return the same instruction if it should not be removed.
FindSubstitute(HInstruction * instruction) const747 HInstruction* FindSubstitute(HInstruction* instruction) const {
748 size_t id = static_cast<size_t>(instruction->GetId());
749 if (id >= substitute_instructions_for_loads_.size()) {
750 // New Phi (may not be in the graph yet), default value or PredicatedInstanceFieldGet.
751 DCHECK(!IsLoad(instruction) || instruction->IsPredicatedInstanceFieldGet());
752 return instruction;
753 }
754 HInstruction* substitute = substitute_instructions_for_loads_[id];
755 DCHECK(substitute == nullptr || IsLoad(instruction));
756 return (substitute != nullptr) ? substitute : instruction;
757 }
758
AddRemovedLoad(HInstruction * load,HInstruction * heap_value)759 void AddRemovedLoad(HInstruction* load, HInstruction* heap_value) {
760 DCHECK(IsLoad(load));
761 DCHECK_EQ(FindSubstitute(load), load);
762 DCHECK_EQ(FindSubstitute(heap_value), heap_value) <<
763 "Unexpected heap_value that has a substitute " << heap_value->DebugName();
764
765 // The load expects to load the heap value as type load->GetType().
766 // However the tracked heap value may not be of that type. An explicit
767 // type conversion may be needed.
768 // There are actually three types involved here:
769 // (1) tracked heap value's type (type A)
770 // (2) heap location (field or element)'s type (type B)
771 // (3) load's type (type C)
772 // We guarantee that type A stored as type B and then fetched out as
773 // type C is the same as casting from type A to type C directly, since
774 // type B and type C will have the same size which is guaranteed in
775 // HInstanceFieldGet/HStaticFieldGet/HArrayGet/HVecLoad's SetType().
776 // So we only need one type conversion from type A to type C.
777 HTypeConversion* type_conversion = FindOrAddTypeConversionIfNecessary(
778 load, heap_value, load->GetType());
779
780 substitute_instructions_for_loads_[load->GetId()] =
781 type_conversion != nullptr ? type_conversion : heap_value;
782 }
783
IsLoad(HInstruction * instruction)784 static bool IsLoad(HInstruction* instruction) {
785 // Unresolved load is not treated as a load.
786 return instruction->IsInstanceFieldGet() ||
787 instruction->IsPredicatedInstanceFieldGet() ||
788 instruction->IsStaticFieldGet() ||
789 instruction->IsVecLoad() ||
790 instruction->IsArrayGet();
791 }
792
IsStore(HInstruction * instruction)793 static bool IsStore(HInstruction* instruction) {
794 // Unresolved store is not treated as a store.
795 return instruction->IsInstanceFieldSet() ||
796 instruction->IsArraySet() ||
797 instruction->IsVecStore() ||
798 instruction->IsStaticFieldSet();
799 }
800
801 // Check if it is allowed to use default values or Phis for the specified load.
IsDefaultOrPhiAllowedForLoad(HInstruction * instruction)802 static bool IsDefaultOrPhiAllowedForLoad(HInstruction* instruction) {
803 DCHECK(IsLoad(instruction));
804 // Using defaults for VecLoads requires to create additional vector operations.
805 // As there are some issues with scheduling vector operations it is better to avoid creating
806 // them.
807 return !instruction->IsVecOperation();
808 }
809
810 // Keep the store referenced by the instruction, or all stores that feed a Phi placeholder.
811 // This is necessary if the stored heap value can be observed.
KeepStores(Value value)812 void KeepStores(Value value) {
813 if (value.IsPureUnknown() || value.IsPartialUnknown()) {
814 return;
815 }
816 if (value.IsMergedUnknown()) {
817 kept_merged_unknowns_.SetBit(PhiPlaceholderIndex(value));
818 phi_placeholders_to_search_for_kept_stores_.SetBit(PhiPlaceholderIndex(value));
819 return;
820 }
821 if (value.NeedsPhi()) {
822 phi_placeholders_to_search_for_kept_stores_.SetBit(PhiPlaceholderIndex(value));
823 } else {
824 HInstruction* instruction = value.GetInstruction();
825 DCHECK(IsStore(instruction));
826 kept_stores_.SetBit(instruction->GetId());
827 }
828 }
829
830 // If a heap location X may alias with heap location at `loc_index`
831 // and heap_values of that heap location X holds a store, keep that store.
832 // It's needed for a dependent load that's not eliminated since any store
833 // that may put value into the load's heap location needs to be kept.
KeepStoresIfAliasedToLocation(ScopedArenaVector<ValueRecord> & heap_values,size_t loc_index)834 void KeepStoresIfAliasedToLocation(ScopedArenaVector<ValueRecord>& heap_values,
835 size_t loc_index) {
836 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
837 if (i == loc_index) {
838 // We use this function when reading a location with unknown value and
839 // therefore we cannot know what exact store wrote that unknown value.
840 // But we can have a phi placeholder here marking multiple stores to keep.
841 DCHECK(
842 !heap_values[i].stored_by.IsInstruction() ||
843 heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo()->IsPartialSingleton());
844 KeepStores(heap_values[i].stored_by);
845 heap_values[i].stored_by = Value::PureUnknown();
846 } else if (heap_location_collector_.MayAlias(i, loc_index)) {
847 KeepStores(heap_values[i].stored_by);
848 heap_values[i].stored_by = Value::PureUnknown();
849 }
850 }
851 }
852
853 // `instruction` is being removed. Try to see if the null check on it
854 // can be removed. This can happen if the same value is set in two branches
855 // but not in dominators. Such as:
856 // int[] a = foo();
857 // if () {
858 // a[0] = 2;
859 // } else {
860 // a[0] = 2;
861 // }
862 // // a[0] can now be replaced with constant 2, and the null check on it can be removed.
TryRemovingNullCheck(HInstruction * instruction)863 void TryRemovingNullCheck(HInstruction* instruction) {
864 HInstruction* prev = instruction->GetPrevious();
865 if ((prev != nullptr) && prev->IsNullCheck() && (prev == instruction->InputAt(0))) {
866 // Previous instruction is a null check for this instruction. Remove the null check.
867 prev->ReplaceWith(prev->InputAt(0));
868 prev->GetBlock()->RemoveInstruction(prev);
869 }
870 }
871
GetDefaultValue(DataType::Type type)872 HInstruction* GetDefaultValue(DataType::Type type) {
873 switch (type) {
874 case DataType::Type::kReference:
875 return GetGraph()->GetNullConstant();
876 case DataType::Type::kBool:
877 case DataType::Type::kUint8:
878 case DataType::Type::kInt8:
879 case DataType::Type::kUint16:
880 case DataType::Type::kInt16:
881 case DataType::Type::kInt32:
882 return GetGraph()->GetIntConstant(0);
883 case DataType::Type::kInt64:
884 return GetGraph()->GetLongConstant(0);
885 case DataType::Type::kFloat32:
886 return GetGraph()->GetFloatConstant(0);
887 case DataType::Type::kFloat64:
888 return GetGraph()->GetDoubleConstant(0);
889 default:
890 UNREACHABLE();
891 }
892 }
893
CanValueBeKeptIfSameAsNew(Value value,HInstruction * new_value,HInstruction * new_value_set_instr)894 bool CanValueBeKeptIfSameAsNew(Value value,
895 HInstruction* new_value,
896 HInstruction* new_value_set_instr) {
897 // For field/array set location operations, if the value is the same as the new_value
898 // it can be kept even if aliasing happens. All aliased operations will access the same memory
899 // range.
900 // For vector values, this is not true. For example:
901 // packed_data = [0xA, 0xB, 0xC, 0xD]; <-- Different values in each lane.
902 // VecStore array[i ,i+1,i+2,i+3] = packed_data;
903 // VecStore array[i+1,i+2,i+3,i+4] = packed_data; <-- We are here (partial overlap).
904 // VecLoad vx = array[i,i+1,i+2,i+3]; <-- Cannot be eliminated because the value
905 // here is not packed_data anymore.
906 //
907 // TODO: to allow such 'same value' optimization on vector data,
908 // LSA needs to report more fine-grain MAY alias information:
909 // (1) May alias due to two vector data partial overlap.
910 // e.g. a[i..i+3] and a[i+1,..,i+4].
911 // (2) May alias due to two vector data may complete overlap each other.
912 // e.g. a[i..i+3] and b[i..i+3].
913 // (3) May alias but the exact relationship between two locations is unknown.
914 // e.g. a[i..i+3] and b[j..j+3], where values of a,b,i,j are all unknown.
915 // This 'same value' optimization can apply only on case (2).
916 if (new_value_set_instr->IsVecOperation()) {
917 return false;
918 }
919
920 return value.Equals(new_value);
921 }
922
923 Value PrepareLoopValue(HBasicBlock* block, size_t idx);
924 Value PrepareLoopStoredBy(HBasicBlock* block, size_t idx);
925 void PrepareLoopRecords(HBasicBlock* block);
926 Value MergePredecessorValues(HBasicBlock* block, size_t idx);
927 void MergePredecessorRecords(HBasicBlock* block);
928
929 void MaterializeNonLoopPhis(PhiPlaceholder phi_placeholder, DataType::Type type);
930
931 void VisitGetLocation(HInstruction* instruction, size_t idx);
932 void VisitSetLocation(HInstruction* instruction, size_t idx, HInstruction* value);
RecordFieldInfo(const FieldInfo * info,size_t heap_loc)933 void RecordFieldInfo(const FieldInfo* info, size_t heap_loc) {
934 field_infos_[heap_loc] = info;
935 }
936
937 void VisitBasicBlock(HBasicBlock* block) override;
938
939 enum class Phase {
940 kLoadElimination,
941 kStoreElimination,
942 kPartialElimination,
943 };
944
945 bool MayAliasOnBackEdge(HBasicBlock* loop_header, size_t idx1, size_t idx2) const;
946
947 bool TryReplacingLoopPhiPlaceholderWithDefault(
948 PhiPlaceholder phi_placeholder,
949 DataType::Type type,
950 /*inout*/ ArenaBitVector* phi_placeholders_to_materialize);
951 bool TryReplacingLoopPhiPlaceholderWithSingleInput(
952 PhiPlaceholder phi_placeholder,
953 /*inout*/ ArenaBitVector* phi_placeholders_to_materialize);
954 std::optional<PhiPlaceholder> FindLoopPhisToMaterialize(
955 PhiPlaceholder phi_placeholder,
956 /*out*/ ArenaBitVector* phi_placeholders_to_materialize,
957 DataType::Type type,
958 bool can_use_default_or_phi);
959 bool MaterializeLoopPhis(const ScopedArenaVector<size_t>& phi_placeholder_indexes,
960 DataType::Type type);
961 bool MaterializeLoopPhis(ArrayRef<const size_t> phi_placeholder_indexes, DataType::Type type);
962 bool MaterializeLoopPhis(const ArenaBitVector& phi_placeholders_to_materialize,
963 DataType::Type type);
964 bool FullyMaterializePhi(PhiPlaceholder phi_placeholder, DataType::Type type);
965 std::optional<PhiPlaceholder> TryToMaterializeLoopPhis(PhiPlaceholder phi_placeholder,
966 HInstruction* load);
967 void ProcessLoopPhiWithUnknownInput(PhiPlaceholder loop_phi_with_unknown_input);
968 void ProcessLoadsRequiringLoopPhis();
969
970 void SearchPhiPlaceholdersForKeptStores();
971 void UpdateValueRecordForStoreElimination(/*inout*/ValueRecord* value_record);
972 void FindOldValueForPhiPlaceholder(PhiPlaceholder phi_placeholder, DataType::Type type);
973 void FindStoresWritingOldValues();
974 void FinishFullLSE();
975 void PrepareForPartialPhiComputation();
976 // Create materialization block and materialization object for the given predecessor of entry.
977 HInstruction* SetupPartialMaterialization(PartialLoadStoreEliminationHelper& helper,
978 HeapRefHolder&& holder,
979 size_t pred_idx,
980 HBasicBlock* blk);
981 // Returns the value that would be read by the 'read' instruction on
982 // 'orig_new_inst' if 'orig_new_inst' has not escaped.
983 HInstruction* GetPartialValueAt(HNewInstance* orig_new_inst, HInstruction* read);
984 void MovePartialEscapes();
985
VisitPredicatedInstanceFieldGet(HPredicatedInstanceFieldGet * instruction)986 void VisitPredicatedInstanceFieldGet(HPredicatedInstanceFieldGet* instruction) override {
987 LOG(FATAL) << "Visited instruction " << instruction->DumpWithoutArgs()
988 << " but LSE should be the only source of predicated-ifield-gets!";
989 }
990
VisitInstanceFieldGet(HInstanceFieldGet * instruction)991 void VisitInstanceFieldGet(HInstanceFieldGet* instruction) override {
992 HInstruction* object = instruction->InputAt(0);
993 const FieldInfo& field = instruction->GetFieldInfo();
994 VisitGetLocation(instruction, heap_location_collector_.GetFieldHeapLocation(object, &field));
995 }
996
VisitInstanceFieldSet(HInstanceFieldSet * instruction)997 void VisitInstanceFieldSet(HInstanceFieldSet* instruction) override {
998 HInstruction* object = instruction->InputAt(0);
999 const FieldInfo& field = instruction->GetFieldInfo();
1000 HInstruction* value = instruction->InputAt(1);
1001 size_t idx = heap_location_collector_.GetFieldHeapLocation(object, &field);
1002 VisitSetLocation(instruction, idx, value);
1003 }
1004
VisitStaticFieldGet(HStaticFieldGet * instruction)1005 void VisitStaticFieldGet(HStaticFieldGet* instruction) override {
1006 HInstruction* cls = instruction->InputAt(0);
1007 const FieldInfo& field = instruction->GetFieldInfo();
1008 VisitGetLocation(instruction, heap_location_collector_.GetFieldHeapLocation(cls, &field));
1009 }
1010
VisitStaticFieldSet(HStaticFieldSet * instruction)1011 void VisitStaticFieldSet(HStaticFieldSet* instruction) override {
1012 HInstruction* cls = instruction->InputAt(0);
1013 const FieldInfo& field = instruction->GetFieldInfo();
1014 HInstruction* value = instruction->InputAt(1);
1015 size_t idx = heap_location_collector_.GetFieldHeapLocation(cls, &field);
1016 VisitSetLocation(instruction, idx, value);
1017 }
1018
VisitArrayGet(HArrayGet * instruction)1019 void VisitArrayGet(HArrayGet* instruction) override {
1020 VisitGetLocation(instruction, heap_location_collector_.GetArrayHeapLocation(instruction));
1021 }
1022
VisitArraySet(HArraySet * instruction)1023 void VisitArraySet(HArraySet* instruction) override {
1024 size_t idx = heap_location_collector_.GetArrayHeapLocation(instruction);
1025 VisitSetLocation(instruction, idx, instruction->GetValue());
1026 }
1027
VisitVecLoad(HVecLoad * instruction)1028 void VisitVecLoad(HVecLoad* instruction) override {
1029 VisitGetLocation(instruction, heap_location_collector_.GetArrayHeapLocation(instruction));
1030 }
1031
VisitVecStore(HVecStore * instruction)1032 void VisitVecStore(HVecStore* instruction) override {
1033 size_t idx = heap_location_collector_.GetArrayHeapLocation(instruction);
1034 VisitSetLocation(instruction, idx, instruction->GetValue());
1035 }
1036
VisitDeoptimize(HDeoptimize * instruction)1037 void VisitDeoptimize(HDeoptimize* instruction) override {
1038 ScopedArenaVector<ValueRecord>& heap_values =
1039 heap_values_for_[instruction->GetBlock()->GetBlockId()];
1040 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
1041 Value* stored_by = &heap_values[i].stored_by;
1042 if (stored_by->IsUnknown()) {
1043 continue;
1044 }
1045 // Stores are generally observeable after deoptimization, except
1046 // for singletons that don't escape in the deoptimization environment.
1047 bool observable = true;
1048 ReferenceInfo* info = heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo();
1049 if (info->IsSingleton()) {
1050 HInstruction* reference = info->GetReference();
1051 // Finalizable objects always escape.
1052 if (!reference->IsNewInstance() || !reference->AsNewInstance()->IsFinalizable()) {
1053 // Check whether the reference for a store is used by an environment local of
1054 // the HDeoptimize. If not, the singleton is not observed after deoptimization.
1055 const HUseList<HEnvironment*>& env_uses = reference->GetEnvUses();
1056 observable = std::any_of(
1057 env_uses.begin(),
1058 env_uses.end(),
1059 [instruction](const HUseListNode<HEnvironment*>& use) {
1060 return use.GetUser()->GetHolder() == instruction;
1061 });
1062 }
1063 }
1064 if (observable) {
1065 KeepStores(*stored_by);
1066 *stored_by = Value::PureUnknown();
1067 }
1068 }
1069 }
1070
1071 // Keep necessary stores before exiting a method via return/throw.
HandleExit(HBasicBlock * block)1072 void HandleExit(HBasicBlock* block) {
1073 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[block->GetBlockId()];
1074 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
1075 ReferenceInfo* ref_info = heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo();
1076 if (!ref_info->IsSingletonAndRemovable() &&
1077 !(ref_info->IsPartialSingleton() && IsPartialNoEscape(block, i))) {
1078 KeepStores(heap_values[i].stored_by);
1079 heap_values[i].stored_by = Value::PureUnknown();
1080 }
1081 }
1082 }
1083
VisitReturn(HReturn * instruction)1084 void VisitReturn(HReturn* instruction) override {
1085 HandleExit(instruction->GetBlock());
1086 }
1087
VisitReturnVoid(HReturnVoid * return_void)1088 void VisitReturnVoid(HReturnVoid* return_void) override {
1089 HandleExit(return_void->GetBlock());
1090 }
1091
VisitThrow(HThrow * throw_instruction)1092 void VisitThrow(HThrow* throw_instruction) override {
1093 HandleExit(throw_instruction->GetBlock());
1094 }
1095
HandleInvoke(HInstruction * instruction)1096 void HandleInvoke(HInstruction* instruction) {
1097 SideEffects side_effects = instruction->GetSideEffects();
1098 ScopedArenaVector<ValueRecord>& heap_values =
1099 heap_values_for_[instruction->GetBlock()->GetBlockId()];
1100 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
1101 ReferenceInfo* ref_info = heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo();
1102 HBasicBlock* blk = instruction->GetBlock();
1103 // We don't need to do anything if the reference has not escaped at this point.
1104 // This is true if either we (1) never escape or (2) sometimes escape but
1105 // there is no possible execution where we have done so at this time. NB
1106 // We count being in the excluded cohort as escaping. Technically, this is
1107 // a bit over-conservative (since we can have multiple non-escaping calls
1108 // before a single escaping one) but this simplifies everything greatly.
1109 auto partial_singleton_did_not_escape = [](ReferenceInfo* ref_info, HBasicBlock* blk) {
1110 DCHECK(ref_info->IsPartialSingleton());
1111 if (!ref_info->GetNoEscapeSubgraph()->ContainsBlock(blk)) {
1112 return false;
1113 }
1114 ArrayRef<const ExecutionSubgraph::ExcludedCohort> cohorts =
1115 ref_info->GetNoEscapeSubgraph()->GetExcludedCohorts();
1116 return std::none_of(cohorts.begin(),
1117 cohorts.end(),
1118 [&](const ExecutionSubgraph::ExcludedCohort& cohort) {
1119 return cohort.PrecedesBlock(blk);
1120 });
1121 };
1122 if (ref_info->IsSingleton() ||
1123 // partial and we aren't currently escaping and we haven't escaped yet.
1124 (ref_info->IsPartialSingleton() && partial_singleton_did_not_escape(ref_info, blk))) {
1125 // Singleton references cannot be seen by the callee.
1126 } else {
1127 if (side_effects.DoesAnyRead() || side_effects.DoesAnyWrite()) {
1128 // Previous stores may become visible (read) and/or impossible for LSE to track (write).
1129 KeepStores(heap_values[i].stored_by);
1130 heap_values[i].stored_by = Value::PureUnknown();
1131 }
1132 if (side_effects.DoesAnyWrite()) {
1133 // The value may be clobbered.
1134 heap_values[i].value = Value::PartialUnknown(heap_values[i].value);
1135 }
1136 }
1137 }
1138 }
1139
VisitInvoke(HInvoke * invoke)1140 void VisitInvoke(HInvoke* invoke) override {
1141 HandleInvoke(invoke);
1142 }
1143
VisitClinitCheck(HClinitCheck * clinit)1144 void VisitClinitCheck(HClinitCheck* clinit) override {
1145 // Class initialization check can result in class initializer calling arbitrary methods.
1146 HandleInvoke(clinit);
1147 }
1148
VisitUnresolvedInstanceFieldGet(HUnresolvedInstanceFieldGet * instruction)1149 void VisitUnresolvedInstanceFieldGet(HUnresolvedInstanceFieldGet* instruction) override {
1150 // Conservatively treat it as an invocation.
1151 HandleInvoke(instruction);
1152 }
1153
VisitUnresolvedInstanceFieldSet(HUnresolvedInstanceFieldSet * instruction)1154 void VisitUnresolvedInstanceFieldSet(HUnresolvedInstanceFieldSet* instruction) override {
1155 // Conservatively treat it as an invocation.
1156 HandleInvoke(instruction);
1157 }
1158
VisitUnresolvedStaticFieldGet(HUnresolvedStaticFieldGet * instruction)1159 void VisitUnresolvedStaticFieldGet(HUnresolvedStaticFieldGet* instruction) override {
1160 // Conservatively treat it as an invocation.
1161 HandleInvoke(instruction);
1162 }
1163
VisitUnresolvedStaticFieldSet(HUnresolvedStaticFieldSet * instruction)1164 void VisitUnresolvedStaticFieldSet(HUnresolvedStaticFieldSet* instruction) override {
1165 // Conservatively treat it as an invocation.
1166 HandleInvoke(instruction);
1167 }
1168
VisitNewInstance(HNewInstance * new_instance)1169 void VisitNewInstance(HNewInstance* new_instance) override {
1170 ReferenceInfo* ref_info = heap_location_collector_.FindReferenceInfoOf(new_instance);
1171 if (ref_info == nullptr) {
1172 // new_instance isn't used for field accesses. No need to process it.
1173 return;
1174 }
1175 if (ref_info->IsSingletonAndRemovable() && !new_instance->NeedsChecks()) {
1176 DCHECK(!new_instance->IsFinalizable());
1177 // new_instance can potentially be eliminated.
1178 singleton_new_instances_.push_back(new_instance);
1179 }
1180 ScopedArenaVector<ValueRecord>& heap_values =
1181 heap_values_for_[new_instance->GetBlock()->GetBlockId()];
1182 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
1183 HInstruction* ref =
1184 heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo()->GetReference();
1185 size_t offset = heap_location_collector_.GetHeapLocation(i)->GetOffset();
1186 if (ref == new_instance) {
1187 if (offset >= mirror::kObjectHeaderSize ||
1188 MemberOffset(offset) == mirror::Object::MonitorOffset()) {
1189 // Instance fields except the header fields are set to default heap values.
1190 // The shadow$_monitor_ field is set to the default value however.
1191 heap_values[i].value = Value::Default();
1192 heap_values[i].stored_by = Value::PureUnknown();
1193 } else if (MemberOffset(offset) == mirror::Object::ClassOffset()) {
1194 // The shadow$_klass_ field is special and has an actual value however.
1195 heap_values[i].value = Value::ForInstruction(new_instance->GetLoadClass());
1196 heap_values[i].stored_by = Value::PureUnknown();
1197 }
1198 }
1199 }
1200 }
1201
VisitNewArray(HNewArray * new_array)1202 void VisitNewArray(HNewArray* new_array) override {
1203 ReferenceInfo* ref_info = heap_location_collector_.FindReferenceInfoOf(new_array);
1204 if (ref_info == nullptr) {
1205 // new_array isn't used for array accesses. No need to process it.
1206 return;
1207 }
1208 if (ref_info->IsSingletonAndRemovable()) {
1209 if (new_array->GetLength()->IsIntConstant() &&
1210 new_array->GetLength()->AsIntConstant()->GetValue() >= 0) {
1211 // new_array can potentially be eliminated.
1212 singleton_new_instances_.push_back(new_array);
1213 } else {
1214 // new_array may throw NegativeArraySizeException. Keep it.
1215 }
1216 }
1217 ScopedArenaVector<ValueRecord>& heap_values =
1218 heap_values_for_[new_array->GetBlock()->GetBlockId()];
1219 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
1220 HeapLocation* location = heap_location_collector_.GetHeapLocation(i);
1221 HInstruction* ref = location->GetReferenceInfo()->GetReference();
1222 if (ref == new_array && location->GetIndex() != nullptr) {
1223 // Array elements are set to default heap values.
1224 heap_values[i].value = Value::Default();
1225 heap_values[i].stored_by = Value::PureUnknown();
1226 }
1227 }
1228 }
1229
ShouldPerformPartialLSE() const1230 bool ShouldPerformPartialLSE() const {
1231 return perform_partial_lse_ && !GetGraph()->IsCompilingOsr();
1232 }
1233
1234 bool perform_partial_lse_;
1235
1236 const HeapLocationCollector& heap_location_collector_;
1237
1238 // Use local allocator for allocating memory.
1239 ScopedArenaAllocator allocator_;
1240
1241 // The number of unique phi_placeholders there possibly are
1242 size_t num_phi_placeholders_;
1243
1244 // One array of heap value records for each block.
1245 ScopedArenaVector<ScopedArenaVector<ValueRecord>> heap_values_for_;
1246
1247 // We record loads and stores for re-processing when we find a loop Phi placeholder
1248 // with unknown value from a predecessor, and also for removing stores that are
1249 // found to be dead, i.e. not marked in `kept_stores_` at the end.
1250 struct LoadStoreRecord {
1251 HInstruction* load_or_store;
1252 size_t heap_location_index;
1253 };
1254 ScopedArenaVector<LoadStoreRecord> loads_and_stores_;
1255
1256 // We record the substitute instructions for loads that should be
1257 // eliminated but may be used by heap locations. They'll be removed
1258 // in the end. These are indexed by the load's id.
1259 ScopedArenaVector<HInstruction*> substitute_instructions_for_loads_;
1260
1261 // Value at the start of the given instruction for instructions which directly
1262 // read from a heap-location (i.e. FieldGet). The mapping to heap-location is
1263 // implicit through the fact that each instruction can only directly refer to
1264 // a single heap-location.
1265 ScopedArenaHashMap<HInstruction*, Value> intermediate_values_;
1266
1267 // Record stores to keep in a bit vector indexed by instruction ID.
1268 ArenaBitVector kept_stores_;
1269 // When we need to keep all stores that feed a Phi placeholder, we just record the
1270 // index of that placeholder for processing after graph traversal.
1271 ArenaBitVector phi_placeholders_to_search_for_kept_stores_;
1272
1273 // Loads that would require a loop Phi to replace are recorded for processing
1274 // later as we do not have enough information from back-edges to determine if
1275 // a suitable Phi can be found or created when we visit these loads.
1276 ScopedArenaHashMap<HInstruction*, ValueRecord> loads_requiring_loop_phi_;
1277
1278 // For stores, record the old value records that were replaced and the stored values.
1279 struct StoreRecord {
1280 ValueRecord old_value_record;
1281 HInstruction* stored_value;
1282 };
1283 // Small pre-allocated initial buffer avoids initializing a large one until it's really needed.
1284 static constexpr size_t kStoreRecordsInitialBufferSize = 16;
1285 std::pair<HInstruction*, StoreRecord> store_records_buffer_[kStoreRecordsInitialBufferSize];
1286 ScopedArenaHashMap<HInstruction*, StoreRecord> store_records_;
1287
1288 // Replacements for Phi placeholders.
1289 // The invalid heap value is used to mark Phi placeholders that cannot be replaced.
1290 ScopedArenaVector<Value> phi_placeholder_replacements_;
1291
1292 // Merged-unknowns that must have their predecessor values kept to ensure
1293 // partially escaped values are written
1294 ArenaBitVector kept_merged_unknowns_;
1295
1296 ScopedArenaVector<HInstruction*> singleton_new_instances_;
1297
1298 // The field infos for each heap location (if relevant).
1299 ScopedArenaVector<const FieldInfo*> field_infos_;
1300
1301 Phase current_phase_;
1302
1303 friend class PartialLoadStoreEliminationHelper;
1304 friend struct ScopedRestoreHeapValues;
1305
1306 friend std::ostream& operator<<(std::ostream& os, const Value& v);
1307 friend std::ostream& operator<<(std::ostream& os, const PriorValueHolder& v);
1308 friend std::ostream& operator<<(std::ostream& oss, const LSEVisitor::Phase& phase);
1309
1310 DISALLOW_COPY_AND_ASSIGN(LSEVisitor);
1311 };
1312
operator <<(std::ostream & oss,const LSEVisitor::PriorValueHolder & p)1313 std::ostream& operator<<(std::ostream& oss, const LSEVisitor::PriorValueHolder& p) {
1314 p.Dump(oss);
1315 return oss;
1316 }
1317
operator <<(std::ostream & oss,const LSEVisitor::Phase & phase)1318 std::ostream& operator<<(std::ostream& oss, const LSEVisitor::Phase& phase) {
1319 switch (phase) {
1320 case LSEVisitor::Phase::kLoadElimination:
1321 return oss << "kLoadElimination";
1322 case LSEVisitor::Phase::kStoreElimination:
1323 return oss << "kStoreElimination";
1324 case LSEVisitor::Phase::kPartialElimination:
1325 return oss << "kPartialElimination";
1326 }
1327 }
1328
Dump(std::ostream & oss) const1329 void LSEVisitor::PriorValueHolder::Dump(std::ostream& oss) const {
1330 if (IsDefault()) {
1331 oss << "Default";
1332 } else if (IsPhi()) {
1333 oss << "Phi: " << GetPhiPlaceholder();
1334 } else {
1335 oss << "Instruction: " << *GetInstruction();
1336 }
1337 }
1338
PriorValueHolder(Value val)1339 constexpr LSEVisitor::PriorValueHolder::PriorValueHolder(Value val)
1340 : value_(Marker{}) {
1341 DCHECK(!val.IsInvalid() && !val.IsPureUnknown());
1342 if (val.IsPartialUnknown()) {
1343 value_ = val.GetPriorValue().value_;
1344 } else if (val.IsMergedUnknown() || val.NeedsPhi()) {
1345 value_ = val.GetPhiPlaceholder();
1346 } else if (val.IsInstruction()) {
1347 value_ = val.GetInstruction();
1348 } else {
1349 DCHECK(val.IsDefault());
1350 }
1351 }
1352
operator ==(const LSEVisitor::Marker &,const LSEVisitor::Marker &)1353 constexpr bool operator==(const LSEVisitor::Marker&, const LSEVisitor::Marker&) {
1354 return true;
1355 }
1356
operator ==(const LSEVisitor::PriorValueHolder & p1,const LSEVisitor::PriorValueHolder & p2)1357 constexpr bool operator==(const LSEVisitor::PriorValueHolder& p1,
1358 const LSEVisitor::PriorValueHolder& p2) {
1359 return p1.Equals(p2);
1360 }
1361
operator ==(const LSEVisitor::PhiPlaceholder & p1,const LSEVisitor::PhiPlaceholder & p2)1362 constexpr bool operator==(const LSEVisitor::PhiPlaceholder& p1,
1363 const LSEVisitor::PhiPlaceholder& p2) {
1364 return p1.Equals(p2);
1365 }
1366
operator ==(const LSEVisitor::Value::NeedsLoopPhiMarker & p1,const LSEVisitor::Value::NeedsLoopPhiMarker & p2)1367 constexpr bool operator==(const LSEVisitor::Value::NeedsLoopPhiMarker& p1,
1368 const LSEVisitor::Value::NeedsLoopPhiMarker& p2) {
1369 return p1.phi_ == p2.phi_;
1370 }
1371
operator ==(const LSEVisitor::Value::NeedsNonLoopPhiMarker & p1,const LSEVisitor::Value::NeedsNonLoopPhiMarker & p2)1372 constexpr bool operator==(const LSEVisitor::Value::NeedsNonLoopPhiMarker& p1,
1373 const LSEVisitor::Value::NeedsNonLoopPhiMarker& p2) {
1374 return p1.phi_ == p2.phi_;
1375 }
1376
operator ==(const LSEVisitor::Value::MergedUnknownMarker & p1,const LSEVisitor::Value::MergedUnknownMarker & p2)1377 constexpr bool operator==(const LSEVisitor::Value::MergedUnknownMarker& p1,
1378 const LSEVisitor::Value::MergedUnknownMarker& p2) {
1379 return p1.phi_ == p2.phi_;
1380 }
1381
operator <<(std::ostream & oss,const LSEVisitor::PhiPlaceholder & p)1382 std::ostream& operator<<(std::ostream& oss, const LSEVisitor::PhiPlaceholder& p) {
1383 p.Dump(oss);
1384 return oss;
1385 }
1386
ToValue() const1387 LSEVisitor::Value LSEVisitor::PriorValueHolder::ToValue() const {
1388 if (IsDefault()) {
1389 return Value::Default();
1390 } else if (IsPhi()) {
1391 return Value::ForLoopPhiPlaceholder(GetPhiPlaceholder());
1392 } else {
1393 return Value::ForInstruction(GetInstruction());
1394 }
1395 }
1396
ExactEquals(LSEVisitor::Value other) const1397 constexpr bool LSEVisitor::Value::ExactEquals(LSEVisitor::Value other) const {
1398 return value_ == other.value_;
1399 }
1400
Equals(LSEVisitor::Value other) const1401 constexpr bool LSEVisitor::Value::Equals(LSEVisitor::Value other) const {
1402 // Only valid values can be compared.
1403 DCHECK(IsValid());
1404 DCHECK(other.IsValid());
1405 if (value_ == other.value_) {
1406 // Note: Two unknown values are considered different.
1407 return !IsUnknown();
1408 } else {
1409 // Default is considered equal to zero-bit-pattern instructions.
1410 return (IsDefault() && other.IsInstruction() && IsZeroBitPattern(other.GetInstruction())) ||
1411 (other.IsDefault() && IsInstruction() && IsZeroBitPattern(GetInstruction()));
1412 }
1413 }
1414
Dump(std::ostream & os) const1415 std::ostream& LSEVisitor::Value::Dump(std::ostream& os) const {
1416 if (std::holds_alternative<LSEVisitor::Value::ValuelessType>(value_)) {
1417 switch (GetValuelessType()) {
1418 case ValuelessType::kDefault:
1419 return os << "Default";
1420 case ValuelessType::kPureUnknown:
1421 return os << "PureUnknown";
1422 case ValuelessType::kInvalid:
1423 return os << "Invalid";
1424 }
1425 } else if (IsPartialUnknown()) {
1426 return os << "PartialUnknown[" << GetPriorValue() << "]";
1427 } else if (IsInstruction()) {
1428 return os << "Instruction[id: " << GetInstruction()->GetId()
1429 << ", block: " << GetInstruction()->GetBlock()->GetBlockId() << "]";
1430 } else if (IsMergedUnknown()) {
1431 return os << "MergedUnknown[block: " << GetPhiPlaceholder().GetBlockId()
1432 << ", heap_loc: " << GetPhiPlaceholder().GetHeapLocation() << "]";
1433
1434 } else if (NeedsLoopPhi()) {
1435 return os << "NeedsLoopPhi[block: " << GetPhiPlaceholder().GetBlockId()
1436 << ", heap_loc: " << GetPhiPlaceholder().GetHeapLocation() << "]";
1437 } else {
1438 return os << "NeedsNonLoopPhi[block: " << GetPhiPlaceholder().GetBlockId()
1439 << ", heap_loc: " << GetPhiPlaceholder().GetHeapLocation() << "]";
1440 }
1441 }
1442
operator <<(std::ostream & os,const LSEVisitor::Value & v)1443 std::ostream& operator<<(std::ostream& os, const LSEVisitor::Value& v) {
1444 return v.Dump(os);
1445 }
1446
LSEVisitor(HGraph * graph,const HeapLocationCollector & heap_location_collector,bool perform_partial_lse,OptimizingCompilerStats * stats)1447 LSEVisitor::LSEVisitor(HGraph* graph,
1448 const HeapLocationCollector& heap_location_collector,
1449 bool perform_partial_lse,
1450 OptimizingCompilerStats* stats)
1451 : HGraphDelegateVisitor(graph, stats),
1452 perform_partial_lse_(perform_partial_lse),
1453 heap_location_collector_(heap_location_collector),
1454 allocator_(graph->GetArenaStack()),
1455 num_phi_placeholders_(GetGraph()->GetBlocks().size() *
1456 heap_location_collector_.GetNumberOfHeapLocations()),
1457 heap_values_for_(graph->GetBlocks().size(),
1458 ScopedArenaVector<ValueRecord>(allocator_.Adapter(kArenaAllocLSE)),
1459 allocator_.Adapter(kArenaAllocLSE)),
1460 loads_and_stores_(allocator_.Adapter(kArenaAllocLSE)),
1461 // We may add new instructions (default values, Phis) but we're not adding loads
1462 // or stores, so we shall not need to resize following vector and BitVector.
1463 substitute_instructions_for_loads_(graph->GetCurrentInstructionId(),
1464 nullptr,
1465 allocator_.Adapter(kArenaAllocLSE)),
1466 intermediate_values_(allocator_.Adapter(kArenaAllocLSE)),
1467 kept_stores_(&allocator_,
1468 /*start_bits=*/graph->GetCurrentInstructionId(),
1469 /*expandable=*/false,
1470 kArenaAllocLSE),
1471 phi_placeholders_to_search_for_kept_stores_(&allocator_,
1472 num_phi_placeholders_,
1473 /*expandable=*/false,
1474 kArenaAllocLSE),
1475 loads_requiring_loop_phi_(allocator_.Adapter(kArenaAllocLSE)),
1476 store_records_(store_records_buffer_,
1477 kStoreRecordsInitialBufferSize,
1478 allocator_.Adapter(kArenaAllocLSE)),
1479 phi_placeholder_replacements_(num_phi_placeholders_,
1480 Value::Invalid(),
1481 allocator_.Adapter(kArenaAllocLSE)),
1482 kept_merged_unknowns_(&allocator_,
1483 /*start_bits=*/num_phi_placeholders_,
1484 /*expandable=*/false,
1485 kArenaAllocLSE),
1486 singleton_new_instances_(allocator_.Adapter(kArenaAllocLSE)),
1487 field_infos_(heap_location_collector_.GetNumberOfHeapLocations(),
1488 allocator_.Adapter(kArenaAllocLSE)),
1489 current_phase_(Phase::kLoadElimination) {
1490 // Clear bit vectors.
1491 phi_placeholders_to_search_for_kept_stores_.ClearAllBits();
1492 kept_stores_.ClearAllBits();
1493 }
1494
PrepareLoopValue(HBasicBlock * block,size_t idx)1495 LSEVisitor::Value LSEVisitor::PrepareLoopValue(HBasicBlock* block, size_t idx) {
1496 // If the pre-header value is known (which implies that the reference dominates this
1497 // block), use a Phi placeholder for the value in the loop header. If all predecessors
1498 // are later found to have a known value, we can replace loads from this location,
1499 // either with the pre-header value or with a new Phi. For array locations, the index
1500 // may be defined inside the loop but the only known value in that case should be the
1501 // default value or a Phi placeholder that can be replaced only with the default value.
1502 HLoopInformation* loop_info = block->GetLoopInformation();
1503 uint32_t pre_header_block_id = loop_info->GetPreHeader()->GetBlockId();
1504 Value pre_header_value = ReplacementOrValue(heap_values_for_[pre_header_block_id][idx].value);
1505 if (pre_header_value.IsUnknown()) {
1506 return pre_header_value;
1507 }
1508 if (kIsDebugBuild) {
1509 // Check that the reference indeed dominates this loop.
1510 HeapLocation* location = heap_location_collector_.GetHeapLocation(idx);
1511 HInstruction* ref = location->GetReferenceInfo()->GetReference();
1512 CHECK(ref->GetBlock() != block && ref->GetBlock()->Dominates(block))
1513 << GetGraph()->PrettyMethod();
1514 // Check that the index, if defined inside the loop, tracks a default value
1515 // or a Phi placeholder requiring a loop Phi.
1516 HInstruction* index = location->GetIndex();
1517 if (index != nullptr && loop_info->Contains(*index->GetBlock())) {
1518 CHECK(pre_header_value.NeedsLoopPhi() || pre_header_value.Equals(Value::Default()))
1519 << GetGraph()->PrettyMethod() << " blk: " << block->GetBlockId() << " "
1520 << pre_header_value;
1521 }
1522 }
1523 PhiPlaceholder phi_placeholder = GetPhiPlaceholder(block->GetBlockId(), idx);
1524 return ReplacementOrValue(Value::ForLoopPhiPlaceholder(phi_placeholder));
1525 }
1526
PrepareLoopStoredBy(HBasicBlock * block,size_t idx)1527 LSEVisitor::Value LSEVisitor::PrepareLoopStoredBy(HBasicBlock* block, size_t idx) {
1528 // Use the Phi placeholder for `stored_by` to make sure all incoming stores are kept
1529 // if the value in the location escapes. This is not applicable to singletons that are
1530 // defined inside the loop as they shall be dead in the loop header.
1531 ReferenceInfo* ref_info = heap_location_collector_.GetHeapLocation(idx)->GetReferenceInfo();
1532 if (ref_info->IsSingleton() &&
1533 block->GetLoopInformation()->Contains(*ref_info->GetReference()->GetBlock())) {
1534 return Value::PureUnknown();
1535 }
1536 PhiPlaceholder phi_placeholder = GetPhiPlaceholder(block->GetBlockId(), idx);
1537 return Value::ForLoopPhiPlaceholder(phi_placeholder);
1538 }
1539
PrepareLoopRecords(HBasicBlock * block)1540 void LSEVisitor::PrepareLoopRecords(HBasicBlock* block) {
1541 DCHECK(block->IsLoopHeader());
1542 int block_id = block->GetBlockId();
1543 HBasicBlock* pre_header = block->GetLoopInformation()->GetPreHeader();
1544 ScopedArenaVector<ValueRecord>& pre_header_heap_values =
1545 heap_values_for_[pre_header->GetBlockId()];
1546 size_t num_heap_locations = heap_location_collector_.GetNumberOfHeapLocations();
1547 DCHECK_EQ(num_heap_locations, pre_header_heap_values.size());
1548 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[block_id];
1549 DCHECK(heap_values.empty());
1550
1551 // Don't eliminate loads in irreducible loops.
1552 if (block->GetLoopInformation()->IsIrreducible()) {
1553 heap_values.resize(num_heap_locations,
1554 {/*value=*/Value::Invalid(), /*stored_by=*/Value::PureUnknown()});
1555 // Also keep the stores before the loop header, including in blocks that were not visited yet.
1556 bool is_osr = GetGraph()->IsCompilingOsr();
1557 for (size_t idx = 0u; idx != num_heap_locations; ++idx) {
1558 heap_values[idx].value =
1559 is_osr ? Value::PureUnknown()
1560 : Value::MergedUnknown(GetPhiPlaceholder(block->GetBlockId(), idx));
1561 KeepStores(Value::ForLoopPhiPlaceholder(GetPhiPlaceholder(block->GetBlockId(), idx)));
1562 }
1563 return;
1564 }
1565
1566 // Fill `heap_values` based on values from pre-header.
1567 heap_values.reserve(num_heap_locations);
1568 for (size_t idx = 0u; idx != num_heap_locations; ++idx) {
1569 heap_values.push_back({ PrepareLoopValue(block, idx), PrepareLoopStoredBy(block, idx) });
1570 }
1571 }
1572
MergePredecessorValues(HBasicBlock * block,size_t idx)1573 LSEVisitor::Value LSEVisitor::MergePredecessorValues(HBasicBlock* block, size_t idx) {
1574 ArrayRef<HBasicBlock* const> predecessors(block->GetPredecessors());
1575 DCHECK(!predecessors.empty());
1576 Value merged_value =
1577 ReplacementOrValue(heap_values_for_[predecessors[0]->GetBlockId()][idx].value);
1578 for (size_t i = 1u, size = predecessors.size(); i != size; ++i) {
1579 Value pred_value =
1580 ReplacementOrValue(heap_values_for_[predecessors[i]->GetBlockId()][idx].value);
1581 if (pred_value.Equals(merged_value)) {
1582 // Value is the same. No need to update our merged value.
1583 continue;
1584 } else if (pred_value.IsUnknown() || merged_value.IsUnknown()) {
1585 // If one is unknown and the other is a different type of unknown
1586 PhiPlaceholder phi_placeholder = GetPhiPlaceholder(block->GetBlockId(), idx);
1587 merged_value = Value::MergedUnknown(phi_placeholder);
1588 // We know that at least one of the merge points is unknown (and both are
1589 // not pure-unknowns since that's captured above). This means that the
1590 // overall value needs to be a MergedUnknown. Just return that.
1591 break;
1592 } else {
1593 // There are conflicting known values. We may still be able to replace loads with a Phi.
1594 PhiPlaceholder phi_placeholder = GetPhiPlaceholder(block->GetBlockId(), idx);
1595 // Propagate the need for a new loop Phi from all predecessors.
1596 bool needs_loop_phi = merged_value.NeedsLoopPhi() || pred_value.NeedsLoopPhi();
1597 merged_value = ReplacementOrValue(Value::ForPhiPlaceholder(phi_placeholder, needs_loop_phi));
1598 }
1599 }
1600 DCHECK(!merged_value.IsPureUnknown() || block->GetPredecessors().size() <= 1)
1601 << merged_value << " in " << GetGraph()->PrettyMethod();
1602 return merged_value;
1603 }
1604
MergePredecessorRecords(HBasicBlock * block)1605 void LSEVisitor::MergePredecessorRecords(HBasicBlock* block) {
1606 if (block->IsExitBlock()) {
1607 // Exit block doesn't really merge values since the control flow ends in
1608 // its predecessors. Each predecessor needs to make sure stores are kept
1609 // if necessary.
1610 return;
1611 }
1612
1613 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[block->GetBlockId()];
1614 DCHECK(heap_values.empty());
1615 size_t num_heap_locations = heap_location_collector_.GetNumberOfHeapLocations();
1616 if (block->GetPredecessors().empty()) {
1617 DCHECK(block->IsEntryBlock());
1618 heap_values.resize(num_heap_locations,
1619 {/*value=*/Value::PureUnknown(), /*stored_by=*/Value::PureUnknown()});
1620 return;
1621 }
1622
1623 heap_values.reserve(num_heap_locations);
1624 for (size_t idx = 0u; idx != num_heap_locations; ++idx) {
1625 Value merged_value = MergePredecessorValues(block, idx);
1626 if (kIsDebugBuild) {
1627 if (merged_value.NeedsPhi()) {
1628 uint32_t block_id = merged_value.GetPhiPlaceholder().GetBlockId();
1629 CHECK(GetGraph()->GetBlocks()[block_id]->Dominates(block));
1630 } else if (merged_value.IsInstruction()) {
1631 CHECK(merged_value.GetInstruction()->GetBlock()->Dominates(block));
1632 }
1633 }
1634 ArrayRef<HBasicBlock* const> predecessors(block->GetPredecessors());
1635 Value merged_stored_by = heap_values_for_[predecessors[0]->GetBlockId()][idx].stored_by;
1636 for (size_t predecessor_idx = 1u; predecessor_idx != predecessors.size(); ++predecessor_idx) {
1637 uint32_t predecessor_block_id = predecessors[predecessor_idx]->GetBlockId();
1638 Value stored_by = heap_values_for_[predecessor_block_id][idx].stored_by;
1639 if ((!stored_by.IsUnknown() || !merged_stored_by.IsUnknown()) &&
1640 !merged_stored_by.Equals(stored_by)) {
1641 // Use the Phi placeholder to track that we need to keep stores from all predecessors.
1642 PhiPlaceholder phi_placeholder = GetPhiPlaceholder(block->GetBlockId(), idx);
1643 merged_stored_by = Value::ForNonLoopPhiPlaceholder(phi_placeholder);
1644 break;
1645 }
1646 }
1647 heap_values.push_back({ merged_value, merged_stored_by });
1648 }
1649 }
1650
FindOrConstructNonLoopPhi(HBasicBlock * block,const ScopedArenaVector<HInstruction * > & phi_inputs,DataType::Type type)1651 static HInstruction* FindOrConstructNonLoopPhi(
1652 HBasicBlock* block,
1653 const ScopedArenaVector<HInstruction*>& phi_inputs,
1654 DataType::Type type) {
1655 for (HInstructionIterator phi_it(block->GetPhis()); !phi_it.Done(); phi_it.Advance()) {
1656 HInstruction* phi = phi_it.Current();
1657 DCHECK_EQ(phi->InputCount(), phi_inputs.size());
1658 auto cmp = [](HInstruction* lhs, const HUserRecord<HInstruction*>& rhs) {
1659 return lhs == rhs.GetInstruction();
1660 };
1661 if (std::equal(phi_inputs.begin(), phi_inputs.end(), phi->GetInputRecords().begin(), cmp)) {
1662 return phi;
1663 }
1664 }
1665 ArenaAllocator* allocator = block->GetGraph()->GetAllocator();
1666 HPhi* phi = new (allocator) HPhi(allocator, kNoRegNumber, phi_inputs.size(), type);
1667 for (size_t i = 0, size = phi_inputs.size(); i != size; ++i) {
1668 DCHECK_NE(phi_inputs[i]->GetType(), DataType::Type::kVoid) << phi_inputs[i]->DebugName();
1669 phi->SetRawInputAt(i, phi_inputs[i]);
1670 }
1671 block->AddPhi(phi);
1672 if (type == DataType::Type::kReference) {
1673 // Update reference type information. Pass invalid handles, these are not used for Phis.
1674 ReferenceTypePropagation rtp_fixup(block->GetGraph(),
1675 Handle<mirror::ClassLoader>(),
1676 Handle<mirror::DexCache>(),
1677 /* is_first_run= */ false);
1678 rtp_fixup.Visit(phi);
1679 }
1680 return phi;
1681 }
1682
MaterializeNonLoopPhis(PhiPlaceholder phi_placeholder,DataType::Type type)1683 void LSEVisitor::MaterializeNonLoopPhis(PhiPlaceholder phi_placeholder, DataType::Type type) {
1684 DCHECK(phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)].IsInvalid());
1685 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
1686 size_t idx = phi_placeholder.GetHeapLocation();
1687
1688 // Use local allocator to reduce peak memory usage.
1689 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
1690 // Reuse the same vector for collecting phi inputs.
1691 ScopedArenaVector<HInstruction*> phi_inputs(allocator.Adapter(kArenaAllocLSE));
1692
1693 ScopedArenaVector<PhiPlaceholder> work_queue(allocator.Adapter(kArenaAllocLSE));
1694 work_queue.push_back(phi_placeholder);
1695 while (!work_queue.empty()) {
1696 PhiPlaceholder current_phi_placeholder = work_queue.back();
1697 if (phi_placeholder_replacements_[PhiPlaceholderIndex(current_phi_placeholder)].IsValid()) {
1698 // This Phi placeholder was pushed to the `work_queue` followed by another Phi placeholder
1699 // that directly or indirectly depends on it, so it was already processed as part of the
1700 // other Phi placeholder's dependencies before this one got back to the top of the stack.
1701 work_queue.pop_back();
1702 continue;
1703 }
1704 uint32_t current_block_id = current_phi_placeholder.GetBlockId();
1705 HBasicBlock* current_block = blocks[current_block_id];
1706 DCHECK_GE(current_block->GetPredecessors().size(), 2u);
1707
1708 // Non-loop Phis cannot depend on a loop Phi, so we should not see any loop header here.
1709 // And the only way for such merged value to reach a different heap location is through
1710 // a load at which point we materialize the Phi. Therefore all non-loop Phi placeholders
1711 // seen here are tied to one heap location.
1712 DCHECK(!current_block->IsLoopHeader())
1713 << current_phi_placeholder << " phase: " << current_phase_;
1714 DCHECK_EQ(current_phi_placeholder.GetHeapLocation(), idx);
1715
1716 phi_inputs.clear();
1717 for (HBasicBlock* predecessor : current_block->GetPredecessors()) {
1718 Value pred_value = ReplacementOrValue(heap_values_for_[predecessor->GetBlockId()][idx].value);
1719 DCHECK(!pred_value.IsPureUnknown()) << pred_value << " block " << current_block->GetBlockId()
1720 << " pred: " << predecessor->GetBlockId();
1721 if (pred_value.NeedsNonLoopPhi() ||
1722 (current_phase_ == Phase::kPartialElimination && pred_value.IsMergedUnknown())) {
1723 // We need to process the Phi placeholder first.
1724 work_queue.push_back(pred_value.GetPhiPlaceholder());
1725 } else if (pred_value.IsDefault()) {
1726 phi_inputs.push_back(GetDefaultValue(type));
1727 } else {
1728 DCHECK(pred_value.IsInstruction()) << pred_value << " block " << current_block->GetBlockId()
1729 << " pred: " << predecessor->GetBlockId();
1730 phi_inputs.push_back(pred_value.GetInstruction());
1731 }
1732 }
1733 if (phi_inputs.size() == current_block->GetPredecessors().size()) {
1734 // All inputs are available. Find or construct the Phi replacement.
1735 phi_placeholder_replacements_[PhiPlaceholderIndex(current_phi_placeholder)] =
1736 Value::ForInstruction(FindOrConstructNonLoopPhi(current_block, phi_inputs, type));
1737 // Remove the block from the queue.
1738 DCHECK_EQ(current_phi_placeholder, work_queue.back());
1739 work_queue.pop_back();
1740 }
1741 }
1742 }
1743
VisitGetLocation(HInstruction * instruction,size_t idx)1744 void LSEVisitor::VisitGetLocation(HInstruction* instruction, size_t idx) {
1745 DCHECK_NE(idx, HeapLocationCollector::kHeapLocationNotFound);
1746 uint32_t block_id = instruction->GetBlock()->GetBlockId();
1747 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[block_id];
1748 ValueRecord& record = heap_values[idx];
1749 if (instruction->IsFieldAccess()) {
1750 RecordFieldInfo(&instruction->GetFieldInfo(), idx);
1751 }
1752 DCHECK(record.value.IsUnknown() || record.value.Equals(ReplacementOrValue(record.value)));
1753 // If we are unknown, we either come from somewhere untracked or we can reconstruct the partial
1754 // value.
1755 DCHECK(!record.value.IsPureUnknown() ||
1756 heap_location_collector_.GetHeapLocation(idx)->GetReferenceInfo() == nullptr ||
1757 !heap_location_collector_.GetHeapLocation(idx)->GetReferenceInfo()->IsPartialSingleton())
1758 << "In " << GetGraph()->PrettyMethod() << ": " << record.value << " for " << *instruction;
1759 intermediate_values_.insert({instruction, record.value});
1760 loads_and_stores_.push_back({ instruction, idx });
1761 if ((record.value.IsDefault() || record.value.NeedsNonLoopPhi()) &&
1762 !IsDefaultOrPhiAllowedForLoad(instruction)) {
1763 record.value = Value::PureUnknown();
1764 }
1765 if (record.value.IsDefault()) {
1766 KeepStores(record.stored_by);
1767 HInstruction* constant = GetDefaultValue(instruction->GetType());
1768 AddRemovedLoad(instruction, constant);
1769 record.value = Value::ForInstruction(constant);
1770 } else if (record.value.IsUnknown()) {
1771 // Load isn't eliminated. Put the load as the value into the HeapLocation.
1772 // This acts like GVN but with better aliasing analysis.
1773 Value old_value = record.value;
1774 record.value = Value::ForInstruction(instruction);
1775 KeepStoresIfAliasedToLocation(heap_values, idx);
1776 KeepStores(old_value);
1777 } else if (record.value.NeedsLoopPhi()) {
1778 // We do not know yet if the value is known for all back edges. Record for future processing.
1779 loads_requiring_loop_phi_.insert(std::make_pair(instruction, record));
1780 } else {
1781 // This load can be eliminated but we may need to construct non-loop Phis.
1782 if (record.value.NeedsNonLoopPhi()) {
1783 MaterializeNonLoopPhis(record.value.GetPhiPlaceholder(), instruction->GetType());
1784 record.value = Replacement(record.value);
1785 }
1786 HInstruction* heap_value = FindSubstitute(record.value.GetInstruction());
1787 AddRemovedLoad(instruction, heap_value);
1788 TryRemovingNullCheck(instruction);
1789 }
1790 }
1791
VisitSetLocation(HInstruction * instruction,size_t idx,HInstruction * value)1792 void LSEVisitor::VisitSetLocation(HInstruction* instruction, size_t idx, HInstruction* value) {
1793 DCHECK_NE(idx, HeapLocationCollector::kHeapLocationNotFound);
1794 DCHECK(!IsStore(value)) << value->DebugName();
1795 if (instruction->IsFieldAccess()) {
1796 RecordFieldInfo(&instruction->GetFieldInfo(), idx);
1797 }
1798 // value may already have a substitute.
1799 value = FindSubstitute(value);
1800 HBasicBlock* block = instruction->GetBlock();
1801 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[block->GetBlockId()];
1802 ValueRecord& record = heap_values[idx];
1803 DCHECK(!record.value.IsInstruction() ||
1804 FindSubstitute(record.value.GetInstruction()) == record.value.GetInstruction());
1805
1806 if (record.value.Equals(value)) {
1807 // Store into the heap location with the same value.
1808 // This store can be eliminated right away.
1809 block->RemoveInstruction(instruction);
1810 return;
1811 }
1812
1813 store_records_.insert(std::make_pair(instruction, StoreRecord{record, value}));
1814 loads_and_stores_.push_back({ instruction, idx });
1815
1816 // If the `record.stored_by` specified a store from this block, it shall be removed
1817 // at the end, except for throwing ArraySet; it cannot be marked for keeping in
1818 // `kept_stores_` anymore after we update the `record.stored_by` below.
1819 DCHECK(!record.stored_by.IsInstruction() ||
1820 record.stored_by.GetInstruction()->GetBlock() != block ||
1821 record.stored_by.GetInstruction()->CanThrow() ||
1822 !kept_stores_.IsBitSet(record.stored_by.GetInstruction()->GetId()));
1823
1824 if (instruction->CanThrow()) {
1825 // Previous stores can become visible.
1826 HandleExit(instruction->GetBlock());
1827 // We cannot remove a possibly throwing store.
1828 // After marking it as kept, it does not matter if we track it in `stored_by` or not.
1829 kept_stores_.SetBit(instruction->GetId());
1830 }
1831
1832 // Update the record.
1833 auto it = loads_requiring_loop_phi_.find(value);
1834 if (it != loads_requiring_loop_phi_.end()) {
1835 // Propapate the Phi placeholder to the record.
1836 record.value = it->second.value;
1837 DCHECK(record.value.NeedsLoopPhi());
1838 } else {
1839 record.value = Value::ForInstruction(value);
1840 }
1841 // Track the store in the value record. If the value is loaded or needed after
1842 // return/deoptimization later, this store isn't really redundant.
1843 record.stored_by = Value::ForInstruction(instruction);
1844
1845 // This store may kill values in other heap locations due to aliasing.
1846 for (size_t i = 0u, size = heap_values.size(); i != size; ++i) {
1847 if (i == idx ||
1848 heap_values[i].value.IsUnknown() ||
1849 CanValueBeKeptIfSameAsNew(heap_values[i].value, value, instruction) ||
1850 !heap_location_collector_.MayAlias(i, idx)) {
1851 continue;
1852 }
1853 // Kill heap locations that may alias and keep previous stores to these locations.
1854 KeepStores(heap_values[i].stored_by);
1855 heap_values[i].stored_by = Value::PureUnknown();
1856 heap_values[i].value = Value::PartialUnknown(heap_values[i].value);
1857 }
1858 }
1859
VisitBasicBlock(HBasicBlock * block)1860 void LSEVisitor::VisitBasicBlock(HBasicBlock* block) {
1861 // Populate the heap_values array for this block.
1862 // TODO: try to reuse the heap_values array from one predecessor if possible.
1863 if (block->IsLoopHeader()) {
1864 PrepareLoopRecords(block);
1865 } else {
1866 MergePredecessorRecords(block);
1867 }
1868 // Visit instructions.
1869 HGraphVisitor::VisitBasicBlock(block);
1870 }
1871
MayAliasOnBackEdge(HBasicBlock * loop_header,size_t idx1,size_t idx2) const1872 bool LSEVisitor::MayAliasOnBackEdge(HBasicBlock* loop_header, size_t idx1, size_t idx2) const {
1873 DCHECK_NE(idx1, idx2);
1874 DCHECK(loop_header->IsLoopHeader());
1875 if (heap_location_collector_.MayAlias(idx1, idx2)) {
1876 return true;
1877 }
1878 // For array locations with index defined inside the loop, include
1879 // all other locations in the array, even those that LSA declares
1880 // non-aliasing, such as `a[i]` and `a[i + 1]`, as they may actually
1881 // refer to the same locations for different iterations. (LSA's
1882 // `ComputeMayAlias()` does not consider different loop iterations.)
1883 HeapLocation* loc1 = heap_location_collector_.GetHeapLocation(idx1);
1884 HeapLocation* loc2 = heap_location_collector_.GetHeapLocation(idx2);
1885 if (loc1->IsArray() &&
1886 loc2->IsArray() &&
1887 HeapLocationCollector::CanReferencesAlias(loc1->GetReferenceInfo(),
1888 loc2->GetReferenceInfo())) {
1889 HLoopInformation* loop_info = loop_header->GetLoopInformation();
1890 if (loop_info->Contains(*loc1->GetIndex()->GetBlock()) ||
1891 loop_info->Contains(*loc2->GetIndex()->GetBlock())) {
1892 // Consider the locations aliasing. Do not optimize the case where both indexes
1893 // are loop invariants defined inside the loop, rely on LICM to pull them out.
1894 return true;
1895 }
1896 }
1897 return false;
1898 }
1899
TryReplacingLoopPhiPlaceholderWithDefault(PhiPlaceholder phi_placeholder,DataType::Type type,ArenaBitVector * phi_placeholders_to_materialize)1900 bool LSEVisitor::TryReplacingLoopPhiPlaceholderWithDefault(
1901 PhiPlaceholder phi_placeholder,
1902 DataType::Type type,
1903 /*inout*/ ArenaBitVector* phi_placeholders_to_materialize) {
1904 // Use local allocator to reduce peak memory usage.
1905 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
1906 ArenaBitVector visited(&allocator,
1907 /*start_bits=*/ num_phi_placeholders_,
1908 /*expandable=*/ false,
1909 kArenaAllocLSE);
1910 visited.ClearAllBits();
1911 ScopedArenaVector<PhiPlaceholder> work_queue(allocator.Adapter(kArenaAllocLSE));
1912
1913 // Use depth first search to check if any non-Phi input is unknown.
1914 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
1915 size_t num_heap_locations = heap_location_collector_.GetNumberOfHeapLocations();
1916 visited.SetBit(PhiPlaceholderIndex(phi_placeholder));
1917 work_queue.push_back(phi_placeholder);
1918 while (!work_queue.empty()) {
1919 PhiPlaceholder current_phi_placeholder = work_queue.back();
1920 work_queue.pop_back();
1921 HBasicBlock* block = blocks[current_phi_placeholder.GetBlockId()];
1922 DCHECK_GE(block->GetPredecessors().size(), 2u);
1923 size_t idx = current_phi_placeholder.GetHeapLocation();
1924 for (HBasicBlock* predecessor : block->GetPredecessors()) {
1925 Value value = ReplacementOrValue(heap_values_for_[predecessor->GetBlockId()][idx].value);
1926 if (value.NeedsPhi()) {
1927 // Visit the predecessor Phi placeholder if it's not visited yet.
1928 if (!visited.IsBitSet(PhiPlaceholderIndex(value))) {
1929 visited.SetBit(PhiPlaceholderIndex(value));
1930 work_queue.push_back(value.GetPhiPlaceholder());
1931 }
1932 } else if (!value.Equals(Value::Default())) {
1933 return false; // Report failure.
1934 }
1935 }
1936 if (block->IsLoopHeader()) {
1937 // For back-edges we need to check all locations that write to the same array,
1938 // even those that LSA declares non-aliasing, such as `a[i]` and `a[i + 1]`
1939 // as they may actually refer to the same locations for different iterations.
1940 for (size_t i = 0; i != num_heap_locations; ++i) {
1941 if (i == idx ||
1942 heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo() !=
1943 heap_location_collector_.GetHeapLocation(idx)->GetReferenceInfo()) {
1944 continue;
1945 }
1946 for (HBasicBlock* predecessor : block->GetPredecessors()) {
1947 // Check if there were any writes to this location.
1948 // Note: We could simply process the values but due to the vector operation
1949 // carve-out (see `IsDefaultOrPhiAllowedForLoad()`), a vector load can cause
1950 // the value to change and not be equal to default. To work around this and
1951 // allow replacing the non-vector load of loop-invariant default values
1952 // anyway, skip over paths that do not have any writes.
1953 ValueRecord record = heap_values_for_[predecessor->GetBlockId()][i];
1954 while (record.stored_by.NeedsLoopPhi() &&
1955 blocks[record.stored_by.GetPhiPlaceholder().GetBlockId()]->IsLoopHeader()) {
1956 HLoopInformation* loop_info =
1957 blocks[record.stored_by.GetPhiPlaceholder().GetBlockId()]->GetLoopInformation();
1958 record = heap_values_for_[loop_info->GetPreHeader()->GetBlockId()][i];
1959 }
1960 Value value = ReplacementOrValue(record.value);
1961 if (value.NeedsPhi()) {
1962 // Visit the predecessor Phi placeholder if it's not visited yet.
1963 if (!visited.IsBitSet(PhiPlaceholderIndex(value))) {
1964 visited.SetBit(PhiPlaceholderIndex(value));
1965 work_queue.push_back(value.GetPhiPlaceholder());
1966 }
1967 } else if (!value.Equals(Value::Default())) {
1968 return false; // Report failure.
1969 }
1970 }
1971 }
1972 }
1973 }
1974
1975 // Record replacement and report success.
1976 HInstruction* replacement = GetDefaultValue(type);
1977 for (uint32_t phi_placeholder_index : visited.Indexes()) {
1978 DCHECK(phi_placeholder_replacements_[phi_placeholder_index].IsInvalid());
1979 phi_placeholder_replacements_[phi_placeholder_index] = Value::ForInstruction(replacement);
1980 }
1981 phi_placeholders_to_materialize->Subtract(&visited);
1982 return true;
1983 }
1984
TryReplacingLoopPhiPlaceholderWithSingleInput(PhiPlaceholder phi_placeholder,ArenaBitVector * phi_placeholders_to_materialize)1985 bool LSEVisitor::TryReplacingLoopPhiPlaceholderWithSingleInput(
1986 PhiPlaceholder phi_placeholder,
1987 /*inout*/ ArenaBitVector* phi_placeholders_to_materialize) {
1988 // Use local allocator to reduce peak memory usage.
1989 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
1990 ArenaBitVector visited(&allocator,
1991 /*start_bits=*/ num_phi_placeholders_,
1992 /*expandable=*/ false,
1993 kArenaAllocLSE);
1994 visited.ClearAllBits();
1995 ScopedArenaVector<PhiPlaceholder> work_queue(allocator.Adapter(kArenaAllocLSE));
1996
1997 // Use depth first search to check if any non-Phi input is unknown.
1998 HInstruction* replacement = nullptr;
1999 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
2000 visited.SetBit(PhiPlaceholderIndex(phi_placeholder));
2001 work_queue.push_back(phi_placeholder);
2002 while (!work_queue.empty()) {
2003 PhiPlaceholder current_phi_placeholder = work_queue.back();
2004 work_queue.pop_back();
2005 HBasicBlock* current_block = blocks[current_phi_placeholder.GetBlockId()];
2006 DCHECK_GE(current_block->GetPredecessors().size(), 2u);
2007 size_t idx = current_phi_placeholder.GetHeapLocation();
2008 for (HBasicBlock* predecessor : current_block->GetPredecessors()) {
2009 Value value = ReplacementOrValue(heap_values_for_[predecessor->GetBlockId()][idx].value);
2010 if (value.NeedsPhi()) {
2011 // Visit the predecessor Phi placeholder if it's not visited yet.
2012 if (!visited.IsBitSet(PhiPlaceholderIndex(value))) {
2013 visited.SetBit(PhiPlaceholderIndex(value));
2014 work_queue.push_back(value.GetPhiPlaceholder());
2015 }
2016 } else {
2017 if (!value.IsInstruction() ||
2018 (replacement != nullptr && replacement != value.GetInstruction())) {
2019 return false; // Report failure.
2020 }
2021 replacement = value.GetInstruction();
2022 }
2023 }
2024 // While `TryReplacingLoopPhiPlaceholderWithDefault()` has special treatment
2025 // for back-edges, it is not needed here. When looking for a single input
2026 // instruction coming from before the loop, the array index must also be
2027 // defined before the loop and the aliasing analysis done by LSA is sufficient.
2028 // Any writes of a different value with an index that is not loop invariant
2029 // would invalidate the heap location in `VisitSetLocation()`.
2030 }
2031
2032 // Record replacement and report success.
2033 DCHECK(replacement != nullptr);
2034 for (uint32_t phi_placeholder_index : visited.Indexes()) {
2035 DCHECK(phi_placeholder_replacements_[phi_placeholder_index].IsInvalid());
2036 phi_placeholder_replacements_[phi_placeholder_index] = Value::ForInstruction(replacement);
2037 }
2038 phi_placeholders_to_materialize->Subtract(&visited);
2039 return true;
2040 }
2041
FindLoopPhisToMaterialize(PhiPlaceholder phi_placeholder,ArenaBitVector * phi_placeholders_to_materialize,DataType::Type type,bool can_use_default_or_phi)2042 std::optional<LSEVisitor::PhiPlaceholder> LSEVisitor::FindLoopPhisToMaterialize(
2043 PhiPlaceholder phi_placeholder,
2044 /*inout*/ ArenaBitVector* phi_placeholders_to_materialize,
2045 DataType::Type type,
2046 bool can_use_default_or_phi) {
2047 DCHECK(phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)].IsInvalid());
2048
2049 // Use local allocator to reduce peak memory usage.
2050 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
2051 ScopedArenaVector<PhiPlaceholder> work_queue(allocator.Adapter(kArenaAllocLSE));
2052
2053 // Use depth first search to check if any non-Phi input is unknown.
2054 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
2055 phi_placeholders_to_materialize->ClearAllBits();
2056 phi_placeholders_to_materialize->SetBit(PhiPlaceholderIndex(phi_placeholder));
2057 work_queue.push_back(phi_placeholder);
2058 while (!work_queue.empty()) {
2059 PhiPlaceholder current_phi_placeholder = work_queue.back();
2060 work_queue.pop_back();
2061 if (!phi_placeholders_to_materialize->IsBitSet(PhiPlaceholderIndex(current_phi_placeholder))) {
2062 // Replaced by `TryReplacingLoopPhiPlaceholderWith{Default,SingleInput}()`.
2063 DCHECK(phi_placeholder_replacements_[PhiPlaceholderIndex(current_phi_placeholder)].Equals(
2064 Value::Default()));
2065 continue;
2066 }
2067 HBasicBlock* current_block = blocks[current_phi_placeholder.GetBlockId()];
2068 DCHECK_GE(current_block->GetPredecessors().size(), 2u);
2069 size_t idx = current_phi_placeholder.GetHeapLocation();
2070 if (current_block->IsLoopHeader()) {
2071 // If the index is defined inside the loop, it may reference different elements of the
2072 // array on each iteration. Since we do not track if all elements of an array are set
2073 // to the same value explicitly, the only known value in pre-header can be the default
2074 // value from NewArray or a Phi placeholder depending on a default value from some outer
2075 // loop pre-header. This Phi placeholder can be replaced only by the default value.
2076 HInstruction* index = heap_location_collector_.GetHeapLocation(idx)->GetIndex();
2077 if (index != nullptr && current_block->GetLoopInformation()->Contains(*index->GetBlock())) {
2078 if (can_use_default_or_phi &&
2079 TryReplacingLoopPhiPlaceholderWithDefault(current_phi_placeholder,
2080 type,
2081 phi_placeholders_to_materialize)) {
2082 continue;
2083 } else {
2084 return current_phi_placeholder; // Report the loop Phi placeholder.
2085 }
2086 }
2087 // A similar situation arises with the index defined outside the loop if we cannot use
2088 // default values or Phis, i.e. for vector loads, as we can only replace the Phi
2089 // placeholder with a single instruction defined before the loop.
2090 if (!can_use_default_or_phi) {
2091 DCHECK(index != nullptr); // Vector operations are array operations.
2092 if (TryReplacingLoopPhiPlaceholderWithSingleInput(current_phi_placeholder,
2093 phi_placeholders_to_materialize)) {
2094 continue;
2095 } else {
2096 return current_phi_placeholder; // Report the loop Phi placeholder.
2097 }
2098 }
2099 }
2100 for (HBasicBlock* predecessor : current_block->GetPredecessors()) {
2101 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[predecessor->GetBlockId()];
2102 Value value = ReplacementOrValue(heap_values[idx].value);
2103 if (value.IsUnknown()) {
2104 // We cannot create a Phi for this loop Phi placeholder.
2105 return current_phi_placeholder; // Report the loop Phi placeholder.
2106 }
2107 // For arrays, the location may have been clobbered by writes to other locations
2108 // in a loop that LSA does not consider aliasing, such as `a[i]` and `a[i + 1]`.
2109 if (current_block->IsLoopHeader() &&
2110 predecessor != current_block->GetLoopInformation()->GetPreHeader() &&
2111 heap_location_collector_.GetHeapLocation(idx)->GetIndex() != nullptr) {
2112 for (size_t i = 0, size = heap_values.size(); i != size; ++i) {
2113 if (i != idx &&
2114 !heap_values[i].stored_by.IsUnknown() &&
2115 MayAliasOnBackEdge(current_block, idx, i)) {
2116 // We cannot create a Phi for this loop Phi placeholder.
2117 return current_phi_placeholder;
2118 }
2119 }
2120 }
2121 if (value.NeedsLoopPhi()) {
2122 // Visit the predecessor Phi placeholder if it's not visited yet.
2123 if (!phi_placeholders_to_materialize->IsBitSet(PhiPlaceholderIndex(value))) {
2124 phi_placeholders_to_materialize->SetBit(PhiPlaceholderIndex(value));
2125 work_queue.push_back(value.GetPhiPlaceholder());
2126 LSE_VLOG << "For materialization of " << phi_placeholder
2127 << " we need to materialize " << value;
2128 }
2129 }
2130 }
2131 }
2132
2133 // There are no unknown values feeding this Phi, so we can construct the Phis if needed.
2134 return std::nullopt;
2135 }
2136
MaterializeLoopPhis(const ScopedArenaVector<size_t> & phi_placeholder_indexes,DataType::Type type)2137 bool LSEVisitor::MaterializeLoopPhis(const ScopedArenaVector<size_t>& phi_placeholder_indexes,
2138 DataType::Type type) {
2139 return MaterializeLoopPhis(ArrayRef<const size_t>(phi_placeholder_indexes), type);
2140 }
2141
MaterializeLoopPhis(ArrayRef<const size_t> phi_placeholder_indexes,DataType::Type type)2142 bool LSEVisitor::MaterializeLoopPhis(ArrayRef<const size_t> phi_placeholder_indexes,
2143 DataType::Type type) {
2144 // Materialize all predecessors that do not need a loop Phi and determine if all inputs
2145 // other than loop Phis are the same.
2146 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
2147 std::optional<Value> other_value = std::nullopt;
2148 for (size_t phi_placeholder_index : phi_placeholder_indexes) {
2149 PhiPlaceholder phi_placeholder = GetPhiPlaceholderAt(phi_placeholder_index);
2150 HBasicBlock* block = blocks[phi_placeholder.GetBlockId()];
2151 DCHECK_GE(block->GetPredecessors().size(), 2u);
2152 size_t idx = phi_placeholder.GetHeapLocation();
2153 for (HBasicBlock* predecessor : block->GetPredecessors()) {
2154 Value value = ReplacementOrValue(heap_values_for_[predecessor->GetBlockId()][idx].value);
2155 if (value.NeedsNonLoopPhi()) {
2156 DCHECK(current_phase_ == Phase::kLoadElimination ||
2157 current_phase_ == Phase::kPartialElimination)
2158 << current_phase_;
2159 MaterializeNonLoopPhis(value.GetPhiPlaceholder(), type);
2160 value = Replacement(value);
2161 }
2162 if (!value.NeedsLoopPhi()) {
2163 if (!other_value) {
2164 // The first other value we found.
2165 other_value = value;
2166 } else if (!other_value->IsInvalid()) {
2167 // Check if the current `value` differs from the previous `other_value`.
2168 if (!value.Equals(*other_value)) {
2169 other_value = Value::Invalid();
2170 }
2171 }
2172 }
2173 }
2174 }
2175
2176 DCHECK(other_value.has_value());
2177 if (!other_value->IsInvalid()) {
2178 HInstruction* replacement =
2179 (other_value->IsDefault()) ? GetDefaultValue(type) : other_value->GetInstruction();
2180 for (size_t phi_placeholder_index : phi_placeholder_indexes) {
2181 phi_placeholder_replacements_[phi_placeholder_index] = Value::ForInstruction(replacement);
2182 }
2183 return true;
2184 }
2185
2186 // If we're materializing only a single Phi, try to match it with an existing Phi.
2187 // (Matching multiple Phis would need investigation. It may be prohibitively slow.)
2188 // This also covers the case when after replacing a previous set of Phi placeholders,
2189 // we continue with a Phi placeholder that does not really need a loop Phi anymore.
2190 if (phi_placeholder_indexes.size() == 1u) {
2191 PhiPlaceholder phi_placeholder = GetPhiPlaceholderAt(phi_placeholder_indexes[0]);
2192 size_t idx = phi_placeholder.GetHeapLocation();
2193 HBasicBlock* block = GetGraph()->GetBlocks()[phi_placeholder.GetBlockId()];
2194 ArrayRef<HBasicBlock* const> predecessors(block->GetPredecessors());
2195 for (HInstructionIterator phi_it(block->GetPhis()); !phi_it.Done(); phi_it.Advance()) {
2196 HInstruction* phi = phi_it.Current();
2197 DCHECK_EQ(phi->InputCount(), predecessors.size());
2198 ArrayRef<HUserRecord<HInstruction*>> phi_inputs = phi->GetInputRecords();
2199 auto cmp = [=](const HUserRecord<HInstruction*>& lhs, HBasicBlock* rhs) {
2200 Value value = ReplacementOrValue(heap_values_for_[rhs->GetBlockId()][idx].value);
2201 if (value.NeedsPhi()) {
2202 DCHECK(value.GetPhiPlaceholder() == phi_placeholder);
2203 return lhs.GetInstruction() == phi;
2204 } else {
2205 DCHECK(value.IsDefault() || value.IsInstruction());
2206 return value.Equals(lhs.GetInstruction());
2207 }
2208 };
2209 if (std::equal(phi_inputs.begin(), phi_inputs.end(), predecessors.begin(), cmp)) {
2210 phi_placeholder_replacements_[phi_placeholder_indexes[0]] = Value::ForInstruction(phi);
2211 return true;
2212 }
2213 }
2214 }
2215
2216 if (current_phase_ == Phase::kStoreElimination) {
2217 // We're not creating Phis during the final store elimination phase.
2218 return false;
2219 }
2220
2221 // There are different inputs to the Phi chain. Create the Phis.
2222 ArenaAllocator* allocator = GetGraph()->GetAllocator();
2223 for (size_t phi_placeholder_index : phi_placeholder_indexes) {
2224 PhiPlaceholder phi_placeholder = GetPhiPlaceholderAt(phi_placeholder_index);
2225 HBasicBlock* block = blocks[phi_placeholder.GetBlockId()];
2226 CHECK_GE(block->GetPredecessors().size(), 2u);
2227 phi_placeholder_replacements_[phi_placeholder_index] = Value::ForInstruction(
2228 new (allocator) HPhi(allocator, kNoRegNumber, block->GetPredecessors().size(), type));
2229 }
2230 // Fill the Phi inputs.
2231 for (size_t phi_placeholder_index : phi_placeholder_indexes) {
2232 PhiPlaceholder phi_placeholder = GetPhiPlaceholderAt(phi_placeholder_index);
2233 HBasicBlock* block = blocks[phi_placeholder.GetBlockId()];
2234 size_t idx = phi_placeholder.GetHeapLocation();
2235 HInstruction* phi = phi_placeholder_replacements_[phi_placeholder_index].GetInstruction();
2236 DCHECK(DataType::IsTypeConversionImplicit(type, phi->GetType()))
2237 << "type=" << type << " vs phi-type=" << phi->GetType();
2238 for (size_t i = 0, size = block->GetPredecessors().size(); i != size; ++i) {
2239 HBasicBlock* predecessor = block->GetPredecessors()[i];
2240 Value value = ReplacementOrValue(heap_values_for_[predecessor->GetBlockId()][idx].value);
2241 HInstruction* input = value.IsDefault() ? GetDefaultValue(type) : value.GetInstruction();
2242 DCHECK_NE(input->GetType(), DataType::Type::kVoid);
2243 phi->SetRawInputAt(i, input);
2244 DCHECK(DataType::IsTypeConversionImplicit(input->GetType(), phi->GetType()))
2245 << " input: " << input->GetType() << value << " phi: " << phi->GetType()
2246 << " request: " << type;
2247 }
2248 }
2249 // Add the Phis to their blocks.
2250 for (size_t phi_placeholder_index : phi_placeholder_indexes) {
2251 PhiPlaceholder phi_placeholder = GetPhiPlaceholderAt(phi_placeholder_index);
2252 HBasicBlock* block = blocks[phi_placeholder.GetBlockId()];
2253 block->AddPhi(phi_placeholder_replacements_[phi_placeholder_index].GetInstruction()->AsPhi());
2254 }
2255 if (type == DataType::Type::kReference) {
2256 ScopedArenaAllocator local_allocator(allocator_.GetArenaStack());
2257 ScopedArenaVector<HInstruction*> phis(local_allocator.Adapter(kArenaAllocLSE));
2258 for (size_t phi_placeholder_index : phi_placeholder_indexes) {
2259 phis.push_back(phi_placeholder_replacements_[phi_placeholder_index].GetInstruction());
2260 }
2261 // Update reference type information. Pass invalid handles, these are not used for Phis.
2262 ReferenceTypePropagation rtp_fixup(GetGraph(),
2263 Handle<mirror::ClassLoader>(),
2264 Handle<mirror::DexCache>(),
2265 /* is_first_run= */ false);
2266 rtp_fixup.Visit(ArrayRef<HInstruction* const>(phis));
2267 }
2268
2269 return true;
2270 }
2271
MaterializeLoopPhis(const ArenaBitVector & phi_placeholders_to_materialize,DataType::Type type)2272 bool LSEVisitor::MaterializeLoopPhis(const ArenaBitVector& phi_placeholders_to_materialize,
2273 DataType::Type type) {
2274 // Use local allocator to reduce peak memory usage.
2275 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
2276
2277 // We want to recognize when a subset of these loop Phis that do not need other
2278 // loop Phis, i.e. a transitive closure, has only one other instruction as an input,
2279 // i.e. that instruction can be used instead of each Phi in the set. See for example
2280 // Main.testLoop{5,6,7,8}() in the test 530-checker-lse. To do that, we shall
2281 // materialize these loop Phis from the smallest transitive closure.
2282
2283 // Construct a matrix of loop phi placeholder dependencies. To reduce the memory usage,
2284 // assign new indexes to the Phi placeholders, making the matrix dense.
2285 ScopedArenaVector<size_t> matrix_indexes(num_phi_placeholders_,
2286 static_cast<size_t>(-1), // Invalid.
2287 allocator.Adapter(kArenaAllocLSE));
2288 ScopedArenaVector<size_t> phi_placeholder_indexes(allocator.Adapter(kArenaAllocLSE));
2289 size_t num_phi_placeholders = phi_placeholders_to_materialize.NumSetBits();
2290 phi_placeholder_indexes.reserve(num_phi_placeholders);
2291 for (uint32_t marker_index : phi_placeholders_to_materialize.Indexes()) {
2292 matrix_indexes[marker_index] = phi_placeholder_indexes.size();
2293 phi_placeholder_indexes.push_back(marker_index);
2294 }
2295 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
2296 ScopedArenaVector<ArenaBitVector*> dependencies(allocator.Adapter(kArenaAllocLSE));
2297 dependencies.reserve(num_phi_placeholders);
2298 for (size_t matrix_index = 0; matrix_index != num_phi_placeholders; ++matrix_index) {
2299 static constexpr bool kExpandable = false;
2300 dependencies.push_back(
2301 ArenaBitVector::Create(&allocator, num_phi_placeholders, kExpandable, kArenaAllocLSE));
2302 ArenaBitVector* current_dependencies = dependencies.back();
2303 current_dependencies->ClearAllBits();
2304 current_dependencies->SetBit(matrix_index); // Count the Phi placeholder as its own dependency.
2305 PhiPlaceholder current_phi_placeholder =
2306 GetPhiPlaceholderAt(phi_placeholder_indexes[matrix_index]);
2307 HBasicBlock* current_block = blocks[current_phi_placeholder.GetBlockId()];
2308 DCHECK_GE(current_block->GetPredecessors().size(), 2u);
2309 size_t idx = current_phi_placeholder.GetHeapLocation();
2310 for (HBasicBlock* predecessor : current_block->GetPredecessors()) {
2311 Value pred_value = ReplacementOrValue(heap_values_for_[predecessor->GetBlockId()][idx].value);
2312 if (pred_value.NeedsLoopPhi()) {
2313 size_t pred_value_index = PhiPlaceholderIndex(pred_value);
2314 DCHECK(phi_placeholder_replacements_[pred_value_index].IsInvalid());
2315 DCHECK_NE(matrix_indexes[pred_value_index], static_cast<size_t>(-1));
2316 current_dependencies->SetBit(matrix_indexes[PhiPlaceholderIndex(pred_value)]);
2317 }
2318 }
2319 }
2320
2321 // Use the Floyd-Warshall algorithm to determine all transitive dependencies.
2322 for (size_t k = 0; k != num_phi_placeholders; ++k) {
2323 for (size_t i = 0; i != num_phi_placeholders; ++i) {
2324 for (size_t j = 0; j != num_phi_placeholders; ++j) {
2325 if (dependencies[i]->IsBitSet(k) && dependencies[k]->IsBitSet(j)) {
2326 dependencies[i]->SetBit(j);
2327 }
2328 }
2329 }
2330 }
2331
2332 // Count the number of transitive dependencies for each replaceable Phi placeholder.
2333 ScopedArenaVector<size_t> num_dependencies(allocator.Adapter(kArenaAllocLSE));
2334 num_dependencies.reserve(num_phi_placeholders);
2335 for (size_t matrix_index = 0; matrix_index != num_phi_placeholders; ++matrix_index) {
2336 num_dependencies.push_back(dependencies[matrix_index]->NumSetBits());
2337 }
2338
2339 // Pick a Phi placeholder with the smallest number of transitive dependencies and
2340 // materialize it and its dependencies. Repeat until we have materialized all.
2341 ScopedArenaVector<size_t> current_subset(allocator.Adapter(kArenaAllocLSE));
2342 current_subset.reserve(num_phi_placeholders);
2343 size_t remaining_phi_placeholders = num_phi_placeholders;
2344 while (remaining_phi_placeholders != 0u) {
2345 auto it = std::min_element(num_dependencies.begin(), num_dependencies.end());
2346 DCHECK_LE(*it, remaining_phi_placeholders);
2347 size_t current_matrix_index = std::distance(num_dependencies.begin(), it);
2348 ArenaBitVector* current_dependencies = dependencies[current_matrix_index];
2349 size_t current_num_dependencies = num_dependencies[current_matrix_index];
2350 current_subset.clear();
2351 for (uint32_t matrix_index : current_dependencies->Indexes()) {
2352 current_subset.push_back(phi_placeholder_indexes[matrix_index]);
2353 }
2354 if (!MaterializeLoopPhis(current_subset, type)) {
2355 DCHECK_EQ(current_phase_, Phase::kStoreElimination);
2356 // This is the final store elimination phase and we shall not be able to eliminate any
2357 // stores that depend on the current subset, so mark these Phi placeholders unreplaceable.
2358 for (uint32_t matrix_index = 0; matrix_index != num_phi_placeholders; ++matrix_index) {
2359 if (dependencies[matrix_index]->IsBitSet(current_matrix_index)) {
2360 DCHECK(phi_placeholder_replacements_[phi_placeholder_indexes[matrix_index]].IsInvalid());
2361 phi_placeholder_replacements_[phi_placeholder_indexes[matrix_index]] =
2362 Value::PureUnknown();
2363 }
2364 }
2365 return false;
2366 }
2367 for (uint32_t matrix_index = 0; matrix_index != num_phi_placeholders; ++matrix_index) {
2368 if (current_dependencies->IsBitSet(matrix_index)) {
2369 // Mark all dependencies as done by incrementing their `num_dependencies[.]`,
2370 // so that they shall never be the minimum again.
2371 num_dependencies[matrix_index] = num_phi_placeholders;
2372 } else if (dependencies[matrix_index]->IsBitSet(current_matrix_index)) {
2373 // Remove dependencies from other Phi placeholders.
2374 dependencies[matrix_index]->Subtract(current_dependencies);
2375 num_dependencies[matrix_index] -= current_num_dependencies;
2376 }
2377 }
2378 remaining_phi_placeholders -= current_num_dependencies;
2379 }
2380 return true;
2381 }
2382
FullyMaterializePhi(PhiPlaceholder phi_placeholder,DataType::Type type)2383 bool LSEVisitor::FullyMaterializePhi(PhiPlaceholder phi_placeholder, DataType::Type type) {
2384 ScopedArenaAllocator saa(GetGraph()->GetArenaStack());
2385 ArenaBitVector abv(&saa, num_phi_placeholders_, false, ArenaAllocKind::kArenaAllocLSE);
2386 auto res =
2387 FindLoopPhisToMaterialize(phi_placeholder, &abv, type, /* can_use_default_or_phi=*/true);
2388 CHECK(!res.has_value()) << *res;
2389 return MaterializeLoopPhis(abv, type);
2390 }
2391
TryToMaterializeLoopPhis(PhiPlaceholder phi_placeholder,HInstruction * load)2392 std::optional<LSEVisitor::PhiPlaceholder> LSEVisitor::TryToMaterializeLoopPhis(
2393 PhiPlaceholder phi_placeholder, HInstruction* load) {
2394 DCHECK(phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)].IsInvalid());
2395
2396 // Use local allocator to reduce peak memory usage.
2397 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
2398
2399 // Find Phi placeholders to materialize.
2400 ArenaBitVector phi_placeholders_to_materialize(
2401 &allocator, num_phi_placeholders_, /*expandable=*/ false, kArenaAllocLSE);
2402 phi_placeholders_to_materialize.ClearAllBits();
2403 DataType::Type type = load->GetType();
2404 bool can_use_default_or_phi = IsDefaultOrPhiAllowedForLoad(load);
2405 std::optional<PhiPlaceholder> loop_phi_with_unknown_input = FindLoopPhisToMaterialize(
2406 phi_placeholder, &phi_placeholders_to_materialize, type, can_use_default_or_phi);
2407 if (loop_phi_with_unknown_input) {
2408 DCHECK_GE(GetGraph()
2409 ->GetBlocks()[loop_phi_with_unknown_input->GetBlockId()]
2410 ->GetPredecessors()
2411 .size(),
2412 2u);
2413 return loop_phi_with_unknown_input; // Return failure.
2414 }
2415
2416 DCHECK_EQ(current_phase_, Phase::kLoadElimination);
2417 bool success = MaterializeLoopPhis(phi_placeholders_to_materialize, type);
2418 DCHECK(success);
2419
2420 // Report success.
2421 return std::nullopt;
2422 }
2423
2424 // Re-process loads and stores in successors from the `loop_phi_with_unknown_input`. This may
2425 // find one or more loads from `loads_requiring_loop_phi_` which cannot be replaced by Phis and
2426 // propagate the load(s) as the new value(s) to successors; this may uncover new elimination
2427 // opportunities. If we find no such load, we shall at least propagate an unknown value to some
2428 // heap location that is needed by another loop Phi placeholder.
ProcessLoopPhiWithUnknownInput(PhiPlaceholder loop_phi_with_unknown_input)2429 void LSEVisitor::ProcessLoopPhiWithUnknownInput(PhiPlaceholder loop_phi_with_unknown_input) {
2430 size_t loop_phi_with_unknown_input_index = PhiPlaceholderIndex(loop_phi_with_unknown_input);
2431 DCHECK(phi_placeholder_replacements_[loop_phi_with_unknown_input_index].IsInvalid());
2432 phi_placeholder_replacements_[loop_phi_with_unknown_input_index] =
2433 Value::MergedUnknown(loop_phi_with_unknown_input);
2434
2435 uint32_t block_id = loop_phi_with_unknown_input.GetBlockId();
2436 const ArenaVector<HBasicBlock*> reverse_post_order = GetGraph()->GetReversePostOrder();
2437 size_t reverse_post_order_index = 0;
2438 size_t reverse_post_order_size = reverse_post_order.size();
2439 size_t loads_and_stores_index = 0u;
2440 size_t loads_and_stores_size = loads_and_stores_.size();
2441
2442 // Skip blocks and instructions before the block containing the loop phi with unknown input.
2443 DCHECK_NE(reverse_post_order_index, reverse_post_order_size);
2444 while (reverse_post_order[reverse_post_order_index]->GetBlockId() != block_id) {
2445 HBasicBlock* block = reverse_post_order[reverse_post_order_index];
2446 while (loads_and_stores_index != loads_and_stores_size &&
2447 loads_and_stores_[loads_and_stores_index].load_or_store->GetBlock() == block) {
2448 ++loads_and_stores_index;
2449 }
2450 ++reverse_post_order_index;
2451 DCHECK_NE(reverse_post_order_index, reverse_post_order_size);
2452 }
2453
2454 // Use local allocator to reduce peak memory usage.
2455 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
2456 // Reuse one temporary vector for all remaining blocks.
2457 size_t num_heap_locations = heap_location_collector_.GetNumberOfHeapLocations();
2458 ScopedArenaVector<Value> local_heap_values(allocator.Adapter(kArenaAllocLSE));
2459
2460 auto get_initial_value = [this](HBasicBlock* block, size_t idx) {
2461 Value value;
2462 if (block->IsLoopHeader()) {
2463 if (block->GetLoopInformation()->IsIrreducible()) {
2464 PhiPlaceholder placeholder = GetPhiPlaceholder(block->GetBlockId(), idx);
2465 value = Value::MergedUnknown(placeholder);
2466 } else {
2467 value = PrepareLoopValue(block, idx);
2468 }
2469 } else {
2470 value = MergePredecessorValues(block, idx);
2471 }
2472 DCHECK(value.IsUnknown() || ReplacementOrValue(value).Equals(value));
2473 return value;
2474 };
2475
2476 // Process remaining blocks and instructions.
2477 bool found_unreplaceable_load = false;
2478 bool replaced_heap_value_with_unknown = false;
2479 for (; reverse_post_order_index != reverse_post_order_size; ++reverse_post_order_index) {
2480 HBasicBlock* block = reverse_post_order[reverse_post_order_index];
2481 if (block->IsExitBlock()) {
2482 continue;
2483 }
2484
2485 // We shall reconstruct only the heap values that we need for processing loads and stores.
2486 local_heap_values.clear();
2487 local_heap_values.resize(num_heap_locations, Value::Invalid());
2488
2489 for (; loads_and_stores_index != loads_and_stores_size; ++loads_and_stores_index) {
2490 HInstruction* load_or_store = loads_and_stores_[loads_and_stores_index].load_or_store;
2491 size_t idx = loads_and_stores_[loads_and_stores_index].heap_location_index;
2492 if (load_or_store->GetBlock() != block) {
2493 break; // End of instructions from the current block.
2494 }
2495 bool is_store = load_or_store->GetSideEffects().DoesAnyWrite();
2496 DCHECK_EQ(is_store, IsStore(load_or_store));
2497 HInstruction* stored_value = nullptr;
2498 if (is_store) {
2499 auto it = store_records_.find(load_or_store);
2500 DCHECK(it != store_records_.end());
2501 stored_value = it->second.stored_value;
2502 }
2503 auto it = loads_requiring_loop_phi_.find(
2504 stored_value != nullptr ? stored_value : load_or_store);
2505 if (it == loads_requiring_loop_phi_.end()) {
2506 continue; // This load or store never needed a loop Phi.
2507 }
2508 ValueRecord& record = it->second;
2509 if (is_store) {
2510 // Process the store by updating `local_heap_values[idx]`. The last update shall
2511 // be propagated to the `heap_values[idx].value` if it previously needed a loop Phi
2512 // at the end of the block.
2513 Value replacement = ReplacementOrValue(record.value);
2514 if (replacement.NeedsLoopPhi()) {
2515 // No replacement yet, use the Phi placeholder from the load.
2516 DCHECK(record.value.NeedsLoopPhi());
2517 local_heap_values[idx] = record.value;
2518 } else {
2519 // If the load fetched a known value, use it, otherwise use the load.
2520 local_heap_values[idx] = Value::ForInstruction(
2521 replacement.IsUnknown() ? stored_value : replacement.GetInstruction());
2522 }
2523 } else {
2524 // Process the load unless it has previously been marked unreplaceable.
2525 if (record.value.NeedsLoopPhi()) {
2526 if (local_heap_values[idx].IsInvalid()) {
2527 local_heap_values[idx] = get_initial_value(block, idx);
2528 }
2529 if (local_heap_values[idx].IsUnknown()) {
2530 // This load cannot be replaced. Keep stores that feed the Phi placeholder
2531 // (no aliasing since then, otherwise the Phi placeholder would not have been
2532 // propagated as a value to this load) and store the load as the new heap value.
2533 found_unreplaceable_load = true;
2534 KeepStores(record.value);
2535 record.value = Value::MergedUnknown(record.value.GetPhiPlaceholder());
2536 local_heap_values[idx] = Value::ForInstruction(load_or_store);
2537 } else if (local_heap_values[idx].NeedsLoopPhi()) {
2538 // The load may still be replaced with a Phi later.
2539 DCHECK(local_heap_values[idx].Equals(record.value));
2540 } else {
2541 // This load can be eliminated but we may need to construct non-loop Phis.
2542 if (local_heap_values[idx].NeedsNonLoopPhi()) {
2543 MaterializeNonLoopPhis(local_heap_values[idx].GetPhiPlaceholder(),
2544 load_or_store->GetType());
2545 local_heap_values[idx] = Replacement(local_heap_values[idx]);
2546 }
2547 record.value = local_heap_values[idx];
2548 HInstruction* heap_value = local_heap_values[idx].GetInstruction();
2549 AddRemovedLoad(load_or_store, heap_value);
2550 TryRemovingNullCheck(load_or_store);
2551 }
2552 }
2553 }
2554 }
2555
2556 // All heap values that previously needed a loop Phi at the end of the block
2557 // need to be updated for processing successors.
2558 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[block->GetBlockId()];
2559 for (size_t idx = 0; idx != num_heap_locations; ++idx) {
2560 if (heap_values[idx].value.NeedsLoopPhi()) {
2561 if (local_heap_values[idx].IsValid()) {
2562 heap_values[idx].value = local_heap_values[idx];
2563 } else {
2564 heap_values[idx].value = get_initial_value(block, idx);
2565 }
2566 if (heap_values[idx].value.IsUnknown()) {
2567 replaced_heap_value_with_unknown = true;
2568 }
2569 }
2570 }
2571 }
2572 DCHECK(found_unreplaceable_load || replaced_heap_value_with_unknown);
2573 }
2574
ProcessLoadsRequiringLoopPhis()2575 void LSEVisitor::ProcessLoadsRequiringLoopPhis() {
2576 // Note: The vector operations carve-out (see `IsDefaultOrPhiAllowedForLoad()`) can possibly
2577 // make the result of the processing depend on the order in which we process these loads.
2578 // To make sure the result is deterministic, iterate over `loads_and_stores_` instead of the
2579 // `loads_requiring_loop_phi_` indexed by non-deterministic pointers.
2580 for (const LoadStoreRecord& load_store_record : loads_and_stores_) {
2581 auto it = loads_requiring_loop_phi_.find(load_store_record.load_or_store);
2582 if (it == loads_requiring_loop_phi_.end()) {
2583 continue;
2584 }
2585 HInstruction* load = it->first;
2586 ValueRecord& record = it->second;
2587 while (record.value.NeedsLoopPhi() &&
2588 phi_placeholder_replacements_[PhiPlaceholderIndex(record.value)].IsInvalid()) {
2589 std::optional<PhiPlaceholder> loop_phi_with_unknown_input =
2590 TryToMaterializeLoopPhis(record.value.GetPhiPlaceholder(), load);
2591 DCHECK_EQ(loop_phi_with_unknown_input.has_value(),
2592 phi_placeholder_replacements_[PhiPlaceholderIndex(record.value)].IsInvalid());
2593 if (loop_phi_with_unknown_input) {
2594 DCHECK_GE(GetGraph()
2595 ->GetBlocks()[loop_phi_with_unknown_input->GetBlockId()]
2596 ->GetPredecessors()
2597 .size(),
2598 2u);
2599 ProcessLoopPhiWithUnknownInput(*loop_phi_with_unknown_input);
2600 }
2601 }
2602 // The load could have been marked as unreplaceable (and stores marked for keeping)
2603 // or marked for replacement with an instruction in ProcessLoopPhiWithUnknownInput().
2604 DCHECK(record.value.IsUnknown() || record.value.IsInstruction() || record.value.NeedsLoopPhi());
2605 if (record.value.NeedsLoopPhi()) {
2606 record.value = Replacement(record.value);
2607 HInstruction* heap_value = record.value.GetInstruction();
2608 AddRemovedLoad(load, heap_value);
2609 TryRemovingNullCheck(load);
2610 }
2611 }
2612 }
2613
SearchPhiPlaceholdersForKeptStores()2614 void LSEVisitor::SearchPhiPlaceholdersForKeptStores() {
2615 ScopedArenaVector<uint32_t> work_queue(allocator_.Adapter(kArenaAllocLSE));
2616 size_t start_size = phi_placeholders_to_search_for_kept_stores_.NumSetBits();
2617 work_queue.reserve(((start_size * 3u) + 1u) / 2u); // Reserve 1.5x start size, rounded up.
2618 for (uint32_t index : phi_placeholders_to_search_for_kept_stores_.Indexes()) {
2619 work_queue.push_back(index);
2620 }
2621 const ArenaVector<HBasicBlock*>& blocks = GetGraph()->GetBlocks();
2622 std::optional<ArenaBitVector> not_kept_stores;
2623 if (stats_) {
2624 not_kept_stores.emplace(GetGraph()->GetAllocator(),
2625 kept_stores_.GetBitSizeOf(),
2626 false,
2627 ArenaAllocKind::kArenaAllocLSE);
2628 }
2629 while (!work_queue.empty()) {
2630 uint32_t cur_phi_idx = work_queue.back();
2631 PhiPlaceholder phi_placeholder = GetPhiPlaceholderAt(cur_phi_idx);
2632 // Only writes to partial-escapes need to be specifically kept.
2633 bool is_partial_kept_merged_unknown =
2634 kept_merged_unknowns_.IsBitSet(cur_phi_idx) &&
2635 heap_location_collector_.GetHeapLocation(phi_placeholder.GetHeapLocation())
2636 ->GetReferenceInfo()
2637 ->IsPartialSingleton();
2638 work_queue.pop_back();
2639 size_t idx = phi_placeholder.GetHeapLocation();
2640 HBasicBlock* block = blocks[phi_placeholder.GetBlockId()];
2641 DCHECK(block != nullptr) << cur_phi_idx << " phi: " << phi_placeholder
2642 << " (blocks: " << blocks.size() << ")";
2643 for (HBasicBlock* predecessor : block->GetPredecessors()) {
2644 ScopedArenaVector<ValueRecord>& heap_values = heap_values_for_[predecessor->GetBlockId()];
2645 // For loop back-edges we must also preserve all stores to locations that
2646 // may alias with the location `idx`.
2647 // TODO: Add tests cases around this.
2648 bool is_back_edge =
2649 block->IsLoopHeader() && predecessor != block->GetLoopInformation()->GetPreHeader();
2650 size_t start = is_back_edge ? 0u : idx;
2651 size_t end = is_back_edge ? heap_values.size() : idx + 1u;
2652 for (size_t i = start; i != end; ++i) {
2653 Value stored_by = heap_values[i].stored_by;
2654 if (!stored_by.IsUnknown() && (i == idx || MayAliasOnBackEdge(block, idx, i))) {
2655 if (stored_by.NeedsPhi()) {
2656 size_t phi_placeholder_index = PhiPlaceholderIndex(stored_by);
2657 if (is_partial_kept_merged_unknown) {
2658 // Propagate merged-unknown keep since otherwise this might look
2659 // like a partial escape we can remove.
2660 kept_merged_unknowns_.SetBit(phi_placeholder_index);
2661 }
2662 if (!phi_placeholders_to_search_for_kept_stores_.IsBitSet(phi_placeholder_index)) {
2663 phi_placeholders_to_search_for_kept_stores_.SetBit(phi_placeholder_index);
2664 work_queue.push_back(phi_placeholder_index);
2665 }
2666 } else {
2667 DCHECK(IsStore(stored_by.GetInstruction()));
2668 ReferenceInfo* ri = heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo();
2669 DCHECK(ri != nullptr) << "No heap value for " << stored_by.GetInstruction()->DebugName()
2670 << " id: " << stored_by.GetInstruction()->GetId() << " block: "
2671 << stored_by.GetInstruction()->GetBlock()->GetBlockId();
2672 if (!is_partial_kept_merged_unknown && IsPartialNoEscape(predecessor, idx)) {
2673 if (not_kept_stores) {
2674 not_kept_stores->SetBit(stored_by.GetInstruction()->GetId());
2675 }
2676 } else {
2677 kept_stores_.SetBit(stored_by.GetInstruction()->GetId());
2678 }
2679 }
2680 }
2681 }
2682 }
2683 }
2684 if (not_kept_stores) {
2685 // a - b := (a & ~b)
2686 not_kept_stores->Subtract(&kept_stores_);
2687 auto num_removed = not_kept_stores->NumSetBits();
2688 MaybeRecordStat(stats_, MethodCompilationStat::kPartialStoreRemoved, num_removed);
2689 }
2690 }
2691
UpdateValueRecordForStoreElimination(ValueRecord * value_record)2692 void LSEVisitor::UpdateValueRecordForStoreElimination(/*inout*/ValueRecord* value_record) {
2693 while (value_record->stored_by.IsInstruction() &&
2694 !kept_stores_.IsBitSet(value_record->stored_by.GetInstruction()->GetId())) {
2695 auto it = store_records_.find(value_record->stored_by.GetInstruction());
2696 DCHECK(it != store_records_.end());
2697 *value_record = it->second.old_value_record;
2698 }
2699 if (value_record->stored_by.NeedsPhi() &&
2700 !phi_placeholders_to_search_for_kept_stores_.IsBitSet(
2701 PhiPlaceholderIndex(value_record->stored_by))) {
2702 // Some stores feeding this heap location may have been eliminated. Use the `stored_by`
2703 // Phi placeholder to recalculate the actual value.
2704 value_record->value = value_record->stored_by;
2705 }
2706 value_record->value = ReplacementOrValue(value_record->value);
2707 if (value_record->value.NeedsNonLoopPhi()) {
2708 // Treat all Phi placeholders as requiring loop Phis at this point.
2709 // We do not want MaterializeLoopPhis() to call MaterializeNonLoopPhis().
2710 value_record->value = Value::ForLoopPhiPlaceholder(value_record->value.GetPhiPlaceholder());
2711 }
2712 }
2713
FindOldValueForPhiPlaceholder(PhiPlaceholder phi_placeholder,DataType::Type type)2714 void LSEVisitor::FindOldValueForPhiPlaceholder(PhiPlaceholder phi_placeholder,
2715 DataType::Type type) {
2716 DCHECK(phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)].IsInvalid());
2717
2718 // Use local allocator to reduce peak memory usage.
2719 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
2720 ArenaBitVector visited(&allocator,
2721 /*start_bits=*/ num_phi_placeholders_,
2722 /*expandable=*/ false,
2723 kArenaAllocLSE);
2724 visited.ClearAllBits();
2725
2726 // Find Phi placeholders to try and match against existing Phis or other replacement values.
2727 ArenaBitVector phi_placeholders_to_materialize(
2728 &allocator, num_phi_placeholders_, /*expandable=*/ false, kArenaAllocLSE);
2729 phi_placeholders_to_materialize.ClearAllBits();
2730 std::optional<PhiPlaceholder> loop_phi_with_unknown_input = FindLoopPhisToMaterialize(
2731 phi_placeholder, &phi_placeholders_to_materialize, type, /*can_use_default_or_phi=*/true);
2732 if (loop_phi_with_unknown_input) {
2733 DCHECK_GE(GetGraph()
2734 ->GetBlocks()[loop_phi_with_unknown_input->GetBlockId()]
2735 ->GetPredecessors()
2736 .size(),
2737 2u);
2738 // Mark the unreplacable placeholder as well as the input Phi placeholder as unreplaceable.
2739 phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)] = Value::PureUnknown();
2740 phi_placeholder_replacements_[PhiPlaceholderIndex(*loop_phi_with_unknown_input)] =
2741 Value::PureUnknown();
2742 return;
2743 }
2744
2745 DCHECK_EQ(current_phase_, Phase::kStoreElimination);
2746 bool success = MaterializeLoopPhis(phi_placeholders_to_materialize, type);
2747 DCHECK(phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)].IsValid());
2748 DCHECK_EQ(phi_placeholder_replacements_[PhiPlaceholderIndex(phi_placeholder)].IsUnknown(),
2749 !success);
2750 }
2751
2752 struct ScopedRestoreHeapValues {
2753 public:
ScopedRestoreHeapValuesart::ScopedRestoreHeapValues2754 ScopedRestoreHeapValues(ArenaStack* alloc,
2755 size_t num_heap_locs,
2756 ScopedArenaVector<ScopedArenaVector<LSEVisitor::ValueRecord>>& to_restore)
2757 : alloc_(alloc),
2758 updated_values_(alloc_.Adapter(kArenaAllocLSE)),
2759 to_restore_(to_restore) {
2760 updated_values_.reserve(num_heap_locs * to_restore_.size());
2761 }
2762
~ScopedRestoreHeapValuesart::ScopedRestoreHeapValues2763 ~ScopedRestoreHeapValues() {
2764 for (const auto& rec : updated_values_) {
2765 to_restore_[rec.blk_id][rec.heap_loc].value = rec.val_;
2766 }
2767 }
2768
2769 template<typename Func>
ForEachRecordart::ScopedRestoreHeapValues2770 void ForEachRecord(Func func) {
2771 for (size_t blk_id : Range(to_restore_.size())) {
2772 for (size_t heap_loc : Range(to_restore_[blk_id].size())) {
2773 LSEVisitor::ValueRecord* vr = &to_restore_[blk_id][heap_loc];
2774 LSEVisitor::Value initial = vr->value;
2775 func(vr);
2776 if (!vr->value.ExactEquals(initial)) {
2777 updated_values_.push_back({blk_id, heap_loc, initial});
2778 }
2779 }
2780 }
2781 }
2782
2783 private:
2784 struct UpdateRecord {
2785 size_t blk_id;
2786 size_t heap_loc;
2787 LSEVisitor::Value val_;
2788 };
2789 ScopedArenaAllocator alloc_;
2790 ScopedArenaVector<UpdateRecord> updated_values_;
2791 ScopedArenaVector<ScopedArenaVector<LSEVisitor::ValueRecord>>& to_restore_;
2792
2793 DISALLOW_COPY_AND_ASSIGN(ScopedRestoreHeapValues);
2794 };
2795
FindStoresWritingOldValues()2796 void LSEVisitor::FindStoresWritingOldValues() {
2797 // Partial LSE relies on knowing the real heap-values not the
2798 // store-replacement versions so we need to restore the map after removing
2799 // stores.
2800 ScopedRestoreHeapValues heap_vals(allocator_.GetArenaStack(),
2801 heap_location_collector_.GetNumberOfHeapLocations(),
2802 heap_values_for_);
2803 // The Phi placeholder replacements have so far been used for eliminating loads,
2804 // tracking values that would be stored if all stores were kept. As we want to
2805 // compare actual old values after removing unmarked stores, prune the Phi
2806 // placeholder replacements that can be fed by values we may not actually store.
2807 // Replacements marked as unknown can be kept as they are fed by some unknown
2808 // value and would end up as unknown again if we recalculated them.
2809 for (size_t i = 0, size = phi_placeholder_replacements_.size(); i != size; ++i) {
2810 if (!phi_placeholder_replacements_[i].IsUnknown() &&
2811 !phi_placeholders_to_search_for_kept_stores_.IsBitSet(i)) {
2812 phi_placeholder_replacements_[i] = Value::Invalid();
2813 }
2814 }
2815
2816 // Update heap values at end of blocks.
2817 heap_vals.ForEachRecord([&](ValueRecord* rec) {
2818 UpdateValueRecordForStoreElimination(rec);
2819 });
2820
2821 if (kIsDebugBuild) {
2822 heap_vals.ForEachRecord([](ValueRecord* rec) {
2823 DCHECK(!rec->value.NeedsNonLoopPhi()) << rec->value;
2824 });
2825 }
2826
2827 // Use local allocator to reduce peak memory usage.
2828 ScopedArenaAllocator allocator(allocator_.GetArenaStack());
2829 // Mark the stores we want to eliminate in a separate bit vector.
2830 ArenaBitVector eliminated_stores(&allocator,
2831 /*start_bits=*/ GetGraph()->GetCurrentInstructionId(),
2832 /*expandable=*/ false,
2833 kArenaAllocLSE);
2834 eliminated_stores.ClearAllBits();
2835
2836 for (auto& entry : store_records_) {
2837 HInstruction* store = entry.first;
2838 StoreRecord& store_record = entry.second;
2839 if (!kept_stores_.IsBitSet(store->GetId())) {
2840 continue; // Ignore stores that are not kept.
2841 }
2842 UpdateValueRecordForStoreElimination(&store_record.old_value_record);
2843 if (store_record.old_value_record.value.NeedsPhi()) {
2844 DataType::Type type = store_record.stored_value->GetType();
2845 FindOldValueForPhiPlaceholder(store_record.old_value_record.value.GetPhiPlaceholder(), type);
2846 store_record.old_value_record.value = ReplacementOrValue(store_record.old_value_record.value);
2847 }
2848 DCHECK(!store_record.old_value_record.value.NeedsPhi());
2849 HInstruction* stored_value = FindSubstitute(store_record.stored_value);
2850 if (store_record.old_value_record.value.Equals(stored_value)) {
2851 eliminated_stores.SetBit(store->GetId());
2852 }
2853 }
2854
2855 // Commit the stores to eliminate by removing them from `kept_stores_`.
2856 kept_stores_.Subtract(&eliminated_stores);
2857 }
2858
Run()2859 void LSEVisitor::Run() {
2860 // 1. Process blocks and instructions in reverse post order.
2861 for (HBasicBlock* block : GetGraph()->GetReversePostOrder()) {
2862 VisitBasicBlock(block);
2863 }
2864
2865 // 2. Process loads that require loop Phis, trying to find/create replacements.
2866 current_phase_ = Phase::kLoadElimination;
2867 ProcessLoadsRequiringLoopPhis();
2868
2869 // 3. Determine which stores to keep and which to eliminate.
2870 current_phase_ = Phase::kStoreElimination;
2871 // Finish marking stores for keeping.
2872 SearchPhiPlaceholdersForKeptStores();
2873
2874 // Find stores that write the same value as is already present in the location.
2875 FindStoresWritingOldValues();
2876
2877 // 4. Replace loads and remove unnecessary stores and singleton allocations.
2878 FinishFullLSE();
2879
2880 // 5. Move partial escapes down and fixup with PHIs.
2881 current_phase_ = Phase::kPartialElimination;
2882 MovePartialEscapes();
2883 }
2884
2885 // Clear unknown loop-phi results. Here we'll be able to use partial-unknowns so we need to
2886 // retry all of them with more information about where they come from.
PrepareForPartialPhiComputation()2887 void LSEVisitor::PrepareForPartialPhiComputation() {
2888 std::replace_if(
2889 phi_placeholder_replacements_.begin(),
2890 phi_placeholder_replacements_.end(),
2891 [](const Value& val) { return !val.IsDefault() && !val.IsInstruction(); },
2892 Value::Invalid());
2893 }
2894
2895 class PartialLoadStoreEliminationHelper {
2896 public:
PartialLoadStoreEliminationHelper(LSEVisitor * lse,ScopedArenaAllocator * alloc)2897 PartialLoadStoreEliminationHelper(LSEVisitor* lse, ScopedArenaAllocator* alloc)
2898 : lse_(lse),
2899 alloc_(alloc),
2900 new_ref_phis_(alloc_->Adapter(kArenaAllocLSE)),
2901 heap_refs_(alloc_->Adapter(kArenaAllocLSE)),
2902 max_preds_per_block_((*std::max_element(GetGraph()->GetActiveBlocks().begin(),
2903 GetGraph()->GetActiveBlocks().end(),
2904 [](HBasicBlock* a, HBasicBlock* b) {
2905 return a->GetNumberOfPredecessors() <
2906 b->GetNumberOfPredecessors();
2907 }))
2908 ->GetNumberOfPredecessors()),
2909 materialization_blocks_(GetGraph()->GetBlocks().size() * max_preds_per_block_,
2910 nullptr,
2911 alloc_->Adapter(kArenaAllocLSE)),
2912 first_materialization_block_id_(GetGraph()->GetBlocks().size()) {
2913 size_t num_partial_singletons = lse_->heap_location_collector_.CountPartialSingletons();
2914 heap_refs_.reserve(num_partial_singletons);
2915 new_ref_phis_.reserve(num_partial_singletons * GetGraph()->GetBlocks().size());
2916 CollectInterestingHeapRefs();
2917 }
2918
~PartialLoadStoreEliminationHelper()2919 ~PartialLoadStoreEliminationHelper() {
2920 if (heap_refs_.empty()) {
2921 return;
2922 }
2923 ReferenceTypePropagation rtp_fixup(GetGraph(),
2924 Handle<mirror::ClassLoader>(),
2925 Handle<mirror::DexCache>(),
2926 /* is_first_run= */ false);
2927 rtp_fixup.Visit(ArrayRef<HInstruction* const>(new_ref_phis_));
2928 GetGraph()->ClearLoopInformation();
2929 GetGraph()->ClearDominanceInformation();
2930 GetGraph()->ClearReachabilityInformation();
2931 GetGraph()->BuildDominatorTree();
2932 GetGraph()->ComputeReachabilityInformation();
2933 }
2934
2935 class IdxToHeapLoc {
2936 public:
IdxToHeapLoc(const HeapLocationCollector * hlc)2937 explicit IdxToHeapLoc(const HeapLocationCollector* hlc) : collector_(hlc) {}
operator ()(size_t idx) const2938 HeapLocation* operator()(size_t idx) const {
2939 return collector_->GetHeapLocation(idx);
2940 }
2941
2942 private:
2943 const HeapLocationCollector* collector_;
2944 };
2945
2946
2947 class HeapReferenceData {
2948 public:
2949 using LocIterator = IterationRange<TransformIterator<BitVector::IndexIterator, IdxToHeapLoc>>;
HeapReferenceData(PartialLoadStoreEliminationHelper * helper,HNewInstance * new_inst,const ExecutionSubgraph * subgraph,ScopedArenaAllocator * alloc)2950 HeapReferenceData(PartialLoadStoreEliminationHelper* helper,
2951 HNewInstance* new_inst,
2952 const ExecutionSubgraph* subgraph,
2953 ScopedArenaAllocator* alloc)
2954 : new_instance_(new_inst),
2955 helper_(helper),
2956 heap_locs_(alloc,
2957 helper->lse_->heap_location_collector_.GetNumberOfHeapLocations(),
2958 /* expandable= */ false,
2959 kArenaAllocLSE),
2960 materializations_(
2961 // We generally won't need to create too many materialization blocks and we can expand
2962 // this as needed so just start off with 2x.
2963 2 * helper->lse_->GetGraph()->GetBlocks().size(),
2964 nullptr,
2965 alloc->Adapter(kArenaAllocLSE)),
2966 collector_(helper->lse_->heap_location_collector_),
2967 subgraph_(subgraph) {}
2968
IterateLocations()2969 LocIterator IterateLocations() {
2970 auto idxs = heap_locs_.Indexes();
2971 return MakeTransformRange(idxs, IdxToHeapLoc(&collector_));
2972 }
2973
AddHeapLocation(size_t idx)2974 void AddHeapLocation(size_t idx) {
2975 heap_locs_.SetBit(idx);
2976 }
2977
GetNoEscapeSubgraph() const2978 const ExecutionSubgraph* GetNoEscapeSubgraph() const {
2979 return subgraph_;
2980 }
2981
IsPostEscape(HBasicBlock * blk)2982 bool IsPostEscape(HBasicBlock* blk) {
2983 return std::any_of(
2984 subgraph_->GetExcludedCohorts().cbegin(),
2985 subgraph_->GetExcludedCohorts().cend(),
2986 [&](const ExecutionSubgraph::ExcludedCohort& ec) { return ec.PrecedesBlock(blk); });
2987 }
2988
InEscapeCohort(HBasicBlock * blk)2989 bool InEscapeCohort(HBasicBlock* blk) {
2990 return std::any_of(
2991 subgraph_->GetExcludedCohorts().cbegin(),
2992 subgraph_->GetExcludedCohorts().cend(),
2993 [&](const ExecutionSubgraph::ExcludedCohort& ec) { return ec.ContainsBlock(blk); });
2994 }
2995
BeforeAllEscapes(HBasicBlock * b)2996 bool BeforeAllEscapes(HBasicBlock* b) {
2997 return std::none_of(subgraph_->GetExcludedCohorts().cbegin(),
2998 subgraph_->GetExcludedCohorts().cend(),
2999 [&](const ExecutionSubgraph::ExcludedCohort& ec) {
3000 return ec.PrecedesBlock(b) || ec.ContainsBlock(b);
3001 });
3002 }
3003
OriginalNewInstance() const3004 HNewInstance* OriginalNewInstance() const {
3005 return new_instance_;
3006 }
3007
3008 // Collect and replace all uses. We need to perform this twice since we will
3009 // generate PHIs and additional uses as we create the default-values for
3010 // pred-gets. These values might be other references that are also being
3011 // partially eliminated. By running just the replacement part again we are
3012 // able to avoid having to keep another whole in-progress partial map
3013 // around. Since we will have already handled all the other uses in the
3014 // first pass the second one will be quite fast.
FixupUses(bool first_pass)3015 void FixupUses(bool first_pass) {
3016 ScopedArenaAllocator saa(GetGraph()->GetArenaStack());
3017 // Replace uses with materialized values.
3018 ScopedArenaVector<InstructionUse<HInstruction>> to_replace(saa.Adapter(kArenaAllocLSE));
3019 ScopedArenaVector<HInstruction*> to_remove(saa.Adapter(kArenaAllocLSE));
3020 // Do we need to add a constructor-fence.
3021 ScopedArenaVector<InstructionUse<HConstructorFence>> constructor_fences(
3022 saa.Adapter(kArenaAllocLSE));
3023 ScopedArenaVector<InstructionUse<HInstruction>> to_predicate(saa.Adapter(kArenaAllocLSE));
3024
3025 CollectReplacements(to_replace, to_remove, constructor_fences, to_predicate);
3026
3027 if (!first_pass) {
3028 // If another partial creates new references they can only be in Phis or pred-get defaults
3029 // so they must be in the to_replace group.
3030 DCHECK(to_predicate.empty());
3031 DCHECK(constructor_fences.empty());
3032 DCHECK(to_remove.empty());
3033 }
3034
3035 ReplaceInput(to_replace);
3036 RemoveAndReplaceInputs(to_remove);
3037 CreateConstructorFences(constructor_fences);
3038 PredicateInstructions(to_predicate);
3039
3040 CHECK(OriginalNewInstance()->GetUses().empty())
3041 << OriginalNewInstance()->GetUses() << ", " << OriginalNewInstance()->GetEnvUses();
3042 }
3043
AddMaterialization(HBasicBlock * blk,HInstruction * ins)3044 void AddMaterialization(HBasicBlock* blk, HInstruction* ins) {
3045 if (blk->GetBlockId() >= materializations_.size()) {
3046 // Make sure the materialization array is large enough, try to avoid
3047 // re-sizing too many times by giving extra space.
3048 materializations_.resize(blk->GetBlockId() * 2, nullptr);
3049 }
3050 DCHECK(materializations_[blk->GetBlockId()] == nullptr)
3051 << "Already have a materialization in block " << blk->GetBlockId() << ": "
3052 << *materializations_[blk->GetBlockId()] << " when trying to set materialization to "
3053 << *ins;
3054 materializations_[blk->GetBlockId()] = ins;
3055 LSE_VLOG << "In block " << blk->GetBlockId() << " materialization is " << *ins;
3056 helper_->NotifyNewMaterialization(ins);
3057 }
3058
HasMaterialization(HBasicBlock * blk) const3059 bool HasMaterialization(HBasicBlock* blk) const {
3060 return blk->GetBlockId() < materializations_.size() &&
3061 materializations_[blk->GetBlockId()] != nullptr;
3062 }
3063
GetMaterialization(HBasicBlock * blk) const3064 HInstruction* GetMaterialization(HBasicBlock* blk) const {
3065 if (materializations_.size() <= blk->GetBlockId() ||
3066 materializations_[blk->GetBlockId()] == nullptr) {
3067 // This must be a materialization block added after the partial LSE of
3068 // the current reference finished. Since every edge can only have at
3069 // most one materialization block added to it we can just check the
3070 // blocks predecessor.
3071 DCHECK(helper_->IsMaterializationBlock(blk));
3072 blk = helper_->FindDominatingNonMaterializationBlock(blk);
3073 DCHECK(!helper_->IsMaterializationBlock(blk));
3074 }
3075 DCHECK_GT(materializations_.size(), blk->GetBlockId());
3076 DCHECK(materializations_[blk->GetBlockId()] != nullptr);
3077 return materializations_[blk->GetBlockId()];
3078 }
3079
GenerateMaterializationValueFromPredecessors(HBasicBlock * blk)3080 void GenerateMaterializationValueFromPredecessors(HBasicBlock* blk) {
3081 DCHECK(std::none_of(GetNoEscapeSubgraph()->GetExcludedCohorts().begin(),
3082 GetNoEscapeSubgraph()->GetExcludedCohorts().end(),
3083 [&](const ExecutionSubgraph::ExcludedCohort& cohort) {
3084 return cohort.IsEntryBlock(blk);
3085 }));
3086 DCHECK(!HasMaterialization(blk));
3087 if (blk->IsExitBlock()) {
3088 return;
3089 } else if (blk->IsLoopHeader()) {
3090 // See comment in execution_subgraph.h. Currently we act as though every
3091 // allocation for partial elimination takes place in the entry block.
3092 // This simplifies the analysis by making it so any escape cohort
3093 // expands to contain any loops it is a part of. This is something that
3094 // we should rectify at some point. In either case however we can still
3095 // special case the loop-header since (1) currently the loop can't have
3096 // any merges between different cohort entries since the pre-header will
3097 // be the earliest place entry can happen and (2) even if the analysis
3098 // is improved to consider lifetime of the object WRT loops any values
3099 // which would require loop-phis would have to make the whole loop
3100 // escape anyway.
3101 // This all means we can always use value from the pre-header when the
3102 // block is the loop-header and we didn't already create a
3103 // materialization block. (NB when we do improve the analysis we will
3104 // need to modify the materialization creation code to deal with this
3105 // correctly.)
3106 HInstruction* pre_header_val =
3107 GetMaterialization(blk->GetLoopInformation()->GetPreHeader());
3108 AddMaterialization(blk, pre_header_val);
3109 return;
3110 }
3111 ScopedArenaAllocator saa(GetGraph()->GetArenaStack());
3112 ScopedArenaVector<HInstruction*> pred_vals(saa.Adapter(kArenaAllocLSE));
3113 pred_vals.reserve(blk->GetNumberOfPredecessors());
3114 for (HBasicBlock* pred : blk->GetPredecessors()) {
3115 DCHECK(HasMaterialization(pred));
3116 pred_vals.push_back(GetMaterialization(pred));
3117 }
3118 GenerateMaterializationValueFromPredecessorsDirect(blk, pred_vals);
3119 }
3120
GenerateMaterializationValueFromPredecessorsForEntry(HBasicBlock * entry,const ScopedArenaVector<HInstruction * > & pred_vals)3121 void GenerateMaterializationValueFromPredecessorsForEntry(
3122 HBasicBlock* entry, const ScopedArenaVector<HInstruction*>& pred_vals) {
3123 DCHECK(std::any_of(GetNoEscapeSubgraph()->GetExcludedCohorts().begin(),
3124 GetNoEscapeSubgraph()->GetExcludedCohorts().end(),
3125 [&](const ExecutionSubgraph::ExcludedCohort& cohort) {
3126 return cohort.IsEntryBlock(entry);
3127 }));
3128 GenerateMaterializationValueFromPredecessorsDirect(entry, pred_vals);
3129 }
3130
3131 private:
3132 template <typename InstructionType>
3133 struct InstructionUse {
3134 InstructionType* instruction_;
3135 size_t index_;
3136 };
3137
ReplaceInput(const ScopedArenaVector<InstructionUse<HInstruction>> & to_replace)3138 void ReplaceInput(const ScopedArenaVector<InstructionUse<HInstruction>>& to_replace) {
3139 for (auto& [ins, idx] : to_replace) {
3140 HInstruction* merged_inst = GetMaterialization(ins->GetBlock());
3141 if (ins->IsPhi() && merged_inst->IsPhi() && ins->GetBlock() == merged_inst->GetBlock()) {
3142 // Phis we just pass through the appropriate inputs.
3143 ins->ReplaceInput(merged_inst->InputAt(idx), idx);
3144 } else {
3145 ins->ReplaceInput(merged_inst, idx);
3146 }
3147 }
3148 }
3149
RemoveAndReplaceInputs(const ScopedArenaVector<HInstruction * > & to_remove)3150 void RemoveAndReplaceInputs(const ScopedArenaVector<HInstruction*>& to_remove) {
3151 for (HInstruction* ins : to_remove) {
3152 if (ins->GetBlock() == nullptr) {
3153 // Already dealt with.
3154 continue;
3155 }
3156 DCHECK(BeforeAllEscapes(ins->GetBlock())) << *ins;
3157 if (ins->IsInstanceFieldGet() || ins->IsInstanceFieldSet()) {
3158 bool instruction_has_users =
3159 ins->IsInstanceFieldGet() && (!ins->GetUses().empty() || !ins->GetEnvUses().empty());
3160 if (instruction_has_users) {
3161 // Make sure any remaining users of read are replaced.
3162 HInstruction* replacement =
3163 helper_->lse_->GetPartialValueAt(OriginalNewInstance(), ins);
3164 // NB ReplaceInput will remove a use from the list so this is
3165 // guaranteed to finish eventually.
3166 while (!ins->GetUses().empty()) {
3167 const HUseListNode<HInstruction*>& use = ins->GetUses().front();
3168 use.GetUser()->ReplaceInput(replacement, use.GetIndex());
3169 }
3170 while (!ins->GetEnvUses().empty()) {
3171 const HUseListNode<HEnvironment*>& use = ins->GetEnvUses().front();
3172 use.GetUser()->ReplaceInput(replacement, use.GetIndex());
3173 }
3174 } else {
3175 DCHECK(ins->GetUses().empty())
3176 << "Instruction has users!\n"
3177 << ins->DumpWithArgs() << "\nUsers are " << ins->GetUses();
3178 DCHECK(ins->GetEnvUses().empty())
3179 << "Instruction has users!\n"
3180 << ins->DumpWithArgs() << "\nUsers are " << ins->GetEnvUses();
3181 }
3182 ins->GetBlock()->RemoveInstruction(ins);
3183 } else {
3184 // Can only be obj == other, obj != other, obj == obj (!?) or, obj != obj (!?)
3185 // Since PHIs are escapes as far as LSE is concerned and we are before
3186 // any escapes these are the only 4 options.
3187 DCHECK(ins->IsEqual() || ins->IsNotEqual()) << *ins;
3188 HInstruction* replacement;
3189 if (UNLIKELY(ins->InputAt(0) == ins->InputAt(1))) {
3190 replacement = ins->IsEqual() ? GetGraph()->GetIntConstant(1)
3191 : GetGraph()->GetIntConstant(0);
3192 } else {
3193 replacement = ins->IsEqual() ? GetGraph()->GetIntConstant(0)
3194 : GetGraph()->GetIntConstant(1);
3195 }
3196 ins->ReplaceWith(replacement);
3197 ins->GetBlock()->RemoveInstruction(ins);
3198 }
3199 }
3200 }
3201
CreateConstructorFences(const ScopedArenaVector<InstructionUse<HConstructorFence>> & constructor_fences)3202 void CreateConstructorFences(
3203 const ScopedArenaVector<InstructionUse<HConstructorFence>>& constructor_fences) {
3204 if (!constructor_fences.empty()) {
3205 uint32_t pc = constructor_fences.front().instruction_->GetDexPc();
3206 for (auto& [cf, idx] : constructor_fences) {
3207 if (cf->GetInputs().size() == 1) {
3208 cf->GetBlock()->RemoveInstruction(cf);
3209 } else {
3210 cf->RemoveInputAt(idx);
3211 }
3212 }
3213 for (const ExecutionSubgraph::ExcludedCohort& ec :
3214 GetNoEscapeSubgraph()->GetExcludedCohorts()) {
3215 for (HBasicBlock* blk : ec.EntryBlocks()) {
3216 for (HBasicBlock* materializer :
3217 Filter(MakeIterationRange(blk->GetPredecessors()),
3218 [&](HBasicBlock* blk) { return helper_->IsMaterializationBlock(blk); })) {
3219 HInstruction* new_cf = new (GetGraph()->GetAllocator()) HConstructorFence(
3220 GetMaterialization(materializer), pc, GetGraph()->GetAllocator());
3221 materializer->InsertInstructionBefore(new_cf, materializer->GetLastInstruction());
3222 }
3223 }
3224 }
3225 }
3226 }
3227
PredicateInstructions(const ScopedArenaVector<InstructionUse<HInstruction>> & to_predicate)3228 void PredicateInstructions(
3229 const ScopedArenaVector<InstructionUse<HInstruction>>& to_predicate) {
3230 for (auto& [ins, idx] : to_predicate) {
3231 if (UNLIKELY(ins->GetBlock() == nullptr)) {
3232 // Already handled due to obj == obj;
3233 continue;
3234 } else if (ins->IsInstanceFieldGet()) {
3235 // IFieldGet[obj] => PredicatedIFieldGet[PartialValue, obj]
3236 HInstruction* new_fget = new (GetGraph()->GetAllocator()) HPredicatedInstanceFieldGet(
3237 ins->AsInstanceFieldGet(),
3238 GetMaterialization(ins->GetBlock()),
3239 helper_->lse_->GetPartialValueAt(OriginalNewInstance(), ins));
3240 MaybeRecordStat(helper_->lse_->stats_, MethodCompilationStat::kPredicatedLoadAdded);
3241 ins->GetBlock()->InsertInstructionBefore(new_fget, ins);
3242 if (ins->GetType() == DataType::Type::kReference) {
3243 // Reference info is the same
3244 new_fget->SetReferenceTypeInfo(ins->GetReferenceTypeInfo());
3245 }
3246 // In this phase, substitute instructions are used only for the predicated get
3247 // default values which are used only if the partial singleton did not escape,
3248 // so the out value of the `new_fget` for the relevant cases is the same as
3249 // the default value.
3250 // TODO: Use the default value for materializing default values used by
3251 // other predicated loads to avoid some unnecessary Phis. (This shall
3252 // complicate the search for replacement in `ReplacementOrValue()`.)
3253 DCHECK(helper_->lse_->substitute_instructions_for_loads_[ins->GetId()] == nullptr);
3254 helper_->lse_->substitute_instructions_for_loads_[ins->GetId()] = new_fget;
3255 ins->ReplaceWith(new_fget);
3256 ins->ReplaceEnvUsesDominatedBy(ins, new_fget);
3257 CHECK(ins->GetEnvUses().empty() && ins->GetUses().empty())
3258 << "Instruction: " << *ins << " uses: " << ins->GetUses()
3259 << ", env: " << ins->GetEnvUses();
3260 ins->GetBlock()->RemoveInstruction(ins);
3261 } else if (ins->IsInstanceFieldSet()) {
3262 // Any predicated sets shouldn't require movement.
3263 ins->AsInstanceFieldSet()->SetIsPredicatedSet();
3264 MaybeRecordStat(helper_->lse_->stats_, MethodCompilationStat::kPredicatedStoreAdded);
3265 HInstruction* merged_inst = GetMaterialization(ins->GetBlock());
3266 ins->ReplaceInput(merged_inst, idx);
3267 } else {
3268 // comparisons need to be split into 2.
3269 DCHECK(ins->IsEqual() || ins->IsNotEqual()) << "bad instruction " << *ins;
3270 bool this_is_first = idx == 0;
3271 if (ins->InputAt(0) == ins->InputAt(1)) {
3272 // This is a obj == obj or obj != obj.
3273 // No idea why anyone would do this but whatever.
3274 ins->ReplaceWith(GetGraph()->GetIntConstant(ins->IsEqual() ? 1 : 0));
3275 ins->GetBlock()->RemoveInstruction(ins);
3276 continue;
3277 } else {
3278 HInstruction* is_escaped = new (GetGraph()->GetAllocator())
3279 HNotEqual(GetMaterialization(ins->GetBlock()), GetGraph()->GetNullConstant());
3280 HInstruction* combine_inst =
3281 ins->IsEqual() ? static_cast<HInstruction*>(new (GetGraph()->GetAllocator()) HAnd(
3282 DataType::Type::kBool, is_escaped, ins))
3283 : static_cast<HInstruction*>(new (GetGraph()->GetAllocator()) HOr(
3284 DataType::Type::kBool, is_escaped, ins));
3285 ins->ReplaceInput(GetMaterialization(ins->GetBlock()), this_is_first ? 0 : 1);
3286 ins->GetBlock()->InsertInstructionBefore(is_escaped, ins);
3287 ins->GetBlock()->InsertInstructionAfter(combine_inst, ins);
3288 ins->ReplaceWith(combine_inst);
3289 combine_inst->ReplaceInput(ins, 1);
3290 }
3291 }
3292 }
3293 }
3294
3295 // Figure out all the instructions we need to
3296 // fixup/replace/remove/duplicate. Since this requires an iteration of an
3297 // intrusive linked list we want to do it only once and collect all the data
3298 // here.
CollectReplacements(ScopedArenaVector<InstructionUse<HInstruction>> & to_replace,ScopedArenaVector<HInstruction * > & to_remove,ScopedArenaVector<InstructionUse<HConstructorFence>> & constructor_fences,ScopedArenaVector<InstructionUse<HInstruction>> & to_predicate)3299 void CollectReplacements(
3300 ScopedArenaVector<InstructionUse<HInstruction>>& to_replace,
3301 ScopedArenaVector<HInstruction*>& to_remove,
3302 ScopedArenaVector<InstructionUse<HConstructorFence>>& constructor_fences,
3303 ScopedArenaVector<InstructionUse<HInstruction>>& to_predicate) {
3304 size_t size = new_instance_->GetUses().SizeSlow();
3305 to_replace.reserve(size);
3306 to_remove.reserve(size);
3307 constructor_fences.reserve(size);
3308 to_predicate.reserve(size);
3309 for (auto& use : new_instance_->GetUses()) {
3310 HBasicBlock* blk =
3311 helper_->FindDominatingNonMaterializationBlock(use.GetUser()->GetBlock());
3312 if (InEscapeCohort(blk)) {
3313 LSE_VLOG << "Replacing " << *new_instance_ << " use in " << *use.GetUser() << " with "
3314 << *GetMaterialization(blk);
3315 to_replace.push_back({use.GetUser(), use.GetIndex()});
3316 } else if (IsPostEscape(blk)) {
3317 LSE_VLOG << "User " << *use.GetUser() << " after escapes!";
3318 // The fields + cmp are normal uses. Phi can only be here if it was
3319 // generated by full LSE so whatever store+load that created the phi
3320 // is the escape.
3321 if (use.GetUser()->IsPhi()) {
3322 to_replace.push_back({use.GetUser(), use.GetIndex()});
3323 } else {
3324 DCHECK(use.GetUser()->IsFieldAccess() ||
3325 use.GetUser()->IsEqual() ||
3326 use.GetUser()->IsNotEqual())
3327 << *use.GetUser() << "@" << use.GetIndex();
3328 to_predicate.push_back({use.GetUser(), use.GetIndex()});
3329 }
3330 } else if (use.GetUser()->IsConstructorFence()) {
3331 LSE_VLOG << "User " << *use.GetUser() << " being moved to materialization!";
3332 constructor_fences.push_back({use.GetUser()->AsConstructorFence(), use.GetIndex()});
3333 } else {
3334 LSE_VLOG << "User " << *use.GetUser() << " not contained in cohort!";
3335 to_remove.push_back(use.GetUser());
3336 }
3337 }
3338 DCHECK_EQ(
3339 to_replace.size() + to_remove.size() + constructor_fences.size() + to_predicate.size(),
3340 size);
3341 }
3342
GenerateMaterializationValueFromPredecessorsDirect(HBasicBlock * blk,const ScopedArenaVector<HInstruction * > & pred_vals)3343 void GenerateMaterializationValueFromPredecessorsDirect(
3344 HBasicBlock* blk, const ScopedArenaVector<HInstruction*>& pred_vals) {
3345 DCHECK(!pred_vals.empty());
3346 bool all_equal = std::all_of(pred_vals.begin() + 1, pred_vals.end(), [&](HInstruction* val) {
3347 return val == pred_vals.front();
3348 });
3349 if (LIKELY(all_equal)) {
3350 AddMaterialization(blk, pred_vals.front());
3351 } else {
3352 // Make a PHI for the predecessors.
3353 HPhi* phi = new (GetGraph()->GetAllocator()) HPhi(
3354 GetGraph()->GetAllocator(), kNoRegNumber, pred_vals.size(), DataType::Type::kReference);
3355 for (const auto& [ins, off] : ZipCount(MakeIterationRange(pred_vals))) {
3356 phi->SetRawInputAt(off, ins);
3357 }
3358 blk->AddPhi(phi);
3359 AddMaterialization(blk, phi);
3360 }
3361 }
3362
GetGraph() const3363 HGraph* GetGraph() const {
3364 return helper_->GetGraph();
3365 }
3366
3367 HNewInstance* new_instance_;
3368 PartialLoadStoreEliminationHelper* helper_;
3369 ArenaBitVector heap_locs_;
3370 ScopedArenaVector<HInstruction*> materializations_;
3371 const HeapLocationCollector& collector_;
3372 const ExecutionSubgraph* subgraph_;
3373 };
3374
GetHeapRefs()3375 ArrayRef<HeapReferenceData> GetHeapRefs() {
3376 return ArrayRef<HeapReferenceData>(heap_refs_);
3377 }
3378
IsMaterializationBlock(HBasicBlock * blk) const3379 bool IsMaterializationBlock(HBasicBlock* blk) const {
3380 return blk->GetBlockId() >= first_materialization_block_id_;
3381 }
3382
GetOrCreateMaterializationBlock(HBasicBlock * entry,size_t pred_num)3383 HBasicBlock* GetOrCreateMaterializationBlock(HBasicBlock* entry, size_t pred_num) {
3384 size_t idx = GetMaterializationBlockIndex(entry, pred_num);
3385 HBasicBlock* blk = materialization_blocks_[idx];
3386 if (blk == nullptr) {
3387 blk = new (GetGraph()->GetAllocator()) HBasicBlock(GetGraph());
3388 GetGraph()->AddBlock(blk);
3389 LSE_VLOG << "creating materialization block " << blk->GetBlockId() << " on edge "
3390 << entry->GetPredecessors()[pred_num]->GetBlockId() << "->" << entry->GetBlockId();
3391 blk->AddInstruction(new (GetGraph()->GetAllocator()) HGoto());
3392 materialization_blocks_[idx] = blk;
3393 }
3394 return blk;
3395 }
3396
GetMaterializationBlock(HBasicBlock * entry,size_t pred_num)3397 HBasicBlock* GetMaterializationBlock(HBasicBlock* entry, size_t pred_num) {
3398 HBasicBlock* out = materialization_blocks_[GetMaterializationBlockIndex(entry, pred_num)];
3399 DCHECK(out != nullptr) << "No materialization block for edge " << entry->GetBlockId() << "->"
3400 << entry->GetPredecessors()[pred_num]->GetBlockId();
3401 return out;
3402 }
3403
IterateMaterializationBlocks()3404 IterationRange<ArenaVector<HBasicBlock*>::const_iterator> IterateMaterializationBlocks() {
3405 return MakeIterationRange(GetGraph()->GetBlocks().begin() + first_materialization_block_id_,
3406 GetGraph()->GetBlocks().end());
3407 }
3408
FixupPartialObjectUsers()3409 void FixupPartialObjectUsers() {
3410 for (PartialLoadStoreEliminationHelper::HeapReferenceData& ref_data : GetHeapRefs()) {
3411 // Use the materialized instances to replace original instance
3412 ref_data.FixupUses(/*first_pass=*/true);
3413 CHECK(ref_data.OriginalNewInstance()->GetUses().empty())
3414 << ref_data.OriginalNewInstance()->GetUses() << ", "
3415 << ref_data.OriginalNewInstance()->GetEnvUses();
3416 }
3417 // This can cause new uses to be created due to the creation of phis/pred-get defaults
3418 for (PartialLoadStoreEliminationHelper::HeapReferenceData& ref_data : GetHeapRefs()) {
3419 // Only need to handle new phis/pred-get defaults. DCHECK that's all we find.
3420 ref_data.FixupUses(/*first_pass=*/false);
3421 CHECK(ref_data.OriginalNewInstance()->GetUses().empty())
3422 << ref_data.OriginalNewInstance()->GetUses() << ", "
3423 << ref_data.OriginalNewInstance()->GetEnvUses();
3424 }
3425 }
3426
3427 // Finds the first block which either is or dominates the given block which is
3428 // not a materialization block
FindDominatingNonMaterializationBlock(HBasicBlock * blk)3429 HBasicBlock* FindDominatingNonMaterializationBlock(HBasicBlock* blk) {
3430 if (LIKELY(!IsMaterializationBlock(blk))) {
3431 // Not a materialization block so itself.
3432 return blk;
3433 } else if (blk->GetNumberOfPredecessors() != 0) {
3434 // We're far enough along that the materialization blocks have been
3435 // inserted into the graph so no need to go searching.
3436 return blk->GetSinglePredecessor();
3437 }
3438 // Search through the materialization blocks to find where it will be
3439 // inserted.
3440 for (auto [mat, idx] : ZipCount(MakeIterationRange(materialization_blocks_))) {
3441 if (mat == blk) {
3442 size_t cur_pred_idx = idx % max_preds_per_block_;
3443 HBasicBlock* entry = GetGraph()->GetBlocks()[idx / max_preds_per_block_];
3444 return entry->GetPredecessors()[cur_pred_idx];
3445 }
3446 }
3447 LOG(FATAL) << "Unable to find materialization block position for " << blk->GetBlockId() << "!";
3448 return nullptr;
3449 }
3450
InsertMaterializationBlocks()3451 void InsertMaterializationBlocks() {
3452 for (auto [mat, idx] : ZipCount(MakeIterationRange(materialization_blocks_))) {
3453 if (mat == nullptr) {
3454 continue;
3455 }
3456 size_t cur_pred_idx = idx % max_preds_per_block_;
3457 HBasicBlock* entry = GetGraph()->GetBlocks()[idx / max_preds_per_block_];
3458 HBasicBlock* pred = entry->GetPredecessors()[cur_pred_idx];
3459 mat->InsertBetween(pred, entry);
3460 LSE_VLOG << "Adding materialization block " << mat->GetBlockId() << " on edge "
3461 << pred->GetBlockId() << "->" << entry->GetBlockId();
3462 }
3463 }
3464
3465 // Replace any env-uses remaining of the partial singletons with the
3466 // appropriate phis and remove the instructions.
RemoveReplacedInstructions()3467 void RemoveReplacedInstructions() {
3468 for (HeapReferenceData& ref_data : GetHeapRefs()) {
3469 CHECK(ref_data.OriginalNewInstance()->GetUses().empty())
3470 << ref_data.OriginalNewInstance()->GetUses() << ", "
3471 << ref_data.OriginalNewInstance()->GetEnvUses()
3472 << " inst is: " << ref_data.OriginalNewInstance();
3473 const auto& env_uses = ref_data.OriginalNewInstance()->GetEnvUses();
3474 while (!env_uses.empty()) {
3475 const HUseListNode<HEnvironment*>& use = env_uses.front();
3476 HInstruction* merged_inst =
3477 ref_data.GetMaterialization(use.GetUser()->GetHolder()->GetBlock());
3478 LSE_VLOG << "Replacing env use of " << *use.GetUser()->GetHolder() << "@" << use.GetIndex()
3479 << " with " << *merged_inst;
3480 use.GetUser()->ReplaceInput(merged_inst, use.GetIndex());
3481 }
3482 ref_data.OriginalNewInstance()->GetBlock()->RemoveInstruction(ref_data.OriginalNewInstance());
3483 }
3484 }
3485
3486 // We need to make sure any allocations dominate their environment uses.
3487 // Technically we could probably remove the env-uses and be fine but this is easy.
ReorderMaterializationsForEnvDominance()3488 void ReorderMaterializationsForEnvDominance() {
3489 for (HBasicBlock* blk : IterateMaterializationBlocks()) {
3490 ScopedArenaAllocator alloc(alloc_->GetArenaStack());
3491 ArenaBitVector still_unsorted(
3492 &alloc, GetGraph()->GetCurrentInstructionId(), false, kArenaAllocLSE);
3493 // This is guaranteed to be very short (since we will abandon LSE if there
3494 // are >= kMaxNumberOfHeapLocations (32) heap locations so that is the
3495 // absolute maximum size this list can be) so doing a selection sort is
3496 // fine. This avoids the need to do a complicated recursive check to
3497 // ensure transitivity for std::sort.
3498 ScopedArenaVector<HNewInstance*> materializations(alloc.Adapter(kArenaAllocLSE));
3499 materializations.reserve(GetHeapRefs().size());
3500 for (HInstruction* ins :
3501 MakeSTLInstructionIteratorRange(HInstructionIterator(blk->GetInstructions()))) {
3502 if (ins->IsNewInstance()) {
3503 materializations.push_back(ins->AsNewInstance());
3504 still_unsorted.SetBit(ins->GetId());
3505 }
3506 }
3507 using Iter = ScopedArenaVector<HNewInstance*>::iterator;
3508 Iter unsorted_start = materializations.begin();
3509 Iter unsorted_end = materializations.end();
3510 // selection sort. Required since the only check we can easily perform a
3511 // is-before-all-unsorted check.
3512 while (unsorted_start != unsorted_end) {
3513 bool found_instruction = false;
3514 for (Iter candidate = unsorted_start; candidate != unsorted_end; ++candidate) {
3515 HNewInstance* ni = *candidate;
3516 if (std::none_of(ni->GetAllEnvironments().cbegin(),
3517 ni->GetAllEnvironments().cend(),
3518 [&](const HEnvironment* env) {
3519 return std::any_of(
3520 env->GetEnvInputs().cbegin(),
3521 env->GetEnvInputs().cend(),
3522 [&](const HInstruction* env_element) {
3523 return env_element != nullptr &&
3524 still_unsorted.IsBitSet(env_element->GetId());
3525 });
3526 })) {
3527 still_unsorted.ClearBit(ni->GetId());
3528 std::swap(*unsorted_start, *candidate);
3529 ++unsorted_start;
3530 found_instruction = true;
3531 break;
3532 }
3533 }
3534 CHECK(found_instruction) << "Unable to select next materialization instruction."
3535 << " Environments have a dependency loop!";
3536 }
3537 // Reverse so we as we prepend them we end up with the correct order.
3538 auto reverse_iter = MakeIterationRange(materializations.rbegin(), materializations.rend());
3539 for (HNewInstance* ins : reverse_iter) {
3540 if (blk->GetFirstInstruction() != ins) {
3541 // Don't do checks since that makes sure the move is safe WRT
3542 // ins->CanBeMoved which for NewInstance is false.
3543 ins->MoveBefore(blk->GetFirstInstruction(), /*do_checks=*/false);
3544 }
3545 }
3546 }
3547 }
3548
3549 private:
CollectInterestingHeapRefs()3550 void CollectInterestingHeapRefs() {
3551 // Get all the partials we need to move around.
3552 for (size_t i = 0; i < lse_->heap_location_collector_.GetNumberOfHeapLocations(); ++i) {
3553 ReferenceInfo* ri = lse_->heap_location_collector_.GetHeapLocation(i)->GetReferenceInfo();
3554 if (ri->IsPartialSingleton() &&
3555 ri->GetReference()->GetBlock() != nullptr &&
3556 ri->GetNoEscapeSubgraph()->ContainsBlock(ri->GetReference()->GetBlock())) {
3557 RecordHeapRefField(ri->GetReference()->AsNewInstance(), i);
3558 }
3559 }
3560 }
3561
RecordHeapRefField(HNewInstance * ni,size_t loc)3562 void RecordHeapRefField(HNewInstance* ni, size_t loc) {
3563 DCHECK(ni != nullptr);
3564 // This is likely to be very short so just do a linear search.
3565 auto it = std::find_if(heap_refs_.begin(), heap_refs_.end(), [&](HeapReferenceData& data) {
3566 return data.OriginalNewInstance() == ni;
3567 });
3568 HeapReferenceData& cur_ref =
3569 (it == heap_refs_.end())
3570 ? heap_refs_.emplace_back(this,
3571 ni,
3572 lse_->heap_location_collector_.GetHeapLocation(loc)
3573 ->GetReferenceInfo()
3574 ->GetNoEscapeSubgraph(),
3575 alloc_)
3576 : *it;
3577 cur_ref.AddHeapLocation(loc);
3578 }
3579
3580
NotifyNewMaterialization(HInstruction * ins)3581 void NotifyNewMaterialization(HInstruction* ins) {
3582 if (ins->IsPhi()) {
3583 new_ref_phis_.push_back(ins->AsPhi());
3584 }
3585 }
3586
GetMaterializationBlockIndex(HBasicBlock * blk,size_t pred_num) const3587 size_t GetMaterializationBlockIndex(HBasicBlock* blk, size_t pred_num) const {
3588 DCHECK_LT(blk->GetBlockId(), first_materialization_block_id_)
3589 << "block is a materialization block!";
3590 DCHECK_LT(pred_num, max_preds_per_block_);
3591 return blk->GetBlockId() * max_preds_per_block_ + pred_num;
3592 }
3593
GetGraph() const3594 HGraph* GetGraph() const {
3595 return lse_->GetGraph();
3596 }
3597
3598 LSEVisitor* lse_;
3599 ScopedArenaAllocator* alloc_;
3600 ScopedArenaVector<HInstruction*> new_ref_phis_;
3601 ScopedArenaVector<HeapReferenceData> heap_refs_;
3602 size_t max_preds_per_block_;
3603 // An array of (# of non-materialization blocks) * max_preds_per_block
3604 // arranged in block-id major order. Since we can only have at most one
3605 // materialization block on each edge this is the maximum possible number of
3606 // materialization blocks.
3607 ScopedArenaVector<HBasicBlock*> materialization_blocks_;
3608 size_t first_materialization_block_id_;
3609
3610 friend void LSEVisitor::MovePartialEscapes();
3611 };
3612
3613 // Work around c++ type checking annoyances with not being able to forward-declare inner types.
3614 class HeapRefHolder
3615 : public std::reference_wrapper<PartialLoadStoreEliminationHelper::HeapReferenceData> {};
3616
SetupPartialMaterialization(PartialLoadStoreEliminationHelper & helper,HeapRefHolder && holder,size_t pred_idx,HBasicBlock * entry)3617 HInstruction* LSEVisitor::SetupPartialMaterialization(PartialLoadStoreEliminationHelper& helper,
3618 HeapRefHolder&& holder,
3619 size_t pred_idx,
3620 HBasicBlock* entry) {
3621 PartialLoadStoreEliminationHelper::HeapReferenceData& ref_data = holder.get();
3622 HBasicBlock* old_pred = entry->GetPredecessors()[pred_idx];
3623 HInstruction* new_inst = ref_data.OriginalNewInstance();
3624 if (UNLIKELY(!new_inst->GetBlock()->Dominates(entry))) {
3625 LSE_VLOG << "Initial materialization in non-dominating block " << entry->GetBlockId()
3626 << " is null!";
3627 return GetGraph()->GetNullConstant();
3628 }
3629 HBasicBlock* bb = helper.GetOrCreateMaterializationBlock(entry, pred_idx);
3630 CHECK(bb != nullptr) << "entry " << entry->GetBlockId() << " -> " << old_pred->GetBlockId();
3631 HNewInstance* repl_create = new_inst->Clone(GetGraph()->GetAllocator())->AsNewInstance();
3632 repl_create->SetPartialMaterialization();
3633 bb->InsertInstructionBefore(repl_create, bb->GetLastInstruction());
3634 repl_create->CopyEnvironmentFrom(new_inst->GetEnvironment());
3635 MaybeRecordStat(stats_, MethodCompilationStat::kPartialAllocationMoved);
3636 LSE_VLOG << "In blk " << bb->GetBlockId() << " initial materialization is " << *repl_create;
3637 ref_data.AddMaterialization(bb, repl_create);
3638 const FieldInfo* info = nullptr;
3639 for (const HeapLocation* loc : ref_data.IterateLocations()) {
3640 size_t loc_off = heap_location_collector_.GetHeapLocationIndex(loc);
3641 info = field_infos_[loc_off];
3642 DCHECK(loc->GetIndex() == nullptr);
3643 Value value = ReplacementOrValue(heap_values_for_[old_pred->GetBlockId()][loc_off].value);
3644 if (value.NeedsLoopPhi() || value.IsMergedUnknown()) {
3645 Value repl = phi_placeholder_replacements_[PhiPlaceholderIndex(value.GetPhiPlaceholder())];
3646 DCHECK(repl.IsDefault() || repl.IsInvalid() || repl.IsInstruction())
3647 << repl << " from " << value << " pred is " << old_pred->GetBlockId();
3648 if (!repl.IsInvalid()) {
3649 value = repl;
3650 } else {
3651 FullyMaterializePhi(value.GetPhiPlaceholder(), info->GetFieldType());
3652 value = phi_placeholder_replacements_[PhiPlaceholderIndex(value.GetPhiPlaceholder())];
3653 }
3654 } else if (value.NeedsNonLoopPhi()) {
3655 Value repl = phi_placeholder_replacements_[PhiPlaceholderIndex(value.GetPhiPlaceholder())];
3656 DCHECK(repl.IsDefault() || repl.IsInvalid() || repl.IsInstruction())
3657 << repl << " from " << value << " pred is " << old_pred->GetBlockId();
3658 if (!repl.IsInvalid()) {
3659 value = repl;
3660 } else {
3661 MaterializeNonLoopPhis(value.GetPhiPlaceholder(), info->GetFieldType());
3662 value = phi_placeholder_replacements_[PhiPlaceholderIndex(value.GetPhiPlaceholder())];
3663 }
3664 }
3665 DCHECK(value.IsDefault() || value.IsInstruction())
3666 << GetGraph()->PrettyMethod() << ": " << value;
3667
3668 if (!value.IsDefault() &&
3669 // shadow$_klass_ doesn't need to be manually initialized.
3670 MemberOffset(loc->GetOffset()) != mirror::Object::ClassOffset()) {
3671 CHECK(info != nullptr);
3672 HInstruction* set_value =
3673 new (GetGraph()->GetAllocator()) HInstanceFieldSet(repl_create,
3674 value.GetInstruction(),
3675 field_infos_[loc_off]->GetField(),
3676 loc->GetType(),
3677 MemberOffset(loc->GetOffset()),
3678 false,
3679 field_infos_[loc_off]->GetFieldIndex(),
3680 loc->GetDeclaringClassDefIndex(),
3681 field_infos_[loc_off]->GetDexFile(),
3682 0u);
3683 bb->InsertInstructionAfter(set_value, repl_create);
3684 LSE_VLOG << "Adding " << *set_value << " for materialization setup!";
3685 }
3686 }
3687 return repl_create;
3688 }
3689
GetPartialValueAt(HNewInstance * orig_new_inst,HInstruction * read)3690 HInstruction* LSEVisitor::GetPartialValueAt(HNewInstance* orig_new_inst, HInstruction* read) {
3691 size_t loc = heap_location_collector_.GetFieldHeapLocation(orig_new_inst, &read->GetFieldInfo());
3692 Value pred = ReplacementOrValue(intermediate_values_.find(read)->second);
3693 LSE_VLOG << "using " << pred << " as default value for " << *read;
3694 if (pred.IsInstruction()) {
3695 return pred.GetInstruction();
3696 } else if (pred.IsMergedUnknown() || pred.NeedsPhi()) {
3697 FullyMaterializePhi(pred.GetPhiPlaceholder(),
3698 heap_location_collector_.GetHeapLocation(loc)->GetType());
3699 HInstruction* res = Replacement(pred).GetInstruction();
3700 LSE_VLOG << pred << " materialized to " << res->DumpWithArgs();
3701 return res;
3702 } else if (pred.IsDefault()) {
3703 HInstruction* res = GetDefaultValue(read->GetType());
3704 LSE_VLOG << pred << " materialized to " << res->DumpWithArgs();
3705 return res;
3706 }
3707 LOG(FATAL) << "Unable to find unescaped value at " << read->DumpWithArgs()
3708 << "! This should be impossible! Value is " << pred;
3709 UNREACHABLE();
3710 }
3711
MovePartialEscapes()3712 void LSEVisitor::MovePartialEscapes() {
3713 if (!ShouldPerformPartialLSE()) {
3714 return;
3715 }
3716
3717 ScopedArenaAllocator saa(allocator_.GetArenaStack());
3718 PartialLoadStoreEliminationHelper helper(this, &saa);
3719
3720 // Since for PHIs we now will have more information (since we know the object
3721 // hasn't escaped) we need to clear the old phi-replacements where we weren't
3722 // able to find the value.
3723 PrepareForPartialPhiComputation();
3724
3725 for (PartialLoadStoreEliminationHelper::HeapReferenceData& ref_data : helper.GetHeapRefs()) {
3726 LSE_VLOG << "Creating materializations for " << *ref_data.OriginalNewInstance();
3727 // Setup entry and exit blocks.
3728 for (const auto& excluded_cohort : ref_data.GetNoEscapeSubgraph()->GetExcludedCohorts()) {
3729 // Setup materialization blocks.
3730 for (HBasicBlock* entry : excluded_cohort.EntryBlocksReversePostOrder()) {
3731 // Setup entries.
3732 // TODO Assuming we correctly break critical edges every entry block
3733 // must have only a single predecessor so we could just put all this
3734 // stuff in there. OTOH simplifier can do it for us and this is simpler
3735 // to implement - giving clean separation between the original graph and
3736 // materialization blocks - so for now we might as well have these new
3737 // blocks.
3738 ScopedArenaAllocator pred_alloc(saa.GetArenaStack());
3739 ScopedArenaVector<HInstruction*> pred_vals(pred_alloc.Adapter(kArenaAllocLSE));
3740 pred_vals.reserve(entry->GetNumberOfPredecessors());
3741 for (const auto& [pred, pred_idx] :
3742 ZipCount(MakeIterationRange(entry->GetPredecessors()))) {
3743 DCHECK(!helper.IsMaterializationBlock(pred));
3744 if (excluded_cohort.IsEntryBlock(pred)) {
3745 pred_vals.push_back(ref_data.GetMaterialization(pred));
3746 continue;
3747 } else {
3748 pred_vals.push_back(SetupPartialMaterialization(helper, {ref_data}, pred_idx, entry));
3749 }
3750 }
3751 ref_data.GenerateMaterializationValueFromPredecessorsForEntry(entry, pred_vals);
3752 }
3753
3754 // Setup exit block heap-values for later phi-generation.
3755 for (HBasicBlock* exit : excluded_cohort.ExitBlocks()) {
3756 // mark every exit of cohorts as having a value so we can easily
3757 // materialize the PHIs.
3758 // TODO By setting this we can easily use the normal MaterializeLoopPhis
3759 // (via FullyMaterializePhis) in order to generate the default-values
3760 // for predicated-gets. This has the unfortunate side effect of creating
3761 // somewhat more phis than are really needed (in some cases). We really
3762 // should try to eventually know that we can lower these PHIs to only
3763 // the non-escaping value in cases where it is possible. Currently this
3764 // is done to some extent in instruction_simplifier but we have more
3765 // information here to do the right thing.
3766 for (const HeapLocation* loc : ref_data.IterateLocations()) {
3767 size_t loc_off = heap_location_collector_.GetHeapLocationIndex(loc);
3768 // This Value::Default() is only used to fill in PHIs used as the
3769 // default value for PredicatedInstanceFieldGets. The actual value
3770 // stored there is meaningless since the Predicated-iget will use the
3771 // actual field value instead on these paths.
3772 heap_values_for_[exit->GetBlockId()][loc_off].value = Value::Default();
3773 }
3774 }
3775 }
3776
3777 // string materialization through the graph.
3778 // // Visit RPO to PHI the materialized object through the cohort.
3779 for (HBasicBlock* blk : GetGraph()->GetReversePostOrder()) {
3780 // NB This doesn't include materialization blocks.
3781 DCHECK(!helper.IsMaterializationBlock(blk))
3782 << "Materialization blocks should not be in RPO yet.";
3783 if (ref_data.HasMaterialization(blk)) {
3784 continue;
3785 } else if (ref_data.BeforeAllEscapes(blk)) {
3786 ref_data.AddMaterialization(blk, GetGraph()->GetNullConstant());
3787 continue;
3788 } else {
3789 ref_data.GenerateMaterializationValueFromPredecessors(blk);
3790 }
3791 }
3792 }
3793
3794 // Once we've generated all the materializations we can update the users.
3795 helper.FixupPartialObjectUsers();
3796
3797 // Actually put materialization blocks into the graph
3798 helper.InsertMaterializationBlocks();
3799
3800 // Get rid of the original instructions.
3801 helper.RemoveReplacedInstructions();
3802
3803 // Ensure everything is ordered correctly in the materialization blocks. This
3804 // involves moving every NewInstance to the top and ordering them so that any
3805 // required env-uses are correctly ordered.
3806 helper.ReorderMaterializationsForEnvDominance();
3807 }
3808
FinishFullLSE()3809 void LSEVisitor::FinishFullLSE() {
3810 // Remove recorded load instructions that should be eliminated.
3811 for (const LoadStoreRecord& record : loads_and_stores_) {
3812 size_t id = dchecked_integral_cast<size_t>(record.load_or_store->GetId());
3813 HInstruction* substitute = substitute_instructions_for_loads_[id];
3814 if (substitute == nullptr) {
3815 continue;
3816 }
3817 HInstruction* load = record.load_or_store;
3818 DCHECK(load != nullptr);
3819 DCHECK(IsLoad(load));
3820 DCHECK(load->GetBlock() != nullptr) << load->DebugName() << "@" << load->GetDexPc();
3821 // We proactively retrieve the substitute for a removed load, so
3822 // a load that has a substitute should not be observed as a heap
3823 // location value.
3824 DCHECK_EQ(FindSubstitute(substitute), substitute);
3825
3826 load->ReplaceWith(substitute);
3827 load->GetBlock()->RemoveInstruction(load);
3828 }
3829
3830 // Remove all the stores we can.
3831 for (const LoadStoreRecord& record : loads_and_stores_) {
3832 bool is_store = record.load_or_store->GetSideEffects().DoesAnyWrite();
3833 DCHECK_EQ(is_store, IsStore(record.load_or_store));
3834 if (is_store && !kept_stores_.IsBitSet(record.load_or_store->GetId())) {
3835 record.load_or_store->GetBlock()->RemoveInstruction(record.load_or_store);
3836 }
3837 }
3838
3839 // Eliminate singleton-classified instructions:
3840 // * - Constructor fences (they never escape this thread).
3841 // * - Allocations (if they are unused).
3842 for (HInstruction* new_instance : singleton_new_instances_) {
3843 size_t removed = HConstructorFence::RemoveConstructorFences(new_instance);
3844 MaybeRecordStat(stats_,
3845 MethodCompilationStat::kConstructorFenceRemovedLSE,
3846 removed);
3847
3848 if (!new_instance->HasNonEnvironmentUses()) {
3849 new_instance->RemoveEnvironmentUsers();
3850 new_instance->GetBlock()->RemoveInstruction(new_instance);
3851 MaybeRecordStat(stats_, MethodCompilationStat::kFullLSEAllocationRemoved);
3852 }
3853 }
3854 }
3855
3856 // The LSEVisitor is a ValueObject (indirectly through base classes) and therefore
3857 // cannot be directly allocated with an arena allocator, so we need to wrap it.
3858 class LSEVisitorWrapper : public DeletableArenaObject<kArenaAllocLSE> {
3859 public:
LSEVisitorWrapper(HGraph * graph,const HeapLocationCollector & heap_location_collector,bool perform_partial_lse,OptimizingCompilerStats * stats)3860 LSEVisitorWrapper(HGraph* graph,
3861 const HeapLocationCollector& heap_location_collector,
3862 bool perform_partial_lse,
3863 OptimizingCompilerStats* stats)
3864 : lse_visitor_(graph, heap_location_collector, perform_partial_lse, stats) {}
3865
Run()3866 void Run() {
3867 lse_visitor_.Run();
3868 }
3869
3870 private:
3871 LSEVisitor lse_visitor_;
3872 };
3873
Run(bool enable_partial_lse)3874 bool LoadStoreElimination::Run(bool enable_partial_lse) {
3875 if (graph_->IsDebuggable() || graph_->HasTryCatch()) {
3876 // Debugger may set heap values or trigger deoptimization of callers.
3877 // Try/catch support not implemented yet.
3878 // Skip this optimization.
3879 return false;
3880 }
3881 // We need to be able to determine reachability. Clear it just to be safe but
3882 // this should initially be empty.
3883 graph_->ClearReachabilityInformation();
3884 // This is O(blocks^3) time complexity. It means we can query reachability in
3885 // O(1) though.
3886 graph_->ComputeReachabilityInformation();
3887 ScopedArenaAllocator allocator(graph_->GetArenaStack());
3888 LoadStoreAnalysis lsa(graph_,
3889 stats_,
3890 &allocator,
3891 enable_partial_lse ? LoadStoreAnalysisType::kFull
3892 : LoadStoreAnalysisType::kNoPredicatedInstructions);
3893 lsa.Run();
3894 const HeapLocationCollector& heap_location_collector = lsa.GetHeapLocationCollector();
3895 if (heap_location_collector.GetNumberOfHeapLocations() == 0) {
3896 // No HeapLocation information from LSA, skip this optimization.
3897 return false;
3898 }
3899
3900 std::unique_ptr<LSEVisitorWrapper> lse_visitor(new (&allocator) LSEVisitorWrapper(
3901 graph_, heap_location_collector, enable_partial_lse, stats_));
3902 lse_visitor->Run();
3903 return true;
3904 }
3905
3906 #undef LSE_VLOG
3907
3908 } // namespace art
3909