1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "gc/shared/tlab_globals.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/objArrayKlass.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/compile.hpp"
41 #include "opto/connode.hpp"
42 #include "opto/convertnode.hpp"
43 #include "opto/inlinetypenode.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/matcher.hpp"
47 #include "opto/memnode.hpp"
48 #include "opto/mempointer.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/phaseX.hpp"
52 #include "opto/regmask.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/traceMergeStoresTag.hpp"
55 #include "opto/vectornode.hpp"
56 #include "utilities/align.hpp"
57 #include "utilities/copy.hpp"
58 #include "utilities/macros.hpp"
59 #include "utilities/powerOfTwo.hpp"
60 #include "utilities/vmError.hpp"
61
62 // Portions of code courtesy of Clifford Click
63
64 // Optimization - Graph Style
65
66 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
67
68 //=============================================================================
69 uint MemNode::size_of() const { return sizeof(*this); }
70
71 const TypePtr *MemNode::adr_type() const {
72 Node* adr = in(Address);
73 if (adr == nullptr) return nullptr; // node is dead
74 const TypePtr* cross_check = nullptr;
75 DEBUG_ONLY(cross_check = _adr_type);
76 return calculate_adr_type(adr->bottom_type(), cross_check);
77 }
78
79 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
80 if (adr != nullptr) {
81 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
82 return true;
83 }
84 }
85 return false;
86 }
87
88 #ifndef PRODUCT
89 void MemNode::dump_spec(outputStream *st) const {
90 if (in(Address) == nullptr) return; // node is dead
91 #ifndef ASSERT
92 // fake the missing field
93 const TypePtr* _adr_type = nullptr;
94 if (in(Address) != nullptr)
95 _adr_type = in(Address)->bottom_type()->isa_ptr();
96 #endif
97 dump_adr_type(this, _adr_type, st);
98
99 Compile* C = Compile::current();
100 if (C->alias_type(_adr_type)->is_volatile()) {
101 st->print(" Volatile!");
102 }
103 if (_unaligned_access) {
104 st->print(" unaligned");
105 }
106 if (_mismatched_access) {
107 st->print(" mismatched");
108 }
109 if (_unsafe_access) {
110 st->print(" unsafe");
111 }
112 }
113
114 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
115 st->print(" @");
116 if (adr_type == nullptr) {
117 st->print("null");
118 } else {
119 adr_type->dump_on(st);
120 Compile* C = Compile::current();
121 Compile::AliasType* atp = nullptr;
122 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
123 if (atp == nullptr)
124 st->print(", idx=?\?;");
125 else if (atp->index() == Compile::AliasIdxBot)
126 st->print(", idx=Bot;");
127 else if (atp->index() == Compile::AliasIdxTop)
128 st->print(", idx=Top;");
129 else if (atp->index() == Compile::AliasIdxRaw)
130 st->print(", idx=Raw;");
131 else {
132 ciField* field = atp->field();
133 if (field) {
134 st->print(", name=");
135 field->print_name_on(st);
136 }
137 st->print(", idx=%d;", atp->index());
138 }
139 }
140 }
141
142 extern void print_alias_types();
143
144 #endif
145
146 // If call is a constructor call on receiver, returns the class which declares the target method,
147 // else returns nullptr. This information can then be used to deduce if call modifies a field of
148 // receiver. Specifically, if the field is declared in a class that is a subclass of the one
149 // declaring the constructor, then the field is set inside the constructor, else the field must be
150 // set before the constructor invocation. E.g. A field Super.x will be set during the execution of
151 // Sub::<init>, while a field Sub.y must be set before Super::<init> is invoked.
152 static ciInstanceKlass* find_constructor_call_method_holder(Node* call, Node* receiver) {
153 if (!call->is_CallJava()) {
154 return nullptr;
155 }
156
157 ciMethod* target = call->as_CallJava()->method();
158 if (target == nullptr || !target->is_object_constructor()) {
159 return nullptr;
160 }
161
162 assert(call->req() > TypeFunc::Parms, "constructor must have at least 1 argument");
163 Node* parm = call->in(TypeFunc::Parms)->uncast();
164 receiver = receiver->uncast();
165 if (parm == receiver || (parm->is_InlineType() && parm->as_InlineType()->get_oop()->uncast() == receiver)) {
166 return target->holder();
167 }
168
169 return nullptr;
170 }
171
172 // Find the memory output corresponding to the fall-through path of a call
173 static Node* find_call_fallthrough_mem_output(CallNode* call) {
174 ResourceMark rm;
175 CallProjections* projs = call->extract_projections(false, false);
176 Node* res = projs->fallthrough_memproj;
177 assert(res != nullptr, "must have a fallthrough mem output");
178 return res;
179 }
180
181 // Try to find a better memory input for a load from a strict final field of an object that is
182 // allocated in the current compilation unit, or is the first parameter when we are in a
183 // constructor
184 static Node* optimize_strict_final_load_memory_from_local_object(ciField* field, ProjNode* base_uncasted) {
185 if (!EnableValhalla) {
186 // In this method we depends on the fact that strict fields are set before the invocation of
187 // super(), I'm not sure if this is true without Valhalla
188 return nullptr;
189 }
190
191 // The node that can be passed into a constructor
192 Node* base = base_uncasted;
193 if (!base_uncasted->is_Parm()) {
194 assert(base_uncasted->_con == AllocateNode::RawAddress && base_uncasted->in(0)->is_Allocate(), "must be the RawAddress of an AllocateNode");
195 base = base_uncasted->in(0)->as_Allocate()->result_cast();
196 assert(base != nullptr && base->in(1) == base_uncasted, "must find a valid base");
197 }
198
199 // Try to see if there is a constructor call on the base
200 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) {
201 Node* out = base->fast_out(i);
202 ciInstanceKlass* target_holder = find_constructor_call_method_holder(out, base);
203 if (target_holder == nullptr) {
204 continue;
205 } else if (target_holder->is_subclass_of(field->holder())) {
206 return find_call_fallthrough_mem_output(out->as_CallJava());
207 } else {
208 Node* res = out->in(TypeFunc::Memory);
209 assert(res != nullptr, "should have a memory input");
210 return res;
211 }
212 }
213
214 return nullptr;
215 }
216
217 // Try to find a better memory input for a load from a strict final field
218 static Node* try_optimize_strict_final_load_memory(PhaseGVN* phase, ciField* field, Node* adr, ProjNode*& base_local) {
219 intptr_t offset = 0;
220 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
221 if (base == nullptr) {
222 return nullptr;
223 }
224
225 Node* base_uncasted = base->uncast();
226 if (base_uncasted->is_Proj()) {
227 MultiNode* multi = base_uncasted->in(0)->as_Multi();
228 if (multi->is_Allocate()) {
229 // The result of an AllocateNode, try to find the constructor call
230 base_local = base_uncasted->as_Proj();
231 return optimize_strict_final_load_memory_from_local_object(field, base_uncasted->as_Proj());
232 } else if (multi->is_Call()) {
233 // The oop is returned from a call, the memory can be the fallthrough output of the call
234 return find_call_fallthrough_mem_output(multi->as_Call());
235 } else if (multi->is_Start()) {
236 // The oop is a parameter
237 if (phase->C->method()->is_object_constructor() && base_uncasted->as_Proj()->_con == TypeFunc::Parms) {
238 // The receiver of a constructor is similar to the result of an AllocateNode
239 base_local = base_uncasted->as_Proj();
240 return optimize_strict_final_load_memory_from_local_object(field, base_uncasted->as_Proj());
241 } else {
242 // Use the start memory otherwise
243 return multi->proj_out(TypeFunc::Memory);
244 }
245 }
246 }
247
248 return nullptr;
249 }
250
251 // Whether a call can modify a strict final field of base_local, given that base_local is allocated
252 // inside the current compilation unit, or is the first parameter when the compilation root is a
253 // constructor. This is equivalent to asking whether base_local is the receiver of the constructor
254 // invocation call and the class declaring the target method is a subclass of the class declaring
255 // field.
256 static bool call_can_modify_local_object(ciField* field, CallNode* call, Node* base_local) {
257 ciInstanceKlass* target_holder = find_constructor_call_method_holder(call, base_local);
258 return target_holder != nullptr && target_holder->is_subclass_of(field->holder());
259 }
260
261 Node* MemNode::optimize_simple_memory_chain(Node* mchain, const TypeOopPtr* t_oop, Node* load, PhaseGVN* phase) {
262 assert(t_oop != nullptr, "sanity");
263 bool is_instance = t_oop->is_known_instance_field();
264
265 ciField* field = phase->C->alias_type(t_oop)->field();
266 bool is_strict_final_load = false;
267
268 // After macro expansion, an allocation may become a call, changing the memory input to the
269 // memory output of that call would be illegal. As a result, disallow this transformation after
270 // macro expansion.
271 if (phase->is_IterGVN() && phase->C->allow_macro_nodes() && load != nullptr && load->is_Load() && !load->as_Load()->is_mismatched_access()) {
272 if (EnableValhalla) {
273 if (field != nullptr && (field->holder()->is_inlinetype() || field->holder()->is_abstract_value_klass())) {
274 is_strict_final_load = true;
275 }
276 #ifdef ASSERT
277 if (t_oop->is_inlinetypeptr() && t_oop->inline_klass()->contains_field_offset(t_oop->offset())) {
278 assert(is_strict_final_load, "sanity check for basic cases");
279 }
280 #endif
281 } else {
282 is_strict_final_load = field != nullptr && t_oop->is_ptr_to_boxed_value();
283 }
284 }
285
286 if (!is_instance && !is_strict_final_load) {
287 return mchain;
288 }
289
290 Node* result = mchain;
291 ProjNode* base_local = nullptr;
292
293 if (is_strict_final_load) {
294 Node* adr = load->in(MemNode::Address);
295 assert(phase->type(adr) == t_oop, "inconsistent type");
296 Node* tmp = try_optimize_strict_final_load_memory(phase, field, adr, base_local);
297 if (tmp != nullptr) {
298 result = tmp;
299 }
300 }
301
302 uint instance_id = t_oop->instance_id();
303 Node* start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
304 Node* prev = nullptr;
305 while (prev != result) {
306 prev = result;
307 if (result == start_mem) {
308 // start_mem is the earliest memory possible
309 break;
310 }
311
312 // skip over a call which does not affect this memory slice
313 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
314 Node* proj_in = result->in(0);
315 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
316 // This is the allocation that creates the object from which we are loading from
317 break;
318 } else if (proj_in->is_Call()) {
319 // ArrayCopyNodes processed here as well
320 CallNode* call = proj_in->as_Call();
321 if (!call->may_modify(t_oop, phase)) {
322 result = call->in(TypeFunc::Memory);
323 } else if (is_strict_final_load && base_local != nullptr && !call_can_modify_local_object(field, call, base_local)) {
324 result = call->in(TypeFunc::Memory);
325 }
326 } else if (proj_in->is_Initialize()) {
327 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
328 // Stop if this is the initialization for the object instance which
329 // contains this memory slice, otherwise skip over it.
330 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
331 break;
332 }
333 if (is_instance) {
334 result = proj_in->in(TypeFunc::Memory);
335 } else if (is_strict_final_load) {
336 Node* klass = alloc->in(AllocateNode::KlassNode);
337 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
338 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
339 // Allocation of another type, must be another object
340 result = proj_in->in(TypeFunc::Memory);
341 } else if (base_local != nullptr && (base_local->is_Parm() || base_local->in(0) != alloc)) {
342 // Allocation of another object
343 result = proj_in->in(TypeFunc::Memory);
344 }
345 }
346 } else if (proj_in->is_MemBar()) {
347 ArrayCopyNode* ac = nullptr;
348 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
349 break;
350 }
351 result = proj_in->in(TypeFunc::Memory);
352 } else if (proj_in->is_top()) {
353 break; // dead code
354 } else {
355 assert(false, "unexpected projection");
356 }
357 } else if (result->is_ClearArray()) {
358 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
359 // Can not bypass initialization of the instance
360 // we are looking for.
361 break;
362 }
363 // Otherwise skip it (the call updated 'result' value).
364 } else if (result->is_MergeMem()) {
365 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
366 }
367 }
368 return result;
369 }
370
371 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
372 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
373 if (t_oop == nullptr)
374 return mchain; // don't try to optimize non-oop types
375 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
376 bool is_instance = t_oop->is_known_instance_field();
377 PhaseIterGVN *igvn = phase->is_IterGVN();
378 if (is_instance && igvn != nullptr && result->is_Phi()) {
379 PhiNode *mphi = result->as_Phi();
380 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
381 const TypePtr *t = mphi->adr_type();
382 bool do_split = false;
383 // In the following cases, Load memory input can be further optimized based on
384 // its precise address type
385 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
386 do_split = true;
387 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
388 const TypeOopPtr* mem_t =
389 t->is_oopptr()->cast_to_exactness(true)
390 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
391 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
392 if (t_oop->isa_aryptr()) {
393 mem_t = mem_t->is_aryptr()
394 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
395 ->cast_to_size(t_oop->is_aryptr()->size())
396 ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
397 ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
398 ->with_offset(t_oop->is_aryptr()->offset())
399 ->is_aryptr();
400 }
401 do_split = mem_t == t_oop;
402 }
403 if (do_split) {
404 // clone the Phi with our address type
405 result = mphi->split_out_instance(t_adr, igvn);
406 } else {
407 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
408 }
409 }
410 return result;
411 }
412
413 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
414 uint alias_idx = phase->C->get_alias_index(tp);
415 Node *mem = mmem;
416 #ifdef ASSERT
417 {
418 // Check that current type is consistent with the alias index used during graph construction
419 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
420 bool consistent = adr_check == nullptr || adr_check->empty() ||
421 phase->C->must_alias(adr_check, alias_idx );
422 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
423 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
424 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
425 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
426 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
427 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
428 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
429 // don't assert if it is dead code.
430 consistent = true;
431 }
432 if( !consistent ) {
433 st->print("alias_idx==%d, adr_check==", alias_idx);
434 if( adr_check == nullptr ) {
435 st->print("null");
436 } else {
437 adr_check->dump();
438 }
439 st->cr();
440 print_alias_types();
441 assert(consistent, "adr_check must match alias idx");
442 }
443 }
444 #endif
445 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
446 // means an array I have not precisely typed yet. Do not do any
447 // alias stuff with it any time soon.
448 const TypeOopPtr *toop = tp->isa_oopptr();
449 if (tp->base() != Type::AnyPtr &&
450 !(toop &&
451 toop->isa_instptr() &&
452 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
453 toop->offset() == Type::OffsetBot)) {
454 // compress paths and change unreachable cycles to TOP
455 // If not, we can update the input infinitely along a MergeMem cycle
456 // Equivalent code in PhiNode::Ideal
457 Node* m = phase->transform(mmem);
458 // If transformed to a MergeMem, get the desired slice
459 // Otherwise the returned node represents memory for every slice
460 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
461 // Update input if it is progress over what we have now
462 }
463 return mem;
464 }
465
466 //--------------------------Ideal_common---------------------------------------
467 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
468 // Unhook non-raw memories from complete (macro-expanded) initializations.
469 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
470 // If our control input is a dead region, kill all below the region
471 Node *ctl = in(MemNode::Control);
472 if (ctl && remove_dead_region(phase, can_reshape))
473 return this;
474 ctl = in(MemNode::Control);
475 // Don't bother trying to transform a dead node
476 if (ctl && ctl->is_top()) return NodeSentinel;
477
478 PhaseIterGVN *igvn = phase->is_IterGVN();
479 // Wait if control on the worklist.
480 if (ctl && can_reshape && igvn != nullptr) {
481 Node* bol = nullptr;
482 Node* cmp = nullptr;
483 if (ctl->in(0)->is_If()) {
484 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
485 bol = ctl->in(0)->in(1);
486 if (bol->is_Bool())
487 cmp = ctl->in(0)->in(1)->in(1);
488 }
489 if (igvn->_worklist.member(ctl) ||
490 (bol != nullptr && igvn->_worklist.member(bol)) ||
491 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
492 // This control path may be dead.
493 // Delay this memory node transformation until the control is processed.
494 igvn->_worklist.push(this);
495 return NodeSentinel; // caller will return null
496 }
497 }
498 // Ignore if memory is dead, or self-loop
499 Node *mem = in(MemNode::Memory);
500 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
501 assert(mem != this, "dead loop in MemNode::Ideal");
502
503 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
504 // This memory slice may be dead.
505 // Delay this mem node transformation until the memory is processed.
506 igvn->_worklist.push(this);
507 return NodeSentinel; // caller will return null
508 }
509
510 Node *address = in(MemNode::Address);
511 const Type *t_adr = phase->type(address);
512 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
513
514 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
515 // Unsafe off-heap access with zero address. Remove access and other control users
516 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
517 assert(ctl != nullptr, "unsafe accesses should be control dependent");
518 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
519 Node* u = ctl->fast_out(i);
520 if (u != ctl) {
521 igvn->rehash_node_delayed(u);
522 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
523 --i, imax -= nb;
524 }
525 }
526 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
527 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
528 phase->C->root()->add_req(halt);
529 return this;
530 }
531
532 if (can_reshape && igvn != nullptr &&
533 (igvn->_worklist.member(address) ||
534 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
535 // The address's base and type may change when the address is processed.
536 // Delay this mem node transformation until the address is processed.
537 igvn->_worklist.push(this);
538 return NodeSentinel; // caller will return null
539 }
540
541 // Do NOT remove or optimize the next lines: ensure a new alias index
542 // is allocated for an oop pointer type before Escape Analysis.
543 // Note: C++ will not remove it since the call has side effect.
544 if (t_adr->isa_oopptr()) {
545 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
546 }
547
548 Node* base = nullptr;
549 if (address->is_AddP()) {
550 base = address->in(AddPNode::Base);
551 }
552 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
553 !t_adr->isa_rawptr()) {
554 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
555 // Skip this node optimization if its address has TOP base.
556 return NodeSentinel; // caller will return null
557 }
558
559 // Avoid independent memory operations
560 Node* old_mem = mem;
561
562 // The code which unhooks non-raw memories from complete (macro-expanded)
563 // initializations was removed. After macro-expansion all stores caught
564 // by Initialize node became raw stores and there is no information
565 // which memory slices they modify. So it is unsafe to move any memory
566 // operation above these stores. Also in most cases hooked non-raw memories
567 // were already unhooked by using information from detect_ptr_independence()
568 // and find_previous_store().
569
570 if (mem->is_MergeMem()) {
571 MergeMemNode* mmem = mem->as_MergeMem();
572 const TypePtr *tp = t_adr->is_ptr();
573
574 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
575 }
576
577 if (mem != old_mem) {
578 set_req_X(MemNode::Memory, mem, phase);
579 if (phase->type(mem) == Type::TOP) return NodeSentinel;
580 return this;
581 }
582
583 // let the subclass continue analyzing...
584 return nullptr;
585 }
586
587 // Helper function for proving some simple control dominations.
588 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
589 // Already assumes that 'dom' is available at 'sub', and that 'sub'
590 // is not a constant (dominated by the method's StartNode).
591 // Used by MemNode::find_previous_store to prove that the
592 // control input of a memory operation predates (dominates)
593 // an allocation it wants to look past.
594 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
595 // dominate 'sub', 'DomResult::NotDominate' if not,
596 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
597 // dead code, but at the end of IGVN, we know the definite result
598 // once the dead code is cleaned up.
599 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
600 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
601 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
602 }
603
604 // Check 'dom'. Skip Proj and CatchProj nodes.
605 dom = dom->find_exact_control(dom);
606 if (dom == nullptr || dom->is_top()) {
607 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
608 }
609
610 if (dom == sub) {
611 // For the case when, for example, 'sub' is Initialize and the original
612 // 'dom' is Proj node of the 'sub'.
613 return DomResult::NotDominate;
614 }
615
616 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
617 return DomResult::Dominate;
618 }
619
620 // 'dom' dominates 'sub' if its control edge and control edges
621 // of all its inputs dominate or equal to sub's control edge.
622
623 // Currently 'sub' is either Allocate, Initialize or Start nodes.
624 // Or Region for the check in LoadNode::Ideal();
625 // 'sub' should have sub->in(0) != nullptr.
626 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
627 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
628
629 // Get control edge of 'sub'.
630 Node* orig_sub = sub;
631 sub = sub->find_exact_control(sub->in(0));
632 if (sub == nullptr || sub->is_top()) {
633 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
634 }
635
636 assert(sub->is_CFG(), "expecting control");
637
638 if (sub == dom) {
639 return DomResult::Dominate;
640 }
641
642 if (sub->is_Start() || sub->is_Root()) {
643 return DomResult::NotDominate;
644 }
645
646 {
647 // Check all control edges of 'dom'.
648
649 ResourceMark rm;
650 Node_List nlist;
651 Unique_Node_List dom_list;
652
653 dom_list.push(dom);
654 bool only_dominating_controls = false;
655
656 for (uint next = 0; next < dom_list.size(); next++) {
657 Node* n = dom_list.at(next);
658 if (n == orig_sub) {
659 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
660 }
661 if (!n->is_CFG() && n->pinned()) {
662 // Check only own control edge for pinned non-control nodes.
663 n = n->find_exact_control(n->in(0));
664 if (n == nullptr || n->is_top()) {
665 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
666 }
667 assert(n->is_CFG(), "expecting control");
668 dom_list.push(n);
669 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
670 only_dominating_controls = true;
671 } else if (n->is_CFG()) {
672 DomResult dom_result = n->dominates(sub, nlist);
673 if (dom_result == DomResult::Dominate) {
674 only_dominating_controls = true;
675 } else {
676 return dom_result;
677 }
678 } else {
679 // First, own control edge.
680 Node* m = n->find_exact_control(n->in(0));
681 if (m != nullptr) {
682 if (m->is_top()) {
683 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
684 }
685 dom_list.push(m);
686 }
687 // Now, the rest of edges.
688 uint cnt = n->req();
689 for (uint i = 1; i < cnt; i++) {
690 m = n->find_exact_control(n->in(i));
691 if (m == nullptr || m->is_top()) {
692 continue;
693 }
694 dom_list.push(m);
695 }
696 }
697 }
698 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
699 }
700 }
701
702 //---------------------detect_ptr_independence---------------------------------
703 // Used by MemNode::find_previous_store to prove that two base
704 // pointers are never equal.
705 // The pointers are accompanied by their associated allocations,
706 // if any, which have been previously discovered by the caller.
707 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
708 Node* p2, AllocateNode* a2,
709 PhaseTransform* phase) {
710 // Attempt to prove that these two pointers cannot be aliased.
711 // They may both manifestly be allocations, and they should differ.
712 // Or, if they are not both allocations, they can be distinct constants.
713 // Otherwise, one is an allocation and the other a pre-existing value.
714 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
715 return (p1 != p2) && p1->is_Con() && p2->is_Con();
716 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
717 return (a1 != a2);
718 } else if (a1 != nullptr) { // one allocation a1
719 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
720 return all_controls_dominate(p2, a1);
721 } else { //(a2 != null) // one allocation a2
722 return all_controls_dominate(p1, a2);
723 }
724 return false;
725 }
726
727
728 // Find an arraycopy ac that produces the memory state represented by parameter mem.
729 // Return ac if
730 // (a) can_see_stored_value=true and ac must have set the value for this load or if
731 // (b) can_see_stored_value=false and ac could have set the value for this load or if
732 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
733 // In case (c) change the parameter mem to the memory input of ac to skip it
734 // when searching stored value.
735 // Otherwise return null.
736 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
737 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
738 if (ac != nullptr) {
739 Node* ld_addp = in(MemNode::Address);
740 Node* src = ac->in(ArrayCopyNode::Src);
741 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
742
743 // This is a load from a cloned array. The corresponding arraycopy ac must
744 // have set the value for the load and we can return ac but only if the load
745 // is known to be within bounds. This is checked below.
746 if (ary_t != nullptr && ld_addp->is_AddP()) {
747 Node* ld_offs = ld_addp->in(AddPNode::Offset);
748 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
749 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
750 jlong elemsize = type2aelembytes(ary_elem);
751
752 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
753 const TypeInt* sizetype = ary_t->size();
754
755 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
756 // The load is known to be within bounds. It receives its value from ac.
757 return ac;
758 }
759 // The load is known to be out-of-bounds.
760 }
761 // The load could be out-of-bounds. It must not be hoisted but must remain
762 // dependent on the runtime range check. This is achieved by returning null.
763 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
764 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
765
766 if (ac->is_arraycopy_validated() ||
767 ac->is_copyof_validated() ||
768 ac->is_copyofrange_validated()) {
769 Node* ld_addp = in(MemNode::Address);
770 if (ld_addp->is_AddP()) {
771 Node* ld_base = ld_addp->in(AddPNode::Address);
772 Node* ld_offs = ld_addp->in(AddPNode::Offset);
773
774 Node* dest = ac->in(ArrayCopyNode::Dest);
775
776 if (dest == ld_base) {
777 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
778 assert(!ld_offs_t->empty(), "dead reference should be checked already");
779 // Take into account vector or unsafe access size
780 jlong ld_size_in_bytes = (jlong)memory_size();
781 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
782 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
783 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
784 return ac;
785 }
786 if (!can_see_stored_value) {
787 mem = ac->in(TypeFunc::Memory);
788 return ac;
789 }
790 }
791 }
792 }
793 }
794 return nullptr;
795 }
796
797 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
798 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
799 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
800 if (ld_alloc != nullptr) {
801 // Check if there is an array copy for a clone
802 Node* mb = mem->in(0);
803 ArrayCopyNode* ac = nullptr;
804 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
805 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
806 ac = mb->in(0)->in(0)->as_ArrayCopy();
807 } else {
808 // Step over GC barrier when ReduceInitialCardMarks is disabled
809 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
810 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
811
812 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
813 ac = control_proj_ac->in(0)->as_ArrayCopy();
814 }
815 }
816
817 if (ac != nullptr && ac->is_clonebasic()) {
818 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
819 if (alloc != nullptr && alloc == ld_alloc) {
820 return ac;
821 }
822 }
823 }
824 }
825 return nullptr;
826 }
827
828 // The logic for reordering loads and stores uses four steps:
829 // (a) Walk carefully past stores and initializations which we
830 // can prove are independent of this load.
831 // (b) Observe that the next memory state makes an exact match
832 // with self (load or store), and locate the relevant store.
833 // (c) Ensure that, if we were to wire self directly to the store,
834 // the optimizer would fold it up somehow.
835 // (d) Do the rewiring, and return, depending on some other part of
836 // the optimizer to fold up the load.
837 // This routine handles steps (a) and (b). Steps (c) and (d) are
838 // specific to loads and stores, so they are handled by the callers.
839 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
840 //
841 Node* MemNode::find_previous_store(PhaseValues* phase) {
842 Node* ctrl = in(MemNode::Control);
843 Node* adr = in(MemNode::Address);
844 intptr_t offset = 0;
845 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
846 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
847
848 if (offset == Type::OffsetBot)
849 return nullptr; // cannot unalias unless there are precise offsets
850
851 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
852 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
853
854 intptr_t size_in_bytes = memory_size();
855
856 Node* mem = in(MemNode::Memory); // start searching here...
857
858 int cnt = 50; // Cycle limiter
859 for (;;) { // While we can dance past unrelated stores...
860 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
861
862 Node* prev = mem;
863 if (mem->is_Store()) {
864 Node* st_adr = mem->in(MemNode::Address);
865 intptr_t st_offset = 0;
866 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
867 if (st_base == nullptr)
868 break; // inscrutable pointer
869
870 // For raw accesses it's not enough to prove that constant offsets don't intersect.
871 // We need the bases to be the equal in order for the offset check to make sense.
872 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
873 break;
874 }
875
876 if (st_offset != offset && st_offset != Type::OffsetBot) {
877 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
878 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
879 if (st_offset >= offset + size_in_bytes ||
880 st_offset <= offset - MAX_STORE ||
881 st_offset <= offset - mem->as_Store()->memory_size()) {
882 // Success: The offsets are provably independent.
883 // (You may ask, why not just test st_offset != offset and be done?
884 // The answer is that stores of different sizes can co-exist
885 // in the same sequence of RawMem effects. We sometimes initialize
886 // a whole 'tile' of array elements with a single jint or jlong.)
887 mem = mem->in(MemNode::Memory);
888 continue; // (a) advance through independent store memory
889 }
890 }
891 if (st_base != base &&
892 detect_ptr_independence(base, alloc,
893 st_base,
894 AllocateNode::Ideal_allocation(st_base),
895 phase)) {
896 // Success: The bases are provably independent.
897 mem = mem->in(MemNode::Memory);
898 continue; // (a) advance through independent store memory
899 }
900
901 // (b) At this point, if the bases or offsets do not agree, we lose,
902 // since we have not managed to prove 'this' and 'mem' independent.
903 if (st_base == base && st_offset == offset) {
904 return mem; // let caller handle steps (c), (d)
905 }
906
907 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
908 InitializeNode* st_init = mem->in(0)->as_Initialize();
909 AllocateNode* st_alloc = st_init->allocation();
910 if (st_alloc == nullptr) {
911 break; // something degenerated
912 }
913 bool known_identical = false;
914 bool known_independent = false;
915 if (alloc == st_alloc) {
916 known_identical = true;
917 } else if (alloc != nullptr) {
918 known_independent = true;
919 } else if (all_controls_dominate(this, st_alloc)) {
920 known_independent = true;
921 }
922
923 if (known_independent) {
924 // The bases are provably independent: Either they are
925 // manifestly distinct allocations, or else the control
926 // of this load dominates the store's allocation.
927 int alias_idx = phase->C->get_alias_index(adr_type());
928 if (alias_idx == Compile::AliasIdxRaw) {
929 mem = st_alloc->in(TypeFunc::Memory);
930 } else {
931 mem = st_init->memory(alias_idx);
932 }
933 continue; // (a) advance through independent store memory
934 }
935
936 // (b) at this point, if we are not looking at a store initializing
937 // the same allocation we are loading from, we lose.
938 if (known_identical) {
939 // From caller, can_see_stored_value will consult find_captured_store.
940 return mem; // let caller handle steps (c), (d)
941 }
942
943 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
944 if (prev != mem) {
945 // Found an arraycopy but it doesn't affect that load
946 continue;
947 }
948 // Found an arraycopy that may affect that load
949 return mem;
950 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
951 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
952 if (mem->is_Proj() && mem->in(0)->is_Call()) {
953 // ArrayCopyNodes processed here as well.
954 CallNode *call = mem->in(0)->as_Call();
955 if (!call->may_modify(addr_t, phase)) {
956 mem = call->in(TypeFunc::Memory);
957 continue; // (a) advance through independent call memory
958 }
959 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
960 ArrayCopyNode* ac = nullptr;
961 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
962 break;
963 }
964 mem = mem->in(0)->in(TypeFunc::Memory);
965 continue; // (a) advance through independent MemBar memory
966 } else if (mem->is_ClearArray()) {
967 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
968 // (the call updated 'mem' value)
969 continue; // (a) advance through independent allocation memory
970 } else {
971 // Can not bypass initialization of the instance
972 // we are looking for.
973 return mem;
974 }
975 } else if (mem->is_MergeMem()) {
976 int alias_idx = phase->C->get_alias_index(adr_type());
977 mem = mem->as_MergeMem()->memory_at(alias_idx);
978 continue; // (a) advance through independent MergeMem memory
979 }
980 }
981
982 // Unless there is an explicit 'continue', we must bail out here,
983 // because 'mem' is an inscrutable memory state (e.g., a call).
984 break;
985 }
986
987 return nullptr; // bail out
988 }
989
990 //----------------------calculate_adr_type-------------------------------------
991 // Helper function. Notices when the given type of address hits top or bottom.
992 // Also, asserts a cross-check of the type against the expected address type.
993 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
994 if (t == Type::TOP) return nullptr; // does not touch memory any more?
995 #ifdef ASSERT
996 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
997 #endif
998 const TypePtr* tp = t->isa_ptr();
999 if (tp == nullptr) {
1000 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
1001 return TypePtr::BOTTOM; // touches lots of memory
1002 } else {
1003 #ifdef ASSERT
1004 // %%%% [phh] We don't check the alias index if cross_check is
1005 // TypeRawPtr::BOTTOM. Needs to be investigated.
1006 if (cross_check != nullptr &&
1007 cross_check != TypePtr::BOTTOM &&
1008 cross_check != TypeRawPtr::BOTTOM) {
1009 // Recheck the alias index, to see if it has changed (due to a bug).
1010 Compile* C = Compile::current();
1011 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
1012 "must stay in the original alias category");
1013 // The type of the address must be contained in the adr_type,
1014 // disregarding "null"-ness.
1015 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
1016 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
1017 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
1018 "real address must not escape from expected memory type");
1019 }
1020 #endif
1021 return tp;
1022 }
1023 }
1024
1025 uint8_t MemNode::barrier_data(const Node* n) {
1026 if (n->is_LoadStore()) {
1027 return n->as_LoadStore()->barrier_data();
1028 } else if (n->is_Mem()) {
1029 return n->as_Mem()->barrier_data();
1030 }
1031 return 0;
1032 }
1033
1034 //=============================================================================
1035 // Should LoadNode::Ideal() attempt to remove control edges?
1036 bool LoadNode::can_remove_control() const {
1037 return !has_pinned_control_dependency();
1038 }
1039 uint LoadNode::size_of() const { return sizeof(*this); }
1040 bool LoadNode::cmp(const Node &n) const {
1041 LoadNode& load = (LoadNode &)n;
1042 return Type::equals(_type, load._type) &&
1043 _control_dependency == load._control_dependency &&
1044 _mo == load._mo;
1045 }
1046 const Type *LoadNode::bottom_type() const { return _type; }
1047 uint LoadNode::ideal_reg() const {
1048 return _type->ideal_reg();
1049 }
1050
1051 #ifndef PRODUCT
1052 void LoadNode::dump_spec(outputStream *st) const {
1053 MemNode::dump_spec(st);
1054 if( !Verbose && !WizardMode ) {
1055 // standard dump does this in Verbose and WizardMode
1056 st->print(" #"); _type->dump_on(st);
1057 }
1058 if (!depends_only_on_test()) {
1059 st->print(" (does not depend only on test, ");
1060 if (control_dependency() == UnknownControl) {
1061 st->print("unknown control");
1062 } else if (control_dependency() == Pinned) {
1063 st->print("pinned");
1064 } else if (adr_type() == TypeRawPtr::BOTTOM) {
1065 st->print("raw access");
1066 } else {
1067 st->print("unknown reason");
1068 }
1069 st->print(")");
1070 }
1071 }
1072 #endif
1073
1074 #ifdef ASSERT
1075 //----------------------------is_immutable_value-------------------------------
1076 // Helper function to allow a raw load without control edge for some cases
1077 bool LoadNode::is_immutable_value(Node* adr) {
1078 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
1079 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
1080
1081 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
1082 int offsets[] = {
1083 in_bytes(JavaThread::osthread_offset()),
1084 in_bytes(JavaThread::threadObj_offset()),
1085 in_bytes(JavaThread::vthread_offset()),
1086 in_bytes(JavaThread::scopedValueCache_offset()),
1087 };
1088
1089 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
1090 if (offset == offsets[i]) {
1091 return true;
1092 }
1093 }
1094 }
1095
1096 return false;
1097 }
1098 #endif
1099
1100 //----------------------------LoadNode::make-----------------------------------
1101 // Polymorphic factory method:
1102 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1103 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1104 Compile* C = gvn.C;
1105
1106 // sanity check the alias category against the created node type
1107 assert(!(adr_type->isa_oopptr() &&
1108 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1109 "use LoadKlassNode instead");
1110 assert(!(adr_type->isa_aryptr() &&
1111 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1112 "use LoadRangeNode instead");
1113 // Check control edge of raw loads
1114 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1115 // oop will be recorded in oop map if load crosses safepoint
1116 rt->isa_oopptr() || is_immutable_value(adr),
1117 "raw memory operations should have control edge");
1118 LoadNode* load = nullptr;
1119 switch (bt) {
1120 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1121 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1122 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1123 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1124 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1125 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1126 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1127 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1128 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1129 case T_OBJECT:
1130 case T_NARROWOOP:
1131 #ifdef _LP64
1132 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1133 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1134 } else
1135 #endif
1136 {
1137 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1138 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1139 }
1140 break;
1141 default:
1142 ShouldNotReachHere();
1143 break;
1144 }
1145 assert(load != nullptr, "LoadNode should have been created");
1146 if (unaligned) {
1147 load->set_unaligned_access();
1148 }
1149 if (mismatched) {
1150 load->set_mismatched_access();
1151 }
1152 if (unsafe) {
1153 load->set_unsafe_access();
1154 }
1155 load->set_barrier_data(barrier_data);
1156 if (load->Opcode() == Op_LoadN) {
1157 Node* ld = gvn.transform(load);
1158 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1159 }
1160
1161 return load;
1162 }
1163
1164 //------------------------------hash-------------------------------------------
1165 uint LoadNode::hash() const {
1166 // unroll addition of interesting fields
1167 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1168 }
1169
1170 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1171 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1172 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1173 bool is_stable_ary = FoldStableValues &&
1174 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1175 tp->isa_aryptr()->is_stable();
1176
1177 return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1178 }
1179
1180 return false;
1181 }
1182
1183 LoadNode* LoadNode::pin_array_access_node() const {
1184 const TypePtr* adr_type = this->adr_type();
1185 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1186 return clone_pinned();
1187 }
1188 return nullptr;
1189 }
1190
1191 // Is the value loaded previously stored by an arraycopy? If so return
1192 // a load node that reads from the source array so we may be able to
1193 // optimize out the ArrayCopy node later.
1194 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1195 Node* ld_adr = in(MemNode::Address);
1196 intptr_t ld_off = 0;
1197 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1198 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1199 if (ac != nullptr) {
1200 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1201
1202 Node* mem = ac->in(TypeFunc::Memory);
1203 Node* ctl = ac->in(0);
1204 Node* src = ac->in(ArrayCopyNode::Src);
1205
1206 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1207 return nullptr;
1208 }
1209
1210 // load depends on the tests that validate the arraycopy
1211 LoadNode* ld = clone_pinned();
1212 Node* addp = in(MemNode::Address)->clone();
1213 if (ac->as_ArrayCopy()->is_clonebasic()) {
1214 assert(ld_alloc != nullptr, "need an alloc");
1215 assert(addp->is_AddP(), "address must be addp");
1216 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1217 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1218 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1219 addp->set_req(AddPNode::Base, src);
1220 addp->set_req(AddPNode::Address, src);
1221 } else {
1222 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1223 ac->as_ArrayCopy()->is_copyof_validated() ||
1224 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1225 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1226 addp->set_req(AddPNode::Base, src);
1227 addp->set_req(AddPNode::Address, src);
1228
1229 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1230 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1231 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1232
1233 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1234 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1235
1236 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1237 #ifdef _LP64
1238 diff = phase->transform(new ConvI2LNode(diff));
1239 #endif
1240 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1241
1242 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1243 addp->set_req(AddPNode::Offset, offset);
1244 }
1245 addp = phase->transform(addp);
1246 #ifdef ASSERT
1247 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1248 ld->_adr_type = adr_type;
1249 #endif
1250 ld->set_req(MemNode::Address, addp);
1251 ld->set_req(0, ctl);
1252 ld->set_req(MemNode::Memory, mem);
1253 return ld;
1254 }
1255 return nullptr;
1256 }
1257
1258 static Node* see_through_inline_type(PhaseValues* phase, const MemNode* load, Node* base, int offset) {
1259 if (!load->is_mismatched_access() && base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) {
1260 InlineTypeNode* vt = base->as_InlineType();
1261 assert(!vt->is_larval(), "must not load from a larval object");
1262 Node* value = vt->field_value_by_offset(offset, true);
1263 assert(value != nullptr, "must see some value");
1264 return value;
1265 }
1266
1267 return nullptr;
1268 }
1269
1270 //---------------------------can_see_stored_value------------------------------
1271 // This routine exists to make sure this set of tests is done the same
1272 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1273 // will change the graph shape in a way which makes memory alive twice at the
1274 // same time (uses the Oracle model of aliasing), then some
1275 // LoadXNode::Identity will fold things back to the equivalence-class model
1276 // of aliasing.
1277 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1278 Node* ld_adr = in(MemNode::Address);
1279 intptr_t ld_off = 0;
1280 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1281 // Try to see through an InlineTypeNode
1282 // LoadN is special because the input is not compressed
1283 if (Opcode() != Op_LoadN) {
1284 Node* value = see_through_inline_type(phase, this, ld_base, ld_off);
1285 if (value != nullptr) {
1286 return value;
1287 }
1288 }
1289
1290 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1291 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1292 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1293 // This is more general than load from boxing objects.
1294 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1295 uint alias_idx = atp->index();
1296 Node* result = nullptr;
1297 Node* current = st;
1298 // Skip through chains of MemBarNodes checking the MergeMems for
1299 // new states for the slice of this load. Stop once any other
1300 // kind of node is encountered. Loads from final memory can skip
1301 // through any kind of MemBar but normal loads shouldn't skip
1302 // through MemBarAcquire since the could allow them to move out of
1303 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1304 // because alias info above them may be inaccurate (e.g., due to
1305 // mixed/mismatched unsafe accesses).
1306 bool is_final_mem = !atp->is_rewritable();
1307 while (current->is_Proj()) {
1308 int opc = current->in(0)->Opcode();
1309 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1310 opc == Op_MemBarAcquireLock ||
1311 opc == Op_LoadFence)) ||
1312 opc == Op_MemBarRelease ||
1313 opc == Op_StoreFence ||
1314 opc == Op_MemBarReleaseLock ||
1315 opc == Op_MemBarStoreStore ||
1316 opc == Op_StoreStoreFence) {
1317 Node* mem = current->in(0)->in(TypeFunc::Memory);
1318 if (mem->is_MergeMem()) {
1319 MergeMemNode* merge = mem->as_MergeMem();
1320 Node* new_st = merge->memory_at(alias_idx);
1321 if (new_st == merge->base_memory()) {
1322 // Keep searching
1323 current = new_st;
1324 continue;
1325 }
1326 // Save the new memory state for the slice and fall through
1327 // to exit.
1328 result = new_st;
1329 }
1330 }
1331 break;
1332 }
1333 if (result != nullptr) {
1334 st = result;
1335 }
1336 }
1337
1338 // Loop around twice in the case Load -> Initialize -> Store.
1339 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1340 for (int trip = 0; trip <= 1; trip++) {
1341
1342 if (st->is_Store()) {
1343 Node* st_adr = st->in(MemNode::Address);
1344 if (st_adr != ld_adr) {
1345 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1346 intptr_t st_off = 0;
1347 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1348 if (ld_base == nullptr) return nullptr;
1349 if (st_base == nullptr) return nullptr;
1350 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1351 if (ld_off != st_off) return nullptr;
1352 if (ld_off == Type::OffsetBot) return nullptr;
1353 // Same base, same offset.
1354 // Possible improvement for arrays: check index value instead of absolute offset.
1355
1356 // At this point we have proven something like this setup:
1357 // B = << base >>
1358 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1359 // S = StoreQ(AddP( B , #Off), V)
1360 // (Actually, we haven't yet proven the Q's are the same.)
1361 // In other words, we are loading from a casted version of
1362 // the same pointer-and-offset that we stored to.
1363 // Casted version may carry a dependency and it is respected.
1364 // Thus, we are able to replace L by V.
1365 }
1366 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1367 if (store_Opcode() != st->Opcode()) {
1368 return nullptr;
1369 }
1370 // LoadVector/StoreVector needs additional check to ensure the types match.
1371 if (st->is_StoreVector()) {
1372 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1373 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1374 if (in_vt != out_vt) {
1375 return nullptr;
1376 }
1377 }
1378 return st->in(MemNode::ValueIn);
1379 }
1380
1381 // A load from a freshly-created object always returns zero.
1382 // (This can happen after LoadNode::Ideal resets the load's memory input
1383 // to find_captured_store, which returned InitializeNode::zero_memory.)
1384 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1385 (st->in(0) == ld_alloc) &&
1386 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1387 // return a zero value for the load's basic type
1388 // (This is one of the few places where a generic PhaseTransform
1389 // can create new nodes. Think of it as lazily manifesting
1390 // virtually pre-existing constants.)
1391 Node* init_value = ld_alloc->in(AllocateNode::InitValue);
1392 if (init_value != nullptr) {
1393 // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
1394 return init_value;
1395 }
1396 assert(ld_alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
1397 if (memory_type() != T_VOID) {
1398 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1399 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1400 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1401 // by the ArrayCopyNode.
1402 return phase->zerocon(memory_type());
1403 }
1404 } else {
1405 // TODO: materialize all-zero vector constant
1406 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1407 }
1408 }
1409
1410 // A load from an initialization barrier can match a captured store.
1411 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1412 InitializeNode* init = st->in(0)->as_Initialize();
1413 AllocateNode* alloc = init->allocation();
1414 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1415 // examine a captured store value
1416 st = init->find_captured_store(ld_off, memory_size(), phase);
1417 if (st != nullptr) {
1418 continue; // take one more trip around
1419 }
1420 }
1421 }
1422
1423 // Load boxed value from result of valueOf() call is input parameter.
1424 if (this->is_Load() && ld_adr->is_AddP() &&
1425 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1426 intptr_t ignore = 0;
1427 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1428 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1429 base = bs->step_over_gc_barrier(base);
1430 if (base != nullptr && base->is_Proj() &&
1431 base->as_Proj()->_con == TypeFunc::Parms &&
1432 base->in(0)->is_CallStaticJava() &&
1433 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1434 return base->in(0)->in(TypeFunc::Parms);
1435 }
1436 }
1437
1438 break;
1439 }
1440
1441 return nullptr;
1442 }
1443
1444 //----------------------is_instance_field_load_with_local_phi------------------
1445 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1446 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1447 in(Address)->is_AddP() ) {
1448 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1449 // Only instances and boxed values.
1450 if( t_oop != nullptr &&
1451 (t_oop->is_ptr_to_boxed_value() ||
1452 t_oop->is_known_instance_field()) &&
1453 t_oop->offset() != Type::OffsetBot &&
1454 t_oop->offset() != Type::OffsetTop) {
1455 return true;
1456 }
1457 }
1458 return false;
1459 }
1460
1461 //------------------------------Identity---------------------------------------
1462 // Loads are identity if previous store is to same address
1463 Node* LoadNode::Identity(PhaseGVN* phase) {
1464 // If the previous store-maker is the right kind of Store, and the store is
1465 // to the same address, then we are equal to the value stored.
1466 Node* mem = in(Memory);
1467 Node* value = can_see_stored_value(mem, phase);
1468 if( value ) {
1469 // byte, short & char stores truncate naturally.
1470 // A load has to load the truncated value which requires
1471 // some sort of masking operation and that requires an
1472 // Ideal call instead of an Identity call.
1473 if (memory_size() < BytesPerInt) {
1474 // If the input to the store does not fit with the load's result type,
1475 // it must be truncated via an Ideal call.
1476 if (!phase->type(value)->higher_equal(phase->type(this)))
1477 return this;
1478 }
1479 // (This works even when value is a Con, but LoadNode::Value
1480 // usually runs first, producing the singleton type of the Con.)
1481 if (!has_pinned_control_dependency() || value->is_Con()) {
1482 return value;
1483 } else {
1484 return this;
1485 }
1486 }
1487
1488 if (has_pinned_control_dependency()) {
1489 return this;
1490 }
1491 // Search for an existing data phi which was generated before for the same
1492 // instance's field to avoid infinite generation of phis in a loop.
1493 Node *region = mem->in(0);
1494 if (is_instance_field_load_with_local_phi(region)) {
1495 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1496 int this_index = phase->C->get_alias_index(addr_t);
1497 int this_offset = addr_t->offset();
1498 int this_iid = addr_t->instance_id();
1499 if (!addr_t->is_known_instance() &&
1500 addr_t->is_ptr_to_boxed_value()) {
1501 // Use _idx of address base (could be Phi node) for boxed values.
1502 intptr_t ignore = 0;
1503 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1504 if (base == nullptr) {
1505 return this;
1506 }
1507 this_iid = base->_idx;
1508 }
1509 const Type* this_type = bottom_type();
1510 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1511 Node* phi = region->fast_out(i);
1512 if (phi->is_Phi() && phi != mem &&
1513 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1514 return phi;
1515 }
1516 }
1517 }
1518
1519 return this;
1520 }
1521
1522 // Construct an equivalent unsigned load.
1523 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1524 BasicType bt = T_ILLEGAL;
1525 const Type* rt = nullptr;
1526 switch (Opcode()) {
1527 case Op_LoadUB: return this;
1528 case Op_LoadUS: return this;
1529 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1530 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1531 default:
1532 assert(false, "no unsigned variant: %s", Name());
1533 return nullptr;
1534 }
1535 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1536 raw_adr_type(), rt, bt, _mo, _control_dependency,
1537 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1538 }
1539
1540 // Construct an equivalent signed load.
1541 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1542 BasicType bt = T_ILLEGAL;
1543 const Type* rt = nullptr;
1544 switch (Opcode()) {
1545 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1546 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1547 case Op_LoadB: // fall through
1548 case Op_LoadS: // fall through
1549 case Op_LoadI: // fall through
1550 case Op_LoadL: return this;
1551 default:
1552 assert(false, "no signed variant: %s", Name());
1553 return nullptr;
1554 }
1555 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1556 raw_adr_type(), rt, bt, _mo, _control_dependency,
1557 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1558 }
1559
1560 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1561 BasicType bt = rt->basic_type();
1562 switch (Opcode()) {
1563 case Op_LoadI: return (bt == T_FLOAT);
1564 case Op_LoadL: return (bt == T_DOUBLE);
1565 case Op_LoadF: return (bt == T_INT);
1566 case Op_LoadD: return (bt == T_LONG);
1567
1568 default: return false;
1569 }
1570 }
1571
1572 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1573 BasicType bt = rt->basic_type();
1574 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1575 bool is_mismatched = is_mismatched_access();
1576 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1577 if (raw_type == nullptr) {
1578 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1579 }
1580 const int op = Opcode();
1581 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1582 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1583 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1584 raw_adr_type(), rt, bt, _mo, _control_dependency,
1585 require_atomic_access, is_unaligned_access(), is_mismatched);
1586 }
1587
1588 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1589 BasicType bt = vt->basic_type();
1590 switch (Opcode()) {
1591 case Op_StoreI: return (bt == T_FLOAT);
1592 case Op_StoreL: return (bt == T_DOUBLE);
1593 case Op_StoreF: return (bt == T_INT);
1594 case Op_StoreD: return (bt == T_LONG);
1595
1596 default: return false;
1597 }
1598 }
1599
1600 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1601 BasicType bt = vt->basic_type();
1602 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1603 const int op = Opcode();
1604 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1605 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1606 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1607 raw_adr_type(), val, bt, _mo, require_atomic_access);
1608
1609 bool is_mismatched = is_mismatched_access();
1610 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1611 if (raw_type == nullptr) {
1612 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1613 }
1614 if (is_mismatched) {
1615 st->set_mismatched_access();
1616 }
1617 return st;
1618 }
1619
1620 // We're loading from an object which has autobox behaviour.
1621 // If this object is result of a valueOf call we'll have a phi
1622 // merging a newly allocated object and a load from the cache.
1623 // We want to replace this load with the original incoming
1624 // argument to the valueOf call.
1625 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1626 assert(igvn->C->eliminate_boxing(), "sanity");
1627 intptr_t ignore = 0;
1628 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1629 if ((base == nullptr) || base->is_Phi()) {
1630 // Push the loads from the phi that comes from valueOf up
1631 // through it to allow elimination of the loads and the recovery
1632 // of the original value. It is done in split_through_phi().
1633 return nullptr;
1634 } else if (base->is_Load() ||
1635 (base->is_DecodeN() && base->in(1)->is_Load())) {
1636 // Eliminate the load of boxed value for integer types from the cache
1637 // array by deriving the value from the index into the array.
1638 // Capture the offset of the load and then reverse the computation.
1639
1640 // Get LoadN node which loads a boxing object from 'cache' array.
1641 if (base->is_DecodeN()) {
1642 base = base->in(1);
1643 }
1644 if (!base->in(Address)->is_AddP()) {
1645 return nullptr; // Complex address
1646 }
1647 AddPNode* address = base->in(Address)->as_AddP();
1648 Node* cache_base = address->in(AddPNode::Base);
1649 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1650 // Get ConP node which is static 'cache' field.
1651 cache_base = cache_base->in(1);
1652 }
1653 if ((cache_base != nullptr) && cache_base->is_Con()) {
1654 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1655 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1656 Node* elements[4];
1657 int shift = exact_log2(type2aelembytes(T_OBJECT));
1658 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1659 if (count > 0 && elements[0]->is_Con() &&
1660 (count == 1 ||
1661 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1662 elements[1]->in(2) == igvn->intcon(shift)))) {
1663 ciObjArray* array = base_type->const_oop()->as_obj_array();
1664 // Fetch the box object cache[0] at the base of the array and get its value
1665 ciInstance* box = array->obj_at(0)->as_instance();
1666 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1667 assert(ik->is_box_klass(), "sanity");
1668 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1669 if (ik->nof_nonstatic_fields() == 1) {
1670 // This should be true nonstatic_field_at requires calling
1671 // nof_nonstatic_fields so check it anyway
1672 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1673 BasicType bt = c.basic_type();
1674 // Only integer types have boxing cache.
1675 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1676 bt == T_BYTE || bt == T_SHORT ||
1677 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1678 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1679 if (cache_low != (int)cache_low) {
1680 return nullptr; // should not happen since cache is array indexed by value
1681 }
1682 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1683 if (offset != (int)offset) {
1684 return nullptr; // should not happen since cache is array indexed by value
1685 }
1686 // Add up all the offsets making of the address of the load
1687 Node* result = elements[0];
1688 for (int i = 1; i < count; i++) {
1689 result = igvn->transform(new AddXNode(result, elements[i]));
1690 }
1691 // Remove the constant offset from the address and then
1692 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1693 // remove the scaling of the offset to recover the original index.
1694 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1695 // Peel the shift off directly but wrap it in a dummy node
1696 // since Ideal can't return existing nodes
1697 igvn->_worklist.push(result); // remove dead node later
1698 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1699 } else if (result->is_Add() && result->in(2)->is_Con() &&
1700 result->in(1)->Opcode() == Op_LShiftX &&
1701 result->in(1)->in(2) == igvn->intcon(shift)) {
1702 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1703 // but for boxing cache access we know that X<<Z will not overflow
1704 // (there is range check) so we do this optimizatrion by hand here.
1705 igvn->_worklist.push(result); // remove dead node later
1706 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1707 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1708 } else {
1709 result = new RShiftXNode(result, igvn->intcon(shift));
1710 }
1711 #ifdef _LP64
1712 if (bt != T_LONG) {
1713 result = new ConvL2INode(igvn->transform(result));
1714 }
1715 #else
1716 if (bt == T_LONG) {
1717 result = new ConvI2LNode(igvn->transform(result));
1718 }
1719 #endif
1720 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1721 // Need to preserve unboxing load type if it is unsigned.
1722 switch(this->Opcode()) {
1723 case Op_LoadUB:
1724 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1725 break;
1726 case Op_LoadUS:
1727 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1728 break;
1729 }
1730 return result;
1731 }
1732 }
1733 }
1734 }
1735 }
1736 return nullptr;
1737 }
1738
1739 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1740 Node* region = phi->in(0);
1741 if (region == nullptr) {
1742 return false; // Wait stable graph
1743 }
1744 uint cnt = phi->req();
1745 for (uint i = 1; i < cnt; i++) {
1746 Node* rc = region->in(i);
1747 if (rc == nullptr || phase->type(rc) == Type::TOP)
1748 return false; // Wait stable graph
1749 Node* in = phi->in(i);
1750 if (in == nullptr || phase->type(in) == Type::TOP)
1751 return false; // Wait stable graph
1752 }
1753 return true;
1754 }
1755
1756 //------------------------------split_through_phi------------------------------
1757 // Check whether a call to 'split_through_phi' would split this load through the
1758 // Phi *base*. This method is essentially a copy of the validations performed
1759 // by 'split_through_phi'. The first use of this method was in EA code as part
1760 // of simplification of allocation merges.
1761 // Some differences from original method (split_through_phi):
1762 // - If base->is_CastPP(): base = base->in(1)
1763 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1764 Node* mem = in(Memory);
1765 Node* address = in(Address);
1766 intptr_t ignore = 0;
1767 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1768
1769 if (base->is_CastPP()) {
1770 base = base->in(1);
1771 }
1772
1773 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1774 return false;
1775 }
1776
1777 if (!mem->is_Phi()) {
1778 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1779 return false;
1780 }
1781 } else if (base->in(0) != mem->in(0)) {
1782 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1783 return false;
1784 }
1785 }
1786
1787 return true;
1788 }
1789
1790 //------------------------------split_through_phi------------------------------
1791 // Split instance or boxed field load through Phi.
1792 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1793 if (req() > 3) {
1794 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1795 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1796 return nullptr;
1797 }
1798 Node* mem = in(Memory);
1799 Node* address = in(Address);
1800 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1801
1802 assert((t_oop != nullptr) &&
1803 (ignore_missing_instance_id ||
1804 t_oop->is_known_instance_field() ||
1805 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1806
1807 Compile* C = phase->C;
1808 intptr_t ignore = 0;
1809 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1810 bool base_is_phi = (base != nullptr) && base->is_Phi();
1811 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1812 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1813 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1814
1815 if (!((mem->is_Phi() || base_is_phi) &&
1816 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1817 return nullptr; // Neither memory or base are Phi
1818 }
1819
1820 if (mem->is_Phi()) {
1821 if (!stable_phi(mem->as_Phi(), phase)) {
1822 return nullptr; // Wait stable graph
1823 }
1824 uint cnt = mem->req();
1825 // Check for loop invariant memory.
1826 if (cnt == 3) {
1827 for (uint i = 1; i < cnt; i++) {
1828 Node* in = mem->in(i);
1829 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1830 if (m == mem) {
1831 if (i == 1) {
1832 // if the first edge was a loop, check second edge too.
1833 // If both are replaceable - we are in an infinite loop
1834 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1835 if (n == mem) {
1836 break;
1837 }
1838 }
1839 set_req(Memory, mem->in(cnt - i));
1840 return this; // made change
1841 }
1842 }
1843 }
1844 }
1845 if (base_is_phi) {
1846 if (!stable_phi(base->as_Phi(), phase)) {
1847 return nullptr; // Wait stable graph
1848 }
1849 uint cnt = base->req();
1850 // Check for loop invariant memory.
1851 if (cnt == 3) {
1852 for (uint i = 1; i < cnt; i++) {
1853 if (base->in(i) == base) {
1854 return nullptr; // Wait stable graph
1855 }
1856 }
1857 }
1858 }
1859
1860 // Split through Phi (see original code in loopopts.cpp).
1861 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1862
1863 // Do nothing here if Identity will find a value
1864 // (to avoid infinite chain of value phis generation).
1865 if (this != Identity(phase)) {
1866 return nullptr;
1867 }
1868
1869 // Select Region to split through.
1870 Node* region;
1871 DomResult dom_result = DomResult::Dominate;
1872 if (!base_is_phi) {
1873 assert(mem->is_Phi(), "sanity");
1874 region = mem->in(0);
1875 // Skip if the region dominates some control edge of the address.
1876 // We will check `dom_result` later.
1877 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1878 } else if (!mem->is_Phi()) {
1879 assert(base_is_phi, "sanity");
1880 region = base->in(0);
1881 // Skip if the region dominates some control edge of the memory.
1882 // We will check `dom_result` later.
1883 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1884 } else if (base->in(0) != mem->in(0)) {
1885 assert(base_is_phi && mem->is_Phi(), "sanity");
1886 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1887 if (dom_result == DomResult::Dominate) {
1888 region = base->in(0);
1889 } else {
1890 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1891 if (dom_result == DomResult::Dominate) {
1892 region = mem->in(0);
1893 }
1894 // Otherwise we encountered a complex graph.
1895 }
1896 } else {
1897 assert(base->in(0) == mem->in(0), "sanity");
1898 region = mem->in(0);
1899 }
1900
1901 PhaseIterGVN* igvn = phase->is_IterGVN();
1902 if (dom_result != DomResult::Dominate) {
1903 if (dom_result == DomResult::EncounteredDeadCode) {
1904 // There is some dead code which eventually will be removed in IGVN.
1905 // Once this is the case, we get an unambiguous dominance result.
1906 // Push the node to the worklist again until the dead code is removed.
1907 igvn->_worklist.push(this);
1908 }
1909 return nullptr;
1910 }
1911
1912 Node* phi = nullptr;
1913 const Type* this_type = this->bottom_type();
1914 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1915 int this_index = C->get_alias_index(t_oop);
1916 int this_offset = t_oop->offset();
1917 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1918 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1919 } else if (ignore_missing_instance_id) {
1920 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1921 } else {
1922 return nullptr;
1923 }
1924
1925 for (uint i = 1; i < region->req(); i++) {
1926 Node* x;
1927 Node* the_clone = nullptr;
1928 Node* in = region->in(i);
1929 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1930 in != nullptr && in->is_OuterStripMinedLoop()) {
1931 // No node should go in the outer strip mined loop
1932 in = in->in(LoopNode::EntryControl);
1933 }
1934 if (in == nullptr || in == C->top()) {
1935 x = C->top(); // Dead path? Use a dead data op
1936 } else {
1937 x = this->clone(); // Else clone up the data op
1938 the_clone = x; // Remember for possible deletion.
1939 // Alter data node to use pre-phi inputs
1940 if (this->in(0) == region) {
1941 x->set_req(0, in);
1942 } else {
1943 x->set_req(0, nullptr);
1944 }
1945 if (mem->is_Phi() && (mem->in(0) == region)) {
1946 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1947 }
1948 if (address->is_Phi() && address->in(0) == region) {
1949 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1950 }
1951 if (base_is_phi && (base->in(0) == region)) {
1952 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1953 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1954 x->set_req(Address, adr_x);
1955 }
1956 }
1957 // Check for a 'win' on some paths
1958 const Type *t = x->Value(igvn);
1959
1960 bool singleton = t->singleton();
1961
1962 // See comments in PhaseIdealLoop::split_thru_phi().
1963 if (singleton && t == Type::TOP) {
1964 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1965 }
1966
1967 if (singleton) {
1968 x = igvn->makecon(t);
1969 } else {
1970 // We now call Identity to try to simplify the cloned node.
1971 // Note that some Identity methods call phase->type(this).
1972 // Make sure that the type array is big enough for
1973 // our new node, even though we may throw the node away.
1974 // (This tweaking with igvn only works because x is a new node.)
1975 igvn->set_type(x, t);
1976 // If x is a TypeNode, capture any more-precise type permanently into Node
1977 // otherwise it will be not updated during igvn->transform since
1978 // igvn->type(x) is set to x->Value() already.
1979 x->raise_bottom_type(t);
1980 Node* y = x->Identity(igvn);
1981 if (y != x) {
1982 x = y;
1983 } else {
1984 y = igvn->hash_find_insert(x);
1985 if (y) {
1986 x = y;
1987 } else {
1988 // Else x is a new node we are keeping
1989 // We do not need register_new_node_with_optimizer
1990 // because set_type has already been called.
1991 igvn->_worklist.push(x);
1992 }
1993 }
1994 }
1995 if (x != the_clone && the_clone != nullptr) {
1996 igvn->remove_dead_node(the_clone);
1997 }
1998 phi->set_req(i, x);
1999 }
2000 // Record Phi
2001 igvn->register_new_node_with_optimizer(phi);
2002 return phi;
2003 }
2004
2005 AllocateNode* LoadNode::is_new_object_mark_load() const {
2006 if (Opcode() == Op_LoadX) {
2007 Node* address = in(MemNode::Address);
2008 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
2009 Node* mem = in(MemNode::Memory);
2010 if (alloc != nullptr && mem->is_Proj() &&
2011 mem->in(0) != nullptr &&
2012 mem->in(0) == alloc->initialization() &&
2013 alloc->initialization()->proj_out_or_null(0) != nullptr) {
2014 return alloc;
2015 }
2016 }
2017 return nullptr;
2018 }
2019
2020
2021 //------------------------------Ideal------------------------------------------
2022 // If the load is from Field memory and the pointer is non-null, it might be possible to
2023 // zero out the control input.
2024 // If the offset is constant and the base is an object allocation,
2025 // try to hook me up to the exact initializing store.
2026 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2027 if (has_pinned_control_dependency()) {
2028 return nullptr;
2029 }
2030 Node* p = MemNode::Ideal_common(phase, can_reshape);
2031 if (p) return (p == NodeSentinel) ? nullptr : p;
2032
2033 Node* ctrl = in(MemNode::Control);
2034 Node* address = in(MemNode::Address);
2035 bool progress = false;
2036
2037 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
2038 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
2039
2040 // Skip up past a SafePoint control. Cannot do this for Stores because
2041 // pointer stores & cardmarks must stay on the same side of a SafePoint.
2042 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
2043 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
2044 !addr_mark &&
2045 (depends_only_on_test() || has_unknown_control_dependency())) {
2046 ctrl = ctrl->in(0);
2047 set_req(MemNode::Control,ctrl);
2048 progress = true;
2049 }
2050
2051 intptr_t ignore = 0;
2052 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
2053 if (base != nullptr
2054 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
2055 // Check for useless control edge in some common special cases
2056 if (in(MemNode::Control) != nullptr
2057 && !(phase->type(address)->is_inlinetypeptr() && is_mismatched_access())
2058 && can_remove_control()
2059 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
2060 && all_controls_dominate(base, phase->C->start())) {
2061 // A method-invariant, non-null address (constant or 'this' argument).
2062 set_req(MemNode::Control, nullptr);
2063 progress = true;
2064 }
2065 }
2066
2067 Node* mem = in(MemNode::Memory);
2068 const TypePtr *addr_t = phase->type(address)->isa_ptr();
2069
2070 if (can_reshape && (addr_t != nullptr)) {
2071 // try to optimize our memory input
2072 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
2073 if (opt_mem != mem) {
2074 set_req_X(MemNode::Memory, opt_mem, phase);
2075 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
2076 return this;
2077 }
2078 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
2079 if ((t_oop != nullptr) &&
2080 (t_oop->is_known_instance_field() ||
2081 t_oop->is_ptr_to_boxed_value())) {
2082 PhaseIterGVN *igvn = phase->is_IterGVN();
2083 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
2084 if (igvn->_worklist.member(opt_mem)) {
2085 // Delay this transformation until memory Phi is processed.
2086 igvn->_worklist.push(this);
2087 return nullptr;
2088 }
2089 // Split instance field load through Phi.
2090 Node* result = split_through_phi(phase);
2091 if (result != nullptr) return result;
2092
2093 if (t_oop->is_ptr_to_boxed_value()) {
2094 Node* result = eliminate_autobox(igvn);
2095 if (result != nullptr) return result;
2096 }
2097 }
2098 }
2099
2100 // Is there a dominating load that loads the same value? Leave
2101 // anything that is not a load of a field/array element (like
2102 // barriers etc.) alone
2103 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
2104 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2105 Node *use = mem->fast_out(i);
2106 if (use != this &&
2107 use->Opcode() == Opcode() &&
2108 use->in(0) != nullptr &&
2109 use->in(0) != in(0) &&
2110 use->in(Address) == in(Address)) {
2111 Node* ctl = in(0);
2112 for (int i = 0; i < 10 && ctl != nullptr; i++) {
2113 ctl = IfNode::up_one_dom(ctl);
2114 if (ctl == use->in(0)) {
2115 set_req(0, use->in(0));
2116 return this;
2117 }
2118 }
2119 }
2120 }
2121 }
2122
2123 // Check for prior store with a different base or offset; make Load
2124 // independent. Skip through any number of them. Bail out if the stores
2125 // are in an endless dead cycle and report no progress. This is a key
2126 // transform for Reflection. However, if after skipping through the Stores
2127 // we can't then fold up against a prior store do NOT do the transform as
2128 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
2129 // array memory alive twice: once for the hoisted Load and again after the
2130 // bypassed Store. This situation only works if EVERYBODY who does
2131 // anti-dependence work knows how to bypass. I.e. we need all
2132 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
2133 // the alias index stuff. So instead, peek through Stores and IFF we can
2134 // fold up, do so.
2135 Node* prev_mem = find_previous_store(phase);
2136 if (prev_mem != nullptr) {
2137 Node* value = can_see_arraycopy_value(prev_mem, phase);
2138 if (value != nullptr) {
2139 return value;
2140 }
2141 }
2142 // Steps (a), (b): Walk past independent stores to find an exact match.
2143 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2144 // (c) See if we can fold up on the spot, but don't fold up here.
2145 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2146 // just return a prior value, which is done by Identity calls.
2147 if (can_see_stored_value(prev_mem, phase)) {
2148 // Make ready for step (d):
2149 set_req_X(MemNode::Memory, prev_mem, phase);
2150 return this;
2151 }
2152 }
2153
2154 if (progress) {
2155 return this;
2156 }
2157
2158 if (!can_reshape) {
2159 phase->record_for_igvn(this);
2160 }
2161 return nullptr;
2162 }
2163
2164 // Helper to recognize certain Klass fields which are invariant across
2165 // some group of array types (e.g., int[] or all T[] where T < Object).
2166 const Type*
2167 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2168 ciKlass* klass) const {
2169 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2170 "must not happen");
2171 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
2172 // The field is Klass::_access_flags. Return its (constant) value.
2173 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
2174 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2175 return TypeInt::make(klass->access_flags());
2176 }
2177 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2178 // The field is Klass::_misc_flags. Return its (constant) value.
2179 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
2180 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2181 return TypeInt::make(klass->misc_flags());
2182 }
2183 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2184 // The field is Klass::_layout_helper. Return its constant value if known.
2185 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2186 return TypeInt::make(klass->layout_helper());
2187 }
2188
2189 // No match.
2190 return nullptr;
2191 }
2192
2193 //------------------------------Value-----------------------------------------
2194 const Type* LoadNode::Value(PhaseGVN* phase) const {
2195 // Either input is TOP ==> the result is TOP
2196 Node* mem = in(MemNode::Memory);
2197 const Type *t1 = phase->type(mem);
2198 if (t1 == Type::TOP) return Type::TOP;
2199 Node* adr = in(MemNode::Address);
2200 const TypePtr* tp = phase->type(adr)->isa_ptr();
2201 if (tp == nullptr || tp->empty()) return Type::TOP;
2202 int off = tp->offset();
2203 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2204 Compile* C = phase->C;
2205
2206 // If we are loading from a freshly-allocated object, produce a zero,
2207 // if the load is provably beyond the header of the object.
2208 // (Also allow a variable load from a fresh array to produce zero.)
2209 const TypeOopPtr* tinst = tp->isa_oopptr();
2210 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2211 Node* value = can_see_stored_value(mem, phase);
2212 if (value != nullptr && value->is_Con()) {
2213 assert(value->bottom_type()->higher_equal(_type), "sanity");
2214 return value->bottom_type();
2215 }
2216
2217 // Try to guess loaded type from pointer type
2218 if (tp->isa_aryptr()) {
2219 const TypeAryPtr* ary = tp->is_aryptr();
2220 const Type* t = ary->elem();
2221
2222 // Determine whether the reference is beyond the header or not, by comparing
2223 // the offset against the offset of the start of the array's data.
2224 // Different array types begin at slightly different offsets (12 vs. 16).
2225 // We choose T_BYTE as an example base type that is least restrictive
2226 // as to alignment, which will therefore produce the smallest
2227 // possible base offset.
2228 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2229 const bool off_beyond_header = (off >= min_base_off);
2230
2231 // Try to constant-fold a stable array element.
2232 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2233 // Make sure the reference is not into the header and the offset is constant
2234 ciObject* aobj = ary->const_oop();
2235 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2236 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2237 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2238 stable_dimension,
2239 memory_type(), is_unsigned());
2240 if (con_type != nullptr) {
2241 return con_type;
2242 }
2243 }
2244 }
2245
2246 // Don't do this for integer types. There is only potential profit if
2247 // the element type t is lower than _type; that is, for int types, if _type is
2248 // more restrictive than t. This only happens here if one is short and the other
2249 // char (both 16 bits), and in those cases we've made an intentional decision
2250 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2251 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2252 //
2253 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2254 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2255 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2256 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2257 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2258 // In fact, that could have been the original type of p1, and p1 could have
2259 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2260 // expression (LShiftL quux 3) independently optimized to the constant 8.
2261 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2262 && (_type->isa_vect() == nullptr)
2263 && !ary->is_flat()
2264 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2265 // t might actually be lower than _type, if _type is a unique
2266 // concrete subclass of abstract class t.
2267 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2268 const Type* jt = t->join_speculative(_type);
2269 // In any case, do not allow the join, per se, to empty out the type.
2270 if (jt->empty() && !t->empty()) {
2271 // This can happen if a interface-typed array narrows to a class type.
2272 jt = _type;
2273 }
2274 #ifdef ASSERT
2275 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2276 // The pointers in the autobox arrays are always non-null
2277 Node* base = adr->in(AddPNode::Base);
2278 if ((base != nullptr) && base->is_DecodeN()) {
2279 // Get LoadN node which loads IntegerCache.cache field
2280 base = base->in(1);
2281 }
2282 if ((base != nullptr) && base->is_Con()) {
2283 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2284 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2285 // It could be narrow oop
2286 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2287 }
2288 }
2289 }
2290 #endif
2291 return jt;
2292 }
2293 }
2294 } else if (tp->base() == Type::InstPtr) {
2295 assert( off != Type::OffsetBot ||
2296 // arrays can be cast to Objects
2297 !tp->isa_instptr() ||
2298 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2299 // Default value load
2300 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2301 // unsafe field access may not have a constant offset
2302 C->has_unsafe_access(),
2303 "Field accesses must be precise" );
2304 // For oop loads, we expect the _type to be precise.
2305
2306 const TypeInstPtr* tinst = tp->is_instptr();
2307 BasicType bt = memory_type();
2308
2309 // Optimize loads from constant fields.
2310 ciObject* const_oop = tinst->const_oop();
2311 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2312 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2313 if (con_type != nullptr) {
2314 return con_type;
2315 }
2316 }
2317 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2318 assert(off != Type::OffsetBot ||
2319 !tp->isa_instklassptr() ||
2320 // arrays can be cast to Objects
2321 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2322 // also allow array-loading from the primary supertype
2323 // array during subtype checks
2324 Opcode() == Op_LoadKlass,
2325 "Field accesses must be precise");
2326 // For klass/static loads, we expect the _type to be precise
2327 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2328 /* With mirrors being an indirect in the Klass*
2329 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2330 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2331 *
2332 * So check the type and klass of the node before the LoadP.
2333 */
2334 Node* adr2 = adr->in(MemNode::Address);
2335 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2336 if (tkls != nullptr && !StressReflectiveCode) {
2337 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2338 ciKlass* klass = tkls->exact_klass();
2339 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2340 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2341 return TypeInstPtr::make(klass->java_mirror());
2342 }
2343 }
2344 }
2345
2346 const TypeKlassPtr *tkls = tp->isa_klassptr();
2347 if (tkls != nullptr) {
2348 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2349 ciKlass* klass = tkls->exact_klass();
2350 // We are loading a field from a Klass metaobject whose identity
2351 // is known at compile time (the type is "exact" or "precise").
2352 // Check for fields we know are maintained as constants by the VM.
2353 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2354 // The field is Klass::_super_check_offset. Return its (constant) value.
2355 // (Folds up type checking code.)
2356 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2357 return TypeInt::make(klass->super_check_offset());
2358 }
2359 if (UseCompactObjectHeaders) { // TODO: Should EnableValhalla also take this path ?
2360 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2361 // The field is Klass::_prototype_header. Return its (constant) value.
2362 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2363 return TypeX::make(klass->prototype_header());
2364 }
2365 }
2366 // Compute index into primary_supers array
2367 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2368 // Check for overflowing; use unsigned compare to handle the negative case.
2369 if( depth < ciKlass::primary_super_limit() ) {
2370 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2371 // (Folds up type checking code.)
2372 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2373 ciKlass *ss = klass->super_of_depth(depth);
2374 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2375 }
2376 const Type* aift = load_array_final_field(tkls, klass);
2377 if (aift != nullptr) return aift;
2378 }
2379
2380 // We can still check if we are loading from the primary_supers array at a
2381 // shallow enough depth. Even though the klass is not exact, entries less
2382 // than or equal to its super depth are correct.
2383 if (tkls->is_loaded()) {
2384 ciKlass* klass = nullptr;
2385 if (tkls->isa_instklassptr()) {
2386 klass = tkls->is_instklassptr()->instance_klass();
2387 } else {
2388 int dims;
2389 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2390 if (inner->isa_instklassptr()) {
2391 klass = inner->is_instklassptr()->instance_klass();
2392 klass = ciObjArrayKlass::make(klass, dims);
2393 }
2394 }
2395 if (klass != nullptr) {
2396 // Compute index into primary_supers array
2397 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2398 // Check for overflowing; use unsigned compare to handle the negative case.
2399 if (depth < ciKlass::primary_super_limit() &&
2400 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2401 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2402 // (Folds up type checking code.)
2403 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2404 ciKlass *ss = klass->super_of_depth(depth);
2405 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2406 }
2407 }
2408 }
2409
2410 // If the type is enough to determine that the thing is not an array,
2411 // we can give the layout_helper a positive interval type.
2412 // This will help short-circuit some reflective code.
2413 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2414 tkls->isa_instklassptr() && // not directly typed as an array
2415 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2416 ) {
2417 // Note: When interfaces are reliable, we can narrow the interface
2418 // test to (klass != Serializable && klass != Cloneable).
2419 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2420 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2421 // The key property of this type is that it folds up tests
2422 // for array-ness, since it proves that the layout_helper is positive.
2423 // Thus, a generic value like the basic object layout helper works fine.
2424 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2425 }
2426 }
2427
2428 bool is_vect = (_type->isa_vect() != nullptr);
2429 if (is_instance && !is_vect) {
2430 // If we have an instance type and our memory input is the
2431 // programs's initial memory state, there is no matching store,
2432 // so just return a zero of the appropriate type -
2433 // except if it is vectorized - then we have no zero constant.
2434 Node *mem = in(MemNode::Memory);
2435 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2436 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2437 // TODO 8350865 This is needed for flat array accesses, somehow the memory of the loads bypasses the intrinsic
2438 // Run TestArrays.test6 in Scenario4, we need more tests for this. TestBasicFunctionality::test20 also needs this.
2439 if (tp->isa_aryptr() && tp->is_aryptr()->is_flat() && !UseFieldFlattening) {
2440 return _type;
2441 }
2442 return Type::get_zero_type(_type->basic_type());
2443 }
2444 }
2445 if (!UseCompactObjectHeaders) {
2446 Node* alloc = is_new_object_mark_load();
2447 if (alloc != nullptr) {
2448 if (EnableValhalla) {
2449 // The mark word may contain property bits (inline, flat, null-free)
2450 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2451 const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2452 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2453 return TypeX::make(tkls->exact_klass()->prototype_header());
2454 }
2455 } else {
2456 return TypeX::make(markWord::prototype().value());
2457 }
2458 }
2459 }
2460
2461 return _type;
2462 }
2463
2464 //------------------------------match_edge-------------------------------------
2465 // Do we Match on this edge index or not? Match only the address.
2466 uint LoadNode::match_edge(uint idx) const {
2467 return idx == MemNode::Address;
2468 }
2469
2470 //--------------------------LoadBNode::Ideal--------------------------------------
2471 //
2472 // If the previous store is to the same address as this load,
2473 // and the value stored was larger than a byte, replace this load
2474 // with the value stored truncated to a byte. If no truncation is
2475 // needed, the replacement is done in LoadNode::Identity().
2476 //
2477 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2478 Node* mem = in(MemNode::Memory);
2479 Node* value = can_see_stored_value(mem,phase);
2480 if (value != nullptr) {
2481 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2482 if (narrow != value) {
2483 return narrow;
2484 }
2485 }
2486 // Identity call will handle the case where truncation is not needed.
2487 return LoadNode::Ideal(phase, can_reshape);
2488 }
2489
2490 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2491 Node* mem = in(MemNode::Memory);
2492 Node* value = can_see_stored_value(mem,phase);
2493 if (value != nullptr && value->is_Con() &&
2494 !value->bottom_type()->higher_equal(_type)) {
2495 // If the input to the store does not fit with the load's result type,
2496 // it must be truncated. We can't delay until Ideal call since
2497 // a singleton Value is needed for split_thru_phi optimization.
2498 int con = value->get_int();
2499 return TypeInt::make((con << 24) >> 24);
2500 }
2501 return LoadNode::Value(phase);
2502 }
2503
2504 //--------------------------LoadUBNode::Ideal-------------------------------------
2505 //
2506 // If the previous store is to the same address as this load,
2507 // and the value stored was larger than a byte, replace this load
2508 // with the value stored truncated to a byte. If no truncation is
2509 // needed, the replacement is done in LoadNode::Identity().
2510 //
2511 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2512 Node* mem = in(MemNode::Memory);
2513 Node* value = can_see_stored_value(mem, phase);
2514 if (value != nullptr) {
2515 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2516 if (narrow != value) {
2517 return narrow;
2518 }
2519 }
2520 // Identity call will handle the case where truncation is not needed.
2521 return LoadNode::Ideal(phase, can_reshape);
2522 }
2523
2524 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2525 Node* mem = in(MemNode::Memory);
2526 Node* value = can_see_stored_value(mem,phase);
2527 if (value != nullptr && value->is_Con() &&
2528 !value->bottom_type()->higher_equal(_type)) {
2529 // If the input to the store does not fit with the load's result type,
2530 // it must be truncated. We can't delay until Ideal call since
2531 // a singleton Value is needed for split_thru_phi optimization.
2532 int con = value->get_int();
2533 return TypeInt::make(con & 0xFF);
2534 }
2535 return LoadNode::Value(phase);
2536 }
2537
2538 //--------------------------LoadUSNode::Ideal-------------------------------------
2539 //
2540 // If the previous store is to the same address as this load,
2541 // and the value stored was larger than a char, replace this load
2542 // with the value stored truncated to a char. If no truncation is
2543 // needed, the replacement is done in LoadNode::Identity().
2544 //
2545 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2546 Node* mem = in(MemNode::Memory);
2547 Node* value = can_see_stored_value(mem,phase);
2548 if (value != nullptr) {
2549 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2550 if (narrow != value) {
2551 return narrow;
2552 }
2553 }
2554 // Identity call will handle the case where truncation is not needed.
2555 return LoadNode::Ideal(phase, can_reshape);
2556 }
2557
2558 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2559 Node* mem = in(MemNode::Memory);
2560 Node* value = can_see_stored_value(mem,phase);
2561 if (value != nullptr && value->is_Con() &&
2562 !value->bottom_type()->higher_equal(_type)) {
2563 // If the input to the store does not fit with the load's result type,
2564 // it must be truncated. We can't delay until Ideal call since
2565 // a singleton Value is needed for split_thru_phi optimization.
2566 int con = value->get_int();
2567 return TypeInt::make(con & 0xFFFF);
2568 }
2569 return LoadNode::Value(phase);
2570 }
2571
2572 //--------------------------LoadSNode::Ideal--------------------------------------
2573 //
2574 // If the previous store is to the same address as this load,
2575 // and the value stored was larger than a short, replace this load
2576 // with the value stored truncated to a short. If no truncation is
2577 // needed, the replacement is done in LoadNode::Identity().
2578 //
2579 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2580 Node* mem = in(MemNode::Memory);
2581 Node* value = can_see_stored_value(mem,phase);
2582 if (value != nullptr) {
2583 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2584 if (narrow != value) {
2585 return narrow;
2586 }
2587 }
2588 // Identity call will handle the case where truncation is not needed.
2589 return LoadNode::Ideal(phase, can_reshape);
2590 }
2591
2592 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2593 Node* mem = in(MemNode::Memory);
2594 Node* value = can_see_stored_value(mem,phase);
2595 if (value != nullptr && value->is_Con() &&
2596 !value->bottom_type()->higher_equal(_type)) {
2597 // If the input to the store does not fit with the load's result type,
2598 // it must be truncated. We can't delay until Ideal call since
2599 // a singleton Value is needed for split_thru_phi optimization.
2600 int con = value->get_int();
2601 return TypeInt::make((con << 16) >> 16);
2602 }
2603 return LoadNode::Value(phase);
2604 }
2605
2606 Node* LoadNNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2607 // Loading from an InlineType, find the input and make an EncodeP
2608 Node* addr = in(Address);
2609 intptr_t offset;
2610 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
2611 Node* value = see_through_inline_type(phase, this, base, offset);
2612 if (value != nullptr) {
2613 return new EncodePNode(value, type());
2614 }
2615
2616 return LoadNode::Ideal(phase, can_reshape);
2617 }
2618
2619 //=============================================================================
2620 //----------------------------LoadKlassNode::make------------------------------
2621 // Polymorphic factory method:
2622 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2623 // sanity check the alias category against the created node type
2624 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2625 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2626 #ifdef _LP64
2627 if (adr_type->is_ptr_to_narrowklass()) {
2628 assert(UseCompressedClassPointers, "no compressed klasses");
2629 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2630 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2631 }
2632 #endif
2633 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2634 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2635 }
2636
2637 //------------------------------Value------------------------------------------
2638 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2639 return klass_value_common(phase);
2640 }
2641
2642 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2643 // Either input is TOP ==> the result is TOP
2644 const Type *t1 = phase->type( in(MemNode::Memory) );
2645 if (t1 == Type::TOP) return Type::TOP;
2646 Node *adr = in(MemNode::Address);
2647 const Type *t2 = phase->type( adr );
2648 if (t2 == Type::TOP) return Type::TOP;
2649 const TypePtr *tp = t2->is_ptr();
2650 if (TypePtr::above_centerline(tp->ptr()) ||
2651 tp->ptr() == TypePtr::Null) return Type::TOP;
2652
2653 // Return a more precise klass, if possible
2654 const TypeInstPtr *tinst = tp->isa_instptr();
2655 if (tinst != nullptr) {
2656 ciInstanceKlass* ik = tinst->instance_klass();
2657 int offset = tinst->offset();
2658 if (ik == phase->C->env()->Class_klass()
2659 && (offset == java_lang_Class::klass_offset() ||
2660 offset == java_lang_Class::array_klass_offset())) {
2661 // We are loading a special hidden field from a Class mirror object,
2662 // the field which points to the VM's Klass metaobject.
2663 bool is_null_free_array = false;
2664 ciType* t = tinst->java_mirror_type(&is_null_free_array);
2665 // java_mirror_type returns non-null for compile-time Class constants.
2666 if (t != nullptr) {
2667 // constant oop => constant klass
2668 if (offset == java_lang_Class::array_klass_offset()) {
2669 if (t->is_void()) {
2670 // We cannot create a void array. Since void is a primitive type return null
2671 // klass. Users of this result need to do a null check on the returned klass.
2672 return TypePtr::NULL_PTR;
2673 }
2674 const TypeKlassPtr* tklass = TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2675 if (is_null_free_array) {
2676 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2677 }
2678 return tklass;
2679 }
2680 if (!t->is_klass()) {
2681 // a primitive Class (e.g., int.class) has null for a klass field
2682 return TypePtr::NULL_PTR;
2683 }
2684 // Fold up the load of the hidden field
2685 const TypeKlassPtr* tklass = TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2686 if (is_null_free_array) {
2687 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2688 }
2689 return tklass;
2690 }
2691 // non-constant mirror, so we can't tell what's going on
2692 }
2693 if (!tinst->is_loaded())
2694 return _type; // Bail out if not loaded
2695 if (offset == oopDesc::klass_offset_in_bytes()) {
2696 return tinst->as_klass_type(true);
2697 }
2698 }
2699
2700 // Check for loading klass from an array
2701 const TypeAryPtr* tary = tp->isa_aryptr();
2702 if (tary != nullptr &&
2703 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2704 return tary->as_klass_type(true);
2705 }
2706
2707 // Check for loading klass from an array klass
2708 const TypeKlassPtr *tkls = tp->isa_klassptr();
2709 if (tkls != nullptr && !StressReflectiveCode) {
2710 if (!tkls->is_loaded())
2711 return _type; // Bail out if not loaded
2712 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2713 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2714 // // Always returning precise element type is incorrect,
2715 // // e.g., element type could be object and array may contain strings
2716 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2717
2718 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2719 // according to the element type's subclassing.
2720 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2721 }
2722 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2723 tkls->offset() == in_bytes(Klass::super_offset())) {
2724 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2725 // The field is Klass::_super. Return its (constant) value.
2726 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2727 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2728 }
2729 }
2730
2731 if (tkls != nullptr && !UseSecondarySupersCache
2732 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2733 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2734 return TypePtr::NULL_PTR;
2735 }
2736
2737 // Bailout case
2738 return LoadNode::Value(phase);
2739 }
2740
2741 //------------------------------Identity---------------------------------------
2742 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2743 // Also feed through the klass in Allocate(...klass...)._klass.
2744 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2745 return klass_identity_common(phase);
2746 }
2747
2748 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2749 Node* x = LoadNode::Identity(phase);
2750 if (x != this) return x;
2751
2752 // Take apart the address into an oop and offset.
2753 // Return 'this' if we cannot.
2754 Node* adr = in(MemNode::Address);
2755 intptr_t offset = 0;
2756 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2757 if (base == nullptr) return this;
2758 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2759 if (toop == nullptr) return this;
2760
2761 // Step over potential GC barrier for OopHandle resolve
2762 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2763 if (bs->is_gc_barrier_node(base)) {
2764 base = bs->step_over_gc_barrier(base);
2765 }
2766
2767 // We can fetch the klass directly through an AllocateNode.
2768 // This works even if the klass is not constant (clone or newArray).
2769 if (offset == oopDesc::klass_offset_in_bytes()) {
2770 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2771 if (allocated_klass != nullptr) {
2772 return allocated_klass;
2773 }
2774 }
2775
2776 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2777 // See inline_native_Class_query for occurrences of these patterns.
2778 // Java Example: x.getClass().isAssignableFrom(y)
2779 //
2780 // This improves reflective code, often making the Class
2781 // mirror go completely dead. (Current exception: Class
2782 // mirrors may appear in debug info, but we could clean them out by
2783 // introducing a new debug info operator for Klass.java_mirror).
2784
2785 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2786 && offset == java_lang_Class::klass_offset()) {
2787 if (base->is_Load()) {
2788 Node* base2 = base->in(MemNode::Address);
2789 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2790 Node* adr2 = base2->in(MemNode::Address);
2791 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2792 if (tkls != nullptr && !tkls->empty()
2793 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2794 && adr2->is_AddP()
2795 ) {
2796 int mirror_field = in_bytes(Klass::java_mirror_offset());
2797 if (tkls->offset() == mirror_field) {
2798 return adr2->in(AddPNode::Base);
2799 }
2800 }
2801 }
2802 }
2803 }
2804
2805 return this;
2806 }
2807
2808 LoadNode* LoadNode::clone_pinned() const {
2809 LoadNode* ld = clone()->as_Load();
2810 ld->_control_dependency = UnknownControl;
2811 return ld;
2812 }
2813
2814
2815 //------------------------------Value------------------------------------------
2816 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2817 const Type *t = klass_value_common(phase);
2818 if (t == Type::TOP)
2819 return t;
2820
2821 return t->make_narrowklass();
2822 }
2823
2824 //------------------------------Identity---------------------------------------
2825 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2826 // Also feed through the klass in Allocate(...klass...)._klass.
2827 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2828 Node *x = klass_identity_common(phase);
2829
2830 const Type *t = phase->type( x );
2831 if( t == Type::TOP ) return x;
2832 if( t->isa_narrowklass()) return x;
2833 assert (!t->isa_narrowoop(), "no narrow oop here");
2834
2835 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2836 }
2837
2838 //------------------------------Value-----------------------------------------
2839 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2840 // Either input is TOP ==> the result is TOP
2841 const Type *t1 = phase->type( in(MemNode::Memory) );
2842 if( t1 == Type::TOP ) return Type::TOP;
2843 Node *adr = in(MemNode::Address);
2844 const Type *t2 = phase->type( adr );
2845 if( t2 == Type::TOP ) return Type::TOP;
2846 const TypePtr *tp = t2->is_ptr();
2847 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2848 const TypeAryPtr *tap = tp->isa_aryptr();
2849 if( !tap ) return _type;
2850 return tap->size();
2851 }
2852
2853 //-------------------------------Ideal---------------------------------------
2854 // Feed through the length in AllocateArray(...length...)._length.
2855 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2856 Node* p = MemNode::Ideal_common(phase, can_reshape);
2857 if (p) return (p == NodeSentinel) ? nullptr : p;
2858
2859 // Take apart the address into an oop and offset.
2860 // Return 'this' if we cannot.
2861 Node* adr = in(MemNode::Address);
2862 intptr_t offset = 0;
2863 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2864 if (base == nullptr) return nullptr;
2865 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2866 if (tary == nullptr) return nullptr;
2867
2868 // We can fetch the length directly through an AllocateArrayNode.
2869 // This works even if the length is not constant (clone or newArray).
2870 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2871 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2872 if (alloc != nullptr) {
2873 Node* allocated_length = alloc->Ideal_length();
2874 Node* len = alloc->make_ideal_length(tary, phase);
2875 if (allocated_length != len) {
2876 // New CastII improves on this.
2877 return len;
2878 }
2879 }
2880 }
2881
2882 return nullptr;
2883 }
2884
2885 //------------------------------Identity---------------------------------------
2886 // Feed through the length in AllocateArray(...length...)._length.
2887 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2888 Node* x = LoadINode::Identity(phase);
2889 if (x != this) return x;
2890
2891 // Take apart the address into an oop and offset.
2892 // Return 'this' if we cannot.
2893 Node* adr = in(MemNode::Address);
2894 intptr_t offset = 0;
2895 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2896 if (base == nullptr) return this;
2897 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2898 if (tary == nullptr) return this;
2899
2900 // We can fetch the length directly through an AllocateArrayNode.
2901 // This works even if the length is not constant (clone or newArray).
2902 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2903 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2904 if (alloc != nullptr) {
2905 Node* allocated_length = alloc->Ideal_length();
2906 // Do not allow make_ideal_length to allocate a CastII node.
2907 Node* len = alloc->make_ideal_length(tary, phase, false);
2908 if (allocated_length == len) {
2909 // Return allocated_length only if it would not be improved by a CastII.
2910 return allocated_length;
2911 }
2912 }
2913 }
2914
2915 return this;
2916
2917 }
2918
2919 //=============================================================================
2920 //---------------------------StoreNode::make-----------------------------------
2921 // Polymorphic factory method:
2922 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo, bool require_atomic_access) {
2923 assert((mo == unordered || mo == release), "unexpected");
2924 Compile* C = gvn.C;
2925 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2926 ctl != nullptr, "raw memory operations should have control edge");
2927
2928 switch (bt) {
2929 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2930 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2931 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2932 case T_CHAR:
2933 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2934 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2935 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2936 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2937 case T_METADATA:
2938 case T_ADDRESS:
2939 case T_OBJECT:
2940 #ifdef _LP64
2941 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2942 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2943 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2944 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2945 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2946 adr->bottom_type()->isa_rawptr())) {
2947 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2948 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2949 }
2950 #endif
2951 {
2952 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2953 }
2954 default:
2955 ShouldNotReachHere();
2956 return (StoreNode*)nullptr;
2957 }
2958 }
2959
2960 //--------------------------bottom_type----------------------------------------
2961 const Type *StoreNode::bottom_type() const {
2962 return Type::MEMORY;
2963 }
2964
2965 //------------------------------hash-------------------------------------------
2966 uint StoreNode::hash() const {
2967 // unroll addition of interesting fields
2968 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2969
2970 // Since they are not commoned, do not hash them:
2971 return NO_HASH;
2972 }
2973
2974 // Link together multiple stores (B/S/C/I) into a longer one.
2975 //
2976 // Example: _store = StoreB[i+3]
2977 //
2978 // RangeCheck[i+0] RangeCheck[i+0]
2979 // StoreB[i+0]
2980 // RangeCheck[i+3] RangeCheck[i+3]
2981 // StoreB[i+1] --> pass: fail:
2982 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2983 // StoreB[i+3]
2984 //
2985 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2986 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2987 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2988 // failing path.
2989 //
2990 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2991 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2992 // and possibly no RangeChecks remain between the stores.
2993 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2994 // and those later RangeChecks can be removed. Example:
2995 //
2996 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2997 // StoreB[i+0] StoreB[i+0] <- first store
2998 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2999 // StoreB[i+1] StoreB[i+1] <- second store
3000 // RangeCheck[i+2] --> removed
3001 // StoreB[i+2] StoreB[i+2]
3002 // RangeCheck[i+3] --> removed
3003 // StoreB[i+3] StoreB[i+3] <- last store
3004 //
3005 // Thus, it is a common pattern that between the first and last store in a chain
3006 // of adjacent stores there remains exactly one RangeCheck, located between the
3007 // first and the second store (e.g. RangeCheck[i+3]).
3008 //
3009 class MergePrimitiveStores : public StackObj {
3010 private:
3011 PhaseGVN* const _phase;
3012 StoreNode* const _store;
3013 // State machine with initial state Unknown
3014 // Allowed transitions:
3015 // Unknown -> Const
3016 // Unknown -> Platform
3017 // Unknown -> Reverse
3018 // Unknown -> NotAdjacent
3019 // Const -> Const
3020 // Const -> NotAdjacent
3021 // Platform -> Platform
3022 // Platform -> NotAdjacent
3023 // Reverse -> Reverse
3024 // Reverse -> NotAdjacent
3025 // NotAdjacent -> NotAdjacent
3026 enum ValueOrder : uint8_t {
3027 Unknown, // Initial state
3028 Const, // Input values are const
3029 Platform, // Platform order
3030 Reverse, // Reverse platform order
3031 NotAdjacent // Not adjacent
3032 };
3033 ValueOrder _value_order;
3034
3035 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
3036
3037 public:
3038 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
3039 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
3040 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
3041 {}
3042
3043 StoreNode* run();
3044
3045 private:
3046 bool is_compatible_store(const StoreNode* other_store) const;
3047 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
3048 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
3049 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
3050 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
3051 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
3052
3053 class Status {
3054 private:
3055 StoreNode* _found_store;
3056 bool _found_range_check;
3057
3058 Status(StoreNode* found_store, bool found_range_check)
3059 : _found_store(found_store), _found_range_check(found_range_check) {}
3060
3061 public:
3062 StoreNode* found_store() const { return _found_store; }
3063 bool found_range_check() const { return _found_range_check; }
3064 static Status make_failure() { return Status(nullptr, false); }
3065
3066 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
3067 if (cfg_status == CFGStatus::Failure) {
3068 return Status::make_failure();
3069 }
3070 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
3071 }
3072
3073 #ifndef PRODUCT
3074 void print_on(outputStream* st) const {
3075 if (_found_store == nullptr) {
3076 st->print_cr("None");
3077 } else {
3078 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
3079 _found_range_check ? "RC" : "no-RC");
3080 }
3081 }
3082 #endif
3083 };
3084
3085 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
3086 Status find_adjacent_use_store(const StoreNode* def_store) const;
3087 Status find_adjacent_def_store(const StoreNode* use_store) const;
3088 Status find_use_store(const StoreNode* def_store) const;
3089 Status find_def_store(const StoreNode* use_store) const;
3090 Status find_use_store_unidirectional(const StoreNode* def_store) const;
3091 Status find_def_store_unidirectional(const StoreNode* use_store) const;
3092
3093 void collect_merge_list(Node_List& merge_list) const;
3094 Node* make_merged_input_value(const Node_List& merge_list);
3095 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3096
3097 #ifndef PRODUCT
3098 // Access to TraceMergeStores tags
3099 bool is_trace(TraceMergeStores::Tag tag) const {
3100 return _trace_tags.at(tag);
3101 }
3102
3103 bool is_trace_basic() const {
3104 return is_trace(TraceMergeStores::Tag::BASIC);
3105 }
3106
3107 bool is_trace_pointer_parsing() const {
3108 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
3109 }
3110
3111 bool is_trace_pointer_aliasing() const {
3112 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
3113 }
3114
3115 bool is_trace_pointer_adjacency() const {
3116 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
3117 }
3118
3119 bool is_trace_success() const {
3120 return is_trace(TraceMergeStores::Tag::SUCCESS);
3121 }
3122 #endif
3123
3124 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3125 };
3126
3127 StoreNode* MergePrimitiveStores::run() {
3128 // Check for B/S/C/I
3129 int opc = _store->Opcode();
3130 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3131 return nullptr;
3132 }
3133
3134 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3135
3136 // The _store must be the "last" store in a chain. If we find a use we could merge with
3137 // then that use or a store further down is the "last" store.
3138 Status status_use = find_adjacent_use_store(_store);
3139 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3140 if (status_use.found_store() != nullptr) {
3141 return nullptr;
3142 }
3143
3144 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3145 Status status_def = find_adjacent_def_store(_store);
3146 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3147 Node* def_store = status_def.found_store();
3148 if (def_store == nullptr) {
3149 return nullptr;
3150 }
3151
3152 // Initialize value order
3153 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3154 _store->in(MemNode::ValueIn),
3155 _store->memory_size());
3156 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3157
3158 ResourceMark rm;
3159 Node_List merge_list;
3160 collect_merge_list(merge_list);
3161
3162 Node* merged_input_value = make_merged_input_value(merge_list);
3163 if (merged_input_value == nullptr) { return nullptr; }
3164
3165 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3166
3167 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3168
3169 return merged_store;
3170 }
3171
3172 // Check compatibility between _store and other_store.
3173 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3174 int opc = _store->Opcode();
3175 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3176
3177 if (other_store == nullptr ||
3178 _store->Opcode() != other_store->Opcode()) {
3179 return false;
3180 }
3181
3182 return true;
3183 }
3184
3185 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3186 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3187 use_store->in(MemNode::ValueIn),
3188 def_store->memory_size())) {
3189 return false;
3190 }
3191
3192 ResourceMark rm;
3193 #ifndef PRODUCT
3194 const TraceMemPointer trace(is_trace_pointer_parsing(),
3195 is_trace_pointer_aliasing(),
3196 is_trace_pointer_adjacency(),
3197 true);
3198 #endif
3199 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3200 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3201 return pointer_def.is_adjacent_to_and_before(pointer_use);
3202 }
3203
3204 // Check input values n1 and n2 can be merged and return the value order
3205 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3206 const int memory_size) const {
3207 // Pattern: [n1 = ConI, n2 = ConI]
3208 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3209 return ValueOrder::Const;
3210 }
3211
3212 Node const *base_n2;
3213 jint shift_n2;
3214 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3215 return ValueOrder::NotAdjacent;
3216 }
3217 Node const *base_n1;
3218 jint shift_n1;
3219 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3220 return ValueOrder::NotAdjacent;
3221 }
3222
3223 int bits_per_store = memory_size * 8;
3224 if (base_n1 != base_n2 ||
3225 abs(shift_n1 - shift_n2) != bits_per_store ||
3226 shift_n1 % bits_per_store != 0) {
3227 // Values are not adjacent
3228 return ValueOrder::NotAdjacent;
3229 }
3230
3231 // Detect value order
3232 #ifdef VM_LITTLE_ENDIAN
3233 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3234 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3235 #else
3236 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3237 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3238 #endif
3239 }
3240
3241 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3242 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3243
3244 switch (input_value_order) {
3245 case ValueOrder::NotAdjacent:
3246 return false;
3247 case ValueOrder::Reverse:
3248 if (memory_size != 1 ||
3249 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3250 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3251 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3252 // ReverseBytes are not supported by platform
3253 return false;
3254 }
3255 // fall-through.
3256 case ValueOrder::Const:
3257 case ValueOrder::Platform:
3258 if (_value_order == ValueOrder::Unknown) {
3259 // Initial state is Unknown, and we find a valid input value order
3260 return true;
3261 }
3262 // The value order can not be changed
3263 return _value_order == input_value_order;
3264 case ValueOrder::Unknown:
3265 default:
3266 ShouldNotReachHere();
3267 }
3268 return false;
3269 }
3270
3271 // Detect pattern: n = base_out >> shift_out
3272 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3273 assert(n != nullptr, "precondition");
3274
3275 int opc = n->Opcode();
3276 if (opc == Op_ConvL2I) {
3277 n = n->in(1);
3278 opc = n->Opcode();
3279 }
3280
3281 if ((opc == Op_RShiftI ||
3282 opc == Op_RShiftL ||
3283 opc == Op_URShiftI ||
3284 opc == Op_URShiftL) &&
3285 n->in(2)->is_ConI()) {
3286 base_out = n->in(1);
3287 shift_out = n->in(2)->get_int();
3288 // The shift must be positive:
3289 return shift_out >= 0;
3290 }
3291
3292 if (phase->type(n)->isa_int() != nullptr ||
3293 phase->type(n)->isa_long() != nullptr) {
3294 // (base >> 0)
3295 base_out = n;
3296 shift_out = 0;
3297 return true;
3298 }
3299 return false;
3300 }
3301
3302 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3303 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3304 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3305
3306 Node* ctrl_use = use_store->in(MemNode::Control);
3307 Node* ctrl_def = def_store->in(MemNode::Control);
3308 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3309 return CFGStatus::Failure;
3310 }
3311
3312 if (ctrl_use == ctrl_def) {
3313 // Same ctrl -> no RangeCheck in between.
3314 // Check: use_store must be the only use of def_store.
3315 if (def_store->outcnt() > 1) {
3316 return CFGStatus::Failure;
3317 }
3318 return CFGStatus::SuccessNoRangeCheck;
3319 }
3320
3321 // Different ctrl -> could have RangeCheck in between.
3322 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3323 // 2. ctrl separated by RangeCheck.
3324 if (def_store->outcnt() != 2) {
3325 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3326 }
3327 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3328 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3329 if (merge_mem == nullptr ||
3330 merge_mem->outcnt() != 1) {
3331 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3332 }
3333 if (!ctrl_use->is_IfProj() ||
3334 !ctrl_use->in(0)->is_RangeCheck() ||
3335 ctrl_use->in(0)->outcnt() != 2) {
3336 return CFGStatus::Failure; // Not RangeCheck.
3337 }
3338 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3339 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3340 if (trap != merge_mem->unique_out() ||
3341 ctrl_use->in(0)->in(0) != ctrl_def) {
3342 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3343 }
3344
3345 return CFGStatus::SuccessWithRangeCheck;
3346 }
3347
3348 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3349 Status status_use = find_use_store(def_store);
3350 StoreNode* use_store = status_use.found_store();
3351 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3352 return Status::make_failure();
3353 }
3354 return status_use;
3355 }
3356
3357 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3358 Status status_def = find_def_store(use_store);
3359 StoreNode* def_store = status_def.found_store();
3360 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3361 return Status::make_failure();
3362 }
3363 return status_def;
3364 }
3365
3366 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3367 Status status_use = find_use_store_unidirectional(def_store);
3368
3369 #ifdef ASSERT
3370 StoreNode* use_store = status_use.found_store();
3371 if (use_store != nullptr) {
3372 Status status_def = find_def_store_unidirectional(use_store);
3373 assert(status_def.found_store() == def_store &&
3374 status_def.found_range_check() == status_use.found_range_check(),
3375 "find_use_store and find_def_store must be symmetric");
3376 }
3377 #endif
3378
3379 return status_use;
3380 }
3381
3382 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3383 Status status_def = find_def_store_unidirectional(use_store);
3384
3385 #ifdef ASSERT
3386 StoreNode* def_store = status_def.found_store();
3387 if (def_store != nullptr) {
3388 Status status_use = find_use_store_unidirectional(def_store);
3389 assert(status_use.found_store() == use_store &&
3390 status_use.found_range_check() == status_def.found_range_check(),
3391 "find_use_store and find_def_store must be symmetric");
3392 }
3393 #endif
3394
3395 return status_def;
3396 }
3397
3398 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3399 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3400
3401 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3402 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3403 if (is_compatible_store(use_store)) {
3404 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3405 }
3406 }
3407
3408 return Status::make_failure();
3409 }
3410
3411 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3412 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3413
3414 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3415 if (!is_compatible_store(def_store)) {
3416 return Status::make_failure();
3417 }
3418
3419 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3420 }
3421
3422 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3423 // The merged store can be at most 8 bytes.
3424 const uint merge_list_max_size = 8 / _store->memory_size();
3425 assert(merge_list_max_size >= 2 &&
3426 merge_list_max_size <= 8 &&
3427 is_power_of_2(merge_list_max_size),
3428 "must be 2, 4 or 8");
3429
3430 // Traverse up the chain of adjacent def stores.
3431 StoreNode* current = _store;
3432 merge_list.push(current);
3433 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3434 Status status = find_adjacent_def_store(current);
3435 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3436
3437 current = status.found_store();
3438 if (current != nullptr) {
3439 merge_list.push(current);
3440
3441 // We can have at most one RangeCheck.
3442 if (status.found_range_check()) {
3443 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3444 break;
3445 }
3446 }
3447 }
3448
3449 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3450
3451 // Truncate the merge_list to a power of 2.
3452 const uint pow2size = round_down_power_of_2(merge_list.size());
3453 assert(pow2size >= 2, "must be merging at least 2 stores");
3454 while (merge_list.size() > pow2size) { merge_list.pop(); }
3455
3456 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3457 }
3458
3459 // Merge the input values of the smaller stores to a single larger input value.
3460 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3461 int new_memory_size = _store->memory_size() * merge_list.size();
3462 Node* first = merge_list.at(merge_list.size()-1);
3463 Node* merged_input_value = nullptr;
3464 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3465 assert(_value_order == ValueOrder::Const, "must match");
3466 // Pattern: [ConI, ConI, ...] -> new constant
3467 jlong con = 0;
3468 jlong bits_per_store = _store->memory_size() * 8;
3469 jlong mask = (((jlong)1) << bits_per_store) - 1;
3470 for (uint i = 0; i < merge_list.size(); i++) {
3471 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3472 #ifdef VM_LITTLE_ENDIAN
3473 con = con << bits_per_store;
3474 con = con | (mask & con_i);
3475 #else // VM_LITTLE_ENDIAN
3476 con_i = (mask & con_i) << (i * bits_per_store);
3477 con = con | con_i;
3478 #endif // VM_LITTLE_ENDIAN
3479 }
3480 merged_input_value = _phase->longcon(con);
3481 } else {
3482 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3483 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3484 // | |
3485 // _store first
3486 //
3487 Node* hi = _store->in(MemNode::ValueIn);
3488 Node* lo = first->in(MemNode::ValueIn);
3489 #ifndef VM_LITTLE_ENDIAN
3490 // `_store` and `first` are swapped in the diagram above
3491 swap(hi, lo);
3492 #endif // !VM_LITTLE_ENDIAN
3493 if (_value_order == ValueOrder::Reverse) {
3494 swap(hi, lo);
3495 }
3496 Node const* hi_base;
3497 jint hi_shift;
3498 merged_input_value = lo;
3499 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3500 assert(is_true, "must detect con RShift");
3501 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3502 // look through
3503 merged_input_value = merged_input_value->in(1);
3504 }
3505 if (merged_input_value != hi_base) {
3506 // merged_input_value is not the base
3507 return nullptr;
3508 }
3509 }
3510
3511 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3512 // Example:
3513 //
3514 // long base = ...;
3515 // a[0] = (byte)(base >> 0);
3516 // a[1] = (byte)(base >> 8);
3517 //
3518 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3519 }
3520
3521 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3522 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3523 "merged_input_value is either int or long, and new_memory_size is small enough");
3524
3525 if (_value_order == ValueOrder::Reverse) {
3526 assert(_store->memory_size() == 1, "only implemented for bytes");
3527 if (new_memory_size == 8) {
3528 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3529 } else if (new_memory_size == 4) {
3530 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3531 } else {
3532 assert(new_memory_size == 2, "sanity check");
3533 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3534 }
3535 }
3536 return merged_input_value;
3537 }
3538
3539 // //
3540 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3541 // | | | | | | //
3542 // | | | | +---------------+ | //
3543 // | | | | | | | //
3544 // | | +---------+ | | +---------------+ //
3545 // | | | | | | | | //
3546 // +--------------+ | | v1 +------------------------------+ | | v1 //
3547 // | | | | | | | | | | | | //
3548 // RangeCheck first_store RangeCheck | | first_store //
3549 // | | | | | | | //
3550 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3551 // | | ===> | | | //
3552 // +--------------+ | a2 v2 | | | //
3553 // | | | | | | | | //
3554 // | second_store | | | //
3555 // | | | | | [v1 v2 ... vn] //
3556 // ... ... | | | | //
3557 // | | | | | v //
3558 // +--------------+ | an vn +--------------+ | | merged_input_value //
3559 // | | | | | | | | //
3560 // last_store (= _store) merged_store //
3561 // //
3562 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3563 Node* first_store = merge_list.at(merge_list.size()-1);
3564 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3565 Node* first_mem = first_store->in(MemNode::Memory);
3566 Node* first_adr = first_store->in(MemNode::Address);
3567
3568 const TypePtr* new_adr_type = _store->adr_type();
3569
3570 int new_memory_size = _store->memory_size() * merge_list.size();
3571 BasicType bt = T_ILLEGAL;
3572 switch (new_memory_size) {
3573 case 2: bt = T_SHORT; break;
3574 case 4: bt = T_INT; break;
3575 case 8: bt = T_LONG; break;
3576 }
3577
3578 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3579 new_adr_type, merged_input_value, bt, MemNode::unordered);
3580
3581 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3582 // are all on the same slice. Hence, we need no barriers.
3583 merged_store->set_mismatched_access();
3584
3585 // Constants above may now also be be packed -> put candidate on worklist
3586 _phase->is_IterGVN()->_worklist.push(first_mem);
3587
3588 return merged_store;
3589 }
3590
3591 #ifndef PRODUCT
3592 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3593 stringStream ss;
3594 ss.print_cr("[TraceMergeStores]: Replace");
3595 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3596 merge_list.at(i)->dump("\n", false, &ss);
3597 }
3598 ss.print_cr("[TraceMergeStores]: with");
3599 merged_input_value->dump("\n", false, &ss);
3600 merged_store->dump("\n", false, &ss);
3601 tty->print("%s", ss.as_string());
3602 }
3603 #endif
3604
3605 //------------------------------Ideal------------------------------------------
3606 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3607 // When a store immediately follows a relevant allocation/initialization,
3608 // try to capture it into the initialization, or hoist it above.
3609 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3610 Node* p = MemNode::Ideal_common(phase, can_reshape);
3611 if (p) return (p == NodeSentinel) ? nullptr : p;
3612
3613 Node* mem = in(MemNode::Memory);
3614 Node* address = in(MemNode::Address);
3615 Node* value = in(MemNode::ValueIn);
3616 // Back-to-back stores to same address? Fold em up. Generally
3617 // unsafe if I have intervening uses...
3618 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3619 Node* st = mem;
3620 // If Store 'st' has more than one use, we cannot fold 'st' away.
3621 // For example, 'st' might be the final state at a conditional
3622 // return. Or, 'st' might be used by some node which is live at
3623 // the same time 'st' is live, which might be unschedulable. So,
3624 // require exactly ONE user until such time as we clone 'mem' for
3625 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3626 // true).
3627 while (st->is_Store() && st->outcnt() == 1) {
3628 // Looking at a dead closed cycle of memory?
3629 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3630 assert(Opcode() == st->Opcode() ||
3631 st->Opcode() == Op_StoreVector ||
3632 Opcode() == Op_StoreVector ||
3633 st->Opcode() == Op_StoreVectorScatter ||
3634 Opcode() == Op_StoreVectorScatter ||
3635 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3636 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3637 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3638 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3639 (st->adr_type()->isa_aryptr() && st->adr_type()->is_aryptr()->is_flat()) || // TODO 8343835
3640 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3641 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3642
3643 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3644 st->as_Store()->memory_size() <= this->memory_size()) {
3645 Node* use = st->raw_out(0);
3646 if (phase->is_IterGVN()) {
3647 phase->is_IterGVN()->rehash_node_delayed(use);
3648 }
3649 // It's OK to do this in the parser, since DU info is always accurate,
3650 // and the parser always refers to nodes via SafePointNode maps.
3651 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3652 return this;
3653 }
3654 st = st->in(MemNode::Memory);
3655 }
3656 }
3657
3658
3659 // Capture an unaliased, unconditional, simple store into an initializer.
3660 // Or, if it is independent of the allocation, hoist it above the allocation.
3661 if (ReduceFieldZeroing && /*can_reshape &&*/
3662 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3663 InitializeNode* init = mem->in(0)->as_Initialize();
3664 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3665 if (offset > 0) {
3666 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3667 // If the InitializeNode captured me, it made a raw copy of me,
3668 // and I need to disappear.
3669 if (moved != nullptr) {
3670 // %%% hack to ensure that Ideal returns a new node:
3671 mem = MergeMemNode::make(mem);
3672 return mem; // fold me away
3673 }
3674 }
3675 }
3676
3677 // Fold reinterpret cast into memory operation:
3678 // StoreX mem (MoveY2X v) => StoreY mem v
3679 if (value->is_Move()) {
3680 const Type* vt = value->in(1)->bottom_type();
3681 if (has_reinterpret_variant(vt)) {
3682 if (phase->C->post_loop_opts_phase()) {
3683 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3684 } else {
3685 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3686 }
3687 }
3688 }
3689
3690 if (MergeStores && UseUnalignedAccesses) {
3691 if (phase->C->merge_stores_phase()) {
3692 MergePrimitiveStores merge(phase, this);
3693 Node* progress = merge.run();
3694 if (progress != nullptr) { return progress; }
3695 } else {
3696 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3697 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3698 // the stores, and we merge the wrong sequence of stores.
3699 // Example:
3700 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3701 // Apply MergeStores:
3702 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3703 // Remove more RangeChecks:
3704 // StoreI [ StoreL ] StoreI
3705 // But now it would have been better to do this instead:
3706 // [ StoreL ] [ StoreL ]
3707 phase->C->record_for_merge_stores_igvn(this);
3708 }
3709 }
3710
3711 return nullptr; // No further progress
3712 }
3713
3714 //------------------------------Value-----------------------------------------
3715 const Type* StoreNode::Value(PhaseGVN* phase) const {
3716 // Either input is TOP ==> the result is TOP
3717 const Type *t1 = phase->type( in(MemNode::Memory) );
3718 if( t1 == Type::TOP ) return Type::TOP;
3719 const Type *t2 = phase->type( in(MemNode::Address) );
3720 if( t2 == Type::TOP ) return Type::TOP;
3721 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3722 if( t3 == Type::TOP ) return Type::TOP;
3723 return Type::MEMORY;
3724 }
3725
3726 //------------------------------Identity---------------------------------------
3727 // Remove redundant stores:
3728 // Store(m, p, Load(m, p)) changes to m.
3729 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3730 Node* StoreNode::Identity(PhaseGVN* phase) {
3731 Node* mem = in(MemNode::Memory);
3732 Node* adr = in(MemNode::Address);
3733 Node* val = in(MemNode::ValueIn);
3734
3735 Node* result = this;
3736
3737 // Load then Store? Then the Store is useless
3738 if (val->is_Load() &&
3739 val->in(MemNode::Address)->eqv_uncast(adr) &&
3740 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3741 val->as_Load()->store_Opcode() == Opcode()) {
3742 // Ensure vector type is the same
3743 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3744 result = mem;
3745 }
3746 }
3747
3748 // Two stores in a row of the same value?
3749 if (result == this &&
3750 mem->is_Store() &&
3751 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3752 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3753 mem->Opcode() == Opcode()) {
3754 if (!is_StoreVector()) {
3755 result = mem;
3756 } else {
3757 const StoreVectorNode* store_vector = as_StoreVector();
3758 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3759 const Node* store_indices = store_vector->indices();
3760 const Node* mem_indices = mem_vector->indices();
3761 const Node* store_mask = store_vector->mask();
3762 const Node* mem_mask = mem_vector->mask();
3763 // Ensure types, indices, and masks match
3764 if (store_vector->vect_type() == mem_vector->vect_type() &&
3765 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3766 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3767 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3768 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3769 result = mem;
3770 }
3771 }
3772 }
3773
3774 // Store of zero anywhere into a freshly-allocated object?
3775 // Then the store is useless.
3776 // (It must already have been captured by the InitializeNode.)
3777 if (result == this && ReduceFieldZeroing) {
3778 // a newly allocated object is already all-zeroes everywhere
3779 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3780 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::InitValue) == val)) {
3781 result = mem;
3782 }
3783
3784 if (result == this && phase->type(val)->is_zero_type()) {
3785 // the store may also apply to zero-bits in an earlier object
3786 Node* prev_mem = find_previous_store(phase);
3787 // Steps (a), (b): Walk past independent stores to find an exact match.
3788 if (prev_mem != nullptr) {
3789 Node* prev_val = can_see_stored_value(prev_mem, phase);
3790 if (prev_val != nullptr && prev_val == val) {
3791 // prev_val and val might differ by a cast; it would be good
3792 // to keep the more informative of the two.
3793 result = mem;
3794 }
3795 }
3796 }
3797 }
3798
3799 PhaseIterGVN* igvn = phase->is_IterGVN();
3800 if (result != this && igvn != nullptr) {
3801 MemBarNode* trailing = trailing_membar();
3802 if (trailing != nullptr) {
3803 #ifdef ASSERT
3804 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3805 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3806 #endif
3807 trailing->remove(igvn);
3808 }
3809 }
3810
3811 return result;
3812 }
3813
3814 //------------------------------match_edge-------------------------------------
3815 // Do we Match on this edge index or not? Match only memory & value
3816 uint StoreNode::match_edge(uint idx) const {
3817 return idx == MemNode::Address || idx == MemNode::ValueIn;
3818 }
3819
3820 //------------------------------cmp--------------------------------------------
3821 // Do not common stores up together. They generally have to be split
3822 // back up anyways, so do not bother.
3823 bool StoreNode::cmp( const Node &n ) const {
3824 return (&n == this); // Always fail except on self
3825 }
3826
3827 //------------------------------Ideal_masked_input-----------------------------
3828 // Check for a useless mask before a partial-word store
3829 // (StoreB ... (AndI valIn conIa) )
3830 // If (conIa & mask == mask) this simplifies to
3831 // (StoreB ... (valIn) )
3832 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3833 Node *val = in(MemNode::ValueIn);
3834 if( val->Opcode() == Op_AndI ) {
3835 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3836 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3837 set_req_X(MemNode::ValueIn, val->in(1), phase);
3838 return this;
3839 }
3840 }
3841 return nullptr;
3842 }
3843
3844
3845 //------------------------------Ideal_sign_extended_input----------------------
3846 // Check for useless sign-extension before a partial-word store
3847 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3848 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3849 // (StoreB ... (v))
3850 // If (conIL > conIR) under some conditions, it can be simplified into
3851 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3852 // This case happens when the value of the store was itself a left shift, that
3853 // gets merged into the inner left shift of the sign-extension. For instance,
3854 // if we have
3855 // array_of_shorts[0] = (short)(v << 2)
3856 // We get a structure such as:
3857 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3858 // that is simplified into
3859 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3860 // It is thus useful to handle cases where conIL > conIR. But this simplification
3861 // does not always hold. Let's see in which cases it's valid.
3862 //
3863 // Let's assume we have the following 32 bits integer v:
3864 // +----------------------------------+
3865 // | v[0..31] |
3866 // +----------------------------------+
3867 // 31 0
3868 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3869 // potentially different magnitudes.
3870 // We denote num_rejected_bits the number of bits of the discarded part. In this
3871 // case, num_rejected_bits == 24.
3872 //
3873 // Statement (proved further below in case analysis):
3874 // Given:
3875 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3876 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3877 // - conIL >= conIR
3878 // - num_rejected_bits >= conIR
3879 // Then this form:
3880 // (RShiftI _ (LShiftI _ v conIL) conIR)
3881 // can be replaced with this form:
3882 // (LShiftI _ v (conIL-conIR))
3883 //
3884 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3885 // have to be correct, as the higher bits are rejected / truncated by the store.
3886 //
3887 // The hypotheses
3888 // 0 <= conIL < BitsPerJavaInteger
3889 // 0 <= conIR < BitsPerJavaInteger
3890 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3891 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3892 // to v << 1).
3893 //
3894 //
3895 // If you don't like case analysis, jump after the conclusion.
3896 // ### Case 1 : conIL == conIR
3897 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3898 // If we do the shift left then right by 24 bits, we get:
3899 // after: v << 24
3900 // +---------+------------------------+
3901 // | v[0..7] | 0 |
3902 // +---------+------------------------+
3903 // 31 24 23 0
3904 // after: (v << 24) >> 24
3905 // +------------------------+---------+
3906 // | sign bit | v[0..7] |
3907 // +------------------------+---------+
3908 // 31 8 7 0
3909 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3910 // result) are the same before and after, so, indeed, simplifying is correct.
3911
3912 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3913 // If we do the shift left then right by 22 bits, we get:
3914 // after: v << 22
3915 // +---------+------------------------+
3916 // | v[0..9] | 0 |
3917 // +---------+------------------------+
3918 // 31 22 21 0
3919 // after: (v << 22) >> 22
3920 // +------------------------+---------+
3921 // | sign bit | v[0..9] |
3922 // +------------------------+---------+
3923 // 31 10 9 0
3924 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3925 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3926 // still correct.
3927
3928 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3929 // If we do the shift left then right by 26 bits, we get:
3930 // after: v << 26
3931 // +---------+------------------------+
3932 // | v[0..5] | 0 |
3933 // +---------+------------------------+
3934 // 31 26 25 0
3935 // after: (v << 26) >> 26
3936 // +------------------------+---------+
3937 // | sign bit | v[0..5] |
3938 // +------------------------+---------+
3939 // 31 6 5 0
3940 // The non-rejected bits are made of
3941 // - 0-5 => the bits 0 to 5 of v
3942 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3943 // Simplifying this as v is not correct.
3944 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3945 //
3946 // ### Case 2: conIL > conIR
3947 // ###### Case 2.1: num_rejected_bits == conIR
3948 // We take conIL == 26 for this example.
3949 // after: v << 26
3950 // +---------+------------------------+
3951 // | v[0..5] | 0 |
3952 // +---------+------------------------+
3953 // 31 26 25 0
3954 // after: (v << 26) >> 24
3955 // +------------------+---------+-----+
3956 // | sign bit | v[0..5] | 0 |
3957 // +------------------+---------+-----+
3958 // 31 8 7 2 1 0
3959 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3960 // The bits 6 and 7 of v have been thrown away after the shift left.
3961 // The simplification is still correct.
3962 //
3963 // ###### Case 2.2: num_rejected_bits > conIR.
3964 // Let's say conIL == 26 and conIR == 22.
3965 // after: v << 26
3966 // +---------+------------------------+
3967 // | v[0..5] | 0 |
3968 // +---------+------------------------+
3969 // 31 26 25 0
3970 // after: (v << 26) >> 22
3971 // +------------------+---------+-----+
3972 // | sign bit | v[0..5] | 0 |
3973 // +------------------+---------+-----+
3974 // 31 10 9 4 3 0
3975 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3976 // - 0-3 => 0
3977 // - 4-7 => bits 0 to 3 of v
3978 // The simplification is still correct.
3979 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3980 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3981 //
3982 // ###### But! Case 2.3: num_rejected_bits < conIR.
3983 // Let's see that this case is not as easy to simplify.
3984 // Let's say conIL == 28 and conIR == 26.
3985 // after: v << 28
3986 // +---------+------------------------+
3987 // | v[0..3] | 0 |
3988 // +---------+------------------------+
3989 // 31 28 27 0
3990 // after: (v << 28) >> 26
3991 // +------------------+---------+-----+
3992 // | sign bit | v[0..3] | 0 |
3993 // +------------------+---------+-----+
3994 // 31 6 5 2 1 0
3995 // The non-rejected bits are made of
3996 // - 0-1 => 0
3997 // - 2-5 => the bits 0 to 3 of v
3998 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3999 // Simplifying this as (v << 2) is not correct.
4000 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
4001 //
4002 // ### Conclusion:
4003 // Our hypotheses are indeed sufficient:
4004 // - 0 <= conIL < BitsPerJavaInteger
4005 // - 0 <= conIR < BitsPerJavaInteger
4006 // - conIL >= conIR
4007 // - num_rejected_bits >= conIR
4008 //
4009 // ### A rationale without case analysis:
4010 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
4011 // zeroes are added. After the shift right, conIR lower bits of the previous result
4012 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
4013 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
4014 // the point of the simplification is to get an expression of the form
4015 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
4016 // simplification is not correct if conIL < conIR.
4017 //
4018 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
4019 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
4020 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
4021 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
4022 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
4023 //
4024 // We do not treat the case conIL < conIR here since the point of this function is
4025 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
4026 // of treating conIL > conIR comes from the cases where the sign-extended value is
4027 // also left-shift expression. Computing the sign-extension of a right-shift expression
4028 // doesn't yield a situation such as
4029 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
4030 // where conIL < conIR.
4031 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
4032 Node* shr = in(MemNode::ValueIn);
4033 if (shr->Opcode() == Op_RShiftI) {
4034 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
4035 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
4036 Node* shl = shr->in(1);
4037 if (shl->Opcode() == Op_LShiftI) {
4038 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
4039 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
4040 if (conIL->get_con() == conIR->get_con()) {
4041 set_req_X(MemNode::ValueIn, shl->in(1), phase);
4042 return this;
4043 }
4044 if (conIL->get_con() > conIR->get_con()) {
4045 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
4046 set_req_X(MemNode::ValueIn, new_shl, phase);
4047 return this;
4048 }
4049 }
4050 }
4051 }
4052 }
4053 return nullptr;
4054 }
4055
4056 //------------------------------value_never_loaded-----------------------------------
4057 // Determine whether there are any possible loads of the value stored.
4058 // For simplicity, we actually check if there are any loads from the
4059 // address stored to, not just for loads of the value stored by this node.
4060 //
4061 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
4062 Node *adr = in(Address);
4063 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
4064 if (adr_oop == nullptr)
4065 return false;
4066 if (!adr_oop->is_known_instance_field())
4067 return false; // if not a distinct instance, there may be aliases of the address
4068 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
4069 Node *use = adr->fast_out(i);
4070 if (use->is_Load() || use->is_LoadStore()) {
4071 return false;
4072 }
4073 }
4074 return true;
4075 }
4076
4077 MemBarNode* StoreNode::trailing_membar() const {
4078 if (is_release()) {
4079 MemBarNode* trailing_mb = nullptr;
4080 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4081 Node* u = fast_out(i);
4082 if (u->is_MemBar()) {
4083 if (u->as_MemBar()->trailing_store()) {
4084 assert(u->Opcode() == Op_MemBarVolatile, "");
4085 assert(trailing_mb == nullptr, "only one");
4086 trailing_mb = u->as_MemBar();
4087 #ifdef ASSERT
4088 Node* leading = u->as_MemBar()->leading_membar();
4089 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4090 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
4091 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
4092 #endif
4093 } else {
4094 assert(u->as_MemBar()->standalone(), "");
4095 }
4096 }
4097 }
4098 return trailing_mb;
4099 }
4100 return nullptr;
4101 }
4102
4103
4104 //=============================================================================
4105 //------------------------------Ideal------------------------------------------
4106 // If the store is from an AND mask that leaves the low bits untouched, then
4107 // we can skip the AND operation. If the store is from a sign-extension
4108 // (a left shift, then right shift) we can skip both.
4109 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
4110 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
4111 if( progress != nullptr ) return progress;
4112
4113 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
4114 if( progress != nullptr ) return progress;
4115
4116 // Finally check the default case
4117 return StoreNode::Ideal(phase, can_reshape);
4118 }
4119
4120 //=============================================================================
4121 //------------------------------Ideal------------------------------------------
4122 // If the store is from an AND mask that leaves the low bits untouched, then
4123 // we can skip the AND operation
4124 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
4125 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
4126 if( progress != nullptr ) return progress;
4127
4128 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4129 if( progress != nullptr ) return progress;
4130
4131 // Finally check the default case
4132 return StoreNode::Ideal(phase, can_reshape);
4133 }
4134
4135 //=============================================================================
4136 //----------------------------------SCMemProjNode------------------------------
4137 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4138 {
4139 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4140 return Type::TOP;
4141 }
4142 return bottom_type();
4143 }
4144
4145 //=============================================================================
4146 //----------------------------------LoadStoreNode------------------------------
4147 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4148 : Node(required),
4149 _type(rt),
4150 _adr_type(at),
4151 _barrier_data(0)
4152 {
4153 init_req(MemNode::Control, c );
4154 init_req(MemNode::Memory , mem);
4155 init_req(MemNode::Address, adr);
4156 init_req(MemNode::ValueIn, val);
4157 init_class_id(Class_LoadStore);
4158 }
4159
4160 //------------------------------Value-----------------------------------------
4161 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4162 // Either input is TOP ==> the result is TOP
4163 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4164 return Type::TOP;
4165 }
4166 const Type* t = phase->type(in(MemNode::Memory));
4167 if (t == Type::TOP) {
4168 return Type::TOP;
4169 }
4170 t = phase->type(in(MemNode::Address));
4171 if (t == Type::TOP) {
4172 return Type::TOP;
4173 }
4174 t = phase->type(in(MemNode::ValueIn));
4175 if (t == Type::TOP) {
4176 return Type::TOP;
4177 }
4178 return bottom_type();
4179 }
4180
4181 uint LoadStoreNode::ideal_reg() const {
4182 return _type->ideal_reg();
4183 }
4184
4185 // This method conservatively checks if the result of a LoadStoreNode is
4186 // used, that is, if it returns true, then it is definitely the case that
4187 // the result of the node is not needed.
4188 // For example, GetAndAdd can be matched into a lock_add instead of a
4189 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4190 bool LoadStoreNode::result_not_used() const {
4191 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4192 Node *x = fast_out(i);
4193 if (x->Opcode() == Op_SCMemProj) {
4194 continue;
4195 }
4196 if (x->bottom_type() == TypeTuple::MEMBAR &&
4197 !x->is_Call() &&
4198 x->Opcode() != Op_Blackhole) {
4199 continue;
4200 }
4201 return false;
4202 }
4203 return true;
4204 }
4205
4206 MemBarNode* LoadStoreNode::trailing_membar() const {
4207 MemBarNode* trailing = nullptr;
4208 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4209 Node* u = fast_out(i);
4210 if (u->is_MemBar()) {
4211 if (u->as_MemBar()->trailing_load_store()) {
4212 assert(u->Opcode() == Op_MemBarAcquire, "");
4213 assert(trailing == nullptr, "only one");
4214 trailing = u->as_MemBar();
4215 #ifdef ASSERT
4216 Node* leading = trailing->leading_membar();
4217 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4218 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4219 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4220 #endif
4221 } else {
4222 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4223 }
4224 }
4225 }
4226
4227 return trailing;
4228 }
4229
4230 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4231
4232 //=============================================================================
4233 //----------------------------------LoadStoreConditionalNode--------------------
4234 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4235 init_req(ExpectedIn, ex );
4236 }
4237
4238 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4239 // Either input is TOP ==> the result is TOP
4240 const Type* t = phase->type(in(ExpectedIn));
4241 if (t == Type::TOP) {
4242 return Type::TOP;
4243 }
4244 return LoadStoreNode::Value(phase);
4245 }
4246
4247 //=============================================================================
4248 //-------------------------------adr_type--------------------------------------
4249 const TypePtr* ClearArrayNode::adr_type() const {
4250 Node *adr = in(3);
4251 if (adr == nullptr) return nullptr; // node is dead
4252 return MemNode::calculate_adr_type(adr->bottom_type());
4253 }
4254
4255 //------------------------------match_edge-------------------------------------
4256 // Do we Match on this edge index or not? Do not match memory
4257 uint ClearArrayNode::match_edge(uint idx) const {
4258 return idx > 1;
4259 }
4260
4261 //------------------------------Identity---------------------------------------
4262 // Clearing a zero length array does nothing
4263 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4264 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4265 }
4266
4267 //------------------------------Idealize---------------------------------------
4268 // Clearing a short array is faster with stores
4269 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4270 // Already know this is a large node, do not try to ideal it
4271 if (_is_large) return nullptr;
4272
4273 const int unit = BytesPerLong;
4274 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4275 if (!t) return nullptr;
4276 if (!t->is_con()) return nullptr;
4277 intptr_t raw_count = t->get_con();
4278 intptr_t size = raw_count;
4279 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4280 // Clearing nothing uses the Identity call.
4281 // Negative clears are possible on dead ClearArrays
4282 // (see jck test stmt114.stmt11402.val).
4283 if (size <= 0 || size % unit != 0) return nullptr;
4284 intptr_t count = size / unit;
4285 // Length too long; communicate this to matchers and assemblers.
4286 // Assemblers are responsible to produce fast hardware clears for it.
4287 if (size > InitArrayShortSize) {
4288 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
4289 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4290 return nullptr;
4291 }
4292 if (!IdealizeClearArrayNode) return nullptr;
4293 Node *mem = in(1);
4294 if( phase->type(mem)==Type::TOP ) return nullptr;
4295 Node *adr = in(3);
4296 const Type* at = phase->type(adr);
4297 if( at==Type::TOP ) return nullptr;
4298 const TypePtr* atp = at->isa_ptr();
4299 // adjust atp to be the correct array element address type
4300 if (atp == nullptr) atp = TypePtr::BOTTOM;
4301 else atp = atp->add_offset(Type::OffsetBot);
4302 // Get base for derived pointer purposes
4303 if( adr->Opcode() != Op_AddP ) Unimplemented();
4304 Node *base = adr->in(1);
4305
4306 Node *val = in(4);
4307 Node *off = phase->MakeConX(BytesPerLong);
4308 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4309 count--;
4310 while( count-- ) {
4311 mem = phase->transform(mem);
4312 adr = phase->transform(new AddPNode(base,adr,off));
4313 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4314 }
4315 return mem;
4316 }
4317
4318 //----------------------------step_through----------------------------------
4319 // Return allocation input memory edge if it is different instance
4320 // or itself if it is the one we are looking for.
4321 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4322 Node* n = *np;
4323 assert(n->is_ClearArray(), "sanity");
4324 intptr_t offset;
4325 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4326 // This method is called only before Allocate nodes are expanded
4327 // during macro nodes expansion. Before that ClearArray nodes are
4328 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4329 // Allocate nodes are expanded) which follows allocations.
4330 assert(alloc != nullptr, "should have allocation");
4331 if (alloc->_idx == instance_id) {
4332 // Can not bypass initialization of the instance we are looking for.
4333 return false;
4334 }
4335 // Otherwise skip it.
4336 InitializeNode* init = alloc->initialization();
4337 if (init != nullptr)
4338 *np = init->in(TypeFunc::Memory);
4339 else
4340 *np = alloc->in(TypeFunc::Memory);
4341 return true;
4342 }
4343
4344 //----------------------------clear_memory-------------------------------------
4345 // Generate code to initialize object storage to zero.
4346 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4347 Node* val,
4348 Node* raw_val,
4349 intptr_t start_offset,
4350 Node* end_offset,
4351 PhaseGVN* phase) {
4352 intptr_t offset = start_offset;
4353
4354 int unit = BytesPerLong;
4355 if ((offset % unit) != 0) {
4356 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4357 adr = phase->transform(adr);
4358 const TypePtr* atp = TypeRawPtr::BOTTOM;
4359 if (val != nullptr) {
4360 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4361 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4362 } else {
4363 assert(raw_val == nullptr, "val may not be null");
4364 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4365 }
4366 mem = phase->transform(mem);
4367 offset += BytesPerInt;
4368 }
4369 assert((offset % unit) == 0, "");
4370
4371 // Initialize the remaining stuff, if any, with a ClearArray.
4372 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
4373 }
4374
4375 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4376 Node* raw_val,
4377 Node* start_offset,
4378 Node* end_offset,
4379 PhaseGVN* phase) {
4380 if (start_offset == end_offset) {
4381 // nothing to do
4382 return mem;
4383 }
4384
4385 int unit = BytesPerLong;
4386 Node* zbase = start_offset;
4387 Node* zend = end_offset;
4388
4389 // Scale to the unit required by the CPU:
4390 if (!Matcher::init_array_count_is_in_bytes) {
4391 Node* shift = phase->intcon(exact_log2(unit));
4392 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4393 zend = phase->transform(new URShiftXNode(zend, shift) );
4394 }
4395
4396 // Bulk clear double-words
4397 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4398 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4399 if (raw_val == nullptr) {
4400 raw_val = phase->MakeConX(0);
4401 }
4402 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4403 return phase->transform(mem);
4404 }
4405
4406 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4407 Node* val,
4408 Node* raw_val,
4409 intptr_t start_offset,
4410 intptr_t end_offset,
4411 PhaseGVN* phase) {
4412 if (start_offset == end_offset) {
4413 // nothing to do
4414 return mem;
4415 }
4416
4417 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4418 intptr_t done_offset = end_offset;
4419 if ((done_offset % BytesPerLong) != 0) {
4420 done_offset -= BytesPerInt;
4421 }
4422 if (done_offset > start_offset) {
4423 mem = clear_memory(ctl, mem, dest, val, raw_val,
4424 start_offset, phase->MakeConX(done_offset), phase);
4425 }
4426 if (done_offset < end_offset) { // emit the final 32-bit store
4427 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4428 adr = phase->transform(adr);
4429 const TypePtr* atp = TypeRawPtr::BOTTOM;
4430 if (val != nullptr) {
4431 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4432 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4433 } else {
4434 assert(raw_val == nullptr, "val may not be null");
4435 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4436 }
4437 mem = phase->transform(mem);
4438 done_offset += BytesPerInt;
4439 }
4440 assert(done_offset == end_offset, "");
4441 return mem;
4442 }
4443
4444 //=============================================================================
4445 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4446 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4447 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4448 #ifdef ASSERT
4449 , _pair_idx(0)
4450 #endif
4451 {
4452 init_class_id(Class_MemBar);
4453 Node* top = C->top();
4454 init_req(TypeFunc::I_O,top);
4455 init_req(TypeFunc::FramePtr,top);
4456 init_req(TypeFunc::ReturnAdr,top);
4457 if (precedent != nullptr)
4458 init_req(TypeFunc::Parms, precedent);
4459 }
4460
4461 //------------------------------cmp--------------------------------------------
4462 uint MemBarNode::hash() const { return NO_HASH; }
4463 bool MemBarNode::cmp( const Node &n ) const {
4464 return (&n == this); // Always fail except on self
4465 }
4466
4467 //------------------------------make-------------------------------------------
4468 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4469 switch (opcode) {
4470 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4471 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4472 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4473 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4474 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4475 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4476 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4477 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4478 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4479 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4480 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4481 case Op_Initialize: return new InitializeNode(C, atp, pn);
4482 default: ShouldNotReachHere(); return nullptr;
4483 }
4484 }
4485
4486 void MemBarNode::remove(PhaseIterGVN *igvn) {
4487 if (outcnt() != 2) {
4488 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4489 assert(outcnt() == 1, "Only control then");
4490 }
4491 if (trailing_store() || trailing_load_store()) {
4492 MemBarNode* leading = leading_membar();
4493 if (leading != nullptr) {
4494 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4495 leading->remove(igvn);
4496 }
4497 }
4498 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4499 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4500 }
4501 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4502 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4503 }
4504 }
4505
4506 //------------------------------Ideal------------------------------------------
4507 // Return a node which is more "ideal" than the current node. Strip out
4508 // control copies
4509 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4510 if (remove_dead_region(phase, can_reshape)) return this;
4511 // Don't bother trying to transform a dead node
4512 if (in(0) && in(0)->is_top()) {
4513 return nullptr;
4514 }
4515
4516 bool progress = false;
4517 // Eliminate volatile MemBars for scalar replaced objects.
4518 if (can_reshape && req() == (Precedent+1)) {
4519 bool eliminate = false;
4520 int opc = Opcode();
4521 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4522 // Volatile field loads and stores.
4523 Node* my_mem = in(MemBarNode::Precedent);
4524 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4525 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4526 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4527 // replace this Precedent (decodeN) with the Load instead.
4528 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4529 Node* load_node = my_mem->in(1);
4530 set_req(MemBarNode::Precedent, load_node);
4531 phase->is_IterGVN()->_worklist.push(my_mem);
4532 my_mem = load_node;
4533 } else {
4534 assert(my_mem->unique_out() == this, "sanity");
4535 assert(!trailing_load_store(), "load store node can't be eliminated");
4536 del_req(Precedent);
4537 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4538 my_mem = nullptr;
4539 }
4540 progress = true;
4541 }
4542 if (my_mem != nullptr && my_mem->is_Mem()) {
4543 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4544 // Check for scalar replaced object reference.
4545 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4546 t_oop->offset() != Type::OffsetBot &&
4547 t_oop->offset() != Type::OffsetTop) {
4548 eliminate = true;
4549 }
4550 }
4551 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4552 // Final field stores.
4553 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4554 if ((alloc != nullptr) && alloc->is_Allocate() &&
4555 alloc->as_Allocate()->does_not_escape_thread()) {
4556 // The allocated object does not escape.
4557 eliminate = true;
4558 }
4559 }
4560 if (eliminate) {
4561 // Replace MemBar projections by its inputs.
4562 PhaseIterGVN* igvn = phase->is_IterGVN();
4563 remove(igvn);
4564 // Must return either the original node (now dead) or a new node
4565 // (Do not return a top here, since that would break the uniqueness of top.)
4566 return new ConINode(TypeInt::ZERO);
4567 }
4568 }
4569 return progress ? this : nullptr;
4570 }
4571
4572 //------------------------------Value------------------------------------------
4573 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4574 if( !in(0) ) return Type::TOP;
4575 if( phase->type(in(0)) == Type::TOP )
4576 return Type::TOP;
4577 return TypeTuple::MEMBAR;
4578 }
4579
4580 //------------------------------match------------------------------------------
4581 // Construct projections for memory.
4582 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4583 switch (proj->_con) {
4584 case TypeFunc::Control:
4585 case TypeFunc::Memory:
4586 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4587 }
4588 ShouldNotReachHere();
4589 return nullptr;
4590 }
4591
4592 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4593 trailing->_kind = TrailingStore;
4594 leading->_kind = LeadingStore;
4595 #ifdef ASSERT
4596 trailing->_pair_idx = leading->_idx;
4597 leading->_pair_idx = leading->_idx;
4598 #endif
4599 }
4600
4601 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4602 trailing->_kind = TrailingLoadStore;
4603 leading->_kind = LeadingLoadStore;
4604 #ifdef ASSERT
4605 trailing->_pair_idx = leading->_idx;
4606 leading->_pair_idx = leading->_idx;
4607 #endif
4608 }
4609
4610 MemBarNode* MemBarNode::trailing_membar() const {
4611 ResourceMark rm;
4612 Node* trailing = (Node*)this;
4613 VectorSet seen;
4614 Node_Stack multis(0);
4615 do {
4616 Node* c = trailing;
4617 uint i = 0;
4618 do {
4619 trailing = nullptr;
4620 for (; i < c->outcnt(); i++) {
4621 Node* next = c->raw_out(i);
4622 if (next != c && next->is_CFG()) {
4623 if (c->is_MultiBranch()) {
4624 if (multis.node() == c) {
4625 multis.set_index(i+1);
4626 } else {
4627 multis.push(c, i+1);
4628 }
4629 }
4630 trailing = next;
4631 break;
4632 }
4633 }
4634 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4635 break;
4636 }
4637 while (multis.size() > 0) {
4638 c = multis.node();
4639 i = multis.index();
4640 if (i < c->req()) {
4641 break;
4642 }
4643 multis.pop();
4644 }
4645 } while (multis.size() > 0);
4646 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4647
4648 MemBarNode* mb = trailing->as_MemBar();
4649 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4650 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4651 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4652 return mb;
4653 }
4654
4655 MemBarNode* MemBarNode::leading_membar() const {
4656 ResourceMark rm;
4657 VectorSet seen;
4658 Node_Stack regions(0);
4659 Node* leading = in(0);
4660 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4661 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4662 leading = nullptr;
4663 while (regions.size() > 0 && leading == nullptr) {
4664 Node* r = regions.node();
4665 uint i = regions.index();
4666 if (i < r->req()) {
4667 leading = r->in(i);
4668 regions.set_index(i+1);
4669 } else {
4670 regions.pop();
4671 }
4672 }
4673 if (leading == nullptr) {
4674 assert(regions.size() == 0, "all paths should have been tried");
4675 return nullptr;
4676 }
4677 }
4678 if (leading->is_Region()) {
4679 regions.push(leading, 2);
4680 leading = leading->in(1);
4681 } else {
4682 leading = leading->in(0);
4683 }
4684 }
4685 #ifdef ASSERT
4686 Unique_Node_List wq;
4687 wq.push((Node*)this);
4688 uint found = 0;
4689 for (uint i = 0; i < wq.size(); i++) {
4690 Node* n = wq.at(i);
4691 if (n->is_Region()) {
4692 for (uint j = 1; j < n->req(); j++) {
4693 Node* in = n->in(j);
4694 if (in != nullptr && !in->is_top()) {
4695 wq.push(in);
4696 }
4697 }
4698 } else {
4699 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4700 assert(n == leading, "consistency check failed");
4701 found++;
4702 } else {
4703 Node* in = n->in(0);
4704 if (in != nullptr && !in->is_top()) {
4705 wq.push(in);
4706 }
4707 }
4708 }
4709 }
4710 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4711 #endif
4712 if (leading == nullptr) {
4713 return nullptr;
4714 }
4715 MemBarNode* mb = leading->as_MemBar();
4716 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4717 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4718 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4719 return mb;
4720 }
4721
4722
4723 //===========================InitializeNode====================================
4724 // SUMMARY:
4725 // This node acts as a memory barrier on raw memory, after some raw stores.
4726 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4727 // The Initialize can 'capture' suitably constrained stores as raw inits.
4728 // It can coalesce related raw stores into larger units (called 'tiles').
4729 // It can avoid zeroing new storage for memory units which have raw inits.
4730 // At macro-expansion, it is marked 'complete', and does not optimize further.
4731 //
4732 // EXAMPLE:
4733 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4734 // ctl = incoming control; mem* = incoming memory
4735 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4736 // First allocate uninitialized memory and fill in the header:
4737 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4738 // ctl := alloc.Control; mem* := alloc.Memory*
4739 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4740 // Then initialize to zero the non-header parts of the raw memory block:
4741 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4742 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4743 // After the initialize node executes, the object is ready for service:
4744 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4745 // Suppose its body is immediately initialized as {1,2}:
4746 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4747 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4748 // mem.SLICE(#short[*]) := store2
4749 //
4750 // DETAILS:
4751 // An InitializeNode collects and isolates object initialization after
4752 // an AllocateNode and before the next possible safepoint. As a
4753 // memory barrier (MemBarNode), it keeps critical stores from drifting
4754 // down past any safepoint or any publication of the allocation.
4755 // Before this barrier, a newly-allocated object may have uninitialized bits.
4756 // After this barrier, it may be treated as a real oop, and GC is allowed.
4757 //
4758 // The semantics of the InitializeNode include an implicit zeroing of
4759 // the new object from object header to the end of the object.
4760 // (The object header and end are determined by the AllocateNode.)
4761 //
4762 // Certain stores may be added as direct inputs to the InitializeNode.
4763 // These stores must update raw memory, and they must be to addresses
4764 // derived from the raw address produced by AllocateNode, and with
4765 // a constant offset. They must be ordered by increasing offset.
4766 // The first one is at in(RawStores), the last at in(req()-1).
4767 // Unlike most memory operations, they are not linked in a chain,
4768 // but are displayed in parallel as users of the rawmem output of
4769 // the allocation.
4770 //
4771 // (See comments in InitializeNode::capture_store, which continue
4772 // the example given above.)
4773 //
4774 // When the associated Allocate is macro-expanded, the InitializeNode
4775 // may be rewritten to optimize collected stores. A ClearArrayNode
4776 // may also be created at that point to represent any required zeroing.
4777 // The InitializeNode is then marked 'complete', prohibiting further
4778 // capturing of nearby memory operations.
4779 //
4780 // During macro-expansion, all captured initializations which store
4781 // constant values of 32 bits or smaller are coalesced (if advantageous)
4782 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4783 // initialized in fewer memory operations. Memory words which are
4784 // covered by neither tiles nor non-constant stores are pre-zeroed
4785 // by explicit stores of zero. (The code shape happens to do all
4786 // zeroing first, then all other stores, with both sequences occurring
4787 // in order of ascending offsets.)
4788 //
4789 // Alternatively, code may be inserted between an AllocateNode and its
4790 // InitializeNode, to perform arbitrary initialization of the new object.
4791 // E.g., the object copying intrinsics insert complex data transfers here.
4792 // The initialization must then be marked as 'complete' disable the
4793 // built-in zeroing semantics and the collection of initializing stores.
4794 //
4795 // While an InitializeNode is incomplete, reads from the memory state
4796 // produced by it are optimizable if they match the control edge and
4797 // new oop address associated with the allocation/initialization.
4798 // They return a stored value (if the offset matches) or else zero.
4799 // A write to the memory state, if it matches control and address,
4800 // and if it is to a constant offset, may be 'captured' by the
4801 // InitializeNode. It is cloned as a raw memory operation and rewired
4802 // inside the initialization, to the raw oop produced by the allocation.
4803 // Operations on addresses which are provably distinct (e.g., to
4804 // other AllocateNodes) are allowed to bypass the initialization.
4805 //
4806 // The effect of all this is to consolidate object initialization
4807 // (both arrays and non-arrays, both piecewise and bulk) into a
4808 // single location, where it can be optimized as a unit.
4809 //
4810 // Only stores with an offset less than TrackedInitializationLimit words
4811 // will be considered for capture by an InitializeNode. This puts a
4812 // reasonable limit on the complexity of optimized initializations.
4813
4814 //---------------------------InitializeNode------------------------------------
4815 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4816 : MemBarNode(C, adr_type, rawoop),
4817 _is_complete(Incomplete), _does_not_escape(false)
4818 {
4819 init_class_id(Class_Initialize);
4820
4821 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4822 assert(in(RawAddress) == rawoop, "proper init");
4823 // Note: allocation() can be null, for secondary initialization barriers
4824 }
4825
4826 // Since this node is not matched, it will be processed by the
4827 // register allocator. Declare that there are no constraints
4828 // on the allocation of the RawAddress edge.
4829 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4830 // This edge should be set to top, by the set_complete. But be conservative.
4831 if (idx == InitializeNode::RawAddress)
4832 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4833 return RegMask::Empty;
4834 }
4835
4836 Node* InitializeNode::memory(uint alias_idx) {
4837 Node* mem = in(Memory);
4838 if (mem->is_MergeMem()) {
4839 return mem->as_MergeMem()->memory_at(alias_idx);
4840 } else {
4841 // incoming raw memory is not split
4842 return mem;
4843 }
4844 }
4845
4846 bool InitializeNode::is_non_zero() {
4847 if (is_complete()) return false;
4848 remove_extra_zeroes();
4849 return (req() > RawStores);
4850 }
4851
4852 void InitializeNode::set_complete(PhaseGVN* phase) {
4853 assert(!is_complete(), "caller responsibility");
4854 _is_complete = Complete;
4855
4856 // After this node is complete, it contains a bunch of
4857 // raw-memory initializations. There is no need for
4858 // it to have anything to do with non-raw memory effects.
4859 // Therefore, tell all non-raw users to re-optimize themselves,
4860 // after skipping the memory effects of this initialization.
4861 PhaseIterGVN* igvn = phase->is_IterGVN();
4862 if (igvn) igvn->add_users_to_worklist(this);
4863 }
4864
4865 // convenience function
4866 // return false if the init contains any stores already
4867 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4868 InitializeNode* init = initialization();
4869 if (init == nullptr || init->is_complete()) {
4870 return false;
4871 }
4872 init->remove_extra_zeroes();
4873 // for now, if this allocation has already collected any inits, bail:
4874 if (init->is_non_zero()) return false;
4875 init->set_complete(phase);
4876 return true;
4877 }
4878
4879 void InitializeNode::remove_extra_zeroes() {
4880 if (req() == RawStores) return;
4881 Node* zmem = zero_memory();
4882 uint fill = RawStores;
4883 for (uint i = fill; i < req(); i++) {
4884 Node* n = in(i);
4885 if (n->is_top() || n == zmem) continue; // skip
4886 if (fill < i) set_req(fill, n); // compact
4887 ++fill;
4888 }
4889 // delete any empty spaces created:
4890 while (fill < req()) {
4891 del_req(fill);
4892 }
4893 }
4894
4895 // Helper for remembering which stores go with which offsets.
4896 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4897 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4898 intptr_t offset = -1;
4899 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4900 phase, offset);
4901 if (base == nullptr) return -1; // something is dead,
4902 if (offset < 0) return -1; // dead, dead
4903 return offset;
4904 }
4905
4906 // Helper for proving that an initialization expression is
4907 // "simple enough" to be folded into an object initialization.
4908 // Attempts to prove that a store's initial value 'n' can be captured
4909 // within the initialization without creating a vicious cycle, such as:
4910 // { Foo p = new Foo(); p.next = p; }
4911 // True for constants and parameters and small combinations thereof.
4912 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4913 ResourceMark rm;
4914 Unique_Node_List worklist;
4915 worklist.push(value);
4916
4917 uint complexity_limit = 20;
4918 for (uint j = 0; j < worklist.size(); j++) {
4919 if (j >= complexity_limit) {
4920 return false; // Bail out if processed too many nodes
4921 }
4922
4923 Node* n = worklist.at(j);
4924 if (n == nullptr) continue; // (can this really happen?)
4925 if (n->is_Proj()) n = n->in(0);
4926 if (n == this) return false; // found a cycle
4927 if (n->is_Con()) continue;
4928 if (n->is_Start()) continue; // params, etc., are OK
4929 if (n->is_Root()) continue; // even better
4930
4931 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4932 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4933 continue;
4934 }
4935
4936 Node* ctl = n->in(0);
4937 if (ctl != nullptr && !ctl->is_top()) {
4938 if (ctl->is_Proj()) ctl = ctl->in(0);
4939 if (ctl == this) return false;
4940
4941 // If we already know that the enclosing memory op is pinned right after
4942 // the init, then any control flow that the store has picked up
4943 // must have preceded the init, or else be equal to the init.
4944 // Even after loop optimizations (which might change control edges)
4945 // a store is never pinned *before* the availability of its inputs.
4946 if (!MemNode::all_controls_dominate(n, this)) {
4947 return false; // failed to prove a good control
4948 }
4949 }
4950
4951 // Check data edges for possible dependencies on 'this'.
4952 for (uint i = 1; i < n->req(); i++) {
4953 Node* m = n->in(i);
4954 if (m == nullptr || m == n || m->is_top()) continue;
4955
4956 // Only process data inputs once
4957 worklist.push(m);
4958 }
4959 }
4960
4961 return true;
4962 }
4963
4964 // Here are all the checks a Store must pass before it can be moved into
4965 // an initialization. Returns zero if a check fails.
4966 // On success, returns the (constant) offset to which the store applies,
4967 // within the initialized memory.
4968 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4969 const int FAIL = 0;
4970 if (st->req() != MemNode::ValueIn + 1)
4971 return FAIL; // an inscrutable StoreNode (card mark?)
4972 Node* ctl = st->in(MemNode::Control);
4973 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4974 return FAIL; // must be unconditional after the initialization
4975 Node* mem = st->in(MemNode::Memory);
4976 if (!(mem->is_Proj() && mem->in(0) == this))
4977 return FAIL; // must not be preceded by other stores
4978 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4979 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4980 !bs->can_initialize_object(st)) {
4981 return FAIL;
4982 }
4983 Node* adr = st->in(MemNode::Address);
4984 intptr_t offset;
4985 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4986 if (alloc == nullptr)
4987 return FAIL; // inscrutable address
4988 if (alloc != allocation())
4989 return FAIL; // wrong allocation! (store needs to float up)
4990 int size_in_bytes = st->memory_size();
4991 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4992 return FAIL; // mismatched access
4993 }
4994 Node* val = st->in(MemNode::ValueIn);
4995
4996 if (!detect_init_independence(val, phase))
4997 return FAIL; // stored value must be 'simple enough'
4998
4999 // The Store can be captured only if nothing after the allocation
5000 // and before the Store is using the memory location that the store
5001 // overwrites.
5002 bool failed = false;
5003 // If is_complete_with_arraycopy() is true the shape of the graph is
5004 // well defined and is safe so no need for extra checks.
5005 if (!is_complete_with_arraycopy()) {
5006 // We are going to look at each use of the memory state following
5007 // the allocation to make sure nothing reads the memory that the
5008 // Store writes.
5009 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
5010 int alias_idx = phase->C->get_alias_index(t_adr);
5011 ResourceMark rm;
5012 Unique_Node_List mems;
5013 mems.push(mem);
5014 Node* unique_merge = nullptr;
5015 for (uint next = 0; next < mems.size(); ++next) {
5016 Node *m = mems.at(next);
5017 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5018 Node *n = m->fast_out(j);
5019 if (n->outcnt() == 0) {
5020 continue;
5021 }
5022 if (n == st) {
5023 continue;
5024 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
5025 // If the control of this use is different from the control
5026 // of the Store which is right after the InitializeNode then
5027 // this node cannot be between the InitializeNode and the
5028 // Store.
5029 continue;
5030 } else if (n->is_MergeMem()) {
5031 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
5032 // We can hit a MergeMemNode (that will likely go away
5033 // later) that is a direct use of the memory state
5034 // following the InitializeNode on the same slice as the
5035 // store node that we'd like to capture. We need to check
5036 // the uses of the MergeMemNode.
5037 mems.push(n);
5038 }
5039 } else if (n->is_Mem()) {
5040 Node* other_adr = n->in(MemNode::Address);
5041 if (other_adr == adr) {
5042 failed = true;
5043 break;
5044 } else {
5045 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
5046 if (other_t_adr != nullptr) {
5047 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
5048 if (other_alias_idx == alias_idx) {
5049 // A load from the same memory slice as the store right
5050 // after the InitializeNode. We check the control of the
5051 // object/array that is loaded from. If it's the same as
5052 // the store control then we cannot capture the store.
5053 assert(!n->is_Store(), "2 stores to same slice on same control?");
5054 Node* base = other_adr;
5055 if (base->is_Phi()) {
5056 // In rare case, base may be a PhiNode and it may read
5057 // the same memory slice between InitializeNode and store.
5058 failed = true;
5059 break;
5060 }
5061 assert(base->is_AddP(), "should be addp but is %s", base->Name());
5062 base = base->in(AddPNode::Base);
5063 if (base != nullptr) {
5064 base = base->uncast();
5065 if (base->is_Proj() && base->in(0) == alloc) {
5066 failed = true;
5067 break;
5068 }
5069 }
5070 }
5071 }
5072 }
5073 } else {
5074 failed = true;
5075 break;
5076 }
5077 }
5078 }
5079 }
5080 if (failed) {
5081 if (!can_reshape) {
5082 // We decided we couldn't capture the store during parsing. We
5083 // should try again during the next IGVN once the graph is
5084 // cleaner.
5085 phase->C->record_for_igvn(st);
5086 }
5087 return FAIL;
5088 }
5089
5090 return offset; // success
5091 }
5092
5093 // Find the captured store in(i) which corresponds to the range
5094 // [start..start+size) in the initialized object.
5095 // If there is one, return its index i. If there isn't, return the
5096 // negative of the index where it should be inserted.
5097 // Return 0 if the queried range overlaps an initialization boundary
5098 // or if dead code is encountered.
5099 // If size_in_bytes is zero, do not bother with overlap checks.
5100 int InitializeNode::captured_store_insertion_point(intptr_t start,
5101 int size_in_bytes,
5102 PhaseValues* phase) {
5103 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
5104
5105 if (is_complete())
5106 return FAIL; // arraycopy got here first; punt
5107
5108 assert(allocation() != nullptr, "must be present");
5109
5110 // no negatives, no header fields:
5111 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
5112
5113 // after a certain size, we bail out on tracking all the stores:
5114 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5115 if (start >= ti_limit) return FAIL;
5116
5117 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
5118 if (i >= limit) return -(int)i; // not found; here is where to put it
5119
5120 Node* st = in(i);
5121 intptr_t st_off = get_store_offset(st, phase);
5122 if (st_off < 0) {
5123 if (st != zero_memory()) {
5124 return FAIL; // bail out if there is dead garbage
5125 }
5126 } else if (st_off > start) {
5127 // ...we are done, since stores are ordered
5128 if (st_off < start + size_in_bytes) {
5129 return FAIL; // the next store overlaps
5130 }
5131 return -(int)i; // not found; here is where to put it
5132 } else if (st_off < start) {
5133 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
5134 if (size_in_bytes != 0 &&
5135 start < st_off + MAX_STORE &&
5136 start < st_off + st->as_Store()->memory_size()) {
5137 return FAIL; // the previous store overlaps
5138 }
5139 } else {
5140 if (size_in_bytes != 0 &&
5141 st->as_Store()->memory_size() != size_in_bytes) {
5142 return FAIL; // mismatched store size
5143 }
5144 return i;
5145 }
5146
5147 ++i;
5148 }
5149 }
5150
5151 // Look for a captured store which initializes at the offset 'start'
5152 // with the given size. If there is no such store, and no other
5153 // initialization interferes, then return zero_memory (the memory
5154 // projection of the AllocateNode).
5155 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5156 PhaseValues* phase) {
5157 assert(stores_are_sane(phase), "");
5158 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5159 if (i == 0) {
5160 return nullptr; // something is dead
5161 } else if (i < 0) {
5162 return zero_memory(); // just primordial zero bits here
5163 } else {
5164 Node* st = in(i); // here is the store at this position
5165 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5166 return st;
5167 }
5168 }
5169
5170 // Create, as a raw pointer, an address within my new object at 'offset'.
5171 Node* InitializeNode::make_raw_address(intptr_t offset,
5172 PhaseGVN* phase) {
5173 Node* addr = in(RawAddress);
5174 if (offset != 0) {
5175 Compile* C = phase->C;
5176 addr = phase->transform( new AddPNode(C->top(), addr,
5177 phase->MakeConX(offset)) );
5178 }
5179 return addr;
5180 }
5181
5182 // Clone the given store, converting it into a raw store
5183 // initializing a field or element of my new object.
5184 // Caller is responsible for retiring the original store,
5185 // with subsume_node or the like.
5186 //
5187 // From the example above InitializeNode::InitializeNode,
5188 // here are the old stores to be captured:
5189 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5190 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5191 //
5192 // Here is the changed code; note the extra edges on init:
5193 // alloc = (Allocate ...)
5194 // rawoop = alloc.RawAddress
5195 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5196 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5197 // init = (Initialize alloc.Control alloc.Memory rawoop
5198 // rawstore1 rawstore2)
5199 //
5200 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5201 PhaseGVN* phase, bool can_reshape) {
5202 assert(stores_are_sane(phase), "");
5203
5204 if (start < 0) return nullptr;
5205 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5206
5207 Compile* C = phase->C;
5208 int size_in_bytes = st->memory_size();
5209 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5210 if (i == 0) return nullptr; // bail out
5211 Node* prev_mem = nullptr; // raw memory for the captured store
5212 if (i > 0) {
5213 prev_mem = in(i); // there is a pre-existing store under this one
5214 set_req(i, C->top()); // temporarily disconnect it
5215 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5216 } else {
5217 i = -i; // no pre-existing store
5218 prev_mem = zero_memory(); // a slice of the newly allocated object
5219 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5220 set_req(--i, C->top()); // reuse this edge; it has been folded away
5221 else
5222 ins_req(i, C->top()); // build a new edge
5223 }
5224 Node* new_st = st->clone();
5225 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5226 new_st->set_req(MemNode::Control, in(Control));
5227 new_st->set_req(MemNode::Memory, prev_mem);
5228 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5229 bs->eliminate_gc_barrier_data(new_st);
5230 new_st = phase->transform(new_st);
5231
5232 // At this point, new_st might have swallowed a pre-existing store
5233 // at the same offset, or perhaps new_st might have disappeared,
5234 // if it redundantly stored the same value (or zero to fresh memory).
5235
5236 // In any case, wire it in:
5237 PhaseIterGVN* igvn = phase->is_IterGVN();
5238 if (igvn) {
5239 igvn->rehash_node_delayed(this);
5240 }
5241 set_req(i, new_st);
5242
5243 // The caller may now kill the old guy.
5244 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5245 assert(check_st == new_st || check_st == nullptr, "must be findable");
5246 assert(!is_complete(), "");
5247 return new_st;
5248 }
5249
5250 static bool store_constant(jlong* tiles, int num_tiles,
5251 intptr_t st_off, int st_size,
5252 jlong con) {
5253 if ((st_off & (st_size-1)) != 0)
5254 return false; // strange store offset (assume size==2**N)
5255 address addr = (address)tiles + st_off;
5256 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5257 switch (st_size) {
5258 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5259 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5260 case sizeof(jint): *(jint*) addr = (jint) con; break;
5261 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5262 default: return false; // strange store size (detect size!=2**N here)
5263 }
5264 return true; // return success to caller
5265 }
5266
5267 // Coalesce subword constants into int constants and possibly
5268 // into long constants. The goal, if the CPU permits,
5269 // is to initialize the object with a small number of 64-bit tiles.
5270 // Also, convert floating-point constants to bit patterns.
5271 // Non-constants are not relevant to this pass.
5272 //
5273 // In terms of the running example on InitializeNode::InitializeNode
5274 // and InitializeNode::capture_store, here is the transformation
5275 // of rawstore1 and rawstore2 into rawstore12:
5276 // alloc = (Allocate ...)
5277 // rawoop = alloc.RawAddress
5278 // tile12 = 0x00010002
5279 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5280 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5281 //
5282 void
5283 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5284 Node* size_in_bytes,
5285 PhaseGVN* phase) {
5286 Compile* C = phase->C;
5287
5288 assert(stores_are_sane(phase), "");
5289 // Note: After this pass, they are not completely sane,
5290 // since there may be some overlaps.
5291
5292 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5293
5294 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5295 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5296 size_limit = MIN2(size_limit, ti_limit);
5297 size_limit = align_up(size_limit, BytesPerLong);
5298 int num_tiles = size_limit / BytesPerLong;
5299
5300 // allocate space for the tile map:
5301 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5302 jlong tiles_buf[small_len];
5303 Node* nodes_buf[small_len];
5304 jlong inits_buf[small_len];
5305 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5306 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5307 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5308 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5309 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5310 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5311 // tiles: exact bitwise model of all primitive constants
5312 // nodes: last constant-storing node subsumed into the tiles model
5313 // inits: which bytes (in each tile) are touched by any initializations
5314
5315 //// Pass A: Fill in the tile model with any relevant stores.
5316
5317 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5318 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5319 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5320 Node* zmem = zero_memory(); // initially zero memory state
5321 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5322 Node* st = in(i);
5323 intptr_t st_off = get_store_offset(st, phase);
5324
5325 // Figure out the store's offset and constant value:
5326 if (st_off < header_size) continue; //skip (ignore header)
5327 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5328 int st_size = st->as_Store()->memory_size();
5329 if (st_off + st_size > size_limit) break;
5330
5331 // Record which bytes are touched, whether by constant or not.
5332 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5333 continue; // skip (strange store size)
5334
5335 const Type* val = phase->type(st->in(MemNode::ValueIn));
5336 if (!val->singleton()) continue; //skip (non-con store)
5337 BasicType type = val->basic_type();
5338
5339 jlong con = 0;
5340 switch (type) {
5341 case T_INT: con = val->is_int()->get_con(); break;
5342 case T_LONG: con = val->is_long()->get_con(); break;
5343 case T_FLOAT: con = jint_cast(val->getf()); break;
5344 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5345 default: continue; //skip (odd store type)
5346 }
5347
5348 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5349 st->Opcode() == Op_StoreL) {
5350 continue; // This StoreL is already optimal.
5351 }
5352
5353 // Store down the constant.
5354 store_constant(tiles, num_tiles, st_off, st_size, con);
5355
5356 intptr_t j = st_off >> LogBytesPerLong;
5357
5358 if (type == T_INT && st_size == BytesPerInt
5359 && (st_off & BytesPerInt) == BytesPerInt) {
5360 jlong lcon = tiles[j];
5361 if (!Matcher::isSimpleConstant64(lcon) &&
5362 st->Opcode() == Op_StoreI) {
5363 // This StoreI is already optimal by itself.
5364 jint* intcon = (jint*) &tiles[j];
5365 intcon[1] = 0; // undo the store_constant()
5366
5367 // If the previous store is also optimal by itself, back up and
5368 // undo the action of the previous loop iteration... if we can.
5369 // But if we can't, just let the previous half take care of itself.
5370 st = nodes[j];
5371 st_off -= BytesPerInt;
5372 con = intcon[0];
5373 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5374 assert(st_off >= header_size, "still ignoring header");
5375 assert(get_store_offset(st, phase) == st_off, "must be");
5376 assert(in(i-1) == zmem, "must be");
5377 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5378 assert(con == tcon->is_int()->get_con(), "must be");
5379 // Undo the effects of the previous loop trip, which swallowed st:
5380 intcon[0] = 0; // undo store_constant()
5381 set_req(i-1, st); // undo set_req(i, zmem)
5382 nodes[j] = nullptr; // undo nodes[j] = st
5383 --old_subword; // undo ++old_subword
5384 }
5385 continue; // This StoreI is already optimal.
5386 }
5387 }
5388
5389 // This store is not needed.
5390 set_req(i, zmem);
5391 nodes[j] = st; // record for the moment
5392 if (st_size < BytesPerLong) // something has changed
5393 ++old_subword; // includes int/float, but who's counting...
5394 else ++old_long;
5395 }
5396
5397 if ((old_subword + old_long) == 0)
5398 return; // nothing more to do
5399
5400 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5401 // Be sure to insert them before overlapping non-constant stores.
5402 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5403 for (int j = 0; j < num_tiles; j++) {
5404 jlong con = tiles[j];
5405 jlong init = inits[j];
5406 if (con == 0) continue;
5407 jint con0, con1; // split the constant, address-wise
5408 jint init0, init1; // split the init map, address-wise
5409 { union { jlong con; jint intcon[2]; } u;
5410 u.con = con;
5411 con0 = u.intcon[0];
5412 con1 = u.intcon[1];
5413 u.con = init;
5414 init0 = u.intcon[0];
5415 init1 = u.intcon[1];
5416 }
5417
5418 Node* old = nodes[j];
5419 assert(old != nullptr, "need the prior store");
5420 intptr_t offset = (j * BytesPerLong);
5421
5422 bool split = !Matcher::isSimpleConstant64(con);
5423
5424 if (offset < header_size) {
5425 assert(offset + BytesPerInt >= header_size, "second int counts");
5426 assert(*(jint*)&tiles[j] == 0, "junk in header");
5427 split = true; // only the second word counts
5428 // Example: int a[] = { 42 ... }
5429 } else if (con0 == 0 && init0 == -1) {
5430 split = true; // first word is covered by full inits
5431 // Example: int a[] = { ... foo(), 42 ... }
5432 } else if (con1 == 0 && init1 == -1) {
5433 split = true; // second word is covered by full inits
5434 // Example: int a[] = { ... 42, foo() ... }
5435 }
5436
5437 // Here's a case where init0 is neither 0 nor -1:
5438 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5439 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5440 // In this case the tile is not split; it is (jlong)42.
5441 // The big tile is stored down, and then the foo() value is inserted.
5442 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5443
5444 Node* ctl = old->in(MemNode::Control);
5445 Node* adr = make_raw_address(offset, phase);
5446 const TypePtr* atp = TypeRawPtr::BOTTOM;
5447
5448 // One or two coalesced stores to plop down.
5449 Node* st[2];
5450 intptr_t off[2];
5451 int nst = 0;
5452 if (!split) {
5453 ++new_long;
5454 off[nst] = offset;
5455 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5456 phase->longcon(con), T_LONG, MemNode::unordered);
5457 } else {
5458 // Omit either if it is a zero.
5459 if (con0 != 0) {
5460 ++new_int;
5461 off[nst] = offset;
5462 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5463 phase->intcon(con0), T_INT, MemNode::unordered);
5464 }
5465 if (con1 != 0) {
5466 ++new_int;
5467 offset += BytesPerInt;
5468 adr = make_raw_address(offset, phase);
5469 off[nst] = offset;
5470 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5471 phase->intcon(con1), T_INT, MemNode::unordered);
5472 }
5473 }
5474
5475 // Insert second store first, then the first before the second.
5476 // Insert each one just before any overlapping non-constant stores.
5477 while (nst > 0) {
5478 Node* st1 = st[--nst];
5479 C->copy_node_notes_to(st1, old);
5480 st1 = phase->transform(st1);
5481 offset = off[nst];
5482 assert(offset >= header_size, "do not smash header");
5483 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5484 guarantee(ins_idx != 0, "must re-insert constant store");
5485 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5486 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5487 set_req(--ins_idx, st1);
5488 else
5489 ins_req(ins_idx, st1);
5490 }
5491 }
5492
5493 if (PrintCompilation && WizardMode)
5494 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5495 old_subword, old_long, new_int, new_long);
5496 if (C->log() != nullptr)
5497 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5498 old_subword, old_long, new_int, new_long);
5499
5500 // Clean up any remaining occurrences of zmem:
5501 remove_extra_zeroes();
5502 }
5503
5504 // Explore forward from in(start) to find the first fully initialized
5505 // word, and return its offset. Skip groups of subword stores which
5506 // together initialize full words. If in(start) is itself part of a
5507 // fully initialized word, return the offset of in(start). If there
5508 // are no following full-word stores, or if something is fishy, return
5509 // a negative value.
5510 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5511 int int_map = 0;
5512 intptr_t int_map_off = 0;
5513 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5514
5515 for (uint i = start, limit = req(); i < limit; i++) {
5516 Node* st = in(i);
5517
5518 intptr_t st_off = get_store_offset(st, phase);
5519 if (st_off < 0) break; // return conservative answer
5520
5521 int st_size = st->as_Store()->memory_size();
5522 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5523 return st_off; // we found a complete word init
5524 }
5525
5526 // update the map:
5527
5528 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5529 if (this_int_off != int_map_off) {
5530 // reset the map:
5531 int_map = 0;
5532 int_map_off = this_int_off;
5533 }
5534
5535 int subword_off = st_off - this_int_off;
5536 int_map |= right_n_bits(st_size) << subword_off;
5537 if ((int_map & FULL_MAP) == FULL_MAP) {
5538 return this_int_off; // we found a complete word init
5539 }
5540
5541 // Did this store hit or cross the word boundary?
5542 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5543 if (next_int_off == this_int_off + BytesPerInt) {
5544 // We passed the current int, without fully initializing it.
5545 int_map_off = next_int_off;
5546 int_map >>= BytesPerInt;
5547 } else if (next_int_off > this_int_off + BytesPerInt) {
5548 // We passed the current and next int.
5549 return this_int_off + BytesPerInt;
5550 }
5551 }
5552
5553 return -1;
5554 }
5555
5556
5557 // Called when the associated AllocateNode is expanded into CFG.
5558 // At this point, we may perform additional optimizations.
5559 // Linearize the stores by ascending offset, to make memory
5560 // activity as coherent as possible.
5561 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5562 intptr_t header_size,
5563 Node* size_in_bytes,
5564 PhaseIterGVN* phase) {
5565 assert(!is_complete(), "not already complete");
5566 assert(stores_are_sane(phase), "");
5567 assert(allocation() != nullptr, "must be present");
5568
5569 remove_extra_zeroes();
5570
5571 if (ReduceFieldZeroing || ReduceBulkZeroing)
5572 // reduce instruction count for common initialization patterns
5573 coalesce_subword_stores(header_size, size_in_bytes, phase);
5574
5575 Node* zmem = zero_memory(); // initially zero memory state
5576 Node* inits = zmem; // accumulating a linearized chain of inits
5577 #ifdef ASSERT
5578 intptr_t first_offset = allocation()->minimum_header_size();
5579 intptr_t last_init_off = first_offset; // previous init offset
5580 intptr_t last_init_end = first_offset; // previous init offset+size
5581 intptr_t last_tile_end = first_offset; // previous tile offset+size
5582 #endif
5583 intptr_t zeroes_done = header_size;
5584
5585 bool do_zeroing = true; // we might give up if inits are very sparse
5586 int big_init_gaps = 0; // how many large gaps have we seen?
5587
5588 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5589 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5590
5591 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5592 Node* st = in(i);
5593 intptr_t st_off = get_store_offset(st, phase);
5594 if (st_off < 0)
5595 break; // unknown junk in the inits
5596 if (st->in(MemNode::Memory) != zmem)
5597 break; // complicated store chains somehow in list
5598
5599 int st_size = st->as_Store()->memory_size();
5600 intptr_t next_init_off = st_off + st_size;
5601
5602 if (do_zeroing && zeroes_done < next_init_off) {
5603 // See if this store needs a zero before it or under it.
5604 intptr_t zeroes_needed = st_off;
5605
5606 if (st_size < BytesPerInt) {
5607 // Look for subword stores which only partially initialize words.
5608 // If we find some, we must lay down some word-level zeroes first,
5609 // underneath the subword stores.
5610 //
5611 // Examples:
5612 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5613 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5614 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5615 //
5616 // Note: coalesce_subword_stores may have already done this,
5617 // if it was prompted by constant non-zero subword initializers.
5618 // But this case can still arise with non-constant stores.
5619
5620 intptr_t next_full_store = find_next_fullword_store(i, phase);
5621
5622 // In the examples above:
5623 // in(i) p q r s x y z
5624 // st_off 12 13 14 15 12 13 14
5625 // st_size 1 1 1 1 1 1 1
5626 // next_full_s. 12 16 16 16 16 16 16
5627 // z's_done 12 16 16 16 12 16 12
5628 // z's_needed 12 16 16 16 16 16 16
5629 // zsize 0 0 0 0 4 0 4
5630 if (next_full_store < 0) {
5631 // Conservative tack: Zero to end of current word.
5632 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5633 } else {
5634 // Zero to beginning of next fully initialized word.
5635 // Or, don't zero at all, if we are already in that word.
5636 assert(next_full_store >= zeroes_needed, "must go forward");
5637 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5638 zeroes_needed = next_full_store;
5639 }
5640 }
5641
5642 if (zeroes_needed > zeroes_done) {
5643 intptr_t zsize = zeroes_needed - zeroes_done;
5644 // Do some incremental zeroing on rawmem, in parallel with inits.
5645 zeroes_done = align_down(zeroes_done, BytesPerInt);
5646 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5647 allocation()->in(AllocateNode::InitValue),
5648 allocation()->in(AllocateNode::RawInitValue),
5649 zeroes_done, zeroes_needed,
5650 phase);
5651 zeroes_done = zeroes_needed;
5652 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5653 do_zeroing = false; // leave the hole, next time
5654 }
5655 }
5656
5657 // Collect the store and move on:
5658 phase->replace_input_of(st, MemNode::Memory, inits);
5659 inits = st; // put it on the linearized chain
5660 set_req(i, zmem); // unhook from previous position
5661
5662 if (zeroes_done == st_off)
5663 zeroes_done = next_init_off;
5664
5665 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5666
5667 #ifdef ASSERT
5668 // Various order invariants. Weaker than stores_are_sane because
5669 // a large constant tile can be filled in by smaller non-constant stores.
5670 assert(st_off >= last_init_off, "inits do not reverse");
5671 last_init_off = st_off;
5672 const Type* val = nullptr;
5673 if (st_size >= BytesPerInt &&
5674 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5675 (int)val->basic_type() < (int)T_OBJECT) {
5676 assert(st_off >= last_tile_end, "tiles do not overlap");
5677 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5678 last_tile_end = MAX2(last_tile_end, next_init_off);
5679 } else {
5680 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5681 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5682 assert(st_off >= last_init_end, "inits do not overlap");
5683 last_init_end = next_init_off; // it's a non-tile
5684 }
5685 #endif //ASSERT
5686 }
5687
5688 remove_extra_zeroes(); // clear out all the zmems left over
5689 add_req(inits);
5690
5691 if (!(UseTLAB && ZeroTLAB)) {
5692 // If anything remains to be zeroed, zero it all now.
5693 zeroes_done = align_down(zeroes_done, BytesPerInt);
5694 // if it is the last unused 4 bytes of an instance, forget about it
5695 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5696 if (zeroes_done + BytesPerLong >= size_limit) {
5697 AllocateNode* alloc = allocation();
5698 assert(alloc != nullptr, "must be present");
5699 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5700 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5701 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5702 if (zeroes_done == k->layout_helper())
5703 zeroes_done = size_limit;
5704 }
5705 }
5706 if (zeroes_done < size_limit) {
5707 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5708 allocation()->in(AllocateNode::InitValue),
5709 allocation()->in(AllocateNode::RawInitValue),
5710 zeroes_done, size_in_bytes, phase);
5711 }
5712 }
5713
5714 set_complete(phase);
5715 return rawmem;
5716 }
5717
5718
5719 #ifdef ASSERT
5720 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5721 if (is_complete())
5722 return true; // stores could be anything at this point
5723 assert(allocation() != nullptr, "must be present");
5724 intptr_t last_off = allocation()->minimum_header_size();
5725 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5726 Node* st = in(i);
5727 intptr_t st_off = get_store_offset(st, phase);
5728 if (st_off < 0) continue; // ignore dead garbage
5729 if (last_off > st_off) {
5730 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5731 this->dump(2);
5732 assert(false, "ascending store offsets");
5733 return false;
5734 }
5735 last_off = st_off + st->as_Store()->memory_size();
5736 }
5737 return true;
5738 }
5739 #endif //ASSERT
5740
5741
5742
5743
5744 //============================MergeMemNode=====================================
5745 //
5746 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5747 // contributing store or call operations. Each contributor provides the memory
5748 // state for a particular "alias type" (see Compile::alias_type). For example,
5749 // if a MergeMem has an input X for alias category #6, then any memory reference
5750 // to alias category #6 may use X as its memory state input, as an exact equivalent
5751 // to using the MergeMem as a whole.
5752 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5753 //
5754 // (Here, the <N> notation gives the index of the relevant adr_type.)
5755 //
5756 // In one special case (and more cases in the future), alias categories overlap.
5757 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5758 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5759 // it is exactly equivalent to that state W:
5760 // MergeMem(<Bot>: W) <==> W
5761 //
5762 // Usually, the merge has more than one input. In that case, where inputs
5763 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5764 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5765 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5766 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5767 //
5768 // A merge can take a "wide" memory state as one of its narrow inputs.
5769 // This simply means that the merge observes out only the relevant parts of
5770 // the wide input. That is, wide memory states arriving at narrow merge inputs
5771 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5772 //
5773 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5774 // and that memory slices "leak through":
5775 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5776 //
5777 // But, in such a cascade, repeated memory slices can "block the leak":
5778 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5779 //
5780 // In the last example, Y is not part of the combined memory state of the
5781 // outermost MergeMem. The system must, of course, prevent unschedulable
5782 // memory states from arising, so you can be sure that the state Y is somehow
5783 // a precursor to state Y'.
5784 //
5785 //
5786 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5787 // of each MergeMemNode array are exactly the numerical alias indexes, including
5788 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5789 // Compile::alias_type (and kin) produce and manage these indexes.
5790 //
5791 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5792 // (Note that this provides quick access to the top node inside MergeMem methods,
5793 // without the need to reach out via TLS to Compile::current.)
5794 //
5795 // As a consequence of what was just described, a MergeMem that represents a full
5796 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5797 // containing all alias categories.
5798 //
5799 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5800 //
5801 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5802 // a memory state for the alias type <N>, or else the top node, meaning that
5803 // there is no particular input for that alias type. Note that the length of
5804 // a MergeMem is variable, and may be extended at any time to accommodate new
5805 // memory states at larger alias indexes. When merges grow, they are of course
5806 // filled with "top" in the unused in() positions.
5807 //
5808 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5809 // (Top was chosen because it works smoothly with passes like GCM.)
5810 //
5811 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5812 // the type of random VM bits like TLS references.) Since it is always the
5813 // first non-Bot memory slice, some low-level loops use it to initialize an
5814 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5815 //
5816 //
5817 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5818 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5819 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5820 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5821 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5822 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5823 //
5824 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5825 // really that different from the other memory inputs. An abbreviation called
5826 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5827 //
5828 //
5829 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5830 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5831 // that "emerges though" the base memory will be marked as excluding the alias types
5832 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5833 //
5834 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5835 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5836 //
5837 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5838 // (It is currently unimplemented.) As you can see, the resulting merge is
5839 // actually a disjoint union of memory states, rather than an overlay.
5840 //
5841
5842 //------------------------------MergeMemNode-----------------------------------
5843 Node* MergeMemNode::make_empty_memory() {
5844 Node* empty_memory = (Node*) Compile::current()->top();
5845 assert(empty_memory->is_top(), "correct sentinel identity");
5846 return empty_memory;
5847 }
5848
5849 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5850 init_class_id(Class_MergeMem);
5851 // all inputs are nullified in Node::Node(int)
5852 // set_input(0, nullptr); // no control input
5853
5854 // Initialize the edges uniformly to top, for starters.
5855 Node* empty_mem = make_empty_memory();
5856 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5857 init_req(i,empty_mem);
5858 }
5859 assert(empty_memory() == empty_mem, "");
5860
5861 if( new_base != nullptr && new_base->is_MergeMem() ) {
5862 MergeMemNode* mdef = new_base->as_MergeMem();
5863 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5864 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5865 mms.set_memory(mms.memory2());
5866 }
5867 assert(base_memory() == mdef->base_memory(), "");
5868 } else {
5869 set_base_memory(new_base);
5870 }
5871 }
5872
5873 // Make a new, untransformed MergeMem with the same base as 'mem'.
5874 // If mem is itself a MergeMem, populate the result with the same edges.
5875 MergeMemNode* MergeMemNode::make(Node* mem) {
5876 return new MergeMemNode(mem);
5877 }
5878
5879 //------------------------------cmp--------------------------------------------
5880 uint MergeMemNode::hash() const { return NO_HASH; }
5881 bool MergeMemNode::cmp( const Node &n ) const {
5882 return (&n == this); // Always fail except on self
5883 }
5884
5885 //------------------------------Identity---------------------------------------
5886 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5887 // Identity if this merge point does not record any interesting memory
5888 // disambiguations.
5889 Node* base_mem = base_memory();
5890 Node* empty_mem = empty_memory();
5891 if (base_mem != empty_mem) { // Memory path is not dead?
5892 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5893 Node* mem = in(i);
5894 if (mem != empty_mem && mem != base_mem) {
5895 return this; // Many memory splits; no change
5896 }
5897 }
5898 }
5899 return base_mem; // No memory splits; ID on the one true input
5900 }
5901
5902 //------------------------------Ideal------------------------------------------
5903 // This method is invoked recursively on chains of MergeMem nodes
5904 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5905 // Remove chain'd MergeMems
5906 //
5907 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5908 // relative to the "in(Bot)". Since we are patching both at the same time,
5909 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5910 // but rewrite each "in(i)" relative to the new "in(Bot)".
5911 Node *progress = nullptr;
5912
5913
5914 Node* old_base = base_memory();
5915 Node* empty_mem = empty_memory();
5916 if (old_base == empty_mem)
5917 return nullptr; // Dead memory path.
5918
5919 MergeMemNode* old_mbase;
5920 if (old_base != nullptr && old_base->is_MergeMem())
5921 old_mbase = old_base->as_MergeMem();
5922 else
5923 old_mbase = nullptr;
5924 Node* new_base = old_base;
5925
5926 // simplify stacked MergeMems in base memory
5927 if (old_mbase) new_base = old_mbase->base_memory();
5928
5929 // the base memory might contribute new slices beyond my req()
5930 if (old_mbase) grow_to_match(old_mbase);
5931
5932 // Note: We do not call verify_sparse on entry, because inputs
5933 // can normalize to the base_memory via subsume_node or similar
5934 // mechanisms. This method repairs that damage.
5935
5936 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5937
5938 // Look at each slice.
5939 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5940 Node* old_in = in(i);
5941 // calculate the old memory value
5942 Node* old_mem = old_in;
5943 if (old_mem == empty_mem) old_mem = old_base;
5944 assert(old_mem == memory_at(i), "");
5945
5946 // maybe update (reslice) the old memory value
5947
5948 // simplify stacked MergeMems
5949 Node* new_mem = old_mem;
5950 MergeMemNode* old_mmem;
5951 if (old_mem != nullptr && old_mem->is_MergeMem())
5952 old_mmem = old_mem->as_MergeMem();
5953 else
5954 old_mmem = nullptr;
5955 if (old_mmem == this) {
5956 // This can happen if loops break up and safepoints disappear.
5957 // A merge of BotPtr (default) with a RawPtr memory derived from a
5958 // safepoint can be rewritten to a merge of the same BotPtr with
5959 // the BotPtr phi coming into the loop. If that phi disappears
5960 // also, we can end up with a self-loop of the mergemem.
5961 // In general, if loops degenerate and memory effects disappear,
5962 // a mergemem can be left looking at itself. This simply means
5963 // that the mergemem's default should be used, since there is
5964 // no longer any apparent effect on this slice.
5965 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5966 // from start. Update the input to TOP.
5967 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5968 }
5969 else if (old_mmem != nullptr) {
5970 new_mem = old_mmem->memory_at(i);
5971 }
5972 // else preceding memory was not a MergeMem
5973
5974 // maybe store down a new value
5975 Node* new_in = new_mem;
5976 if (new_in == new_base) new_in = empty_mem;
5977
5978 if (new_in != old_in) {
5979 // Warning: Do not combine this "if" with the previous "if"
5980 // A memory slice might have be be rewritten even if it is semantically
5981 // unchanged, if the base_memory value has changed.
5982 set_req_X(i, new_in, phase);
5983 progress = this; // Report progress
5984 }
5985 }
5986
5987 if (new_base != old_base) {
5988 set_req_X(Compile::AliasIdxBot, new_base, phase);
5989 // Don't use set_base_memory(new_base), because we need to update du.
5990 assert(base_memory() == new_base, "");
5991 progress = this;
5992 }
5993
5994 if( base_memory() == this ) {
5995 // a self cycle indicates this memory path is dead
5996 set_req(Compile::AliasIdxBot, empty_mem);
5997 }
5998
5999 // Resolve external cycles by calling Ideal on a MergeMem base_memory
6000 // Recursion must occur after the self cycle check above
6001 if( base_memory()->is_MergeMem() ) {
6002 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
6003 Node *m = phase->transform(new_mbase); // Rollup any cycles
6004 if( m != nullptr &&
6005 (m->is_top() ||
6006 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
6007 // propagate rollup of dead cycle to self
6008 set_req(Compile::AliasIdxBot, empty_mem);
6009 }
6010 }
6011
6012 if( base_memory() == empty_mem ) {
6013 progress = this;
6014 // Cut inputs during Parse phase only.
6015 // During Optimize phase a dead MergeMem node will be subsumed by Top.
6016 if( !can_reshape ) {
6017 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6018 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
6019 }
6020 }
6021 }
6022
6023 if( !progress && base_memory()->is_Phi() && can_reshape ) {
6024 // Check if PhiNode::Ideal's "Split phis through memory merges"
6025 // transform should be attempted. Look for this->phi->this cycle.
6026 uint merge_width = req();
6027 if (merge_width > Compile::AliasIdxRaw) {
6028 PhiNode* phi = base_memory()->as_Phi();
6029 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
6030 if (phi->in(i) == this) {
6031 phase->is_IterGVN()->_worklist.push(phi);
6032 break;
6033 }
6034 }
6035 }
6036 }
6037
6038 assert(progress || verify_sparse(), "please, no dups of base");
6039 return progress;
6040 }
6041
6042 //-------------------------set_base_memory-------------------------------------
6043 void MergeMemNode::set_base_memory(Node *new_base) {
6044 Node* empty_mem = empty_memory();
6045 set_req(Compile::AliasIdxBot, new_base);
6046 assert(memory_at(req()) == new_base, "must set default memory");
6047 // Clear out other occurrences of new_base:
6048 if (new_base != empty_mem) {
6049 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6050 if (in(i) == new_base) set_req(i, empty_mem);
6051 }
6052 }
6053 }
6054
6055 //------------------------------out_RegMask------------------------------------
6056 const RegMask &MergeMemNode::out_RegMask() const {
6057 return RegMask::Empty;
6058 }
6059
6060 //------------------------------dump_spec--------------------------------------
6061 #ifndef PRODUCT
6062 void MergeMemNode::dump_spec(outputStream *st) const {
6063 st->print(" {");
6064 Node* base_mem = base_memory();
6065 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
6066 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
6067 if (mem == base_mem) { st->print(" -"); continue; }
6068 st->print( " N%d:", mem->_idx );
6069 Compile::current()->get_adr_type(i)->dump_on(st);
6070 }
6071 st->print(" }");
6072 }
6073 #endif // !PRODUCT
6074
6075
6076 #ifdef ASSERT
6077 static bool might_be_same(Node* a, Node* b) {
6078 if (a == b) return true;
6079 if (!(a->is_Phi() || b->is_Phi())) return false;
6080 // phis shift around during optimization
6081 return true; // pretty stupid...
6082 }
6083
6084 // verify a narrow slice (either incoming or outgoing)
6085 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
6086 if (!VerifyAliases) return; // don't bother to verify unless requested
6087 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
6088 if (Node::in_dump()) return; // muzzle asserts when printing
6089 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
6090 assert(n != nullptr, "");
6091 // Elide intervening MergeMem's
6092 while (n->is_MergeMem()) {
6093 n = n->as_MergeMem()->memory_at(alias_idx);
6094 }
6095 Compile* C = Compile::current();
6096 const TypePtr* n_adr_type = n->adr_type();
6097 if (n == m->empty_memory()) {
6098 // Implicit copy of base_memory()
6099 } else if (n_adr_type != TypePtr::BOTTOM) {
6100 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6101 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6102 } else {
6103 // A few places like make_runtime_call "know" that VM calls are narrow,
6104 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6105 bool expected_wide_mem = false;
6106 if (n == m->base_memory()) {
6107 expected_wide_mem = true;
6108 } else if (alias_idx == Compile::AliasIdxRaw ||
6109 n == m->memory_at(Compile::AliasIdxRaw)) {
6110 expected_wide_mem = true;
6111 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6112 // memory can "leak through" calls on channels that
6113 // are write-once. Allow this also.
6114 expected_wide_mem = true;
6115 }
6116 assert(expected_wide_mem, "expected narrow slice replacement");
6117 }
6118 }
6119 #else // !ASSERT
6120 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
6121 #endif
6122
6123
6124 //-----------------------------memory_at---------------------------------------
6125 Node* MergeMemNode::memory_at(uint alias_idx) const {
6126 assert(alias_idx >= Compile::AliasIdxRaw ||
6127 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6128 "must avoid base_memory and AliasIdxTop");
6129
6130 // Otherwise, it is a narrow slice.
6131 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6132 if (is_empty_memory(n)) {
6133 // the array is sparse; empty slots are the "top" node
6134 n = base_memory();
6135 assert(Node::in_dump()
6136 || n == nullptr || n->bottom_type() == Type::TOP
6137 || n->adr_type() == nullptr // address is TOP
6138 || n->adr_type() == TypePtr::BOTTOM
6139 || n->adr_type() == TypeRawPtr::BOTTOM
6140 || !Compile::current()->do_aliasing(),
6141 "must be a wide memory");
6142 // do_aliasing == false if we are organizing the memory states manually.
6143 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6144 } else {
6145 // make sure the stored slice is sane
6146 #ifdef ASSERT
6147 if (VMError::is_error_reported() || Node::in_dump()) {
6148 } else if (might_be_same(n, base_memory())) {
6149 // Give it a pass: It is a mostly harmless repetition of the base.
6150 // This can arise normally from node subsumption during optimization.
6151 } else {
6152 verify_memory_slice(this, alias_idx, n);
6153 }
6154 #endif
6155 }
6156 return n;
6157 }
6158
6159 //---------------------------set_memory_at-------------------------------------
6160 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6161 verify_memory_slice(this, alias_idx, n);
6162 Node* empty_mem = empty_memory();
6163 if (n == base_memory()) n = empty_mem; // collapse default
6164 uint need_req = alias_idx+1;
6165 if (req() < need_req) {
6166 if (n == empty_mem) return; // already the default, so do not grow me
6167 // grow the sparse array
6168 do {
6169 add_req(empty_mem);
6170 } while (req() < need_req);
6171 }
6172 set_req( alias_idx, n );
6173 }
6174
6175
6176
6177 //--------------------------iteration_setup------------------------------------
6178 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6179 if (other != nullptr) {
6180 grow_to_match(other);
6181 // invariant: the finite support of mm2 is within mm->req()
6182 #ifdef ASSERT
6183 for (uint i = req(); i < other->req(); i++) {
6184 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6185 }
6186 #endif
6187 }
6188 // Replace spurious copies of base_memory by top.
6189 Node* base_mem = base_memory();
6190 if (base_mem != nullptr && !base_mem->is_top()) {
6191 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6192 if (in(i) == base_mem)
6193 set_req(i, empty_memory());
6194 }
6195 }
6196 }
6197
6198 //---------------------------grow_to_match-------------------------------------
6199 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6200 Node* empty_mem = empty_memory();
6201 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6202 // look for the finite support of the other memory
6203 for (uint i = other->req(); --i >= req(); ) {
6204 if (other->in(i) != empty_mem) {
6205 uint new_len = i+1;
6206 while (req() < new_len) add_req(empty_mem);
6207 break;
6208 }
6209 }
6210 }
6211
6212 //---------------------------verify_sparse-------------------------------------
6213 #ifndef PRODUCT
6214 bool MergeMemNode::verify_sparse() const {
6215 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6216 Node* base_mem = base_memory();
6217 // The following can happen in degenerate cases, since empty==top.
6218 if (is_empty_memory(base_mem)) return true;
6219 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6220 assert(in(i) != nullptr, "sane slice");
6221 if (in(i) == base_mem) return false; // should have been the sentinel value!
6222 }
6223 return true;
6224 }
6225
6226 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6227 Node* n;
6228 n = mm->in(idx);
6229 if (mem == n) return true; // might be empty_memory()
6230 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6231 if (mem == n) return true;
6232 return false;
6233 }
6234 #endif // !PRODUCT