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