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