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() == oopDesc::klass_offset_in_bytes() ||
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() == oopDesc::klass_offset_in_bytes()),
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 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1975 // The field is Klass::_modifier_flags. Return its (constant) value.
1976 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1977 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1978 return TypeInt::make(klass->modifier_flags());
1979 }
1980 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1981 // The field is Klass::_access_flags. Return its (constant) value.
1982 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1983 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1984 return TypeInt::make(klass->access_flags());
1985 }
1986 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
1987 // The field is Klass::_misc_flags. Return its (constant) value.
1988 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1989 assert(this->Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
1990 return TypeInt::make(klass->misc_flags());
1991 }
1992 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1993 // The field is Klass::_layout_helper. Return its constant value if known.
1994 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1995 return TypeInt::make(klass->layout_helper());
1996 }
1997
1998 // No match.
1999 return nullptr;
2000 }
2001
2002 //------------------------------Value-----------------------------------------
2003 const Type* LoadNode::Value(PhaseGVN* phase) const {
2004 // Either input is TOP ==> the result is TOP
2005 Node* mem = in(MemNode::Memory);
2006 const Type *t1 = phase->type(mem);
2007 if (t1 == Type::TOP) return Type::TOP;
2008 Node* adr = in(MemNode::Address);
2009 const TypePtr* tp = phase->type(adr)->isa_ptr();
2010 if (tp == nullptr || tp->empty()) return Type::TOP;
2011 int off = tp->offset();
2012 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2013 Compile* C = phase->C;
2014
2015 // Try to guess loaded type from pointer type
2016 if (tp->isa_aryptr()) {
2017 const TypeAryPtr* ary = tp->is_aryptr();
2018 const Type* t = ary->elem();
2019
2020 // Determine whether the reference is beyond the header or not, by comparing
2021 // the offset against the offset of the start of the array's data.
2022 // Different array types begin at slightly different offsets (12 vs. 16).
2023 // We choose T_BYTE as an example base type that is least restrictive
2024 // as to alignment, which will therefore produce the smallest
2025 // possible base offset.
2026 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2027 const bool off_beyond_header = (off >= min_base_off);
2028
2029 // Try to constant-fold a stable array element.
2030 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2031 // Make sure the reference is not into the header and the offset is constant
2032 ciObject* aobj = ary->const_oop();
2033 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2034 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2035 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2036 stable_dimension,
2037 memory_type(), is_unsigned());
2038 if (con_type != nullptr) {
2039 return con_type;
2040 }
2041 }
2042 }
2043
2044 // Don't do this for integer types. There is only potential profit if
2045 // the element type t is lower than _type; that is, for int types, if _type is
2046 // more restrictive than t. This only happens here if one is short and the other
2047 // char (both 16 bits), and in those cases we've made an intentional decision
2048 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2049 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2050 //
2051 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2052 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2053 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2054 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2055 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2056 // In fact, that could have been the original type of p1, and p1 could have
2057 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2058 // expression (LShiftL quux 3) independently optimized to the constant 8.
2059 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2060 && (_type->isa_vect() == nullptr)
2061 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2062 // t might actually be lower than _type, if _type is a unique
2063 // concrete subclass of abstract class t.
2064 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2065 const Type* jt = t->join_speculative(_type);
2066 // In any case, do not allow the join, per se, to empty out the type.
2067 if (jt->empty() && !t->empty()) {
2068 // This can happen if a interface-typed array narrows to a class type.
2069 jt = _type;
2070 }
2071 #ifdef ASSERT
2072 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2073 // The pointers in the autobox arrays are always non-null
2074 Node* base = adr->in(AddPNode::Base);
2075 if ((base != nullptr) && base->is_DecodeN()) {
2076 // Get LoadN node which loads IntegerCache.cache field
2077 base = base->in(1);
2078 }
2079 if ((base != nullptr) && base->is_Con()) {
2080 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2081 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2082 // It could be narrow oop
2083 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2084 }
2085 }
2086 }
2087 #endif
2088 return jt;
2089 }
2090 }
2091 } else if (tp->base() == Type::InstPtr) {
2092 assert( off != Type::OffsetBot ||
2093 // arrays can be cast to Objects
2094 !tp->isa_instptr() ||
2095 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2096 // unsafe field access may not have a constant offset
2097 C->has_unsafe_access(),
2098 "Field accesses must be precise" );
2099 // For oop loads, we expect the _type to be precise.
2100
2101 // Optimize loads from constant fields.
2102 const TypeInstPtr* tinst = tp->is_instptr();
2103 ciObject* const_oop = tinst->const_oop();
2104 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2105 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
2106 if (con_type != nullptr) {
2107 return con_type;
2108 }
2109 }
2110 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2111 assert(off != Type::OffsetBot ||
2112 !tp->isa_instklassptr() ||
2113 // arrays can be cast to Objects
2114 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2115 // also allow array-loading from the primary supertype
2116 // array during subtype checks
2117 Opcode() == Op_LoadKlass,
2118 "Field accesses must be precise");
2119 // For klass/static loads, we expect the _type to be precise
2120 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2121 /* With mirrors being an indirect in the Klass*
2122 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2123 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2124 *
2125 * So check the type and klass of the node before the LoadP.
2126 */
2127 Node* adr2 = adr->in(MemNode::Address);
2128 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2129 if (tkls != nullptr && !StressReflectiveCode) {
2130 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2131 ciKlass* klass = tkls->exact_klass();
2132 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2133 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2134 return TypeInstPtr::make(klass->java_mirror());
2135 }
2136 }
2137 }
2138
2139 const TypeKlassPtr *tkls = tp->isa_klassptr();
2140 if (tkls != nullptr) {
2141 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2142 ciKlass* klass = tkls->exact_klass();
2143 // We are loading a field from a Klass metaobject whose identity
2144 // is known at compile time (the type is "exact" or "precise").
2145 // Check for fields we know are maintained as constants by the VM.
2146 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2147 // The field is Klass::_super_check_offset. Return its (constant) value.
2148 // (Folds up type checking code.)
2149 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2150 return TypeInt::make(klass->super_check_offset());
2151 }
2152 // Compute index into primary_supers array
2153 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2154 // Check for overflowing; use unsigned compare to handle the negative case.
2155 if( depth < ciKlass::primary_super_limit() ) {
2156 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2157 // (Folds up type checking code.)
2158 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2159 ciKlass *ss = klass->super_of_depth(depth);
2160 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2161 }
2162 const Type* aift = load_array_final_field(tkls, klass);
2163 if (aift != nullptr) return aift;
2164 }
2165
2166 // We can still check if we are loading from the primary_supers array at a
2167 // shallow enough depth. Even though the klass is not exact, entries less
2168 // than or equal to its super depth are correct.
2169 if (tkls->is_loaded()) {
2170 ciKlass* klass = nullptr;
2171 if (tkls->isa_instklassptr()) {
2172 klass = tkls->is_instklassptr()->instance_klass();
2173 } else {
2174 int dims;
2175 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2176 if (inner->isa_instklassptr()) {
2177 klass = inner->is_instklassptr()->instance_klass();
2178 klass = ciObjArrayKlass::make(klass, dims);
2179 }
2180 }
2181 if (klass != nullptr) {
2182 // Compute index into primary_supers array
2183 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2184 // Check for overflowing; use unsigned compare to handle the negative case.
2185 if (depth < ciKlass::primary_super_limit() &&
2186 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2187 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2188 // (Folds up type checking code.)
2189 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2190 ciKlass *ss = klass->super_of_depth(depth);
2191 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2192 }
2193 }
2194 }
2195
2196 // If the type is enough to determine that the thing is not an array,
2197 // we can give the layout_helper a positive interval type.
2198 // This will help short-circuit some reflective code.
2199 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2200 tkls->isa_instklassptr() && // not directly typed as an array
2201 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2202 ) {
2203 // Note: When interfaces are reliable, we can narrow the interface
2204 // test to (klass != Serializable && klass != Cloneable).
2205 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2206 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2207 // The key property of this type is that it folds up tests
2208 // for array-ness, since it proves that the layout_helper is positive.
2209 // Thus, a generic value like the basic object layout helper works fine.
2210 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2211 }
2212 }
2213
2214 // If we are loading from a freshly-allocated object, produce a zero,
2215 // if the load is provably beyond the header of the object.
2216 // (Also allow a variable load from a fresh array to produce zero.)
2217 const TypeOopPtr *tinst = tp->isa_oopptr();
2218 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2219 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value();
2220 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2221 Node* value = can_see_stored_value(mem,phase);
2222 if (value != nullptr && value->is_Con()) {
2223 assert(value->bottom_type()->higher_equal(_type),"sanity");
2224 return value->bottom_type();
2225 }
2226 }
2227
2228 bool is_vect = (_type->isa_vect() != nullptr);
2229 if (is_instance && !is_vect) {
2230 // If we have an instance type and our memory input is the
2231 // programs's initial memory state, there is no matching store,
2232 // so just return a zero of the appropriate type -
2233 // except if it is vectorized - then we have no zero constant.
2234 Node *mem = in(MemNode::Memory);
2235 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2236 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2237 return Type::get_zero_type(_type->basic_type());
2238 }
2239 }
2240
2241 Node* alloc = is_new_object_mark_load();
2242 if (alloc != nullptr) {
2243 return TypeX::make(markWord::prototype().value());
2244 }
2245
2246 return _type;
2247 }
2248
2249 //------------------------------match_edge-------------------------------------
2250 // Do we Match on this edge index or not? Match only the address.
2251 uint LoadNode::match_edge(uint idx) const {
2252 return idx == MemNode::Address;
2253 }
2254
2255 //--------------------------LoadBNode::Ideal--------------------------------------
2256 //
2257 // If the previous store is to the same address as this load,
2258 // and the value stored was larger than a byte, replace this load
2259 // with the value stored truncated to a byte. If no truncation is
2260 // needed, the replacement is done in LoadNode::Identity().
2261 //
2262 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2263 Node* mem = in(MemNode::Memory);
2264 Node* value = can_see_stored_value(mem,phase);
2265 if (value != nullptr) {
2266 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2267 if (narrow != value) {
2268 return narrow;
2269 }
2270 }
2271 // Identity call will handle the case where truncation is not needed.
2272 return LoadNode::Ideal(phase, can_reshape);
2273 }
2274
2275 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2276 Node* mem = in(MemNode::Memory);
2277 Node* value = can_see_stored_value(mem,phase);
2278 if (value != nullptr && value->is_Con() &&
2279 !value->bottom_type()->higher_equal(_type)) {
2280 // If the input to the store does not fit with the load's result type,
2281 // it must be truncated. We can't delay until Ideal call since
2282 // a singleton Value is needed for split_thru_phi optimization.
2283 int con = value->get_int();
2284 return TypeInt::make((con << 24) >> 24);
2285 }
2286 return LoadNode::Value(phase);
2287 }
2288
2289 //--------------------------LoadUBNode::Ideal-------------------------------------
2290 //
2291 // If the previous store is to the same address as this load,
2292 // and the value stored was larger than a byte, replace this load
2293 // with the value stored truncated to a byte. If no truncation is
2294 // needed, the replacement is done in LoadNode::Identity().
2295 //
2296 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2297 Node* mem = in(MemNode::Memory);
2298 Node* value = can_see_stored_value(mem, phase);
2299 if (value != nullptr) {
2300 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2301 if (narrow != value) {
2302 return narrow;
2303 }
2304 }
2305 // Identity call will handle the case where truncation is not needed.
2306 return LoadNode::Ideal(phase, can_reshape);
2307 }
2308
2309 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2310 Node* mem = in(MemNode::Memory);
2311 Node* value = can_see_stored_value(mem,phase);
2312 if (value != nullptr && value->is_Con() &&
2313 !value->bottom_type()->higher_equal(_type)) {
2314 // If the input to the store does not fit with the load's result type,
2315 // it must be truncated. We can't delay until Ideal call since
2316 // a singleton Value is needed for split_thru_phi optimization.
2317 int con = value->get_int();
2318 return TypeInt::make(con & 0xFF);
2319 }
2320 return LoadNode::Value(phase);
2321 }
2322
2323 //--------------------------LoadUSNode::Ideal-------------------------------------
2324 //
2325 // If the previous store is to the same address as this load,
2326 // and the value stored was larger than a char, replace this load
2327 // with the value stored truncated to a char. If no truncation is
2328 // needed, the replacement is done in LoadNode::Identity().
2329 //
2330 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2331 Node* mem = in(MemNode::Memory);
2332 Node* value = can_see_stored_value(mem,phase);
2333 if (value != nullptr) {
2334 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2335 if (narrow != value) {
2336 return narrow;
2337 }
2338 }
2339 // Identity call will handle the case where truncation is not needed.
2340 return LoadNode::Ideal(phase, can_reshape);
2341 }
2342
2343 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2344 Node* mem = in(MemNode::Memory);
2345 Node* value = can_see_stored_value(mem,phase);
2346 if (value != nullptr && value->is_Con() &&
2347 !value->bottom_type()->higher_equal(_type)) {
2348 // If the input to the store does not fit with the load's result type,
2349 // it must be truncated. We can't delay until Ideal call since
2350 // a singleton Value is needed for split_thru_phi optimization.
2351 int con = value->get_int();
2352 return TypeInt::make(con & 0xFFFF);
2353 }
2354 return LoadNode::Value(phase);
2355 }
2356
2357 //--------------------------LoadSNode::Ideal--------------------------------------
2358 //
2359 // If the previous store is to the same address as this load,
2360 // and the value stored was larger than a short, replace this load
2361 // with the value stored truncated to a short. If no truncation is
2362 // needed, the replacement is done in LoadNode::Identity().
2363 //
2364 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2365 Node* mem = in(MemNode::Memory);
2366 Node* value = can_see_stored_value(mem,phase);
2367 if (value != nullptr) {
2368 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2369 if (narrow != value) {
2370 return narrow;
2371 }
2372 }
2373 // Identity call will handle the case where truncation is not needed.
2374 return LoadNode::Ideal(phase, can_reshape);
2375 }
2376
2377 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2378 Node* mem = in(MemNode::Memory);
2379 Node* value = can_see_stored_value(mem,phase);
2380 if (value != nullptr && value->is_Con() &&
2381 !value->bottom_type()->higher_equal(_type)) {
2382 // If the input to the store does not fit with the load's result type,
2383 // it must be truncated. We can't delay until Ideal call since
2384 // a singleton Value is needed for split_thru_phi optimization.
2385 int con = value->get_int();
2386 return TypeInt::make((con << 16) >> 16);
2387 }
2388 return LoadNode::Value(phase);
2389 }
2390
2391 //=============================================================================
2392 //----------------------------LoadKlassNode::make------------------------------
2393 // Polymorphic factory method:
2394 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2395 // sanity check the alias category against the created node type
2396 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2397 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2398 #ifdef _LP64
2399 if (adr_type->is_ptr_to_narrowklass()) {
2400 assert(UseCompressedClassPointers, "no compressed klasses");
2401 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2402 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2403 }
2404 #endif
2405 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2406 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2407 }
2408
2409 //------------------------------Value------------------------------------------
2410 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2411 return klass_value_common(phase);
2412 }
2413
2414 // In most cases, LoadKlassNode does not have the control input set. If the control
2415 // input is set, it must not be removed (by LoadNode::Ideal()).
2416 bool LoadKlassNode::can_remove_control() const {
2417 return false;
2418 }
2419
2420 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2421 // Either input is TOP ==> the result is TOP
2422 const Type *t1 = phase->type( in(MemNode::Memory) );
2423 if (t1 == Type::TOP) return Type::TOP;
2424 Node *adr = in(MemNode::Address);
2425 const Type *t2 = phase->type( adr );
2426 if (t2 == Type::TOP) return Type::TOP;
2427 const TypePtr *tp = t2->is_ptr();
2428 if (TypePtr::above_centerline(tp->ptr()) ||
2429 tp->ptr() == TypePtr::Null) return Type::TOP;
2430
2431 // Return a more precise klass, if possible
2432 const TypeInstPtr *tinst = tp->isa_instptr();
2433 if (tinst != nullptr) {
2434 ciInstanceKlass* ik = tinst->instance_klass();
2435 int offset = tinst->offset();
2436 if (ik == phase->C->env()->Class_klass()
2437 && (offset == java_lang_Class::klass_offset() ||
2438 offset == java_lang_Class::array_klass_offset())) {
2439 // We are loading a special hidden field from a Class mirror object,
2440 // the field which points to the VM's Klass metaobject.
2441 ciType* t = tinst->java_mirror_type();
2442 // java_mirror_type returns non-null for compile-time Class constants.
2443 if (t != nullptr) {
2444 // constant oop => constant klass
2445 if (offset == java_lang_Class::array_klass_offset()) {
2446 if (t->is_void()) {
2447 // We cannot create a void array. Since void is a primitive type return null
2448 // klass. Users of this result need to do a null check on the returned klass.
2449 return TypePtr::NULL_PTR;
2450 }
2451 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2452 }
2453 if (!t->is_klass()) {
2454 // a primitive Class (e.g., int.class) has null for a klass field
2455 return TypePtr::NULL_PTR;
2456 }
2457 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2458 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2459 }
2460 // non-constant mirror, so we can't tell what's going on
2461 }
2462 if (!tinst->is_loaded())
2463 return _type; // Bail out if not loaded
2464 if (offset == oopDesc::klass_offset_in_bytes()) {
2465 return tinst->as_klass_type(true);
2466 }
2467 }
2468
2469 // Check for loading klass from an array
2470 const TypeAryPtr *tary = tp->isa_aryptr();
2471 if (tary != nullptr &&
2472 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2473 return tary->as_klass_type(true);
2474 }
2475
2476 // Check for loading klass from an array klass
2477 const TypeKlassPtr *tkls = tp->isa_klassptr();
2478 if (tkls != nullptr && !StressReflectiveCode) {
2479 if (!tkls->is_loaded())
2480 return _type; // Bail out if not loaded
2481 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2482 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2483 // // Always returning precise element type is incorrect,
2484 // // e.g., element type could be object and array may contain strings
2485 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2486
2487 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2488 // according to the element type's subclassing.
2489 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2490 }
2491 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2492 tkls->offset() == in_bytes(Klass::super_offset())) {
2493 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2494 // The field is Klass::_super. Return its (constant) value.
2495 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2496 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2497 }
2498 }
2499
2500 if (tkls != nullptr && !UseSecondarySupersCache
2501 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2502 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2503 return TypePtr::NULL_PTR;
2504 }
2505
2506 // Bailout case
2507 return LoadNode::Value(phase);
2508 }
2509
2510 //------------------------------Identity---------------------------------------
2511 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2512 // Also feed through the klass in Allocate(...klass...)._klass.
2513 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2514 return klass_identity_common(phase);
2515 }
2516
2517 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2518 Node* x = LoadNode::Identity(phase);
2519 if (x != this) return x;
2520
2521 // Take apart the address into an oop and offset.
2522 // Return 'this' if we cannot.
2523 Node* adr = in(MemNode::Address);
2524 intptr_t offset = 0;
2525 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2526 if (base == nullptr) return this;
2527 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2528 if (toop == nullptr) return this;
2529
2530 // Step over potential GC barrier for OopHandle resolve
2531 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2532 if (bs->is_gc_barrier_node(base)) {
2533 base = bs->step_over_gc_barrier(base);
2534 }
2535
2536 // We can fetch the klass directly through an AllocateNode.
2537 // This works even if the klass is not constant (clone or newArray).
2538 if (offset == oopDesc::klass_offset_in_bytes()) {
2539 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2540 if (allocated_klass != nullptr) {
2541 return allocated_klass;
2542 }
2543 }
2544
2545 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2546 // See inline_native_Class_query for occurrences of these patterns.
2547 // Java Example: x.getClass().isAssignableFrom(y)
2548 //
2549 // This improves reflective code, often making the Class
2550 // mirror go completely dead. (Current exception: Class
2551 // mirrors may appear in debug info, but we could clean them out by
2552 // introducing a new debug info operator for Klass.java_mirror).
2553
2554 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2555 && offset == java_lang_Class::klass_offset()) {
2556 if (base->is_Load()) {
2557 Node* base2 = base->in(MemNode::Address);
2558 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2559 Node* adr2 = base2->in(MemNode::Address);
2560 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2561 if (tkls != nullptr && !tkls->empty()
2562 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2563 && adr2->is_AddP()
2564 ) {
2565 int mirror_field = in_bytes(Klass::java_mirror_offset());
2566 if (tkls->offset() == mirror_field) {
2567 return adr2->in(AddPNode::Base);
2568 }
2569 }
2570 }
2571 }
2572 }
2573
2574 return this;
2575 }
2576
2577 LoadNode* LoadNode::clone_pinned() const {
2578 LoadNode* ld = clone()->as_Load();
2579 ld->_control_dependency = UnknownControl;
2580 return ld;
2581 }
2582
2583
2584 //------------------------------Value------------------------------------------
2585 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2586 const Type *t = klass_value_common(phase);
2587 if (t == Type::TOP)
2588 return t;
2589
2590 return t->make_narrowklass();
2591 }
2592
2593 //------------------------------Identity---------------------------------------
2594 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2595 // Also feed through the klass in Allocate(...klass...)._klass.
2596 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2597 Node *x = klass_identity_common(phase);
2598
2599 const Type *t = phase->type( x );
2600 if( t == Type::TOP ) return x;
2601 if( t->isa_narrowklass()) return x;
2602 assert (!t->isa_narrowoop(), "no narrow oop here");
2603
2604 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2605 }
2606
2607 //------------------------------Value-----------------------------------------
2608 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2609 // Either input is TOP ==> the result is TOP
2610 const Type *t1 = phase->type( in(MemNode::Memory) );
2611 if( t1 == Type::TOP ) return Type::TOP;
2612 Node *adr = in(MemNode::Address);
2613 const Type *t2 = phase->type( adr );
2614 if( t2 == Type::TOP ) return Type::TOP;
2615 const TypePtr *tp = t2->is_ptr();
2616 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2617 const TypeAryPtr *tap = tp->isa_aryptr();
2618 if( !tap ) return _type;
2619 return tap->size();
2620 }
2621
2622 //-------------------------------Ideal---------------------------------------
2623 // Feed through the length in AllocateArray(...length...)._length.
2624 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2625 Node* p = MemNode::Ideal_common(phase, can_reshape);
2626 if (p) return (p == NodeSentinel) ? nullptr : p;
2627
2628 // Take apart the address into an oop and offset.
2629 // Return 'this' if we cannot.
2630 Node* adr = in(MemNode::Address);
2631 intptr_t offset = 0;
2632 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2633 if (base == nullptr) return nullptr;
2634 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2635 if (tary == nullptr) return nullptr;
2636
2637 // We can fetch the length directly through an AllocateArrayNode.
2638 // This works even if the length is not constant (clone or newArray).
2639 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2640 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2641 if (alloc != nullptr) {
2642 Node* allocated_length = alloc->Ideal_length();
2643 Node* len = alloc->make_ideal_length(tary, phase);
2644 if (allocated_length != len) {
2645 // New CastII improves on this.
2646 return len;
2647 }
2648 }
2649 }
2650
2651 return nullptr;
2652 }
2653
2654 //------------------------------Identity---------------------------------------
2655 // Feed through the length in AllocateArray(...length...)._length.
2656 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2657 Node* x = LoadINode::Identity(phase);
2658 if (x != this) return x;
2659
2660 // Take apart the address into an oop and offset.
2661 // Return 'this' if we cannot.
2662 Node* adr = in(MemNode::Address);
2663 intptr_t offset = 0;
2664 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2665 if (base == nullptr) return this;
2666 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2667 if (tary == nullptr) return this;
2668
2669 // We can fetch the length directly through an AllocateArrayNode.
2670 // This works even if the length is not constant (clone or newArray).
2671 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2672 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2673 if (alloc != nullptr) {
2674 Node* allocated_length = alloc->Ideal_length();
2675 // Do not allow make_ideal_length to allocate a CastII node.
2676 Node* len = alloc->make_ideal_length(tary, phase, false);
2677 if (allocated_length == len) {
2678 // Return allocated_length only if it would not be improved by a CastII.
2679 return allocated_length;
2680 }
2681 }
2682 }
2683
2684 return this;
2685
2686 }
2687
2688 //=============================================================================
2689 //---------------------------StoreNode::make-----------------------------------
2690 // Polymorphic factory method:
2691 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) {
2692 assert((mo == unordered || mo == release), "unexpected");
2693 Compile* C = gvn.C;
2694 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2695 ctl != nullptr, "raw memory operations should have control edge");
2696
2697 switch (bt) {
2698 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2699 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2700 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2701 case T_CHAR:
2702 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2703 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2704 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2705 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2706 case T_METADATA:
2707 case T_ADDRESS:
2708 case T_OBJECT:
2709 #ifdef _LP64
2710 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2711 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2712 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2713 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2714 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2715 adr->bottom_type()->isa_rawptr())) {
2716 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2717 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2718 }
2719 #endif
2720 {
2721 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2722 }
2723 default:
2724 ShouldNotReachHere();
2725 return (StoreNode*)nullptr;
2726 }
2727 }
2728
2729 //--------------------------bottom_type----------------------------------------
2730 const Type *StoreNode::bottom_type() const {
2731 return Type::MEMORY;
2732 }
2733
2734 //------------------------------hash-------------------------------------------
2735 uint StoreNode::hash() const {
2736 // unroll addition of interesting fields
2737 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2738
2739 // Since they are not commoned, do not hash them:
2740 return NO_HASH;
2741 }
2742
2743 // Link together multiple stores (B/S/C/I) into a longer one.
2744 //
2745 // Example: _store = StoreB[i+3]
2746 //
2747 // RangeCheck[i+0] RangeCheck[i+0]
2748 // StoreB[i+0]
2749 // RangeCheck[i+3] RangeCheck[i+3]
2750 // StoreB[i+1] --> pass: fail:
2751 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2752 // StoreB[i+3]
2753 //
2754 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2755 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2756 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2757 // failing path.
2758 //
2759 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2760 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2761 // and possibly no RangeChecks remain between the stores.
2762 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2763 // and those later RangeChecks can be removed. Example:
2764 //
2765 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2766 // StoreB[i+0] StoreB[i+0] <- first store
2767 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2768 // StoreB[i+1] StoreB[i+1] <- second store
2769 // RangeCheck[i+2] --> removed
2770 // StoreB[i+2] StoreB[i+2]
2771 // RangeCheck[i+3] --> removed
2772 // StoreB[i+3] StoreB[i+3] <- last store
2773 //
2774 // Thus, it is a common pattern that between the first and last store in a chain
2775 // of adjacent stores there remains exactly one RangeCheck, located between the
2776 // first and the second store (e.g. RangeCheck[i+3]).
2777 //
2778 class MergePrimitiveStores : public StackObj {
2779 private:
2780 PhaseGVN* const _phase;
2781 StoreNode* const _store;
2782
2783 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2784
2785 public:
2786 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2787 _phase(phase), _store(store)
2788 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2789 {}
2790
2791 StoreNode* run();
2792
2793 private:
2794 bool is_compatible_store(const StoreNode* other_store) const;
2795 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2796 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2797 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out);
2798 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2799 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2800
2801 class Status {
2802 private:
2803 StoreNode* _found_store;
2804 bool _found_range_check;
2805
2806 Status(StoreNode* found_store, bool found_range_check)
2807 : _found_store(found_store), _found_range_check(found_range_check) {}
2808
2809 public:
2810 StoreNode* found_store() const { return _found_store; }
2811 bool found_range_check() const { return _found_range_check; }
2812 static Status make_failure() { return Status(nullptr, false); }
2813
2814 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2815 if (cfg_status == CFGStatus::Failure) {
2816 return Status::make_failure();
2817 }
2818 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2819 }
2820
2821 #ifndef PRODUCT
2822 void print_on(outputStream* st) const {
2823 if (_found_store == nullptr) {
2824 st->print_cr("None");
2825 } else {
2826 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2827 _found_range_check ? "RC" : "no-RC");
2828 }
2829 }
2830 #endif
2831 };
2832
2833 Status find_adjacent_use_store(const StoreNode* def_store) const;
2834 Status find_adjacent_def_store(const StoreNode* use_store) const;
2835 Status find_use_store(const StoreNode* def_store) const;
2836 Status find_def_store(const StoreNode* use_store) const;
2837 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2838 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2839
2840 void collect_merge_list(Node_List& merge_list) const;
2841 Node* make_merged_input_value(const Node_List& merge_list);
2842 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2843
2844 #ifndef PRODUCT
2845 // Access to TraceMergeStores tags
2846 bool is_trace(TraceMergeStores::Tag tag) const {
2847 return _trace_tags.at(tag);
2848 }
2849
2850 bool is_trace_basic() const {
2851 return is_trace(TraceMergeStores::Tag::BASIC);
2852 }
2853
2854 bool is_trace_pointer() const {
2855 return is_trace(TraceMergeStores::Tag::POINTER);
2856 }
2857
2858 bool is_trace_aliasing() const {
2859 return is_trace(TraceMergeStores::Tag::ALIASING);
2860 }
2861
2862 bool is_trace_adjacency() const {
2863 return is_trace(TraceMergeStores::Tag::ADJACENCY);
2864 }
2865
2866 bool is_trace_success() const {
2867 return is_trace(TraceMergeStores::Tag::SUCCESS);
2868 }
2869 #endif
2870
2871 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2872 };
2873
2874 StoreNode* MergePrimitiveStores::run() {
2875 // Check for B/S/C/I
2876 int opc = _store->Opcode();
2877 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2878 return nullptr;
2879 }
2880
2881 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
2882
2883 // The _store must be the "last" store in a chain. If we find a use we could merge with
2884 // then that use or a store further down is the "last" store.
2885 Status status_use = find_adjacent_use_store(_store);
2886 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
2887 if (status_use.found_store() != nullptr) {
2888 return nullptr;
2889 }
2890
2891 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
2892 Status status_def = find_adjacent_def_store(_store);
2893 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
2894 if (status_def.found_store() == nullptr) {
2895 return nullptr;
2896 }
2897
2898 ResourceMark rm;
2899 Node_List merge_list;
2900 collect_merge_list(merge_list);
2901
2902 Node* merged_input_value = make_merged_input_value(merge_list);
2903 if (merged_input_value == nullptr) { return nullptr; }
2904
2905 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
2906
2907 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
2908
2909 return merged_store;
2910 }
2911
2912 // Check compatibility between _store and other_store.
2913 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
2914 int opc = _store->Opcode();
2915 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
2916
2917 if (other_store == nullptr ||
2918 _store->Opcode() != other_store->Opcode()) {
2919 return false;
2920 }
2921
2922 return true;
2923 }
2924
2925 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
2926 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
2927 use_store->in(MemNode::ValueIn),
2928 def_store->memory_size())) {
2929 return false;
2930 }
2931
2932 ResourceMark rm;
2933 #ifndef PRODUCT
2934 const TraceMemPointer trace(is_trace_pointer(),
2935 is_trace_aliasing(),
2936 is_trace_adjacency());
2937 #endif
2938 const MemPointer pointer_use(use_store NOT_PRODUCT( COMMA trace ));
2939 const MemPointer pointer_def(def_store NOT_PRODUCT( COMMA trace ));
2940 return pointer_def.is_adjacent_to_and_before(pointer_use);
2941 }
2942
2943 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
2944 // Pattern: [n1 = ConI, n2 = ConI]
2945 if (n1->Opcode() == Op_ConI) {
2946 return n2->Opcode() == Op_ConI;
2947 }
2948
2949 // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
2950 #ifndef VM_LITTLE_ENDIAN
2951 // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
2952 // Swapping n1 with n2 gives same pattern as on little endian platforms.
2953 swap(n1, n2);
2954 #endif // !VM_LITTLE_ENDIAN
2955 Node const* base_n2;
2956 jint shift_n2;
2957 if (!is_con_RShift(n2, base_n2, shift_n2)) {
2958 return false;
2959 }
2960 if (n1->Opcode() == Op_ConvL2I) {
2961 // look through
2962 n1 = n1->in(1);
2963 }
2964 Node const* base_n1;
2965 jint shift_n1;
2966 if (n1 == base_n2) {
2967 // n1 = base = base >> 0
2968 base_n1 = n1;
2969 shift_n1 = 0;
2970 } else if (!is_con_RShift(n1, base_n1, shift_n1)) {
2971 return false;
2972 }
2973 int bits_per_store = memory_size * 8;
2974 if (base_n1 != base_n2 ||
2975 shift_n1 + bits_per_store != shift_n2 ||
2976 shift_n1 % bits_per_store != 0) {
2977 return false;
2978 }
2979
2980 // both load from same value with correct shift
2981 return true;
2982 }
2983
2984 // Detect pattern: n = base_out >> shift_out
2985 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out) {
2986 assert(n != nullptr, "precondition");
2987
2988 int opc = n->Opcode();
2989 if (opc == Op_ConvL2I) {
2990 n = n->in(1);
2991 opc = n->Opcode();
2992 }
2993
2994 if ((opc == Op_RShiftI ||
2995 opc == Op_RShiftL ||
2996 opc == Op_URShiftI ||
2997 opc == Op_URShiftL) &&
2998 n->in(2)->is_ConI()) {
2999 base_out = n->in(1);
3000 shift_out = n->in(2)->get_int();
3001 // The shift must be positive:
3002 return shift_out >= 0;
3003 }
3004 return false;
3005 }
3006
3007 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3008 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3009 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3010
3011 Node* ctrl_use = use_store->in(MemNode::Control);
3012 Node* ctrl_def = def_store->in(MemNode::Control);
3013 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3014 return CFGStatus::Failure;
3015 }
3016
3017 if (ctrl_use == ctrl_def) {
3018 // Same ctrl -> no RangeCheck in between.
3019 // Check: use_store must be the only use of def_store.
3020 if (def_store->outcnt() > 1) {
3021 return CFGStatus::Failure;
3022 }
3023 return CFGStatus::SuccessNoRangeCheck;
3024 }
3025
3026 // Different ctrl -> could have RangeCheck in between.
3027 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3028 // 2. ctrl separated by RangeCheck.
3029 if (def_store->outcnt() != 2) {
3030 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3031 }
3032 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3033 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3034 if (merge_mem == nullptr ||
3035 merge_mem->outcnt() != 1) {
3036 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3037 }
3038 if (!ctrl_use->is_IfProj() ||
3039 !ctrl_use->in(0)->is_RangeCheck() ||
3040 ctrl_use->in(0)->outcnt() != 2) {
3041 return CFGStatus::Failure; // Not RangeCheck.
3042 }
3043 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3044 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3045 if (trap != merge_mem->unique_out() ||
3046 ctrl_use->in(0)->in(0) != ctrl_def) {
3047 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3048 }
3049
3050 return CFGStatus::SuccessWithRangeCheck;
3051 }
3052
3053 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3054 Status status_use = find_use_store(def_store);
3055 StoreNode* use_store = status_use.found_store();
3056 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3057 return Status::make_failure();
3058 }
3059 return status_use;
3060 }
3061
3062 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3063 Status status_def = find_def_store(use_store);
3064 StoreNode* def_store = status_def.found_store();
3065 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3066 return Status::make_failure();
3067 }
3068 return status_def;
3069 }
3070
3071 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3072 Status status_use = find_use_store_unidirectional(def_store);
3073
3074 #ifdef ASSERT
3075 StoreNode* use_store = status_use.found_store();
3076 if (use_store != nullptr) {
3077 Status status_def = find_def_store_unidirectional(use_store);
3078 assert(status_def.found_store() == def_store &&
3079 status_def.found_range_check() == status_use.found_range_check(),
3080 "find_use_store and find_def_store must be symmetric");
3081 }
3082 #endif
3083
3084 return status_use;
3085 }
3086
3087 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3088 Status status_def = find_def_store_unidirectional(use_store);
3089
3090 #ifdef ASSERT
3091 StoreNode* def_store = status_def.found_store();
3092 if (def_store != nullptr) {
3093 Status status_use = find_use_store_unidirectional(def_store);
3094 assert(status_use.found_store() == use_store &&
3095 status_use.found_range_check() == status_def.found_range_check(),
3096 "find_use_store and find_def_store must be symmetric");
3097 }
3098 #endif
3099
3100 return status_def;
3101 }
3102
3103 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3104 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3105
3106 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3107 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3108 if (is_compatible_store(use_store)) {
3109 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3110 }
3111 }
3112
3113 return Status::make_failure();
3114 }
3115
3116 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3117 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3118
3119 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3120 if (!is_compatible_store(def_store)) {
3121 return Status::make_failure();
3122 }
3123
3124 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3125 }
3126
3127 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3128 // The merged store can be at most 8 bytes.
3129 const uint merge_list_max_size = 8 / _store->memory_size();
3130 assert(merge_list_max_size >= 2 &&
3131 merge_list_max_size <= 8 &&
3132 is_power_of_2(merge_list_max_size),
3133 "must be 2, 4 or 8");
3134
3135 // Traverse up the chain of adjacent def stores.
3136 StoreNode* current = _store;
3137 merge_list.push(current);
3138 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3139 Status status = find_adjacent_def_store(current);
3140 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3141
3142 current = status.found_store();
3143 if (current != nullptr) {
3144 merge_list.push(current);
3145
3146 // We can have at most one RangeCheck.
3147 if (status.found_range_check()) {
3148 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3149 break;
3150 }
3151 }
3152 }
3153
3154 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3155
3156 // Truncate the merge_list to a power of 2.
3157 const uint pow2size = round_down_power_of_2(merge_list.size());
3158 assert(pow2size >= 2, "must be merging at least 2 stores");
3159 while (merge_list.size() > pow2size) { merge_list.pop(); }
3160
3161 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3162 }
3163
3164 // Merge the input values of the smaller stores to a single larger input value.
3165 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3166 int new_memory_size = _store->memory_size() * merge_list.size();
3167 Node* first = merge_list.at(merge_list.size()-1);
3168 Node* merged_input_value = nullptr;
3169 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3170 // Pattern: [ConI, ConI, ...] -> new constant
3171 jlong con = 0;
3172 jlong bits_per_store = _store->memory_size() * 8;
3173 jlong mask = (((jlong)1) << bits_per_store) - 1;
3174 for (uint i = 0; i < merge_list.size(); i++) {
3175 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3176 #ifdef VM_LITTLE_ENDIAN
3177 con = con << bits_per_store;
3178 con = con | (mask & con_i);
3179 #else // VM_LITTLE_ENDIAN
3180 con_i = (mask & con_i) << (i * bits_per_store);
3181 con = con | con_i;
3182 #endif // VM_LITTLE_ENDIAN
3183 }
3184 merged_input_value = _phase->longcon(con);
3185 } else {
3186 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3187 // | |
3188 // _store first
3189 //
3190 Node* hi = _store->in(MemNode::ValueIn);
3191 Node* lo = first->in(MemNode::ValueIn);
3192 #ifndef VM_LITTLE_ENDIAN
3193 // `_store` and `first` are swapped in the diagram above
3194 swap(hi, lo);
3195 #endif // !VM_LITTLE_ENDIAN
3196 Node const* hi_base;
3197 jint hi_shift;
3198 merged_input_value = lo;
3199 bool is_true = is_con_RShift(hi, hi_base, hi_shift);
3200 assert(is_true, "must detect con RShift");
3201 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3202 // look through
3203 merged_input_value = merged_input_value->in(1);
3204 }
3205 if (merged_input_value != hi_base) {
3206 // merged_input_value is not the base
3207 return nullptr;
3208 }
3209 }
3210
3211 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3212 // Example:
3213 //
3214 // long base = ...;
3215 // a[0] = (byte)(base >> 0);
3216 // a[1] = (byte)(base >> 8);
3217 //
3218 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3219 }
3220
3221 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3222 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3223 "merged_input_value is either int or long, and new_memory_size is small enough");
3224
3225 return merged_input_value;
3226 }
3227
3228 // //
3229 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3230 // | | | | | | //
3231 // | | | | +---------------+ | //
3232 // | | | | | | | //
3233 // | | +---------+ | | +---------------+ //
3234 // | | | | | | | | //
3235 // +--------------+ | | v1 +------------------------------+ | | v1 //
3236 // | | | | | | | | | | | | //
3237 // RangeCheck first_store RangeCheck | | first_store //
3238 // | | | | | | | //
3239 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3240 // | | ===> | | | //
3241 // +--------------+ | a2 v2 | | | //
3242 // | | | | | | | | //
3243 // | second_store | | | //
3244 // | | | | | [v1 v2 ... vn] //
3245 // ... ... | | | | //
3246 // | | | | | v //
3247 // +--------------+ | an vn +--------------+ | | merged_input_value //
3248 // | | | | | | | | //
3249 // last_store (= _store) merged_store //
3250 // //
3251 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3252 Node* first_store = merge_list.at(merge_list.size()-1);
3253 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3254 Node* first_mem = first_store->in(MemNode::Memory);
3255 Node* first_adr = first_store->in(MemNode::Address);
3256
3257 const TypePtr* new_adr_type = _store->adr_type();
3258
3259 int new_memory_size = _store->memory_size() * merge_list.size();
3260 BasicType bt = T_ILLEGAL;
3261 switch (new_memory_size) {
3262 case 2: bt = T_SHORT; break;
3263 case 4: bt = T_INT; break;
3264 case 8: bt = T_LONG; break;
3265 }
3266
3267 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3268 new_adr_type, merged_input_value, bt, MemNode::unordered);
3269
3270 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3271 // are all on the same slice. Hence, we need no barriers.
3272 merged_store->set_mismatched_access();
3273
3274 // Constants above may now also be be packed -> put candidate on worklist
3275 _phase->is_IterGVN()->_worklist.push(first_mem);
3276
3277 return merged_store;
3278 }
3279
3280 #ifndef PRODUCT
3281 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3282 stringStream ss;
3283 ss.print_cr("[TraceMergeStores]: Replace");
3284 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3285 merge_list.at(i)->dump("\n", false, &ss);
3286 }
3287 ss.print_cr("[TraceMergeStores]: with");
3288 merged_input_value->dump("\n", false, &ss);
3289 merged_store->dump("\n", false, &ss);
3290 tty->print("%s", ss.as_string());
3291 }
3292 #endif
3293
3294 //------------------------------Ideal------------------------------------------
3295 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3296 // When a store immediately follows a relevant allocation/initialization,
3297 // try to capture it into the initialization, or hoist it above.
3298 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3299 Node* p = MemNode::Ideal_common(phase, can_reshape);
3300 if (p) return (p == NodeSentinel) ? nullptr : p;
3301
3302 Node* mem = in(MemNode::Memory);
3303 Node* address = in(MemNode::Address);
3304 Node* value = in(MemNode::ValueIn);
3305 // Back-to-back stores to same address? Fold em up. Generally
3306 // unsafe if I have intervening uses.
3307 {
3308 Node* st = mem;
3309 // If Store 'st' has more than one use, we cannot fold 'st' away.
3310 // For example, 'st' might be the final state at a conditional
3311 // return. Or, 'st' might be used by some node which is live at
3312 // the same time 'st' is live, which might be unschedulable. So,
3313 // require exactly ONE user until such time as we clone 'mem' for
3314 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3315 // true).
3316 while (st->is_Store() && st->outcnt() == 1) {
3317 // Looking at a dead closed cycle of memory?
3318 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3319 assert(Opcode() == st->Opcode() ||
3320 st->Opcode() == Op_StoreVector ||
3321 Opcode() == Op_StoreVector ||
3322 st->Opcode() == Op_StoreVectorScatter ||
3323 Opcode() == Op_StoreVectorScatter ||
3324 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3325 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3326 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3327 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3328 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3329
3330 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3331 st->as_Store()->memory_size() <= this->memory_size()) {
3332 Node* use = st->raw_out(0);
3333 if (phase->is_IterGVN()) {
3334 phase->is_IterGVN()->rehash_node_delayed(use);
3335 }
3336 // It's OK to do this in the parser, since DU info is always accurate,
3337 // and the parser always refers to nodes via SafePointNode maps.
3338 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3339 return this;
3340 }
3341 st = st->in(MemNode::Memory);
3342 }
3343 }
3344
3345
3346 // Capture an unaliased, unconditional, simple store into an initializer.
3347 // Or, if it is independent of the allocation, hoist it above the allocation.
3348 if (ReduceFieldZeroing && /*can_reshape &&*/
3349 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3350 InitializeNode* init = mem->in(0)->as_Initialize();
3351 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3352 if (offset > 0) {
3353 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3354 // If the InitializeNode captured me, it made a raw copy of me,
3355 // and I need to disappear.
3356 if (moved != nullptr) {
3357 // %%% hack to ensure that Ideal returns a new node:
3358 mem = MergeMemNode::make(mem);
3359 return mem; // fold me away
3360 }
3361 }
3362 }
3363
3364 // Fold reinterpret cast into memory operation:
3365 // StoreX mem (MoveY2X v) => StoreY mem v
3366 if (value->is_Move()) {
3367 const Type* vt = value->in(1)->bottom_type();
3368 if (has_reinterpret_variant(vt)) {
3369 if (phase->C->post_loop_opts_phase()) {
3370 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3371 } else {
3372 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3373 }
3374 }
3375 }
3376
3377 if (MergeStores && UseUnalignedAccesses) {
3378 if (phase->C->post_loop_opts_phase()) {
3379 MergePrimitiveStores merge(phase, this);
3380 Node* progress = merge.run();
3381 if (progress != nullptr) { return progress; }
3382 } else {
3383 phase->C->record_for_post_loop_opts_igvn(this);
3384 }
3385 }
3386
3387 return nullptr; // No further progress
3388 }
3389
3390 //------------------------------Value-----------------------------------------
3391 const Type* StoreNode::Value(PhaseGVN* phase) const {
3392 // Either input is TOP ==> the result is TOP
3393 const Type *t1 = phase->type( in(MemNode::Memory) );
3394 if( t1 == Type::TOP ) return Type::TOP;
3395 const Type *t2 = phase->type( in(MemNode::Address) );
3396 if( t2 == Type::TOP ) return Type::TOP;
3397 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3398 if( t3 == Type::TOP ) return Type::TOP;
3399 return Type::MEMORY;
3400 }
3401
3402 //------------------------------Identity---------------------------------------
3403 // Remove redundant stores:
3404 // Store(m, p, Load(m, p)) changes to m.
3405 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3406 Node* StoreNode::Identity(PhaseGVN* phase) {
3407 Node* mem = in(MemNode::Memory);
3408 Node* adr = in(MemNode::Address);
3409 Node* val = in(MemNode::ValueIn);
3410
3411 Node* result = this;
3412
3413 // Load then Store? Then the Store is useless
3414 if (val->is_Load() &&
3415 val->in(MemNode::Address)->eqv_uncast(adr) &&
3416 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3417 val->as_Load()->store_Opcode() == Opcode()) {
3418 // Ensure vector type is the same
3419 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3420 result = mem;
3421 }
3422 }
3423
3424 // Two stores in a row of the same value?
3425 if (result == this &&
3426 mem->is_Store() &&
3427 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3428 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3429 mem->Opcode() == Opcode()) {
3430 if (!is_StoreVector()) {
3431 result = mem;
3432 } else {
3433 const StoreVectorNode* store_vector = as_StoreVector();
3434 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3435 const Node* store_indices = store_vector->indices();
3436 const Node* mem_indices = mem_vector->indices();
3437 const Node* store_mask = store_vector->mask();
3438 const Node* mem_mask = mem_vector->mask();
3439 // Ensure types, indices, and masks match
3440 if (store_vector->vect_type() == mem_vector->vect_type() &&
3441 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3442 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3443 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3444 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3445 result = mem;
3446 }
3447 }
3448 }
3449
3450 // Store of zero anywhere into a freshly-allocated object?
3451 // Then the store is useless.
3452 // (It must already have been captured by the InitializeNode.)
3453 if (result == this &&
3454 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3455 // a newly allocated object is already all-zeroes everywhere
3456 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3457 result = mem;
3458 }
3459
3460 if (result == this) {
3461 // the store may also apply to zero-bits in an earlier object
3462 Node* prev_mem = find_previous_store(phase);
3463 // Steps (a), (b): Walk past independent stores to find an exact match.
3464 if (prev_mem != nullptr) {
3465 Node* prev_val = can_see_stored_value(prev_mem, phase);
3466 if (prev_val != nullptr && prev_val == val) {
3467 // prev_val and val might differ by a cast; it would be good
3468 // to keep the more informative of the two.
3469 result = mem;
3470 }
3471 }
3472 }
3473 }
3474
3475 PhaseIterGVN* igvn = phase->is_IterGVN();
3476 if (result != this && igvn != nullptr) {
3477 MemBarNode* trailing = trailing_membar();
3478 if (trailing != nullptr) {
3479 #ifdef ASSERT
3480 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3481 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3482 #endif
3483 trailing->remove(igvn);
3484 }
3485 }
3486
3487 return result;
3488 }
3489
3490 //------------------------------match_edge-------------------------------------
3491 // Do we Match on this edge index or not? Match only memory & value
3492 uint StoreNode::match_edge(uint idx) const {
3493 return idx == MemNode::Address || idx == MemNode::ValueIn;
3494 }
3495
3496 //------------------------------cmp--------------------------------------------
3497 // Do not common stores up together. They generally have to be split
3498 // back up anyways, so do not bother.
3499 bool StoreNode::cmp( const Node &n ) const {
3500 return (&n == this); // Always fail except on self
3501 }
3502
3503 //------------------------------Ideal_masked_input-----------------------------
3504 // Check for a useless mask before a partial-word store
3505 // (StoreB ... (AndI valIn conIa) )
3506 // If (conIa & mask == mask) this simplifies to
3507 // (StoreB ... (valIn) )
3508 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3509 Node *val = in(MemNode::ValueIn);
3510 if( val->Opcode() == Op_AndI ) {
3511 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3512 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3513 set_req_X(MemNode::ValueIn, val->in(1), phase);
3514 return this;
3515 }
3516 }
3517 return nullptr;
3518 }
3519
3520
3521 //------------------------------Ideal_sign_extended_input----------------------
3522 // Check for useless sign-extension before a partial-word store
3523 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
3524 // If (conIL == conIR && conIR <= num_bits) this simplifies to
3525 // (StoreB ... (valIn) )
3526 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
3527 Node *val = in(MemNode::ValueIn);
3528 if( val->Opcode() == Op_RShiftI ) {
3529 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3530 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
3531 Node *shl = val->in(1);
3532 if( shl->Opcode() == Op_LShiftI ) {
3533 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
3534 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
3535 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3536 return this;
3537 }
3538 }
3539 }
3540 }
3541 return nullptr;
3542 }
3543
3544 //------------------------------value_never_loaded-----------------------------------
3545 // Determine whether there are any possible loads of the value stored.
3546 // For simplicity, we actually check if there are any loads from the
3547 // address stored to, not just for loads of the value stored by this node.
3548 //
3549 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3550 Node *adr = in(Address);
3551 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3552 if (adr_oop == nullptr)
3553 return false;
3554 if (!adr_oop->is_known_instance_field())
3555 return false; // if not a distinct instance, there may be aliases of the address
3556 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3557 Node *use = adr->fast_out(i);
3558 if (use->is_Load() || use->is_LoadStore()) {
3559 return false;
3560 }
3561 }
3562 return true;
3563 }
3564
3565 MemBarNode* StoreNode::trailing_membar() const {
3566 if (is_release()) {
3567 MemBarNode* trailing_mb = nullptr;
3568 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3569 Node* u = fast_out(i);
3570 if (u->is_MemBar()) {
3571 if (u->as_MemBar()->trailing_store()) {
3572 assert(u->Opcode() == Op_MemBarVolatile, "");
3573 assert(trailing_mb == nullptr, "only one");
3574 trailing_mb = u->as_MemBar();
3575 #ifdef ASSERT
3576 Node* leading = u->as_MemBar()->leading_membar();
3577 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3578 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3579 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3580 #endif
3581 } else {
3582 assert(u->as_MemBar()->standalone(), "");
3583 }
3584 }
3585 }
3586 return trailing_mb;
3587 }
3588 return nullptr;
3589 }
3590
3591
3592 //=============================================================================
3593 //------------------------------Ideal------------------------------------------
3594 // If the store is from an AND mask that leaves the low bits untouched, then
3595 // we can skip the AND operation. If the store is from a sign-extension
3596 // (a left shift, then right shift) we can skip both.
3597 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3598 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3599 if( progress != nullptr ) return progress;
3600
3601 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3602 if( progress != nullptr ) return progress;
3603
3604 // Finally check the default case
3605 return StoreNode::Ideal(phase, can_reshape);
3606 }
3607
3608 //=============================================================================
3609 //------------------------------Ideal------------------------------------------
3610 // If the store is from an AND mask that leaves the low bits untouched, then
3611 // we can skip the AND operation
3612 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3613 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3614 if( progress != nullptr ) return progress;
3615
3616 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3617 if( progress != nullptr ) return progress;
3618
3619 // Finally check the default case
3620 return StoreNode::Ideal(phase, can_reshape);
3621 }
3622
3623 //=============================================================================
3624 //----------------------------------SCMemProjNode------------------------------
3625 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3626 {
3627 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3628 return Type::TOP;
3629 }
3630 return bottom_type();
3631 }
3632
3633 //=============================================================================
3634 //----------------------------------LoadStoreNode------------------------------
3635 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3636 : Node(required),
3637 _type(rt),
3638 _adr_type(at),
3639 _barrier_data(0)
3640 {
3641 init_req(MemNode::Control, c );
3642 init_req(MemNode::Memory , mem);
3643 init_req(MemNode::Address, adr);
3644 init_req(MemNode::ValueIn, val);
3645 init_class_id(Class_LoadStore);
3646 }
3647
3648 //------------------------------Value-----------------------------------------
3649 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3650 // Either input is TOP ==> the result is TOP
3651 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3652 return Type::TOP;
3653 }
3654 const Type* t = phase->type(in(MemNode::Memory));
3655 if (t == Type::TOP) {
3656 return Type::TOP;
3657 }
3658 t = phase->type(in(MemNode::Address));
3659 if (t == Type::TOP) {
3660 return Type::TOP;
3661 }
3662 t = phase->type(in(MemNode::ValueIn));
3663 if (t == Type::TOP) {
3664 return Type::TOP;
3665 }
3666 return bottom_type();
3667 }
3668
3669 uint LoadStoreNode::ideal_reg() const {
3670 return _type->ideal_reg();
3671 }
3672
3673 // This method conservatively checks if the result of a LoadStoreNode is
3674 // used, that is, if it returns true, then it is definitely the case that
3675 // the result of the node is not needed.
3676 // For example, GetAndAdd can be matched into a lock_add instead of a
3677 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
3678 bool LoadStoreNode::result_not_used() const {
3679 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3680 Node *x = fast_out(i);
3681 if (x->Opcode() == Op_SCMemProj) {
3682 continue;
3683 }
3684 if (x->bottom_type() == TypeTuple::MEMBAR &&
3685 !x->is_Call() &&
3686 x->Opcode() != Op_Blackhole) {
3687 continue;
3688 }
3689 return false;
3690 }
3691 return true;
3692 }
3693
3694 MemBarNode* LoadStoreNode::trailing_membar() const {
3695 MemBarNode* trailing = nullptr;
3696 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3697 Node* u = fast_out(i);
3698 if (u->is_MemBar()) {
3699 if (u->as_MemBar()->trailing_load_store()) {
3700 assert(u->Opcode() == Op_MemBarAcquire, "");
3701 assert(trailing == nullptr, "only one");
3702 trailing = u->as_MemBar();
3703 #ifdef ASSERT
3704 Node* leading = trailing->leading_membar();
3705 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3706 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3707 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3708 #endif
3709 } else {
3710 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3711 }
3712 }
3713 }
3714
3715 return trailing;
3716 }
3717
3718 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3719
3720 //=============================================================================
3721 //----------------------------------LoadStoreConditionalNode--------------------
3722 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
3723 init_req(ExpectedIn, ex );
3724 }
3725
3726 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3727 // Either input is TOP ==> the result is TOP
3728 const Type* t = phase->type(in(ExpectedIn));
3729 if (t == Type::TOP) {
3730 return Type::TOP;
3731 }
3732 return LoadStoreNode::Value(phase);
3733 }
3734
3735 //=============================================================================
3736 //-------------------------------adr_type--------------------------------------
3737 const TypePtr* ClearArrayNode::adr_type() const {
3738 Node *adr = in(3);
3739 if (adr == nullptr) return nullptr; // node is dead
3740 return MemNode::calculate_adr_type(adr->bottom_type());
3741 }
3742
3743 //------------------------------match_edge-------------------------------------
3744 // Do we Match on this edge index or not? Do not match memory
3745 uint ClearArrayNode::match_edge(uint idx) const {
3746 return idx > 1;
3747 }
3748
3749 //------------------------------Identity---------------------------------------
3750 // Clearing a zero length array does nothing
3751 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3752 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3753 }
3754
3755 //------------------------------Idealize---------------------------------------
3756 // Clearing a short array is faster with stores
3757 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3758 // Already know this is a large node, do not try to ideal it
3759 if (_is_large) return nullptr;
3760
3761 const int unit = BytesPerLong;
3762 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3763 if (!t) return nullptr;
3764 if (!t->is_con()) return nullptr;
3765 intptr_t raw_count = t->get_con();
3766 intptr_t size = raw_count;
3767 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3768 // Clearing nothing uses the Identity call.
3769 // Negative clears are possible on dead ClearArrays
3770 // (see jck test stmt114.stmt11402.val).
3771 if (size <= 0 || size % unit != 0) return nullptr;
3772 intptr_t count = size / unit;
3773 // Length too long; communicate this to matchers and assemblers.
3774 // Assemblers are responsible to produce fast hardware clears for it.
3775 if (size > InitArrayShortSize) {
3776 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3777 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3778 return nullptr;
3779 }
3780 if (!IdealizeClearArrayNode) return nullptr;
3781 Node *mem = in(1);
3782 if( phase->type(mem)==Type::TOP ) return nullptr;
3783 Node *adr = in(3);
3784 const Type* at = phase->type(adr);
3785 if( at==Type::TOP ) return nullptr;
3786 const TypePtr* atp = at->isa_ptr();
3787 // adjust atp to be the correct array element address type
3788 if (atp == nullptr) atp = TypePtr::BOTTOM;
3789 else atp = atp->add_offset(Type::OffsetBot);
3790 // Get base for derived pointer purposes
3791 if( adr->Opcode() != Op_AddP ) Unimplemented();
3792 Node *base = adr->in(1);
3793
3794 Node *zero = phase->makecon(TypeLong::ZERO);
3795 Node *off = phase->MakeConX(BytesPerLong);
3796 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3797 count--;
3798 while( count-- ) {
3799 mem = phase->transform(mem);
3800 adr = phase->transform(new AddPNode(base,adr,off));
3801 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3802 }
3803 return mem;
3804 }
3805
3806 //----------------------------step_through----------------------------------
3807 // Return allocation input memory edge if it is different instance
3808 // or itself if it is the one we are looking for.
3809 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
3810 Node* n = *np;
3811 assert(n->is_ClearArray(), "sanity");
3812 intptr_t offset;
3813 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3814 // This method is called only before Allocate nodes are expanded
3815 // during macro nodes expansion. Before that ClearArray nodes are
3816 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3817 // Allocate nodes are expanded) which follows allocations.
3818 assert(alloc != nullptr, "should have allocation");
3819 if (alloc->_idx == instance_id) {
3820 // Can not bypass initialization of the instance we are looking for.
3821 return false;
3822 }
3823 // Otherwise skip it.
3824 InitializeNode* init = alloc->initialization();
3825 if (init != nullptr)
3826 *np = init->in(TypeFunc::Memory);
3827 else
3828 *np = alloc->in(TypeFunc::Memory);
3829 return true;
3830 }
3831
3832 //----------------------------clear_memory-------------------------------------
3833 // Generate code to initialize object storage to zero.
3834 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3835 intptr_t start_offset,
3836 Node* end_offset,
3837 PhaseGVN* phase) {
3838 intptr_t offset = start_offset;
3839
3840 int unit = BytesPerLong;
3841 if ((offset % unit) != 0) {
3842 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3843 adr = phase->transform(adr);
3844 const TypePtr* atp = TypeRawPtr::BOTTOM;
3845 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3846 mem = phase->transform(mem);
3847 offset += BytesPerInt;
3848 }
3849 assert((offset % unit) == 0, "");
3850
3851 // Initialize the remaining stuff, if any, with a ClearArray.
3852 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3853 }
3854
3855 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3856 Node* start_offset,
3857 Node* end_offset,
3858 PhaseGVN* phase) {
3859 if (start_offset == end_offset) {
3860 // nothing to do
3861 return mem;
3862 }
3863
3864 int unit = BytesPerLong;
3865 Node* zbase = start_offset;
3866 Node* zend = end_offset;
3867
3868 // Scale to the unit required by the CPU:
3869 if (!Matcher::init_array_count_is_in_bytes) {
3870 Node* shift = phase->intcon(exact_log2(unit));
3871 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3872 zend = phase->transform(new URShiftXNode(zend, shift) );
3873 }
3874
3875 // Bulk clear double-words
3876 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3877 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3878 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3879 return phase->transform(mem);
3880 }
3881
3882 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3883 intptr_t start_offset,
3884 intptr_t end_offset,
3885 PhaseGVN* phase) {
3886 if (start_offset == end_offset) {
3887 // nothing to do
3888 return mem;
3889 }
3890
3891 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3892 intptr_t done_offset = end_offset;
3893 if ((done_offset % BytesPerLong) != 0) {
3894 done_offset -= BytesPerInt;
3895 }
3896 if (done_offset > start_offset) {
3897 mem = clear_memory(ctl, mem, dest,
3898 start_offset, phase->MakeConX(done_offset), phase);
3899 }
3900 if (done_offset < end_offset) { // emit the final 32-bit store
3901 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3902 adr = phase->transform(adr);
3903 const TypePtr* atp = TypeRawPtr::BOTTOM;
3904 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3905 mem = phase->transform(mem);
3906 done_offset += BytesPerInt;
3907 }
3908 assert(done_offset == end_offset, "");
3909 return mem;
3910 }
3911
3912 //=============================================================================
3913 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3914 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
3915 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3916 #ifdef ASSERT
3917 , _pair_idx(0)
3918 #endif
3919 {
3920 init_class_id(Class_MemBar);
3921 Node* top = C->top();
3922 init_req(TypeFunc::I_O,top);
3923 init_req(TypeFunc::FramePtr,top);
3924 init_req(TypeFunc::ReturnAdr,top);
3925 if (precedent != nullptr)
3926 init_req(TypeFunc::Parms, precedent);
3927 }
3928
3929 //------------------------------cmp--------------------------------------------
3930 uint MemBarNode::hash() const { return NO_HASH; }
3931 bool MemBarNode::cmp( const Node &n ) const {
3932 return (&n == this); // Always fail except on self
3933 }
3934
3935 //------------------------------make-------------------------------------------
3936 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3937 switch (opcode) {
3938 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3939 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3940 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3941 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3942 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3943 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
3944 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3945 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3946 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3947 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3948 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3949 case Op_Initialize: return new InitializeNode(C, atp, pn);
3950 default: ShouldNotReachHere(); return nullptr;
3951 }
3952 }
3953
3954 void MemBarNode::remove(PhaseIterGVN *igvn) {
3955 if (outcnt() != 2) {
3956 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
3957 assert(outcnt() == 1, "Only control then");
3958 }
3959 if (trailing_store() || trailing_load_store()) {
3960 MemBarNode* leading = leading_membar();
3961 if (leading != nullptr) {
3962 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3963 leading->remove(igvn);
3964 }
3965 }
3966 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
3967 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3968 }
3969 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
3970 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3971 }
3972 }
3973
3974 //------------------------------Ideal------------------------------------------
3975 // Return a node which is more "ideal" than the current node. Strip out
3976 // control copies
3977 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3978 if (remove_dead_region(phase, can_reshape)) return this;
3979 // Don't bother trying to transform a dead node
3980 if (in(0) && in(0)->is_top()) {
3981 return nullptr;
3982 }
3983
3984 bool progress = false;
3985 // Eliminate volatile MemBars for scalar replaced objects.
3986 if (can_reshape && req() == (Precedent+1)) {
3987 bool eliminate = false;
3988 int opc = Opcode();
3989 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3990 // Volatile field loads and stores.
3991 Node* my_mem = in(MemBarNode::Precedent);
3992 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3993 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3994 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3995 // replace this Precedent (decodeN) with the Load instead.
3996 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3997 Node* load_node = my_mem->in(1);
3998 set_req(MemBarNode::Precedent, load_node);
3999 phase->is_IterGVN()->_worklist.push(my_mem);
4000 my_mem = load_node;
4001 } else {
4002 assert(my_mem->unique_out() == this, "sanity");
4003 assert(!trailing_load_store(), "load store node can't be eliminated");
4004 del_req(Precedent);
4005 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4006 my_mem = nullptr;
4007 }
4008 progress = true;
4009 }
4010 if (my_mem != nullptr && my_mem->is_Mem()) {
4011 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4012 // Check for scalar replaced object reference.
4013 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4014 t_oop->offset() != Type::OffsetBot &&
4015 t_oop->offset() != Type::OffsetTop) {
4016 eliminate = true;
4017 }
4018 }
4019 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4020 // Final field stores.
4021 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4022 if ((alloc != nullptr) && alloc->is_Allocate() &&
4023 alloc->as_Allocate()->does_not_escape_thread()) {
4024 // The allocated object does not escape.
4025 eliminate = true;
4026 }
4027 }
4028 if (eliminate) {
4029 // Replace MemBar projections by its inputs.
4030 PhaseIterGVN* igvn = phase->is_IterGVN();
4031 remove(igvn);
4032 // Must return either the original node (now dead) or a new node
4033 // (Do not return a top here, since that would break the uniqueness of top.)
4034 return new ConINode(TypeInt::ZERO);
4035 }
4036 }
4037 return progress ? this : nullptr;
4038 }
4039
4040 //------------------------------Value------------------------------------------
4041 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4042 if( !in(0) ) return Type::TOP;
4043 if( phase->type(in(0)) == Type::TOP )
4044 return Type::TOP;
4045 return TypeTuple::MEMBAR;
4046 }
4047
4048 //------------------------------match------------------------------------------
4049 // Construct projections for memory.
4050 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4051 switch (proj->_con) {
4052 case TypeFunc::Control:
4053 case TypeFunc::Memory:
4054 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4055 }
4056 ShouldNotReachHere();
4057 return nullptr;
4058 }
4059
4060 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4061 trailing->_kind = TrailingStore;
4062 leading->_kind = LeadingStore;
4063 #ifdef ASSERT
4064 trailing->_pair_idx = leading->_idx;
4065 leading->_pair_idx = leading->_idx;
4066 #endif
4067 }
4068
4069 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4070 trailing->_kind = TrailingLoadStore;
4071 leading->_kind = LeadingLoadStore;
4072 #ifdef ASSERT
4073 trailing->_pair_idx = leading->_idx;
4074 leading->_pair_idx = leading->_idx;
4075 #endif
4076 }
4077
4078 MemBarNode* MemBarNode::trailing_membar() const {
4079 ResourceMark rm;
4080 Node* trailing = (Node*)this;
4081 VectorSet seen;
4082 Node_Stack multis(0);
4083 do {
4084 Node* c = trailing;
4085 uint i = 0;
4086 do {
4087 trailing = nullptr;
4088 for (; i < c->outcnt(); i++) {
4089 Node* next = c->raw_out(i);
4090 if (next != c && next->is_CFG()) {
4091 if (c->is_MultiBranch()) {
4092 if (multis.node() == c) {
4093 multis.set_index(i+1);
4094 } else {
4095 multis.push(c, i+1);
4096 }
4097 }
4098 trailing = next;
4099 break;
4100 }
4101 }
4102 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4103 break;
4104 }
4105 while (multis.size() > 0) {
4106 c = multis.node();
4107 i = multis.index();
4108 if (i < c->req()) {
4109 break;
4110 }
4111 multis.pop();
4112 }
4113 } while (multis.size() > 0);
4114 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4115
4116 MemBarNode* mb = trailing->as_MemBar();
4117 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4118 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4119 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4120 return mb;
4121 }
4122
4123 MemBarNode* MemBarNode::leading_membar() const {
4124 ResourceMark rm;
4125 VectorSet seen;
4126 Node_Stack regions(0);
4127 Node* leading = in(0);
4128 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4129 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4130 leading = nullptr;
4131 while (regions.size() > 0 && leading == nullptr) {
4132 Node* r = regions.node();
4133 uint i = regions.index();
4134 if (i < r->req()) {
4135 leading = r->in(i);
4136 regions.set_index(i+1);
4137 } else {
4138 regions.pop();
4139 }
4140 }
4141 if (leading == nullptr) {
4142 assert(regions.size() == 0, "all paths should have been tried");
4143 return nullptr;
4144 }
4145 }
4146 if (leading->is_Region()) {
4147 regions.push(leading, 2);
4148 leading = leading->in(1);
4149 } else {
4150 leading = leading->in(0);
4151 }
4152 }
4153 #ifdef ASSERT
4154 Unique_Node_List wq;
4155 wq.push((Node*)this);
4156 uint found = 0;
4157 for (uint i = 0; i < wq.size(); i++) {
4158 Node* n = wq.at(i);
4159 if (n->is_Region()) {
4160 for (uint j = 1; j < n->req(); j++) {
4161 Node* in = n->in(j);
4162 if (in != nullptr && !in->is_top()) {
4163 wq.push(in);
4164 }
4165 }
4166 } else {
4167 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4168 assert(n == leading, "consistency check failed");
4169 found++;
4170 } else {
4171 Node* in = n->in(0);
4172 if (in != nullptr && !in->is_top()) {
4173 wq.push(in);
4174 }
4175 }
4176 }
4177 }
4178 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4179 #endif
4180 if (leading == nullptr) {
4181 return nullptr;
4182 }
4183 MemBarNode* mb = leading->as_MemBar();
4184 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4185 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4186 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4187 return mb;
4188 }
4189
4190
4191 //===========================InitializeNode====================================
4192 // SUMMARY:
4193 // This node acts as a memory barrier on raw memory, after some raw stores.
4194 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4195 // The Initialize can 'capture' suitably constrained stores as raw inits.
4196 // It can coalesce related raw stores into larger units (called 'tiles').
4197 // It can avoid zeroing new storage for memory units which have raw inits.
4198 // At macro-expansion, it is marked 'complete', and does not optimize further.
4199 //
4200 // EXAMPLE:
4201 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4202 // ctl = incoming control; mem* = incoming memory
4203 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4204 // First allocate uninitialized memory and fill in the header:
4205 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4206 // ctl := alloc.Control; mem* := alloc.Memory*
4207 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4208 // Then initialize to zero the non-header parts of the raw memory block:
4209 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4210 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4211 // After the initialize node executes, the object is ready for service:
4212 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4213 // Suppose its body is immediately initialized as {1,2}:
4214 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4215 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4216 // mem.SLICE(#short[*]) := store2
4217 //
4218 // DETAILS:
4219 // An InitializeNode collects and isolates object initialization after
4220 // an AllocateNode and before the next possible safepoint. As a
4221 // memory barrier (MemBarNode), it keeps critical stores from drifting
4222 // down past any safepoint or any publication of the allocation.
4223 // Before this barrier, a newly-allocated object may have uninitialized bits.
4224 // After this barrier, it may be treated as a real oop, and GC is allowed.
4225 //
4226 // The semantics of the InitializeNode include an implicit zeroing of
4227 // the new object from object header to the end of the object.
4228 // (The object header and end are determined by the AllocateNode.)
4229 //
4230 // Certain stores may be added as direct inputs to the InitializeNode.
4231 // These stores must update raw memory, and they must be to addresses
4232 // derived from the raw address produced by AllocateNode, and with
4233 // a constant offset. They must be ordered by increasing offset.
4234 // The first one is at in(RawStores), the last at in(req()-1).
4235 // Unlike most memory operations, they are not linked in a chain,
4236 // but are displayed in parallel as users of the rawmem output of
4237 // the allocation.
4238 //
4239 // (See comments in InitializeNode::capture_store, which continue
4240 // the example given above.)
4241 //
4242 // When the associated Allocate is macro-expanded, the InitializeNode
4243 // may be rewritten to optimize collected stores. A ClearArrayNode
4244 // may also be created at that point to represent any required zeroing.
4245 // The InitializeNode is then marked 'complete', prohibiting further
4246 // capturing of nearby memory operations.
4247 //
4248 // During macro-expansion, all captured initializations which store
4249 // constant values of 32 bits or smaller are coalesced (if advantageous)
4250 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4251 // initialized in fewer memory operations. Memory words which are
4252 // covered by neither tiles nor non-constant stores are pre-zeroed
4253 // by explicit stores of zero. (The code shape happens to do all
4254 // zeroing first, then all other stores, with both sequences occurring
4255 // in order of ascending offsets.)
4256 //
4257 // Alternatively, code may be inserted between an AllocateNode and its
4258 // InitializeNode, to perform arbitrary initialization of the new object.
4259 // E.g., the object copying intrinsics insert complex data transfers here.
4260 // The initialization must then be marked as 'complete' disable the
4261 // built-in zeroing semantics and the collection of initializing stores.
4262 //
4263 // While an InitializeNode is incomplete, reads from the memory state
4264 // produced by it are optimizable if they match the control edge and
4265 // new oop address associated with the allocation/initialization.
4266 // They return a stored value (if the offset matches) or else zero.
4267 // A write to the memory state, if it matches control and address,
4268 // and if it is to a constant offset, may be 'captured' by the
4269 // InitializeNode. It is cloned as a raw memory operation and rewired
4270 // inside the initialization, to the raw oop produced by the allocation.
4271 // Operations on addresses which are provably distinct (e.g., to
4272 // other AllocateNodes) are allowed to bypass the initialization.
4273 //
4274 // The effect of all this is to consolidate object initialization
4275 // (both arrays and non-arrays, both piecewise and bulk) into a
4276 // single location, where it can be optimized as a unit.
4277 //
4278 // Only stores with an offset less than TrackedInitializationLimit words
4279 // will be considered for capture by an InitializeNode. This puts a
4280 // reasonable limit on the complexity of optimized initializations.
4281
4282 //---------------------------InitializeNode------------------------------------
4283 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4284 : MemBarNode(C, adr_type, rawoop),
4285 _is_complete(Incomplete), _does_not_escape(false)
4286 {
4287 init_class_id(Class_Initialize);
4288
4289 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4290 assert(in(RawAddress) == rawoop, "proper init");
4291 // Note: allocation() can be null, for secondary initialization barriers
4292 }
4293
4294 // Since this node is not matched, it will be processed by the
4295 // register allocator. Declare that there are no constraints
4296 // on the allocation of the RawAddress edge.
4297 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4298 // This edge should be set to top, by the set_complete. But be conservative.
4299 if (idx == InitializeNode::RawAddress)
4300 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4301 return RegMask::Empty;
4302 }
4303
4304 Node* InitializeNode::memory(uint alias_idx) {
4305 Node* mem = in(Memory);
4306 if (mem->is_MergeMem()) {
4307 return mem->as_MergeMem()->memory_at(alias_idx);
4308 } else {
4309 // incoming raw memory is not split
4310 return mem;
4311 }
4312 }
4313
4314 bool InitializeNode::is_non_zero() {
4315 if (is_complete()) return false;
4316 remove_extra_zeroes();
4317 return (req() > RawStores);
4318 }
4319
4320 void InitializeNode::set_complete(PhaseGVN* phase) {
4321 assert(!is_complete(), "caller responsibility");
4322 _is_complete = Complete;
4323
4324 // After this node is complete, it contains a bunch of
4325 // raw-memory initializations. There is no need for
4326 // it to have anything to do with non-raw memory effects.
4327 // Therefore, tell all non-raw users to re-optimize themselves,
4328 // after skipping the memory effects of this initialization.
4329 PhaseIterGVN* igvn = phase->is_IterGVN();
4330 if (igvn) igvn->add_users_to_worklist(this);
4331 }
4332
4333 // convenience function
4334 // return false if the init contains any stores already
4335 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4336 InitializeNode* init = initialization();
4337 if (init == nullptr || init->is_complete()) return false;
4338 init->remove_extra_zeroes();
4339 // for now, if this allocation has already collected any inits, bail:
4340 if (init->is_non_zero()) return false;
4341 init->set_complete(phase);
4342 return true;
4343 }
4344
4345 void InitializeNode::remove_extra_zeroes() {
4346 if (req() == RawStores) return;
4347 Node* zmem = zero_memory();
4348 uint fill = RawStores;
4349 for (uint i = fill; i < req(); i++) {
4350 Node* n = in(i);
4351 if (n->is_top() || n == zmem) continue; // skip
4352 if (fill < i) set_req(fill, n); // compact
4353 ++fill;
4354 }
4355 // delete any empty spaces created:
4356 while (fill < req()) {
4357 del_req(fill);
4358 }
4359 }
4360
4361 // Helper for remembering which stores go with which offsets.
4362 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4363 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4364 intptr_t offset = -1;
4365 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4366 phase, offset);
4367 if (base == nullptr) return -1; // something is dead,
4368 if (offset < 0) return -1; // dead, dead
4369 return offset;
4370 }
4371
4372 // Helper for proving that an initialization expression is
4373 // "simple enough" to be folded into an object initialization.
4374 // Attempts to prove that a store's initial value 'n' can be captured
4375 // within the initialization without creating a vicious cycle, such as:
4376 // { Foo p = new Foo(); p.next = p; }
4377 // True for constants and parameters and small combinations thereof.
4378 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4379 ResourceMark rm;
4380 Unique_Node_List worklist;
4381 worklist.push(value);
4382
4383 uint complexity_limit = 20;
4384 for (uint j = 0; j < worklist.size(); j++) {
4385 if (j >= complexity_limit) {
4386 return false; // Bail out if processed too many nodes
4387 }
4388
4389 Node* n = worklist.at(j);
4390 if (n == nullptr) continue; // (can this really happen?)
4391 if (n->is_Proj()) n = n->in(0);
4392 if (n == this) return false; // found a cycle
4393 if (n->is_Con()) continue;
4394 if (n->is_Start()) continue; // params, etc., are OK
4395 if (n->is_Root()) continue; // even better
4396
4397 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4398 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4399 continue;
4400 }
4401
4402 Node* ctl = n->in(0);
4403 if (ctl != nullptr && !ctl->is_top()) {
4404 if (ctl->is_Proj()) ctl = ctl->in(0);
4405 if (ctl == this) return false;
4406
4407 // If we already know that the enclosing memory op is pinned right after
4408 // the init, then any control flow that the store has picked up
4409 // must have preceded the init, or else be equal to the init.
4410 // Even after loop optimizations (which might change control edges)
4411 // a store is never pinned *before* the availability of its inputs.
4412 if (!MemNode::all_controls_dominate(n, this)) {
4413 return false; // failed to prove a good control
4414 }
4415 }
4416
4417 // Check data edges for possible dependencies on 'this'.
4418 for (uint i = 1; i < n->req(); i++) {
4419 Node* m = n->in(i);
4420 if (m == nullptr || m == n || m->is_top()) continue;
4421
4422 // Only process data inputs once
4423 worklist.push(m);
4424 }
4425 }
4426
4427 return true;
4428 }
4429
4430 // Here are all the checks a Store must pass before it can be moved into
4431 // an initialization. Returns zero if a check fails.
4432 // On success, returns the (constant) offset to which the store applies,
4433 // within the initialized memory.
4434 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4435 const int FAIL = 0;
4436 if (st->req() != MemNode::ValueIn + 1)
4437 return FAIL; // an inscrutable StoreNode (card mark?)
4438 Node* ctl = st->in(MemNode::Control);
4439 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4440 return FAIL; // must be unconditional after the initialization
4441 Node* mem = st->in(MemNode::Memory);
4442 if (!(mem->is_Proj() && mem->in(0) == this))
4443 return FAIL; // must not be preceded by other stores
4444 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4445 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4446 !bs->can_initialize_object(st)) {
4447 return FAIL;
4448 }
4449 Node* adr = st->in(MemNode::Address);
4450 intptr_t offset;
4451 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4452 if (alloc == nullptr)
4453 return FAIL; // inscrutable address
4454 if (alloc != allocation())
4455 return FAIL; // wrong allocation! (store needs to float up)
4456 int size_in_bytes = st->memory_size();
4457 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4458 return FAIL; // mismatched access
4459 }
4460 Node* val = st->in(MemNode::ValueIn);
4461
4462 if (!detect_init_independence(val, phase))
4463 return FAIL; // stored value must be 'simple enough'
4464
4465 // The Store can be captured only if nothing after the allocation
4466 // and before the Store is using the memory location that the store
4467 // overwrites.
4468 bool failed = false;
4469 // If is_complete_with_arraycopy() is true the shape of the graph is
4470 // well defined and is safe so no need for extra checks.
4471 if (!is_complete_with_arraycopy()) {
4472 // We are going to look at each use of the memory state following
4473 // the allocation to make sure nothing reads the memory that the
4474 // Store writes.
4475 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4476 int alias_idx = phase->C->get_alias_index(t_adr);
4477 ResourceMark rm;
4478 Unique_Node_List mems;
4479 mems.push(mem);
4480 Node* unique_merge = nullptr;
4481 for (uint next = 0; next < mems.size(); ++next) {
4482 Node *m = mems.at(next);
4483 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4484 Node *n = m->fast_out(j);
4485 if (n->outcnt() == 0) {
4486 continue;
4487 }
4488 if (n == st) {
4489 continue;
4490 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4491 // If the control of this use is different from the control
4492 // of the Store which is right after the InitializeNode then
4493 // this node cannot be between the InitializeNode and the
4494 // Store.
4495 continue;
4496 } else if (n->is_MergeMem()) {
4497 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4498 // We can hit a MergeMemNode (that will likely go away
4499 // later) that is a direct use of the memory state
4500 // following the InitializeNode on the same slice as the
4501 // store node that we'd like to capture. We need to check
4502 // the uses of the MergeMemNode.
4503 mems.push(n);
4504 }
4505 } else if (n->is_Mem()) {
4506 Node* other_adr = n->in(MemNode::Address);
4507 if (other_adr == adr) {
4508 failed = true;
4509 break;
4510 } else {
4511 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4512 if (other_t_adr != nullptr) {
4513 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4514 if (other_alias_idx == alias_idx) {
4515 // A load from the same memory slice as the store right
4516 // after the InitializeNode. We check the control of the
4517 // object/array that is loaded from. If it's the same as
4518 // the store control then we cannot capture the store.
4519 assert(!n->is_Store(), "2 stores to same slice on same control?");
4520 Node* base = other_adr;
4521 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4522 base = base->in(AddPNode::Base);
4523 if (base != nullptr) {
4524 base = base->uncast();
4525 if (base->is_Proj() && base->in(0) == alloc) {
4526 failed = true;
4527 break;
4528 }
4529 }
4530 }
4531 }
4532 }
4533 } else {
4534 failed = true;
4535 break;
4536 }
4537 }
4538 }
4539 }
4540 if (failed) {
4541 if (!can_reshape) {
4542 // We decided we couldn't capture the store during parsing. We
4543 // should try again during the next IGVN once the graph is
4544 // cleaner.
4545 phase->C->record_for_igvn(st);
4546 }
4547 return FAIL;
4548 }
4549
4550 return offset; // success
4551 }
4552
4553 // Find the captured store in(i) which corresponds to the range
4554 // [start..start+size) in the initialized object.
4555 // If there is one, return its index i. If there isn't, return the
4556 // negative of the index where it should be inserted.
4557 // Return 0 if the queried range overlaps an initialization boundary
4558 // or if dead code is encountered.
4559 // If size_in_bytes is zero, do not bother with overlap checks.
4560 int InitializeNode::captured_store_insertion_point(intptr_t start,
4561 int size_in_bytes,
4562 PhaseValues* phase) {
4563 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4564
4565 if (is_complete())
4566 return FAIL; // arraycopy got here first; punt
4567
4568 assert(allocation() != nullptr, "must be present");
4569
4570 // no negatives, no header fields:
4571 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4572
4573 // after a certain size, we bail out on tracking all the stores:
4574 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4575 if (start >= ti_limit) return FAIL;
4576
4577 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4578 if (i >= limit) return -(int)i; // not found; here is where to put it
4579
4580 Node* st = in(i);
4581 intptr_t st_off = get_store_offset(st, phase);
4582 if (st_off < 0) {
4583 if (st != zero_memory()) {
4584 return FAIL; // bail out if there is dead garbage
4585 }
4586 } else if (st_off > start) {
4587 // ...we are done, since stores are ordered
4588 if (st_off < start + size_in_bytes) {
4589 return FAIL; // the next store overlaps
4590 }
4591 return -(int)i; // not found; here is where to put it
4592 } else if (st_off < start) {
4593 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4594 if (size_in_bytes != 0 &&
4595 start < st_off + MAX_STORE &&
4596 start < st_off + st->as_Store()->memory_size()) {
4597 return FAIL; // the previous store overlaps
4598 }
4599 } else {
4600 if (size_in_bytes != 0 &&
4601 st->as_Store()->memory_size() != size_in_bytes) {
4602 return FAIL; // mismatched store size
4603 }
4604 return i;
4605 }
4606
4607 ++i;
4608 }
4609 }
4610
4611 // Look for a captured store which initializes at the offset 'start'
4612 // with the given size. If there is no such store, and no other
4613 // initialization interferes, then return zero_memory (the memory
4614 // projection of the AllocateNode).
4615 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4616 PhaseValues* phase) {
4617 assert(stores_are_sane(phase), "");
4618 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4619 if (i == 0) {
4620 return nullptr; // something is dead
4621 } else if (i < 0) {
4622 return zero_memory(); // just primordial zero bits here
4623 } else {
4624 Node* st = in(i); // here is the store at this position
4625 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4626 return st;
4627 }
4628 }
4629
4630 // Create, as a raw pointer, an address within my new object at 'offset'.
4631 Node* InitializeNode::make_raw_address(intptr_t offset,
4632 PhaseGVN* phase) {
4633 Node* addr = in(RawAddress);
4634 if (offset != 0) {
4635 Compile* C = phase->C;
4636 addr = phase->transform( new AddPNode(C->top(), addr,
4637 phase->MakeConX(offset)) );
4638 }
4639 return addr;
4640 }
4641
4642 // Clone the given store, converting it into a raw store
4643 // initializing a field or element of my new object.
4644 // Caller is responsible for retiring the original store,
4645 // with subsume_node or the like.
4646 //
4647 // From the example above InitializeNode::InitializeNode,
4648 // here are the old stores to be captured:
4649 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4650 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4651 //
4652 // Here is the changed code; note the extra edges on init:
4653 // alloc = (Allocate ...)
4654 // rawoop = alloc.RawAddress
4655 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4656 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4657 // init = (Initialize alloc.Control alloc.Memory rawoop
4658 // rawstore1 rawstore2)
4659 //
4660 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4661 PhaseGVN* phase, bool can_reshape) {
4662 assert(stores_are_sane(phase), "");
4663
4664 if (start < 0) return nullptr;
4665 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4666
4667 Compile* C = phase->C;
4668 int size_in_bytes = st->memory_size();
4669 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4670 if (i == 0) return nullptr; // bail out
4671 Node* prev_mem = nullptr; // raw memory for the captured store
4672 if (i > 0) {
4673 prev_mem = in(i); // there is a pre-existing store under this one
4674 set_req(i, C->top()); // temporarily disconnect it
4675 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4676 } else {
4677 i = -i; // no pre-existing store
4678 prev_mem = zero_memory(); // a slice of the newly allocated object
4679 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4680 set_req(--i, C->top()); // reuse this edge; it has been folded away
4681 else
4682 ins_req(i, C->top()); // build a new edge
4683 }
4684 Node* new_st = st->clone();
4685 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4686 new_st->set_req(MemNode::Control, in(Control));
4687 new_st->set_req(MemNode::Memory, prev_mem);
4688 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4689 bs->eliminate_gc_barrier_data(new_st);
4690 new_st = phase->transform(new_st);
4691
4692 // At this point, new_st might have swallowed a pre-existing store
4693 // at the same offset, or perhaps new_st might have disappeared,
4694 // if it redundantly stored the same value (or zero to fresh memory).
4695
4696 // In any case, wire it in:
4697 PhaseIterGVN* igvn = phase->is_IterGVN();
4698 if (igvn) {
4699 igvn->rehash_node_delayed(this);
4700 }
4701 set_req(i, new_st);
4702
4703 // The caller may now kill the old guy.
4704 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4705 assert(check_st == new_st || check_st == nullptr, "must be findable");
4706 assert(!is_complete(), "");
4707 return new_st;
4708 }
4709
4710 static bool store_constant(jlong* tiles, int num_tiles,
4711 intptr_t st_off, int st_size,
4712 jlong con) {
4713 if ((st_off & (st_size-1)) != 0)
4714 return false; // strange store offset (assume size==2**N)
4715 address addr = (address)tiles + st_off;
4716 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4717 switch (st_size) {
4718 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4719 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4720 case sizeof(jint): *(jint*) addr = (jint) con; break;
4721 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4722 default: return false; // strange store size (detect size!=2**N here)
4723 }
4724 return true; // return success to caller
4725 }
4726
4727 // Coalesce subword constants into int constants and possibly
4728 // into long constants. The goal, if the CPU permits,
4729 // is to initialize the object with a small number of 64-bit tiles.
4730 // Also, convert floating-point constants to bit patterns.
4731 // Non-constants are not relevant to this pass.
4732 //
4733 // In terms of the running example on InitializeNode::InitializeNode
4734 // and InitializeNode::capture_store, here is the transformation
4735 // of rawstore1 and rawstore2 into rawstore12:
4736 // alloc = (Allocate ...)
4737 // rawoop = alloc.RawAddress
4738 // tile12 = 0x00010002
4739 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4740 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4741 //
4742 void
4743 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4744 Node* size_in_bytes,
4745 PhaseGVN* phase) {
4746 Compile* C = phase->C;
4747
4748 assert(stores_are_sane(phase), "");
4749 // Note: After this pass, they are not completely sane,
4750 // since there may be some overlaps.
4751
4752 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4753
4754 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4755 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4756 size_limit = MIN2(size_limit, ti_limit);
4757 size_limit = align_up(size_limit, BytesPerLong);
4758 int num_tiles = size_limit / BytesPerLong;
4759
4760 // allocate space for the tile map:
4761 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4762 jlong tiles_buf[small_len];
4763 Node* nodes_buf[small_len];
4764 jlong inits_buf[small_len];
4765 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4766 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4767 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4768 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4769 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4770 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4771 // tiles: exact bitwise model of all primitive constants
4772 // nodes: last constant-storing node subsumed into the tiles model
4773 // inits: which bytes (in each tile) are touched by any initializations
4774
4775 //// Pass A: Fill in the tile model with any relevant stores.
4776
4777 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4778 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4779 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4780 Node* zmem = zero_memory(); // initially zero memory state
4781 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4782 Node* st = in(i);
4783 intptr_t st_off = get_store_offset(st, phase);
4784
4785 // Figure out the store's offset and constant value:
4786 if (st_off < header_size) continue; //skip (ignore header)
4787 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4788 int st_size = st->as_Store()->memory_size();
4789 if (st_off + st_size > size_limit) break;
4790
4791 // Record which bytes are touched, whether by constant or not.
4792 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4793 continue; // skip (strange store size)
4794
4795 const Type* val = phase->type(st->in(MemNode::ValueIn));
4796 if (!val->singleton()) continue; //skip (non-con store)
4797 BasicType type = val->basic_type();
4798
4799 jlong con = 0;
4800 switch (type) {
4801 case T_INT: con = val->is_int()->get_con(); break;
4802 case T_LONG: con = val->is_long()->get_con(); break;
4803 case T_FLOAT: con = jint_cast(val->getf()); break;
4804 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4805 default: continue; //skip (odd store type)
4806 }
4807
4808 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4809 st->Opcode() == Op_StoreL) {
4810 continue; // This StoreL is already optimal.
4811 }
4812
4813 // Store down the constant.
4814 store_constant(tiles, num_tiles, st_off, st_size, con);
4815
4816 intptr_t j = st_off >> LogBytesPerLong;
4817
4818 if (type == T_INT && st_size == BytesPerInt
4819 && (st_off & BytesPerInt) == BytesPerInt) {
4820 jlong lcon = tiles[j];
4821 if (!Matcher::isSimpleConstant64(lcon) &&
4822 st->Opcode() == Op_StoreI) {
4823 // This StoreI is already optimal by itself.
4824 jint* intcon = (jint*) &tiles[j];
4825 intcon[1] = 0; // undo the store_constant()
4826
4827 // If the previous store is also optimal by itself, back up and
4828 // undo the action of the previous loop iteration... if we can.
4829 // But if we can't, just let the previous half take care of itself.
4830 st = nodes[j];
4831 st_off -= BytesPerInt;
4832 con = intcon[0];
4833 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
4834 assert(st_off >= header_size, "still ignoring header");
4835 assert(get_store_offset(st, phase) == st_off, "must be");
4836 assert(in(i-1) == zmem, "must be");
4837 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4838 assert(con == tcon->is_int()->get_con(), "must be");
4839 // Undo the effects of the previous loop trip, which swallowed st:
4840 intcon[0] = 0; // undo store_constant()
4841 set_req(i-1, st); // undo set_req(i, zmem)
4842 nodes[j] = nullptr; // undo nodes[j] = st
4843 --old_subword; // undo ++old_subword
4844 }
4845 continue; // This StoreI is already optimal.
4846 }
4847 }
4848
4849 // This store is not needed.
4850 set_req(i, zmem);
4851 nodes[j] = st; // record for the moment
4852 if (st_size < BytesPerLong) // something has changed
4853 ++old_subword; // includes int/float, but who's counting...
4854 else ++old_long;
4855 }
4856
4857 if ((old_subword + old_long) == 0)
4858 return; // nothing more to do
4859
4860 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4861 // Be sure to insert them before overlapping non-constant stores.
4862 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4863 for (int j = 0; j < num_tiles; j++) {
4864 jlong con = tiles[j];
4865 jlong init = inits[j];
4866 if (con == 0) continue;
4867 jint con0, con1; // split the constant, address-wise
4868 jint init0, init1; // split the init map, address-wise
4869 { union { jlong con; jint intcon[2]; } u;
4870 u.con = con;
4871 con0 = u.intcon[0];
4872 con1 = u.intcon[1];
4873 u.con = init;
4874 init0 = u.intcon[0];
4875 init1 = u.intcon[1];
4876 }
4877
4878 Node* old = nodes[j];
4879 assert(old != nullptr, "need the prior store");
4880 intptr_t offset = (j * BytesPerLong);
4881
4882 bool split = !Matcher::isSimpleConstant64(con);
4883
4884 if (offset < header_size) {
4885 assert(offset + BytesPerInt >= header_size, "second int counts");
4886 assert(*(jint*)&tiles[j] == 0, "junk in header");
4887 split = true; // only the second word counts
4888 // Example: int a[] = { 42 ... }
4889 } else if (con0 == 0 && init0 == -1) {
4890 split = true; // first word is covered by full inits
4891 // Example: int a[] = { ... foo(), 42 ... }
4892 } else if (con1 == 0 && init1 == -1) {
4893 split = true; // second word is covered by full inits
4894 // Example: int a[] = { ... 42, foo() ... }
4895 }
4896
4897 // Here's a case where init0 is neither 0 nor -1:
4898 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4899 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4900 // In this case the tile is not split; it is (jlong)42.
4901 // The big tile is stored down, and then the foo() value is inserted.
4902 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4903
4904 Node* ctl = old->in(MemNode::Control);
4905 Node* adr = make_raw_address(offset, phase);
4906 const TypePtr* atp = TypeRawPtr::BOTTOM;
4907
4908 // One or two coalesced stores to plop down.
4909 Node* st[2];
4910 intptr_t off[2];
4911 int nst = 0;
4912 if (!split) {
4913 ++new_long;
4914 off[nst] = offset;
4915 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4916 phase->longcon(con), T_LONG, MemNode::unordered);
4917 } else {
4918 // Omit either if it is a zero.
4919 if (con0 != 0) {
4920 ++new_int;
4921 off[nst] = offset;
4922 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4923 phase->intcon(con0), T_INT, MemNode::unordered);
4924 }
4925 if (con1 != 0) {
4926 ++new_int;
4927 offset += BytesPerInt;
4928 adr = make_raw_address(offset, phase);
4929 off[nst] = offset;
4930 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4931 phase->intcon(con1), T_INT, MemNode::unordered);
4932 }
4933 }
4934
4935 // Insert second store first, then the first before the second.
4936 // Insert each one just before any overlapping non-constant stores.
4937 while (nst > 0) {
4938 Node* st1 = st[--nst];
4939 C->copy_node_notes_to(st1, old);
4940 st1 = phase->transform(st1);
4941 offset = off[nst];
4942 assert(offset >= header_size, "do not smash header");
4943 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4944 guarantee(ins_idx != 0, "must re-insert constant store");
4945 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4946 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4947 set_req(--ins_idx, st1);
4948 else
4949 ins_req(ins_idx, st1);
4950 }
4951 }
4952
4953 if (PrintCompilation && WizardMode)
4954 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4955 old_subword, old_long, new_int, new_long);
4956 if (C->log() != nullptr)
4957 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4958 old_subword, old_long, new_int, new_long);
4959
4960 // Clean up any remaining occurrences of zmem:
4961 remove_extra_zeroes();
4962 }
4963
4964 // Explore forward from in(start) to find the first fully initialized
4965 // word, and return its offset. Skip groups of subword stores which
4966 // together initialize full words. If in(start) is itself part of a
4967 // fully initialized word, return the offset of in(start). If there
4968 // are no following full-word stores, or if something is fishy, return
4969 // a negative value.
4970 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4971 int int_map = 0;
4972 intptr_t int_map_off = 0;
4973 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4974
4975 for (uint i = start, limit = req(); i < limit; i++) {
4976 Node* st = in(i);
4977
4978 intptr_t st_off = get_store_offset(st, phase);
4979 if (st_off < 0) break; // return conservative answer
4980
4981 int st_size = st->as_Store()->memory_size();
4982 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4983 return st_off; // we found a complete word init
4984 }
4985
4986 // update the map:
4987
4988 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4989 if (this_int_off != int_map_off) {
4990 // reset the map:
4991 int_map = 0;
4992 int_map_off = this_int_off;
4993 }
4994
4995 int subword_off = st_off - this_int_off;
4996 int_map |= right_n_bits(st_size) << subword_off;
4997 if ((int_map & FULL_MAP) == FULL_MAP) {
4998 return this_int_off; // we found a complete word init
4999 }
5000
5001 // Did this store hit or cross the word boundary?
5002 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5003 if (next_int_off == this_int_off + BytesPerInt) {
5004 // We passed the current int, without fully initializing it.
5005 int_map_off = next_int_off;
5006 int_map >>= BytesPerInt;
5007 } else if (next_int_off > this_int_off + BytesPerInt) {
5008 // We passed the current and next int.
5009 return this_int_off + BytesPerInt;
5010 }
5011 }
5012
5013 return -1;
5014 }
5015
5016
5017 // Called when the associated AllocateNode is expanded into CFG.
5018 // At this point, we may perform additional optimizations.
5019 // Linearize the stores by ascending offset, to make memory
5020 // activity as coherent as possible.
5021 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5022 intptr_t header_size,
5023 Node* size_in_bytes,
5024 PhaseIterGVN* phase) {
5025 assert(!is_complete(), "not already complete");
5026 assert(stores_are_sane(phase), "");
5027 assert(allocation() != nullptr, "must be present");
5028
5029 remove_extra_zeroes();
5030
5031 if (ReduceFieldZeroing || ReduceBulkZeroing)
5032 // reduce instruction count for common initialization patterns
5033 coalesce_subword_stores(header_size, size_in_bytes, phase);
5034
5035 Node* zmem = zero_memory(); // initially zero memory state
5036 Node* inits = zmem; // accumulating a linearized chain of inits
5037 #ifdef ASSERT
5038 intptr_t first_offset = allocation()->minimum_header_size();
5039 intptr_t last_init_off = first_offset; // previous init offset
5040 intptr_t last_init_end = first_offset; // previous init offset+size
5041 intptr_t last_tile_end = first_offset; // previous tile offset+size
5042 #endif
5043 intptr_t zeroes_done = header_size;
5044
5045 bool do_zeroing = true; // we might give up if inits are very sparse
5046 int big_init_gaps = 0; // how many large gaps have we seen?
5047
5048 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5049 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5050
5051 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5052 Node* st = in(i);
5053 intptr_t st_off = get_store_offset(st, phase);
5054 if (st_off < 0)
5055 break; // unknown junk in the inits
5056 if (st->in(MemNode::Memory) != zmem)
5057 break; // complicated store chains somehow in list
5058
5059 int st_size = st->as_Store()->memory_size();
5060 intptr_t next_init_off = st_off + st_size;
5061
5062 if (do_zeroing && zeroes_done < next_init_off) {
5063 // See if this store needs a zero before it or under it.
5064 intptr_t zeroes_needed = st_off;
5065
5066 if (st_size < BytesPerInt) {
5067 // Look for subword stores which only partially initialize words.
5068 // If we find some, we must lay down some word-level zeroes first,
5069 // underneath the subword stores.
5070 //
5071 // Examples:
5072 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5073 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5074 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5075 //
5076 // Note: coalesce_subword_stores may have already done this,
5077 // if it was prompted by constant non-zero subword initializers.
5078 // But this case can still arise with non-constant stores.
5079
5080 intptr_t next_full_store = find_next_fullword_store(i, phase);
5081
5082 // In the examples above:
5083 // in(i) p q r s x y z
5084 // st_off 12 13 14 15 12 13 14
5085 // st_size 1 1 1 1 1 1 1
5086 // next_full_s. 12 16 16 16 16 16 16
5087 // z's_done 12 16 16 16 12 16 12
5088 // z's_needed 12 16 16 16 16 16 16
5089 // zsize 0 0 0 0 4 0 4
5090 if (next_full_store < 0) {
5091 // Conservative tack: Zero to end of current word.
5092 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5093 } else {
5094 // Zero to beginning of next fully initialized word.
5095 // Or, don't zero at all, if we are already in that word.
5096 assert(next_full_store >= zeroes_needed, "must go forward");
5097 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5098 zeroes_needed = next_full_store;
5099 }
5100 }
5101
5102 if (zeroes_needed > zeroes_done) {
5103 intptr_t zsize = zeroes_needed - zeroes_done;
5104 // Do some incremental zeroing on rawmem, in parallel with inits.
5105 zeroes_done = align_down(zeroes_done, BytesPerInt);
5106 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5107 zeroes_done, zeroes_needed,
5108 phase);
5109 zeroes_done = zeroes_needed;
5110 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5111 do_zeroing = false; // leave the hole, next time
5112 }
5113 }
5114
5115 // Collect the store and move on:
5116 phase->replace_input_of(st, MemNode::Memory, inits);
5117 inits = st; // put it on the linearized chain
5118 set_req(i, zmem); // unhook from previous position
5119
5120 if (zeroes_done == st_off)
5121 zeroes_done = next_init_off;
5122
5123 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5124
5125 #ifdef ASSERT
5126 // Various order invariants. Weaker than stores_are_sane because
5127 // a large constant tile can be filled in by smaller non-constant stores.
5128 assert(st_off >= last_init_off, "inits do not reverse");
5129 last_init_off = st_off;
5130 const Type* val = nullptr;
5131 if (st_size >= BytesPerInt &&
5132 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5133 (int)val->basic_type() < (int)T_OBJECT) {
5134 assert(st_off >= last_tile_end, "tiles do not overlap");
5135 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5136 last_tile_end = MAX2(last_tile_end, next_init_off);
5137 } else {
5138 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5139 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5140 assert(st_off >= last_init_end, "inits do not overlap");
5141 last_init_end = next_init_off; // it's a non-tile
5142 }
5143 #endif //ASSERT
5144 }
5145
5146 remove_extra_zeroes(); // clear out all the zmems left over
5147 add_req(inits);
5148
5149 if (!(UseTLAB && ZeroTLAB)) {
5150 // If anything remains to be zeroed, zero it all now.
5151 zeroes_done = align_down(zeroes_done, BytesPerInt);
5152 // if it is the last unused 4 bytes of an instance, forget about it
5153 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5154 if (zeroes_done + BytesPerLong >= size_limit) {
5155 AllocateNode* alloc = allocation();
5156 assert(alloc != nullptr, "must be present");
5157 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5158 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5159 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5160 if (zeroes_done == k->layout_helper())
5161 zeroes_done = size_limit;
5162 }
5163 }
5164 if (zeroes_done < size_limit) {
5165 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5166 zeroes_done, size_in_bytes, phase);
5167 }
5168 }
5169
5170 set_complete(phase);
5171 return rawmem;
5172 }
5173
5174
5175 #ifdef ASSERT
5176 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5177 if (is_complete())
5178 return true; // stores could be anything at this point
5179 assert(allocation() != nullptr, "must be present");
5180 intptr_t last_off = allocation()->minimum_header_size();
5181 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5182 Node* st = in(i);
5183 intptr_t st_off = get_store_offset(st, phase);
5184 if (st_off < 0) continue; // ignore dead garbage
5185 if (last_off > st_off) {
5186 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
5187 this->dump(2);
5188 assert(false, "ascending store offsets");
5189 return false;
5190 }
5191 last_off = st_off + st->as_Store()->memory_size();
5192 }
5193 return true;
5194 }
5195 #endif //ASSERT
5196
5197
5198
5199
5200 //============================MergeMemNode=====================================
5201 //
5202 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5203 // contributing store or call operations. Each contributor provides the memory
5204 // state for a particular "alias type" (see Compile::alias_type). For example,
5205 // if a MergeMem has an input X for alias category #6, then any memory reference
5206 // to alias category #6 may use X as its memory state input, as an exact equivalent
5207 // to using the MergeMem as a whole.
5208 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5209 //
5210 // (Here, the <N> notation gives the index of the relevant adr_type.)
5211 //
5212 // In one special case (and more cases in the future), alias categories overlap.
5213 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5214 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5215 // it is exactly equivalent to that state W:
5216 // MergeMem(<Bot>: W) <==> W
5217 //
5218 // Usually, the merge has more than one input. In that case, where inputs
5219 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5220 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5221 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5222 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5223 //
5224 // A merge can take a "wide" memory state as one of its narrow inputs.
5225 // This simply means that the merge observes out only the relevant parts of
5226 // the wide input. That is, wide memory states arriving at narrow merge inputs
5227 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5228 //
5229 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5230 // and that memory slices "leak through":
5231 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5232 //
5233 // But, in such a cascade, repeated memory slices can "block the leak":
5234 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5235 //
5236 // In the last example, Y is not part of the combined memory state of the
5237 // outermost MergeMem. The system must, of course, prevent unschedulable
5238 // memory states from arising, so you can be sure that the state Y is somehow
5239 // a precursor to state Y'.
5240 //
5241 //
5242 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5243 // of each MergeMemNode array are exactly the numerical alias indexes, including
5244 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5245 // Compile::alias_type (and kin) produce and manage these indexes.
5246 //
5247 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5248 // (Note that this provides quick access to the top node inside MergeMem methods,
5249 // without the need to reach out via TLS to Compile::current.)
5250 //
5251 // As a consequence of what was just described, a MergeMem that represents a full
5252 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5253 // containing all alias categories.
5254 //
5255 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5256 //
5257 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5258 // a memory state for the alias type <N>, or else the top node, meaning that
5259 // there is no particular input for that alias type. Note that the length of
5260 // a MergeMem is variable, and may be extended at any time to accommodate new
5261 // memory states at larger alias indexes. When merges grow, they are of course
5262 // filled with "top" in the unused in() positions.
5263 //
5264 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5265 // (Top was chosen because it works smoothly with passes like GCM.)
5266 //
5267 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5268 // the type of random VM bits like TLS references.) Since it is always the
5269 // first non-Bot memory slice, some low-level loops use it to initialize an
5270 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5271 //
5272 //
5273 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5274 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5275 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5276 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5277 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5278 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5279 //
5280 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5281 // really that different from the other memory inputs. An abbreviation called
5282 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5283 //
5284 //
5285 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5286 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5287 // that "emerges though" the base memory will be marked as excluding the alias types
5288 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5289 //
5290 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5291 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5292 //
5293 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5294 // (It is currently unimplemented.) As you can see, the resulting merge is
5295 // actually a disjoint union of memory states, rather than an overlay.
5296 //
5297
5298 //------------------------------MergeMemNode-----------------------------------
5299 Node* MergeMemNode::make_empty_memory() {
5300 Node* empty_memory = (Node*) Compile::current()->top();
5301 assert(empty_memory->is_top(), "correct sentinel identity");
5302 return empty_memory;
5303 }
5304
5305 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5306 init_class_id(Class_MergeMem);
5307 // all inputs are nullified in Node::Node(int)
5308 // set_input(0, nullptr); // no control input
5309
5310 // Initialize the edges uniformly to top, for starters.
5311 Node* empty_mem = make_empty_memory();
5312 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5313 init_req(i,empty_mem);
5314 }
5315 assert(empty_memory() == empty_mem, "");
5316
5317 if( new_base != nullptr && new_base->is_MergeMem() ) {
5318 MergeMemNode* mdef = new_base->as_MergeMem();
5319 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5320 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5321 mms.set_memory(mms.memory2());
5322 }
5323 assert(base_memory() == mdef->base_memory(), "");
5324 } else {
5325 set_base_memory(new_base);
5326 }
5327 }
5328
5329 // Make a new, untransformed MergeMem with the same base as 'mem'.
5330 // If mem is itself a MergeMem, populate the result with the same edges.
5331 MergeMemNode* MergeMemNode::make(Node* mem) {
5332 return new MergeMemNode(mem);
5333 }
5334
5335 //------------------------------cmp--------------------------------------------
5336 uint MergeMemNode::hash() const { return NO_HASH; }
5337 bool MergeMemNode::cmp( const Node &n ) const {
5338 return (&n == this); // Always fail except on self
5339 }
5340
5341 //------------------------------Identity---------------------------------------
5342 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5343 // Identity if this merge point does not record any interesting memory
5344 // disambiguations.
5345 Node* base_mem = base_memory();
5346 Node* empty_mem = empty_memory();
5347 if (base_mem != empty_mem) { // Memory path is not dead?
5348 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5349 Node* mem = in(i);
5350 if (mem != empty_mem && mem != base_mem) {
5351 return this; // Many memory splits; no change
5352 }
5353 }
5354 }
5355 return base_mem; // No memory splits; ID on the one true input
5356 }
5357
5358 //------------------------------Ideal------------------------------------------
5359 // This method is invoked recursively on chains of MergeMem nodes
5360 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5361 // Remove chain'd MergeMems
5362 //
5363 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5364 // relative to the "in(Bot)". Since we are patching both at the same time,
5365 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5366 // but rewrite each "in(i)" relative to the new "in(Bot)".
5367 Node *progress = nullptr;
5368
5369
5370 Node* old_base = base_memory();
5371 Node* empty_mem = empty_memory();
5372 if (old_base == empty_mem)
5373 return nullptr; // Dead memory path.
5374
5375 MergeMemNode* old_mbase;
5376 if (old_base != nullptr && old_base->is_MergeMem())
5377 old_mbase = old_base->as_MergeMem();
5378 else
5379 old_mbase = nullptr;
5380 Node* new_base = old_base;
5381
5382 // simplify stacked MergeMems in base memory
5383 if (old_mbase) new_base = old_mbase->base_memory();
5384
5385 // the base memory might contribute new slices beyond my req()
5386 if (old_mbase) grow_to_match(old_mbase);
5387
5388 // Note: We do not call verify_sparse on entry, because inputs
5389 // can normalize to the base_memory via subsume_node or similar
5390 // mechanisms. This method repairs that damage.
5391
5392 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5393
5394 // Look at each slice.
5395 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5396 Node* old_in = in(i);
5397 // calculate the old memory value
5398 Node* old_mem = old_in;
5399 if (old_mem == empty_mem) old_mem = old_base;
5400 assert(old_mem == memory_at(i), "");
5401
5402 // maybe update (reslice) the old memory value
5403
5404 // simplify stacked MergeMems
5405 Node* new_mem = old_mem;
5406 MergeMemNode* old_mmem;
5407 if (old_mem != nullptr && old_mem->is_MergeMem())
5408 old_mmem = old_mem->as_MergeMem();
5409 else
5410 old_mmem = nullptr;
5411 if (old_mmem == this) {
5412 // This can happen if loops break up and safepoints disappear.
5413 // A merge of BotPtr (default) with a RawPtr memory derived from a
5414 // safepoint can be rewritten to a merge of the same BotPtr with
5415 // the BotPtr phi coming into the loop. If that phi disappears
5416 // also, we can end up with a self-loop of the mergemem.
5417 // In general, if loops degenerate and memory effects disappear,
5418 // a mergemem can be left looking at itself. This simply means
5419 // that the mergemem's default should be used, since there is
5420 // no longer any apparent effect on this slice.
5421 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5422 // from start. Update the input to TOP.
5423 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5424 }
5425 else if (old_mmem != nullptr) {
5426 new_mem = old_mmem->memory_at(i);
5427 }
5428 // else preceding memory was not a MergeMem
5429
5430 // maybe store down a new value
5431 Node* new_in = new_mem;
5432 if (new_in == new_base) new_in = empty_mem;
5433
5434 if (new_in != old_in) {
5435 // Warning: Do not combine this "if" with the previous "if"
5436 // A memory slice might have be be rewritten even if it is semantically
5437 // unchanged, if the base_memory value has changed.
5438 set_req_X(i, new_in, phase);
5439 progress = this; // Report progress
5440 }
5441 }
5442
5443 if (new_base != old_base) {
5444 set_req_X(Compile::AliasIdxBot, new_base, phase);
5445 // Don't use set_base_memory(new_base), because we need to update du.
5446 assert(base_memory() == new_base, "");
5447 progress = this;
5448 }
5449
5450 if( base_memory() == this ) {
5451 // a self cycle indicates this memory path is dead
5452 set_req(Compile::AliasIdxBot, empty_mem);
5453 }
5454
5455 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5456 // Recursion must occur after the self cycle check above
5457 if( base_memory()->is_MergeMem() ) {
5458 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5459 Node *m = phase->transform(new_mbase); // Rollup any cycles
5460 if( m != nullptr &&
5461 (m->is_top() ||
5462 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5463 // propagate rollup of dead cycle to self
5464 set_req(Compile::AliasIdxBot, empty_mem);
5465 }
5466 }
5467
5468 if( base_memory() == empty_mem ) {
5469 progress = this;
5470 // Cut inputs during Parse phase only.
5471 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5472 if( !can_reshape ) {
5473 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5474 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5475 }
5476 }
5477 }
5478
5479 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5480 // Check if PhiNode::Ideal's "Split phis through memory merges"
5481 // transform should be attempted. Look for this->phi->this cycle.
5482 uint merge_width = req();
5483 if (merge_width > Compile::AliasIdxRaw) {
5484 PhiNode* phi = base_memory()->as_Phi();
5485 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5486 if (phi->in(i) == this) {
5487 phase->is_IterGVN()->_worklist.push(phi);
5488 break;
5489 }
5490 }
5491 }
5492 }
5493
5494 assert(progress || verify_sparse(), "please, no dups of base");
5495 return progress;
5496 }
5497
5498 //-------------------------set_base_memory-------------------------------------
5499 void MergeMemNode::set_base_memory(Node *new_base) {
5500 Node* empty_mem = empty_memory();
5501 set_req(Compile::AliasIdxBot, new_base);
5502 assert(memory_at(req()) == new_base, "must set default memory");
5503 // Clear out other occurrences of new_base:
5504 if (new_base != empty_mem) {
5505 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5506 if (in(i) == new_base) set_req(i, empty_mem);
5507 }
5508 }
5509 }
5510
5511 //------------------------------out_RegMask------------------------------------
5512 const RegMask &MergeMemNode::out_RegMask() const {
5513 return RegMask::Empty;
5514 }
5515
5516 //------------------------------dump_spec--------------------------------------
5517 #ifndef PRODUCT
5518 void MergeMemNode::dump_spec(outputStream *st) const {
5519 st->print(" {");
5520 Node* base_mem = base_memory();
5521 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5522 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5523 if (mem == base_mem) { st->print(" -"); continue; }
5524 st->print( " N%d:", mem->_idx );
5525 Compile::current()->get_adr_type(i)->dump_on(st);
5526 }
5527 st->print(" }");
5528 }
5529 #endif // !PRODUCT
5530
5531
5532 #ifdef ASSERT
5533 static bool might_be_same(Node* a, Node* b) {
5534 if (a == b) return true;
5535 if (!(a->is_Phi() || b->is_Phi())) return false;
5536 // phis shift around during optimization
5537 return true; // pretty stupid...
5538 }
5539
5540 // verify a narrow slice (either incoming or outgoing)
5541 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5542 if (!VerifyAliases) return; // don't bother to verify unless requested
5543 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5544 if (Node::in_dump()) return; // muzzle asserts when printing
5545 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5546 assert(n != nullptr, "");
5547 // Elide intervening MergeMem's
5548 while (n->is_MergeMem()) {
5549 n = n->as_MergeMem()->memory_at(alias_idx);
5550 }
5551 Compile* C = Compile::current();
5552 const TypePtr* n_adr_type = n->adr_type();
5553 if (n == m->empty_memory()) {
5554 // Implicit copy of base_memory()
5555 } else if (n_adr_type != TypePtr::BOTTOM) {
5556 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5557 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5558 } else {
5559 // A few places like make_runtime_call "know" that VM calls are narrow,
5560 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5561 bool expected_wide_mem = false;
5562 if (n == m->base_memory()) {
5563 expected_wide_mem = true;
5564 } else if (alias_idx == Compile::AliasIdxRaw ||
5565 n == m->memory_at(Compile::AliasIdxRaw)) {
5566 expected_wide_mem = true;
5567 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5568 // memory can "leak through" calls on channels that
5569 // are write-once. Allow this also.
5570 expected_wide_mem = true;
5571 }
5572 assert(expected_wide_mem, "expected narrow slice replacement");
5573 }
5574 }
5575 #else // !ASSERT
5576 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5577 #endif
5578
5579
5580 //-----------------------------memory_at---------------------------------------
5581 Node* MergeMemNode::memory_at(uint alias_idx) const {
5582 assert(alias_idx >= Compile::AliasIdxRaw ||
5583 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5584 "must avoid base_memory and AliasIdxTop");
5585
5586 // Otherwise, it is a narrow slice.
5587 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5588 if (is_empty_memory(n)) {
5589 // the array is sparse; empty slots are the "top" node
5590 n = base_memory();
5591 assert(Node::in_dump()
5592 || n == nullptr || n->bottom_type() == Type::TOP
5593 || n->adr_type() == nullptr // address is TOP
5594 || n->adr_type() == TypePtr::BOTTOM
5595 || n->adr_type() == TypeRawPtr::BOTTOM
5596 || !Compile::current()->do_aliasing(),
5597 "must be a wide memory");
5598 // do_aliasing == false if we are organizing the memory states manually.
5599 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5600 } else {
5601 // make sure the stored slice is sane
5602 #ifdef ASSERT
5603 if (VMError::is_error_reported() || Node::in_dump()) {
5604 } else if (might_be_same(n, base_memory())) {
5605 // Give it a pass: It is a mostly harmless repetition of the base.
5606 // This can arise normally from node subsumption during optimization.
5607 } else {
5608 verify_memory_slice(this, alias_idx, n);
5609 }
5610 #endif
5611 }
5612 return n;
5613 }
5614
5615 //---------------------------set_memory_at-------------------------------------
5616 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5617 verify_memory_slice(this, alias_idx, n);
5618 Node* empty_mem = empty_memory();
5619 if (n == base_memory()) n = empty_mem; // collapse default
5620 uint need_req = alias_idx+1;
5621 if (req() < need_req) {
5622 if (n == empty_mem) return; // already the default, so do not grow me
5623 // grow the sparse array
5624 do {
5625 add_req(empty_mem);
5626 } while (req() < need_req);
5627 }
5628 set_req( alias_idx, n );
5629 }
5630
5631
5632
5633 //--------------------------iteration_setup------------------------------------
5634 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5635 if (other != nullptr) {
5636 grow_to_match(other);
5637 // invariant: the finite support of mm2 is within mm->req()
5638 #ifdef ASSERT
5639 for (uint i = req(); i < other->req(); i++) {
5640 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5641 }
5642 #endif
5643 }
5644 // Replace spurious copies of base_memory by top.
5645 Node* base_mem = base_memory();
5646 if (base_mem != nullptr && !base_mem->is_top()) {
5647 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5648 if (in(i) == base_mem)
5649 set_req(i, empty_memory());
5650 }
5651 }
5652 }
5653
5654 //---------------------------grow_to_match-------------------------------------
5655 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5656 Node* empty_mem = empty_memory();
5657 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5658 // look for the finite support of the other memory
5659 for (uint i = other->req(); --i >= req(); ) {
5660 if (other->in(i) != empty_mem) {
5661 uint new_len = i+1;
5662 while (req() < new_len) add_req(empty_mem);
5663 break;
5664 }
5665 }
5666 }
5667
5668 //---------------------------verify_sparse-------------------------------------
5669 #ifndef PRODUCT
5670 bool MergeMemNode::verify_sparse() const {
5671 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5672 Node* base_mem = base_memory();
5673 // The following can happen in degenerate cases, since empty==top.
5674 if (is_empty_memory(base_mem)) return true;
5675 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5676 assert(in(i) != nullptr, "sane slice");
5677 if (in(i) == base_mem) return false; // should have been the sentinel value!
5678 }
5679 return true;
5680 }
5681
5682 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5683 Node* n;
5684 n = mm->in(idx);
5685 if (mem == n) return true; // might be empty_memory()
5686 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5687 if (mem == n) return true;
5688 return false;
5689 }
5690 #endif // !PRODUCT